Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * pg_buffercache_pages.c
4 : * display some contents of the buffer cache
5 : *
6 : * contrib/pg_buffercache/pg_buffercache_pages.c
7 : *-------------------------------------------------------------------------
8 : */
9 : #include "postgres.h"
10 :
11 : #include "access/htup_details.h"
12 : #include "access/relation.h"
13 : #include "catalog/pg_type.h"
14 : #include "funcapi.h"
15 : #include "port/pg_numa.h"
16 : #include "storage/buf_internals.h"
17 : #include "storage/bufmgr.h"
18 : #include "utils/rel.h"
19 : #include "utils/tuplestore.h"
20 :
21 :
22 : #define NUM_BUFFERCACHE_PAGES_MIN_ELEM 8
23 : #define NUM_BUFFERCACHE_PAGES_ELEM 9
24 : #define NUM_BUFFERCACHE_SUMMARY_ELEM 5
25 : #define NUM_BUFFERCACHE_USAGE_COUNTS_ELEM 4
26 : #define NUM_BUFFERCACHE_EVICT_ELEM 2
27 : #define NUM_BUFFERCACHE_EVICT_RELATION_ELEM 3
28 : #define NUM_BUFFERCACHE_EVICT_ALL_ELEM 3
29 : #define NUM_BUFFERCACHE_MARK_DIRTY_ELEM 2
30 : #define NUM_BUFFERCACHE_MARK_DIRTY_RELATION_ELEM 3
31 : #define NUM_BUFFERCACHE_MARK_DIRTY_ALL_ELEM 3
32 :
33 : #define NUM_BUFFERCACHE_OS_PAGES_ELEM 3
34 :
35 1 : PG_MODULE_MAGIC_EXT(
36 : .name = "pg_buffercache",
37 : .version = PG_VERSION
38 : );
39 :
40 : /*
41 : * Record structure holding the to be exposed cache data for OS pages. This
42 : * structure is used by pg_buffercache_os_pages(), where NUMA information may
43 : * or may not be included.
44 : */
45 : typedef struct
46 : {
47 : uint32 bufferid;
48 : int64 page_num;
49 : int32 numa_node;
50 : } BufferCacheOsPagesRec;
51 :
52 : /*
53 : * Function context for data persisting over repeated calls.
54 : */
55 : typedef struct
56 : {
57 : TupleDesc tupdesc;
58 : bool include_numa;
59 : BufferCacheOsPagesRec *record;
60 : } BufferCacheOsPagesContext;
61 :
62 :
63 : /*
64 : * Function returning data from the shared buffer cache - buffer number,
65 : * relation node/tablespace/database/blocknum and dirty indicator.
66 : */
67 2 : PG_FUNCTION_INFO_V1(pg_buffercache_pages);
68 2 : PG_FUNCTION_INFO_V1(pg_buffercache_os_pages);
69 1 : PG_FUNCTION_INFO_V1(pg_buffercache_numa_pages);
70 2 : PG_FUNCTION_INFO_V1(pg_buffercache_summary);
71 2 : PG_FUNCTION_INFO_V1(pg_buffercache_usage_counts);
72 3 : PG_FUNCTION_INFO_V1(pg_buffercache_evict);
73 2 : PG_FUNCTION_INFO_V1(pg_buffercache_evict_relation);
74 2 : PG_FUNCTION_INFO_V1(pg_buffercache_evict_all);
75 2 : PG_FUNCTION_INFO_V1(pg_buffercache_mark_dirty);
76 2 : PG_FUNCTION_INFO_V1(pg_buffercache_mark_dirty_relation);
77 2 : PG_FUNCTION_INFO_V1(pg_buffercache_mark_dirty_all);
78 :
79 :
80 : /* Only need to touch memory once per backend process lifetime */
81 : static bool firstNumaTouch = true;
82 :
83 :
84 : Datum
85 2 : pg_buffercache_pages(PG_FUNCTION_ARGS)
86 : {
87 2 : ReturnSetInfo *rsinfo = (ReturnSetInfo *) fcinfo->resultinfo;
88 : TupleDesc expected_tupledesc;
89 : int i;
90 :
91 : /*
92 : * To smoothly support upgrades from version 1.0 of this extension
93 : * transparently handle the (non-)existence of the pinning_backends
94 : * column. We unfortunately have to get the result type for that... - we
95 : * can't use the result type determined by the function definition without
96 : * potentially crashing when somebody uses the old (or even wrong)
97 : * function definition though.
98 : */
99 2 : if (get_call_result_type(fcinfo, NULL, &expected_tupledesc) != TYPEFUNC_COMPOSITE)
100 0 : elog(ERROR, "return type must be a row type");
101 :
102 2 : if (expected_tupledesc->natts < NUM_BUFFERCACHE_PAGES_MIN_ELEM ||
103 2 : expected_tupledesc->natts > NUM_BUFFERCACHE_PAGES_ELEM)
104 0 : elog(ERROR, "incorrect number of output arguments");
105 :
106 2 : InitMaterializedSRF(fcinfo, 0);
107 :
108 : /*
109 : * Scan through all the buffers, adding one row for each of the buffers to
110 : * the tuplestore.
