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) 1996-2018, 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                     outer_user_id;
  79         Oid                     temp_namespace_id;
  80         Oid                     temp_toast_namespace_id;
  81         int                     sec_context;
  82         bool            is_superuser;
  83         PGPROC     *parallel_master_pgproc;
  84         pid_t           parallel_master_pid;
  85         BackendId       parallel_master_backend_id;
  86
  87         /* Mutex protects remaining fields. */
  88         slock_t         mutex;
  89
  90         /* Maximum XactLastRecEnd of any worker. */
  91         XLogRecPtr      last_xlog_end;
  92 } FixedParallelState;
  93
  94 /*
  95  * Our parallel worker number.  We initialize this to -1, meaning that we are
  96  * not a parallel worker.  In parallel workers, it will be set to a value >= 0
  97  * and < the number of workers before any user code is invoked; each parallel
  98  * worker will get a different parallel worker number.
  99  */
 100 int                     ParallelWorkerNumber = -1;
 101
 102 /* Is there a parallel message pending which we need to receive? */
 103 volatile bool ParallelMessagePending = false;
 104
 105 /* Are we initializing a parallel worker? */
 106 bool            InitializingParallelWorker = false;
 107
 108 /* Pointer to our fixed parallel state. */
 109 static FixedParallelState *MyFixedParallelState;
 110
 111 /* List of active parallel contexts. */
 112 static dlist_head pcxt_list = DLIST_STATIC_INIT(pcxt_list);
 113
 114 /*
 115  * List of internal parallel worker entry points.  We need this for
 116  * reasons explained in LookupParallelWorkerFunction(), below.
 117  */
 118 static const struct
 119 {
 120         const char *fn_name;
 121         parallel_worker_main_type fn_addr;
 122 }                       InternalParallelWorkers[] =
 123
 124 {
 125         {
 126                 "ParallelQueryMain", ParallelQueryMain
 127         }
 128 };
 129
 130 /* Private functions. */
 131 static void HandleParallelMessage(ParallelContext *pcxt, int i, StringInfo msg);
 132 static void WaitForParallelWorkersToExit(ParallelContext *pcxt);
 133 static parallel_worker_main_type LookupParallelWorkerFunction(const char *libraryname, const char *funcname);
 134
 135
 136 /*
 137  * Establish a new parallel context.  This should be done after entering
 138  * parallel mode, and (unless there is an error) the context should be
 139  * destroyed before exiting the current subtransaction.
 140  */
 141 ParallelContext *
 142 CreateParallelContext(const char *library_name, const char *function_name,
 143                                           int nworkers)
 144 {
 145         MemoryContext oldcontext;
 146         ParallelContext *pcxt;
 147
 148         /* It is unsafe to create a parallel context if not in parallel mode. */
 149         Assert(IsInParallelMode());
 150
 151         /* Number of workers should be non-negative. */
 152         Assert(nworkers >= 0);
 153
 154         /*
 155          * If dynamic shared memory is not available, we won't be able to use
 156          * background workers.
 157          */
 158         if (dynamic_shared_memory_type == DSM_IMPL_NONE)
 159                 nworkers = 0;
 160
 161         /*
 162          * If we are running under serializable isolation, we can't use parallel
 163          * workers, at least not until somebody enhances that mechanism to be
 164          * parallel-aware.
 165          */
 166         if (IsolationIsSerializable())
 167                 nworkers = 0;
 168
 169         /* We might be running in a short-lived memory context. */
 170         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 171
 172         /* Initialize a new ParallelContext. */
 173         pcxt = palloc0(sizeof(ParallelContext));
 174         pcxt->subid = GetCurrentSubTransactionId();
 175         pcxt->nworkers = nworkers;
 176         pcxt->library_name = pstrdup(library_name);
 177         pcxt->function_name = pstrdup(function_name);
 178         pcxt->error_context_stack = error_context_stack;
 179         shm_toc_initialize_estimator(&pcxt->estimator);
 180         dlist_push_head(&pcxt_list, &pcxt->node);
 181
 182         /* Restore previous memory context. */
 183         MemoryContextSwitchTo(oldcontext);
 184
 185         return pcxt;
 186 }
 187
 188 /*
 189  * Establish the dynamic shared memory segment for a parallel context and
 190  * copy state and other bookkeeping information that will be needed by
 191  * parallel workers into it.
 192  */
 193 void
 194 InitializeParallelDSM(ParallelContext *pcxt)
 195 {
 196         MemoryContext oldcontext;
 197         Size            library_len = 0;
 198         Size            guc_len = 0;
 199         Size            combocidlen = 0;
 200         Size            tsnaplen = 0;
 201         Size            asnaplen = 0;
 202         Size            tstatelen = 0;
 203         Size            segsize = 0;
 204         int                     i;
 205         FixedParallelState *fps;
 206         dsm_handle      session_dsm_handle = DSM_HANDLE_INVALID;
 207         Snapshot        transaction_snapshot = GetTransactionSnapshot();
 208         Snapshot        active_snapshot = GetActiveSnapshot();
 209
 210         /* We might be running in a very short-lived memory context. */
 211         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 212
 213         /* Allow space to store the fixed-size parallel state. */
 214         shm_toc_estimate_chunk(&pcxt->estimator, sizeof(FixedParallelState));
 215         shm_toc_estimate_keys(&pcxt->estimator, 1);
 216
 217         /*
 218          * Normally, the user will have requested at least one worker process, but
 219          * if by chance they have not, we can skip a bunch of things here.
 220          */
 221         if (pcxt->nworkers > 0)
 222         {
 223                 /* Get (or create) the per-session DSM segment's handle. */
 224                 session_dsm_handle = GetSessionDsmHandle();
 225
 226                 /*
 227                  * If we weren't able to create a per-session DSM segment, then we can
 228                  * continue but we can't safely launch any workers because their
 229                  * record typmods would be incompatible so they couldn't exchange
 230                  * tuples.
