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 : static TupleDesc build_buffercache_pages_tupledesc(int natts);
63 :
64 :
65 : /*
66 : * Function returning data from the shared buffer cache - buffer number,
67 : * relation node/tablespace/database/blocknum and dirty indicator.
68 : */
69 2 : PG_FUNCTION_INFO_V1(pg_buffercache_pages);
70 2 : PG_FUNCTION_INFO_V1(pg_buffercache_os_pages);
71 1 : PG_FUNCTION_INFO_V1(pg_buffercache_numa_pages);
72 2 : PG_FUNCTION_INFO_V1(pg_buffercache_summary);
73 2 : PG_FUNCTION_INFO_V1(pg_buffercache_usage_counts);
74 3 : PG_FUNCTION_INFO_V1(pg_buffercache_evict);
75 2 : PG_FUNCTION_INFO_V1(pg_buffercache_evict_relation);
76 2 : PG_FUNCTION_INFO_V1(pg_buffercache_evict_all);
77 2 : PG_FUNCTION_INFO_V1(pg_buffercache_mark_dirty);
78 2 : PG_FUNCTION_INFO_V1(pg_buffercache_mark_dirty_relation);
79 2 : PG_FUNCTION_INFO_V1(pg_buffercache_mark_dirty_all);
80 :
81 :
82 : /* Only need to touch memory once per backend process lifetime */
83 : static bool firstNumaTouch = true;
84 :
85 :
86 : Datum
87 3 : pg_buffercache_pages(PG_FUNCTION_ARGS)
88 : {
89 3 : ReturnSetInfo *rsinfo = (ReturnSetInfo *) fcinfo->resultinfo;
90 : TupleDesc expected_tupledesc;
91 : TupleDesc actual_tupledesc;
92 : MemoryContext oldcontext;
93 : int i;
94 :
95 : /*
96 : * To smoothly support upgrades from version 1.0 of this extension
97 : * transparently handle the (non-)existence of the pinning_backends
98 : * column. We unfortunately have to get the result type for that... - we
99 : * can't use the result type determined by the function definition without
100 : * potentially crashing when somebody uses the old (or even wrong)
101 : * function definition though.
102 : */
103 3 : if (get_call_result_type(fcinfo, NULL, &expected_tupledesc) != TYPEFUNC_COMPOSITE)
104 0 : elog(ERROR, "return type must be a row type");
105 :
106 3 : if (expected_tupledesc->natts < NUM_BUFFERCACHE_PAGES_MIN_ELEM ||
107 3 : expected_tupledesc->natts > NUM_BUFFERCACHE_PAGES_ELEM)
108 0 : elog(ERROR, "incorrect number of output arguments");
109 :
110 3 : InitMaterializedSRF(fcinfo, 0);
111 :
112 3 : oldcontext = MemoryContextSwitchTo(rsinfo->econtext->ecxt_per_query_memory);
113 3 : actual_tupledesc = build_buffercache_pages_tupledesc(expected_tupledesc->natts);
114 3 : MemoryContextSwitchTo(oldcontext);
115 :
116 : /*
117 : * Override the caller-supplied descriptor with the tuple descriptor that
118 : * matches the values we actually return, so executor-side
119 : * tupledesc_match() can verify the caller's row definition.
120 : *
121 : * Do not free the previous rsinfo->setDesc here: for RECORD results it
122 : * can alias rsinfo->expectedDesc, which the executor still needs to
123 : * reference.
124 : */
125 3 : rsinfo->setDesc = actual_tupledesc;
126 :
127 : /*
128 : * Scan through all the buffers, adding one row for each of the buffers to
129 : * the tuplestore.
130 : *
131 : * We don't hold the partition locks, so we don't get a consistent
132 : * snapshot across all buffers, but we do grab the buffer header locks, so
133 : * the information of each buffer is self-consistent.
