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 "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         /* Maximum XactLastRecEnd of any worker. */
  81         XLogRecPtr      last_xlog_end;
  82 } FixedParallelState;
  83
  84 /*
  85  * Our parallel worker number.  We initialize this to -1, meaning that we are
  86  * not a parallel worker.  In parallel workers, it will be set to a value >= 0
  87  * and < the number of workers before any user code is invoked; each parallel
  88  * worker will get a different parallel worker number.
  89  */
  90 int                     ParallelWorkerNumber = -1;
  91
  92 /* Is there a parallel message pending which we need to receive? */
  93 bool            ParallelMessagePending = false;
  94
  95 /* Are we initializing a parallel worker? */
  96 bool            InitializingParallelWorker = false;
  97
  98 /* Pointer to our fixed parallel state. */
  99 static FixedParallelState *MyFixedParallelState;
 100
 101 /* List of active parallel contexts. */
 102 static dlist_head pcxt_list = DLIST_STATIC_INIT(pcxt_list);
 103
 104 /* Private functions. */
 105 static void HandleParallelMessage(ParallelContext *, int, StringInfo msg);
 106 static void ParallelErrorContext(void *arg);
 107 static void ParallelExtensionTrampoline(dsm_segment *seg, shm_toc *toc);
 108 static void ParallelWorkerMain(Datum main_arg);
 109 static void WaitForParallelWorkersToExit(ParallelContext *pcxt);
 110
 111 /*
 112  * Establish a new parallel context.  This should be done after entering
 113  * parallel mode, and (unless there is an error) the context should be
 114  * destroyed before exiting the current subtransaction.
 115  */
 116 ParallelContext *
 117 CreateParallelContext(parallel_worker_main_type entrypoint, int nworkers)
 118 {
 119         MemoryContext oldcontext;
 120         ParallelContext *pcxt;
 121
 122         /* It is unsafe to create a parallel context if not in parallel mode. */
 123         Assert(IsInParallelMode());
 124
 125         /* Number of workers should be non-negative. */
 126         Assert(nworkers >= 0);
 127
 128         /*
 129          * If dynamic shared memory is not available, we won't be able to use
 130          * background workers.
 131          */
 132         if (dynamic_shared_memory_type == DSM_IMPL_NONE)
 133                 nworkers = 0;
 134
 135         /*
 136          * If we are running under serializable isolation, we can't use
 137          * parallel workers, at least not until somebody enhances that mechanism
 138          * to be parallel-aware.
 139          */
 140         if (IsolationIsSerializable())
 141                 nworkers = 0;
 142
 143         /* We might be running in a short-lived memory context. */
 144         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 145
 146         /* Initialize a new ParallelContext. */
 147         pcxt = palloc0(sizeof(ParallelContext));
 148         pcxt->subid = GetCurrentSubTransactionId();
 149         pcxt->nworkers = nworkers;
 150         pcxt->entrypoint = entrypoint;
 151         pcxt->error_context_stack = error_context_stack;
 152         shm_toc_initialize_estimator(&pcxt->estimator);
 153         dlist_push_head(&pcxt_list, &pcxt->node);
 154
 155         /* Restore previous memory context. */
 156         MemoryContextSwitchTo(oldcontext);
 157
 158         return pcxt;
 159 }
 160
 161 /*
 162  * Establish a new parallel context that calls a function provided by an
 163  * extension.  This works around the fact that the library might get mapped
 164  * at a different address in each backend.
 165  */
 166 ParallelContext *
 167 CreateParallelContextForExternalFunction(char *library_name,
 168                                                                                  char *function_name,
 169                                                                                  int nworkers)
 170 {
 171         MemoryContext oldcontext;
 172         ParallelContext *pcxt;
 173
 174         /* We might be running in a very short-lived memory context. */
 175         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 176
 177         /* Create the context. */
 178         pcxt = CreateParallelContext(ParallelExtensionTrampoline, nworkers);
 179         pcxt->library_name = pstrdup(library_name);
 180         pcxt->function_name = pstrdup(function_name);
 181
 182         /* Restore previous memory context. */
 183         MemoryContextSwitchTo(oldcontext);
 184
 185         return pcxt;
 186 }
 187
 188 /*
 189  * Establish the dynamic shared memory segment for a parallel context and
 190  * copied state and other bookkeeping information that will need by parallel
 191  * workers into it.
 192  */
 193 void
 194 InitializeParallelDSM(ParallelContext *pcxt)
 195 {
 196         MemoryContext oldcontext;
 197         Size            library_len = 0;
 198         Size            guc_len = 0;
 199         Size            combocidlen = 0;
 200         Size            tsnaplen = 0;
 201         Size            asnaplen = 0;
 202         Size            tstatelen = 0;
 203         Size            segsize = 0;
 204         int                     i;
 205         FixedParallelState *fps;
 206         Snapshot        transaction_snapshot = GetTransactionSnapshot();
 207         Snapshot        active_snapshot = GetActiveSnapshot();
 208
 209         /* We might be running in a very short-lived memory context. */
 210         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 211
 212         /* Allow space to store the fixed-size parallel state. */
 213         shm_toc_estimate_chunk(&pcxt->estimator, sizeof(FixedParallelState));
 214         shm_toc_estimate_keys(&pcxt->estimator, 1);
 215
 216         /*
 217          * Normally, the user will have requested at least one worker process, but
 218          * if by chance they have not, we can skip a bunch of things here.
