1 /*-------------------------------------------------------------------------
4 * Infrastructure for launching parallel workers
6 * Portions Copyright (c) 1996-2017, PostgreSQL Global Development Group
7 * Portions Copyright (c) 1994, Regents of the University of California
10 * src/backend/access/transam/parallel.c
12 *-------------------------------------------------------------------------
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"
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"
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.
48 #define PARALLEL_ERROR_QUEUE_SIZE 16384
50 /* Magic number for parallel context TOC. */
51 #define PARALLEL_MAGIC 0x50477c7c
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.
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)
67 /* Fixed-size parallel state. */
68 typedef struct FixedParallelState
70 /* Fixed-size state that workers must restore. */
72 Oid authenticated_user_id;
74 Oid temp_namespace_id;
75 Oid temp_toast_namespace_id;
77 PGPROC *parallel_master_pgproc;
78 pid_t parallel_master_pid;
79 BackendId parallel_master_backend_id;
81 /* Mutex protects remaining fields. */
84 /* Maximum XactLastRecEnd of any worker. */
85 XLogRecPtr last_xlog_end;
89 * Our parallel worker number. We initialize this to -1, meaning that we are
90 * not a parallel worker. In parallel workers, it will be set to a value >= 0
91 * and < the number of workers before any user code is invoked; each parallel
92 * worker will get a different parallel worker number.
94 int ParallelWorkerNumber = -1;
96 /* Is there a parallel message pending which we need to receive? */
97 volatile bool ParallelMessagePending = false;
99 /* Are we initializing a parallel worker? */
100 bool InitializingParallelWorker = false;
102 /* Pointer to our fixed parallel state. */
103 static FixedParallelState *MyFixedParallelState;
105 /* List of active parallel contexts. */
106 static dlist_head pcxt_list = DLIST_STATIC_INIT(pcxt_list);
109 * List of internal parallel worker entry points. We need this for
110 * reasons explained in LookupParallelWorkerFunction(), below.
115 parallel_worker_main_type fn_addr;
116 } InternalParallelWorkers[] =
120 "ParallelQueryMain", ParallelQueryMain
124 /* Private functions. */
125 static void HandleParallelMessage(ParallelContext *pcxt, int i, StringInfo msg);
126 static void WaitForParallelWorkersToExit(ParallelContext *pcxt);
127 static parallel_worker_main_type LookupParallelWorkerFunction(const char *libraryname, const char *funcname);
131 * Establish a new parallel context. This should be done after entering
132 * parallel mode, and (unless there is an error) the context should be
133 * destroyed before exiting the current subtransaction.
136 CreateParallelContext(const char *library_name, const char *function_name,
139 MemoryContext oldcontext;
140 ParallelContext *pcxt;
142 /* It is unsafe to create a parallel context if not in parallel mode. */
143 Assert(IsInParallelMode());
145 /* Number of workers should be non-negative. */
146 Assert(nworkers >= 0);
149 * If dynamic shared memory is not available, we won't be able to use
150 * background workers.
152 if (dynamic_shared_memory_type == DSM_IMPL_NONE)
156 * If we are running under serializable isolation, we can't use parallel
157 * workers, at least not until somebody enhances that mechanism to be
160 if (IsolationIsSerializable())
163 /* We might be running in a short-lived memory context. */
164 oldcontext = MemoryContextSwitchTo(TopTransactionContext);
166 /* Initialize a new ParallelContext. */
167 pcxt = palloc0(sizeof(ParallelContext));
168 pcxt->subid = GetCurrentSubTransactionId();
169 pcxt->nworkers = nworkers;
170 pcxt->library_name = pstrdup(library_name);
171 pcxt->function_name = pstrdup(function_name);
172 pcxt->error_context_stack = error_context_stack;
173 shm_toc_initialize_estimator(&pcxt->estimator);
174 dlist_push_head(&pcxt_list, &pcxt->node);
176 /* Restore previous memory context. */
177 MemoryContextSwitchTo(oldcontext);
183 * Establish the dynamic shared memory segment for a parallel context and
184 * copy state and other bookkeeping information that will be needed by
185 * parallel workers into it.
