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 :
20 :
21 : #define NUM_BUFFERCACHE_PAGES_MIN_ELEM 8
22 : #define NUM_BUFFERCACHE_PAGES_ELEM 9
23 : #define NUM_BUFFERCACHE_SUMMARY_ELEM 5
24 : #define NUM_BUFFERCACHE_USAGE_COUNTS_ELEM 4
25 : #define NUM_BUFFERCACHE_EVICT_ELEM 2
26 : #define NUM_BUFFERCACHE_EVICT_RELATION_ELEM 3
27 : #define NUM_BUFFERCACHE_EVICT_ALL_ELEM 3
28 :
29 : #define NUM_BUFFERCACHE_NUMA_ELEM 3
30 :
31 2 : PG_MODULE_MAGIC_EXT(
32 : .name = "pg_buffercache",
33 : .version = PG_VERSION
34 : );
35 :
36 : /*
37 : * Record structure holding the to be exposed cache data.
38 : */
39 : typedef struct
40 : {
41 : uint32 bufferid;
42 : RelFileNumber relfilenumber;
43 : Oid reltablespace;
44 : Oid reldatabase;
45 : ForkNumber forknum;
46 : BlockNumber blocknum;
47 : bool isvalid;
48 : bool isdirty;
49 : uint16 usagecount;
50 :
51 : /*
52 : * An int32 is sufficiently large, as MAX_BACKENDS prevents a buffer from
53 : * being pinned by too many backends and each backend will only pin once
54 : * because of bufmgr.c's PrivateRefCount infrastructure.
55 : */
56 : int32 pinning_backends;
57 : } BufferCachePagesRec;
58 :
59 :
60 : /*
61 : * Function context for data persisting over repeated calls.
62 : */
63 : typedef struct
64 : {
65 : TupleDesc tupdesc;
66 : BufferCachePagesRec *record;
67 : } BufferCachePagesContext;
68 :
69 : /*
70 : * Record structure holding the to be exposed cache data.
71 : */
72 : typedef struct
73 : {
74 : uint32 bufferid;
75 : int64 page_num;
76 : int32 numa_node;
77 : } BufferCacheNumaRec;
78 :
79 : /*
80 : * Function context for data persisting over repeated calls.
81 : */
82 : typedef struct
83 : {
84 : TupleDesc tupdesc;
85 : int buffers_per_page;
86 : int pages_per_buffer;
87 : int os_page_size;
88 : BufferCacheNumaRec *record;
89 : } BufferCacheNumaContext;
90 :
91 :
92 : /*
93 : * Function returning data from the shared buffer cache - buffer number,
94 : * relation node/tablespace/database/blocknum and dirty indicator.
95 : */
96 4 : PG_FUNCTION_INFO_V1(pg_buffercache_pages);
97 2 : PG_FUNCTION_INFO_V1(pg_buffercache_numa_pages);
98 4 : PG_FUNCTION_INFO_V1(pg_buffercache_summary);
99 4 : PG_FUNCTION_INFO_V1(pg_buffercache_usage_counts);
100 6 : PG_FUNCTION_INFO_V1(pg_buffercache_evict);
101 4 : PG_FUNCTION_INFO_V1(pg_buffercache_evict_relation);
102 4 : PG_FUNCTION_INFO_V1(pg_buffercache_evict_all);
103 :
104 :
105 : /* Only need to touch memory once per backend process lifetime */
106 : static bool firstNumaTouch = true;
107 :
108 :
109 : Datum
110 65540 : pg_buffercache_pages(PG_FUNCTION_ARGS)
111 : {
112 : FuncCallContext *funcctx;
113 : Datum result;
114 : MemoryContext oldcontext;
115 : BufferCachePagesContext *fctx; /* User function context. */
116 : TupleDesc tupledesc;
117 : TupleDesc expected_tupledesc;
118 : HeapTuple tuple;
119 :
120 65540 : if (SRF_IS_FIRSTCALL())
121 : {
122 : int i;
123 :
124 4 : funcctx = SRF_FIRSTCALL_INIT();
125 :
126 : /* Switch context when allocating stuff to be used in later calls */
127 4 : oldcontext = MemoryContextSwitchTo(funcctx->multi_call_memory_ctx);
128 :
129 : /* Create a user function context for cross-call persistence */
130 4 : fctx = (BufferCachePagesContext *) palloc(sizeof(BufferCachePagesContext));
131 :
132 : /*
133 : * To smoothly support upgrades from version 1.0 of this extension
134 : * transparently handle the (non-)existence of the pinning_backends
135 : * column. We unfortunately have to get the result type for that... -
136 : * we can't use the result type determined by the function definition
137 : * without potentially crashing when somebody uses the old (or even
138 : * wrong) function definition though.
