granicus.if.org Git - postgresql/blob - src/backend/access/transam/parallel.c

   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/session.h"
  19 #include "access/xact.h"
  20 #include "access/xlog.h"
  21 #include "catalog/namespace.h"
  22 #include "commands/async.h"
  23 #include "executor/execParallel.h"
  24 #include "libpq/libpq.h"
  25 #include "libpq/pqformat.h"
  26 #include "libpq/pqmq.h"
  27 #include "miscadmin.h"
  28 #include "optimizer/planmain.h"
  29 #include "pgstat.h"
  30 #include "storage/ipc.h"
  31 #include "storage/sinval.h"
  32 #include "storage/spin.h"
  33 #include "tcop/tcopprot.h"
  34 #include "utils/combocid.h"
  35 #include "utils/guc.h"
  36 #include "utils/inval.h"
  37 #include "utils/memutils.h"
  38 #include "utils/resowner.h"
  39 #include "utils/snapmgr.h"
  40 #include "utils/typcache.h"
  41
  42
  43 /*
  44  * We don't want to waste a lot of memory on an error queue which, most of
  45  * the time, will process only a handful of small messages.  However, it is
  46  * desirable to make it large enough that a typical ErrorResponse can be sent
  47  * without blocking.  That way, a worker that errors out can write the whole
  48  * message into the queue and terminate without waiting for the user backend.
  49  */
  50 #define PARALLEL_ERROR_QUEUE_SIZE                       16384
  51
  52 /* Magic number for parallel context TOC. */
  53 #define PARALLEL_MAGIC                                          0x50477c7c
  54
  55 /*
  56  * Magic numbers for per-context parallel state sharing.  Higher-level code
  57  * should use smaller values, leaving these very large ones for use by this
  58  * module.
  59  */
  60 #define PARALLEL_KEY_FIXED                                      UINT64CONST(0xFFFFFFFFFFFF0001)
  61 #define PARALLEL_KEY_ERROR_QUEUE                        UINT64CONST(0xFFFFFFFFFFFF0002)
  62 #define PARALLEL_KEY_LIBRARY                            UINT64CONST(0xFFFFFFFFFFFF0003)
  63 #define PARALLEL_KEY_GUC                                        UINT64CONST(0xFFFFFFFFFFFF0004)
  64 #define PARALLEL_KEY_COMBO_CID                          UINT64CONST(0xFFFFFFFFFFFF0005)
  65 #define PARALLEL_KEY_TRANSACTION_SNAPSHOT       UINT64CONST(0xFFFFFFFFFFFF0006)
  66 #define PARALLEL_KEY_ACTIVE_SNAPSHOT            UINT64CONST(0xFFFFFFFFFFFF0007)
  67 #define PARALLEL_KEY_TRANSACTION_STATE          UINT64CONST(0xFFFFFFFFFFFF0008)
  68 #define PARALLEL_KEY_ENTRYPOINT                         UINT64CONST(0xFFFFFFFFFFFF0009)
  69 #define PARALLEL_KEY_SESSION_DSM                        UINT64CONST(0xFFFFFFFFFFFF000A)
  70
  71 /* Fixed-size parallel state. */
  72 typedef struct FixedParallelState
  73 {
  74         /* Fixed-size state that workers must restore. */
  75         Oid                     database_id;
  76         Oid                     authenticated_user_id;
  77         Oid                     current_user_id;
  78         Oid                     temp_namespace_id;
  79         Oid                     temp_toast_namespace_id;
  80         int                     sec_context;
  81         PGPROC     *parallel_master_pgproc;
  82         pid_t           parallel_master_pid;
  83         BackendId       parallel_master_backend_id;
  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 /*
 113  * List of internal parallel worker entry points.  We need this for
 114  * reasons explained in LookupParallelWorkerFunction(), below.
 115  */
 116 static const struct
 117 {
 118         const char *fn_name;
 119         parallel_worker_main_type fn_addr;
 120 }                       InternalParallelWorkers[] =
 121
 122 {
 123         {
 124                 "ParallelQueryMain", ParallelQueryMain
 125         }
 126 };
 127
 128 /* Private functions. */
 129 static void HandleParallelMessage(ParallelContext *pcxt, int i, StringInfo msg);
 130 static void WaitForParallelWorkersToExit(ParallelContext *pcxt);
 131 static parallel_worker_main_type LookupParallelWorkerFunction(const char *libraryname, const char *funcname);
 132
 133
 134 /*
 135  * Establish a new parallel context.  This should be done after entering
 136  * parallel mode, and (unless there is an error) the context should be
 137  * destroyed before exiting the current subtransaction.
 138  */
 139 ParallelContext *
 140 CreateParallelContext(const char *library_name, const char *function_name,
 141                                           int nworkers)
 142 {
 143         MemoryContext oldcontext;
 144         ParallelContext *pcxt;
 145
 146         /* It is unsafe to create a parallel context if not in parallel mode. */
 147         Assert(IsInParallelMode());
 148
 149         /* Number of workers should be non-negative. */
 150         Assert(nworkers >= 0);
 151
 152         /*
 153          * If dynamic shared memory is not available, we won't be able to use
 154          * background workers.
 155          */
 156         if (dynamic_shared_memory_type == DSM_IMPL_NONE)
 157                 nworkers = 0;
 158
 159         /*
 160          * If we are running under serializable isolation, we can't use parallel
 161          * workers, at least not until somebody enhances that mechanism to be
 162          * parallel-aware.
 163          */
 164         if (IsolationIsSerializable())
 165                 nworkers = 0;
 166
 167         /* We might be running in a short-lived memory context. */
 168         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 169
 170         /* Initialize a new ParallelContext. */
 171         pcxt = palloc0(sizeof(ParallelContext));
 172         pcxt->subid = GetCurrentSubTransactionId();
 173         pcxt->nworkers = nworkers;
 174         pcxt->library_name = pstrdup(library_name);
 175         pcxt->function_name = pstrdup(function_name);
 176         pcxt->error_context_stack = error_context_stack;
 177         shm_toc_initialize_estimator(&pcxt->estimator);
 178         dlist_push_head(&pcxt_list, &pcxt->node);
 179
 180         /* Restore previous memory context. */
 181         MemoryContextSwitchTo(oldcontext);
 182
 183         return pcxt;
 184 }
 185
 186 /*
 187  * Establish the dynamic shared memory segment for a parallel context and
 188  * copy state and other bookkeeping information that will be needed by
 189  * parallel workers into it.
