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) 2016-2025, 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 "pgstat.h"
34 : #include "postmaster/bgworker.h"
35 : #include "postmaster/interrupt.h"
36 : #include "storage/buf_internals.h"
37 : #include "storage/dsm.h"
38 : #include "storage/dsm_registry.h"
39 : #include "storage/fd.h"
40 : #include "storage/ipc.h"
41 : #include "storage/latch.h"
42 : #include "storage/lwlock.h"
43 : #include "storage/procsignal.h"
44 : #include "storage/read_stream.h"
45 : #include "storage/smgr.h"
46 : #include "tcop/tcopprot.h"
47 : #include "utils/guc.h"
48 : #include "utils/rel.h"
49 : #include "utils/relfilenumbermap.h"
50 : #include "utils/timestamp.h"
51 :
52 : #define AUTOPREWARM_FILE "autoprewarm.blocks"
53 :
54 : /* Metadata for each block we dump. */
55 : typedef struct BlockInfoRecord
56 : {
57 : Oid database;
58 : Oid tablespace;
59 : RelFileNumber filenumber;
60 : ForkNumber forknum;
61 : BlockNumber blocknum;
62 : } BlockInfoRecord;
63 :
64 : /* Shared state information for autoprewarm bgworker. */
65 : typedef struct AutoPrewarmSharedState
66 : {
67 : LWLock lock; /* mutual exclusion */
68 : pid_t bgworker_pid; /* for main bgworker */
69 : pid_t pid_using_dumpfile; /* for autoprewarm or block dump */
70 :
71 : /* Following items are for communication with per-database worker */
72 : dsm_handle block_info_handle;
73 : Oid database;
74 : int prewarm_start_idx;
75 : int prewarm_stop_idx;
76 : int prewarmed_blocks;
77 : } AutoPrewarmSharedState;
78 :
79 : /*
80 : * Private data passed through the read stream API for our use in the
81 : * callback.
82 : */
83 : typedef struct AutoPrewarmReadStreamData
84 : {
85 : /* The array of records containing the blocks we should prewarm. */
86 : BlockInfoRecord *block_info;
87 :
88 : /*
89 : * pos is the read stream callback's index into block_info. Because the
90 : * read stream may read ahead, pos is likely to be ahead of the index in
91 : * the main loop in autoprewarm_database_main().
92 : */
93 : int pos;
94 : Oid tablespace;
95 : RelFileNumber filenumber;
96 : ForkNumber forknum;
97 : BlockNumber nblocks;
98 : } AutoPrewarmReadStreamData;
99 :
100 :
101 : PGDLLEXPORT void autoprewarm_main(Datum main_arg);
102 : PGDLLEXPORT void autoprewarm_database_main(Datum main_arg);
103 :
104 4 : PG_FUNCTION_INFO_V1(autoprewarm_start_worker);
105 6 : PG_FUNCTION_INFO_V1(autoprewarm_dump_now);
106 :
107 : static void apw_load_buffers(void);
108 : static int apw_dump_now(bool is_bgworker, bool dump_unlogged);
109 : static void apw_start_leader_worker(void);
110 : static void apw_start_database_worker(void);
111 : static bool apw_init_shmem(void);
112 : static void apw_detach_shmem(int code, Datum arg);
113 : static int apw_compare_blockinfo(const void *p, const void *q);
114 :
115 : /* Pointer to shared-memory state. */
116 : static AutoPrewarmSharedState *apw_state = NULL;
117 :
118 : /* GUC variables. */
119 : static bool autoprewarm = true; /* start worker? */
120 : static int autoprewarm_interval = 300; /* dump interval */
121 :
122 : /*
123 : * Module load callback.
124 : */
125 : void
126 10 : _PG_init(void)
127 : {
128 10 : DefineCustomIntVariable("pg_prewarm.autoprewarm_interval",
129 : "Sets the interval between dumps of shared buffers",
130 : "If set to zero, time-based dumping is disabled.",
131 : &autoprewarm_interval,
132 : 300,
133 : 0, INT_MAX / 1000,
134 : PGC_SIGHUP,
135 : GUC_UNIT_S,
136 : NULL,
137 : NULL,
138 : NULL);
139 :
140 10 : if (!process_shared_preload_libraries_in_progress)
141 6 : return;
142 :
143 : /* can't define PGC_POSTMASTER variable after startup */
144 4 : DefineCustomBoolVariable("pg_prewarm.autoprewarm",
145 : "Starts the autoprewarm worker.",
146 : NULL,
147 : &autoprewarm,
148 : true,
149 : PGC_POSTMASTER,
150 : 0,
151 : NULL,
152 : NULL,
153 : NULL);
154 :
155 4 : MarkGUCPrefixReserved("pg_prewarm");
156 :
157 : /* Register autoprewarm worker, if enabled. */
158 4 : if (autoprewarm)
159 4 : apw_start_leader_worker();
160 : }
161 :
162 : /*
163 : * Main entry point for the leader autoprewarm process. Per-database workers
164 : * have a separate entry point.
