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) 2016-2026, 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 0 : PG_FUNCTION_INFO_V1(autoprewarm_start_worker);
105 0 : 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 0 : _PG_init(void)
127 : {
128 0 : 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 0 : if (!process_shared_preload_libraries_in_progress)
141 0 : return;
142 :
143 : /* can't define PGC_POSTMASTER variable after startup */
144 0 : 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 0 : MarkGUCPrefixReserved("pg_prewarm");
156 :
157 : /* Register autoprewarm worker, if enabled. */
158 0 : if (autoprewarm)
159 0 : apw_start_leader_worker();
160 0 : }
161 :
162 : /*
163 : * Main entry point for the leader autoprewarm process. Per-database workers
164 : * have a separate entry point.
165 : */
166 : void
167 0 : autoprewarm_main(Datum main_arg)
168 : {
169 0 : bool first_time = true;
170 0 : bool final_dump_allowed = true;
171 0 : TimestampTz last_dump_time = 0;
172 :
173 : /* Establish signal handlers; once that's done, unblock signals. */
174 0 : pqsignal(SIGTERM, SignalHandlerForShutdownRequest);
175 0 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
176 0 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
177 0 : BackgroundWorkerUnblockSignals();
178 :
179 : /* Create (if necessary) and attach to our shared memory area. */
180 0 : 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 0 : 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 0 : LWLockAcquire(&apw_state->lock, LW_EXCLUSIVE);
197 0 : 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 0 : apw_state->bgworker_pid = MyProcPid;
206 0 : 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 0 : if (first_time)
223 : {
224 0 : apw_load_buffers();
225 0 : final_dump_allowed = !ShutdownRequestPending;
226 0 : last_dump_time = GetCurrentTimestamp();
227 0 : }
228 :
229 : /* Periodically dump buffers until terminated. */
230 0 : while (!ShutdownRequestPending)
231 : {
232 : /* In case of a SIGHUP, just reload the configuration. */
233 0 : if (ConfigReloadPending)
234 : {
235 0 : ConfigReloadPending = false;
236 0 : ProcessConfigFile(PGC_SIGHUP);
237 0 : }
238 :
239 0 : if (autoprewarm_interval <= 0)
240 : {
241 : /* We're only dumping at shutdown, so just wait forever. */
242 0 : (void) WaitLatch(MyLatch,
243 : WL_LATCH_SET | WL_EXIT_ON_PM_DEATH,
244 : -1L,
245 : PG_WAIT_EXTENSION);
246 0 : }
247 : else
248 : {
249 0 : TimestampTz next_dump_time;
250 0 : 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 0 : delay_in_ms =
257 0 : TimestampDifferenceMilliseconds(GetCurrentTimestamp(),
258 0 : 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 0 : delay_in_ms,
272 : PG_WAIT_EXTENSION);
273 0 : }
274 :
275 : /* Reset the latch, loop. */
276 0 : 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 0 : if (final_dump_allowed)
284 0 : apw_dump_now(true, true);
285 0 : }
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 0 : apw_load_buffers(void)
293 : {
294 0 : FILE *file = NULL;
295 0 : int num_elements,
296 : i;
297 0 : BlockInfoRecord *blkinfo;
298 0 : 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 0 : LWLockAcquire(&apw_state->lock, LW_EXCLUSIVE);
305 0 : if (apw_state->pid_using_dumpfile == InvalidPid)
306 0 : 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 0 : return;
314 : }
315 0 : 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 0 : file = AllocateFile(AUTOPREWARM_FILE, "r");
322 0 : if (!file)
323 : {
324 0 : if (errno == ENOENT)
325 : {
326 0 : LWLockAcquire(&apw_state->lock, LW_EXCLUSIVE);
327 0 : apw_state->pid_using_dumpfile = InvalidPid;
328 0 : LWLockRelease(&apw_state->lock);
329 0 : 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 0 : }
336 :
337 : /* First line of the file is a record count. */
