1/*--------------------------------------------------------------------
2 * bgworker.c
3 * POSTGRES pluggable background workers implementation
4 *
5 * Portions Copyright (c) 1996-2019, PostgreSQL Global Development Group
6 *
7 * IDENTIFICATION
8 * src/backend/postmaster/bgworker.c
9 *
10 *-------------------------------------------------------------------------
11 */
12
13#include "postgres.h"
14
15#include <unistd.h>
16
17#include "libpq/pqsignal.h"
18#include "access/parallel.h"
19#include "miscadmin.h"
20#include "pgstat.h"
21#include "port/atomics.h"
22#include "postmaster/bgworker_internals.h"
23#include "postmaster/postmaster.h"
24#include "replication/logicallauncher.h"
25#include "replication/logicalworker.h"
26#include "storage/dsm.h"
27#include "storage/ipc.h"
28#include "storage/latch.h"
29#include "storage/lwlock.h"
30#include "storage/pg_shmem.h"
31#include "storage/pmsignal.h"
32#include "storage/proc.h"
33#include "storage/procsignal.h"
34#include "storage/shmem.h"
35#include "tcop/tcopprot.h"
36#include "utils/ascii.h"
37#include "utils/ps_status.h"
38#include "utils/timeout.h"
39
40/*
41 * The postmaster's list of registered background workers, in private memory.
42 */
43slist_head BackgroundWorkerList = SLIST_STATIC_INIT(BackgroundWorkerList);
44
45/*
46 * BackgroundWorkerSlots exist in shared memory and can be accessed (via
47 * the BackgroundWorkerArray) by both the postmaster and by regular backends.
48 * However, the postmaster cannot take locks, even spinlocks, because this
49 * might allow it to crash or become wedged if shared memory gets corrupted.
50 * Such an outcome is intolerable. Therefore, we need a lockless protocol
51 * for coordinating access to this data.
52 *
53 * The 'in_use' flag is used to hand off responsibility for the slot between
54 * the postmaster and the rest of the system. When 'in_use' is false,
55 * the postmaster will ignore the slot entirely, except for the 'in_use' flag
56 * itself, which it may read. In this state, regular backends may modify the
57 * slot. Once a backend sets 'in_use' to true, the slot becomes the
58 * responsibility of the postmaster. Regular backends may no longer modify it,
59 * but the postmaster may examine it. Thus, a backend initializing a slot
60 * must fully initialize the slot - and insert a write memory barrier - before
61 * marking it as in use.
62 *
63 * As an exception, however, even when the slot is in use, regular backends
64 * may set the 'terminate' flag for a slot, telling the postmaster not
65 * to restart it. Once the background worker is no longer running, the slot
66 * will be released for reuse.
67 *
68 * In addition to coordinating with the postmaster, backends modifying this
69 * data structure must coordinate with each other. Since they can take locks,
70 * this is straightforward: any backend wishing to manipulate a slot must
71 * take BackgroundWorkerLock in exclusive mode. Backends wishing to read
72 * data that might get concurrently modified by other backends should take
73 * this lock in shared mode. No matter what, backends reading this data
74 * structure must be able to tolerate concurrent modifications by the
75 * postmaster.
76 */
77typedef struct BackgroundWorkerSlot
78{
79 bool in_use;
80 bool terminate;
81 pid_t pid; /* InvalidPid = not started yet; 0 = dead */
82 uint64 generation; /* incremented when slot is recycled */
83 BackgroundWorker worker;
84} BackgroundWorkerSlot;
85
86/*
87 * In order to limit the total number of parallel workers (according to
88 * max_parallel_workers GUC), we maintain the number of active parallel
89 * workers. Since the postmaster cannot take locks, two variables are used for
90 * this purpose: the number of registered parallel workers (modified by the
91 * backends, protected by BackgroundWorkerLock) and the number of terminated
92 * parallel workers (modified only by the postmaster, lockless). The active
93 * number of parallel workers is the number of registered workers minus the
94 * terminated ones. These counters can of course overflow, but it's not
95 * important here since the subtraction will still give the right number.
96 */
97typedef struct BackgroundWorkerArray
98{
99 int total_slots;
100 uint32 parallel_register_count;
101 uint32 parallel_terminate_count;
102 BackgroundWorkerSlot slot[FLEXIBLE_ARRAY_MEMBER];
103} BackgroundWorkerArray;
104
105struct BackgroundWorkerHandle
106{
107 int slot;
108 uint64 generation;
109};
110
111static BackgroundWorkerArray *BackgroundWorkerData;
112
113/*
114 * List of internal background worker entry points. We need this for
115 * reasons explained in LookupBackgroundWorkerFunction(), below.
116 */
117static const struct
118{
119 const char *fn_name;
120 bgworker_main_type fn_addr;
121} InternalBGWorkers[] =
122
123{
124 {
125 "ParallelWorkerMain", ParallelWorkerMain
126 },
127 {
128 "ApplyLauncherMain", ApplyLauncherMain
129 },
130 {
131 "ApplyWorkerMain", ApplyWorkerMain
132 }
133};
134
135/* Private functions. */
136static bgworker_main_type LookupBackgroundWorkerFunction(const char *libraryname, const char *funcname);
137
138
139/*
140 * Calculate shared memory needed.
141 */
142Size
143BackgroundWorkerShmemSize(void)
144{
145 Size size;
146
147 /* Array of workers is variably sized. */
148 size = offsetof(BackgroundWorkerArray, slot);
149 size = add_size(size, mul_size(max_worker_processes,
150 sizeof(BackgroundWorkerSlot)));
151
152 return size;
153}
154
155/*
156 * Initialize shared memory.
