1 /*-------------------------------------------------------------------------
2 *
3 * parallel.c
4 * Infrastructure for launching parallel workers
5 *
6 * Portions Copyright (c) 1996-2017, 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 "pgstat.h"
29 #include "storage/ipc.h"
30 #include "storage/sinval.h"
31 #include "storage/spin.h"
32 #include "tcop/tcopprot.h"
33 #include "utils/combocid.h"
34 #include "utils/guc.h"
35 #include "utils/inval.h"
36 #include "utils/memutils.h"
37 #include "utils/resowner.h"
38 #include "utils/snapmgr.h"
39
40
41 /*
42 * We don't want to waste a lot of memory on an error queue which, most of
43 * the time, will process only a handful of small messages. However, it is
44 * desirable to make it large enough that a typical ErrorResponse can be sent
45 * without blocking. That way, a worker that errors out can write the whole
46 * message into the queue and terminate without waiting for the user backend.
47 */
48 #define PARALLEL_ERROR_QUEUE_SIZE 16384
49
50 /* Magic number for parallel context TOC. */
51 #define PARALLEL_MAGIC 0x50477c7c
52
53 /*
54 * Magic numbers for parallel state sharing. Higher-level code should use
55 * smaller values, leaving these very large ones for use by this module.
56 */
57 #define PARALLEL_KEY_FIXED UINT64CONST(0xFFFFFFFFFFFF0001)
58 #define PARALLEL_KEY_ERROR_QUEUE UINT64CONST(0xFFFFFFFFFFFF0002)
59 #define PARALLEL_KEY_LIBRARY UINT64CONST(0xFFFFFFFFFFFF0003)
60 #define PARALLEL_KEY_GUC UINT64CONST(0xFFFFFFFFFFFF0004)
61 #define PARALLEL_KEY_COMBO_CID UINT64CONST(0xFFFFFFFFFFFF0005)
62 #define PARALLEL_KEY_TRANSACTION_SNAPSHOT UINT64CONST(0xFFFFFFFFFFFF0006)
63 #define PARALLEL_KEY_ACTIVE_SNAPSHOT UINT64CONST(0xFFFFFFFFFFFF0007)
64 #define PARALLEL_KEY_TRANSACTION_STATE UINT64CONST(0xFFFFFFFFFFFF0008)
65 #define PARALLEL_KEY_ENTRYPOINT UINT64CONST(0xFFFFFFFFFFFF0009)
66
67 /* Fixed-size parallel state. */
68 typedef struct FixedParallelState
69 {
70 /* Fixed-size state that workers must restore. */
71 Oid database_id;
72 Oid authenticated_user_id;
73 Oid current_user_id;
74 Oid outer_user_id;
75 Oid temp_namespace_id;
76 Oid temp_toast_namespace_id;
77 int sec_context;
78 bool is_superuser;
79 PGPROC *parallel_master_pgproc;
80 pid_t parallel_master_pid;
81 BackendId parallel_master_backend_id;
82 TimestampTz xact_ts;
83 TimestampTz stmt_ts;
84
85 /* Mutex protects remaining fields. */
86 slock_t mutex;
87
88 /* Maximum XactLastRecEnd of any worker. */
89 XLogRecPtr last_xlog_end;
90 } FixedParallelState;
91
92 /*
93 * Our parallel worker number. We initialize this to -1, meaning that we are
94 * not a parallel worker. In parallel workers, it will be set to a value >= 0
95 * and < the number of workers before any user code is invoked; each parallel
96 * worker will get a different parallel worker number.
97 */
98 int ParallelWorkerNumber = -1;
99
100 /* Is there a parallel message pending which we need to receive? */
101 volatile bool ParallelMessagePending = false;
102
103 /* Are we initializing a parallel worker? */
104 bool InitializingParallelWorker = false;
105
106 /* Pointer to our fixed parallel state. */
107 static FixedParallelState *MyFixedParallelState;
108
109 /* List of active parallel contexts. */
110 static dlist_head pcxt_list = DLIST_STATIC_INIT(pcxt_list);
111
112 /* Backend-local copy of data from FixedParallelState. */
113 static pid_t ParallelMasterPid;
114
115 /*
116 * List of internal parallel worker entry points. We need this for
117 * reasons explained in LookupParallelWorkerFunction(), below.
118 */
119 static const struct
120 {
121 const char *fn_name;
122 parallel_worker_main_type fn_addr;
123 } InternalParallelWorkers[] =
124
125 {
126 {
127 "ParallelQueryMain", ParallelQueryMain
128 }
129 };
130
131 /* Private functions. */
132 static void HandleParallelMessage(ParallelContext *pcxt, int i, StringInfo msg);
133 static void WaitForParallelWorkersToExit(ParallelContext *pcxt);
134 static parallel_worker_main_type LookupParallelWorkerFunction(const char *libraryname, const char *funcname);
135 static void ParallelWorkerShutdown(int code, Datum arg);
136
137
138 /*
139 * Establish a new parallel context. This should be done after entering
140 * parallel mode, and (unless there is an error) the context should be
141 * destroyed before exiting the current subtransaction.
142 */
143 ParallelContext *
CreateParallelContext(const char * library_name,const char * function_name,int nworkers)144 CreateParallelContext(const char *library_name, const char *function_name,
145 int nworkers)
146 {
147 MemoryContext oldcontext;
148 ParallelContext *pcxt;
149
150 /* It is unsafe to create a parallel context if not in parallel mode. */
151 Assert(IsInParallelMode());
152
153 /* Number of workers should be non-negative. */
154 Assert(nworkers >= 0);
155
156 /*
157 * If dynamic shared memory is not available, we won't be able to use
158 * background workers.
159 */
160 if (dynamic_shared_memory_type == DSM_IMPL_NONE)
161 nworkers = 0;
162
163 /*
164 * If we are running under serializable isolation, we can't use parallel
165 * workers, at least not until somebody enhances that mechanism to be
166 * parallel-aware.
