PostgreSQL Source Code  git master
bgworker.c
Go to the documentation of this file.
1 /*--------------------------------------------------------------------
2  * bgworker.c
3  * POSTGRES pluggable background workers implementation
4  *
5  * Portions Copyright (c) 1996-2020, PostgreSQL Global Development Group
6  *
7  * IDENTIFICATION
8  * src/backend/postmaster/bgworker.c
9  *
10  *-------------------------------------------------------------------------
11  */
12 
13 #include "postgres.h"
14 
15 #include "access/parallel.h"
16 #include "libpq/pqsignal.h"
17 #include "miscadmin.h"
18 #include "pgstat.h"
19 #include "port/atomics.h"
21 #include "postmaster/interrupt.h"
22 #include "postmaster/postmaster.h"
25 #include "storage/dsm.h"
26 #include "storage/ipc.h"
27 #include "storage/latch.h"
28 #include "storage/lwlock.h"
29 #include "storage/pg_shmem.h"
30 #include "storage/pmsignal.h"
31 #include "storage/proc.h"
32 #include "storage/procsignal.h"
33 #include "storage/shmem.h"
34 #include "tcop/tcopprot.h"
35 #include "utils/ascii.h"
36 #include "utils/ps_status.h"
37 #include "utils/timeout.h"
38 
39 /*
40  * The postmaster's list of registered background workers, in private memory.
41  */
43 
44 /*
45  * BackgroundWorkerSlots exist in shared memory and can be accessed (via
46  * the BackgroundWorkerArray) by both the postmaster and by regular backends.
47  * However, the postmaster cannot take locks, even spinlocks, because this
48  * might allow it to crash or become wedged if shared memory gets corrupted.
49  * Such an outcome is intolerable. Therefore, we need a lockless protocol
50  * for coordinating access to this data.
51  *
52  * The 'in_use' flag is used to hand off responsibility for the slot between
53  * the postmaster and the rest of the system. When 'in_use' is false,
54  * the postmaster will ignore the slot entirely, except for the 'in_use' flag
55  * itself, which it may read. In this state, regular backends may modify the
56  * slot. Once a backend sets 'in_use' to true, the slot becomes the
57  * responsibility of the postmaster. Regular backends may no longer modify it,
58  * but the postmaster may examine it. Thus, a backend initializing a slot
59  * must fully initialize the slot - and insert a write memory barrier - before
60  * marking it as in use.
61  *
62  * As an exception, however, even when the slot is in use, regular backends
63  * may set the 'terminate' flag for a slot, telling the postmaster not
64  * to restart it. Once the background worker is no longer running, the slot
65  * will be released for reuse.
66  *
67  * In addition to coordinating with the postmaster, backends modifying this
68  * data structure must coordinate with each other. Since they can take locks,
69  * this is straightforward: any backend wishing to manipulate a slot must
70  * take BackgroundWorkerLock in exclusive mode. Backends wishing to read
71  * data that might get concurrently modified by other backends should take
72  * this lock in shared mode. No matter what, backends reading this data
73  * structure must be able to tolerate concurrent modifications by the
74  * postmaster.
75  */
76 typedef struct BackgroundWorkerSlot
77 {
78  bool in_use;
79  bool terminate;
80  pid_t pid; /* InvalidPid = not started yet; 0 = dead */
81  uint64 generation; /* incremented when slot is recycled */
84 
85 /*
86  * In order to limit the total number of parallel workers (according to
87  * max_parallel_workers GUC), we maintain the number of active parallel
88  * workers. Since the postmaster cannot take locks, two variables are used for
89  * this purpose: the number of registered parallel workers (modified by the
90  * backends, protected by BackgroundWorkerLock) and the number of terminated
91  * parallel workers (modified only by the postmaster, lockless). The active
92  * number of parallel workers is the number of registered workers minus the
93  * terminated ones. These counters can of course overflow, but it's not
94  * important here since the subtraction will still give the right number.
95  */
96 typedef struct BackgroundWorkerArray
97 {
103 
105 {
106  int slot;
107  uint64 generation;
108 };
109 
111 
112 /*
113  * List of internal background worker entry points. We need this for
114  * reasons explained in LookupBackgroundWorkerFunction(), below.
115  */
116 static const struct
117 {
118  const char *fn_name;
120 } InternalBGWorkers[] =
121 
122 {
123  {
124  "ParallelWorkerMain", ParallelWorkerMain
125  },
126  {
127  "ApplyLauncherMain", ApplyLauncherMain
128  },
129  {
130  "ApplyWorkerMain", ApplyWorkerMain
131  }
132 };
133 
134 /* Private functions. */
135 static bgworker_main_type LookupBackgroundWorkerFunction(const char *libraryname, const char *funcname);
136 
137 
138 /*
139  * Calculate shared memory needed.
140  */
141 Size
143 {
144  Size size;
145 
146  /* Array of workers is variably sized. */
147  size = offsetof(BackgroundWorkerArray, slot);
149  sizeof(BackgroundWorkerSlot)));
150 
151  return size;
152 }
153 
154 /*
155  * Initialize shared memory.
156  */
157 void
159 {
160  bool found;
161 
162  BackgroundWorkerData = ShmemInitStruct("Background Worker Data",
164  &found);
165  if (!IsUnderPostmaster)
166  {
167  slist_iter siter;
168  int slotno = 0;
169 
170  BackgroundWorkerData->total_slots = max_worker_processes;
171  BackgroundWorkerData->parallel_register_count = 0;
172  BackgroundWorkerData->parallel_terminate_count = 0;
173 
174  /*
175  * Copy contents of worker list into shared memory. Record the shared
176  * memory slot assigned to each worker. This ensures a 1-to-1
177  * correspondence between the postmaster's private list and the array
178  * in shared memory.
179  */
180  slist_foreach(siter, &BackgroundWorkerList)
181  {
182  BackgroundWorkerSlot *slot = &BackgroundWorkerData->slot[slotno];
183  RegisteredBgWorker *rw;
184 
185  rw = slist_container(RegisteredBgWorker, rw_lnode, siter.cur);
186  Assert(slotno < max_worker_processes);
187  slot->in_use = true;
188  slot->terminate = false;
189  slot->pid = InvalidPid;
190  slot->generation = 0;
191  rw->rw_shmem_slot = slotno;
192  rw->rw_worker.bgw_notify_pid = 0; /* might be reinit after crash */
193  memcpy(&slot->worker, &rw->rw_worker, sizeof(BackgroundWorker));
194  ++slotno;
195  }
196 
197  /*
198  * Mark any remaining slots as not in use.
