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

   1 /*-------------------------------------------------------------------------
   2  *
   3  * parallel.c
   4  *        Infrastructure for launching parallel workers
   5  *
   6  * Portions Copyright (c) 1996-2014, PostgreSQL Global Development Group
   7  * Portions Copyright (c) 1994, Regents of the University of California
   8  *
   9  * IDENTIFICATION
  10  *        src/backend/access/transam/parallel.c
  11  *
  12  *-------------------------------------------------------------------------
  13  */
  14
  15 #include "postgres.h"
  16
  17 #include "access/xact.h"
  18 #include "access/xlog.h"
  19 #include "access/parallel.h"
  20 #include "commands/async.h"
  21 #include "libpq/libpq.h"
  22 #include "libpq/pqformat.h"
  23 #include "libpq/pqmq.h"
  24 #include "miscadmin.h"
  25 #include "storage/ipc.h"
  26 #include "storage/sinval.h"
  27 #include "storage/spin.h"
  28 #include "tcop/tcopprot.h"
  29 #include "utils/combocid.h"
  30 #include "utils/guc.h"
  31 #include "utils/inval.h"
  32 #include "utils/memutils.h"
  33 #include "utils/resowner.h"
  34 #include "utils/snapmgr.h"
  35
  36 /*
  37  * We don't want to waste a lot of memory on an error queue which, most of
  38  * the time, will process only a handful of small messages.  However, it is
  39  * desirable to make it large enough that a typical ErrorResponse can be sent
  40  * without blocking.  That way, a worker that errors out can write the whole
  41  * message into the queue and terminate without waiting for the user backend.
  42  */
  43 #define PARALLEL_ERROR_QUEUE_SIZE                       16384
  44
  45 /* Magic number for parallel context TOC. */
  46 #define PARALLEL_MAGIC                                          0x50477c7c
  47
  48 /*
  49  * Magic numbers for parallel state sharing.  Higher-level code should use
  50  * smaller values, leaving these very large ones for use by this module.
  51  */
  52 #define PARALLEL_KEY_FIXED                                      UINT64CONST(0xFFFFFFFFFFFF0001)
  53 #define PARALLEL_KEY_ERROR_QUEUE                        UINT64CONST(0xFFFFFFFFFFFF0002)
  54 #define PARALLEL_KEY_LIBRARY                            UINT64CONST(0xFFFFFFFFFFFF0003)
  55 #define PARALLEL_KEY_GUC                                        UINT64CONST(0xFFFFFFFFFFFF0004)
  56 #define PARALLEL_KEY_COMBO_CID                          UINT64CONST(0xFFFFFFFFFFFF0005)
  57 #define PARALLEL_KEY_TRANSACTION_SNAPSHOT       UINT64CONST(0xFFFFFFFFFFFF0006)
  58 #define PARALLEL_KEY_ACTIVE_SNAPSHOT            UINT64CONST(0xFFFFFFFFFFFF0007)
  59 #define PARALLEL_KEY_TRANSACTION_STATE          UINT64CONST(0xFFFFFFFFFFFF0008)
  60 #define PARALLEL_KEY_EXTENSION_TRAMPOLINE       UINT64CONST(0xFFFFFFFFFFFF0009)
  61
  62 /* Fixed-size parallel state. */
  63 typedef struct FixedParallelState
  64 {
  65         /* Fixed-size state that workers must restore. */
  66         Oid                     database_id;
  67         Oid                     authenticated_user_id;
  68         Oid                     current_user_id;
  69         int                     sec_context;
  70         PGPROC     *parallel_master_pgproc;
  71         pid_t           parallel_master_pid;
  72         BackendId       parallel_master_backend_id;
  73
  74         /* Entrypoint for parallel workers. */
  75         parallel_worker_main_type entrypoint;
  76
  77         /* Mutex protects remaining fields. */
  78         slock_t         mutex;
  79
  80         /* Track whether workers have attached. */
  81         int                     workers_expected;
  82         int                     workers_attached;
  83
  84         /* Maximum XactLastRecEnd of any worker. */
  85         XLogRecPtr      last_xlog_end;
  86 } FixedParallelState;
  87
  88 /*
  89  * Our parallel worker number.  We initialize this to -1, meaning that we are
  90  * not a parallel worker.  In parallel workers, it will be set to a value >= 0
  91  * and < the number of workers before any user code is invoked; each parallel
  92  * worker will get a different parallel worker number.
  93  */
  94 int                     ParallelWorkerNumber = -1;
  95
  96 /* Is there a parallel message pending which we need to receive? */
  97 bool            ParallelMessagePending = false;
  98
  99 /* Are we initializing a parallel worker? */
 100 bool            InitializingParallelWorker = false;
 101
 102 /* Pointer to our fixed parallel state. */
 103 static FixedParallelState *MyFixedParallelState;
 104
 105 /* List of active parallel contexts. */
 106 static dlist_head pcxt_list = DLIST_STATIC_INIT(pcxt_list);
 107
 108 /* Private functions. */
 109 static void HandleParallelMessage(ParallelContext *, int, StringInfo msg);
 110 static void ParallelErrorContext(void *arg);
 111 static void ParallelExtensionTrampoline(dsm_segment *seg, shm_toc *toc);
 112 static void ParallelWorkerMain(Datum main_arg);
 113 static void WaitForParallelWorkersToExit(ParallelContext *pcxt);
 114
 115 /*
 116  * Establish a new parallel context.  This should be done after entering
 117  * parallel mode, and (unless there is an error) the context should be
 118  * destroyed before exiting the current subtransaction.