167 */
168 if (IsolationIsSerializable())
169 nworkers = 0;
170
171 /* We might be running in a short-lived memory context. */
172 oldcontext = MemoryContextSwitchTo(TopTransactionContext);
173
174 /* Initialize a new ParallelContext. */
175 pcxt = palloc0(sizeof(ParallelContext));
176 pcxt->subid = GetCurrentSubTransactionId();
177 pcxt->nworkers = nworkers;
178 pcxt->library_name = pstrdup(library_name);
179 pcxt->function_name = pstrdup(function_name);
180 pcxt->error_context_stack = error_context_stack;
181 shm_toc_initialize_estimator(&pcxt->estimator);
182 dlist_push_head(&pcxt_list, &pcxt->node);
183
184 /* Restore previous memory context. */
185 MemoryContextSwitchTo(oldcontext);
186
187 return pcxt;
188 }
189
190 /*
191 * Establish the dynamic shared memory segment for a parallel context and
192 * copy state and other bookkeeping information that will be needed by
193 * parallel workers into it.
194 */
195 void
InitializeParallelDSM(ParallelContext * pcxt)196 InitializeParallelDSM(ParallelContext *pcxt)
197 {
198 MemoryContext oldcontext;
199 Size library_len = 0;
200 Size guc_len = 0;
201 Size combocidlen = 0;
202 Size tsnaplen = 0;
203 Size asnaplen = 0;
204 Size tstatelen = 0;
205 Size segsize = 0;
206 int i;
207 FixedParallelState *fps;
208 Snapshot transaction_snapshot = GetTransactionSnapshot();
209 Snapshot active_snapshot = GetActiveSnapshot();
210
211 /* We might be running in a very short-lived memory context. */
212 oldcontext = MemoryContextSwitchTo(TopTransactionContext);
213
214 /* Allow space to store the fixed-size parallel state. */
215 shm_toc_estimate_chunk(&pcxt->estimator, sizeof(FixedParallelState));
216 shm_toc_estimate_keys(&pcxt->estimator, 1);
217
218 /*
219 * Normally, the user will have requested at least one worker process, but
220 * if by chance they have not, we can skip a bunch of things here.
221 */
222 if (pcxt->nworkers > 0)
223 {
224 /* Estimate space for various kinds of state sharing. */
225 library_len = EstimateLibraryStateSpace();
226 shm_toc_estimate_chunk(&pcxt->estimator, library_len);
227 guc_len = EstimateGUCStateSpace();
228 shm_toc_estimate_chunk(&pcxt->estimator, guc_len);
229 combocidlen = EstimateComboCIDStateSpace();
230 shm_toc_estimate_chunk(&pcxt->estimator, combocidlen);
231 if (IsolationUsesXactSnapshot())
232 {
233 tsnaplen = EstimateSnapshotSpace(transaction_snapshot);
234 shm_toc_estimate_chunk(&pcxt->estimator, tsnaplen);
235 }
236 asnaplen = EstimateSnapshotSpace(active_snapshot);
237 shm_toc_estimate_chunk(&pcxt->estimator, asnaplen);
238 tstatelen = EstimateTransactionStateSpace();
239 shm_toc_estimate_chunk(&pcxt->estimator, tstatelen);
240 /* If you add more chunks here, you probably need to add keys. */
241 shm_toc_estimate_keys(&pcxt->estimator, 6);
242
243 /* Estimate space need for error queues. */
244 StaticAssertStmt(BUFFERALIGN(PARALLEL_ERROR_QUEUE_SIZE) ==
245 PARALLEL_ERROR_QUEUE_SIZE,
246 "parallel error queue size not buffer-aligned");
247 shm_toc_estimate_chunk(&pcxt->estimator,
248 mul_size(PARALLEL_ERROR_QUEUE_SIZE,
249 pcxt->nworkers));
250 shm_toc_estimate_keys(&pcxt->estimator, 1);
251
252 /* Estimate how much we'll need for the entrypoint info. */
253 shm_toc_estimate_chunk(&pcxt->estimator, strlen(pcxt->library_name) +
254 strlen(pcxt->function_name) + 2);
255 shm_toc_estimate_keys(&pcxt->estimator, 1);
256 }
257
258 /*
259 * Create DSM and initialize with new table of contents. But if the user
260 * didn't request any workers, then don't bother creating a dynamic shared
261 * memory segment; instead, just use backend-private memory.
262 *
263 * Also, if we can't create a dynamic shared memory segment because the
264 * maximum number of segments have already been created, then fall back to
265 * backend-private memory, and plan not to use any workers. We hope this
266 * won't happen very often, but it's better to abandon the use of
267 * parallelism than to fail outright.
268 */
269 segsize = shm_toc_estimate(&pcxt->estimator);
270 if (pcxt->nworkers > 0)
271 pcxt->seg = dsm_create(segsize, DSM_CREATE_NULL_IF_MAXSEGMENTS);
272 if (pcxt->seg != NULL)
273 pcxt->toc = shm_toc_create(PARALLEL_MAGIC,
274 dsm_segment_address(pcxt->seg),
275 segsize);
276 else
277 {
278 pcxt->nworkers = 0;
279 pcxt->private_memory = MemoryContextAlloc(TopMemoryContext, segsize);
280 pcxt->toc = shm_toc_create(PARALLEL_MAGIC, pcxt->private_memory,
281 segsize);
282 }
283
284 /* Initialize fixed-size state in shared memory. */
285 fps = (FixedParallelState *)
286 shm_toc_allocate(pcxt->toc, sizeof(FixedParallelState));
287 fps->database_id = MyDatabaseId;
288 fps->authenticated_user_id = GetAuthenticatedUserId();
289 fps->outer_user_id = GetCurrentRoleId();
290 fps->is_superuser = session_auth_is_superuser;
291 GetUserIdAndSecContext(&fps->current_user_id, &fps->sec_context);
292 GetTempNamespaceState(&fps->temp_namespace_id,
293 &fps->temp_toast_namespace_id);
294 fps->parallel_master_pgproc = MyProc;
295 fps->parallel_master_pid = MyProcPid;
296 fps->parallel_master_backend_id = MyBackendId;
297 fps->xact_ts = GetCurrentTransactionStartTimestamp();
298 fps->stmt_ts = GetCurrentStatementStartTimestamp();
299 SpinLockInit(&fps->mutex);
300 fps->last_xlog_end = 0;
301 shm_toc_insert(pcxt->toc, PARALLEL_KEY_FIXED, fps);
302
303 /* We can skip the rest of this if we're not budgeting for any workers. */
304 if (pcxt->nworkers > 0)
305 {
306 char *libraryspace;
307 char *gucspace;
308 char *combocidspace;
309 char *tsnapspace;
310 char *asnapspace;
311 char *tstatespace;
312 char *error_queue_space;
313 char *entrypointstate;
314 Size lnamelen;
315
316 /* Serialize shared libraries we have loaded. */
317 libraryspace = shm_toc_allocate(pcxt->toc, library_len);
318 SerializeLibraryState(library_len, libraryspace);
319 shm_toc_insert(pcxt->toc, PARALLEL_KEY_LIBRARY, libraryspace);
320
321 /* Serialize GUC settings. */
322 gucspace = shm_toc_allocate(pcxt->toc, guc_len);
323 SerializeGUCState(guc_len, gucspace);
324 shm_toc_insert(pcxt->toc, PARALLEL_KEY_GUC, gucspace);
325
326 /* Serialize combo CID state. */
327 combocidspace = shm_toc_allocate(pcxt->toc, combocidlen);
328 SerializeComboCIDState(combocidlen, combocidspace);
329 shm_toc_insert(pcxt->toc, PARALLEL_KEY_COMBO_CID, combocidspace);
330
331 /*
332 * Serialize the transaction snapshot if the transaction
333 * isolation-level uses a transaction snapshot.