199  */
200  while (slotno < max_worker_processes)
201  {
202  BackgroundWorkerSlot *slot = &BackgroundWorkerData->slot[slotno];
203 
204  slot->in_use = false;
205  ++slotno;
206  }
207  }
208  else
209  Assert(found);
210 }
211 
212 /*
213  * Search the postmaster's backend-private list of RegisteredBgWorker objects
214  * for the one that maps to the given slot number.
215  */
216 static RegisteredBgWorker *
218 {
219  slist_iter siter;
220 
221  slist_foreach(siter, &BackgroundWorkerList)
222  {
223  RegisteredBgWorker *rw;
224 
225  rw = slist_container(RegisteredBgWorker, rw_lnode, siter.cur);
226  if (rw->rw_shmem_slot == slotno)
227  return rw;
228  }
229 
230  return NULL;
231 }
232 
233 /*
234  * Notice changes to shared memory made by other backends. This code
235  * runs in the postmaster, so we must be very careful not to assume that
236  * shared memory contents are sane. Otherwise, a rogue backend could take
237  * out the postmaster.
238  */
239 void
241 {
242  int slotno;
243 
244  /*
245  * The total number of slots stored in shared memory should match our
246  * notion of max_worker_processes. If it does not, something is very
247  * wrong. Further down, we always refer to this value as
248  * max_worker_processes, in case shared memory gets corrupted while we're
249  * looping.
250  */
251  if (max_worker_processes != BackgroundWorkerData->total_slots)
252  {
253  elog(LOG,
254  "inconsistent background worker state (max_worker_processes=%d, total_slots=%d",
256  BackgroundWorkerData->total_slots);
257  return;
258  }
259 
260  /*
261  * Iterate through slots, looking for newly-registered workers or workers
262  * who must die.
263  */
264  for (slotno = 0; slotno < max_worker_processes; ++slotno)
265  {
266  BackgroundWorkerSlot *slot = &BackgroundWorkerData->slot[slotno];
267  RegisteredBgWorker *rw;
268 
269  if (!slot->in_use)
270  continue;
271 
272  /*
273  * Make sure we don't see the in_use flag before the updated slot
274  * contents.
275  */
276  pg_read_barrier();
277 
278  /* See whether we already know about this worker. */
280  if (rw != NULL)
281  {
282  /*
283  * In general, the worker data can't change after it's initially
284  * registered. However, someone can set the terminate flag.
285  */
286  if (slot->terminate && !rw->rw_terminate)
287  {
288  rw->rw_terminate = true;
289  if (rw->rw_pid != 0)
290  kill(rw->rw_pid, SIGTERM);
291  else
292  {
293  /* Report never-started, now-terminated worker as dead. */
295  }
296  }
297  continue;
298  }
299 
300  /*
301  * If the worker is marked for termination, we don't need to add it to
302  * the registered workers list; we can just free the slot. However, if
303  * bgw_notify_pid is set, the process that registered the worker may
304  * need to know that we've processed the terminate request, so be sure
305  * to signal it.
306  */
307  if (slot->terminate)
308  {
309  int notify_pid;
310 
311  /*
312  * We need a memory barrier here to make sure that the load of
313  * bgw_notify_pid and the update of parallel_terminate_count
314  * complete before the store to in_use.
315  */
316  notify_pid = slot->worker.bgw_notify_pid;
317  if ((slot->worker.bgw_flags & BGWORKER_CLASS_PARALLEL) != 0)
318  BackgroundWorkerData->parallel_terminate_count++;
320  slot->pid = 0;
321  slot->in_use = false;
322  if (notify_pid != 0)
323  kill(notify_pid, SIGUSR1);
324 
325  continue;
326  }
327 
328  /*
329  * Copy the registration data into the registered workers list.
330  */
331  rw = malloc(sizeof(RegisteredBgWorker));
332  if (rw == NULL)
333  {
334  ereport(LOG,
335  (errcode(ERRCODE_OUT_OF_MEMORY),
336  errmsg("out of memory")));
337  return;
338  }
339 
340  /*
341  * Copy strings in a paranoid way. If shared memory is corrupted, the
342  * source data might not even be NUL-terminated.
343  */
345  slot->worker.bgw_name, BGW_MAXLEN);
347  slot->worker.bgw_type, BGW_MAXLEN);
352 
353  /*
354  * Copy various fixed-size fields.
355  *
356  * flags, start_time, and restart_time are examined by the postmaster,
357  * but nothing too bad will happen if they are corrupted. The
358  * remaining fields will only be examined by the child process. It
359  * might crash, but we won't.
360  */
361  rw->rw_worker.bgw_flags = slot->worker.bgw_flags;
365  memcpy(rw->rw_worker.bgw_extra, slot->worker.bgw_extra, BGW_EXTRALEN);
366 
367  /*
368  * Copy the PID to be notified about state changes, but only if the
369  * postmaster knows about a backend with that PID. It isn't an error
370  * if the postmaster doesn't know about the PID, because the backend
371  * that requested the worker could have died (or been killed) just
372  * after doing so. Nonetheless, at least until we get some experience
373  * with how this plays out in the wild, log a message at a relative
374  * high debug level.
375  */
378  {
379  elog(DEBUG1, "worker notification PID %lu is not valid",
380  (long) rw->rw_worker.bgw_notify_pid);
381  rw->rw_worker.bgw_notify_pid = 0;
382  }
383 
384  /* Initialize postmaster bookkeeping. */
385  rw->rw_backend = NULL;
386  rw->rw_pid = 0;
387  rw->rw_child_slot = 0;
388  rw->rw_crashed_at = 0;
389  rw->rw_shmem_slot = slotno;
390  rw->rw_terminate = false;
391 
392  /* Log it! */
393  ereport(DEBUG1,
394  (errmsg("registering background worker \"%s\"",
395  rw->rw_worker.bgw_name)));
396 
397  slist_push_head(&BackgroundWorkerList, &rw->rw_lnode);
398  }
399 }
400 
401 /*
402  * Forget about a background worker that's no longer needed.
