Line data Source code
1 : /*--------------------------------------------------------------------
2 : * bgworker.c
3 : * POSTGRES pluggable background workers implementation
4 : *
5 : * Portions Copyright (c) 1996-2026, 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"
20 : #include "postmaster/bgworker_internals.h"
21 : #include "postmaster/postmaster.h"
22 : #include "replication/logicallauncher.h"
23 : #include "replication/logicalworker.h"
24 : #include "storage/ipc.h"
25 : #include "storage/latch.h"
26 : #include "storage/lwlock.h"
27 : #include "storage/pmsignal.h"
28 : #include "storage/proc.h"
29 : #include "storage/procarray.h"
30 : #include "storage/procsignal.h"
31 : #include "storage/shmem.h"
32 : #include "tcop/tcopprot.h"
33 : #include "utils/ascii.h"
34 : #include "utils/memutils.h"
35 : #include "utils/ps_status.h"
36 : #include "utils/timeout.h"
37 : #include "utils/wait_event.h"
38 :
39 : /*
40 : * The postmaster's list of registered background workers, in private memory.
41 : */
42 : dlist_head BackgroundWorkerList = DLIST_STATIC_INIT(BackgroundWorkerList);
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 */
82 : BackgroundWorker worker;
83 : } BackgroundWorkerSlot;
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 : {
98 : int total_slots;
99 : uint32 parallel_register_count;
100 : uint32 parallel_terminate_count;
101 : BackgroundWorkerSlot slot[FLEXIBLE_ARRAY_MEMBER];
102 : } BackgroundWorkerArray;
103 :
104 : struct BackgroundWorkerHandle
105 : {
106 : int slot;
107 : uint64 generation;
108 : };
109 :
110 : static BackgroundWorkerArray *BackgroundWorkerData;
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;
119 : bgworker_main_type fn_addr;
120 : } InternalBGWorkers[] =
121 :
122 : {
123 : {
124 : .fn_name = "ParallelWorkerMain",
125 : .fn_addr = ParallelWorkerMain
126 : },
127 : {
128 : .fn_name = "ApplyLauncherMain",
129 : .fn_addr = ApplyLauncherMain
130 : },
131 : {
132 : .fn_name = "ApplyWorkerMain",
133 : .fn_addr = ApplyWorkerMain
134 : },
135 : {
136 : .fn_name = "ParallelApplyWorkerMain",
137 : .fn_addr = ParallelApplyWorkerMain
138 : },
139 : {
140 : .fn_name = "TableSyncWorkerMain",
141 : .fn_addr = TableSyncWorkerMain
142 : },
143 : {
144 : .fn_name = "SequenceSyncWorkerMain",
145 : .fn_addr = SequenceSyncWorkerMain
146 : }
147 : };
148 :
149 : /* Private functions. */
150 : static bgworker_main_type LookupBackgroundWorkerFunction(const char *libraryname, const char *funcname);
151 :
152 :
153 : /*
154 : * Calculate shared memory needed.
155 : */
156 : Size
157 3324 : BackgroundWorkerShmemSize(void)
158 : {
159 : Size size;
160 :
161 : /* Array of workers is variably sized. */
162 3324 : size = offsetof(BackgroundWorkerArray, slot);
163 3324 : size = add_size(size, mul_size(max_worker_processes,
164 : sizeof(BackgroundWorkerSlot)));
165 :
166 3324 : return size;
167 : }
168 :
169 : /*
170 : * Initialize shared memory.
171 : */
172 : void
173 1159 : BackgroundWorkerShmemInit(void)
174 : {
175 : bool found;
176 :
177 1159 : BackgroundWorkerData = ShmemInitStruct("Background Worker Data",
178 : BackgroundWorkerShmemSize(),
179 : &found);
180 1159 : if (!IsUnderPostmaster)
181 : {
182 : dlist_iter iter;
183 1159 : int slotno = 0;
184 :
185 1159 : BackgroundWorkerData->total_slots = max_worker_processes;
186 1159 : BackgroundWorkerData->parallel_register_count = 0;
187 1159 : BackgroundWorkerData->parallel_terminate_count = 0;
188 :
189 : /*
190 : * Copy contents of worker list into shared memory. Record the shared
191 : * memory slot assigned to each worker. This ensures a 1-to-1
192 : * correspondence between the postmaster's private list and the array
193 : * in shared memory.
194 : */
195 2043 : dlist_foreach(iter, &BackgroundWorkerList)
196 : {
197 884 : BackgroundWorkerSlot *slot = &BackgroundWorkerData->slot[slotno];
198 : RegisteredBgWorker *rw;
199 :
200 884 : rw = dlist_container(RegisteredBgWorker, rw_lnode, iter.cur);
201 : Assert(slotno < max_worker_processes);
202 884 : slot->in_use = true;
203 884 : slot->terminate = false;
204 884 : slot->pid = InvalidPid;
205 884 : slot->generation = 0;
206 884 : rw->rw_shmem_slot = slotno;
207 884 : rw->rw_worker.bgw_notify_pid = 0; /* might be reinit after crash */
208 884 : memcpy(&slot->worker, &rw->rw_worker, sizeof(BackgroundWorker));
209 884 : ++slotno;
210 : }
211 :
212 : /*
213 : * Mark any remaining slots as not in use.
214 : */
215 9538 : while (slotno < max_worker_processes)
216 : {
217 8379 : BackgroundWorkerSlot *slot = &BackgroundWorkerData->slot[slotno];
218 :
219 8379 : slot->in_use = false;
220 8379 : ++slotno;
221 : }
222 : }
223 : else
224 : Assert(found);
225 1159 : }
226 :
227 : /*
228 : * Search the postmaster's backend-private list of RegisteredBgWorker objects
229 : * for the one that maps to the given slot number.
230 : */
231 : static RegisteredBgWorker *
232 4521 : FindRegisteredWorkerBySlotNumber(int slotno)
233 : {
234 : dlist_iter iter;
235 :
236 10985 : dlist_foreach(iter, &BackgroundWorkerList)
237 : {
238 : RegisteredBgWorker *rw;
239 :
240 8874 : rw = dlist_container(RegisteredBgWorker, rw_lnode, iter.cur);
241 8874 : if (rw->rw_shmem_slot == slotno)
242 2410 : return rw;
243 : }
244 :
245 2111 : return NULL;
246 : }
247 :
248 : /*
249 : * Notice changes to shared memory made by other backends.
250 : * Accept new worker requests only if allow_new_workers is true.
251 : *
252 : * This code runs in the postmaster, so we must be very careful not to assume
253 : * that shared memory contents are sane. Otherwise, a rogue backend could
254 : * take out the postmaster.
255 : */
256 : void
257 1374 : BackgroundWorkerStateChange(bool allow_new_workers)
258 : {
259 : int slotno;
260 :
261 : /*
262 : * The total number of slots stored in shared memory should match our
263 : * notion of max_worker_processes. If it does not, something is very
264 : * wrong. Further down, we always refer to this value as
265 : * max_worker_processes, in case shared memory gets corrupted while we're
266 : * looping.