 219          */
 220         if (pcxt->nworkers > 0)
 221         {
 222                 /* Estimate space for various kinds of state sharing. */
 223                 library_len = EstimateLibraryStateSpace();
 224                 shm_toc_estimate_chunk(&pcxt->estimator, library_len);
 225                 guc_len = EstimateGUCStateSpace();
 226                 shm_toc_estimate_chunk(&pcxt->estimator, guc_len);
 227                 combocidlen = EstimateComboCIDStateSpace();
 228                 shm_toc_estimate_chunk(&pcxt->estimator, combocidlen);
 229                 tsnaplen = EstimateSnapshotSpace(transaction_snapshot);
 230                 shm_toc_estimate_chunk(&pcxt->estimator, tsnaplen);
 231                 asnaplen = EstimateSnapshotSpace(active_snapshot);
 232                 shm_toc_estimate_chunk(&pcxt->estimator, asnaplen);
 233                 tstatelen = EstimateTransactionStateSpace();
 234                 shm_toc_estimate_chunk(&pcxt->estimator, tstatelen);
 235                 /* If you add more chunks here, you probably need to add keys. */
 236                 shm_toc_estimate_keys(&pcxt->estimator, 6);
 237
 238                 /* Estimate space need for error queues. */
 239                 StaticAssertStmt(BUFFERALIGN(PARALLEL_ERROR_QUEUE_SIZE) ==
 240                                                  PARALLEL_ERROR_QUEUE_SIZE,
 241                                                  "parallel error queue size not buffer-aligned");
 242                 shm_toc_estimate_chunk(&pcxt->estimator,
 243                                                            PARALLEL_ERROR_QUEUE_SIZE * pcxt->nworkers);
 244                 shm_toc_estimate_keys(&pcxt->estimator, 1);
 245
 246                 /* Estimate how much we'll need for extension entrypoint info. */
 247                 if (pcxt->library_name != NULL)
 248                 {
 249                         Assert(pcxt->entrypoint == ParallelExtensionTrampoline);
 250                         Assert(pcxt->function_name != NULL);
 251                         shm_toc_estimate_chunk(&pcxt->estimator, strlen(pcxt->library_name)
 252                                                                    + strlen(pcxt->function_name) + 2);
 253                         shm_toc_estimate_keys(&pcxt->estimator, 1);
 254                 }
 255         }
 256
 257         /*
 258          * Create DSM and initialize with new table of contents.  But if the user
 259          * didn't request any workers, then don't bother creating a dynamic shared
 260          * memory segment; instead, just use backend-private memory.
 261          *
 262          * Also, if we can't create a dynamic shared memory segment because the
 263          * maximum number of segments have already been created, then fall back to
 264          * backend-private memory, and plan not to use any workers.  We hope this
 265          * won't happen very often, but it's better to abandon the use of
 266          * parallelism than to fail outright.
 267          */
 268         segsize = shm_toc_estimate(&pcxt->estimator);
 269         if (pcxt->nworkers != 0)
 270                 pcxt->seg = dsm_create(segsize, DSM_CREATE_NULL_IF_MAXSEGMENTS);
 271         if (pcxt->seg != NULL)
 272                 pcxt->toc = shm_toc_create(PARALLEL_MAGIC,
 273                                                                    dsm_segment_address(pcxt->seg),
 274                                                                    segsize);
 275         else
 276         {
 277                 pcxt->nworkers = 0;
 278                 pcxt->private_memory = MemoryContextAlloc(TopMemoryContext, segsize);
 279                 pcxt->toc = shm_toc_create(PARALLEL_MAGIC, pcxt->private_memory,
 280                                                                    segsize);
 281         }
 282
 283         /* Initialize fixed-size state in shared memory. */
 284         fps = (FixedParallelState *)
 285                 shm_toc_allocate(pcxt->toc, sizeof(FixedParallelState));
 286         fps->database_id = MyDatabaseId;
 287         fps->authenticated_user_id = GetAuthenticatedUserId();
 288         GetUserIdAndSecContext(&fps->current_user_id, &fps->sec_context);
 289         fps->parallel_master_pgproc = MyProc;
 290         fps->parallel_master_pid = MyProcPid;
 291         fps->parallel_master_backend_id = MyBackendId;
 292         fps->entrypoint = pcxt->entrypoint;
 293         SpinLockInit(&fps->mutex);
 294         fps->last_xlog_end = 0;
 295         shm_toc_insert(pcxt->toc, PARALLEL_KEY_FIXED, fps);
 296
 297         /* We can skip the rest of this if we're not budgeting for any workers. */
 298         if (pcxt->nworkers > 0)
 299         {
 300                 char       *libraryspace;
 301                 char       *gucspace;
 302                 char       *combocidspace;
 303                 char       *tsnapspace;
 304                 char       *asnapspace;
 305                 char       *tstatespace;
 306                 char       *error_queue_space;
 307
 308                 /* Serialize shared libraries we have loaded. */
 309                 libraryspace = shm_toc_allocate(pcxt->toc, library_len);
 310                 SerializeLibraryState(library_len, libraryspace);
 311                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_LIBRARY, libraryspace);
