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

   1 /*-------------------------------------------------------------------------
   2  *
   3  * parallel.c
   4  *        Infrastructure for launching parallel workers
   5  *
   6  * Portions Copyright (c) 1996-2016, PostgreSQL Global Development Group
   7  * Portions Copyright (c) 1994, Regents of the University of California
   8  *
   9  * IDENTIFICATION
  10  *        src/backend/access/transam/parallel.c
  11  *
  12  *-------------------------------------------------------------------------
  13  */
  14
  15 #include "postgres.h"
  16
  17 #include "access/xact.h"
  18 #include "access/xlog.h"
  19 #include "access/parallel.h"
  20 #include "commands/async.h"
  21 #include "libpq/libpq.h"
  22 #include "libpq/pqformat.h"
  23 #include "libpq/pqmq.h"
  24 #include "miscadmin.h"
  25 #include "optimizer/planmain.h"
  26 #include "storage/ipc.h"
  27 #include "storage/sinval.h"
  28 #include "storage/spin.h"
  29 #include "tcop/tcopprot.h"
  30 #include "utils/combocid.h"
  31 #include "utils/guc.h"
  32 #include "utils/inval.h"
  33 #include "utils/memutils.h"
  34 #include "utils/resowner.h"
  35 #include "utils/snapmgr.h"
  36
  37 /*
  38  * We don't want to waste a lot of memory on an error queue which, most of
  39  * the time, will process only a handful of small messages.  However, it is
  40  * desirable to make it large enough that a typical ErrorResponse can be sent
  41  * without blocking.  That way, a worker that errors out can write the whole
  42  * message into the queue and terminate without waiting for the user backend.
  43  */
  44 #define PARALLEL_ERROR_QUEUE_SIZE                       16384
  45
  46 /* Magic number for parallel context TOC. */
  47 #define PARALLEL_MAGIC                                          0x50477c7c
  48
  49 /*
  50  * Magic numbers for parallel state sharing.  Higher-level code should use
  51  * smaller values, leaving these very large ones for use by this module.
  52  */
  53 #define PARALLEL_KEY_FIXED                                      UINT64CONST(0xFFFFFFFFFFFF0001)
  54 #define PARALLEL_KEY_ERROR_QUEUE                        UINT64CONST(0xFFFFFFFFFFFF0002)
  55 #define PARALLEL_KEY_LIBRARY                            UINT64CONST(0xFFFFFFFFFFFF0003)
  56 #define PARALLEL_KEY_GUC                                        UINT64CONST(0xFFFFFFFFFFFF0004)
  57 #define PARALLEL_KEY_COMBO_CID                          UINT64CONST(0xFFFFFFFFFFFF0005)
  58 #define PARALLEL_KEY_TRANSACTION_SNAPSHOT       UINT64CONST(0xFFFFFFFFFFFF0006)
  59 #define PARALLEL_KEY_ACTIVE_SNAPSHOT            UINT64CONST(0xFFFFFFFFFFFF0007)
  60 #define PARALLEL_KEY_TRANSACTION_STATE          UINT64CONST(0xFFFFFFFFFFFF0008)
  61 #define PARALLEL_KEY_EXTENSION_TRAMPOLINE       UINT64CONST(0xFFFFFFFFFFFF0009)
  62
  63 /* Fixed-size parallel state. */
  64 typedef struct FixedParallelState
  65 {
  66         /* Fixed-size state that workers must restore. */
  67         Oid                     database_id;
  68         Oid                     authenticated_user_id;
  69         Oid                     current_user_id;
  70         int                     sec_context;
  71         PGPROC     *parallel_master_pgproc;
  72         pid_t           parallel_master_pid;
  73         BackendId       parallel_master_backend_id;
  74
  75         /* Entrypoint for parallel workers. */
  76         parallel_worker_main_type entrypoint;
  77
  78         /* Mutex protects remaining fields. */
  79         slock_t         mutex;
  80
  81         /* Maximum XactLastRecEnd of any worker. */
  82         XLogRecPtr      last_xlog_end;
  83 } FixedParallelState;
  84
  85 /*
  86  * Our parallel worker number.  We initialize this to -1, meaning that we are
  87  * not a parallel worker.  In parallel workers, it will be set to a value >= 0
  88  * and < the number of workers before any user code is invoked; each parallel
  89  * worker will get a different parallel worker number.
  90  */
  91 int                     ParallelWorkerNumber = -1;
  92
  93 /* Is there a parallel message pending which we need to receive? */
  94 bool            ParallelMessagePending = false;
  95
  96 /* Are we initializing a parallel worker? */
  97 bool            InitializingParallelWorker = false;
  98
  99 /* Pointer to our fixed parallel state. */
 100 static FixedParallelState *MyFixedParallelState;
 101
 102 /* List of active parallel contexts. */
 103 static dlist_head pcxt_list = DLIST_STATIC_INIT(pcxt_list);
 104
 105 /* Private functions. */
 106 static void HandleParallelMessage(ParallelContext *, int, StringInfo msg);
 107 static void ParallelErrorContext(void *arg);
 108 static void ParallelExtensionTrampoline(dsm_segment *seg, shm_toc *toc);
 109 static void ParallelWorkerMain(Datum main_arg);
 110 static void WaitForParallelWorkersToExit(ParallelContext *pcxt);
 111
 112 /*
 113  * Establish a new parallel context.  This should be done after entering
 114  * parallel mode, and (unless there is an error) the context should be
 115  * destroyed before exiting the current subtransaction.
