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

   1 /*-------------------------------------------------------------------------
   2  *
   3  * parallel.c
   4  *        Infrastructure for launching parallel workers
   5  *
   6  * Portions Copyright (c) 1996-2017, PostgreSQL Global Development Group
   7  * Portions Copyright (c) 1994, Regents of the University of California
   8  *
   9  * IDENTIFICATION
  10  *        src/backend/access/transam/parallel.c
  11  *
  12  *-------------------------------------------------------------------------
  13  */
  14
  15 #include "postgres.h"
  16
  17 #include "access/parallel.h"
  18 #include "access/xact.h"
  19 #include "access/xlog.h"
  20 #include "catalog/namespace.h"
  21 #include "commands/async.h"
  22 #include "libpq/libpq.h"
  23 #include "libpq/pqformat.h"
  24 #include "libpq/pqmq.h"
  25 #include "miscadmin.h"
  26 #include "optimizer/planmain.h"
  27 #include "pgstat.h"
  28 #include "storage/ipc.h"
  29 #include "storage/sinval.h"
  30 #include "storage/spin.h"
  31 #include "tcop/tcopprot.h"
  32 #include "utils/combocid.h"
  33 #include "utils/guc.h"
  34 #include "utils/inval.h"
  35 #include "utils/memutils.h"
  36 #include "utils/resowner.h"
  37 #include "utils/snapmgr.h"
  38
  39
  40 /*
  41  * We don't want to waste a lot of memory on an error queue which, most of
  42  * the time, will process only a handful of small messages.  However, it is
  43  * desirable to make it large enough that a typical ErrorResponse can be sent
  44  * without blocking.  That way, a worker that errors out can write the whole
  45  * message into the queue and terminate without waiting for the user backend.
  46  */
  47 #define PARALLEL_ERROR_QUEUE_SIZE                       16384
  48
  49 /* Magic number for parallel context TOC. */
  50 #define PARALLEL_MAGIC                                          0x50477c7c
  51
  52 /*
  53  * Magic numbers for parallel state sharing.  Higher-level code should use
  54  * smaller values, leaving these very large ones for use by this module.
  55  */
  56 #define PARALLEL_KEY_FIXED                                      UINT64CONST(0xFFFFFFFFFFFF0001)
  57 #define PARALLEL_KEY_ERROR_QUEUE                        UINT64CONST(0xFFFFFFFFFFFF0002)
  58 #define PARALLEL_KEY_LIBRARY                            UINT64CONST(0xFFFFFFFFFFFF0003)
  59 #define PARALLEL_KEY_GUC                                        UINT64CONST(0xFFFFFFFFFFFF0004)
  60 #define PARALLEL_KEY_COMBO_CID                          UINT64CONST(0xFFFFFFFFFFFF0005)
  61 #define PARALLEL_KEY_TRANSACTION_SNAPSHOT       UINT64CONST(0xFFFFFFFFFFFF0006)
  62 #define PARALLEL_KEY_ACTIVE_SNAPSHOT            UINT64CONST(0xFFFFFFFFFFFF0007)
  63 #define PARALLEL_KEY_TRANSACTION_STATE          UINT64CONST(0xFFFFFFFFFFFF0008)
  64 #define PARALLEL_KEY_EXTENSION_TRAMPOLINE       UINT64CONST(0xFFFFFFFFFFFF0009)
  65
  66 /* Fixed-size parallel state. */
  67 typedef struct FixedParallelState
  68 {
  69         /* Fixed-size state that workers must restore. */
  70         Oid                     database_id;
  71         Oid                     authenticated_user_id;
  72         Oid                     current_user_id;
  73         Oid                     temp_namespace_id;
  74         Oid                     temp_toast_namespace_id;
  75         int                     sec_context;
  76         PGPROC     *parallel_master_pgproc;
  77         pid_t           parallel_master_pid;
  78         BackendId       parallel_master_backend_id;
  79
  80         /* Entrypoint for parallel workers. */
  81         parallel_worker_main_type entrypoint;
  82
  83         /* Mutex protects remaining fields. */
  84         slock_t         mutex;
  85
  86         /* Maximum XactLastRecEnd of any worker. */
  87         XLogRecPtr      last_xlog_end;
  88 } FixedParallelState;
  89
  90 /*
  91  * Our parallel worker number.  We initialize this to -1, meaning that we are
  92  * not a parallel worker.  In parallel workers, it will be set to a value >= 0
  93  * and < the number of workers before any user code is invoked; each parallel
  94  * worker will get a different parallel worker number.
  95  */
  96 int                     ParallelWorkerNumber = -1;
  97
  98 /* Is there a parallel message pending which we need to receive? */
  99 volatile bool ParallelMessagePending = false;
 100
 101 /* Are we initializing a parallel worker? */
 102 bool            InitializingParallelWorker = false;
 103
 104 /* Pointer to our fixed parallel state. */
 105 static FixedParallelState *MyFixedParallelState;
 106
 107 /* List of active parallel contexts. */
 108 static dlist_head pcxt_list = DLIST_STATIC_INIT(pcxt_list);
 109
 110 /* Private functions. */
 111 static void HandleParallelMessage(ParallelContext *pcxt, int i, StringInfo msg);
 112 static void ParallelExtensionTrampoline(dsm_segment *seg, shm_toc *toc);
 113 static void ParallelWorkerMain(Datum main_arg);
 114 static void WaitForParallelWorkersToExit(ParallelContext *pcxt);
 115
 116
 117 /*
 118  * Establish a new parallel context.  This should be done after entering
 119  * parallel mode, and (unless there is an error) the context should be
 120  * destroyed before exiting the current subtransaction.