 312
 313                 /* Serialize GUC settings. */
 314                 gucspace = shm_toc_allocate(pcxt->toc, guc_len);
 315                 SerializeGUCState(guc_len, gucspace);
 316                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_GUC, gucspace);
 317
 318                 /* Serialize combo CID state. */
 319                 combocidspace = shm_toc_allocate(pcxt->toc, combocidlen);
 320                 SerializeComboCIDState(combocidlen, combocidspace);
 321                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_COMBO_CID, combocidspace);
 322
 323                 /* Serialize transaction snapshot and active snapshot. */
 324                 tsnapspace = shm_toc_allocate(pcxt->toc, tsnaplen);
 325                 SerializeSnapshot(transaction_snapshot, tsnapspace);
 326                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT,
 327                                            tsnapspace);
 328                 asnapspace = shm_toc_allocate(pcxt->toc, asnaplen);
 329                 SerializeSnapshot(active_snapshot, asnapspace);
 330                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ACTIVE_SNAPSHOT, asnapspace);
 331
 332                 /* Serialize transaction state. */
 333                 tstatespace = shm_toc_allocate(pcxt->toc, tstatelen);
 334                 SerializeTransactionState(tstatelen, tstatespace);
 335                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_STATE, tstatespace);
 336
 337                 /* Allocate space for worker information. */
 338                 pcxt->worker = palloc0(sizeof(ParallelWorkerInfo) * pcxt->nworkers);
 339
 340                 /*
 341                  * Establish error queues in dynamic shared memory.
 342                  *
 343                  * These queues should be used only for transmitting ErrorResponse,
 344                  * NoticeResponse, and NotifyResponse protocol messages.  Tuple data
 345                  * should be transmitted via separate (possibly larger?) queues.
 346                  */
 347                 error_queue_space =
 348                         shm_toc_allocate(pcxt->toc,
 349                                                          PARALLEL_ERROR_QUEUE_SIZE * pcxt->nworkers);
 350                 for (i = 0; i < pcxt->nworkers; ++i)
 351                 {
 352                         char       *start;
 353                         shm_mq     *mq;
 354
 355                         start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
 356                         mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
 357                         shm_mq_set_receiver(mq, MyProc);
 358                         pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
 359                 }
 360                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE, error_queue_space);
 361
 362                 /* Serialize extension entrypoint information. */
 363                 if (pcxt->library_name != NULL)
 364                 {
 365                         Size            lnamelen = strlen(pcxt->library_name);
 366                         char       *extensionstate;
 367
 368                         extensionstate = shm_toc_allocate(pcxt->toc, lnamelen
 369                                                                                   + strlen(pcxt->function_name) + 2);
 370                         strcpy(extensionstate, pcxt->library_name);
 371                         strcpy(extensionstate + lnamelen + 1, pcxt->function_name);
 372                         shm_toc_insert(pcxt->toc, PARALLEL_KEY_EXTENSION_TRAMPOLINE,
 373                                                    extensionstate);
 374                 }
 375         }
 376
 377         /* Restore previous memory context. */
 378         MemoryContextSwitchTo(oldcontext);
 379 }
 380
 381 /*
 382  * Reinitialize the dynamic shared memory segment for a parallel context such
 383  * that we could launch workers for it again.
 384  */
 385 void
 386 ReinitializeParallelDSM(ParallelContext *pcxt)
 387 {
 388         FixedParallelState *fps;
 389         char       *error_queue_space;
 390         int                     i;
 391
 392         if (pcxt->nworkers_launched == 0)
 393                 return;
 394
 395         WaitForParallelWorkersToFinish(pcxt);
 396         WaitForParallelWorkersToExit(pcxt);
 397
 398         /* Reset a few bits of fixed parallel state to a clean state. */
 399         fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED);
 400         fps->last_xlog_end = 0;
 401
 402         /* Recreate error queues. */
 403         error_queue_space =
 404                 shm_toc_lookup(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE);
 405         for (i = 0; i < pcxt->nworkers; ++i)
 406         {
 407                 char       *start;
 408                 shm_mq     *mq;
 409
 410                 start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
 411                 mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
 412                 shm_mq_set_receiver(mq, MyProc);
 413                 pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
 414         }
 415
 416         /* Reset number of workers launched. */
 417         pcxt->nworkers_launched = 0;
 418 }
 419
 420 /*
 421  * Launch parallel workers.