111 : *
112 : * We don't hold the partition locks, so we don't get a consistent
113 : * snapshot across all buffers, but we do grab the buffer header locks, so
114 : * the information of each buffer is self-consistent.
115 : */
116 32770 : for (i = 0; i < NBuffers; i++)
117 : {
118 : BufferDesc *bufHdr;
119 : uint64 buf_state;
120 : uint32 bufferid;
121 : RelFileNumber relfilenumber;
122 : Oid reltablespace;
123 : Oid reldatabase;
124 : ForkNumber forknum;
125 : BlockNumber blocknum;
126 : bool isvalid;
127 : bool isdirty;
128 : uint16 usagecount;
129 : int32 pinning_backends;
130 : Datum values[NUM_BUFFERCACHE_PAGES_ELEM];
131 : bool nulls[NUM_BUFFERCACHE_PAGES_ELEM];
132 :
133 32768 : CHECK_FOR_INTERRUPTS();
134 :
135 32768 : bufHdr = GetBufferDescriptor(i);
136 : /* Lock each buffer header before inspecting. */
137 32768 : buf_state = LockBufHdr(bufHdr);
138 :
139 32768 : bufferid = BufferDescriptorGetBuffer(bufHdr);
140 32768 : relfilenumber = BufTagGetRelNumber(&bufHdr->tag);
141 32768 : reltablespace = bufHdr->tag.spcOid;
142 32768 : reldatabase = bufHdr->tag.dbOid;
143 32768 : forknum = BufTagGetForkNum(&bufHdr->tag);
144 32768 : blocknum = bufHdr->tag.blockNum;
145 32768 : usagecount = BUF_STATE_GET_USAGECOUNT(buf_state);
146 32768 : pinning_backends = BUF_STATE_GET_REFCOUNT(buf_state);
147 :
148 32768 : if (buf_state & BM_DIRTY)
149 1950 : isdirty = true;
150 : else
151 30818 : isdirty = false;
152 :
153 : /* Note if the buffer is valid, and has storage created */
154 32768 : if ((buf_state & BM_VALID) && (buf_state & BM_TAG_VALID))
155 4100 : isvalid = true;
156 : else
157 28668 : isvalid = false;
158 :
159 32768 : UnlockBufHdr(bufHdr);
160 :
161 : /* Build the tuple and add it to tuplestore */
162 32768 : values[0] = Int32GetDatum(bufferid);
163 32768 : nulls[0] = false;
164 :
165 : /*
166 : * Set all fields except the bufferid to null if the buffer is unused
167 : * or not valid.
168 : */
169 32768 : if (blocknum == InvalidBlockNumber || isvalid == false)
170 : {
171 28668 : nulls[1] = true;
172 28668 : nulls[2] = true;
173 28668 : nulls[3] = true;
174 28668 : nulls[4] = true;
175 28668 : nulls[5] = true;
176 28668 : nulls[6] = true;
177 28668 : nulls[7] = true;
178 : /* unused for v1.0 callers, but the array is always long enough */
179 28668 : nulls[8] = true;
180 : }
181 : else
182 : {
183 4100 : values[1] = ObjectIdGetDatum(relfilenumber);
184 4100 : nulls[1] = false;
185 4100 : values[2] = ObjectIdGetDatum(reltablespace);
186 4100 : nulls[2] = false;
187 4100 : values[3] = ObjectIdGetDatum(reldatabase);
188 4100 : nulls[3] = false;
189 4100 : values[4] = Int16GetDatum(forknum);
190 4100 : nulls[4] = false;
191 4100 : values[5] = Int64GetDatum((int64) blocknum);
192 4100 : nulls[5] = false;
193 4100 : values[6] = BoolGetDatum(isdirty);
194 4100 : nulls[6] = false;
195 4100 : values[7] = UInt16GetDatum(usagecount);
196 4100 : nulls[7] = false;
197 : /* unused for v1.0 callers, but the array is always long enough */
198 4100 : values[8] = Int32GetDatum(pinning_backends);
199 4100 : nulls[8] = false;
200 : }
201 :
202 32768 : tuplestore_putvalues(rsinfo->setResult, rsinfo->setDesc, values, nulls);
203 : }
204 :
205 2 : return (Datum) 0;
206 : }
207 :
208 : /*
209 : * Inquire about OS pages mappings for shared buffers, with NUMA information,
210 : * optionally.
211 : *
212 : * When "include_numa" is false, this routines ignores everything related
213 : * to NUMA (returned as NULL values), returning mapping information between
214 : * shared buffers and OS pages.