334 */
335 if (IsolationUsesXactSnapshot())
336 {
337 tsnapspace = shm_toc_allocate(pcxt->toc, tsnaplen);
338 SerializeSnapshot(transaction_snapshot, tsnapspace);
339 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT,
340 tsnapspace);
341 }
342
343 /* Serialize the active snapshot. */
344 asnapspace = shm_toc_allocate(pcxt->toc, asnaplen);
345 SerializeSnapshot(active_snapshot, asnapspace);
346 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ACTIVE_SNAPSHOT, asnapspace);
347
348 /* Serialize transaction state. */
349 tstatespace = shm_toc_allocate(pcxt->toc, tstatelen);
350 SerializeTransactionState(tstatelen, tstatespace);
351 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_STATE, tstatespace);
352
353 /* Allocate space for worker information. */
354 pcxt->worker = palloc0(sizeof(ParallelWorkerInfo) * pcxt->nworkers);
355
356 /*
357 * Establish error queues in dynamic shared memory.
358 *
359 * These queues should be used only for transmitting ErrorResponse,
360 * NoticeResponse, and NotifyResponse protocol messages. Tuple data
361 * should be transmitted via separate (possibly larger?) queues.
362 */
363 error_queue_space =
364 shm_toc_allocate(pcxt->toc,
365 mul_size(PARALLEL_ERROR_QUEUE_SIZE,
366 pcxt->nworkers));
367 for (i = 0; i < pcxt->nworkers; ++i)
368 {
369 char *start;
370 shm_mq *mq;
371
372 start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
373 mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
374 shm_mq_set_receiver(mq, MyProc);
375 pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
376 }
377 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE, error_queue_space);
378
379 /*
380 * Serialize entrypoint information. It's unsafe to pass function
381 * pointers across processes, as the function pointer may be different
382 * in each process in EXEC_BACKEND builds, so we always pass library
383 * and function name. (We use library name "postgres" for functions
384 * in the core backend.)
385 */
386 lnamelen = strlen(pcxt->library_name);
387 entrypointstate = shm_toc_allocate(pcxt->toc, lnamelen +
388 strlen(pcxt->function_name) + 2);
389 strcpy(entrypointstate, pcxt->library_name);
390 strcpy(entrypointstate + lnamelen + 1, pcxt->function_name);
391 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ENTRYPOINT, entrypointstate);
392 }
393
394 /* Restore previous memory context. */
395 MemoryContextSwitchTo(oldcontext);
396 }
397
398 /*
399 * Reinitialize the dynamic shared memory segment for a parallel context such
400 * that we could launch workers for it again.
401 */
402 void
ReinitializeParallelDSM(ParallelContext * pcxt)403 ReinitializeParallelDSM(ParallelContext *pcxt)
404 {
405 FixedParallelState *fps;
406
407 /* Wait for any old workers to exit. */
408 if (pcxt->nworkers_launched > 0)
409 {
410 WaitForParallelWorkersToFinish(pcxt);
411 WaitForParallelWorkersToExit(pcxt);
412 pcxt->nworkers_launched = 0;
413 if (pcxt->any_message_received)
414 {
415 pfree(pcxt->any_message_received);
416 pcxt->any_message_received = NULL;
417 }
418 }
419
420 /* Reset a few bits of fixed parallel state to a clean state. */
421 fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED, false);
422 fps->last_xlog_end = 0;
423
424 /* Recreate error queues (if they exist). */
425 if (pcxt->nworkers > 0)
426 {
427 char *error_queue_space;
428 int i;
429
430 error_queue_space =
431 shm_toc_lookup(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE, false);
432 for (i = 0; i < pcxt->nworkers; ++i)
433 {
434 char *start;
435 shm_mq *mq;
436
437 start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
438 mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
439 shm_mq_set_receiver(mq, MyProc);
440 pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
441 }
442 }
443 }
444
445 /*
446 * Launch parallel workers.
447 */
448 void
LaunchParallelWorkers(ParallelContext * pcxt)449 LaunchParallelWorkers(ParallelContext *pcxt)
450 {
451 MemoryContext oldcontext;
452 BackgroundWorker worker;
453 int i;
454 bool any_registrations_failed = false;
455
456 /* Skip this if we have no workers. */
457 if (pcxt->nworkers == 0)
458 return;
459
460 /* We need to be a lock group leader. */
461 BecomeLockGroupLeader();
462
463 /* If we do have workers, we'd better have a DSM segment. */
464 Assert(pcxt->seg != NULL);
465
466 /* We might be running in a short-lived memory context. */
467 oldcontext = MemoryContextSwitchTo(TopTransactionContext);
468
469 /* Configure a worker. */
470 memset(&worker, 0, sizeof(worker));
471 snprintf(worker.bgw_name, BGW_MAXLEN, "parallel worker for PID %d",
472 MyProcPid);
473 worker.bgw_flags =
474 BGWORKER_SHMEM_ACCESS | BGWORKER_BACKEND_DATABASE_CONNECTION
475 | BGWORKER_CLASS_PARALLEL;
476 worker.bgw_start_time = BgWorkerStart_ConsistentState;
477 worker.bgw_restart_time = BGW_NEVER_RESTART;
478 sprintf(worker.bgw_library_name, "postgres");
479 sprintf(worker.bgw_function_name, "ParallelWorkerMain");
480 worker.bgw_main_arg = UInt32GetDatum(dsm_segment_handle(pcxt->seg));
481 worker.bgw_notify_pid = MyProcPid;
482
483 /*
484 * Start workers.