 119  */
 120 ParallelContext *
 121 CreateParallelContext(parallel_worker_main_type entrypoint, int nworkers)
 122 {
 123         MemoryContext oldcontext;
 124         ParallelContext *pcxt;
 125
 126         /* It is unsafe to create a parallel context if not in parallel mode. */
 127         Assert(IsInParallelMode());
 128
 129         /* Number of workers should be non-negative. */
 130         Assert(nworkers >= 0);
 131
 132         /*
 133          * If dynamic shared memory is not available, we won't be able to use
 134          * background workers.
 135          */
 136         if (dynamic_shared_memory_type == DSM_IMPL_NONE)
 137                 nworkers = 0;
 138
 139         /*
 140          * If we are running under serializable isolation, we can't use
 141          * parallel workers, at least not until somebody enhances that mechanism
 142          * to be parallel-aware.
 143          */
 144         if (IsolationIsSerializable())
 145                 nworkers = 0;
 146
 147         /* We might be running in a short-lived memory context. */
 148         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 149
 150         /* Initialize a new ParallelContext. */
 151         pcxt = palloc0(sizeof(ParallelContext));
 152         pcxt->subid = GetCurrentSubTransactionId();
 153         pcxt->nworkers = nworkers;
 154         pcxt->entrypoint = entrypoint;
 155         pcxt->error_context_stack = error_context_stack;
 156         shm_toc_initialize_estimator(&pcxt->estimator);
 157         dlist_push_head(&pcxt_list, &pcxt->node);
 158
 159         /* Restore previous memory context. */
 160         MemoryContextSwitchTo(oldcontext);
 161
 162         return pcxt;
 163 }
 164
 165 /*
 166  * Establish a new parallel context that calls a function provided by an
 167  * extension.  This works around the fact that the library might get mapped
 168  * at a different address in each backend.
 169  */
 170 ParallelContext *
 171 CreateParallelContextForExternalFunction(char *library_name,
 172                                                                                  char *function_name,
 173                                                                                  int nworkers)
 174 {
 175         MemoryContext oldcontext;
 176         ParallelContext *pcxt;
 177
 178         /* We might be running in a very short-lived memory context. */
 179         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 180
 181         /* Create the context. */
 182         pcxt = CreateParallelContext(ParallelExtensionTrampoline, nworkers);
 183         pcxt->library_name = pstrdup(library_name);
 184         pcxt->function_name = pstrdup(function_name);
 185
 186         /* Restore previous memory context. */
 187         MemoryContextSwitchTo(oldcontext);
 188
 189         return pcxt;
 190 }
 191
 192 /*
 193  * Establish the dynamic shared memory segment for a parallel context and
 194  * copied state and other bookkeeping information that will need by parallel
 195  * workers into it.
 196  */
 197 void
 198 InitializeParallelDSM(ParallelContext *pcxt)
 199 {
 200         MemoryContext oldcontext;
 201         Size            library_len = 0;
 202         Size            guc_len = 0;
 203         Size            combocidlen = 0;
 204         Size            tsnaplen = 0;
 205         Size            asnaplen = 0;
 206         Size            tstatelen = 0;
 207         Size            segsize = 0;
 208         int                     i;
 209         FixedParallelState *fps;
 210         Snapshot        transaction_snapshot = GetTransactionSnapshot();
 211         Snapshot        active_snapshot = GetActiveSnapshot();
 212
 213         /* We might be running in a very short-lived memory context. */
 214         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 215
 216         /* Allow space to store the fixed-size parallel state. */
 217         shm_toc_estimate_chunk(&pcxt->estimator, sizeof(FixedParallelState));
 218         shm_toc_estimate_keys(&pcxt->estimator, 1);
 219
 220         /*
 221          * Normally, the user will have requested at least one worker process, but
 222          * if by chance they have not, we can skip a bunch of things here.
 223          */
 224         if (pcxt->nworkers > 0)
 225         {
 226                 /* Estimate space for various kinds of state sharing. */
 227                 library_len = EstimateLibraryStateSpace();
 228                 shm_toc_estimate_chunk(&pcxt->estimator, library_len);
 229                 guc_len = EstimateGUCStateSpace();
 230                 shm_toc_estimate_chunk(&pcxt->estimator, guc_len);
 231                 combocidlen = EstimateComboCIDStateSpace();
 232                 shm_toc_estimate_chunk(&pcxt->estimator, combocidlen);
 233                 tsnaplen = EstimateSnapshotSpace(transaction_snapshot);
 234                 shm_toc_estimate_chunk(&pcxt->estimator, tsnaplen);
 235                 asnaplen = EstimateSnapshotSpace(active_snapshot);
 236                 shm_toc_estimate_chunk(&pcxt->estimator, asnaplen);
 237                 tstatelen = EstimateTransactionStateSpace();
 238                 shm_toc_estimate_chunk(&pcxt->estimator, tstatelen);
 239                 /* If you add more chunks here, you probably need to add keys. */
 240                 shm_toc_estimate_keys(&pcxt->estimator, 6);
 241
 242                 /* Estimate space need for error queues. */
 243                 StaticAssertStmt(BUFFERALIGN(PARALLEL_ERROR_QUEUE_SIZE) ==
 244                                                  PARALLEL_ERROR_QUEUE_SIZE,
 245                                                  "parallel error queue size not buffer-aligned");
 246                 shm_toc_estimate_chunk(&pcxt->estimator,
 247                                                            PARALLEL_ERROR_QUEUE_SIZE * pcxt->nworkers);
 248                 shm_toc_estimate_keys(&pcxt->estimator, 1);
 249
 250                 /* Estimate how much we'll need for extension entrypoint info. */
 251                 if (pcxt->library_name != NULL)
 252                 {
 253                         Assert(pcxt->entrypoint == ParallelExtensionTrampoline);
 254                         Assert(pcxt->function_name != NULL);
 255                         shm_toc_estimate_chunk(&pcxt->estimator, strlen(pcxt->library_name)
 256                                                                    + strlen(pcxt->function_name) + 2);
 257                         shm_toc_estimate_keys(&pcxt->estimator, 1);
 258                 }
 259         }
 260
 261         /*
 262          * Create DSM and initialize with new table of contents.  But if the user
 263          * didn't request any workers, then don't bother creating a dynamic shared
 264          * memory segment; instead, just use backend-private memory.