 231                  */
 232                 if (session_dsm_handle == DSM_HANDLE_INVALID)
 233                         pcxt->nworkers = 0;
 234         }
 235
 236         if (pcxt->nworkers > 0)
 237         {
 238                 /* Estimate space for various kinds of state sharing. */
 239                 library_len = EstimateLibraryStateSpace();
 240                 shm_toc_estimate_chunk(&pcxt->estimator, library_len);
 241                 guc_len = EstimateGUCStateSpace();
 242                 shm_toc_estimate_chunk(&pcxt->estimator, guc_len);
 243                 combocidlen = EstimateComboCIDStateSpace();
 244                 shm_toc_estimate_chunk(&pcxt->estimator, combocidlen);
 245                 tsnaplen = EstimateSnapshotSpace(transaction_snapshot);
 246                 shm_toc_estimate_chunk(&pcxt->estimator, tsnaplen);
 247                 asnaplen = EstimateSnapshotSpace(active_snapshot);
 248                 shm_toc_estimate_chunk(&pcxt->estimator, asnaplen);
 249                 tstatelen = EstimateTransactionStateSpace();
 250                 shm_toc_estimate_chunk(&pcxt->estimator, tstatelen);
 251                 shm_toc_estimate_chunk(&pcxt->estimator, sizeof(dsm_handle));
 252                 /* If you add more chunks here, you probably need to add keys. */
 253                 shm_toc_estimate_keys(&pcxt->estimator, 7);
 254
 255                 /* Estimate space need for error queues. */
 256                 StaticAssertStmt(BUFFERALIGN(PARALLEL_ERROR_QUEUE_SIZE) ==
 257                                                  PARALLEL_ERROR_QUEUE_SIZE,
 258                                                  "parallel error queue size not buffer-aligned");
 259                 shm_toc_estimate_chunk(&pcxt->estimator,
 260                                                            mul_size(PARALLEL_ERROR_QUEUE_SIZE,
 261                                                                                 pcxt->nworkers));
 262                 shm_toc_estimate_keys(&pcxt->estimator, 1);
 263
 264                 /* Estimate how much we'll need for the entrypoint info. */
 265                 shm_toc_estimate_chunk(&pcxt->estimator, strlen(pcxt->library_name) +
 266                                                            strlen(pcxt->function_name) + 2);
 267                 shm_toc_estimate_keys(&pcxt->estimator, 1);
 268         }
 269
 270         /*
 271          * Create DSM and initialize with new table of contents.  But if the user
 272          * didn't request any workers, then don't bother creating a dynamic shared
 273          * memory segment; instead, just use backend-private memory.
 274          *
 275          * Also, if we can't create a dynamic shared memory segment because the
 276          * maximum number of segments have already been created, then fall back to
 277          * backend-private memory, and plan not to use any workers.  We hope this
 278          * won't happen very often, but it's better to abandon the use of
 279          * parallelism than to fail outright.
 280          */
 281         segsize = shm_toc_estimate(&pcxt->estimator);
 282         if (pcxt->nworkers > 0)
 283                 pcxt->seg = dsm_create(segsize, DSM_CREATE_NULL_IF_MAXSEGMENTS);
 284         if (pcxt->seg != NULL)
 285                 pcxt->toc = shm_toc_create(PARALLEL_MAGIC,
 286                                                                    dsm_segment_address(pcxt->seg),
 287                                                                    segsize);
 288         else
 289         {
 290                 pcxt->nworkers = 0;
 291                 pcxt->private_memory = MemoryContextAlloc(TopMemoryContext, segsize);
 292                 pcxt->toc = shm_toc_create(PARALLEL_MAGIC, pcxt->private_memory,
 293                                                                    segsize);
 294         }
 295
 296         /* Initialize fixed-size state in shared memory. */
 297         fps = (FixedParallelState *)
 298                 shm_toc_allocate(pcxt->toc, sizeof(FixedParallelState));
 299         fps->database_id = MyDatabaseId;
 300         fps->authenticated_user_id = GetAuthenticatedUserId();
 301         fps->outer_user_id = GetCurrentRoleId();
 302         fps->is_superuser = session_auth_is_superuser;
 303         GetUserIdAndSecContext(&fps->current_user_id, &fps->sec_context);
 304         GetTempNamespaceState(&fps->temp_namespace_id,
 305                                                   &fps->temp_toast_namespace_id);
 306         fps->parallel_master_pgproc = MyProc;
 307         fps->parallel_master_pid = MyProcPid;
 308         fps->parallel_master_backend_id = MyBackendId;
 309         SpinLockInit(&fps->mutex);
 310         fps->last_xlog_end = 0;
 311         shm_toc_insert(pcxt->toc, PARALLEL_KEY_FIXED, fps);
 312
 313         /* We can skip the rest of this if we're not budgeting for any workers. */
 314         if (pcxt->nworkers > 0)
 315         {
 316                 char       *libraryspace;
 317                 char       *gucspace;
 318                 char       *combocidspace;
 319                 char       *tsnapspace;
 320                 char       *asnapspace;
 321                 char       *tstatespace;
 322                 char       *error_queue_space;
 323                 char       *session_dsm_handle_space;
 324                 char       *entrypointstate;
 325                 Size            lnamelen;
 326
 327                 /* Serialize shared libraries we have loaded. */
 328                 libraryspace = shm_toc_allocate(pcxt->toc, library_len);
 329                 SerializeLibraryState(library_len, libraryspace);
 330                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_LIBRARY, libraryspace);
 331
 332                 /* Serialize GUC settings. */
 333                 gucspace = shm_toc_allocate(pcxt->toc, guc_len);
 334                 SerializeGUCState(guc_len, gucspace);
 335                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_GUC, gucspace);
 336
 337                 /* Serialize combo CID state. */
 338                 combocidspace = shm_toc_allocate(pcxt->toc, combocidlen);
 339                 SerializeComboCIDState(combocidlen, combocidspace);
 340                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_COMBO_CID, combocidspace);
 341
 342                 /* Serialize transaction snapshot and active snapshot. */
 343                 tsnapspace = shm_toc_allocate(pcxt->toc, tsnaplen);
 344                 SerializeSnapshot(transaction_snapshot, tsnapspace);
 345                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT,
 346                                            tsnapspace);
 347                 asnapspace = shm_toc_allocate(pcxt->toc, asnaplen);
 348                 SerializeSnapshot(active_snapshot, asnapspace);
 349                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ACTIVE_SNAPSHOT, asnapspace);
 350
 351                 /* Provide the handle for per-session segment. */
 352                 session_dsm_handle_space = shm_toc_allocate(pcxt->toc,
 353                                                                                                         sizeof(dsm_handle));
 354                 *(dsm_handle *) session_dsm_handle_space = session_dsm_handle;
 355                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_SESSION_DSM,
 356                                            session_dsm_handle_space);
 357
 358                 /* Serialize transaction state. */
 359                 tstatespace = shm_toc_allocate(pcxt->toc, tstatelen);
 360                 SerializeTransactionState(tstatelen, tstatespace);
 361                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_STATE, tstatespace);
 362
 363                 /* Allocate space for worker information. */
 364                 pcxt->worker = palloc0(sizeof(ParallelWorkerInfo) * pcxt->nworkers);
 365
 366                 /*
 367                  * Establish error queues in dynamic shared memory.