 190  */
 191 void
 192 InitializeParallelDSM(ParallelContext *pcxt)
 193 {
 194         MemoryContext oldcontext;
 195         Size            library_len = 0;
 196         Size            guc_len = 0;
 197         Size            combocidlen = 0;
 198         Size            tsnaplen = 0;
 199         Size            asnaplen = 0;
 200         Size            tstatelen = 0;
 201         Size            segsize = 0;
 202         int                     i;
 203         FixedParallelState *fps;
 204         dsm_handle      session_dsm_handle = DSM_HANDLE_INVALID;
 205         Snapshot        transaction_snapshot = GetTransactionSnapshot();
 206         Snapshot        active_snapshot = GetActiveSnapshot();
 207
 208         /* We might be running in a very short-lived memory context. */
 209         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 210
 211         /* Allow space to store the fixed-size parallel state. */
 212         shm_toc_estimate_chunk(&pcxt->estimator, sizeof(FixedParallelState));
 213         shm_toc_estimate_keys(&pcxt->estimator, 1);
 214
 215         /*
 216          * Normally, the user will have requested at least one worker process, but
 217          * if by chance they have not, we can skip a bunch of things here.
 218          */
 219         if (pcxt->nworkers > 0)
 220         {
 221                 /* Get (or create) the per-session DSM segment's handle. */
 222                 session_dsm_handle = GetSessionDsmHandle();
 223
 224                 /*
 225                  * If we weren't able to create a per-session DSM segment, then we can
 226                  * continue but we can't safely launch any workers because their
 227                  * record typmods would be incompatible so they couldn't exchange
 228                  * tuples.
 229                  */
 230                 if (session_dsm_handle == DSM_HANDLE_INVALID)
 231                         pcxt->nworkers = 0;
 232         }
 233
 234         if (pcxt->nworkers > 0)
 235         {
 236                 /* Estimate space for various kinds of state sharing. */
 237                 library_len = EstimateLibraryStateSpace();
 238                 shm_toc_estimate_chunk(&pcxt->estimator, library_len);
 239                 guc_len = EstimateGUCStateSpace();
 240                 shm_toc_estimate_chunk(&pcxt->estimator, guc_len);
 241                 combocidlen = EstimateComboCIDStateSpace();
 242                 shm_toc_estimate_chunk(&pcxt->estimator, combocidlen);
 243                 tsnaplen = EstimateSnapshotSpace(transaction_snapshot);
 244                 shm_toc_estimate_chunk(&pcxt->estimator, tsnaplen);
 245                 asnaplen = EstimateSnapshotSpace(active_snapshot);
 246                 shm_toc_estimate_chunk(&pcxt->estimator, asnaplen);
 247                 tstatelen = EstimateTransactionStateSpace();
 248                 shm_toc_estimate_chunk(&pcxt->estimator, tstatelen);
 249                 shm_toc_estimate_chunk(&pcxt->estimator, sizeof(dsm_handle));
 250                 /* If you add more chunks here, you probably need to add keys. */
 251                 shm_toc_estimate_keys(&pcxt->estimator, 7);
 252
 253                 /* Estimate space need for error queues. */
 254                 StaticAssertStmt(BUFFERALIGN(PARALLEL_ERROR_QUEUE_SIZE) ==
 255                                                  PARALLEL_ERROR_QUEUE_SIZE,
 256                                                  "parallel error queue size not buffer-aligned");
 257                 shm_toc_estimate_chunk(&pcxt->estimator,
 258                                                            mul_size(PARALLEL_ERROR_QUEUE_SIZE,
 259                                                                                 pcxt->nworkers));
 260                 shm_toc_estimate_keys(&pcxt->estimator, 1);
 261
 262                 /* Estimate how much we'll need for the entrypoint info. */
 263                 shm_toc_estimate_chunk(&pcxt->estimator, strlen(pcxt->library_name) +
 264                                                            strlen(pcxt->function_name) + 2);
 265                 shm_toc_estimate_keys(&pcxt->estimator, 1);
 266         }
 267
 268         /*
 269          * Create DSM and initialize with new table of contents.  But if the user
 270          * didn't request any workers, then don't bother creating a dynamic shared
 271          * memory segment; instead, just use backend-private memory.
 272          *
 273          * Also, if we can't create a dynamic shared memory segment because the
 274          * maximum number of segments have already been created, then fall back to
 275          * backend-private memory, and plan not to use any workers.  We hope this
 276          * won't happen very often, but it's better to abandon the use of
 277          * parallelism than to fail outright.
