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) 1996-2014, 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/xact.h"
  18 #include "access/xlog.h"
  19 #include "access/parallel.h"
  20 #include "commands/async.h"
  21 #include "libpq/libpq.h"
  22 #include "libpq/pqformat.h"
  23 #include "libpq/pqmq.h"
  24 #include "miscadmin.h"
  25 #include "storage/ipc.h"
  26 #include "storage/sinval.h"
  27 #include "storage/spin.h"
  28 #include "tcop/tcopprot.h"
  29 #include "utils/combocid.h"
  30 #include "utils/guc.h"
  31 #include "utils/inval.h"
  32 #include "utils/memutils.h"
  33 #include "utils/resowner.h"
  34 #include "utils/snapmgr.h"
  35
  36 /*
  37  * We don't want to waste a lot of memory on an error queue which, most of
  38  * the time, will process only a handful of small messages.  However, it is
  39  * desirable to make it large enough that a typical ErrorResponse can be sent
  40  * without blocking.  That way, a worker that errors out can write the whole
  41  * message into the queue and terminate without waiting for the user backend.
  42  */
  43 #define PARALLEL_ERROR_QUEUE_SIZE                       16384
  44
  45 /* Magic number for parallel context TOC. */
  46 #define PARALLEL_MAGIC                                          0x50477c7c
  47
  48 /*
  49  * Magic numbers for parallel state sharing.  Higher-level code should use
  50  * smaller values, leaving these very large ones for use by this module.
  51  */
  52 #define PARALLEL_KEY_FIXED                                      UINT64CONST(0xFFFFFFFFFFFF0001)
  53 #define PARALLEL_KEY_ERROR_QUEUE                        UINT64CONST(0xFFFFFFFFFFFF0002)
  54 #define PARALLEL_KEY_LIBRARY                            UINT64CONST(0xFFFFFFFFFFFF0003)
  55 #define PARALLEL_KEY_GUC                                        UINT64CONST(0xFFFFFFFFFFFF0004)
  56 #define PARALLEL_KEY_COMBO_CID                          UINT64CONST(0xFFFFFFFFFFFF0005)
  57 #define PARALLEL_KEY_TRANSACTION_SNAPSHOT       UINT64CONST(0xFFFFFFFFFFFF0006)
  58 #define PARALLEL_KEY_ACTIVE_SNAPSHOT            UINT64CONST(0xFFFFFFFFFFFF0007)
  59 #define PARALLEL_KEY_TRANSACTION_STATE          UINT64CONST(0xFFFFFFFFFFFF0008)
  60 #define PARALLEL_KEY_EXTENSION_TRAMPOLINE       UINT64CONST(0xFFFFFFFFFFFF0009)
  61
  62 /* Fixed-size parallel state. */
  63 typedef struct FixedParallelState
  64 {
  65         /* Fixed-size state that workers must restore. */
  66         Oid                     database_id;
  67         Oid                     authenticated_user_id;
  68         Oid                     current_user_id;
  69         int                     sec_context;
  70         PGPROC     *parallel_master_pgproc;
  71         pid_t           parallel_master_pid;
  72         BackendId       parallel_master_backend_id;
  73
  74         /* Entrypoint for parallel workers. */
  75         parallel_worker_main_type entrypoint;
  76
  77         /* Mutex protects remaining fields. */
  78         slock_t         mutex;
  79
  80         /* Track whether workers have attached. */
  81         int                     workers_expected;
  82         int                     workers_attached;
  83
  84         /* Maximum XactLastRecEnd of any worker. */
  85         XLogRecPtr      last_xlog_end;
  86 } FixedParallelState;
  87
  88 /*
  89  * Our parallel worker number.  We initialize this to -1, meaning that we are
  90  * not a parallel worker.  In parallel workers, it will be set to a value >= 0
  91  * and < the number of workers before any user code is invoked; each parallel
  92  * worker will get a different parallel worker number.
  93  */
  94 int                     ParallelWorkerNumber = -1;
  95
  96 /* Is there a parallel message pending which we need to receive? */
  97 bool            ParallelMessagePending = false;
  98
  99 /* Are we initializing a parallel worker? */
 100 bool            InitializingParallelWorker = false;
 101
 102 /* Pointer to our fixed parallel state. */
 103 static FixedParallelState *MyFixedParallelState;
 104
 105 /* List of active parallel contexts. */
 106 static dlist_head pcxt_list = DLIST_STATIC_INIT(pcxt_list);
 107
 108 /* Private functions. */
 109 static void HandleParallelMessage(ParallelContext *, int, StringInfo msg);
 110 static void ParallelErrorContext(void *arg);
 111 static void ParallelExtensionTrampoline(dsm_segment *seg, shm_toc *toc);
 112 static void ParallelWorkerMain(Datum main_arg);
 113
 114 /*
 115  * Establish a new parallel context.  This should be done after entering
 116  * parallel mode, and (unless there is an error) the context should be
 117  * destroyed before exiting the current subtransaction.
 118  */
 119 ParallelContext *
 120 CreateParallelContext(parallel_worker_main_type entrypoint, int nworkers)
 121 {
 122         MemoryContext oldcontext;
 123         ParallelContext *pcxt;
 124
 125         /* It is unsafe to create a parallel context if not in parallel mode. */
 126         Assert(IsInParallelMode());
 127
 128         /* Number of workers should be non-negative. */
 129         Assert(nworkers >= 0);
 130
 131         /*
 132          * If dynamic shared memory is not available, we won't be able to use
 133          * background workers.
