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) 1996-2016, 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 "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 "storage/ipc.h"
  28 #include "storage/sinval.h"
  29 #include "storage/spin.h"
  30 #include "tcop/tcopprot.h"
  31 #include "utils/combocid.h"
  32 #include "utils/guc.h"
  33 #include "utils/inval.h"
  34 #include "utils/memutils.h"
  35 #include "utils/resowner.h"
  36 #include "utils/snapmgr.h"
  37
  38 /*
  39  * We don't want to waste a lot of memory on an error queue which, most of
  40  * the time, will process only a handful of small messages.  However, it is
  41  * desirable to make it large enough that a typical ErrorResponse can be sent
  42  * without blocking.  That way, a worker that errors out can write the whole
  43  * message into the queue and terminate without waiting for the user backend.
  44  */
  45 #define PARALLEL_ERROR_QUEUE_SIZE                       16384
  46
  47 /* Magic number for parallel context TOC. */
  48 #define PARALLEL_MAGIC                                          0x50477c7c
  49
  50 /*
  51  * Magic numbers for parallel state sharing.  Higher-level code should use
  52  * smaller values, leaving these very large ones for use by this module.
  53  */
  54 #define PARALLEL_KEY_FIXED                                      UINT64CONST(0xFFFFFFFFFFFF0001)
  55 #define PARALLEL_KEY_ERROR_QUEUE                        UINT64CONST(0xFFFFFFFFFFFF0002)
  56 #define PARALLEL_KEY_LIBRARY                            UINT64CONST(0xFFFFFFFFFFFF0003)
  57 #define PARALLEL_KEY_GUC                                        UINT64CONST(0xFFFFFFFFFFFF0004)
  58 #define PARALLEL_KEY_COMBO_CID                          UINT64CONST(0xFFFFFFFFFFFF0005)
  59 #define PARALLEL_KEY_TRANSACTION_SNAPSHOT       UINT64CONST(0xFFFFFFFFFFFF0006)
  60 #define PARALLEL_KEY_ACTIVE_SNAPSHOT            UINT64CONST(0xFFFFFFFFFFFF0007)
  61 #define PARALLEL_KEY_TRANSACTION_STATE          UINT64CONST(0xFFFFFFFFFFFF0008)
  62 #define PARALLEL_KEY_EXTENSION_TRAMPOLINE       UINT64CONST(0xFFFFFFFFFFFF0009)
  63
  64 /* Fixed-size parallel state. */
  65 typedef struct FixedParallelState
  66 {
  67         /* Fixed-size state that workers must restore. */
  68         Oid                     database_id;
  69         Oid                     authenticated_user_id;
  70         Oid                     current_user_id;
  71         Oid                     temp_namespace_id;
  72         Oid                     temp_toast_namespace_id;
  73         int                     sec_context;
  74         PGPROC     *parallel_master_pgproc;
  75         pid_t           parallel_master_pid;
  76         BackendId       parallel_master_backend_id;
  77
  78         /* Entrypoint for parallel workers. */
  79         parallel_worker_main_type entrypoint;
  80
  81         /* Mutex protects remaining fields. */
  82         slock_t         mutex;
  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 static void WaitForParallelWorkersToExit(ParallelContext *pcxt);
 114
 115 /*
 116  * Establish a new parallel context.  This should be done after entering
 117  * parallel mode, and (unless there is an error) the context should be
 118  * destroyed before exiting the current subtransaction.
 119  */
 120 ParallelContext *
 121 CreateParallelContext(parallel_worker_main_type entrypoint, int nworkers)
 122 {
 123         MemoryContext oldcontext;
 124         ParallelContext *pcxt;
 125
 126         /* It is unsafe to create a parallel context if not in parallel mode. */
 127         Assert(IsInParallelMode());
 128
 129         /* Number of workers should be non-negative. */
 130         Assert(nworkers >= 0);
 131
 132         /*
 133          * If dynamic shared memory is not available, we won't be able to use
 134          * background workers.
 135          */
 136         if (dynamic_shared_memory_type == DSM_IMPL_NONE)
 137                 nworkers = 0;
 138
 139         /*
 140          * If we are running under serializable isolation, we can't use parallel
 141          * workers, at least not until somebody enhances that mechanism to be
 142          * parallel-aware.
 143          */
 144         if (IsolationIsSerializable())
 145                 nworkers = 0;
 146
 147         /* We might be running in a short-lived memory context. */
 148         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 149
 150         /* Initialize a new ParallelContext. */
 151         pcxt = palloc0(sizeof(ParallelContext));
 152         pcxt->subid = GetCurrentSubTransactionId();
 153         pcxt->nworkers = nworkers;
 154         pcxt->entrypoint = entrypoint;
 155         pcxt->error_context_stack = error_context_stack;
 156         shm_toc_initialize_estimator(&pcxt->estimator);
 157         dlist_push_head(&pcxt_list, &pcxt->node);
 158
 159         /* Restore previous memory context. */
 160         MemoryContextSwitchTo(oldcontext);
 161
 162         return pcxt;
 163 }
 164
 165 /*
 166  * Establish a new parallel context that calls a function provided by an
 167  * extension.  This works around the fact that the library might get mapped
 168  * at a different address in each backend.