485 *
486 * The caller must be able to tolerate ending up with fewer workers than
487 * expected, so there is no need to throw an error here if registration
488 * fails. It wouldn't help much anyway, because registering the worker in
489 * no way guarantees that it will start up and initialize successfully.
490 */
491 for (i = 0; i < pcxt->nworkers; ++i)
492 {
493 memcpy(worker.bgw_extra, &i, sizeof(int));
494 if (!any_registrations_failed &&
495 RegisterDynamicBackgroundWorker(&worker,
496 &pcxt->worker[i].bgwhandle))
497 {
498 shm_mq_set_handle(pcxt->worker[i].error_mqh,
499 pcxt->worker[i].bgwhandle);
500 pcxt->nworkers_launched++;
501 }
502 else
503 {
504 /*
505 * If we weren't able to register the worker, then we've bumped up
506 * against the max_worker_processes limit, and future
507 * registrations will probably fail too, so arrange to skip them.
508 * But we still have to execute this code for the remaining slots
509 * to make sure that we forget about the error queues we budgeted
510 * for those workers. Otherwise, we'll wait for them to start,
511 * but they never will.
512 */
513 any_registrations_failed = true;
514 pcxt->worker[i].bgwhandle = NULL;
515 shm_mq_detach(pcxt->worker[i].error_mqh);
516 pcxt->worker[i].error_mqh = NULL;
517 }
518 }
519
520 /*
521 * Now that nworkers_launched has taken its final value, we can initialize
522 * any_message_received.
523 */
524 if (pcxt->nworkers_launched > 0)
525 pcxt->any_message_received =
526 palloc0(sizeof(bool) * pcxt->nworkers_launched);
527
528 /* Restore previous memory context. */
529 MemoryContextSwitchTo(oldcontext);
530 }
531
532 /*
533 * Wait for all workers to finish computing.
534 *
535 * Even if the parallel operation seems to have completed successfully, it's
536 * important to call this function afterwards. We must not miss any errors
537 * the workers may have thrown during the parallel operation, or any that they
538 * may yet throw while shutting down.
539 *
540 * Also, we want to update our notion of XactLastRecEnd based on worker
541 * feedback.
542 */
543 void
WaitForParallelWorkersToFinish(ParallelContext * pcxt)544 WaitForParallelWorkersToFinish(ParallelContext *pcxt)
545 {
546 for (;;)
547 {
548 bool anyone_alive = false;
549 int nfinished = 0;
550 int i;
551
552 /*
553 * This will process any parallel messages that are pending, which may
554 * change the outcome of the loop that follows. It may also throw an
555 * error propagated from a worker.
556 */
557 CHECK_FOR_INTERRUPTS();
558
559 for (i = 0; i < pcxt->nworkers_launched; ++i)
560 {
561 /*
562 * If error_mqh is NULL, then the worker has already exited
563 * cleanly. If we have received a message through error_mqh from
564 * the worker, we know it started up cleanly, and therefore we're
565 * certain to be notified when it exits.
566 */
567 if (pcxt->worker[i].error_mqh == NULL)
568 ++nfinished;
569 else if (pcxt->any_message_received[i])
570 {
571 anyone_alive = true;
572 break;
573 }
574 }
575
576 if (!anyone_alive)
577 {
578 /* If all workers are known to have finished, we're done. */
579 if (nfinished >= pcxt->nworkers_launched)
580 {
581 Assert(nfinished == pcxt->nworkers_launched);
582 break;
583 }
584
585 /*
586 * We didn't detect any living workers, but not all workers are
587 * known to have exited cleanly. Either not all workers have
588 * launched yet, or maybe some of them failed to start or
589 * terminated abnormally.
590 */
591 for (i = 0; i < pcxt->nworkers_launched; ++i)
592 {
593 pid_t pid;
594 shm_mq *mq;
595
596 /*
597 * If the worker is BGWH_NOT_YET_STARTED or BGWH_STARTED, we
598 * should just keep waiting. If it is BGWH_STOPPED, then
599 * further investigation is needed.
600 */
601 if (pcxt->worker[i].error_mqh == NULL ||
602 pcxt->worker[i].bgwhandle == NULL ||
603 GetBackgroundWorkerPid(pcxt->worker[i].bgwhandle,
604 &pid) != BGWH_STOPPED)
605 continue;
606
607 /*
608 * Check whether the worker ended up stopped without ever
609 * attaching to the error queue. If so, the postmaster was
610 * unable to fork the worker or it exited without initializing
611 * properly. We must throw an error, since the caller may
612 * have been expecting the worker to do some work before
613 * exiting.
614 */
615 mq = shm_mq_get_queue(pcxt->worker[i].error_mqh);
616 if (shm_mq_get_sender(mq) == NULL)
617 ereport(ERROR,
618 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
619 errmsg("parallel worker failed to initialize"),
620 errhint("More details may be available in the server log.")));
621
622 /*
623 * The worker is stopped, but is attached to the error queue.
624 * Unless there's a bug somewhere, this will only happen when
625 * the worker writes messages and terminates after the
626 * CHECK_FOR_INTERRUPTS() near the top of this function and
627 * before the call to GetBackgroundWorkerPid(). In that case,
628 * or latch should have been set as well and the right things
629 * will happen on the next pass through the loop.
630 */
631 }
632 }
633
634 WaitLatch(MyLatch, WL_LATCH_SET, -1,
635 WAIT_EVENT_PARALLEL_FINISH);
636 ResetLatch(MyLatch);
637 }
638
639 if (pcxt->toc != NULL)
640 {
641 FixedParallelState *fps;
642
643 fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED, false);
644 if (fps->last_xlog_end > XactLastRecEnd)
645 XactLastRecEnd = fps->last_xlog_end;
646 }
647 }
648
649 /*
650 * Wait for all workers to exit.