139 : */
140 4 : if (get_call_result_type(fcinfo, NULL, &expected_tupledesc) != TYPEFUNC_COMPOSITE)
141 0 : elog(ERROR, "return type must be a row type");
142 :
143 4 : if (expected_tupledesc->natts < NUM_BUFFERCACHE_PAGES_MIN_ELEM ||
144 4 : expected_tupledesc->natts > NUM_BUFFERCACHE_PAGES_ELEM)
145 0 : elog(ERROR, "incorrect number of output arguments");
146 :
147 : /* Construct a tuple descriptor for the result rows. */
148 4 : tupledesc = CreateTemplateTupleDesc(expected_tupledesc->natts);
149 4 : TupleDescInitEntry(tupledesc, (AttrNumber) 1, "bufferid",
150 : INT4OID, -1, 0);
151 4 : TupleDescInitEntry(tupledesc, (AttrNumber) 2, "relfilenode",
152 : OIDOID, -1, 0);
153 4 : TupleDescInitEntry(tupledesc, (AttrNumber) 3, "reltablespace",
154 : OIDOID, -1, 0);
155 4 : TupleDescInitEntry(tupledesc, (AttrNumber) 4, "reldatabase",
156 : OIDOID, -1, 0);
157 4 : TupleDescInitEntry(tupledesc, (AttrNumber) 5, "relforknumber",
158 : INT2OID, -1, 0);
159 4 : TupleDescInitEntry(tupledesc, (AttrNumber) 6, "relblocknumber",
160 : INT8OID, -1, 0);
161 4 : TupleDescInitEntry(tupledesc, (AttrNumber) 7, "isdirty",
162 : BOOLOID, -1, 0);
163 4 : TupleDescInitEntry(tupledesc, (AttrNumber) 8, "usage_count",
164 : INT2OID, -1, 0);
165 :
166 4 : if (expected_tupledesc->natts == NUM_BUFFERCACHE_PAGES_ELEM)
167 4 : TupleDescInitEntry(tupledesc, (AttrNumber) 9, "pinning_backends",
168 : INT4OID, -1, 0);
169 :
170 4 : fctx->tupdesc = BlessTupleDesc(tupledesc);
171 :
172 : /* Allocate NBuffers worth of BufferCachePagesRec records. */
173 4 : fctx->record = (BufferCachePagesRec *)
174 4 : MemoryContextAllocHuge(CurrentMemoryContext,
175 : sizeof(BufferCachePagesRec) * NBuffers);
176 :
177 : /* Set max calls and remember the user function context. */
178 4 : funcctx->max_calls = NBuffers;
179 4 : funcctx->user_fctx = fctx;
180 :
181 : /* Return to original context when allocating transient memory */
182 4 : MemoryContextSwitchTo(oldcontext);
183 :
184 : /*
185 : * Scan through all the buffers, saving the relevant fields in the
186 : * fctx->record structure.
187 : *
188 : * We don't hold the partition locks, so we don't get a consistent
189 : * snapshot across all buffers, but we do grab the buffer header
190 : * locks, so the information of each buffer is self-consistent.
