PostgreSQL Source Code  git master
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros
bgworker.c
Go to the documentation of this file.
1 /*--------------------------------------------------------------------
2  * bgworker.c
3  * POSTGRES pluggable background workers implementation
4  *
5  * Portions Copyright (c) 1996-2017, 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"
23 #include "postmaster/postmaster.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  */
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  */
77 typedef 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 */
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  */
97 typedef struct BackgroundWorkerArray
98 {
102  BackgroundWorkerSlot slot[FLEXIBLE_ARRAY_MEMBER];
104 
106 {
107  int slot;
108  uint64 generation;
109 };
110 
112 
113 /*
114  * List of internal background worker entry points. We need this for
115  * reasons explained in LookupBackgroundWorkerFunction(), below.
116  */
117 static const struct
118 {
119  const char *fn_name;
121 } InternalBGWorkers[] =
122 
123 {
124  {
125  "ParallelWorkerMain", ParallelWorkerMain
126  },
127  {
128  "ApplyLauncherMain", ApplyLauncherMain
129  },
130  {
131  "ApplyWorkerMain", ApplyWorkerMain
132  }
133 };
134 
135 /* Private functions. */
136 static bgworker_main_type LookupBackgroundWorkerFunction(const char *libraryname, const char *funcname);
137 
138 
139 /*
140  * Calculate shared memory needed.
141  */
142 Size
144 {
145  Size size;
146 
147  /* Array of workers is variably sized. */
148  size = offsetof(BackgroundWorkerArray, slot);
150  sizeof(BackgroundWorkerSlot)));
151 
152  return size;
153 }
154 
155 /*
156  * Initialize shared memory.
157  */
158 void
160 {
161  bool found;
162 
163  BackgroundWorkerData = ShmemInitStruct("Background Worker Data",
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  */
217 static RegisteredBgWorker *
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  */
240 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",
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. */
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. */
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++;
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  */
346  slot->worker.bgw_name, BGW_MAXLEN);
351 
352  /*
353  * Copy various fixed-size fields.
354  *
355  * flags, start_time, and restart_time are examined by the postmaster,
356  * but nothing too bad will happen if they are corrupted. The
357  * remaining fields will only be examined by the child process. It
358  * might crash, but we won't.
359  */
360  rw->rw_worker.bgw_flags = slot->worker.bgw_flags;
364  memcpy(rw->rw_worker.bgw_extra, slot->worker.bgw_extra, BGW_EXTRALEN);
365 
366  /*
367  * Copy the PID to be notified about state changes, but only if the
368  * postmaster knows about a backend with that PID. It isn't an error
369  * if the postmaster doesn't know about the PID, because the backend
370  * that requested the worker could have died (or been killed) just
371  * after doing so. Nonetheless, at least until we get some experience
372  * with how this plays out in the wild, log a message at a relative
373  * high debug level.
374  */
377  {
378  elog(DEBUG1, "worker notification PID %lu is not valid",
379  (long) rw->rw_worker.bgw_notify_pid);
380  rw->rw_worker.bgw_notify_pid = 0;
381  }
382 
383  /* Initialize postmaster bookkeeping. */
384  rw->rw_backend = NULL;
385  rw->rw_pid = 0;
386  rw->rw_child_slot = 0;
387  rw->rw_crashed_at = 0;
388  rw->rw_shmem_slot = slotno;
389  rw->rw_terminate = false;
390 
391  /* Log it! */
392  ereport(DEBUG1,
393  (errmsg("registering background worker \"%s\"",
394  rw->rw_worker.bgw_name)));
395 
396  slist_push_head(&BackgroundWorkerList, &rw->rw_lnode);
397  }
398 }
399 
400 /*
401  * Forget about a background worker that's no longer needed.
402  *
403  * The worker must be identified by passing an slist_mutable_iter that
404  * points to it. This convention allows deletion of workers during
405  * searches of the worker list, and saves having to search the list again.
406  *
407  * This function must be invoked only in the postmaster.
408  */
409 void
411 {
412  RegisteredBgWorker *rw;
413  BackgroundWorkerSlot *slot;
414 
415  rw = slist_container(RegisteredBgWorker, rw_lnode, cur->cur);
416 
418  slot = &BackgroundWorkerData->slot[rw->rw_shmem_slot];
420  BackgroundWorkerData->parallel_terminate_count++;
421 
422  slot->in_use = false;
423 
424  ereport(DEBUG1,
425  (errmsg("unregistering background worker \"%s\"",
426  rw->rw_worker.bgw_name)));
427 
429  free(rw);
430 }
431 
432 /*
433  * Report the PID of a newly-launched background worker in shared memory.
434  *
435  * This function should only be called from the postmaster.
436  */
437 void
439 {
440  BackgroundWorkerSlot *slot;
441 
443  slot = &BackgroundWorkerData->slot[rw->rw_shmem_slot];
444  slot->pid = rw->rw_pid;
445 
446  if (rw->rw_worker.bgw_notify_pid != 0)
447  kill(rw->rw_worker.bgw_notify_pid, SIGUSR1);
448 }
449 
450 /*
451  * Report that the PID of a background worker is now zero because a
452  * previously-running background worker has exited.
