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/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 "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 /*
  40  * We don't want to waste a lot of memory on an error queue which, most of
  41  * the time, will process only a handful of small messages.  However, it is
  42  * desirable to make it large enough that a typical ErrorResponse can be sent
  43  * without blocking.  That way, a worker that errors out can write the whole
  44  * message into the queue and terminate without waiting for the user backend.
  45  */
  46 #define PARALLEL_ERROR_QUEUE_SIZE                       16384
  47
  48 /* Magic number for parallel context TOC. */
  49 #define PARALLEL_MAGIC                                          0x50477c7c
  50
  51 /*
  52  * Magic numbers for parallel state sharing.  Higher-level code should use
  53  * smaller values, leaving these very large ones for use by this module.
  54  */
  55 #define PARALLEL_KEY_FIXED                                      UINT64CONST(0xFFFFFFFFFFFF0001)
  56 #define PARALLEL_KEY_ERROR_QUEUE                        UINT64CONST(0xFFFFFFFFFFFF0002)
  57 #define PARALLEL_KEY_LIBRARY                            UINT64CONST(0xFFFFFFFFFFFF0003)
  58 #define PARALLEL_KEY_GUC                                        UINT64CONST(0xFFFFFFFFFFFF0004)
  59 #define PARALLEL_KEY_COMBO_CID                          UINT64CONST(0xFFFFFFFFFFFF0005)
  60 #define PARALLEL_KEY_TRANSACTION_SNAPSHOT       UINT64CONST(0xFFFFFFFFFFFF0006)
  61 #define PARALLEL_KEY_ACTIVE_SNAPSHOT            UINT64CONST(0xFFFFFFFFFFFF0007)
  62 #define PARALLEL_KEY_TRANSACTION_STATE          UINT64CONST(0xFFFFFFFFFFFF0008)
  63 #define PARALLEL_KEY_EXTENSION_TRAMPOLINE       UINT64CONST(0xFFFFFFFFFFFF0009)
  64
  65 /* Fixed-size parallel state. */
  66 typedef struct FixedParallelState
  67 {
  68         /* Fixed-size state that workers must restore. */
  69         Oid                     database_id;
  70         Oid                     authenticated_user_id;
  71         Oid                     current_user_id;
  72         Oid                     temp_namespace_id;
  73         Oid                     temp_toast_namespace_id;
  74         int                     sec_context;
  75         PGPROC     *parallel_master_pgproc;
  76         pid_t           parallel_master_pid;
  77         BackendId       parallel_master_backend_id;
  78
  79         /* Entrypoint for parallel workers. */
  80         parallel_worker_main_type entrypoint;
  81
  82         /* Mutex protects remaining fields. */
  83         slock_t         mutex;
  84
  85         /* Maximum XactLastRecEnd of any worker. */
  86         XLogRecPtr      last_xlog_end;
  87 } FixedParallelState;
  88
  89 /*
  90  * Our parallel worker number.  We initialize this to -1, meaning that we are
  91  * not a parallel worker.  In parallel workers, it will be set to a value >= 0
  92  * and < the number of workers before any user code is invoked; each parallel
  93  * worker will get a different parallel worker number.
  94  */
  95 int                     ParallelWorkerNumber = -1;
  96
  97 /* Is there a parallel message pending which we need to receive? */
  98 volatile bool ParallelMessagePending = false;
  99
 100 /* Are we initializing a parallel worker? */
 101 bool            InitializingParallelWorker = false;
 102
 103 /* Pointer to our fixed parallel state. */
 104 static FixedParallelState *MyFixedParallelState;
 105
 106 /* List of active parallel contexts. */
 107 static dlist_head pcxt_list = DLIST_STATIC_INIT(pcxt_list);
 108
 109 /* Private functions. */
 110 static void HandleParallelMessage(ParallelContext *pcxt, int i, StringInfo msg);
 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 /*
 117  * Establish a new parallel context.  This should be done after entering
 118  * parallel mode, and (unless there is an error) the context should be
 119  * destroyed before exiting the current subtransaction.
 120  */
 121 ParallelContext *
 122 CreateParallelContext(parallel_worker_main_type entrypoint, int nworkers)
 123 {
 124         MemoryContext oldcontext;
 125         ParallelContext *pcxt;
 126
 127         /* It is unsafe to create a parallel context if not in parallel mode. */
 128         Assert(IsInParallelMode());
 129
 130         /* Number of workers should be non-negative. */
 131         Assert(nworkers >= 0);
 132
 133         /*
 134          * If dynamic shared memory is not available, we won't be able to use
 135          * background workers.
 136          */
 137         if (dynamic_shared_memory_type == DSM_IMPL_NONE)
 138                 nworkers = 0;
 139
 140         /*
 141          * If we are running under serializable isolation, we can't use parallel
 142          * workers, at least not until somebody enhances that mechanism to be
 143          * parallel-aware.
 144          */
 145         if (IsolationIsSerializable())
 146                 nworkers = 0;
 147
 148         /* We might be running in a short-lived memory context. */
 149         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 150
 151         /* Initialize a new ParallelContext. */
 152         pcxt = palloc0(sizeof(ParallelContext));
 153         pcxt->subid = GetCurrentSubTransactionId();
 154         pcxt->nworkers = nworkers;
 155         pcxt->entrypoint = entrypoint;
 156         pcxt->error_context_stack = error_context_stack;
 157         shm_toc_initialize_estimator(&pcxt->estimator);
 158         dlist_push_head(&pcxt_list, &pcxt->node);
 159
 160         /* Restore previous memory context. */
 161         MemoryContextSwitchTo(oldcontext);
 162
 163         return pcxt;
 164 }
 165
 166 /*
 167  * Establish a new parallel context that calls a function provided by an
 168  * extension.  This works around the fact that the library might get mapped
 169  * at a different address in each backend.