165 : */
166 : void
167 4 : autoprewarm_main(Datum main_arg)
168 : {
169 4 : bool first_time = true;
170 4 : bool final_dump_allowed = true;
171 4 : TimestampTz last_dump_time = 0;
172 :
173 : /* Establish signal handlers; once that's done, unblock signals. */
174 4 : pqsignal(SIGTERM, SignalHandlerForShutdownRequest);
175 4 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
176 4 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
177 4 : BackgroundWorkerUnblockSignals();
178 :
179 : /* Create (if necessary) and attach to our shared memory area. */
180 4 : if (apw_init_shmem())
181 0 : first_time = false;
182 :
183 : /*
184 : * Set on-detach hook so that our PID will be cleared on exit.
185 : *
186 : * NB: Autoprewarm's state is stored in a DSM segment, and DSM segments
187 : * are detached before calling the on_shmem_exit callbacks, so we must put
188 : * apw_detach_shmem in the before_shmem_exit callback list.
189 : */
190 4 : before_shmem_exit(apw_detach_shmem, 0);
191 :
192 : /*
193 : * Store our PID in the shared memory area --- unless there's already
194 : * another worker running, in which case just exit.
195 : */
196 4 : LWLockAcquire(&apw_state->lock, LW_EXCLUSIVE);
197 4 : if (apw_state->bgworker_pid != InvalidPid)
198 : {
199 0 : LWLockRelease(&apw_state->lock);
200 0 : ereport(LOG,
201 : (errmsg("autoprewarm worker is already running under PID %d",
202 : (int) apw_state->bgworker_pid)));
203 0 : return;
204 : }
205 4 : apw_state->bgworker_pid = MyProcPid;
206 4 : LWLockRelease(&apw_state->lock);
207 :
208 : /*
209 : * Preload buffers from the dump file only if we just created the shared
210 : * memory region. Otherwise, it's either already been done or shouldn't
211 : * be done - e.g. because the old dump file has been overwritten since the
212 : * server was started.
213 : *
214 : * There's not much point in performing a dump immediately after we finish
215 : * preloading; so, if we do end up preloading, consider the last dump time
216 : * to be equal to the current time.
217 : *
218 : * If apw_load_buffers() is terminated early by a shutdown request,
219 : * prevent dumping out our state below the loop, because we'd effectively
220 : * just truncate the saved state to however much we'd managed to preload.
221 : */
222 4 : if (first_time)
223 : {
224 4 : apw_load_buffers();
225 4 : final_dump_allowed = !ShutdownRequestPending;
226 4 : last_dump_time = GetCurrentTimestamp();
227 : }
228 :
229 : /* Periodically dump buffers until terminated. */
230 10 : while (!ShutdownRequestPending)
231 : {
232 : /* In case of a SIGHUP, just reload the configuration. */
233 6 : if (ConfigReloadPending)
234 : {
235 0 : ConfigReloadPending = false;
236 0 : ProcessConfigFile(PGC_SIGHUP);
237 : }
238 :
239 6 : if (autoprewarm_interval <= 0)
240 : {
241 : /* We're only dumping at shutdown, so just wait forever. */
242 6 : (void) WaitLatch(MyLatch,
243 : WL_LATCH_SET | WL_EXIT_ON_PM_DEATH,
244 : -1L,
245 : PG_WAIT_EXTENSION);
246 : }
247 : else
248 : {
249 : TimestampTz next_dump_time;
250 : long delay_in_ms;
251 :
252 : /* Compute the next dump time. */
253 0 : next_dump_time =
254 0 : TimestampTzPlusMilliseconds(last_dump_time,
255 : autoprewarm_interval * 1000);
256 : delay_in_ms =
257 0 : TimestampDifferenceMilliseconds(GetCurrentTimestamp(),
258 : next_dump_time);
259 :
260 : /* Perform a dump if it's time. */
261 0 : if (delay_in_ms <= 0)
262 : {
263 0 : last_dump_time = GetCurrentTimestamp();
264 0 : apw_dump_now(true, false);
265 0 : continue;
266 : }
267 :
268 : /* Sleep until the next dump time. */
269 0 : (void) WaitLatch(MyLatch,
270 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
271 : delay_in_ms,
272 : PG_WAIT_EXTENSION);
273 : }
274 :
275 : /* Reset the latch, loop. */
276 6 : ResetLatch(MyLatch);
277 : }
278 :
279 : /*
280 : * Dump one last time. We assume this is probably the result of a system
281 : * shutdown, although it's possible that we've merely been terminated.