 121  */
 122 ParallelContext *
 123 CreateParallelContext(parallel_worker_main_type entrypoint, int nworkers)
 124 {
 125         MemoryContext oldcontext;
 126         ParallelContext *pcxt;
 127
 128         /* It is unsafe to create a parallel context if not in parallel mode. */
 129         Assert(IsInParallelMode());
 130
 131         /* Number of workers should be non-negative. */
 132         Assert(nworkers >= 0);
 133
 134         /*
 135          * If dynamic shared memory is not available, we won't be able to use
 136          * background workers.
 137          */
 138         if (dynamic_shared_memory_type == DSM_IMPL_NONE)
 139                 nworkers = 0;
 140
 141         /*
 142          * If we are running under serializable isolation, we can't use parallel
 143          * workers, at least not until somebody enhances that mechanism to be
 144          * parallel-aware.
 145          */
 146         if (IsolationIsSerializable())
 147                 nworkers = 0;
 148
 149         /* We might be running in a short-lived memory context. */
 150         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 151
 152         /* Initialize a new ParallelContext. */
 153         pcxt = palloc0(sizeof(ParallelContext));
 154         pcxt->subid = GetCurrentSubTransactionId();
 155         pcxt->nworkers = nworkers;
 156         pcxt->entrypoint = entrypoint;
 157         pcxt->error_context_stack = error_context_stack;
 158         shm_toc_initialize_estimator(&pcxt->estimator);
 159         dlist_push_head(&pcxt_list, &pcxt->node);
 160
 161         /* Restore previous memory context. */
 162         MemoryContextSwitchTo(oldcontext);
 163
 164         return pcxt;
 165 }
 166
 167 /*
 168  * Establish a new parallel context that calls a function provided by an
 169  * extension.  This works around the fact that the library might get mapped
 170  * at a different address in each backend.
 171  */
 172 ParallelContext *
 173 CreateParallelContextForExternalFunction(char *library_name,
 174                                                                                  char *function_name,
 175                                                                                  int nworkers)
 176 {
 177         MemoryContext oldcontext;
 178         ParallelContext *pcxt;
 179
 180         /* We might be running in a very short-lived memory context. */
 181         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 182
 183         /* Create the context. */
 184         pcxt = CreateParallelContext(ParallelExtensionTrampoline, nworkers);
 185         pcxt->library_name = pstrdup(library_name);
 186         pcxt->function_name = pstrdup(function_name);
 187
 188         /* Restore previous memory context. */
 189         MemoryContextSwitchTo(oldcontext);
 190
 191         return pcxt;
 192 }
 193
 194 /*
 195  * Establish the dynamic shared memory segment for a parallel context and
 196  * copy state and other bookkeeping information that will be needed by
 197  * parallel workers into it.
 198  */
 199 void
 200 InitializeParallelDSM(ParallelContext *pcxt)
 201 {
 202         MemoryContext oldcontext;
 203         Size            library_len = 0;
 204         Size            guc_len = 0;
 205         Size            combocidlen = 0;
 206         Size            tsnaplen = 0;
 207         Size            asnaplen = 0;
 208         Size            tstatelen = 0;
 209         Size            segsize = 0;
 210         int                     i;
 211         FixedParallelState *fps;
 212         Snapshot        transaction_snapshot = GetTransactionSnapshot();
 213         Snapshot        active_snapshot = GetActiveSnapshot();
 214
 215         /* We might be running in a very short-lived memory context. */
 216         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 217
 218         /* Allow space to store the fixed-size parallel state. */
 219         shm_toc_estimate_chunk(&pcxt->estimator, sizeof(FixedParallelState));
 220         shm_toc_estimate_keys(&pcxt->estimator, 1);
 221
 222         /*
 223          * Normally, the user will have requested at least one worker process, but
 224          * if by chance they have not, we can skip a bunch of things here.
 225          */
 226         if (pcxt->nworkers > 0)
 227         {
 228                 /* Estimate space for various kinds of state sharing. */
 229                 library_len = EstimateLibraryStateSpace();
 230                 shm_toc_estimate_chunk(&pcxt->estimator, library_len);
 231                 guc_len = EstimateGUCStateSpace();
 232                 shm_toc_estimate_chunk(&pcxt->estimator, guc_len);
 233                 combocidlen = EstimateComboCIDStateSpace();
 234                 shm_toc_estimate_chunk(&pcxt->estimator, combocidlen);
 235                 tsnaplen = EstimateSnapshotSpace(transaction_snapshot);
 236                 shm_toc_estimate_chunk(&pcxt->estimator, tsnaplen);
 237                 asnaplen = EstimateSnapshotSpace(active_snapshot);
 238                 shm_toc_estimate_chunk(&pcxt->estimator, asnaplen);
 239                 tstatelen = EstimateTransactionStateSpace();
 240                 shm_toc_estimate_chunk(&pcxt->estimator, tstatelen);
 241                 /* If you add more chunks here, you probably need to add keys. */
 242                 shm_toc_estimate_keys(&pcxt->estimator, 6);
 243
 244                 /* Estimate space need for error queues. */
 245                 StaticAssertStmt(BUFFERALIGN(PARALLEL_ERROR_QUEUE_SIZE) ==
 246                                                  PARALLEL_ERROR_QUEUE_SIZE,
 247                                                  "parallel error queue size not buffer-aligned");
 248                 shm_toc_estimate_chunk(&pcxt->estimator,
 249                                                            mul_size(PARALLEL_ERROR_QUEUE_SIZE,
 250                                                                                 pcxt->nworkers));
 251                 shm_toc_estimate_keys(&pcxt->estimator, 1);
 252
 253                 /* Estimate how much we'll need for extension entrypoint info. */
 254                 if (pcxt->library_name != NULL)
 255                 {
 256                         Assert(pcxt->entrypoint == ParallelExtensionTrampoline);
 257                         Assert(pcxt->function_name != NULL);
 258                         shm_toc_estimate_chunk(&pcxt->estimator, strlen(pcxt->library_name)
 259                                                                    + strlen(pcxt->function_name) + 2);
 260                         shm_toc_estimate_keys(&pcxt->estimator, 1);
 261                 }
 262         }
 263
 264         /*
 265          * Create DSM and initialize with new table of contents.  But if the user
 266          * didn't request any workers, then don't bother creating a dynamic shared
 267          * memory segment; instead, just use backend-private memory.