338 0 : 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 0 : seg = dsm_create(sizeof(BlockInfoRecord) * num_elements, 0);
346 0 : blkinfo = (BlockInfoRecord *) dsm_segment_address(seg);
347 :
348 : /* Read records, one per line. */
349 0 : for (i = 0; i < num_elements; i++)
350 : {
351 0 : unsigned forknum;
352 :
353 0 : if (fscanf(file, "%u,%u,%u,%u,%u\n", &blkinfo[i].database,
354 0 : &blkinfo[i].tablespace, &blkinfo[i].filenumber,
355 0 : &forknum, &blkinfo[i].blocknum) != 5)
356 0 : ereport(ERROR,
357 : (errmsg("autoprewarm block dump file is corrupted at line %d",
358 : i + 1)));
359 0 : blkinfo[i].forknum = forknum;
360 0 : }
361 :
362 0 : FreeFile(file);
363 :
364 : /* Sort the blocks to be loaded. */
365 0 : qsort(blkinfo, num_elements, sizeof(BlockInfoRecord),
366 : apw_compare_blockinfo);
367 :
368 : /* Populate shared memory state. */
369 0 : apw_state->block_info_handle = dsm_segment_handle(seg);
370 0 : apw_state->prewarm_start_idx = apw_state->prewarm_stop_idx = 0;
371 0 : apw_state->prewarmed_blocks = 0;
372 :
373 : /* Don't prewarm more than we can fit. */
374 0 : if (num_elements > NBuffers)
375 : {
376 0 : num_elements = NBuffers;
377 0 : ereport(LOG,
378 : (errmsg("autoprewarm capping prewarmed blocks to %d (shared_buffers size)",
379 : NBuffers)));
380 0 : }
381 :
382 : /* Get the info position of the first block of the next database. */
383 0 : while (apw_state->prewarm_start_idx < num_elements)
384 : {
385 0 : int j = apw_state->prewarm_start_idx;
386 0 : Oid current_db = blkinfo[j].database;
387 :
388 : /*
389 : * Advance the prewarm_stop_idx to the first BlockInfoRecord that does
390 : * not belong to this database.
391 : */
392 0 : j++;
393 0 : while (j < num_elements)
394 : {
395 0 : if (current_db != blkinfo[j].database)
396 : {
397 : /*
398 : * Combine BlockInfoRecords for global objects with those of
399 : * the database.
400 : */
401 0 : if (current_db != InvalidOid)
402 0 : break;
403 0 : current_db = blkinfo[j].database;
404 0 : }
405 :
406 0 : j++;
407 : }
408 :
409 : /*
410 : * If we reach this point with current_db == InvalidOid, then only
411 : * BlockInfoRecords belonging to global objects exist. We can't
412 : * prewarm without a database connection, so just bail out.
413 : */
414 0 : if (current_db == InvalidOid)
415 0 : break;
416 :
417 : /* Configure stop point and database for next per-database worker. */
418 0 : apw_state->prewarm_stop_idx = j;
419 0 : apw_state->database = current_db;
420 0 : Assert(apw_state->prewarm_start_idx < apw_state->prewarm_stop_idx);
421 :
422 : /*
423 : * Likewise, don't launch if we've already been told to shut down.
424 : * (The launch would fail anyway, but we might as well skip it.)
425 : */
426 0 : if (ShutdownRequestPending)
427 0 : break;
428 :
429 : /*
430 : * Start a per-database worker to load blocks for this database; this
431 : * function will return once the per-database worker exits.
432 : */
433 0 : apw_start_database_worker();
434 :
435 : /* Prepare for next database. */
436 0 : apw_state->prewarm_start_idx = apw_state->prewarm_stop_idx;
437 0 : }
438 :
439 : /* Clean up. */
440 0 : dsm_detach(seg);
441 0 : LWLockAcquire(&apw_state->lock, LW_EXCLUSIVE);
442 0 : apw_state->block_info_handle = DSM_HANDLE_INVALID;
443 0 : apw_state->pid_using_dumpfile = InvalidPid;
444 0 : LWLockRelease(&apw_state->lock);
445 :
446 : /* Report our success, if we were able to finish. */
447 0 : if (!ShutdownRequestPending)
448 0 : ereport(LOG,
449 : (errmsg("autoprewarm successfully prewarmed %d of %d previously-loaded blocks",
450 : apw_state->prewarmed_blocks, num_elements)));
451 0 : }
452 :
453 : /*
454 : * Return the next block number of a specific relation and fork to read
455 : * according to the array of BlockInfoRecord.
