1 /*-------------------------------------------------------------------------
2 *
3 * parallel.c
4 * Infrastructure for launching parallel workers
5 *
6 * Portions Copyright (c) 1996-2016, PostgreSQL Global Development Group
7 * Portions Copyright (c) 1994, Regents of the University of California
8 *
9 * IDENTIFICATION
10 * src/backend/access/transam/parallel.c
11 *
12 *-------------------------------------------------------------------------
13 */
14
15 #include "postgres.h"
16
17 #include "access/parallel.h"
18 #include "access/xact.h"
19 #include "access/xlog.h"
20 #include "catalog/namespace.h"
21 #include "commands/async.h"
22 #include "executor/execParallel.h"
23 #include "libpq/libpq.h"
24 #include "libpq/pqformat.h"
25 #include "libpq/pqmq.h"
26 #include "miscadmin.h"
27 #include "optimizer/planmain.h"
28 #include "storage/ipc.h"
29 #include "storage/sinval.h"
30 #include "storage/spin.h"
31 #include "tcop/tcopprot.h"
32 #include "utils/combocid.h"
33 #include "utils/guc.h"
34 #include "utils/inval.h"
35 #include "utils/memutils.h"
36 #include "utils/resowner.h"
37 #include "utils/snapmgr.h"
38
39
40 /*
41 * We don't want to waste a lot of memory on an error queue which, most of
42 * the time, will process only a handful of small messages. However, it is
43 * desirable to make it large enough that a typical ErrorResponse can be sent
44 * without blocking. That way, a worker that errors out can write the whole
45 * message into the queue and terminate without waiting for the user backend.
46 */
47 #define PARALLEL_ERROR_QUEUE_SIZE 16384
48
49 /* Magic number for parallel context TOC. */
50 #define PARALLEL_MAGIC 0x50477c7c
51
52 /*
53 * Magic numbers for parallel state sharing. Higher-level code should use
54 * smaller values, leaving these very large ones for use by this module.
55 */
56 #define PARALLEL_KEY_FIXED UINT64CONST(0xFFFFFFFFFFFF0001)
57 #define PARALLEL_KEY_ERROR_QUEUE UINT64CONST(0xFFFFFFFFFFFF0002)
58 #define PARALLEL_KEY_LIBRARY UINT64CONST(0xFFFFFFFFFFFF0003)
59 #define PARALLEL_KEY_GUC UINT64CONST(0xFFFFFFFFFFFF0004)
60 #define PARALLEL_KEY_COMBO_CID UINT64CONST(0xFFFFFFFFFFFF0005)
61 #define PARALLEL_KEY_TRANSACTION_SNAPSHOT UINT64CONST(0xFFFFFFFFFFFF0006)
62 #define PARALLEL_KEY_ACTIVE_SNAPSHOT UINT64CONST(0xFFFFFFFFFFFF0007)
63 #define PARALLEL_KEY_TRANSACTION_STATE UINT64CONST(0xFFFFFFFFFFFF0008)
64 #define PARALLEL_KEY_ENTRYPOINT UINT64CONST(0xFFFFFFFFFFFF0009)
65
66 /* Fixed-size parallel state. */
67 typedef struct FixedParallelState
68 {
69 /* Fixed-size state that workers must restore. */
70 Oid database_id;
71 Oid authenticated_user_id;
72 Oid current_user_id;
73 Oid outer_user_id;
74 Oid temp_namespace_id;
75 Oid temp_toast_namespace_id;
76 int sec_context;
77 bool is_superuser;
78 PGPROC *parallel_master_pgproc;
79 pid_t parallel_master_pid;
80 BackendId parallel_master_backend_id;
81 TimestampTz xact_ts;
82 TimestampTz stmt_ts;
83
84 /* Entrypoint for parallel workers (deprecated)! */
85 parallel_worker_main_type entrypoint;
86
87 /* Mutex protects remaining fields. */
88 slock_t mutex;
89
90 /* Maximum XactLastRecEnd of any worker. */
91 XLogRecPtr last_xlog_end;
92 } FixedParallelState;
93
94 /*
95 * Our parallel worker number. We initialize this to -1, meaning that we are
96 * not a parallel worker. In parallel workers, it will be set to a value >= 0
97 * and < the number of workers before any user code is invoked; each parallel
98 * worker will get a different parallel worker number.
99 */
100 int ParallelWorkerNumber = -1;
101
102 /* Is there a parallel message pending which we need to receive? */
103 volatile bool ParallelMessagePending = false;
104
105 /* Are we initializing a parallel worker? */
106 bool InitializingParallelWorker = false;
107
108 /* Pointer to our fixed parallel state. */
109 static FixedParallelState *MyFixedParallelState;
110
111 /* List of active parallel contexts. */
112 static dlist_head pcxt_list = DLIST_STATIC_INIT(pcxt_list);
113
114 /* Backend-local copy of data from FixedParallelState. */
115 static pid_t ParallelMasterPid;
116
117 /*
118 * List of internal parallel worker entry points. We need this for
119 * reasons explained in LookupParallelWorkerFunction(), below.
120 */
121 static const struct
122 {
123 const char *fn_name;
124 parallel_worker_main_type fn_addr;
125 } InternalParallelWorkers[] =
126
127 {
128 {
129 "ParallelQueryMain", ParallelQueryMain
130 }
131 };
132
133 /* Private functions. */
134 static void HandleParallelMessage(ParallelContext *pcxt, int i, StringInfo msg);
135 static void WaitForParallelWorkersToExit(ParallelContext *pcxt);
136 static parallel_worker_main_type LookupParallelWorkerFunction(char *libraryname, char *funcname);
137 static void ParallelWorkerShutdown(int code, Datum arg);
138
139
140 /*
141 * Establish a new parallel context. This should be done after entering
142 * parallel mode, and (unless there is an error) the context should be
143 * destroyed before exiting the current subtransaction.
144 *
145 * NB: specifying the entrypoint as a function address is unportable.
146 * This will go away in Postgres 10, in favor of the API provided by
147 * CreateParallelContextForExternalFunction; in the meantime use that.
148 */
149 ParallelContext *
CreateParallelContext(parallel_worker_main_type entrypoint,int nworkers)150 CreateParallelContext(parallel_worker_main_type entrypoint, int nworkers)
151 {
152 MemoryContext oldcontext;
153 ParallelContext *pcxt;
154
155 /* It is unsafe to create a parallel context if not in parallel mode. */
156 Assert(IsInParallelMode());
157
158 /* Number of workers should be non-negative. */
159 Assert(nworkers >= 0);
160
161 /*
162 * If dynamic shared memory is not available, we won't be able to use
163 * background workers.