215 : *
216 : * When "include_numa" is true, NUMA is initialized and numa_node values
217 : * are generated. In order to get reliable results we also need to touch
218 : * memory pages, so that the inquiry about NUMA memory node does not return
219 : * -2, indicating unmapped/unallocated pages.
220 : *
221 : * Buffers may be smaller or larger than OS memory pages. For each buffer we
222 : * return one entry for each memory page used by the buffer (if the buffer is
223 : * smaller, it only uses a part of one memory page).
224 : *
225 : * We expect both sizes (for buffers and memory pages) to be a power-of-2, so
226 : * one is always a multiple of the other.
227 : *
228 : */
229 : static Datum
230 65538 : pg_buffercache_os_pages_internal(FunctionCallInfo fcinfo, bool include_numa)
231 : {
232 : FuncCallContext *funcctx;
233 : MemoryContext oldcontext;
234 : BufferCacheOsPagesContext *fctx; /* User function context. */
235 : TupleDesc tupledesc;
236 : TupleDesc expected_tupledesc;
237 : HeapTuple tuple;
238 : Datum result;
239 :
240 65538 : if (SRF_IS_FIRSTCALL())
241 : {
242 : int i,
243 : idx;
244 : Size os_page_size;
245 : int pages_per_buffer;
246 2 : int *os_page_status = NULL;
247 2 : uint64 os_page_count = 0;
248 : int max_entries;
249 : char *startptr,
250 : *endptr;
251 :
252 : /* If NUMA information is requested, initialize NUMA support. */
253 2 : if (include_numa && pg_numa_init() == -1)
254 0 : elog(ERROR, "libnuma initialization failed or NUMA is not supported on this platform");
255 :
256 : /*
257 : * The database block size and OS memory page size are unlikely to be
258 : * the same. The block size is 1-32KB, the memory page size depends on
259 : * platform. On x86 it's usually 4KB, on ARM it's 4KB or 64KB, but
260 : * there are also features like THP etc. Moreover, we don't quite know
261 : * how the pages and buffers "align" in memory - the buffers may be
262 : * shifted in some way, using more memory pages than necessary.
263 : *
264 : * So we need to be careful about mapping buffers to memory pages. We
265 : * calculate the maximum number of pages a buffer might use, so that
266 : * we allocate enough space for the entries. And then we count the
267 : * actual number of entries as we scan the buffers.
268 : *
269 : * This information is needed before calling move_pages() for NUMA
270 : * node id inquiry.
271 : */
272 2 : os_page_size = pg_get_shmem_pagesize();
273 :
274 : /*
275 : * The pages and block size is expected to be 2^k, so one divides the
276 : * other (we don't know in which direction). This does not say
277 : * anything about relative alignment of pages/buffers.
278 : */
279 : Assert((os_page_size % BLCKSZ == 0) || (BLCKSZ % os_page_size == 0));
280 :
281 2 : if (include_numa)
282 : {
283 0 : void **os_page_ptrs = NULL;
284 :
285 : /*
286 : * How many addresses we are going to query? Simply get the page
287 : * for the first buffer, and first page after the last buffer, and
288 : * count the pages from that.
289 : */
290 0 : startptr = (char *) TYPEALIGN_DOWN(os_page_size,
291 : BufferGetBlock(1));
292 0 : endptr = (char *) TYPEALIGN(os_page_size,
293 : (char *) BufferGetBlock(NBuffers) + BLCKSZ);
294 0 : os_page_count = (endptr - startptr) / os_page_size;
295 :
296 : /* Used to determine the NUMA node for all OS pages at once */
297 0 : os_page_ptrs = palloc0_array(void *, os_page_count);
298 0 : os_page_status = palloc_array(int, os_page_count);
299 :
300 : /*
301 : * Fill pointers for all the memory pages. This loop stores and
302 : * touches (if needed) addresses into os_page_ptrs[] as input to
303 : * one big move_pages(2) inquiry system call, as done in
304 : * pg_numa_query_pages().
305 : */
306 0 : idx = 0;
307 0 : for (char *ptr = startptr; ptr < endptr; ptr += os_page_size)
308 : {
309 0 : os_page_ptrs[idx++] = ptr;
310 :
311 : /* Only need to touch memory once per backend process lifetime */
312 0 : if (firstNumaTouch)
313 : pg_numa_touch_mem_if_required(ptr);
314 : }
315 :
316 : Assert(idx == os_page_count);
317 :
318 0 : elog(DEBUG1, "NUMA: NBuffers=%d os_page_count=" UINT64_FORMAT " "
319 : "os_page_size=%zu", NBuffers, os_page_count, os_page_size);
320 :
321 : /*
322 : * If we ever get 0xff back from kernel inquiry, then we probably
323 : * have bug in our buffers to OS page mapping code here.