651 *
652 * This function ensures that workers have been completely shutdown. The
653 * difference between WaitForParallelWorkersToFinish and this function is
654 * that former just ensures that last message sent by worker backend is
655 * received by master backend whereas this ensures the complete shutdown.
656 */
657 static void
WaitForParallelWorkersToExit(ParallelContext * pcxt)658 WaitForParallelWorkersToExit(ParallelContext *pcxt)
659 {
660 int i;
661
662 /* Wait until the workers actually die. */
663 for (i = 0; i < pcxt->nworkers_launched; ++i)
664 {
665 BgwHandleStatus status;
666
667 if (pcxt->worker == NULL || pcxt->worker[i].bgwhandle == NULL)
668 continue;
669
670 status = WaitForBackgroundWorkerShutdown(pcxt->worker[i].bgwhandle);
671
672 /*
673 * If the postmaster kicked the bucket, we have no chance of cleaning
674 * up safely -- we won't be able to tell when our workers are actually
675 * dead. This doesn't necessitate a PANIC since they will all abort
676 * eventually, but we can't safely continue this session.
677 */
678 if (status == BGWH_POSTMASTER_DIED)
679 ereport(FATAL,
680 (errcode(ERRCODE_ADMIN_SHUTDOWN),
681 errmsg("postmaster exited during a parallel transaction")));
682
683 /* Release memory. */
684 pfree(pcxt->worker[i].bgwhandle);
685 pcxt->worker[i].bgwhandle = NULL;
686 }
687 }
688
689 /*
690 * Destroy a parallel context.
691 *
692 * If expecting a clean exit, you should use WaitForParallelWorkersToFinish()
693 * first, before calling this function. When this function is invoked, any
694 * remaining workers are forcibly killed; the dynamic shared memory segment
695 * is unmapped; and we then wait (uninterruptibly) for the workers to exit.
696 */
697 void
DestroyParallelContext(ParallelContext * pcxt)698 DestroyParallelContext(ParallelContext *pcxt)
699 {
700 int i;
701
702 /*
703 * Be careful about order of operations here! We remove the parallel
704 * context from the list before we do anything else; otherwise, if an
705 * error occurs during a subsequent step, we might try to nuke it again
706 * from AtEOXact_Parallel or AtEOSubXact_Parallel.
707 */
708 dlist_delete(&pcxt->node);
709
710 /* Kill each worker in turn, and forget their error queues. */
711 if (pcxt->worker != NULL)
712 {
713 for (i = 0; i < pcxt->nworkers_launched; ++i)
714 {
715 if (pcxt->worker[i].error_mqh != NULL)
716 {
717 TerminateBackgroundWorker(pcxt->worker[i].bgwhandle);
718
719 shm_mq_detach(pcxt->worker[i].error_mqh);
720 pcxt->worker[i].error_mqh = NULL;
721 }
722 }
723 }
724
725 /*
726 * If we have allocated a shared memory segment, detach it. This will
727 * implicitly detach the error queues, and any other shared memory queues,
728 * stored there.
729 */
730 if (pcxt->seg != NULL)
731 {
732 dsm_detach(pcxt->seg);
733 pcxt->seg = NULL;
734 }
735
736 /*
737 * If this parallel context is actually in backend-private memory rather
738 * than shared memory, free that memory instead.
739 */
740 if (pcxt->private_memory != NULL)
741 {
742 pfree(pcxt->private_memory);
743 pcxt->private_memory = NULL;
744 }
745
746 /*
747 * We can't finish transaction commit or abort until all of the workers
748 * have exited. This means, in particular, that we can't respond to
749 * interrupts at this stage.
750 */
751 HOLD_INTERRUPTS();
752 WaitForParallelWorkersToExit(pcxt);
753 RESUME_INTERRUPTS();
754
755 /* Free the worker array itself. */
756 if (pcxt->worker != NULL)
757 {
758 pfree(pcxt->worker);
759 pcxt->worker = NULL;
760 }
761
762 /* Free memory. */
763 pfree(pcxt->library_name);
764 pfree(pcxt->function_name);
765 pfree(pcxt);
766 }
767
768 /*
769 * Are there any parallel contexts currently active?
770 */
771 bool
ParallelContextActive(void)772 ParallelContextActive(void)
773 {
774 return !dlist_is_empty(&pcxt_list);
775 }
776
777 /*
778 * Handle receipt of an interrupt indicating a parallel worker message.
779 *
780 * Note: this is called within a signal handler! All we can do is set
781 * a flag that will cause the next CHECK_FOR_INTERRUPTS() to invoke
782 * HandleParallelMessages().
783 */
784 void
HandleParallelMessageInterrupt(void)785 HandleParallelMessageInterrupt(void)
786 {
787 InterruptPending = true;
788 ParallelMessagePending = true;
789 SetLatch(MyLatch);
790 }
791
792 /*
793 * Handle any queued protocol messages received from parallel workers.
794 */
795 void
HandleParallelMessages(void)796 HandleParallelMessages(void)
797 {
798 dlist_iter iter;
799 MemoryContext oldcontext;
800
801 static MemoryContext hpm_context = NULL;
802
803 /*
804 * This is invoked from ProcessInterrupts(), and since some of the
805 * functions it calls contain CHECK_FOR_INTERRUPTS(), there is a potential
806 * for recursive calls if more signals are received while this runs. It's
807 * unclear that recursive entry would be safe, and it doesn't seem useful
808 * even if it is safe, so let's block interrupts until done.
809 */
810 HOLD_INTERRUPTS();
811
812 /*
813 * Moreover, CurrentMemoryContext might be pointing almost anywhere. We
814 * don't want to risk leaking data into long-lived contexts, so let's do
815 * our work here in a private context that we can reset on each use.
816 */
817 if (hpm_context == NULL) /* first time through? */
818 hpm_context = AllocSetContextCreate(TopMemoryContext,
819 "HandleParallelMessages",
820 ALLOCSET_DEFAULT_SIZES);
821 else
822 MemoryContextReset(hpm_context);
823
824 oldcontext = MemoryContextSwitchTo(hpm_context);
825
826 /* OK to process messages. Reset the flag saying there are more to do. */
827 ParallelMessagePending = false;
828
829 dlist_foreach(iter, &pcxt_list)
830 {
831 ParallelContext *pcxt;
832 int i;
833
834 pcxt = dlist_container(ParallelContext, node, iter.cur);
835 if (pcxt->worker == NULL)
836 continue;
837
838 for (i = 0; i < pcxt->nworkers_launched; ++i)
839 {
840 /*
841 * Read as many messages as we can from each worker, but stop when
842 * either (1) the worker's error queue goes away, which can happen
843 * if we receive a Terminate message from the worker; or (2) no
844 * more messages can be read from the worker without blocking.
