1 /*-------------------------------------------------------------------------
4 * Infrastructure for launching parallel workers
6 * Portions Copyright (c) 1996-2024, 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/brin.h"
18 #include "access/nbtree.h"
19 #include "access/parallel.h"
20 #include "access/session.h"
21 #include "access/xact.h"
22 #include "access/xlog.h"
23 #include "catalog/index.h"
24 #include "catalog/namespace.h"
25 #include "catalog/pg_enum.h"
26 #include "catalog/storage.h"
27 #include "commands/async.h"
28 #include "commands/vacuum.h"
29 #include "executor/execParallel.h"
30 #include "libpq/libpq.h"
31 #include "libpq/pqformat.h"
32 #include "libpq/pqmq.h"
33 #include "miscadmin.h"
34 #include "optimizer/optimizer.h"
36 #include "storage/ipc.h"
37 #include "storage/predicate.h"
38 #include "storage/spin.h"
39 #include "tcop/tcopprot.h"
40 #include "utils/combocid.h"
41 #include "utils/guc.h"
42 #include "utils/inval.h"
43 #include "utils/memutils.h"
44 #include "utils/relmapper.h"
45 #include "utils/snapmgr.h"
48 * We don't want to waste a lot of memory on an error queue which, most of
49 * the time, will process only a handful of small messages. However, it is
50 * desirable to make it large enough that a typical ErrorResponse can be sent
51 * without blocking. That way, a worker that errors out can write the whole
52 * message into the queue and terminate without waiting for the user backend.
54 #define PARALLEL_ERROR_QUEUE_SIZE 16384
56 /* Magic number for parallel context TOC. */
57 #define PARALLEL_MAGIC 0x50477c7c
60 * Magic numbers for per-context parallel state sharing. Higher-level code
61 * should use smaller values, leaving these very large ones for use by this
64 #define PARALLEL_KEY_FIXED UINT64CONST(0xFFFFFFFFFFFF0001)
65 #define PARALLEL_KEY_ERROR_QUEUE UINT64CONST(0xFFFFFFFFFFFF0002)
66 #define PARALLEL_KEY_LIBRARY UINT64CONST(0xFFFFFFFFFFFF0003)
67 #define PARALLEL_KEY_GUC UINT64CONST(0xFFFFFFFFFFFF0004)
68 #define PARALLEL_KEY_COMBO_CID UINT64CONST(0xFFFFFFFFFFFF0005)
69 #define PARALLEL_KEY_TRANSACTION_SNAPSHOT UINT64CONST(0xFFFFFFFFFFFF0006)
70 #define PARALLEL_KEY_ACTIVE_SNAPSHOT UINT64CONST(0xFFFFFFFFFFFF0007)
71 #define PARALLEL_KEY_TRANSACTION_STATE UINT64CONST(0xFFFFFFFFFFFF0008)
72 #define PARALLEL_KEY_ENTRYPOINT UINT64CONST(0xFFFFFFFFFFFF0009)
73 #define PARALLEL_KEY_SESSION_DSM UINT64CONST(0xFFFFFFFFFFFF000A)
74 #define PARALLEL_KEY_PENDING_SYNCS UINT64CONST(0xFFFFFFFFFFFF000B)
75 #define PARALLEL_KEY_REINDEX_STATE UINT64CONST(0xFFFFFFFFFFFF000C)
76 #define PARALLEL_KEY_RELMAPPER_STATE UINT64CONST(0xFFFFFFFFFFFF000D)
77 #define PARALLEL_KEY_UNCOMMITTEDENUMS UINT64CONST(0xFFFFFFFFFFFF000E)
78 #define PARALLEL_KEY_CLIENTCONNINFO UINT64CONST(0xFFFFFFFFFFFF000F)
80 /* Fixed-size parallel state. */
81 typedef struct FixedParallelState
83 /* Fixed-size state that workers must restore. */
85 Oid authenticated_user_id
;
88 Oid temp_namespace_id
;
89 Oid temp_toast_namespace_id
;
92 PGPROC
*parallel_leader_pgproc
;
93 pid_t parallel_leader_pid
;
94 ProcNumber parallel_leader_proc_number
;
97 SerializableXactHandle serializable_xact_handle
;
99 /* Mutex protects remaining fields. */
102 /* Maximum XactLastRecEnd of any worker. */
103 XLogRecPtr last_xlog_end
;
104 } FixedParallelState
;
107 * Our parallel worker number. We initialize this to -1, meaning that we are
108 * not a parallel worker. In parallel workers, it will be set to a value >= 0
109 * and < the number of workers before any user code is invoked; each parallel
110 * worker will get a different parallel worker number.
112 int ParallelWorkerNumber
= -1;
114 /* Is there a parallel message pending which we need to receive? */
115 volatile sig_atomic_t ParallelMessagePending
= false;
117 /* Are we initializing a parallel worker? */
118 bool InitializingParallelWorker
= false;
120 /* Pointer to our fixed parallel state. */
121 static FixedParallelState
*MyFixedParallelState
;
123 /* List of active parallel contexts. */
124 static dlist_head pcxt_list
= DLIST_STATIC_INIT(pcxt_list
);
126 /* Backend-local copy of data from FixedParallelState. */
127 static pid_t ParallelLeaderPid
;
130 * List of internal parallel worker entry points. We need this for
131 * reasons explained in LookupParallelWorkerFunction(), below.
136 parallel_worker_main_type fn_addr
;
137 } InternalParallelWorkers
[] =
141 "ParallelQueryMain", ParallelQueryMain
144 "_bt_parallel_build_main", _bt_parallel_build_main
147 "_brin_parallel_build_main", _brin_parallel_build_main
150 "parallel_vacuum_main", parallel_vacuum_main
154 /* Private functions. */
155 static void HandleParallelMessage(ParallelContext
*pcxt
, int i
, StringInfo msg
);
156 static void WaitForParallelWorkersToExit(ParallelContext
*pcxt
);
157 static parallel_worker_main_type
LookupParallelWorkerFunction(const char *libraryname
, const char *funcname
);
158 static void ParallelWorkerShutdown(int code
, Datum arg
);
162 * Establish a new parallel context. This should be done after entering
163 * parallel mode, and (unless there is an error) the context should be
164 * destroyed before exiting the current subtransaction.
167 CreateParallelContext(const char *library_name
, const char *function_name
,
170 MemoryContext oldcontext
;
171 ParallelContext
*pcxt
;
173 /* It is unsafe to create a parallel context if not in parallel mode. */
174 Assert(IsInParallelMode());
176 /* Number of workers should be non-negative. */
177 Assert(nworkers
>= 0);
179 /* We might be running in a short-lived memory context. */
180 oldcontext
= MemoryContextSwitchTo(TopTransactionContext
);
182 /* Initialize a new ParallelContext. */
183 pcxt
= palloc0(sizeof(ParallelContext
));
184 pcxt
->subid
= GetCurrentSubTransactionId();
185 pcxt
->nworkers
= nworkers
;
186 pcxt
->nworkers_to_launch
= nworkers
;
187 pcxt
->library_name
= pstrdup(library_name
);
188 pcxt
->function_name
= pstrdup(function_name
);
189 pcxt
->error_context_stack
= error_context_stack
;
190 shm_toc_initialize_estimator(&pcxt
->estimator
);
191 dlist_push_head(&pcxt_list
, &pcxt
->node
);
193 /* Restore previous memory context. */
194 MemoryContextSwitchTo(oldcontext
);
200 * Establish the dynamic shared memory segment for a parallel context and
201 * copy state and other bookkeeping information that will be needed by
202 * parallel workers into it.