456 : */
457 : static BlockNumber
458 0 : apw_read_stream_next_block(ReadStream *stream,
459 : void *callback_private_data,
460 : void *per_buffer_data)
461 : {
462 0 : AutoPrewarmReadStreamData *p = callback_private_data;
463 :
464 0 : CHECK_FOR_INTERRUPTS();
465 :
466 0 : while (p->pos < apw_state->prewarm_stop_idx)
467 : {
468 0 : BlockInfoRecord blk = p->block_info[p->pos];
469 :
470 0 : if (blk.tablespace != p->tablespace)
471 0 : return InvalidBlockNumber;
472 :
473 0 : if (blk.filenumber != p->filenumber)
474 0 : return InvalidBlockNumber;
475 :
476 0 : if (blk.forknum != p->forknum)
477 0 : return InvalidBlockNumber;
478 :
479 0 : p->pos++;
480 :
481 : /*
482 : * Check whether blocknum is valid and within fork file size.
483 : * Fast-forward through any invalid blocks. We want p->pos to reflect
484 : * the location of the next relation or fork before ending the stream.
485 : */
486 0 : if (blk.blocknum >= p->nblocks)
487 0 : continue;
488 :
489 0 : return blk.blocknum;
490 0 : }
491 :
492 0 : return InvalidBlockNumber;
493 0 : }
494 :
495 : /*
496 : * Prewarm all blocks for one database (and possibly also global objects, if
497 : * those got grouped with this database).
498 : */
499 : void
500 0 : autoprewarm_database_main(Datum main_arg)
501 : {
502 0 : BlockInfoRecord *block_info;
503 0 : int i;
504 0 : BlockInfoRecord blk;
505 0 : dsm_segment *seg;
506 :
507 : /* Establish signal handlers; once that's done, unblock signals. */
508 0 : pqsignal(SIGTERM, die);
509 0 : BackgroundWorkerUnblockSignals();
510 :
511 : /* Connect to correct database and get block information. */
512 0 : apw_init_shmem();
513 0 : seg = dsm_attach(apw_state->block_info_handle);
514 0 : if (seg == NULL)
515 0 : ereport(ERROR,
516 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
517 : errmsg("could not map dynamic shared memory segment")));
518 0 : BackgroundWorkerInitializeConnectionByOid(apw_state->database, InvalidOid, 0);
519 0 : block_info = (BlockInfoRecord *) dsm_segment_address(seg);
520 :
521 0 : i = apw_state->prewarm_start_idx;
522 0 : blk = block_info[i];
523 :
524 : /*
525 : * Loop until we run out of blocks to prewarm or until we run out of
526 : * buffers.
527 : */
528 0 : while (i < apw_state->prewarm_stop_idx)
529 : {
530 0 : Oid tablespace = blk.tablespace;
531 0 : RelFileNumber filenumber = blk.filenumber;
532 0 : Oid reloid;
533 0 : Relation rel;
534 :
535 : /*
536 : * All blocks between prewarm_start_idx and prewarm_stop_idx should
537 : * belong either to global objects or the same database.
538 : */
539 0 : Assert(blk.database == apw_state->database || blk.database == 0);
540 :
541 0 : StartTransactionCommand();
542 :
543 0 : reloid = RelidByRelfilenumber(blk.tablespace, blk.filenumber);
544 0 : if (!OidIsValid(reloid) ||
545 0 : (rel = try_relation_open(reloid, AccessShareLock)) == NULL)
546 : {
547 : /* We failed to open the relation, so there is nothing to close. */
548 0 : CommitTransactionCommand();
549 :
550 : /*
551 : * Fast-forward to the next relation. We want to skip all of the
552 : * other records referencing this relation since we know we can't
553 : * open it. That way, we avoid repeatedly trying and failing to
554 : * open the same relation.
555 : */
556 0 : for (; i < apw_state->prewarm_stop_idx; i++)
557 : {
558 0 : blk = block_info[i];
559 0 : if (blk.tablespace != tablespace ||
560 0 : blk.filenumber != filenumber)
561 0 : break;
562 0 : }
563 :
564 : /* Time to try and open our newfound relation */
565 0 : continue;
566 : }
567 :
568 : /*
569 : * We have a relation; now let's loop until we find a valid fork of
570 : * the relation or we run out of buffers. Once we've read from all
571 : * valid forks or run out of options, we'll close the relation and
572 : * move on.