324 : */
325 0 : memset(os_page_status, 0xff, sizeof(int) * os_page_count);
326 :
327 : /* Query NUMA status for all the pointers */
328 0 : if (pg_numa_query_pages(0, os_page_count, os_page_ptrs, os_page_status) == -1)
329 0 : elog(ERROR, "failed NUMA pages inquiry: %m");
330 : }
331 :
332 : /* Initialize the multi-call context, load entries about buffers */
333 :
334 2 : funcctx = SRF_FIRSTCALL_INIT();
335 :
336 : /* Switch context when allocating stuff to be used in later calls */
337 2 : oldcontext = MemoryContextSwitchTo(funcctx->multi_call_memory_ctx);
338 :
339 : /* Create a user function context for cross-call persistence */
340 2 : fctx = palloc_object(BufferCacheOsPagesContext);
341 :
342 2 : if (get_call_result_type(fcinfo, NULL, &expected_tupledesc) != TYPEFUNC_COMPOSITE)
343 0 : elog(ERROR, "return type must be a row type");
344 :
345 2 : if (expected_tupledesc->natts != NUM_BUFFERCACHE_OS_PAGES_ELEM)
346 0 : elog(ERROR, "incorrect number of output arguments");
347 :
348 : /* Construct a tuple descriptor for the result rows. */
349 2 : tupledesc = CreateTemplateTupleDesc(expected_tupledesc->natts);
350 2 : TupleDescInitEntry(tupledesc, (AttrNumber) 1, "bufferid",
351 : INT4OID, -1, 0);
352 2 : TupleDescInitEntry(tupledesc, (AttrNumber) 2, "os_page_num",
353 : INT8OID, -1, 0);
354 2 : TupleDescInitEntry(tupledesc, (AttrNumber) 3, "numa_node",
355 : INT4OID, -1, 0);
356 :
357 2 : TupleDescFinalize(tupledesc);
358 2 : fctx->tupdesc = BlessTupleDesc(tupledesc);
359 2 : fctx->include_numa = include_numa;
360 :
361 : /*
362 : * Each buffer needs at least one entry, but it might be offset in
363 : * some way, and use one extra entry. So we allocate space for the
364 : * maximum number of entries we might need, and then count the exact
365 : * number as we're walking buffers. That way we can do it in one pass,
366 : * without reallocating memory.
367 : */
368 2 : pages_per_buffer = Max(1, BLCKSZ / os_page_size) + 1;
369 2 : max_entries = NBuffers * pages_per_buffer;
370 :
371 : /* Allocate entries for BufferCacheOsPagesRec records. */
372 2 : fctx->record = (BufferCacheOsPagesRec *)
373 2 : MemoryContextAllocHuge(CurrentMemoryContext,
374 : sizeof(BufferCacheOsPagesRec) * max_entries);
375 :
376 : /* Return to original context when allocating transient memory */
377 2 : MemoryContextSwitchTo(oldcontext);
378 :
379 2 : if (include_numa && firstNumaTouch)
380 0 : elog(DEBUG1, "NUMA: page-faulting the buffercache for proper NUMA readouts");
381 :
382 : /*
383 : * Scan through all the buffers, saving the relevant fields in the
384 : * fctx->record structure.
385 : *
386 : * We don't hold the partition locks, so we don't get a consistent
387 : * snapshot across all buffers, but we do grab the buffer header
388 : * locks, so the information of each buffer is self-consistent.
389 : */
390 2 : startptr = (char *) TYPEALIGN_DOWN(os_page_size, (char *) BufferGetBlock(1));
391 2 : idx = 0;
392 32770 : for (i = 0; i < NBuffers; i++)
393 : {
394 32768 : char *buffptr = (char *) BufferGetBlock(i + 1);
395 : BufferDesc *bufHdr;
396 : uint32 bufferid;
397 : int32 page_num;
398 : char *startptr_buff,
399 : *endptr_buff;
400 :
401 32768 : CHECK_FOR_INTERRUPTS();
402 :
403 32768 : bufHdr = GetBufferDescriptor(i);
404 :
405 : /* Lock each buffer header before inspecting. */
406 32768 : LockBufHdr(bufHdr);
407 32768 : bufferid = BufferDescriptorGetBuffer(bufHdr);
408 32768 : UnlockBufHdr(bufHdr);
409 :
410 : /* start of the first page of this buffer */
411 32768 : startptr_buff = (char *) TYPEALIGN_DOWN(os_page_size, buffptr);
412 :
413 : /* end of the buffer (no need to align to memory page) */
414 32768 : endptr_buff = buffptr + BLCKSZ;
415 :
416 : Assert(startptr_buff < endptr_buff);
417 :
418 : /* calculate ID of the first page for this buffer */
419 32768 : page_num = (startptr_buff - startptr) / os_page_size;