157 */
158void
159BackgroundWorkerShmemInit(void)
160{
161 bool found;
162
163 BackgroundWorkerData = ShmemInitStruct("Background Worker Data",
164 BackgroundWorkerShmemSize(),
165 &found);
166 if (!IsUnderPostmaster)
167 {
168 slist_iter siter;
169 int slotno = 0;
170
171 BackgroundWorkerData->total_slots = max_worker_processes;
172 BackgroundWorkerData->parallel_register_count = 0;
173 BackgroundWorkerData->parallel_terminate_count = 0;
174
175 /*
176 * Copy contents of worker list into shared memory. Record the shared
177 * memory slot assigned to each worker. This ensures a 1-to-1
178 * correspondence between the postmaster's private list and the array
179 * in shared memory.
180 */
181 slist_foreach(siter, &BackgroundWorkerList)
182 {
183 BackgroundWorkerSlot *slot = &BackgroundWorkerData->slot[slotno];
184 RegisteredBgWorker *rw;
185
186 rw = slist_container(RegisteredBgWorker, rw_lnode, siter.cur);
187 Assert(slotno < max_worker_processes);
188 slot->in_use = true;
189 slot->terminate = false;
190 slot->pid = InvalidPid;
191 slot->generation = 0;
192 rw->rw_shmem_slot = slotno;
193 rw->rw_worker.bgw_notify_pid = 0; /* might be reinit after crash */
194 memcpy(&slot->worker, &rw->rw_worker, sizeof(BackgroundWorker));
195 ++slotno;
196 }
197
198 /*
199 * Mark any remaining slots as not in use.
200 */
201 while (slotno < max_worker_processes)
202 {
203 BackgroundWorkerSlot *slot = &BackgroundWorkerData->slot[slotno];
204
205 slot->in_use = false;
206 ++slotno;
207 }
208 }
209 else
210 Assert(found);
211}
212
213/*
214 * Search the postmaster's backend-private list of RegisteredBgWorker objects
215 * for the one that maps to the given slot number.
216 */
217static RegisteredBgWorker *
218FindRegisteredWorkerBySlotNumber(int slotno)
219{
220 slist_iter siter;
221
222 slist_foreach(siter, &BackgroundWorkerList)
223 {
224 RegisteredBgWorker *rw;
225
226 rw = slist_container(RegisteredBgWorker, rw_lnode, siter.cur);
227 if (rw->rw_shmem_slot == slotno)
228 return rw;
229 }
230
231 return NULL;
232}
233
234/*
235 * Notice changes to shared memory made by other backends. This code
236 * runs in the postmaster, so we must be very careful not to assume that
237 * shared memory contents are sane. Otherwise, a rogue backend could take
238 * out the postmaster.
239 */
240void
241BackgroundWorkerStateChange(void)
242{
243 int slotno;
244
245 /*
246 * The total number of slots stored in shared memory should match our
247 * notion of max_worker_processes. If it does not, something is very
248 * wrong. Further down, we always refer to this value as
249 * max_worker_processes, in case shared memory gets corrupted while we're
250 * looping.
251 */
252 if (max_worker_processes != BackgroundWorkerData->total_slots)
253 {
254 elog(LOG,
255 "inconsistent background worker state (max_worker_processes=%d, total_slots=%d",
256 max_worker_processes,
257 BackgroundWorkerData->total_slots);
258 return;
259 }
260
261 /*
262 * Iterate through slots, looking for newly-registered workers or workers
263 * who must die.
264 */
265 for (slotno = 0; slotno < max_worker_processes; ++slotno)
266 {
267 BackgroundWorkerSlot *slot = &BackgroundWorkerData->slot[slotno];
268 RegisteredBgWorker *rw;
269
270 if (!slot->in_use)
271 continue;
272
273 /*
274 * Make sure we don't see the in_use flag before the updated slot
275 * contents.
276 */
277 pg_read_barrier();
278
279 /* See whether we already know about this worker. */
280 rw = FindRegisteredWorkerBySlotNumber(slotno);
281 if (rw != NULL)
282 {
283 /*
284 * In general, the worker data can't change after it's initially
285 * registered. However, someone can set the terminate flag.
286 */
287 if (slot->terminate && !rw->rw_terminate)
288 {
289 rw->rw_terminate = true;
290 if (rw->rw_pid != 0)
291 kill(rw->rw_pid, SIGTERM);
292 else
293 {
294 /* Report never-started, now-terminated worker as dead. */
295 ReportBackgroundWorkerPID(rw);
296 }
297 }
298 continue;
299 }
300
301 /*
302 * If the worker is marked for termination, we don't need to add it to
303 * the registered workers list; we can just free the slot. However, if
304 * bgw_notify_pid is set, the process that registered the worker may
305 * need to know that we've processed the terminate request, so be sure
306 * to signal it.
307 */
308 if (slot->terminate)
309 {
310 int notify_pid;
311
312 /*
313 * We need a memory barrier here to make sure that the load of
314 * bgw_notify_pid and the update of parallel_terminate_count
315 * complete before the store to in_use.
316 */
317 notify_pid = slot->worker.bgw_notify_pid;
318 if ((slot->worker.bgw_flags & BGWORKER_CLASS_PARALLEL) != 0)
319 BackgroundWorkerData->parallel_terminate_count++;
320 pg_memory_barrier();
321 slot->pid = 0;
322 slot->in_use = false;
323 if (notify_pid != 0)
324 kill(notify_pid, SIGUSR1);
325
326 continue;
327 }
328
329 /*
330 * Copy the registration data into the registered workers list.