 116  */
 117 ParallelContext *
 118 CreateParallelContext(parallel_worker_main_type entrypoint, int nworkers)
 119 {
 120         MemoryContext oldcontext;
 121         ParallelContext *pcxt;
 122
 123         /* It is unsafe to create a parallel context if not in parallel mode. */
 124         Assert(IsInParallelMode());
 125
 126         /* Number of workers should be non-negative. */
 127         Assert(nworkers >= 0);
 128
 129         /*
 130          * If dynamic shared memory is not available, we won't be able to use
 131          * background workers.
 132          */
 133         if (dynamic_shared_memory_type == DSM_IMPL_NONE)
 134                 nworkers = 0;
 135
 136         /*
 137          * If we are running under serializable isolation, we can't use parallel
 138          * workers, at least not until somebody enhances that mechanism to be
 139          * parallel-aware.
 140          */
 141         if (IsolationIsSerializable())
 142                 nworkers = 0;
 143
 144         /* We might be running in a short-lived memory context. */
 145         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 146
 147         /* Initialize a new ParallelContext. */
 148         pcxt = palloc0(sizeof(ParallelContext));
 149         pcxt->subid = GetCurrentSubTransactionId();
 150         pcxt->nworkers = nworkers;
 151         pcxt->entrypoint = entrypoint;
 152         pcxt->error_context_stack = error_context_stack;
 153         shm_toc_initialize_estimator(&pcxt->estimator);
 154         dlist_push_head(&pcxt_list, &pcxt->node);
 155
 156         /* Restore previous memory context. */
 157         MemoryContextSwitchTo(oldcontext);
 158
 159         return pcxt;
 160 }
 161
 162 /*
 163  * Establish a new parallel context that calls a function provided by an
 164  * extension.  This works around the fact that the library might get mapped
 165  * at a different address in each backend.
 166  */
 167 ParallelContext *
 168 CreateParallelContextForExternalFunction(char *library_name,
 169                                                                                  char *function_name,
 170                                                                                  int nworkers)
 171 {
 172         MemoryContext oldcontext;
 173         ParallelContext *pcxt;
 174
 175         /* We might be running in a very short-lived memory context. */
 176         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 177
 178         /* Create the context. */
 179         pcxt = CreateParallelContext(ParallelExtensionTrampoline, nworkers);
 180         pcxt->library_name = pstrdup(library_name);
 181         pcxt->function_name = pstrdup(function_name);
 182
 183         /* Restore previous memory context. */
 184         MemoryContextSwitchTo(oldcontext);
 185
 186         return pcxt;
 187 }
 188
 189 /*
 190  * Establish the dynamic shared memory segment for a parallel context and
 191  * copied state and other bookkeeping information that will need by parallel
 192  * workers into it.
 193  */
 194 void
 195 InitializeParallelDSM(ParallelContext *pcxt)
 196 {
 197         MemoryContext oldcontext;
 198         Size            library_len = 0;
 199         Size            guc_len = 0;
 200         Size            combocidlen = 0;
 201         Size            tsnaplen = 0;
 202         Size            asnaplen = 0;
 203         Size            tstatelen = 0;
 204         Size            segsize = 0;
 205         int                     i;
 206         FixedParallelState *fps;
 207         Snapshot        transaction_snapshot = GetTransactionSnapshot();
 208         Snapshot        active_snapshot = GetActiveSnapshot();
 209
 210         /* We might be running in a very short-lived memory context. */
 211         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 212
 213         /* Allow space to store the fixed-size parallel state. */
 214         shm_toc_estimate_chunk(&pcxt->estimator, sizeof(FixedParallelState));
 215         shm_toc_estimate_keys(&pcxt->estimator, 1);
 216
 217         /*
 218          * Normally, the user will have requested at least one worker process, but
 219          * if by chance they have not, we can skip a bunch of things here.
 220          */
 221         if (pcxt->nworkers > 0)
 222         {
 223                 /* Estimate space for various kinds of state sharing. */
 224                 library_len = EstimateLibraryStateSpace();
 225                 shm_toc_estimate_chunk(&pcxt->estimator, library_len);
 226                 guc_len = EstimateGUCStateSpace();
 227                 shm_toc_estimate_chunk(&pcxt->estimator, guc_len);
 228                 combocidlen = EstimateComboCIDStateSpace();
 229                 shm_toc_estimate_chunk(&pcxt->estimator, combocidlen);
 230                 tsnaplen = EstimateSnapshotSpace(transaction_snapshot);
 231                 shm_toc_estimate_chunk(&pcxt->estimator, tsnaplen);
 232                 asnaplen = EstimateSnapshotSpace(active_snapshot);
 233                 shm_toc_estimate_chunk(&pcxt->estimator, asnaplen);
 234                 tstatelen = EstimateTransactionStateSpace();
 235                 shm_toc_estimate_chunk(&pcxt->estimator, tstatelen);
 236                 /* If you add more chunks here, you probably need to add keys. */
 237                 shm_toc_estimate_keys(&pcxt->estimator, 6);
 238
 239                 /* Estimate space need for error queues. */
 240                 StaticAssertStmt(BUFFERALIGN(PARALLEL_ERROR_QUEUE_SIZE) ==
 241                                                  PARALLEL_ERROR_QUEUE_SIZE,
 242                                                  "parallel error queue size not buffer-aligned");
 243                 shm_toc_estimate_chunk(&pcxt->estimator,
 244                                                            mul_size(PARALLEL_ERROR_QUEUE_SIZE,
 245                                                                                 pcxt->nworkers));
 246                 shm_toc_estimate_keys(&pcxt->estimator, 1);
 247
 248                 /* Estimate how much we'll need for extension entrypoint info. */
 249                 if (pcxt->library_name != NULL)
 250                 {
 251                         Assert(pcxt->entrypoint == ParallelExtensionTrampoline);
 252                         Assert(pcxt->function_name != NULL);
 253                         shm_toc_estimate_chunk(&pcxt->estimator, strlen(pcxt->library_name)
 254                                                                    + strlen(pcxt->function_name) + 2);
 255                         shm_toc_estimate_keys(&pcxt->estimator, 1);
 256                 }
 257         }
 258
 259         /*
 260          * Create DSM and initialize with new table of contents.  But if the user
 261          * didn't request any workers, then don't bother creating a dynamic shared
 262          * memory segment; instead, just use backend-private memory.