282 : */
283 4 : if (final_dump_allowed)
284 4 : apw_dump_now(true, true);
285 : }
286 :
287 : /*
288 : * Read the dump file and launch per-database workers one at a time to
289 : * prewarm the buffers found there.
290 : */
291 : static void
292 4 : apw_load_buffers(void)
293 : {
294 4 : FILE *file = NULL;
295 : int num_elements,
296 : i;
297 : BlockInfoRecord *blkinfo;
298 : dsm_segment *seg;
299 :
300 : /*
301 : * Skip the prewarm if the dump file is in use; otherwise, prevent any
302 : * other process from writing it while we're using it.
303 : */
304 4 : LWLockAcquire(&apw_state->lock, LW_EXCLUSIVE);
305 4 : if (apw_state->pid_using_dumpfile == InvalidPid)
306 4 : apw_state->pid_using_dumpfile = MyProcPid;
307 : else
308 : {
309 0 : LWLockRelease(&apw_state->lock);
310 0 : ereport(LOG,
311 : (errmsg("skipping prewarm because block dump file is being written by PID %d",
312 : (int) apw_state->pid_using_dumpfile)));
313 2 : return;
314 : }
315 4 : LWLockRelease(&apw_state->lock);
316 :
317 : /*
318 : * Open the block dump file. Exit quietly if it doesn't exist, but report
319 : * any other error.
320 : */
321 4 : file = AllocateFile(AUTOPREWARM_FILE, "r");
322 4 : if (!file)
323 : {
324 2 : if (errno == ENOENT)
325 : {
326 2 : LWLockAcquire(&apw_state->lock, LW_EXCLUSIVE);
327 2 : apw_state->pid_using_dumpfile = InvalidPid;
328 2 : LWLockRelease(&apw_state->lock);
329 2 : return; /* No file to load. */
330 : }
331 0 : ereport(ERROR,
332 : (errcode_for_file_access(),
333 : errmsg("could not read file \"%s\": %m",
334 : AUTOPREWARM_FILE)));
335 : }
336 :
337 : /* First line of the file is a record count. */
338 2 : if (fscanf(file, "<<%d>>\n", &num_elements) != 1)
339 0 : ereport(ERROR,
340 : (errcode_for_file_access(),
341 : errmsg("could not read from file \"%s\": %m",
342 : AUTOPREWARM_FILE)));
343 :
344 : /* Allocate a dynamic shared memory segment to store the record data. */
345 2 : seg = dsm_create(sizeof(BlockInfoRecord) * num_elements, 0);
346 2 : blkinfo = (BlockInfoRecord *) dsm_segment_address(seg);
347 :
348 : /* Read records, one per line. */
349 428 : for (i = 0; i < num_elements; i++)
350 : {
351 : unsigned forknum;
352 :
353 426 : if (fscanf(file, "%u,%u,%u,%u,%u\n", &blkinfo[i].database,
354 426 : &blkinfo[i].tablespace, &blkinfo[i].filenumber,
355 426 : &forknum, &blkinfo[i].blocknum) != 5)
356 0 : ereport(ERROR,
357 : (errmsg("autoprewarm block dump file is corrupted at line %d",
358 : i + 1)));
359 426 : blkinfo[i].forknum = forknum;
360 : }
361 :
362 2 : FreeFile(file);
363 :
364 : /* Sort the blocks to be loaded. */
365 2 : qsort(blkinfo, num_elements, sizeof(BlockInfoRecord),
366 : apw_compare_blockinfo);
367 :
368 : /* Populate shared memory state. */
369 2 : apw_state->block_info_handle = dsm_segment_handle(seg);
370 2 : apw_state->prewarm_start_idx = apw_state->prewarm_stop_idx = 0;
371 2 : apw_state->prewarmed_blocks = 0;
372 :
373 : /* Get the info position of the first block of the next database. */
374 4 : while (apw_state->prewarm_start_idx < num_elements)
375 : {
376 2 : int j = apw_state->prewarm_start_idx;
377 2 : Oid current_db = blkinfo[j].database;
378 :
379 : /*
380 : * Advance the prewarm_stop_idx to the first BlockInfoRecord that does
381 : * not belong to this database.
382 : */
383 2 : j++;
384 426 : while (j < num_elements)
385 : {
386 424 : if (current_db != blkinfo[j].database)
387 : {
388 : /*
389 : * Combine BlockInfoRecords for global objects with those of
390 : * the database.
391 : */
392 2 : if (current_db != InvalidOid)
393 0 : break;
394 2 : current_db = blkinfo[j].database;
395 : }
396 :
397 424 : j++;
398 : }
399 :
400 : /*
401 : * If we reach this point with current_db == InvalidOid, then only
402 : * BlockInfoRecords belonging to global objects exist. We can't
403 : * prewarm without a database connection, so just bail out.