 278          */
 279         segsize = shm_toc_estimate(&pcxt->estimator);
 280         if (pcxt->nworkers > 0)
 281                 pcxt->seg = dsm_create(segsize, DSM_CREATE_NULL_IF_MAXSEGMENTS);
 282         if (pcxt->seg != NULL)
 283                 pcxt->toc = shm_toc_create(PARALLEL_MAGIC,
 284                                                                    dsm_segment_address(pcxt->seg),
 285                                                                    segsize);
 286         else
 287         {
 288                 pcxt->nworkers = 0;
 289                 pcxt->private_memory = MemoryContextAlloc(TopMemoryContext, segsize);
 290                 pcxt->toc = shm_toc_create(PARALLEL_MAGIC, pcxt->private_memory,
 291                                                                    segsize);
 292         }
 293
 294         /* Initialize fixed-size state in shared memory. */
 295         fps = (FixedParallelState *)
 296                 shm_toc_allocate(pcxt->toc, sizeof(FixedParallelState));
 297         fps->database_id = MyDatabaseId;
 298         fps->authenticated_user_id = GetAuthenticatedUserId();
 299         GetUserIdAndSecContext(&fps->current_user_id, &fps->sec_context);
 300         GetTempNamespaceState(&fps->temp_namespace_id,
 301                                                   &fps->temp_toast_namespace_id);
 302         fps->parallel_master_pgproc = MyProc;
 303         fps->parallel_master_pid = MyProcPid;
 304         fps->parallel_master_backend_id = MyBackendId;
 305         SpinLockInit(&fps->mutex);
 306         fps->last_xlog_end = 0;
 307         shm_toc_insert(pcxt->toc, PARALLEL_KEY_FIXED, fps);
 308
 309         /* We can skip the rest of this if we're not budgeting for any workers. */
 310         if (pcxt->nworkers > 0)
 311         {
 312                 char       *libraryspace;
 313                 char       *gucspace;
 314                 char       *combocidspace;
 315                 char       *tsnapspace;
 316                 char       *asnapspace;
 317                 char       *tstatespace;
 318                 char       *error_queue_space;
 319                 char       *session_dsm_handle_space;
 320                 char       *entrypointstate;
 321                 Size            lnamelen;
 322
 323                 /* Serialize shared libraries we have loaded. */
 324                 libraryspace = shm_toc_allocate(pcxt->toc, library_len);
 325                 SerializeLibraryState(library_len, libraryspace);
 326                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_LIBRARY, libraryspace);
 327
 328                 /* Serialize GUC settings. */
 329                 gucspace = shm_toc_allocate(pcxt->toc, guc_len);
 330                 SerializeGUCState(guc_len, gucspace);
 331                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_GUC, gucspace);
 332
 333                 /* Serialize combo CID state. */
 334                 combocidspace = shm_toc_allocate(pcxt->toc, combocidlen);
 335                 SerializeComboCIDState(combocidlen, combocidspace);
 336                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_COMBO_CID, combocidspace);
 337
 338                 /* Serialize transaction snapshot and active snapshot. */
 339                 tsnapspace = shm_toc_allocate(pcxt->toc, tsnaplen);
 340                 SerializeSnapshot(transaction_snapshot, tsnapspace);
 341                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT,
 342                                            tsnapspace);
 343                 asnapspace = shm_toc_allocate(pcxt->toc, asnaplen);
 344                 SerializeSnapshot(active_snapshot, asnapspace);
 345                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ACTIVE_SNAPSHOT, asnapspace);
 346
 347                 /* Provide the handle for per-session segment. */
 348                 session_dsm_handle_space = shm_toc_allocate(pcxt->toc,
 349                                                                                                         sizeof(dsm_handle));
 350                 *(dsm_handle *) session_dsm_handle_space = session_dsm_handle;
 351                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_SESSION_DSM,
 352                                            session_dsm_handle_space);
 353
 354                 /* Serialize transaction state. */
 355                 tstatespace = shm_toc_allocate(pcxt->toc, tstatelen);
 356                 SerializeTransactionState(tstatelen, tstatespace);
 357                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_STATE, tstatespace);
 358
 359                 /* Allocate space for worker information. */
 360                 pcxt->worker = palloc0(sizeof(ParallelWorkerInfo) * pcxt->nworkers);
 361
 362                 /*
 363                  * Establish error queues in dynamic shared memory.
 364                  *
 365                  * These queues should be used only for transmitting ErrorResponse,
 366                  * NoticeResponse, and NotifyResponse protocol messages.  Tuple data
 367                  * should be transmitted via separate (possibly larger?) queues.
 368                  */
 369                 error_queue_space =
 370                         shm_toc_allocate(pcxt->toc,
 371                                                          mul_size(PARALLEL_ERROR_QUEUE_SIZE,
 372                                                                           pcxt->nworkers));
 373                 for (i = 0; i < pcxt->nworkers; ++i)
 374                 {
 375                         char       *start;
 376                         shm_mq     *mq;
 377
 378                         start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
 379                         mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
 380                         shm_mq_set_receiver(mq, MyProc);
 381                         pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
 382                 }
 383                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE, error_queue_space);
 384
 385                 /*
 386                  * Serialize entrypoint information.  It's unsafe to pass function
 387                  * pointers across processes, as the function pointer may be different
 388                  * in each process in EXEC_BACKEND builds, so we always pass library
 389                  * and function name.  (We use library name "postgres" for functions
 390                  * in the core backend.)
 391                  */
 392                 lnamelen = strlen(pcxt->library_name);
 393                 entrypointstate = shm_toc_allocate(pcxt->toc, lnamelen +
 394                                                                                    strlen(pcxt->function_name) + 2);
 395                 strcpy(entrypointstate, pcxt->library_name);
 396                 strcpy(entrypointstate + lnamelen + 1, pcxt->function_name);
 397                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ENTRYPOINT, entrypointstate);
 398         }
 399
 400         /* Restore previous memory context. */
 401         MemoryContextSwitchTo(oldcontext);
 402 }
 403
 404 /*
 405  * Reinitialize the dynamic shared memory segment for a parallel context such
 406  * that we could launch workers for it again.
 407  */
 408 void
 409 ReinitializeParallelDSM(ParallelContext *pcxt)
 410 {
 411         FixedParallelState *fps;
 412         char       *error_queue_space;
 413         int                     i;
 414
 415         /* Wait for any old workers to exit. */
 416         if (pcxt->nworkers_launched > 0)
 417         {
 418                 WaitForParallelWorkersToFinish(pcxt);
 419                 WaitForParallelWorkersToExit(pcxt);
 420                 pcxt->nworkers_launched = 0;
 421         }
 422
 423         /* Reset a few bits of fixed parallel state to a clean state. */
 424         fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED, false);
 425         fps->last_xlog_end = 0;
 426
 427         /* Recreate error queues. */
 428         error_queue_space =
 429                 shm_toc_lookup(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE, false);
 430         for (i = 0; i < pcxt->nworkers; ++i)
 431         {
 432                 char       *start;
 433                 shm_mq     *mq;
 434
 435                 start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
 436                 mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
 437                 shm_mq_set_receiver(mq, MyProc);
 438                 pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
 439         }
 440 }
 441
 442 /*
 443  * Launch parallel workers.