 263          *
 264          * Also, if we can't create a dynamic shared memory segment because the
 265          * maximum number of segments have already been created, then fall back to
 266          * backend-private memory, and plan not to use any workers.  We hope this
 267          * won't happen very often, but it's better to abandon the use of
 268          * parallelism than to fail outright.
 269          */
 270         segsize = shm_toc_estimate(&pcxt->estimator);
 271         if (pcxt->nworkers != 0)
 272                 pcxt->seg = dsm_create(segsize, DSM_CREATE_NULL_IF_MAXSEGMENTS);
 273         if (pcxt->seg != NULL)
 274                 pcxt->toc = shm_toc_create(PARALLEL_MAGIC,
 275                                                                    dsm_segment_address(pcxt->seg),
 276                                                                    segsize);
 277         else
 278         {
 279                 pcxt->nworkers = 0;
 280                 pcxt->private_memory = MemoryContextAlloc(TopMemoryContext, segsize);
 281                 pcxt->toc = shm_toc_create(PARALLEL_MAGIC, pcxt->private_memory,
 282                                                                    segsize);
 283         }
 284
 285         /* Initialize fixed-size state in shared memory. */
 286         fps = (FixedParallelState *)
 287                 shm_toc_allocate(pcxt->toc, sizeof(FixedParallelState));
 288         fps->database_id = MyDatabaseId;
 289         fps->authenticated_user_id = GetAuthenticatedUserId();
 290         GetUserIdAndSecContext(&fps->current_user_id, &fps->sec_context);
 291         fps->parallel_master_pgproc = MyProc;
 292         fps->parallel_master_pid = MyProcPid;
 293         fps->parallel_master_backend_id = MyBackendId;
 294         fps->entrypoint = pcxt->entrypoint;
 295         SpinLockInit(&fps->mutex);
 296         fps->last_xlog_end = 0;
 297         shm_toc_insert(pcxt->toc, PARALLEL_KEY_FIXED, fps);
 298
 299         /* We can skip the rest of this if we're not budgeting for any workers. */
 300         if (pcxt->nworkers > 0)
 301         {
 302                 char       *libraryspace;
 303                 char       *gucspace;
 304                 char       *combocidspace;
 305                 char       *tsnapspace;
 306                 char       *asnapspace;
 307                 char       *tstatespace;
 308                 char       *error_queue_space;
 309
 310                 /* Serialize shared libraries we have loaded. */
 311                 libraryspace = shm_toc_allocate(pcxt->toc, library_len);
 312                 SerializeLibraryState(library_len, libraryspace);
 313                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_LIBRARY, libraryspace);
 314
 315                 /* Serialize GUC settings. */
 316                 gucspace = shm_toc_allocate(pcxt->toc, guc_len);
 317                 SerializeGUCState(guc_len, gucspace);
 318                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_GUC, gucspace);
 319
 320                 /* Serialize combo CID state. */
 321                 combocidspace = shm_toc_allocate(pcxt->toc, combocidlen);
 322                 SerializeComboCIDState(combocidlen, combocidspace);
 323                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_COMBO_CID, combocidspace);
 324
 325                 /* Serialize transaction snapshot and active snapshot. */
 326                 tsnapspace = shm_toc_allocate(pcxt->toc, tsnaplen);
 327                 SerializeSnapshot(transaction_snapshot, tsnapspace);
 328                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT,
 329                                            tsnapspace);
 330                 asnapspace = shm_toc_allocate(pcxt->toc, asnaplen);
 331                 SerializeSnapshot(active_snapshot, asnapspace);
 332                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ACTIVE_SNAPSHOT, asnapspace);
 333
 334                 /* Serialize transaction state. */
 335                 tstatespace = shm_toc_allocate(pcxt->toc, tstatelen);
 336                 SerializeTransactionState(tstatelen, tstatespace);
 337                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_STATE, tstatespace);
 338
 339                 /* Allocate space for worker information. */
 340                 pcxt->worker = palloc0(sizeof(ParallelWorkerInfo) * pcxt->nworkers);
 341
 342                 /*
 343                  * Establish error queues in dynamic shared memory.
 344                  *
 345                  * These queues should be used only for transmitting ErrorResponse,
 346                  * NoticeResponse, and NotifyResponse protocol messages.  Tuple data
 347                  * should be transmitted via separate (possibly larger?) queues.