 368                  *
 369                  * These queues should be used only for transmitting ErrorResponse,
 370                  * NoticeResponse, and NotifyResponse protocol messages.  Tuple data
 371                  * should be transmitted via separate (possibly larger?) queues.
 372                  */
 373                 error_queue_space =
 374                         shm_toc_allocate(pcxt->toc,
 375                                                          mul_size(PARALLEL_ERROR_QUEUE_SIZE,
 376                                                                           pcxt->nworkers));
 377                 for (i = 0; i < pcxt->nworkers; ++i)
 378                 {
 379                         char       *start;
 380                         shm_mq     *mq;
 381
 382                         start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
 383                         mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
 384                         shm_mq_set_receiver(mq, MyProc);
 385                         pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
 386                 }
 387                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE, error_queue_space);
 388
 389                 /*
 390                  * Serialize entrypoint information.  It's unsafe to pass function
 391                  * pointers across processes, as the function pointer may be different
 392                  * in each process in EXEC_BACKEND builds, so we always pass library
 393                  * and function name.  (We use library name "postgres" for functions
 394                  * in the core backend.)
 395                  */
 396                 lnamelen = strlen(pcxt->library_name);
 397                 entrypointstate = shm_toc_allocate(pcxt->toc, lnamelen +
 398                                                                                    strlen(pcxt->function_name) + 2);
 399                 strcpy(entrypointstate, pcxt->library_name);
 400                 strcpy(entrypointstate + lnamelen + 1, pcxt->function_name);
 401                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ENTRYPOINT, entrypointstate);
 402         }
 403
 404         /* Restore previous memory context. */
 405         MemoryContextSwitchTo(oldcontext);
 406 }
 407
 408 /*
 409  * Reinitialize the dynamic shared memory segment for a parallel context such
 410  * that we could launch workers for it again.
 411  */
 412 void
 413 ReinitializeParallelDSM(ParallelContext *pcxt)
 414 {
 415         FixedParallelState *fps;
 416         char       *error_queue_space;
 417         int                     i;
 418
 419         /* Wait for any old workers to exit. */
 420         if (pcxt->nworkers_launched > 0)
 421         {
 422                 WaitForParallelWorkersToFinish(pcxt);
 423                 WaitForParallelWorkersToExit(pcxt);
 424                 pcxt->nworkers_launched = 0;
 425         }
 426
 427         /* Reset a few bits of fixed parallel state to a clean state. */
 428         fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED, false);
 429         fps->last_xlog_end = 0;
 430
 431         /* Recreate error queues (if they exist). */
 432         error_queue_space =
 433                 shm_toc_lookup(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE, true);
 434         Assert(pcxt->nworkers == 0 || error_queue_space != NULL);
 435         for (i = 0; i < pcxt->nworkers; ++i)
 436         {
 437                 char       *start;
 438                 shm_mq     *mq;
 439
 440                 start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
 441                 mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
 442                 shm_mq_set_receiver(mq, MyProc);
 443                 pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
 444         }
 445 }
 446
 447 /*
 448  * Launch parallel workers.
 449  */
 450 void
 451 LaunchParallelWorkers(ParallelContext *pcxt)
 452 {
 453         MemoryContext oldcontext;
 454         BackgroundWorker worker;
 455         int                     i;
 456         bool            any_registrations_failed = false;
 457
 458         /* Skip this if we have no workers. */
 459         if (pcxt->nworkers == 0)
 460                 return;
 461
 462         /* We need to be a lock group leader. */
 463         BecomeLockGroupLeader();
 464
 465         /* If we do have workers, we'd better have a DSM segment. */
 466         Assert(pcxt->seg != NULL);
 467
 468         /* We might be running in a short-lived memory context. */
 469         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 470
 471         /* Configure a worker. */
 472         memset(&worker, 0, sizeof(worker));
 473         snprintf(worker.bgw_name, BGW_MAXLEN, "parallel worker for PID %d",
 474                          MyProcPid);
 475         snprintf(worker.bgw_type, BGW_MAXLEN, "parallel worker");
 476         worker.bgw_flags =
 477                 BGWORKER_SHMEM_ACCESS | BGWORKER_BACKEND_DATABASE_CONNECTION
 478                 | BGWORKER_CLASS_PARALLEL;
 479         worker.bgw_start_time = BgWorkerStart_ConsistentState;
 480         worker.bgw_restart_time = BGW_NEVER_RESTART;
 481         sprintf(worker.bgw_library_name, "postgres");
 482         sprintf(worker.bgw_function_name, "ParallelWorkerMain");
 483         worker.bgw_main_arg = UInt32GetDatum(dsm_segment_handle(pcxt->seg));
 484         worker.bgw_notify_pid = MyProcPid;
 485
 486         /*
 487          * Start workers.