188 InitializeParallelDSM(ParallelContext *pcxt)
190 MemoryContext oldcontext;
191 Size library_len = 0;
193 Size combocidlen = 0;
199 FixedParallelState *fps;
200 Snapshot transaction_snapshot = GetTransactionSnapshot();
201 Snapshot active_snapshot = GetActiveSnapshot();
203 /* We might be running in a very short-lived memory context. */
204 oldcontext = MemoryContextSwitchTo(TopTransactionContext);
206 /* Allow space to store the fixed-size parallel state. */
207 shm_toc_estimate_chunk(&pcxt->estimator, sizeof(FixedParallelState));
208 shm_toc_estimate_keys(&pcxt->estimator, 1);
211 * Normally, the user will have requested at least one worker process, but
212 * if by chance they have not, we can skip a bunch of things here.
214 if (pcxt->nworkers > 0)
216 /* Estimate space for various kinds of state sharing. */
217 library_len = EstimateLibraryStateSpace();
218 shm_toc_estimate_chunk(&pcxt->estimator, library_len);
219 guc_len = EstimateGUCStateSpace();
220 shm_toc_estimate_chunk(&pcxt->estimator, guc_len);
221 combocidlen = EstimateComboCIDStateSpace();
222 shm_toc_estimate_chunk(&pcxt->estimator, combocidlen);
223 tsnaplen = EstimateSnapshotSpace(transaction_snapshot);
224 shm_toc_estimate_chunk(&pcxt->estimator, tsnaplen);
225 asnaplen = EstimateSnapshotSpace(active_snapshot);
226 shm_toc_estimate_chunk(&pcxt->estimator, asnaplen);
227 tstatelen = EstimateTransactionStateSpace();
228 shm_toc_estimate_chunk(&pcxt->estimator, tstatelen);
229 /* If you add more chunks here, you probably need to add keys. */
230 shm_toc_estimate_keys(&pcxt->estimator, 6);
232 /* Estimate space need for error queues. */
233 StaticAssertStmt(BUFFERALIGN(PARALLEL_ERROR_QUEUE_SIZE) ==
234 PARALLEL_ERROR_QUEUE_SIZE,
235 "parallel error queue size not buffer-aligned");
236 shm_toc_estimate_chunk(&pcxt->estimator,
237 mul_size(PARALLEL_ERROR_QUEUE_SIZE,
239 shm_toc_estimate_keys(&pcxt->estimator, 1);
241 /* Estimate how much we'll need for the entrypoint info. */
242 shm_toc_estimate_chunk(&pcxt->estimator, strlen(pcxt->library_name) +
243 strlen(pcxt->function_name) + 2);
244 shm_toc_estimate_keys(&pcxt->estimator, 1);
248 * Create DSM and initialize with new table of contents. But if the user
249 * didn't request any workers, then don't bother creating a dynamic shared
250 * memory segment; instead, just use backend-private memory.
252 * Also, if we can't create a dynamic shared memory segment because the
253 * maximum number of segments have already been created, then fall back to
254 * backend-private memory, and plan not to use any workers. We hope this
255 * won't happen very often, but it's better to abandon the use of
256 * parallelism than to fail outright.
258 segsize = shm_toc_estimate(&pcxt->estimator);
259 if (pcxt->nworkers > 0)
260 pcxt->seg = dsm_create(segsize, DSM_CREATE_NULL_IF_MAXSEGMENTS);
261 if (pcxt->seg != NULL)
262 pcxt->toc = shm_toc_create(PARALLEL_MAGIC,
263 dsm_segment_address(pcxt->seg),
268 pcxt->private_memory = MemoryContextAlloc(TopMemoryContext, segsize);
269 pcxt->toc = shm_toc_create(PARALLEL_MAGIC, pcxt->private_memory,
273 /* Initialize fixed-size state in shared memory. */
274 fps = (FixedParallelState *)
275 shm_toc_allocate(pcxt->toc, sizeof(FixedParallelState));
276 fps->database_id = MyDatabaseId;
277 fps->authenticated_user_id = GetAuthenticatedUserId();
278 GetUserIdAndSecContext(&fps->current_user_id, &fps->sec_context);
279 GetTempNamespaceState(&fps->temp_namespace_id,
280 &fps->temp_toast_namespace_id);
281 fps->parallel_master_pgproc = MyProc;
282 fps->parallel_master_pid = MyProcPid;
283 fps->parallel_master_backend_id = MyBackendId;
284 SpinLockInit(&fps->mutex);
285 fps->last_xlog_end = 0;
286 shm_toc_insert(pcxt->toc, PARALLEL_KEY_FIXED, fps);
288 /* We can skip the rest of this if we're not budgeting for any workers. */
289 if (pcxt->nworkers > 0)
297 char *error_queue_space;