331 */
332 rw = malloc(sizeof(RegisteredBgWorker));
333 if (rw == NULL)
334 {
335 ereport(LOG,
336 (errcode(ERRCODE_OUT_OF_MEMORY),
337 errmsg("out of memory")));
338 return;
339 }
340
341 /*
342 * Copy strings in a paranoid way. If shared memory is corrupted, the
343 * source data might not even be NUL-terminated.
344 */
345 ascii_safe_strlcpy(rw->rw_worker.bgw_name,
346 slot->worker.bgw_name, BGW_MAXLEN);
347 ascii_safe_strlcpy(rw->rw_worker.bgw_type,
348 slot->worker.bgw_type, BGW_MAXLEN);
349 ascii_safe_strlcpy(rw->rw_worker.bgw_library_name,
350 slot->worker.bgw_library_name, BGW_MAXLEN);
351 ascii_safe_strlcpy(rw->rw_worker.bgw_function_name,
352 slot->worker.bgw_function_name, BGW_MAXLEN);
353
354 /*
355 * Copy various fixed-size fields.
356 *
357 * flags, start_time, and restart_time are examined by the postmaster,
358 * but nothing too bad will happen if they are corrupted. The
359 * remaining fields will only be examined by the child process. It
360 * might crash, but we won't.
361 */
362 rw->rw_worker.bgw_flags = slot->worker.bgw_flags;
363 rw->rw_worker.bgw_start_time = slot->worker.bgw_start_time;
364 rw->rw_worker.bgw_restart_time = slot->worker.bgw_restart_time;
365 rw->rw_worker.bgw_main_arg = slot->worker.bgw_main_arg;
366 memcpy(rw->rw_worker.bgw_extra, slot->worker.bgw_extra, BGW_EXTRALEN);
367
368 /*
369 * Copy the PID to be notified about state changes, but only if the
370 * postmaster knows about a backend with that PID. It isn't an error
371 * if the postmaster doesn't know about the PID, because the backend
372 * that requested the worker could have died (or been killed) just
373 * after doing so. Nonetheless, at least until we get some experience
374 * with how this plays out in the wild, log a message at a relative
375 * high debug level.
376 */
377 rw->rw_worker.bgw_notify_pid = slot->worker.bgw_notify_pid;
378 if (!PostmasterMarkPIDForWorkerNotify(rw->rw_worker.bgw_notify_pid))
379 {
380 elog(DEBUG1, "worker notification PID %lu is not valid",
381 (long) rw->rw_worker.bgw_notify_pid);
382 rw->rw_worker.bgw_notify_pid = 0;
383 }
384
385 /* Initialize postmaster bookkeeping. */
386 rw->rw_backend = NULL;
387 rw->rw_pid = 0;
388 rw->rw_child_slot = 0;
389 rw->rw_crashed_at = 0;
390 rw->rw_shmem_slot = slotno;
391 rw->rw_terminate = false;
392
393 /* Log it! */
394 ereport(DEBUG1,
395 (errmsg("registering background worker \"%s\"",
396 rw->rw_worker.bgw_name)));
397
398 slist_push_head(&BackgroundWorkerList, &rw->rw_lnode);
399 }
400}
401
402/*
403 * Forget about a background worker that's no longer needed.
404 *
405 * The worker must be identified by passing an slist_mutable_iter that
406 * points to it. This convention allows deletion of workers during
407 * searches of the worker list, and saves having to search the list again.
408 *
409 * This function must be invoked only in the postmaster.
410 */
411void
412ForgetBackgroundWorker(slist_mutable_iter *cur)
413{
414 RegisteredBgWorker *rw;
415 BackgroundWorkerSlot *slot;
416
417 rw = slist_container(RegisteredBgWorker, rw_lnode, cur->cur);
418
419 Assert(rw->rw_shmem_slot < max_worker_processes);
420 slot = &BackgroundWorkerData->slot[rw->rw_shmem_slot];
421 if ((rw->rw_worker.bgw_flags & BGWORKER_CLASS_PARALLEL) != 0)
422 BackgroundWorkerData->parallel_terminate_count++;
423
424 slot->in_use = false;
425
426 ereport(DEBUG1,
427 (errmsg("unregistering background worker \"%s\"",
428 rw->rw_worker.bgw_name)));
429
430 slist_delete_current(cur);
431 free(rw);
432}
433
434/*
435 * Report the PID of a newly-launched background worker in shared memory.
436 *
437 * This function should only be called from the postmaster.
438 */
439void
440ReportBackgroundWorkerPID(RegisteredBgWorker *rw)
441{
442 BackgroundWorkerSlot *slot;
443
444 Assert(rw->rw_shmem_slot < max_worker_processes);
445 slot = &BackgroundWorkerData->slot[rw->rw_shmem_slot];
446 slot->pid = rw->rw_pid;
447
448 if (rw->rw_worker.bgw_notify_pid != 0)
449 kill(rw->rw_worker.bgw_notify_pid, SIGUSR1);
450}
451
452/*
453 * Report that the PID of a background worker is now zero because a
454 * previously-running background worker has exited.
455 *
456 * This function should only be called from the postmaster.
457 */
458void
459ReportBackgroundWorkerExit(slist_mutable_iter *cur)
460{
461 RegisteredBgWorker *rw;
462 BackgroundWorkerSlot *slot;
463 int notify_pid;
464
465 rw = slist_container(RegisteredBgWorker, rw_lnode, cur->cur);
466
467 Assert(rw->rw_shmem_slot < max_worker_processes);
468 slot = &BackgroundWorkerData->slot[rw->rw_shmem_slot];
469 slot->pid = rw->rw_pid;
470 notify_pid = rw->rw_worker.bgw_notify_pid;
471
472 /*
473 * If this worker is slated for deregistration, do that before notifying
474 * the process which started it. Otherwise, if that process tries to
475 * reuse the slot immediately, it might not be available yet. In theory
476 * that could happen anyway if the process checks slot->pid at just the
477 * wrong moment, but this makes the window narrower.