 422  */
 423 void
 424 LaunchParallelWorkers(ParallelContext *pcxt)
 425 {
 426         MemoryContext oldcontext;
 427         BackgroundWorker worker;
 428         int                     i;
 429         bool            any_registrations_failed = false;
 430
 431         /* Skip this if we have no workers. */
 432         if (pcxt->nworkers == 0)
 433                 return;
 434
 435         /* We need to be a lock group leader. */
 436         BecomeLockGroupLeader();
 437
 438         /* If we do have workers, we'd better have a DSM segment. */
 439         Assert(pcxt->seg != NULL);
 440
 441         /* We might be running in a short-lived memory context. */
 442         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 443
 444         /* Configure a worker. */
 445         snprintf(worker.bgw_name, BGW_MAXLEN, "parallel worker for PID %d",
 446                          MyProcPid);
 447         worker.bgw_flags =
 448                 BGWORKER_SHMEM_ACCESS | BGWORKER_BACKEND_DATABASE_CONNECTION;
 449         worker.bgw_start_time = BgWorkerStart_ConsistentState;
 450         worker.bgw_restart_time = BGW_NEVER_RESTART;
 451         worker.bgw_main = ParallelWorkerMain;
 452         worker.bgw_main_arg = UInt32GetDatum(dsm_segment_handle(pcxt->seg));
 453         worker.bgw_notify_pid = MyProcPid;
 454         memset(&worker.bgw_extra, 0, BGW_EXTRALEN);
 455
 456         /*
 457          * Start workers.
 458          *
 459          * The caller must be able to tolerate ending up with fewer workers than
 460          * expected, so there is no need to throw an error here if registration
 461          * fails.  It wouldn't help much anyway, because registering the worker in
 462          * no way guarantees that it will start up and initialize successfully.
 463          */
 464         for (i = 0; i < pcxt->nworkers; ++i)
 465         {
 466                 memcpy(worker.bgw_extra, &i, sizeof(int));
 467                 if (!any_registrations_failed &&
 468                         RegisterDynamicBackgroundWorker(&worker,
 469                                                                                         &pcxt->worker[i].bgwhandle))
 470                 {
 471                         shm_mq_set_handle(pcxt->worker[i].error_mqh,
 472                                                           pcxt->worker[i].bgwhandle);
 473                         pcxt->nworkers_launched++;
 474                 }
 475                 else
 476                 {
 477                         /*
 478                          * If we weren't able to register the worker, then we've bumped up
 479                          * against the max_worker_processes limit, and future
 480                          * registrations will probably fail too, so arrange to skip them.
 481                          * But we still have to execute this code for the remaining slots
 482                          * to make sure that we forget about the error queues we budgeted
 483                          * for those workers.  Otherwise, we'll wait for them to start,
 484                          * but they never will.
 485                          */
 486                         any_registrations_failed = true;
 487                         pcxt->worker[i].bgwhandle = NULL;
 488                         pcxt->worker[i].error_mqh = NULL;
 489                 }
 490         }
 491
 492         /* Restore previous memory context. */
 493         MemoryContextSwitchTo(oldcontext);
 494 }
 495
 496 /*
 497  * Wait for all workers to finish computing.
 498  *
 499  * Even if the parallel operation seems to have completed successfully, it's
 500  * important to call this function afterwards.  We must not miss any errors
 501  * the workers may have thrown during the parallel operation, or any that they
 502  * may yet throw while shutting down.
 503  *
 504  * Also, we want to update our notion of XactLastRecEnd based on worker
 505  * feedback.
 506  */
 507 void
 508 WaitForParallelWorkersToFinish(ParallelContext *pcxt)
 509 {
 510         for (;;)
 511         {
 512                 bool            anyone_alive = false;
 513                 int                     i;
 514
 515                 /*
 516                  * This will process any parallel messages that are pending, which may
 517                  * change the outcome of the loop that follows.  It may also throw an
 518                  * error propagated from a worker.
 519                  */
 520                 CHECK_FOR_INTERRUPTS();
 521
 522                 for (i = 0; i < pcxt->nworkers; ++i)
 523                 {
 524                         if (pcxt->worker[i].error_mqh != NULL)
 525                         {
 526                                 anyone_alive = true;
 527                                 break;
 528                         }
 529                 }
 530
 531                 if (!anyone_alive)
 532                         break;
 533
 534                 WaitLatch(&MyProc->procLatch, WL_LATCH_SET, -1);
 535                 ResetLatch(&MyProc->procLatch);
 536         }
 537
 538         if (pcxt->toc != NULL)
 539         {
 540                 FixedParallelState *fps;
 541
 542                 fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED);
 543                 if (fps->last_xlog_end > XactLastRecEnd)
 544                         XactLastRecEnd = fps->last_xlog_end;
 545         }
 546 }
 547
 548 /*
 549  * Wait for all workers to exit.
 550  *
 551  * This function ensures that workers have been completely shutdown.  The
 552  * difference between WaitForParallelWorkersToFinish and this function is
 553  * that former just ensures that last message sent by worker backend is
 554  * received by master backend whereas this ensures the complete shutdown.