 268          *
 269          * Also, if we can't create a dynamic shared memory segment because the
 270          * maximum number of segments have already been created, then fall back to
 271          * backend-private memory, and plan not to use any workers.  We hope this
 272          * won't happen very often, but it's better to abandon the use of
 273          * parallelism than to fail outright.
 274          */
 275         segsize = shm_toc_estimate(&pcxt->estimator);
 276         if (pcxt->nworkers > 0)
 277                 pcxt->seg = dsm_create(segsize, DSM_CREATE_NULL_IF_MAXSEGMENTS);
 278         if (pcxt->seg != NULL)
 279                 pcxt->toc = shm_toc_create(PARALLEL_MAGIC,
 280                                                                    dsm_segment_address(pcxt->seg),
 281                                                                    segsize);
 282         else
 283         {
 284                 pcxt->nworkers = 0;
 285                 pcxt->private_memory = MemoryContextAlloc(TopMemoryContext, segsize);
 286                 pcxt->toc = shm_toc_create(PARALLEL_MAGIC, pcxt->private_memory,
 287                                                                    segsize);
 288         }
 289
 290         /* Initialize fixed-size state in shared memory. */
 291         fps = (FixedParallelState *)
 292                 shm_toc_allocate(pcxt->toc, sizeof(FixedParallelState));
 293         fps->database_id = MyDatabaseId;
 294         fps->authenticated_user_id = GetAuthenticatedUserId();
 295         GetUserIdAndSecContext(&fps->current_user_id, &fps->sec_context);
 296         GetTempNamespaceState(&fps->temp_namespace_id,
 297                                                   &fps->temp_toast_namespace_id);
 298         fps->parallel_master_pgproc = MyProc;
 299         fps->parallel_master_pid = MyProcPid;
 300         fps->parallel_master_backend_id = MyBackendId;
 301         fps->entrypoint = pcxt->entrypoint;
 302         SpinLockInit(&fps->mutex);
 303         fps->last_xlog_end = 0;
 304         shm_toc_insert(pcxt->toc, PARALLEL_KEY_FIXED, fps);
 305
 306         /* We can skip the rest of this if we're not budgeting for any workers. */
 307         if (pcxt->nworkers > 0)
 308         {
 309                 char       *libraryspace;
 310                 char       *gucspace;
 311                 char       *combocidspace;
 312                 char       *tsnapspace;
 313                 char       *asnapspace;
 314                 char       *tstatespace;
 315                 char       *error_queue_space;
 316
 317                 /* Serialize shared libraries we have loaded. */
 318                 libraryspace = shm_toc_allocate(pcxt->toc, library_len);
 319                 SerializeLibraryState(library_len, libraryspace);
 320                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_LIBRARY, libraryspace);
 321
 322                 /* Serialize GUC settings. */
 323                 gucspace = shm_toc_allocate(pcxt->toc, guc_len);
 324                 SerializeGUCState(guc_len, gucspace);
 325                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_GUC, gucspace);
 326
 327                 /* Serialize combo CID state. */
 328                 combocidspace = shm_toc_allocate(pcxt->toc, combocidlen);
 329                 SerializeComboCIDState(combocidlen, combocidspace);
 330                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_COMBO_CID, combocidspace);
 331
 332                 /* Serialize transaction snapshot and active snapshot. */
 333                 tsnapspace = shm_toc_allocate(pcxt->toc, tsnaplen);
 334                 SerializeSnapshot(transaction_snapshot, tsnapspace);
 335                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT,
 336                                            tsnapspace);
 337                 asnapspace = shm_toc_allocate(pcxt->toc, asnaplen);
 338                 SerializeSnapshot(active_snapshot, asnapspace);
 339                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ACTIVE_SNAPSHOT, asnapspace);
 340
 341                 /* Serialize transaction state. */
 342                 tstatespace = shm_toc_allocate(pcxt->toc, tstatelen);
 343                 SerializeTransactionState(tstatelen, tstatespace);
 344                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_STATE, tstatespace);
 345
 346                 /* Allocate space for worker information. */
 347                 pcxt->worker = palloc0(sizeof(ParallelWorkerInfo) * pcxt->nworkers);
 348
 349                 /*
 350                  * Establish error queues in dynamic shared memory.
 351                  *
 352                  * These queues should be used only for transmitting ErrorResponse,
 353                  * NoticeResponse, and NotifyResponse protocol messages.  Tuple data
 354                  * should be transmitted via separate (possibly larger?) queues.
 355                  */
 356                 error_queue_space =
 357                         shm_toc_allocate(pcxt->toc,
 358                                                          mul_size(PARALLEL_ERROR_QUEUE_SIZE,
 359                                                                           pcxt->nworkers));
 360                 for (i = 0; i < pcxt->nworkers; ++i)
 361                 {
 362                         char       *start;
 363                         shm_mq     *mq;
 364
 365                         start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
 366                         mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
 367                         shm_mq_set_receiver(mq, MyProc);
 368                         pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
 369                 }
 370                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE, error_queue_space);
 371
 372                 /* Serialize extension entrypoint information. */
 373                 if (pcxt->library_name != NULL)
 374                 {
 375                         Size            lnamelen = strlen(pcxt->library_name);
 376                         char       *extensionstate;
 377
 378                         extensionstate = shm_toc_allocate(pcxt->toc, lnamelen
 379                                                                                   + strlen(pcxt->function_name) + 2);
 380                         strcpy(extensionstate, pcxt->library_name);
 381                         strcpy(extensionstate + lnamelen + 1, pcxt->function_name);
 382                         shm_toc_insert(pcxt->toc, PARALLEL_KEY_EXTENSION_TRAMPOLINE,
 383                                                    extensionstate);
 384                 }
 385         }
 386
 387         /* Restore previous memory context. */
 388         MemoryContextSwitchTo(oldcontext);
 389 }
 390
 391 /*
 392  * Reinitialize the dynamic shared memory segment for a parallel context such
 393  * that we could launch workers for it again.