 134          */
 135         if (dynamic_shared_memory_type == DSM_IMPL_NONE)
 136                 nworkers = 0;
 137
 138         /*
 139          * If we are running under serializable isolation, we can't use
 140          * parallel workers, at least not until somebody enhances that mechanism
 141          * to be parallel-aware.
 142          */
 143         if (IsolationIsSerializable())
 144                 nworkers = 0;
 145
 146         /* We might be running in a short-lived memory context. */
 147         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 148
 149         /* Initialize a new ParallelContext. */
 150         pcxt = palloc0(sizeof(ParallelContext));
 151         pcxt->subid = GetCurrentSubTransactionId();
 152         pcxt->nworkers = nworkers;
 153         pcxt->entrypoint = entrypoint;
 154         pcxt->error_context_stack = error_context_stack;
 155         shm_toc_initialize_estimator(&pcxt->estimator);
 156         dlist_push_head(&pcxt_list, &pcxt->node);
 157
 158         /* Restore previous memory context. */
 159         MemoryContextSwitchTo(oldcontext);
 160
 161         return pcxt;
 162 }
 163
 164 /*
 165  * Establish a new parallel context that calls a function provided by an
 166  * extension.  This works around the fact that the library might get mapped
 167  * at a different address in each backend.
 168  */
 169 ParallelContext *
 170 CreateParallelContextForExternalFunction(char *library_name,
 171                                                                                  char *function_name,
 172                                                                                  int nworkers)
 173 {
 174         MemoryContext oldcontext;
 175         ParallelContext *pcxt;
 176
 177         /* We might be running in a very short-lived memory context. */
 178         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 179
 180         /* Create the context. */
 181         pcxt = CreateParallelContext(ParallelExtensionTrampoline, nworkers);
 182         pcxt->library_name = pstrdup(library_name);
 183         pcxt->function_name = pstrdup(function_name);
 184
 185         /* Restore previous memory context. */
 186         MemoryContextSwitchTo(oldcontext);
 187
 188         return pcxt;
 189 }
 190
 191 /*
 192  * Establish the dynamic shared memory segment for a parallel context and
 193  * copied state and other bookkeeping information that will need by parallel
 194  * workers into it.
 195  */
 196 void
 197 InitializeParallelDSM(ParallelContext *pcxt)
 198 {
 199         MemoryContext oldcontext;
 200         Size            library_len = 0;
 201         Size            guc_len = 0;
 202         Size            combocidlen = 0;
 203         Size            tsnaplen = 0;
 204         Size            asnaplen = 0;
 205         Size            tstatelen = 0;
 206         Size            segsize = 0;
 207         int                     i;
 208         FixedParallelState *fps;
 209         Snapshot        transaction_snapshot = GetTransactionSnapshot();
 210         Snapshot        active_snapshot = GetActiveSnapshot();
 211
 212         /* We might be running in a very short-lived memory context. */
 213         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 214
 215         /* Allow space to store the fixed-size parallel state. */
 216         shm_toc_estimate_chunk(&pcxt->estimator, sizeof(FixedParallelState));
 217         shm_toc_estimate_keys(&pcxt->estimator, 1);
 218
 219         /*
 220          * Normally, the user will have requested at least one worker process, but
 221          * if by chance they have not, we can skip a bunch of things here.
 222          */
 223         if (pcxt->nworkers > 0)
 224         {
 225                 /* Estimate space for various kinds of state sharing. */
 226                 library_len = EstimateLibraryStateSpace();
 227                 shm_toc_estimate_chunk(&pcxt->estimator, library_len);
 228                 guc_len = EstimateGUCStateSpace();
 229                 shm_toc_estimate_chunk(&pcxt->estimator, guc_len);
 230                 combocidlen = EstimateComboCIDStateSpace();
 231                 shm_toc_estimate_chunk(&pcxt->estimator, combocidlen);
 232                 tsnaplen = EstimateSnapshotSpace(transaction_snapshot);
 233                 shm_toc_estimate_chunk(&pcxt->estimator, tsnaplen);
 234                 asnaplen = EstimateSnapshotSpace(active_snapshot);
 235                 shm_toc_estimate_chunk(&pcxt->estimator, asnaplen);
 236                 tstatelen = EstimateTransactionStateSpace();
 237                 shm_toc_estimate_chunk(&pcxt->estimator, tstatelen);
 238                 /* If you add more chunks here, you probably need to add keys. */
 239                 shm_toc_estimate_keys(&pcxt->estimator, 6);
 240
 241                 /* Estimate space need for error queues. */
 242                 StaticAssertStmt(BUFFERALIGN(PARALLEL_ERROR_QUEUE_SIZE) ==
 243                                                  PARALLEL_ERROR_QUEUE_SIZE,
 244                                                  "parallel error queue size not buffer-aligned");
 245                 shm_toc_estimate_chunk(&pcxt->estimator,
 246                                                            PARALLEL_ERROR_QUEUE_SIZE * pcxt->nworkers);
 247                 shm_toc_estimate_keys(&pcxt->estimator, 1);
 248
 249                 /* Estimate how much we'll need for extension entrypoint info. */
 250                 if (pcxt->library_name != NULL)
 251                 {
 252                         Assert(pcxt->entrypoint == ParallelExtensionTrampoline);
 253                         Assert(pcxt->function_name != NULL);
 254                         shm_toc_estimate_chunk(&pcxt->estimator, strlen(pcxt->library_name)
 255                                                                    + strlen(pcxt->function_name) + 2);
 256                         shm_toc_estimate_keys(&pcxt->estimator, 1);
 257                 }
 258         }
 259
 260         /*
 261          * Create DSM and initialize with new table of contents.  But if the user
 262          * didn't request any workers, then don't bother creating a dynamic shared
 263          * memory segment; instead, just use backend-private memory.
