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 ParallelErrorContext(void *arg);
 112 static void ParallelExtensionTrampoline(dsm_segment *seg, shm_toc *toc);
 113 static void ParallelWorkerMain(Datum main_arg);
 114 static void WaitForParallelWorkersToExit(ParallelContext *pcxt);
 115
 116
 117 /*
 118  * Establish a new parallel context.  This should be done after entering
 119  * parallel mode, and (unless there is an error) the context should be
 120  * destroyed before exiting the current subtransaction.
 121  */
 122 ParallelContext *
 123 CreateParallelContext(parallel_worker_main_type entrypoint, int nworkers)
 124 {
 125         MemoryContext oldcontext;
 126         ParallelContext *pcxt;
 127
 128         /* It is unsafe to create a parallel context if not in parallel mode. */
 129         Assert(IsInParallelMode());
 130
 131         /* Number of workers should be non-negative. */
 132         Assert(nworkers >= 0);
 133
 134         /*
 135          * If dynamic shared memory is not available, we won't be able to use
 136          * background workers.
 137          */
 138         if (dynamic_shared_memory_type == DSM_IMPL_NONE)
 139                 nworkers = 0;
 140
 141         /*
 142          * If we are running under serializable isolation, we can't use parallel
 143          * workers, at least not until somebody enhances that mechanism to be
 144          * parallel-aware.
 145          */
 146         if (IsolationIsSerializable())
 147                 nworkers = 0;
 148
 149         /* We might be running in a short-lived memory context. */
 150         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 151
 152         /* Initialize a new ParallelContext. */
 153         pcxt = palloc0(sizeof(ParallelContext));
 154         pcxt->subid = GetCurrentSubTransactionId();
 155         pcxt->nworkers = nworkers;
 156         pcxt->entrypoint = entrypoint;
 157         pcxt->error_context_stack = error_context_stack;
 158         shm_toc_initialize_estimator(&pcxt->estimator);
 159         dlist_push_head(&pcxt_list, &pcxt->node);
 160
 161         /* Restore previous memory context. */
 162         MemoryContextSwitchTo(oldcontext);
 163
 164         return pcxt;
 165 }
 166
 167 /*
 168  * Establish a new parallel context that calls a function provided by an
 169  * extension.  This works around the fact that the library might get mapped
 170  * at a different address in each backend.
 171  */
 172 ParallelContext *
 173 CreateParallelContextForExternalFunction(char *library_name,
 174                                                                                  char *function_name,
 175                                                                                  int nworkers)
 176 {
 177         MemoryContext oldcontext;
 178         ParallelContext *pcxt;
 179
 180         /* We might be running in a very short-lived memory context. */
 181         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 182
 183         /* Create the context. */
 184         pcxt = CreateParallelContext(ParallelExtensionTrampoline, nworkers);
 185         pcxt->library_name = pstrdup(library_name);
 186         pcxt->function_name = pstrdup(function_name);
 187
 188         /* Restore previous memory context. */
 189         MemoryContextSwitchTo(oldcontext);
 190
 191         return pcxt;
 192 }
 193
 194 /*
 195  * Establish the dynamic shared memory segment for a parallel context and
 196  * copy state and other bookkeeping information that will be needed by
 197  * parallel workers into it.
 198  */
 199 void
 200 InitializeParallelDSM(ParallelContext *pcxt)
 201 {
 202         MemoryContext oldcontext;
 203         Size            library_len = 0;
 204         Size            guc_len = 0;
 205         Size            combocidlen = 0;
 206         Size            tsnaplen = 0;
 207         Size            asnaplen = 0;
 208         Size            tstatelen = 0;
 209         Size            segsize = 0;
 210         int                     i;
 211         FixedParallelState *fps;
 212         Snapshot        transaction_snapshot = GetTransactionSnapshot();
 213         Snapshot        active_snapshot = GetActiveSnapshot();
 214
 215         /* We might be running in a very short-lived memory context. */
 216         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 217
 218         /* Allow space to store the fixed-size parallel state. */
 219         shm_toc_estimate_chunk(&pcxt->estimator, sizeof(FixedParallelState));
 220         shm_toc_estimate_keys(&pcxt->estimator, 1);
 221
 222         /*
 223          * Normally, the user will have requested at least one worker process, but
 224          * if by chance they have not, we can skip a bunch of things here.