 394  */
 395 void
 396 ReinitializeParallelDSM(ParallelContext *pcxt)
 397 {
 398         FixedParallelState *fps;
 399         char       *error_queue_space;
 400         int                     i;
 401
 402         /* Wait for any old workers to exit. */
 403         if (pcxt->nworkers_launched > 0)
 404         {
 405                 WaitForParallelWorkersToFinish(pcxt);
 406                 WaitForParallelWorkersToExit(pcxt);
 407                 pcxt->nworkers_launched = 0;
 408         }
 409
 410         /* Reset a few bits of fixed parallel state to a clean state. */
 411         fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED);
 412         fps->last_xlog_end = 0;
 413
 414         /* Recreate error queues. */
 415         error_queue_space =
 416                 shm_toc_lookup(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE);
 417         for (i = 0; i < pcxt->nworkers; ++i)
 418         {
 419                 char       *start;
 420                 shm_mq     *mq;
 421
 422                 start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
 423                 mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
 424                 shm_mq_set_receiver(mq, MyProc);
 425                 pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
 426         }
 427 }
 428
 429 /*
 430  * Launch parallel workers.
 431  */
 432 void
 433 LaunchParallelWorkers(ParallelContext *pcxt)
 434 {
 435         MemoryContext oldcontext;
 436         BackgroundWorker worker;
 437         int                     i;
 438         bool            any_registrations_failed = false;
 439
 440         /* Skip this if we have no workers. */
 441         if (pcxt->nworkers == 0)
 442                 return;
 443
 444         /* We need to be a lock group leader. */
 445         BecomeLockGroupLeader();
 446
 447         /* If we do have workers, we'd better have a DSM segment. */
 448         Assert(pcxt->seg != NULL);
 449
 450         /* We might be running in a short-lived memory context. */
 451         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 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                 | BGWORKER_CLASS_PARALLEL;
 459         worker.bgw_start_time = BgWorkerStart_ConsistentState;
 460         worker.bgw_restart_time = BGW_NEVER_RESTART;
 461         worker.bgw_main = ParallelWorkerMain;
 462         worker.bgw_main_arg = UInt32GetDatum(dsm_segment_handle(pcxt->seg));
 463         worker.bgw_notify_pid = MyProcPid;
 464         memset(&worker.bgw_extra, 0, BGW_EXTRALEN);
 465
 466         /*
 467          * Start workers.
 468          *
 469          * The caller must be able to tolerate ending up with fewer workers than
 470          * expected, so there is no need to throw an error here if registration
 471          * fails.  It wouldn't help much anyway, because registering the worker in
 472          * no way guarantees that it will start up and initialize successfully.
 473          */
 474         for (i = 0; i < pcxt->nworkers; ++i)
 475         {
 476                 memcpy(worker.bgw_extra, &i, sizeof(int));
 477                 if (!any_registrations_failed &&
 478                         RegisterDynamicBackgroundWorker(&worker,
 479                                                                                         &pcxt->worker[i].bgwhandle))
 480                 {
 481                         shm_mq_set_handle(pcxt->worker[i].error_mqh,
 482                                                           pcxt->worker[i].bgwhandle);
 483                         pcxt->nworkers_launched++;
 484                 }
 485                 else
 486                 {
 487                         /*
 488                          * If we weren't able to register the worker, then we've bumped up
 489                          * against the max_worker_processes limit, and future
 490                          * registrations will probably fail too, so arrange to skip them.
 491                          * But we still have to execute this code for the remaining slots
 492                          * to make sure that we forget about the error queues we budgeted
 493                          * for those workers.  Otherwise, we'll wait for them to start,
 494                          * but they never will.
 495                          */
 496                         any_registrations_failed = true;
 497                         pcxt->worker[i].bgwhandle = NULL;
 498                         pfree(pcxt->worker[i].error_mqh);
 499                         pcxt->worker[i].error_mqh = NULL;
 500                 }
 501         }
 502
 503         /* Restore previous memory context. */
 504         MemoryContextSwitchTo(oldcontext);
 505 }
 506
 507 /*
 508  * Wait for all workers to finish computing.
 509  *
 510  * Even if the parallel operation seems to have completed successfully, it's
 511  * important to call this function afterwards.  We must not miss any errors
 512  * the workers may have thrown during the parallel operation, or any that they
 513  * may yet throw while shutting down.
 514  *
 515  * Also, we want to update our notion of XactLastRecEnd based on worker
 516  * feedback.
 517  */
 518 void
 519 WaitForParallelWorkersToFinish(ParallelContext *pcxt)
 520 {
 521         for (;;)
 522         {
 523                 bool            anyone_alive = false;
 524                 int                     i;
 525
 526                 /*
 527                  * This will process any parallel messages that are pending, which may
 528                  * change the outcome of the loop that follows.  It may also throw an
 529                  * error propagated from a worker.
 530                  */
 531                 CHECK_FOR_INTERRUPTS();
 532
 533                 for (i = 0; i < pcxt->nworkers_launched; ++i)
 534                 {
 535                         if (pcxt->worker[i].error_mqh != NULL)
 536                         {
 537                                 anyone_alive = true;
 538                                 break;
 539                         }
 540                 }
 541
 542                 if (!anyone_alive)
 543                         break;
 544
 545                 WaitLatch(&MyProc->procLatch, WL_LATCH_SET, -1,
 546                                   WAIT_EVENT_PARALLEL_FINISH);
 547                 ResetLatch(&MyProc->procLatch);
 548         }
 549
 550         if (pcxt->toc != NULL)
 551         {
 552                 FixedParallelState *fps;
 553
 554                 fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED);
 555                 if (fps->last_xlog_end > XactLastRecEnd)
 556                         XactLastRecEnd = fps->last_xlog_end;
 557         }
 558 }
 559
 560 /*
 561  * Wait for all workers to exit.