 444  */
 445 void
 446 LaunchParallelWorkers(ParallelContext *pcxt)
 447 {
 448         MemoryContext oldcontext;
 449         BackgroundWorker worker;
 450         int                     i;
 451         bool            any_registrations_failed = false;
 452
 453         /* Skip this if we have no workers. */
 454         if (pcxt->nworkers == 0)
 455                 return;
 456
 457         /* We need to be a lock group leader. */
 458         BecomeLockGroupLeader();
 459
 460         /* If we do have workers, we'd better have a DSM segment. */
 461         Assert(pcxt->seg != NULL);
 462
 463         /* We might be running in a short-lived memory context. */
 464         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 465
 466         /* Configure a worker. */
 467         memset(&worker, 0, sizeof(worker));
 468         snprintf(worker.bgw_name, BGW_MAXLEN, "parallel worker for PID %d",
 469                          MyProcPid);
 470         worker.bgw_flags =
 471                 BGWORKER_SHMEM_ACCESS | BGWORKER_BACKEND_DATABASE_CONNECTION
 472                 | BGWORKER_CLASS_PARALLEL;
 473         worker.bgw_start_time = BgWorkerStart_ConsistentState;
 474         worker.bgw_restart_time = BGW_NEVER_RESTART;
 475         sprintf(worker.bgw_library_name, "postgres");
 476         sprintf(worker.bgw_function_name, "ParallelWorkerMain");
 477         worker.bgw_main_arg = UInt32GetDatum(dsm_segment_handle(pcxt->seg));
 478         worker.bgw_notify_pid = MyProcPid;
 479
 480         /*
 481          * Start workers.
 482          *
 483          * The caller must be able to tolerate ending up with fewer workers than
 484          * expected, so there is no need to throw an error here if registration
 485          * fails.  It wouldn't help much anyway, because registering the worker in
 486          * no way guarantees that it will start up and initialize successfully.
 487          */
 488         for (i = 0; i < pcxt->nworkers; ++i)
 489         {
 490                 memcpy(worker.bgw_extra, &i, sizeof(int));
 491                 if (!any_registrations_failed &&
 492                         RegisterDynamicBackgroundWorker(&worker,
 493                                                                                         &pcxt->worker[i].bgwhandle))
 494                 {
 495                         shm_mq_set_handle(pcxt->worker[i].error_mqh,
 496                                                           pcxt->worker[i].bgwhandle);
 497                         pcxt->nworkers_launched++;
 498                 }
 499                 else
 500                 {
 501                         /*
 502                          * If we weren't able to register the worker, then we've bumped up
 503                          * against the max_worker_processes limit, and future
 504                          * registrations will probably fail too, so arrange to skip them.
 505                          * But we still have to execute this code for the remaining slots
 506                          * to make sure that we forget about the error queues we budgeted
 507                          * for those workers.  Otherwise, we'll wait for them to start,
 508                          * but they never will.
 509                          */
 510                         any_registrations_failed = true;
 511                         pcxt->worker[i].bgwhandle = NULL;
 512                         shm_mq_detach(pcxt->worker[i].error_mqh);
 513                         pcxt->worker[i].error_mqh = NULL;
 514                 }
 515         }
 516
 517         /* Restore previous memory context. */
 518         MemoryContextSwitchTo(oldcontext);
 519 }
 520
 521 /*
 522  * Wait for all workers to finish computing.
 523  *
 524  * Even if the parallel operation seems to have completed successfully, it's
 525  * important to call this function afterwards.  We must not miss any errors
 526  * the workers may have thrown during the parallel operation, or any that they
 527  * may yet throw while shutting down.
 528  *
 529  * Also, we want to update our notion of XactLastRecEnd based on worker
 530  * feedback.
 531  */
 532 void
 533 WaitForParallelWorkersToFinish(ParallelContext *pcxt)
 534 {
 535         for (;;)
 536         {
 537                 bool            anyone_alive = false;
 538                 int                     i;
 539
 540                 /*
 541                  * This will process any parallel messages that are pending, which may
 542                  * change the outcome of the loop that follows.  It may also throw an
 543                  * error propagated from a worker.
 544                  */
 545                 CHECK_FOR_INTERRUPTS();
 546
 547                 for (i = 0; i < pcxt->nworkers_launched; ++i)
 548                 {
 549                         if (pcxt->worker[i].error_mqh != NULL)
 550                         {
 551                                 anyone_alive = true;
 552                                 break;
 553                         }
 554                 }
 555
 556                 if (!anyone_alive)
 557                         break;
 558
 559                 WaitLatch(MyLatch, WL_LATCH_SET, -1,
 560                                   WAIT_EVENT_PARALLEL_FINISH);
 561                 ResetLatch(MyLatch);
 562         }
 563
 564         if (pcxt->toc != NULL)
 565         {
 566                 FixedParallelState *fps;
 567
 568                 fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED, false);
 569                 if (fps->last_xlog_end > XactLastRecEnd)
 570                         XactLastRecEnd = fps->last_xlog_end;
 571         }
 572 }
 573
 574 /*
 575  * Wait for all workers to exit.
 576  *
 577  * This function ensures that workers have been completely shutdown.  The
 578  * difference between WaitForParallelWorkersToFinish and this function is
 579  * that former just ensures that last message sent by worker backend is
 580  * received by master backend whereas this ensures the complete shutdown.
 581  */
 582 static void
 583 WaitForParallelWorkersToExit(ParallelContext *pcxt)
 584 {
 585         int                     i;
 586
 587         /* Wait until the workers actually die. */
 588         for (i = 0; i < pcxt->nworkers_launched; ++i)
 589         {
 590                 BgwHandleStatus status;
 591
 592                 if (pcxt->worker == NULL || pcxt->worker[i].bgwhandle == NULL)
 593                         continue;
 594
 595                 status = WaitForBackgroundWorkerShutdown(pcxt->worker[i].bgwhandle);
 596
 597                 /*
 598                  * If the postmaster kicked the bucket, we have no chance of cleaning
 599                  * up safely -- we won't be able to tell when our workers are actually
 600                  * dead.  This doesn't necessitate a PANIC since they will all abort
 601                  * eventually, but we can't safely continue this session.