 169  */
 170 ParallelContext *
 171 CreateParallelContextForExternalFunction(char *library_name,
 172                                                                                  char *function_name,
 173                                                                                  int nworkers)
 174 {
 175         MemoryContext oldcontext;
 176         ParallelContext *pcxt;
 177
 178         /* We might be running in a very short-lived memory context. */
 179         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 180
 181         /* Create the context. */
 182         pcxt = CreateParallelContext(ParallelExtensionTrampoline, nworkers);
 183         pcxt->library_name = pstrdup(library_name);
 184         pcxt->function_name = pstrdup(function_name);
 185
 186         /* Restore previous memory context. */
 187         MemoryContextSwitchTo(oldcontext);
 188
 189         return pcxt;
 190 }
 191
 192 /*
 193  * Establish the dynamic shared memory segment for a parallel context and
 194  * copy state and other bookkeeping information that will be needed by
 195  * parallel workers into it.
 196  */
 197 void
 198 InitializeParallelDSM(ParallelContext *pcxt)
 199 {
 200         MemoryContext oldcontext;
 201         Size            library_len = 0;
 202         Size            guc_len = 0;
 203         Size            combocidlen = 0;
 204         Size            tsnaplen = 0;
 205         Size            asnaplen = 0;
 206         Size            tstatelen = 0;
 207         Size            segsize = 0;
 208         int                     i;
 209         FixedParallelState *fps;
 210         Snapshot        transaction_snapshot = GetTransactionSnapshot();
 211         Snapshot        active_snapshot = GetActiveSnapshot();
 212
 213         /* We might be running in a very short-lived memory context. */
 214         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 215
 216         /* Allow space to store the fixed-size parallel state. */
 217         shm_toc_estimate_chunk(&pcxt->estimator, sizeof(FixedParallelState));
 218         shm_toc_estimate_keys(&pcxt->estimator, 1);
 219
 220         /*
 221          * Normally, the user will have requested at least one worker process, but
 222          * if by chance they have not, we can skip a bunch of things here.
 223          */
 224         if (pcxt->nworkers > 0)
 225         {
 226                 /* Estimate space for various kinds of state sharing. */
 227                 library_len = EstimateLibraryStateSpace();
 228                 shm_toc_estimate_chunk(&pcxt->estimator, library_len);
 229                 guc_len = EstimateGUCStateSpace();
 230                 shm_toc_estimate_chunk(&pcxt->estimator, guc_len);
 231                 combocidlen = EstimateComboCIDStateSpace();
 232                 shm_toc_estimate_chunk(&pcxt->estimator, combocidlen);
 233                 tsnaplen = EstimateSnapshotSpace(transaction_snapshot);
 234                 shm_toc_estimate_chunk(&pcxt->estimator, tsnaplen);
 235                 asnaplen = EstimateSnapshotSpace(active_snapshot);
 236                 shm_toc_estimate_chunk(&pcxt->estimator, asnaplen);
 237                 tstatelen = EstimateTransactionStateSpace();
 238                 shm_toc_estimate_chunk(&pcxt->estimator, tstatelen);
 239                 /* If you add more chunks here, you probably need to add keys. */
 240                 shm_toc_estimate_keys(&pcxt->estimator, 6);
 241
 242                 /* Estimate space need for error queues. */
 243                 StaticAssertStmt(BUFFERALIGN(PARALLEL_ERROR_QUEUE_SIZE) ==
 244                                                  PARALLEL_ERROR_QUEUE_SIZE,
 245                                                  "parallel error queue size not buffer-aligned");
 246                 shm_toc_estimate_chunk(&pcxt->estimator,
 247                                                            mul_size(PARALLEL_ERROR_QUEUE_SIZE,
 248                                                                                 pcxt->nworkers));
 249                 shm_toc_estimate_keys(&pcxt->estimator, 1);
 250
 251                 /* Estimate how much we'll need for extension entrypoint info. */
 252                 if (pcxt->library_name != NULL)
 253                 {
 254                         Assert(pcxt->entrypoint == ParallelExtensionTrampoline);
 255                         Assert(pcxt->function_name != NULL);
 256                         shm_toc_estimate_chunk(&pcxt->estimator, strlen(pcxt->library_name)
 257                                                                    + strlen(pcxt->function_name) + 2);
 258                         shm_toc_estimate_keys(&pcxt->estimator, 1);
 259                 }
 260         }
 261
 262         /*
 263          * Create DSM and initialize with new table of contents.  But if the user
 264          * didn't request any workers, then don't bother creating a dynamic shared
 265          * memory segment; instead, just use backend-private memory.
 266          *
 267          * Also, if we can't create a dynamic shared memory segment because the
 268          * maximum number of segments have already been created, then fall back to
 269          * backend-private memory, and plan not to use any workers.  We hope this
 270          * won't happen very often, but it's better to abandon the use of
 271          * parallelism than to fail outright.