 488          *
 489          * The caller must be able to tolerate ending up with fewer workers than
 490          * expected, so there is no need to throw an error here if registration
 491          * fails.  It wouldn't help much anyway, because registering the worker in
 492          * no way guarantees that it will start up and initialize successfully.
 493          */
 494         for (i = 0; i < pcxt->nworkers; ++i)
 495         {
 496                 memcpy(worker.bgw_extra, &i, sizeof(int));
 497                 if (!any_registrations_failed &&
 498                         RegisterDynamicBackgroundWorker(&worker,
 499                                                                                         &pcxt->worker[i].bgwhandle))
 500                 {
 501                         shm_mq_set_handle(pcxt->worker[i].error_mqh,
 502                                                           pcxt->worker[i].bgwhandle);
 503                         pcxt->nworkers_launched++;
 504                 }
 505                 else
 506                 {
 507                         /*
 508                          * If we weren't able to register the worker, then we've bumped up
 509                          * against the max_worker_processes limit, and future
 510                          * registrations will probably fail too, so arrange to skip them.
 511                          * But we still have to execute this code for the remaining slots
 512                          * to make sure that we forget about the error queues we budgeted
 513                          * for those workers.  Otherwise, we'll wait for them to start,
 514                          * but they never will.
 515                          */
 516                         any_registrations_failed = true;
 517                         pcxt->worker[i].bgwhandle = NULL;
 518                         shm_mq_detach(pcxt->worker[i].error_mqh);
 519                         pcxt->worker[i].error_mqh = NULL;
 520                 }
 521         }
 522
 523         /* Restore previous memory context. */
 524         MemoryContextSwitchTo(oldcontext);
 525 }
 526
 527 /*
 528  * Wait for all workers to finish computing.
 529  *
 530  * Even if the parallel operation seems to have completed successfully, it's
 531  * important to call this function afterwards.  We must not miss any errors
 532  * the workers may have thrown during the parallel operation, or any that they
 533  * may yet throw while shutting down.
 534  *
 535  * Also, we want to update our notion of XactLastRecEnd based on worker
 536  * feedback.
 537  */
 538 void
 539 WaitForParallelWorkersToFinish(ParallelContext *pcxt)
 540 {
 541         for (;;)
 542         {
 543                 bool            anyone_alive = false;
 544                 int                     i;
 545
 546                 /*
 547                  * This will process any parallel messages that are pending, which may
 548                  * change the outcome of the loop that follows.  It may also throw an
 549                  * error propagated from a worker.
 550                  */
 551                 CHECK_FOR_INTERRUPTS();
 552
 553                 for (i = 0; i < pcxt->nworkers_launched; ++i)
 554                 {
 555                         if (pcxt->worker[i].error_mqh != NULL)
 556                         {
 557                                 anyone_alive = true;
 558                                 break;
 559                         }
 560                 }
 561
 562                 if (!anyone_alive)
 563                         break;
 564
 565                 WaitLatch(MyLatch, WL_LATCH_SET, -1,
 566                                   WAIT_EVENT_PARALLEL_FINISH);
 567                 ResetLatch(MyLatch);
 568         }
 569
 570         if (pcxt->toc != NULL)
 571         {
 572                 FixedParallelState *fps;
 573
 574                 fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED, false);
 575                 if (fps->last_xlog_end > XactLastRecEnd)
 576                         XactLastRecEnd = fps->last_xlog_end;
 577         }
 578 }
 579
 580 /*
 581  * Wait for all workers to exit.
 582  *
 583  * This function ensures that workers have been completely shutdown.  The
 584  * difference between WaitForParallelWorkersToFinish and this function is
 585  * that former just ensures that last message sent by worker backend is
 586  * received by master backend whereas this ensures the complete shutdown.
 587  */
 588 static void
 589 WaitForParallelWorkersToExit(ParallelContext *pcxt)
 590 {
 591         int                     i;
 592
 593         /* Wait until the workers actually die. */
 594         for (i = 0; i < pcxt->nworkers_launched; ++i)
 595         {
 596                 BgwHandleStatus status;
 597
 598                 if (pcxt->worker == NULL || pcxt->worker[i].bgwhandle == NULL)
 599                         continue;
 600
 601                 status = WaitForBackgroundWorkerShutdown(pcxt->worker[i].bgwhandle);
 602
 603                 /*
 604                  * If the postmaster kicked the bucket, we have no chance of cleaning
 605                  * up safely -- we won't be able to tell when our workers are actually
 606                  * dead.  This doesn't necessitate a PANIC since they will all abort
 607                  * eventually, but we can't safely continue this session.
 608                  */
 609                 if (status == BGWH_POSTMASTER_DIED)
 610                         ereport(FATAL,
 611                                         (errcode(ERRCODE_ADMIN_SHUTDOWN),
 612                                          errmsg("postmaster exited during a parallel transaction")));
 613
 614                 /* Release memory. */
 615                 pfree(pcxt->worker[i].bgwhandle);
 616                 pcxt->worker[i].bgwhandle = NULL;
 617         }
 618 }
 619
 620 /*
 621  * Destroy a parallel context.