403  *
404  * The worker must be identified by passing an slist_mutable_iter that
405  * points to it. This convention allows deletion of workers during
406  * searches of the worker list, and saves having to search the list again.
407  *
408  * This function must be invoked only in the postmaster.
409  */
410 void
412 {
413  RegisteredBgWorker *rw;
414  BackgroundWorkerSlot *slot;
415 
416  rw = slist_container(RegisteredBgWorker, rw_lnode, cur->cur);
417 
419  slot = &BackgroundWorkerData->slot[rw->rw_shmem_slot];
421  BackgroundWorkerData->parallel_terminate_count++;
422 
423  slot->in_use = false;
424 
425  ereport(DEBUG1,
426  (errmsg("unregistering background worker \"%s\"",
427  rw->rw_worker.bgw_name)));
428 
430  free(rw);
431 }
432 
433 /*
434  * Report the PID of a newly-launched background worker in shared memory.
435  *
436  * This function should only be called from the postmaster.
437  */
438 void
440 {
441  BackgroundWorkerSlot *slot;
442 
444  slot = &BackgroundWorkerData->slot[rw->rw_shmem_slot];
445  slot->pid = rw->rw_pid;
446 
447  if (rw->rw_worker.bgw_notify_pid != 0)
449 }
450 
451 /*
452  * Report that the PID of a background worker is now zero because a
453  * previously-running background worker has exited.
454  *
455  * This function should only be called from the postmaster.
456  */
457 void
459 {
460  RegisteredBgWorker *rw;
461  BackgroundWorkerSlot *slot;
462  int notify_pid;
463 
464  rw = slist_container(RegisteredBgWorker, rw_lnode, cur->cur);
465 
467  slot = &BackgroundWorkerData->slot[rw->rw_shmem_slot];
468  slot->pid = rw->rw_pid;
469  notify_pid = rw->rw_worker.bgw_notify_pid;
470 
471  /*
472  * If this worker is slated for deregistration, do that before notifying
473  * the process which started it. Otherwise, if that process tries to
474  * reuse the slot immediately, it might not be available yet. In theory
475  * that could happen anyway if the process checks slot->pid at just the
476  * wrong moment, but this makes the window narrower.
477  */
478  if (rw->rw_terminate ||
481 
482  if (notify_pid != 0)
483  kill(notify_pid, SIGUSR1);
484 }
485 
486 /*
487  * Cancel SIGUSR1 notifications for a PID belonging to an exiting backend.
488  *
489  * This function should only be called from the postmaster.
490  */
491 void
493 {
494  slist_iter siter;
495 
496  slist_foreach(siter, &BackgroundWorkerList)
497  {
498  RegisteredBgWorker *rw;
499 
500  rw = slist_container(RegisteredBgWorker, rw_lnode, siter.cur);
501  if (rw->rw_worker.bgw_notify_pid == pid)
502  rw->rw_worker.bgw_notify_pid = 0;
503  }
504 }
505 
506 /*
507  * Reset background worker crash state.
508  *
509  * We assume that, after a crash-and-restart cycle, background workers without
510  * the never-restart flag should be restarted immediately, instead of waiting
511  * for bgw_restart_time to elapse.
512  */
513 void
515 {
516  slist_mutable_iter iter;
517 
518  slist_foreach_modify(iter, &BackgroundWorkerList)
519  {
520  RegisteredBgWorker *rw;
521 
522  rw = slist_container(RegisteredBgWorker, rw_lnode, iter.cur);
523 
525  {
526  /*
527  * Workers marked BGW_NEVER_RESTART shouldn't get relaunched after
528  * the crash, so forget about them. (If we wait until after the
529  * crash to forget about them, and they are parallel workers,
530  * parallel_terminate_count will get incremented after we've
531  * already zeroed parallel_register_count, which would be bad.)
532  */
533  ForgetBackgroundWorker(&iter);
534  }
535  else
536  {
537  /*
538  * The accounting which we do via parallel_register_count and
539  * parallel_terminate_count would get messed up if a worker marked
540  * parallel could survive a crash and restart cycle. All such
541  * workers should be marked BGW_NEVER_RESTART, and thus control
542  * should never reach this branch.
543  */
545 
546  /*
547  * Allow this worker to be restarted immediately after we finish
548  * resetting.
549  */
550  rw->rw_crashed_at = 0;
551  }
552  }
553 }
554 
555 #ifdef EXEC_BACKEND
556 /*
557  * In EXEC_BACKEND mode, workers use this to retrieve their details from
558  * shared memory.
559  */
561 BackgroundWorkerEntry(int slotno)
562 {
563  static BackgroundWorker myEntry;
564  BackgroundWorkerSlot *slot;
565 
566  Assert(slotno < BackgroundWorkerData->total_slots);
567  slot = &BackgroundWorkerData->slot[slotno];
568  Assert(slot->in_use);
569 
570  /* must copy this in case we don't intend to retain shmem access */
571  memcpy(&myEntry, &slot->worker, sizeof myEntry);
572  return &myEntry;
573 }
574 #endif
575 
576 /*
577  * Complain about the BackgroundWorker definition using error level elevel.
578  * Return true if it looks ok, false if not (unless elevel >= ERROR, in
579  * which case we won't return at all in the not-OK case).
580  */
581 static bool
583 {
584  /* sanity check for flags */
586  {
587  if (!(worker->bgw_flags & BGWORKER_SHMEM_ACCESS))
588  {
589  ereport(elevel,
590  (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
591  errmsg("background worker \"%s\": must attach to shared memory in order to request a database connection",
592  worker->bgw_name)));
593  return false;
594  }
595 
597  {
598  ereport(elevel,
599  (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
600  errmsg("background worker \"%s\": cannot request database access if starting at postmaster start",
601  worker->bgw_name)));
602  return false;
603  }
604 
605  /* XXX other checks? */
606  }
607 
608  if ((worker->bgw_restart_time < 0 &&
609  worker->bgw_restart_time != BGW_NEVER_RESTART) ||
610  (worker->bgw_restart_time > USECS_PER_DAY / 1000))
611  {
612  ereport(elevel,
613  (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
614  errmsg("background worker \"%s\": invalid restart interval",
615  worker->bgw_name)));
616  return false;
617  }
618 
619  /*
620  * Parallel workers may not be configured for restart, because the
621  * parallel_register_count/parallel_terminate_count accounting can't
622  * handle parallel workers lasting through a crash-and-restart cycle.