453  *
454  * This function should only be called from the postmaster.
455  */
456 void
458 {
459  RegisteredBgWorker *rw;
460  BackgroundWorkerSlot *slot;
461  int notify_pid;
462 
463  rw = slist_container(RegisteredBgWorker, rw_lnode, cur->cur);
464 
466  slot = &BackgroundWorkerData->slot[rw->rw_shmem_slot];
467  slot->pid = rw->rw_pid;
468  notify_pid = rw->rw_worker.bgw_notify_pid;
469 
470  /*
471  * If this worker is slated for deregistration, do that before notifying
472  * the process which started it. Otherwise, if that process tries to
473  * reuse the slot immediately, it might not be available yet. In theory
474  * that could happen anyway if the process checks slot->pid at just the
475  * wrong moment, but this makes the window narrower.
476  */
477  if (rw->rw_terminate ||
480 
481  if (notify_pid != 0)
482  kill(notify_pid, SIGUSR1);
483 }
484 
485 /*
486  * Cancel SIGUSR1 notifications for a PID belonging to an exiting backend.
487  *
488  * This function should only be called from the postmaster.
489  */
490 void
492 {
493  slist_iter siter;
494 
495  slist_foreach(siter, &BackgroundWorkerList)
496  {
497  RegisteredBgWorker *rw;
498 
499  rw = slist_container(RegisteredBgWorker, rw_lnode, siter.cur);
500  if (rw->rw_worker.bgw_notify_pid == pid)
501  rw->rw_worker.bgw_notify_pid = 0;
502  }
503 }
504 
505 /*
506  * Reset background worker crash state.
507  *
508  * We assume that, after a crash-and-restart cycle, background workers without
509  * the never-restart flag should be restarted immediately, instead of waiting
510  * for bgw_restart_time to elapse.
511  */
512 void
514 {
515  slist_mutable_iter iter;
516 
517  slist_foreach_modify(iter, &BackgroundWorkerList)
518  {
519  RegisteredBgWorker *rw;
520 
521  rw = slist_container(RegisteredBgWorker, rw_lnode, iter.cur);
522 
524  {
525  /*
526  * Workers marked BGW_NVER_RESTART shouldn't get relaunched after
527  * the crash, so forget about them. (If we wait until after the
528  * crash to forget about them, and they are parallel workers,
529  * parallel_terminate_count will get incremented after we've
530  * already zeroed parallel_register_count, which would be bad.)
531  */
532  ForgetBackgroundWorker(&iter);
533  }
534  else
535  {
536  /*
537  * The accounting which we do via parallel_register_count and
538  * parallel_terminate_count would get messed up if a worker marked
539  * parallel could survive a crash and restart cycle. All such
540  * workers should be marked BGW_NEVER_RESTART, and thus control
541  * should never reach this branch.
542  */
544 
545  /*
546  * Allow this worker to be restarted immediately after we finish
547  * resetting.
548  */
549  rw->rw_crashed_at = 0;
550  }
551  }
552 }
553 
554 #ifdef EXEC_BACKEND
555 /*
556  * In EXEC_BACKEND mode, workers use this to retrieve their details from
557  * shared memory.
558  */
560 BackgroundWorkerEntry(int slotno)
561 {
562  static BackgroundWorker myEntry;
563  BackgroundWorkerSlot *slot;
564 
565  Assert(slotno < BackgroundWorkerData->total_slots);
566  slot = &BackgroundWorkerData->slot[slotno];
567  Assert(slot->in_use);
568 
569  /* must copy this in case we don't intend to retain shmem access */
570  memcpy(&myEntry, &slot->worker, sizeof myEntry);
571  return &myEntry;
572 }
573 #endif
574 
575 /*
576  * Complain about the BackgroundWorker definition using error level elevel.
577  * Return true if it looks ok, false if not (unless elevel >= ERROR, in
578  * which case we won't return at all in the not-OK case).
579  */
580 static bool
582 {
583  /* sanity check for flags */
585  {
586  if (!(worker->bgw_flags & BGWORKER_SHMEM_ACCESS))
587  {
588  ereport(elevel,
589  (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
590  errmsg("background worker \"%s\": must attach to shared memory in order to request a database connection",
591  worker->bgw_name)));
592  return false;
593  }
594 
596  {
597  ereport(elevel,
598  (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
599  errmsg("background worker \"%s\": cannot request database access if starting at postmaster start",
600  worker->bgw_name)));
601  return false;
602  }
603 
604  /* XXX other checks? */
605  }
606 
607  if ((worker->bgw_restart_time < 0 &&
608  worker->bgw_restart_time != BGW_NEVER_RESTART) ||
609  (worker->bgw_restart_time > USECS_PER_DAY / 1000))
610  {
611  ereport(elevel,
612  (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
613  errmsg("background worker \"%s\": invalid restart interval",
614  worker->bgw_name)));
615  return false;
616  }
617 
618  /*
619  * Parallel workers may not be configured for restart, because the
620  * parallel_register_count/parallel_terminate_count accounting can't
621  * handle parallel workers lasting through a crash-and-restart cycle.