404 : */
405 2 : if (current_db == InvalidOid)
406 0 : break;
407 :
408 : /* Configure stop point and database for next per-database worker. */
409 2 : apw_state->prewarm_stop_idx = j;
410 2 : apw_state->database = current_db;
411 : Assert(apw_state->prewarm_start_idx < apw_state->prewarm_stop_idx);
412 :
413 : /* If we've run out of free buffers, don't launch another worker. */
414 2 : if (!have_free_buffer())
415 0 : break;
416 :
417 : /*
418 : * Likewise, don't launch if we've already been told to shut down.
419 : * (The launch would fail anyway, but we might as well skip it.)
420 : */
421 2 : if (ShutdownRequestPending)
422 0 : break;
423 :
424 : /*
425 : * Start a per-database worker to load blocks for this database; this
426 : * function will return once the per-database worker exits.
427 : */
428 2 : apw_start_database_worker();
429 :
430 : /* Prepare for next database. */
431 2 : apw_state->prewarm_start_idx = apw_state->prewarm_stop_idx;
432 : }
433 :
434 : /* Clean up. */
435 2 : dsm_detach(seg);
436 2 : LWLockAcquire(&apw_state->lock, LW_EXCLUSIVE);
437 2 : apw_state->block_info_handle = DSM_HANDLE_INVALID;
438 2 : apw_state->pid_using_dumpfile = InvalidPid;
439 2 : LWLockRelease(&apw_state->lock);
440 :
441 : /* Report our success, if we were able to finish. */
442 2 : if (!ShutdownRequestPending)
443 2 : ereport(LOG,
444 : (errmsg("autoprewarm successfully prewarmed %d of %d previously-loaded blocks",
445 : apw_state->prewarmed_blocks, num_elements)));
446 : }
447 :
448 : /*
449 : * Return the next block number of a specific relation and fork to read
450 : * according to the array of BlockInfoRecord.
451 : */
452 : static BlockNumber
453 552 : apw_read_stream_next_block(ReadStream *stream,
454 : void *callback_private_data,
455 : void *per_buffer_data)
456 : {
457 552 : AutoPrewarmReadStreamData *p = callback_private_data;
458 :
459 552 : CHECK_FOR_INTERRUPTS();
460 :
461 552 : while (p->pos < apw_state->prewarm_stop_idx)
462 : {
463 550 : BlockInfoRecord blk = p->block_info[p->pos];
464 :
465 550 : if (!have_free_buffer())
466 : {
467 0 : p->pos = apw_state->prewarm_stop_idx;
468 550 : return InvalidBlockNumber;
469 : }
470 :
471 550 : if (blk.tablespace != p->tablespace)
472 2 : return InvalidBlockNumber;
473 :
474 548 : if (blk.filenumber != p->filenumber)
475 94 : return InvalidBlockNumber;
476 :
477 454 : if (blk.forknum != p->forknum)
478 28 : return InvalidBlockNumber;
479 :
480 426 : p->pos++;
481 :
482 : /*
483 : * Check whether blocknum is valid and within fork file size.
484 : * Fast-forward through any invalid blocks. We want p->pos to reflect
485 : * the location of the next relation or fork before ending the stream.
486 : */
487 426 : if (blk.blocknum >= p->nblocks)
488 0 : continue;
489 :
490 426 : return blk.blocknum;
491 : }
492 :
493 2 : return InvalidBlockNumber;
494 : }
495 :
496 : /*
497 : * Prewarm all blocks for one database (and possibly also global objects, if
498 : * those got grouped with this database).
499 : */
500 : void
501 2 : autoprewarm_database_main(Datum main_arg)
502 : {
503 : BlockInfoRecord *block_info;
504 : int i;
505 : BlockInfoRecord blk;
506 : dsm_segment *seg;
507 :
508 : /* Establish signal handlers; once that's done, unblock signals. */
509 2 : pqsignal(SIGTERM, die);
510 2 : BackgroundWorkerUnblockSignals();
511 :
512 : /* Connect to correct database and get block information. */
513 2 : apw_init_shmem();
514 2 : seg = dsm_attach(apw_state->block_info_handle);
515 2 : if (seg == NULL)
516 0 : ereport(ERROR,
517 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
518 : errmsg("could not map dynamic shared memory segment")));
519 2 : BackgroundWorkerInitializeConnectionByOid(apw_state->database, InvalidOid, 0);
520 2 : block_info = (BlockInfoRecord *) dsm_segment_address(seg);
521 :
522 2 : i = apw_state->prewarm_start_idx;
523 2 : blk = block_info[i];
524 :
525 : /*
526 : * Loop until we run out of blocks to prewarm or until we run out of free
527 : * buffers.