845 */
846 while (pcxt->worker[i].error_mqh != NULL)
847 {
848 shm_mq_result res;
849 Size nbytes;
850 void *data;
851
852 res = shm_mq_receive(pcxt->worker[i].error_mqh, &nbytes,
853 &data, true);
854 if (res == SHM_MQ_WOULD_BLOCK)
855 break;
856 else if (res == SHM_MQ_SUCCESS)
857 {
858 StringInfoData msg;
859
860 initStringInfo(&msg);
861 appendBinaryStringInfo(&msg, data, nbytes);
862 HandleParallelMessage(pcxt, i, &msg);
863 pfree(msg.data);
864 }
865 else
866 ereport(ERROR,
867 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
868 errmsg("lost connection to parallel worker")));
869 }
870 }
871 }
872
873 MemoryContextSwitchTo(oldcontext);
874
875 /* Might as well clear the context on our way out */
876 MemoryContextReset(hpm_context);
877
878 RESUME_INTERRUPTS();
879 }
880
881 /*
882 * Handle a single protocol message received from a single parallel worker.
883 */
884 static void
HandleParallelMessage(ParallelContext * pcxt,int i,StringInfo msg)885 HandleParallelMessage(ParallelContext *pcxt, int i, StringInfo msg)
886 {
887 char msgtype;
888
889 if (pcxt->any_message_received != NULL)
890 pcxt->any_message_received[i] = true;
891
892 msgtype = pq_getmsgbyte(msg);
893
894 switch (msgtype)
895 {
896 case 'K': /* BackendKeyData */
897 {
898 int32 pid = pq_getmsgint(msg, 4);
899
900 (void) pq_getmsgint(msg, 4); /* discard cancel key */
901 (void) pq_getmsgend(msg);
902 pcxt->worker[i].pid = pid;
903 break;
904 }
905
906 case 'E': /* ErrorResponse */
907 case 'N': /* NoticeResponse */
908 {
909 ErrorData edata;
910 ErrorContextCallback *save_error_context_stack;
911
912 /* Parse ErrorResponse or NoticeResponse. */
913 pq_parse_errornotice(msg, &edata);
914
915 /* Death of a worker isn't enough justification for suicide. */
916 edata.elevel = Min(edata.elevel, ERROR);
917
918 /*
919 * If desired, add a context line to show that this is a
920 * message propagated from a parallel worker. Otherwise, it
921 * can sometimes be confusing to understand what actually
922 * happened. (We don't do this in FORCE_PARALLEL_REGRESS mode
923 * because it causes test-result instability depending on
924 * whether a parallel worker is actually used or not.)
925 */
926 if (force_parallel_mode != FORCE_PARALLEL_REGRESS)
927 {
928 if (edata.context)
929 edata.context = psprintf("%s\n%s", edata.context,
930 _("parallel worker"));
931 else
932 edata.context = pstrdup(_("parallel worker"));
933 }
934
935 /*
936 * Context beyond that should use the error context callbacks
937 * that were in effect when the ParallelContext was created,
938 * not the current ones.
939 */
940 save_error_context_stack = error_context_stack;
941 error_context_stack = pcxt->error_context_stack;
942
943 /* Rethrow error or print notice. */
944 ThrowErrorData(&edata);
945
946 /* Not an error, so restore previous context stack. */
947 error_context_stack = save_error_context_stack;
948
949 break;
950 }
951
952 case 'A': /* NotifyResponse */
953 {
954 /* Propagate NotifyResponse. */
955 int32 pid;
956 const char *channel;
957 const char *payload;
958
959 pid = pq_getmsgint(msg, 4);
960 channel = pq_getmsgrawstring(msg);
961 payload = pq_getmsgrawstring(msg);
962 pq_endmessage(msg);
963
964 NotifyMyFrontEnd(channel, payload, pid);
965
966 break;
967 }
968
969 case 'X': /* Terminate, indicating clean exit */
970 {
971 shm_mq_detach(pcxt->worker[i].error_mqh);
972 pcxt->worker[i].error_mqh = NULL;
973 break;
974 }
975
976 default:
977 {
978 elog(ERROR, "unrecognized message type received from parallel worker: %c (message length %d bytes)",
979 msgtype, msg->len);
980 }
981 }
982 }
983
984 /*
985 * End-of-subtransaction cleanup for parallel contexts.
986 *
987 * Currently, it's forbidden to enter or leave a subtransaction while
988 * parallel mode is in effect, so we could just blow away everything. But
989 * we may want to relax that restriction in the future, so this code
990 * contemplates that there may be multiple subtransaction IDs in pcxt_list.
991 */
992 void
AtEOSubXact_Parallel(bool isCommit,SubTransactionId mySubId)993 AtEOSubXact_Parallel(bool isCommit, SubTransactionId mySubId)
994 {
995 while (!dlist_is_empty(&pcxt_list))
996 {
997 ParallelContext *pcxt;
998
999 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
1000 if (pcxt->subid != mySubId)
1001 break;
1002 if (isCommit)
1003 elog(WARNING, "leaked parallel context");
1004 DestroyParallelContext(pcxt);
1005 }
1006 }
1007
1008 /*
1009 * End-of-transaction cleanup for parallel contexts.
1010 */
1011 void
AtEOXact_Parallel(bool isCommit)1012 AtEOXact_Parallel(bool isCommit)
1013 {
1014 while (!dlist_is_empty(&pcxt_list))
1015 {
1016 ParallelContext *pcxt;
1017
1018 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
1019 if (isCommit)
1020 elog(WARNING, "leaked parallel context");
1021 DestroyParallelContext(pcxt);
1022 }
1023 }
1024
1025 /*
1026 * Main entrypoint for parallel workers.