 562  *
 563  * This function ensures that workers have been completely shutdown.  The
 564  * difference between WaitForParallelWorkersToFinish and this function is
 565  * that former just ensures that last message sent by worker backend is
 566  * received by master backend whereas this ensures the complete shutdown.
 567  */
 568 static void
 569 WaitForParallelWorkersToExit(ParallelContext *pcxt)
 570 {
 571         int                     i;
 572
 573         /* Wait until the workers actually die. */
 574         for (i = 0; i < pcxt->nworkers_launched; ++i)
 575         {
 576                 BgwHandleStatus status;
 577
 578                 if (pcxt->worker == NULL || pcxt->worker[i].bgwhandle == NULL)
 579                         continue;
 580
 581                 status = WaitForBackgroundWorkerShutdown(pcxt->worker[i].bgwhandle);
 582
 583                 /*
 584                  * If the postmaster kicked the bucket, we have no chance of cleaning
 585                  * up safely -- we won't be able to tell when our workers are actually
 586                  * dead.  This doesn't necessitate a PANIC since they will all abort
 587                  * eventually, but we can't safely continue this session.
 588                  */
 589                 if (status == BGWH_POSTMASTER_DIED)
 590                         ereport(FATAL,
 591                                         (errcode(ERRCODE_ADMIN_SHUTDOWN),
 592                                  errmsg("postmaster exited during a parallel transaction")));
 593
 594                 /* Release memory. */
 595                 pfree(pcxt->worker[i].bgwhandle);
 596                 pcxt->worker[i].bgwhandle = NULL;
 597         }
 598 }
 599
 600 /*
 601  * Destroy a parallel context.
 602  *
 603  * If expecting a clean exit, you should use WaitForParallelWorkersToFinish()
 604  * first, before calling this function.  When this function is invoked, any
 605  * remaining workers are forcibly killed; the dynamic shared memory segment
 606  * is unmapped; and we then wait (uninterruptibly) for the workers to exit.
 607  */
 608 void
 609 DestroyParallelContext(ParallelContext *pcxt)
 610 {
 611         int                     i;
 612
 613         /*
 614          * Be careful about order of operations here!  We remove the parallel
 615          * context from the list before we do anything else; otherwise, if an
 616          * error occurs during a subsequent step, we might try to nuke it again
 617          * from AtEOXact_Parallel or AtEOSubXact_Parallel.
 618          */
 619         dlist_delete(&pcxt->node);
 620
 621         /* Kill each worker in turn, and forget their error queues. */
 622         if (pcxt->worker != NULL)
 623         {
 624                 for (i = 0; i < pcxt->nworkers_launched; ++i)
 625                 {
 626                         if (pcxt->worker[i].error_mqh != NULL)
 627                         {
 628                                 TerminateBackgroundWorker(pcxt->worker[i].bgwhandle);
 629
 630                                 pfree(pcxt->worker[i].error_mqh);
 631                                 pcxt->worker[i].error_mqh = NULL;
 632                         }
 633                 }
 634         }
 635
 636         /*
 637          * If we have allocated a shared memory segment, detach it.  This will
 638          * implicitly detach the error queues, and any other shared memory queues,
 639          * stored there.
 640          */
 641         if (pcxt->seg != NULL)
 642         {
 643                 dsm_detach(pcxt->seg);
 644                 pcxt->seg = NULL;
 645         }
 646
 647         /*
 648          * If this parallel context is actually in backend-private memory rather
 649          * than shared memory, free that memory instead.
 650          */
 651         if (pcxt->private_memory != NULL)
 652         {
 653                 pfree(pcxt->private_memory);
 654                 pcxt->private_memory = NULL;
 655         }
 656
 657         /*
 658          * We can't finish transaction commit or abort until all of the workers
 659          * have exited.  This means, in particular, that we can't respond to
 660          * interrupts at this stage.
 661          */
 662         HOLD_INTERRUPTS();
 663         WaitForParallelWorkersToExit(pcxt);
 664         RESUME_INTERRUPTS();
 665
 666         /* Free the worker array itself. */
 667         if (pcxt->worker != NULL)
 668         {
 669                 pfree(pcxt->worker);
 670                 pcxt->worker = NULL;
 671         }
 672
 673         /* Free memory. */
 674         pfree(pcxt);
 675 }
 676
 677 /*
 678  * Are there any parallel contexts currently active?
 679  */
 680 bool
 681 ParallelContextActive(void)
 682 {
 683         return !dlist_is_empty(&pcxt_list);
 684 }
 685
 686 /*
 687  * Handle receipt of an interrupt indicating a parallel worker message.
 688  *
 689  * Note: this is called within a signal handler!  All we can do is set
 690  * a flag that will cause the next CHECK_FOR_INTERRUPTS() to invoke
 691  * HandleParallelMessages().
 692  */
 693 void
 694 HandleParallelMessageInterrupt(void)
 695 {
 696         InterruptPending = true;
 697         ParallelMessagePending = true;
 698         SetLatch(MyLatch);
 699 }
 700
 701 /*
 702  * Handle any queued protocol messages received from parallel workers.
 703  */
 704 void
 705 HandleParallelMessages(void)
 706 {
 707         dlist_iter      iter;
 708         MemoryContext oldcontext;
 709
 710         static MemoryContext hpm_context = NULL;
 711
 712         /*
 713          * This is invoked from ProcessInterrupts(), and since some of the
 714          * functions it calls contain CHECK_FOR_INTERRUPTS(), there is a potential
 715          * for recursive calls if more signals are received while this runs.  It's
 716          * unclear that recursive entry would be safe, and it doesn't seem useful
 717          * even if it is safe, so let's block interrupts until done.