528 : */
529 100 : while (i < apw_state->prewarm_stop_idx && have_free_buffer())
530 : {
531 98 : Oid tablespace = blk.tablespace;
532 98 : RelFileNumber filenumber = blk.filenumber;
533 : Oid reloid;
534 : Relation rel;
535 :
536 : /*
537 : * All blocks between prewarm_start_idx and prewarm_stop_idx should
538 : * belong either to global objects or the same database.
539 : */
540 : Assert(blk.database == apw_state->database || blk.database == 0);
541 :
542 98 : StartTransactionCommand();
543 :
544 98 : reloid = RelidByRelfilenumber(blk.tablespace, blk.filenumber);
545 196 : if (!OidIsValid(reloid) ||
546 98 : (rel = try_relation_open(reloid, AccessShareLock)) == NULL)
547 : {
548 : /* We failed to open the relation, so there is nothing to close. */
549 0 : CommitTransactionCommand();
550 :
551 : /*
552 : * Fast-forward to the next relation. We want to skip all of the
553 : * other records referencing this relation since we know we can't
554 : * open it. That way, we avoid repeatedly trying and failing to
555 : * open the same relation.
556 : */
557 0 : for (; i < apw_state->prewarm_stop_idx; i++)
558 : {
559 0 : blk = block_info[i];
560 0 : if (blk.tablespace != tablespace ||
561 0 : blk.filenumber != filenumber)
562 : break;
563 : }
564 :
565 : /* Time to try and open our newfound relation */
566 0 : continue;
567 : }
568 :
569 : /*
570 : * We have a relation; now let's loop until we find a valid fork of
571 : * the relation or we run out of free buffers. Once we've read from
572 : * all valid forks or run out of options, we'll close the relation and
573 : * move on.
574 : */
575 224 : while (i < apw_state->prewarm_stop_idx &&
576 222 : blk.tablespace == tablespace &&
577 346 : blk.filenumber == filenumber &&
578 126 : have_free_buffer())
579 : {
580 126 : ForkNumber forknum = blk.forknum;
581 : BlockNumber nblocks;
582 : struct AutoPrewarmReadStreamData p;
583 : ReadStream *stream;
584 : Buffer buf;
585 :
586 : /*
587 : * smgrexists is not safe for illegal forknum, hence check whether
588 : * the passed forknum is valid before using it in smgrexists.
589 : */
590 126 : if (blk.forknum <= InvalidForkNumber ||
591 126 : blk.forknum > MAX_FORKNUM ||
592 126 : !smgrexists(RelationGetSmgr(rel), blk.forknum))
593 : {
594 : /*
595 : * Fast-forward to the next fork. We want to skip all of the
596 : * other records referencing this fork since we already know
597 : * it's not valid.
598 : */
599 0 : for (; i < apw_state->prewarm_stop_idx; i++)
600 : {
601 0 : blk = block_info[i];
602 0 : if (blk.tablespace != tablespace ||
603 0 : blk.filenumber != filenumber ||
604 0 : blk.forknum != forknum)
605 : break;
606 : }
607 :
608 : /* Time to check if this newfound fork is valid */
609 0 : continue;
610 : }
611 :
612 126 : nblocks = RelationGetNumberOfBlocksInFork(rel, blk.forknum);
613 :
614 126 : p = (struct AutoPrewarmReadStreamData)
615 : {
616 : .block_info = block_info,
617 : .pos = i,
618 : .tablespace = tablespace,
619 : .filenumber = filenumber,
620 : .forknum = forknum,
621 : .nblocks = nblocks,
622 : };
623 :
624 126 : stream = read_stream_begin_relation(READ_STREAM_MAINTENANCE |
625 : READ_STREAM_DEFAULT |
626 : READ_STREAM_USE_BATCHING,
627 : NULL,
628 : rel,
629 : p.forknum,
630 : apw_read_stream_next_block,
631 : &p,
632 : 0);
633 :
634 : /*
635 : * Loop until we've prewarmed all the blocks from this fork. The
636 : * read stream callback will check that we still have free buffers
637 : * before requesting each block from the read stream API.
638 : */
639 552 : while ((buf = read_stream_next_buffer(stream, NULL)) != InvalidBuffer)
640 : {
641 426 : apw_state->prewarmed_blocks++;
642 426 : ReleaseBuffer(buf);
643 : }
644 :
645 126 : read_stream_end(stream);
646 :
647 : /* Advance i past all the blocks just prewarmed. */
648 126 : i = p.pos;
649 126 : blk = block_info[i];
650 : }
651 :
652 98 : relation_close(rel, AccessShareLock);
653 98 : CommitTransactionCommand();
654 : }
655 :
656 2 : dsm_detach(seg);
657 2 : }
658 :
659 : /*
660 : * Dump information on blocks in shared buffers. We use a text format here
661 : * so that it's easy to understand and even change the file contents if
662 : * necessary.