205 InitializeParallelDSM(ParallelContext
*pcxt
)
207 MemoryContext oldcontext
;
208 Size library_len
= 0;
210 Size combocidlen
= 0;
214 Size pendingsyncslen
= 0;
216 Size relmapperlen
= 0;
217 Size uncommittedenumslen
= 0;
218 Size clientconninfolen
= 0;
221 FixedParallelState
*fps
;
222 dsm_handle session_dsm_handle
= DSM_HANDLE_INVALID
;
223 Snapshot transaction_snapshot
= GetTransactionSnapshot();
224 Snapshot active_snapshot
= GetActiveSnapshot();
226 /* We might be running in a very short-lived memory context. */
227 oldcontext
= MemoryContextSwitchTo(TopTransactionContext
);
229 /* Allow space to store the fixed-size parallel state. */
230 shm_toc_estimate_chunk(&pcxt
->estimator
, sizeof(FixedParallelState
));
231 shm_toc_estimate_keys(&pcxt
->estimator
, 1);
234 * Normally, the user will have requested at least one worker process, but
235 * if by chance they have not, we can skip a bunch of things here.
237 if (pcxt
->nworkers
> 0)
239 /* Get (or create) the per-session DSM segment's handle. */
240 session_dsm_handle
= GetSessionDsmHandle();
243 * If we weren't able to create a per-session DSM segment, then we can
244 * continue but we can't safely launch any workers because their
245 * record typmods would be incompatible so they couldn't exchange
248 if (session_dsm_handle
== DSM_HANDLE_INVALID
)
252 if (pcxt
->nworkers
> 0)
254 /* Estimate space for various kinds of state sharing. */
255 library_len
= EstimateLibraryStateSpace();
256 shm_toc_estimate_chunk(&pcxt
->estimator
, library_len
);
257 guc_len
= EstimateGUCStateSpace();
258 shm_toc_estimate_chunk(&pcxt
->estimator
, guc_len
);
259 combocidlen
= EstimateComboCIDStateSpace();
260 shm_toc_estimate_chunk(&pcxt
->estimator
, combocidlen
);
261 if (IsolationUsesXactSnapshot())
263 tsnaplen
= EstimateSnapshotSpace(transaction_snapshot
);
264 shm_toc_estimate_chunk(&pcxt
->estimator
, tsnaplen
);
266 asnaplen
= EstimateSnapshotSpace(active_snapshot
);
267 shm_toc_estimate_chunk(&pcxt
->estimator
, asnaplen
);
268 tstatelen
= EstimateTransactionStateSpace();
269 shm_toc_estimate_chunk(&pcxt
->estimator
, tstatelen
);
270 shm_toc_estimate_chunk(&pcxt
->estimator
, sizeof(dsm_handle
));
271 pendingsyncslen
= EstimatePendingSyncsSpace();
272 shm_toc_estimate_chunk(&pcxt
->estimator
, pendingsyncslen
);
273 reindexlen
= EstimateReindexStateSpace();
274 shm_toc_estimate_chunk(&pcxt
->estimator
, reindexlen
);
275 relmapperlen
= EstimateRelationMapSpace();
276 shm_toc_estimate_chunk(&pcxt
->estimator
, relmapperlen
);
277 uncommittedenumslen
= EstimateUncommittedEnumsSpace();
278 shm_toc_estimate_chunk(&pcxt
->estimator
, uncommittedenumslen
);
279 clientconninfolen
= EstimateClientConnectionInfoSpace();
280 shm_toc_estimate_chunk(&pcxt
->estimator
, clientconninfolen
);
281 /* If you add more chunks here, you probably need to add keys. */
282 shm_toc_estimate_keys(&pcxt
->estimator
, 12);
284 /* Estimate space need for error queues. */
285 StaticAssertStmt(BUFFERALIGN(PARALLEL_ERROR_QUEUE_SIZE
) ==
286 PARALLEL_ERROR_QUEUE_SIZE
,
287 "parallel error queue size not buffer-aligned");
288 shm_toc_estimate_chunk(&pcxt
->estimator
,
289 mul_size(PARALLEL_ERROR_QUEUE_SIZE
,
291 shm_toc_estimate_keys(&pcxt
->estimator
, 1);
293 /* Estimate how much we'll need for the entrypoint info. */
294 shm_toc_estimate_chunk(&pcxt
->estimator
, strlen(pcxt
->library_name
) +
295 strlen(pcxt
->function_name
) + 2);
296 shm_toc_estimate_keys(&pcxt
->estimator
, 1);
300 * Create DSM and initialize with new table of contents. But if the user
301 * didn't request any workers, then don't bother creating a dynamic shared
302 * memory segment; instead, just use backend-private memory.
304 * Also, if we can't create a dynamic shared memory segment because the
305 * maximum number of segments have already been created, then fall back to
306 * backend-private memory, and plan not to use any workers. We hope this
307 * won't happen very often, but it's better to abandon the use of
308 * parallelism than to fail outright.
310 segsize
= shm_toc_estimate(&pcxt
->estimator
);
311 if (pcxt
->nworkers
> 0)
312 pcxt
->seg
= dsm_create(segsize
, DSM_CREATE_NULL_IF_MAXSEGMENTS
);
313 if (pcxt
->seg
!= NULL
)
314 pcxt
->toc
= shm_toc_create(PARALLEL_MAGIC
,
315 dsm_segment_address(pcxt
->seg
),
320 pcxt
->private_memory
= MemoryContextAlloc(TopMemoryContext
, segsize
);
321 pcxt
->toc
= shm_toc_create(PARALLEL_MAGIC
, pcxt
->private_memory
,
325 /* Initialize fixed-size state in shared memory. */
326 fps
= (FixedParallelState
*)
327 shm_toc_allocate(pcxt
->toc
, sizeof(FixedParallelState
));
328 fps
->database_id
= MyDatabaseId
;
329 fps
->authenticated_user_id
= GetAuthenticatedUserId();
330 fps
->outer_user_id
= GetCurrentRoleId();
331 fps
->is_superuser
= current_role_is_superuser
;
332 GetUserIdAndSecContext(&fps
->current_user_id
, &fps
->sec_context
);
333 GetTempNamespaceState(&fps
->temp_namespace_id
,
334 &fps
->temp_toast_namespace_id
);
335 fps
->parallel_leader_pgproc
= MyProc
;
336 fps
->parallel_leader_pid
= MyProcPid
;
337 fps
->parallel_leader_proc_number
= MyProcNumber
;
338 fps
->xact_ts
= GetCurrentTransactionStartTimestamp();
339 fps
->stmt_ts
= GetCurrentStatementStartTimestamp();
340 fps
->serializable_xact_handle
= ShareSerializableXact();
341 SpinLockInit(&fps
->mutex
);
342 fps
->last_xlog_end
= 0;
343 shm_toc_insert(pcxt
->toc
, PARALLEL_KEY_FIXED
, fps
);
345 /* We can skip the rest of this if we're not budgeting for any workers. */
346 if (pcxt
->nworkers
> 0)
354 char *pendingsyncsspace
;
356 char *relmapperspace
;
357 char *error_queue_space
;
358 char *session_dsm_handle_space
;
359 char *entrypointstate
;
360 char *uncommittedenumsspace
;
361 char *clientconninfospace
;
364 /* Serialize shared libraries we have loaded. */
365 libraryspace
= shm_toc_allocate(pcxt
->toc
, library_len
);
366 SerializeLibraryState(library_len
, libraryspace
);
367 shm_toc_insert(pcxt
->toc
, PARALLEL_KEY_LIBRARY
, libraryspace
);
369 /* Serialize GUC settings. */
370 gucspace
= shm_toc_allocate(pcxt
->toc
, guc_len
);
371 SerializeGUCState(guc_len
, gucspace
);
372 shm_toc_insert(pcxt
->toc
, PARALLEL_KEY_GUC
, gucspace
);
374 /* Serialize combo CID state. */
375 combocidspace
= shm_toc_allocate(pcxt
->toc
, combocidlen
);
376 SerializeComboCIDState(combocidlen
, combocidspace
);
377 shm_toc_insert(pcxt
->toc
, PARALLEL_KEY_COMBO_CID
, combocidspace
);
380 * Serialize the transaction snapshot if the transaction isolation
381 * level uses a transaction snapshot.