420 :
421 : /* Add an entry for each OS page overlapping with this buffer. */
422 98304 : for (char *ptr = startptr_buff; ptr < endptr_buff; ptr += os_page_size)
423 : {
424 65536 : fctx->record[idx].bufferid = bufferid;
425 65536 : fctx->record[idx].page_num = page_num;
426 65536 : fctx->record[idx].numa_node = include_numa ? os_page_status[page_num] : -1;
427 :
428 : /* advance to the next entry/page */
429 65536 : ++idx;
430 65536 : ++page_num;
431 : }
432 : }
433 :
434 : Assert(idx <= max_entries);
435 :
436 : if (include_numa)
437 : Assert(idx >= os_page_count);
438 :
439 : /* Set max calls and remember the user function context. */
440 2 : funcctx->max_calls = idx;
441 2 : funcctx->user_fctx = fctx;
442 :
443 : /* Remember this backend touched the pages (only relevant for NUMA) */
444 2 : if (include_numa)
445 0 : firstNumaTouch = false;
446 : }
447 :
448 65538 : funcctx = SRF_PERCALL_SETUP();
449 :
450 : /* Get the saved state */
451 65538 : fctx = funcctx->user_fctx;
452 :
453 65538 : if (funcctx->call_cntr < funcctx->max_calls)
454 : {
455 65536 : uint32 i = funcctx->call_cntr;
456 : Datum values[NUM_BUFFERCACHE_OS_PAGES_ELEM];
457 : bool nulls[NUM_BUFFERCACHE_OS_PAGES_ELEM];
458 :
459 65536 : values[0] = Int32GetDatum(fctx->record[i].bufferid);
460 65536 : nulls[0] = false;
461 :
462 65536 : values[1] = Int64GetDatum(fctx->record[i].page_num);
463 65536 : nulls[1] = false;
464 :
465 65536 : if (fctx->include_numa)
466 : {
467 : /* status is valid node number */
468 0 : if (fctx->record[i].numa_node >= 0)
469 : {
470 0 : values[2] = Int32GetDatum(fctx->record[i].numa_node);
471 0 : nulls[2] = false;
472 : }
473 : else
474 : {
475 : /* some kind of error (e.g. pages moved to swap) */
476 0 : values[2] = (Datum) 0;
477 0 : nulls[2] = true;
478 : }
479 : }
480 : else
481 : {
482 65536 : values[2] = (Datum) 0;
483 65536 : nulls[2] = true;
484 : }
485 :
486 : /* Build and return the tuple. */
487 65536 : tuple = heap_form_tuple(fctx->tupdesc, values, nulls);
488 65536 : result = HeapTupleGetDatum(tuple);
489 :
490 65536 : SRF_RETURN_NEXT(funcctx, result);
491 : }
492 : else
493 2 : SRF_RETURN_DONE(funcctx);
494 : }
495 :
496 : /*
497 : * pg_buffercache_os_pages
498 : *
499 : * Retrieve information about OS pages, with or without NUMA information.
500 : */
501 : Datum
502 65538 : pg_buffercache_os_pages(PG_FUNCTION_ARGS)
503 : {
504 : bool include_numa;
505 :
506 : /* Get the boolean parameter that controls the NUMA behavior. */
507 65538 : include_numa = PG_GETARG_BOOL(0);
508 :
509 65538 : return pg_buffercache_os_pages_internal(fcinfo, include_numa);
510 : }
511 :
512 : /* Backward-compatible wrapper for v1.6. */
513 : Datum
514 0 : pg_buffercache_numa_pages(PG_FUNCTION_ARGS)
515 : {
516 : /* Call internal function with include_numa=true */
517 0 : return pg_buffercache_os_pages_internal(fcinfo, true);
518 : }
519 :
520 : Datum
521 2 : pg_buffercache_summary(PG_FUNCTION_ARGS)
522 : {
523 : Datum result;
524 : TupleDesc tupledesc;
525 : HeapTuple tuple;
526 : Datum values[NUM_BUFFERCACHE_SUMMARY_ELEM];
527 : bool nulls[NUM_BUFFERCACHE_SUMMARY_ELEM];
528 :
529 2 : int32 buffers_used = 0;
530 2 : int32 buffers_unused = 0;
531 2 : int32 buffers_dirty = 0;
532 2 : int32 buffers_pinned = 0;
533 2 : int64 usagecount_total = 0;
534 :
535 2 : if (get_call_result_type(fcinfo, NULL, &tupledesc) != TYPEFUNC_COMPOSITE)
536 0 : elog(ERROR, "return type must be a row type");
537 :
538 32770 : for (int i = 0; i < NBuffers; i++)
539 : {
540 : BufferDesc *bufHdr;
541 : uint64 buf_state;
542 :
543 32768 : CHECK_FOR_INTERRUPTS();
544 :
545 : /*
546 : * This function summarizes the state of all headers. Locking the
547 : * buffer headers wouldn't provide an improved result as the state of
548 : * the buffer can still change after we release the lock and it'd
549 : * noticeably increase the cost of the function.