623  */
624  if (worker->bgw_restart_time != BGW_NEVER_RESTART &&
625  (worker->bgw_flags & BGWORKER_CLASS_PARALLEL) != 0)
626  {
627  ereport(elevel,
628  (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
629  errmsg("background worker \"%s\": parallel workers may not be configured for restart",
630  worker->bgw_name)));
631  return false;
632  }
633 
634  /*
635  * If bgw_type is not filled in, use bgw_name.
636  */
637  if (strcmp(worker->bgw_type, "") == 0)
638  strcpy(worker->bgw_type, worker->bgw_name);
639 
640  return true;
641 }
642 
643 /*
644  * Standard SIGTERM handler for background workers
645  */
646 static void
648 {
650 
651  ereport(FATAL,
652  (errcode(ERRCODE_ADMIN_SHUTDOWN),
653  errmsg("terminating background worker \"%s\" due to administrator command",
655 }
656 
657 /*
658  * Standard SIGUSR1 handler for unconnected workers
659  *
660  * Here, we want to make sure an unconnected worker will at least heed
661  * latch activity.
662  */
663 static void
665 {
666  int save_errno = errno;
667 
669 
670  errno = save_errno;
671 }
672 
673 /*
674  * Start a new background worker
675  *
676  * This is the main entry point for background worker, to be called from
677  * postmaster.
678  */
679 void
681 {
682  sigjmp_buf local_sigjmp_buf;
684  bgworker_main_type entrypt;
685 
686  if (worker == NULL)
687  elog(FATAL, "unable to find bgworker entry");
688 
689  IsBackgroundWorker = true;
690 
692  init_ps_display(worker->bgw_name);
693 
694  /*
695  * If we're not supposed to have shared memory access, then detach from
696  * shared memory. If we didn't request shared memory access, the
697  * postmaster won't force a cluster-wide restart if we exit unexpectedly,
698  * so we'd better make sure that we don't mess anything up that would
699  * require that sort of cleanup.
700  */
701  if ((worker->bgw_flags & BGWORKER_SHMEM_ACCESS) == 0)
702  {
703  dsm_detach_all();
705  }
706 
708 
709  /* Apply PostAuthDelay */
710  if (PostAuthDelay > 0)
711  pg_usleep(PostAuthDelay * 1000000L);
712 
713  /*
714  * Set up signal handlers.
715  */
717  {
718  /*
719  * SIGINT is used to signal canceling the current action
720  */
724 
725  /* XXX Any other handlers needed here? */
726  }
727  else
728  {
729  pqsignal(SIGINT, SIG_IGN);
731  pqsignal(SIGFPE, SIG_IGN);
732  }
733  pqsignal(SIGTERM, bgworker_die);
734  /* SIGQUIT handler was already set up by InitPostmasterChild */
736 
737  InitializeTimeouts(); /* establishes SIGALRM handler */
738 
742 
743  /*
744  * If an exception is encountered, processing resumes here.
745  *
746  * We just need to clean up, report the error, and go away.
747  */
748  if (sigsetjmp(local_sigjmp_buf, 1) != 0)
749  {
750  /* Since not using PG_TRY, must reset error stack by hand */
751  error_context_stack = NULL;
752 
753  /* Prevent interrupts while cleaning up */
754  HOLD_INTERRUPTS();
755 
756  /*
757  * sigsetjmp will have blocked all signals, but we may need to accept
758  * signals while communicating with our parallel leader. Once we've
759  * done HOLD_INTERRUPTS() it should be safe to unblock signals.
760  */
762 
763  /* Report the error to the parallel leader and the server log */
764  EmitErrorReport();
765 
766  /*
767  * Do we need more cleanup here? For shmem-connected bgworkers, we
768  * will call InitProcess below, which will install ProcKill as exit
769  * callback. That will take care of releasing locks, etc.
770  */
771 
772  /* and go away */
773  proc_exit(1);
774  }
775 
776  /* We can now handle ereport(ERROR) */
777  PG_exception_stack = &local_sigjmp_buf;
778 
779  /*
780  * If the background worker request shared memory access, set that up now;
781  * else, detach all shared memory segments.
782  */
783  if (worker->bgw_flags & BGWORKER_SHMEM_ACCESS)
784  {
785  /*
786  * Early initialization. Some of this could be useful even for
787  * background workers that aren't using shared memory, but they can
788  * call the individual startup routines for those subsystems if
789  * needed.
790  */
791  BaseInit();
792 
793  /*
794  * Create a per-backend PGPROC struct in shared memory, except in the
795  * EXEC_BACKEND case where this was done in SubPostmasterMain. We must
796  * do this before we can use LWLocks (and in the EXEC_BACKEND case we
797  * already had to do some stuff with LWLocks).
798  */
799 #ifndef EXEC_BACKEND
800  InitProcess();
801 #endif
802  }
803 
804  /*
805  * Look up the entry point function, loading its library if necessary.
806  */
808  worker->bgw_function_name);
809 
810  /*
811  * Note that in normal processes, we would call InitPostgres here. For a
812  * worker, however, we don't know what database to connect to, yet; so we
813  * need to wait until the user code does it via
814  * BackgroundWorkerInitializeConnection().
815  */
816 
817  /*
818  * Now invoke the user-defined worker code
819  */
820  entrypt(worker->bgw_main_arg);
821 
822  /* ... and if it returns, we're done */
823  proc_exit(0);
824 }
825 
826 /*
827  * Register a new static background worker.
828  *
829  * This can only be called directly from postmaster or in the _PG_init
830  * function of a module library that's loaded by shared_preload_libraries;
831  * otherwise it will have no effect.