 622  *
 623  * If expecting a clean exit, you should use WaitForParallelWorkersToFinish()
 624  * first, before calling this function.  When this function is invoked, any
 625  * remaining workers are forcibly killed; the dynamic shared memory segment
 626  * is unmapped; and we then wait (uninterruptibly) for the workers to exit.
 627  */
 628 void
 629 DestroyParallelContext(ParallelContext *pcxt)
 630 {
 631         int                     i;
 632
 633         /*
 634          * Be careful about order of operations here!  We remove the parallel
 635          * context from the list before we do anything else; otherwise, if an
 636          * error occurs during a subsequent step, we might try to nuke it again
 637          * from AtEOXact_Parallel or AtEOSubXact_Parallel.
 638          */
 639         dlist_delete(&pcxt->node);
 640
 641         /* Kill each worker in turn, and forget their error queues. */
 642         if (pcxt->worker != NULL)
 643         {
 644                 for (i = 0; i < pcxt->nworkers_launched; ++i)
 645                 {
 646                         if (pcxt->worker[i].error_mqh != NULL)
 647                         {
 648                                 TerminateBackgroundWorker(pcxt->worker[i].bgwhandle);
 649
 650                                 shm_mq_detach(pcxt->worker[i].error_mqh);
 651                                 pcxt->worker[i].error_mqh = NULL;
 652                         }
 653                 }
 654         }
 655
 656         /*
 657          * If we have allocated a shared memory segment, detach it.  This will
 658          * implicitly detach the error queues, and any other shared memory queues,
 659          * stored there.
 660          */
 661         if (pcxt->seg != NULL)
 662         {
 663                 dsm_detach(pcxt->seg);
 664                 pcxt->seg = NULL;
 665         }
 666
 667         /*
 668          * If this parallel context is actually in backend-private memory rather
 669          * than shared memory, free that memory instead.
 670          */
 671         if (pcxt->private_memory != NULL)
 672         {
 673                 pfree(pcxt->private_memory);
 674                 pcxt->private_memory = NULL;
 675         }
 676
 677         /*
 678          * We can't finish transaction commit or abort until all of the workers
 679          * have exited.  This means, in particular, that we can't respond to
 680          * interrupts at this stage.
 681          */
 682         HOLD_INTERRUPTS();
 683         WaitForParallelWorkersToExit(pcxt);
 684         RESUME_INTERRUPTS();
 685
 686         /* Free the worker array itself. */
 687         if (pcxt->worker != NULL)
 688         {
 689                 pfree(pcxt->worker);
 690                 pcxt->worker = NULL;
 691         }
 692
 693         /* Free memory. */
 694         pfree(pcxt->library_name);
 695         pfree(pcxt->function_name);
 696         pfree(pcxt);
 697 }
 698
 699 /*
 700  * Are there any parallel contexts currently active?
 701  */
 702 bool
 703 ParallelContextActive(void)
 704 {
 705         return !dlist_is_empty(&pcxt_list);
 706 }
 707
 708 /*
 709  * Handle receipt of an interrupt indicating a parallel worker message.
 710  *
 711  * Note: this is called within a signal handler!  All we can do is set
 712  * a flag that will cause the next CHECK_FOR_INTERRUPTS() to invoke
 713  * HandleParallelMessages().
 714  */
 715 void
 716 HandleParallelMessageInterrupt(void)
 717 {
 718         InterruptPending = true;
 719         ParallelMessagePending = true;
 720         SetLatch(MyLatch);
 721 }
 722
 723 /*
 724  * Handle any queued protocol messages received from parallel workers.
 725  */
 726 void
 727 HandleParallelMessages(void)
 728 {
 729         dlist_iter      iter;
 730         MemoryContext oldcontext;
 731
 732         static MemoryContext hpm_context = NULL;
 733
 734         /*
 735          * This is invoked from ProcessInterrupts(), and since some of the
 736          * functions it calls contain CHECK_FOR_INTERRUPTS(), there is a potential
 737          * for recursive calls if more signals are received while this runs.  It's
 738          * unclear that recursive entry would be safe, and it doesn't seem useful
 739          * even if it is safe, so let's block interrupts until done.
 740          */
 741         HOLD_INTERRUPTS();
 742
 743         /*
 744          * Moreover, CurrentMemoryContext might be pointing almost anywhere.  We
 745          * don't want to risk leaking data into long-lived contexts, so let's do
 746          * our work here in a private context that we can reset on each use.
 747          */
 748         if (hpm_context == NULL)        /* first time through? */
 749                 hpm_context = AllocSetContextCreate(TopMemoryContext,
 750                                                                                         "HandleParallelMessages",
 751                                                                                         ALLOCSET_DEFAULT_SIZES);
 752         else
 753                 MemoryContextReset(hpm_context);
 754
 755         oldcontext = MemoryContextSwitchTo(hpm_context);
 756
 757         /* OK to process messages.  Reset the flag saying there are more to do. */
 758         ParallelMessagePending = false;
 759
 760         dlist_foreach(iter, &pcxt_list)
 761         {
 762                 ParallelContext *pcxt;
 763                 int                     i;
 764
 765                 pcxt = dlist_container(ParallelContext, node, iter.cur);
 766                 if (pcxt->worker == NULL)
 767                         continue;
 768
 769                 for (i = 0; i < pcxt->nworkers_launched; ++i)
 770                 {
 771                         /*
 772                          * Read as many messages as we can from each worker, but stop when
 773                          * either (1) the worker's error queue goes away, which can happen
 774                          * if we receive a Terminate message from the worker; or (2) no
 775                          * more messages can be read from the worker without blocking.