298 char *entrypointstate;
301 /* Serialize shared libraries we have loaded. */
302 libraryspace = shm_toc_allocate(pcxt->toc, library_len);
303 SerializeLibraryState(library_len, libraryspace);
304 shm_toc_insert(pcxt->toc, PARALLEL_KEY_LIBRARY, libraryspace);
306 /* Serialize GUC settings. */
307 gucspace = shm_toc_allocate(pcxt->toc, guc_len);
308 SerializeGUCState(guc_len, gucspace);
309 shm_toc_insert(pcxt->toc, PARALLEL_KEY_GUC, gucspace);
311 /* Serialize combo CID state. */
312 combocidspace = shm_toc_allocate(pcxt->toc, combocidlen);
313 SerializeComboCIDState(combocidlen, combocidspace);
314 shm_toc_insert(pcxt->toc, PARALLEL_KEY_COMBO_CID, combocidspace);
316 /* Serialize transaction snapshot and active snapshot. */
317 tsnapspace = shm_toc_allocate(pcxt->toc, tsnaplen);
318 SerializeSnapshot(transaction_snapshot, tsnapspace);
319 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT,
321 asnapspace = shm_toc_allocate(pcxt->toc, asnaplen);
322 SerializeSnapshot(active_snapshot, asnapspace);
323 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ACTIVE_SNAPSHOT, asnapspace);
325 /* Serialize transaction state. */
326 tstatespace = shm_toc_allocate(pcxt->toc, tstatelen);
327 SerializeTransactionState(tstatelen, tstatespace);
328 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_STATE, tstatespace);
330 /* Allocate space for worker information. */
331 pcxt->worker = palloc0(sizeof(ParallelWorkerInfo) * pcxt->nworkers);
334 * Establish error queues in dynamic shared memory.
336 * These queues should be used only for transmitting ErrorResponse,
337 * NoticeResponse, and NotifyResponse protocol messages. Tuple data
338 * should be transmitted via separate (possibly larger?) queues.
341 shm_toc_allocate(pcxt->toc,
342 mul_size(PARALLEL_ERROR_QUEUE_SIZE,
344 for (i = 0; i < pcxt->nworkers; ++i)
349 start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
350 mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
351 shm_mq_set_receiver(mq, MyProc);
352 pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
354 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE, error_queue_space);
357 * Serialize entrypoint information. It's unsafe to pass function
358 * pointers across processes, as the function pointer may be different
359 * in each process in EXEC_BACKEND builds, so we always pass library
360 * and function name. (We use library name "postgres" for functions
361 * in the core backend.)
363 lnamelen = strlen(pcxt->library_name);
364 entrypointstate = shm_toc_allocate(pcxt->toc, lnamelen +
365 strlen(pcxt->function_name) + 2);
366 strcpy(entrypointstate, pcxt->library_name);
367 strcpy(entrypointstate + lnamelen + 1, pcxt->function_name);
368 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ENTRYPOINT, entrypointstate);
371 /* Restore previous memory context. */
372 MemoryContextSwitchTo(oldcontext);
376 * Reinitialize the dynamic shared memory segment for a parallel context such
377 * that we could launch workers for it again.
380 ReinitializeParallelDSM(ParallelContext *pcxt)
382 FixedParallelState *fps;
383 char *error_queue_space;
386 /* Wait for any old workers to exit. */
387 if (pcxt->nworkers_launched > 0)
389 WaitForParallelWorkersToFinish(pcxt);
390 WaitForParallelWorkersToExit(pcxt);
391 pcxt->nworkers_launched = 0;
394 /* Reset a few bits of fixed parallel state to a clean state. */
395 fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED, false);
396 fps->last_xlog_end = 0;
398 /* Recreate error queues. */
400 shm_toc_lookup(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE, false);
401 for (i = 0; i < pcxt->nworkers; ++i)
406 start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
407 mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
408 shm_mq_set_receiver(mq, MyProc);
409 pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
414 * Launch parallel workers.