 170  */
 171 ParallelContext *
 172 CreateParallelContextForExternalFunction(char *library_name,
 173                                                                                  char *function_name,
 174                                                                                  int nworkers)
 175 {
 176         MemoryContext oldcontext;
 177         ParallelContext *pcxt;
 178
 179         /* We might be running in a very short-lived memory context. */
 180         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 181
 182         /* Create the context. */
 183         pcxt = CreateParallelContext(ParallelExtensionTrampoline, nworkers);
 184         pcxt->library_name = pstrdup(library_name);
 185         pcxt->function_name = pstrdup(function_name);
 186
 187         /* Restore previous memory context. */
 188         MemoryContextSwitchTo(oldcontext);
 189
 190         return pcxt;
 191 }
 192
 193 /*
 194  * Establish the dynamic shared memory segment for a parallel context and
 195  * copy state and other bookkeeping information that will be needed by
 196  * parallel workers into it.
 197  */
 198 void
 199 InitializeParallelDSM(ParallelContext *pcxt)
 200 {
 201         MemoryContext oldcontext;
 202         Size            library_len = 0;
 203         Size            guc_len = 0;
 204         Size            combocidlen = 0;
 205         Size            tsnaplen = 0;
 206         Size            asnaplen = 0;
 207         Size            tstatelen = 0;
 208         Size            segsize = 0;
 209         int                     i;
 210         FixedParallelState *fps;
 211         Snapshot        transaction_snapshot = GetTransactionSnapshot();
 212         Snapshot        active_snapshot = GetActiveSnapshot();
 213
 214         /* We might be running in a very short-lived memory context. */
 215         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 216
 217         /* Allow space to store the fixed-size parallel state. */
 218         shm_toc_estimate_chunk(&pcxt->estimator, sizeof(FixedParallelState));
 219         shm_toc_estimate_keys(&pcxt->estimator, 1);
 220
 221         /*
 222          * Normally, the user will have requested at least one worker process, but
 223          * if by chance they have not, we can skip a bunch of things here.
 224          */
 225         if (pcxt->nworkers > 0)
 226         {
 227                 /* Estimate space for various kinds of state sharing. */
 228                 library_len = EstimateLibraryStateSpace();
 229                 shm_toc_estimate_chunk(&pcxt->estimator, library_len);
 230                 guc_len = EstimateGUCStateSpace();
 231                 shm_toc_estimate_chunk(&pcxt->estimator, guc_len);
 232                 combocidlen = EstimateComboCIDStateSpace();
 233                 shm_toc_estimate_chunk(&pcxt->estimator, combocidlen);
 234                 tsnaplen = EstimateSnapshotSpace(transaction_snapshot);
 235                 shm_toc_estimate_chunk(&pcxt->estimator, tsnaplen);
 236                 asnaplen = EstimateSnapshotSpace(active_snapshot);
 237                 shm_toc_estimate_chunk(&pcxt->estimator, asnaplen);
 238                 tstatelen = EstimateTransactionStateSpace();
 239                 shm_toc_estimate_chunk(&pcxt->estimator, tstatelen);
 240                 /* If you add more chunks here, you probably need to add keys. */
 241                 shm_toc_estimate_keys(&pcxt->estimator, 6);
 242
 243                 /* Estimate space need for error queues. */
 244                 StaticAssertStmt(BUFFERALIGN(PARALLEL_ERROR_QUEUE_SIZE) ==
 245                                                  PARALLEL_ERROR_QUEUE_SIZE,
 246                                                  "parallel error queue size not buffer-aligned");
 247                 shm_toc_estimate_chunk(&pcxt->estimator,
 248                                                            mul_size(PARALLEL_ERROR_QUEUE_SIZE,
 249                                                                                 pcxt->nworkers));
 250                 shm_toc_estimate_keys(&pcxt->estimator, 1);
 251
 252                 /* Estimate how much we'll need for extension entrypoint info. */
 253                 if (pcxt->library_name != NULL)
 254                 {
 255                         Assert(pcxt->entrypoint == ParallelExtensionTrampoline);
 256                         Assert(pcxt->function_name != NULL);
 257                         shm_toc_estimate_chunk(&pcxt->estimator, strlen(pcxt->library_name)
 258                                                                    + strlen(pcxt->function_name) + 2);
 259                         shm_toc_estimate_keys(&pcxt->estimator, 1);
 260                 }
 261         }
 262
 263         /*
 264          * Create DSM and initialize with new table of contents.  But if the user
 265          * didn't request any workers, then don't bother creating a dynamic shared
 266          * memory segment; instead, just use backend-private memory.
 267          *
 268          * Also, if we can't create a dynamic shared memory segment because the
 269          * maximum number of segments have already been created, then fall back to
 270          * backend-private memory, and plan not to use any workers.  We hope this
 271          * won't happen very often, but it's better to abandon the use of
 272          * parallelism than to fail outright.
