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

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