622  */
623  if (worker->bgw_restart_time != BGW_NEVER_RESTART &&
624  (worker->bgw_flags & BGWORKER_CLASS_PARALLEL) != 0)
625  {
626  ereport(elevel,
627  (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
628  errmsg("background worker \"%s\": parallel workers may not be configured for restart",
629  worker->bgw_name)));
630  return false;
631  }
632 
633  return true;
634 }
635 
636 static void
638 {
639  sigaddset(&BlockSig, SIGQUIT); /* prevent nested calls */
641 
642  /*
643  * We DO NOT want to run proc_exit() callbacks -- we're here because
644  * shared memory may be corrupted, so we don't want to try to clean up our
645  * transaction. Just nail the windows shut and get out of town. Now that
646  * there's an atexit callback to prevent third-party code from breaking
647  * things by calling exit() directly, we have to reset the callbacks
648  * explicitly to make this work as intended.
649  */
650  on_exit_reset();
651 
652  /*
653  * Note we do exit(2) not exit(0). This is to force the postmaster into a
654  * system reset cycle if some idiot DBA sends a manual SIGQUIT to a random
655  * backend. This is necessary precisely because we don't clean up our
656  * shared memory state. (The "dead man switch" mechanism in pmsignal.c
657  * should ensure the postmaster sees this as a crash, too, but no harm in
658  * being doubly sure.)
659  */
660  exit(2);
661 }
662 
663 /*
664  * Standard SIGTERM handler for background workers
665  */
666 static void
668 {
670 
671  ereport(FATAL,
672  (errcode(ERRCODE_ADMIN_SHUTDOWN),
673  errmsg("terminating background worker \"%s\" due to administrator command",
675 }
676 
677 /*
678  * Standard SIGUSR1 handler for unconnected workers
679  *
680  * Here, we want to make sure an unconnected worker will at least heed
681  * latch activity.
682  */
683 static void
685 {
686  int save_errno = errno;
687 
689 
690  errno = save_errno;
691 }
692 
693 /*
694  * Start a new background worker
695  *
696  * This is the main entry point for background worker, to be called from
697  * postmaster.
698  */
699 void
701 {
702  sigjmp_buf local_sigjmp_buf;
703  char buf[MAXPGPATH];
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  snprintf(buf, MAXPGPATH, "bgworker: %s", worker->bgw_name);
714  init_ps_display(buf, "", "", "");
715 
716  /*
717  * If we're not supposed to have shared memory access, then detach from
718  * shared memory. If we didn't request shared memory access, the
719  * postmaster won't force a cluster-wide restart if we exit unexpectedly,
720  * so we'd better make sure that we don't mess anything up that would
721  * require that sort of cleanup.
722  */
723  if ((worker->bgw_flags & BGWORKER_SHMEM_ACCESS) == 0)
724  {
725  dsm_detach_all();
727  }
728 
730 
731  /* Apply PostAuthDelay */
732  if (PostAuthDelay > 0)
733  pg_usleep(PostAuthDelay * 1000000L);
734 
735  /*
736  * Set up signal handlers.
737  */
739  {
740  /*
741  * SIGINT is used to signal canceling the current action
742  */
746 
747  /* XXX Any other handlers needed here? */
748  }
749  else
750  {
751  pqsignal(SIGINT, SIG_IGN);
753  pqsignal(SIGFPE, SIG_IGN);
754  }
755  pqsignal(SIGTERM, bgworker_die);
757 
759  InitializeTimeouts(); /* establishes SIGALRM handler */
760 
764 
765  /*
766  * If an exception is encountered, processing resumes here.
767  *
768  * See notes in postgres.c about the design of this coding.
769  */
770  if (sigsetjmp(local_sigjmp_buf, 1) != 0)
771  {
772  /* Since not using PG_TRY, must reset error stack by hand */
774 
775  /* Prevent interrupts while cleaning up */
776  HOLD_INTERRUPTS();
777 
778  /* Report the error to the server log */
779  EmitErrorReport();
780 
781  /*
782  * Do we need more cleanup here? For shmem-connected bgworkers, we
783  * will call InitProcess below, which will install ProcKill as exit
784  * callback. That will take care of releasing locks, etc.
785  */
786 
787  /* and go away */
788  proc_exit(1);
789  }
790 
791  /* We can now handle ereport(ERROR) */
792  PG_exception_stack = &local_sigjmp_buf;
793 
794  /*
795  * If the background worker request shared memory access, set that up now;
796  * else, detach all shared memory segments.