 272          */
 273         segsize = shm_toc_estimate(&pcxt->estimator);
 274         if (pcxt->nworkers > 0)
 275                 pcxt->seg = dsm_create(segsize, DSM_CREATE_NULL_IF_MAXSEGMENTS);
 276         if (pcxt->seg != NULL)
 277                 pcxt->toc = shm_toc_create(PARALLEL_MAGIC,
 278                                                                    dsm_segment_address(pcxt->seg),
 279                                                                    segsize);
 280         else
 281         {
 282                 pcxt->nworkers = 0;
 283                 pcxt->private_memory = MemoryContextAlloc(TopMemoryContext, segsize);
 284                 pcxt->toc = shm_toc_create(PARALLEL_MAGIC, pcxt->private_memory,
 285                                                                    segsize);
 286         }
 287
 288         /* Initialize fixed-size state in shared memory. */
 289         fps = (FixedParallelState *)
 290                 shm_toc_allocate(pcxt->toc, sizeof(FixedParallelState));
 291         fps->database_id = MyDatabaseId;
 292         fps->authenticated_user_id = GetAuthenticatedUserId();
 293         GetUserIdAndSecContext(&fps->current_user_id, &fps->sec_context);
 294         GetTempNamespaceState(&fps->temp_namespace_id,
 295                                                   &fps->temp_toast_namespace_id);
 296         fps->parallel_master_pgproc = MyProc;
 297         fps->parallel_master_pid = MyProcPid;
 298         fps->parallel_master_backend_id = MyBackendId;
 299         fps->entrypoint = pcxt->entrypoint;
 300         SpinLockInit(&fps->mutex);
 301         fps->last_xlog_end = 0;
 302         shm_toc_insert(pcxt->toc, PARALLEL_KEY_FIXED, fps);
 303
 304         /* We can skip the rest of this if we're not budgeting for any workers. */
 305         if (pcxt->nworkers > 0)
 306         {
 307                 char       *libraryspace;
 308                 char       *gucspace;
 309                 char       *combocidspace;
 310                 char       *tsnapspace;
 311                 char       *asnapspace;
 312                 char       *tstatespace;
 313                 char       *error_queue_space;
 314
 315                 /* Serialize shared libraries we have loaded. */
 316                 libraryspace = shm_toc_allocate(pcxt->toc, library_len);
 317                 SerializeLibraryState(library_len, libraryspace);
 318                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_LIBRARY, libraryspace);
 319
 320                 /* Serialize GUC settings. */
 321                 gucspace = shm_toc_allocate(pcxt->toc, guc_len);
 322                 SerializeGUCState(guc_len, gucspace);
 323                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_GUC, gucspace);
 324
 325                 /* Serialize combo CID state. */
 326                 combocidspace = shm_toc_allocate(pcxt->toc, combocidlen);
 327                 SerializeComboCIDState(combocidlen, combocidspace);
 328                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_COMBO_CID, combocidspace);
 329
 330                 /* Serialize transaction snapshot and active snapshot. */
 331                 tsnapspace = shm_toc_allocate(pcxt->toc, tsnaplen);
 332                 SerializeSnapshot(transaction_snapshot, tsnapspace);
 333                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT,
 334                                            tsnapspace);
 335                 asnapspace = shm_toc_allocate(pcxt->toc, asnaplen);
 336                 SerializeSnapshot(active_snapshot, asnapspace);
 337                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ACTIVE_SNAPSHOT, asnapspace);
 338
 339                 /* Serialize transaction state. */
 340                 tstatespace = shm_toc_allocate(pcxt->toc, tstatelen);
 341                 SerializeTransactionState(tstatelen, tstatespace);
 342                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_STATE, tstatespace);
 343
 344                 /* Allocate space for worker information. */
 345                 pcxt->worker = palloc0(sizeof(ParallelWorkerInfo) * pcxt->nworkers);
 346
 347                 /*
 348                  * Establish error queues in dynamic shared memory.
 349                  *
 350                  * These queues should be used only for transmitting ErrorResponse,
 351                  * NoticeResponse, and NotifyResponse protocol messages.  Tuple data
 352                  * should be transmitted via separate (possibly larger?) queues.
 353                  */
 354                 error_queue_space =
 355                         shm_toc_allocate(pcxt->toc,
 356                                                          mul_size(PARALLEL_ERROR_QUEUE_SIZE,
 357                                                                           pcxt->nworkers));
 358                 for (i = 0; i < pcxt->nworkers; ++i)
 359                 {
 360                         char       *start;
 361                         shm_mq     *mq;
 362
 363                         start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
 364                         mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
 365                         shm_mq_set_receiver(mq, MyProc);
 366                         pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
 367                 }
 368                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE, error_queue_space);
 369
 370                 /* Serialize extension entrypoint information. */
 371                 if (pcxt->library_name != NULL)
 372                 {
 373                         Size            lnamelen = strlen(pcxt->library_name);
 374                         char       *extensionstate;
 375
 376                         extensionstate = shm_toc_allocate(pcxt->toc, lnamelen
 377                                                                                   + strlen(pcxt->function_name) + 2);
 378                         strcpy(extensionstate, pcxt->library_name);
 379                         strcpy(extensionstate + lnamelen + 1, pcxt->function_name);
 380                         shm_toc_insert(pcxt->toc, PARALLEL_KEY_EXTENSION_TRAMPOLINE,
 381                                                    extensionstate);
 382                 }
 383         }
 384
 385         /* Restore previous memory context. */
 386         MemoryContextSwitchTo(oldcontext);
 387 }
 388
 389 /*
 390  * Reinitialize the dynamic shared memory segment for a parallel context such
 391  * that we could launch workers for it again.