 776                          */
 777                         while (pcxt->worker[i].error_mqh != NULL)
 778                         {
 779                                 shm_mq_result res;
 780                                 Size            nbytes;
 781                                 void       *data;
 782
 783                                 res = shm_mq_receive(pcxt->worker[i].error_mqh, &nbytes,
 784                                                                          &data, true);
 785                                 if (res == SHM_MQ_WOULD_BLOCK)
 786                                         break;
 787                                 else if (res == SHM_MQ_SUCCESS)
 788                                 {
 789                                         StringInfoData msg;
 790
 791                                         initStringInfo(&msg);
 792                                         appendBinaryStringInfo(&msg, data, nbytes);
 793                                         HandleParallelMessage(pcxt, i, &msg);
 794                                         pfree(msg.data);
 795                                 }
 796                                 else
 797                                         ereport(ERROR,
 798                                                         (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 799                                                          errmsg("lost connection to parallel worker")));
 800                         }
 801                 }
 802         }
 803
 804         MemoryContextSwitchTo(oldcontext);
 805
 806         /* Might as well clear the context on our way out */
 807         MemoryContextReset(hpm_context);
 808
 809         RESUME_INTERRUPTS();
 810 }
 811
 812 /*
 813  * Handle a single protocol message received from a single parallel worker.
 814  */
 815 static void
 816 HandleParallelMessage(ParallelContext *pcxt, int i, StringInfo msg)
 817 {
 818         char            msgtype;
 819
 820         msgtype = pq_getmsgbyte(msg);
 821
 822         switch (msgtype)
 823         {
 824                 case 'K':                               /* BackendKeyData */
 825                         {
 826                                 int32           pid = pq_getmsgint(msg, 4);
 827
 828                                 (void) pq_getmsgint(msg, 4);    /* discard cancel key */
 829                                 (void) pq_getmsgend(msg);
 830                                 pcxt->worker[i].pid = pid;
 831                                 break;
 832                         }
 833
 834                 case 'E':                               /* ErrorResponse */
 835                 case 'N':                               /* NoticeResponse */
 836                         {
 837                                 ErrorData       edata;
 838                                 ErrorContextCallback *save_error_context_stack;
 839
 840                                 /* Parse ErrorResponse or NoticeResponse. */
 841                                 pq_parse_errornotice(msg, &edata);
 842
 843                                 /* Death of a worker isn't enough justification for suicide. */
 844                                 edata.elevel = Min(edata.elevel, ERROR);
 845
 846                                 /*
 847                                  * If desired, add a context line to show that this is a
 848                                  * message propagated from a parallel worker.  Otherwise, it
 849                                  * can sometimes be confusing to understand what actually
 850                                  * happened.  (We don't do this in FORCE_PARALLEL_REGRESS mode
 851                                  * because it causes test-result instability depending on
 852                                  * whether a parallel worker is actually used or not.)
 853                                  */
 854                                 if (force_parallel_mode != FORCE_PARALLEL_REGRESS)
 855                                 {
 856                                         if (edata.context)
 857                                                 edata.context = psprintf("%s\n%s", edata.context,
 858                                                                                                  _("parallel worker"));
 859                                         else
 860                                                 edata.context = pstrdup(_("parallel worker"));
 861                                 }
 862
 863                                 /*
 864                                  * Context beyond that should use the error context callbacks
 865                                  * that were in effect when the ParallelContext was created,
 866                                  * not the current ones.
 867                                  */
 868                                 save_error_context_stack = error_context_stack;
 869                                 error_context_stack = pcxt->error_context_stack;
 870
 871                                 /* Rethrow error or print notice. */
 872                                 ThrowErrorData(&edata);
 873
 874                                 /* Not an error, so restore previous context stack. */
 875                                 error_context_stack = save_error_context_stack;
 876
 877                                 break;
 878                         }
 879
 880                 case 'A':                               /* NotifyResponse */
 881                         {
 882                                 /* Propagate NotifyResponse. */
 883                                 int32           pid;
 884                                 const char *channel;
 885                                 const char *payload;
 886
 887                                 pid = pq_getmsgint(msg, 4);
 888                                 channel = pq_getmsgrawstring(msg);
 889                                 payload = pq_getmsgrawstring(msg);
 890                                 pq_endmessage(msg);
 891
 892                                 NotifyMyFrontEnd(channel, payload, pid);
 893
 894                                 break;
 895                         }
 896
 897                 case 'X':                               /* Terminate, indicating clean exit */
 898                         {
 899                                 shm_mq_detach(pcxt->worker[i].error_mqh);
 900                                 pcxt->worker[i].error_mqh = NULL;
 901                                 break;
 902                         }
 903
 904                 default:
 905                         {
 906                                 elog(ERROR, "unrecognized message type received from parallel worker: %c (message length %d bytes)",
 907                                          msgtype, msg->len);
 908                         }
 909         }
 910 }
 911
 912 /*
 913  * End-of-subtransaction cleanup for parallel contexts.
 914  *
 915  * Currently, it's forbidden to enter or leave a subtransaction while
 916  * parallel mode is in effect, so we could just blow away everything.  But
 917  * we may want to relax that restriction in the future, so this code
 918  * contemplates that there may be multiple subtransaction IDs in pcxt_list.
 919  */
 920 void
 921 AtEOSubXact_Parallel(bool isCommit, SubTransactionId mySubId)
 922 {
 923         while (!dlist_is_empty(&pcxt_list))
 924         {
 925                 ParallelContext *pcxt;
 926
 927                 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
 928                 if (pcxt->subid != mySubId)
 929                         break;
 930                 if (isCommit)
 931                         elog(WARNING, "leaked parallel context");
 932                 DestroyParallelContext(pcxt);
 933         }
 934 }
 935
 936 /*
 937  * End-of-transaction cleanup for parallel contexts.
 938  */
 939 void
 940 AtEOXact_Parallel(bool isCommit)
 941 {
 942         while (!dlist_is_empty(&pcxt_list))
 943         {
 944                 ParallelContext *pcxt;
 945
 946                 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
 947                 if (isCommit)
 948                         elog(WARNING, "leaked parallel context");
 949                 DestroyParallelContext(pcxt);
 950         }
 951 }
 952
 953 /*
 954  * Main entrypoint for parallel workers.