 265          *
 266          * Also, if we can't create a dynamic shared memory segment because the
 267          * maximum number of segments have already been created, then fall back to
 268          * backend-private memory, and plan not to use any workers.  We hope this
 269          * won't happen very often, but it's better to abandon the use of
 270          * parallelism than to fail outright.
 271          */
 272         segsize = shm_toc_estimate(&pcxt->estimator);
 273         if (pcxt->nworkers != 0)
 274                 pcxt->seg = dsm_create(segsize, DSM_CREATE_NULL_IF_MAXSEGMENTS);
 275         if (pcxt->seg != NULL)
 276                 pcxt->toc = shm_toc_create(PARALLEL_MAGIC,
 277                                                                    dsm_segment_address(pcxt->seg),
 278                                                                    segsize);
 279         else
 280         {
 281                 pcxt->nworkers = 0;
 282                 pcxt->private_memory = MemoryContextAlloc(TopMemoryContext, segsize);
 283                 pcxt->toc = shm_toc_create(PARALLEL_MAGIC, pcxt->private_memory,
 284                                                                    segsize);
 285         }
 286
 287         /* Initialize fixed-size state in shared memory. */
 288         fps = (FixedParallelState *)
 289                 shm_toc_allocate(pcxt->toc, sizeof(FixedParallelState));
 290         fps->database_id = MyDatabaseId;
 291         fps->authenticated_user_id = GetAuthenticatedUserId();
 292         GetUserIdAndSecContext(&fps->current_user_id, &fps->sec_context);
 293         fps->parallel_master_pgproc = MyProc;
 294         fps->parallel_master_pid = MyProcPid;
 295         fps->parallel_master_backend_id = MyBackendId;
 296         fps->entrypoint = pcxt->entrypoint;
 297         SpinLockInit(&fps->mutex);
 298         fps->workers_expected = pcxt->nworkers;
 299         fps->workers_attached = 0;
 300         fps->last_xlog_end = 0;
 301         shm_toc_insert(pcxt->toc, PARALLEL_KEY_FIXED, fps);
 302
 303         /* We can skip the rest of this if we're not budgeting for any workers. */
 304         if (pcxt->nworkers > 0)
 305         {
 306                 char       *libraryspace;
 307                 char       *gucspace;
 308                 char       *combocidspace;
 309                 char       *tsnapspace;
 310                 char       *asnapspace;
 311                 char       *tstatespace;
 312                 char       *error_queue_space;
 313
 314                 /* Serialize shared libraries we have loaded. */
 315                 libraryspace = shm_toc_allocate(pcxt->toc, library_len);
 316                 SerializeLibraryState(library_len, libraryspace);
 317                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_LIBRARY, libraryspace);
 318
 319                 /* Serialize GUC settings. */
 320                 gucspace = shm_toc_allocate(pcxt->toc, guc_len);
 321                 SerializeGUCState(guc_len, gucspace);
 322                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_GUC, gucspace);
 323
 324                 /* Serialize combo CID state. */
 325                 combocidspace = shm_toc_allocate(pcxt->toc, combocidlen);
 326                 SerializeComboCIDState(combocidlen, combocidspace);
 327                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_COMBO_CID, combocidspace);
 328
 329                 /* Serialize transaction snapshot and active snapshot. */
 330                 tsnapspace = shm_toc_allocate(pcxt->toc, tsnaplen);
 331                 SerializeSnapshot(transaction_snapshot, tsnapspace);
 332                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT,
 333                                            tsnapspace);
 334                 asnapspace = shm_toc_allocate(pcxt->toc, asnaplen);
 335                 SerializeSnapshot(active_snapshot, asnapspace);
 336                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ACTIVE_SNAPSHOT, asnapspace);
 337
 338                 /* Serialize transaction state. */
 339                 tstatespace = shm_toc_allocate(pcxt->toc, tstatelen);
 340                 SerializeTransactionState(tstatelen, tstatespace);
 341                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_STATE, tstatespace);
 342
 343                 /* Allocate space for worker information. */
 344                 pcxt->worker = palloc0(sizeof(ParallelWorkerInfo) * pcxt->nworkers);
 345
 346                 /*
 347                  * Establish error queues in dynamic shared memory.
 348                  *
 349                  * These queues should be used only for transmitting ErrorResponse,
 350                  * NoticeResponse, and NotifyResponse protocol messages.  Tuple data
 351                  * should be transmitted via separate (possibly larger?) queues.