134 : */
135 49155 : for (i = 0; i < NBuffers; i++)
136 : {
137 : BufferDesc *bufHdr;
138 : uint64 buf_state;
139 : uint32 bufferid;
140 : RelFileNumber relfilenumber;
141 : Oid reltablespace;
142 : Oid reldatabase;
143 : ForkNumber forknum;
144 : BlockNumber blocknum;
145 : bool isvalid;
146 : bool isdirty;
147 : uint16 usagecount;
148 : int32 pinning_backends;
149 : Datum values[NUM_BUFFERCACHE_PAGES_ELEM];
150 : bool nulls[NUM_BUFFERCACHE_PAGES_ELEM];
151 :
152 49152 : CHECK_FOR_INTERRUPTS();
153 :
154 49152 : bufHdr = GetBufferDescriptor(i);
155 : /* Lock each buffer header before inspecting. */
156 49152 : buf_state = LockBufHdr(bufHdr);
157 :
158 49152 : bufferid = BufferDescriptorGetBuffer(bufHdr);
159 49152 : relfilenumber = BufTagGetRelNumber(&bufHdr->tag);
160 49152 : reltablespace = bufHdr->tag.spcOid;
161 49152 : reldatabase = bufHdr->tag.dbOid;
162 49152 : forknum = BufTagGetForkNum(&bufHdr->tag);
163 49152 : blocknum = bufHdr->tag.blockNum;
164 49152 : usagecount = BUF_STATE_GET_USAGECOUNT(buf_state);
165 49152 : pinning_backends = BUF_STATE_GET_REFCOUNT(buf_state);
166 :
167 49152 : if (buf_state & BM_DIRTY)
168 2925 : isdirty = true;
169 : else
170 46227 : isdirty = false;
171 :
172 : /* Note if the buffer is valid, and has storage created */
173 49152 : if ((buf_state & BM_VALID) && (buf_state & BM_TAG_VALID))
174 6070 : isvalid = true;
175 : else
176 43082 : isvalid = false;
177 :
178 49152 : UnlockBufHdr(bufHdr);
179 :
180 : /* Build the tuple and add it to tuplestore */
181 49152 : values[0] = Int32GetDatum(bufferid);
182 49152 : nulls[0] = false;
183 :
184 : /*
185 : * Set all fields except the bufferid to null if the buffer is unused
186 : * or not valid.
187 : */
188 49152 : if (blocknum == InvalidBlockNumber || isvalid == false)
189 : {
190 43082 : nulls[1] = true;
191 43082 : nulls[2] = true;
192 43082 : nulls[3] = true;
193 43082 : nulls[4] = true;
194 43082 : nulls[5] = true;
195 43082 : nulls[6] = true;
196 43082 : nulls[7] = true;
197 : /* unused for v1.0 callers, but the array is always long enough */
198 43082 : nulls[8] = true;
199 : }
200 : else
201 : {
202 6070 : values[1] = ObjectIdGetDatum(relfilenumber);
203 6070 : nulls[1] = false;
204 6070 : values[2] = ObjectIdGetDatum(reltablespace);
205 6070 : nulls[2] = false;
206 6070 : values[3] = ObjectIdGetDatum(reldatabase);
207 6070 : nulls[3] = false;
208 6070 : values[4] = Int16GetDatum(forknum);
209 6070 : nulls[4] = false;
210 6070 : values[5] = Int64GetDatum((int64) blocknum);
211 6070 : nulls[5] = false;
212 6070 : values[6] = BoolGetDatum(isdirty);
213 6070 : nulls[6] = false;
214 6070 : values[7] = Int16GetDatum(usagecount);
215 6070 : nulls[7] = false;
216 : /* unused for v1.0 callers, but the array is always long enough */
217 6070 : values[8] = Int32GetDatum(pinning_backends);
218 6070 : nulls[8] = false;
219 : }
220 :
221 49152 : tuplestore_putvalues(rsinfo->setResult, rsinfo->setDesc, values, nulls);
222 : }
223 :
224 3 : return (Datum) 0;
225 : }
226 :
227 : static TupleDesc
228 3 : build_buffercache_pages_tupledesc(int natts)
229 : {
230 : TupleDesc tupledesc;
231 :
232 3 : tupledesc = CreateTemplateTupleDesc(natts);
233 3 : TupleDescInitEntry(tupledesc, (AttrNumber) 1, "bufferid",
234 : INT4OID, -1, 0);
235 3 : TupleDescInitEntry(tupledesc, (AttrNumber) 2, "relfilenode",
236 : OIDOID, -1, 0);
237 3 : TupleDescInitEntry(tupledesc, (AttrNumber) 3, "reltablespace",
238 : OIDOID, -1, 0);
239 3 : TupleDescInitEntry(tupledesc, (AttrNumber) 4, "reldatabase",
240 : OIDOID, -1, 0);
241 3 : TupleDescInitEntry(tupledesc, (AttrNumber) 5, "relforknumber",
242 : INT2OID, -1, 0);
243 3 : TupleDescInitEntry(tupledesc, (AttrNumber) 6, "relblocknumber",
244 : INT8OID, -1, 0);
245 3 : TupleDescInitEntry(tupledesc, (AttrNumber) 7, "isdirty",
246 : BOOLOID, -1, 0);
247 3 : TupleDescInitEntry(tupledesc, (AttrNumber) 8, "usagecount",
248 : INT2OID, -1, 0);
249 :
250 3 : if (natts == NUM_BUFFERCACHE_PAGES_ELEM)
251 2 : TupleDescInitEntry(tupledesc, (AttrNumber) 9, "pinning_backends",
252 : INT4OID, -1, 0);
253 :
254 3 : TupleDescFinalize(tupledesc);
255 :
256 3 : return BlessTupleDesc(tupledesc);
257 : }
258 :
259 : /*
260 : * Inquire about OS pages mappings for shared buffers, with NUMA information,
261 : * optionally.