 955  */
 956 void
 957 ParallelWorkerMain(Datum main_arg)
 958 {
 959         dsm_segment *seg;
 960         shm_toc    *toc;
 961         FixedParallelState *fps;
 962         char       *error_queue_space;
 963         shm_mq     *mq;
 964         shm_mq_handle *mqh;
 965         char       *libraryspace;
 966         char       *entrypointstate;
 967         char       *library_name;
 968         char       *function_name;
 969         parallel_worker_main_type entrypt;
 970         char       *gucspace;
 971         char       *combocidspace;
 972         char       *tsnapspace;
 973         char       *asnapspace;
 974         char       *tstatespace;
 975         StringInfoData msgbuf;
 976         char       *session_dsm_handle_space;
 977
 978         /* Set flag to indicate that we're initializing a parallel worker. */
 979         InitializingParallelWorker = true;
 980
 981         /* Establish signal handlers. */
 982         pqsignal(SIGTERM, die);
 983         BackgroundWorkerUnblockSignals();
 984
 985         /* Determine and set our parallel worker number. */
 986         Assert(ParallelWorkerNumber == -1);
 987         memcpy(&ParallelWorkerNumber, MyBgworkerEntry->bgw_extra, sizeof(int));
 988
 989         /* Set up a memory context and resource owner. */
 990         Assert(CurrentResourceOwner == NULL);
 991         CurrentResourceOwner = ResourceOwnerCreate(NULL, "parallel toplevel");
 992         CurrentMemoryContext = AllocSetContextCreate(TopMemoryContext,
 993                                                                                                  "Parallel worker",
 994                                                                                                  ALLOCSET_DEFAULT_SIZES);
 995
 996         /*
 997          * Now that we have a resource owner, we can attach to the dynamic shared
 998          * memory segment and read the table of contents.
 999          */
1000         seg = dsm_attach(DatumGetUInt32(main_arg));
1001         if (seg == NULL)
1002                 ereport(ERROR,
1003                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1004                                  errmsg("could not map dynamic shared memory segment")));
1005         toc = shm_toc_attach(PARALLEL_MAGIC, dsm_segment_address(seg));
1006         if (toc == NULL)
1007                 ereport(ERROR,
1008                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1009                                  errmsg("invalid magic number in dynamic shared memory segment")));
1010
1011         /* Look up fixed parallel state. */
1012         fps = shm_toc_lookup(toc, PARALLEL_KEY_FIXED, false);
1013         MyFixedParallelState = fps;
1014
1015         /*
1016          * Now that we have a worker number, we can find and attach to the error
1017          * queue provided for us.  That's good, because until we do that, any
1018          * errors that happen here will not be reported back to the process that
1019          * requested that this worker be launched.
1020          */
1021         error_queue_space = shm_toc_lookup(toc, PARALLEL_KEY_ERROR_QUEUE, false);
1022         mq = (shm_mq *) (error_queue_space +
1023                                          ParallelWorkerNumber * PARALLEL_ERROR_QUEUE_SIZE);
1024         shm_mq_set_sender(mq, MyProc);
1025         mqh = shm_mq_attach(mq, seg, NULL);
1026         pq_redirect_to_shm_mq(seg, mqh);
1027         pq_set_parallel_master(fps->parallel_master_pid,
1028                                                    fps->parallel_master_backend_id);
1029
1030         /*
1031          * Send a BackendKeyData message to the process that initiated parallelism
1032          * so that it has access to our PID before it receives any other messages
1033          * from us.  Our cancel key is sent, too, since that's the way the
1034          * protocol message is defined, but it won't actually be used for anything
1035          * in this case.
1036          */
1037         pq_beginmessage(&msgbuf, 'K');
1038         pq_sendint32(&msgbuf, (int32) MyProcPid);
1039         pq_sendint32(&msgbuf, (int32) MyCancelKey);
1040         pq_endmessage(&msgbuf);
1041
1042         /*
1043          * Hooray! Primary initialization is complete.  Now, we need to set up our
1044          * backend-local state to match the original backend.
1045          */
1046
1047         /*
1048          * Join locking group.  We must do this before anything that could try to
1049          * acquire a heavyweight lock, because any heavyweight locks acquired to
1050          * this point could block either directly against the parallel group
1051          * leader or against some process which in turn waits for a lock that
1052          * conflicts with the parallel group leader, causing an undetected
1053          * deadlock.  (If we can't join the lock group, the leader has gone away,
1054          * so just exit quietly.)
1055          */
1056         if (!BecomeLockGroupMember(fps->parallel_master_pgproc,
1057                                                            fps->parallel_master_pid))
1058                 return;
1059
1060         /*
1061          * Load libraries that were loaded by original backend.  We want to do
1062          * this before restoring GUCs, because the libraries might define custom
1063          * variables.
1064          */
1065         libraryspace = shm_toc_lookup(toc, PARALLEL_KEY_LIBRARY, false);
1066         RestoreLibraryState(libraryspace);
1067
1068         /*
1069          * Identify the entry point to be called.  In theory this could result in
1070          * loading an additional library, though most likely the entry point is in
1071          * the core backend or in a library we just loaded.
1072          */
1073         entrypointstate = shm_toc_lookup(toc, PARALLEL_KEY_ENTRYPOINT, false);
1074         library_name = entrypointstate;
1075         function_name = entrypointstate + strlen(library_name) + 1;
1076
1077         entrypt = LookupParallelWorkerFunction(library_name, function_name);
1078
1079         /* Restore database connection. */
1080         BackgroundWorkerInitializeConnectionByOid(fps->database_id,
1081                                                                                           fps->authenticated_user_id);
1082
1083         /*
1084          * Set the client encoding to the database encoding, since that is what
1085          * the leader will expect.