797  */
798  if (worker->bgw_flags & BGWORKER_SHMEM_ACCESS)
799  {
800  /*
801  * Early initialization. Some of this could be useful even for
802  * background workers that aren't using shared memory, but they can
803  * call the individual startup routines for those subsystems if
804  * needed.
805  */
806  BaseInit();
807 
808  /*
809  * Create a per-backend PGPROC struct in shared memory, except in the
810  * EXEC_BACKEND case where this was done in SubPostmasterMain. We must
811  * do this before we can use LWLocks (and in the EXEC_BACKEND case we
812  * already had to do some stuff with LWLocks).
813  */
814 #ifndef EXEC_BACKEND
815  InitProcess();
816 #endif
817  }
818 
819  /*
820  * Look up the entry point function, loading its library if necessary.
821  */
823  worker->bgw_function_name);
824 
825  /*
826  * Note that in normal processes, we would call InitPostgres here. For a
827  * worker, however, we don't know what database to connect to, yet; so we
828  * need to wait until the user code does it via
829  * BackgroundWorkerInitializeConnection().
830  */
831 
832  /*
833  * Now invoke the user-defined worker code
834  */
835  entrypt(worker->bgw_main_arg);
836 
837  /* ... and if it returns, we're done */
838  proc_exit(0);
839 }
840 
841 /*
842  * Register a new static background worker.
843  *
844  * This can only be called directly from postmaster or in the _PG_init
845  * function of a module library that's loaded by shared_preload_libraries;
846  * otherwise it will have no effect.
847  */
848 void
850 {
851  RegisteredBgWorker *rw;
852  static int numworkers = 0;
853 
854  if (!IsUnderPostmaster)
855  ereport(DEBUG1,
856  (errmsg("registering background worker \"%s\"", worker->bgw_name)));
857 
859  strcmp(worker->bgw_library_name, "postgres") != 0)
860  {
861  if (!IsUnderPostmaster)
862  ereport(LOG,
863  (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
864  errmsg("background worker \"%s\": must be registered in shared_preload_libraries",
865  worker->bgw_name)));
866  return;
867  }
868 
869  if (!SanityCheckBackgroundWorker(worker, LOG))
870  return;
871 
872  if (worker->bgw_notify_pid != 0)
873  {
874  ereport(LOG,
875  (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
876  errmsg("background worker \"%s\": only dynamic background workers can request notification",
877  worker->bgw_name)));
878  return;
879  }
880 
881  /*
882  * Enforce maximum number of workers. Note this is overly restrictive: we
883  * could allow more non-shmem-connected workers, because these don't count
884  * towards the MAX_BACKENDS limit elsewhere. For now, it doesn't seem
885  * important to relax this restriction.
886  */
887  if (++numworkers > max_worker_processes)
888  {
889  ereport(LOG,
890  (errcode(ERRCODE_CONFIGURATION_LIMIT_EXCEEDED),
891  errmsg("too many background workers"),
892  errdetail_plural("Up to %d background worker can be registered with the current settings.",
893  "Up to %d background workers can be registered with the current settings.",
896  errhint("Consider increasing the configuration parameter \"max_worker_processes\".")));
897  return;
898  }
899 
900  /*
901  * Copy the registration data into the registered workers list.
902  */
903  rw = malloc(sizeof(RegisteredBgWorker));
904  if (rw == NULL)
905  {
906  ereport(LOG,
907  (errcode(ERRCODE_OUT_OF_MEMORY),
908  errmsg("out of memory")));
909  return;
910  }
911 
912  rw->rw_worker = *worker;
913  rw->rw_backend = NULL;
914  rw->rw_pid = 0;
915  rw->rw_child_slot = 0;
916  rw->rw_crashed_at = 0;
917  rw->rw_terminate = false;
918 
919  slist_push_head(&BackgroundWorkerList, &rw->rw_lnode);
920 }
921 
922 /*
923  * Register a new background worker from a regular backend.
924  *
925  * Returns true on success and false on failure. Failure typically indicates
926  * that no background worker slots are currently available.
927  *
928  * If handle != NULL, we'll set *handle to a pointer that can subsequently
929  * be used as an argument to GetBackgroundWorkerPid(). The caller can
930  * free this pointer using pfree(), if desired.
931  */
932 bool
934  BackgroundWorkerHandle **handle)
935 {
936  int slotno;
937  bool success = false;
938  bool parallel;
939  uint64 generation = 0;
940 
941  /*
942  * We can't register dynamic background workers from the postmaster. If
943  * this is a standalone backend, we're the only process and can't start
944  * any more. In a multi-process environment, it might be theoretically
945  * possible, but we don't currently support it due to locking
946  * considerations; see comments on the BackgroundWorkerSlot data
947  * structure.