550 : */
551 32768 : bufHdr = GetBufferDescriptor(i);
552 32768 : buf_state = pg_atomic_read_u64(&bufHdr->state);
553 :
554 32768 : if (buf_state & BM_VALID)
555 : {
556 4100 : buffers_used++;
557 4100 : usagecount_total += BUF_STATE_GET_USAGECOUNT(buf_state);
558 :
559 4100 : if (buf_state & BM_DIRTY)
560 1950 : buffers_dirty++;
561 : }
562 : else
563 28668 : buffers_unused++;
564 :
565 32768 : if (BUF_STATE_GET_REFCOUNT(buf_state) > 0)
566 0 : buffers_pinned++;
567 : }
568 :
569 2 : memset(nulls, 0, sizeof(nulls));
570 2 : values[0] = Int32GetDatum(buffers_used);
571 2 : values[1] = Int32GetDatum(buffers_unused);
572 2 : values[2] = Int32GetDatum(buffers_dirty);
573 2 : values[3] = Int32GetDatum(buffers_pinned);
574 :
575 2 : if (buffers_used != 0)
576 2 : values[4] = Float8GetDatum((double) usagecount_total / buffers_used);
577 : else
578 0 : nulls[4] = true;
579 :
580 : /* Build and return the tuple. */
581 2 : tuple = heap_form_tuple(tupledesc, values, nulls);
582 2 : result = HeapTupleGetDatum(tuple);
583 :
584 2 : PG_RETURN_DATUM(result);
585 : }
586 :
587 : Datum
588 2 : pg_buffercache_usage_counts(PG_FUNCTION_ARGS)
589 : {
590 2 : ReturnSetInfo *rsinfo = (ReturnSetInfo *) fcinfo->resultinfo;
591 2 : int usage_counts[BM_MAX_USAGE_COUNT + 1] = {0};
592 2 : int dirty[BM_MAX_USAGE_COUNT + 1] = {0};
593 2 : int pinned[BM_MAX_USAGE_COUNT + 1] = {0};
594 : Datum values[NUM_BUFFERCACHE_USAGE_COUNTS_ELEM];
595 2 : bool nulls[NUM_BUFFERCACHE_USAGE_COUNTS_ELEM] = {0};
596 :
597 2 : InitMaterializedSRF(fcinfo, 0);
598 :
599 32770 : for (int i = 0; i < NBuffers; i++)
600 : {
601 32768 : BufferDesc *bufHdr = GetBufferDescriptor(i);
602 32768 : uint64 buf_state = pg_atomic_read_u64(&bufHdr->state);
603 : int usage_count;
604 :
605 32768 : CHECK_FOR_INTERRUPTS();
606 :
607 32768 : usage_count = BUF_STATE_GET_USAGECOUNT(buf_state);
608 32768 : usage_counts[usage_count]++;
609 :
610 32768 : if (buf_state & BM_DIRTY)
611 1950 : dirty[usage_count]++;
612 :
613 32768 : if (BUF_STATE_GET_REFCOUNT(buf_state) > 0)
614 0 : pinned[usage_count]++;
615 : }
616 :
617 14 : for (int i = 0; i < BM_MAX_USAGE_COUNT + 1; i++)
618 : {
619 12 : values[0] = Int32GetDatum(i);
620 12 : values[1] = Int32GetDatum(usage_counts[i]);
621 12 : values[2] = Int32GetDatum(dirty[i]);
622 12 : values[3] = Int32GetDatum(pinned[i]);
623 :
624 12 : tuplestore_putvalues(rsinfo->setResult, rsinfo->setDesc, values, nulls);
625 : }
626 :
627 2 : return (Datum) 0;
628 : }
629 :
630 : /*
631 : * Helper function to check if the user has superuser privileges.
632 : */
633 : static void
634 20 : pg_buffercache_superuser_check(char *func_name)
635 : {
636 20 : if (!superuser())
637 6 : ereport(ERROR,
638 : (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
639 : errmsg("must be superuser to use %s()",
640 : func_name)));
641 14 : }
642 :
643 : /*
644 : * Try to evict a shared buffer.