 273          */
 274         segsize = shm_toc_estimate(&pcxt->estimator);
 275         if (pcxt->nworkers > 0)
 276                 pcxt->seg = dsm_create(segsize, DSM_CREATE_NULL_IF_MAXSEGMENTS);
 277         if (pcxt->seg != NULL)
 278                 pcxt->toc = shm_toc_create(PARALLEL_MAGIC,
 279                                                                    dsm_segment_address(pcxt->seg),
 280                                                                    segsize);
 281         else
 282         {
 283                 pcxt->nworkers = 0;
 284                 pcxt->private_memory = MemoryContextAlloc(TopMemoryContext, segsize);
 285                 pcxt->toc = shm_toc_create(PARALLEL_MAGIC, pcxt->private_memory,
 286                                                                    segsize);
 287         }
 288
 289         /* Initialize fixed-size state in shared memory. */
 290         fps = (FixedParallelState *)
 291                 shm_toc_allocate(pcxt->toc, sizeof(FixedParallelState));
 292         fps->database_id = MyDatabaseId;
 293         fps->authenticated_user_id = GetAuthenticatedUserId();
 294         GetUserIdAndSecContext(&fps->current_user_id, &fps->sec_context);
 295         GetTempNamespaceState(&fps->temp_namespace_id,
 296                                                   &fps->temp_toast_namespace_id);
 297         fps->parallel_master_pgproc = MyProc;
 298         fps->parallel_master_pid = MyProcPid;
 299         fps->parallel_master_backend_id = MyBackendId;
 300         fps->entrypoint = pcxt->entrypoint;
 301         SpinLockInit(&fps->mutex);
 302         fps->last_xlog_end = 0;
 303         shm_toc_insert(pcxt->toc, PARALLEL_KEY_FIXED, fps);
 304
 305         /* We can skip the rest of this if we're not budgeting for any workers. */
 306         if (pcxt->nworkers > 0)
 307         {
 308                 char       *libraryspace;
 309                 char       *gucspace;
 310                 char       *combocidspace;
 311                 char       *tsnapspace;
 312                 char       *asnapspace;
 313                 char       *tstatespace;
 314                 char       *error_queue_space;
 315
 316                 /* Serialize shared libraries we have loaded. */
 317                 libraryspace = shm_toc_allocate(pcxt->toc, library_len);
 318                 SerializeLibraryState(library_len, libraryspace);
 319                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_LIBRARY, libraryspace);
 320
 321                 /* Serialize GUC settings. */
 322                 gucspace = shm_toc_allocate(pcxt->toc, guc_len);
 323                 SerializeGUCState(guc_len, gucspace);
 324                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_GUC, gucspace);
 325
 326                 /* Serialize combo CID state. */
 327                 combocidspace = shm_toc_allocate(pcxt->toc, combocidlen);
 328                 SerializeComboCIDState(combocidlen, combocidspace);
 329                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_COMBO_CID, combocidspace);
 330
 331                 /* Serialize transaction snapshot and active snapshot. */
 332                 tsnapspace = shm_toc_allocate(pcxt->toc, tsnaplen);
 333                 SerializeSnapshot(transaction_snapshot, tsnapspace);
 334                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT,
 335                                            tsnapspace);
 336                 asnapspace = shm_toc_allocate(pcxt->toc, asnaplen);
 337                 SerializeSnapshot(active_snapshot, asnapspace);
 338                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ACTIVE_SNAPSHOT, asnapspace);
 339
 340                 /* Serialize transaction state. */
 341                 tstatespace = shm_toc_allocate(pcxt->toc, tstatelen);
 342                 SerializeTransactionState(tstatelen, tstatespace);
 343                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_STATE, tstatespace);
 344
 345                 /* Allocate space for worker information. */
 346                 pcxt->worker = palloc0(sizeof(ParallelWorkerInfo) * pcxt->nworkers);
 347
 348                 /*
 349                  * Establish error queues in dynamic shared memory.
 350                  *
 351                  * These queues should be used only for transmitting ErrorResponse,
 352                  * NoticeResponse, and NotifyResponse protocol messages.  Tuple data
 353                  * should be transmitted via separate (possibly larger?) queues.
 354                  */
 355                 error_queue_space =
 356                         shm_toc_allocate(pcxt->toc,
 357                                                          mul_size(PARALLEL_ERROR_QUEUE_SIZE,
 358                                                                           pcxt->nworkers));
 359                 for (i = 0; i < pcxt->nworkers; ++i)
 360                 {
 361                         char       *start;
 362                         shm_mq     *mq;
 363
 364                         start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
 365                         mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
 366                         shm_mq_set_receiver(mq, MyProc);
 367                         pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
 368                 }
 369                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE, error_queue_space);
 370
 371                 /* Serialize extension entrypoint information. */
 372                 if (pcxt->library_name != NULL)
 373                 {
 374                         Size            lnamelen = strlen(pcxt->library_name);
 375                         char       *extensionstate;
 376
 377                         extensionstate = shm_toc_allocate(pcxt->toc, lnamelen
 378                                                                                   + strlen(pcxt->function_name) + 2);
 379                         strcpy(extensionstate, pcxt->library_name);
 380                         strcpy(extensionstate + lnamelen + 1, pcxt->function_name);
 381                         shm_toc_insert(pcxt->toc, PARALLEL_KEY_EXTENSION_TRAMPOLINE,
 382                                                    extensionstate);
 383                 }
 384         }
 385
 386         /* Restore previous memory context. */
 387         MemoryContextSwitchTo(oldcontext);
 388 }
 389
 390 /*
 391  * Reinitialize the dynamic shared memory segment for a parallel context such
 392  * that we could launch workers for it again.
 393  */
 394 void
 395 ReinitializeParallelDSM(ParallelContext *pcxt)
 396 {
 397         FixedParallelState *fps;
 398         char       *error_queue_space;
 399         int                     i;
 400
 401         /* Wait for any old workers to exit. */
 402         if (pcxt->nworkers_launched > 0)
 403         {
 404                 WaitForParallelWorkersToFinish(pcxt);
 405                 WaitForParallelWorkersToExit(pcxt);
 406                 pcxt->nworkers_launched = 0;
 407         }
 408
 409         /* Reset a few bits of fixed parallel state to a clean state. */
 410         fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED);
 411         fps->last_xlog_end = 0;
 412
 413         /* Recreate error queues. */
 414         error_queue_space =
 415                 shm_toc_lookup(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE);
 416         for (i = 0; i < pcxt->nworkers; ++i)
 417         {
 418                 char       *start;
 419                 shm_mq     *mq;
 420
 421                 start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
 422                 mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
 423                 shm_mq_set_receiver(mq, MyProc);
 424                 pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
 425         }
 426 }
 427
 428 /*
 429  * Launch parallel workers.