417 LaunchParallelWorkers(ParallelContext *pcxt)
419 MemoryContext oldcontext;
420 BackgroundWorker worker;
422 bool any_registrations_failed = false;
424 /* Skip this if we have no workers. */
425 if (pcxt->nworkers == 0)
428 /* We need to be a lock group leader. */
429 BecomeLockGroupLeader();
431 /* If we do have workers, we'd better have a DSM segment. */
432 Assert(pcxt->seg != NULL);
434 /* We might be running in a short-lived memory context. */
435 oldcontext = MemoryContextSwitchTo(TopTransactionContext);
437 /* Configure a worker. */
438 memset(&worker, 0, sizeof(worker));
439 snprintf(worker.bgw_name, BGW_MAXLEN, "parallel worker for PID %d",
442 BGWORKER_SHMEM_ACCESS | BGWORKER_BACKEND_DATABASE_CONNECTION
443 | BGWORKER_CLASS_PARALLEL;
444 worker.bgw_start_time = BgWorkerStart_ConsistentState;
445 worker.bgw_restart_time = BGW_NEVER_RESTART;
446 sprintf(worker.bgw_library_name, "postgres");
447 sprintf(worker.bgw_function_name, "ParallelWorkerMain");
448 worker.bgw_main_arg = UInt32GetDatum(dsm_segment_handle(pcxt->seg));
449 worker.bgw_notify_pid = MyProcPid;
454 * The caller must be able to tolerate ending up with fewer workers than
455 * expected, so there is no need to throw an error here if registration
456 * fails. It wouldn't help much anyway, because registering the worker in
457 * no way guarantees that it will start up and initialize successfully.
459 for (i = 0; i < pcxt->nworkers; ++i)
461 memcpy(worker.bgw_extra, &i, sizeof(int));
462 if (!any_registrations_failed &&
463 RegisterDynamicBackgroundWorker(&worker,
464 &pcxt->worker[i].bgwhandle))
466 shm_mq_set_handle(pcxt->worker[i].error_mqh,
467 pcxt->worker[i].bgwhandle);
468 pcxt->nworkers_launched++;
473 * If we weren't able to register the worker, then we've bumped up
474 * against the max_worker_processes limit, and future
475 * registrations will probably fail too, so arrange to skip them.
476 * But we still have to execute this code for the remaining slots
477 * to make sure that we forget about the error queues we budgeted
478 * for those workers. Otherwise, we'll wait for them to start,
479 * but they never will.
481 any_registrations_failed = true;
482 pcxt->worker[i].bgwhandle = NULL;
483 pfree(pcxt->worker[i].error_mqh);
484 pcxt->worker[i].error_mqh = NULL;
488 /* Restore previous memory context. */
489 MemoryContextSwitchTo(oldcontext);
493 * Wait for all workers to finish computing.
495 * Even if the parallel operation seems to have completed successfully, it's
496 * important to call this function afterwards. We must not miss any errors
497 * the workers may have thrown during the parallel operation, or any that they
498 * may yet throw while shutting down.
500 * Also, we want to update our notion of XactLastRecEnd based on worker
504 WaitForParallelWorkersToFinish(ParallelContext *pcxt)
508 bool anyone_alive = false;
512 * This will process any parallel messages that are pending, which may
513 * change the outcome of the loop that follows. It may also throw an
514 * error propagated from a worker.
516 CHECK_FOR_INTERRUPTS();
518 for (i = 0; i < pcxt->nworkers_launched; ++i)
520 if (pcxt->worker[i].error_mqh != NULL)
530 WaitLatch(&MyProc->procLatch, WL_LATCH_SET, -1,
531 WAIT_EVENT_PARALLEL_FINISH);
532 ResetLatch(&MyProc->procLatch);
535 if (pcxt->toc != NULL)
537 FixedParallelState *fps;
539 fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED, false);
540 if (fps->last_xlog_end > XactLastRecEnd)
541 XactLastRecEnd = fps->last_xlog_end;
546 * Wait for all workers to exit.
548 * This function ensures that workers have been completely shutdown. The
549 * difference between WaitForParallelWorkersToFinish and this function is
550 * that former just ensures that last message sent by worker backend is
551 * received by master backend whereas this ensures the complete shutdown.
554 WaitForParallelWorkersToExit(ParallelContext *pcxt)
558 /* Wait until the workers actually die. */
559 for (i = 0; i < pcxt->nworkers_launched; ++i)
561 BgwHandleStatus status;
563 if (pcxt->worker == NULL || pcxt->worker[i].bgwhandle == NULL)
566 status = WaitForBackgroundWorkerShutdown(pcxt->worker[i].bgwhandle);
569 * If the postmaster kicked the bucket, we have no chance of cleaning
570 * up safely -- we won't be able to tell when our workers are actually
571 * dead. This doesn't necessitate a PANIC since they will all abort
572 * eventually, but we can't safely continue this session.