478 */
479 if (rw->rw_terminate ||
480 rw->rw_worker.bgw_restart_time == BGW_NEVER_RESTART)
481 ForgetBackgroundWorker(cur);
482
483 if (notify_pid != 0)
484 kill(notify_pid, SIGUSR1);
485}
486
487/*
488 * Cancel SIGUSR1 notifications for a PID belonging to an exiting backend.
489 *
490 * This function should only be called from the postmaster.
491 */
492void
493BackgroundWorkerStopNotifications(pid_t pid)
494{
495 slist_iter siter;
496
497 slist_foreach(siter, &BackgroundWorkerList)
498 {
499 RegisteredBgWorker *rw;
500
501 rw = slist_container(RegisteredBgWorker, rw_lnode, siter.cur);
502 if (rw->rw_worker.bgw_notify_pid == pid)
503 rw->rw_worker.bgw_notify_pid = 0;
504 }
505}
506
507/*
508 * Reset background worker crash state.
509 *
510 * We assume that, after a crash-and-restart cycle, background workers without
511 * the never-restart flag should be restarted immediately, instead of waiting
512 * for bgw_restart_time to elapse.
513 */
514void
515ResetBackgroundWorkerCrashTimes(void)
516{
517 slist_mutable_iter iter;
518
519 slist_foreach_modify(iter, &BackgroundWorkerList)
520 {
521 RegisteredBgWorker *rw;
522
523 rw = slist_container(RegisteredBgWorker, rw_lnode, iter.cur);
524
525 if (rw->rw_worker.bgw_restart_time == BGW_NEVER_RESTART)
526 {
527 /*
528 * Workers marked BGW_NEVER_RESTART shouldn't get relaunched after
529 * the crash, so forget about them. (If we wait until after the
530 * crash to forget about them, and they are parallel workers,
531 * parallel_terminate_count will get incremented after we've
532 * already zeroed parallel_register_count, which would be bad.)
533 */
534 ForgetBackgroundWorker(&iter);
535 }
536 else
537 {
538 /*
539 * The accounting which we do via parallel_register_count and
540 * parallel_terminate_count would get messed up if a worker marked
541 * parallel could survive a crash and restart cycle. All such
542 * workers should be marked BGW_NEVER_RESTART, and thus control
543 * should never reach this branch.
544 */
545 Assert((rw->rw_worker.bgw_flags & BGWORKER_CLASS_PARALLEL) == 0);
546
547 /*
548 * Allow this worker to be restarted immediately after we finish
549 * resetting.
550 */
551 rw->rw_crashed_at = 0;
552 }
553 }
554}
555
556#ifdef EXEC_BACKEND
557/*
558 * In EXEC_BACKEND mode, workers use this to retrieve their details from
559 * shared memory.
560 */
561BackgroundWorker *
562BackgroundWorkerEntry(int slotno)
563{
564 static BackgroundWorker myEntry;
565 BackgroundWorkerSlot *slot;
566
567 Assert(slotno < BackgroundWorkerData->total_slots);
568 slot = &BackgroundWorkerData->slot[slotno];
569 Assert(slot->in_use);
570
571 /* must copy this in case we don't intend to retain shmem access */
572 memcpy(&myEntry, &slot->worker, sizeof myEntry);
573 return &myEntry;
574}
575#endif
576
577/*
578 * Complain about the BackgroundWorker definition using error level elevel.
579 * Return true if it looks ok, false if not (unless elevel >= ERROR, in
580 * which case we won't return at all in the not-OK case).
581 */
582static bool
583SanityCheckBackgroundWorker(BackgroundWorker *worker, int elevel)
584{
585 /* sanity check for flags */
586 if (worker->bgw_flags & BGWORKER_BACKEND_DATABASE_CONNECTION)
587 {
588 if (!(worker->bgw_flags & BGWORKER_SHMEM_ACCESS))
589 {
590 ereport(elevel,
591 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
592 errmsg("background worker \"%s\": must attach to shared memory in order to request a database connection",
593 worker->bgw_name)));
594 return false;
595 }
596
597 if (worker->bgw_start_time == BgWorkerStart_PostmasterStart)
598 {
599 ereport(elevel,
600 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
601 errmsg("background worker \"%s\": cannot request database access if starting at postmaster start",
602 worker->bgw_name)));
603 return false;
604 }
605
606 /* XXX other checks? */
607 }
608
609 if ((worker->bgw_restart_time < 0 &&
610 worker->bgw_restart_time != BGW_NEVER_RESTART) ||
611 (worker->bgw_restart_time > USECS_PER_DAY / 1000))
612 {
613 ereport(elevel,
614 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
615 errmsg("background worker \"%s\": invalid restart interval",
616 worker->bgw_name)));
617 return false;
618 }
619
620 /*
621 * Parallel workers may not be configured for restart, because the
622 * parallel_register_count/parallel_terminate_count accounting can't
623 * handle parallel workers lasting through a crash-and-restart cycle.
624 */
625 if (worker->bgw_restart_time != BGW_NEVER_RESTART &&
626 (worker->bgw_flags & BGWORKER_CLASS_PARALLEL) != 0)
627 {
628 ereport(elevel,
629 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
630 errmsg("background worker \"%s\": parallel workers may not be configured for restart",
631 worker->bgw_name)));
632 return false;
633 }
634
635 /*
636 * If bgw_type is not filled in, use bgw_name.