 225          */
 226         if (pcxt->nworkers > 0)
 227         {
 228                 /* Estimate space for various kinds of state sharing. */
 229                 library_len = EstimateLibraryStateSpace();
 230                 shm_toc_estimate_chunk(&pcxt->estimator, library_len);
 231                 guc_len = EstimateGUCStateSpace();
 232                 shm_toc_estimate_chunk(&pcxt->estimator, guc_len);
 233                 combocidlen = EstimateComboCIDStateSpace();
 234                 shm_toc_estimate_chunk(&pcxt->estimator, combocidlen);
 235                 tsnaplen = EstimateSnapshotSpace(transaction_snapshot);
 236                 shm_toc_estimate_chunk(&pcxt->estimator, tsnaplen);
 237                 asnaplen = EstimateSnapshotSpace(active_snapshot);
 238                 shm_toc_estimate_chunk(&pcxt->estimator, asnaplen);
 239                 tstatelen = EstimateTransactionStateSpace();
 240                 shm_toc_estimate_chunk(&pcxt->estimator, tstatelen);
 241                 /* If you add more chunks here, you probably need to add keys. */
 242                 shm_toc_estimate_keys(&pcxt->estimator, 6);
 243
 244                 /* Estimate space need for error queues. */
 245                 StaticAssertStmt(BUFFERALIGN(PARALLEL_ERROR_QUEUE_SIZE) ==
 246                                                  PARALLEL_ERROR_QUEUE_SIZE,
 247                                                  "parallel error queue size not buffer-aligned");
 248                 shm_toc_estimate_chunk(&pcxt->estimator,
 249                                                            mul_size(PARALLEL_ERROR_QUEUE_SIZE,
 250                                                                                 pcxt->nworkers));
 251                 shm_toc_estimate_keys(&pcxt->estimator, 1);
 252
 253                 /* Estimate how much we'll need for extension entrypoint info. */
 254                 if (pcxt->library_name != NULL)
 255                 {
 256                         Assert(pcxt->entrypoint == ParallelExtensionTrampoline);
 257                         Assert(pcxt->function_name != NULL);
 258                         shm_toc_estimate_chunk(&pcxt->estimator, strlen(pcxt->library_name)
 259                                                                    + strlen(pcxt->function_name) + 2);
 260                         shm_toc_estimate_keys(&pcxt->estimator, 1);
 261                 }
 262         }
 263
 264         /*
 265          * Create DSM and initialize with new table of contents.  But if the user
 266          * didn't request any workers, then don't bother creating a dynamic shared
 267          * memory segment; instead, just use backend-private memory.
 268          *
 269          * Also, if we can't create a dynamic shared memory segment because the
 270          * maximum number of segments have already been created, then fall back to
 271          * backend-private memory, and plan not to use any workers.  We hope this
 272          * won't happen very often, but it's better to abandon the use of
 273          * parallelism than to fail outright.
 274          */
 275         segsize = shm_toc_estimate(&pcxt->estimator);
 276         if (pcxt->nworkers > 0)
 277                 pcxt->seg = dsm_create(segsize, DSM_CREATE_NULL_IF_MAXSEGMENTS);
 278         if (pcxt->seg != NULL)
 279                 pcxt->toc = shm_toc_create(PARALLEL_MAGIC,
 280                                                                    dsm_segment_address(pcxt->seg),
 281                                                                    segsize);
 282         else
 283         {
 284                 pcxt->nworkers = 0;
 285                 pcxt->private_memory = MemoryContextAlloc(TopMemoryContext, segsize);
 286                 pcxt->toc = shm_toc_create(PARALLEL_MAGIC, pcxt->private_memory,
 287                                                                    segsize);
 288         }
 289
 290         /* Initialize fixed-size state in shared memory. */
 291         fps = (FixedParallelState *)
 292                 shm_toc_allocate(pcxt->toc, sizeof(FixedParallelState));
 293         fps->database_id = MyDatabaseId;
 294         fps->authenticated_user_id = GetAuthenticatedUserId();
 295         GetUserIdAndSecContext(&fps->current_user_id, &fps->sec_context);
 296         GetTempNamespaceState(&fps->temp_namespace_id,
 297                                                   &fps->temp_toast_namespace_id);
 298         fps->parallel_master_pgproc = MyProc;
 299         fps->parallel_master_pid = MyProcPid;
 300         fps->parallel_master_backend_id = MyBackendId;
 301         fps->entrypoint = pcxt->entrypoint;
 302         SpinLockInit(&fps->mutex);
 303         fps->last_xlog_end = 0;
 304         shm_toc_insert(pcxt->toc, PARALLEL_KEY_FIXED, fps);
 305
 306         /* We can skip the rest of this if we're not budgeting for any workers. */
 307         if (pcxt->nworkers > 0)
 308         {
 309                 char       *libraryspace;
 310                 char       *gucspace;
 311                 char       *combocidspace;
 312                 char       *tsnapspace;
 313                 char       *asnapspace;
 314                 char       *tstatespace;
 315                 char       *error_queue_space;
 316
 317                 /* Serialize shared libraries we have loaded. */
 318                 libraryspace = shm_toc_allocate(pcxt->toc, library_len);
 319                 SerializeLibraryState(library_len, libraryspace);
 320                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_LIBRARY, libraryspace);
 321
 322                 /* Serialize GUC settings. */
 323                 gucspace = shm_toc_allocate(pcxt->toc, guc_len);
 324                 SerializeGUCState(guc_len, gucspace);
 325                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_GUC, gucspace);
 326
 327                 /* Serialize combo CID state. */
 328                 combocidspace = shm_toc_allocate(pcxt->toc, combocidlen);
 329                 SerializeComboCIDState(combocidlen, combocidspace);
 330                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_COMBO_CID, combocidspace);
 331
 332                 /* Serialize transaction snapshot and active snapshot. */
 333                 tsnapspace = shm_toc_allocate(pcxt->toc, tsnaplen);
 334                 SerializeSnapshot(transaction_snapshot, tsnapspace);
 335                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT,
 336                                            tsnapspace);
 337                 asnapspace = shm_toc_allocate(pcxt->toc, asnaplen);
 338                 SerializeSnapshot(active_snapshot, asnapspace);
 339                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ACTIVE_SNAPSHOT, asnapspace);
 340
 341                 /* Serialize transaction state. */
 342                 tstatespace = shm_toc_allocate(pcxt->toc, tstatelen);
 343                 SerializeTransactionState(tstatelen, tstatespace);
 344                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_STATE, tstatespace);
 345
 346                 /* Allocate space for worker information. */
 347                 pcxt->worker = palloc0(sizeof(ParallelWorkerInfo) * pcxt->nworkers);
 348
 349                 /*
 350                  * Establish error queues in dynamic shared memory.