164 */
165 if (dynamic_shared_memory_type == DSM_IMPL_NONE)
166 nworkers = 0;
167
168 /*
169 * If we are running under serializable isolation, we can't use parallel
170 * workers, at least not until somebody enhances that mechanism to be
171 * parallel-aware.
172 */
173 if (IsolationIsSerializable())
174 nworkers = 0;
175
176 /* We might be running in a short-lived memory context. */
177 oldcontext = MemoryContextSwitchTo(TopTransactionContext);
178
179 /* Initialize a new ParallelContext. */
180 pcxt = palloc0(sizeof(ParallelContext));
181 pcxt->subid = GetCurrentSubTransactionId();
182 pcxt->nworkers = nworkers;
183 pcxt->entrypoint = entrypoint;
184 pcxt->error_context_stack = error_context_stack;
185 shm_toc_initialize_estimator(&pcxt->estimator);
186 dlist_push_head(&pcxt_list, &pcxt->node);
187
188 /* Restore previous memory context. */
189 MemoryContextSwitchTo(oldcontext);
190
191 return pcxt;
192 }
193
194 /*
195 * Establish a new parallel context that calls a function specified by name.
196 * Unlike CreateParallelContext, this is robust against possible differences
197 * in address space layout between different processes.
198 */
199 ParallelContext *
CreateParallelContextForExternalFunction(char * library_name,char * function_name,int nworkers)200 CreateParallelContextForExternalFunction(char *library_name,
201 char *function_name,
202 int nworkers)
203 {
204 MemoryContext oldcontext;
205 ParallelContext *pcxt;
206
207 /* We might be running in a very short-lived memory context. */
208 oldcontext = MemoryContextSwitchTo(TopTransactionContext);
209
210 /* Create the context. */
211 pcxt = CreateParallelContext(NULL, nworkers);
212 pcxt->library_name = pstrdup(library_name);
213 pcxt->function_name = pstrdup(function_name);
214
215 /* Restore previous memory context. */
216 MemoryContextSwitchTo(oldcontext);
217
218 return pcxt;
219 }
220
221 /*
222 * Establish the dynamic shared memory segment for a parallel context and
223 * copy state and other bookkeeping information that will be needed by
224 * parallel workers into it.
225 */
226 void
InitializeParallelDSM(ParallelContext * pcxt)227 InitializeParallelDSM(ParallelContext *pcxt)
228 {
229 MemoryContext oldcontext;
230 Size library_len = 0;
231 Size guc_len = 0;
232 Size combocidlen = 0;
233 Size tsnaplen = 0;
234 Size asnaplen = 0;
235 Size tstatelen = 0;
236 Size segsize = 0;
237 int i;
238 FixedParallelState *fps;
239 Snapshot transaction_snapshot = GetTransactionSnapshot();
240 Snapshot active_snapshot = GetActiveSnapshot();
241
242 /* We might be running in a very short-lived memory context. */
243 oldcontext = MemoryContextSwitchTo(TopTransactionContext);
244
245 /* Allow space to store the fixed-size parallel state. */
246 shm_toc_estimate_chunk(&pcxt->estimator, sizeof(FixedParallelState));
247 shm_toc_estimate_keys(&pcxt->estimator, 1);
248
249 /*
250 * Normally, the user will have requested at least one worker process, but
251 * if by chance they have not, we can skip a bunch of things here.
252 */
253 if (pcxt->nworkers > 0)
254 {
255 /* Estimate space for various kinds of state sharing. */
256 library_len = EstimateLibraryStateSpace();
257 shm_toc_estimate_chunk(&pcxt->estimator, library_len);
258 guc_len = EstimateGUCStateSpace();
259 shm_toc_estimate_chunk(&pcxt->estimator, guc_len);
260 combocidlen = EstimateComboCIDStateSpace();
261 shm_toc_estimate_chunk(&pcxt->estimator, combocidlen);
262 tsnaplen = EstimateSnapshotSpace(transaction_snapshot);
263 shm_toc_estimate_chunk(&pcxt->estimator, tsnaplen);
264 asnaplen = EstimateSnapshotSpace(active_snapshot);
265 shm_toc_estimate_chunk(&pcxt->estimator, asnaplen);
266 tstatelen = EstimateTransactionStateSpace();
267 shm_toc_estimate_chunk(&pcxt->estimator, tstatelen);
268 /* If you add more chunks here, you probably need to add keys. */
269 shm_toc_estimate_keys(&pcxt->estimator, 6);
270
271 /* Estimate space need for error queues. */
272 StaticAssertStmt(BUFFERALIGN(PARALLEL_ERROR_QUEUE_SIZE) ==
273 PARALLEL_ERROR_QUEUE_SIZE,
274 "parallel error queue size not buffer-aligned");
275 shm_toc_estimate_chunk(&pcxt->estimator,
276 mul_size(PARALLEL_ERROR_QUEUE_SIZE,
277 pcxt->nworkers));
278 shm_toc_estimate_keys(&pcxt->estimator, 1);
279
280 /* Estimate how much we'll need for entrypoint info. */
281 if (pcxt->library_name != NULL)
282 {
283 Assert(pcxt->function_name != NULL);
284 shm_toc_estimate_chunk(&pcxt->estimator, strlen(pcxt->library_name)
285 + strlen(pcxt->function_name) + 2);
286 shm_toc_estimate_keys(&pcxt->estimator, 1);
287 }
288 }
289
290 /*
291 * Create DSM and initialize with new table of contents. But if the user
292 * didn't request any workers, then don't bother creating a dynamic shared
293 * memory segment; instead, just use backend-private memory.
294 *
295 * Also, if we can't create a dynamic shared memory segment because the
296 * maximum number of segments have already been created, then fall back to
297 * backend-private memory, and plan not to use any workers. We hope this
298 * won't happen very often, but it's better to abandon the use of
299 * parallelism than to fail outright.