383 if (IsolationUsesXactSnapshot())
385 tsnapspace
= shm_toc_allocate(pcxt
->toc
, tsnaplen
);
386 SerializeSnapshot(transaction_snapshot
, tsnapspace
);
387 shm_toc_insert(pcxt
->toc
, PARALLEL_KEY_TRANSACTION_SNAPSHOT
,
391 /* Serialize the active snapshot. */
392 asnapspace
= shm_toc_allocate(pcxt
->toc
, asnaplen
);
393 SerializeSnapshot(active_snapshot
, asnapspace
);
394 shm_toc_insert(pcxt
->toc
, PARALLEL_KEY_ACTIVE_SNAPSHOT
, asnapspace
);
396 /* Provide the handle for per-session segment. */
397 session_dsm_handle_space
= shm_toc_allocate(pcxt
->toc
,
399 *(dsm_handle
*) session_dsm_handle_space
= session_dsm_handle
;
400 shm_toc_insert(pcxt
->toc
, PARALLEL_KEY_SESSION_DSM
,
401 session_dsm_handle_space
);
403 /* Serialize transaction state. */
404 tstatespace
= shm_toc_allocate(pcxt
->toc
, tstatelen
);
405 SerializeTransactionState(tstatelen
, tstatespace
);
406 shm_toc_insert(pcxt
->toc
, PARALLEL_KEY_TRANSACTION_STATE
, tstatespace
);
408 /* Serialize pending syncs. */
409 pendingsyncsspace
= shm_toc_allocate(pcxt
->toc
, pendingsyncslen
);
410 SerializePendingSyncs(pendingsyncslen
, pendingsyncsspace
);
411 shm_toc_insert(pcxt
->toc
, PARALLEL_KEY_PENDING_SYNCS
,
414 /* Serialize reindex state. */
415 reindexspace
= shm_toc_allocate(pcxt
->toc
, reindexlen
);
416 SerializeReindexState(reindexlen
, reindexspace
);
417 shm_toc_insert(pcxt
->toc
, PARALLEL_KEY_REINDEX_STATE
, reindexspace
);
419 /* Serialize relmapper state. */
420 relmapperspace
= shm_toc_allocate(pcxt
->toc
, relmapperlen
);
421 SerializeRelationMap(relmapperlen
, relmapperspace
);
422 shm_toc_insert(pcxt
->toc
, PARALLEL_KEY_RELMAPPER_STATE
,
425 /* Serialize uncommitted enum state. */
426 uncommittedenumsspace
= shm_toc_allocate(pcxt
->toc
,
427 uncommittedenumslen
);
428 SerializeUncommittedEnums(uncommittedenumsspace
, uncommittedenumslen
);
429 shm_toc_insert(pcxt
->toc
, PARALLEL_KEY_UNCOMMITTEDENUMS
,
430 uncommittedenumsspace
);
432 /* Serialize our ClientConnectionInfo. */
433 clientconninfospace
= shm_toc_allocate(pcxt
->toc
, clientconninfolen
);
434 SerializeClientConnectionInfo(clientconninfolen
, clientconninfospace
);
435 shm_toc_insert(pcxt
->toc
, PARALLEL_KEY_CLIENTCONNINFO
,
436 clientconninfospace
);
438 /* Allocate space for worker information. */
439 pcxt
->worker
= palloc0(sizeof(ParallelWorkerInfo
) * pcxt
->nworkers
);
442 * Establish error queues in dynamic shared memory.
444 * These queues should be used only for transmitting ErrorResponse,
445 * NoticeResponse, and NotifyResponse protocol messages. Tuple data
446 * should be transmitted via separate (possibly larger?) queues.
449 shm_toc_allocate(pcxt
->toc
,
450 mul_size(PARALLEL_ERROR_QUEUE_SIZE
,
452 for (i
= 0; i
< pcxt
->nworkers
; ++i
)
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
);
462 shm_toc_insert(pcxt
->toc
, PARALLEL_KEY_ERROR_QUEUE
, error_queue_space
);
465 * Serialize entrypoint information. It's unsafe to pass function
466 * pointers across processes, as the function pointer may be different
467 * in each process in EXEC_BACKEND builds, so we always pass library
468 * and function name. (We use library name "postgres" for functions
469 * in the core backend.)
471 lnamelen
= strlen(pcxt
->library_name
);
472 entrypointstate
= shm_toc_allocate(pcxt
->toc
, lnamelen
+
473 strlen(pcxt
->function_name
) + 2);
474 strcpy(entrypointstate
, pcxt
->library_name
);
475 strcpy(entrypointstate
+ lnamelen
+ 1, pcxt
->function_name
);
476 shm_toc_insert(pcxt
->toc
, PARALLEL_KEY_ENTRYPOINT
, entrypointstate
);
479 /* Restore previous memory context. */
480 MemoryContextSwitchTo(oldcontext
);
484 * Reinitialize the dynamic shared memory segment for a parallel context such
485 * that we could launch workers for it again.
488 ReinitializeParallelDSM(ParallelContext
*pcxt
)
490 FixedParallelState
*fps
;
492 /* Wait for any old workers to exit. */
493 if (pcxt
->nworkers_launched
> 0)
495 WaitForParallelWorkersToFinish(pcxt
);
496 WaitForParallelWorkersToExit(pcxt
);
497 pcxt
->nworkers_launched
= 0;
498 if (pcxt
->known_attached_workers
)
500 pfree(pcxt
->known_attached_workers
);
501 pcxt
->known_attached_workers
= NULL
;
502 pcxt
->nknown_attached_workers
= 0;
506 /* Reset a few bits of fixed parallel state to a clean state. */
507 fps
= shm_toc_lookup(pcxt
->toc
, PARALLEL_KEY_FIXED
, false);
508 fps
->last_xlog_end
= 0;
510 /* Recreate error queues (if they exist). */
511 if (pcxt
->nworkers
> 0)
513 char *error_queue_space
;
517 shm_toc_lookup(pcxt
->toc
, PARALLEL_KEY_ERROR_QUEUE
, false);
518 for (i
= 0; i
< pcxt
->nworkers
; ++i
)
523 start
= error_queue_space
+ i
* PARALLEL_ERROR_QUEUE_SIZE
;
524 mq
= shm_mq_create(start
, PARALLEL_ERROR_QUEUE_SIZE
);
525 shm_mq_set_receiver(mq
, MyProc
);
526 pcxt
->worker
[i
].error_mqh
= shm_mq_attach(mq
, pcxt
->seg
, NULL
);
532 * Reinitialize parallel workers for a parallel context such that we could
533 * launch a different number of workers. This is required for cases where
534 * we need to reuse the same DSM segment, but the number of workers can
535 * vary from run-to-run.