 351                  *
 352                  * These queues should be used only for transmitting ErrorResponse,
 353                  * NoticeResponse, and NotifyResponse protocol messages.  Tuple data
 354                  * should be transmitted via separate (possibly larger?) queues.
 355                  */
 356                 error_queue_space =
 357                         shm_toc_allocate(pcxt->toc,
 358                                                          mul_size(PARALLEL_ERROR_QUEUE_SIZE,
 359                                                                           pcxt->nworkers));
 360                 for (i = 0; i < pcxt->nworkers; ++i)
 361                 {
 362                         char       *start;
 363                         shm_mq     *mq;
 364
 365                         start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
 366                         mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
 367                         shm_mq_set_receiver(mq, MyProc);
 368                         pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
 369                 }
 370                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE, error_queue_space);
 371
 372                 /* Serialize extension entrypoint information. */
 373                 if (pcxt->library_name != NULL)
 374                 {
 375                         Size            lnamelen = strlen(pcxt->library_name);
 376                         char       *extensionstate;
 377
 378                         extensionstate = shm_toc_allocate(pcxt->toc, lnamelen
 379                                                                                   + strlen(pcxt->function_name) + 2);
 380                         strcpy(extensionstate, pcxt->library_name);
 381                         strcpy(extensionstate + lnamelen + 1, pcxt->function_name);
 382                         shm_toc_insert(pcxt->toc, PARALLEL_KEY_EXTENSION_TRAMPOLINE,
 383                                                    extensionstate);
 384                 }
 385         }
 386
 387         /* Restore previous memory context. */
 388         MemoryContextSwitchTo(oldcontext);
 389 }
 390
 391 /*
 392  * Reinitialize the dynamic shared memory segment for a parallel context such
 393  * that we could launch workers for it again.
 394  */
 395 void
 396 ReinitializeParallelDSM(ParallelContext *pcxt)
 397 {
 398         FixedParallelState *fps;
 399         char       *error_queue_space;
 400         int                     i;
 401
 402         /* Wait for any old workers to exit. */
 403         if (pcxt->nworkers_launched > 0)
 404         {
 405                 WaitForParallelWorkersToFinish(pcxt);
 406                 WaitForParallelWorkersToExit(pcxt);
 407                 pcxt->nworkers_launched = 0;
 408         }
 409
 410         /* Reset a few bits of fixed parallel state to a clean state. */
 411         fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED);
 412         fps->last_xlog_end = 0;
 413
 414         /* Recreate error queues. */
 415         error_queue_space =
 416                 shm_toc_lookup(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE);
 417         for (i = 0; i < pcxt->nworkers; ++i)
 418         {
 419                 char       *start;
 420                 shm_mq     *mq;
 421
 422                 start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
 423                 mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
 424                 shm_mq_set_receiver(mq, MyProc);
 425                 pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
 426         }
 427 }
 428
 429 /*
 430  * Launch parallel workers.
 431  */
 432 void
 433 LaunchParallelWorkers(ParallelContext *pcxt)
 434 {
 435         MemoryContext oldcontext;
 436         BackgroundWorker worker;
 437         int                     i;
 438         bool            any_registrations_failed = false;
 439
 440         /* Skip this if we have no workers. */
 441         if (pcxt->nworkers == 0)
 442                 return;
 443
 444         /* We need to be a lock group leader. */
 445         BecomeLockGroupLeader();
 446
 447         /* If we do have workers, we'd better have a DSM segment. */
 448         Assert(pcxt->seg != NULL);
 449
 450         /* We might be running in a short-lived memory context. */
 451         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 452
 453         /* Configure a worker. */
 454         snprintf(worker.bgw_name, BGW_MAXLEN, "parallel worker for PID %d",
 455                          MyProcPid);
 456         worker.bgw_flags =
 457                 BGWORKER_SHMEM_ACCESS | BGWORKER_BACKEND_DATABASE_CONNECTION;
 458         worker.bgw_start_time = BgWorkerStart_ConsistentState;
 459         worker.bgw_restart_time = BGW_NEVER_RESTART;
 460         worker.bgw_main = ParallelWorkerMain;
 461         worker.bgw_main_arg = UInt32GetDatum(dsm_segment_handle(pcxt->seg));
 462         worker.bgw_notify_pid = MyProcPid;
 463         memset(&worker.bgw_extra, 0, BGW_EXTRALEN);
 464
 465         /*
 466          * Start workers.