832  */
833 void
835 {
836  RegisteredBgWorker *rw;
837  static int numworkers = 0;
838 
839  if (!IsUnderPostmaster)
840  ereport(DEBUG1,
841  (errmsg("registering background worker \"%s\"", worker->bgw_name)));
842 
844  strcmp(worker->bgw_library_name, "postgres") != 0)
845  {
846  if (!IsUnderPostmaster)
847  ereport(LOG,
848  (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
849  errmsg("background worker \"%s\": must be registered in shared_preload_libraries",
850  worker->bgw_name)));
851  return;
852  }
853 
854  if (!SanityCheckBackgroundWorker(worker, LOG))
855  return;
856 
857  if (worker->bgw_notify_pid != 0)
858  {
859  ereport(LOG,
860  (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
861  errmsg("background worker \"%s\": only dynamic background workers can request notification",
862  worker->bgw_name)));
863  return;
864  }
865 
866  /*
867  * Enforce maximum number of workers. Note this is overly restrictive: we
868  * could allow more non-shmem-connected workers, because these don't count
869  * towards the MAX_BACKENDS limit elsewhere. For now, it doesn't seem
870  * important to relax this restriction.
871  */
872  if (++numworkers > max_worker_processes)
873  {
874  ereport(LOG,
875  (errcode(ERRCODE_CONFIGURATION_LIMIT_EXCEEDED),
876  errmsg("too many background workers"),
877  errdetail_plural("Up to %d background worker can be registered with the current settings.",
878  "Up to %d background workers can be registered with the current settings.",
881  errhint("Consider increasing the configuration parameter \"max_worker_processes\".")));
882  return;
883  }
884 
885  /*
886  * Copy the registration data into the registered workers list.
887  */
888  rw = malloc(sizeof(RegisteredBgWorker));
889  if (rw == NULL)
890  {
891  ereport(LOG,
892  (errcode(ERRCODE_OUT_OF_MEMORY),
893  errmsg("out of memory")));
894  return;
895  }
896 
897  rw->rw_worker = *worker;
898  rw->rw_backend = NULL;
899  rw->rw_pid = 0;
900  rw->rw_child_slot = 0;
901  rw->rw_crashed_at = 0;
902  rw->rw_terminate = false;
903 
904  slist_push_head(&BackgroundWorkerList, &rw->rw_lnode);
905 }
906 
907 /*
908  * Register a new background worker from a regular backend.
909  *
910  * Returns true on success and false on failure. Failure typically indicates
911  * that no background worker slots are currently available.
912  *
913  * If handle != NULL, we'll set *handle to a pointer that can subsequently
914  * be used as an argument to GetBackgroundWorkerPid(). The caller can
915  * free this pointer using pfree(), if desired.
916  */
917 bool
919  BackgroundWorkerHandle **handle)
920 {
921  int slotno;
922  bool success = false;
923  bool parallel;
924  uint64 generation = 0;
925 
926  /*
927  * We can't register dynamic background workers from the postmaster. If
928  * this is a standalone backend, we're the only process and can't start
929  * any more. In a multi-process environment, it might be theoretically
930  * possible, but we don't currently support it due to locking
931  * considerations; see comments on the BackgroundWorkerSlot data
932  * structure.
933  */
934  if (!IsUnderPostmaster)
935  return false;
936 
937  if (!SanityCheckBackgroundWorker(worker, ERROR))
938  return false;
939 
940  parallel = (worker->bgw_flags & BGWORKER_CLASS_PARALLEL) != 0;
941 
942  LWLockAcquire(BackgroundWorkerLock, LW_EXCLUSIVE);
943 
944  /*
945  * If this is a parallel worker, check whether there are already too many
946  * parallel workers; if so, don't register another one. Our view of
947  * parallel_terminate_count may be slightly stale, but that doesn't really
948  * matter: we would have gotten the same result if we'd arrived here
949  * slightly earlier anyway. There's no help for it, either, since the
950  * postmaster must not take locks; a memory barrier wouldn't guarantee
951  * anything useful.
952  */
953  if (parallel && (BackgroundWorkerData->parallel_register_count -
954  BackgroundWorkerData->parallel_terminate_count) >=
956  {
957  Assert(BackgroundWorkerData->parallel_register_count -
958  BackgroundWorkerData->parallel_terminate_count <=
960  LWLockRelease(BackgroundWorkerLock);
961  return false;
962  }
963 
964  /*
965  * Look for an unused slot. If we find one, grab it.
966  */
967  for (slotno = 0; slotno < BackgroundWorkerData->total_slots; ++slotno)
968  {
969  BackgroundWorkerSlot *slot = &BackgroundWorkerData->slot[slotno];
970 
971  if (!slot->in_use)
972  {
973  memcpy(&slot->worker, worker, sizeof(BackgroundWorker));
974  slot->pid = InvalidPid; /* indicates not started yet */
975  slot->generation++;
976  slot->terminate = false;
977  generation = slot->generation;
978  if (parallel)
979  BackgroundWorkerData->parallel_register_count++;
980 
981  /*
982  * Make sure postmaster doesn't see the slot as in use before it
983  * sees the new contents.
984  */
986 
987  slot->in_use = true;
988  success = true;
989  break;
990  }
991  }
992 
993  LWLockRelease(BackgroundWorkerLock);
994 
995  /* If we found a slot, tell the postmaster to notice the change. */
996  if (success)
998 
999  /*
1000  * If we found a slot and the user has provided a handle, initialize it.
1001  */
1002  if (success && handle)
1003  {
1004  *handle = palloc(sizeof(BackgroundWorkerHandle));
1005  (*handle)->slot = slotno;
1006  (*handle)->generation = generation;
1007  }
1008 
1009  return success;
1010 }
1011 
1012 /*
1013  * Get the PID of a dynamically-registered background worker.
1014  *
1015  * If the worker is determined to be running, the return value will be
1016  * BGWH_STARTED and *pidp will get the PID of the worker process. If the
1017  * postmaster has not yet attempted to start the worker, the return value will
1018  * be BGWH_NOT_YET_STARTED. Otherwise, the return value is BGWH_STOPPED.
1019  *
1020  * BGWH_STOPPED can indicate either that the worker is temporarily stopped
1021  * (because it is configured for automatic restart and exited non-zero),
1022  * or that the worker is permanently stopped (because it exited with exit
1023  * code 0, or was not configured for automatic restart), or even that the
1024  * worker was unregistered without ever starting (either because startup
1025  * failed and the worker is not configured for automatic restart, or because
1026  * TerminateBackgroundWorker was used before the worker was successfully
1027  * started).