300 */
301 segsize = shm_toc_estimate(&pcxt->estimator);
302 if (pcxt->nworkers > 0)
303 pcxt->seg = dsm_create(segsize, DSM_CREATE_NULL_IF_MAXSEGMENTS);
304 if (pcxt->seg != NULL)
305 pcxt->toc = shm_toc_create(PARALLEL_MAGIC,
306 dsm_segment_address(pcxt->seg),
307 segsize);
308 else
309 {
310 pcxt->nworkers = 0;
311 pcxt->private_memory = MemoryContextAlloc(TopMemoryContext, segsize);
312 pcxt->toc = shm_toc_create(PARALLEL_MAGIC, pcxt->private_memory,
313 segsize);
314 }
315
316 /* Initialize fixed-size state in shared memory. */
317 fps = (FixedParallelState *)
318 shm_toc_allocate(pcxt->toc, sizeof(FixedParallelState));
319 fps->database_id = MyDatabaseId;
320 fps->authenticated_user_id = GetAuthenticatedUserId();
321 fps->outer_user_id = GetCurrentRoleId();
322 fps->is_superuser = session_auth_is_superuser;
323 GetUserIdAndSecContext(&fps->current_user_id, &fps->sec_context);
324 GetTempNamespaceState(&fps->temp_namespace_id,
325 &fps->temp_toast_namespace_id);
326 fps->parallel_master_pgproc = MyProc;
327 fps->parallel_master_pid = MyProcPid;
328 fps->parallel_master_backend_id = MyBackendId;
329 fps->xact_ts = GetCurrentTransactionStartTimestamp();
330 fps->stmt_ts = GetCurrentStatementStartTimestamp();
331 fps->entrypoint = pcxt->entrypoint;
332 SpinLockInit(&fps->mutex);
333 fps->last_xlog_end = 0;
334 shm_toc_insert(pcxt->toc, PARALLEL_KEY_FIXED, fps);
335
336 /* We can skip the rest of this if we're not budgeting for any workers. */
337 if (pcxt->nworkers > 0)
338 {
339 char *libraryspace;
340 char *gucspace;
341 char *combocidspace;
342 char *tsnapspace;
343 char *asnapspace;
344 char *tstatespace;
345 char *error_queue_space;
346
347 /* Serialize shared libraries we have loaded. */
348 libraryspace = shm_toc_allocate(pcxt->toc, library_len);
349 SerializeLibraryState(library_len, libraryspace);
350 shm_toc_insert(pcxt->toc, PARALLEL_KEY_LIBRARY, libraryspace);
351
352 /* Serialize GUC settings. */
353 gucspace = shm_toc_allocate(pcxt->toc, guc_len);
354 SerializeGUCState(guc_len, gucspace);
355 shm_toc_insert(pcxt->toc, PARALLEL_KEY_GUC, gucspace);
356
357 /* Serialize combo CID state. */
358 combocidspace = shm_toc_allocate(pcxt->toc, combocidlen);
359 SerializeComboCIDState(combocidlen, combocidspace);
360 shm_toc_insert(pcxt->toc, PARALLEL_KEY_COMBO_CID, combocidspace);
361
362 /* Serialize transaction snapshot and active snapshot. */
363 tsnapspace = shm_toc_allocate(pcxt->toc, tsnaplen);
364 SerializeSnapshot(transaction_snapshot, tsnapspace);
365 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT,
366 tsnapspace);
367 asnapspace = shm_toc_allocate(pcxt->toc, asnaplen);
368 SerializeSnapshot(active_snapshot, asnapspace);
369 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ACTIVE_SNAPSHOT, asnapspace);
370
371 /* Serialize transaction state. */
372 tstatespace = shm_toc_allocate(pcxt->toc, tstatelen);
373 SerializeTransactionState(tstatelen, tstatespace);
374 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_STATE, tstatespace);
375
376 /* Allocate space for worker information. */
377 pcxt->worker = palloc0(sizeof(ParallelWorkerInfo) * pcxt->nworkers);
378
379 /*
380 * Establish error queues in dynamic shared memory.
381 *
382 * These queues should be used only for transmitting ErrorResponse,
383 * NoticeResponse, and NotifyResponse protocol messages. Tuple data
384 * should be transmitted via separate (possibly larger?) queues.
385 */
386 error_queue_space =
387 shm_toc_allocate(pcxt->toc,
388 mul_size(PARALLEL_ERROR_QUEUE_SIZE,
389 pcxt->nworkers));
390 for (i = 0; i < pcxt->nworkers; ++i)
391 {
392 char *start;
393 shm_mq *mq;
394
395 start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
396 mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
397 shm_mq_set_receiver(mq, MyProc);
398 pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
399 }
400 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE, error_queue_space);
401
402 /* Serialize entrypoint information. */
403 if (pcxt->library_name != NULL)
404 {
405 Size lnamelen = strlen(pcxt->library_name);
406 char *extensionstate;
407
408 extensionstate = shm_toc_allocate(pcxt->toc, lnamelen
409 + strlen(pcxt->function_name) + 2);
410 strcpy(extensionstate, pcxt->library_name);
411 strcpy(extensionstate + lnamelen + 1, pcxt->function_name);
412 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ENTRYPOINT,
413 extensionstate);
414 }
415 }
416
417 /* Restore previous memory context. */
418 MemoryContextSwitchTo(oldcontext);
419 }
420
421 /*
422 * Reinitialize the dynamic shared memory segment for a parallel context such
423 * that we could launch workers for it again.
424 */
425 void
ReinitializeParallelDSM(ParallelContext * pcxt)426 ReinitializeParallelDSM(ParallelContext *pcxt)
427 {
428 FixedParallelState *fps;
429 char *error_queue_space;
430 int i;
431
432 /* Wait for any old workers to exit. */
433 if (pcxt->nworkers_launched > 0)
434 {
435 WaitForParallelWorkersToFinish(pcxt);
436 WaitForParallelWorkersToExit(pcxt);
437 pcxt->nworkers_launched = 0;
438 if (pcxt->any_message_received)
439 {
440 pfree(pcxt->any_message_received);
441 pcxt->any_message_received = NULL;
442 }
443 }
444
445 /* Reset a few bits of fixed parallel state to a clean state. */
446 fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED);
447 fps->last_xlog_end = 0;
448
449 /* Recreate error queues. */
450 error_queue_space =
451 shm_toc_lookup(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE);
452 for (i = 0; i < pcxt->nworkers; ++i)
453 {
454 char *start;
455 shm_mq *mq;
456
457 start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
458 mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
459 shm_mq_set_receiver(mq, MyProc);
460 pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
461 }
462 }
463
464 /*
465 * Launch parallel workers.