 264          *
 265          * Also, if we can't create a dynamic shared memory segment because the
 266          * maximum number of segments have already been created, then fall back to
 267          * backend-private memory, and plan not to use any workers.  We hope this
 268          * won't happen very often, but it's better to abandon the use of
 269          * parallelism than to fail outright.
 270          */
 271         segsize = shm_toc_estimate(&pcxt->estimator);
 272         if (pcxt->nworkers != 0)
 273                 pcxt->seg = dsm_create(segsize, DSM_CREATE_NULL_IF_MAXSEGMENTS);
 274         if (pcxt->seg != NULL)
 275                 pcxt->toc = shm_toc_create(PARALLEL_MAGIC,
 276                                                                    dsm_segment_address(pcxt->seg),
 277                                                                    segsize);
 278         else
 279         {
 280                 pcxt->nworkers = 0;
 281                 pcxt->private_memory = MemoryContextAlloc(TopMemoryContext, segsize);
 282                 pcxt->toc = shm_toc_create(PARALLEL_MAGIC, pcxt->private_memory,
 283                                                                    segsize);
 284         }
 285
 286         /* Initialize fixed-size state in shared memory. */
 287         fps = (FixedParallelState *)
 288                 shm_toc_allocate(pcxt->toc, sizeof(FixedParallelState));
 289         fps->database_id = MyDatabaseId;
 290         fps->authenticated_user_id = GetAuthenticatedUserId();
 291         GetUserIdAndSecContext(&fps->current_user_id, &fps->sec_context);
 292         fps->parallel_master_pgproc = MyProc;
 293         fps->parallel_master_pid = MyProcPid;
 294         fps->parallel_master_backend_id = MyBackendId;
 295         fps->entrypoint = pcxt->entrypoint;
 296         SpinLockInit(&fps->mutex);
 297         fps->workers_expected = pcxt->nworkers;
 298         fps->workers_attached = 0;
 299         fps->last_xlog_end = 0;
 300         shm_toc_insert(pcxt->toc, PARALLEL_KEY_FIXED, fps);
 301
 302         /* We can skip the rest of this if we're not budgeting for any workers. */
 303         if (pcxt->nworkers > 0)
 304         {
 305                 char       *libraryspace;
 306                 char       *gucspace;
 307                 char       *combocidspace;
 308                 char       *tsnapspace;
 309                 char       *asnapspace;
 310                 char       *tstatespace;
 311                 char       *error_queue_space;
 312
 313                 /* Serialize shared libraries we have loaded. */
 314                 libraryspace = shm_toc_allocate(pcxt->toc, library_len);
 315                 SerializeLibraryState(library_len, libraryspace);
 316                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_LIBRARY, libraryspace);
 317
 318                 /* Serialize GUC settings. */
 319                 gucspace = shm_toc_allocate(pcxt->toc, guc_len);
 320                 SerializeGUCState(guc_len, gucspace);
 321                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_GUC, gucspace);
 322
 323                 /* Serialize combo CID state. */
 324                 combocidspace = shm_toc_allocate(pcxt->toc, combocidlen);
 325                 SerializeComboCIDState(combocidlen, combocidspace);
 326                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_COMBO_CID, combocidspace);
 327
 328                 /* Serialize transaction snapshot and active snapshot. */
 329                 tsnapspace = shm_toc_allocate(pcxt->toc, tsnaplen);
 330                 SerializeSnapshot(transaction_snapshot, tsnapspace);
 331                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT,
 332                                            tsnapspace);
 333                 asnapspace = shm_toc_allocate(pcxt->toc, asnaplen);
 334                 SerializeSnapshot(active_snapshot, asnapspace);
 335                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ACTIVE_SNAPSHOT, asnapspace);
 336
 337                 /* Serialize transaction state. */
 338                 tstatespace = shm_toc_allocate(pcxt->toc, tstatelen);
 339                 SerializeTransactionState(tstatelen, tstatespace);
 340                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_STATE, tstatespace);
 341
 342                 /* Allocate space for worker information. */
 343                 pcxt->worker = palloc0(sizeof(ParallelWorkerInfo) * pcxt->nworkers);
 344
 345                 /*
 346                  * Establish error queues in dynamic shared memory.
 347                  *
 348                  * These queues should be used only for transmitting ErrorResponse,
 349                  * NoticeResponse, and NotifyResponse protocol messages.  Tuple data
 350                  * should be transmitted via separate (possibly larger?) queues.