262 : *
263 : * When "include_numa" is false, this routines ignores everything related
264 : * to NUMA (returned as NULL values), returning mapping information between
265 : * shared buffers and OS pages.
266 : *
267 : * When "include_numa" is true, NUMA is initialized and numa_node values
268 : * are generated. In order to get reliable results we also need to touch
269 : * memory pages, so that the inquiry about NUMA memory node does not return
270 : * -2, indicating unmapped/unallocated pages.
271 : *
272 : * Buffers may be smaller or larger than OS memory pages. For each buffer we
273 : * return one entry for each memory page used by the buffer (if the buffer is
274 : * smaller, it only uses a part of one memory page).
275 : *
276 : * We expect both sizes (for buffers and memory pages) to be a power-of-2, so
277 : * one is always a multiple of the other.
278 : *
279 : */
280 : static Datum
281 65538 : pg_buffercache_os_pages_internal(FunctionCallInfo fcinfo, bool include_numa)
282 : {
283 : FuncCallContext *funcctx;
284 : MemoryContext oldcontext;
285 : BufferCacheOsPagesContext *fctx; /* User function context. */
286 : TupleDesc tupledesc;
287 : TupleDesc expected_tupledesc;
288 : HeapTuple tuple;
289 : Datum result;
290 :
291 65538 : if (SRF_IS_FIRSTCALL())
292 : {
293 : int i,
294 : idx;
295 : Size os_page_size;
296 : int pages_per_buffer;
297 2 : int *os_page_status = NULL;
298 2 : uint64 os_page_count = 0;
299 : int max_entries;
300 : char *startptr,
301 : *endptr;
302 :
303 : /* If NUMA information is requested, initialize NUMA support. */
304 2 : if (include_numa && pg_numa_init() == -1)
305 0 : elog(ERROR, "libnuma initialization failed or NUMA is not supported on this platform");
306 :
307 : /*
308 : * The database block size and OS memory page size are unlikely to be
309 : * the same. The block size is 1-32KB, the memory page size depends on
310 : * platform. On x86 it's usually 4KB, on ARM it's 4KB or 64KB, but
311 : * there are also features like THP etc. Moreover, we don't quite know
312 : * how the pages and buffers "align" in memory - the buffers may be
313 : * shifted in some way, using more memory pages than necessary.
314 : *
315 : * So we need to be careful about mapping buffers to memory pages. We
316 : * calculate the maximum number of pages a buffer might use, so that
317 : * we allocate enough space for the entries. And then we count the
318 : * actual number of entries as we scan the buffers.
319 : *
320 : * This information is needed before calling move_pages() for NUMA
321 : * node id inquiry.
322 : */
323 2 : os_page_size = pg_get_shmem_pagesize();
324 :
325 : /*
326 : * The pages and block size is expected to be 2^k, so one divides the
327 : * other (we don't know in which direction). This does not say
328 : * anything about relative alignment of pages/buffers.
329 : */
330 : Assert((os_page_size % BLCKSZ == 0) || (BLCKSZ % os_page_size == 0));
331 :
332 2 : if (include_numa)
333 : {
334 0 : void **os_page_ptrs = NULL;
335 :
336 : /*
337 : * How many addresses we are going to query? Simply get the page
338 : * for the first buffer, and first page after the last buffer, and
339 : * count the pages from that.
340 : */
341 0 : startptr = (char *) TYPEALIGN_DOWN(os_page_size,
342 : BufferGetBlock(1));
343 0 : endptr = (char *) TYPEALIGN(os_page_size,
344 : (char *) BufferGetBlock(NBuffers) + BLCKSZ);
345 0 : os_page_count = (endptr - startptr) / os_page_size;
346 :
347 : /* Used to determine the NUMA node for all OS pages at once */
348 0 : os_page_ptrs = palloc0_array(void *, os_page_count);
349 0 : os_page_status = palloc_array(int, os_page_count);
350 :
351 : /*
352 : * Fill pointers for all the memory pages. This loop stores and
353 : * touches (if needed) addresses into os_page_ptrs[] as input to
354 : * one big move_pages(2) inquiry system call, as done in
355 : * pg_numa_query_pages().
356 : */
357 0 : idx = 0;
358 0 : for (char *ptr = startptr; ptr < endptr; ptr += os_page_size)
359 : {
360 0 : os_page_ptrs[idx++] = ptr;
361 :
362 : /* Only need to touch memory once per backend process lifetime */
363 0 : if (firstNumaTouch)
364 : pg_numa_touch_mem_if_required(ptr);
365 : }
366 :
367 : Assert(idx == os_page_count);
368 :
369 0 : elog(DEBUG1, "NUMA: NBuffers=%d os_page_count=" UINT64_FORMAT " "
370 : "os_page_size=%zu", NBuffers, os_page_count, os_page_size);
371 :
372 : /*
373 : * If we ever get 0xff back from kernel inquiry, then we probably
374 : * have bug in our buffers to OS page mapping code here.