1028  */
1031 {
1032  BackgroundWorkerSlot *slot;
1033  pid_t pid;
1034 
1035  Assert(handle->slot < max_worker_processes);
1036  slot = &BackgroundWorkerData->slot[handle->slot];
1037 
1038  /*
1039  * We could probably arrange to synchronize access to data using memory
1040  * barriers only, but for now, let's just keep it simple and grab the
1041  * lock. It seems unlikely that there will be enough traffic here to
1042  * result in meaningful contention.
1043  */
1044  LWLockAcquire(BackgroundWorkerLock, LW_SHARED);
1045 
1046  /*
1047  * The generation number can't be concurrently changed while we hold the
1048  * lock. The pid, which is updated by the postmaster, can change at any
1049  * time, but we assume such changes are atomic. So the value we read
1050  * won't be garbage, but it might be out of date by the time the caller
1051  * examines it (but that's unavoidable anyway).
1052  *
1053  * The in_use flag could be in the process of changing from true to false,
1054  * but if it is already false then it can't change further.
1055  */
1056  if (handle->generation != slot->generation || !slot->in_use)
1057  pid = 0;
1058  else
1059  pid = slot->pid;
1060 
1061  /* All done. */
1062  LWLockRelease(BackgroundWorkerLock);
1063 
1064  if (pid == 0)
1065  return BGWH_STOPPED;
1066  else if (pid == InvalidPid)
1067  return BGWH_NOT_YET_STARTED;
1068  *pidp = pid;
1069  return BGWH_STARTED;
1070 }
1071 
1072 /*
1073  * Wait for a background worker to start up.
1074  *
1075  * This is like GetBackgroundWorkerPid(), except that if the worker has not
1076  * yet started, we wait for it to do so; thus, BGWH_NOT_YET_STARTED is never
1077  * returned. However, if the postmaster has died, we give up and return
1078  * BGWH_POSTMASTER_DIED, since it that case we know that startup will not
1079  * take place.
1080  */
1083 {
1085  int rc;
1086 
1087  for (;;)
1088  {
1089  pid_t pid;
1090 
1092 
1093  status = GetBackgroundWorkerPid(handle, &pid);
1094  if (status == BGWH_STARTED)
1095  *pidp = pid;
1096  if (status != BGWH_NOT_YET_STARTED)
1097  break;
1098 
1099  rc = WaitLatch(MyLatch,
1102 
1103  if (rc & WL_POSTMASTER_DEATH)
1104  {
1105  status = BGWH_POSTMASTER_DIED;
1106  break;
1107  }
1108 
1110  }
1111 
1112  return status;
1113 }
1114 
1115 /*
1116  * Wait for a background worker to stop.
1117  *
1118  * If the worker hasn't yet started, or is running, we wait for it to stop
1119  * and then return BGWH_STOPPED. However, if the postmaster has died, we give
1120  * up and return BGWH_POSTMASTER_DIED, because it's the postmaster that
1121  * notifies us when a worker's state changes.
1122  */
1125 {
1127  int rc;
1128 
1129  for (;;)
1130  {
1131  pid_t pid;
1132 
1134 
1135  status = GetBackgroundWorkerPid(handle, &pid);
1136  if (status == BGWH_STOPPED)
1137  break;
1138 
1139  rc = WaitLatch(MyLatch,
1142 
1143  if (rc & WL_POSTMASTER_DEATH)
1144  {
1145  status = BGWH_POSTMASTER_DIED;
1146  break;
1147  }
1148 
1150  }
1151 
1152  return status;
1153 }
1154 
1155 /*
1156  * Instruct the postmaster to terminate a background worker.
1157  *
1158  * Note that it's safe to do this without regard to whether the worker is
1159  * still running, or even if the worker may already have exited and been
1160  * unregistered.
1161  */
1162 void
1164 {
1165  BackgroundWorkerSlot *slot;
1166  bool signal_postmaster = false;
1167 
1168  Assert(handle->slot < max_worker_processes);
1169  slot = &BackgroundWorkerData->slot[handle->slot];
1170 
1171  /* Set terminate flag in shared memory, unless slot has been reused. */
1172  LWLockAcquire(BackgroundWorkerLock, LW_EXCLUSIVE);
1173  if (handle->generation == slot->generation)
1174  {
1175  slot->terminate = true;
1176  signal_postmaster = true;
1177  }
1178  LWLockRelease(BackgroundWorkerLock);
1179 
1180  /* Make sure the postmaster notices the change to shared memory. */
1181  if (signal_postmaster)
1183 }
1184 
1185 /*
1186  * Look up (and possibly load) a bgworker entry point function.
1187  *
1188  * For functions contained in the core code, we use library name "postgres"
1189  * and consult the InternalBGWorkers array. External functions are
1190  * looked up, and loaded if necessary, using load_external_function().
1191  *
1192  * The point of this is to pass function names as strings across process
1193  * boundaries. We can't pass actual function addresses because of the
1194  * possibility that the function has been loaded at a different address
1195  * in a different process. This is obviously a hazard for functions in
1196  * loadable libraries, but it can happen even for functions in the core code
1197  * on platforms using EXEC_BACKEND (e.g., Windows).
1198  *
1199  * At some point it might be worthwhile to get rid of InternalBGWorkers[]
1200  * in favor of applying load_external_function() for core functions too;
1201  * but that raises portability issues that are not worth addressing now.
1202  */
1203 static bgworker_main_type
1204 LookupBackgroundWorkerFunction(const char *libraryname, const char *funcname)
1205 {
1206  /*
1207  * If the function is to be loaded from postgres itself, search the
1208  * InternalBGWorkers array.