267 : */
268 1374 : if (max_worker_processes != BackgroundWorkerData->total_slots)
269 : {
270 0 : ereport(LOG,
271 : (errmsg("inconsistent background worker state (\"max_worker_processes\"=%d, total slots=%d)",
272 : max_worker_processes,
273 : BackgroundWorkerData->total_slots)));
274 0 : return;
275 : }
276 :
277 : /*
278 : * Iterate through slots, looking for newly-registered workers or workers
279 : * who must die.
280 : */
281 12510 : for (slotno = 0; slotno < max_worker_processes; ++slotno)
282 : {
283 11136 : BackgroundWorkerSlot *slot = &BackgroundWorkerData->slot[slotno];
284 : RegisteredBgWorker *rw;
285 :
286 11136 : if (!slot->in_use)
287 6615 : continue;
288 :
289 : /*
290 : * Make sure we don't see the in_use flag before the updated slot
291 : * contents.
292 : */
293 4521 : pg_read_barrier();
294 :
295 : /* See whether we already know about this worker. */
296 4521 : rw = FindRegisteredWorkerBySlotNumber(slotno);
297 4521 : if (rw != NULL)
298 : {
299 : /*
300 : * In general, the worker data can't change after it's initially
301 : * registered. However, someone can set the terminate flag.
302 : */
303 2410 : if (slot->terminate && !rw->rw_terminate)
304 : {
305 10 : rw->rw_terminate = true;
306 10 : if (rw->rw_pid != 0)
307 10 : kill(rw->rw_pid, SIGTERM);
308 : else
309 : {
310 : /* Report never-started, now-terminated worker as dead. */
311 0 : ReportBackgroundWorkerPID(rw);
312 : }
313 : }
314 2410 : continue;
315 : }
316 :
317 : /*
318 : * If we aren't allowing new workers, then immediately mark it for
319 : * termination; the next stanza will take care of cleaning it up.
320 : * Doing this ensures that any process waiting for the worker will get
321 : * awoken, even though the worker will never be allowed to run.
322 : */
323 2111 : if (!allow_new_workers)
324 2 : slot->terminate = true;
325 :
326 : /*
327 : * If the worker is marked for termination, we don't need to add it to
328 : * the registered workers list; we can just free the slot. However, if
329 : * bgw_notify_pid is set, the process that registered the worker may
330 : * need to know that we've processed the terminate request, so be sure
331 : * to signal it.
332 : */
333 2111 : if (slot->terminate)
334 2 : {
335 : int notify_pid;
336 :
337 : /*
338 : * We need a memory barrier here to make sure that the load of
339 : * bgw_notify_pid and the update of parallel_terminate_count
340 : * complete before the store to in_use.
341 : */
342 2 : notify_pid = slot->worker.bgw_notify_pid;
343 2 : if ((slot->worker.bgw_flags & BGWORKER_CLASS_PARALLEL) != 0)
344 0 : BackgroundWorkerData->parallel_terminate_count++;
345 2 : slot->pid = 0;
346 :
347 2 : pg_memory_barrier();
348 2 : slot->in_use = false;
349 :
350 2 : if (notify_pid != 0)
351 2 : kill(notify_pid, SIGUSR1);
352 :
353 2 : continue;
354 : }
355 :
356 : /*
357 : * Copy the registration data into the registered workers list.
358 : */
359 2109 : rw = MemoryContextAllocExtended(PostmasterContext,
360 : sizeof(RegisteredBgWorker),
361 : MCXT_ALLOC_NO_OOM | MCXT_ALLOC_ZERO);
362 2109 : if (rw == NULL)
363 : {
364 0 : ereport(LOG,
365 : (errcode(ERRCODE_OUT_OF_MEMORY),
366 : errmsg("out of memory")));
367 0 : return;
368 : }
369 :
370 : /*
371 : * Copy strings in a paranoid way. If shared memory is corrupted, the
372 : * source data might not even be NUL-terminated.
373 : */
374 2109 : ascii_safe_strlcpy(rw->rw_worker.bgw_name,
375 2109 : slot->worker.bgw_name, BGW_MAXLEN);
376 2109 : ascii_safe_strlcpy(rw->rw_worker.bgw_type,
377 2109 : slot->worker.bgw_type, BGW_MAXLEN);
378 2109 : ascii_safe_strlcpy(rw->rw_worker.bgw_library_name,
379 2109 : slot->worker.bgw_library_name, MAXPGPATH);
380 2109 : ascii_safe_strlcpy(rw->rw_worker.bgw_function_name,
381 2109 : slot->worker.bgw_function_name, BGW_MAXLEN);
382 :
383 : /*
384 : * Copy various fixed-size fields.
385 : *
386 : * flags, start_time, and restart_time are examined by the postmaster,
387 : * but nothing too bad will happen if they are corrupted. The
388 : * remaining fields will only be examined by the child process. It
389 : * might crash, but we won't.
390 : */
391 2109 : rw->rw_worker.bgw_flags = slot->worker.bgw_flags;
392 2109 : rw->rw_worker.bgw_start_time = slot->worker.bgw_start_time;
393 2109 : rw->rw_worker.bgw_restart_time = slot->worker.bgw_restart_time;
394 2109 : rw->rw_worker.bgw_main_arg = slot->worker.bgw_main_arg;
395 2109 : memcpy(rw->rw_worker.bgw_extra, slot->worker.bgw_extra, BGW_EXTRALEN);
396 :
397 : /*
398 : * Copy the PID to be notified about state changes, but only if the
399 : * postmaster knows about a backend with that PID. It isn't an error
400 : * if the postmaster doesn't know about the PID, because the backend
401 : * that requested the worker could have died (or been killed) just
402 : * after doing so. Nonetheless, at least until we get some experience
403 : * with how this plays out in the wild, log a message at a relative
404 : * high debug level.
405 : */
406 2109 : rw->rw_worker.bgw_notify_pid = slot->worker.bgw_notify_pid;
407 2109 : if (!PostmasterMarkPIDForWorkerNotify(rw->rw_worker.bgw_notify_pid))
408 : {
409 0 : elog(DEBUG1, "worker notification PID %d is not valid",
410 : (int) rw->rw_worker.bgw_notify_pid);
411 0 : rw->rw_worker.bgw_notify_pid = 0;
412 : }
413 :
414 : /* Initialize postmaster bookkeeping. */
415 2109 : rw->rw_pid = 0;
416 2109 : rw->rw_crashed_at = 0;
417 2109 : rw->rw_shmem_slot = slotno;
418 2109 : rw->rw_terminate = false;
419 :
420 : /* Log it! */
421 2109 : ereport(DEBUG1,
422 : (errmsg_internal("registering background worker \"%s\"",
423 : rw->rw_worker.bgw_name)));
424 :
425 2109 : dlist_push_head(&BackgroundWorkerList, &rw->rw_lnode);
426 : }
427 : }
428 :
429 : /*
430 : * Forget about a background worker that's no longer needed.