375 : */
376 0 : memset(os_page_status, 0xff, sizeof(int) * os_page_count);
377 :
378 : /* Query NUMA status for all the pointers */
379 0 : if (pg_numa_query_pages(0, os_page_count, os_page_ptrs, os_page_status) == -1)
380 0 : elog(ERROR, "failed NUMA pages inquiry: %m");
381 : }
382 :
383 : /* Initialize the multi-call context, load entries about buffers */
384 :
385 2 : funcctx = SRF_FIRSTCALL_INIT();
386 :
387 : /* Switch context when allocating stuff to be used in later calls */
388 2 : oldcontext = MemoryContextSwitchTo(funcctx->multi_call_memory_ctx);
389 :
390 : /* Create a user function context for cross-call persistence */
391 2 : fctx = palloc_object(BufferCacheOsPagesContext);
392 :
393 2 : if (get_call_result_type(fcinfo, NULL, &expected_tupledesc) != TYPEFUNC_COMPOSITE)
394 0 : elog(ERROR, "return type must be a row type");
395 :
396 2 : if (expected_tupledesc->natts != NUM_BUFFERCACHE_OS_PAGES_ELEM)
397 0 : elog(ERROR, "incorrect number of output arguments");
398 :
399 : /* Construct a tuple descriptor for the result rows. */
400 2 : tupledesc = CreateTemplateTupleDesc(expected_tupledesc->natts);
401 2 : TupleDescInitEntry(tupledesc, (AttrNumber) 1, "bufferid",
402 : INT4OID, -1, 0);
403 2 : TupleDescInitEntry(tupledesc, (AttrNumber) 2, "os_page_num",
404 : INT8OID, -1, 0);
405 2 : TupleDescInitEntry(tupledesc, (AttrNumber) 3, "numa_node",
406 : INT4OID, -1, 0);
407 :
408 2 : TupleDescFinalize(tupledesc);
409 2 : fctx->tupdesc = BlessTupleDesc(tupledesc);
410 2 : fctx->include_numa = include_numa;
411 :
412 : /*
413 : * Each buffer needs at least one entry, but it might be offset in
414 : * some way, and use one extra entry. So we allocate space for the
415 : * maximum number of entries we might need, and then count the exact
416 : * number as we're walking buffers. That way we can do it in one pass,
417 : * without reallocating memory.
418 : */
419 2 : pages_per_buffer = Max(1, BLCKSZ / os_page_size) + 1;
420 2 : max_entries = NBuffers * pages_per_buffer;
421 :
422 : /* Allocate entries for BufferCacheOsPagesRec records. */
423 2 : fctx->record = (BufferCacheOsPagesRec *)
424 2 : MemoryContextAllocHuge(CurrentMemoryContext,
425 : sizeof(BufferCacheOsPagesRec) * max_entries);
426 :
427 : /* Return to original context when allocating transient memory */
428 2 : MemoryContextSwitchTo(oldcontext);
429 :
430 2 : if (include_numa && firstNumaTouch)
431 0 : elog(DEBUG1, "NUMA: page-faulting the buffercache for proper NUMA readouts");
432 :
433 : /*
434 : * Scan through all the buffers, saving the relevant fields in the
435 : * fctx->record structure.
436 : *
437 : * We don't hold the partition locks, so we don't get a consistent
438 : * snapshot across all buffers, but we do grab the buffer header
439 : * locks, so the information of each buffer is self-consistent.