191 : */
192 65540 : for (i = 0; i < NBuffers; i++)
193 : {
194 : BufferDesc *bufHdr;
195 : uint32 buf_state;
196 :
197 65536 : bufHdr = GetBufferDescriptor(i);
198 : /* Lock each buffer header before inspecting. */
199 65536 : buf_state = LockBufHdr(bufHdr);
200 :
201 65536 : fctx->record[i].bufferid = BufferDescriptorGetBuffer(bufHdr);
202 65536 : fctx->record[i].relfilenumber = BufTagGetRelNumber(&bufHdr->tag);
203 65536 : fctx->record[i].reltablespace = bufHdr->tag.spcOid;
204 65536 : fctx->record[i].reldatabase = bufHdr->tag.dbOid;
205 65536 : fctx->record[i].forknum = BufTagGetForkNum(&bufHdr->tag);
206 65536 : fctx->record[i].blocknum = bufHdr->tag.blockNum;
207 65536 : fctx->record[i].usagecount = BUF_STATE_GET_USAGECOUNT(buf_state);
208 65536 : fctx->record[i].pinning_backends = BUF_STATE_GET_REFCOUNT(buf_state);
209 :
210 65536 : if (buf_state & BM_DIRTY)
211 3780 : fctx->record[i].isdirty = true;
212 : else
213 61756 : fctx->record[i].isdirty = false;
214 :
215 : /* Note if the buffer is valid, and has storage created */
216 65536 : if ((buf_state & BM_VALID) && (buf_state & BM_TAG_VALID))
217 7960 : fctx->record[i].isvalid = true;
218 : else
219 57576 : fctx->record[i].isvalid = false;
220 :
221 65536 : UnlockBufHdr(bufHdr, buf_state);
222 : }
223 : }
224 :
225 65540 : funcctx = SRF_PERCALL_SETUP();
226 :
227 : /* Get the saved state */
228 65540 : fctx = funcctx->user_fctx;
229 :
230 65540 : if (funcctx->call_cntr < funcctx->max_calls)
231 : {
232 65536 : uint32 i = funcctx->call_cntr;
233 : Datum values[NUM_BUFFERCACHE_PAGES_ELEM];
234 : bool nulls[NUM_BUFFERCACHE_PAGES_ELEM];
235 :
236 65536 : values[0] = Int32GetDatum(fctx->record[i].bufferid);
237 65536 : nulls[0] = false;
238 :
239 : /*
240 : * Set all fields except the bufferid to null if the buffer is unused
241 : * or not valid.
242 : */
243 65536 : if (fctx->record[i].blocknum == InvalidBlockNumber ||
244 7960 : fctx->record[i].isvalid == false)
245 : {
246 57576 : nulls[1] = true;
247 57576 : nulls[2] = true;
248 57576 : nulls[3] = true;
249 57576 : nulls[4] = true;
250 57576 : nulls[5] = true;
251 57576 : nulls[6] = true;
252 57576 : nulls[7] = true;
253 : /* unused for v1.0 callers, but the array is always long enough */
254 57576 : nulls[8] = true;
255 : }
256 : else
257 : {
258 7960 : values[1] = ObjectIdGetDatum(fctx->record[i].relfilenumber);
259 7960 : nulls[1] = false;
260 7960 : values[2] = ObjectIdGetDatum(fctx->record[i].reltablespace);
261 7960 : nulls[2] = false;
262 7960 : values[3] = ObjectIdGetDatum(fctx->record[i].reldatabase);
263 7960 : nulls[3] = false;
264 7960 : values[4] = ObjectIdGetDatum(fctx->record[i].forknum);
265 7960 : nulls[4] = false;
266 7960 : values[5] = Int64GetDatum((int64) fctx->record[i].blocknum);
267 7960 : nulls[5] = false;
268 7960 : values[6] = BoolGetDatum(fctx->record[i].isdirty);
269 7960 : nulls[6] = false;
270 7960 : values[7] = Int16GetDatum(fctx->record[i].usagecount);
271 7960 : nulls[7] = false;
272 : /* unused for v1.0 callers, but the array is always long enough */
273 7960 : values[8] = Int32GetDatum(fctx->record[i].pinning_backends);
274 7960 : nulls[8] = false;
275 : }
276 :
277 : /* Build and return the tuple. */
278 65536 : tuple = heap_form_tuple(fctx->tupdesc, values, nulls);
279 65536 : result = HeapTupleGetDatum(tuple);
280 :
281 65536 : SRF_RETURN_NEXT(funcctx, result);
282 : }
283 : else
284 4 : SRF_RETURN_DONE(funcctx);
285 : }
286 :
287 : /*
288 : * Inquire about NUMA memory mappings for shared buffers.
289 : *
290 : * Returns NUMA node ID for each memory page used by the buffer. Buffers may
291 : * be smaller or larger than OS memory pages. For each buffer we return one
292 : * entry for each memory page used by the buffer (if the buffer is smaller,
293 : * it only uses a part of one memory page).