 430  */
 431 void
 432 LaunchParallelWorkers(ParallelContext *pcxt)
 433 {
 434         MemoryContext oldcontext;
 435         BackgroundWorker worker;
 436         int                     i;
 437         bool            any_registrations_failed = false;
 438
 439         /* Skip this if we have no workers. */
 440         if (pcxt->nworkers == 0)
 441                 return;
 442
 443         /* We need to be a lock group leader. */
 444         BecomeLockGroupLeader();
 445
 446         /* If we do have workers, we'd better have a DSM segment. */
 447         Assert(pcxt->seg != NULL);
 448
 449         /* We might be running in a short-lived memory context. */
 450         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 451
 452         /* Configure a worker. */
 453         snprintf(worker.bgw_name, BGW_MAXLEN, "parallel worker for PID %d",
 454                          MyProcPid);
 455         worker.bgw_flags =
 456                 BGWORKER_SHMEM_ACCESS | BGWORKER_BACKEND_DATABASE_CONNECTION;
 457         worker.bgw_start_time = BgWorkerStart_ConsistentState;
 458         worker.bgw_restart_time = BGW_NEVER_RESTART;
 459         worker.bgw_main = ParallelWorkerMain;
 460         worker.bgw_main_arg = UInt32GetDatum(dsm_segment_handle(pcxt->seg));
 461         worker.bgw_notify_pid = MyProcPid;
 462         memset(&worker.bgw_extra, 0, BGW_EXTRALEN);
 463
 464         /*
 465          * Start workers.
 466          *
 467          * The caller must be able to tolerate ending up with fewer workers than
 468          * expected, so there is no need to throw an error here if registration
 469          * fails.  It wouldn't help much anyway, because registering the worker in
 470          * no way guarantees that it will start up and initialize successfully.
 471          */
 472         for (i = 0; i < pcxt->nworkers; ++i)
 473         {
 474                 memcpy(worker.bgw_extra, &i, sizeof(int));
 475                 if (!any_registrations_failed &&
 476                         RegisterDynamicBackgroundWorker(&worker,
 477                                                                                         &pcxt->worker[i].bgwhandle))
 478                 {
 479                         shm_mq_set_handle(pcxt->worker[i].error_mqh,
 480                                                           pcxt->worker[i].bgwhandle);
 481                         pcxt->nworkers_launched++;
 482                 }
 483                 else
 484                 {
 485                         /*
 486                          * If we weren't able to register the worker, then we've bumped up
 487                          * against the max_worker_processes limit, and future
 488                          * registrations will probably fail too, so arrange to skip them.
 489                          * But we still have to execute this code for the remaining slots
 490                          * to make sure that we forget about the error queues we budgeted
 491                          * for those workers.  Otherwise, we'll wait for them to start,
 492                          * but they never will.
 493                          */
 494                         any_registrations_failed = true;
 495                         pcxt->worker[i].bgwhandle = NULL;
 496                         pfree(pcxt->worker[i].error_mqh);
 497                         pcxt->worker[i].error_mqh = NULL;
 498                 }
 499         }
 500
 501         /* Restore previous memory context. */
 502         MemoryContextSwitchTo(oldcontext);
 503 }
 504
 505 /*
 506  * Wait for all workers to finish computing.
 507  *
 508  * Even if the parallel operation seems to have completed successfully, it's
 509  * important to call this function afterwards.  We must not miss any errors
 510  * the workers may have thrown during the parallel operation, or any that they
 511  * may yet throw while shutting down.
 512  *
 513  * Also, we want to update our notion of XactLastRecEnd based on worker
 514  * feedback.
 515  */
 516 void
 517 WaitForParallelWorkersToFinish(ParallelContext *pcxt)
 518 {
 519         for (;;)
 520         {
 521                 bool            anyone_alive = false;
 522                 int                     i;
 523
 524                 /*
 525                  * This will process any parallel messages that are pending, which may
 526                  * change the outcome of the loop that follows.  It may also throw an
 527                  * error propagated from a worker.
 528                  */
 529                 CHECK_FOR_INTERRUPTS();
 530
 531                 for (i = 0; i < pcxt->nworkers_launched; ++i)
 532                 {
 533                         if (pcxt->worker[i].error_mqh != NULL)
 534                         {
 535                                 anyone_alive = true;
 536                                 break;
 537                         }
 538                 }
 539
 540                 if (!anyone_alive)
 541                         break;
 542
 543                 WaitLatch(&MyProc->procLatch, WL_LATCH_SET, -1);
 544                 ResetLatch(&MyProc->procLatch);
 545         }
 546
 547         if (pcxt->toc != NULL)
 548         {
 549                 FixedParallelState *fps;
 550
 551                 fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED);
 552                 if (fps->last_xlog_end > XactLastRecEnd)
 553                         XactLastRecEnd = fps->last_xlog_end;
 554         }
 555 }
 556
 557 /*
 558  * Wait for all workers to exit.
 559  *
 560  * This function ensures that workers have been completely shutdown.  The
 561  * difference between WaitForParallelWorkersToFinish and this function is
 562  * that former just ensures that last message sent by worker backend is
 563  * received by master backend whereas this ensures the complete shutdown.