440 : */
441 2 : startptr = (char *) TYPEALIGN_DOWN(os_page_size, (char *) BufferGetBlock(1));
442 2 : idx = 0;
443 32770 : for (i = 0; i < NBuffers; i++)
444 : {
445 32768 : char *buffptr = (char *) BufferGetBlock(i + 1);
446 : BufferDesc *bufHdr;
447 : uint32 bufferid;
448 : int32 page_num;
449 : char *startptr_buff,
450 : *endptr_buff;
451 :
452 32768 : CHECK_FOR_INTERRUPTS();
453 :
454 32768 : bufHdr = GetBufferDescriptor(i);
455 :
456 : /* Lock each buffer header before inspecting. */
457 32768 : LockBufHdr(bufHdr);
458 32768 : bufferid = BufferDescriptorGetBuffer(bufHdr);
459 32768 : UnlockBufHdr(bufHdr);
460 :
461 : /* start of the first page of this buffer */
462 32768 : startptr_buff = (char *) TYPEALIGN_DOWN(os_page_size, buffptr);
463 :
464 : /* end of the buffer (no need to align to memory page) */
465 32768 : endptr_buff = buffptr + BLCKSZ;
466 :
467 : Assert(startptr_buff < endptr_buff);
468 :
469 : /* calculate ID of the first page for this buffer */
470 32768 : page_num = (startptr_buff - startptr) / os_page_size;
471 :
472 : /* Add an entry for each OS page overlapping with this buffer. */
473 98304 : for (char *ptr = startptr_buff; ptr < endptr_buff; ptr += os_page_size)
474 : {
475 65536 : fctx->record[idx].bufferid = bufferid;
476 65536 : fctx->record[idx].page_num = page_num;
477 65536 : fctx->record[idx].numa_node = include_numa ? os_page_status[page_num] : -1;
478 :
479 : /* advance to the next entry/page */
480 65536 : ++idx;
481 65536 : ++page_num;
482 : }
483 : }
484 :
485 : Assert(idx <= max_entries);
486 :
487 : if (include_numa)
488 : Assert(idx >= os_page_count);
489 :
490 : /* Set max calls and remember the user function context. */
491 2 : funcctx->max_calls = idx;
492 2 : funcctx->user_fctx = fctx;
493 :
494 : /* Remember this backend touched the pages (only relevant for NUMA) */
495 2 : if (include_numa)
496 0 : firstNumaTouch = false;
497 : }
498 :
499 65538 : funcctx = SRF_PERCALL_SETUP();
500 :
501 : /* Get the saved state */
502 65538 : fctx = funcctx->user_fctx;
503 :
504 65538 : if (funcctx->call_cntr < funcctx->max_calls)
505 : {
506 65536 : uint32 i = funcctx->call_cntr;
507 : Datum values[NUM_BUFFERCACHE_OS_PAGES_ELEM];
508 : bool nulls[NUM_BUFFERCACHE_OS_PAGES_ELEM];
509 :
510 65536 : values[0] = Int32GetDatum(fctx->record[i].bufferid);
511 65536 : nulls[0] = false;
512 :
513 65536 : values[1] = Int64GetDatum(fctx->record[i].page_num);
514 65536 : nulls[1] = false;
515 :
516 65536 : if (fctx->include_numa)
517 : {
518 : /* status is valid node number */
519 0 : if (fctx->record[i].numa_node >= 0)
520 : {
521 0 : values[2] = Int32GetDatum(fctx->record[i].numa_node);
522 0 : nulls[2] = false;
523 : }
524 : else
525 : {
526 : /* some kind of error (e.g. pages moved to swap) */
527 0 : values[2] = (Datum) 0;
528 0 : nulls[2] = true;
529 : }
530 : }
531 : else
532 : {
533 65536 : values[2] = (Datum) 0;
534 65536 : nulls[2] = true;
535 : }
536 :
537 : /* Build and return the tuple. */
538 65536 : tuple = heap_form_tuple(fctx->tupdesc, values, nulls);
539 65536 : result = HeapTupleGetDatum(tuple);
540 :
541 65536 : SRF_RETURN_NEXT(funcctx, result);
542 : }
543 : else
544 2 : SRF_RETURN_DONE(funcctx);
545 : }
546 :
547 : /*
548 : * pg_buffercache_os_pages
549 : *
550 : * Retrieve information about OS pages, with or without NUMA information.
551 : */
552 : Datum
553 65538 : pg_buffercache_os_pages(PG_FUNCTION_ARGS)
554 : {
555 : bool include_numa;
556 :
557 : /* Get the boolean parameter that controls the NUMA behavior. */
558 65538 : include_numa = PG_GETARG_BOOL(0);
559 :
560 65538 : return pg_buffercache_os_pages_internal(fcinfo, include_numa);
561 : }
562 :
563 : /* Backward-compatible wrapper for v1.6. */
564 : Datum
565 0 : pg_buffercache_numa_pages(PG_FUNCTION_ARGS)
566 : {
567 : /* Call internal function with include_numa=true */
568 0 : return pg_buffercache_os_pages_internal(fcinfo, true);
569 : }
570 :
571 : Datum
572 2 : pg_buffercache_summary(PG_FUNCTION_ARGS)
573 : {
574 : Datum result;
575 : TupleDesc tupledesc;
576 : HeapTuple tuple;
577 : Datum values[NUM_BUFFERCACHE_SUMMARY_ELEM];
578 : bool nulls[NUM_BUFFERCACHE_SUMMARY_ELEM];
579 :
580 2 : int32 buffers_used = 0;
581 2 : int32 buffers_unused = 0;
582 2 : int32 buffers_dirty = 0;
583 2 : int32 buffers_pinned = 0;
584 2 : int64 usagecount_total = 0;
585 :
586 2 : if (get_call_result_type(fcinfo, NULL, &tupledesc) != TYPEFUNC_COMPOSITE)
587 0 : elog(ERROR, "return type must be a row type");
588 :
589 32770 : for (int i = 0; i < NBuffers; i++)
590 : {
591 : BufferDesc *bufHdr;
592 : uint64 buf_state;
593 :
594 32768 : CHECK_FOR_INTERRUPTS();
595 :
596 : /*
597 : * This function summarizes the state of all headers. Locking the
598 : * buffer headers wouldn't provide an improved result as the state of
599 : * the buffer can still change after we release the lock and it'd
600 : * noticeably increase the cost of the function.