294 : *
295 : * We expect both sizes (for buffers and memory pages) to be a power-of-2, so
296 : * one is always a multiple of the other.
297 : *
298 : * In order to get reliable results we also need to touch memory pages, so
299 : * that the inquiry about NUMA memory node doesn't return -2 (which indicates
300 : * unmapped/unallocated pages).
301 : */
302 : Datum
303 0 : pg_buffercache_numa_pages(PG_FUNCTION_ARGS)
304 : {
305 : FuncCallContext *funcctx;
306 : MemoryContext oldcontext;
307 : BufferCacheNumaContext *fctx; /* User function context. */
308 : TupleDesc tupledesc;
309 : TupleDesc expected_tupledesc;
310 : HeapTuple tuple;
311 : Datum result;
312 :
313 0 : if (SRF_IS_FIRSTCALL())
314 : {
315 : int i,
316 : idx;
317 : Size os_page_size;
318 : void **os_page_ptrs;
319 : int *os_page_status;
320 : uint64 os_page_count;
321 : int pages_per_buffer;
322 : int max_entries;
323 : volatile uint64 touch pg_attribute_unused();
324 : char *startptr,
325 : *endptr;
326 :
327 0 : if (pg_numa_init() == -1)
328 0 : elog(ERROR, "libnuma initialization failed or NUMA is not supported on this platform");
329 :
330 : /*
331 : * The database block size and OS memory page size are unlikely to be
332 : * the same. The block size is 1-32KB, the memory page size depends on
333 : * platform. On x86 it's usually 4KB, on ARM it's 4KB or 64KB, but
334 : * there are also features like THP etc. Moreover, we don't quite know
335 : * how the pages and buffers "align" in memory - the buffers may be
336 : * shifted in some way, using more memory pages than necessary.
337 : *
338 : * So we need to be careful about mapping buffers to memory pages. We
339 : * calculate the maximum number of pages a buffer might use, so that
340 : * we allocate enough space for the entries. And then we count the
341 : * actual number of entries as we scan the buffers.
342 : *
343 : * This information is needed before calling move_pages() for NUMA
344 : * node id inquiry.
345 : */
346 0 : os_page_size = pg_get_shmem_pagesize();
347 :
348 : /*
349 : * The pages and block size is expected to be 2^k, so one divides the
350 : * other (we don't know in which direction). This does not say
351 : * anything about relative alignment of pages/buffers.
352 : */
353 : Assert((os_page_size % BLCKSZ == 0) || (BLCKSZ % os_page_size == 0));
354 :
355 : /*
356 : * How many addresses we are going to query? Simply get the page for
357 : * the first buffer, and first page after the last buffer, and count
358 : * the pages from that.
359 : */
360 0 : startptr = (char *) TYPEALIGN_DOWN(os_page_size,
361 : BufferGetBlock(1));
362 0 : endptr = (char *) TYPEALIGN(os_page_size,
363 : (char *) BufferGetBlock(NBuffers) + BLCKSZ);
364 0 : os_page_count = (endptr - startptr) / os_page_size;
365 :
366 : /* Used to determine the NUMA node for all OS pages at once */
367 0 : os_page_ptrs = palloc0(sizeof(void *) * os_page_count);
368 0 : os_page_status = palloc(sizeof(uint64) * os_page_count);
369 :
370 : /* Fill pointers for all the memory pages. */
371 0 : idx = 0;
372 0 : for (char *ptr = startptr; ptr < endptr; ptr += os_page_size)
373 : {
374 0 : os_page_ptrs[idx++] = ptr;
375 :
376 : /* Only need to touch memory once per backend process lifetime */
377 0 : if (firstNumaTouch)
378 : pg_numa_touch_mem_if_required(touch, ptr);
379 : }
380 :
381 : Assert(idx == os_page_count);
382 :
383 0 : elog(DEBUG1, "NUMA: NBuffers=%d os_page_count=" UINT64_FORMAT " "
384 : "os_page_size=%zu", NBuffers, os_page_count, os_page_size);
385 :
386 : /*
387 : * If we ever get 0xff back from kernel inquiry, then we probably have
388 : * bug in our buffers to OS page mapping code here.