 467          *
 468          * The caller must be able to tolerate ending up with fewer workers than
 469          * expected, so there is no need to throw an error here if registration
 470          * fails.  It wouldn't help much anyway, because registering the worker in
 471          * no way guarantees that it will start up and initialize successfully.
 472          */
 473         for (i = 0; i < pcxt->nworkers; ++i)
 474         {
 475                 memcpy(worker.bgw_extra, &i, sizeof(int));
 476                 if (!any_registrations_failed &&
 477                         RegisterDynamicBackgroundWorker(&worker,
 478                                                                                         &pcxt->worker[i].bgwhandle))
 479                 {
 480                         shm_mq_set_handle(pcxt->worker[i].error_mqh,
 481                                                           pcxt->worker[i].bgwhandle);
 482                         pcxt->nworkers_launched++;
 483                 }
 484                 else
 485                 {
 486                         /*
 487                          * If we weren't able to register the worker, then we've bumped up
 488                          * against the max_worker_processes limit, and future
 489                          * registrations will probably fail too, so arrange to skip them.
 490                          * But we still have to execute this code for the remaining slots
 491                          * to make sure that we forget about the error queues we budgeted
 492                          * for those workers.  Otherwise, we'll wait for them to start,
 493                          * but they never will.
 494                          */
 495                         any_registrations_failed = true;
 496                         pcxt->worker[i].bgwhandle = NULL;
 497                         pfree(pcxt->worker[i].error_mqh);
 498                         pcxt->worker[i].error_mqh = NULL;
 499                 }
 500         }
 501
 502         /* Restore previous memory context. */
 503         MemoryContextSwitchTo(oldcontext);
 504 }
 505
 506 /*
 507  * Wait for all workers to finish computing.
 508  *
 509  * Even if the parallel operation seems to have completed successfully, it's
 510  * important to call this function afterwards.  We must not miss any errors
 511  * the workers may have thrown during the parallel operation, or any that they
 512  * may yet throw while shutting down.
 513  *
 514  * Also, we want to update our notion of XactLastRecEnd based on worker
 515  * feedback.
 516  */
 517 void
 518 WaitForParallelWorkersToFinish(ParallelContext *pcxt)
 519 {
 520         for (;;)
 521         {
 522                 bool            anyone_alive = false;
 523                 int                     i;
 524
 525                 /*
 526                  * This will process any parallel messages that are pending, which may
 527                  * change the outcome of the loop that follows.  It may also throw an
 528                  * error propagated from a worker.
 529                  */
 530                 CHECK_FOR_INTERRUPTS();
 531
 532                 for (i = 0; i < pcxt->nworkers_launched; ++i)
 533                 {
 534                         if (pcxt->worker[i].error_mqh != NULL)
 535                         {
 536                                 anyone_alive = true;
 537                                 break;
 538                         }
 539                 }
 540
 541                 if (!anyone_alive)
 542                         break;
 543
 544                 WaitLatch(&MyProc->procLatch, WL_LATCH_SET, -1);
 545                 ResetLatch(&MyProc->procLatch);
 546         }
 547
 548         if (pcxt->toc != NULL)
 549         {
 550                 FixedParallelState *fps;
 551
 552                 fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED);
 553                 if (fps->last_xlog_end > XactLastRecEnd)
 554                         XactLastRecEnd = fps->last_xlog_end;
 555         }
 556 }
 557
 558 /*
 559  * Wait for all workers to exit.
 560  *
 561  * This function ensures that workers have been completely shutdown.  The
 562  * difference between WaitForParallelWorkersToFinish and this function is
 563  * that former just ensures that last message sent by worker backend is
 564  * received by master backend whereas this ensures the complete shutdown.
 565  */
 566 static void
 567 WaitForParallelWorkersToExit(ParallelContext *pcxt)
 568 {
 569         int                     i;
 570
 571         /* Wait until the workers actually die. */
 572         for (i = 0; i < pcxt->nworkers_launched; ++i)
 573         {
 574                 BgwHandleStatus status;
 575
 576                 if (pcxt->worker == NULL || pcxt->worker[i].bgwhandle == NULL)
 577                         continue;
 578
 579                 status = WaitForBackgroundWorkerShutdown(pcxt->worker[i].bgwhandle);
 580
 581                 /*
 582                  * If the postmaster kicked the bucket, we have no chance of cleaning
 583                  * up safely -- we won't be able to tell when our workers are actually
 584                  * dead.  This doesn't necessitate a PANIC since they will all abort
 585                  * eventually, but we can't safely continue this session.
 586                  */
 587                 if (status == BGWH_POSTMASTER_DIED)
 588                         ereport(FATAL,
 589                                         (errcode(ERRCODE_ADMIN_SHUTDOWN),
 590                                  errmsg("postmaster exited during a parallel transaction")));
 591
 592                 /* Release memory. */
 593                 pfree(pcxt->worker[i].bgwhandle);
 594                 pcxt->worker[i].bgwhandle = NULL;
 595         }
 596 }
 597
 598 /*
 599  * Destroy a parallel context.