 564  */
 565 static void
 566 WaitForParallelWorkersToExit(ParallelContext *pcxt)
 567 {
 568         int                     i;
 569
 570         /* Wait until the workers actually die. */
 571         for (i = 0; i < pcxt->nworkers_launched; ++i)
 572         {
 573                 BgwHandleStatus status;
 574
 575                 if (pcxt->worker == NULL || pcxt->worker[i].bgwhandle == NULL)
 576                         continue;
 577
 578                 status = WaitForBackgroundWorkerShutdown(pcxt->worker[i].bgwhandle);
 579
 580                 /*
 581                  * If the postmaster kicked the bucket, we have no chance of cleaning
 582                  * up safely -- we won't be able to tell when our workers are actually
 583                  * dead.  This doesn't necessitate a PANIC since they will all abort
 584                  * eventually, but we can't safely continue this session.
 585                  */
 586                 if (status == BGWH_POSTMASTER_DIED)
 587                         ereport(FATAL,
 588                                         (errcode(ERRCODE_ADMIN_SHUTDOWN),
 589                                  errmsg("postmaster exited during a parallel transaction")));
 590
 591                 /* Release memory. */
 592                 pfree(pcxt->worker[i].bgwhandle);
 593                 pcxt->worker[i].bgwhandle = NULL;
 594         }
 595 }
 596
 597 /*
 598  * Destroy a parallel context.
 599  *
 600  * If expecting a clean exit, you should use WaitForParallelWorkersToFinish()
 601  * first, before calling this function.  When this function is invoked, any
 602  * remaining workers are forcibly killed; the dynamic shared memory segment
 603  * is unmapped; and we then wait (uninterruptibly) for the workers to exit.
 604  */
 605 void
 606 DestroyParallelContext(ParallelContext *pcxt)
 607 {
 608         int                     i;
 609
 610         /*
 611          * Be careful about order of operations here!  We remove the parallel
 612          * context from the list before we do anything else; otherwise, if an
 613          * error occurs during a subsequent step, we might try to nuke it again
 614          * from AtEOXact_Parallel or AtEOSubXact_Parallel.
 615          */
 616         dlist_delete(&pcxt->node);
 617
 618         /* Kill each worker in turn, and forget their error queues. */
 619         if (pcxt->worker != NULL)
 620         {
 621                 for (i = 0; i < pcxt->nworkers_launched; ++i)
 622                 {
 623                         if (pcxt->worker[i].error_mqh != NULL)
 624                         {
 625                                 TerminateBackgroundWorker(pcxt->worker[i].bgwhandle);
 626
 627                                 pfree(pcxt->worker[i].error_mqh);
 628                                 pcxt->worker[i].error_mqh = NULL;
 629                         }
 630                 }
 631         }
 632
 633         /*
 634          * If we have allocated a shared memory segment, detach it.  This will
 635          * implicitly detach the error queues, and any other shared memory queues,
 636          * stored there.
 637          */
 638         if (pcxt->seg != NULL)
 639         {
 640                 dsm_detach(pcxt->seg);
 641                 pcxt->seg = NULL;
 642         }
 643
 644         /*
 645          * If this parallel context is actually in backend-private memory rather
 646          * than shared memory, free that memory instead.
 647          */
 648         if (pcxt->private_memory != NULL)
 649         {
 650                 pfree(pcxt->private_memory);
 651                 pcxt->private_memory = NULL;
 652         }
 653
 654         /*
 655          * We can't finish transaction commit or abort until all of the workers
 656          * have exited.  This means, in particular, that we can't respond to
 657          * interrupts at this stage.
 658          */
 659         HOLD_INTERRUPTS();
 660         WaitForParallelWorkersToExit(pcxt);
 661         RESUME_INTERRUPTS();
 662
 663         /* Free the worker array itself. */
 664         if (pcxt->worker != NULL)
 665         {
 666                 pfree(pcxt->worker);
 667                 pcxt->worker = NULL;
 668         }
 669
 670         /* Free memory. */
 671         pfree(pcxt);
 672 }
 673
 674 /*
 675  * Are there any parallel contexts currently active?
 676  */
 677 bool
 678 ParallelContextActive(void)
 679 {
 680         return !dlist_is_empty(&pcxt_list);
 681 }
 682
 683 /*
 684  * Handle receipt of an interrupt indicating a parallel worker message.
 685  *
 686  * Note: this is called within a signal handler!  All we can do is set
 687  * a flag that will cause the next CHECK_FOR_INTERRUPTS() to invoke
 688  * HandleParallelMessages().
 689  */
 690 void
 691 HandleParallelMessageInterrupt(void)
 692 {
 693         InterruptPending = true;
 694         ParallelMessagePending = true;
 695         SetLatch(MyLatch);
 696 }
 697
 698 /*
 699  * Handle any queued protocol messages received from parallel workers.
 700  */
 701 void
 702 HandleParallelMessages(void)
 703 {
 704         dlist_iter      iter;
 705         MemoryContext oldcontext;
 706
 707         static MemoryContext hpm_context = NULL;
 708
 709         /*
 710          * This is invoked from ProcessInterrupts(), and since some of the
 711          * functions it calls contain CHECK_FOR_INTERRUPTS(), there is a potential
 712          * for recursive calls if more signals are received while this runs.  It's
 713          * unclear that recursive entry would be safe, and it doesn't seem useful
 714          * even if it is safe, so let's block interrupts until done.
 715          */
 716         HOLD_INTERRUPTS();
 717
 718         /*
 719          * Moreover, CurrentMemoryContext might be pointing almost anywhere.  We
 720          * don't want to risk leaking data into long-lived contexts, so let's do
 721          * our work here in a private context that we can reset on each use.