601 : */
602 32768 : bufHdr = GetBufferDescriptor(i);
603 32768 : buf_state = pg_atomic_read_u64(&bufHdr->state);
604 :
605 32768 : if (buf_state & BM_VALID)
606 : {
607 4046 : buffers_used++;
608 4046 : usagecount_total += BUF_STATE_GET_USAGECOUNT(buf_state);
609 :
610 4046 : if (buf_state & BM_DIRTY)
611 1950 : buffers_dirty++;
612 : }
613 : else
614 28722 : buffers_unused++;
615 :
616 32768 : if (BUF_STATE_GET_REFCOUNT(buf_state) > 0)
617 0 : buffers_pinned++;
618 : }
619 :
620 2 : memset(nulls, 0, sizeof(nulls));
621 2 : values[0] = Int32GetDatum(buffers_used);
622 2 : values[1] = Int32GetDatum(buffers_unused);
623 2 : values[2] = Int32GetDatum(buffers_dirty);
624 2 : values[3] = Int32GetDatum(buffers_pinned);
625 :
626 2 : if (buffers_used != 0)
627 2 : values[4] = Float8GetDatum((double) usagecount_total / buffers_used);
628 : else
629 0 : nulls[4] = true;
630 :
631 : /* Build and return the tuple. */
632 2 : tuple = heap_form_tuple(tupledesc, values, nulls);
633 2 : result = HeapTupleGetDatum(tuple);
634 :
635 2 : PG_RETURN_DATUM(result);
636 : }
637 :
638 : Datum
639 2 : pg_buffercache_usage_counts(PG_FUNCTION_ARGS)
640 : {
641 2 : ReturnSetInfo *rsinfo = (ReturnSetInfo *) fcinfo->resultinfo;
642 2 : int usage_counts[BM_MAX_USAGE_COUNT + 1] = {0};
643 2 : int dirty[BM_MAX_USAGE_COUNT + 1] = {0};
644 2 : int pinned[BM_MAX_USAGE_COUNT + 1] = {0};
645 : Datum values[NUM_BUFFERCACHE_USAGE_COUNTS_ELEM];
646 2 : bool nulls[NUM_BUFFERCACHE_USAGE_COUNTS_ELEM] = {0};
647 :
648 2 : InitMaterializedSRF(fcinfo, 0);
649 :
650 32770 : for (int i = 0; i < NBuffers; i++)
651 : {
652 32768 : BufferDesc *bufHdr = GetBufferDescriptor(i);
653 32768 : uint64 buf_state = pg_atomic_read_u64(&bufHdr->state);
654 : int usage_count;
655 :
656 32768 : CHECK_FOR_INTERRUPTS();
657 :
658 32768 : usage_count = BUF_STATE_GET_USAGECOUNT(buf_state);
659 32768 : usage_counts[usage_count]++;
660 :
661 32768 : if (buf_state & BM_DIRTY)
662 1950 : dirty[usage_count]++;
663 :
664 32768 : if (BUF_STATE_GET_REFCOUNT(buf_state) > 0)
665 0 : pinned[usage_count]++;
666 : }
667 :
668 14 : for (int i = 0; i < BM_MAX_USAGE_COUNT + 1; i++)
669 : {
670 12 : values[0] = Int32GetDatum(i);
671 12 : values[1] = Int32GetDatum(usage_counts[i]);
672 12 : values[2] = Int32GetDatum(dirty[i]);
673 12 : values[3] = Int32GetDatum(pinned[i]);
674 :
675 12 : tuplestore_putvalues(rsinfo->setResult, rsinfo->setDesc, values, nulls);
676 : }
677 :
678 2 : return (Datum) 0;
679 : }
680 :
681 : /*
682 : * Helper function to check if the user has superuser privileges.
683 : */
684 : static void
685 20 : pg_buffercache_superuser_check(char *func_name)
686 : {
687 20 : if (!superuser())
688 6 : ereport(ERROR,
689 : (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
690 : errmsg("must be superuser to use %s()",
691 : func_name)));
692 14 : }
693 :
694 : /*
695 : * Try to evict a shared buffer.
