Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * autoprewarm.c
4 : * Periodically dump information about the blocks present in
5 : * shared_buffers, and reload them on server restart.
6 : *
7 : * Due to locking considerations, we can't actually begin prewarming
8 : * until the server reaches a consistent state. We need the catalogs
9 : * to be consistent so that we can figure out which relation to lock,
10 : * and we need to lock the relations so that we don't try to prewarm
11 : * pages from a relation that is in the process of being dropped.
12 : *
13 : * While prewarming, autoprewarm will use two workers. There's a
14 : * leader worker that reads and sorts the list of blocks to be
15 : * prewarmed and then launches a per-database worker for each
16 : * relevant database in turn. The former keeps running after the
17 : * initial prewarm is complete to update the dump file periodically.
18 : *
19 : * Copyright (c) 2016-2024, PostgreSQL Global Development Group
20 : *
21 : * IDENTIFICATION
22 : * contrib/pg_prewarm/autoprewarm.c
23 : *
24 : *-------------------------------------------------------------------------
25 : */
26 :
27 : #include "postgres.h"
28 :
29 : #include <unistd.h>
30 :
31 : #include "access/relation.h"
32 : #include "access/xact.h"
33 : #include "catalog/pg_class.h"
34 : #include "catalog/pg_type.h"
35 : #include "pgstat.h"
36 : #include "postmaster/bgworker.h"
37 : #include "postmaster/interrupt.h"
38 : #include "storage/buf_internals.h"
39 : #include "storage/dsm.h"
40 : #include "storage/dsm_registry.h"
41 : #include "storage/fd.h"
42 : #include "storage/ipc.h"
43 : #include "storage/latch.h"
44 : #include "storage/lwlock.h"
45 : #include "storage/proc.h"
46 : #include "storage/procsignal.h"
47 : #include "storage/shmem.h"
48 : #include "storage/smgr.h"
49 : #include "tcop/tcopprot.h"
50 : #include "utils/acl.h"
51 : #include "utils/datetime.h"
52 : #include "utils/guc.h"
53 : #include "utils/memutils.h"
54 : #include "utils/rel.h"
55 : #include "utils/relfilenumbermap.h"
56 : #include "utils/resowner.h"
57 :
58 : #define AUTOPREWARM_FILE "autoprewarm.blocks"
59 :
60 : /* Metadata for each block we dump. */
61 : typedef struct BlockInfoRecord
62 : {
63 : Oid database;
64 : Oid tablespace;
65 : RelFileNumber filenumber;
66 : ForkNumber forknum;
67 : BlockNumber blocknum;
68 : } BlockInfoRecord;
69 :
70 : /* Shared state information for autoprewarm bgworker. */
71 : typedef struct AutoPrewarmSharedState
72 : {
73 : LWLock lock; /* mutual exclusion */
74 : pid_t bgworker_pid; /* for main bgworker */
75 : pid_t pid_using_dumpfile; /* for autoprewarm or block dump */
76 :
77 : /* Following items are for communication with per-database worker */
78 : dsm_handle block_info_handle;
79 : Oid database;
80 : int prewarm_start_idx;
81 : int prewarm_stop_idx;
82 : int prewarmed_blocks;
83 : } AutoPrewarmSharedState;
84 :
85 : PGDLLEXPORT void autoprewarm_main(Datum main_arg);
86 : PGDLLEXPORT void autoprewarm_database_main(Datum main_arg);
87 :
88 4 : PG_FUNCTION_INFO_V1(autoprewarm_start_worker);
89 6 : PG_FUNCTION_INFO_V1(autoprewarm_dump_now);
90 :
91 : static void apw_load_buffers(void);
92 : static int apw_dump_now(bool is_bgworker, bool dump_unlogged);
93 : static void apw_start_leader_worker(void);
94 : static void apw_start_database_worker(void);
95 : static bool apw_init_shmem(void);
96 : static void apw_detach_shmem(int code, Datum arg);
97 : static int apw_compare_blockinfo(const void *p, const void *q);
98 :
99 : /* Pointer to shared-memory state. */
100 : static AutoPrewarmSharedState *apw_state = NULL;
101 :
102 : /* GUC variables. */
103 : static bool autoprewarm = true; /* start worker? */
104 : static int autoprewarm_interval = 300; /* dump interval */
105 :
106 : /*
107 : * Module load callback.