466 */
467 void
LaunchParallelWorkers(ParallelContext * pcxt)468 LaunchParallelWorkers(ParallelContext *pcxt)
469 {
470 MemoryContext oldcontext;
471 BackgroundWorker worker;
472 int i;
473 bool any_registrations_failed = false;
474
475 /* Skip this if we have no workers. */
476 if (pcxt->nworkers == 0)
477 return;
478
479 /* We need to be a lock group leader. */
480 BecomeLockGroupLeader();
481
482 /* If we do have workers, we'd better have a DSM segment. */
483 Assert(pcxt->seg != NULL);
484
485 /* We might be running in a short-lived memory context. */
486 oldcontext = MemoryContextSwitchTo(TopTransactionContext);
487
488 /* Configure a worker. */
489 snprintf(worker.bgw_name, BGW_MAXLEN, "parallel worker for PID %d",
490 MyProcPid);
491 worker.bgw_flags =
492 BGWORKER_SHMEM_ACCESS | BGWORKER_BACKEND_DATABASE_CONNECTION;
493 worker.bgw_start_time = BgWorkerStart_ConsistentState;
494 worker.bgw_restart_time = BGW_NEVER_RESTART;
495 worker.bgw_main = NULL;
496 sprintf(worker.bgw_library_name, "postgres");
497 sprintf(worker.bgw_function_name, "ParallelWorkerMain");
498 worker.bgw_main_arg = UInt32GetDatum(dsm_segment_handle(pcxt->seg));
499 worker.bgw_notify_pid = MyProcPid;
500 memset(&worker.bgw_extra, 0, BGW_EXTRALEN);
501
502 /*
503 * Start workers.
504 *
505 * The caller must be able to tolerate ending up with fewer workers than
506 * expected, so there is no need to throw an error here if registration
507 * fails. It wouldn't help much anyway, because registering the worker in
508 * no way guarantees that it will start up and initialize successfully.
509 */
510 for (i = 0; i < pcxt->nworkers; ++i)
511 {
512 memcpy(worker.bgw_extra, &i, sizeof(int));
513 if (!any_registrations_failed &&
514 RegisterDynamicBackgroundWorker(&worker,
515 &pcxt->worker[i].bgwhandle))
516 {
517 shm_mq_set_handle(pcxt->worker[i].error_mqh,
518 pcxt->worker[i].bgwhandle);
519 pcxt->nworkers_launched++;
520 }
521 else
522 {
523 /*
524 * If we weren't able to register the worker, then we've bumped up
525 * against the max_worker_processes limit, and future
526 * registrations will probably fail too, so arrange to skip them.
527 * But we still have to execute this code for the remaining slots
528 * to make sure that we forget about the error queues we budgeted
529 * for those workers. Otherwise, we'll wait for them to start,
530 * but they never will.
531 */
532 any_registrations_failed = true;
533 pcxt->worker[i].bgwhandle = NULL;
534 pfree(pcxt->worker[i].error_mqh);
535 pcxt->worker[i].error_mqh = NULL;
536 }
537 }
538
539 /*
540 * Now that nworkers_launched has taken its final value, we can initialize
541 * any_message_received.
542 */
543 if (pcxt->nworkers_launched > 0)
544 pcxt->any_message_received =
545 palloc0(sizeof(bool) * pcxt->nworkers_launched);
546
547 /* Restore previous memory context. */
548 MemoryContextSwitchTo(oldcontext);
549 }
550
551 /*
552 * Wait for all workers to finish computing.
553 *
554 * Even if the parallel operation seems to have completed successfully, it's
555 * important to call this function afterwards. We must not miss any errors
556 * the workers may have thrown during the parallel operation, or any that they
557 * may yet throw while shutting down.
558 *
559 * Also, we want to update our notion of XactLastRecEnd based on worker
560 * feedback.
561 */
562 void
WaitForParallelWorkersToFinish(ParallelContext * pcxt)563 WaitForParallelWorkersToFinish(ParallelContext *pcxt)
564 {
565 for (;;)
566 {
567 bool anyone_alive = false;
568 int nfinished = 0;
569 int i;
570
571 /*
572 * This will process any parallel messages that are pending, which may
573 * change the outcome of the loop that follows. It may also throw an
574 * error propagated from a worker.
575 */
576 CHECK_FOR_INTERRUPTS();
577
578 for (i = 0; i < pcxt->nworkers_launched; ++i)
579 {
580 /*
581 * If error_mqh is NULL, then the worker has already exited
582 * cleanly. If we have received a message through error_mqh from
583 * the worker, we know it started up cleanly, and therefore we're
584 * certain to be notified when it exits.
585 */
586 if (pcxt->worker[i].error_mqh == NULL)
587 ++nfinished;
588 else if (pcxt->any_message_received[i])
589 {
590 anyone_alive = true;
591 break;
592 }
593 }
594
595 if (!anyone_alive)
596 {
597 /* If all workers are known to have finished, we're done. */
598 if (nfinished >= pcxt->nworkers_launched)
599 {
600 Assert(nfinished == pcxt->nworkers_launched);
601 break;
602 }
603
604 /*
605 * We didn't detect any living workers, but not all workers are
606 * known to have exited cleanly. Either not all workers have
607 * launched yet, or maybe some of them failed to start or
608 * terminated abnormally.
609 */
610 for (i = 0; i < pcxt->nworkers_launched; ++i)
611 {
612 pid_t pid;
613 shm_mq *mq;
614
615 /*
616 * If the worker is BGWH_NOT_YET_STARTED or BGWH_STARTED, we
617 * should just keep waiting. If it is BGWH_STOPPED, then
618 * further investigation is needed.
619 */
620 if (pcxt->worker[i].error_mqh == NULL ||
621 pcxt->worker[i].bgwhandle == NULL ||
622 GetBackgroundWorkerPid(pcxt->worker[i].bgwhandle,
623 &pid) != BGWH_STOPPED)
624 continue;
625
626 /*
627 * Check whether the worker ended up stopped without ever
628 * attaching to the error queue. If so, the postmaster was
629 * unable to fork the worker or it exited without initializing
630 * properly. We must throw an error, since the caller may
631 * have been expecting the worker to do some work before
632 * exiting.
633 */
634 mq = shm_mq_get_queue(pcxt->worker[i].error_mqh);
635 if (shm_mq_get_sender(mq) == NULL)
636 ereport(ERROR,
637 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
638 errmsg("parallel worker failed to initialize"),
639 errhint("More details may be available in the server log.")));
640
641 /*
642 * The worker is stopped, but is attached to the error queue.