 602                  */
 603                 if (status == BGWH_POSTMASTER_DIED)
 604                         ereport(FATAL,
 605                                         (errcode(ERRCODE_ADMIN_SHUTDOWN),
 606                                          errmsg("postmaster exited during a parallel transaction")));
 607
 608                 /* Release memory. */
 609                 pfree(pcxt->worker[i].bgwhandle);
 610                 pcxt->worker[i].bgwhandle = NULL;
 611         }
 612 }
 613
 614 /*
 615  * Destroy a parallel context.
 616  *
 617  * If expecting a clean exit, you should use WaitForParallelWorkersToFinish()
 618  * first, before calling this function.  When this function is invoked, any
 619  * remaining workers are forcibly killed; the dynamic shared memory segment
 620  * is unmapped; and we then wait (uninterruptibly) for the workers to exit.
 621  */
 622 void
 623 DestroyParallelContext(ParallelContext *pcxt)
 624 {
 625         int                     i;
 626
 627         /*
 628          * Be careful about order of operations here!  We remove the parallel
 629          * context from the list before we do anything else; otherwise, if an
 630          * error occurs during a subsequent step, we might try to nuke it again
 631          * from AtEOXact_Parallel or AtEOSubXact_Parallel.
 632          */
 633         dlist_delete(&pcxt->node);
 634
 635         /* Kill each worker in turn, and forget their error queues. */
 636         if (pcxt->worker != NULL)
 637         {
 638                 for (i = 0; i < pcxt->nworkers_launched; ++i)
 639                 {
 640                         if (pcxt->worker[i].error_mqh != NULL)
 641                         {
 642                                 TerminateBackgroundWorker(pcxt->worker[i].bgwhandle);
 643
 644                                 shm_mq_detach(pcxt->worker[i].error_mqh);
 645                                 pcxt->worker[i].error_mqh = NULL;
 646                         }
 647                 }
 648         }
 649
 650         /*
 651          * If we have allocated a shared memory segment, detach it.  This will
 652          * implicitly detach the error queues, and any other shared memory queues,
 653          * stored there.
 654          */
 655         if (pcxt->seg != NULL)
 656         {
 657                 dsm_detach(pcxt->seg);
 658                 pcxt->seg = NULL;
 659         }
 660
 661         /*
 662          * If this parallel context is actually in backend-private memory rather
 663          * than shared memory, free that memory instead.
 664          */
 665         if (pcxt->private_memory != NULL)
 666         {
 667                 pfree(pcxt->private_memory);
 668                 pcxt->private_memory = NULL;
 669         }
 670
 671         /*
 672          * We can't finish transaction commit or abort until all of the workers
 673          * have exited.  This means, in particular, that we can't respond to
 674          * interrupts at this stage.
 675          */
 676         HOLD_INTERRUPTS();
 677         WaitForParallelWorkersToExit(pcxt);
 678         RESUME_INTERRUPTS();
 679
 680         /* Free the worker array itself. */
 681         if (pcxt->worker != NULL)
 682         {
 683                 pfree(pcxt->worker);
 684                 pcxt->worker = NULL;
 685         }
 686
 687         /* Free memory. */
 688         pfree(pcxt->library_name);
 689         pfree(pcxt->function_name);
 690         pfree(pcxt);
 691 }
 692
 693 /*
 694  * Are there any parallel contexts currently active?
 695  */
 696 bool
 697 ParallelContextActive(void)
 698 {
 699         return !dlist_is_empty(&pcxt_list);
 700 }
 701
 702 /*
 703  * Handle receipt of an interrupt indicating a parallel worker message.
 704  *
 705  * Note: this is called within a signal handler!  All we can do is set
 706  * a flag that will cause the next CHECK_FOR_INTERRUPTS() to invoke
 707  * HandleParallelMessages().
 708  */
 709 void
 710 HandleParallelMessageInterrupt(void)
 711 {
 712         InterruptPending = true;
 713         ParallelMessagePending = true;
 714         SetLatch(MyLatch);
 715 }
 716
 717 /*
 718  * Handle any queued protocol messages received from parallel workers.
 719  */
 720 void
 721 HandleParallelMessages(void)
 722 {
 723         dlist_iter      iter;
 724         MemoryContext oldcontext;
 725
 726         static MemoryContext hpm_context = NULL;
 727
 728         /*
 729          * This is invoked from ProcessInterrupts(), and since some of the
 730          * functions it calls contain CHECK_FOR_INTERRUPTS(), there is a potential
 731          * for recursive calls if more signals are received while this runs.  It's
 732          * unclear that recursive entry would be safe, and it doesn't seem useful
 733          * even if it is safe, so let's block interrupts until done.
 734          */
 735         HOLD_INTERRUPTS();
 736
 737         /*
 738          * Moreover, CurrentMemoryContext might be pointing almost anywhere.  We
 739          * don't want to risk leaking data into long-lived contexts, so let's do
 740          * our work here in a private context that we can reset on each use.
 741          */
 742         if (hpm_context == NULL)        /* first time through? */
 743                 hpm_context = AllocSetContextCreate(TopMemoryContext,
 744                                                                                         "HandleParallelMessages",
 745                                                                                         ALLOCSET_DEFAULT_SIZES);
 746         else
 747                 MemoryContextReset(hpm_context);
 748
 749         oldcontext = MemoryContextSwitchTo(hpm_context);
 750
 751         /* OK to process messages.  Reset the flag saying there are more to do. */
 752         ParallelMessagePending = false;
 753
 754         dlist_foreach(iter, &pcxt_list)
 755         {
 756                 ParallelContext *pcxt;
 757                 int                     i;
 758
 759                 pcxt = dlist_container(ParallelContext, node, iter.cur);
 760                 if (pcxt->worker == NULL)
 761                         continue;
 762
 763                 for (i = 0; i < pcxt->nworkers_launched; ++i)
 764                 {
 765                         /*
 766                          * Read as many messages as we can from each worker, but stop when
 767                          * either (1) the worker's error queue goes away, which can happen
 768                          * if we receive a Terminate message from the worker; or (2) no
 769                          * more messages can be read from the worker without blocking.