389 : */
390 0 : memset(os_page_status, 0xff, sizeof(int) * os_page_count);
391 :
392 : /* Query NUMA status for all the pointers */
393 0 : if (pg_numa_query_pages(0, os_page_count, os_page_ptrs, os_page_status) == -1)
394 0 : elog(ERROR, "failed NUMA pages inquiry: %m");
395 :
396 : /* Initialize the multi-call context, load entries about buffers */
397 :
398 0 : funcctx = SRF_FIRSTCALL_INIT();
399 :
400 : /* Switch context when allocating stuff to be used in later calls */
401 0 : oldcontext = MemoryContextSwitchTo(funcctx->multi_call_memory_ctx);
402 :
403 : /* Create a user function context for cross-call persistence */
404 0 : fctx = (BufferCacheNumaContext *) palloc(sizeof(BufferCacheNumaContext));
405 :
406 0 : if (get_call_result_type(fcinfo, NULL, &expected_tupledesc) != TYPEFUNC_COMPOSITE)
407 0 : elog(ERROR, "return type must be a row type");
408 :
409 0 : if (expected_tupledesc->natts != NUM_BUFFERCACHE_NUMA_ELEM)
410 0 : elog(ERROR, "incorrect number of output arguments");
411 :
412 : /* Construct a tuple descriptor for the result rows. */
413 0 : tupledesc = CreateTemplateTupleDesc(expected_tupledesc->natts);
414 0 : TupleDescInitEntry(tupledesc, (AttrNumber) 1, "bufferid",
415 : INT4OID, -1, 0);
416 0 : TupleDescInitEntry(tupledesc, (AttrNumber) 2, "os_page_num",
417 : INT8OID, -1, 0);
418 0 : TupleDescInitEntry(tupledesc, (AttrNumber) 3, "numa_node",
419 : INT4OID, -1, 0);
420 :
421 0 : fctx->tupdesc = BlessTupleDesc(tupledesc);
422 :
423 : /*
424 : * Each buffer needs at least one entry, but it might be offset in
425 : * some way, and use one extra entry. So we allocate space for the
426 : * maximum number of entries we might need, and then count the exact
427 : * number as we're walking buffers. That way we can do it in one pass,
428 : * without reallocating memory.
429 : */
430 0 : pages_per_buffer = Max(1, BLCKSZ / os_page_size) + 1;
431 0 : max_entries = NBuffers * pages_per_buffer;
432 :
433 : /* Allocate entries for BufferCachePagesRec records. */
434 0 : fctx->record = (BufferCacheNumaRec *)
435 0 : MemoryContextAllocHuge(CurrentMemoryContext,
436 : sizeof(BufferCacheNumaRec) * max_entries);
437 :
438 : /* Return to original context when allocating transient memory */
439 0 : MemoryContextSwitchTo(oldcontext);
440 :
441 0 : if (firstNumaTouch)
442 0 : elog(DEBUG1, "NUMA: page-faulting the buffercache for proper NUMA readouts");
443 :
444 : /*
445 : * Scan through all the buffers, saving the relevant fields in the
446 : * fctx->record structure.
447 : *
448 : * We don't hold the partition locks, so we don't get a consistent
449 : * snapshot across all buffers, but we do grab the buffer header
450 : * locks, so the information of each buffer is self-consistent.
451 : *
452 : * This loop touches and stores addresses into os_page_ptrs[] as input
453 : * to one big move_pages(2) inquiry system call. Basically we ask for
454 : * all memory pages for NBuffers.