 351                  */
 352                 error_queue_space =
 353                         shm_toc_allocate(pcxt->toc,
 354                                                          PARALLEL_ERROR_QUEUE_SIZE * pcxt->nworkers);
 355                 for (i = 0; i < pcxt->nworkers; ++i)
 356                 {
 357                         char       *start;
 358                         shm_mq     *mq;
 359
 360                         start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
 361                         mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
 362                         shm_mq_set_receiver(mq, MyProc);
 363                         pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
 364                 }
 365                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE, error_queue_space);
 366
 367                 /* Serialize extension entrypoint information. */
 368                 if (pcxt->library_name != NULL)
 369                 {
 370                         Size            lnamelen = strlen(pcxt->library_name);
 371                         char       *extensionstate;
 372
 373                         extensionstate = shm_toc_allocate(pcxt->toc, lnamelen
 374                                                                                   + strlen(pcxt->function_name) + 2);
 375                         strcpy(extensionstate, pcxt->library_name);
 376                         strcpy(extensionstate + lnamelen + 1, pcxt->function_name);
 377                         shm_toc_insert(pcxt->toc, PARALLEL_KEY_EXTENSION_TRAMPOLINE,
 378                                                    extensionstate);
 379                 }
 380         }
 381
 382         /* Restore previous memory context. */
 383         MemoryContextSwitchTo(oldcontext);
 384 }
 385
 386 /*
 387  * Launch parallel workers.
 388  */
 389 void
 390 LaunchParallelWorkers(ParallelContext *pcxt)
 391 {
 392         MemoryContext oldcontext;
 393         BackgroundWorker worker;
 394         int                     i;
 395         bool            any_registrations_failed = false;
 396
 397         /* Skip this if we have no workers. */
 398         if (pcxt->nworkers == 0)
 399                 return;
 400
 401         /* If we do have workers, we'd better have a DSM segment. */
 402         Assert(pcxt->seg != NULL);
 403
 404         /* We might be running in a short-lived memory context. */
 405         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 406
 407         /*
 408          * This function can be called for a parallel context for which it has
 409          * already been called previously, but only if all of the old workers
 410          * have already exited.  When this case arises, we need to do some extra
 411          * reinitialization.
 412          */
 413         if (pcxt->nworkers_launched > 0)
 414         {
 415                 FixedParallelState *fps;
 416                 char       *error_queue_space;
 417
 418                 /* Clean out old worker handles. */
 419                 for (i = 0; i < pcxt->nworkers; ++i)
 420                 {
 421                         if (pcxt->worker[i].error_mqh != NULL)
 422                                 elog(ERROR, "previously launched worker still alive");
 423                         if (pcxt->worker[i].bgwhandle != NULL)
 424                         {
 425                                 pfree(pcxt->worker[i].bgwhandle);
 426                                 pcxt->worker[i].bgwhandle = NULL;
 427                         }
 428                 }
 429
 430                 /* Reset a few bits of fixed parallel state to a clean state. */
 431                 fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED);
 432                 fps->workers_attached = 0;
 433                 fps->last_xlog_end = 0;
 434
 435                 /* Recreate error queues. */
 436                 error_queue_space =
 437                         shm_toc_lookup(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE);
 438                 for (i = 0; i < pcxt->nworkers; ++i)
 439                 {
 440                         char       *start;
 441                         shm_mq     *mq;
 442
 443                         start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
 444                         mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
 445                         shm_mq_set_receiver(mq, MyProc);
 446                         pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
 447                 }
 448
 449                 /* Reset number of workers launched. */
 450                 pcxt->nworkers_launched = 0;
 451         }
 452
 453         /* Configure a worker. */
 454         snprintf(worker.bgw_name, BGW_MAXLEN, "parallel worker for PID %d",
 455                          MyProcPid);
 456         worker.bgw_flags =
 457                 BGWORKER_SHMEM_ACCESS | BGWORKER_BACKEND_DATABASE_CONNECTION;
 458         worker.bgw_start_time = BgWorkerStart_ConsistentState;
 459         worker.bgw_restart_time = BGW_NEVER_RESTART;
 460         worker.bgw_main = ParallelWorkerMain;
 461         worker.bgw_main_arg = UInt32GetDatum(dsm_segment_handle(pcxt->seg));
 462         worker.bgw_notify_pid = MyProcPid;
 463
 464         /*
 465          * Start workers.
 466          *
 467          * The caller must be able to tolerate ending up with fewer workers than
 468          * expected, so there is no need to throw an error here if registration
 469          * fails.  It wouldn't help much anyway, because registering the worker in
 470          * no way guarantees that it will start up and initialize successfully.
 471          */
 472         for (i = 0; i < pcxt->nworkers; ++i)
 473         {
 474                 if (!any_registrations_failed &&
 475                         RegisterDynamicBackgroundWorker(&worker,
 476                                                                                         &pcxt->worker[i].bgwhandle))
 477                 {
 478                         shm_mq_set_handle(pcxt->worker[i].error_mqh,
 479                                                           pcxt->worker[i].bgwhandle);
 480                         pcxt->nworkers_launched++;
 481                 }
 482                 else
 483                 {
 484                         /*
 485                          * If we weren't able to register the worker, then we've bumped up
 486                          * against the max_worker_processes limit, and future
 487                          * registrations will probably fail too, so arrange to skip them.
 488                          * But we still have to execute this code for the remaining slots
 489                          * to make sure that we forget about the error queues we budgeted
 490                          * for those workers.  Otherwise, we'll wait for them to start,
 491                          * but they never will.
 492                          */
 493                         any_registrations_failed = true;
 494                         pcxt->worker[i].bgwhandle = NULL;
 495                         pcxt->worker[i].error_mqh = NULL;
 496                 }
 497         }
 498
 499         /* Restore previous memory context. */
 500         MemoryContextSwitchTo(oldcontext);
 501 }
 502
 503 /*
 504  * Wait for all workers to exit.