696 : */
697 : Datum
698 5 : pg_buffercache_evict(PG_FUNCTION_ARGS)
699 : {
700 : Datum result;
701 : TupleDesc tupledesc;
702 : HeapTuple tuple;
703 : Datum values[NUM_BUFFERCACHE_EVICT_ELEM];
704 5 : bool nulls[NUM_BUFFERCACHE_EVICT_ELEM] = {0};
705 :
706 5 : Buffer buf = PG_GETARG_INT32(0);
707 : bool buffer_flushed;
708 :
709 5 : if (get_call_result_type(fcinfo, NULL, &tupledesc) != TYPEFUNC_COMPOSITE)
710 0 : elog(ERROR, "return type must be a row type");
711 :
712 5 : pg_buffercache_superuser_check("pg_buffercache_evict");
713 :
714 4 : if (buf < 1 || buf > NBuffers)
715 3 : elog(ERROR, "bad buffer ID: %d", buf);
716 :
717 1 : values[0] = BoolGetDatum(EvictUnpinnedBuffer(buf, &buffer_flushed));
718 1 : values[1] = BoolGetDatum(buffer_flushed);
719 :
720 1 : tuple = heap_form_tuple(tupledesc, values, nulls);
721 1 : result = HeapTupleGetDatum(tuple);
722 :
723 1 : PG_RETURN_DATUM(result);
724 : }
725 :
726 : /*
727 : * Try to evict specified relation.
728 : */
729 : Datum
730 3 : pg_buffercache_evict_relation(PG_FUNCTION_ARGS)
731 : {
732 : Datum result;
733 : TupleDesc tupledesc;
734 : HeapTuple tuple;
735 : Datum values[NUM_BUFFERCACHE_EVICT_RELATION_ELEM];
736 3 : bool nulls[NUM_BUFFERCACHE_EVICT_RELATION_ELEM] = {0};
737 :
738 : Oid relOid;
739 : Relation rel;
740 :
741 3 : int32 buffers_evicted = 0;
742 3 : int32 buffers_flushed = 0;
743 3 : int32 buffers_skipped = 0;
744 :
745 3 : if (get_call_result_type(fcinfo, NULL, &tupledesc) != TYPEFUNC_COMPOSITE)
746 0 : elog(ERROR, "return type must be a row type");
747 :
748 3 : pg_buffercache_superuser_check("pg_buffercache_evict_relation");
749 :
750 2 : relOid = PG_GETARG_OID(0);
751 :
752 2 : rel = relation_open(relOid, AccessShareLock);
753 :
754 2 : if (RelationUsesLocalBuffers(rel))
755 1 : ereport(ERROR,
756 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
757 : errmsg("relation uses local buffers, %s() is intended to be used for shared buffers only",
758 : "pg_buffercache_evict_relation")));
759 :
760 1 : EvictRelUnpinnedBuffers(rel, &buffers_evicted, &buffers_flushed,
761 : &buffers_skipped);
762 :
763 1 : relation_close(rel, AccessShareLock);
764 :
765 1 : values[0] = Int32GetDatum(buffers_evicted);
766 1 : values[1] = Int32GetDatum(buffers_flushed);
767 1 : values[2] = Int32GetDatum(buffers_skipped);
768 :
769 1 : tuple = heap_form_tuple(tupledesc, values, nulls);
770 1 : result = HeapTupleGetDatum(tuple);
771 :
772 1 : PG_RETURN_DATUM(result);
773 : }
774 :
775 :
776 : /*
777 : * Try to evict all shared buffers.
778 : */
779 : Datum
780 2 : pg_buffercache_evict_all(PG_FUNCTION_ARGS)
781 : {
782 : Datum result;
783 : TupleDesc tupledesc;
784 : HeapTuple tuple;
785 : Datum values[NUM_BUFFERCACHE_EVICT_ALL_ELEM];
786 2 : bool nulls[NUM_BUFFERCACHE_EVICT_ALL_ELEM] = {0};
787 :
788 2 : int32 buffers_evicted = 0;
789 2 : int32 buffers_flushed = 0;
790 2 : int32 buffers_skipped = 0;
791 :
792 2 : if (get_call_result_type(fcinfo, NULL, &tupledesc) != TYPEFUNC_COMPOSITE)
793 0 : elog(ERROR, "return type must be a row type");
794 :
795 2 : pg_buffercache_superuser_check("pg_buffercache_evict_all");
796 :
797 1 : EvictAllUnpinnedBuffers(&buffers_evicted, &buffers_flushed,
798 : &buffers_skipped);
799 :
800 1 : values[0] = Int32GetDatum(buffers_evicted);
801 1 : values[1] = Int32GetDatum(buffers_flushed);
802 1 : values[2] = Int32GetDatum(buffers_skipped);
803 :
804 1 : tuple = heap_form_tuple(tupledesc, values, nulls);
805 1 : result = HeapTupleGetDatum(tuple);
806 :
807 1 : PG_RETURN_DATUM(result);
808 : }
809 :
810 : /*
811 : * Try to mark a shared buffer as dirty.