455 : */
456 0 : startptr = (char *) TYPEALIGN_DOWN(os_page_size, (char *) BufferGetBlock(1));
457 0 : idx = 0;
458 0 : for (i = 0; i < NBuffers; i++)
459 : {
460 0 : char *buffptr = (char *) BufferGetBlock(i + 1);
461 : BufferDesc *bufHdr;
462 : uint32 buf_state;
463 : uint32 bufferid;
464 : int32 page_num;
465 : char *startptr_buff,
466 : *endptr_buff;
467 :
468 0 : CHECK_FOR_INTERRUPTS();
469 :
470 0 : bufHdr = GetBufferDescriptor(i);
471 :
472 : /* Lock each buffer header before inspecting. */
473 0 : buf_state = LockBufHdr(bufHdr);
474 0 : bufferid = BufferDescriptorGetBuffer(bufHdr);
475 0 : UnlockBufHdr(bufHdr, buf_state);
476 :
477 : /* start of the first page of this buffer */
478 0 : startptr_buff = (char *) TYPEALIGN_DOWN(os_page_size, buffptr);
479 :
480 : /* end of the buffer (no need to align to memory page) */
481 0 : endptr_buff = buffptr + BLCKSZ;
482 :
483 : Assert(startptr_buff < endptr_buff);
484 :
485 : /* calculate ID of the first page for this buffer */
486 0 : page_num = (startptr_buff - startptr) / os_page_size;
487 :
488 : /* Add an entry for each OS page overlapping with this buffer. */
489 0 : for (char *ptr = startptr_buff; ptr < endptr_buff; ptr += os_page_size)
490 : {
491 0 : fctx->record[idx].bufferid = bufferid;
492 0 : fctx->record[idx].page_num = page_num;
493 0 : fctx->record[idx].numa_node = os_page_status[page_num];
494 :
495 : /* advance to the next entry/page */
496 0 : ++idx;
497 0 : ++page_num;
498 : }
499 : }
500 :
501 : Assert((idx >= os_page_count) && (idx <= max_entries));
502 :
503 : /* Set max calls and remember the user function context. */
504 0 : funcctx->max_calls = idx;
505 0 : funcctx->user_fctx = fctx;
506 :
507 : /* Remember this backend touched the pages */
508 0 : firstNumaTouch = false;
509 : }
510 :
511 0 : funcctx = SRF_PERCALL_SETUP();
512 :
513 : /* Get the saved state */
514 0 : fctx = funcctx->user_fctx;
515 :
516 0 : if (funcctx->call_cntr < funcctx->max_calls)
517 : {
518 0 : uint32 i = funcctx->call_cntr;
519 : Datum values[NUM_BUFFERCACHE_NUMA_ELEM];
520 : bool nulls[NUM_BUFFERCACHE_NUMA_ELEM];
521 :
522 0 : values[0] = Int32GetDatum(fctx->record[i].bufferid);
523 0 : nulls[0] = false;
524 :
525 0 : values[1] = Int64GetDatum(fctx->record[i].page_num);
526 0 : nulls[1] = false;
527 :
528 0 : values[2] = Int32GetDatum(fctx->record[i].numa_node);
529 0 : nulls[2] = false;
530 :
531 : /* Build and return the tuple. */
532 0 : tuple = heap_form_tuple(fctx->tupdesc, values, nulls);
533 0 : result = HeapTupleGetDatum(tuple);
534 :
535 0 : SRF_RETURN_NEXT(funcctx, result);
536 : }
537 : else
538 0 : SRF_RETURN_DONE(funcctx);
539 : }
540 :
541 : Datum
542 4 : pg_buffercache_summary(PG_FUNCTION_ARGS)
543 : {
544 : Datum result;
545 : TupleDesc tupledesc;
546 : HeapTuple tuple;
547 : Datum values[NUM_BUFFERCACHE_SUMMARY_ELEM];
548 : bool nulls[NUM_BUFFERCACHE_SUMMARY_ELEM];
549 :
550 4 : int32 buffers_used = 0;
551 4 : int32 buffers_unused = 0;
552 4 : int32 buffers_dirty = 0;
553 4 : int32 buffers_pinned = 0;
554 4 : int64 usagecount_total = 0;
555 :
556 4 : if (get_call_result_type(fcinfo, NULL, &tupledesc) != TYPEFUNC_COMPOSITE)
557 0 : elog(ERROR, "return type must be a row type");
558 :
559 65540 : for (int i = 0; i < NBuffers; i++)
560 : {
561 : BufferDesc *bufHdr;
562 : uint32 buf_state;
563 :
564 : /*
565 : * This function summarizes the state of all headers. Locking the
566 : * buffer headers wouldn't provide an improved result as the state of
567 : * the buffer can still change after we release the lock and it'd
568 : * noticeably increase the cost of the function.