948  */
949  if (!IsUnderPostmaster)
950  return false;
951 
952  if (!SanityCheckBackgroundWorker(worker, ERROR))
953  return false;
954 
955  parallel = (worker->bgw_flags & BGWORKER_CLASS_PARALLEL) != 0;
956 
957  LWLockAcquire(BackgroundWorkerLock, LW_EXCLUSIVE);
958 
959  /*
960  * If this is a parallel worker, check whether there are already too many
961  * parallel workers; if so, don't register another one. Our view of
962  * parallel_terminate_count may be slightly stale, but that doesn't really
963  * matter: we would have gotten the same result if we'd arrived here
964  * slightly earlier anyway. There's no help for it, either, since the
965  * postmaster must not take locks; a memory barrier wouldn't guarantee
966  * anything useful.
967  */
968  if (parallel && (BackgroundWorkerData->parallel_register_count -
969  BackgroundWorkerData->parallel_terminate_count) >=
971  {
972  Assert(BackgroundWorkerData->parallel_register_count -
973  BackgroundWorkerData->parallel_terminate_count <=
975  LWLockRelease(BackgroundWorkerLock);
976  return false;
977  }
978 
979  /*
980  * Look for an unused slot. If we find one, grab it.
981  */
982  for (slotno = 0; slotno < BackgroundWorkerData->total_slots; ++slotno)
983  {
984  BackgroundWorkerSlot *slot = &BackgroundWorkerData->slot[slotno];
985 
986  if (!slot->in_use)
987  {
988  memcpy(&slot->worker, worker, sizeof(BackgroundWorker));
989  slot->pid = InvalidPid; /* indicates not started yet */
990  slot->generation++;
991  slot->terminate = false;
992  generation = slot->generation;
993  if (parallel)
994  BackgroundWorkerData->parallel_register_count++;
995 
996  /*
997  * Make sure postmaster doesn't see the slot as in use before it
998  * sees the new contents.
999  */
1000  pg_write_barrier();
1001 
1002  slot->in_use = true;
1003  success = true;
1004  break;
1005  }
1006  }
1007 
1008  LWLockRelease(BackgroundWorkerLock);
1009 
1010  /* If we found a slot, tell the postmaster to notice the change. */
1011  if (success)
1013 
1014  /*
1015  * If we found a slot and the user has provided a handle, initialize it.
1016  */
1017  if (success && handle)
1018  {
1019  *handle = palloc(sizeof(BackgroundWorkerHandle));
1020  (*handle)->slot = slotno;
1021  (*handle)->generation = generation;
1022  }
1023 
1024  return success;
1025 }
1026 
1027 /*
1028  * Get the PID of a dynamically-registered background worker.
1029  *
1030  * If the worker is determined to be running, the return value will be
1031  * BGWH_STARTED and *pidp will get the PID of the worker process.
1032  * Otherwise, the return value will be BGWH_NOT_YET_STARTED if the worker
1033  * hasn't been started yet, and BGWH_STOPPED if the worker was previously
1034  * running but is no longer.
1035  *
1036  * In the latter case, the worker may be stopped temporarily (if it is
1037  * configured for automatic restart and exited non-zero) or gone for
1038  * good (if it exited with code 0 or if it is configured not to restart).
1039  */
1042 {
1043  BackgroundWorkerSlot *slot;
1044  pid_t pid;
1045 
1046  Assert(handle->slot < max_worker_processes);
1047  slot = &BackgroundWorkerData->slot[handle->slot];
1048 
1049  /*
1050  * We could probably arrange to synchronize access to data using memory
1051  * barriers only, but for now, let's just keep it simple and grab the
1052  * lock. It seems unlikely that there will be enough traffic here to
1053  * result in meaningful contention.
1054  */
1055  LWLockAcquire(BackgroundWorkerLock, LW_SHARED);
1056 
1057  /*
1058  * The generation number can't be concurrently changed while we hold the
1059  * lock. The pid, which is updated by the postmaster, can change at any
1060  * time, but we assume such changes are atomic. So the value we read
1061  * won't be garbage, but it might be out of date by the time the caller
1062  * examines it (but that's unavoidable anyway).
1063  */
1064  if (handle->generation != slot->generation)
1065  pid = 0;
1066  else
1067  pid = slot->pid;
1068 
1069  /* All done. */
1070  LWLockRelease(BackgroundWorkerLock);
1071 
1072  if (pid == 0)
1073  return BGWH_STOPPED;
1074  else if (pid == InvalidPid)
1075  return BGWH_NOT_YET_STARTED;
1076  *pidp = pid;
1077  return BGWH_STARTED;
1078 }
1079 
1080 /*
1081  * Wait for a background worker to start up.
1082  *
1083  * This is like GetBackgroundWorkerPid(), except that if the worker has not
1084  * yet started, we wait for it to do so; thus, BGWH_NOT_YET_STARTED is never
1085  * returned. However, if the postmaster has died, we give up and return
1086  * BGWH_POSTMASTER_DIED, since it that case we know that startup will not
1087  * take place.