538 ReinitializeParallelWorkers(ParallelContext
*pcxt
, int nworkers_to_launch
)
541 * The number of workers that need to be launched must be less than the
542 * number of workers with which the parallel context is initialized.
544 Assert(pcxt
->nworkers
>= nworkers_to_launch
);
545 pcxt
->nworkers_to_launch
= nworkers_to_launch
;
549 * Launch parallel workers.
552 LaunchParallelWorkers(ParallelContext
*pcxt
)
554 MemoryContext oldcontext
;
555 BackgroundWorker worker
;
557 bool any_registrations_failed
= false;
559 /* Skip this if we have no workers. */
560 if (pcxt
->nworkers
== 0 || pcxt
->nworkers_to_launch
== 0)
563 /* We need to be a lock group leader. */
564 BecomeLockGroupLeader();
566 /* If we do have workers, we'd better have a DSM segment. */
567 Assert(pcxt
->seg
!= NULL
);
569 /* We might be running in a short-lived memory context. */
570 oldcontext
= MemoryContextSwitchTo(TopTransactionContext
);
572 /* Configure a worker. */
573 memset(&worker
, 0, sizeof(worker
));
574 snprintf(worker
.bgw_name
, BGW_MAXLEN
, "parallel worker for PID %d",
576 snprintf(worker
.bgw_type
, BGW_MAXLEN
, "parallel worker");
578 BGWORKER_SHMEM_ACCESS
| BGWORKER_BACKEND_DATABASE_CONNECTION
579 | BGWORKER_CLASS_PARALLEL
;
580 worker
.bgw_start_time
= BgWorkerStart_ConsistentState
;
581 worker
.bgw_restart_time
= BGW_NEVER_RESTART
;
582 sprintf(worker
.bgw_library_name
, "postgres");
583 sprintf(worker
.bgw_function_name
, "ParallelWorkerMain");
584 worker
.bgw_main_arg
= UInt32GetDatum(dsm_segment_handle(pcxt
->seg
));
585 worker
.bgw_notify_pid
= MyProcPid
;
590 * The caller must be able to tolerate ending up with fewer workers than
591 * expected, so there is no need to throw an error here if registration
592 * fails. It wouldn't help much anyway, because registering the worker in
593 * no way guarantees that it will start up and initialize successfully.
595 for (i
= 0; i
< pcxt
->nworkers_to_launch
; ++i
)
597 memcpy(worker
.bgw_extra
, &i
, sizeof(int));
598 if (!any_registrations_failed
&&
599 RegisterDynamicBackgroundWorker(&worker
,
600 &pcxt
->worker
[i
].bgwhandle
))
602 shm_mq_set_handle(pcxt
->worker
[i
].error_mqh
,
603 pcxt
->worker
[i
].bgwhandle
);
604 pcxt
->nworkers_launched
++;
609 * If we weren't able to register the worker, then we've bumped up
610 * against the max_worker_processes limit, and future
611 * registrations will probably fail too, so arrange to skip them.
612 * But we still have to execute this code for the remaining slots
613 * to make sure that we forget about the error queues we budgeted
614 * for those workers. Otherwise, we'll wait for them to start,
615 * but they never will.
617 any_registrations_failed
= true;
618 pcxt
->worker
[i
].bgwhandle
= NULL
;
619 shm_mq_detach(pcxt
->worker
[i
].error_mqh
);
620 pcxt
->worker
[i
].error_mqh
= NULL
;
625 * Now that nworkers_launched has taken its final value, we can initialize
626 * known_attached_workers.
628 if (pcxt
->nworkers_launched
> 0)
630 pcxt
->known_attached_workers
=
631 palloc0(sizeof(bool) * pcxt
->nworkers_launched
);
632 pcxt
->nknown_attached_workers
= 0;
635 /* Restore previous memory context. */
636 MemoryContextSwitchTo(oldcontext
);
640 * Wait for all workers to attach to their error queues, and throw an error if
641 * any worker fails to do this.
643 * Callers can assume that if this function returns successfully, then the
644 * number of workers given by pcxt->nworkers_launched have initialized and
645 * attached to their error queues. Whether or not these workers are guaranteed
646 * to still be running depends on what code the caller asked them to run;
647 * this function does not guarantee that they have not exited. However, it
648 * does guarantee that any workers which exited must have done so cleanly and
649 * after successfully performing the work with which they were tasked.
651 * If this function is not called, then some of the workers that were launched
652 * may not have been started due to a fork() failure, or may have exited during
653 * early startup prior to attaching to the error queue, so nworkers_launched
654 * cannot be viewed as completely reliable. It will never be less than the
655 * number of workers which actually started, but it might be more. Any workers
656 * that failed to start will still be discovered by
657 * WaitForParallelWorkersToFinish and an error will be thrown at that time,
658 * provided that function is eventually reached.
660 * In general, the leader process should do as much work as possible before
661 * calling this function. fork() failures and other early-startup failures
662 * are very uncommon, and having the leader sit idle when it could be doing
663 * useful work is undesirable. However, if the leader needs to wait for
664 * all of its workers or for a specific worker, it may want to call this
665 * function before doing so. If not, it must make some other provision for
666 * the failure-to-start case, lest it wait forever. On the other hand, a
667 * leader which never waits for a worker that might not be started yet, or
668 * at least never does so prior to WaitForParallelWorkersToFinish(), need not
669 * call this function at all.
672 WaitForParallelWorkersToAttach(ParallelContext
*pcxt
)
676 /* Skip this if we have no launched workers. */
677 if (pcxt
->nworkers_launched
== 0)
683 * This will process any parallel messages that are pending and it may
684 * also throw an error propagated from a worker.
686 CHECK_FOR_INTERRUPTS();
688 for (i
= 0; i
< pcxt
->nworkers_launched
; ++i
)
690 BgwHandleStatus status
;
695 if (pcxt
->known_attached_workers
[i
])
699 * If error_mqh is NULL, then the worker has already exited
702 if (pcxt
->worker
[i
].error_mqh
== NULL
)
704 pcxt
->known_attached_workers
[i
] = true;
705 ++pcxt
->nknown_attached_workers
;
709 status
= GetBackgroundWorkerPid(pcxt
->worker
[i
].bgwhandle
, &pid
);
710 if (status
== BGWH_STARTED
)
712 /* Has the worker attached to the error queue? */
713 mq
= shm_mq_get_queue(pcxt
->worker
[i
].error_mqh
);
714 if (shm_mq_get_sender(mq
) != NULL
)
716 /* Yes, so it is known to be attached. */
717 pcxt
->known_attached_workers
[i
] = true;
718 ++pcxt
->nknown_attached_workers
;
721 else if (status
== BGWH_STOPPED
)
724 * If the worker stopped without attaching to the error queue,
727 mq
= shm_mq_get_queue(pcxt
->worker
[i
].error_mqh
);
728 if (shm_mq_get_sender(mq
) == NULL
)
730 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE
),
731 errmsg("parallel worker failed to initialize"),
732 errhint("More details may be available in the server log.")));
734 pcxt
->known_attached_workers
[i
] = true;
735 ++pcxt
->nknown_attached_workers
;
740 * Worker not yet started, so we must wait. The postmaster
741 * will notify us if the worker's state changes. Our latch
742 * might also get set for some other reason, but if so we'll
743 * just end up waiting for the same worker again.