645 : */
646 : Datum
647 5 : pg_buffercache_evict(PG_FUNCTION_ARGS)
648 : {
649 : Datum result;
650 : TupleDesc tupledesc;
651 : HeapTuple tuple;
652 : Datum values[NUM_BUFFERCACHE_EVICT_ELEM];
653 5 : bool nulls[NUM_BUFFERCACHE_EVICT_ELEM] = {0};
654 :
655 5 : Buffer buf = PG_GETARG_INT32(0);
656 : bool buffer_flushed;
657 :
658 5 : if (get_call_result_type(fcinfo, NULL, &tupledesc) != TYPEFUNC_COMPOSITE)
659 0 : elog(ERROR, "return type must be a row type");
660 :
661 5 : pg_buffercache_superuser_check("pg_buffercache_evict");
662 :
663 4 : if (buf < 1 || buf > NBuffers)
664 3 : elog(ERROR, "bad buffer ID: %d", buf);
665 :
666 1 : values[0] = BoolGetDatum(EvictUnpinnedBuffer(buf, &buffer_flushed));
667 1 : values[1] = BoolGetDatum(buffer_flushed);
668 :
669 1 : tuple = heap_form_tuple(tupledesc, values, nulls);
670 1 : result = HeapTupleGetDatum(tuple);
671 :
672 1 : PG_RETURN_DATUM(result);
673 : }
674 :
675 : /*
676 : * Try to evict specified relation.
677 : */
678 : Datum
679 3 : pg_buffercache_evict_relation(PG_FUNCTION_ARGS)
680 : {
681 : Datum result;
682 : TupleDesc tupledesc;
683 : HeapTuple tuple;
684 : Datum values[NUM_BUFFERCACHE_EVICT_RELATION_ELEM];
685 3 : bool nulls[NUM_BUFFERCACHE_EVICT_RELATION_ELEM] = {0};
686 :
687 : Oid relOid;
688 : Relation rel;
689 :
690 3 : int32 buffers_evicted = 0;
691 3 : int32 buffers_flushed = 0;
692 3 : int32 buffers_skipped = 0;
693 :
694 3 : if (get_call_result_type(fcinfo, NULL, &tupledesc) != TYPEFUNC_COMPOSITE)
695 0 : elog(ERROR, "return type must be a row type");
696 :
697 3 : pg_buffercache_superuser_check("pg_buffercache_evict_relation");
698 :
699 2 : relOid = PG_GETARG_OID(0);
700 :
701 2 : rel = relation_open(relOid, AccessShareLock);
702 :
703 2 : if (RelationUsesLocalBuffers(rel))
704 1 : ereport(ERROR,
705 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
706 : errmsg("relation uses local buffers, %s() is intended to be used for shared buffers only",
707 : "pg_buffercache_evict_relation")));
708 :
709 1 : EvictRelUnpinnedBuffers(rel, &buffers_evicted, &buffers_flushed,
710 : &buffers_skipped);
711 :
712 1 : relation_close(rel, AccessShareLock);
713 :
714 1 : values[0] = Int32GetDatum(buffers_evicted);
715 1 : values[1] = Int32GetDatum(buffers_flushed);
716 1 : values[2] = Int32GetDatum(buffers_skipped);
717 :
718 1 : tuple = heap_form_tuple(tupledesc, values, nulls);
719 1 : result = HeapTupleGetDatum(tuple);
720 :
721 1 : PG_RETURN_DATUM(result);
722 : }
723 :
724 :
725 : /*
726 : * Try to evict all shared buffers.
727 : */
728 : Datum
729 2 : pg_buffercache_evict_all(PG_FUNCTION_ARGS)
730 : {
731 : Datum result;
732 : TupleDesc tupledesc;
733 : HeapTuple tuple;
734 : Datum values[NUM_BUFFERCACHE_EVICT_ALL_ELEM];
735 2 : bool nulls[NUM_BUFFERCACHE_EVICT_ALL_ELEM] = {0};
736 :
737 2 : int32 buffers_evicted = 0;
738 2 : int32 buffers_flushed = 0;
739 2 : int32 buffers_skipped = 0;
740 :
741 2 : if (get_call_result_type(fcinfo, NULL, &tupledesc) != TYPEFUNC_COMPOSITE)
742 0 : elog(ERROR, "return type must be a row type");
743 :
744 2 : pg_buffercache_superuser_check("pg_buffercache_evict_all");
745 :
746 1 : EvictAllUnpinnedBuffers(&buffers_evicted, &buffers_flushed,
747 : &buffers_skipped);
748 :
749 1 : values[0] = Int32GetDatum(buffers_evicted);
750 1 : values[1] = Int32GetDatum(buffers_flushed);
751 1 : values[2] = Int32GetDatum(buffers_skipped);
752 :
753 1 : tuple = heap_form_tuple(tupledesc, values, nulls);
754 1 : result = HeapTupleGetDatum(tuple);
755 :
756 1 : PG_RETURN_DATUM(result);
757 : }
758 :
759 : /*
760 : * Try to mark a shared buffer as dirty.