1088  */
1091 {
1093  int rc;
1094 
1095  for (;;)
1096  {
1097  pid_t pid;
1098 
1100 
1101  status = GetBackgroundWorkerPid(handle, &pid);
1102  if (status == BGWH_STARTED)
1103  *pidp = pid;
1104  if (status != BGWH_NOT_YET_STARTED)
1105  break;
1106 
1107  rc = WaitLatch(MyLatch,
1110 
1111  if (rc & WL_POSTMASTER_DEATH)
1112  {
1113  status = BGWH_POSTMASTER_DIED;
1114  break;
1115  }
1116 
1118  }
1119 
1120  return status;
1121 }
1122 
1123 /*
1124  * Wait for a background worker to stop.
1125  *
1126  * If the worker hasn't yet started, or is running, we wait for it to stop
1127  * and then return BGWH_STOPPED. However, if the postmaster has died, we give
1128  * up and return BGWH_POSTMASTER_DIED, because it's the postmaster that
1129  * notifies us when a worker's state changes.
1130  */
1133 {
1135  int rc;
1136 
1137  for (;;)
1138  {
1139  pid_t pid;
1140 
1142 
1143  status = GetBackgroundWorkerPid(handle, &pid);
1144  if (status == BGWH_STOPPED)
1145  break;
1146 
1147  rc = WaitLatch(&MyProc->procLatch,
1150 
1151  if (rc & WL_POSTMASTER_DEATH)
1152  {
1153  status = BGWH_POSTMASTER_DIED;
1154  break;
1155  }
1156 
1158  }
1159 
1160  return status;
1161 }
1162 
1163 /*
1164  * Instruct the postmaster to terminate a background worker.
1165  *
1166  * Note that it's safe to do this without regard to whether the worker is
1167  * still running, or even if the worker may already have existed and been
1168  * unregistered.
1169  */
1170 void
1172 {
1173  BackgroundWorkerSlot *slot;
1174  bool signal_postmaster = false;
1175 
1176  Assert(handle->slot < max_worker_processes);
1177  slot = &BackgroundWorkerData->slot[handle->slot];
1178 
1179  /* Set terminate flag in shared memory, unless slot has been reused. */
1180  LWLockAcquire(BackgroundWorkerLock, LW_EXCLUSIVE);
1181  if (handle->generation == slot->generation)
1182  {
1183  slot->terminate = true;
1184  signal_postmaster = true;
1185  }
1186  LWLockRelease(BackgroundWorkerLock);
1187 
1188  /* Make sure the postmaster notices the change to shared memory. */
1189  if (signal_postmaster)
1191 }
1192 
1193 /*
1194  * Look up (and possibly load) a bgworker entry point function.
1195  *
1196  * For functions contained in the core code, we use library name "postgres"
1197  * and consult the InternalBGWorkers array. External functions are
1198  * looked up, and loaded if necessary, using load_external_function().
1199  *
1200  * The point of this is to pass function names as strings across process
1201  * boundaries. We can't pass actual function addresses because of the
1202  * possibility that the function has been loaded at a different address
1203  * in a different process. This is obviously a hazard for functions in
1204  * loadable libraries, but it can happen even for functions in the core code
1205  * on platforms using EXEC_BACKEND (e.g., Windows).
1206  *
1207  * At some point it might be worthwhile to get rid of InternalBGWorkers[]
1208  * in favor of applying load_external_function() for core functions too;
1209  * but that raises portability issues that are not worth addressing now.
1210  */
1211 static bgworker_main_type
1212 LookupBackgroundWorkerFunction(const char *libraryname, const char *funcname)
1213 {
1214  /*
1215  * If the function is to be loaded from postgres itself, search the
1216  * InternalBGWorkers array.
1217  */
1218  if (strcmp(libraryname, "postgres") == 0)
1219  {
1220  int i;
1221 
1222  for (i = 0; i < lengthof(InternalBGWorkers); i++)
1223  {
1224  if (strcmp(InternalBGWorkers[i].fn_name, funcname) == 0)
1225  return InternalBGWorkers[i].fn_addr;
1226  }
1227 
1228  /* We can only reach this by programming error. */
1229  elog(ERROR, "internal function \"%s\" not found", funcname);
1230  }
1231 
1232  /* Otherwise load from external library. */
1233  return (bgworker_main_type)
1234  load_external_function(libraryname, funcname, true, NULL);
1235 }
char bgw_extra[BGW_EXTRALEN]
Definition: bgworker.h:97
void InitializeTimeouts(void)
Definition: timeout.c:340
static void bgworker_sigusr1_handler(SIGNAL_ARGS)
Definition: bgworker.c:684
struct bkend * rw_backend
int max_parallel_workers
Definition: globals.c:125
BgwHandleStatus WaitForBackgroundWorkerStartup(BackgroundWorkerHandle *handle, pid_t *pidp)
Definition: bgworker.c:1090
#define SIGUSR1
Definition: win32.h:202
void RegisterBackgroundWorker(BackgroundWorker *worker)
Definition: bgworker.c:849
#define DEBUG1
Definition: elog.h:25
int errhint(const char *fmt,...)