574 if (status == BGWH_POSTMASTER_DIED)
576 (errcode(ERRCODE_ADMIN_SHUTDOWN),
577 errmsg("postmaster exited during a parallel transaction")));
579 /* Release memory. */
580 pfree(pcxt->worker[i].bgwhandle);
581 pcxt->worker[i].bgwhandle = NULL;
586 * Destroy a parallel context.
588 * If expecting a clean exit, you should use WaitForParallelWorkersToFinish()
589 * first, before calling this function. When this function is invoked, any
590 * remaining workers are forcibly killed; the dynamic shared memory segment
591 * is unmapped; and we then wait (uninterruptibly) for the workers to exit.
594 DestroyParallelContext(ParallelContext *pcxt)
599 * Be careful about order of operations here! We remove the parallel
600 * context from the list before we do anything else; otherwise, if an
601 * error occurs during a subsequent step, we might try to nuke it again
602 * from AtEOXact_Parallel or AtEOSubXact_Parallel.
604 dlist_delete(&pcxt->node);
606 /* Kill each worker in turn, and forget their error queues. */
607 if (pcxt->worker != NULL)
609 for (i = 0; i < pcxt->nworkers_launched; ++i)
611 if (pcxt->worker[i].error_mqh != NULL)
613 TerminateBackgroundWorker(pcxt->worker[i].bgwhandle);
615 pfree(pcxt->worker[i].error_mqh);
616 pcxt->worker[i].error_mqh = NULL;
622 * If we have allocated a shared memory segment, detach it. This will
623 * implicitly detach the error queues, and any other shared memory queues,
626 if (pcxt->seg != NULL)
628 dsm_detach(pcxt->seg);
633 * If this parallel context is actually in backend-private memory rather
634 * than shared memory, free that memory instead.
636 if (pcxt->private_memory != NULL)
638 pfree(pcxt->private_memory);
639 pcxt->private_memory = NULL;
643 * We can't finish transaction commit or abort until all of the workers
644 * have exited. This means, in particular, that we can't respond to
645 * interrupts at this stage.
648 WaitForParallelWorkersToExit(pcxt);
651 /* Free the worker array itself. */
652 if (pcxt->worker != NULL)
659 pfree(pcxt->library_name);
660 pfree(pcxt->function_name);
665 * Are there any parallel contexts currently active?
668 ParallelContextActive(void)
670 return !dlist_is_empty(&pcxt_list);
674 * Handle receipt of an interrupt indicating a parallel worker message.
676 * Note: this is called within a signal handler! All we can do is set
677 * a flag that will cause the next CHECK_FOR_INTERRUPTS() to invoke
678 * HandleParallelMessages().
681 HandleParallelMessageInterrupt(void)
683 InterruptPending = true;
684 ParallelMessagePending = true;
689 * Handle any queued protocol messages received from parallel workers.
692 HandleParallelMessages(void)
695 MemoryContext oldcontext;
697 static MemoryContext hpm_context = NULL;
700 * This is invoked from ProcessInterrupts(), and since some of the
701 * functions it calls contain CHECK_FOR_INTERRUPTS(), there is a potential
702 * for recursive calls if more signals are received while this runs. It's
703 * unclear that recursive entry would be safe, and it doesn't seem useful
704 * even if it is safe, so let's block interrupts until done.
709 * Moreover, CurrentMemoryContext might be pointing almost anywhere. We
710 * don't want to risk leaking data into long-lived contexts, so let's do
711 * our work here in a private context that we can reset on each use.
713 if (hpm_context == NULL) /* first time through? */
714 hpm_context = AllocSetContextCreate(TopMemoryContext,
715 "HandleParallelMessages",
716 ALLOCSET_DEFAULT_SIZES);
718 MemoryContextReset(hpm_context);
720 oldcontext = MemoryContextSwitchTo(hpm_context);
722 /* OK to process messages. Reset the flag saying there are more to do. */
723 ParallelMessagePending = false;
725 dlist_foreach(iter, &pcxt_list)
727 ParallelContext *pcxt;
730 pcxt = dlist_container(ParallelContext, node, iter.cur);
731 if (pcxt->worker == NULL)
734 for (i = 0; i < pcxt->nworkers_launched; ++i)
737 * Read as many messages as we can from each worker, but stop when
738 * either (1) the worker's error queue goes away, which can happen
739 * if we receive a Terminate message from the worker; or (2) no
740 * more messages can be read from the worker without blocking.