637 */
638 if (strcmp(worker->bgw_type, "") == 0)
639 strcpy(worker->bgw_type, worker->bgw_name);
640
641 return true;
642}
643
644static void
645bgworker_quickdie(SIGNAL_ARGS)
646{
647 /*
648 * We DO NOT want to run proc_exit() or atexit() callbacks -- we're here
649 * because shared memory may be corrupted, so we don't want to try to
650 * clean up our transaction. Just nail the windows shut and get out of
651 * town. The callbacks wouldn't be safe to run from a signal handler,
652 * anyway.
653 *
654 * Note we do _exit(2) not _exit(0). This is to force the postmaster into
655 * a system reset cycle if someone sends a manual SIGQUIT to a random
656 * backend. This is necessary precisely because we don't clean up our
657 * shared memory state. (The "dead man switch" mechanism in pmsignal.c
658 * should ensure the postmaster sees this as a crash, too, but no harm in
659 * being doubly sure.)
660 */
661 _exit(2);
662}
663
664/*
665 * Standard SIGTERM handler for background workers
666 */
667static void
668bgworker_die(SIGNAL_ARGS)
669{
670 PG_SETMASK(&BlockSig);
671
672 ereport(FATAL,
673 (errcode(ERRCODE_ADMIN_SHUTDOWN),
674 errmsg("terminating background worker \"%s\" due to administrator command",
675 MyBgworkerEntry->bgw_type)));
676}
677
678/*
679 * Standard SIGUSR1 handler for unconnected workers
680 *
681 * Here, we want to make sure an unconnected worker will at least heed
682 * latch activity.
683 */
684static void
685bgworker_sigusr1_handler(SIGNAL_ARGS)
686{
687 int save_errno = errno;
688
689 latch_sigusr1_handler();
690
691 errno = save_errno;
692}
693
694/*
695 * Start a new background worker
696 *
697 * This is the main entry point for background worker, to be called from
698 * postmaster.
699 */
700void
701StartBackgroundWorker(void)
702{
703 sigjmp_buf local_sigjmp_buf;
704 BackgroundWorker *worker = MyBgworkerEntry;
705 bgworker_main_type entrypt;
706
707 if (worker == NULL)
708 elog(FATAL, "unable to find bgworker entry");
709
710 IsBackgroundWorker = true;
711
712 /* Identify myself via ps */
713 init_ps_display(worker->bgw_name, "", "", "");
714
715 /*
716 * If we're not supposed to have shared memory access, then detach from
717 * shared memory. If we didn't request shared memory access, the
718 * postmaster won't force a cluster-wide restart if we exit unexpectedly,
719 * so we'd better make sure that we don't mess anything up that would
720 * require that sort of cleanup.
721 */
722 if ((worker->bgw_flags & BGWORKER_SHMEM_ACCESS) == 0)
723 {
724 dsm_detach_all();
725 PGSharedMemoryDetach();
726 }
727
728 SetProcessingMode(InitProcessing);
729
730 /* Apply PostAuthDelay */
731 if (PostAuthDelay > 0)
732 pg_usleep(PostAuthDelay * 1000000L);
733
734 /*
735 * Set up signal handlers.
736 */
737 if (worker->bgw_flags & BGWORKER_BACKEND_DATABASE_CONNECTION)
738 {
739 /*
740 * SIGINT is used to signal canceling the current action
741 */
742 pqsignal(SIGINT, StatementCancelHandler);
743 pqsignal(SIGUSR1, procsignal_sigusr1_handler);
744 pqsignal(SIGFPE, FloatExceptionHandler);
745
746 /* XXX Any other handlers needed here? */
747 }
748 else
749 {
750 pqsignal(SIGINT, SIG_IGN);
751 pqsignal(SIGUSR1, bgworker_sigusr1_handler);
752 pqsignal(SIGFPE, SIG_IGN);
753 }
754 pqsignal(SIGTERM, bgworker_die);
755 pqsignal(SIGHUP, SIG_IGN);
756
757 pqsignal(SIGQUIT, bgworker_quickdie);
758 InitializeTimeouts(); /* establishes SIGALRM handler */
759
760 pqsignal(SIGPIPE, SIG_IGN);
761 pqsignal(SIGUSR2, SIG_IGN);
762 pqsignal(SIGCHLD, SIG_DFL);
763
764 /*
765 * If an exception is encountered, processing resumes here.
766 *
767 * See notes in postgres.c about the design of this coding.
768 */
769 if (sigsetjmp(local_sigjmp_buf, 1) != 0)
770 {
771 /* Since not using PG_TRY, must reset error stack by hand */
772 error_context_stack = NULL;
773
774 /* Prevent interrupts while cleaning up */
775 HOLD_INTERRUPTS();
776
777 /* Report the error to the server log */
778 EmitErrorReport();
779
780 /*
781 * Do we need more cleanup here? For shmem-connected bgworkers, we
782 * will call InitProcess below, which will install ProcKill as exit
783 * callback. That will take care of releasing locks, etc.
784 */
785
786 /* and go away */
787 proc_exit(1);
788 }
789
790 /* We can now handle ereport(ERROR) */
791 PG_exception_stack = &local_sigjmp_buf;
792
793 /*
794 * If the background worker request shared memory access, set that up now;
795 * else, detach all shared memory segments.
796 */
797 if (worker->bgw_flags & BGWORKER_SHMEM_ACCESS)
798 {
799 /*
800 * Early initialization. Some of this could be useful even for
801 * background workers that aren't using shared memory, but they can
802 * call the individual startup routines for those subsystems if
803 * needed.