 392  */
 393 void
 394 ReinitializeParallelDSM(ParallelContext *pcxt)
 395 {
 396         FixedParallelState *fps;
 397         char       *error_queue_space;
 398         int                     i;
 399
 400         /* Wait for any old workers to exit. */
 401         if (pcxt->nworkers_launched > 0)
 402         {
 403                 WaitForParallelWorkersToFinish(pcxt);
 404                 WaitForParallelWorkersToExit(pcxt);
 405                 pcxt->nworkers_launched = 0;
 406         }
 407
 408         /* Reset a few bits of fixed parallel state to a clean state. */
 409         fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED);
 410         fps->last_xlog_end = 0;
 411
 412         /* Recreate error queues. */
 413         error_queue_space =
 414                 shm_toc_lookup(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE);
 415         for (i = 0; i < pcxt->nworkers; ++i)
 416         {
 417                 char       *start;
 418                 shm_mq     *mq;
 419
 420                 start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
 421                 mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
 422                 shm_mq_set_receiver(mq, MyProc);
 423                 pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
 424         }
 425 }
 426
 427 /*
 428  * Launch parallel workers.
 429  */
 430 void
 431 LaunchParallelWorkers(ParallelContext *pcxt)
 432 {
 433         MemoryContext oldcontext;
 434         BackgroundWorker worker;
 435         int                     i;
 436         bool            any_registrations_failed = false;
 437
 438         /* Skip this if we have no workers. */
 439         if (pcxt->nworkers == 0)
 440                 return;
 441
 442         /* We need to be a lock group leader. */
 443         BecomeLockGroupLeader();
 444
 445         /* If we do have workers, we'd better have a DSM segment. */
 446         Assert(pcxt->seg != NULL);
 447
 448         /* We might be running in a short-lived memory context. */
 449         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 450
 451         /* Configure a worker. */
 452         snprintf(worker.bgw_name, BGW_MAXLEN, "parallel worker for PID %d",
 453                          MyProcPid);
 454         worker.bgw_flags =
 455                 BGWORKER_SHMEM_ACCESS | BGWORKER_BACKEND_DATABASE_CONNECTION;
 456         worker.bgw_start_time = BgWorkerStart_ConsistentState;
 457         worker.bgw_restart_time = BGW_NEVER_RESTART;
 458         worker.bgw_main = ParallelWorkerMain;
 459         worker.bgw_main_arg = UInt32GetDatum(dsm_segment_handle(pcxt->seg));
 460         worker.bgw_notify_pid = MyProcPid;
 461         memset(&worker.bgw_extra, 0, BGW_EXTRALEN);
 462
 463         /*
 464          * Start workers.
 465          *
 466          * The caller must be able to tolerate ending up with fewer workers than
 467          * expected, so there is no need to throw an error here if registration
 468          * fails.  It wouldn't help much anyway, because registering the worker in
 469          * no way guarantees that it will start up and initialize successfully.
 470          */
 471         for (i = 0; i < pcxt->nworkers; ++i)
 472         {
 473                 memcpy(worker.bgw_extra, &i, sizeof(int));
 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                         pfree(pcxt->worker[i].error_mqh);
 496                         pcxt->worker[i].error_mqh = NULL;
 497                 }
 498         }
 499
 500         /* Restore previous memory context. */
 501         MemoryContextSwitchTo(oldcontext);
 502 }
 503
 504 /*
 505  * Wait for all workers to finish computing.
 506  *
 507  * Even if the parallel operation seems to have completed successfully, it's
 508  * important to call this function afterwards.  We must not miss any errors
 509  * the workers may have thrown during the parallel operation, or any that they
 510  * may yet throw while shutting down.
 511  *
 512  * Also, we want to update our notion of XactLastRecEnd based on worker
 513  * feedback.
 514  */
 515 void
 516 WaitForParallelWorkersToFinish(ParallelContext *pcxt)
 517 {
 518         for (;;)
 519         {
 520                 bool            anyone_alive = false;
 521                 int                     i;
 522
 523                 /*
 524                  * This will process any parallel messages that are pending, which may
 525                  * change the outcome of the loop that follows.  It may also throw an
 526                  * error propagated from a worker.
 527                  */
 528                 CHECK_FOR_INTERRUPTS();
 529
 530                 for (i = 0; i < pcxt->nworkers_launched; ++i)
 531                 {
 532                         if (pcxt->worker[i].error_mqh != NULL)
 533                         {
 534                                 anyone_alive = true;
 535                                 break;
 536                         }
 537                 }
 538
 539                 if (!anyone_alive)
 540                         break;
 541
 542                 WaitLatch(&MyProc->procLatch, WL_LATCH_SET, -1);
 543                 ResetLatch(&MyProc->procLatch);
 544         }
 545
 546         if (pcxt->toc != NULL)
 547         {
 548                 FixedParallelState *fps;
 549
 550                 fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED);
 551                 if (fps->last_xlog_end > XactLastRecEnd)
 552                         XactLastRecEnd = fps->last_xlog_end;
 553         }
 554 }
 555
 556 /*
 557  * Wait for all workers to exit.
 558  *
 559  * This function ensures that workers have been completely shutdown.  The
 560  * difference between WaitForParallelWorkersToFinish and this function is
 561  * that former just ensures that last message sent by worker backend is
 562  * received by master backend whereas this ensures the complete shutdown.