108 : */
109 : void
110 10 : _PG_init(void)
111 : {
112 10 : DefineCustomIntVariable("pg_prewarm.autoprewarm_interval",
113 : "Sets the interval between dumps of shared buffers",
114 : "If set to zero, time-based dumping is disabled.",
115 : &autoprewarm_interval,
116 : 300,
117 : 0, INT_MAX / 1000,
118 : PGC_SIGHUP,
119 : GUC_UNIT_S,
120 : NULL,
121 : NULL,
122 : NULL);
123 :
124 10 : if (!process_shared_preload_libraries_in_progress)
125 6 : return;
126 :
127 : /* can't define PGC_POSTMASTER variable after startup */
128 4 : DefineCustomBoolVariable("pg_prewarm.autoprewarm",
129 : "Starts the autoprewarm worker.",
130 : NULL,
131 : &autoprewarm,
132 : true,
133 : PGC_POSTMASTER,
134 : 0,
135 : NULL,
136 : NULL,
137 : NULL);
138 :
139 4 : MarkGUCPrefixReserved("pg_prewarm");
140 :
141 : /* Register autoprewarm worker, if enabled. */
142 4 : if (autoprewarm)
143 4 : apw_start_leader_worker();
144 : }
145 :
146 : /*
147 : * Main entry point for the leader autoprewarm process. Per-database workers
148 : * have a separate entry point.
149 : */
150 : void
151 4 : autoprewarm_main(Datum main_arg)
152 : {
153 4 : bool first_time = true;
154 4 : bool final_dump_allowed = true;
155 4 : TimestampTz last_dump_time = 0;
156 :
157 : /* Establish signal handlers; once that's done, unblock signals. */
158 4 : pqsignal(SIGTERM, SignalHandlerForShutdownRequest);
159 4 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
160 4 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
161 4 : BackgroundWorkerUnblockSignals();
162 :
163 : /* Create (if necessary) and attach to our shared memory area. */
164 4 : if (apw_init_shmem())
165 0 : first_time = false;
166 :
167 : /*
168 : * Set on-detach hook so that our PID will be cleared on exit.
169 : *
170 : * NB: Autoprewarm's state is stored in a DSM segment, and DSM segments
171 : * are detached before calling the on_shmem_exit callbacks, so we must put
172 : * apw_detach_shmem in the before_shmem_exit callback list.
173 : */
174 4 : before_shmem_exit(apw_detach_shmem, 0);
175 :
176 : /*
177 : * Store our PID in the shared memory area --- unless there's already
178 : * another worker running, in which case just exit.
179 : */
180 4 : LWLockAcquire(&apw_state->lock, LW_EXCLUSIVE);
181 4 : if (apw_state->bgworker_pid != InvalidPid)
182 : {
183 0 : LWLockRelease(&apw_state->lock);
184 0 : ereport(LOG,
185 : (errmsg("autoprewarm worker is already running under PID %d",
186 : (int) apw_state->bgworker_pid)));
187 0 : return;
188 : }
189 4 : apw_state->bgworker_pid = MyProcPid;
190 4 : LWLockRelease(&apw_state->lock);
191 :
192 : /*
193 : * Preload buffers from the dump file only if we just created the shared
194 : * memory region. Otherwise, it's either already been done or shouldn't
195 : * be done - e.g. because the old dump file has been overwritten since the
196 : * server was started.
197 : *
198 : * There's not much point in performing a dump immediately after we finish
199 : * preloading; so, if we do end up preloading, consider the last dump time
200 : * to be equal to the current time.
201 : *
202 : * If apw_load_buffers() is terminated early by a shutdown request,
203 : * prevent dumping out our state below the loop, because we'd effectively
204 : * just truncate the saved state to however much we'd managed to preload.