643 * Unless there's a bug somewhere, this will only happen when
644 * the worker writes messages and terminates after the
645 * CHECK_FOR_INTERRUPTS() near the top of this function and
646 * before the call to GetBackgroundWorkerPid(). In that case,
647 * or latch should have been set as well and the right things
648 * will happen on the next pass through the loop.
649 */
650 }
651 }
652
653 WaitLatch(&MyProc->procLatch, WL_LATCH_SET, -1);
654 ResetLatch(&MyProc->procLatch);
655 }
656
657 if (pcxt->toc != NULL)
658 {
659 FixedParallelState *fps;
660
661 fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED);
662 if (fps->last_xlog_end > XactLastRecEnd)
663 XactLastRecEnd = fps->last_xlog_end;
664 }
665 }
666
667 /*
668 * Wait for all workers to exit.
669 *
670 * This function ensures that workers have been completely shutdown. The
671 * difference between WaitForParallelWorkersToFinish and this function is
672 * that former just ensures that last message sent by worker backend is
673 * received by master backend whereas this ensures the complete shutdown.
674 */
675 static void
WaitForParallelWorkersToExit(ParallelContext * pcxt)676 WaitForParallelWorkersToExit(ParallelContext *pcxt)
677 {
678 int i;
679
680 /* Wait until the workers actually die. */
681 for (i = 0; i < pcxt->nworkers_launched; ++i)
682 {
683 BgwHandleStatus status;
684
685 if (pcxt->worker == NULL || pcxt->worker[i].bgwhandle == NULL)
686 continue;
687
688 status = WaitForBackgroundWorkerShutdown(pcxt->worker[i].bgwhandle);
689
690 /*
691 * If the postmaster kicked the bucket, we have no chance of cleaning
692 * up safely -- we won't be able to tell when our workers are actually
693 * dead. This doesn't necessitate a PANIC since they will all abort
694 * eventually, but we can't safely continue this session.
695 */
696 if (status == BGWH_POSTMASTER_DIED)
697 ereport(FATAL,
698 (errcode(ERRCODE_ADMIN_SHUTDOWN),
699 errmsg("postmaster exited during a parallel transaction")));
700
701 /* Release memory. */
702 pfree(pcxt->worker[i].bgwhandle);
703 pcxt->worker[i].bgwhandle = NULL;
704 }
705 }
706
707 /*
708 * Destroy a parallel context.
709 *
710 * If expecting a clean exit, you should use WaitForParallelWorkersToFinish()
711 * first, before calling this function. When this function is invoked, any
712 * remaining workers are forcibly killed; the dynamic shared memory segment
713 * is unmapped; and we then wait (uninterruptibly) for the workers to exit.
714 */
715 void
DestroyParallelContext(ParallelContext * pcxt)716 DestroyParallelContext(ParallelContext *pcxt)
717 {
718 int i;
719
720 /*
721 * Be careful about order of operations here! We remove the parallel
722 * context from the list before we do anything else; otherwise, if an
723 * error occurs during a subsequent step, we might try to nuke it again
724 * from AtEOXact_Parallel or AtEOSubXact_Parallel.
725 */
726 dlist_delete(&pcxt->node);
727
728 /* Kill each worker in turn, and forget their error queues. */
729 if (pcxt->worker != NULL)
730 {
731 for (i = 0; i < pcxt->nworkers_launched; ++i)
732 {
733 if (pcxt->worker[i].error_mqh != NULL)
734 {
735 TerminateBackgroundWorker(pcxt->worker[i].bgwhandle);
736
737 pfree(pcxt->worker[i].error_mqh);
738 pcxt->worker[i].error_mqh = NULL;
739 }
740 }
741 }
742
743 /*
744 * If we have allocated a shared memory segment, detach it. This will
745 * implicitly detach the error queues, and any other shared memory queues,
746 * stored there.
747 */
748 if (pcxt->seg != NULL)
749 {
750 dsm_detach(pcxt->seg);
751 pcxt->seg = NULL;
752 }
753
754 /*
755 * If this parallel context is actually in backend-private memory rather
756 * than shared memory, free that memory instead.
757 */
758 if (pcxt->private_memory != NULL)
759 {
760 pfree(pcxt->private_memory);
761 pcxt->private_memory = NULL;
762 }
763
764 /*
765 * We can't finish transaction commit or abort until all of the workers
766 * have exited. This means, in particular, that we can't respond to
767 * interrupts at this stage.
768 */
769 HOLD_INTERRUPTS();
770 WaitForParallelWorkersToExit(pcxt);
771 RESUME_INTERRUPTS();
772
773 /* Free the worker array itself. */
774 if (pcxt->worker != NULL)
775 {
776 pfree(pcxt->worker);
777 pcxt->worker = NULL;
778 }
779
780 /* Free memory. */
781 if (pcxt->library_name)
782 pfree(pcxt->library_name);
783 if (pcxt->function_name)
784 pfree(pcxt->function_name);
785 pfree(pcxt);
786 }
787
788 /*
789 * Are there any parallel contexts currently active?
790 */
791 bool
ParallelContextActive(void)792 ParallelContextActive(void)
793 {
794 return !dlist_is_empty(&pcxt_list);
795 }
796
797 /*
798 * Handle receipt of an interrupt indicating a parallel worker message.
799 *
800 * Note: this is called within a signal handler! All we can do is set
801 * a flag that will cause the next CHECK_FOR_INTERRUPTS() to invoke
802 * HandleParallelMessages().
803 */
804 void
HandleParallelMessageInterrupt(void)805 HandleParallelMessageInterrupt(void)
806 {
807 InterruptPending = true;
808 ParallelMessagePending = true;
809 SetLatch(MyLatch);
810 }
811
812 /*
813 * Handle any queued protocol messages received from parallel workers.
814 */
815 void
HandleParallelMessages(void)816 HandleParallelMessages(void)
817 {
818 dlist_iter iter;
819 MemoryContext oldcontext;
820
821 static MemoryContext hpm_context = NULL;
822
823 /*
824 * This is invoked from ProcessInterrupts(), and since some of the
825 * functions it calls contain CHECK_FOR_INTERRUPTS(), there is a potential
826 * for recursive calls if more signals are received while this runs. It's
827 * unclear that recursive entry would be safe, and it doesn't seem useful
828 * even if it is safe, so let's block interrupts until done.
829 */
830 HOLD_INTERRUPTS();
831
832 /*
833 * Moreover, CurrentMemoryContext might be pointing almost anywhere. We
834 * don't want to risk leaking data into long-lived contexts, so let's do
835 * our work here in a private context that we can reset on each use.