742 while (pcxt->worker[i].error_mqh != NULL)
748 res = shm_mq_receive(pcxt->worker[i].error_mqh, &nbytes,
750 if (res == SHM_MQ_WOULD_BLOCK)
752 else if (res == SHM_MQ_SUCCESS)
756 initStringInfo(&msg);
757 appendBinaryStringInfo(&msg, data, nbytes);
758 HandleParallelMessage(pcxt, i, &msg);
763 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
764 errmsg("lost connection to parallel worker")));
769 MemoryContextSwitchTo(oldcontext);
771 /* Might as well clear the context on our way out */
772 MemoryContextReset(hpm_context);
778 * Handle a single protocol message received from a single parallel worker.
781 HandleParallelMessage(ParallelContext *pcxt, int i, StringInfo msg)
785 msgtype = pq_getmsgbyte(msg);
789 case 'K': /* BackendKeyData */
791 int32 pid = pq_getmsgint(msg, 4);
793 (void) pq_getmsgint(msg, 4); /* discard cancel key */
794 (void) pq_getmsgend(msg);
795 pcxt->worker[i].pid = pid;
799 case 'E': /* ErrorResponse */
800 case 'N': /* NoticeResponse */
803 ErrorContextCallback *save_error_context_stack;
805 /* Parse ErrorResponse or NoticeResponse. */
806 pq_parse_errornotice(msg, &edata);
808 /* Death of a worker isn't enough justification for suicide. */
809 edata.elevel = Min(edata.elevel, ERROR);
812 * If desired, add a context line to show that this is a
813 * message propagated from a parallel worker. Otherwise, it
814 * can sometimes be confusing to understand what actually
815 * happened. (We don't do this in FORCE_PARALLEL_REGRESS mode
816 * because it causes test-result instability depending on
817 * whether a parallel worker is actually used or not.)
819 if (force_parallel_mode != FORCE_PARALLEL_REGRESS)
822 edata.context = psprintf("%s\n%s", edata.context,
823 _("parallel worker"));
825 edata.context = pstrdup(_("parallel worker"));
829 * Context beyond that should use the error context callbacks
830 * that were in effect when the ParallelContext was created,
831 * not the current ones.
833 save_error_context_stack = error_context_stack;
834 error_context_stack = pcxt->error_context_stack;
836 /* Rethrow error or print notice. */
837 ThrowErrorData(&edata);
839 /* Not an error, so restore previous context stack. */
840 error_context_stack = save_error_context_stack;
845 case 'A': /* NotifyResponse */
847 /* Propagate NotifyResponse. */
852 pid = pq_getmsgint(msg, 4);
853 channel = pq_getmsgrawstring(msg);
854 payload = pq_getmsgrawstring(msg);
857 NotifyMyFrontEnd(channel, payload, pid);
862 case 'X': /* Terminate, indicating clean exit */
864 pfree(pcxt->worker[i].error_mqh);
865 pcxt->worker[i].error_mqh = NULL;
871 elog(ERROR, "unrecognized message type received from parallel worker: %c (message length %d bytes)",
878 * End-of-subtransaction cleanup for parallel contexts.
880 * Currently, it's forbidden to enter or leave a subtransaction while
881 * parallel mode is in effect, so we could just blow away everything. But
882 * we may want to relax that restriction in the future, so this code
883 * contemplates that there may be multiple subtransaction IDs in pcxt_list.
886 AtEOSubXact_Parallel(bool isCommit, SubTransactionId mySubId)
888 while (!dlist_is_empty(&pcxt_list))
890 ParallelContext *pcxt;
892 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
893 if (pcxt->subid != mySubId)
896 elog(WARNING, "leaked parallel context");
897 DestroyParallelContext(pcxt);
902 * End-of-transaction cleanup for parallel contexts.
905 AtEOXact_Parallel(bool isCommit)
907 while (!dlist_is_empty(&pcxt_list))
909 ParallelContext *pcxt;
911 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
913 elog(WARNING, "leaked parallel context");
914 DestroyParallelContext(pcxt);
919 * Main entrypoint for parallel workers.