804 */
805 BaseInit();
806
807 /*
808 * Create a per-backend PGPROC struct in shared memory, except in the
809 * EXEC_BACKEND case where this was done in SubPostmasterMain. We must
810 * do this before we can use LWLocks (and in the EXEC_BACKEND case we
811 * already had to do some stuff with LWLocks).
812 */
813#ifndef EXEC_BACKEND
814 InitProcess();
815#endif
816 }
817
818 /*
819 * Look up the entry point function, loading its library if necessary.
820 */
821 entrypt = LookupBackgroundWorkerFunction(worker->bgw_library_name,
822 worker->bgw_function_name);
823
824 /*
825 * Note that in normal processes, we would call InitPostgres here. For a
826 * worker, however, we don't know what database to connect to, yet; so we
827 * need to wait until the user code does it via
828 * BackgroundWorkerInitializeConnection().
829 */
830
831 /*
832 * Now invoke the user-defined worker code
833 */
834 entrypt(worker->bgw_main_arg);
835
836 /* ... and if it returns, we're done */
837 proc_exit(0);
838}
839
840/*
841 * Register a new static background worker.
842 *
843 * This can only be called directly from postmaster or in the _PG_init
844 * function of a module library that's loaded by shared_preload_libraries;
845 * otherwise it will have no effect.
846 */
847void
848RegisterBackgroundWorker(BackgroundWorker *worker)
849{
850 RegisteredBgWorker *rw;
851 static int numworkers = 0;
852
853 if (!IsUnderPostmaster)
854 ereport(DEBUG1,
855 (errmsg("registering background worker \"%s\"", worker->bgw_name)));
856
857 if (!process_shared_preload_libraries_in_progress &&
858 strcmp(worker->bgw_library_name, "postgres") != 0)
859 {
860 if (!IsUnderPostmaster)
861 ereport(LOG,
862 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
863 errmsg("background worker \"%s\": must be registered in shared_preload_libraries",
864 worker->bgw_name)));
865 return;
866 }
867
868 if (!SanityCheckBackgroundWorker(worker, LOG))
869 return;
870
871 if (worker->bgw_notify_pid != 0)
872 {
873 ereport(LOG,
874 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
875 errmsg("background worker \"%s\": only dynamic background workers can request notification",
876 worker->bgw_name)));
877 return;
878 }
879
880 /*
881 * Enforce maximum number of workers. Note this is overly restrictive: we
882 * could allow more non-shmem-connected workers, because these don't count
883 * towards the MAX_BACKENDS limit elsewhere. For now, it doesn't seem
884 * important to relax this restriction.
885 */
886 if (++numworkers > max_worker_processes)
887 {
888 ereport(LOG,
889 (errcode(ERRCODE_CONFIGURATION_LIMIT_EXCEEDED),
890 errmsg("too many background workers"),
891 errdetail_plural("Up to %d background worker can be registered with the current settings.",
892 "Up to %d background workers can be registered with the current settings.",
893 max_worker_processes,
894 max_worker_processes),
895 errhint("Consider increasing the configuration parameter \"max_worker_processes\".")));
896 return;
897 }
898
899 /*
900 * Copy the registration data into the registered workers list.
901 */
902 rw = malloc(sizeof(RegisteredBgWorker));
903 if (rw == NULL)
904 {
905 ereport(LOG,
906 (errcode(ERRCODE_OUT_OF_MEMORY),
907 errmsg("out of memory")));
908 return;
909 }
910
911 rw->rw_worker = *worker;
912 rw->rw_backend = NULL;
913 rw->rw_pid = 0;
914 rw->rw_child_slot = 0;
915 rw->rw_crashed_at = 0;
916 rw->rw_terminate = false;
917
918 slist_push_head(&BackgroundWorkerList, &rw->rw_lnode);
919}
920
921/*
922 * Register a new background worker from a regular backend.
923 *
924 * Returns true on success and false on failure. Failure typically indicates
925 * that no background worker slots are currently available.
926 *
927 * If handle != NULL, we'll set *handle to a pointer that can subsequently
928 * be used as an argument to GetBackgroundWorkerPid(). The caller can
929 * free this pointer using pfree(), if desired.
930 */
931bool
932RegisterDynamicBackgroundWorker(BackgroundWorker *worker,
933 BackgroundWorkerHandle **handle)
934{
935 int slotno;
936 bool success = false;
937 bool parallel;
938 uint64 generation = 0;
939
940 /*
941 * We can't register dynamic background workers from the postmaster. If
942 * this is a standalone backend, we're the only process and can't start
943 * any more. In a multi-process environment, it might be theoretically
944 * possible, but we don't currently support it due to locking
945 * considerations; see comments on the BackgroundWorkerSlot data
946 * structure.
947 */
948 if (!IsUnderPostmaster)
949 return false;
950
951 if (!SanityCheckBackgroundWorker(worker, ERROR))
952 return false;
953
954 parallel = (worker->bgw_flags & BGWORKER_CLASS_PARALLEL) != 0;
955
956 LWLockAcquire(BackgroundWorkerLock, LW_EXCLUSIVE);
957
958 /*
959 * If this is a parallel worker, check whether there are already too many
960 * parallel workers; if so, don't register another one. Our view of
961 * parallel_terminate_count may be slightly stale, but that doesn't really
962 * matter: we would have gotten the same result if we'd arrived here
963 * slightly earlier anyway. There's no help for it, either, since the
964 * postmaster must not take locks; a memory barrier wouldn't guarantee
965 * anything useful.