 348                  */
 349                 error_queue_space =
 350                         shm_toc_allocate(pcxt->toc,
 351                                                          mul_size(PARALLEL_ERROR_QUEUE_SIZE,
 352                                                                           pcxt->nworkers));
 353                 for (i = 0; i < pcxt->nworkers; ++i)
 354                 {
 355                         char       *start;
 356                         shm_mq     *mq;
 357
 358                         start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
 359                         mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
 360                         shm_mq_set_receiver(mq, MyProc);
 361                         pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
 362                 }
 363                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE, error_queue_space);
 364
 365                 /* Serialize extension entrypoint information. */
 366                 if (pcxt->library_name != NULL)
 367                 {
 368                         Size            lnamelen = strlen(pcxt->library_name);
 369                         char       *extensionstate;
 370
 371                         extensionstate = shm_toc_allocate(pcxt->toc, lnamelen
 372                                                                                   + strlen(pcxt->function_name) + 2);
 373                         strcpy(extensionstate, pcxt->library_name);
 374                         strcpy(extensionstate + lnamelen + 1, pcxt->function_name);
 375                         shm_toc_insert(pcxt->toc, PARALLEL_KEY_EXTENSION_TRAMPOLINE,
 376                                                    extensionstate);
 377                 }
 378         }
 379
 380         /* Restore previous memory context. */
 381         MemoryContextSwitchTo(oldcontext);
 382 }
 383
 384 /*
 385  * Reinitialize the dynamic shared memory segment for a parallel context such
 386  * that we could launch workers for it again.
 387  */
 388 void
 389 ReinitializeParallelDSM(ParallelContext *pcxt)
 390 {
 391         FixedParallelState *fps;
 392         char       *error_queue_space;
 393         int                     i;
 394
 395         if (pcxt->nworkers_launched == 0)
 396                 return;
 397
 398         WaitForParallelWorkersToFinish(pcxt);
 399         WaitForParallelWorkersToExit(pcxt);
 400
 401         /* Reset a few bits of fixed parallel state to a clean state. */
 402         fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED);
 403         fps->last_xlog_end = 0;
 404
 405         /* Recreate error queues. */
 406         error_queue_space =
 407                 shm_toc_lookup(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE);
 408         for (i = 0; i < pcxt->nworkers; ++i)
 409         {
 410                 char       *start;
 411                 shm_mq     *mq;
 412
 413                 start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
 414                 mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
 415                 shm_mq_set_receiver(mq, MyProc);
 416                 pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
 417         }
 418
 419         /* Reset number of workers launched. */
 420         pcxt->nworkers_launched = 0;
 421 }
 422
 423 /*
 424  * Launch parallel workers.
 425  */
 426 void
 427 LaunchParallelWorkers(ParallelContext *pcxt)
 428 {
 429         MemoryContext oldcontext;
 430         BackgroundWorker worker;
 431         int                     i;
 432         bool            any_registrations_failed = false;
 433
 434         /* Skip this if we have no workers. */
 435         if (pcxt->nworkers == 0)
 436                 return;
 437
 438         /* We need to be a lock group leader. */
 439         BecomeLockGroupLeader();
 440
 441         /* If we do have workers, we'd better have a DSM segment. */
 442         Assert(pcxt->seg != NULL);
 443
 444         /* We might be running in a short-lived memory context. */
 445         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 446
 447         /* Configure a worker. */
 448         snprintf(worker.bgw_name, BGW_MAXLEN, "parallel worker for PID %d",
 449                          MyProcPid);
 450         worker.bgw_flags =
 451                 BGWORKER_SHMEM_ACCESS | BGWORKER_BACKEND_DATABASE_CONNECTION;
 452         worker.bgw_start_time = BgWorkerStart_ConsistentState;
 453         worker.bgw_restart_time = BGW_NEVER_RESTART;
 454         worker.bgw_main = ParallelWorkerMain;
 455         worker.bgw_main_arg = UInt32GetDatum(dsm_segment_handle(pcxt->seg));
 456         worker.bgw_notify_pid = MyProcPid;
 457         memset(&worker.bgw_extra, 0, BGW_EXTRALEN);
 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                 memcpy(worker.bgw_extra, &i, sizeof(int));
 470                 if (!any_registrations_failed &&
 471                         RegisterDynamicBackgroundWorker(&worker,
 472                                                                                         &pcxt->worker[i].bgwhandle))
 473                 {
 474                         shm_mq_set_handle(pcxt->worker[i].error_mqh,
 475                                                           pcxt->worker[i].bgwhandle);
 476                         pcxt->nworkers_launched++;
 477                 }
 478                 else
 479                 {
 480                         /*
 481                          * If we weren't able to register the worker, then we've bumped up
 482                          * against the max_worker_processes limit, and future
 483                          * registrations will probably fail too, so arrange to skip them.
 484                          * But we still have to execute this code for the remaining slots
 485                          * to make sure that we forget about the error queues we budgeted
 486                          * for those workers.  Otherwise, we'll wait for them to start,
 487                          * but they never will.
 488                          */
 489                         any_registrations_failed = true;
 490                         pcxt->worker[i].bgwhandle = NULL;
 491                         pcxt->worker[i].error_mqh = NULL;
 492                 }
 493         }
 494
 495         /* Restore previous memory context. */
 496         MemoryContextSwitchTo(oldcontext);
 497 }
 498
 499 /*
 500  * Wait for all workers to finish computing.
 501  *
 502  * Even if the parallel operation seems to have completed successfully, it's
 503  * important to call this function afterwards.  We must not miss any errors
 504  * the workers may have thrown during the parallel operation, or any that they
 505  * may yet throw while shutting down.
 506  *
 507  * Also, we want to update our notion of XactLastRecEnd based on worker
 508  * feedback.