 563  */
 564 static void
 565 WaitForParallelWorkersToExit(ParallelContext *pcxt)
 566 {
 567         int                     i;
 568
 569         /* Wait until the workers actually die. */
 570         for (i = 0; i < pcxt->nworkers_launched; ++i)
 571         {
 572                 BgwHandleStatus status;
 573
 574                 if (pcxt->worker == NULL || pcxt->worker[i].bgwhandle == NULL)
 575                         continue;
 576
 577                 status = WaitForBackgroundWorkerShutdown(pcxt->worker[i].bgwhandle);
 578
 579                 /*
 580                  * If the postmaster kicked the bucket, we have no chance of cleaning
 581                  * up safely -- we won't be able to tell when our workers are actually
 582                  * dead.  This doesn't necessitate a PANIC since they will all abort
 583                  * eventually, but we can't safely continue this session.
 584                  */
 585                 if (status == BGWH_POSTMASTER_DIED)
 586                         ereport(FATAL,
 587                                         (errcode(ERRCODE_ADMIN_SHUTDOWN),
 588                                  errmsg("postmaster exited during a parallel transaction")));
 589
 590                 /* Release memory. */
 591                 pfree(pcxt->worker[i].bgwhandle);
 592                 pcxt->worker[i].bgwhandle = NULL;
 593         }
 594 }
 595
 596 /*
 597  * Destroy a parallel context.
 598  *
 599  * If expecting a clean exit, you should use WaitForParallelWorkersToFinish()
 600  * first, before calling this function.  When this function is invoked, any
 601  * remaining workers are forcibly killed; the dynamic shared memory segment
 602  * is unmapped; and we then wait (uninterruptibly) for the workers to exit.
 603  */
 604 void
 605 DestroyParallelContext(ParallelContext *pcxt)
 606 {
 607         int                     i;
 608
 609         /*
 610          * Be careful about order of operations here!  We remove the parallel
 611          * context from the list before we do anything else; otherwise, if an
 612          * error occurs during a subsequent step, we might try to nuke it again
 613          * from AtEOXact_Parallel or AtEOSubXact_Parallel.
 614          */
 615         dlist_delete(&pcxt->node);
 616
 617         /* Kill each worker in turn, and forget their error queues. */
 618         if (pcxt->worker != NULL)
 619         {
 620                 for (i = 0; i < pcxt->nworkers_launched; ++i)
 621                 {
 622                         if (pcxt->worker[i].error_mqh != NULL)
 623                         {
 624                                 TerminateBackgroundWorker(pcxt->worker[i].bgwhandle);
 625
 626                                 pfree(pcxt->worker[i].error_mqh);
 627                                 pcxt->worker[i].error_mqh = NULL;
 628                         }
 629                 }
 630         }
 631
 632         /*
 633          * If we have allocated a shared memory segment, detach it.  This will
 634          * implicitly detach the error queues, and any other shared memory queues,
 635          * stored there.
 636          */
 637         if (pcxt->seg != NULL)
 638         {
 639                 dsm_detach(pcxt->seg);
 640                 pcxt->seg = NULL;
 641         }
 642
 643         /*
 644          * If this parallel context is actually in backend-private memory rather
 645          * than shared memory, free that memory instead.
 646          */
 647         if (pcxt->private_memory != NULL)
 648         {
 649                 pfree(pcxt->private_memory);
 650                 pcxt->private_memory = NULL;
 651         }
 652
 653         /*
 654          * We can't finish transaction commit or abort until all of the workers
 655          * have exited.  This means, in particular, that we can't respond to
 656          * interrupts at this stage.
 657          */
 658         HOLD_INTERRUPTS();
 659         WaitForParallelWorkersToExit(pcxt);
 660         RESUME_INTERRUPTS();
 661
 662         /* Free the worker array itself. */
 663         if (pcxt->worker != NULL)
 664         {
 665                 pfree(pcxt->worker);
 666                 pcxt->worker = NULL;
 667         }
 668
 669         /* Free memory. */
 670         pfree(pcxt);
 671 }
 672
 673 /*
 674  * Are there any parallel contexts currently active?
 675  */
 676 bool
 677 ParallelContextActive(void)
 678 {
 679         return !dlist_is_empty(&pcxt_list);
 680 }
 681
 682 /*
 683  * Handle receipt of an interrupt indicating a parallel worker message.
 684  */
 685 void
 686 HandleParallelMessageInterrupt(void)
 687 {
 688         int                     save_errno = errno;
 689
 690         InterruptPending = true;
 691         ParallelMessagePending = true;
 692         SetLatch(MyLatch);
 693
 694         errno = save_errno;
 695 }
 696
 697 /*
 698  * Handle any queued protocol messages received from parallel workers.
 699  */
 700 void
 701 HandleParallelMessages(void)
 702 {
 703         dlist_iter      iter;
 704
 705         ParallelMessagePending = false;
 706
 707         dlist_foreach(iter, &pcxt_list)
 708         {
 709                 ParallelContext *pcxt;
 710                 int                     i;
 711                 Size            nbytes;
 712                 void       *data;
 713
 714                 pcxt = dlist_container(ParallelContext, node, iter.cur);
 715                 if (pcxt->worker == NULL)
 716                         continue;
 717
 718                 for (i = 0; i < pcxt->nworkers_launched; ++i)
 719                 {
 720                         /*
 721                          * Read as many messages as we can from each worker, but stop when
 722                          * either (1) the error queue goes away, which can happen if we
 723                          * receive a Terminate message from the worker; or (2) no more
 724                          * messages can be read from the worker without blocking.