836 */
837 if (hpm_context == NULL) /* first time through? */
838 hpm_context = AllocSetContextCreate(TopMemoryContext,
839 "HandleParallelMessages",
840 ALLOCSET_DEFAULT_SIZES);
841 else
842 MemoryContextReset(hpm_context);
843
844 oldcontext = MemoryContextSwitchTo(hpm_context);
845
846 /* OK to process messages. Reset the flag saying there are more to do. */
847 ParallelMessagePending = false;
848
849 dlist_foreach(iter, &pcxt_list)
850 {
851 ParallelContext *pcxt;
852 int i;
853
854 pcxt = dlist_container(ParallelContext, node, iter.cur);
855 if (pcxt->worker == NULL)
856 continue;
857
858 for (i = 0; i < pcxt->nworkers_launched; ++i)
859 {
860 /*
861 * Read as many messages as we can from each worker, but stop when
862 * either (1) the worker's error queue goes away, which can happen
863 * if we receive a Terminate message from the worker; or (2) no
864 * more messages can be read from the worker without blocking.
865 */
866 while (pcxt->worker[i].error_mqh != NULL)
867 {
868 shm_mq_result res;
869 Size nbytes;
870 void *data;
871
872 res = shm_mq_receive(pcxt->worker[i].error_mqh, &nbytes,
873 &data, true);
874 if (res == SHM_MQ_WOULD_BLOCK)
875 break;
876 else if (res == SHM_MQ_SUCCESS)
877 {
878 StringInfoData msg;
879
880 initStringInfo(&msg);
881 appendBinaryStringInfo(&msg, data, nbytes);
882 HandleParallelMessage(pcxt, i, &msg);
883 pfree(msg.data);
884 }
885 else
886 ereport(ERROR,
887 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
888 errmsg("lost connection to parallel worker")));
889 }
890 }
891 }
892
893 MemoryContextSwitchTo(oldcontext);
894
895 /* Might as well clear the context on our way out */
896 MemoryContextReset(hpm_context);
897
898 RESUME_INTERRUPTS();
899 }
900
901 /*
902 * Handle a single protocol message received from a single parallel worker.
903 */
904 static void
HandleParallelMessage(ParallelContext * pcxt,int i,StringInfo msg)905 HandleParallelMessage(ParallelContext *pcxt, int i, StringInfo msg)
906 {
907 char msgtype;
908
909 if (pcxt->any_message_received != NULL)
910 pcxt->any_message_received[i] = true;
911
912 msgtype = pq_getmsgbyte(msg);
913
914 switch (msgtype)
915 {
916 case 'K': /* BackendKeyData */
917 {
918 int32 pid = pq_getmsgint(msg, 4);
919
920 (void) pq_getmsgint(msg, 4); /* discard cancel key */
921 (void) pq_getmsgend(msg);
922 pcxt->worker[i].pid = pid;
923 break;
924 }
925
926 case 'E': /* ErrorResponse */
927 case 'N': /* NoticeResponse */
928 {
929 ErrorData edata;
930 ErrorContextCallback *save_error_context_stack;
931
932 /* Parse ErrorResponse or NoticeResponse. */
933 pq_parse_errornotice(msg, &edata);
934
935 /* Death of a worker isn't enough justification for suicide. */
936 edata.elevel = Min(edata.elevel, ERROR);
937
938 /*
939 * If desired, add a context line to show that this is a
940 * message propagated from a parallel worker. Otherwise, it
941 * can sometimes be confusing to understand what actually
942 * happened. (We don't do this in FORCE_PARALLEL_REGRESS mode
943 * because it causes test-result instability depending on
944 * whether a parallel worker is actually used or not.)
945 */
946 if (force_parallel_mode != FORCE_PARALLEL_REGRESS)
947 {
948 if (edata.context)
949 edata.context = psprintf("%s\n%s", edata.context,
950 _("parallel worker"));
951 else
952 edata.context = pstrdup(_("parallel worker"));
953 }
954
955 /*
956 * Context beyond that should use the error context callbacks
957 * that were in effect when the ParallelContext was created,
958 * not the current ones.
959 */
960 save_error_context_stack = error_context_stack;
961 error_context_stack = pcxt->error_context_stack;
962
963 /* Rethrow error or print notice. */
964 ThrowErrorData(&edata);
965
966 /* Not an error, so restore previous context stack. */
967 error_context_stack = save_error_context_stack;
968
969 break;
970 }
971
972 case 'A': /* NotifyResponse */
973 {
974 /* Propagate NotifyResponse. */
975 int32 pid;
976 const char *channel;
977 const char *payload;
978
979 pid = pq_getmsgint(msg, 4);
980 channel = pq_getmsgrawstring(msg);
981 payload = pq_getmsgrawstring(msg);
982 pq_endmessage(msg);
983
984 NotifyMyFrontEnd(channel, payload, pid);
985
986 break;
987 }
988
989 case 'X': /* Terminate, indicating clean exit */
990 {
991 pfree(pcxt->worker[i].error_mqh);
992 pcxt->worker[i].error_mqh = NULL;
993 break;
994 }
995
996 default:
997 {
998 elog(ERROR, "unrecognized message type received from parallel worker: %c (message length %d bytes)",
999 msgtype, msg->len);
1000 }
1001 }
1002 }
1003
1004 /*
1005 * End-of-subtransaction cleanup for parallel contexts.
1006 *
1007 * Currently, it's forbidden to enter or leave a subtransaction while
1008 * parallel mode is in effect, so we could just blow away everything. But
1009 * we may want to relax that restriction in the future, so this code
1010 * contemplates that there may be multiple subtransaction IDs in pcxt_list.
1011 */
1012 void
AtEOSubXact_Parallel(bool isCommit,SubTransactionId mySubId)1013 AtEOSubXact_Parallel(bool isCommit, SubTransactionId mySubId)
1014 {
1015 while (!dlist_is_empty(&pcxt_list))
1016 {
1017 ParallelContext *pcxt;
1018
1019 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
1020 if (pcxt->subid != mySubId)
1021 break;
1022 if (isCommit)
1023 elog(WARNING, "leaked parallel context");
1024 DestroyParallelContext(pcxt);
1025 }
1026 }
1027
1028 /*
1029 * End-of-transaction cleanup for parallel contexts.