431 : *
432 : * NOTE: The entry is unlinked from BackgroundWorkerList. If the caller is
433 : * iterating through it, better use a mutable iterator!
434 : *
435 : * Caller is responsible for notifying bgw_notify_pid, if appropriate.
436 : *
437 : * This function must be invoked only in the postmaster.
438 : */
439 : void
440 2090 : ForgetBackgroundWorker(RegisteredBgWorker *rw)
441 : {
442 : BackgroundWorkerSlot *slot;
443 :
444 : Assert(rw->rw_shmem_slot < max_worker_processes);
445 2090 : slot = &BackgroundWorkerData->slot[rw->rw_shmem_slot];
446 : Assert(slot->in_use);
447 :
448 : /*
449 : * We need a memory barrier here to make sure that the update of
450 : * parallel_terminate_count completes before the store to in_use.
451 : */
452 2090 : if ((rw->rw_worker.bgw_flags & BGWORKER_CLASS_PARALLEL) != 0)
453 1489 : BackgroundWorkerData->parallel_terminate_count++;
454 :
455 2090 : pg_memory_barrier();
456 2090 : slot->in_use = false;
457 :
458 2090 : ereport(DEBUG1,
459 : (errmsg_internal("unregistering background worker \"%s\"",
460 : rw->rw_worker.bgw_name)));
461 :
462 2090 : dlist_delete(&rw->rw_lnode);
463 2090 : pfree(rw);
464 2090 : }
465 :
466 : /*
467 : * Report the PID of a newly-launched background worker in shared memory.
468 : *
469 : * This function should only be called from the postmaster.
470 : */
471 : void
472 2884 : ReportBackgroundWorkerPID(RegisteredBgWorker *rw)
473 : {
474 : BackgroundWorkerSlot *slot;
475 :
476 : Assert(rw->rw_shmem_slot < max_worker_processes);
477 2884 : slot = &BackgroundWorkerData->slot[rw->rw_shmem_slot];
478 2884 : slot->pid = rw->rw_pid;
479 :
480 2884 : if (rw->rw_worker.bgw_notify_pid != 0)
481 2109 : kill(rw->rw_worker.bgw_notify_pid, SIGUSR1);
482 2884 : }
483 :
484 : /*
485 : * Report that the PID of a background worker is now zero because a
486 : * previously-running background worker has exited.
487 : *
488 : * NOTE: The entry may be unlinked from BackgroundWorkerList. If the caller
489 : * is iterating through it, better use a mutable iterator!
490 : *
491 : * This function should only be called from the postmaster.
492 : */
493 : void
494 2557 : ReportBackgroundWorkerExit(RegisteredBgWorker *rw)
495 : {
496 : BackgroundWorkerSlot *slot;
497 : int notify_pid;
498 :
499 : Assert(rw->rw_shmem_slot < max_worker_processes);
500 2557 : slot = &BackgroundWorkerData->slot[rw->rw_shmem_slot];
501 2557 : slot->pid = rw->rw_pid;
502 2557 : notify_pid = rw->rw_worker.bgw_notify_pid;
503 :
504 : /*
505 : * If this worker is slated for deregistration, do that before notifying
506 : * the process which started it. Otherwise, if that process tries to
507 : * reuse the slot immediately, it might not be available yet. In theory
508 : * that could happen anyway if the process checks slot->pid at just the
509 : * wrong moment, but this makes the window narrower.
510 : */
511 2557 : if (rw->rw_terminate ||
512 737 : rw->rw_worker.bgw_restart_time == BGW_NEVER_RESTART)
513 2090 : ForgetBackgroundWorker(rw);
514 :
515 2557 : if (notify_pid != 0)
516 2060 : kill(notify_pid, SIGUSR1);
517 2557 : }
518 :
519 : /*
520 : * Cancel SIGUSR1 notifications for a PID belonging to an exiting backend.
521 : *
522 : * This function should only be called from the postmaster.
523 : */
524 : void
525 280 : BackgroundWorkerStopNotifications(pid_t pid)
526 : {
527 : dlist_iter iter;
528 :
529 901 : dlist_foreach(iter, &BackgroundWorkerList)
530 : {
531 : RegisteredBgWorker *rw;
532 :
533 621 : rw = dlist_container(RegisteredBgWorker, rw_lnode, iter.cur);
534 621 : if (rw->rw_worker.bgw_notify_pid == pid)
535 32 : rw->rw_worker.bgw_notify_pid = 0;
536 : }
537 280 : }
538 :
539 : /*
540 : * Cancel any not-yet-started worker requests that have waiting processes.
541 : *
542 : * This is called during a normal ("smart" or "fast") database shutdown.
543 : * After this point, no new background workers will be started, so anything
544 : * that might be waiting for them needs to be kicked off its wait. We do
545 : * that by canceling the bgworker registration entirely, which is perhaps
546 : * overkill, but since we're shutting down it does not matter whether the
547 : * registration record sticks around.
548 : *
549 : * This function should only be called from the postmaster.
550 : */
551 : void
552 580 : ForgetUnstartedBackgroundWorkers(void)
553 : {
554 : dlist_mutable_iter iter;
555 :
556 1148 : dlist_foreach_modify(iter, &BackgroundWorkerList)
557 : {
558 : RegisteredBgWorker *rw;
559 : BackgroundWorkerSlot *slot;
560 :
561 568 : rw = dlist_container(RegisteredBgWorker, rw_lnode, iter.cur);
562 : Assert(rw->rw_shmem_slot < max_worker_processes);
563 568 : slot = &BackgroundWorkerData->slot[rw->rw_shmem_slot];
564 :
565 : /* If it's not yet started, and there's someone waiting ... */
566 568 : if (slot->pid == InvalidPid &&
567 56 : rw->rw_worker.bgw_notify_pid != 0)
568 : {
569 : /* ... then zap it, and notify the waiter */
570 0 : int notify_pid = rw->rw_worker.bgw_notify_pid;
571 :
572 0 : ForgetBackgroundWorker(rw);
573 0 : if (notify_pid != 0)
574 0 : kill(notify_pid, SIGUSR1);
575 : }
576 : }
577 580 : }
578 :
579 : /*
580 : * Reset background worker crash state.
581 : *
582 : * We assume that, after a crash-and-restart cycle, background workers without
583 : * the never-restart flag should be restarted immediately, instead of waiting
584 : * for bgw_restart_time to elapse. On the other hand, workers with that flag
585 : * should be forgotten immediately, since we won't ever restart them.
586 : *
587 : * This function should only be called from the postmaster.