 509  */
 510 void
 511 WaitForParallelWorkersToFinish(ParallelContext *pcxt)
 512 {
 513         for (;;)
 514         {
 515                 bool            anyone_alive = false;
 516                 int                     i;
 517
 518                 /*
 519                  * This will process any parallel messages that are pending, which may
 520                  * change the outcome of the loop that follows.  It may also throw an
 521                  * error propagated from a worker.
 522                  */
 523                 CHECK_FOR_INTERRUPTS();
 524
 525                 for (i = 0; i < pcxt->nworkers_launched; ++i)
 526                 {
 527                         if (pcxt->worker[i].error_mqh != NULL)
 528                         {
 529                                 anyone_alive = true;
 530                                 break;
 531                         }
 532                 }
 533
 534                 if (!anyone_alive)
 535                         break;
 536
 537                 WaitLatch(&MyProc->procLatch, WL_LATCH_SET, -1);
 538                 ResetLatch(&MyProc->procLatch);
 539         }
 540
 541         if (pcxt->toc != NULL)
 542         {
 543                 FixedParallelState *fps;
 544
 545                 fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED);
 546                 if (fps->last_xlog_end > XactLastRecEnd)
 547                         XactLastRecEnd = fps->last_xlog_end;
 548         }
 549 }
 550
 551 /*
 552  * Wait for all workers to exit.
 553  *
 554  * This function ensures that workers have been completely shutdown.  The
 555  * difference between WaitForParallelWorkersToFinish and this function is
 556  * that former just ensures that last message sent by worker backend is
 557  * received by master backend whereas this ensures the complete shutdown.
 558  */
 559 static void
 560 WaitForParallelWorkersToExit(ParallelContext *pcxt)
 561 {
 562         int                     i;
 563
 564         /* Wait until the workers actually die. */
 565         for (i = 0; i < pcxt->nworkers_launched; ++i)
 566         {
 567                 BgwHandleStatus status;
 568
 569                 if (pcxt->worker == NULL || pcxt->worker[i].bgwhandle == NULL)
 570                         continue;
 571
 572                 status = WaitForBackgroundWorkerShutdown(pcxt->worker[i].bgwhandle);
 573
 574                 /*
 575                  * If the postmaster kicked the bucket, we have no chance of cleaning
 576                  * up safely -- we won't be able to tell when our workers are actually
 577                  * dead.  This doesn't necessitate a PANIC since they will all abort
 578                  * eventually, but we can't safely continue this session.
 579                  */
 580                 if (status == BGWH_POSTMASTER_DIED)
 581                         ereport(FATAL,
 582                                         (errcode(ERRCODE_ADMIN_SHUTDOWN),
 583                                  errmsg("postmaster exited during a parallel transaction")));
 584
 585                 /* Release memory. */
 586                 pfree(pcxt->worker[i].bgwhandle);
 587                 pcxt->worker[i].bgwhandle = NULL;
 588         }
 589 }
 590
 591 /*
 592  * Destroy a parallel context.
 593  *
 594  * If expecting a clean exit, you should use WaitForParallelWorkersToFinish()
 595  * first, before calling this function.  When this function is invoked, any
 596  * remaining workers are forcibly killed; the dynamic shared memory segment
 597  * is unmapped; and we then wait (uninterruptibly) for the workers to exit.
 598  */
 599 void
 600 DestroyParallelContext(ParallelContext *pcxt)
 601 {
 602         int                     i;
 603
 604         /*
 605          * Be careful about order of operations here!  We remove the parallel
 606          * context from the list before we do anything else; otherwise, if an
 607          * error occurs during a subsequent step, we might try to nuke it again
 608          * from AtEOXact_Parallel or AtEOSubXact_Parallel.
 609          */
 610         dlist_delete(&pcxt->node);
 611
 612         /* Kill each worker in turn, and forget their error queues. */
 613         if (pcxt->worker != NULL)
 614         {
 615                 for (i = 0; i < pcxt->nworkers_launched; ++i)
 616                 {
 617                         if (pcxt->worker[i].error_mqh != NULL)
 618                         {
 619                                 TerminateBackgroundWorker(pcxt->worker[i].bgwhandle);
 620
 621                                 pfree(pcxt->worker[i].error_mqh);
 622                                 pcxt->worker[i].error_mqh = NULL;
 623                         }
 624                 }
 625         }
 626
 627         /*
 628          * If we have allocated a shared memory segment, detach it.  This will
 629          * implicitly detach the error queues, and any other shared memory queues,
 630          * stored there.
 631          */
 632         if (pcxt->seg != NULL)
 633         {
 634                 dsm_detach(pcxt->seg);
 635                 pcxt->seg = NULL;
 636         }
 637
 638         /*
 639          * If this parallel context is actually in backend-private memory rather
 640          * than shared memory, free that memory instead.
 641          */
 642         if (pcxt->private_memory != NULL)
 643         {
 644                 pfree(pcxt->private_memory);
 645                 pcxt->private_memory = NULL;
 646         }
 647
 648         /*
 649          * We can't finish transaction commit or abort until all of the workers
 650          * have exited.  This means, in particular, that we can't respond to
 651          * interrupts at this stage.
 652          */
 653         HOLD_INTERRUPTS();
 654         WaitForParallelWorkersToExit(pcxt);
 655         RESUME_INTERRUPTS();
 656
 657         /* Free the worker array itself. */
 658         if (pcxt->worker != NULL)
 659         {
 660                 pfree(pcxt->worker);
 661                 pcxt->worker = NULL;
 662         }
 663
 664         /* Free memory. */
 665         pfree(pcxt);
 666 }
 667
 668 /*
 669  * Are there any parallel contexts currently active?