922 ParallelWorkerMain(Datum main_arg)
926 FixedParallelState *fps;
927 char *error_queue_space;
931 char *entrypointstate;
934 parallel_worker_main_type entrypt;
940 StringInfoData msgbuf;
942 /* Set flag to indicate that we're initializing a parallel worker. */
943 InitializingParallelWorker = true;
945 /* Establish signal handlers. */
946 pqsignal(SIGTERM, die);
947 BackgroundWorkerUnblockSignals();
949 /* Determine and set our parallel worker number. */
950 Assert(ParallelWorkerNumber == -1);
951 memcpy(&ParallelWorkerNumber, MyBgworkerEntry->bgw_extra, sizeof(int));
953 /* Set up a memory context and resource owner. */
954 Assert(CurrentResourceOwner == NULL);
955 CurrentResourceOwner = ResourceOwnerCreate(NULL, "parallel toplevel");
956 CurrentMemoryContext = AllocSetContextCreate(TopMemoryContext,
958 ALLOCSET_DEFAULT_SIZES);
961 * Now that we have a resource owner, we can attach to the dynamic shared
962 * memory segment and read the table of contents.
964 seg = dsm_attach(DatumGetUInt32(main_arg));
967 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
968 errmsg("could not map dynamic shared memory segment")));
969 toc = shm_toc_attach(PARALLEL_MAGIC, dsm_segment_address(seg));
972 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
973 errmsg("invalid magic number in dynamic shared memory segment")));
975 /* Look up fixed parallel state. */
976 fps = shm_toc_lookup(toc, PARALLEL_KEY_FIXED, false);
977 MyFixedParallelState = fps;
980 * Now that we have a worker number, we can find and attach to the error
981 * queue provided for us. That's good, because until we do that, any
982 * errors that happen here will not be reported back to the process that
983 * requested that this worker be launched.
985 error_queue_space = shm_toc_lookup(toc, PARALLEL_KEY_ERROR_QUEUE, false);
986 mq = (shm_mq *) (error_queue_space +
987 ParallelWorkerNumber * PARALLEL_ERROR_QUEUE_SIZE);
988 shm_mq_set_sender(mq, MyProc);
989 mqh = shm_mq_attach(mq, seg, NULL);
990 pq_redirect_to_shm_mq(seg, mqh);
991 pq_set_parallel_master(fps->parallel_master_pid,
992 fps->parallel_master_backend_id);
995 * Send a BackendKeyData message to the process that initiated parallelism
996 * so that it has access to our PID before it receives any other messages
997 * from us. Our cancel key is sent, too, since that's the way the
998 * protocol message is defined, but it won't actually be used for anything
1001 pq_beginmessage(&msgbuf, 'K');
1002 pq_sendint(&msgbuf, (int32) MyProcPid, sizeof(int32));
1003 pq_sendint(&msgbuf, (int32) MyCancelKey, sizeof(int32));
1004 pq_endmessage(&msgbuf);
1007 * Hooray! Primary initialization is complete. Now, we need to set up our
1008 * backend-local state to match the original backend.
1012 * Join locking group. We must do this before anything that could try to
1013 * acquire a heavyweight lock, because any heavyweight locks acquired to
1014 * this point could block either directly against the parallel group
1015 * leader or against some process which in turn waits for a lock that
1016 * conflicts with the parallel group leader, causing an undetected
1017 * deadlock. (If we can't join the lock group, the leader has gone away,
1018 * so just exit quietly.)
1020 if (!BecomeLockGroupMember(fps->parallel_master_pgproc,
1021 fps->parallel_master_pid))
1025 * Load libraries that were loaded by original backend. We want to do
1026 * this before restoring GUCs, because the libraries might define custom
1029 libraryspace = shm_toc_lookup(toc, PARALLEL_KEY_LIBRARY, false);
1030 RestoreLibraryState(libraryspace);
1033 * Identify the entry point to be called. In theory this could result in
1034 * loading an additional library, though most likely the entry point is in
1035 * the core backend or in a library we just loaded.
1037 entrypointstate = shm_toc_lookup(toc, PARALLEL_KEY_ENTRYPOINT, false);
1038 library_name = entrypointstate;
1039 function_name = entrypointstate + strlen(library_name) + 1;
1041 entrypt = LookupParallelWorkerFunction(library_name, function_name);
1043 /* Restore database connection. */
1044 BackgroundWorkerInitializeConnectionByOid(fps->database_id,
1045 fps->authenticated_user_id);
1048 * Set the client encoding to the database encoding, since that is what
1049 * the leader will expect.