 600  *
 601  * If expecting a clean exit, you should use WaitForParallelWorkersToFinish()
 602  * first, before calling this function.  When this function is invoked, any
 603  * remaining workers are forcibly killed; the dynamic shared memory segment
 604  * is unmapped; and we then wait (uninterruptibly) for the workers to exit.
 605  */
 606 void
 607 DestroyParallelContext(ParallelContext *pcxt)
 608 {
 609         int                     i;
 610
 611         /*
 612          * Be careful about order of operations here!  We remove the parallel
 613          * context from the list before we do anything else; otherwise, if an
 614          * error occurs during a subsequent step, we might try to nuke it again
 615          * from AtEOXact_Parallel or AtEOSubXact_Parallel.
 616          */
 617         dlist_delete(&pcxt->node);
 618
 619         /* Kill each worker in turn, and forget their error queues. */
 620         if (pcxt->worker != NULL)
 621         {
 622                 for (i = 0; i < pcxt->nworkers_launched; ++i)
 623                 {
 624                         if (pcxt->worker[i].error_mqh != NULL)
 625                         {
 626                                 TerminateBackgroundWorker(pcxt->worker[i].bgwhandle);
 627
 628                                 pfree(pcxt->worker[i].error_mqh);
 629                                 pcxt->worker[i].error_mqh = NULL;
 630                         }
 631                 }
 632         }
 633
 634         /*
 635          * If we have allocated a shared memory segment, detach it.  This will
 636          * implicitly detach the error queues, and any other shared memory queues,
 637          * stored there.
 638          */
 639         if (pcxt->seg != NULL)
 640         {
 641                 dsm_detach(pcxt->seg);
 642                 pcxt->seg = NULL;
 643         }
 644
 645         /*
 646          * If this parallel context is actually in backend-private memory rather
 647          * than shared memory, free that memory instead.
 648          */
 649         if (pcxt->private_memory != NULL)
 650         {
 651                 pfree(pcxt->private_memory);
 652                 pcxt->private_memory = NULL;
 653         }
 654
 655         /*
 656          * We can't finish transaction commit or abort until all of the workers
 657          * have exited.  This means, in particular, that we can't respond to
 658          * interrupts at this stage.
 659          */
 660         HOLD_INTERRUPTS();
 661         WaitForParallelWorkersToExit(pcxt);
 662         RESUME_INTERRUPTS();
 663
 664         /* Free the worker array itself. */
 665         if (pcxt->worker != NULL)
 666         {
 667                 pfree(pcxt->worker);
 668                 pcxt->worker = NULL;
 669         }
 670
 671         /* Free memory. */
 672         pfree(pcxt);
 673 }
 674
 675 /*
 676  * Are there any parallel contexts currently active?
 677  */
 678 bool
 679 ParallelContextActive(void)
 680 {
 681         return !dlist_is_empty(&pcxt_list);
 682 }
 683
 684 /*
 685  * Handle receipt of an interrupt indicating a parallel worker message.
 686  *
 687  * Note: this is called within a signal handler!  All we can do is set
 688  * a flag that will cause the next CHECK_FOR_INTERRUPTS() to invoke
 689  * HandleParallelMessages().
 690  */
 691 void
 692 HandleParallelMessageInterrupt(void)
 693 {
 694         InterruptPending = true;
 695         ParallelMessagePending = true;
 696         SetLatch(MyLatch);
 697 }
 698
 699 /*
 700  * Handle any queued protocol messages received from parallel workers.
 701  */
 702 void
 703 HandleParallelMessages(void)
 704 {
 705         dlist_iter      iter;
 706
 707         ParallelMessagePending = false;
 708
 709         dlist_foreach(iter, &pcxt_list)
 710         {
 711                 ParallelContext *pcxt;
 712                 int                     i;
 713                 Size            nbytes;
 714                 void       *data;
 715
 716                 pcxt = dlist_container(ParallelContext, node, iter.cur);
 717                 if (pcxt->worker == NULL)
 718                         continue;
 719
 720                 for (i = 0; i < pcxt->nworkers_launched; ++i)
 721                 {
 722                         /*
 723                          * Read as many messages as we can from each worker, but stop when
 724                          * either (1) the error queue goes away, which can happen if we
 725                          * receive a Terminate message from the worker; or (2) no more
 726                          * messages can be read from the worker without blocking.
 727                          */
 728                         while (pcxt->worker[i].error_mqh != NULL)
 729                         {
 730                                 shm_mq_result res;
 731
 732                                 res = shm_mq_receive(pcxt->worker[i].error_mqh, &nbytes,
 733                                                                          &data, true);
 734                                 if (res == SHM_MQ_WOULD_BLOCK)
 735                                         break;
 736                                 else if (res == SHM_MQ_SUCCESS)
 737                                 {
 738                                         StringInfoData msg;
 739
 740                                         initStringInfo(&msg);
 741                                         appendBinaryStringInfo(&msg, data, nbytes);
 742                                         HandleParallelMessage(pcxt, i, &msg);
 743                                         pfree(msg.data);
 744                                 }
 745                                 else
 746                                         ereport(ERROR,
 747                                                   (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 748                                                    errmsg("lost connection to parallel worker")));
 749                         }
 750                 }
 751         }
 752 }
 753
 754 /*
 755  * Handle a single protocol message received from a single parallel worker.