 722          */
 723         if (hpm_context == NULL)        /* first time through? */
 724                 hpm_context = AllocSetContextCreate(TopMemoryContext,
 725                                                                                         "HandleParallelMessages context",
 726                                                                                         ALLOCSET_DEFAULT_MINSIZE,
 727                                                                                         ALLOCSET_DEFAULT_INITSIZE,
 728                                                                                         ALLOCSET_DEFAULT_MAXSIZE);
 729         else
 730                 MemoryContextReset(hpm_context);
 731
 732         oldcontext = MemoryContextSwitchTo(hpm_context);
 733
 734         /* OK to process messages.  Reset the flag saying there are more to do. */
 735         ParallelMessagePending = false;
 736
 737         dlist_foreach(iter, &pcxt_list)
 738         {
 739                 ParallelContext *pcxt;
 740                 int                     i;
 741
 742                 pcxt = dlist_container(ParallelContext, node, iter.cur);
 743                 if (pcxt->worker == NULL)
 744                         continue;
 745
 746                 for (i = 0; i < pcxt->nworkers_launched; ++i)
 747                 {
 748                         /*
 749                          * Read as many messages as we can from each worker, but stop when
 750                          * either (1) the worker's error queue goes away, which can happen
 751                          * if we receive a Terminate message from the worker; or (2) no
 752                          * more messages can be read from the worker without blocking.
 753                          */
 754                         while (pcxt->worker[i].error_mqh != NULL)
 755                         {
 756                                 shm_mq_result res;
 757                                 Size            nbytes;
 758                                 void       *data;
 759
 760                                 res = shm_mq_receive(pcxt->worker[i].error_mqh, &nbytes,
 761                                                                          &data, true);
 762                                 if (res == SHM_MQ_WOULD_BLOCK)
 763                                         break;
 764                                 else if (res == SHM_MQ_SUCCESS)
 765                                 {
 766                                         StringInfoData msg;
 767
 768                                         initStringInfo(&msg);
 769                                         appendBinaryStringInfo(&msg, data, nbytes);
 770                                         HandleParallelMessage(pcxt, i, &msg);
 771                                         pfree(msg.data);
 772                                 }
 773                                 else
 774                                         ereport(ERROR,
 775                                                   (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 776                                                    errmsg("lost connection to parallel worker")));
 777                         }
 778                 }
 779         }
 780
 781         MemoryContextSwitchTo(oldcontext);
 782
 783         /* Might as well clear the context on our way out */
 784         MemoryContextReset(hpm_context);
 785
 786         RESUME_INTERRUPTS();
 787 }
 788
 789 /*
 790  * Handle a single protocol message received from a single parallel worker.
 791  */
 792 static void
 793 HandleParallelMessage(ParallelContext *pcxt, int i, StringInfo msg)
 794 {
 795         char            msgtype;
 796
 797         msgtype = pq_getmsgbyte(msg);
 798
 799         switch (msgtype)
 800         {
 801                 case 'K':                               /* BackendKeyData */
 802                         {
 803                                 int32           pid = pq_getmsgint(msg, 4);
 804
 805                                 (void) pq_getmsgint(msg, 4);    /* discard cancel key */
 806                                 (void) pq_getmsgend(msg);
 807                                 pcxt->worker[i].pid = pid;
 808                                 break;
 809                         }
 810
 811                 case 'E':                               /* ErrorResponse */
 812                 case 'N':                               /* NoticeResponse */
 813                         {
 814                                 ErrorData       edata;
 815                                 ErrorContextCallback *save_error_context_stack;
 816
 817                                 /* Parse ErrorResponse or NoticeResponse. */
 818                                 pq_parse_errornotice(msg, &edata);
 819
 820                                 /* Death of a worker isn't enough justification for suicide. */
 821                                 edata.elevel = Min(edata.elevel, ERROR);
 822
 823                                 /*
 824                                  * If desired, add a context line to show that this is a
 825                                  * message propagated from a parallel worker.  Otherwise, it
 826                                  * can sometimes be confusing to understand what actually
 827                                  * happened.  (We don't do this in FORCE_PARALLEL_REGRESS mode
 828                                  * because it causes test-result instability depending on
 829                                  * whether a parallel worker is actually used or not.)
 830                                  */
 831                                 if (force_parallel_mode != FORCE_PARALLEL_REGRESS)
 832                                 {
 833                                         if (edata.context)
 834                                                 edata.context = psprintf("%s\n%s", edata.context,
 835                                                                                                  _("parallel worker"));
 836                                         else
 837                                                 edata.context = pstrdup(_("parallel worker"));
 838                                 }
 839
 840                                 /*
 841                                  * Context beyond that should use the error context callbacks
 842                                  * that were in effect when the ParallelContext was created,
 843                                  * not the current ones.
 844                                  */
 845                                 save_error_context_stack = error_context_stack;
 846                                 error_context_stack = pcxt->error_context_stack;
 847
 848                                 /* Rethrow error or print notice. */
 849                                 ThrowErrorData(&edata);
 850
 851                                 /* Not an error, so restore previous context stack. */
 852                                 error_context_stack = save_error_context_stack;
 853
 854                                 break;
 855                         }
 856
 857                 case 'A':                               /* NotifyResponse */
 858                         {
 859                                 /* Propagate NotifyResponse. */
 860                                 int32           pid;
 861                                 const char *channel;
 862                                 const char *payload;
 863
 864                                 pid = pq_getmsgint(msg, 4);
 865                                 channel = pq_getmsgrawstring(msg);
 866                                 payload = pq_getmsgrawstring(msg);
 867                                 pq_endmessage(msg);
 868
 869                                 NotifyMyFrontEnd(channel, payload, pid);
 870
 871                                 break;
 872                         }
 873
 874                 case 'X':                               /* Terminate, indicating clean exit */
 875                         {
 876                                 pfree(pcxt->worker[i].error_mqh);
 877                                 pcxt->worker[i].error_mqh = NULL;
 878                                 break;
 879                         }
 880
 881                 default:
 882                         {
 883                                 elog(ERROR, "unrecognized message type received from parallel worker: %c (message length %d bytes)",
 884                                          msgtype, msg->len);
 885                         }
 886         }
 887 }
 888
 889 /*
 890  * End-of-subtransaction cleanup for parallel contexts.