 770                          */
 771                         while (pcxt->worker[i].error_mqh != NULL)
 772                         {
 773                                 shm_mq_result res;
 774                                 Size            nbytes;
 775                                 void       *data;
 776
 777                                 res = shm_mq_receive(pcxt->worker[i].error_mqh, &nbytes,
 778                                                                          &data, true);
 779                                 if (res == SHM_MQ_WOULD_BLOCK)
 780                                         break;
 781                                 else if (res == SHM_MQ_SUCCESS)
 782                                 {
 783                                         StringInfoData msg;
 784
 785                                         initStringInfo(&msg);
 786                                         appendBinaryStringInfo(&msg, data, nbytes);
 787                                         HandleParallelMessage(pcxt, i, &msg);
 788                                         pfree(msg.data);
 789                                 }
 790                                 else
 791                                         ereport(ERROR,
 792                                                         (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 793                                                          errmsg("lost connection to parallel worker")));
 794                         }
 795                 }
 796         }
 797
 798         MemoryContextSwitchTo(oldcontext);
 799
 800         /* Might as well clear the context on our way out */
 801         MemoryContextReset(hpm_context);
 802
 803         RESUME_INTERRUPTS();
 804 }
 805
 806 /*
 807  * Handle a single protocol message received from a single parallel worker.
 808  */
 809 static void
 810 HandleParallelMessage(ParallelContext *pcxt, int i, StringInfo msg)
 811 {
 812         char            msgtype;
 813
 814         msgtype = pq_getmsgbyte(msg);
 815
 816         switch (msgtype)
 817         {
 818                 case 'K':                               /* BackendKeyData */
 819                         {
 820                                 int32           pid = pq_getmsgint(msg, 4);
 821
 822                                 (void) pq_getmsgint(msg, 4);    /* discard cancel key */
 823                                 (void) pq_getmsgend(msg);
 824                                 pcxt->worker[i].pid = pid;
 825                                 break;
 826                         }
 827
 828                 case 'E':                               /* ErrorResponse */
 829                 case 'N':                               /* NoticeResponse */
 830                         {
 831                                 ErrorData       edata;
 832                                 ErrorContextCallback *save_error_context_stack;
 833
 834                                 /* Parse ErrorResponse or NoticeResponse. */
 835                                 pq_parse_errornotice(msg, &edata);
 836
 837                                 /* Death of a worker isn't enough justification for suicide. */
 838                                 edata.elevel = Min(edata.elevel, ERROR);
 839
 840                                 /*
 841                                  * If desired, add a context line to show that this is a
 842                                  * message propagated from a parallel worker.  Otherwise, it
 843                                  * can sometimes be confusing to understand what actually
 844                                  * happened.  (We don't do this in FORCE_PARALLEL_REGRESS mode
 845                                  * because it causes test-result instability depending on
 846                                  * whether a parallel worker is actually used or not.)
 847                                  */
 848                                 if (force_parallel_mode != FORCE_PARALLEL_REGRESS)
 849                                 {
 850                                         if (edata.context)
 851                                                 edata.context = psprintf("%s\n%s", edata.context,
 852                                                                                                  _("parallel worker"));
 853                                         else
 854                                                 edata.context = pstrdup(_("parallel worker"));
 855                                 }
 856
 857                                 /*
 858                                  * Context beyond that should use the error context callbacks
 859                                  * that were in effect when the ParallelContext was created,
 860                                  * not the current ones.
 861                                  */
 862                                 save_error_context_stack = error_context_stack;
 863                                 error_context_stack = pcxt->error_context_stack;
 864
 865                                 /* Rethrow error or print notice. */
 866                                 ThrowErrorData(&edata);
 867
 868                                 /* Not an error, so restore previous context stack. */
 869                                 error_context_stack = save_error_context_stack;
 870
 871                                 break;
 872                         }
 873
 874                 case 'A':                               /* NotifyResponse */
 875                         {
 876                                 /* Propagate NotifyResponse. */
 877                                 int32           pid;
 878                                 const char *channel;
 879                                 const char *payload;
 880
 881                                 pid = pq_getmsgint(msg, 4);
 882                                 channel = pq_getmsgrawstring(msg);
 883                                 payload = pq_getmsgrawstring(msg);
 884                                 pq_endmessage(msg);
 885
 886                                 NotifyMyFrontEnd(channel, payload, pid);
 887
 888                                 break;
 889                         }
 890
 891                 case 'X':                               /* Terminate, indicating clean exit */
 892                         {
 893                                 shm_mq_detach(pcxt->worker[i].error_mqh);
 894                                 pcxt->worker[i].error_mqh = NULL;
 895                                 break;
 896                         }
 897
 898                 default:
 899                         {
 900                                 elog(ERROR, "unrecognized message type received from parallel worker: %c (message length %d bytes)",
 901                                          msgtype, msg->len);
 902                         }
 903         }
 904 }
 905
 906 /*
 907  * End-of-subtransaction cleanup for parallel contexts.
 908  *
 909  * Currently, it's forbidden to enter or leave a subtransaction while
 910  * parallel mode is in effect, so we could just blow away everything.  But
 911  * we may want to relax that restriction in the future, so this code
 912  * contemplates that there may be multiple subtransaction IDs in pcxt_list.