1209  */
1210  if (strcmp(libraryname, "postgres") == 0)
1211  {
1212  int i;
1213 
1214  for (i = 0; i < lengthof(InternalBGWorkers); i++)
1215  {
1216  if (strcmp(InternalBGWorkers[i].fn_name, funcname) == 0)
1217  return InternalBGWorkers[i].fn_addr;
1218  }
1219 
1220  /* We can only reach this by programming error. */
1221  elog(ERROR, "internal function \"%s\" not found", funcname);
1222  }
1223 
1224  /* Otherwise load from external library. */
1225  return (bgworker_main_type)
1226  load_external_function(libraryname, funcname, true, NULL);
1227 }
1228 
1229 /*
1230  * Given a PID, get the bgw_type of the background worker. Returns NULL if
1231  * not a valid background worker.
1232  *
1233  * The return value is in static memory belonging to this function, so it has
1234  * to be used before calling this function again. This is so that the caller
1235  * doesn't have to worry about the background worker locking protocol.
1236  */
1237 const char *
1239 {
1240  int slotno;
1241  bool found = false;
1242  static char result[BGW_MAXLEN];
1243 
1244  LWLockAcquire(BackgroundWorkerLock, LW_SHARED);
1245 
1246  for (slotno = 0; slotno < BackgroundWorkerData->total_slots; slotno++)
1247  {
1248  BackgroundWorkerSlot *slot = &BackgroundWorkerData->slot[slotno];
1249 
1250  if (slot->pid > 0 && slot->pid == pid)
1251  {
1252  strcpy(result, slot->worker.bgw_type);
1253  found = true;
1254  break;
1255  }
1256  }
1257 
1258  LWLockRelease(BackgroundWorkerLock);
1259 
1260  if (!found)
1261  return NULL;
1262 
1263  return result;
1264 }
char bgw_extra[BGW_EXTRALEN]
Definition: bgworker.h:98
const char * GetBackgroundWorkerTypeByPid(pid_t pid)
Definition: bgworker.c:1238
void InitializeTimeouts(void)
Definition: timeout.c:346
static void bgworker_sigusr1_handler(SIGNAL_ARGS)
Definition: bgworker.c:664
struct bkend * rw_backend
int max_parallel_workers
Definition: globals.c:135
BgwHandleStatus WaitForBackgroundWorkerStartup(BackgroundWorkerHandle *handle, pid_t *pidp)
Definition: bgworker.c:1082
void RegisterBackgroundWorker(BackgroundWorker *worker)
Definition: bgworker.c:834
#define DEBUG1
Definition: elog.h:25
int errhint(const char *fmt,...)
Definition: elog.c:1068
void BackgroundWorkerStateChange(void)
Definition: bgworker.c:240
void ApplyLauncherMain(Datum main_arg)
Definition: launcher.c:950
bool process_shared_preload_libraries_in_progress
Definition: miscinit.c:1596
#define SIGUSR1
Definition: win32_port.h:171
BackendType MyBackendType
Definition: miscinit.c:62
#define SIGCHLD
Definition: win32_port.h:169
slist_node * cur
Definition: ilist.h:226
static BackgroundWorkerArray * BackgroundWorkerData
Definition: bgworker.c:110
slist_node * cur
Definition: ilist.h:241
static void slist_push_head(slist_head *head, slist_node *node)
Definition: ilist.h:574
#define FLEXIBLE_ARRAY_MEMBER
Definition: c.h:284
void ReportBackgroundWorkerExit(slist_mutable_iter *cur)
Definition: bgworker.c:458
int bgw_restart_time
Definition: bgworker.h:94
struct cursor * cur
Definition: ecpg.c:28
void proc_exit(int code)
Definition: ipc.c:104
int errcode(int sqlerrcode)
Definition: elog.c:610
#define BGWORKER_CLASS_PARALLEL
Definition: bgworker.h:67
static const struct @21 InternalBGWorkers[]
#define SetProcessingMode(mode)
Definition: miscadmin.h:399
BackgroundWorker * MyBgworkerEntry
Definition: postmaster.c:195
#define kill(pid, sig)
Definition: win32_port.h:454
void BaseInit(void)
Definition: postinit.c:548
void ascii_safe_strlcpy(char *dest, const char *src, size_t destsiz)
Definition: ascii.c:173
#define lengthof(array)
Definition: c.h:676
#define SIGPIPE
Definition: win32_port.h:164
const char * fn_name
Definition: bgworker.c:118
#define SIGUSR2
Definition: win32_port.h:172
#define LOG
Definition: elog.h:26
#define BGWORKER_SHMEM_ACCESS
Definition: bgworker.h:52
void BackgroundWorkerStopNotifications(pid_t pid)
Definition: bgworker.c:492
void ResetBackgroundWorkerCrashTimes(void)
Definition: bgworker.c:514
char bgw_function_name[BGW_MAXLEN]
Definition: bgworker.h:96
Size BackgroundWorkerShmemSize(void)
Definition: bgworker.c:142
void PGSharedMemoryDetach(void)
Definition: sysv_shmem.c:909
void ResetLatch(Latch *latch)
Definition: latch.c:588
#define PG_SETMASK(mask)
Definition: pqsignal.h:19
void StatementCancelHandler(SIGNAL_ARGS)
Definition: postgres.c:2857
int WaitLatch(Latch *latch, int wakeEvents, long timeout, uint32 wait_event_info)
Definition: latch.c:390
bool IsBackgroundWorker
Definition: globals.c:111
void StartBackgroundWorker(void)
Definition: bgworker.c:680
#define slist_foreach_modify(iter, lhead)
Definition: ilist.h:716
#define malloc(a)
Definition: header.h:50
void LWLockRelease(LWLock *lock)
Definition: lwlock.c:1812
ErrorContextCallback * error_context_stack
Definition: elog.c:92
void pg_usleep(long microsec)
Definition: signal.c:53
Datum bgw_main_arg
Definition: bgworker.h:97
void BackgroundWorkerShmemInit(void)
Definition: bgworker.c:158
void ReportBackgroundWorkerPID(RegisteredBgWorker *rw)
Definition: bgworker.c:439
#define ERROR
Definition: elog.h:43
BgwHandleStatus WaitForBackgroundWorkerShutdown(BackgroundWorkerHandle *handle)
Definition: bgworker.c:1124