 555  */
 556 static void
 557 WaitForParallelWorkersToExit(ParallelContext *pcxt)
 558 {
 559         int                     i;
 560
 561         /* Wait until the workers actually die. */
 562         for (i = 0; i < pcxt->nworkers; ++i)
 563         {
 564                 BgwHandleStatus status;
 565
 566                 if (pcxt->worker == NULL || pcxt->worker[i].bgwhandle == NULL)
 567                         continue;
 568
 569                 status = WaitForBackgroundWorkerShutdown(pcxt->worker[i].bgwhandle);
 570
 571                 /*
 572                  * If the postmaster kicked the bucket, we have no chance of cleaning
 573                  * up safely -- we won't be able to tell when our workers are actually
 574                  * dead.  This doesn't necessitate a PANIC since they will all abort
 575                  * eventually, but we can't safely continue this session.
 576                  */
 577                 if (status == BGWH_POSTMASTER_DIED)
 578                         ereport(FATAL,
 579                                         (errcode(ERRCODE_ADMIN_SHUTDOWN),
 580                                  errmsg("postmaster exited during a parallel transaction")));
 581
 582                 /* Release memory. */
 583                 pfree(pcxt->worker[i].bgwhandle);
 584                 pcxt->worker[i].bgwhandle = NULL;
 585         }
 586 }
 587
 588 /*
 589  * Destroy a parallel context.
 590  *
 591  * If expecting a clean exit, you should use WaitForParallelWorkersToFinish()
 592  * first, before calling this function.  When this function is invoked, any
 593  * remaining workers are forcibly killed; the dynamic shared memory segment
 594  * is unmapped; and we then wait (uninterruptibly) for the workers to exit.
 595  */
 596 void
 597 DestroyParallelContext(ParallelContext *pcxt)
 598 {
 599         int                     i;
 600
 601         /*
 602          * Be careful about order of operations here!  We remove the parallel
 603          * context from the list before we do anything else; otherwise, if an
 604          * error occurs during a subsequent step, we might try to nuke it again
 605          * from AtEOXact_Parallel or AtEOSubXact_Parallel.
 606          */
 607         dlist_delete(&pcxt->node);
 608
 609         /* Kill each worker in turn, and forget their error queues. */
 610         if (pcxt->worker != NULL)
 611         {
 612                 for (i = 0; i < pcxt->nworkers; ++i)
 613                 {
 614                         if (pcxt->worker[i].error_mqh != NULL)
 615                         {
 616                                 TerminateBackgroundWorker(pcxt->worker[i].bgwhandle);
 617
 618                                 pfree(pcxt->worker[i].error_mqh);
 619                                 pcxt->worker[i].error_mqh = NULL;
 620                         }
 621                 }
 622         }
 623
 624         /*
 625          * If we have allocated a shared memory segment, detach it.  This will
 626          * implicitly detach the error queues, and any other shared memory queues,
 627          * stored there.
 628          */
 629         if (pcxt->seg != NULL)
 630         {
 631                 dsm_detach(pcxt->seg);
 632                 pcxt->seg = NULL;
 633         }
 634
 635         /*
 636          * If this parallel context is actually in backend-private memory rather
 637          * than shared memory, free that memory instead.
 638          */
 639         if (pcxt->private_memory != NULL)
 640         {
 641                 pfree(pcxt->private_memory);
 642                 pcxt->private_memory = NULL;
 643         }
 644
 645         /*
 646          * We can't finish transaction commit or abort until all of the
 647          * workers have exited.  This means, in particular, that we can't respond
 648          * to interrupts at this stage.
 649          */
 650         HOLD_INTERRUPTS();
 651         WaitForParallelWorkersToExit(pcxt);
 652         RESUME_INTERRUPTS();
 653
 654         /* Free the worker array itself. */
 655         if (pcxt->worker != NULL)
 656         {
 657                 pfree(pcxt->worker);
 658                 pcxt->worker = NULL;
 659         }
 660
 661         /* Free memory. */
 662         pfree(pcxt);
 663 }
 664
 665 /*
 666  * Are there any parallel contexts currently active?
 667  */
 668 bool
 669 ParallelContextActive(void)
 670 {
 671         return !dlist_is_empty(&pcxt_list);
 672 }
 673
 674 /*
 675  * Handle receipt of an interrupt indicating a parallel worker message.
 676  */
 677 void
 678 HandleParallelMessageInterrupt(void)
 679 {
 680         int                     save_errno = errno;
 681
 682         InterruptPending = true;
 683         ParallelMessagePending = true;
 684         SetLatch(MyLatch);
 685
 686         errno = save_errno;
 687 }
 688
 689 /*
 690  * Handle any queued protocol messages received from parallel workers.
 691  */
 692 void
 693 HandleParallelMessages(void)
 694 {
 695         dlist_iter      iter;
 696
 697         ParallelMessagePending = false;
 698
 699         dlist_foreach(iter, &pcxt_list)
 700         {
 701                 ParallelContext *pcxt;
 702                 int                     i;
 703                 Size            nbytes;
 704                 void       *data;
 705
 706                 pcxt = dlist_container(ParallelContext, node, iter.cur);
 707                 if (pcxt->worker == NULL)
 708                         continue;
 709
 710                 for (i = 0; i < pcxt->nworkers; ++i)
 711                 {
 712                         /*
 713                          * Read as many messages as we can from each worker, but stop when
 714                          * either (1) the error queue goes away, which can happen if we
 715                          * receive a Terminate message from the worker; or (2) no more
 716                          * messages can be read from the worker without blocking.