588 : */
589 : void
590 5 : ResetBackgroundWorkerCrashTimes(void)
591 : {
592 : dlist_mutable_iter iter;
593 :
594 10 : dlist_foreach_modify(iter, &BackgroundWorkerList)
595 : {
596 : RegisteredBgWorker *rw;
597 :
598 5 : rw = dlist_container(RegisteredBgWorker, rw_lnode, iter.cur);
599 :
600 5 : if (rw->rw_worker.bgw_restart_time == BGW_NEVER_RESTART)
601 : {
602 : /*
603 : * Workers marked BGW_NEVER_RESTART shouldn't get relaunched after
604 : * the crash, so forget about them. (If we wait until after the
605 : * crash to forget about them, and they are parallel workers,
606 : * parallel_terminate_count will get incremented after we've
607 : * already zeroed parallel_register_count, which would be bad.)
608 : */
609 0 : ForgetBackgroundWorker(rw);
610 : }
611 : else
612 : {
613 : /*
614 : * The accounting which we do via parallel_register_count and
615 : * parallel_terminate_count would get messed up if a worker marked
616 : * parallel could survive a crash and restart cycle. All such
617 : * workers should be marked BGW_NEVER_RESTART, and thus control
618 : * should never reach this branch.
619 : */
620 : Assert((rw->rw_worker.bgw_flags & BGWORKER_CLASS_PARALLEL) == 0);
621 :
622 : /*
623 : * Allow this worker to be restarted immediately after we finish
624 : * resetting.
625 : */
626 5 : rw->rw_crashed_at = 0;
627 5 : rw->rw_pid = 0;
628 :
629 : /*
630 : * If there was anyone waiting for it, they're history.
631 : */
632 5 : rw->rw_worker.bgw_notify_pid = 0;
633 : }
634 : }
635 5 : }
636 :
637 : /*
638 : * Complain about the BackgroundWorker definition using error level elevel.
639 : * Return true if it looks ok, false if not (unless elevel >= ERROR, in
640 : * which case we won't return at all in the not-OK case).
641 : */
642 : static bool
643 2859 : SanityCheckBackgroundWorker(BackgroundWorker *worker, int elevel)
644 : {
645 : /* sanity check for flags */
646 :
647 : /*
648 : * We used to support workers not connected to shared memory, but don't
649 : * anymore. Thus this is a required flag now. We're not removing the flag
650 : * for compatibility reasons and because the flag still provides some
651 : * signal when reading code.
652 : */
653 2859 : if (!(worker->bgw_flags & BGWORKER_SHMEM_ACCESS))
654 : {
655 0 : ereport(elevel,
656 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
657 : errmsg("background worker \"%s\": background workers without shared memory access are not supported",
658 : worker->bgw_name)));
659 0 : return false;
660 : }
661 :
662 2859 : if (worker->bgw_flags & BGWORKER_BACKEND_DATABASE_CONNECTION)
663 : {
664 2850 : if (worker->bgw_start_time == BgWorkerStart_PostmasterStart)
665 : {
666 0 : ereport(elevel,
667 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
668 : errmsg("background worker \"%s\": cannot request database access if starting at postmaster start",
669 : worker->bgw_name)));
670 0 : return false;
671 : }
672 :
673 : /* XXX other checks? */
674 : }
675 :
676 : /* Interruptible workers require a database connection */
677 2859 : if ((worker->bgw_flags & BGWORKER_INTERRUPTIBLE) &&
678 4 : !(worker->bgw_flags & BGWORKER_BACKEND_DATABASE_CONNECTION))
679 : {
680 0 : ereport(elevel,
681 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
682 : errmsg("background worker \"%s\": cannot make background workers interruptible without database access",
683 : worker->bgw_name)));
684 0 : return false;
685 : }
686 :
687 2859 : if ((worker->bgw_restart_time < 0 &&
688 1976 : worker->bgw_restart_time != BGW_NEVER_RESTART) ||
689 2859 : (worker->bgw_restart_time > USECS_PER_DAY / 1000))
690 : {
691 0 : ereport(elevel,
692 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
693 : errmsg("background worker \"%s\": invalid restart interval",
694 : worker->bgw_name)));
695 0 : return false;
696 : }
697 :
698 : /*
699 : * Parallel workers may not be configured for restart, because the
700 : * parallel_register_count/parallel_terminate_count accounting can't
701 : * handle parallel workers lasting through a crash-and-restart cycle.
702 : */
703 2859 : if (worker->bgw_restart_time != BGW_NEVER_RESTART &&
704 883 : (worker->bgw_flags & BGWORKER_CLASS_PARALLEL) != 0)
705 : {
706 0 : ereport(elevel,
707 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
708 : errmsg("background worker \"%s\": parallel workers may not be configured for restart",
709 : worker->bgw_name)));
710 0 : return false;
711 : }
712 :
713 : /*
714 : * If bgw_type is not filled in, use bgw_name.
715 : */
716 2859 : if (strcmp(worker->bgw_type, "") == 0)
717 0 : strcpy(worker->bgw_type, worker->bgw_name);
718 :
719 2859 : return true;
720 : }
721 :
722 : /*
723 : * Main entry point for background worker processes.
724 : */
725 : void
726 2565 : BackgroundWorkerMain(const void *startup_data, size_t startup_data_len)
727 : {
728 : sigjmp_buf local_sigjmp_buf;
729 : BackgroundWorker *worker;
730 : bgworker_main_type entrypt;
731 :
732 2565 : if (startup_data == NULL)
733 0 : elog(FATAL, "unable to find bgworker entry");
734 : Assert(startup_data_len == sizeof(BackgroundWorker));
735 2565 : worker = MemoryContextAlloc(TopMemoryContext, sizeof(BackgroundWorker));
736 2565 : memcpy(worker, startup_data, sizeof(BackgroundWorker));
737 :
738 : /*
739 : * Now that we're done reading the startup data, release postmaster's
740 : * working memory context.
741 : */
742 2565 : if (PostmasterContext)
743 : {
744 2565 : MemoryContextDelete(PostmasterContext);
745 2565 : PostmasterContext = NULL;
746 : }
747 :
748 2565 : MyBgworkerEntry = worker;
749 2565 : init_ps_display(worker->bgw_name);
750 :
751 : Assert(GetProcessingMode() == InitProcessing);
752 :
753 : /* Apply PostAuthDelay */
754 2565 : if (PostAuthDelay > 0)
755 0 : pg_usleep(PostAuthDelay * 1000000L);
756 :
757 : /*
758 : * Set up signal handlers.
