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