 891  *
 892  * Currently, it's forbidden to enter or leave a subtransaction while
 893  * parallel mode is in effect, so we could just blow away everything.  But
 894  * we may want to relax that restriction in the future, so this code
 895  * contemplates that there may be multiple subtransaction IDs in pcxt_list.
 896  */
 897 void
 898 AtEOSubXact_Parallel(bool isCommit, SubTransactionId mySubId)
 899 {
 900         while (!dlist_is_empty(&pcxt_list))
 901         {
 902                 ParallelContext *pcxt;
 903
 904                 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
 905                 if (pcxt->subid != mySubId)
 906                         break;
 907                 if (isCommit)
 908                         elog(WARNING, "leaked parallel context");
 909                 DestroyParallelContext(pcxt);
 910         }
 911 }
 912
 913 /*
 914  * End-of-transaction cleanup for parallel contexts.
 915  */
 916 void
 917 AtEOXact_Parallel(bool isCommit)
 918 {
 919         while (!dlist_is_empty(&pcxt_list))
 920         {
 921                 ParallelContext *pcxt;
 922
 923                 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
 924                 if (isCommit)
 925                         elog(WARNING, "leaked parallel context");
 926                 DestroyParallelContext(pcxt);
 927         }
 928 }
 929
 930 /*
 931  * Main entrypoint for parallel workers.
 932  */
 933 static void
 934 ParallelWorkerMain(Datum main_arg)
 935 {
 936         dsm_segment *seg;
 937         shm_toc    *toc;
 938         FixedParallelState *fps;
 939         char       *error_queue_space;
 940         shm_mq     *mq;
 941         shm_mq_handle *mqh;
 942         char       *libraryspace;
 943         char       *gucspace;
 944         char       *combocidspace;
 945         char       *tsnapspace;
 946         char       *asnapspace;
 947         char       *tstatespace;
 948         StringInfoData msgbuf;
 949
 950         /* Set flag to indicate that we're initializing a parallel worker. */
 951         InitializingParallelWorker = true;
 952
 953         /* Establish signal handlers. */
 954         pqsignal(SIGTERM, die);
 955         BackgroundWorkerUnblockSignals();
 956
 957         /* Determine and set our parallel worker number. */
 958         Assert(ParallelWorkerNumber == -1);
 959         memcpy(&ParallelWorkerNumber, MyBgworkerEntry->bgw_extra, sizeof(int));
 960
 961         /* Set up a memory context and resource owner. */
 962         Assert(CurrentResourceOwner == NULL);
 963         CurrentResourceOwner = ResourceOwnerCreate(NULL, "parallel toplevel");
 964         CurrentMemoryContext = AllocSetContextCreate(TopMemoryContext,
 965                                                                                                  "parallel worker",
 966                                                                                                  ALLOCSET_DEFAULT_MINSIZE,
 967                                                                                                  ALLOCSET_DEFAULT_INITSIZE,
 968                                                                                                  ALLOCSET_DEFAULT_MAXSIZE);
 969
 970         /*
 971          * Now that we have a resource owner, we can attach to the dynamic shared
 972          * memory segment and read the table of contents.
 973          */
 974         seg = dsm_attach(DatumGetUInt32(main_arg));
 975         if (seg == NULL)
 976                 ereport(ERROR,
 977                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 978                                  errmsg("could not map dynamic shared memory segment")));
 979         toc = shm_toc_attach(PARALLEL_MAGIC, dsm_segment_address(seg));
 980         if (toc == NULL)
 981                 ereport(ERROR,
 982                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 983                    errmsg("invalid magic number in dynamic shared memory segment")));
 984
 985         /* Look up fixed parallel state. */
 986         fps = shm_toc_lookup(toc, PARALLEL_KEY_FIXED);
 987         Assert(fps != NULL);
 988         MyFixedParallelState = fps;
 989
 990         /*
 991          * Now that we have a worker number, we can find and attach to the error
 992          * queue provided for us.  That's good, because until we do that, any
 993          * errors that happen here will not be reported back to the process that
 994          * requested that this worker be launched.
 995          */
 996         error_queue_space = shm_toc_lookup(toc, PARALLEL_KEY_ERROR_QUEUE);
 997         mq = (shm_mq *) (error_queue_space +
 998                                          ParallelWorkerNumber * PARALLEL_ERROR_QUEUE_SIZE);
 999         shm_mq_set_sender(mq, MyProc);
1000         mqh = shm_mq_attach(mq, seg, NULL);
1001         pq_redirect_to_shm_mq(seg, mqh);
1002         pq_set_parallel_master(fps->parallel_master_pid,
1003                                                    fps->parallel_master_backend_id);
1004
1005         /*
1006          * Send a BackendKeyData message to the process that initiated parallelism
1007          * so that it has access to our PID before it receives any other messages
1008          * from us.  Our cancel key is sent, too, since that's the way the
1009          * protocol message is defined, but it won't actually be used for anything
1010          * in this case.
1011          */
1012         pq_beginmessage(&msgbuf, 'K');
1013         pq_sendint(&msgbuf, (int32) MyProcPid, sizeof(int32));
1014         pq_sendint(&msgbuf, (int32) MyCancelKey, sizeof(int32));
1015         pq_endmessage(&msgbuf);
1016
1017         /*
1018          * Hooray! Primary initialization is complete.  Now, we need to set up our
1019          * backend-local state to match the original backend.