 913  */
 914 void
 915 AtEOSubXact_Parallel(bool isCommit, SubTransactionId mySubId)
 916 {
 917         while (!dlist_is_empty(&pcxt_list))
 918         {
 919                 ParallelContext *pcxt;
 920
 921                 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
 922                 if (pcxt->subid != mySubId)
 923                         break;
 924                 if (isCommit)
 925                         elog(WARNING, "leaked parallel context");
 926                 DestroyParallelContext(pcxt);
 927         }
 928 }
 929
 930 /*
 931  * End-of-transaction cleanup for parallel contexts.
 932  */
 933 void
 934 AtEOXact_Parallel(bool isCommit)
 935 {
 936         while (!dlist_is_empty(&pcxt_list))
 937         {
 938                 ParallelContext *pcxt;
 939
 940                 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
 941                 if (isCommit)
 942                         elog(WARNING, "leaked parallel context");
 943                 DestroyParallelContext(pcxt);
 944         }
 945 }
 946
 947 /*
 948  * Main entrypoint for parallel workers.
 949  */
 950 void
 951 ParallelWorkerMain(Datum main_arg)
 952 {
 953         dsm_segment *seg;
 954         shm_toc    *toc;
 955         FixedParallelState *fps;
 956         char       *error_queue_space;
 957         shm_mq     *mq;
 958         shm_mq_handle *mqh;
 959         char       *libraryspace;
 960         char       *entrypointstate;
 961         char       *library_name;
 962         char       *function_name;
 963         parallel_worker_main_type entrypt;
 964         char       *gucspace;
 965         char       *combocidspace;
 966         char       *tsnapspace;
 967         char       *asnapspace;
 968         char       *tstatespace;
 969         StringInfoData msgbuf;
 970         char       *session_dsm_handle_space;
 971
 972         /* Set flag to indicate that we're initializing a parallel worker. */
 973         InitializingParallelWorker = true;
 974
 975         /* Establish signal handlers. */
 976         pqsignal(SIGTERM, die);
 977         BackgroundWorkerUnblockSignals();
 978
 979         /* Determine and set our parallel worker number. */
 980         Assert(ParallelWorkerNumber == -1);
 981         memcpy(&ParallelWorkerNumber, MyBgworkerEntry->bgw_extra, sizeof(int));
 982
 983         /* Set up a memory context and resource owner. */
 984         Assert(CurrentResourceOwner == NULL);
 985         CurrentResourceOwner = ResourceOwnerCreate(NULL, "parallel toplevel");
 986         CurrentMemoryContext = AllocSetContextCreate(TopMemoryContext,
 987                                                                                                  "Parallel worker",
 988                                                                                                  ALLOCSET_DEFAULT_SIZES);
 989
 990         /*
 991          * Now that we have a resource owner, we can attach to the dynamic shared
 992          * memory segment and read the table of contents.
 993          */
 994         seg = dsm_attach(DatumGetUInt32(main_arg));
 995         if (seg == NULL)
 996                 ereport(ERROR,
 997                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 998                                  errmsg("could not map dynamic shared memory segment")));
 999         toc = shm_toc_attach(PARALLEL_MAGIC, dsm_segment_address(seg));
1000         if (toc == NULL)
1001                 ereport(ERROR,
1002                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1003                                  errmsg("invalid magic number in dynamic shared memory segment")));
1004
1005         /* Look up fixed parallel state. */
1006         fps = shm_toc_lookup(toc, PARALLEL_KEY_FIXED, false);
1007         MyFixedParallelState = fps;
1008
1009         /*
1010          * Now that we have a worker number, we can find and attach to the error
1011          * queue provided for us.  That's good, because until we do that, any
1012          * errors that happen here will not be reported back to the process that
1013          * requested that this worker be launched.
1014          */
1015         error_queue_space = shm_toc_lookup(toc, PARALLEL_KEY_ERROR_QUEUE, false);
1016         mq = (shm_mq *) (error_queue_space +
1017                                          ParallelWorkerNumber * PARALLEL_ERROR_QUEUE_SIZE);
1018         shm_mq_set_sender(mq, MyProc);
1019         mqh = shm_mq_attach(mq, seg, NULL);
1020         pq_redirect_to_shm_mq(seg, mqh);
1021         pq_set_parallel_master(fps->parallel_master_pid,
1022                                                    fps->parallel_master_backend_id);
1023
1024         /*
1025          * Send a BackendKeyData message to the process that initiated parallelism
1026          * so that it has access to our PID before it receives any other messages
1027          * from us.  Our cancel key is sent, too, since that's the way the
1028          * protocol message is defined, but it won't actually be used for anything
1029          * in this case.
1030          */
1031         pq_beginmessage(&msgbuf, 'K');
1032         pq_sendint(&msgbuf, (int32) MyProcPid, sizeof(int32));
1033         pq_sendint(&msgbuf, (int32) MyCancelKey, sizeof(int32));
1034         pq_endmessage(&msgbuf);
1035
1036         /*
1037          * Hooray! Primary initialization is complete.  Now, we need to set up our
1038          * backend-local state to match the original backend.
1039          */
1040
1041         /*
1042          * Join locking group.  We must do this before anything that could try to
1043          * acquire a heavyweight lock, because any heavyweight locks acquired to
1044          * this point could block either directly against the parallel group
1045          * leader or against some process which in turn waits for a lock that
1046          * conflicts with the parallel group leader, causing an undetected
1047          * deadlock.  (If we can't join the lock group, the leader has gone away,
1048          * so just exit quietly.)
1049          */
1050         if (!BecomeLockGroupMember(fps->parallel_master_pgproc,
1051                                                            fps->parallel_master_pid))
1052                 return;
1053
1054         /*
1055          * Load libraries that were loaded by original backend.  We want to do
1056          * this before restoring GUCs, because the libraries might define custom
1057          * variables.
1058          */
1059         libraryspace = shm_toc_lookup(toc, PARALLEL_KEY_LIBRARY, false);
1060         RestoreLibraryState(libraryspace);
1061
1062         /*
1063          * Identify the entry point to be called.  In theory this could result in
1064          * loading an additional library, though most likely the entry point is in
1065          * the core backend or in a library we just loaded.