573 : */
574 0 : while (i < apw_state->prewarm_stop_idx &&
575 0 : blk.tablespace == tablespace &&
576 0 : blk.filenumber == filenumber)
577 : {
578 0 : ForkNumber forknum = blk.forknum;
579 0 : BlockNumber nblocks;
580 0 : struct AutoPrewarmReadStreamData p;
581 0 : ReadStream *stream;
582 0 : Buffer buf;
583 :
584 : /*
585 : * smgrexists is not safe for illegal forknum, hence check whether
586 : * the passed forknum is valid before using it in smgrexists.
587 : */
588 0 : if (blk.forknum <= InvalidForkNumber ||
589 0 : blk.forknum > MAX_FORKNUM ||
590 0 : !smgrexists(RelationGetSmgr(rel), blk.forknum))
591 : {
592 : /*
593 : * Fast-forward to the next fork. We want to skip all of the
594 : * other records referencing this fork since we already know
595 : * it's not valid.
596 : */
597 0 : for (; i < apw_state->prewarm_stop_idx; i++)
598 : {
599 0 : blk = block_info[i];
600 0 : if (blk.tablespace != tablespace ||
601 0 : blk.filenumber != filenumber ||
602 0 : blk.forknum != forknum)
603 0 : break;
604 0 : }
605 :
606 : /* Time to check if this newfound fork is valid */
607 0 : continue;
608 : }
609 :
610 0 : nblocks = RelationGetNumberOfBlocksInFork(rel, blk.forknum);
611 :
612 0 : p = (struct AutoPrewarmReadStreamData)
613 0 : {
614 0 : .block_info = block_info,
615 0 : .pos = i,
616 0 : .tablespace = tablespace,
617 0 : .filenumber = filenumber,
618 0 : .forknum = forknum,
619 0 : .nblocks = nblocks,
620 : };
621 :
622 0 : stream = read_stream_begin_relation(READ_STREAM_MAINTENANCE |
623 : READ_STREAM_DEFAULT |
624 : READ_STREAM_USE_BATCHING,
625 : NULL,
626 0 : rel,
627 0 : p.forknum,
628 : apw_read_stream_next_block,
629 : &p,
630 : 0);
631 :
632 : /*
633 : * Loop until we've prewarmed all the blocks from this fork. The
634 : * read stream callback will check that we still have free buffers
635 : * before requesting each block from the read stream API.
636 : */
637 0 : while ((buf = read_stream_next_buffer(stream, NULL)) != InvalidBuffer)
638 : {
639 0 : apw_state->prewarmed_blocks++;
640 0 : ReleaseBuffer(buf);
641 : }
642 :
643 0 : read_stream_end(stream);
644 :
645 : /* Advance i past all the blocks just prewarmed. */
646 0 : i = p.pos;
647 0 : blk = block_info[i];
648 0 : }
649 :
650 0 : relation_close(rel, AccessShareLock);
651 0 : CommitTransactionCommand();
652 0 : }
653 :
654 0 : dsm_detach(seg);
655 0 : }
656 :
657 : /*
658 : * Dump information on blocks in shared buffers. We use a text format here
659 : * so that it's easy to understand and even change the file contents if
660 : * necessary.
661 : * Returns the number of blocks dumped.
662 : */
663 : static int
664 0 : apw_dump_now(bool is_bgworker, bool dump_unlogged)
665 : {
666 0 : int num_blocks;
667 0 : int i;
668 0 : int ret;
669 0 : BlockInfoRecord *block_info_array;
670 0 : BufferDesc *bufHdr;
671 0 : FILE *file;
672 0 : char transient_dump_file_path[MAXPGPATH];
673 0 : pid_t pid;
674 :
675 0 : LWLockAcquire(&apw_state->lock, LW_EXCLUSIVE);
676 0 : pid = apw_state->pid_using_dumpfile;
677 0 : if (apw_state->pid_using_dumpfile == InvalidPid)
678 0 : apw_state->pid_using_dumpfile = MyProcPid;
679 0 : LWLockRelease(&apw_state->lock);
680 :
681 0 : if (pid != InvalidPid)
682 : {
683 0 : if (!is_bgworker)
684 0 : ereport(ERROR,
685 : (errmsg("could not perform block dump because dump file is being used by PID %d",
686 : (int) apw_state->pid_using_dumpfile)));
687 :
688 0 : ereport(LOG,
689 : (errmsg("skipping block dump because it is already being performed by PID %d",
690 : (int) apw_state->pid_using_dumpfile)));
691 0 : return 0;
692 : }
693 :
694 : /*
695 : * With sufficiently large shared_buffers, allocation will exceed 1GB, so
696 : * allow for a huge allocation to prevent outright failure.