 505  *
 506  * Even if the parallel operation seems to have completed successfully, it's
 507  * important to call this function afterwards.  We must not miss any errors
 508  * the workers may have thrown during the parallel operation, or any that they
 509  * may yet throw while shutting down.
 510  *
 511  * Also, we want to update our notion of XactLastRecEnd based on worker
 512  * feedback.
 513  */
 514 void
 515 WaitForParallelWorkersToFinish(ParallelContext *pcxt)
 516 {
 517         for (;;)
 518         {
 519                 bool            anyone_alive = false;
 520                 int                     i;
 521
 522                 /*
 523                  * This will process any parallel messages that are pending, which may
 524                  * change the outcome of the loop that follows.  It may also throw an
 525                  * error propagated from a worker.
 526                  */
 527                 CHECK_FOR_INTERRUPTS();
 528
 529                 for (i = 0; i < pcxt->nworkers; ++i)
 530                 {
 531                         if (pcxt->worker[i].error_mqh != NULL)
 532                         {
 533                                 anyone_alive = true;
 534                                 break;
 535                         }
 536                 }
 537
 538                 if (!anyone_alive)
 539                         break;
 540
 541                 WaitLatch(&MyProc->procLatch, WL_LATCH_SET, -1);
 542                 ResetLatch(&MyProc->procLatch);
 543         }
 544
 545         if (pcxt->toc != NULL)
 546         {
 547                 FixedParallelState *fps;
 548
 549                 fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED);
 550                 if (fps->last_xlog_end > XactLastRecEnd)
 551                         XactLastRecEnd = fps->last_xlog_end;
 552         }
 553 }
 554
 555 /*
 556  * Destroy a parallel context.
 557  *
 558  * If expecting a clean exit, you should use WaitForParallelWorkersToFinish()
 559  * first, before calling this function.  When this function is invoked, any
 560  * remaining workers are forcibly killed; the dynamic shared memory segment
 561  * is unmapped; and we then wait (uninterruptibly) for the workers to exit.
 562  */
 563 void
 564 DestroyParallelContext(ParallelContext *pcxt)
 565 {
 566         int                     i;
 567
 568         /*
 569          * Be careful about order of operations here!  We remove the parallel
 570          * context from the list before we do anything else; otherwise, if an
 571          * error occurs during a subsequent step, we might try to nuke it again
 572          * from AtEOXact_Parallel or AtEOSubXact_Parallel.
 573          */
 574         dlist_delete(&pcxt->node);
 575
 576         /* Kill each worker in turn, and forget their error queues. */
 577         if (pcxt->worker != NULL)
 578         {
 579                 for (i = 0; i < pcxt->nworkers; ++i)
 580                 {
 581                         if (pcxt->worker[i].bgwhandle != NULL)
 582                                 TerminateBackgroundWorker(pcxt->worker[i].bgwhandle);
 583                         if (pcxt->worker[i].error_mqh != NULL)
 584                         {
 585                                 pfree(pcxt->worker[i].error_mqh);
 586                                 pcxt->worker[i].error_mqh = NULL;
 587                         }
 588                 }
 589         }
 590
 591         /*
 592          * If we have allocated a shared memory segment, detach it.  This will
 593          * implicitly detach the error queues, and any other shared memory queues,
 594          * stored there.
 595          */
 596         if (pcxt->seg != NULL)
 597         {
 598                 dsm_detach(pcxt->seg);
 599                 pcxt->seg = NULL;
 600         }
 601
 602         /*
 603          * If this parallel context is actually in backend-private memory rather
 604          * than shared memory, free that memory instead.
 605          */
 606         if (pcxt->private_memory != NULL)
 607         {
 608                 pfree(pcxt->private_memory);
 609                 pcxt->private_memory = NULL;
 610         }
 611
 612         /* Wait until the workers actually die. */
 613         for (i = 0; i < pcxt->nworkers; ++i)
 614         {
 615                 BgwHandleStatus status;
 616
 617                 if (pcxt->worker == NULL || pcxt->worker[i].bgwhandle == NULL)
 618                         continue;
 619
 620                 /*
 621                  * We can't finish transaction commit or abort until all of the
 622                  * workers are dead.  This means, in particular, that we can't respond
 623                  * to interrupts at this stage.
 624                  */
 625                 HOLD_INTERRUPTS();
 626                 status = WaitForBackgroundWorkerShutdown(pcxt->worker[i].bgwhandle);
 627                 RESUME_INTERRUPTS();
 628
 629                 /*
 630                  * If the postmaster kicked the bucket, we have no chance of cleaning
 631                  * up safely -- we won't be able to tell when our workers are actually
 632                  * dead.  This doesn't necessitate a PANIC since they will all abort
 633                  * eventually, but we can't safely continue this session.
 634                  */
 635                 if (status == BGWH_POSTMASTER_DIED)
 636                         ereport(FATAL,
 637                                         (errcode(ERRCODE_ADMIN_SHUTDOWN),
 638                                  errmsg("postmaster exited during a parallel transaction")));
 639
 640                 /* Release memory. */
 641                 pfree(pcxt->worker[i].bgwhandle);
 642                 pcxt->worker[i].bgwhandle = NULL;
 643         }
 644
 645         /* Free the worker array itself. */
 646         if (pcxt->worker != NULL)
 647         {
 648                 pfree(pcxt->worker);
 649                 pcxt->worker = NULL;
 650         }
 651
 652         /* Free memory. */
 653         pfree(pcxt);
 654 }
 655
 656 /*
 657  * Are there any parallel contexts currently active?