205 : */
206 4 : if (first_time)
207 : {
208 4 : apw_load_buffers();
209 4 : final_dump_allowed = !ShutdownRequestPending;
210 4 : last_dump_time = GetCurrentTimestamp();
211 : }
212 :
213 : /* Periodically dump buffers until terminated. */
214 10 : while (!ShutdownRequestPending)
215 : {
216 : /* In case of a SIGHUP, just reload the configuration. */
217 6 : if (ConfigReloadPending)
218 : {
219 0 : ConfigReloadPending = false;
220 0 : ProcessConfigFile(PGC_SIGHUP);
221 : }
222 :
223 6 : if (autoprewarm_interval <= 0)
224 : {
225 : /* We're only dumping at shutdown, so just wait forever. */
226 6 : (void) WaitLatch(MyLatch,
227 : WL_LATCH_SET | WL_EXIT_ON_PM_DEATH,
228 : -1L,
229 : WAIT_EVENT_EXTENSION);
230 : }
231 : else
232 : {
233 : TimestampTz next_dump_time;
234 : long delay_in_ms;
235 :
236 : /* Compute the next dump time. */
237 0 : next_dump_time =
238 0 : TimestampTzPlusMilliseconds(last_dump_time,
239 : autoprewarm_interval * 1000);
240 : delay_in_ms =
241 0 : TimestampDifferenceMilliseconds(GetCurrentTimestamp(),
242 : next_dump_time);
243 :
244 : /* Perform a dump if it's time. */
245 0 : if (delay_in_ms <= 0)
246 : {
247 0 : last_dump_time = GetCurrentTimestamp();
248 0 : apw_dump_now(true, false);
249 0 : continue;
250 : }
251 :
252 : /* Sleep until the next dump time. */
253 0 : (void) WaitLatch(MyLatch,
254 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
255 : delay_in_ms,
256 : WAIT_EVENT_EXTENSION);
257 : }
258 :
259 : /* Reset the latch, loop. */
260 6 : ResetLatch(MyLatch);
261 : }
262 :
263 : /*
264 : * Dump one last time. We assume this is probably the result of a system
265 : * shutdown, although it's possible that we've merely been terminated.
266 : */
267 4 : if (final_dump_allowed)
268 4 : apw_dump_now(true, true);
269 : }
270 :
271 : /*
272 : * Read the dump file and launch per-database workers one at a time to
273 : * prewarm the buffers found there.
274 : */
275 : static void
276 4 : apw_load_buffers(void)
277 : {
278 4 : FILE *file = NULL;
279 : int num_elements,
280 : i;
281 : BlockInfoRecord *blkinfo;
282 : dsm_segment *seg;
283 :
284 : /*
285 : * Skip the prewarm if the dump file is in use; otherwise, prevent any
286 : * other process from writing it while we're using it.
287 : */
288 4 : LWLockAcquire(&apw_state->lock, LW_EXCLUSIVE);
289 4 : if (apw_state->pid_using_dumpfile == InvalidPid)
290 4 : apw_state->pid_using_dumpfile = MyProcPid;
291 : else
292 : {
293 0 : LWLockRelease(&apw_state->lock);
294 0 : ereport(LOG,
295 : (errmsg("skipping prewarm because block dump file is being written by PID %d",
296 : (int) apw_state->pid_using_dumpfile)));
297 2 : return;
298 : }
299 4 : LWLockRelease(&apw_state->lock);
300 :
301 : /*
302 : * Open the block dump file. Exit quietly if it doesn't exist, but report
303 : * any other error.
304 : */
305 4 : file = AllocateFile(AUTOPREWARM_FILE, "r");
306 4 : if (!file)
307 : {
308 2 : if (errno == ENOENT)
309 : {
310 2 : LWLockAcquire(&apw_state->lock, LW_EXCLUSIVE);
311 2 : apw_state->pid_using_dumpfile = InvalidPid;
312 2 : LWLockRelease(&apw_state->lock);
313 2 : return; /* No file to load. */
314 : }
315 0 : ereport(ERROR,
316 : (errcode_for_file_access(),
317 : errmsg("could not read file \"%s\": %m",
318 : AUTOPREWARM_FILE)));
319 : }
320 :
321 : /* First line of the file is a record count. */
322 2 : if (fscanf(file, "<<%d>>\n", &num_elements) != 1)
323 0 : ereport(ERROR,
324 : (errcode_for_file_access(),
325 : errmsg("could not read from file \"%s\": %m",
326 : AUTOPREWARM_FILE)));
327 :
328 : /* Allocate a dynamic shared memory segment to store the record data. */
329 2 : seg = dsm_create(sizeof(BlockInfoRecord) * num_elements, 0);
330 2 : blkinfo = (BlockInfoRecord *) dsm_segment_address(seg);
331 :
332 : /* Read records, one per line. */
333 392 : for (i = 0; i < num_elements; i++)
334 : {
335 : unsigned forknum;
336 :
337 390 : if (fscanf(file, "%u,%u,%u,%u,%u\n", &blkinfo[i].database,
338 390 : &blkinfo[i].tablespace, &blkinfo[i].filenumber,
339 390 : &forknum, &blkinfo[i].blocknum) != 5)
340 0 : ereport(ERROR,
341 : (errmsg("autoprewarm block dump file is corrupted at line %d",
342 : i + 1)));
343 390 : blkinfo[i].forknum = forknum;
344 : }
345 :
346 2 : FreeFile(file);
347 :
348 : /* Sort the blocks to be loaded. */
349 2 : qsort(blkinfo, num_elements, sizeof(BlockInfoRecord),
350 : apw_compare_blockinfo);
351 :
352 : /* Populate shared memory state. */
353 2 : apw_state->block_info_handle = dsm_segment_handle(seg);
354 2 : apw_state->prewarm_start_idx = apw_state->prewarm_stop_idx = 0;
355 2 : apw_state->prewarmed_blocks = 0;
356 :
357 : /* Get the info position of the first block of the next database. */
358 4 : while (apw_state->prewarm_start_idx < num_elements)
359 : {
360 2 : int j = apw_state->prewarm_start_idx;
361 2 : Oid current_db = blkinfo[j].database;
362 :
363 : /*
364 : * Advance the prewarm_stop_idx to the first BlockInfoRecord that does
365 : * not belong to this database.