 717                          */
 718                         while (pcxt->worker[i].error_mqh != NULL)
 719                         {
 720                                 shm_mq_result res;
 721
 722                                 res = shm_mq_receive(pcxt->worker[i].error_mqh, &nbytes,
 723                                                                          &data, true);
 724                                 if (res == SHM_MQ_WOULD_BLOCK)
 725                                         break;
 726                                 else if (res == SHM_MQ_SUCCESS)
 727                                 {
 728                                         StringInfoData msg;
 729
 730                                         initStringInfo(&msg);
 731                                         appendBinaryStringInfo(&msg, data, nbytes);
 732                                         HandleParallelMessage(pcxt, i, &msg);
 733                                         pfree(msg.data);
 734                                 }
 735                                 else
 736                                         ereport(ERROR,
 737                                                         (errcode(ERRCODE_INTERNAL_ERROR),       /* XXX: wrong errcode? */
 738                                                          errmsg("lost connection to parallel worker")));
 739
 740                                 /* This might make the error queue go away. */
 741                                 CHECK_FOR_INTERRUPTS();
 742                         }
 743                 }
 744         }
 745 }
 746
 747 /*
 748  * Handle a single protocol message received from a single parallel worker.
 749  */
 750 static void
 751 HandleParallelMessage(ParallelContext *pcxt, int i, StringInfo msg)
 752 {
 753         char            msgtype;
 754
 755         msgtype = pq_getmsgbyte(msg);
 756
 757         switch (msgtype)
 758         {
 759                 case 'K':                               /* BackendKeyData */
 760                         {
 761                                 int32           pid = pq_getmsgint(msg, 4);
 762
 763                                 (void) pq_getmsgint(msg, 4);    /* discard cancel key */
 764                                 (void) pq_getmsgend(msg);
 765                                 pcxt->worker[i].pid = pid;
 766                                 break;
 767                         }
 768
 769                 case 'E':                               /* ErrorResponse */
 770                 case 'N':                               /* NoticeResponse */
 771                         {
 772                                 ErrorData       edata;
 773                                 ErrorContextCallback errctx;
 774                                 ErrorContextCallback *save_error_context_stack;
 775
 776                                 /*
 777                                  * Rethrow the error using the error context callbacks that
 778                                  * were in effect when the context was created, not the
 779                                  * current ones.
 780                                  */
 781                                 save_error_context_stack = error_context_stack;
 782                                 errctx.callback = ParallelErrorContext;
 783                                 errctx.arg = &pcxt->worker[i].pid;
 784                                 errctx.previous = pcxt->error_context_stack;
 785                                 error_context_stack = &errctx;
 786
 787                                 /* Parse ErrorResponse or NoticeResponse. */
 788                                 pq_parse_errornotice(msg, &edata);
 789
 790                                 /* Death of a worker isn't enough justification for suicide. */
 791                                 edata.elevel = Min(edata.elevel, ERROR);
 792
 793                                 /* Rethrow error or notice. */
 794                                 ThrowErrorData(&edata);
 795
 796                                 /* Restore previous context. */
 797                                 error_context_stack = save_error_context_stack;
 798
 799                                 break;
 800                         }
 801
 802                 case 'A':                               /* NotifyResponse */
 803                         {
 804                                 /* Propagate NotifyResponse. */
 805                                 pq_putmessage(msg->data[0], &msg->data[1], msg->len - 1);
 806                                 break;
 807                         }
 808
 809                 case 'X':                               /* Terminate, indicating clean exit */
 810                         {
 811                                 pfree(pcxt->worker[i].error_mqh);
 812                                 pcxt->worker[i].error_mqh = NULL;
 813                                 break;
 814                         }
 815
 816                 default:
 817                         {
 818                                 elog(ERROR, "unknown message type: %c (%d bytes)",
 819                                          msgtype, msg->len);
 820                         }
 821         }
 822 }
 823
 824 /*
 825  * End-of-subtransaction cleanup for parallel contexts.
 826  *
 827  * Currently, it's forbidden to enter or leave a subtransaction while
 828  * parallel mode is in effect, so we could just blow away everything.  But
 829  * we may want to relax that restriction in the future, so this code
 830  * contemplates that there may be multiple subtransaction IDs in pcxt_list.
 831  */
 832 void
 833 AtEOSubXact_Parallel(bool isCommit, SubTransactionId mySubId)
 834 {
 835         while (!dlist_is_empty(&pcxt_list))
 836         {
 837                 ParallelContext *pcxt;
 838
 839                 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
 840                 if (pcxt->subid != mySubId)
 841                         break;
 842                 if (isCommit)
 843                         elog(WARNING, "leaked parallel context");
 844                 DestroyParallelContext(pcxt);
 845         }
 846 }
 847
 848 /*
 849  * End-of-transaction cleanup for parallel contexts.