745 rc
= WaitLatch(MyLatch
,
746 WL_LATCH_SET
| WL_EXIT_ON_PM_DEATH
,
747 -1, WAIT_EVENT_BGWORKER_STARTUP
);
749 if (rc
& WL_LATCH_SET
)
754 /* If all workers are known to have started, we're done. */
755 if (pcxt
->nknown_attached_workers
>= pcxt
->nworkers_launched
)
757 Assert(pcxt
->nknown_attached_workers
== pcxt
->nworkers_launched
);
764 * Wait for all workers to finish computing.
766 * Even if the parallel operation seems to have completed successfully, it's
767 * important to call this function afterwards. We must not miss any errors
768 * the workers may have thrown during the parallel operation, or any that they
769 * may yet throw while shutting down.
771 * Also, we want to update our notion of XactLastRecEnd based on worker
775 WaitForParallelWorkersToFinish(ParallelContext
*pcxt
)
779 bool anyone_alive
= false;
784 * This will process any parallel messages that are pending, which may
785 * change the outcome of the loop that follows. It may also throw an
786 * error propagated from a worker.
788 CHECK_FOR_INTERRUPTS();
790 for (i
= 0; i
< pcxt
->nworkers_launched
; ++i
)
793 * If error_mqh is NULL, then the worker has already exited
794 * cleanly. If we have received a message through error_mqh from
795 * the worker, we know it started up cleanly, and therefore we're
796 * certain to be notified when it exits.
798 if (pcxt
->worker
[i
].error_mqh
== NULL
)
800 else if (pcxt
->known_attached_workers
[i
])
809 /* If all workers are known to have finished, we're done. */
810 if (nfinished
>= pcxt
->nworkers_launched
)
812 Assert(nfinished
== pcxt
->nworkers_launched
);
817 * We didn't detect any living workers, but not all workers are
818 * known to have exited cleanly. Either not all workers have
819 * launched yet, or maybe some of them failed to start or
820 * terminated abnormally.
822 for (i
= 0; i
< pcxt
->nworkers_launched
; ++i
)
828 * If the worker is BGWH_NOT_YET_STARTED or BGWH_STARTED, we
829 * should just keep waiting. If it is BGWH_STOPPED, then
830 * further investigation is needed.
832 if (pcxt
->worker
[i
].error_mqh
== NULL
||
833 pcxt
->worker
[i
].bgwhandle
== NULL
||
834 GetBackgroundWorkerPid(pcxt
->worker
[i
].bgwhandle
,
835 &pid
) != BGWH_STOPPED
)
839 * Check whether the worker ended up stopped without ever
840 * attaching to the error queue. If so, the postmaster was
841 * unable to fork the worker or it exited without initializing
842 * properly. We must throw an error, since the caller may
843 * have been expecting the worker to do some work before
846 mq
= shm_mq_get_queue(pcxt
->worker
[i
].error_mqh
);
847 if (shm_mq_get_sender(mq
) == NULL
)
849 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE
),
850 errmsg("parallel worker failed to initialize"),
851 errhint("More details may be available in the server log.")));
854 * The worker is stopped, but is attached to the error queue.
855 * Unless there's a bug somewhere, this will only happen when
856 * the worker writes messages and terminates after the
857 * CHECK_FOR_INTERRUPTS() near the top of this function and
858 * before the call to GetBackgroundWorkerPid(). In that case,
859 * or latch should have been set as well and the right things
860 * will happen on the next pass through the loop.
865 (void) WaitLatch(MyLatch
, WL_LATCH_SET
| WL_EXIT_ON_PM_DEATH
, -1,
866 WAIT_EVENT_PARALLEL_FINISH
);
870 if (pcxt
->toc
!= NULL
)
872 FixedParallelState
*fps
;
874 fps
= shm_toc_lookup(pcxt
->toc
, PARALLEL_KEY_FIXED
, false);
875 if (fps
->last_xlog_end
> XactLastRecEnd
)
876 XactLastRecEnd
= fps
->last_xlog_end
;
881 * Wait for all workers to exit.
883 * This function ensures that workers have been completely shutdown. The
884 * difference between WaitForParallelWorkersToFinish and this function is
885 * that the former just ensures that last message sent by a worker backend is
886 * received by the leader backend whereas this ensures the complete shutdown.
889 WaitForParallelWorkersToExit(ParallelContext
*pcxt
)
893 /* Wait until the workers actually die. */
894 for (i
= 0; i
< pcxt
->nworkers_launched
; ++i
)
896 BgwHandleStatus status
;
898 if (pcxt
->worker
== NULL
|| pcxt
->worker
[i
].bgwhandle
== NULL
)
901 status
= WaitForBackgroundWorkerShutdown(pcxt
->worker
[i
].bgwhandle
);
904 * If the postmaster kicked the bucket, we have no chance of cleaning
905 * up safely -- we won't be able to tell when our workers are actually
906 * dead. This doesn't necessitate a PANIC since they will all abort
907 * eventually, but we can't safely continue this session.
909 if (status
== BGWH_POSTMASTER_DIED
)
911 (errcode(ERRCODE_ADMIN_SHUTDOWN
),
912 errmsg("postmaster exited during a parallel transaction")));
914 /* Release memory. */
915 pfree(pcxt
->worker
[i
].bgwhandle
);
916 pcxt
->worker
[i
].bgwhandle
= NULL
;
921 * Destroy a parallel context.
923 * If expecting a clean exit, you should use WaitForParallelWorkersToFinish()
924 * first, before calling this function. When this function is invoked, any
925 * remaining workers are forcibly killed; the dynamic shared memory segment
926 * is unmapped; and we then wait (uninterruptibly) for the workers to exit.
929 DestroyParallelContext(ParallelContext
*pcxt
)
934 * Be careful about order of operations here! We remove the parallel
935 * context from the list before we do anything else; otherwise, if an
936 * error occurs during a subsequent step, we might try to nuke it again
937 * from AtEOXact_Parallel or AtEOSubXact_Parallel.
939 dlist_delete(&pcxt
->node
);
941 /* Kill each worker in turn, and forget their error queues. */
942 if (pcxt
->worker
!= NULL
)
944 for (i
= 0; i
< pcxt
->nworkers_launched
; ++i
)
946 if (pcxt
->worker
[i
].error_mqh
!= NULL
)
948 TerminateBackgroundWorker(pcxt
->worker
[i
].bgwhandle
);
950 shm_mq_detach(pcxt
->worker
[i
].error_mqh
);
951 pcxt
->worker
[i
].error_mqh
= NULL
;
957 * If we have allocated a shared memory segment, detach it. This will
958 * implicitly detach the error queues, and any other shared memory queues,
961 if (pcxt
->seg
!= NULL
)
963 dsm_detach(pcxt
->seg
);
968 * If this parallel context is actually in backend-private memory rather
969 * than shared memory, free that memory instead.