 725                          */
 726                         while (pcxt->worker[i].error_mqh != NULL)
 727                         {
 728                                 shm_mq_result res;
 729
 730                                 res = shm_mq_receive(pcxt->worker[i].error_mqh, &nbytes,
 731                                                                          &data, true);
 732                                 if (res == SHM_MQ_WOULD_BLOCK)
 733                                         break;
 734                                 else if (res == SHM_MQ_SUCCESS)
 735                                 {
 736                                         StringInfoData msg;
 737
 738                                         initStringInfo(&msg);
 739                                         appendBinaryStringInfo(&msg, data, nbytes);
 740                                         HandleParallelMessage(pcxt, i, &msg);
 741                                         pfree(msg.data);
 742                                 }
 743                                 else
 744                                         ereport(ERROR,
 745                                                         (errcode(ERRCODE_INTERNAL_ERROR),       /* XXX: wrong errcode? */
 746                                                          errmsg("lost connection to parallel worker")));
 747
 748                                 /* This might make the error queue go away. */
 749                                 CHECK_FOR_INTERRUPTS();
 750                         }
 751                 }
 752         }
 753 }
 754
 755 /*
 756  * Handle a single protocol message received from a single parallel worker.
 757  */
 758 static void
 759 HandleParallelMessage(ParallelContext *pcxt, int i, StringInfo msg)
 760 {
 761         char            msgtype;
 762
 763         msgtype = pq_getmsgbyte(msg);
 764
 765         switch (msgtype)
 766         {
 767                 case 'K':                               /* BackendKeyData */
 768                         {
 769                                 int32           pid = pq_getmsgint(msg, 4);
 770
 771                                 (void) pq_getmsgint(msg, 4);    /* discard cancel key */
 772                                 (void) pq_getmsgend(msg);
 773                                 pcxt->worker[i].pid = pid;
 774                                 break;
 775                         }
 776
 777                 case 'E':                               /* ErrorResponse */
 778                 case 'N':                               /* NoticeResponse */
 779                         {
 780                                 ErrorData       edata;
 781                                 ErrorContextCallback errctx;
 782                                 ErrorContextCallback *save_error_context_stack;
 783
 784                                 /*
 785                                  * Rethrow the error using the error context callbacks that
 786                                  * were in effect when the context was created, not the
 787                                  * current ones.
 788                                  */
 789                                 save_error_context_stack = error_context_stack;
 790                                 errctx.callback = ParallelErrorContext;
 791                                 errctx.arg = NULL;
 792                                 errctx.previous = pcxt->error_context_stack;
 793                                 error_context_stack = &errctx;
 794
 795                                 /* Parse ErrorResponse or NoticeResponse. */
 796                                 pq_parse_errornotice(msg, &edata);
 797
 798                                 /* Death of a worker isn't enough justification for suicide. */
 799                                 edata.elevel = Min(edata.elevel, ERROR);
 800
 801                                 /* Rethrow error or notice. */
 802                                 ThrowErrorData(&edata);
 803
 804                                 /* Restore previous context. */
 805                                 error_context_stack = save_error_context_stack;
 806
 807                                 break;
 808                         }
 809
 810                 case 'A':                               /* NotifyResponse */
 811                         {
 812                                 /* Propagate NotifyResponse. */
 813                                 int32           pid;
 814                                 const char *channel;
 815                                 const char *payload;
 816
 817                                 pid = pq_getmsgint(msg, 4);
 818                                 channel = pq_getmsgrawstring(msg);
 819                                 payload = pq_getmsgrawstring(msg);
 820                                 pq_endmessage(msg);
 821
 822                                 NotifyMyFrontEnd(channel, payload, pid);
 823
 824                                 break;
 825                         }
 826
 827                 case 'X':                               /* Terminate, indicating clean exit */
 828                         {
 829                                 pfree(pcxt->worker[i].error_mqh);
 830                                 pcxt->worker[i].error_mqh = NULL;
 831                                 break;
 832                         }
 833
 834                 default:
 835                         {
 836                                 elog(ERROR, "unknown message type: %c (%d bytes)",
 837                                          msgtype, msg->len);
 838                         }
 839         }
 840 }
 841
 842 /*
 843  * End-of-subtransaction cleanup for parallel contexts.
 844  *
 845  * Currently, it's forbidden to enter or leave a subtransaction while
 846  * parallel mode is in effect, so we could just blow away everything.  But
 847  * we may want to relax that restriction in the future, so this code
 848  * contemplates that there may be multiple subtransaction IDs in pcxt_list.
 849  */
 850 void
 851 AtEOSubXact_Parallel(bool isCommit, SubTransactionId mySubId)
 852 {
 853         while (!dlist_is_empty(&pcxt_list))
 854         {
 855                 ParallelContext *pcxt;
 856
 857                 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
 858                 if (pcxt->subid != mySubId)
 859                         break;
 860                 if (isCommit)
 861                         elog(WARNING, "leaked parallel context");
 862                 DestroyParallelContext(pcxt);
 863         }
 864 }
 865
 866 /*
 867  * End-of-transaction cleanup for parallel contexts.