 352                  */
 353                 error_queue_space =
 354                         shm_toc_allocate(pcxt->toc,
 355                                                          PARALLEL_ERROR_QUEUE_SIZE * pcxt->nworkers);
 356                 for (i = 0; i < pcxt->nworkers; ++i)
 357                 {
 358                         char       *start;
 359                         shm_mq     *mq;
 360
 361                         start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
 362                         mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
 363                         shm_mq_set_receiver(mq, MyProc);
 364                         pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
 365                 }
 366                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE, error_queue_space);
 367
 368                 /* Serialize extension entrypoint information. */
 369                 if (pcxt->library_name != NULL)
 370                 {
 371                         Size            lnamelen = strlen(pcxt->library_name);
 372                         char       *extensionstate;
 373
 374                         extensionstate = shm_toc_allocate(pcxt->toc, lnamelen
 375                                                                                   + strlen(pcxt->function_name) + 2);
 376                         strcpy(extensionstate, pcxt->library_name);
 377                         strcpy(extensionstate + lnamelen + 1, pcxt->function_name);
 378                         shm_toc_insert(pcxt->toc, PARALLEL_KEY_EXTENSION_TRAMPOLINE,
 379                                                    extensionstate);
 380                 }
 381         }
 382
 383         /* Restore previous memory context. */
 384         MemoryContextSwitchTo(oldcontext);
 385 }
 386
 387 /*
 388  * Reinitialize the dynamic shared memory segment for a parallel context such
 389  * that we could launch workers for it again.
 390  */
 391 void
 392 ReinitializeParallelDSM(ParallelContext *pcxt)
 393 {
 394         FixedParallelState *fps;
 395         char       *error_queue_space;
 396         int                     i;
 397
 398         if (pcxt->nworkers_launched == 0)
 399                 return;
 400
 401         WaitForParallelWorkersToFinish(pcxt);
 402         WaitForParallelWorkersToExit(pcxt);
 403
 404         /* Reset a few bits of fixed parallel state to a clean state. */
 405         fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED);
 406         fps->workers_attached = 0;
 407         fps->last_xlog_end = 0;
 408
 409         /* Recreate error queues. */
 410         error_queue_space =
 411                 shm_toc_lookup(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE);
 412         for (i = 0; i < pcxt->nworkers; ++i)
 413         {
 414                 char       *start;
 415                 shm_mq     *mq;
 416
 417                 start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
 418                 mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
 419                 shm_mq_set_receiver(mq, MyProc);
 420                 pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
 421         }
 422
 423         /* Reset number of workers launched. */
 424         pcxt->nworkers_launched = 0;
 425 }
 426
 427 /*
 428  * Launch parallel workers.
 429  */
 430 void
 431 LaunchParallelWorkers(ParallelContext *pcxt)
 432 {
 433         MemoryContext oldcontext;
 434         BackgroundWorker worker;
 435         int                     i;
 436         bool            any_registrations_failed = false;
 437
 438         /* Skip this if we have no workers. */
 439         if (pcxt->nworkers == 0)
 440                 return;
 441
 442         /* If we do have workers, we'd better have a DSM segment. */
 443         Assert(pcxt->seg != NULL);
 444
 445         /* We might be running in a short-lived memory context. */
 446         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 447
 448         /* Configure a worker. */
 449         snprintf(worker.bgw_name, BGW_MAXLEN, "parallel worker for PID %d",
 450                          MyProcPid);
 451         worker.bgw_flags =
 452                 BGWORKER_SHMEM_ACCESS | BGWORKER_BACKEND_DATABASE_CONNECTION;
 453         worker.bgw_start_time = BgWorkerStart_ConsistentState;
 454         worker.bgw_restart_time = BGW_NEVER_RESTART;
 455         worker.bgw_main = ParallelWorkerMain;
 456         worker.bgw_main_arg = UInt32GetDatum(dsm_segment_handle(pcxt->seg));
 457         worker.bgw_notify_pid = MyProcPid;
 458
 459         /*
 460          * Start workers.
 461          *
 462          * The caller must be able to tolerate ending up with fewer workers than
 463          * expected, so there is no need to throw an error here if registration
 464          * fails.  It wouldn't help much anyway, because registering the worker in
 465          * no way guarantees that it will start up and initialize successfully.
 466          */
 467         for (i = 0; i < pcxt->nworkers; ++i)
 468         {
 469                 if (!any_registrations_failed &&
 470                         RegisterDynamicBackgroundWorker(&worker,
 471                                                                                         &pcxt->worker[i].bgwhandle))
 472                 {
 473                         shm_mq_set_handle(pcxt->worker[i].error_mqh,
 474                                                           pcxt->worker[i].bgwhandle);
 475                         pcxt->nworkers_launched++;
 476                 }
 477                 else
 478                 {
 479                         /*
 480                          * If we weren't able to register the worker, then we've bumped up
 481                          * against the max_worker_processes limit, and future
 482                          * registrations will probably fail too, so arrange to skip them.
 483                          * But we still have to execute this code for the remaining slots
 484                          * to make sure that we forget about the error queues we budgeted
 485                          * for those workers.  Otherwise, we'll wait for them to start,
 486                          * but they never will.
 487                          */
 488                         any_registrations_failed = true;
 489                         pcxt->worker[i].bgwhandle = NULL;
 490                         pcxt->worker[i].error_mqh = NULL;
 491                 }
 492         }
 493
 494         /* Restore previous memory context. */
 495         MemoryContextSwitchTo(oldcontext);
 496 }
 497
 498 /*
 499  * Wait for all workers to finish computing.
 500  *
 501  * Even if the parallel operation seems to have completed successfully, it's
 502  * important to call this function afterwards.  We must not miss any errors
 503  * the workers may have thrown during the parallel operation, or any that they
 504  * may yet throw while shutting down.