366 : */
367 2 : j++;
368 390 : while (j < num_elements)
369 : {
370 388 : if (current_db != blkinfo[j].database)
371 : {
372 : /*
373 : * Combine BlockInfoRecords for global objects with those of
374 : * the database.
375 : */
376 2 : if (current_db != InvalidOid)
377 0 : break;
378 2 : current_db = blkinfo[j].database;
379 : }
380 :
381 388 : j++;
382 : }
383 :
384 : /*
385 : * If we reach this point with current_db == InvalidOid, then only
386 : * BlockInfoRecords belonging to global objects exist. We can't
387 : * prewarm without a database connection, so just bail out.
388 : */
389 2 : if (current_db == InvalidOid)
390 0 : break;
391 :
392 : /* Configure stop point and database for next per-database worker. */
393 2 : apw_state->prewarm_stop_idx = j;
394 2 : apw_state->database = current_db;
395 : Assert(apw_state->prewarm_start_idx < apw_state->prewarm_stop_idx);
396 :
397 : /* If we've run out of free buffers, don't launch another worker. */
398 2 : if (!have_free_buffer())
399 0 : break;
400 :
401 : /*
402 : * Likewise, don't launch if we've already been told to shut down.
403 : * (The launch would fail anyway, but we might as well skip it.)
404 : */
405 2 : if (ShutdownRequestPending)
406 0 : break;
407 :
408 : /*
409 : * Start a per-database worker to load blocks for this database; this
410 : * function will return once the per-database worker exits.
411 : */
412 2 : apw_start_database_worker();
413 :
414 : /* Prepare for next database. */
415 2 : apw_state->prewarm_start_idx = apw_state->prewarm_stop_idx;
416 : }
417 :
418 : /* Clean up. */
419 2 : dsm_detach(seg);
420 2 : LWLockAcquire(&apw_state->lock, LW_EXCLUSIVE);
421 2 : apw_state->block_info_handle = DSM_HANDLE_INVALID;
422 2 : apw_state->pid_using_dumpfile = InvalidPid;
423 2 : LWLockRelease(&apw_state->lock);
424 :
425 : /* Report our success, if we were able to finish. */
426 2 : if (!ShutdownRequestPending)
427 2 : ereport(LOG,
428 : (errmsg("autoprewarm successfully prewarmed %d of %d previously-loaded blocks",
429 : apw_state->prewarmed_blocks, num_elements)));
430 : }
431 :
432 : /*
433 : * Prewarm all blocks for one database (and possibly also global objects, if
434 : * those got grouped with this database).
435 : */
436 : void
437 2 : autoprewarm_database_main(Datum main_arg)
438 : {
439 : int pos;
440 : BlockInfoRecord *block_info;
441 2 : Relation rel = NULL;
442 2 : BlockNumber nblocks = 0;
443 2 : BlockInfoRecord *old_blk = NULL;
444 : dsm_segment *seg;
445 :
446 : /* Establish signal handlers; once that's done, unblock signals. */
447 2 : pqsignal(SIGTERM, die);
448 2 : BackgroundWorkerUnblockSignals();
449 :
450 : /* Connect to correct database and get block information. */
451 2 : apw_init_shmem();
452 2 : seg = dsm_attach(apw_state->block_info_handle);
453 2 : if (seg == NULL)
454 0 : ereport(ERROR,
455 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
456 : errmsg("could not map dynamic shared memory segment")));
457 2 : BackgroundWorkerInitializeConnectionByOid(apw_state->database, InvalidOid, 0);
458 2 : block_info = (BlockInfoRecord *) dsm_segment_address(seg);
459 2 : pos = apw_state->prewarm_start_idx;
460 :
461 : /*
462 : * Loop until we run out of blocks to prewarm or until we run out of free
463 : * buffers.