759 : */
760 2565 : if (worker->bgw_flags & BGWORKER_BACKEND_DATABASE_CONNECTION)
761 : {
762 : /*
763 : * SIGINT is used to signal canceling the current action
764 : */
765 2556 : pqsignal(SIGINT, StatementCancelHandler);
766 2556 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
767 2556 : pqsignal(SIGFPE, FloatExceptionHandler);
768 :
769 : /* XXX Any other handlers needed here? */
770 : }
771 : else
772 : {
773 9 : pqsignal(SIGINT, SIG_IGN);
774 9 : pqsignal(SIGUSR1, SIG_IGN);
775 9 : pqsignal(SIGFPE, SIG_IGN);
776 : }
777 2565 : pqsignal(SIGTERM, die);
778 : /* SIGQUIT handler was already set up by InitPostmasterChild */
779 2565 : pqsignal(SIGHUP, SIG_IGN);
780 :
781 2565 : InitializeTimeouts(); /* establishes SIGALRM handler */
782 :
783 2565 : pqsignal(SIGPIPE, SIG_IGN);
784 2565 : pqsignal(SIGUSR2, SIG_IGN);
785 2565 : pqsignal(SIGCHLD, SIG_DFL);
786 :
787 : /*
788 : * If an exception is encountered, processing resumes here.
789 : *
790 : * We just need to clean up, report the error, and go away.
791 : */
792 2565 : if (sigsetjmp(local_sigjmp_buf, 1) != 0)
793 : {
794 : /* Since not using PG_TRY, must reset error stack by hand */
795 159 : error_context_stack = NULL;
796 :
797 : /* Prevent interrupts while cleaning up */
798 159 : HOLD_INTERRUPTS();
799 :
800 : /*
801 : * sigsetjmp will have blocked all signals, but we may need to accept
802 : * signals while communicating with our parallel leader. Once we've
803 : * done HOLD_INTERRUPTS() it should be safe to unblock signals.
804 : */
805 159 : BackgroundWorkerUnblockSignals();
806 :
807 : /* Report the error to the parallel leader and the server log */
808 159 : EmitErrorReport();
809 :
810 : /*
811 : * Do we need more cleanup here? For shmem-connected bgworkers, we
812 : * will call InitProcess below, which will install ProcKill as exit
813 : * callback. That will take care of releasing locks, etc.
814 : */
815 :
816 : /* and go away */
817 159 : proc_exit(1);
818 : }
819 :
820 : /* We can now handle ereport(ERROR) */
821 2565 : PG_exception_stack = &local_sigjmp_buf;
822 :
823 : /*
824 : * Create a per-backend PGPROC struct in shared memory. We must do this
825 : * before we can use LWLocks or access any shared memory.
826 : */
827 2565 : InitProcess();
828 :
829 : /*
830 : * Early initialization.
831 : */
832 2565 : BaseInit();
833 :
834 : /*
835 : * Look up the entry point function, loading its library if necessary.
836 : */
837 5130 : entrypt = LookupBackgroundWorkerFunction(worker->bgw_library_name,
838 2565 : worker->bgw_function_name);
839 :
840 : /*
841 : * Note that in normal processes, we would call InitPostgres here. For a
842 : * worker, however, we don't know what database to connect to, yet; so we
843 : * need to wait until the user code does it via
844 : * BackgroundWorkerInitializeConnection().
845 : */
846 :
847 : /*
848 : * Now invoke the user-defined worker code
849 : */
850 2565 : entrypt(worker->bgw_main_arg);
851 :
852 : /* ... and if it returns, we're done */
853 1486 : proc_exit(0);
854 : }
855 :
856 : /*
857 : * Connect background worker to a database.
858 : */
859 : void
860 470 : BackgroundWorkerInitializeConnection(const char *dbname, const char *username, uint32 flags)
861 : {
862 470 : BackgroundWorker *worker = MyBgworkerEntry;
863 470 : bits32 init_flags = 0; /* never honor session_preload_libraries */
864 :
865 : /* ignore datallowconn and ACL_CONNECT? */
866 470 : if (flags & BGWORKER_BYPASS_ALLOWCONN)
867 0 : init_flags |= INIT_PG_OVERRIDE_ALLOW_CONNS;
868 : /* ignore rolcanlogin? */
869 470 : if (flags & BGWORKER_BYPASS_ROLELOGINCHECK)
870 0 : init_flags |= INIT_PG_OVERRIDE_ROLE_LOGIN;
871 :
872 : /* XXX is this the right errcode? */
873 470 : if (!(worker->bgw_flags & BGWORKER_BACKEND_DATABASE_CONNECTION))
874 0 : ereport(FATAL,
875 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
876 : errmsg("database connection requirement not indicated during registration")));
877 :
878 470 : InitPostgres(dbname, InvalidOid, /* database to connect to */
879 : username, InvalidOid, /* role to connect as */
880 : init_flags,
881 : NULL); /* no out_dbname */
882 :
883 : /* it had better not gotten out of "init" mode yet */
884 470 : if (!IsInitProcessingMode())
885 0 : ereport(ERROR,
886 : (errmsg("invalid processing mode in background worker")));
887 470 : SetProcessingMode(NormalProcessing);
888 470 : }
889 :
890 : /*
891 : * Connect background worker to a database using OIDs.
892 : */
893 : void
894 2086 : BackgroundWorkerInitializeConnectionByOid(Oid dboid, Oid useroid, uint32 flags)
895 : {
896 2086 : BackgroundWorker *worker = MyBgworkerEntry;
897 2086 : bits32 init_flags = 0; /* never honor session_preload_libraries */
898 :
899 : /* ignore datallowconn and ACL_CONNECT? */
900 2086 : if (flags & BGWORKER_BYPASS_ALLOWCONN)
901 1489 : init_flags |= INIT_PG_OVERRIDE_ALLOW_CONNS;
902 : /* ignore rolcanlogin? */
903 2086 : if (flags & BGWORKER_BYPASS_ROLELOGINCHECK)
904 1489 : init_flags |= INIT_PG_OVERRIDE_ROLE_LOGIN;
905 :
906 : /* XXX is this the right errcode? */
907 2086 : if (!(worker->bgw_flags & BGWORKER_BACKEND_DATABASE_CONNECTION))
908 0 : ereport(FATAL,
909 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
910 : errmsg("database connection requirement not indicated during registration")));
911 :
912 2086 : InitPostgres(NULL, dboid, /* database to connect to */
913 : NULL, useroid, /* role to connect as */
914 : init_flags,
915 : NULL); /* no out_dbname */
916 :
917 : /* it had better not gotten out of "init" mode yet */
918 2082 : if (!IsInitProcessingMode())
919 0 : ereport(ERROR,
920 : (errmsg("invalid processing mode in background worker")));
921 2082 : SetProcessingMode(NormalProcessing);
922 2082 : }
923 :
924 : /*
925 : * Block/unblock signals in a background worker
926 : */
927 : void
928 0 : BackgroundWorkerBlockSignals(void)
929 : {
930 0 : sigprocmask(SIG_SETMASK, &BlockSig, NULL);
931 0 : }
932 :
933 : void
934 2724 : BackgroundWorkerUnblockSignals(void)
935 : {
936 2724 : sigprocmask(SIG_SETMASK, &UnBlockSig, NULL);
937 2724 : }
938 :
939 : /*
940 : * Register a new static background worker.
941 : *
942 : * This can only be called directly from postmaster or in the _PG_init
943 : * function of a module library that's loaded by shared_preload_libraries;
944 : * otherwise it will have no effect.