 505  *
 506  * Also, we want to update our notion of XactLastRecEnd based on worker
 507  * feedback.
 508  */
 509 void
 510 WaitForParallelWorkersToFinish(ParallelContext *pcxt)
 511 {
 512         for (;;)
 513         {
 514                 bool            anyone_alive = false;
 515                 int                     i;
 516
 517                 /*
 518                  * This will process any parallel messages that are pending, which may
 519                  * change the outcome of the loop that follows.  It may also throw an
 520                  * error propagated from a worker.
 521                  */
 522                 CHECK_FOR_INTERRUPTS();
 523
 524                 for (i = 0; i < pcxt->nworkers; ++i)
 525                 {
 526                         if (pcxt->worker[i].error_mqh != NULL)
 527                         {
 528                                 anyone_alive = true;
 529                                 break;
 530                         }
 531                 }
 532
 533                 if (!anyone_alive)
 534                         break;
 535
 536                 WaitLatch(&MyProc->procLatch, WL_LATCH_SET, -1);
 537                 ResetLatch(&MyProc->procLatch);
 538         }
 539
 540         if (pcxt->toc != NULL)
 541         {
 542                 FixedParallelState *fps;
 543
 544                 fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED);
 545                 if (fps->last_xlog_end > XactLastRecEnd)
 546                         XactLastRecEnd = fps->last_xlog_end;
 547         }
 548 }
 549
 550 /*
 551  * Wait for all workers to exit.
 552  *
 553  * This function ensures that workers have been completely shutdown.  The
 554  * difference between WaitForParallelWorkersToFinish and this function is
 555  * that former just ensures that last message sent by worker backend is
 556  * received by master backend whereas this ensures the complete shutdown.
 557  */
 558 static void
 559 WaitForParallelWorkersToExit(ParallelContext *pcxt)
 560 {
 561         int                     i;
 562
 563         /* Wait until the workers actually die. */
 564         for (i = 0; i < pcxt->nworkers; ++i)
 565         {
 566                 BgwHandleStatus status;
 567
 568                 if (pcxt->worker == NULL || pcxt->worker[i].bgwhandle == NULL)
 569                         continue;
 570
 571                 status = WaitForBackgroundWorkerShutdown(pcxt->worker[i].bgwhandle);
 572
 573                 /*
 574                  * If the postmaster kicked the bucket, we have no chance of cleaning
 575                  * up safely -- we won't be able to tell when our workers are actually
 576                  * dead.  This doesn't necessitate a PANIC since they will all abort
 577                  * eventually, but we can't safely continue this session.
 578                  */
 579                 if (status == BGWH_POSTMASTER_DIED)
 580                         ereport(FATAL,
 581                                         (errcode(ERRCODE_ADMIN_SHUTDOWN),
 582                                  errmsg("postmaster exited during a parallel transaction")));
 583
 584                 /* Release memory. */
 585                 pfree(pcxt->worker[i].bgwhandle);
 586                 pcxt->worker[i].bgwhandle = NULL;
 587         }
 588 }
 589
 590 /*
 591  * Destroy a parallel context.
 592  *
 593  * If expecting a clean exit, you should use WaitForParallelWorkersToFinish()
 594  * first, before calling this function.  When this function is invoked, any
 595  * remaining workers are forcibly killed; the dynamic shared memory segment
 596  * is unmapped; and we then wait (uninterruptibly) for the workers to exit.
 597  */
 598 void
 599 DestroyParallelContext(ParallelContext *pcxt)
 600 {
 601         int                     i;
 602
 603         /*
 604          * Be careful about order of operations here!  We remove the parallel
 605          * context from the list before we do anything else; otherwise, if an
 606          * error occurs during a subsequent step, we might try to nuke it again
 607          * from AtEOXact_Parallel or AtEOSubXact_Parallel.
 608          */
 609         dlist_delete(&pcxt->node);
 610
 611         /* Kill each worker in turn, and forget their error queues. */
 612         if (pcxt->worker != NULL)
 613         {
 614                 for (i = 0; i < pcxt->nworkers; ++i)
 615                 {
 616                         if (pcxt->worker[i].error_mqh != NULL)
 617                         {
 618                                 TerminateBackgroundWorker(pcxt->worker[i].bgwhandle);
 619
 620                                 pfree(pcxt->worker[i].error_mqh);
 621                                 pcxt->worker[i].error_mqh = NULL;
 622                         }
 623                 }
 624         }
 625
 626         /*
 627          * If we have allocated a shared memory segment, detach it.  This will
 628          * implicitly detach the error queues, and any other shared memory queues,
 629          * stored there.
 630          */
 631         if (pcxt->seg != NULL)
 632         {
 633                 dsm_detach(pcxt->seg);
 634                 pcxt->seg = NULL;
 635         }
 636
 637         /*
 638          * If this parallel context is actually in backend-private memory rather
 639          * than shared memory, free that memory instead.
 640          */
 641         if (pcxt->private_memory != NULL)
 642         {
 643                 pfree(pcxt->private_memory);
 644                 pcxt->private_memory = NULL;
 645         }
 646
 647         /*
 648          * We can't finish transaction commit or abort until all of the
 649          * workers have exited.  This means, in particular, that we can't respond
 650          * to interrupts at this stage.