663 : * Returns the number of blocks dumped.
664 : */
665 : static int
666 6 : apw_dump_now(bool is_bgworker, bool dump_unlogged)
667 : {
668 : int num_blocks;
669 : int i;
670 : int ret;
671 : BlockInfoRecord *block_info_array;
672 : BufferDesc *bufHdr;
673 : FILE *file;
674 : char transient_dump_file_path[MAXPGPATH];
675 : pid_t pid;
676 :
677 6 : LWLockAcquire(&apw_state->lock, LW_EXCLUSIVE);
678 6 : pid = apw_state->pid_using_dumpfile;
679 6 : if (apw_state->pid_using_dumpfile == InvalidPid)
680 6 : apw_state->pid_using_dumpfile = MyProcPid;
681 6 : LWLockRelease(&apw_state->lock);
682 :
683 6 : if (pid != InvalidPid)
684 : {
685 0 : if (!is_bgworker)
686 0 : ereport(ERROR,
687 : (errmsg("could not perform block dump because dump file is being used by PID %d",
688 : (int) apw_state->pid_using_dumpfile)));
689 :
690 0 : ereport(LOG,
691 : (errmsg("skipping block dump because it is already being performed by PID %d",
692 : (int) apw_state->pid_using_dumpfile)));
693 0 : return 0;
694 : }
695 :
696 : block_info_array =
697 6 : (BlockInfoRecord *) palloc(sizeof(BlockInfoRecord) * NBuffers);
698 :
699 98310 : for (num_blocks = 0, i = 0; i < NBuffers; i++)
700 : {
701 : uint32 buf_state;
702 :
703 98304 : CHECK_FOR_INTERRUPTS();
704 :
705 98304 : bufHdr = GetBufferDescriptor(i);
706 :
707 : /* Lock each buffer header before inspecting. */
708 98304 : buf_state = LockBufHdr(bufHdr);
709 :
710 : /*
711 : * Unlogged tables will be automatically truncated after a crash or
712 : * unclean shutdown. In such cases we need not prewarm them. Dump them
713 : * only if requested by caller.
714 : */
715 98304 : if (buf_state & BM_TAG_VALID &&
716 1278 : ((buf_state & BM_PERMANENT) || dump_unlogged))
717 : {
718 1278 : block_info_array[num_blocks].database = bufHdr->tag.dbOid;
719 1278 : block_info_array[num_blocks].tablespace = bufHdr->tag.spcOid;
720 2556 : block_info_array[num_blocks].filenumber =
721 1278 : BufTagGetRelNumber(&bufHdr->tag);
722 2556 : block_info_array[num_blocks].forknum =
723 1278 : BufTagGetForkNum(&bufHdr->tag);
724 1278 : block_info_array[num_blocks].blocknum = bufHdr->tag.blockNum;
725 1278 : ++num_blocks;
726 : }
727 :
728 98304 : UnlockBufHdr(bufHdr, buf_state);
729 : }
730 :
731 6 : snprintf(transient_dump_file_path, MAXPGPATH, "%s.tmp", AUTOPREWARM_FILE);
732 6 : file = AllocateFile(transient_dump_file_path, "w");
733 6 : if (!file)
734 0 : ereport(ERROR,
735 : (errcode_for_file_access(),
736 : errmsg("could not open file \"%s\": %m",
737 : transient_dump_file_path)));
738 :
739 6 : ret = fprintf(file, "<<%d>>\n", num_blocks);
740 6 : if (ret < 0)
741 : {
742 0 : int save_errno = errno;
743 :
744 0 : FreeFile(file);
745 0 : unlink(transient_dump_file_path);
746 0 : errno = save_errno;
747 0 : ereport(ERROR,
748 : (errcode_for_file_access(),
749 : errmsg("could not write to file \"%s\": %m",
750 : transient_dump_file_path)));
751 : }
752 :
753 1284 : for (i = 0; i < num_blocks; i++)
754 : {
755 1278 : CHECK_FOR_INTERRUPTS();
756 :
757 1278 : ret = fprintf(file, "%u,%u,%u,%u,%u\n",
758 1278 : block_info_array[i].database,
759 1278 : block_info_array[i].tablespace,
760 1278 : block_info_array[i].filenumber,
761 1278 : (uint32) block_info_array[i].forknum,
762 1278 : block_info_array[i].blocknum);
763 1278 : if (ret < 0)
764 : {
765 0 : int save_errno = errno;
766 :
767 0 : FreeFile(file);
768 0 : unlink(transient_dump_file_path);
769 0 : errno = save_errno;
770 0 : ereport(ERROR,
771 : (errcode_for_file_access(),
772 : errmsg("could not write to file \"%s\": %m",
773 : transient_dump_file_path)));
774 : }
775 : }
776 :
777 6 : pfree(block_info_array);
778 :
779 : /*
780 : * Rename transient_dump_file_path to AUTOPREWARM_FILE to make things
781 : * permanent.