 658  */
 659 bool
 660 ParallelContextActive(void)
 661 {
 662         return !dlist_is_empty(&pcxt_list);
 663 }
 664
 665 /*
 666  * Handle receipt of an interrupt indicating a parallel worker message.
 667  */
 668 void
 669 HandleParallelMessageInterrupt(void)
 670 {
 671         int                     save_errno = errno;
 672
 673         InterruptPending = true;
 674         ParallelMessagePending = true;
 675         SetLatch(MyLatch);
 676
 677         errno = save_errno;
 678 }
 679
 680 /*
 681  * Handle any queued protocol messages received from parallel workers.
 682  */
 683 void
 684 HandleParallelMessages(void)
 685 {
 686         dlist_iter      iter;
 687
 688         ParallelMessagePending = false;
 689
 690         dlist_foreach(iter, &pcxt_list)
 691         {
 692                 ParallelContext *pcxt;
 693                 int                     i;
 694                 Size            nbytes;
 695                 void       *data;
 696
 697                 pcxt = dlist_container(ParallelContext, node, iter.cur);
 698                 if (pcxt->worker == NULL)
 699                         continue;
 700
 701                 for (i = 0; i < pcxt->nworkers; ++i)
 702                 {
 703                         /*
 704                          * Read as many messages as we can from each worker, but stop when
 705                          * either (1) the error queue goes away, which can happen if we
 706                          * receive a Terminate message from the worker; or (2) no more
 707                          * messages can be read from the worker without blocking.
 708                          */
 709                         while (pcxt->worker[i].error_mqh != NULL)
 710                         {
 711                                 shm_mq_result res;
 712
 713                                 res = shm_mq_receive(pcxt->worker[i].error_mqh, &nbytes,
 714                                                                          &data, true);
 715                                 if (res == SHM_MQ_WOULD_BLOCK)
 716                                         break;
 717                                 else if (res == SHM_MQ_SUCCESS)
 718                                 {
 719                                         StringInfoData msg;
 720
 721                                         initStringInfo(&msg);
 722                                         appendBinaryStringInfo(&msg, data, nbytes);
 723                                         HandleParallelMessage(pcxt, i, &msg);
 724                                         pfree(msg.data);
 725                                 }
 726                                 else
 727                                         ereport(ERROR,
 728                                                         (errcode(ERRCODE_INTERNAL_ERROR),       /* XXX: wrong errcode? */
 729                                                          errmsg("lost connection to parallel worker")));
 730
 731                                 /* This might make the error queue go away. */
 732                                 CHECK_FOR_INTERRUPTS();
 733                         }
 734                 }
 735         }
 736 }
 737
 738 /*
 739  * Handle a single protocol message received from a single parallel worker.
 740  */
 741 static void
 742 HandleParallelMessage(ParallelContext *pcxt, int i, StringInfo msg)
 743 {
 744         char            msgtype;
 745
 746         msgtype = pq_getmsgbyte(msg);
 747
 748         switch (msgtype)
 749         {
 750                 case 'K':                               /* BackendKeyData */
 751                         {
 752                                 int32           pid = pq_getmsgint(msg, 4);
 753
 754                                 (void) pq_getmsgint(msg, 4);    /* discard cancel key */
 755                                 (void) pq_getmsgend(msg);
 756                                 pcxt->worker[i].pid = pid;
 757                                 break;
 758                         }
 759
 760                 case 'E':                               /* ErrorResponse */
 761                 case 'N':                               /* NoticeResponse */
 762                         {
 763                                 ErrorData       edata;
 764                                 ErrorContextCallback errctx;
 765                                 ErrorContextCallback *save_error_context_stack;
 766
 767                                 /*
 768                                  * Rethrow the error using the error context callbacks that
 769                                  * were in effect when the context was created, not the
 770                                  * current ones.
 771                                  */
 772                                 save_error_context_stack = error_context_stack;
 773                                 errctx.callback = ParallelErrorContext;
 774                                 errctx.arg = &pcxt->worker[i].pid;
 775                                 errctx.previous = pcxt->error_context_stack;
 776                                 error_context_stack = &errctx;
 777
 778                                 /* Parse ErrorReponse or NoticeResponse. */
 779                                 pq_parse_errornotice(msg, &edata);
 780
 781                                 /* Death of a worker isn't enough justification for suicide. */
 782                                 edata.elevel = Min(edata.elevel, ERROR);
 783
 784                                 /* Rethrow error or notice. */
 785                                 ThrowErrorData(&edata);
 786
 787                                 /* Restore previous context. */
 788                                 error_context_stack = save_error_context_stack;
 789
 790                                 break;
 791                         }
 792
 793                 case 'A':                               /* NotifyResponse */
 794                         {
 795                                 /* Propagate NotifyResponse. */
 796                                 pq_putmessage(msg->data[0], &msg->data[1], msg->len - 1);
 797                                 break;
 798                         }
 799
 800                 case 'X':                               /* Terminate, indicating clean exit */
 801                         {
 802                                 pfree(pcxt->worker[i].bgwhandle);
 803                                 pfree(pcxt->worker[i].error_mqh);
 804                                 pcxt->worker[i].bgwhandle = NULL;
 805                                 pcxt->worker[i].error_mqh = NULL;
 806                                 break;
 807                         }
 808
 809                 default:
 810                         {
 811                                 elog(ERROR, "unknown message type: %c (%d bytes)",
 812                                          msgtype, msg->len);
 813                         }
 814         }
 815 }
 816
 817 /*
 818  * End-of-subtransaction cleanup for parallel contexts.