 651          */
 652         HOLD_INTERRUPTS();
 653         WaitForParallelWorkersToExit(pcxt);
 654         RESUME_INTERRUPTS();
 655
 656         /* Free the worker array itself. */
 657         if (pcxt->worker != NULL)
 658         {
 659                 pfree(pcxt->worker);
 660                 pcxt->worker = NULL;
 661         }
 662
 663         /* Free memory. */
 664         pfree(pcxt);
 665 }
 666
 667 /*
 668  * Are there any parallel contexts currently active?
 669  */
 670 bool
 671 ParallelContextActive(void)
 672 {
 673         return !dlist_is_empty(&pcxt_list);
 674 }
 675
 676 /*
 677  * Handle receipt of an interrupt indicating a parallel worker message.
 678  */
 679 void
 680 HandleParallelMessageInterrupt(void)
 681 {
 682         int                     save_errno = errno;
 683
 684         InterruptPending = true;
 685         ParallelMessagePending = true;
 686         SetLatch(MyLatch);
 687
 688         errno = save_errno;
 689 }
 690
 691 /*
 692  * Handle any queued protocol messages received from parallel workers.
 693  */
 694 void
 695 HandleParallelMessages(void)
 696 {
 697         dlist_iter      iter;
 698
 699         ParallelMessagePending = false;
 700
 701         dlist_foreach(iter, &pcxt_list)
 702         {
 703                 ParallelContext *pcxt;
 704                 int                     i;
 705                 Size            nbytes;
 706                 void       *data;
 707
 708                 pcxt = dlist_container(ParallelContext, node, iter.cur);
 709                 if (pcxt->worker == NULL)
 710                         continue;
 711
 712                 for (i = 0; i < pcxt->nworkers; ++i)
 713                 {
 714                         /*
 715                          * Read as many messages as we can from each worker, but stop when
 716                          * either (1) the error queue goes away, which can happen if we
 717                          * receive a Terminate message from the worker; or (2) no more
 718                          * messages can be read from the worker without blocking.
 719                          */
 720                         while (pcxt->worker[i].error_mqh != NULL)
 721                         {
 722                                 shm_mq_result res;
 723
 724                                 res = shm_mq_receive(pcxt->worker[i].error_mqh, &nbytes,
 725                                                                          &data, true);
 726                                 if (res == SHM_MQ_WOULD_BLOCK)
 727                                         break;
 728                                 else if (res == SHM_MQ_SUCCESS)
 729                                 {
 730                                         StringInfoData msg;
 731
 732                                         initStringInfo(&msg);
 733                                         appendBinaryStringInfo(&msg, data, nbytes);
 734                                         HandleParallelMessage(pcxt, i, &msg);
 735                                         pfree(msg.data);
 736                                 }
 737                                 else
 738                                         ereport(ERROR,
 739                                                         (errcode(ERRCODE_INTERNAL_ERROR),       /* XXX: wrong errcode? */
 740                                                          errmsg("lost connection to parallel worker")));
 741
 742                                 /* This might make the error queue go away. */
 743                                 CHECK_FOR_INTERRUPTS();
 744                         }
 745                 }
 746         }
 747 }
 748
 749 /*
 750  * Handle a single protocol message received from a single parallel worker.
 751  */
 752 static void
 753 HandleParallelMessage(ParallelContext *pcxt, int i, StringInfo msg)
 754 {
 755         char            msgtype;
 756
 757         msgtype = pq_getmsgbyte(msg);
 758
 759         switch (msgtype)
 760         {
 761                 case 'K':                               /* BackendKeyData */
 762                         {
 763                                 int32           pid = pq_getmsgint(msg, 4);
 764
 765                                 (void) pq_getmsgint(msg, 4);    /* discard cancel key */
 766                                 (void) pq_getmsgend(msg);
 767                                 pcxt->worker[i].pid = pid;
 768                                 break;
 769                         }
 770
 771                 case 'E':                               /* ErrorResponse */
 772                 case 'N':                               /* NoticeResponse */
 773                         {
 774                                 ErrorData       edata;
 775                                 ErrorContextCallback errctx;
 776                                 ErrorContextCallback *save_error_context_stack;
 777
 778                                 /*
 779                                  * Rethrow the error using the error context callbacks that
 780                                  * were in effect when the context was created, not the
 781                                  * current ones.
 782                                  */
 783                                 save_error_context_stack = error_context_stack;
 784                                 errctx.callback = ParallelErrorContext;
 785                                 errctx.arg = &pcxt->worker[i].pid;
 786                                 errctx.previous = pcxt->error_context_stack;
 787                                 error_context_stack = &errctx;
 788
 789                                 /* Parse ErrorResponse or NoticeResponse. */
 790                                 pq_parse_errornotice(msg, &edata);
 791
 792                                 /* Death of a worker isn't enough justification for suicide. */
 793                                 edata.elevel = Min(edata.elevel, ERROR);
 794
 795                                 /* Rethrow error or notice. */
 796                                 ThrowErrorData(&edata);
 797
 798                                 /* Restore previous context. */
 799                                 error_context_stack = save_error_context_stack;
 800
 801                                 break;
 802                         }
 803
 804                 case 'A':                               /* NotifyResponse */
 805                         {
 806                                 /* Propagate NotifyResponse. */
 807                                 pq_putmessage(msg->data[0], &msg->data[1], msg->len - 1);
 808                                 break;
 809                         }
 810
 811                 case 'X':                               /* Terminate, indicating clean exit */
 812                         {
 813                                 pfree(pcxt->worker[i].error_mqh);
 814                                 pcxt->worker[i].error_mqh = NULL;
 815                                 break;
 816                         }
 817
 818                 default:
 819                         {
 820                                 elog(ERROR, "unknown message type: %c (%d bytes)",
 821                                          msgtype, msg->len);
 822                         }
 823         }
 824 }
 825
 826 /*
 827  * End-of-subtransaction cleanup for parallel contexts.