 718          */
 719         HOLD_INTERRUPTS();
 720
 721         /*
 722          * Moreover, CurrentMemoryContext might be pointing almost anywhere.  We
 723          * don't want to risk leaking data into long-lived contexts, so let's do
 724          * our work here in a private context that we can reset on each use.
 725          */
 726         if (hpm_context == NULL)        /* first time through? */
 727                 hpm_context = AllocSetContextCreate(TopMemoryContext,
 728                                                                                         "HandleParallelMessages",
 729                                                                                         ALLOCSET_DEFAULT_SIZES);
 730         else
 731                 MemoryContextReset(hpm_context);
 732
 733         oldcontext = MemoryContextSwitchTo(hpm_context);
 734
 735         /* OK to process messages.  Reset the flag saying there are more to do. */
 736         ParallelMessagePending = false;
 737
 738         dlist_foreach(iter, &pcxt_list)
 739         {
 740                 ParallelContext *pcxt;
 741                 int                     i;
 742
 743                 pcxt = dlist_container(ParallelContext, node, iter.cur);
 744                 if (pcxt->worker == NULL)
 745                         continue;
 746
 747                 for (i = 0; i < pcxt->nworkers_launched; ++i)
 748                 {
 749                         /*
 750                          * Read as many messages as we can from each worker, but stop when
 751                          * either (1) the worker's error queue goes away, which can happen
 752                          * if we receive a Terminate message from the worker; or (2) no
 753                          * more messages can be read from the worker without blocking.
 754                          */
 755                         while (pcxt->worker[i].error_mqh != NULL)
 756                         {
 757                                 shm_mq_result res;
 758                                 Size            nbytes;
 759                                 void       *data;
 760
 761                                 res = shm_mq_receive(pcxt->worker[i].error_mqh, &nbytes,
 762                                                                          &data, true);
 763                                 if (res == SHM_MQ_WOULD_BLOCK)
 764                                         break;
 765                                 else if (res == SHM_MQ_SUCCESS)
 766                                 {
 767                                         StringInfoData msg;
 768
 769                                         initStringInfo(&msg);
 770                                         appendBinaryStringInfo(&msg, data, nbytes);
 771                                         HandleParallelMessage(pcxt, i, &msg);
 772                                         pfree(msg.data);
 773                                 }
 774                                 else
 775                                         ereport(ERROR,
 776                                                   (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 777                                                    errmsg("lost connection to parallel worker")));
 778                         }
 779                 }
 780         }
 781
 782         MemoryContextSwitchTo(oldcontext);
 783
 784         /* Might as well clear the context on our way out */
 785         MemoryContextReset(hpm_context);
 786
 787         RESUME_INTERRUPTS();
 788 }
 789
 790 /*
 791  * Handle a single protocol message received from a single parallel worker.
 792  */
 793 static void
 794 HandleParallelMessage(ParallelContext *pcxt, int i, StringInfo msg)
 795 {
 796         char            msgtype;
 797
 798         msgtype = pq_getmsgbyte(msg);
 799
 800         switch (msgtype)
 801         {
 802                 case 'K':                               /* BackendKeyData */
 803                         {
 804                                 int32           pid = pq_getmsgint(msg, 4);
 805
 806                                 (void) pq_getmsgint(msg, 4);    /* discard cancel key */
 807                                 (void) pq_getmsgend(msg);
 808                                 pcxt->worker[i].pid = pid;
 809                                 break;
 810                         }
 811
 812                 case 'E':                               /* ErrorResponse */
 813                 case 'N':                               /* NoticeResponse */
 814                         {
 815                                 ErrorData       edata;
 816                                 ErrorContextCallback *save_error_context_stack;
 817
 818                                 /* Parse ErrorResponse or NoticeResponse. */
 819                                 pq_parse_errornotice(msg, &edata);
 820
 821                                 /* Death of a worker isn't enough justification for suicide. */
 822                                 edata.elevel = Min(edata.elevel, ERROR);
 823
 824                                 /*
 825                                  * If desired, add a context line to show that this is a
 826                                  * message propagated from a parallel worker.  Otherwise, it
 827                                  * can sometimes be confusing to understand what actually
 828                                  * happened.  (We don't do this in FORCE_PARALLEL_REGRESS mode
 829                                  * because it causes test-result instability depending on
 830                                  * whether a parallel worker is actually used or not.)
 831                                  */
 832                                 if (force_parallel_mode != FORCE_PARALLEL_REGRESS)
 833                                 {
 834                                         if (edata.context)
 835                                                 edata.context = psprintf("%s\n%s", edata.context,
 836                                                                                                  _("parallel worker"));
 837                                         else
 838                                                 edata.context = pstrdup(_("parallel worker"));
 839                                 }
 840
 841                                 /*
 842                                  * Context beyond that should use the error context callbacks
 843                                  * that were in effect when the ParallelContext was created,
 844                                  * not the current ones.
 845                                  */
 846                                 save_error_context_stack = error_context_stack;
 847                                 error_context_stack = pcxt->error_context_stack;
 848
 849                                 /* Rethrow error or print notice. */
 850                                 ThrowErrorData(&edata);
 851
 852                                 /* Not an error, so restore previous context stack. */
 853                                 error_context_stack = save_error_context_stack;
 854
 855                                 break;
 856                         }
 857
 858                 case 'A':                               /* NotifyResponse */
 859                         {
 860                                 /* Propagate NotifyResponse. */
 861                                 int32           pid;
 862                                 const char *channel;
 863                                 const char *payload;
 864
 865                                 pid = pq_getmsgint(msg, 4);
 866                                 channel = pq_getmsgrawstring(msg);
 867                                 payload = pq_getmsgrawstring(msg);
 868                                 pq_endmessage(msg);
 869
 870                                 NotifyMyFrontEnd(channel, payload, pid);
 871
 872                                 break;
 873                         }
 874
 875                 case 'X':                               /* Terminate, indicating clean exit */
 876                         {
 877                                 pfree(pcxt->worker[i].error_mqh);
 878                                 pcxt->worker[i].error_mqh = NULL;
 879                                 break;
 880                         }
 881
 882                 default:
 883                         {
 884                                 elog(ERROR, "unrecognized message type received from parallel worker: %c (message length %d bytes)",
 885                                          msgtype, msg->len);
 886                         }
 887         }
 888 }
 889
 890 /*
 891  * End-of-subtransaction cleanup for parallel contexts.