1030 */
1031 void
AtEOXact_Parallel(bool isCommit)1032 AtEOXact_Parallel(bool isCommit)
1033 {
1034 while (!dlist_is_empty(&pcxt_list))
1035 {
1036 ParallelContext *pcxt;
1037
1038 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
1039 if (isCommit)
1040 elog(WARNING, "leaked parallel context");
1041 DestroyParallelContext(pcxt);
1042 }
1043 }
1044
1045 /*
1046 * Main entrypoint for parallel workers.
1047 */
1048 void
ParallelWorkerMain(Datum main_arg)1049 ParallelWorkerMain(Datum main_arg)
1050 {
1051 dsm_segment *seg;
1052 shm_toc *toc;
1053 FixedParallelState *fps;
1054 char *error_queue_space;
1055 shm_mq *mq;
1056 shm_mq_handle *mqh;
1057 char *libraryspace;
1058 char *entrypointstate;
1059 parallel_worker_main_type entrypt;
1060 char *gucspace;
1061 char *combocidspace;
1062 char *tsnapspace;
1063 char *asnapspace;
1064 char *tstatespace;
1065 StringInfoData msgbuf;
1066
1067 /* Set flag to indicate that we're initializing a parallel worker. */
1068 InitializingParallelWorker = true;
1069
1070 /* Establish signal handlers. */
1071 pqsignal(SIGTERM, die);
1072 BackgroundWorkerUnblockSignals();
1073
1074 /* Determine and set our parallel worker number. */
1075 Assert(ParallelWorkerNumber == -1);
1076 memcpy(&ParallelWorkerNumber, MyBgworkerEntry->bgw_extra, sizeof(int));
1077
1078 /* Set up a memory context and resource owner. */
1079 Assert(CurrentResourceOwner == NULL);
1080 CurrentResourceOwner = ResourceOwnerCreate(NULL, "parallel toplevel");
1081 CurrentMemoryContext = AllocSetContextCreate(TopMemoryContext,
1082 "Parallel worker",
1083 ALLOCSET_DEFAULT_SIZES);
1084
1085 /*
1086 * Now that we have a resource owner, we can attach to the dynamic shared
1087 * memory segment and read the table of contents.
1088 */
1089 seg = dsm_attach(DatumGetUInt32(main_arg));
1090 if (seg == NULL)
1091 ereport(ERROR,
1092 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1093 errmsg("could not map dynamic shared memory segment")));
1094 toc = shm_toc_attach(PARALLEL_MAGIC, dsm_segment_address(seg));
1095 if (toc == NULL)
1096 ereport(ERROR,
1097 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1098 errmsg("invalid magic number in dynamic shared memory segment")));
1099
1100 /* Look up fixed parallel state. */
1101 fps = shm_toc_lookup(toc, PARALLEL_KEY_FIXED);
1102 Assert(fps != NULL);
1103 MyFixedParallelState = fps;
1104
1105 /* Arrange to signal the leader if we exit. */
1106 ParallelMasterPid = fps->parallel_master_pid;
1107 ParallelMasterBackendId = fps->parallel_master_backend_id;
1108 on_shmem_exit(ParallelWorkerShutdown, (Datum) 0);
1109
1110 /*
1111 * Now we can find and attach to the error queue provided for us. That's
1112 * good, because until we do that, any errors that happen here will not be
1113 * reported back to the process that requested that this worker be
1114 * launched.
1115 */
1116 error_queue_space = shm_toc_lookup(toc, PARALLEL_KEY_ERROR_QUEUE);
1117 mq = (shm_mq *) (error_queue_space +
1118 ParallelWorkerNumber * PARALLEL_ERROR_QUEUE_SIZE);
1119 shm_mq_set_sender(mq, MyProc);
1120 mqh = shm_mq_attach(mq, seg, NULL);
1121 pq_redirect_to_shm_mq(seg, mqh);
1122 pq_set_parallel_master(fps->parallel_master_pid,
1123 fps->parallel_master_backend_id);
1124
1125 /*
1126 * Send a BackendKeyData message to the process that initiated parallelism
1127 * so that it has access to our PID before it receives any other messages
1128 * from us. Our cancel key is sent, too, since that's the way the
1129 * protocol message is defined, but it won't actually be used for anything
1130 * in this case.
1131 */
1132 pq_beginmessage(&msgbuf, 'K');
1133 pq_sendint(&msgbuf, (int32) MyProcPid, sizeof(int32));
1134 pq_sendint(&msgbuf, (int32) MyCancelKey, sizeof(int32));
1135 pq_endmessage(&msgbuf);
1136
1137 /*
1138 * Hooray! Primary initialization is complete. Now, we need to set up our
1139 * backend-local state to match the original backend.
1140 */
1141
1142 /*
1143 * Join locking group. We must do this before anything that could try to
1144 * acquire a heavyweight lock, because any heavyweight locks acquired to
1145 * this point could block either directly against the parallel group
1146 * leader or against some process which in turn waits for a lock that
1147 * conflicts with the parallel group leader, causing an undetected
1148 * deadlock. (If we can't join the lock group, the leader has gone away,
1149 * so just exit quietly.)
1150 */
1151 if (!BecomeLockGroupMember(fps->parallel_master_pgproc,
1152 fps->parallel_master_pid))
1153 return;
1154
1155 /*
1156 * Restore transaction and statement start-time timestamps. This must
1157 * happen before anything that would start a transaction, else asserts in
1158 * xact.c will fire.
1159 */
1160 SetParallelStartTimestamps(fps->xact_ts, fps->stmt_ts);
1161
1162 /*
1163 * Identify the entry point to be called. In theory this could result in
1164 * loading an additional library, though most likely the entry point is in
1165 * the core backend or in a library we just loaded.
1166 */
1167 entrypointstate = shm_toc_lookup(toc, PARALLEL_KEY_ENTRYPOINT);
1168 if (entrypointstate != NULL)
1169 {
1170 char *library_name;
1171 char *function_name;
1172
1173 library_name = entrypointstate;
1174 function_name = entrypointstate + strlen(library_name) + 1;
1175
1176 entrypt = LookupParallelWorkerFunction(library_name, function_name);
1177 }
1178 else
1179 entrypt = fps->entrypoint;
1180
1181 /* Restore database connection. */
1182 BackgroundWorkerInitializeConnectionByOid(fps->database_id,
1183 fps->authenticated_user_id);
1184
1185 /*
1186 * Set the client encoding to the database encoding, since that is what
1187 * the leader will expect.