 828  *
 829  * Currently, it's forbidden to enter or leave a subtransaction while
 830  * parallel mode is in effect, so we could just blow away everything.  But
 831  * we may want to relax that restriction in the future, so this code
 832  * contemplates that there may be multiple subtransaction IDs in pcxt_list.
 833  */
 834 void
 835 AtEOSubXact_Parallel(bool isCommit, SubTransactionId mySubId)
 836 {
 837         while (!dlist_is_empty(&pcxt_list))
 838         {
 839                 ParallelContext *pcxt;
 840
 841                 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
 842                 if (pcxt->subid != mySubId)
 843                         break;
 844                 if (isCommit)
 845                         elog(WARNING, "leaked parallel context");
 846                 DestroyParallelContext(pcxt);
 847         }
 848 }
 849
 850 /*
 851  * End-of-transaction cleanup for parallel contexts.
 852  */
 853 void
 854 AtEOXact_Parallel(bool isCommit)
 855 {
 856         while (!dlist_is_empty(&pcxt_list))
 857         {
 858                 ParallelContext *pcxt;
 859
 860                 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
 861                 if (isCommit)
 862                         elog(WARNING, "leaked parallel context");
 863                 DestroyParallelContext(pcxt);
 864         }
 865 }
 866
 867 /*
 868  * Main entrypoint for parallel workers.
 869  */
 870 static void
 871 ParallelWorkerMain(Datum main_arg)
 872 {
 873         dsm_segment *seg;
 874         shm_toc    *toc;
 875         FixedParallelState *fps;
 876         char       *error_queue_space;
 877         shm_mq     *mq;
 878         shm_mq_handle *mqh;
 879         char       *libraryspace;
 880         char       *gucspace;
 881         char       *combocidspace;
 882         char       *tsnapspace;
 883         char       *asnapspace;
 884         char       *tstatespace;
 885         StringInfoData msgbuf;
 886
 887         /* Set flag to indicate that we're initializing a parallel worker. */
 888         InitializingParallelWorker = true;
 889
 890         /* Establish signal handlers. */
 891         pqsignal(SIGTERM, die);
 892         BackgroundWorkerUnblockSignals();
 893
 894         /* Set up a memory context and resource owner. */
 895         Assert(CurrentResourceOwner == NULL);
 896         CurrentResourceOwner = ResourceOwnerCreate(NULL, "parallel toplevel");
 897         CurrentMemoryContext = AllocSetContextCreate(TopMemoryContext,
 898                                                                                                  "parallel worker",
 899                                                                                                  ALLOCSET_DEFAULT_MINSIZE,
 900                                                                                                  ALLOCSET_DEFAULT_INITSIZE,
 901                                                                                                  ALLOCSET_DEFAULT_MAXSIZE);
 902
 903         /*
 904          * Now that we have a resource owner, we can attach to the dynamic shared
 905          * memory segment and read the table of contents.
 906          */
 907         seg = dsm_attach(DatumGetUInt32(main_arg));
 908         if (seg == NULL)
 909                 ereport(ERROR,
 910                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 911                                  errmsg("unable to map dynamic shared memory segment")));
 912         toc = shm_toc_attach(PARALLEL_MAGIC, dsm_segment_address(seg));
 913         if (toc == NULL)
 914                 ereport(ERROR,
 915                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 916                            errmsg("bad magic number in dynamic shared memory segment")));
 917
 918         /* Determine and set our worker number. */
 919         fps = shm_toc_lookup(toc, PARALLEL_KEY_FIXED);
 920         Assert(fps != NULL);
 921         Assert(ParallelWorkerNumber == -1);
 922         SpinLockAcquire(&fps->mutex);
 923         if (fps->workers_attached < fps->workers_expected)
 924                 ParallelWorkerNumber = fps->workers_attached++;
 925         SpinLockRelease(&fps->mutex);
 926         if (ParallelWorkerNumber < 0)
 927                 ereport(ERROR,
 928                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 929                                  errmsg("too many parallel workers already attached")));
 930         MyFixedParallelState = fps;
 931
 932         /*
 933          * Now that we have a worker number, we can find and attach to the error
 934          * queue provided for us.  That's good, because until we do that, any
 935          * errors that happen here will not be reported back to the process that
 936          * requested that this worker be launched.
 937          */
 938         error_queue_space = shm_toc_lookup(toc, PARALLEL_KEY_ERROR_QUEUE);
 939         mq = (shm_mq *) (error_queue_space +
 940                                          ParallelWorkerNumber * PARALLEL_ERROR_QUEUE_SIZE);
 941         shm_mq_set_sender(mq, MyProc);
 942         mqh = shm_mq_attach(mq, seg, NULL);
 943         pq_redirect_to_shm_mq(seg, mqh);
 944         pq_set_parallel_master(fps->parallel_master_pid,
 945                                                    fps->parallel_master_backend_id);
 946
 947         /*
 948          * Send a BackendKeyData message to the process that initiated parallelism
 949          * so that it has access to our PID before it receives any other messages
 950          * from us.  Our cancel key is sent, too, since that's the way the
 951          * protocol message is defined, but it won't actually be used for anything
 952          * in this case.