 892  *
 893  * Currently, it's forbidden to enter or leave a subtransaction while
 894  * parallel mode is in effect, so we could just blow away everything.  But
 895  * we may want to relax that restriction in the future, so this code
 896  * contemplates that there may be multiple subtransaction IDs in pcxt_list.
 897  */
 898 void
 899 AtEOSubXact_Parallel(bool isCommit, SubTransactionId mySubId)
 900 {
 901         while (!dlist_is_empty(&pcxt_list))
 902         {
 903                 ParallelContext *pcxt;
 904
 905                 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
 906                 if (pcxt->subid != mySubId)
 907                         break;
 908                 if (isCommit)
 909                         elog(WARNING, "leaked parallel context");
 910                 DestroyParallelContext(pcxt);
 911         }
 912 }
 913
 914 /*
 915  * End-of-transaction cleanup for parallel contexts.
 916  */
 917 void
 918 AtEOXact_Parallel(bool isCommit)
 919 {
 920         while (!dlist_is_empty(&pcxt_list))
 921         {
 922                 ParallelContext *pcxt;
 923
 924                 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
 925                 if (isCommit)
 926                         elog(WARNING, "leaked parallel context");
 927                 DestroyParallelContext(pcxt);
 928         }
 929 }
 930
 931 /*
 932  * Main entrypoint for parallel workers.
 933  */
 934 static void
 935 ParallelWorkerMain(Datum main_arg)
 936 {
 937         dsm_segment *seg;
 938         shm_toc    *toc;
 939         FixedParallelState *fps;
 940         char       *error_queue_space;
 941         shm_mq     *mq;
 942         shm_mq_handle *mqh;
 943         char       *libraryspace;
 944         char       *gucspace;
 945         char       *combocidspace;
 946         char       *tsnapspace;
 947         char       *asnapspace;
 948         char       *tstatespace;
 949         StringInfoData msgbuf;
 950
 951         /* Set flag to indicate that we're initializing a parallel worker. */
 952         InitializingParallelWorker = true;
 953
 954         /* Establish signal handlers. */
 955         pqsignal(SIGTERM, die);
 956         BackgroundWorkerUnblockSignals();
 957
 958         /* Determine and set our parallel worker number. */
 959         Assert(ParallelWorkerNumber == -1);
 960         memcpy(&ParallelWorkerNumber, MyBgworkerEntry->bgw_extra, sizeof(int));
 961
 962         /* Set up a memory context and resource owner. */
 963         Assert(CurrentResourceOwner == NULL);
 964         CurrentResourceOwner = ResourceOwnerCreate(NULL, "parallel toplevel");
 965         CurrentMemoryContext = AllocSetContextCreate(TopMemoryContext,
 966                                                                                                  "Parallel worker",
 967                                                                                                  ALLOCSET_DEFAULT_SIZES);
 968
 969         /*
 970          * Now that we have a resource owner, we can attach to the dynamic shared
 971          * memory segment and read the table of contents.
 972          */
 973         seg = dsm_attach(DatumGetUInt32(main_arg));
 974         if (seg == NULL)
 975                 ereport(ERROR,
 976                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 977                                  errmsg("could not map dynamic shared memory segment")));
 978         toc = shm_toc_attach(PARALLEL_MAGIC, dsm_segment_address(seg));
 979         if (toc == NULL)
 980                 ereport(ERROR,
 981                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 982                    errmsg("invalid magic number in dynamic shared memory segment")));
 983
 984         /* Look up fixed parallel state. */
 985         fps = shm_toc_lookup(toc, PARALLEL_KEY_FIXED);
 986         Assert(fps != NULL);
 987         MyFixedParallelState = fps;
 988
 989         /*
 990          * Now that we have a worker number, we can find and attach to the error
 991          * queue provided for us.  That's good, because until we do that, any
 992          * errors that happen here will not be reported back to the process that
 993          * requested that this worker be launched.
 994          */
 995         error_queue_space = shm_toc_lookup(toc, PARALLEL_KEY_ERROR_QUEUE);
 996         mq = (shm_mq *) (error_queue_space +
 997                                          ParallelWorkerNumber * PARALLEL_ERROR_QUEUE_SIZE);
 998         shm_mq_set_sender(mq, MyProc);
 999         mqh = shm_mq_attach(mq, seg, NULL);
1000         pq_redirect_to_shm_mq(seg, mqh);
1001         pq_set_parallel_master(fps->parallel_master_pid,
1002                                                    fps->parallel_master_backend_id);
1003
1004         /*
1005          * Send a BackendKeyData message to the process that initiated parallelism
1006          * so that it has access to our PID before it receives any other messages
1007          * from us.  Our cancel key is sent, too, since that's the way the
1008          * protocol message is defined, but it won't actually be used for anything
1009          * in this case.
1010          */
1011         pq_beginmessage(&msgbuf, 'K');
1012         pq_sendint(&msgbuf, (int32) MyProcPid, sizeof(int32));
1013         pq_sendint(&msgbuf, (int32) MyCancelKey, sizeof(int32));
1014         pq_endmessage(&msgbuf);
1015
1016         /*
1017          * Hooray! Primary initialization is complete.  Now, we need to set up our
1018          * backend-local state to match the original backend.