812 : */
813 : Datum
814 5 : pg_buffercache_mark_dirty(PG_FUNCTION_ARGS)
815 : {
816 :
817 : Datum result;
818 : TupleDesc tupledesc;
819 : HeapTuple tuple;
820 : Datum values[NUM_BUFFERCACHE_MARK_DIRTY_ELEM];
821 5 : bool nulls[NUM_BUFFERCACHE_MARK_DIRTY_ELEM] = {0};
822 :
823 5 : Buffer buf = PG_GETARG_INT32(0);
824 : bool buffer_already_dirty;
825 :
826 5 : if (get_call_result_type(fcinfo, NULL, &tupledesc) != TYPEFUNC_COMPOSITE)
827 0 : elog(ERROR, "return type must be a row type");
828 :
829 5 : pg_buffercache_superuser_check("pg_buffercache_mark_dirty");
830 :
831 4 : if (buf < 1 || buf > NBuffers)
832 3 : elog(ERROR, "bad buffer ID: %d", buf);
833 :
834 1 : values[0] = BoolGetDatum(MarkDirtyUnpinnedBuffer(buf, &buffer_already_dirty));
835 1 : values[1] = BoolGetDatum(buffer_already_dirty);
836 :
837 1 : tuple = heap_form_tuple(tupledesc, values, nulls);
838 1 : result = HeapTupleGetDatum(tuple);
839 :
840 1 : PG_RETURN_DATUM(result);
841 : }
842 :
843 : /*
844 : * Try to mark all the shared buffers of a relation as dirty.
845 : */
846 : Datum
847 3 : pg_buffercache_mark_dirty_relation(PG_FUNCTION_ARGS)
848 : {
849 : Datum result;
850 : TupleDesc tupledesc;
851 : HeapTuple tuple;
852 : Datum values[NUM_BUFFERCACHE_MARK_DIRTY_RELATION_ELEM];
853 3 : bool nulls[NUM_BUFFERCACHE_MARK_DIRTY_RELATION_ELEM] = {0};
854 :
855 : Oid relOid;
856 : Relation rel;
857 :
858 3 : int32 buffers_already_dirty = 0;
859 3 : int32 buffers_dirtied = 0;
860 3 : int32 buffers_skipped = 0;
861 :
862 3 : if (get_call_result_type(fcinfo, NULL, &tupledesc) != TYPEFUNC_COMPOSITE)
863 0 : elog(ERROR, "return type must be a row type");
864 :
865 3 : pg_buffercache_superuser_check("pg_buffercache_mark_dirty_relation");
866 :
867 2 : relOid = PG_GETARG_OID(0);
868 :
869 2 : rel = relation_open(relOid, AccessShareLock);
870 :
871 2 : if (RelationUsesLocalBuffers(rel))
872 1 : ereport(ERROR,
873 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
874 : errmsg("relation uses local buffers, %s() is intended to be used for shared buffers only",
875 : "pg_buffercache_mark_dirty_relation")));
876 :
877 1 : MarkDirtyRelUnpinnedBuffers(rel, &buffers_dirtied, &buffers_already_dirty,
878 : &buffers_skipped);
879 :
880 1 : relation_close(rel, AccessShareLock);
881 :
882 1 : values[0] = Int32GetDatum(buffers_dirtied);
883 1 : values[1] = Int32GetDatum(buffers_already_dirty);
884 1 : values[2] = Int32GetDatum(buffers_skipped);
885 :
886 1 : tuple = heap_form_tuple(tupledesc, values, nulls);
887 1 : result = HeapTupleGetDatum(tuple);
888 :
889 1 : PG_RETURN_DATUM(result);
890 : }
891 :
892 : /*
893 : * Try to mark all the shared buffers as dirty.
894 : */
895 : Datum
896 2 : pg_buffercache_mark_dirty_all(PG_FUNCTION_ARGS)
897 : {
898 : Datum result;
899 : TupleDesc tupledesc;
900 : HeapTuple tuple;
901 : Datum values[NUM_BUFFERCACHE_MARK_DIRTY_ALL_ELEM];
902 2 : bool nulls[NUM_BUFFERCACHE_MARK_DIRTY_ALL_ELEM] = {0};
903 :
904 2 : int32 buffers_already_dirty = 0;
905 2 : int32 buffers_dirtied = 0;
906 2 : int32 buffers_skipped = 0;
907 :
908 2 : if (get_call_result_type(fcinfo, NULL, &tupledesc) != TYPEFUNC_COMPOSITE)
909 0 : elog(ERROR, "return type must be a row type");
910 :
911 2 : pg_buffercache_superuser_check("pg_buffercache_mark_dirty_all");
912 :
913 1 : MarkDirtyAllUnpinnedBuffers(&buffers_dirtied, &buffers_already_dirty,
914 : &buffers_skipped);
915 :
916 1 : values[0] = Int32GetDatum(buffers_dirtied);
917 1 : values[1] = Int32GetDatum(buffers_already_dirty);
918 1 : values[2] = Int32GetDatum(buffers_skipped);
919 :
920 1 : tuple = heap_form_tuple(tupledesc, values, nulls);
921 1 : result = HeapTupleGetDatum(tuple);
922 :
923 1 : PG_RETURN_DATUM(result);
924 : }
|