697 : *
698 : * (In the future, it might be a good idea to redesign this to use a more
699 : * memory-efficient data structure.)
700 : */
701 0 : block_info_array = (BlockInfoRecord *)
702 0 : palloc_extended((sizeof(BlockInfoRecord) * NBuffers), MCXT_ALLOC_HUGE);
703 :
704 0 : for (num_blocks = 0, i = 0; i < NBuffers; i++)
705 : {
706 0 : uint64 buf_state;
707 :
708 0 : CHECK_FOR_INTERRUPTS();
709 :
710 0 : bufHdr = GetBufferDescriptor(i);
711 :
712 : /* Lock each buffer header before inspecting. */
713 0 : buf_state = LockBufHdr(bufHdr);
714 :
715 : /*
716 : * Unlogged tables will be automatically truncated after a crash or
717 : * unclean shutdown. In such cases we need not prewarm them. Dump them
718 : * only if requested by caller.
719 : */
720 0 : if (buf_state & BM_TAG_VALID &&
721 0 : ((buf_state & BM_PERMANENT) || dump_unlogged))
722 : {
723 0 : block_info_array[num_blocks].database = bufHdr->tag.dbOid;
724 0 : block_info_array[num_blocks].tablespace = bufHdr->tag.spcOid;
725 0 : block_info_array[num_blocks].filenumber =
726 0 : BufTagGetRelNumber(&bufHdr->tag);
727 0 : block_info_array[num_blocks].forknum =
728 0 : BufTagGetForkNum(&bufHdr->tag);
729 0 : block_info_array[num_blocks].blocknum = bufHdr->tag.blockNum;
730 0 : ++num_blocks;
731 0 : }
732 :
733 0 : UnlockBufHdr(bufHdr);
734 0 : }
735 :
736 0 : snprintf(transient_dump_file_path, MAXPGPATH, "%s.tmp", AUTOPREWARM_FILE);
737 0 : file = AllocateFile(transient_dump_file_path, "w");
738 0 : if (!file)
739 0 : ereport(ERROR,
740 : (errcode_for_file_access(),
741 : errmsg("could not open file \"%s\": %m",
742 : transient_dump_file_path)));
743 :
744 0 : ret = fprintf(file, "<<%d>>\n", num_blocks);
745 0 : if (ret < 0)
746 : {
747 0 : int save_errno = errno;
748 :
749 0 : FreeFile(file);
750 0 : unlink(transient_dump_file_path);
751 0 : errno = save_errno;
752 0 : ereport(ERROR,
753 : (errcode_for_file_access(),
754 : errmsg("could not write to file \"%s\": %m",
755 : transient_dump_file_path)));
756 0 : }
757 :
758 0 : for (i = 0; i < num_blocks; i++)
759 : {
760 0 : CHECK_FOR_INTERRUPTS();
761 :
762 0 : ret = fprintf(file, "%u,%u,%u,%u,%u\n",
763 0 : block_info_array[i].database,
764 0 : block_info_array[i].tablespace,
765 0 : block_info_array[i].filenumber,
766 0 : (uint32) block_info_array[i].forknum,
767 0 : block_info_array[i].blocknum);
768 0 : if (ret < 0)
769 : {
770 0 : int save_errno = errno;
771 :
772 0 : FreeFile(file);
773 0 : unlink(transient_dump_file_path);
774 0 : errno = save_errno;
775 0 : ereport(ERROR,
776 : (errcode_for_file_access(),
777 : errmsg("could not write to file \"%s\": %m",
778 : transient_dump_file_path)));
779 0 : }
780 0 : }
781 :
782 0 : pfree(block_info_array);
783 :
784 : /*
785 : * Rename transient_dump_file_path to AUTOPREWARM_FILE to make things
786 : * permanent.