 756  */
 757 static void
 758 HandleParallelMessage(ParallelContext *pcxt, int i, StringInfo msg)
 759 {
 760         char            msgtype;
 761
 762         msgtype = pq_getmsgbyte(msg);
 763
 764         switch (msgtype)
 765         {
 766                 case 'K':                               /* BackendKeyData */
 767                         {
 768                                 int32           pid = pq_getmsgint(msg, 4);
 769
 770                                 (void) pq_getmsgint(msg, 4);    /* discard cancel key */
 771                                 (void) pq_getmsgend(msg);
 772                                 pcxt->worker[i].pid = pid;
 773                                 break;
 774                         }
 775
 776                 case 'E':                               /* ErrorResponse */
 777                 case 'N':                               /* NoticeResponse */
 778                         {
 779                                 ErrorData       edata;
 780                                 ErrorContextCallback errctx;
 781                                 ErrorContextCallback *save_error_context_stack;
 782
 783                                 /*
 784                                  * Rethrow the error using the error context callbacks that
 785                                  * were in effect when the context was created, not the
 786                                  * current ones.
 787                                  */
 788                                 save_error_context_stack = error_context_stack;
 789                                 errctx.callback = ParallelErrorContext;
 790                                 errctx.arg = NULL;
 791                                 errctx.previous = pcxt->error_context_stack;
 792                                 error_context_stack = &errctx;
 793
 794                                 /* Parse ErrorResponse or NoticeResponse. */
 795                                 pq_parse_errornotice(msg, &edata);
 796
 797                                 /* Death of a worker isn't enough justification for suicide. */
 798                                 edata.elevel = Min(edata.elevel, ERROR);
 799
 800                                 /* Rethrow error or notice. */
 801                                 ThrowErrorData(&edata);
 802
 803                                 /* Restore previous context. */
 804                                 error_context_stack = save_error_context_stack;
 805
 806                                 break;
 807                         }
 808
 809                 case 'A':                               /* NotifyResponse */
 810                         {
 811                                 /* Propagate NotifyResponse. */
 812                                 int32           pid;
 813                                 const char *channel;
 814                                 const char *payload;
 815
 816                                 pid = pq_getmsgint(msg, 4);
 817                                 channel = pq_getmsgrawstring(msg);
 818                                 payload = pq_getmsgrawstring(msg);
 819                                 pq_endmessage(msg);
 820
 821                                 NotifyMyFrontEnd(channel, payload, pid);
 822
 823                                 break;
 824                         }
 825
 826                 case 'X':                               /* Terminate, indicating clean exit */
 827                         {
 828                                 pfree(pcxt->worker[i].error_mqh);
 829                                 pcxt->worker[i].error_mqh = NULL;
 830                                 break;
 831                         }
 832
 833                 default:
 834                         {
 835                                 elog(ERROR, "unrecognized message type received from parallel worker: %c (message length %d bytes)",
 836                                          msgtype, msg->len);
 837                         }
 838         }
 839 }
 840
 841 /*
 842  * End-of-subtransaction cleanup for parallel contexts.
 843  *
 844  * Currently, it's forbidden to enter or leave a subtransaction while
 845  * parallel mode is in effect, so we could just blow away everything.  But
 846  * we may want to relax that restriction in the future, so this code
 847  * contemplates that there may be multiple subtransaction IDs in pcxt_list.
 848  */
 849 void
 850 AtEOSubXact_Parallel(bool isCommit, SubTransactionId mySubId)
 851 {
 852         while (!dlist_is_empty(&pcxt_list))
 853         {
 854                 ParallelContext *pcxt;
 855
 856                 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
 857                 if (pcxt->subid != mySubId)
 858                         break;
 859                 if (isCommit)
 860                         elog(WARNING, "leaked parallel context");
 861                 DestroyParallelContext(pcxt);
 862         }
 863 }
 864
 865 /*
 866  * End-of-transaction cleanup for parallel contexts.
 867  */
 868 void
 869 AtEOXact_Parallel(bool isCommit)
 870 {
 871         while (!dlist_is_empty(&pcxt_list))
 872         {
 873                 ParallelContext *pcxt;
 874
 875                 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
 876                 if (isCommit)
 877                         elog(WARNING, "leaked parallel context");
 878                 DestroyParallelContext(pcxt);
 879         }
 880 }
 881
 882 /*
 883  * Main entrypoint for parallel workers.