971 if (pcxt
->private_memory
!= NULL
)
973 pfree(pcxt
->private_memory
);
974 pcxt
->private_memory
= NULL
;
978 * We can't finish transaction commit or abort until all of the workers
979 * have exited. This means, in particular, that we can't respond to
980 * interrupts at this stage.
983 WaitForParallelWorkersToExit(pcxt
);
986 /* Free the worker array itself. */
987 if (pcxt
->worker
!= NULL
)
994 pfree(pcxt
->library_name
);
995 pfree(pcxt
->function_name
);
1000 * Are there any parallel contexts currently active?
1003 ParallelContextActive(void)
1005 return !dlist_is_empty(&pcxt_list
);
1009 * Handle receipt of an interrupt indicating a parallel worker message.
1011 * Note: this is called within a signal handler! All we can do is set
1012 * a flag that will cause the next CHECK_FOR_INTERRUPTS() to invoke
1013 * HandleParallelMessages().
1016 HandleParallelMessageInterrupt(void)
1018 InterruptPending
= true;
1019 ParallelMessagePending
= true;
1024 * Handle any queued protocol messages received from parallel workers.
1027 HandleParallelMessages(void)
1030 MemoryContext oldcontext
;
1032 static MemoryContext hpm_context
= NULL
;
1035 * This is invoked from ProcessInterrupts(), and since some of the
1036 * functions it calls contain CHECK_FOR_INTERRUPTS(), there is a potential
1037 * for recursive calls if more signals are received while this runs. It's
1038 * unclear that recursive entry would be safe, and it doesn't seem useful
1039 * even if it is safe, so let's block interrupts until done.
1044 * Moreover, CurrentMemoryContext might be pointing almost anywhere. We
1045 * don't want to risk leaking data into long-lived contexts, so let's do
1046 * our work here in a private context that we can reset on each use.
1048 if (hpm_context
== NULL
) /* first time through? */
1049 hpm_context
= AllocSetContextCreate(TopMemoryContext
,
1050 "HandleParallelMessages",
1051 ALLOCSET_DEFAULT_SIZES
);
1053 MemoryContextReset(hpm_context
);
1055 oldcontext
= MemoryContextSwitchTo(hpm_context
);
1057 /* OK to process messages. Reset the flag saying there are more to do. */
1058 ParallelMessagePending
= false;
1060 dlist_foreach(iter
, &pcxt_list
)
1062 ParallelContext
*pcxt
;
1065 pcxt
= dlist_container(ParallelContext
, node
, iter
.cur
);
1066 if (pcxt
->worker
== NULL
)
1069 for (i
= 0; i
< pcxt
->nworkers_launched
; ++i
)
1072 * Read as many messages as we can from each worker, but stop when
1073 * either (1) the worker's error queue goes away, which can happen
1074 * if we receive a Terminate message from the worker; or (2) no
1075 * more messages can be read from the worker without blocking.
1077 while (pcxt
->worker
[i
].error_mqh
!= NULL
)
1083 res
= shm_mq_receive(pcxt
->worker
[i
].error_mqh
, &nbytes
,
1085 if (res
== SHM_MQ_WOULD_BLOCK
)
1087 else if (res
== SHM_MQ_SUCCESS
)
1091 initStringInfo(&msg
);
1092 appendBinaryStringInfo(&msg
, data
, nbytes
);
1093 HandleParallelMessage(pcxt
, i
, &msg
);
1098 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE
),
1099 errmsg("lost connection to parallel worker")));
1104 MemoryContextSwitchTo(oldcontext
);
1106 /* Might as well clear the context on our way out */
1107 MemoryContextReset(hpm_context
);
1109 RESUME_INTERRUPTS();
1113 * Handle a single protocol message received from a single parallel worker.
1116 HandleParallelMessage(ParallelContext
*pcxt
, int i
, StringInfo msg
)
1120 if (pcxt
->known_attached_workers
!= NULL
&&
1121 !pcxt
->known_attached_workers
[i
])
1123 pcxt
->known_attached_workers
[i
] = true;
1124 pcxt
->nknown_attached_workers
++;
1127 msgtype
= pq_getmsgbyte(msg
);
1131 case PqMsg_ErrorResponse
:
1132 case PqMsg_NoticeResponse
:
1135 ErrorContextCallback
*save_error_context_stack
;
1137 /* Parse ErrorResponse or NoticeResponse. */
1138 pq_parse_errornotice(msg
, &edata
);
1140 /* Death of a worker isn't enough justification for suicide. */
1141 edata
.elevel
= Min(edata
.elevel
, ERROR
);
1144 * If desired, add a context line to show that this is a
1145 * message propagated from a parallel worker. Otherwise, it
1146 * can sometimes be confusing to understand what actually
1147 * happened. (We don't do this in DEBUG_PARALLEL_REGRESS mode
1148 * because it causes test-result instability depending on
1149 * whether a parallel worker is actually used or not.)
1151 if (debug_parallel_query
!= DEBUG_PARALLEL_REGRESS
)
1154 edata
.context
= psprintf("%s\n%s", edata
.context
,
1155 _("parallel worker"));
1157 edata
.context
= pstrdup(_("parallel worker"));
1161 * Context beyond that should use the error context callbacks
1162 * that were in effect when the ParallelContext was created,
1163 * not the current ones.
1165 save_error_context_stack
= error_context_stack
;
1166 error_context_stack
= pcxt
->error_context_stack
;
1168 /* Rethrow error or print notice. */
1169 ThrowErrorData(&edata
);
1171 /* Not an error, so restore previous context stack. */
1172 error_context_stack
= save_error_context_stack
;
1177 case PqMsg_NotificationResponse
:
1179 /* Propagate NotifyResponse. */
1181 const char *channel
;
1182 const char *payload
;
1184 pid
= pq_getmsgint(msg
, 4);
1185 channel
= pq_getmsgrawstring(msg
);
1186 payload
= pq_getmsgrawstring(msg
);
1189 NotifyMyFrontEnd(channel
, payload
, pid
);
1194 case 'P': /* Parallel progress reporting */
1197 * Only incremental progress reporting is currently supported.
1198 * However, it's possible to add more fields to the message to
1199 * allow for handling of other backend progress APIs.
1201 int index
= pq_getmsgint(msg
, 4);
1202 int64 incr
= pq_getmsgint64(msg
);
1206 pgstat_progress_incr_param(index
, incr
);
1211 case PqMsg_Terminate
:
1213 shm_mq_detach(pcxt
->worker
[i
].error_mqh
);
1214 pcxt
->worker
[i
].error_mqh
= NULL
;
1220 elog(ERROR
, "unrecognized message type received from parallel worker: %c (message length %d bytes)",
1227 * End-of-subtransaction cleanup for parallel contexts.