1086          */
1087         SetClientEncoding(GetDatabaseEncoding());
1088
1089         /* Restore GUC values from launching backend. */
1090         gucspace = shm_toc_lookup(toc, PARALLEL_KEY_GUC, false);
1091         StartTransactionCommand();
1092         RestoreGUCState(gucspace);
1093         CommitTransactionCommand();
1094
1095         /* Crank up a transaction state appropriate to a parallel worker. */
1096         tstatespace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_STATE, false);
1097         StartParallelWorkerTransaction(tstatespace);
1098
1099         /* Restore combo CID state. */
1100         combocidspace = shm_toc_lookup(toc, PARALLEL_KEY_COMBO_CID, false);
1101         RestoreComboCIDState(combocidspace);
1102
1103         /* Attach to the per-session DSM segment and contained objects. */
1104         session_dsm_handle_space =
1105                 shm_toc_lookup(toc, PARALLEL_KEY_SESSION_DSM, false);
1106         AttachSession(*(dsm_handle *) session_dsm_handle_space);
1107
1108         /* Restore transaction snapshot. */
1109         tsnapspace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT, false);
1110         RestoreTransactionSnapshot(RestoreSnapshot(tsnapspace),
1111                                                            fps->parallel_master_pgproc);
1112
1113         /* Restore active snapshot. */
1114         asnapspace = shm_toc_lookup(toc, PARALLEL_KEY_ACTIVE_SNAPSHOT, false);
1115         PushActiveSnapshot(RestoreSnapshot(asnapspace));
1116
1117         /*
1118          * We've changed which tuples we can see, and must therefore invalidate
1119          * system caches.
1120          */
1121         InvalidateSystemCaches();
1122
1123         /*
1124          * Restore current role id.  Skip verifying whether session user is
1125          * allowed to become this role and blindly restore the leader's state for
1126          * current role.
1127          */
1128         SetCurrentRoleId(fps->outer_user_id, fps->is_superuser);
1129
1130         /* Restore user ID and security context. */
1131         SetUserIdAndSecContext(fps->current_user_id, fps->sec_context);
1132
1133         /* Restore temp-namespace state to ensure search path matches leader's. */
1134         SetTempNamespaceState(fps->temp_namespace_id,
1135                                                   fps->temp_toast_namespace_id);
1136
1137         /* Set ParallelMasterBackendId so we know how to address temp relations. */
1138         ParallelMasterBackendId = fps->parallel_master_backend_id;
1139
1140         /*
1141          * We've initialized all of our state now; nothing should change
1142          * hereafter.
1143          */
1144         InitializingParallelWorker = false;
1145         EnterParallelMode();
1146
1147         /*
1148          * Time to do the real work: invoke the caller-supplied code.
1149          */
1150         entrypt(seg, toc);
1151
1152         /* Must exit parallel mode to pop active snapshot. */
1153         ExitParallelMode();
1154
1155         /* Must pop active snapshot so resowner.c doesn't complain. */
1156         PopActiveSnapshot();
1157
1158         /* Shut down the parallel-worker transaction. */
1159         EndParallelWorkerTransaction();
1160
1161         /* Detach from the per-session DSM segment. */
1162         DetachSession();
1163
1164         /* Report success. */
1165         pq_putmessage('X', NULL, 0);
1166 }
1167
1168 /*
1169  * Update shared memory with the ending location of the last WAL record we
1170  * wrote, if it's greater than the value already stored there.
1171  */
1172 void
1173 ParallelWorkerReportLastRecEnd(XLogRecPtr last_xlog_end)
1174 {
1175         FixedParallelState *fps = MyFixedParallelState;
1176
1177         Assert(fps != NULL);
1178         SpinLockAcquire(&fps->mutex);
1179         if (fps->last_xlog_end < last_xlog_end)
1180                 fps->last_xlog_end = last_xlog_end;
1181         SpinLockRelease(&fps->mutex);
1182 }
1183
1184 /*
1185  * Look up (and possibly load) a parallel worker entry point function.
1186  *
1187  * For functions contained in the core code, we use library name "postgres"
1188  * and consult the InternalParallelWorkers array.  External functions are
1189  * looked up, and loaded if necessary, using load_external_function().
1190  *
1191  * The point of this is to pass function names as strings across process
1192  * boundaries.  We can't pass actual function addresses because of the
1193  * possibility that the function has been loaded at a different address
1194  * in a different process.  This is obviously a hazard for functions in
1195  * loadable libraries, but it can happen even for functions in the core code
1196  * on platforms using EXEC_BACKEND (e.g., Windows).
1197  *
1198  * At some point it might be worthwhile to get rid of InternalParallelWorkers[]
1199  * in favor of applying load_external_function() for core functions too;
1200  * but that raises portability issues that are not worth addressing now.
1201  */
1202 static parallel_worker_main_type
1203 LookupParallelWorkerFunction(const char *libraryname, const char *funcname)
1204 {
1205         /*
1206          * If the function is to be loaded from postgres itself, search the
1207          * InternalParallelWorkers array.
1208          */
1209         if (strcmp(libraryname, "postgres") == 0)
1210         {
1211                 int                     i;
1212
1213                 for (i = 0; i < lengthof(InternalParallelWorkers); i++)
1214                 {
1215                         if (strcmp(InternalParallelWorkers[i].fn_name, funcname) == 0)
1216                                 return InternalParallelWorkers[i].fn_addr;
1217                 }
1218
1219                 /* We can only reach this by programming error. */
1220                 elog(ERROR, "internal function \"%s\" not found", funcname);
1221         }
1222
1223         /* Otherwise load from external library. */
1224         return (parallel_worker_main_type)
1225                 load_external_function(libraryname, funcname, true, NULL);
1226 }