464 : */
465 392 : while (pos < apw_state->prewarm_stop_idx && have_free_buffer())
466 : {
467 390 : BlockInfoRecord *blk = &block_info[pos++];
468 : Buffer buf;
469 :
470 390 : CHECK_FOR_INTERRUPTS();
471 :
472 : /*
473 : * Quit if we've reached records for another database. If previous
474 : * blocks are of some global objects, then continue pre-warming.
475 : */
476 390 : if (old_blk != NULL && old_blk->database != blk->database &&
477 2 : old_blk->database != 0)
478 0 : break;
479 :
480 : /*
481 : * As soon as we encounter a block of a new relation, close the old
482 : * relation. Note that rel will be NULL if try_relation_open failed
483 : * previously; in that case, there is nothing to close.
484 : */
485 390 : if (old_blk != NULL && old_blk->filenumber != blk->filenumber &&
486 : rel != NULL)
487 : {
488 94 : relation_close(rel, AccessShareLock);
489 94 : rel = NULL;
490 94 : CommitTransactionCommand();
491 : }
492 :
493 : /*
494 : * Try to open each new relation, but only once, when we first
495 : * encounter it. If it's been dropped, skip the associated blocks.
496 : */
497 390 : if (old_blk == NULL || old_blk->filenumber != blk->filenumber)
498 : {
499 : Oid reloid;
500 :
501 : Assert(rel == NULL);
502 96 : StartTransactionCommand();
503 96 : reloid = RelidByRelfilenumber(blk->tablespace, blk->filenumber);
504 96 : if (OidIsValid(reloid))
505 96 : rel = try_relation_open(reloid, AccessShareLock);
506 :
507 96 : if (!rel)
508 0 : CommitTransactionCommand();
509 : }
510 390 : if (!rel)
511 : {
512 0 : old_blk = blk;
513 0 : continue;
514 : }
515 :
516 : /* Once per fork, check for fork existence and size. */
517 390 : if (old_blk == NULL ||
518 388 : old_blk->filenumber != blk->filenumber ||
519 294 : old_blk->forknum != blk->forknum)
520 : {
521 : /*
522 : * smgrexists is not safe for illegal forknum, hence check whether
523 : * the passed forknum is valid before using it in smgrexists.
524 : */
525 124 : if (blk->forknum > InvalidForkNumber &&
526 248 : blk->forknum <= MAX_FORKNUM &&
527 124 : smgrexists(RelationGetSmgr(rel), blk->forknum))
528 124 : nblocks = RelationGetNumberOfBlocksInFork(rel, blk->forknum);
529 : else
530 0 : nblocks = 0;
531 : }
532 :
533 : /* Check whether blocknum is valid and within fork file size. */
534 390 : if (blk->blocknum >= nblocks)
535 : {
536 : /* Move to next forknum. */
537 0 : old_blk = blk;
538 0 : continue;
539 : }
540 :
541 : /* Prewarm buffer. */
542 390 : buf = ReadBufferExtended(rel, blk->forknum, blk->blocknum, RBM_NORMAL,
543 : NULL);
544 390 : if (BufferIsValid(buf))
545 : {
546 390 : apw_state->prewarmed_blocks++;
547 390 : ReleaseBuffer(buf);
548 : }
549 :
550 390 : old_blk = blk;
551 : }
552 :
553 2 : dsm_detach(seg);
554 :
555 : /* Release lock on previous relation. */
556 2 : if (rel)
557 : {
558 2 : relation_close(rel, AccessShareLock);
559 2 : CommitTransactionCommand();
560 : }
561 2 : }
562 :
563 : /*
564 : * Dump information on blocks in shared buffers. We use a text format here
565 : * so that it's easy to understand and even change the file contents if
566 : * necessary.
567 : * Returns the number of blocks dumped.