 850  */
 851 void
 852 AtEOXact_Parallel(bool isCommit)
 853 {
 854         while (!dlist_is_empty(&pcxt_list))
 855         {
 856                 ParallelContext *pcxt;
 857
 858                 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
 859                 if (isCommit)
 860                         elog(WARNING, "leaked parallel context");
 861                 DestroyParallelContext(pcxt);
 862         }
 863 }
 864
 865 /*
 866  * Main entrypoint for parallel workers.
 867  */
 868 static void
 869 ParallelWorkerMain(Datum main_arg)
 870 {
 871         dsm_segment *seg;
 872         shm_toc    *toc;
 873         FixedParallelState *fps;
 874         char       *error_queue_space;
 875         shm_mq     *mq;
 876         shm_mq_handle *mqh;
 877         char       *libraryspace;
 878         char       *gucspace;
 879         char       *combocidspace;
 880         char       *tsnapspace;
 881         char       *asnapspace;
 882         char       *tstatespace;
 883         StringInfoData msgbuf;
 884
 885         /* Set flag to indicate that we're initializing a parallel worker. */
 886         InitializingParallelWorker = true;
 887
 888         /* Establish signal handlers. */
 889         pqsignal(SIGTERM, die);
 890         BackgroundWorkerUnblockSignals();
 891
 892         /* Determine and set our parallel worker number. */
 893         Assert(ParallelWorkerNumber == -1);
 894         memcpy(&ParallelWorkerNumber, MyBgworkerEntry->bgw_extra, sizeof(int));
 895
 896         /* Set up a memory context and resource owner. */
 897         Assert(CurrentResourceOwner == NULL);
 898         CurrentResourceOwner = ResourceOwnerCreate(NULL, "parallel toplevel");
 899         CurrentMemoryContext = AllocSetContextCreate(TopMemoryContext,
 900                                                                                                  "parallel worker",
 901                                                                                                  ALLOCSET_DEFAULT_MINSIZE,
 902                                                                                                  ALLOCSET_DEFAULT_INITSIZE,
 903                                                                                                  ALLOCSET_DEFAULT_MAXSIZE);
 904
 905         /*
 906          * Now that we have a resource owner, we can attach to the dynamic shared
 907          * memory segment and read the table of contents.
 908          */
 909         seg = dsm_attach(DatumGetUInt32(main_arg));
 910         if (seg == NULL)
 911                 ereport(ERROR,
 912                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 913                                  errmsg("could not map dynamic shared memory segment")));
 914         toc = shm_toc_attach(PARALLEL_MAGIC, dsm_segment_address(seg));
 915         if (toc == NULL)
 916                 ereport(ERROR,
 917                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 918                            errmsg("invalid magic number in dynamic shared memory segment")));
 919
 920         /* Look up fixed parallel state. */
 921         fps = shm_toc_lookup(toc, PARALLEL_KEY_FIXED);
 922         Assert(fps != NULL);
 923         MyFixedParallelState = fps;
 924
 925         /*
 926          * Now that we have a worker number, we can find and attach to the error
 927          * queue provided for us.  That's good, because until we do that, any
 928          * errors that happen here will not be reported back to the process that
 929          * requested that this worker be launched.
 930          */
 931         error_queue_space = shm_toc_lookup(toc, PARALLEL_KEY_ERROR_QUEUE);
 932         mq = (shm_mq *) (error_queue_space +
 933                                          ParallelWorkerNumber * PARALLEL_ERROR_QUEUE_SIZE);
 934         shm_mq_set_sender(mq, MyProc);
 935         mqh = shm_mq_attach(mq, seg, NULL);
 936         pq_redirect_to_shm_mq(seg, mqh);
 937         pq_set_parallel_master(fps->parallel_master_pid,
 938                                                    fps->parallel_master_backend_id);
 939
 940         /*
 941          * Send a BackendKeyData message to the process that initiated parallelism
 942          * so that it has access to our PID before it receives any other messages
 943          * from us.  Our cancel key is sent, too, since that's the way the
 944          * protocol message is defined, but it won't actually be used for anything
 945          * in this case.
 946          */
 947         pq_beginmessage(&msgbuf, 'K');
 948         pq_sendint(&msgbuf, (int32) MyProcPid, sizeof(int32));
 949         pq_sendint(&msgbuf, (int32) MyCancelKey, sizeof(int32));
 950         pq_endmessage(&msgbuf);
 951
 952         /*
 953          * Hooray! Primary initialization is complete.  Now, we need to set up our
 954          * backend-local state to match the original backend.
 955          */
 956
 957         /*
 958          * Join locking group.  We must do this before anything that could try
 959          * to acquire a heavyweight lock, because any heavyweight locks acquired
 960          * to this point could block either directly against the parallel group
 961          * leader or against some process which in turn waits for a lock that
 962          * conflicts with the parallel group leader, causing an undetected
 963          * deadlock.  (If we can't join the lock group, the leader has gone away,
 964          * so just exit quietly.)
 965          */
 966         if (!BecomeLockGroupMember(fps->parallel_master_pgproc,
 967                                                            fps->parallel_master_pid))
 968                 return;
 969
 970         /*
 971          * Load libraries that were loaded by original backend.  We want to do
 972          * this before restoring GUCs, because the libraries might define custom
 973          * variables.