Definition: elog.c:987
void BackgroundWorkerStateChange(void)
Definition: bgworker.c:241
void ApplyLauncherMain(Datum main_arg)
Definition: launcher.c:782
bool process_shared_preload_libraries_in_progress
Definition: miscinit.c:1417
slist_node * cur
Definition: ilist.h:226
PGPROC * MyProc
Definition: proc.c:67
static BackgroundWorkerArray * BackgroundWorkerData
Definition: bgworker.c:111
slist_node * cur
Definition: ilist.h:241
static void slist_push_head(slist_head *head, slist_node *node)
Definition: ilist.h:574
void ReportBackgroundWorkerExit(slist_mutable_iter *cur)
Definition: bgworker.c:457
int bgw_restart_time
Definition: bgworker.h:93
struct cursor * cur
Definition: ecpg.c:28
void proc_exit(int code)
Definition: ipc.c:99
int errcode(int sqlerrcode)
Definition: elog.c:575
#define BGWORKER_CLASS_PARALLEL
Definition: bgworker.h:67
#define SetProcessingMode(mode)
Definition: miscadmin.h:371
BackgroundWorker * MyBgworkerEntry
Definition: postmaster.c:189
void BaseInit(void)
Definition: postinit.c:517
int snprintf(char *str, size_t count, const char *fmt,...) pg_attribute_printf(3
void ascii_safe_strlcpy(char *dest, const char *src, size_t destsiz)
Definition: ascii.c:173
#define lengthof(array)
Definition: c.h:562
const char * fn_name
Definition: bgworker.c:119
void ResetLatch(volatile Latch *latch)
Definition: latch.c:450
#define LOG
Definition: elog.h:26
#define BGWORKER_SHMEM_ACCESS
Definition: bgworker.h:52
void BackgroundWorkerStopNotifications(pid_t pid)
Definition: bgworker.c:491
void ResetBackgroundWorkerCrashTimes(void)
Definition: bgworker.c:513
char bgw_function_name[BGW_MAXLEN]
Definition: bgworker.h:95
#define SIGQUIT
Definition: win32.h:189
Size BackgroundWorkerShmemSize(void)
Definition: bgworker.c:143
void PGSharedMemoryDetach(void)
Definition: sysv_shmem.c:789
#define PG_SETMASK(mask)
Definition: pqsignal.h:19
void StatementCancelHandler(SIGNAL_ARGS)
Definition: postgres.c:2654
Latch procLatch
Definition: proc.h:103
bool IsBackgroundWorker
Definition: globals.c:102
void StartBackgroundWorker(void)
Definition: bgworker.c:700
#define slist_foreach_modify(iter, lhead)
Definition: ilist.h:716
#define malloc(a)
Definition: header.h:50
void LWLockRelease(LWLock *lock)
Definition: lwlock.c:1715
ErrorContextCallback * error_context_stack
Definition: elog.c:88
static const struct @23 InternalBGWorkers[]
void pg_usleep(long microsec)
Definition: signal.c:53
Datum bgw_main_arg
Definition: bgworker.h:96
int WaitLatch(volatile Latch *latch, int wakeEvents, long timeout, uint32 wait_event_info)
Definition: latch.c:288
#define SIG_IGN
Definition: win32.h:185
void BackgroundWorkerShmemInit(void)
Definition: bgworker.c:159
void ReportBackgroundWorkerPID(RegisteredBgWorker *rw)
Definition: bgworker.c:438
#define ERROR
Definition: elog.h:43
BgwHandleStatus WaitForBackgroundWorkerShutdown(BackgroundWorkerHandle *handle)
Definition: bgworker.c:1132
#define SLIST_STATIC_INIT(name)
Definition: ilist.h:249
struct BackgroundWorkerSlot BackgroundWorkerSlot
void on_exit_reset(void)
Definition: ipc.c:396
void * ShmemInitStruct(const char *name, Size size, bool *foundPtr)
Definition: shmem.c:372
#define FATAL
Definition: elog.h:52
void(* bgworker_main_type)(Datum main_arg)
Definition: bgworker.h:71
#define MAXPGPATH
bool PostmasterMarkPIDForWorkerNotify(int pid)
Definition: postmaster.c:5850
slist_head BackgroundWorkerList
Definition: bgworker.c:43
void InitProcess(void)
Definition: proc.c:287
static bool success
Definition: pg_basebackup.c:96
uint32 parallel_register_count
Definition: bgworker.c:100
static char * buf
Definition: pg_test_fsync.c:66
bool IsUnderPostmaster
Definition: globals.c:100
void ForgetBackgroundWorker(slist_mutable_iter *cur)
Definition: bgworker.c:410
#define BGW_NEVER_RESTART
Definition: bgworker.h:84
unsigned int uint32
Definition: c.h:268
BgwHandleStatus
Definition: bgworker.h:101
#define USECS_PER_DAY
Definition: timestamp.h:91
PGFunction load_external_function(const char *filename, const char *funcname, bool signalNotFound, void **filehandle)