 868  */
 869 void
 870 AtEOXact_Parallel(bool isCommit)
 871 {
 872         while (!dlist_is_empty(&pcxt_list))
 873         {
 874                 ParallelContext *pcxt;
 875
 876                 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
 877                 if (isCommit)
 878                         elog(WARNING, "leaked parallel context");
 879                 DestroyParallelContext(pcxt);
 880         }
 881 }
 882
 883 /*
 884  * Main entrypoint for parallel workers.
 885  */
 886 static void
 887 ParallelWorkerMain(Datum main_arg)
 888 {
 889         dsm_segment *seg;
 890         shm_toc    *toc;
 891         FixedParallelState *fps;
 892         char       *error_queue_space;
 893         shm_mq     *mq;
 894         shm_mq_handle *mqh;
 895         char       *libraryspace;
 896         char       *gucspace;
 897         char       *combocidspace;
 898         char       *tsnapspace;
 899         char       *asnapspace;
 900         char       *tstatespace;
 901         StringInfoData msgbuf;
 902
 903         /* Set flag to indicate that we're initializing a parallel worker. */
 904         InitializingParallelWorker = true;
 905
 906         /* Establish signal handlers. */
 907         pqsignal(SIGTERM, die);
 908         BackgroundWorkerUnblockSignals();
 909
 910         /* Determine and set our parallel worker number. */
 911         Assert(ParallelWorkerNumber == -1);
 912         memcpy(&ParallelWorkerNumber, MyBgworkerEntry->bgw_extra, sizeof(int));
 913
 914         /* Set up a memory context and resource owner. */
 915         Assert(CurrentResourceOwner == NULL);
 916         CurrentResourceOwner = ResourceOwnerCreate(NULL, "parallel toplevel");
 917         CurrentMemoryContext = AllocSetContextCreate(TopMemoryContext,
 918                                                                                                  "parallel worker",
 919                                                                                                  ALLOCSET_DEFAULT_MINSIZE,
 920                                                                                                  ALLOCSET_DEFAULT_INITSIZE,
 921                                                                                                  ALLOCSET_DEFAULT_MAXSIZE);
 922
 923         /*
 924          * Now that we have a resource owner, we can attach to the dynamic shared
 925          * memory segment and read the table of contents.
 926          */
 927         seg = dsm_attach(DatumGetUInt32(main_arg));
 928         if (seg == NULL)
 929                 ereport(ERROR,
 930                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 931                                  errmsg("could not map dynamic shared memory segment")));
 932         toc = shm_toc_attach(PARALLEL_MAGIC, dsm_segment_address(seg));
 933         if (toc == NULL)
 934                 ereport(ERROR,
 935                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 936                    errmsg("invalid magic number in dynamic shared memory segment")));
 937
 938         /* Look up fixed parallel state. */
 939         fps = shm_toc_lookup(toc, PARALLEL_KEY_FIXED);
 940         Assert(fps != NULL);
 941         MyFixedParallelState = fps;
 942
 943         /*
 944          * Now that we have a worker number, we can find and attach to the error
 945          * queue provided for us.  That's good, because until we do that, any
 946          * errors that happen here will not be reported back to the process that
 947          * requested that this worker be launched.
 948          */
 949         error_queue_space = shm_toc_lookup(toc, PARALLEL_KEY_ERROR_QUEUE);
 950         mq = (shm_mq *) (error_queue_space +
 951                                          ParallelWorkerNumber * PARALLEL_ERROR_QUEUE_SIZE);
 952         shm_mq_set_sender(mq, MyProc);
 953         mqh = shm_mq_attach(mq, seg, NULL);
 954         pq_redirect_to_shm_mq(seg, mqh);
 955         pq_set_parallel_master(fps->parallel_master_pid,
 956                                                    fps->parallel_master_backend_id);
 957
 958         /*
 959          * Send a BackendKeyData message to the process that initiated parallelism
 960          * so that it has access to our PID before it receives any other messages
 961          * from us.  Our cancel key is sent, too, since that's the way the
 962          * protocol message is defined, but it won't actually be used for anything
 963          * in this case.
 964          */
 965         pq_beginmessage(&msgbuf, 'K');
 966         pq_sendint(&msgbuf, (int32) MyProcPid, sizeof(int32));
 967         pq_sendint(&msgbuf, (int32) MyCancelKey, sizeof(int32));
 968         pq_endmessage(&msgbuf);
 969
 970         /*
 971          * Hooray! Primary initialization is complete.  Now, we need to set up our
 972          * backend-local state to match the original backend.
 973          */
 974
 975         /*
 976          * Join locking group.  We must do this before anything that could try to
 977          * acquire a heavyweight lock, because any heavyweight locks acquired to
 978          * this point could block either directly against the parallel group
 979          * leader or against some process which in turn waits for a lock that
 980          * conflicts with the parallel group leader, causing an undetected
 981          * deadlock.  (If we can't join the lock group, the leader has gone away,
 982          * so just exit quietly.)
 983          */
 984         if (!BecomeLockGroupMember(fps->parallel_master_pgproc,
 985                                                            fps->parallel_master_pid))
 986                 return;
 987
 988         /*
 989          * Load libraries that were loaded by original backend.  We want to do
 990          * this before restoring GUCs, because the libraries might define custom
 991          * variables.
 992          */
 993         libraryspace = shm_toc_lookup(toc, PARALLEL_KEY_LIBRARY);
 994         Assert(libraryspace != NULL);
 995         RestoreLibraryState(libraryspace);
 996
 997         /* Restore database connection. */
 998         BackgroundWorkerInitializeConnectionByOid(fps->database_id,
 999                                                                                           fps->authenticated_user_id);
1000
1001         /*
1002          * Set the client encoding to the database encoding, since that is what
1003          * the leader will expect.