#define SLIST_STATIC_INIT(name)
Definition: ilist.h:249
struct BackgroundWorkerSlot BackgroundWorkerSlot
void * ShmemInitStruct(const char *name, Size size, bool *foundPtr)
Definition: shmem.c:392
#define FATAL
Definition: elog.h:52
bool PostmasterMarkPIDForWorkerNotify(int pid)
Definition: postmaster.c:6142
slist_head BackgroundWorkerList
Definition: bgworker.c:42
void InitProcess(void)
Definition: proc.c:301
uint32 parallel_register_count
Definition: bgworker.c:99
bool IsUnderPostmaster
Definition: globals.c:109
#define SIGHUP
Definition: win32_port.h:159
void ForgetBackgroundWorker(slist_mutable_iter *cur)
Definition: bgworker.c:411
#define BGW_NEVER_RESTART
Definition: bgworker.h:84
unsigned int uint32
Definition: c.h:375
BgwHandleStatus
Definition: bgworker.h:102
#define USECS_PER_DAY
Definition: timestamp.h:91
static bgworker_main_type LookupBackgroundWorkerFunction(const char *libraryname, const char *funcname)
Definition: bgworker.c:1204
void * load_external_function(const char *filename, const char *funcname, bool signalNotFound, void **filehandle)
Definition: dfmgr.c:107
static bool SanityCheckBackgroundWorker(BackgroundWorker *worker, int elevel)
Definition: bgworker.c:582
#define SIG_IGN
Definition: win32_port.h:156
static int elevel
Definition: vacuumlazy.c:333
Size mul_size(Size s1, Size s2)
Definition: shmem.c:515
#define slist_container(type, membername, ptr)
Definition: ilist.h:674
sigset_t BlockSig
Definition: pqsignal.c:22
#define WL_POSTMASTER_DEATH
Definition: latch.h:128
struct BackgroundWorkerArray BackgroundWorkerArray
Size add_size(Size s1, Size s2)
Definition: shmem.c:498
int PostAuthDelay
Definition: postgres.c:102
void EmitErrorReport(void)
Definition: elog.c:1420
static RegisteredBgWorker * FindRegisteredWorkerBySlotNumber(int slotno)
Definition: bgworker.c:217
void TerminateBackgroundWorker(BackgroundWorkerHandle *handle)
Definition: bgworker.c:1163
#define ereport(elevel,...)
Definition: elog.h:144
pqsigfunc pqsignal(int signum, pqsigfunc handler)
Definition: signal.c:170
#define free(a)
Definition: header.h:65
#define pg_memory_barrier()
Definition: atomics.h:145
#define SIG_DFL
Definition: win32_port.h:154
#define SIGNAL_ARGS
Definition: c.h:1275
char bgw_name[BGW_MAXLEN]
Definition: bgworker.h:90
#define Assert(condition)
Definition: c.h:746
BackgroundWorker worker
Definition: bgworker.c:82
#define BGWORKER_BACKEND_DATABASE_CONNECTION
Definition: bgworker.h:59
#define pg_read_barrier()
Definition: atomics.h:158
#define BGW_MAXLEN
Definition: bgworker.h:85
size_t Size
Definition: c.h:474
BgWorkerStartTime bgw_start_time
Definition: bgworker.h:93
int errdetail_plural(const char *fmt_singular, const char *fmt_plural, unsigned long n,...)
Definition: elog.c:1046
bool RegisterDynamicBackgroundWorker(BackgroundWorker *worker, BackgroundWorkerHandle **handle)
Definition: bgworker.c:918
bool LWLockAcquire(LWLock *lock, LWLockMode mode)
Definition: lwlock.c:1208
void dsm_detach_all(void)
Definition: dsm.c:741
sigjmp_buf * PG_exception_stack
Definition: elog.c:94
#define slist_foreach(iter, lhead)
Definition: ilist.h:700
bgworker_main_type fn_addr
Definition: bgworker.c:119
char bgw_type[BGW_MAXLEN]
Definition: bgworker.h:91
void * palloc(Size size)
Definition: mcxt.c:950
int errmsg(const char *fmt,...)
Definition: elog.c:821
void ParallelWorkerMain(Datum main_arg)
Definition: parallel.c:1239
uint32 parallel_terminate_count
Definition: bgworker.c:100
pid_t bgw_notify_pid
Definition: bgworker.h:99
#define HOLD_INTERRUPTS()
Definition: miscadmin.h:116
#define elog(elevel,...)
Definition: elog.h:214
int i
void FloatExceptionHandler(SIGNAL_ARGS)
Definition: postgres.c:2878
#define pg_write_barrier()
Definition: atomics.h:159
int max_worker_processes
Definition: globals.c:134
struct Latch * MyLatch
Definition: globals.c:54
static void bgworker_die(SIGNAL_ARGS)
Definition: bgworker.c:647
#define CHECK_FOR_INTERRUPTS()
Definition: miscadmin.h:99
static void slist_delete_current(slist_mutable_iter *iter)
Definition: ilist.h:652
void SendPostmasterSignal(PMSignalReason reason)
Definition: pmsignal.c:146
void init_ps_display(const char *fixed_part)
Definition: ps_status.c:258
void latch_sigusr1_handler(void)
Definition: latch.c:1940
#define MAX_PARALLEL_WORKER_LIMIT
void procsignal_sigusr1_handler(SIGNAL_ARGS)
Definition: procsignal.c:551
static void static void status(const char *fmt,...) pg_attribute_printf(1
Definition: pg_regress.c:227
BackgroundWorker rw_worker
static bool success
Definition: initdb.c:162
BackgroundWorkerSlot slot[FLEXIBLE_ARRAY_MEMBER]
Definition: bgworker.c:101
char bgw_library_name[BGW_MAXLEN]
Definition: bgworker.h:95
#define BGW_EXTRALEN
Definition: bgworker.h:86
#define WL_LATCH_SET
Definition: latch.h:124
void(* bgworker_main_type)(Datum main_arg)
Definition: bgworker.h:71
#define offsetof(type, field)
Definition: c.h:669
BgwHandleStatus GetBackgroundWorkerPid(BackgroundWorkerHandle *handle, pid_t *pidp)
Definition: bgworker.c:1030
void ApplyWorkerMain(Datum main_arg)
Definition: worker.c:2917
#define InvalidPid
Definition: miscadmin.h:32
void BackgroundWorkerUnblockSignals(void)
Definition: postmaster.c:5788