569 : */
570 65536 : bufHdr = GetBufferDescriptor(i);
571 65536 : buf_state = pg_atomic_read_u32(&bufHdr->state);
572 :
573 65536 : if (buf_state & BM_VALID)
574 : {
575 7960 : buffers_used++;
576 7960 : usagecount_total += BUF_STATE_GET_USAGECOUNT(buf_state);
577 :
578 7960 : if (buf_state & BM_DIRTY)
579 3780 : buffers_dirty++;
580 : }
581 : else
582 57576 : buffers_unused++;
583 :
584 65536 : if (BUF_STATE_GET_REFCOUNT(buf_state) > 0)
585 0 : buffers_pinned++;
586 : }
587 :
588 4 : memset(nulls, 0, sizeof(nulls));
589 4 : values[0] = Int32GetDatum(buffers_used);
590 4 : values[1] = Int32GetDatum(buffers_unused);
591 4 : values[2] = Int32GetDatum(buffers_dirty);
592 4 : values[3] = Int32GetDatum(buffers_pinned);
593 :
594 4 : if (buffers_used != 0)
595 4 : values[4] = Float8GetDatum((double) usagecount_total / buffers_used);
596 : else
597 0 : nulls[4] = true;
598 :
599 : /* Build and return the tuple. */
600 4 : tuple = heap_form_tuple(tupledesc, values, nulls);
601 4 : result = HeapTupleGetDatum(tuple);
602 :
603 4 : PG_RETURN_DATUM(result);
604 : }
605 :
606 : Datum
607 4 : pg_buffercache_usage_counts(PG_FUNCTION_ARGS)
608 : {
609 4 : ReturnSetInfo *rsinfo = (ReturnSetInfo *) fcinfo->resultinfo;
610 4 : int usage_counts[BM_MAX_USAGE_COUNT + 1] = {0};
611 4 : int dirty[BM_MAX_USAGE_COUNT + 1] = {0};
612 4 : int pinned[BM_MAX_USAGE_COUNT + 1] = {0};
613 : Datum values[NUM_BUFFERCACHE_USAGE_COUNTS_ELEM];
614 4 : bool nulls[NUM_BUFFERCACHE_USAGE_COUNTS_ELEM] = {0};
615 :
616 4 : InitMaterializedSRF(fcinfo, 0);
617 :
618 65540 : for (int i = 0; i < NBuffers; i++)
619 : {
620 65536 : BufferDesc *bufHdr = GetBufferDescriptor(i);
621 65536 : uint32 buf_state = pg_atomic_read_u32(&bufHdr->state);
622 : int usage_count;
623 :
624 65536 : usage_count = BUF_STATE_GET_USAGECOUNT(buf_state);
625 65536 : usage_counts[usage_count]++;
626 :
627 65536 : if (buf_state & BM_DIRTY)
628 3780 : dirty[usage_count]++;
629 :
630 65536 : if (BUF_STATE_GET_REFCOUNT(buf_state) > 0)
631 0 : pinned[usage_count]++;
632 : }
633 :
634 28 : for (int i = 0; i < BM_MAX_USAGE_COUNT + 1; i++)
635 : {
636 24 : values[0] = Int32GetDatum(i);
637 24 : values[1] = Int32GetDatum(usage_counts[i]);
638 24 : values[2] = Int32GetDatum(dirty[i]);
639 24 : values[3] = Int32GetDatum(pinned[i]);
640 :
641 24 : tuplestore_putvalues(rsinfo->setResult, rsinfo->setDesc, values, nulls);
642 : }
643 :
644 4 : return (Datum) 0;
645 : }
646 :
647 : /*
648 : * Helper function to check if the user has superuser privileges.
649 : */
650 : static void
651 20 : pg_buffercache_superuser_check(char *func_name)
652 : {
653 20 : if (!superuser())
654 6 : ereport(ERROR,
655 : (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
656 : errmsg("must be superuser to use %s()",
657 : func_name)));
658 14 : }
659 :
660 : /*
661 : * Try to evict a shared buffer.