787 : */
788 0 : ret = FreeFile(file);
789 0 : if (ret != 0)
790 : {
791 0 : int save_errno = errno;
792 :
793 0 : unlink(transient_dump_file_path);
794 0 : errno = save_errno;
795 0 : ereport(ERROR,
796 : (errcode_for_file_access(),
797 : errmsg("could not close file \"%s\": %m",
798 : transient_dump_file_path)));
799 0 : }
800 :
801 0 : (void) durable_rename(transient_dump_file_path, AUTOPREWARM_FILE, ERROR);
802 0 : apw_state->pid_using_dumpfile = InvalidPid;
803 :
804 0 : ereport(DEBUG1,
805 : (errmsg_internal("wrote block details for %d blocks", num_blocks)));
806 0 : return num_blocks;
807 0 : }
808 :
809 : /*
810 : * SQL-callable function to launch autoprewarm.
811 : */
812 : Datum
813 0 : autoprewarm_start_worker(PG_FUNCTION_ARGS)
814 : {
815 0 : pid_t pid;
816 :
817 0 : if (!autoprewarm)
818 0 : ereport(ERROR,
819 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
820 : errmsg("autoprewarm is disabled")));
821 :
822 0 : apw_init_shmem();
823 0 : LWLockAcquire(&apw_state->lock, LW_EXCLUSIVE);
824 0 : pid = apw_state->bgworker_pid;
825 0 : LWLockRelease(&apw_state->lock);
826 :
827 0 : if (pid != InvalidPid)
828 0 : ereport(ERROR,
829 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
830 : errmsg("autoprewarm worker is already running under PID %d",
831 : (int) pid)));
832 :
833 0 : apw_start_leader_worker();
834 :
835 0 : PG_RETURN_VOID();
836 0 : }
837 :
838 : /*
839 : * SQL-callable function to perform an immediate block dump.
840 : *
841 : * Note: this is declared to return int8, as insurance against some
842 : * very distant day when we might make NBuffers wider than int.
843 : */
844 : Datum
845 0 : autoprewarm_dump_now(PG_FUNCTION_ARGS)
846 : {
847 0 : int num_blocks;
848 :
849 0 : apw_init_shmem();
850 :
851 0 : PG_ENSURE_ERROR_CLEANUP(apw_detach_shmem, 0);
852 : {
853 0 : num_blocks = apw_dump_now(false, true);
854 : }
855 0 : PG_END_ENSURE_ERROR_CLEANUP(apw_detach_shmem, 0);
856 :
857 0 : PG_RETURN_INT64((int64) num_blocks);
858 0 : }
859 :
860 : static void
861 0 : apw_init_state(void *ptr, void *arg)
862 : {
863 0 : AutoPrewarmSharedState *state = (AutoPrewarmSharedState *) ptr;
864 :
865 0 : LWLockInitialize(&state->lock, LWLockNewTrancheId("autoprewarm"));
866 0 : state->bgworker_pid = InvalidPid;
867 0 : state->pid_using_dumpfile = InvalidPid;
868 0 : }
869 :
870 : /*
871 : * Allocate and initialize autoprewarm related shared memory, if not already
872 : * done, and set up backend-local pointer to that state. Returns true if an
873 : * existing shared memory segment was found.
874 : */
875 : static bool
876 0 : apw_init_shmem(void)
877 : {
878 0 : bool found;
879 :
880 0 : apw_state = GetNamedDSMSegment("autoprewarm",
881 : sizeof(AutoPrewarmSharedState),
882 : apw_init_state,
883 : &found, NULL);
884 :
885 0 : return found;
886 0 : }
887 :
888 : /*
889 : * Clear our PID from autoprewarm shared state.
890 : */
891 : static void
892 0 : apw_detach_shmem(int code, Datum arg)
893 : {
894 0 : LWLockAcquire(&apw_state->lock, LW_EXCLUSIVE);
895 0 : if (apw_state->pid_using_dumpfile == MyProcPid)
896 0 : apw_state->pid_using_dumpfile = InvalidPid;
897 0 : if (apw_state->bgworker_pid == MyProcPid)
898 0 : apw_state->bgworker_pid = InvalidPid;
899 0 : LWLockRelease(&apw_state->lock);
900 0 : }
901 :
902 : /*
903 : * Start autoprewarm leader worker process.