966 */
967 if (parallel && (BackgroundWorkerData->parallel_register_count -
968 BackgroundWorkerData->parallel_terminate_count) >=
969 max_parallel_workers)
970 {
971 Assert(BackgroundWorkerData->parallel_register_count -
972 BackgroundWorkerData->parallel_terminate_count <=
973 MAX_PARALLEL_WORKER_LIMIT);
974 LWLockRelease(BackgroundWorkerLock);
975 return false;
976 }
977
978 /*
979 * Look for an unused slot. If we find one, grab it.
980 */
981 for (slotno = 0; slotno < BackgroundWorkerData->total_slots; ++slotno)
982 {
983 BackgroundWorkerSlot *slot = &BackgroundWorkerData->slot[slotno];
984
985 if (!slot->in_use)
986 {
987 memcpy(&slot->worker, worker, sizeof(BackgroundWorker));
988 slot->pid = InvalidPid; /* indicates not started yet */
989 slot->generation++;
990 slot->terminate = false;
991 generation = slot->generation;
992 if (parallel)
993 BackgroundWorkerData->parallel_register_count++;
994
995 /*
996 * Make sure postmaster doesn't see the slot as in use before it
997 * sees the new contents.
998 */
999 pg_write_barrier();
1000
1001 slot->in_use = true;
1002 success = true;
1003 break;
1004 }
1005 }
1006
1007 LWLockRelease(BackgroundWorkerLock);
1008
1009 /* If we found a slot, tell the postmaster to notice the change. */
1010 if (success)
1011 SendPostmasterSignal(PMSIGNAL_BACKGROUND_WORKER_CHANGE);
1012
1013 /*
1014 * If we found a slot and the user has provided a handle, initialize it.
1015 */
1016 if (success && handle)
1017 {
1018 *handle = palloc(sizeof(BackgroundWorkerHandle));
1019 (*handle)->slot = slotno;
1020 (*handle)->generation = generation;
1021 }
1022
1023 return success;
1024}
1025
1026/*
1027 * Get the PID of a dynamically-registered background worker.
1028 *
1029 * If the worker is determined to be running, the return value will be
1030 * BGWH_STARTED and *pidp will get the PID of the worker process. If the
1031 * postmaster has not yet attempted to start the worker, the return value will
1032 * be BGWH_NOT_YET_STARTED. Otherwise, the return value is BGWH_STOPPED.
1033 *
1034 * BGWH_STOPPED can indicate either that the worker is temporarily stopped
1035 * (because it is configured for automatic restart and exited non-zero),
1036 * or that the worker is permanently stopped (because it exited with exit
1037 * code 0, or was not configured for automatic restart), or even that the
1038 * worker was unregistered without ever starting (either because startup
1039 * failed and the worker is not configured for automatic restart, or because
1040 * TerminateBackgroundWorker was used before the worker was successfully
1041 * started).
1042 */
1043BgwHandleStatus
1044GetBackgroundWorkerPid(BackgroundWorkerHandle *handle, pid_t *pidp)
1045{
1046 BackgroundWorkerSlot *slot;
1047 pid_t pid;
1048
1049 Assert(handle->slot < max_worker_processes);
1050 slot = &BackgroundWorkerData->slot[handle->slot];
1051
1052 /*
1053 * We could probably arrange to synchronize access to data using memory
1054 * barriers only, but for now, let's just keep it simple and grab the
1055 * lock. It seems unlikely that there will be enough traffic here to
1056 * result in meaningful contention.
1057 */
1058 LWLockAcquire(BackgroundWorkerLock, LW_SHARED);
1059
1060 /*
1061 * The generation number can't be concurrently changed while we hold the
1062 * lock. The pid, which is updated by the postmaster, can change at any
1063 * time, but we assume such changes are atomic. So the value we read
1064 * won't be garbage, but it might be out of date by the time the caller
1065 * examines it (but that's unavoidable anyway).
1066 *
1067 * The in_use flag could be in the process of changing from true to false,
1068 * but if it is already false then it can't change further.
1069 */
1070 if (handle->generation != slot->generation || !slot->in_use)
1071 pid = 0;
1072 else
1073 pid = slot->pid;
1074
1075 /* All done. */
1076 LWLockRelease(BackgroundWorkerLock);
1077
1078 if (pid == 0)
1079 return BGWH_STOPPED;
1080 else if (pid == InvalidPid)
1081 return BGWH_NOT_YET_STARTED;
1082 *pidp = pid;
1083 return BGWH_STARTED;
1084}
1085
1086/*
1087 * Wait for a background worker to start up.
1088 *
1089 * This is like GetBackgroundWorkerPid(), except that if the worker has not
1090 * yet started, we wait for it to do so; thus, BGWH_NOT_YET_STARTED is never
1091 * returned. However, if the postmaster has died, we give up and return
1092 * BGWH_POSTMASTER_DIED, since it that case we know that startup will not
1093 * take place.
1094 */
1095BgwHandleStatus
1096WaitForBackgroundWorkerStartup(BackgroundWorkerHandle *handle, pid_t *pidp)
1097{
1098 BgwHandleStatus status;
1099 int rc;
1100
1101 for (;;)
1102 {
1103 pid_t pid;
1104
1105 CHECK_FOR_INTERRUPTS();
1106
1107 status = GetBackgroundWorkerPid(handle, &pid);
1108 if (status == BGWH_STARTED)
1109 *pidp = pid;
1110 if (status != BGWH_NOT_YET_STARTED)
1111 break;
1112
1113 rc = WaitLatch(MyLatch,
1114 WL_LATCH_SET | WL_POSTMASTER_DEATH, 0,
1115 WAIT_EVENT_BGWORKER_STARTUP);
1116
1117 if (rc & WL_POSTMASTER_DEATH)
1118 {
1119 status = BGWH_POSTMASTER_DIED;
1120 break;
1121 }
1122
1123 ResetLatch(MyLatch);
1124 }
1125
1126 return status;
1127}
1128
1129/*
1130 * Wait for a background worker to stop.