568 : */
569 : static int
570 6 : apw_dump_now(bool is_bgworker, bool dump_unlogged)
571 : {
572 : int num_blocks;
573 : int i;
574 : int ret;
575 : BlockInfoRecord *block_info_array;
576 : BufferDesc *bufHdr;
577 : FILE *file;
578 : char transient_dump_file_path[MAXPGPATH];
579 : pid_t pid;
580 :
581 6 : LWLockAcquire(&apw_state->lock, LW_EXCLUSIVE);
582 6 : pid = apw_state->pid_using_dumpfile;
583 6 : if (apw_state->pid_using_dumpfile == InvalidPid)
584 6 : apw_state->pid_using_dumpfile = MyProcPid;
585 6 : LWLockRelease(&apw_state->lock);
586 :
587 6 : if (pid != InvalidPid)
588 : {
589 0 : if (!is_bgworker)
590 0 : ereport(ERROR,
591 : (errmsg("could not perform block dump because dump file is being used by PID %d",
592 : (int) apw_state->pid_using_dumpfile)));
593 :
594 0 : ereport(LOG,
595 : (errmsg("skipping block dump because it is already being performed by PID %d",
596 : (int) apw_state->pid_using_dumpfile)));
597 0 : return 0;
598 : }
599 :
600 : block_info_array =
601 6 : (BlockInfoRecord *) palloc(sizeof(BlockInfoRecord) * NBuffers);
602 :
603 98310 : for (num_blocks = 0, i = 0; i < NBuffers; i++)
604 : {
605 : uint32 buf_state;
606 :
607 98304 : CHECK_FOR_INTERRUPTS();
608 :
609 98304 : bufHdr = GetBufferDescriptor(i);
610 :
611 : /* Lock each buffer header before inspecting. */
612 98304 : buf_state = LockBufHdr(bufHdr);
613 :
614 : /*
615 : * Unlogged tables will be automatically truncated after a crash or
616 : * unclean shutdown. In such cases we need not prewarm them. Dump them
617 : * only if requested by caller.
618 : */
619 98304 : if (buf_state & BM_TAG_VALID &&
620 1170 : ((buf_state & BM_PERMANENT) || dump_unlogged))
621 : {
622 1170 : block_info_array[num_blocks].database = bufHdr->tag.dbOid;
623 1170 : block_info_array[num_blocks].tablespace = bufHdr->tag.spcOid;
624 2340 : block_info_array[num_blocks].filenumber =
625 1170 : BufTagGetRelNumber(&bufHdr->tag);
626 2340 : block_info_array[num_blocks].forknum =
627 1170 : BufTagGetForkNum(&bufHdr->tag);
628 1170 : block_info_array[num_blocks].blocknum = bufHdr->tag.blockNum;
629 1170 : ++num_blocks;
630 : }
631 :
632 98304 : UnlockBufHdr(bufHdr, buf_state);
633 : }
634 :
635 6 : snprintf(transient_dump_file_path, MAXPGPATH, "%s.tmp", AUTOPREWARM_FILE);
636 6 : file = AllocateFile(transient_dump_file_path, "w");
637 6 : if (!file)
638 0 : ereport(ERROR,
639 : (errcode_for_file_access(),
640 : errmsg("could not open file \"%s\": %m",
641 : transient_dump_file_path)));
642 :
643 6 : ret = fprintf(file, "<<%d>>\n", num_blocks);
644 6 : if (ret < 0)
645 : {
646 0 : int save_errno = errno;
647 :
648 0 : FreeFile(file);
649 0 : unlink(transient_dump_file_path);
650 0 : errno = save_errno;
651 0 : ereport(ERROR,
652 : (errcode_for_file_access(),
653 : errmsg("could not write to file \"%s\": %m",
654 : transient_dump_file_path)));
655 : }
656 :
657 1176 : for (i = 0; i < num_blocks; i++)
658 : {
659 1170 : CHECK_FOR_INTERRUPTS();
660 :
661 1170 : ret = fprintf(file, "%u,%u,%u,%u,%u\n",
662 1170 : block_info_array[i].database,
663 1170 : block_info_array[i].tablespace,
664 1170 : block_info_array[i].filenumber,
665 1170 : (uint32) block_info_array[i].forknum,
666 1170 : block_info_array[i].blocknum);
667 1170 : if (ret < 0)
668 : {
669 0 : int save_errno = errno;
670 :
671 0 : FreeFile(file);
672 0 : unlink(transient_dump_file_path);
673 0 : errno = save_errno;
674 0 : ereport(ERROR,
675 : (errcode_for_file_access(),
676 : errmsg("could not write to file \"%s\": %m",
677 : transient_dump_file_path)));
678 : }
679 : }
680 :
681 6 : pfree(block_info_array);
682 :
683 : /*
684 : * Rename transient_dump_file_path to AUTOPREWARM_FILE to make things
685 : * permanent.