 819  *
 820  * Currently, it's forbidden to enter or leave a subtransaction while
 821  * parallel mode is in effect, so we could just blow away everything.  But
 822  * we may want to relax that restriction in the future, so this code
 823  * contemplates that there may be multiple subtransaction IDs in pcxt_list.
 824  */
 825 void
 826 AtEOSubXact_Parallel(bool isCommit, SubTransactionId mySubId)
 827 {
 828         while (!dlist_is_empty(&pcxt_list))
 829         {
 830                 ParallelContext *pcxt;
 831
 832                 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
 833                 if (pcxt->subid != mySubId)
 834                         break;
 835                 if (isCommit)
 836                         elog(WARNING, "leaked parallel context");
 837                 DestroyParallelContext(pcxt);
 838         }
 839 }
 840
 841 /*
 842  * End-of-transaction cleanup for parallel contexts.
 843  */
 844 void
 845 AtEOXact_Parallel(bool isCommit)
 846 {
 847         while (!dlist_is_empty(&pcxt_list))
 848         {
 849                 ParallelContext *pcxt;
 850
 851                 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
 852                 if (isCommit)
 853                         elog(WARNING, "leaked parallel context");
 854                 DestroyParallelContext(pcxt);
 855         }
 856 }
 857
 858 /*
 859  * Main entrypoint for parallel workers.
 860  */
 861 static void
 862 ParallelWorkerMain(Datum main_arg)
 863 {
 864         dsm_segment *seg;
 865         shm_toc    *toc;
 866         FixedParallelState *fps;
 867         char       *error_queue_space;
 868         shm_mq     *mq;
 869         shm_mq_handle *mqh;
 870         char       *libraryspace;
 871         char       *gucspace;
 872         char       *combocidspace;
 873         char       *tsnapspace;
 874         char       *asnapspace;
 875         char       *tstatespace;
 876         StringInfoData msgbuf;
 877
 878         /* Set flag to indicate that we're initializing a parallel worker. */
 879         InitializingParallelWorker = true;
 880
 881         /* Establish signal handlers. */
 882         pqsignal(SIGTERM, die);
 883         BackgroundWorkerUnblockSignals();
 884
 885         /* Set up a memory context and resource owner. */
 886         Assert(CurrentResourceOwner == NULL);
 887         CurrentResourceOwner = ResourceOwnerCreate(NULL, "parallel toplevel");
 888         CurrentMemoryContext = AllocSetContextCreate(TopMemoryContext,
 889                                                                                                  "parallel worker",
 890                                                                                                  ALLOCSET_DEFAULT_MINSIZE,
 891                                                                                                  ALLOCSET_DEFAULT_INITSIZE,
 892                                                                                                  ALLOCSET_DEFAULT_MAXSIZE);
 893
 894         /*
 895          * Now that we have a resource owner, we can attach to the dynamic shared
 896          * memory segment and read the table of contents.
 897          */
 898         seg = dsm_attach(DatumGetUInt32(main_arg));
 899         if (seg == NULL)
 900                 ereport(ERROR,
 901                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 902                                  errmsg("unable to map dynamic shared memory segment")));
 903         toc = shm_toc_attach(PARALLEL_MAGIC, dsm_segment_address(seg));
 904         if (toc == NULL)
 905                 ereport(ERROR,
 906                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 907                            errmsg("bad magic number in dynamic shared memory segment")));
 908
 909         /* Determine and set our worker number. */
 910         fps = shm_toc_lookup(toc, PARALLEL_KEY_FIXED);
 911         Assert(fps != NULL);
 912         Assert(ParallelWorkerNumber == -1);
 913         SpinLockAcquire(&fps->mutex);
 914         if (fps->workers_attached < fps->workers_expected)
 915                 ParallelWorkerNumber = fps->workers_attached++;
 916         SpinLockRelease(&fps->mutex);
 917         if (ParallelWorkerNumber < 0)
 918                 ereport(ERROR,
 919                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 920                                  errmsg("too many parallel workers already attached")));
 921         MyFixedParallelState = fps;
 922
 923         /*
 924          * Now that we have a worker number, we can find and attach to the error
 925          * queue provided for us.  That's good, because until we do that, any
 926          * errors that happen here will not be reported back to the process that
 927          * requested that this worker be launched.
 928          */
 929         error_queue_space = shm_toc_lookup(toc, PARALLEL_KEY_ERROR_QUEUE);
 930         mq = (shm_mq *) (error_queue_space +
 931                                          ParallelWorkerNumber * PARALLEL_ERROR_QUEUE_SIZE);
 932         shm_mq_set_sender(mq, MyProc);
 933         mqh = shm_mq_attach(mq, seg, NULL);
 934         pq_redirect_to_shm_mq(seg, mqh);
 935         pq_set_parallel_master(fps->parallel_master_pid,
 936                                                    fps->parallel_master_backend_id);
 937
 938         /*
 939          * Send a BackendKeyData message to the process that initiated parallelism
 940          * so that it has access to our PID before it receives any other messages
 941          * from us.  Our cancel key is sent, too, since that's the way the
 942          * protocol message is defined, but it won't actually be used for anything
 943          * in this case.