 884  */
 885 static void
 886 ParallelWorkerMain(Datum main_arg)
 887 {
 888         dsm_segment *seg;
 889         shm_toc    *toc;
 890         FixedParallelState *fps;
 891         char       *error_queue_space;
 892         shm_mq     *mq;
 893         shm_mq_handle *mqh;
 894         char       *libraryspace;
 895         char       *gucspace;
 896         char       *combocidspace;
 897         char       *tsnapspace;
 898         char       *asnapspace;
 899         char       *tstatespace;
 900         StringInfoData msgbuf;
 901
 902         /* Set flag to indicate that we're initializing a parallel worker. */
 903         InitializingParallelWorker = true;
 904
 905         /* Establish signal handlers. */
 906         pqsignal(SIGTERM, die);
 907         BackgroundWorkerUnblockSignals();
 908
 909         /* Determine and set our parallel worker number. */
 910         Assert(ParallelWorkerNumber == -1);
 911         memcpy(&ParallelWorkerNumber, MyBgworkerEntry->bgw_extra, sizeof(int));
 912
 913         /* Set up a memory context and resource owner. */
 914         Assert(CurrentResourceOwner == NULL);
 915         CurrentResourceOwner = ResourceOwnerCreate(NULL, "parallel toplevel");
 916         CurrentMemoryContext = AllocSetContextCreate(TopMemoryContext,
 917                                                                                                  "parallel worker",
 918                                                                                                  ALLOCSET_DEFAULT_MINSIZE,
 919                                                                                                  ALLOCSET_DEFAULT_INITSIZE,
 920                                                                                                  ALLOCSET_DEFAULT_MAXSIZE);
 921
 922         /*
 923          * Now that we have a resource owner, we can attach to the dynamic shared
 924          * memory segment and read the table of contents.
 925          */
 926         seg = dsm_attach(DatumGetUInt32(main_arg));
 927         if (seg == NULL)
 928                 ereport(ERROR,
 929                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 930                                  errmsg("could not map dynamic shared memory segment")));
 931         toc = shm_toc_attach(PARALLEL_MAGIC, dsm_segment_address(seg));
 932         if (toc == NULL)
 933                 ereport(ERROR,
 934                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 935                    errmsg("invalid magic number in dynamic shared memory segment")));
 936
 937         /* Look up fixed parallel state. */
 938         fps = shm_toc_lookup(toc, PARALLEL_KEY_FIXED);
 939         Assert(fps != NULL);
 940         MyFixedParallelState = fps;
 941
 942         /*
 943          * Now that we have a worker number, we can find and attach to the error
 944          * queue provided for us.  That's good, because until we do that, any
 945          * errors that happen here will not be reported back to the process that
 946          * requested that this worker be launched.
 947          */
 948         error_queue_space = shm_toc_lookup(toc, PARALLEL_KEY_ERROR_QUEUE);
 949         mq = (shm_mq *) (error_queue_space +
 950                                          ParallelWorkerNumber * PARALLEL_ERROR_QUEUE_SIZE);
 951         shm_mq_set_sender(mq, MyProc);
 952         mqh = shm_mq_attach(mq, seg, NULL);
 953         pq_redirect_to_shm_mq(seg, mqh);
 954         pq_set_parallel_master(fps->parallel_master_pid,
 955                                                    fps->parallel_master_backend_id);
 956
 957         /*
 958          * Send a BackendKeyData message to the process that initiated parallelism
 959          * so that it has access to our PID before it receives any other messages
 960          * from us.  Our cancel key is sent, too, since that's the way the
 961          * protocol message is defined, but it won't actually be used for anything
 962          * in this case.
 963          */
 964         pq_beginmessage(&msgbuf, 'K');
 965         pq_sendint(&msgbuf, (int32) MyProcPid, sizeof(int32));
 966         pq_sendint(&msgbuf, (int32) MyCancelKey, sizeof(int32));
 967         pq_endmessage(&msgbuf);
 968
 969         /*
 970          * Hooray! Primary initialization is complete.  Now, we need to set up our
 971          * backend-local state to match the original backend.
 972          */
 973
 974         /*
 975          * Join locking group.  We must do this before anything that could try to
 976          * acquire a heavyweight lock, because any heavyweight locks acquired to
 977          * this point could block either directly against the parallel group
 978          * leader or against some process which in turn waits for a lock that
 979          * conflicts with the parallel group leader, causing an undetected
 980          * deadlock.  (If we can't join the lock group, the leader has gone away,
 981          * so just exit quietly.)
 982          */
 983         if (!BecomeLockGroupMember(fps->parallel_master_pgproc,
 984                                                            fps->parallel_master_pid))
 985                 return;
 986
 987         /*
 988          * Load libraries that were loaded by original backend.  We want to do
 989          * this before restoring GUCs, because the libraries might define custom
 990          * variables.
 991          */
 992         libraryspace = shm_toc_lookup(toc, PARALLEL_KEY_LIBRARY);
 993         Assert(libraryspace != NULL);
 994         RestoreLibraryState(libraryspace);
 995
 996         /* Restore database connection. */
 997         BackgroundWorkerInitializeConnectionByOid(fps->database_id,
 998                                                                                           fps->authenticated_user_id);
 999
1000         /*
1001          * Set the client encoding to the database encoding, since that is what
1002          * the leader will expect.