904 : */
905 : static void
906 0 : apw_start_leader_worker(void)
907 : {
908 0 : BackgroundWorker worker = {0};
909 0 : BackgroundWorkerHandle *handle;
910 0 : BgwHandleStatus status;
911 0 : pid_t pid;
912 :
913 0 : worker.bgw_flags = BGWORKER_SHMEM_ACCESS;
914 0 : worker.bgw_start_time = BgWorkerStart_ConsistentState;
915 0 : strcpy(worker.bgw_library_name, "pg_prewarm");
916 0 : strcpy(worker.bgw_function_name, "autoprewarm_main");
917 0 : strcpy(worker.bgw_name, "autoprewarm leader");
918 0 : strcpy(worker.bgw_type, "autoprewarm leader");
919 :
920 0 : if (process_shared_preload_libraries_in_progress)
921 : {
922 0 : RegisterBackgroundWorker(&worker);
923 0 : return;
924 : }
925 :
926 : /* must set notify PID to wait for startup */
927 0 : worker.bgw_notify_pid = MyProcPid;
928 :
929 0 : if (!RegisterDynamicBackgroundWorker(&worker, &handle))
930 0 : ereport(ERROR,
931 : (errcode(ERRCODE_INSUFFICIENT_RESOURCES),
932 : errmsg("could not register background process"),
933 : errhint("You may need to increase \"max_worker_processes\".")));
934 :
935 0 : status = WaitForBackgroundWorkerStartup(handle, &pid);
936 0 : if (status != BGWH_STARTED)
937 0 : ereport(ERROR,
938 : (errcode(ERRCODE_INSUFFICIENT_RESOURCES),
939 : errmsg("could not start background process"),
940 : errhint("More details may be available in the server log.")));
941 0 : }
942 :
943 : /*
944 : * Start autoprewarm per-database worker process.
945 : */
946 : static void
947 0 : apw_start_database_worker(void)
948 : {
949 0 : BackgroundWorker worker = {0};
950 0 : BackgroundWorkerHandle *handle;
951 :
952 0 : worker.bgw_flags =
953 : BGWORKER_SHMEM_ACCESS | BGWORKER_BACKEND_DATABASE_CONNECTION;
954 0 : worker.bgw_start_time = BgWorkerStart_ConsistentState;
955 0 : worker.bgw_restart_time = BGW_NEVER_RESTART;
956 0 : strcpy(worker.bgw_library_name, "pg_prewarm");
957 0 : strcpy(worker.bgw_function_name, "autoprewarm_database_main");
958 0 : strcpy(worker.bgw_name, "autoprewarm worker");
959 0 : strcpy(worker.bgw_type, "autoprewarm worker");
960 :
961 : /* must set notify PID to wait for shutdown */
962 0 : worker.bgw_notify_pid = MyProcPid;
963 :
964 0 : if (!RegisterDynamicBackgroundWorker(&worker, &handle))
965 0 : ereport(ERROR,
966 : (errcode(ERRCODE_INSUFFICIENT_RESOURCES),
967 : errmsg("registering dynamic bgworker autoprewarm failed"),
968 : errhint("Consider increasing the configuration parameter \"%s\".", "max_worker_processes")));
969 :
970 : /*
971 : * Ignore return value; if it fails, postmaster has died, but we have
972 : * checks for that elsewhere.
973 : */
974 0 : WaitForBackgroundWorkerShutdown(handle);
975 0 : }
976 :
977 : /* Compare member elements to check whether they are not equal. */
978 : #define cmp_member_elem(fld) \
979 : do { \
980 : if (a->fld < b->fld) \
981 : return -1; \
982 : else if (a->fld > b->fld) \
983 : return 1; \
984 : } while(0)
985 :
986 : /*
987 : * apw_compare_blockinfo
988 : *
989 : * We depend on all records for a particular database being consecutive
990 : * in the dump file; each per-database worker will preload blocks until
991 : * it sees a block for some other database. Sorting by tablespace,
992 : * filenumber, forknum, and blocknum isn't critical for correctness, but
993 : * helps us get a sequential I/O pattern.
994 : */
995 : static int
996 0 : apw_compare_blockinfo(const void *p, const void *q)
997 : {
998 0 : const BlockInfoRecord *a = (const BlockInfoRecord *) p;
999 0 : const BlockInfoRecord *b = (const BlockInfoRecord *) q;
1000 :
1001 0 : cmp_member_elem(database);
1002 0 : cmp_member_elem(tablespace);
1003 0 : cmp_member_elem(filenumber);
1004 0 : cmp_member_elem(forknum);
1005 0 : cmp_member_elem(blocknum);
1006 :
1007 0 : return 0;
1008 0 : }
|