1229 * Here we remove only parallel contexts initiated within the current
1233 AtEOSubXact_Parallel(bool isCommit
, SubTransactionId mySubId
)
1235 while (!dlist_is_empty(&pcxt_list
))
1237 ParallelContext
*pcxt
;
1239 pcxt
= dlist_head_element(ParallelContext
, node
, &pcxt_list
);
1240 if (pcxt
->subid
!= mySubId
)
1243 elog(WARNING
, "leaked parallel context");
1244 DestroyParallelContext(pcxt
);
1249 * End-of-transaction cleanup for parallel contexts.
1251 * We nuke all remaining parallel contexts.
1254 AtEOXact_Parallel(bool isCommit
)
1256 while (!dlist_is_empty(&pcxt_list
))
1258 ParallelContext
*pcxt
;
1260 pcxt
= dlist_head_element(ParallelContext
, node
, &pcxt_list
);
1262 elog(WARNING
, "leaked parallel context");
1263 DestroyParallelContext(pcxt
);
1268 * Main entrypoint for parallel workers.
1271 ParallelWorkerMain(Datum main_arg
)
1275 FixedParallelState
*fps
;
1276 char *error_queue_space
;
1280 char *entrypointstate
;
1282 char *function_name
;
1283 parallel_worker_main_type entrypt
;
1285 char *combocidspace
;
1289 char *pendingsyncsspace
;
1291 char *relmapperspace
;
1292 char *uncommittedenumsspace
;
1293 char *clientconninfospace
;
1294 char *session_dsm_handle_space
;
1298 /* Set flag to indicate that we're initializing a parallel worker. */
1299 InitializingParallelWorker
= true;
1301 /* Establish signal handlers. */
1302 pqsignal(SIGTERM
, die
);
1303 BackgroundWorkerUnblockSignals();
1305 /* Determine and set our parallel worker number. */
1306 Assert(ParallelWorkerNumber
== -1);
1307 memcpy(&ParallelWorkerNumber
, MyBgworkerEntry
->bgw_extra
, sizeof(int));
1309 /* Set up a memory context to work in, just for cleanliness. */
1310 CurrentMemoryContext
= AllocSetContextCreate(TopMemoryContext
,
1312 ALLOCSET_DEFAULT_SIZES
);
1315 * Attach to the dynamic shared memory segment for the parallel query, and
1316 * find its table of contents.
1318 * Note: at this point, we have not created any ResourceOwner in this
1319 * process. This will result in our DSM mapping surviving until process
1320 * exit, which is fine. If there were a ResourceOwner, it would acquire
1321 * ownership of the mapping, but we have no need for that.
1323 seg
= dsm_attach(DatumGetUInt32(main_arg
));
1326 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE
),
1327 errmsg("could not map dynamic shared memory segment")));
1328 toc
= shm_toc_attach(PARALLEL_MAGIC
, dsm_segment_address(seg
));
1331 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE
),
1332 errmsg("invalid magic number in dynamic shared memory segment")));
1334 /* Look up fixed parallel state. */
1335 fps
= shm_toc_lookup(toc
, PARALLEL_KEY_FIXED
, false);
1336 MyFixedParallelState
= fps
;
1338 /* Arrange to signal the leader if we exit. */
1339 ParallelLeaderPid
= fps
->parallel_leader_pid
;
1340 ParallelLeaderProcNumber
= fps
->parallel_leader_proc_number
;
1341 before_shmem_exit(ParallelWorkerShutdown
, PointerGetDatum(seg
));
1344 * Now we can find and attach to the error queue provided for us. That's
1345 * good, because until we do that, any errors that happen here will not be
1346 * reported back to the process that requested that this worker be
1349 error_queue_space
= shm_toc_lookup(toc
, PARALLEL_KEY_ERROR_QUEUE
, false);
1350 mq
= (shm_mq
*) (error_queue_space
+
1351 ParallelWorkerNumber
* PARALLEL_ERROR_QUEUE_SIZE
);
1352 shm_mq_set_sender(mq
, MyProc
);
1353 mqh
= shm_mq_attach(mq
, seg
, NULL
);
1354 pq_redirect_to_shm_mq(seg
, mqh
);
1355 pq_set_parallel_leader(fps
->parallel_leader_pid
,
1356 fps
->parallel_leader_proc_number
);
1359 * Hooray! Primary initialization is complete. Now, we need to set up our
1360 * backend-local state to match the original backend.
1364 * Join locking group. We must do this before anything that could try to
1365 * acquire a heavyweight lock, because any heavyweight locks acquired to
1366 * this point could block either directly against the parallel group
1367 * leader or against some process which in turn waits for a lock that
1368 * conflicts with the parallel group leader, causing an undetected
1369 * deadlock. (If we can't join the lock group, the leader has gone away,
1370 * so just exit quietly.)
1372 if (!BecomeLockGroupMember(fps
->parallel_leader_pgproc
,
1373 fps
->parallel_leader_pid
))
1377 * Restore transaction and statement start-time timestamps. This must
1378 * happen before anything that would start a transaction, else asserts in
1381 SetParallelStartTimestamps(fps
->xact_ts
, fps
->stmt_ts
);
1384 * Identify the entry point to be called. In theory this could result in
1385 * loading an additional library, though most likely the entry point is in
1386 * the core backend or in a library we just loaded.
1388 entrypointstate
= shm_toc_lookup(toc
, PARALLEL_KEY_ENTRYPOINT
, false);
1389 library_name
= entrypointstate
;
1390 function_name
= entrypointstate
+ strlen(library_name
) + 1;
1392 entrypt
= LookupParallelWorkerFunction(library_name
, function_name
);
1394 /* Restore database connection. */
1395 BackgroundWorkerInitializeConnectionByOid(fps
->database_id
,
1396 fps
->authenticated_user_id
,
1400 * Set the client encoding to the database encoding, since that is what
1401 * the leader will expect.
1403 SetClientEncoding(GetDatabaseEncoding());
1406 * Load libraries that were loaded by original backend. We want to do
1407 * this before restoring GUCs, because the libraries might define custom
1410 libraryspace
= shm_toc_lookup(toc
, PARALLEL_KEY_LIBRARY
, false);
1411 StartTransactionCommand();
1412 RestoreLibraryState(libraryspace
);
1414 /* Restore GUC values from launching backend. */
1415 gucspace
= shm_toc_lookup(toc
, PARALLEL_KEY_GUC
, false);
1416 RestoreGUCState(gucspace
);
1417 CommitTransactionCommand();
1419 /* Crank up a transaction state appropriate to a parallel worker. */
1420 tstatespace
= shm_toc_lookup(toc
, PARALLEL_KEY_TRANSACTION_STATE
, false);
1421 StartParallelWorkerTransaction(tstatespace
);
1423 /* Restore combo CID state. */
1424 combocidspace
= shm_toc_lookup(toc
, PARALLEL_KEY_COMBO_CID
, false);
1425 RestoreComboCIDState(combocidspace
);
1427 /* Attach to the per-session DSM segment and contained objects. */
1428 session_dsm_handle_space
=
1429 shm_toc_lookup(toc
, PARALLEL_KEY_SESSION_DSM
, false);
1430 AttachSession(*(dsm_handle
*) session_dsm_handle_space
);
1433 * If the transaction isolation level is REPEATABLE READ or SERIALIZABLE,
1434 * the leader has serialized the transaction snapshot and we must restore
1435 * it. At lower isolation levels, there is no transaction-lifetime
1436 * snapshot, but we need TransactionXmin to get set to a value which is
1437 * less than or equal to the xmin of every snapshot that will be used by
1438 * this worker. The easiest way to accomplish that is to install the
1439 * active snapshot as the transaction snapshot. Code running in this
1440 * parallel worker might take new snapshots via GetTransactionSnapshot()
1441 * or GetLatestSnapshot(), but it shouldn't have any way of acquiring a
1442 * snapshot older than the active snapshot.