 670  */
 671 bool
 672 ParallelContextActive(void)
 673 {
 674         return !dlist_is_empty(&pcxt_list);
 675 }
 676
 677 /*
 678  * Handle receipt of an interrupt indicating a parallel worker message.
 679  */
 680 void
 681 HandleParallelMessageInterrupt(void)
 682 {
 683         int                     save_errno = errno;
 684
 685         InterruptPending = true;
 686         ParallelMessagePending = true;
 687         SetLatch(MyLatch);
 688
 689         errno = save_errno;
 690 }
 691
 692 /*
 693  * Handle any queued protocol messages received from parallel workers.
 694  */
 695 void
 696 HandleParallelMessages(void)
 697 {
 698         dlist_iter      iter;
 699
 700         ParallelMessagePending = false;
 701
 702         dlist_foreach(iter, &pcxt_list)
 703         {
 704                 ParallelContext *pcxt;
 705                 int                     i;
 706                 Size            nbytes;
 707                 void       *data;
 708
 709                 pcxt = dlist_container(ParallelContext, node, iter.cur);
 710                 if (pcxt->worker == NULL)
 711                         continue;
 712
 713                 for (i = 0; i < pcxt->nworkers_launched; ++i)
 714                 {
 715                         /*
 716                          * Read as many messages as we can from each worker, but stop when
 717                          * either (1) the error queue goes away, which can happen if we
 718                          * receive a Terminate message from the worker; or (2) no more
 719                          * messages can be read from the worker without blocking.
 720                          */
 721                         while (pcxt->worker[i].error_mqh != NULL)
 722                         {
 723                                 shm_mq_result res;
 724
 725                                 res = shm_mq_receive(pcxt->worker[i].error_mqh, &nbytes,
 726                                                                          &data, true);
 727                                 if (res == SHM_MQ_WOULD_BLOCK)
 728                                         break;
 729                                 else if (res == SHM_MQ_SUCCESS)
 730                                 {
 731                                         StringInfoData msg;
 732
 733                                         initStringInfo(&msg);
 734                                         appendBinaryStringInfo(&msg, data, nbytes);
 735                                         HandleParallelMessage(pcxt, i, &msg);
 736                                         pfree(msg.data);
 737                                 }
 738                                 else
 739                                         ereport(ERROR,
 740                                                         (errcode(ERRCODE_INTERNAL_ERROR),       /* XXX: wrong errcode? */
 741                                                          errmsg("lost connection to parallel worker")));
 742
 743                                 /* This might make the error queue go away. */
 744                                 CHECK_FOR_INTERRUPTS();
 745                         }
 746                 }
 747         }
 748 }
 749
 750 /*
 751  * Handle a single protocol message received from a single parallel worker.
 752  */
 753 static void
 754 HandleParallelMessage(ParallelContext *pcxt, int i, StringInfo msg)
 755 {
 756         char            msgtype;
 757
 758         msgtype = pq_getmsgbyte(msg);
 759
 760         switch (msgtype)
 761         {
 762                 case 'K':                               /* BackendKeyData */
 763                         {
 764                                 int32           pid = pq_getmsgint(msg, 4);
 765
 766                                 (void) pq_getmsgint(msg, 4);    /* discard cancel key */
 767                                 (void) pq_getmsgend(msg);
 768                                 pcxt->worker[i].pid = pid;
 769                                 break;
 770                         }
 771
 772                 case 'E':                               /* ErrorResponse */
 773                 case 'N':                               /* NoticeResponse */
 774                         {
 775                                 ErrorData       edata;
 776                                 ErrorContextCallback errctx;
 777                                 ErrorContextCallback *save_error_context_stack;
 778
 779                                 /*
 780                                  * Rethrow the error using the error context callbacks that
 781                                  * were in effect when the context was created, not the
 782                                  * current ones.
 783                                  */
 784                                 save_error_context_stack = error_context_stack;
 785                                 errctx.callback = ParallelErrorContext;
 786                                 errctx.arg = &pcxt->worker[i].pid;
 787                                 errctx.previous = pcxt->error_context_stack;
 788                                 error_context_stack = &errctx;
 789
 790                                 /* Parse ErrorResponse or NoticeResponse. */
 791                                 pq_parse_errornotice(msg, &edata);
 792
 793                                 /* Death of a worker isn't enough justification for suicide. */
 794                                 edata.elevel = Min(edata.elevel, ERROR);
 795
 796                                 /* Rethrow error or notice. */
 797                                 ThrowErrorData(&edata);
 798
 799                                 /* Restore previous context. */
 800                                 error_context_stack = save_error_context_stack;
 801
 802                                 break;
 803                         }
 804
 805                 case 'A':                               /* NotifyResponse */
 806                         {
 807                                 /* Propagate NotifyResponse. */
 808                                 pq_putmessage(msg->data[0], &msg->data[1], msg->len - 1);
 809                                 break;
 810                         }
 811
 812                 case 'X':                               /* Terminate, indicating clean exit */
 813                         {
 814                                 pfree(pcxt->worker[i].error_mqh);
 815                                 pcxt->worker[i].error_mqh = NULL;
 816                                 break;
 817                         }
 818
 819                 default:
 820                         {
 821                                 elog(ERROR, "unknown message type: %c (%d bytes)",
 822                                          msgtype, msg->len);
 823                         }
 824         }
 825 }
 826
 827 /*
 828  * End-of-subtransaction cleanup for parallel contexts.
 829  *
 830  * Currently, it's forbidden to enter or leave a subtransaction while
 831  * parallel mode is in effect, so we could just blow away everything.  But
 832  * we may want to relax that restriction in the future, so this code
 833  * contemplates that there may be multiple subtransaction IDs in pcxt_list.