945 : */
946 : void
947 886 : RegisterBackgroundWorker(BackgroundWorker *worker)
948 : {
949 : RegisteredBgWorker *rw;
950 : static int numworkers = 0;
951 :
952 : /*
953 : * Static background workers can only be registered in the postmaster
954 : * process.
955 : */
956 886 : if (IsUnderPostmaster || !IsPostmasterEnvironment)
957 : {
958 : /*
959 : * In EXEC_BACKEND or single-user mode, we process
960 : * shared_preload_libraries in backend processes too. We cannot
961 : * register static background workers at that stage, but many
962 : * libraries' _PG_init() functions don't distinguish whether they're
963 : * being loaded in the postmaster or in a backend, they just check
964 : * process_shared_preload_libraries_in_progress. It's a bit sloppy,
965 : * but for historical reasons we tolerate it. In EXEC_BACKEND mode,
966 : * the background workers should already have been registered when the
967 : * library was loaded in postmaster.
968 : */
969 0 : if (process_shared_preload_libraries_in_progress)
970 0 : return;
971 0 : ereport(LOG,
972 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
973 : errmsg("background worker \"%s\": must be registered in \"shared_preload_libraries\"",
974 : worker->bgw_name)));
975 0 : return;
976 : }
977 :
978 : /*
979 : * Cannot register static background workers after calling
980 : * BackgroundWorkerShmemInit().
981 : */
982 886 : if (BackgroundWorkerData != NULL)
983 0 : elog(ERROR, "cannot register background worker \"%s\" after shmem init",
984 : worker->bgw_name);
985 :
986 886 : ereport(DEBUG1,
987 : (errmsg_internal("registering background worker \"%s\"", worker->bgw_name)));
988 :
989 886 : if (!SanityCheckBackgroundWorker(worker, LOG))
990 0 : return;
991 :
992 886 : if (worker->bgw_notify_pid != 0)
993 : {
994 0 : ereport(LOG,
995 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
996 : errmsg("background worker \"%s\": only dynamic background workers can request notification",
997 : worker->bgw_name)));
998 0 : return;
999 : }
1000 :
1001 : /*
1002 : * Enforce maximum number of workers. Note this is overly restrictive: we
1003 : * could allow more non-shmem-connected workers, because these don't count
1004 : * towards the MAX_BACKENDS limit elsewhere. For now, it doesn't seem
1005 : * important to relax this restriction.
1006 : */
1007 886 : if (++numworkers > max_worker_processes)
1008 : {
1009 0 : ereport(LOG,
1010 : (errcode(ERRCODE_CONFIGURATION_LIMIT_EXCEEDED),
1011 : errmsg("too many background workers"),
1012 : errdetail_plural("Up to %d background worker can be registered with the current settings.",
1013 : "Up to %d background workers can be registered with the current settings.",
1014 : max_worker_processes,
1015 : max_worker_processes),
1016 : errhint("Consider increasing the configuration parameter \"%s\".", "max_worker_processes")));
1017 0 : return;
1018 : }
1019 :
1020 : /*
1021 : * Copy the registration data into the registered workers list.
1022 : */
1023 886 : rw = MemoryContextAllocExtended(PostmasterContext,
1024 : sizeof(RegisteredBgWorker),
1025 : MCXT_ALLOC_NO_OOM);
1026 886 : if (rw == NULL)
1027 : {
1028 0 : ereport(LOG,
1029 : (errcode(ERRCODE_OUT_OF_MEMORY),
1030 : errmsg("out of memory")));
1031 0 : return;
1032 : }
1033 :
1034 886 : rw->rw_worker = *worker;
1035 886 : rw->rw_pid = 0;
1036 886 : rw->rw_crashed_at = 0;
1037 886 : rw->rw_terminate = false;
1038 :
1039 886 : dlist_push_head(&BackgroundWorkerList, &rw->rw_lnode);
1040 : }
1041 :
1042 : /*
1043 : * Register a new background worker from a regular backend.
1044 : *
1045 : * Returns true on success and false on failure. Failure typically indicates
1046 : * that no background worker slots are currently available.
1047 : *
1048 : * If handle != NULL, we'll set *handle to a pointer that can subsequently
1049 : * be used as an argument to GetBackgroundWorkerPid(). The caller can
1050 : * free this pointer using pfree(), if desired.
1051 : */
1052 : bool
1053 1973 : RegisterDynamicBackgroundWorker(BackgroundWorker *worker,
1054 : BackgroundWorkerHandle **handle)
1055 : {
1056 : int slotno;
1057 1973 : bool success = false;
1058 : bool parallel;
1059 1973 : uint64 generation = 0;
1060 :
1061 : /*
1062 : * We can't register dynamic background workers from the postmaster. If
1063 : * this is a standalone backend, we're the only process and can't start
1064 : * any more. In a multi-process environment, it might be theoretically
1065 : * possible, but we don't currently support it due to locking
1066 : * considerations; see comments on the BackgroundWorkerSlot data
1067 : * structure.
1068 : */
1069 1973 : if (!IsUnderPostmaster)
1070 0 : return false;
1071 :
1072 1973 : if (!SanityCheckBackgroundWorker(worker, ERROR))
1073 0 : return false;
1074 :
1075 1973 : parallel = (worker->bgw_flags & BGWORKER_CLASS_PARALLEL) != 0;
1076 :
1077 1973 : LWLockAcquire(BackgroundWorkerLock, LW_EXCLUSIVE);
1078 :
1079 : /*
1080 : * If this is a parallel worker, check whether there are already too many
1081 : * parallel workers; if so, don't register another one. Our view of
1082 : * parallel_terminate_count may be slightly stale, but that doesn't really
1083 : * matter: we would have gotten the same result if we'd arrived here
1084 : * slightly earlier anyway. There's no help for it, either, since the
1085 : * postmaster must not take locks; a memory barrier wouldn't guarantee
1086 : * anything useful.
1087 : */
1088 1973 : if (parallel && (BackgroundWorkerData->parallel_register_count -
1089 1505 : BackgroundWorkerData->parallel_terminate_count) >=
1090 : max_parallel_workers)
1091 : {
1092 : Assert(BackgroundWorkerData->parallel_register_count -
1093 : BackgroundWorkerData->parallel_terminate_count <=
1094 : MAX_PARALLEL_WORKER_LIMIT);
1095 10 : LWLockRelease(BackgroundWorkerLock);
1096 10 : return false;
1097 : }
1098 :
1099 : /*
1100 : * Look for an unused slot. If we find one, grab it.
1101 : */
1102 6222 : for (slotno = 0; slotno < BackgroundWorkerData->total_slots; ++slotno)
1103 : {
1104 6216 : BackgroundWorkerSlot *slot = &BackgroundWorkerData->slot[slotno];
1105 :
1106 6216 : if (!slot->in_use)
1107 : {
1108 1957 : memcpy(&slot->worker, worker, sizeof(BackgroundWorker));
1109 1957 : slot->pid = InvalidPid; /* indicates not started yet */
1110 1957 : slot->generation++;
1111 1957 : slot->terminate = false;
1112 1957 : generation = slot->generation;
1113 1957 : if (parallel)
1114 1489 : BackgroundWorkerData->parallel_register_count++;
1115 :
1116 : /*
1117 : * Make sure postmaster doesn't see the slot as in use before it
1118 : * sees the new contents.