1020          */
1021
1022         /*
1023          * Join locking group.  We must do this before anything that could try to
1024          * acquire a heavyweight lock, because any heavyweight locks acquired to
1025          * this point could block either directly against the parallel group
1026          * leader or against some process which in turn waits for a lock that
1027          * conflicts with the parallel group leader, causing an undetected
1028          * deadlock.  (If we can't join the lock group, the leader has gone away,
1029          * so just exit quietly.)
1030          */
1031         if (!BecomeLockGroupMember(fps->parallel_master_pgproc,
1032                                                            fps->parallel_master_pid))
1033                 return;
1034
1035         /*
1036          * Load libraries that were loaded by original backend.  We want to do
1037          * this before restoring GUCs, because the libraries might define custom
1038          * variables.
1039          */
1040         libraryspace = shm_toc_lookup(toc, PARALLEL_KEY_LIBRARY);
1041         Assert(libraryspace != NULL);
1042         RestoreLibraryState(libraryspace);
1043
1044         /* Restore database connection. */
1045         BackgroundWorkerInitializeConnectionByOid(fps->database_id,
1046                                                                                           fps->authenticated_user_id);
1047
1048         /*
1049          * Set the client encoding to the database encoding, since that is what
1050          * the leader will expect.
1051          */
1052         SetClientEncoding(GetDatabaseEncoding());
1053
1054         /* Restore GUC values from launching backend. */
1055         gucspace = shm_toc_lookup(toc, PARALLEL_KEY_GUC);
1056         Assert(gucspace != NULL);
1057         StartTransactionCommand();
1058         RestoreGUCState(gucspace);
1059         CommitTransactionCommand();
1060
1061         /* Crank up a transaction state appropriate to a parallel worker. */
1062         tstatespace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_STATE);
1063         StartParallelWorkerTransaction(tstatespace);
1064
1065         /* Restore combo CID state. */
1066         combocidspace = shm_toc_lookup(toc, PARALLEL_KEY_COMBO_CID);
1067         Assert(combocidspace != NULL);
1068         RestoreComboCIDState(combocidspace);
1069
1070         /* Restore transaction snapshot. */
1071         tsnapspace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT);
1072         Assert(tsnapspace != NULL);
1073         RestoreTransactionSnapshot(RestoreSnapshot(tsnapspace),
1074                                                            fps->parallel_master_pgproc);
1075
1076         /* Restore active snapshot. */
1077         asnapspace = shm_toc_lookup(toc, PARALLEL_KEY_ACTIVE_SNAPSHOT);
1078         Assert(asnapspace != NULL);
1079         PushActiveSnapshot(RestoreSnapshot(asnapspace));
1080
1081         /*
1082          * We've changed which tuples we can see, and must therefore invalidate
1083          * system caches.
1084          */
1085         InvalidateSystemCaches();
1086
1087         /* Restore user ID and security context. */
1088         SetUserIdAndSecContext(fps->current_user_id, fps->sec_context);
1089
1090         /* Restore temp-namespace state to ensure search path matches leader's. */
1091         SetTempNamespaceState(fps->temp_namespace_id,
1092                                                   fps->temp_toast_namespace_id);
1093
1094         /* Set ParallelMasterBackendId so we know how to address temp relations. */
1095         ParallelMasterBackendId = fps->parallel_master_backend_id;
1096
1097         /*
1098          * We've initialized all of our state now; nothing should change
1099          * hereafter.
1100          */
1101         InitializingParallelWorker = false;
1102         EnterParallelMode();
1103
1104         /*
1105          * Time to do the real work: invoke the caller-supplied code.
1106          *
1107          * If you get a crash at this line, see the comments for
1108          * ParallelExtensionTrampoline.
1109          */
1110         fps->entrypoint(seg, toc);
1111
1112         /* Must exit parallel mode to pop active snapshot. */
1113         ExitParallelMode();
1114
1115         /* Must pop active snapshot so resowner.c doesn't complain. */
1116         PopActiveSnapshot();
1117
1118         /* Shut down the parallel-worker transaction. */
1119         EndParallelWorkerTransaction();
1120
1121         /* Report success. */
1122         pq_putmessage('X', NULL, 0);
1123 }
1124
1125 /*
1126  * It's unsafe for the entrypoint invoked by ParallelWorkerMain to be a
1127  * function living in a dynamically loaded module, because the module might
1128  * not be loaded in every process, or might be loaded but not at the same
1129  * address.  To work around that problem, CreateParallelContextForExtension()
1130  * arranges to call this function rather than calling the extension-provided
1131  * function directly; and this function then looks up the real entrypoint and
1132  * calls it.
1133  */
1134 static void
1135 ParallelExtensionTrampoline(dsm_segment *seg, shm_toc *toc)
1136 {
1137         char       *extensionstate;
1138         char       *library_name;
1139         char       *function_name;
1140         parallel_worker_main_type entrypt;
1141
1142         extensionstate = shm_toc_lookup(toc, PARALLEL_KEY_EXTENSION_TRAMPOLINE);
1143         Assert(extensionstate != NULL);
1144         library_name = extensionstate;
1145         function_name = extensionstate + strlen(library_name) + 1;
1146
1147         entrypt = (parallel_worker_main_type)
1148                 load_external_function(library_name, function_name, true, NULL);
1149         entrypt(seg, toc);
1150 }
1151
1152 /*
1153  * Update shared memory with the ending location of the last WAL record we
1154  * wrote, if it's greater than the value already stored there.
1155  */
1156 void
1157 ParallelWorkerReportLastRecEnd(XLogRecPtr last_xlog_end)
1158 {
1159         FixedParallelState *fps = MyFixedParallelState;
1160
1161         Assert(fps != NULL);
1162         SpinLockAcquire(&fps->mutex);
1163         if (fps->last_xlog_end < last_xlog_end)
1164                 fps->last_xlog_end = last_xlog_end;
1165         SpinLockRelease(&fps->mutex);
1166 }