782 : */
783 6 : ret = FreeFile(file);
784 6 : if (ret != 0)
785 : {
786 0 : int save_errno = errno;
787 :
788 0 : unlink(transient_dump_file_path);
789 0 : errno = save_errno;
790 0 : ereport(ERROR,
791 : (errcode_for_file_access(),
792 : errmsg("could not close file \"%s\": %m",
793 : transient_dump_file_path)));
794 : }
795 :
796 6 : (void) durable_rename(transient_dump_file_path, AUTOPREWARM_FILE, ERROR);
797 6 : apw_state->pid_using_dumpfile = InvalidPid;
798 :
799 6 : ereport(DEBUG1,
800 : (errmsg_internal("wrote block details for %d blocks", num_blocks)));
801 6 : return num_blocks;
802 : }
803 :
804 : /*
805 : * SQL-callable function to launch autoprewarm.
806 : */
807 : Datum
808 0 : autoprewarm_start_worker(PG_FUNCTION_ARGS)
809 : {
810 : pid_t pid;
811 :
812 0 : if (!autoprewarm)
813 0 : ereport(ERROR,
814 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
815 : errmsg("autoprewarm is disabled")));
816 :
817 0 : apw_init_shmem();
818 0 : LWLockAcquire(&apw_state->lock, LW_EXCLUSIVE);
819 0 : pid = apw_state->bgworker_pid;
820 0 : LWLockRelease(&apw_state->lock);
821 :
822 0 : if (pid != InvalidPid)
823 0 : ereport(ERROR,
824 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
825 : errmsg("autoprewarm worker is already running under PID %d",
826 : (int) pid)));
827 :
828 0 : apw_start_leader_worker();
829 :
830 0 : PG_RETURN_VOID();
831 : }
832 :
833 : /*
834 : * SQL-callable function to perform an immediate block dump.
835 : *
836 : * Note: this is declared to return int8, as insurance against some
837 : * very distant day when we might make NBuffers wider than int.
838 : */
839 : Datum
840 2 : autoprewarm_dump_now(PG_FUNCTION_ARGS)
841 : {
842 : int num_blocks;
843 :
844 2 : apw_init_shmem();
845 :
846 2 : PG_ENSURE_ERROR_CLEANUP(apw_detach_shmem, 0);
847 : {
848 2 : num_blocks = apw_dump_now(false, true);
849 : }
850 2 : PG_END_ENSURE_ERROR_CLEANUP(apw_detach_shmem, 0);
851 :
852 2 : PG_RETURN_INT64((int64) num_blocks);
853 : }
854 :
855 : static void
856 4 : apw_init_state(void *ptr)
857 : {
858 4 : AutoPrewarmSharedState *state = (AutoPrewarmSharedState *) ptr;
859 :
860 4 : LWLockInitialize(&state->lock, LWLockNewTrancheId());
861 4 : state->bgworker_pid = InvalidPid;
862 4 : state->pid_using_dumpfile = InvalidPid;
863 4 : }
864 :
865 : /*
866 : * Allocate and initialize autoprewarm related shared memory, if not already
867 : * done, and set up backend-local pointer to that state. Returns true if an
868 : * existing shared memory segment was found.
869 : */
870 : static bool
871 8 : apw_init_shmem(void)
872 : {
873 : bool found;
874 :
875 8 : apw_state = GetNamedDSMSegment("autoprewarm",
876 : sizeof(AutoPrewarmSharedState),
877 : apw_init_state,
878 : &found);
879 8 : LWLockRegisterTranche(apw_state->lock.tranche, "autoprewarm");
880 :
881 8 : return found;
882 : }
883 :
884 : /*
885 : * Clear our PID from autoprewarm shared state.
886 : */
887 : static void
888 4 : apw_detach_shmem(int code, Datum arg)
889 : {
890 4 : LWLockAcquire(&apw_state->lock, LW_EXCLUSIVE);
891 4 : if (apw_state->pid_using_dumpfile == MyProcPid)
892 0 : apw_state->pid_using_dumpfile = InvalidPid;
893 4 : if (apw_state->bgworker_pid == MyProcPid)
894 4 : apw_state->bgworker_pid = InvalidPid;
895 4 : LWLockRelease(&apw_state->lock);
896 4 : }
897 :
898 : /*
899 : * Start autoprewarm leader worker process.