1188 */
1189 SetClientEncoding(GetDatabaseEncoding());
1190
1191 /*
1192 * Load libraries that were loaded by original backend. We want to do
1193 * this before restoring GUCs, because the libraries might define custom
1194 * variables.
1195 */
1196 libraryspace = shm_toc_lookup(toc, PARALLEL_KEY_LIBRARY);
1197 Assert(libraryspace != NULL);
1198 StartTransactionCommand();
1199 RestoreLibraryState(libraryspace);
1200
1201 /* Restore GUC values from launching backend. */
1202 gucspace = shm_toc_lookup(toc, PARALLEL_KEY_GUC);
1203 Assert(gucspace != NULL);
1204 RestoreGUCState(gucspace);
1205 CommitTransactionCommand();
1206
1207 /* Crank up a transaction state appropriate to a parallel worker. */
1208 tstatespace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_STATE);
1209 StartParallelWorkerTransaction(tstatespace);
1210
1211 /* Restore combo CID state. */
1212 combocidspace = shm_toc_lookup(toc, PARALLEL_KEY_COMBO_CID);
1213 Assert(combocidspace != NULL);
1214 RestoreComboCIDState(combocidspace);
1215
1216 /* Restore transaction snapshot. */
1217 tsnapspace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT);
1218 Assert(tsnapspace != NULL);
1219 RestoreTransactionSnapshot(RestoreSnapshot(tsnapspace),
1220 fps->parallel_master_pgproc);
1221
1222 /* Restore active snapshot. */
1223 asnapspace = shm_toc_lookup(toc, PARALLEL_KEY_ACTIVE_SNAPSHOT);
1224 Assert(asnapspace != NULL);
1225 PushActiveSnapshot(RestoreSnapshot(asnapspace));
1226
1227 /*
1228 * We've changed which tuples we can see, and must therefore invalidate
1229 * system caches.
1230 */
1231 InvalidateSystemCaches();
1232
1233 /*
1234 * Restore current role id. Skip verifying whether session user is
1235 * allowed to become this role and blindly restore the leader's state for
1236 * current role.
1237 */
1238 SetCurrentRoleId(fps->outer_user_id, fps->is_superuser);
1239
1240 /* Restore user ID and security context. */
1241 SetUserIdAndSecContext(fps->current_user_id, fps->sec_context);
1242
1243 /* Restore temp-namespace state to ensure search path matches leader's. */
1244 SetTempNamespaceState(fps->temp_namespace_id,
1245 fps->temp_toast_namespace_id);
1246
1247 /*
1248 * We've initialized all of our state now; nothing should change
1249 * hereafter.
1250 */
1251 InitializingParallelWorker = false;
1252 EnterParallelMode();
1253
1254 /*
1255 * Time to do the real work: invoke the caller-supplied code.
1256 *
1257 * If you get a crash at this line, try using
1258 * CreateParallelContextForExternalFunction instead of
1259 * CreateParallelContext.
1260 */
1261 entrypt(seg, toc);
1262
1263 /* Must exit parallel mode to pop active snapshot. */
1264 ExitParallelMode();
1265
1266 /* Must pop active snapshot so resowner.c doesn't complain. */
1267 PopActiveSnapshot();
1268
1269 /* Shut down the parallel-worker transaction. */
1270 EndParallelWorkerTransaction();
1271
1272 /* Report success. */
1273 pq_putmessage('X', NULL, 0);
1274 }
1275
1276 /*
1277 * Update shared memory with the ending location of the last WAL record we
1278 * wrote, if it's greater than the value already stored there.
1279 */
1280 void
ParallelWorkerReportLastRecEnd(XLogRecPtr last_xlog_end)1281 ParallelWorkerReportLastRecEnd(XLogRecPtr last_xlog_end)
1282 {
1283 FixedParallelState *fps = MyFixedParallelState;
1284
1285 Assert(fps != NULL);
1286 SpinLockAcquire(&fps->mutex);
1287 if (fps->last_xlog_end < last_xlog_end)
1288 fps->last_xlog_end = last_xlog_end;
1289 SpinLockRelease(&fps->mutex);
1290 }
1291
1292 /*
1293 * Make sure the leader tries to read from our error queue one more time.
1294 * This guards against the case where we exit uncleanly without sending an
1295 * ErrorResponse to the leader, for example because some code calls proc_exit
1296 * directly.
1297 */
1298 static void
ParallelWorkerShutdown(int code,Datum arg)1299 ParallelWorkerShutdown(int code, Datum arg)
1300 {
1301 SendProcSignal(ParallelMasterPid,
1302 PROCSIG_PARALLEL_MESSAGE,
1303 ParallelMasterBackendId);
1304 }
1305
1306 /*
1307 * Look up (and possibly load) a parallel worker entry point function.
1308 *
1309 * For functions contained in the core code, we use library name "postgres"
1310 * and consult the InternalParallelWorkers array. External functions are
1311 * looked up, and loaded if necessary, using load_external_function().
1312 *
1313 * The point of this is to pass function names as strings across process
1314 * boundaries. We can't pass actual function addresses because of the
1315 * possibility that the function has been loaded at a different address
1316 * in a different process. This is obviously a hazard for functions in
1317 * loadable libraries, but it can happen even for functions in the core code
1318 * on platforms using EXEC_BACKEND (e.g., Windows).
1319 *
1320 * At some point it might be worthwhile to get rid of InternalParallelWorkers[]
1321 * in favor of applying load_external_function() for core functions too;
1322 * but that raises portability issues that are not worth addressing now.
1323 */
1324 static parallel_worker_main_type
LookupParallelWorkerFunction(char * libraryname,char * funcname)1325 LookupParallelWorkerFunction(char *libraryname, char *funcname)
1326 {
1327 /*
1328 * If the function is to be loaded from postgres itself, search the
1329 * InternalParallelWorkers array.
1330 */
1331 if (strcmp(libraryname, "postgres") == 0)
1332 {
1333 int i;
1334
1335 for (i = 0; i < lengthof(InternalParallelWorkers); i++)
1336 {
1337 if (strcmp(InternalParallelWorkers[i].fn_name, funcname) == 0)
1338 return InternalParallelWorkers[i].fn_addr;
1339 }
1340
1341 /* We can only reach this by programming error. */
1342 elog(ERROR, "internal function \"%s\" not found", funcname);
1343 }
1344
1345 /* Otherwise load from external library. */
1346 return (parallel_worker_main_type)
1347 load_external_function(libraryname, funcname, true, NULL);
1348 }
1349