1119 : */
1120 1957 : pg_write_barrier();
1121 :
1122 1957 : slot->in_use = true;
1123 1957 : success = true;
1124 1957 : break;
1125 : }
1126 : }
1127 :
1128 1963 : LWLockRelease(BackgroundWorkerLock);
1129 :
1130 : /* If we found a slot, tell the postmaster to notice the change. */
1131 1963 : if (success)
1132 1957 : SendPostmasterSignal(PMSIGNAL_BACKGROUND_WORKER_CHANGE);
1133 :
1134 : /*
1135 : * If we found a slot and the user has provided a handle, initialize it.
1136 : */
1137 1963 : if (success && handle)
1138 : {
1139 1957 : *handle = palloc_object(BackgroundWorkerHandle);
1140 1957 : (*handle)->slot = slotno;
1141 1957 : (*handle)->generation = generation;
1142 : }
1143 :
1144 1963 : return success;
1145 : }
1146 :
1147 : /*
1148 : * Get the PID of a dynamically-registered background worker.
1149 : *
1150 : * If the worker is determined to be running, the return value will be
1151 : * BGWH_STARTED and *pidp will get the PID of the worker process. If the
1152 : * postmaster has not yet attempted to start the worker, the return value will
1153 : * be BGWH_NOT_YET_STARTED. Otherwise, the return value is BGWH_STOPPED.
1154 : *
1155 : * BGWH_STOPPED can indicate either that the worker is temporarily stopped
1156 : * (because it is configured for automatic restart and exited non-zero),
1157 : * or that the worker is permanently stopped (because it exited with exit
1158 : * code 0, or was not configured for automatic restart), or even that the
1159 : * worker was unregistered without ever starting (either because startup
1160 : * failed and the worker is not configured for automatic restart, or because
1161 : * TerminateBackgroundWorker was used before the worker was successfully
1162 : * started).
1163 : */
1164 : BgwHandleStatus
1165 9845363 : GetBackgroundWorkerPid(BackgroundWorkerHandle *handle, pid_t *pidp)
1166 : {
1167 : BackgroundWorkerSlot *slot;
1168 : pid_t pid;
1169 :
1170 : Assert(handle->slot < max_worker_processes);
1171 9845363 : slot = &BackgroundWorkerData->slot[handle->slot];
1172 :
1173 : /*
1174 : * We could probably arrange to synchronize access to data using memory
1175 : * barriers only, but for now, let's just keep it simple and grab the
1176 : * lock. It seems unlikely that there will be enough traffic here to
1177 : * result in meaningful contention.
1178 : */
1179 9845363 : LWLockAcquire(BackgroundWorkerLock, LW_SHARED);
1180 :
1181 : /*
1182 : * The generation number can't be concurrently changed while we hold the
1183 : * lock. The pid, which is updated by the postmaster, can change at any
1184 : * time, but we assume such changes are atomic. So the value we read
1185 : * won't be garbage, but it might be out of date by the time the caller
1186 : * examines it (but that's unavoidable anyway).
1187 : *
1188 : * The in_use flag could be in the process of changing from true to false,
1189 : * but if it is already false then it can't change further.
1190 : */
1191 9845363 : if (handle->generation != slot->generation || !slot->in_use)
1192 1490 : pid = 0;
1193 : else
1194 9843873 : pid = slot->pid;
1195 :
1196 : /* All done. */
1197 9845363 : LWLockRelease(BackgroundWorkerLock);
1198 :
1199 9845363 : if (pid == 0)
1200 1490 : return BGWH_STOPPED;
1201 9843873 : else if (pid == InvalidPid)
1202 943020 : return BGWH_NOT_YET_STARTED;
1203 8900853 : *pidp = pid;
1204 8900853 : return BGWH_STARTED;
1205 : }
1206 :
1207 : /*
1208 : * Wait for a background worker to start up.
1209 : *
1210 : * This is like GetBackgroundWorkerPid(), except that if the worker has not
1211 : * yet started, we wait for it to do so; thus, BGWH_NOT_YET_STARTED is never
1212 : * returned. However, if the postmaster has died, we give up and return
1213 : * BGWH_POSTMASTER_DIED, since it that case we know that startup will not
1214 : * take place.
1215 : *
1216 : * The caller *must* have set our PID as the worker's bgw_notify_pid,
1217 : * else we will not be awoken promptly when the worker's state changes.
1218 : */
1219 : BgwHandleStatus
1220 12 : WaitForBackgroundWorkerStartup(BackgroundWorkerHandle *handle, pid_t *pidp)
1221 : {
1222 : BgwHandleStatus status;
1223 : int rc;
1224 :
1225 : for (;;)
1226 7 : {
1227 : pid_t pid;
1228 :
1229 19 : CHECK_FOR_INTERRUPTS();
1230 :
1231 19 : status = GetBackgroundWorkerPid(handle, &pid);
1232 19 : if (status == BGWH_STARTED)
1233 12 : *pidp = pid;
1234 19 : if (status != BGWH_NOT_YET_STARTED)
1235 12 : break;
1236 :
1237 7 : rc = WaitLatch(MyLatch,
1238 : WL_LATCH_SET | WL_POSTMASTER_DEATH, 0,
1239 : WAIT_EVENT_BGWORKER_STARTUP);
1240 :
1241 7 : if (rc & WL_POSTMASTER_DEATH)
1242 : {
1243 0 : status = BGWH_POSTMASTER_DIED;
1244 0 : break;
1245 : }
1246 :
1247 7 : ResetLatch(MyLatch);
1248 : }
1249 :
1250 12 : return status;
1251 : }
1252 :
1253 : /*
1254 : * Wait for a background worker to stop.
1255 : *
1256 : * If the worker hasn't yet started, or is running, we wait for it to stop
1257 : * and then return BGWH_STOPPED. However, if the postmaster has died, we give
1258 : * up and return BGWH_POSTMASTER_DIED, because it's the postmaster that
1259 : * notifies us when a worker's state changes.
1260 : *
1261 : * The caller *must* have set our PID as the worker's bgw_notify_pid,
1262 : * else we will not be awoken promptly when the worker's state changes.