1019          */
1020
1021         /*
1022          * Join locking group.  We must do this before anything that could try to
1023          * acquire a heavyweight lock, because any heavyweight locks acquired to
1024          * this point could block either directly against the parallel group
1025          * leader or against some process which in turn waits for a lock that
1026          * conflicts with the parallel group leader, causing an undetected
1027          * deadlock.  (If we can't join the lock group, the leader has gone away,
1028          * so just exit quietly.)
1029          */
1030         if (!BecomeLockGroupMember(fps->parallel_master_pgproc,
1031                                                            fps->parallel_master_pid))
1032                 return;
1033
1034         /*
1035          * Load libraries that were loaded by original backend.  We want to do
1036          * this before restoring GUCs, because the libraries might define custom
1037          * variables.
1038          */
1039         libraryspace = shm_toc_lookup(toc, PARALLEL_KEY_LIBRARY);
1040         Assert(libraryspace != NULL);
1041         RestoreLibraryState(libraryspace);
1042
1043         /* Restore database connection. */
1044         BackgroundWorkerInitializeConnectionByOid(fps->database_id,
1045                                                                                           fps->authenticated_user_id);
1046
1047         /*
1048          * Set the client encoding to the database encoding, since that is what
1049          * the leader will expect.
1050          */
1051         SetClientEncoding(GetDatabaseEncoding());
1052
1053         /* Restore GUC values from launching backend. */
1054         gucspace = shm_toc_lookup(toc, PARALLEL_KEY_GUC);
1055         Assert(gucspace != NULL);
1056         StartTransactionCommand();
1057         RestoreGUCState(gucspace);
1058         CommitTransactionCommand();
1059
1060         /* Crank up a transaction state appropriate to a parallel worker. */
1061         tstatespace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_STATE);
1062         StartParallelWorkerTransaction(tstatespace);
1063
1064         /* Restore combo CID state. */
1065         combocidspace = shm_toc_lookup(toc, PARALLEL_KEY_COMBO_CID);
1066         Assert(combocidspace != NULL);
1067         RestoreComboCIDState(combocidspace);
1068
1069         /* Restore transaction snapshot. */
1070         tsnapspace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT);
1071         Assert(tsnapspace != NULL);
1072         RestoreTransactionSnapshot(RestoreSnapshot(tsnapspace),
1073                                                            fps->parallel_master_pgproc);
1074
1075         /* Restore active snapshot. */
1076         asnapspace = shm_toc_lookup(toc, PARALLEL_KEY_ACTIVE_SNAPSHOT);
1077         Assert(asnapspace != NULL);
1078         PushActiveSnapshot(RestoreSnapshot(asnapspace));
1079
1080         /*
1081          * We've changed which tuples we can see, and must therefore invalidate
1082          * system caches.
1083          */
1084         InvalidateSystemCaches();
1085
1086         /* Restore user ID and security context. */
1087         SetUserIdAndSecContext(fps->current_user_id, fps->sec_context);
1088
1089         /* Restore temp-namespace state to ensure search path matches leader's. */
1090         SetTempNamespaceState(fps->temp_namespace_id,
1091                                                   fps->temp_toast_namespace_id);
1092
1093         /* Set ParallelMasterBackendId so we know how to address temp relations. */
1094         ParallelMasterBackendId = fps->parallel_master_backend_id;
1095
1096         /*
1097          * We've initialized all of our state now; nothing should change
1098          * hereafter.
1099          */
1100         InitializingParallelWorker = false;
1101         EnterParallelMode();
1102
1103         /*
1104          * Time to do the real work: invoke the caller-supplied code.
1105          *
1106          * If you get a crash at this line, see the comments for
1107          * ParallelExtensionTrampoline.
1108          */
1109         fps->entrypoint(seg, toc);
1110
1111         /* Must exit parallel mode to pop active snapshot. */
1112         ExitParallelMode();
1113
1114         /* Must pop active snapshot so resowner.c doesn't complain. */
1115         PopActiveSnapshot();
1116
1117         /* Shut down the parallel-worker transaction. */
1118         EndParallelWorkerTransaction();
1119
1120         /* Report success. */
1121         pq_putmessage('X', NULL, 0);
1122 }
1123
1124 /*
1125  * It's unsafe for the entrypoint invoked by ParallelWorkerMain to be a
1126  * function living in a dynamically loaded module, because the module might
1127  * not be loaded in every process, or might be loaded but not at the same
1128  * address.  To work around that problem, CreateParallelContextForExtension()
1129  * arranges to call this function rather than calling the extension-provided
1130  * function directly; and this function then looks up the real entrypoint and
1131  * calls it.
1132  */
1133 static void
1134 ParallelExtensionTrampoline(dsm_segment *seg, shm_toc *toc)
1135 {
1136         char       *extensionstate;
1137         char       *library_name;
1138         char       *function_name;
1139         parallel_worker_main_type entrypt;
1140
1141         extensionstate = shm_toc_lookup(toc, PARALLEL_KEY_EXTENSION_TRAMPOLINE);
1142         Assert(extensionstate != NULL);
1143         library_name = extensionstate;
1144         function_name = extensionstate + strlen(library_name) + 1;
1145
1146         entrypt = (parallel_worker_main_type)
1147                 load_external_function(library_name, function_name, true, NULL);
1148         entrypt(seg, toc);
1149 }
1150
1151 /*
1152  * Update shared memory with the ending location of the last WAL record we
1153  * wrote, if it's greater than the value already stored there.
1154  */
1155 void
1156 ParallelWorkerReportLastRecEnd(XLogRecPtr last_xlog_end)
1157 {
1158         FixedParallelState *fps = MyFixedParallelState;
1159
1160         Assert(fps != NULL);
1161         SpinLockAcquire(&fps->mutex);
1162         if (fps->last_xlog_end < last_xlog_end)
1163                 fps->last_xlog_end = last_xlog_end;
1164         SpinLockRelease(&fps->mutex);
1165 }