686 : */
687 6 : ret = FreeFile(file);
688 6 : if (ret != 0)
689 : {
690 0 : int save_errno = errno;
691 :
692 0 : unlink(transient_dump_file_path);
693 0 : errno = save_errno;
694 0 : ereport(ERROR,
695 : (errcode_for_file_access(),
696 : errmsg("could not close file \"%s\": %m",
697 : transient_dump_file_path)));
698 : }
699 :
700 6 : (void) durable_rename(transient_dump_file_path, AUTOPREWARM_FILE, ERROR);
701 6 : apw_state->pid_using_dumpfile = InvalidPid;
702 :
703 6 : ereport(DEBUG1,
704 : (errmsg_internal("wrote block details for %d blocks", num_blocks)));
705 6 : return num_blocks;
706 : }
707 :
708 : /*
709 : * SQL-callable function to launch autoprewarm.
710 : */
711 : Datum
712 0 : autoprewarm_start_worker(PG_FUNCTION_ARGS)
713 : {
714 : pid_t pid;
715 :
716 0 : if (!autoprewarm)
717 0 : ereport(ERROR,
718 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
719 : errmsg("autoprewarm is disabled")));
720 :
721 0 : apw_init_shmem();
722 0 : LWLockAcquire(&apw_state->lock, LW_EXCLUSIVE);
723 0 : pid = apw_state->bgworker_pid;
724 0 : LWLockRelease(&apw_state->lock);
725 :
726 0 : if (pid != InvalidPid)
727 0 : ereport(ERROR,
728 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
729 : errmsg("autoprewarm worker is already running under PID %d",
730 : (int) pid)));
731 :
732 0 : apw_start_leader_worker();
733 :
734 0 : PG_RETURN_VOID();
735 : }
736 :
737 : /*
738 : * SQL-callable function to perform an immediate block dump.
739 : *
740 : * Note: this is declared to return int8, as insurance against some
741 : * very distant day when we might make NBuffers wider than int.
742 : */
743 : Datum
744 2 : autoprewarm_dump_now(PG_FUNCTION_ARGS)
745 : {
746 : int num_blocks;
747 :
748 2 : apw_init_shmem();
749 :
750 2 : PG_ENSURE_ERROR_CLEANUP(apw_detach_shmem, 0);
751 : {
752 2 : num_blocks = apw_dump_now(false, true);
753 : }
754 2 : PG_END_ENSURE_ERROR_CLEANUP(apw_detach_shmem, 0);
755 :
756 2 : PG_RETURN_INT64((int64) num_blocks);
757 : }
758 :
759 : static void
760 4 : apw_init_state(void *ptr)
761 : {
762 4 : AutoPrewarmSharedState *state = (AutoPrewarmSharedState *) ptr;
763 :
764 4 : LWLockInitialize(&state->lock, LWLockNewTrancheId());
765 4 : state->bgworker_pid = InvalidPid;
766 4 : state->pid_using_dumpfile = InvalidPid;
767 4 : }
768 :
769 : /*
770 : * Allocate and initialize autoprewarm related shared memory, if not already
771 : * done, and set up backend-local pointer to that state. Returns true if an
772 : * existing shared memory segment was found.
773 : */
774 : static bool
775 8 : apw_init_shmem(void)
776 : {
777 : bool found;
778 :
779 8 : apw_state = GetNamedDSMSegment("autoprewarm",
780 : sizeof(AutoPrewarmSharedState),
781 : apw_init_state,
782 : &found);
783 8 : LWLockRegisterTranche(apw_state->lock.tranche, "autoprewarm");
784 :
785 8 : return found;
786 : }
787 :
788 : /*
789 : * Clear our PID from autoprewarm shared state.
790 : */
791 : static void
792 4 : apw_detach_shmem(int code, Datum arg)
793 : {
794 4 : LWLockAcquire(&apw_state->lock, LW_EXCLUSIVE);
795 4 : if (apw_state->pid_using_dumpfile == MyProcPid)
796 0 : apw_state->pid_using_dumpfile = InvalidPid;
797 4 : if (apw_state->bgworker_pid == MyProcPid)
798 4 : apw_state->bgworker_pid = InvalidPid;
799 4 : LWLockRelease(&apw_state->lock);
800 4 : }
801 :
802 : /*
803 : * Start autoprewarm leader worker process.