1027 */
1028 void
ParallelWorkerMain(Datum main_arg)1029 ParallelWorkerMain(Datum main_arg)
1030 {
1031 dsm_segment *seg;
1032 shm_toc *toc;
1033 FixedParallelState *fps;
1034 char *error_queue_space;
1035 shm_mq *mq;
1036 shm_mq_handle *mqh;
1037 char *libraryspace;
1038 char *entrypointstate;
1039 char *library_name;
1040 char *function_name;
1041 parallel_worker_main_type entrypt;
1042 char *gucspace;
1043 char *combocidspace;
1044 char *tsnapspace;
1045 char *asnapspace;
1046 char *tstatespace;
1047 StringInfoData msgbuf;
1048 Snapshot tsnapshot;
1049 Snapshot asnapshot;
1050
1051 /* Set flag to indicate that we're initializing a parallel worker. */
1052 InitializingParallelWorker = true;
1053
1054 /* Establish signal handlers. */
1055 pqsignal(SIGTERM, die);
1056 BackgroundWorkerUnblockSignals();
1057
1058 /* Determine and set our parallel worker number. */
1059 Assert(ParallelWorkerNumber == -1);
1060 memcpy(&ParallelWorkerNumber, MyBgworkerEntry->bgw_extra, sizeof(int));
1061
1062 /* Set up a memory context and resource owner. */
1063 Assert(CurrentResourceOwner == NULL);
1064 CurrentResourceOwner = ResourceOwnerCreate(NULL, "parallel toplevel");
1065 CurrentMemoryContext = AllocSetContextCreate(TopMemoryContext,
1066 "Parallel worker",
1067 ALLOCSET_DEFAULT_SIZES);
1068
1069 /*
1070 * Now that we have a resource owner, we can attach to the dynamic shared
1071 * memory segment and read the table of contents.
1072 */
1073 seg = dsm_attach(DatumGetUInt32(main_arg));
1074 if (seg == NULL)
1075 ereport(ERROR,
1076 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1077 errmsg("could not map dynamic shared memory segment")));
1078 toc = shm_toc_attach(PARALLEL_MAGIC, dsm_segment_address(seg));
1079 if (toc == NULL)
1080 ereport(ERROR,
1081 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1082 errmsg("invalid magic number in dynamic shared memory segment")));
1083
1084 /* Look up fixed parallel state. */
1085 fps = shm_toc_lookup(toc, PARALLEL_KEY_FIXED, false);
1086 MyFixedParallelState = fps;
1087
1088 /* Arrange to signal the leader if we exit. */
1089 ParallelMasterPid = fps->parallel_master_pid;
1090 ParallelMasterBackendId = fps->parallel_master_backend_id;
1091 on_shmem_exit(ParallelWorkerShutdown, (Datum) 0);
1092
1093 /*
1094 * Now we can find and attach to the error queue provided for us. That's
1095 * good, because until we do that, any errors that happen here will not be
1096 * reported back to the process that requested that this worker be
1097 * launched.
1098 */
1099 error_queue_space = shm_toc_lookup(toc, PARALLEL_KEY_ERROR_QUEUE, false);
1100 mq = (shm_mq *) (error_queue_space +
1101 ParallelWorkerNumber * PARALLEL_ERROR_QUEUE_SIZE);
1102 shm_mq_set_sender(mq, MyProc);
1103 mqh = shm_mq_attach(mq, seg, NULL);
1104 pq_redirect_to_shm_mq(seg, mqh);
1105 pq_set_parallel_master(fps->parallel_master_pid,
1106 fps->parallel_master_backend_id);
1107
1108 /*
1109 * Send a BackendKeyData message to the process that initiated parallelism
1110 * so that it has access to our PID before it receives any other messages
1111 * from us. Our cancel key is sent, too, since that's the way the
1112 * protocol message is defined, but it won't actually be used for anything
1113 * in this case.
1114 */
1115 pq_beginmessage(&msgbuf, 'K');
1116 pq_sendint(&msgbuf, (int32) MyProcPid, sizeof(int32));
1117 pq_sendint(&msgbuf, (int32) MyCancelKey, sizeof(int32));
1118 pq_endmessage(&msgbuf);
1119
1120 /*
1121 * Hooray! Primary initialization is complete. Now, we need to set up our
1122 * backend-local state to match the original backend.
1123 */
1124
1125 /*
1126 * Join locking group. We must do this before anything that could try to
1127 * acquire a heavyweight lock, because any heavyweight locks acquired to
1128 * this point could block either directly against the parallel group
1129 * leader or against some process which in turn waits for a lock that
1130 * conflicts with the parallel group leader, causing an undetected
1131 * deadlock. (If we can't join the lock group, the leader has gone away,
1132 * so just exit quietly.)
1133 */
1134 if (!BecomeLockGroupMember(fps->parallel_master_pgproc,
1135 fps->parallel_master_pid))
1136 return;
1137
1138 /*
1139 * Restore transaction and statement start-time timestamps. This must
1140 * happen before anything that would start a transaction, else asserts in
1141 * xact.c will fire.
1142 */
1143 SetParallelStartTimestamps(fps->xact_ts, fps->stmt_ts);
1144
1145 /*
1146 * Identify the entry point to be called. In theory this could result in
1147 * loading an additional library, though most likely the entry point is in
1148 * the core backend or in a library we just loaded.
1149 */
1150 entrypointstate = shm_toc_lookup(toc, PARALLEL_KEY_ENTRYPOINT, false);
1151 library_name = entrypointstate;
1152 function_name = entrypointstate + strlen(library_name) + 1;
1153
1154 entrypt = LookupParallelWorkerFunction(library_name, function_name);
1155
1156 /* Restore database connection. */
1157 BackgroundWorkerInitializeConnectionByOid(fps->database_id,
1158 fps->authenticated_user_id);
1159
1160 /*
1161 * Set the client encoding to the database encoding, since that is what
1162 * the leader will expect.
1163 */
1164 SetClientEncoding(GetDatabaseEncoding());
1165
1166 /*
1167 * Load libraries that were loaded by original backend. We want to do
1168 * this before restoring GUCs, because the libraries might define custom
1169 * variables.