1003          */
1004         SetClientEncoding(GetDatabaseEncoding());
1005
1006         /* Restore GUC values from launching backend. */
1007         gucspace = shm_toc_lookup(toc, PARALLEL_KEY_GUC);
1008         Assert(gucspace != NULL);
1009         StartTransactionCommand();
1010         RestoreGUCState(gucspace);
1011         CommitTransactionCommand();
1012
1013         /* Crank up a transaction state appropriate to a parallel worker. */
1014         tstatespace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_STATE);
1015         StartParallelWorkerTransaction(tstatespace);
1016
1017         /* Restore combo CID state. */
1018         combocidspace = shm_toc_lookup(toc, PARALLEL_KEY_COMBO_CID);
1019         Assert(combocidspace != NULL);
1020         RestoreComboCIDState(combocidspace);
1021
1022         /* Restore transaction snapshot. */
1023         tsnapspace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT);
1024         Assert(tsnapspace != NULL);
1025         RestoreTransactionSnapshot(RestoreSnapshot(tsnapspace),
1026                                                            fps->parallel_master_pgproc);
1027
1028         /* Restore active snapshot. */
1029         asnapspace = shm_toc_lookup(toc, PARALLEL_KEY_ACTIVE_SNAPSHOT);
1030         Assert(asnapspace != NULL);
1031         PushActiveSnapshot(RestoreSnapshot(asnapspace));
1032
1033         /*
1034          * We've changed which tuples we can see, and must therefore invalidate
1035          * system caches.
1036          */
1037         InvalidateSystemCaches();
1038
1039         /* Restore user ID and security context. */
1040         SetUserIdAndSecContext(fps->current_user_id, fps->sec_context);
1041
1042         /* Restore temp-namespace state to ensure search path matches leader's. */
1043         SetTempNamespaceState(fps->temp_namespace_id,
1044                                                   fps->temp_toast_namespace_id);
1045
1046         /* Set ParallelMasterBackendId so we know how to address temp relations. */
1047         ParallelMasterBackendId = fps->parallel_master_backend_id;
1048
1049         /*
1050          * We've initialized all of our state now; nothing should change
1051          * hereafter.
1052          */
1053         InitializingParallelWorker = false;
1054         EnterParallelMode();
1055
1056         /*
1057          * Time to do the real work: invoke the caller-supplied code.
1058          *
1059          * If you get a crash at this line, see the comments for
1060          * ParallelExtensionTrampoline.
1061          */
1062         fps->entrypoint(seg, toc);
1063
1064         /* Must exit parallel mode to pop active snapshot. */
1065         ExitParallelMode();
1066
1067         /* Must pop active snapshot so resowner.c doesn't complain. */
1068         PopActiveSnapshot();
1069
1070         /* Shut down the parallel-worker transaction. */
1071         EndParallelWorkerTransaction();
1072
1073         /* Report success. */
1074         pq_putmessage('X', NULL, 0);
1075 }
1076
1077 /*
1078  * It's unsafe for the entrypoint invoked by ParallelWorkerMain to be a
1079  * function living in a dynamically loaded module, because the module might
1080  * not be loaded in every process, or might be loaded but not at the same
1081  * address.  To work around that problem, CreateParallelContextForExtension()
1082  * arranges to call this function rather than calling the extension-provided
1083  * function directly; and this function then looks up the real entrypoint and
1084  * calls it.
1085  */
1086 static void
1087 ParallelExtensionTrampoline(dsm_segment *seg, shm_toc *toc)
1088 {
1089         char       *extensionstate;
1090         char       *library_name;
1091         char       *function_name;
1092         parallel_worker_main_type entrypt;
1093
1094         extensionstate = shm_toc_lookup(toc, PARALLEL_KEY_EXTENSION_TRAMPOLINE);
1095         Assert(extensionstate != NULL);
1096         library_name = extensionstate;
1097         function_name = extensionstate + strlen(library_name) + 1;
1098
1099         entrypt = (parallel_worker_main_type)
1100                 load_external_function(library_name, function_name, true, NULL);
1101         entrypt(seg, toc);
1102 }
1103
1104 /*
1105  * Give the user a hint that this is a message propagated from a parallel
1106  * worker.  Otherwise, it can sometimes be confusing to understand what
1107  * actually happened.
1108  */
1109 static void
1110 ParallelErrorContext(void *arg)
1111 {
1112         if (force_parallel_mode != FORCE_PARALLEL_REGRESS)
1113                 errcontext("parallel worker");
1114 }
1115
1116 /*
1117  * Update shared memory with the ending location of the last WAL record we
1118  * wrote, if it's greater than the value already stored there.
1119  */
1120 void
1121 ParallelWorkerReportLastRecEnd(XLogRecPtr last_xlog_end)
1122 {
1123         FixedParallelState *fps = MyFixedParallelState;
1124
1125         Assert(fps != NULL);
1126         SpinLockAcquire(&fps->mutex);
1127         if (fps->last_xlog_end < last_xlog_end)
1128                 fps->last_xlog_end = last_xlog_end;
1129         SpinLockRelease(&fps->mutex);
1130 }