 944          */
 945         pq_beginmessage(&msgbuf, 'K');
 946         pq_sendint(&msgbuf, (int32) MyProcPid, sizeof(int32));
 947         pq_sendint(&msgbuf, (int32) MyCancelKey, sizeof(int32));
 948         pq_endmessage(&msgbuf);
 949
 950         /*
 951          * Hooray! Primary initialization is complete.  Now, we need to set up our
 952          * backend-local state to match the original backend.
 953          */
 954
 955         /*
 956          * Load libraries that were loaded by original backend.  We want to do
 957          * this before restoring GUCs, because the libraries might define custom
 958          * variables.
 959          */
 960         libraryspace = shm_toc_lookup(toc, PARALLEL_KEY_LIBRARY);
 961         Assert(libraryspace != NULL);
 962         RestoreLibraryState(libraryspace);
 963
 964         /* Restore database connection. */
 965         BackgroundWorkerInitializeConnectionByOid(fps->database_id,
 966                                                                                           fps->authenticated_user_id);
 967
 968         /* Restore GUC values from launching backend. */
 969         gucspace = shm_toc_lookup(toc, PARALLEL_KEY_GUC);
 970         Assert(gucspace != NULL);
 971         StartTransactionCommand();
 972         RestoreGUCState(gucspace);
 973         CommitTransactionCommand();
 974
 975         /* Crank up a transaction state appropriate to a parallel worker. */
 976         tstatespace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_STATE);
 977         StartParallelWorkerTransaction(tstatespace);
 978
 979         /* Restore combo CID state. */
 980         combocidspace = shm_toc_lookup(toc, PARALLEL_KEY_COMBO_CID);
 981         Assert(combocidspace != NULL);
 982         RestoreComboCIDState(combocidspace);
 983
 984         /* Restore transaction snapshot. */
 985         tsnapspace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT);
 986         Assert(tsnapspace != NULL);
 987         RestoreTransactionSnapshot(RestoreSnapshot(tsnapspace),
 988                                                            fps->parallel_master_pgproc);
 989
 990         /* Restore active snapshot. */
 991         asnapspace = shm_toc_lookup(toc, PARALLEL_KEY_ACTIVE_SNAPSHOT);
 992         Assert(asnapspace != NULL);
 993         PushActiveSnapshot(RestoreSnapshot(asnapspace));
 994
 995         /*
 996          * We've changed which tuples we can see, and must therefore invalidate
 997          * system caches.
 998          */
 999         InvalidateSystemCaches();
1000
1001         /* Restore user ID and security context. */
1002         SetUserIdAndSecContext(fps->current_user_id, fps->sec_context);
1003
1004         /*
1005          * We've initialized all of our state now; nothing should change
1006          * hereafter.
1007          */
1008         InitializingParallelWorker = false;
1009         EnterParallelMode();
1010
1011         /*
1012          * Time to do the real work: invoke the caller-supplied code.
1013          *
1014          * If you get a crash at this line, see the comments for
1015          * ParallelExtensionTrampoline.
1016          */
1017         fps->entrypoint(seg, toc);
1018
1019         /* Must exit parallel mode to pop active snapshot. */
1020         ExitParallelMode();
1021
1022         /* Must pop active snapshot so resowner.c doesn't complain. */
1023         PopActiveSnapshot();
1024
1025         /* Shut down the parallel-worker transaction. */
1026         EndParallelWorkerTransaction();
1027
1028         /* Report success. */
1029         pq_putmessage('X', NULL, 0);
1030 }
1031
1032 /*
1033  * It's unsafe for the entrypoint invoked by ParallelWorkerMain to be a
1034  * function living in a dynamically loaded module, because the module might
1035  * not be loaded in every process, or might be loaded but not at the same
1036  * address.  To work around that problem, CreateParallelContextForExtension()
1037  * arranges to call this function rather than calling the extension-provided
1038  * function directly; and this function then looks up the real entrypoint and
1039  * calls it.
1040  */
1041 static void
1042 ParallelExtensionTrampoline(dsm_segment *seg, shm_toc *toc)
1043 {
1044         char       *extensionstate;
1045         char       *library_name;
1046         char       *function_name;
1047         parallel_worker_main_type entrypt;
1048
1049         extensionstate = shm_toc_lookup(toc, PARALLEL_KEY_EXTENSION_TRAMPOLINE);
1050         Assert(extensionstate != NULL);
1051         library_name = extensionstate;
1052         function_name = extensionstate + strlen(library_name) + 1;
1053
1054         entrypt = (parallel_worker_main_type)
1055                 load_external_function(library_name, function_name, true, NULL);
1056         entrypt(seg, toc);
1057 }
1058
1059 /*
1060  * Give the user a hint that this is a message propagated from a parallel
1061  * worker.  Otherwise, it can sometimes be confusing to understand what
1062  * actually happened.
1063  */
1064 static void
1065 ParallelErrorContext(void *arg)
1066 {
1067         errcontext("parallel worker, pid %d", *(int32 *) arg);
1068 }
1069
1070 /*
1071  * Update shared memory with the ending location of the last WAL record we
1072  * wrote, if it's greater than the value already stored there.
1073  */
1074 void
1075 ParallelWorkerReportLastRecEnd(XLogRecPtr last_xlog_end)
1076 {
1077         FixedParallelState *fps = MyFixedParallelState;
1078
1079         Assert(fps != NULL);
1080         SpinLockAcquire(&fps->mutex);
1081         if (fps->last_xlog_end < last_xlog_end)
1082                 fps->last_xlog_end = last_xlog_end;
1083         SpinLockRelease(&fps->mutex);
1084 }