1131 *
1132 * If the worker hasn't yet started, or is running, we wait for it to stop
1133 * and then return BGWH_STOPPED. However, if the postmaster has died, we give
1134 * up and return BGWH_POSTMASTER_DIED, because it's the postmaster that
1135 * notifies us when a worker's state changes.
1136 */
1137BgwHandleStatus
1138WaitForBackgroundWorkerShutdown(BackgroundWorkerHandle *handle)
1139{
1140 BgwHandleStatus status;
1141 int rc;
1142
1143 for (;;)
1144 {
1145 pid_t pid;
1146
1147 CHECK_FOR_INTERRUPTS();
1148
1149 status = GetBackgroundWorkerPid(handle, &pid);
1150 if (status == BGWH_STOPPED)
1151 break;
1152
1153 rc = WaitLatch(MyLatch,
1154 WL_LATCH_SET | WL_POSTMASTER_DEATH, 0,
1155 WAIT_EVENT_BGWORKER_SHUTDOWN);
1156
1157 if (rc & WL_POSTMASTER_DEATH)
1158 {
1159 status = BGWH_POSTMASTER_DIED;
1160 break;
1161 }
1162
1163 ResetLatch(MyLatch);
1164 }
1165
1166 return status;
1167}
1168
1169/*
1170 * Instruct the postmaster to terminate a background worker.
1171 *
1172 * Note that it's safe to do this without regard to whether the worker is
1173 * still running, or even if the worker may already have existed and been
1174 * unregistered.
1175 */
1176void
1177TerminateBackgroundWorker(BackgroundWorkerHandle *handle)
1178{
1179 BackgroundWorkerSlot *slot;
1180 bool signal_postmaster = false;
1181
1182 Assert(handle->slot < max_worker_processes);
1183 slot = &BackgroundWorkerData->slot[handle->slot];
1184
1185 /* Set terminate flag in shared memory, unless slot has been reused. */
1186 LWLockAcquire(BackgroundWorkerLock, LW_EXCLUSIVE);
1187 if (handle->generation == slot->generation)
1188 {
1189 slot->terminate = true;
1190 signal_postmaster = true;
1191 }
1192 LWLockRelease(BackgroundWorkerLock);
1193
1194 /* Make sure the postmaster notices the change to shared memory. */
1195 if (signal_postmaster)
1196 SendPostmasterSignal(PMSIGNAL_BACKGROUND_WORKER_CHANGE);
1197}
1198
1199/*
1200 * Look up (and possibly load) a bgworker entry point function.
1201 *
1202 * For functions contained in the core code, we use library name "postgres"
1203 * and consult the InternalBGWorkers array. External functions are
1204 * looked up, and loaded if necessary, using load_external_function().
1205 *
1206 * The point of this is to pass function names as strings across process
1207 * boundaries. We can't pass actual function addresses because of the
1208 * possibility that the function has been loaded at a different address
1209 * in a different process. This is obviously a hazard for functions in
1210 * loadable libraries, but it can happen even for functions in the core code
1211 * on platforms using EXEC_BACKEND (e.g., Windows).
1212 *
1213 * At some point it might be worthwhile to get rid of InternalBGWorkers[]
1214 * in favor of applying load_external_function() for core functions too;
1215 * but that raises portability issues that are not worth addressing now.
1216 */
1217static bgworker_main_type
1218LookupBackgroundWorkerFunction(const char *libraryname, const char *funcname)
1219{
1220 /*
1221 * If the function is to be loaded from postgres itself, search the
1222 * InternalBGWorkers array.
1223 */
1224 if (strcmp(libraryname, "postgres") == 0)
1225 {
1226 int i;
1227
1228 for (i = 0; i < lengthof(InternalBGWorkers); i++)
1229 {
1230 if (strcmp(InternalBGWorkers[i].fn_name, funcname) == 0)
1231 return InternalBGWorkers[i].fn_addr;
1232 }
1233
1234 /* We can only reach this by programming error. */
1235 elog(ERROR, "internal function \"%s\" not found", funcname);
1236 }
1237
1238 /* Otherwise load from external library. */
1239 return (bgworker_main_type)
1240 load_external_function(libraryname, funcname, true, NULL);
1241}
1242
1243/*
1244 * Given a PID, get the bgw_type of the background worker. Returns NULL if
1245 * not a valid background worker.
1246 *
1247 * The return value is in static memory belonging to this function, so it has
1248 * to be used before calling this function again. This is so that the caller
1249 * doesn't have to worry about the background worker locking protocol.
1250 */
1251const char *
1252GetBackgroundWorkerTypeByPid(pid_t pid)
1253{
1254 int slotno;
1255 bool found = false;
1256 static char result[BGW_MAXLEN];
1257
1258 LWLockAcquire(BackgroundWorkerLock, LW_SHARED);
1259
1260 for (slotno = 0; slotno < BackgroundWorkerData->total_slots; slotno++)
1261 {
1262 BackgroundWorkerSlot *slot = &BackgroundWorkerData->slot[slotno];
1263
1264 if (slot->pid > 0 && slot->pid == pid)
1265 {
1266 strcpy(result, slot->worker.bgw_type);
1267 found = true;
1268 break;
1269 }
1270 }
1271
1272 LWLockRelease(BackgroundWorkerLock);
1273
1274 if (!found)
1275 return NULL;
1276
1277 return result;
1278}
1279