900 : */
901 : static void
902 4 : apw_start_leader_worker(void)
903 : {
904 4 : BackgroundWorker worker = {0};
905 : BackgroundWorkerHandle *handle;
906 : BgwHandleStatus status;
907 : pid_t pid;
908 :
909 4 : worker.bgw_flags = BGWORKER_SHMEM_ACCESS;
910 4 : worker.bgw_start_time = BgWorkerStart_ConsistentState;
911 4 : strcpy(worker.bgw_library_name, "pg_prewarm");
912 4 : strcpy(worker.bgw_function_name, "autoprewarm_main");
913 4 : strcpy(worker.bgw_name, "autoprewarm leader");
914 4 : strcpy(worker.bgw_type, "autoprewarm leader");
915 :
916 4 : if (process_shared_preload_libraries_in_progress)
917 : {
918 4 : RegisterBackgroundWorker(&worker);
919 4 : return;
920 : }
921 :
922 : /* must set notify PID to wait for startup */
923 0 : worker.bgw_notify_pid = MyProcPid;
924 :
925 0 : if (!RegisterDynamicBackgroundWorker(&worker, &handle))
926 0 : ereport(ERROR,
927 : (errcode(ERRCODE_INSUFFICIENT_RESOURCES),
928 : errmsg("could not register background process"),
929 : errhint("You may need to increase \"max_worker_processes\".")));
930 :
931 0 : status = WaitForBackgroundWorkerStartup(handle, &pid);
932 0 : if (status != BGWH_STARTED)
933 0 : ereport(ERROR,
934 : (errcode(ERRCODE_INSUFFICIENT_RESOURCES),
935 : errmsg("could not start background process"),
936 : errhint("More details may be available in the server log.")));
937 : }
938 :
939 : /*
940 : * Start autoprewarm per-database worker process.
941 : */
942 : static void
943 2 : apw_start_database_worker(void)
944 : {
945 2 : BackgroundWorker worker = {0};
946 : BackgroundWorkerHandle *handle;
947 :
948 2 : worker.bgw_flags =
949 : BGWORKER_SHMEM_ACCESS | BGWORKER_BACKEND_DATABASE_CONNECTION;
950 2 : worker.bgw_start_time = BgWorkerStart_ConsistentState;
951 2 : worker.bgw_restart_time = BGW_NEVER_RESTART;
952 2 : strcpy(worker.bgw_library_name, "pg_prewarm");
953 2 : strcpy(worker.bgw_function_name, "autoprewarm_database_main");
954 2 : strcpy(worker.bgw_name, "autoprewarm worker");
955 2 : strcpy(worker.bgw_type, "autoprewarm worker");
956 :
957 : /* must set notify PID to wait for shutdown */
958 2 : worker.bgw_notify_pid = MyProcPid;
959 :
960 2 : if (!RegisterDynamicBackgroundWorker(&worker, &handle))
961 0 : ereport(ERROR,
962 : (errcode(ERRCODE_INSUFFICIENT_RESOURCES),
963 : errmsg("registering dynamic bgworker autoprewarm failed"),
964 : errhint("Consider increasing the configuration parameter \"%s\".", "max_worker_processes")));
965 :
966 : /*
967 : * Ignore return value; if it fails, postmaster has died, but we have
968 : * checks for that elsewhere.
969 : */
970 2 : WaitForBackgroundWorkerShutdown(handle);
971 2 : }
972 :
973 : /* Compare member elements to check whether they are not equal. */
974 : #define cmp_member_elem(fld) \
975 : do { \
976 : if (a->fld < b->fld) \
977 : return -1; \
978 : else if (a->fld > b->fld) \
979 : return 1; \
980 : } while(0)
981 :
982 : /*
983 : * apw_compare_blockinfo
984 : *
985 : * We depend on all records for a particular database being consecutive
986 : * in the dump file; each per-database worker will preload blocks until
987 : * it sees a block for some other database. Sorting by tablespace,
988 : * filenumber, forknum, and blocknum isn't critical for correctness, but
989 : * helps us get a sequential I/O pattern.
990 : */
991 : static int
992 3470 : apw_compare_blockinfo(const void *p, const void *q)
993 : {
994 3470 : const BlockInfoRecord *a = (const BlockInfoRecord *) p;
995 3470 : const BlockInfoRecord *b = (const BlockInfoRecord *) q;
996 :
997 3470 : cmp_member_elem(database);
998 3352 : cmp_member_elem(tablespace);
999 3352 : cmp_member_elem(filenumber);
1000 1420 : cmp_member_elem(forknum);
1001 1238 : cmp_member_elem(blocknum);
1002 :
1003 0 : return 0;
1004 : }
|