662 : */
663 : Datum
664 10 : pg_buffercache_evict(PG_FUNCTION_ARGS)
665 : {
666 : Datum result;
667 : TupleDesc tupledesc;
668 : HeapTuple tuple;
669 : Datum values[NUM_BUFFERCACHE_EVICT_ELEM];
670 10 : bool nulls[NUM_BUFFERCACHE_EVICT_ELEM] = {0};
671 :
672 10 : Buffer buf = PG_GETARG_INT32(0);
673 : bool buffer_flushed;
674 :
675 10 : if (get_call_result_type(fcinfo, NULL, &tupledesc) != TYPEFUNC_COMPOSITE)
676 0 : elog(ERROR, "return type must be a row type");
677 :
678 10 : pg_buffercache_superuser_check("pg_buffercache_evict");
679 :
680 8 : if (buf < 1 || buf > NBuffers)
681 6 : elog(ERROR, "bad buffer ID: %d", buf);
682 :
683 2 : values[0] = BoolGetDatum(EvictUnpinnedBuffer(buf, &buffer_flushed));
684 2 : values[1] = BoolGetDatum(buffer_flushed);
685 :
686 2 : tuple = heap_form_tuple(tupledesc, values, nulls);
687 2 : result = HeapTupleGetDatum(tuple);
688 :
689 2 : PG_RETURN_DATUM(result);
690 : }
691 :
692 : /*
693 : * Try to evict specified relation.
694 : */
695 : Datum
696 6 : pg_buffercache_evict_relation(PG_FUNCTION_ARGS)
697 : {
698 : Datum result;
699 : TupleDesc tupledesc;
700 : HeapTuple tuple;
701 : Datum values[NUM_BUFFERCACHE_EVICT_RELATION_ELEM];
702 6 : bool nulls[NUM_BUFFERCACHE_EVICT_RELATION_ELEM] = {0};
703 :
704 : Oid relOid;
705 : Relation rel;
706 :
707 6 : int32 buffers_evicted = 0;
708 6 : int32 buffers_flushed = 0;
709 6 : int32 buffers_skipped = 0;
710 :
711 6 : if (get_call_result_type(fcinfo, NULL, &tupledesc) != TYPEFUNC_COMPOSITE)
712 0 : elog(ERROR, "return type must be a row type");
713 :
714 6 : pg_buffercache_superuser_check("pg_buffercache_evict_relation");
715 :
716 4 : relOid = PG_GETARG_OID(0);
717 :
718 4 : rel = relation_open(relOid, AccessShareLock);
719 :
720 4 : if (RelationUsesLocalBuffers(rel))
721 2 : ereport(ERROR,
722 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
723 : errmsg("relation uses local buffers, %s() is intended to be used for shared buffers only",
724 : "pg_buffercache_evict_relation")));
725 :
726 2 : EvictRelUnpinnedBuffers(rel, &buffers_evicted, &buffers_flushed,
727 : &buffers_skipped);
728 :
729 2 : relation_close(rel, AccessShareLock);
730 :
731 2 : values[0] = Int32GetDatum(buffers_evicted);
732 2 : values[1] = Int32GetDatum(buffers_flushed);
733 2 : values[2] = Int32GetDatum(buffers_skipped);
734 :
735 2 : tuple = heap_form_tuple(tupledesc, values, nulls);
736 2 : result = HeapTupleGetDatum(tuple);
737 :
738 2 : PG_RETURN_DATUM(result);
739 : }
740 :
741 :
742 : /*
743 : * Try to evict all shared buffers.
744 : */
745 : Datum
746 4 : pg_buffercache_evict_all(PG_FUNCTION_ARGS)
747 : {
748 : Datum result;
749 : TupleDesc tupledesc;
750 : HeapTuple tuple;
751 : Datum values[NUM_BUFFERCACHE_EVICT_ALL_ELEM];
752 4 : bool nulls[NUM_BUFFERCACHE_EVICT_ALL_ELEM] = {0};
753 :
754 4 : int32 buffers_evicted = 0;
755 4 : int32 buffers_flushed = 0;
756 4 : int32 buffers_skipped = 0;
757 :
758 4 : if (get_call_result_type(fcinfo, NULL, &tupledesc) != TYPEFUNC_COMPOSITE)
759 0 : elog(ERROR, "return type must be a row type");
760 :
761 4 : pg_buffercache_superuser_check("pg_buffercache_evict_all");
762 :
763 2 : EvictAllUnpinnedBuffers(&buffers_evicted, &buffers_flushed,
764 : &buffers_skipped);
765 :
766 2 : values[0] = Int32GetDatum(buffers_evicted);
767 2 : values[1] = Int32GetDatum(buffers_flushed);
768 2 : values[2] = Int32GetDatum(buffers_skipped);
769 :
770 2 : tuple = heap_form_tuple(tupledesc, values, nulls);
771 2 : result = HeapTupleGetDatum(tuple);
772 :
773 2 : PG_RETURN_DATUM(result);
774 : }
|