1066          */
1067         entrypointstate = shm_toc_lookup(toc, PARALLEL_KEY_ENTRYPOINT, false);
1068         library_name = entrypointstate;
1069         function_name = entrypointstate + strlen(library_name) + 1;
1070
1071         entrypt = LookupParallelWorkerFunction(library_name, function_name);
1072
1073         /* Restore database connection. */
1074         BackgroundWorkerInitializeConnectionByOid(fps->database_id,
1075                                                                                           fps->authenticated_user_id);
1076
1077         /*
1078          * Set the client encoding to the database encoding, since that is what
1079          * the leader will expect.
1080          */
1081         SetClientEncoding(GetDatabaseEncoding());
1082
1083         /* Restore GUC values from launching backend. */
1084         gucspace = shm_toc_lookup(toc, PARALLEL_KEY_GUC, false);
1085         StartTransactionCommand();
1086         RestoreGUCState(gucspace);
1087         CommitTransactionCommand();
1088
1089         /* Crank up a transaction state appropriate to a parallel worker. */
1090         tstatespace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_STATE, false);
1091         StartParallelWorkerTransaction(tstatespace);
1092
1093         /* Restore combo CID state. */
1094         combocidspace = shm_toc_lookup(toc, PARALLEL_KEY_COMBO_CID, false);
1095         RestoreComboCIDState(combocidspace);
1096
1097         /* Attach to the per-session DSM segment and contained objects. */
1098         session_dsm_handle_space =
1099                 shm_toc_lookup(toc, PARALLEL_KEY_SESSION_DSM, false);
1100         AttachSession(*(dsm_handle *) session_dsm_handle_space);
1101
1102         /* Restore transaction snapshot. */
1103         tsnapspace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT, false);
1104         RestoreTransactionSnapshot(RestoreSnapshot(tsnapspace),
1105                                                            fps->parallel_master_pgproc);
1106
1107         /* Restore active snapshot. */
1108         asnapspace = shm_toc_lookup(toc, PARALLEL_KEY_ACTIVE_SNAPSHOT, false);
1109         PushActiveSnapshot(RestoreSnapshot(asnapspace));
1110
1111         /*
1112          * We've changed which tuples we can see, and must therefore invalidate
1113          * system caches.
1114          */
1115         InvalidateSystemCaches();
1116
1117         /* Restore user ID and security context. */
1118         SetUserIdAndSecContext(fps->current_user_id, fps->sec_context);
1119
1120         /* Restore temp-namespace state to ensure search path matches leader's. */
1121         SetTempNamespaceState(fps->temp_namespace_id,
1122                                                   fps->temp_toast_namespace_id);
1123
1124         /* Set ParallelMasterBackendId so we know how to address temp relations. */
1125         ParallelMasterBackendId = fps->parallel_master_backend_id;
1126
1127         /*
1128          * We've initialized all of our state now; nothing should change
1129          * hereafter.
1130          */
1131         InitializingParallelWorker = false;
1132         EnterParallelMode();
1133
1134         /*
1135          * Time to do the real work: invoke the caller-supplied code.
1136          */
1137         entrypt(seg, toc);
1138
1139         /* Must exit parallel mode to pop active snapshot. */
1140         ExitParallelMode();
1141
1142         /* Must pop active snapshot so resowner.c doesn't complain. */
1143         PopActiveSnapshot();
1144
1145         /* Shut down the parallel-worker transaction. */
1146         EndParallelWorkerTransaction();
1147
1148         /* Detach from the per-session DSM segment. */
1149         DetachSession();
1150
1151         /* Report success. */
1152         pq_putmessage('X', NULL, 0);
1153 }
1154
1155 /*
1156  * Update shared memory with the ending location of the last WAL record we
1157  * wrote, if it's greater than the value already stored there.
1158  */
1159 void
1160 ParallelWorkerReportLastRecEnd(XLogRecPtr last_xlog_end)
1161 {
1162         FixedParallelState *fps = MyFixedParallelState;
1163
1164         Assert(fps != NULL);
1165         SpinLockAcquire(&fps->mutex);
1166         if (fps->last_xlog_end < last_xlog_end)
1167                 fps->last_xlog_end = last_xlog_end;
1168         SpinLockRelease(&fps->mutex);
1169 }
1170
1171 /*
1172  * Look up (and possibly load) a parallel worker entry point function.
1173  *
1174  * For functions contained in the core code, we use library name "postgres"
1175  * and consult the InternalParallelWorkers array.  External functions are
1176  * looked up, and loaded if necessary, using load_external_function().
1177  *
1178  * The point of this is to pass function names as strings across process
1179  * boundaries.  We can't pass actual function addresses because of the
1180  * possibility that the function has been loaded at a different address
1181  * in a different process.  This is obviously a hazard for functions in
1182  * loadable libraries, but it can happen even for functions in the core code
1183  * on platforms using EXEC_BACKEND (e.g., Windows).
1184  *
1185  * At some point it might be worthwhile to get rid of InternalParallelWorkers[]
1186  * in favor of applying load_external_function() for core functions too;
1187  * but that raises portability issues that are not worth addressing now.
1188  */
1189 static parallel_worker_main_type
1190 LookupParallelWorkerFunction(const char *libraryname, const char *funcname)
1191 {
1192         /*
1193          * If the function is to be loaded from postgres itself, search the
1194          * InternalParallelWorkers array.
1195          */
1196         if (strcmp(libraryname, "postgres") == 0)
1197         {
1198                 int                     i;
1199
1200                 for (i = 0; i < lengthof(InternalParallelWorkers); i++)
1201                 {
1202                         if (strcmp(InternalParallelWorkers[i].fn_name, funcname) == 0)
1203                                 return InternalParallelWorkers[i].fn_addr;
1204                 }
1205
1206                 /* We can only reach this by programming error. */
1207                 elog(ERROR, "internal function \"%s\" not found", funcname);
1208         }
1209
1210         /* Otherwise load from external library. */
1211         return (parallel_worker_main_type)
1212                 load_external_function(libraryname, funcname, true, NULL);
1213 }