 834  */
 835 void
 836 AtEOSubXact_Parallel(bool isCommit, SubTransactionId mySubId)
 837 {
 838         while (!dlist_is_empty(&pcxt_list))
 839         {
 840                 ParallelContext *pcxt;
 841
 842                 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
 843                 if (pcxt->subid != mySubId)
 844                         break;
 845                 if (isCommit)
 846                         elog(WARNING, "leaked parallel context");
 847                 DestroyParallelContext(pcxt);
 848         }
 849 }
 850
 851 /*
 852  * End-of-transaction cleanup for parallel contexts.
 853  */
 854 void
 855 AtEOXact_Parallel(bool isCommit)
 856 {
 857         while (!dlist_is_empty(&pcxt_list))
 858         {
 859                 ParallelContext *pcxt;
 860
 861                 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
 862                 if (isCommit)
 863                         elog(WARNING, "leaked parallel context");
 864                 DestroyParallelContext(pcxt);
 865         }
 866 }
 867
 868 /*
 869  * Main entrypoint for parallel workers.
 870  */
 871 static void
 872 ParallelWorkerMain(Datum main_arg)
 873 {
 874         dsm_segment *seg;
 875         shm_toc    *toc;
 876         FixedParallelState *fps;
 877         char       *error_queue_space;
 878         shm_mq     *mq;
 879         shm_mq_handle *mqh;
 880         char       *libraryspace;
 881         char       *gucspace;
 882         char       *combocidspace;
 883         char       *tsnapspace;
 884         char       *asnapspace;
 885         char       *tstatespace;
 886         StringInfoData msgbuf;
 887
 888         /* Set flag to indicate that we're initializing a parallel worker. */
 889         InitializingParallelWorker = true;
 890
 891         /* Establish signal handlers. */
 892         pqsignal(SIGTERM, die);
 893         BackgroundWorkerUnblockSignals();
 894
 895         /* Determine and set our parallel worker number. */
 896         Assert(ParallelWorkerNumber == -1);
 897         memcpy(&ParallelWorkerNumber, MyBgworkerEntry->bgw_extra, sizeof(int));
 898
 899         /* Set up a memory context and resource owner. */
 900         Assert(CurrentResourceOwner == NULL);
 901         CurrentResourceOwner = ResourceOwnerCreate(NULL, "parallel toplevel");
 902         CurrentMemoryContext = AllocSetContextCreate(TopMemoryContext,
 903                                                                                                  "parallel worker",
 904                                                                                                  ALLOCSET_DEFAULT_MINSIZE,
 905                                                                                                  ALLOCSET_DEFAULT_INITSIZE,
 906                                                                                                  ALLOCSET_DEFAULT_MAXSIZE);
 907
 908         /*
 909          * Now that we have a resource owner, we can attach to the dynamic shared
 910          * memory segment and read the table of contents.
 911          */
 912         seg = dsm_attach(DatumGetUInt32(main_arg));
 913         if (seg == NULL)
 914                 ereport(ERROR,
 915                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 916                                  errmsg("could not map dynamic shared memory segment")));
 917         toc = shm_toc_attach(PARALLEL_MAGIC, dsm_segment_address(seg));
 918         if (toc == NULL)
 919                 ereport(ERROR,
 920                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 921                    errmsg("invalid magic number in dynamic shared memory segment")));
 922
 923         /* Look up fixed parallel state. */
 924         fps = shm_toc_lookup(toc, PARALLEL_KEY_FIXED);
 925         Assert(fps != NULL);
 926         MyFixedParallelState = fps;
 927
 928         /*
 929          * Now that we have a worker number, we can find and attach to the error
 930          * queue provided for us.  That's good, because until we do that, any
 931          * errors that happen here will not be reported back to the process that
 932          * requested that this worker be launched.
 933          */
 934         error_queue_space = shm_toc_lookup(toc, PARALLEL_KEY_ERROR_QUEUE);
 935         mq = (shm_mq *) (error_queue_space +
 936                                          ParallelWorkerNumber * PARALLEL_ERROR_QUEUE_SIZE);
 937         shm_mq_set_sender(mq, MyProc);
 938         mqh = shm_mq_attach(mq, seg, NULL);
 939         pq_redirect_to_shm_mq(seg, mqh);
 940         pq_set_parallel_master(fps->parallel_master_pid,
 941                                                    fps->parallel_master_backend_id);
 942
 943         /*
 944          * Send a BackendKeyData message to the process that initiated parallelism
 945          * so that it has access to our PID before it receives any other messages
 946          * from us.  Our cancel key is sent, too, since that's the way the
 947          * protocol message is defined, but it won't actually be used for anything
 948          * in this case.
 949          */
 950         pq_beginmessage(&msgbuf, 'K');
 951         pq_sendint(&msgbuf, (int32) MyProcPid, sizeof(int32));
 952         pq_sendint(&msgbuf, (int32) MyCancelKey, sizeof(int32));
 953         pq_endmessage(&msgbuf);
 954
 955         /*
 956          * Hooray! Primary initialization is complete.  Now, we need to set up our
 957          * backend-local state to match the original backend.
 958          */
 959
 960         /*
 961          * Join locking group.  We must do this before anything that could try to
 962          * acquire a heavyweight lock, because any heavyweight locks acquired to
 963          * this point could block either directly against the parallel group
 964          * leader or against some process which in turn waits for a lock that
 965          * conflicts with the parallel group leader, causing an undetected
 966          * deadlock.  (If we can't join the lock group, the leader has gone away,
 967          * so just exit quietly.)