1051 SetClientEncoding(GetDatabaseEncoding());
1053 /* Restore GUC values from launching backend. */
1054 gucspace = shm_toc_lookup(toc, PARALLEL_KEY_GUC, false);
1055 StartTransactionCommand();
1056 RestoreGUCState(gucspace);
1057 CommitTransactionCommand();
1059 /* Crank up a transaction state appropriate to a parallel worker. */
1060 tstatespace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_STATE, false);
1061 StartParallelWorkerTransaction(tstatespace);
1063 /* Restore combo CID state. */
1064 combocidspace = shm_toc_lookup(toc, PARALLEL_KEY_COMBO_CID, false);
1065 RestoreComboCIDState(combocidspace);
1067 /* Restore transaction snapshot. */
1068 tsnapspace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT, false);
1069 RestoreTransactionSnapshot(RestoreSnapshot(tsnapspace),
1070 fps->parallel_master_pgproc);
1072 /* Restore active snapshot. */
1073 asnapspace = shm_toc_lookup(toc, PARALLEL_KEY_ACTIVE_SNAPSHOT, false);
1074 PushActiveSnapshot(RestoreSnapshot(asnapspace));
1077 * We've changed which tuples we can see, and must therefore invalidate
1080 InvalidateSystemCaches();
1082 /* Restore user ID and security context. */
1083 SetUserIdAndSecContext(fps->current_user_id, fps->sec_context);
1085 /* Restore temp-namespace state to ensure search path matches leader's. */
1086 SetTempNamespaceState(fps->temp_namespace_id,
1087 fps->temp_toast_namespace_id);
1089 /* Set ParallelMasterBackendId so we know how to address temp relations. */
1090 ParallelMasterBackendId = fps->parallel_master_backend_id;
1093 * We've initialized all of our state now; nothing should change
1096 InitializingParallelWorker = false;
1097 EnterParallelMode();
1100 * Time to do the real work: invoke the caller-supplied code.
1104 /* Must exit parallel mode to pop active snapshot. */
1107 /* Must pop active snapshot so resowner.c doesn't complain. */
1108 PopActiveSnapshot();
1110 /* Shut down the parallel-worker transaction. */
1111 EndParallelWorkerTransaction();
1113 /* Report success. */
1114 pq_putmessage('X', NULL, 0);
1118 * Update shared memory with the ending location of the last WAL record we
1119 * wrote, if it's greater than the value already stored there.
1122 ParallelWorkerReportLastRecEnd(XLogRecPtr last_xlog_end)
1124 FixedParallelState *fps = MyFixedParallelState;
1126 Assert(fps != NULL);
1127 SpinLockAcquire(&fps->mutex);
1128 if (fps->last_xlog_end < last_xlog_end)
1129 fps->last_xlog_end = last_xlog_end;
1130 SpinLockRelease(&fps->mutex);
1134 * Look up (and possibly load) a parallel worker entry point function.
1136 * For functions contained in the core code, we use library name "postgres"
1137 * and consult the InternalParallelWorkers array. External functions are
1138 * looked up, and loaded if necessary, using load_external_function().
1140 * The point of this is to pass function names as strings across process
1141 * boundaries. We can't pass actual function addresses because of the
1142 * possibility that the function has been loaded at a different address
1143 * in a different process. This is obviously a hazard for functions in
1144 * loadable libraries, but it can happen even for functions in the core code
1145 * on platforms using EXEC_BACKEND (e.g., Windows).
1147 * At some point it might be worthwhile to get rid of InternalParallelWorkers[]
1148 * in favor of applying load_external_function() for core functions too;
1149 * but that raises portability issues that are not worth addressing now.
1151 static parallel_worker_main_type
1152 LookupParallelWorkerFunction(const char *libraryname, const char *funcname)
1155 * If the function is to be loaded from postgres itself, search the
1156 * InternalParallelWorkers array.
1158 if (strcmp(libraryname, "postgres") == 0)
1162 for (i = 0; i < lengthof(InternalParallelWorkers); i++)
1164 if (strcmp(InternalParallelWorkers[i].fn_name, funcname) == 0)
1165 return InternalParallelWorkers[i].fn_addr;
1168 /* We can only reach this by programming error. */
1169 elog(ERROR, "internal function \"%s\" not found", funcname);
1172 /* Otherwise load from external library. */
1173 return (parallel_worker_main_type)
1174 load_external_function(libraryname, funcname, true, NULL);