1004          */
1005         SetClientEncoding(GetDatabaseEncoding());
1006
1007         /* Restore GUC values from launching backend. */
1008         gucspace = shm_toc_lookup(toc, PARALLEL_KEY_GUC);
1009         Assert(gucspace != NULL);
1010         StartTransactionCommand();
1011         RestoreGUCState(gucspace);
1012         CommitTransactionCommand();
1013
1014         /* Crank up a transaction state appropriate to a parallel worker. */
1015         tstatespace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_STATE);
1016         StartParallelWorkerTransaction(tstatespace);
1017
1018         /* Restore combo CID state. */
1019         combocidspace = shm_toc_lookup(toc, PARALLEL_KEY_COMBO_CID);
1020         Assert(combocidspace != NULL);
1021         RestoreComboCIDState(combocidspace);
1022
1023         /* Restore transaction snapshot. */
1024         tsnapspace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT);
1025         Assert(tsnapspace != NULL);
1026         RestoreTransactionSnapshot(RestoreSnapshot(tsnapspace),
1027                                                            fps->parallel_master_pgproc);
1028
1029         /* Restore active snapshot. */
1030         asnapspace = shm_toc_lookup(toc, PARALLEL_KEY_ACTIVE_SNAPSHOT);
1031         Assert(asnapspace != NULL);
1032         PushActiveSnapshot(RestoreSnapshot(asnapspace));
1033
1034         /*
1035          * We've changed which tuples we can see, and must therefore invalidate
1036          * system caches.
1037          */
1038         InvalidateSystemCaches();
1039
1040         /* Restore user ID and security context. */
1041         SetUserIdAndSecContext(fps->current_user_id, fps->sec_context);
1042
1043         /* Restore temp-namespace state to ensure search path matches leader's. */
1044         SetTempNamespaceState(fps->temp_namespace_id,
1045                                                   fps->temp_toast_namespace_id);
1046
1047         /* Set ParallelMasterBackendId so we know how to address temp relations. */
1048         ParallelMasterBackendId = fps->parallel_master_backend_id;
1049
1050         /*
1051          * We've initialized all of our state now; nothing should change
1052          * hereafter.
1053          */
1054         InitializingParallelWorker = false;
1055         EnterParallelMode();
1056
1057         /*
1058          * Time to do the real work: invoke the caller-supplied code.
1059          *
1060          * If you get a crash at this line, see the comments for
1061          * ParallelExtensionTrampoline.
1062          */
1063         fps->entrypoint(seg, toc);
1064
1065         /* Must exit parallel mode to pop active snapshot. */
1066         ExitParallelMode();
1067
1068         /* Must pop active snapshot so resowner.c doesn't complain. */
1069         PopActiveSnapshot();
1070
1071         /* Shut down the parallel-worker transaction. */
1072         EndParallelWorkerTransaction();
1073
1074         /* Report success. */
1075         pq_putmessage('X', NULL, 0);
1076 }
1077
1078 /*
1079  * It's unsafe for the entrypoint invoked by ParallelWorkerMain to be a
1080  * function living in a dynamically loaded module, because the module might
1081  * not be loaded in every process, or might be loaded but not at the same
1082  * address.  To work around that problem, CreateParallelContextForExtension()
1083  * arranges to call this function rather than calling the extension-provided
1084  * function directly; and this function then looks up the real entrypoint and
1085  * calls it.
1086  */
1087 static void
1088 ParallelExtensionTrampoline(dsm_segment *seg, shm_toc *toc)
1089 {
1090         char       *extensionstate;
1091         char       *library_name;
1092         char       *function_name;
1093         parallel_worker_main_type entrypt;
1094
1095         extensionstate = shm_toc_lookup(toc, PARALLEL_KEY_EXTENSION_TRAMPOLINE);
1096         Assert(extensionstate != NULL);
1097         library_name = extensionstate;
1098         function_name = extensionstate + strlen(library_name) + 1;
1099
1100         entrypt = (parallel_worker_main_type)
1101                 load_external_function(library_name, function_name, true, NULL);
1102         entrypt(seg, toc);
1103 }
1104
1105 /*
1106  * Give the user a hint that this is a message propagated from a parallel
1107  * worker.  Otherwise, it can sometimes be confusing to understand what
1108  * actually happened.
1109  */
1110 static void
1111 ParallelErrorContext(void *arg)
1112 {
1113         if (force_parallel_mode != FORCE_PARALLEL_REGRESS)
1114                 errcontext("parallel worker");
1115 }
1116
1117 /*
1118  * Update shared memory with the ending location of the last WAL record we
1119  * wrote, if it's greater than the value already stored there.
1120  */
1121 void
1122 ParallelWorkerReportLastRecEnd(XLogRecPtr last_xlog_end)
1123 {
1124         FixedParallelState *fps = MyFixedParallelState;
1125
1126         Assert(fps != NULL);
1127         SpinLockAcquire(&fps->mutex);
1128         if (fps->last_xlog_end < last_xlog_end)
1129                 fps->last_xlog_end = last_xlog_end;
1130         SpinLockRelease(&fps->mutex);
1131 }