Definition: dfmgr.c:94
#define ereport(elevel, rest)
Definition: elog.h:122
static bgworker_main_type LookupBackgroundWorkerFunction(const char *libraryname, const char *funcname)
Definition: bgworker.c:1212
static bool SanityCheckBackgroundWorker(BackgroundWorker *worker, int elevel)
Definition: bgworker.c:581
static int elevel
Definition: vacuumlazy.c:137
Size mul_size(Size s1, Size s2)
Definition: shmem.c:492
#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:475
int PostAuthDelay
Definition: postgres.c:97
void EmitErrorReport(void)
Definition: elog.c:1446
static RegisteredBgWorker * FindRegisteredWorkerBySlotNumber(int slotno)
Definition: bgworker.c:218
#define SIGPIPE
Definition: win32.h:193
#define SIGHUP
Definition: win32.h:188
void TerminateBackgroundWorker(BackgroundWorkerHandle *handle)
Definition: bgworker.c:1171
#define SIG_DFL
Definition: win32.h:183
pqsigfunc pqsignal(int signum, pqsigfunc handler)
Definition: signal.c:168
#define free(a)
Definition: header.h:65
#define pg_memory_barrier()
Definition: atomics.h:148
#define SIGNAL_ARGS
Definition: c.h:1079
#define NULL
Definition: c.h:229
char bgw_name[BGW_MAXLEN]
Definition: bgworker.h:90
#define Assert(condition)
Definition: c.h:675
BackgroundWorker worker
Definition: bgworker.c:83
#define BGWORKER_BACKEND_DATABASE_CONNECTION
Definition: bgworker.h:59
#define pg_read_barrier()
Definition: atomics.h:161
#define BGW_MAXLEN
Definition: bgworker.h:85
size_t Size
Definition: c.h:356
BgWorkerStartTime bgw_start_time
Definition: bgworker.h:92
int errdetail_plural(const char *fmt_singular, const char *fmt_plural, unsigned long n,...)
Definition: elog.c:965
bool RegisterDynamicBackgroundWorker(BackgroundWorker *worker, BackgroundWorkerHandle **handle)
Definition: bgworker.c:933
bool LWLockAcquire(LWLock *lock, LWLockMode mode)
Definition: lwlock.c:1111
void dsm_detach_all(void)
Definition: dsm.c:666
sigjmp_buf * PG_exception_stack
Definition: elog.c:90
#define slist_foreach(iter, lhead)
Definition: ilist.h:700
bgworker_main_type fn_addr
Definition: bgworker.c:120
void * palloc(Size size)
Definition: mcxt.c:849
int errmsg(const char *fmt,...)
Definition: elog.c:797
void ParallelWorkerMain(Datum main_arg)
Definition: parallel.c:922
uint32 parallel_terminate_count
Definition: bgworker.c:101
pid_t bgw_notify_pid
Definition: bgworker.h:98
#define HOLD_INTERRUPTS()
Definition: miscadmin.h:114
int i
void FloatExceptionHandler(SIGNAL_ARGS)
Definition: postgres.c:2675
#define pg_write_barrier()
Definition: atomics.h:162
int max_worker_processes
Definition: globals.c:124
struct Latch * MyLatch
Definition: globals.c:51
static void bgworker_die(SIGNAL_ARGS)
Definition: bgworker.c:667
#define CHECK_FOR_INTERRUPTS()
Definition: miscadmin.h:97
static void slist_delete_current(slist_mutable_iter *iter)
Definition: ilist.h:652
void SendPostmasterSignal(PMSignalReason reason)
Definition: pmsignal.c:113
#define elog
Definition: elog.h:219
static void bgworker_quickdie(SIGNAL_ARGS)
Definition: bgworker.c:637
void latch_sigusr1_handler(void)
Definition: latch.c:1404
#define MAX_PARALLEL_WORKER_LIMIT
void procsignal_sigusr1_handler(SIGNAL_ARGS)
Definition: procsignal.c:260
static void static void status(const char *fmt,...) pg_attribute_printf(1
Definition: pg_regress.c:224
BackgroundWorker rw_worker
#define SIGCHLD
Definition: win32.h:198
BackgroundWorkerSlot slot[FLEXIBLE_ARRAY_MEMBER]
Definition: bgworker.c:102
char bgw_library_name[BGW_MAXLEN]
Definition: bgworker.h:94
#define BGW_EXTRALEN
Definition: bgworker.h:86
#define WL_LATCH_SET
Definition: latch.h:124
#define SIGUSR2
Definition: win32.h:203
#define offsetof(type, field)
Definition: c.h:555
BgwHandleStatus GetBackgroundWorkerPid(BackgroundWorkerHandle *handle, pid_t *pidp)
Definition: bgworker.c:1041
void ApplyWorkerMain(Datum main_arg)
Definition: worker.c:1442
void init_ps_display(const char *username, const char *dbname, const char *host_info, const char *initial_str)
Definition: ps_status.c:244
#define InvalidPid
Definition: miscadmin.h:31