 968          */
 969         if (!BecomeLockGroupMember(fps->parallel_master_pgproc,
 970                                                            fps->parallel_master_pid))
 971                 return;
 972
 973         /*
 974          * Load libraries that were loaded by original backend.  We want to do
 975          * this before restoring GUCs, because the libraries might define custom
 976          * variables.
 977          */
 978         libraryspace = shm_toc_lookup(toc, PARALLEL_KEY_LIBRARY);
 979         Assert(libraryspace != NULL);
 980         RestoreLibraryState(libraryspace);
 981
 982         /* Restore database connection. */
 983         BackgroundWorkerInitializeConnectionByOid(fps->database_id,
 984                                                                                           fps->authenticated_user_id);
 985
 986         /* Restore GUC values from launching backend. */
 987         gucspace = shm_toc_lookup(toc, PARALLEL_KEY_GUC);
 988         Assert(gucspace != NULL);
 989         StartTransactionCommand();
 990         RestoreGUCState(gucspace);
 991         CommitTransactionCommand();
 992
 993         /* Crank up a transaction state appropriate to a parallel worker. */
 994         tstatespace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_STATE);
 995         StartParallelWorkerTransaction(tstatespace);
 996
 997         /* Restore combo CID state. */
 998         combocidspace = shm_toc_lookup(toc, PARALLEL_KEY_COMBO_CID);
 999         Assert(combocidspace != NULL);
1000         RestoreComboCIDState(combocidspace);
1001
1002         /* Restore transaction snapshot. */
1003         tsnapspace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT);
1004         Assert(tsnapspace != NULL);
1005         RestoreTransactionSnapshot(RestoreSnapshot(tsnapspace),
1006                                                            fps->parallel_master_pgproc);
1007
1008         /* Restore active snapshot. */
1009         asnapspace = shm_toc_lookup(toc, PARALLEL_KEY_ACTIVE_SNAPSHOT);
1010         Assert(asnapspace != NULL);
1011         PushActiveSnapshot(RestoreSnapshot(asnapspace));
1012
1013         /*
1014          * We've changed which tuples we can see, and must therefore invalidate
1015          * system caches.
1016          */
1017         InvalidateSystemCaches();
1018
1019         /* Restore user ID and security context. */
1020         SetUserIdAndSecContext(fps->current_user_id, fps->sec_context);
1021
1022         /*
1023          * We've initialized all of our state now; nothing should change
1024          * hereafter.
1025          */
1026         InitializingParallelWorker = false;
1027         EnterParallelMode();
1028
1029         /*
1030          * Time to do the real work: invoke the caller-supplied code.
1031          *
1032          * If you get a crash at this line, see the comments for
1033          * ParallelExtensionTrampoline.
1034          */
1035         fps->entrypoint(seg, toc);
1036
1037         /* Must exit parallel mode to pop active snapshot. */
1038         ExitParallelMode();
1039
1040         /* Must pop active snapshot so resowner.c doesn't complain. */
1041         PopActiveSnapshot();
1042
1043         /* Shut down the parallel-worker transaction. */
1044         EndParallelWorkerTransaction();
1045
1046         /* Report success. */
1047         pq_putmessage('X', NULL, 0);
1048 }
1049
1050 /*
1051  * It's unsafe for the entrypoint invoked by ParallelWorkerMain to be a
1052  * function living in a dynamically loaded module, because the module might
1053  * not be loaded in every process, or might be loaded but not at the same
1054  * address.  To work around that problem, CreateParallelContextForExtension()
1055  * arranges to call this function rather than calling the extension-provided
1056  * function directly; and this function then looks up the real entrypoint and
1057  * calls it.
1058  */
1059 static void
1060 ParallelExtensionTrampoline(dsm_segment *seg, shm_toc *toc)
1061 {
1062         char       *extensionstate;
1063         char       *library_name;
1064         char       *function_name;
1065         parallel_worker_main_type entrypt;
1066
1067         extensionstate = shm_toc_lookup(toc, PARALLEL_KEY_EXTENSION_TRAMPOLINE);
1068         Assert(extensionstate != NULL);
1069         library_name = extensionstate;
1070         function_name = extensionstate + strlen(library_name) + 1;
1071
1072         entrypt = (parallel_worker_main_type)
1073                 load_external_function(library_name, function_name, true, NULL);
1074         entrypt(seg, toc);
1075 }
1076
1077 /*
1078  * Give the user a hint that this is a message propagated from a parallel
1079  * worker.  Otherwise, it can sometimes be confusing to understand what
1080  * actually happened.
1081  */
1082 static void
1083 ParallelErrorContext(void *arg)
1084 {
1085         if (force_parallel_mode != FORCE_PARALLEL_REGRESS)
1086                 errcontext("parallel worker, PID %d", *(int32 *) arg);
1087 }
1088
1089 /*
1090  * Update shared memory with the ending location of the last WAL record we
1091  * wrote, if it's greater than the value already stored there.
1092  */
1093 void
1094 ParallelWorkerReportLastRecEnd(XLogRecPtr last_xlog_end)
1095 {
1096         FixedParallelState *fps = MyFixedParallelState;
1097
1098         Assert(fps != NULL);
1099         SpinLockAcquire(&fps->mutex);
1100         if (fps->last_xlog_end < last_xlog_end)
1101                 fps->last_xlog_end = last_xlog_end;
1102         SpinLockRelease(&fps->mutex);
1103 }