804 : */
805 : static void
806 4 : apw_start_leader_worker(void)
807 : {
808 : BackgroundWorker worker;
809 : BackgroundWorkerHandle *handle;
810 : BgwHandleStatus status;
811 : pid_t pid;
812 :
813 4 : MemSet(&worker, 0, sizeof(BackgroundWorker));
814 4 : worker.bgw_flags = BGWORKER_SHMEM_ACCESS;
815 4 : worker.bgw_start_time = BgWorkerStart_ConsistentState;
816 4 : strcpy(worker.bgw_library_name, "pg_prewarm");
817 4 : strcpy(worker.bgw_function_name, "autoprewarm_main");
818 4 : strcpy(worker.bgw_name, "autoprewarm leader");
819 4 : strcpy(worker.bgw_type, "autoprewarm leader");
820 :
821 4 : if (process_shared_preload_libraries_in_progress)
822 : {
823 4 : RegisterBackgroundWorker(&worker);
824 4 : return;
825 : }
826 :
827 : /* must set notify PID to wait for startup */
828 0 : worker.bgw_notify_pid = MyProcPid;
829 :
830 0 : if (!RegisterDynamicBackgroundWorker(&worker, &handle))
831 0 : ereport(ERROR,
832 : (errcode(ERRCODE_INSUFFICIENT_RESOURCES),
833 : errmsg("could not register background process"),
834 : errhint("You may need to increase max_worker_processes.")));
835 :
836 0 : status = WaitForBackgroundWorkerStartup(handle, &pid);
837 0 : if (status != BGWH_STARTED)
838 0 : ereport(ERROR,
839 : (errcode(ERRCODE_INSUFFICIENT_RESOURCES),
840 : errmsg("could not start background process"),
841 : errhint("More details may be available in the server log.")));
842 : }
843 :
844 : /*
845 : * Start autoprewarm per-database worker process.
846 : */
847 : static void
848 2 : apw_start_database_worker(void)
849 : {
850 : BackgroundWorker worker;
851 : BackgroundWorkerHandle *handle;
852 :
853 2 : MemSet(&worker, 0, sizeof(BackgroundWorker));
854 2 : worker.bgw_flags =
855 : BGWORKER_SHMEM_ACCESS | BGWORKER_BACKEND_DATABASE_CONNECTION;
856 2 : worker.bgw_start_time = BgWorkerStart_ConsistentState;
857 2 : worker.bgw_restart_time = BGW_NEVER_RESTART;
858 2 : strcpy(worker.bgw_library_name, "pg_prewarm");
859 2 : strcpy(worker.bgw_function_name, "autoprewarm_database_main");
860 2 : strcpy(worker.bgw_name, "autoprewarm worker");
861 2 : strcpy(worker.bgw_type, "autoprewarm worker");
862 :
863 : /* must set notify PID to wait for shutdown */
864 2 : worker.bgw_notify_pid = MyProcPid;
865 :
866 2 : if (!RegisterDynamicBackgroundWorker(&worker, &handle))
867 0 : ereport(ERROR,
868 : (errcode(ERRCODE_INSUFFICIENT_RESOURCES),
869 : errmsg("registering dynamic bgworker autoprewarm failed"),
870 : errhint("Consider increasing configuration parameter max_worker_processes.")));
871 :
872 : /*
873 : * Ignore return value; if it fails, postmaster has died, but we have
874 : * checks for that elsewhere.
875 : */
876 2 : WaitForBackgroundWorkerShutdown(handle);
877 2 : }
878 :
879 : /* Compare member elements to check whether they are not equal. */
880 : #define cmp_member_elem(fld) \
881 : do { \
882 : if (a->fld < b->fld) \
883 : return -1; \
884 : else if (a->fld > b->fld) \
885 : return 1; \
886 : } while(0)
887 :
888 : /*
889 : * apw_compare_blockinfo
890 : *
891 : * We depend on all records for a particular database being consecutive
892 : * in the dump file; each per-database worker will preload blocks until
893 : * it sees a block for some other database. Sorting by tablespace,
894 : * filenumber, forknum, and blocknum isn't critical for correctness, but
895 : * helps us get a sequential I/O pattern.
896 : */
897 : static int
898 3262 : apw_compare_blockinfo(const void *p, const void *q)
899 : {
900 3262 : const BlockInfoRecord *a = (const BlockInfoRecord *) p;
901 3262 : const BlockInfoRecord *b = (const BlockInfoRecord *) q;
902 :
903 3262 : cmp_member_elem(database);
904 3048 : cmp_member_elem(tablespace);
905 3048 : cmp_member_elem(filenumber);
906 1178 : cmp_member_elem(forknum);
907 1024 : cmp_member_elem(blocknum);
908 :
909 0 : return 0;
910 : }
|