1263 : */
1264 : BgwHandleStatus
1265 1490 : WaitForBackgroundWorkerShutdown(BackgroundWorkerHandle *handle)
1266 : {
1267 : BgwHandleStatus status;
1268 : int rc;
1269 :
1270 : for (;;)
1271 1725 : {
1272 : pid_t pid;
1273 :
1274 3215 : CHECK_FOR_INTERRUPTS();
1275 :
1276 3215 : status = GetBackgroundWorkerPid(handle, &pid);
1277 3215 : if (status == BGWH_STOPPED)
1278 1490 : break;
1279 :
1280 1725 : rc = WaitLatch(MyLatch,
1281 : WL_LATCH_SET | WL_POSTMASTER_DEATH, 0,
1282 : WAIT_EVENT_BGWORKER_SHUTDOWN);
1283 :
1284 1725 : if (rc & WL_POSTMASTER_DEATH)
1285 : {
1286 0 : status = BGWH_POSTMASTER_DIED;
1287 0 : break;
1288 : }
1289 :
1290 1725 : ResetLatch(MyLatch);
1291 : }
1292 :
1293 1490 : return status;
1294 : }
1295 :
1296 : /*
1297 : * Instruct the postmaster to terminate a background worker.
1298 : *
1299 : * Note that it's safe to do this without regard to whether the worker is
1300 : * still running, or even if the worker may already have exited and been
1301 : * unregistered.
1302 : */
1303 : void
1304 6 : TerminateBackgroundWorker(BackgroundWorkerHandle *handle)
1305 : {
1306 : BackgroundWorkerSlot *slot;
1307 6 : bool signal_postmaster = false;
1308 :
1309 : Assert(handle->slot < max_worker_processes);
1310 6 : slot = &BackgroundWorkerData->slot[handle->slot];
1311 :
1312 : /* Set terminate flag in shared memory, unless slot has been reused. */
1313 6 : LWLockAcquire(BackgroundWorkerLock, LW_EXCLUSIVE);
1314 6 : if (handle->generation == slot->generation)
1315 : {
1316 6 : slot->terminate = true;
1317 6 : signal_postmaster = true;
1318 : }
1319 6 : LWLockRelease(BackgroundWorkerLock);
1320 :
1321 : /* Make sure the postmaster notices the change to shared memory. */
1322 6 : if (signal_postmaster)
1323 6 : SendPostmasterSignal(PMSIGNAL_BACKGROUND_WORKER_CHANGE);
1324 6 : }
1325 :
1326 : /*
1327 : * Look up (and possibly load) a bgworker entry point function.
1328 : *
1329 : * For functions contained in the core code, we use library name "postgres"
1330 : * and consult the InternalBGWorkers array. External functions are
1331 : * looked up, and loaded if necessary, using load_external_function().
1332 : *
1333 : * The point of this is to pass function names as strings across process
1334 : * boundaries. We can't pass actual function addresses because of the
1335 : * possibility that the function has been loaded at a different address
1336 : * in a different process. This is obviously a hazard for functions in
1337 : * loadable libraries, but it can happen even for functions in the core code
1338 : * on platforms using EXEC_BACKEND (e.g., Windows).
1339 : *
1340 : * At some point it might be worthwhile to get rid of InternalBGWorkers[]
1341 : * in favor of applying load_external_function() for core functions too;
1342 : * but that raises portability issues that are not worth addressing now.
1343 : */
1344 : static bgworker_main_type
1345 2565 : LookupBackgroundWorkerFunction(const char *libraryname, const char *funcname)
1346 : {
1347 : /*
1348 : * If the function is to be loaded from postgres itself, search the
1349 : * InternalBGWorkers array.
1350 : */
1351 2565 : if (strcmp(libraryname, "postgres") == 0)
1352 : {
1353 : int i;
1354 :
1355 4649 : for (i = 0; i < lengthof(InternalBGWorkers); i++)
1356 : {
1357 4649 : if (strcmp(InternalBGWorkers[i].fn_name, funcname) == 0)
1358 2547 : return InternalBGWorkers[i].fn_addr;
1359 : }
1360 :
1361 : /* We can only reach this by programming error. */
1362 0 : elog(ERROR, "internal function \"%s\" not found", funcname);
1363 : }
1364 :
1365 : /* Otherwise load from external library. */
1366 18 : return (bgworker_main_type)
1367 18 : load_external_function(libraryname, funcname, true, NULL);
1368 : }
1369 :
1370 : /*
1371 : * Given a PID, get the bgw_type of the background worker. Returns NULL if
1372 : * not a valid background worker.
1373 : *
1374 : * The return value is in static memory belonging to this function, so it has
1375 : * to be used before calling this function again. This is so that the caller
1376 : * doesn't have to worry about the background worker locking protocol.
1377 : */
1378 : const char *
1379 991 : GetBackgroundWorkerTypeByPid(pid_t pid)
1380 : {
1381 : int slotno;
1382 991 : bool found = false;
1383 : static char result[BGW_MAXLEN];
1384 :
1385 991 : LWLockAcquire(BackgroundWorkerLock, LW_SHARED);
1386 :
1387 1135 : for (slotno = 0; slotno < BackgroundWorkerData->total_slots; slotno++)
1388 : {
1389 1135 : BackgroundWorkerSlot *slot = &BackgroundWorkerData->slot[slotno];
1390 :
1391 1135 : if (slot->pid > 0 && slot->pid == pid)
1392 : {
1393 991 : strcpy(result, slot->worker.bgw_type);
1394 991 : found = true;
1395 991 : break;
1396 : }
1397 : }
1398 :
1399 991 : LWLockRelease(BackgroundWorkerLock);
1400 :
1401 991 : if (!found)
1402 0 : return NULL;
1403 :
1404 991 : return result;
1405 : }
1406 :
1407 : /*
1408 : * Terminate all background workers connected to the given database, if the
1409 : * workers can be interrupted.
1410 : */
1411 : void
1412 7 : TerminateBackgroundWorkersForDatabase(Oid databaseId)
1413 : {
1414 7 : bool signal_postmaster = false;
1415 :
1416 7 : LWLockAcquire(BackgroundWorkerLock, LW_EXCLUSIVE);
1417 :
1418 : /*
1419 : * Iterate through slots, looking for workers connected to the given
1420 : * database.
1421 : */
1422 63 : for (int slotno = 0; slotno < BackgroundWorkerData->total_slots; slotno++)
1423 : {
1424 56 : BackgroundWorkerSlot *slot = &BackgroundWorkerData->slot[slotno];
1425 :
1426 56 : if (slot->in_use &&
1427 14 : (slot->worker.bgw_flags & BGWORKER_INTERRUPTIBLE))
1428 : {
1429 4 : PGPROC *proc = BackendPidGetProc(slot->pid);
1430 :
1431 4 : if (proc && proc->databaseId == databaseId)
1432 : {
1433 4 : slot->terminate = true;
1434 4 : signal_postmaster = true;
1435 : }
1436 : }
1437 : }
1438 :
1439 7 : LWLockRelease(BackgroundWorkerLock);
1440 :
1441 : /* Make sure the postmaster notices the change to shared memory. */
1442 7 : if (signal_postmaster)
1443 4 : SendPostmasterSignal(PMSIGNAL_BACKGROUND_WORKER_CHANGE);
1444 7 : }
|