761 : */
762 : Datum
763 5 : pg_buffercache_mark_dirty(PG_FUNCTION_ARGS)
764 : {
765 :
766 : Datum result;
767 : TupleDesc tupledesc;
768 : HeapTuple tuple;
769 : Datum values[NUM_BUFFERCACHE_MARK_DIRTY_ELEM];
770 5 : bool nulls[NUM_BUFFERCACHE_MARK_DIRTY_ELEM] = {0};
771 :
772 5 : Buffer buf = PG_GETARG_INT32(0);
773 : bool buffer_already_dirty;
774 :
775 5 : if (get_call_result_type(fcinfo, NULL, &tupledesc) != TYPEFUNC_COMPOSITE)
776 0 : elog(ERROR, "return type must be a row type");
777 :
778 5 : pg_buffercache_superuser_check("pg_buffercache_mark_dirty");
779 :
780 4 : if (buf < 1 || buf > NBuffers)
781 3 : elog(ERROR, "bad buffer ID: %d", buf);
782 :
783 1 : values[0] = BoolGetDatum(MarkDirtyUnpinnedBuffer(buf, &buffer_already_dirty));
784 1 : values[1] = BoolGetDatum(buffer_already_dirty);
785 :
786 1 : tuple = heap_form_tuple(tupledesc, values, nulls);
787 1 : result = HeapTupleGetDatum(tuple);
788 :
789 1 : PG_RETURN_DATUM(result);
790 : }
791 :
792 : /*
793 : * Try to mark all the shared buffers of a relation as dirty.
794 : */
795 : Datum
796 3 : pg_buffercache_mark_dirty_relation(PG_FUNCTION_ARGS)
797 : {
798 : Datum result;
799 : TupleDesc tupledesc;
800 : HeapTuple tuple;
801 : Datum values[NUM_BUFFERCACHE_MARK_DIRTY_RELATION_ELEM];
802 3 : bool nulls[NUM_BUFFERCACHE_MARK_DIRTY_RELATION_ELEM] = {0};
803 :
804 : Oid relOid;
805 : Relation rel;
806 :
807 3 : int32 buffers_already_dirty = 0;
808 3 : int32 buffers_dirtied = 0;
809 3 : int32 buffers_skipped = 0;
810 :
811 3 : if (get_call_result_type(fcinfo, NULL, &tupledesc) != TYPEFUNC_COMPOSITE)
812 0 : elog(ERROR, "return type must be a row type");
813 :
814 3 : pg_buffercache_superuser_check("pg_buffercache_mark_dirty_relation");
815 :
816 2 : relOid = PG_GETARG_OID(0);
817 :
818 2 : rel = relation_open(relOid, AccessShareLock);
819 :
820 2 : if (RelationUsesLocalBuffers(rel))
821 1 : ereport(ERROR,
822 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
823 : errmsg("relation uses local buffers, %s() is intended to be used for shared buffers only",
824 : "pg_buffercache_mark_dirty_relation")));
825 :
826 1 : MarkDirtyRelUnpinnedBuffers(rel, &buffers_dirtied, &buffers_already_dirty,
827 : &buffers_skipped);
828 :
829 1 : relation_close(rel, AccessShareLock);
830 :
831 1 : values[0] = Int32GetDatum(buffers_dirtied);
832 1 : values[1] = Int32GetDatum(buffers_already_dirty);
833 1 : values[2] = Int32GetDatum(buffers_skipped);
834 :
835 1 : tuple = heap_form_tuple(tupledesc, values, nulls);
836 1 : result = HeapTupleGetDatum(tuple);
837 :
838 1 : PG_RETURN_DATUM(result);
839 : }
840 :
841 : /*
842 : * Try to mark all the shared buffers as dirty.
843 : */
844 : Datum
845 2 : pg_buffercache_mark_dirty_all(PG_FUNCTION_ARGS)
846 : {
847 : Datum result;
848 : TupleDesc tupledesc;
849 : HeapTuple tuple;
850 : Datum values[NUM_BUFFERCACHE_MARK_DIRTY_ALL_ELEM];
851 2 : bool nulls[NUM_BUFFERCACHE_MARK_DIRTY_ALL_ELEM] = {0};
852 :
853 2 : int32 buffers_already_dirty = 0;
854 2 : int32 buffers_dirtied = 0;
855 2 : int32 buffers_skipped = 0;
856 :
857 2 : if (get_call_result_type(fcinfo, NULL, &tupledesc) != TYPEFUNC_COMPOSITE)
858 0 : elog(ERROR, "return type must be a row type");
859 :
860 2 : pg_buffercache_superuser_check("pg_buffercache_mark_dirty_all");
861 :
862 1 : MarkDirtyAllUnpinnedBuffers(&buffers_dirtied, &buffers_already_dirty,
863 : &buffers_skipped);
864 :
865 1 : values[0] = Int32GetDatum(buffers_dirtied);
866 1 : values[1] = Int32GetDatum(buffers_already_dirty);
867 1 : values[2] = Int32GetDatum(buffers_skipped);
868 :
869 1 : tuple = heap_form_tuple(tupledesc, values, nulls);
870 1 : result = HeapTupleGetDatum(tuple);
871 :
872 1 : PG_RETURN_DATUM(result);
873 : }
|