1444 asnapspace
= shm_toc_lookup(toc
, PARALLEL_KEY_ACTIVE_SNAPSHOT
, false);
1445 tsnapspace
= shm_toc_lookup(toc
, PARALLEL_KEY_TRANSACTION_SNAPSHOT
, true);
1446 asnapshot
= RestoreSnapshot(asnapspace
);
1447 tsnapshot
= tsnapspace
? RestoreSnapshot(tsnapspace
) : asnapshot
;
1448 RestoreTransactionSnapshot(tsnapshot
,
1449 fps
->parallel_leader_pgproc
);
1450 PushActiveSnapshot(asnapshot
);
1453 * We've changed which tuples we can see, and must therefore invalidate
1456 InvalidateSystemCaches();
1459 * Restore current role id. Skip verifying whether session user is
1460 * allowed to become this role and blindly restore the leader's state for
1463 SetCurrentRoleId(fps
->outer_user_id
, fps
->is_superuser
);
1465 /* Restore user ID and security context. */
1466 SetUserIdAndSecContext(fps
->current_user_id
, fps
->sec_context
);
1468 /* Restore temp-namespace state to ensure search path matches leader's. */
1469 SetTempNamespaceState(fps
->temp_namespace_id
,
1470 fps
->temp_toast_namespace_id
);
1472 /* Restore pending syncs. */
1473 pendingsyncsspace
= shm_toc_lookup(toc
, PARALLEL_KEY_PENDING_SYNCS
,
1475 RestorePendingSyncs(pendingsyncsspace
);
1477 /* Restore reindex state. */
1478 reindexspace
= shm_toc_lookup(toc
, PARALLEL_KEY_REINDEX_STATE
, false);
1479 RestoreReindexState(reindexspace
);
1481 /* Restore relmapper state. */
1482 relmapperspace
= shm_toc_lookup(toc
, PARALLEL_KEY_RELMAPPER_STATE
, false);
1483 RestoreRelationMap(relmapperspace
);
1485 /* Restore uncommitted enums. */
1486 uncommittedenumsspace
= shm_toc_lookup(toc
, PARALLEL_KEY_UNCOMMITTEDENUMS
,
1488 RestoreUncommittedEnums(uncommittedenumsspace
);
1490 /* Restore the ClientConnectionInfo. */
1491 clientconninfospace
= shm_toc_lookup(toc
, PARALLEL_KEY_CLIENTCONNINFO
,
1493 RestoreClientConnectionInfo(clientconninfospace
);
1496 * Initialize SystemUser now that MyClientConnectionInfo is restored. Also
1497 * ensure that auth_method is actually valid, aka authn_id is not NULL.
1499 if (MyClientConnectionInfo
.authn_id
)
1500 InitializeSystemUser(MyClientConnectionInfo
.authn_id
,
1501 hba_authname(MyClientConnectionInfo
.auth_method
));
1503 /* Attach to the leader's serializable transaction, if SERIALIZABLE. */
1504 AttachSerializableXact(fps
->serializable_xact_handle
);
1507 * We've initialized all of our state now; nothing should change
1510 InitializingParallelWorker
= false;
1511 EnterParallelMode();
1514 * Time to do the real work: invoke the caller-supplied code.
1518 /* Must exit parallel mode to pop active snapshot. */
1521 /* Must pop active snapshot so snapmgr.c doesn't complain. */
1522 PopActiveSnapshot();
1524 /* Shut down the parallel-worker transaction. */
1525 EndParallelWorkerTransaction();
1527 /* Detach from the per-session DSM segment. */
1530 /* Report success. */
1531 pq_putmessage(PqMsg_Terminate
, NULL
, 0);
1535 * Update shared memory with the ending location of the last WAL record we
1536 * wrote, if it's greater than the value already stored there.
1539 ParallelWorkerReportLastRecEnd(XLogRecPtr last_xlog_end
)
1541 FixedParallelState
*fps
= MyFixedParallelState
;
1543 Assert(fps
!= NULL
);
1544 SpinLockAcquire(&fps
->mutex
);
1545 if (fps
->last_xlog_end
< last_xlog_end
)
1546 fps
->last_xlog_end
= last_xlog_end
;
1547 SpinLockRelease(&fps
->mutex
);
1551 * Make sure the leader tries to read from our error queue one more time.
1552 * This guards against the case where we exit uncleanly without sending an
1553 * ErrorResponse to the leader, for example because some code calls proc_exit
1556 * Also explicitly detach from dsm segment so that subsystems using
1557 * on_dsm_detach() have a chance to send stats before the stats subsystem is
1558 * shut down as part of a before_shmem_exit() hook.
1560 * One might think this could instead be solved by carefully ordering the
1561 * attaching to dsm segments, so that the pgstats segments get detached from
1562 * later than the parallel query one. That turns out to not work because the
1563 * stats hash might need to grow which can cause new segments to be allocated,
1564 * which then will be detached from earlier.
1567 ParallelWorkerShutdown(int code
, Datum arg
)
1569 SendProcSignal(ParallelLeaderPid
,
1570 PROCSIG_PARALLEL_MESSAGE
,
1571 ParallelLeaderProcNumber
);
1573 dsm_detach((dsm_segment
*) DatumGetPointer(arg
));
1577 * Look up (and possibly load) a parallel worker entry point function.
1579 * For functions contained in the core code, we use library name "postgres"
1580 * and consult the InternalParallelWorkers array. External functions are
1581 * looked up, and loaded if necessary, using load_external_function().
1583 * The point of this is to pass function names as strings across process
1584 * boundaries. We can't pass actual function addresses because of the
1585 * possibility that the function has been loaded at a different address
1586 * in a different process. This is obviously a hazard for functions in
1587 * loadable libraries, but it can happen even for functions in the core code
1588 * on platforms using EXEC_BACKEND (e.g., Windows).
1590 * At some point it might be worthwhile to get rid of InternalParallelWorkers[]
1591 * in favor of applying load_external_function() for core functions too;
1592 * but that raises portability issues that are not worth addressing now.
1594 static parallel_worker_main_type
1595 LookupParallelWorkerFunction(const char *libraryname
, const char *funcname
)
1598 * If the function is to be loaded from postgres itself, search the
1599 * InternalParallelWorkers array.
1601 if (strcmp(libraryname
, "postgres") == 0)
1605 for (i
= 0; i
< lengthof(InternalParallelWorkers
); i
++)
1607 if (strcmp(InternalParallelWorkers
[i
].fn_name
, funcname
) == 0)
1608 return InternalParallelWorkers
[i
].fn_addr
;
1611 /* We can only reach this by programming error. */
1612 elog(ERROR
, "internal function \"%s\" not found", funcname
);
1615 /* Otherwise load from external library. */
1616 return (parallel_worker_main_type
)
1617 load_external_function(libraryname
, funcname
, true, NULL
);