 974          */
 975         libraryspace = shm_toc_lookup(toc, PARALLEL_KEY_LIBRARY);
 976         Assert(libraryspace != NULL);
 977         RestoreLibraryState(libraryspace);
 978
 979         /* Restore database connection. */
 980         BackgroundWorkerInitializeConnectionByOid(fps->database_id,
 981                                                                                           fps->authenticated_user_id);
 982
 983         /* Restore GUC values from launching backend. */
 984         gucspace = shm_toc_lookup(toc, PARALLEL_KEY_GUC);
 985         Assert(gucspace != NULL);
 986         StartTransactionCommand();
 987         RestoreGUCState(gucspace);
 988         CommitTransactionCommand();
 989
 990         /* Crank up a transaction state appropriate to a parallel worker. */
 991         tstatespace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_STATE);
 992         StartParallelWorkerTransaction(tstatespace);
 993
 994         /* Restore combo CID state. */
 995         combocidspace = shm_toc_lookup(toc, PARALLEL_KEY_COMBO_CID);
 996         Assert(combocidspace != NULL);
 997         RestoreComboCIDState(combocidspace);
 998
 999         /* Restore transaction snapshot. */
1000         tsnapspace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT);
1001         Assert(tsnapspace != NULL);
1002         RestoreTransactionSnapshot(RestoreSnapshot(tsnapspace),
1003                                                            fps->parallel_master_pgproc);
1004
1005         /* Restore active snapshot. */
1006         asnapspace = shm_toc_lookup(toc, PARALLEL_KEY_ACTIVE_SNAPSHOT);
1007         Assert(asnapspace != NULL);
1008         PushActiveSnapshot(RestoreSnapshot(asnapspace));
1009
1010         /*
1011          * We've changed which tuples we can see, and must therefore invalidate
1012          * system caches.
1013          */
1014         InvalidateSystemCaches();
1015
1016         /* Restore user ID and security context. */
1017         SetUserIdAndSecContext(fps->current_user_id, fps->sec_context);
1018
1019         /*
1020          * We've initialized all of our state now; nothing should change
1021          * hereafter.
1022          */
1023         InitializingParallelWorker = false;
1024         EnterParallelMode();
1025
1026         /*
1027          * Time to do the real work: invoke the caller-supplied code.
1028          *
1029          * If you get a crash at this line, see the comments for
1030          * ParallelExtensionTrampoline.
1031          */
1032         fps->entrypoint(seg, toc);
1033
1034         /* Must exit parallel mode to pop active snapshot. */
1035         ExitParallelMode();
1036
1037         /* Must pop active snapshot so resowner.c doesn't complain. */
1038         PopActiveSnapshot();
1039
1040         /* Shut down the parallel-worker transaction. */
1041         EndParallelWorkerTransaction();
1042
1043         /* Report success. */
1044         pq_putmessage('X', NULL, 0);
1045 }
1046
1047 /*
1048  * It's unsafe for the entrypoint invoked by ParallelWorkerMain to be a
1049  * function living in a dynamically loaded module, because the module might
1050  * not be loaded in every process, or might be loaded but not at the same
1051  * address.  To work around that problem, CreateParallelContextForExtension()
1052  * arranges to call this function rather than calling the extension-provided
1053  * function directly; and this function then looks up the real entrypoint and
1054  * calls it.
1055  */
1056 static void
1057 ParallelExtensionTrampoline(dsm_segment *seg, shm_toc *toc)
1058 {
1059         char       *extensionstate;
1060         char       *library_name;
1061         char       *function_name;
1062         parallel_worker_main_type entrypt;
1063
1064         extensionstate = shm_toc_lookup(toc, PARALLEL_KEY_EXTENSION_TRAMPOLINE);
1065         Assert(extensionstate != NULL);
1066         library_name = extensionstate;
1067         function_name = extensionstate + strlen(library_name) + 1;
1068
1069         entrypt = (parallel_worker_main_type)
1070                 load_external_function(library_name, function_name, true, NULL);
1071         entrypt(seg, toc);
1072 }
1073
1074 /*
1075  * Give the user a hint that this is a message propagated from a parallel
1076  * worker.  Otherwise, it can sometimes be confusing to understand what
1077  * actually happened.
1078  */
1079 static void
1080 ParallelErrorContext(void *arg)
1081 {
1082         errcontext("parallel worker, PID %d", *(int32 *) arg);
1083 }
1084
1085 /*
1086  * Update shared memory with the ending location of the last WAL record we
1087  * wrote, if it's greater than the value already stored there.
1088  */
1089 void
1090 ParallelWorkerReportLastRecEnd(XLogRecPtr last_xlog_end)
1091 {
1092         FixedParallelState *fps = MyFixedParallelState;
1093
1094         Assert(fps != NULL);
1095         SpinLockAcquire(&fps->mutex);
1096         if (fps->last_xlog_end < last_xlog_end)
1097                 fps->last_xlog_end = last_xlog_end;
1098         SpinLockRelease(&fps->mutex);
1099 }