1170 */
1171 libraryspace = shm_toc_lookup(toc, PARALLEL_KEY_LIBRARY, false);
1172 StartTransactionCommand();
1173 RestoreLibraryState(libraryspace);
1174
1175 /* Restore GUC values from launching backend. */
1176 gucspace = shm_toc_lookup(toc, PARALLEL_KEY_GUC, false);
1177 RestoreGUCState(gucspace);
1178 CommitTransactionCommand();
1179
1180 /* Crank up a transaction state appropriate to a parallel worker. */
1181 tstatespace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_STATE, false);
1182 StartParallelWorkerTransaction(tstatespace);
1183
1184 /* Restore combo CID state. */
1185 combocidspace = shm_toc_lookup(toc, PARALLEL_KEY_COMBO_CID, false);
1186 RestoreComboCIDState(combocidspace);
1187
1188 /*
1189 * If the transaction isolation level is REPEATABLE READ or SERIALIZABLE,
1190 * the leader has serialized the transaction snapshot and we must restore
1191 * it. At lower isolation levels, there is no transaction-lifetime
1192 * snapshot, but we need TransactionXmin to get set to a value which is
1193 * less than or equal to the xmin of every snapshot that will be used by
1194 * this worker. The easiest way to accomplish that is to install the
1195 * active snapshot as the transaction snapshot. Code running in this
1196 * parallel worker might take new snapshots via GetTransactionSnapshot()
1197 * or GetLatestSnapshot(), but it shouldn't have any way of acquiring a
1198 * snapshot older than the active snapshot.
1199 */
1200 asnapspace = shm_toc_lookup(toc, PARALLEL_KEY_ACTIVE_SNAPSHOT, false);
1201 tsnapspace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT, true);
1202 asnapshot = RestoreSnapshot(asnapspace);
1203 tsnapshot = tsnapspace ? RestoreSnapshot(tsnapspace) : asnapshot;
1204 RestoreTransactionSnapshot(tsnapshot,
1205 fps->parallel_master_pgproc);
1206 PushActiveSnapshot(asnapshot);
1207
1208 /*
1209 * We've changed which tuples we can see, and must therefore invalidate
1210 * system caches.
1211 */
1212 InvalidateSystemCaches();
1213
1214 /*
1215 * Restore current role id. Skip verifying whether session user is
1216 * allowed to become this role and blindly restore the leader's state for
1217 * current role.
1218 */
1219 SetCurrentRoleId(fps->outer_user_id, fps->is_superuser);
1220
1221 /* Restore user ID and security context. */
1222 SetUserIdAndSecContext(fps->current_user_id, fps->sec_context);
1223
1224 /* Restore temp-namespace state to ensure search path matches leader's. */
1225 SetTempNamespaceState(fps->temp_namespace_id,
1226 fps->temp_toast_namespace_id);
1227
1228 /*
1229 * We've initialized all of our state now; nothing should change
1230 * hereafter.
1231 */
1232 InitializingParallelWorker = false;
1233 EnterParallelMode();
1234
1235 /*
1236 * Time to do the real work: invoke the caller-supplied code.
1237 */
1238 entrypt(seg, toc);
1239
1240 /* Must exit parallel mode to pop active snapshot. */
1241 ExitParallelMode();
1242
1243 /* Must pop active snapshot so resowner.c doesn't complain. */
1244 PopActiveSnapshot();
1245
1246 /* Shut down the parallel-worker transaction. */
1247 EndParallelWorkerTransaction();
1248
1249 /* Report success. */
1250 pq_putmessage('X', NULL, 0);
1251 }
1252
1253 /*
1254 * Update shared memory with the ending location of the last WAL record we
1255 * wrote, if it's greater than the value already stored there.
1256 */
1257 void
ParallelWorkerReportLastRecEnd(XLogRecPtr last_xlog_end)1258 ParallelWorkerReportLastRecEnd(XLogRecPtr last_xlog_end)
1259 {
1260 FixedParallelState *fps = MyFixedParallelState;
1261
1262 Assert(fps != NULL);
1263 SpinLockAcquire(&fps->mutex);
1264 if (fps->last_xlog_end < last_xlog_end)
1265 fps->last_xlog_end = last_xlog_end;
1266 SpinLockRelease(&fps->mutex);
1267 }
1268
1269 /*
1270 * Make sure the leader tries to read from our error queue one more time.
1271 * This guards against the case where we exit uncleanly without sending an
1272 * ErrorResponse to the leader, for example because some code calls proc_exit
1273 * directly.
1274 */
1275 static void
ParallelWorkerShutdown(int code,Datum arg)1276 ParallelWorkerShutdown(int code, Datum arg)
1277 {
1278 SendProcSignal(ParallelMasterPid,
1279 PROCSIG_PARALLEL_MESSAGE,
1280 ParallelMasterBackendId);
1281 }
1282
1283 /*
1284 * Look up (and possibly load) a parallel worker entry point function.
1285 *
1286 * For functions contained in the core code, we use library name "postgres"
1287 * and consult the InternalParallelWorkers array. External functions are
1288 * looked up, and loaded if necessary, using load_external_function().
1289 *
1290 * The point of this is to pass function names as strings across process
1291 * boundaries. We can't pass actual function addresses because of the
1292 * possibility that the function has been loaded at a different address
1293 * in a different process. This is obviously a hazard for functions in
1294 * loadable libraries, but it can happen even for functions in the core code
1295 * on platforms using EXEC_BACKEND (e.g., Windows).
1296 *
1297 * At some point it might be worthwhile to get rid of InternalParallelWorkers[]
1298 * in favor of applying load_external_function() for core functions too;
1299 * but that raises portability issues that are not worth addressing now.
1300 */
1301 static parallel_worker_main_type
LookupParallelWorkerFunction(const char * libraryname,const char * funcname)1302 LookupParallelWorkerFunction(const char *libraryname, const char *funcname)
1303 {
1304 /*
1305 * If the function is to be loaded from postgres itself, search the
1306 * InternalParallelWorkers array.
1307 */
1308 if (strcmp(libraryname, "postgres") == 0)
1309 {
1310 int i;
1311
1312 for (i = 0; i < lengthof(InternalParallelWorkers); i++)
1313 {
1314 if (strcmp(InternalParallelWorkers[i].fn_name, funcname) == 0)
1315 return InternalParallelWorkers[i].fn_addr;
1316 }
1317
1318 /* We can only reach this by programming error. */
1319 elog(ERROR, "internal function \"%s\" not found", funcname);
1320 }
1321
1322 /* Otherwise load from external library. */
1323 return (parallel_worker_main_type)
1324 load_external_function(libraryname, funcname, true, NULL);
1325 }
1326