 953          */
 954         pq_beginmessage(&msgbuf, 'K');
 955         pq_sendint(&msgbuf, (int32) MyProcPid, sizeof(int32));
 956         pq_sendint(&msgbuf, (int32) MyCancelKey, sizeof(int32));
 957         pq_endmessage(&msgbuf);
 958
 959         /*
 960          * Hooray! Primary initialization is complete.  Now, we need to set up our
 961          * backend-local state to match the original backend.
 962          */
 963
 964         /*
 965          * Load libraries that were loaded by original backend.  We want to do
 966          * this before restoring GUCs, because the libraries might define custom
 967          * variables.
 968          */
 969         libraryspace = shm_toc_lookup(toc, PARALLEL_KEY_LIBRARY);
 970         Assert(libraryspace != NULL);
 971         RestoreLibraryState(libraryspace);
 972
 973         /* Restore database connection. */
 974         BackgroundWorkerInitializeConnectionByOid(fps->database_id,
 975                                                                                           fps->authenticated_user_id);
 976
 977         /* Restore GUC values from launching backend. */
 978         gucspace = shm_toc_lookup(toc, PARALLEL_KEY_GUC);
 979         Assert(gucspace != NULL);
 980         StartTransactionCommand();
 981         RestoreGUCState(gucspace);
 982         CommitTransactionCommand();
 983
 984         /* Crank up a transaction state appropriate to a parallel worker. */
 985         tstatespace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_STATE);
 986         StartParallelWorkerTransaction(tstatespace);
 987
 988         /* Restore combo CID state. */
 989         combocidspace = shm_toc_lookup(toc, PARALLEL_KEY_COMBO_CID);
 990         Assert(combocidspace != NULL);
 991         RestoreComboCIDState(combocidspace);
 992
 993         /* Restore transaction snapshot. */
 994         tsnapspace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT);
 995         Assert(tsnapspace != NULL);
 996         RestoreTransactionSnapshot(RestoreSnapshot(tsnapspace),
 997                                                            fps->parallel_master_pgproc);
 998
 999         /* Restore active snapshot. */
1000         asnapspace = shm_toc_lookup(toc, PARALLEL_KEY_ACTIVE_SNAPSHOT);
1001         Assert(asnapspace != NULL);
1002         PushActiveSnapshot(RestoreSnapshot(asnapspace));
1003
1004         /*
1005          * We've changed which tuples we can see, and must therefore invalidate
1006          * system caches.
1007          */
1008         InvalidateSystemCaches();
1009
1010         /* Restore user ID and security context. */
1011         SetUserIdAndSecContext(fps->current_user_id, fps->sec_context);
1012
1013         /*
1014          * We've initialized all of our state now; nothing should change
1015          * hereafter.
1016          */
1017         InitializingParallelWorker = false;
1018         EnterParallelMode();
1019
1020         /*
1021          * Time to do the real work: invoke the caller-supplied code.
1022          *
1023          * If you get a crash at this line, see the comments for
1024          * ParallelExtensionTrampoline.
1025          */
1026         fps->entrypoint(seg, toc);
1027
1028         /* Must exit parallel mode to pop active snapshot. */
1029         ExitParallelMode();
1030
1031         /* Must pop active snapshot so resowner.c doesn't complain. */
1032         PopActiveSnapshot();
1033
1034         /* Shut down the parallel-worker transaction. */
1035         EndParallelWorkerTransaction();
1036
1037         /* Report success. */
1038         pq_putmessage('X', NULL, 0);
1039 }
1040
1041 /*
1042  * It's unsafe for the entrypoint invoked by ParallelWorkerMain to be a
1043  * function living in a dynamically loaded module, because the module might
1044  * not be loaded in every process, or might be loaded but not at the same
1045  * address.  To work around that problem, CreateParallelContextForExtension()
1046  * arranges to call this function rather than calling the extension-provided
1047  * function directly; and this function then looks up the real entrypoint and
1048  * calls it.
1049  */
1050 static void
1051 ParallelExtensionTrampoline(dsm_segment *seg, shm_toc *toc)
1052 {
1053         char       *extensionstate;
1054         char       *library_name;
1055         char       *function_name;
1056         parallel_worker_main_type entrypt;
1057
1058         extensionstate = shm_toc_lookup(toc, PARALLEL_KEY_EXTENSION_TRAMPOLINE);
1059         Assert(extensionstate != NULL);
1060         library_name = extensionstate;
1061         function_name = extensionstate + strlen(library_name) + 1;
1062
1063         entrypt = (parallel_worker_main_type)
1064                 load_external_function(library_name, function_name, true, NULL);
1065         entrypt(seg, toc);
1066 }
1067
1068 /*
1069  * Give the user a hint that this is a message propagated from a parallel
1070  * worker.  Otherwise, it can sometimes be confusing to understand what
1071  * actually happened.
1072  */
1073 static void
1074 ParallelErrorContext(void *arg)
1075 {
1076         errcontext("parallel worker, pid %d", *(int32 *) arg);
1077 }
1078
1079 /*
1080  * Update shared memory with the ending location of the last WAL record we
1081  * wrote, if it's greater than the value already stored there.
1082  */
1083 void
1084 ParallelWorkerReportLastRecEnd(XLogRecPtr last_xlog_end)
1085 {
1086         FixedParallelState *fps = MyFixedParallelState;
1087
1088         Assert(fps != NULL);
1089         SpinLockAcquire(&fps->mutex);
1090         if (fps->last_xlog_end < last_xlog_end)
1091                 fps->last_xlog_end = last_xlog_end;
1092         SpinLockRelease(&fps->mutex);
1093 }