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, but
 209          * 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 to
 255          * backend-private memory, and plan not to use any workers.  We hope this
 256          * 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 in
 412          * 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 up
 425                          * 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 may
 463                  * change the outcome of the loop that follows.  It may also throw an
 464                  * 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 when
 641                          * either (1) the error queue goes away, which can happen if we
 642                          * 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
 690                                 (void) pq_getmsgint(msg, 4);    /* discard cancel key */
 691                                 (void) pq_getmsgend(msg);
 692                                 pcxt->worker[i].pid = pid;
 693                                 break;
 694                         }
 695
 696                 case 'E':                               /* ErrorResponse */
 697                 case 'N':                               /* NoticeResponse */
 698                         {
 699                                 ErrorData       edata;
 700                                 ErrorContextCallback errctx;
 701                                 ErrorContextCallback *save_error_context_stack;
 702
 703                                 /*
 704                                  * Rethrow the error using the error context callbacks that
 705                                  * were in effect when the context was created, not the
 706                                  * current ones.
 707                                  */
 708                                 save_error_context_stack = error_context_stack;
 709                                 errctx.callback = ParallelErrorContext;
 710                                 errctx.arg = &pcxt->worker[i].pid;
 711                                 errctx.previous = pcxt->error_context_stack;
 712                                 error_context_stack = &errctx;
 713
 714                                 /* Parse ErrorReponse or NoticeResponse. */
 715                                 pq_parse_errornotice(msg, &edata);
 716
 717                                 /* Death of a worker isn't enough justification for suicide. */
 718                                 edata.elevel = Min(edata.elevel, ERROR);
 719
 720                                 /* Rethrow error or notice. */
 721                                 ThrowErrorData(&edata);
 722
 723                                 /* Restore previous context. */
 724                                 error_context_stack = save_error_context_stack;
 725
 726                                 break;
 727                         }
 728
 729                 case 'A':                               /* NotifyResponse */
 730                         {
 731                                 /* Propagate NotifyResponse. */
 732                                 pq_putmessage(msg->data[0], &msg->data[1], msg->len - 1);
 733                                 break;
 734                         }
 735
 736                 case 'X':                               /* Terminate, indicating clean exit */
 737                         {
 738                                 pfree(pcxt->worker[i].bgwhandle);
 739                                 pfree(pcxt->worker[i].error_mqh);
 740                                 pcxt->worker[i].bgwhandle = NULL;
 741                                 pcxt->worker[i].error_mqh = NULL;
 742                                 break;
 743                         }
 744
 745                 default:
 746                         {
 747                                 elog(ERROR, "unknown message type: %c (%d bytes)",
 748                                          msgtype, msg->len);
 749                         }
 750         }
 751 }
 752
 753 /*
 754  * End-of-subtransaction cleanup for parallel contexts.
 755  *
 756  * Currently, it's forbidden to enter or leave a subtransaction while
 757  * parallel mode is in effect, so we could just blow away everything.  But
 758  * we may want to relax that restriction in the future, so this code
 759  * contemplates that there may be multiple subtransaction IDs in pcxt_list.
 760  */
 761 void
 762 AtEOSubXact_Parallel(bool isCommit, SubTransactionId mySubId)
 763 {
 764         while (!dlist_is_empty(&pcxt_list))
 765         {
 766                 ParallelContext *pcxt;
 767
 768                 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
 769                 if (pcxt->subid != mySubId)
 770                         break;
 771                 if (isCommit)
 772                         elog(WARNING, "leaked parallel context");
 773                 DestroyParallelContext(pcxt);
 774         }
 775 }
 776
 777 /*
 778  * End-of-transaction cleanup for parallel contexts.
 779  */
 780 void
 781 AtEOXact_Parallel(bool isCommit)
 782 {
 783         while (!dlist_is_empty(&pcxt_list))
 784         {
 785                 ParallelContext *pcxt;
 786
 787                 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
 788                 if (isCommit)
 789                         elog(WARNING, "leaked parallel context");
 790                 DestroyParallelContext(pcxt);
 791         }
 792 }
 793
 794 /*
 795  * Main entrypoint for parallel workers.
 796  */
 797 static void
 798 ParallelWorkerMain(Datum main_arg)
 799 {
 800         dsm_segment *seg;
 801         shm_toc    *toc;
 802         FixedParallelState *fps;
 803         char       *error_queue_space;
 804         shm_mq     *mq;
 805         shm_mq_handle *mqh;
 806         char       *libraryspace;
 807         char       *gucspace;
 808         char       *combocidspace;
 809         char       *tsnapspace;
 810         char       *asnapspace;
 811         char       *tstatespace;
 812         StringInfoData msgbuf;
 813
 814         /* Establish signal handlers. */
 815         pqsignal(SIGTERM, die);
 816         BackgroundWorkerUnblockSignals();
 817
 818         /* Set up a memory context and resource owner. */
 819         Assert(CurrentResourceOwner == NULL);
 820         CurrentResourceOwner = ResourceOwnerCreate(NULL, "parallel toplevel");
 821         CurrentMemoryContext = AllocSetContextCreate(TopMemoryContext,
 822                                                                                                  "parallel worker",
 823                                                                                                  ALLOCSET_DEFAULT_MINSIZE,
 824                                                                                                  ALLOCSET_DEFAULT_INITSIZE,
 825                                                                                                  ALLOCSET_DEFAULT_MAXSIZE);
 826
 827         /*
 828          * Now that we have a resource owner, we can attach to the dynamic shared
 829          * memory segment and read the table of contents.
 830          */
 831         seg = dsm_attach(DatumGetUInt32(main_arg));
 832         if (seg == NULL)
 833                 ereport(ERROR,
 834                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 835                                  errmsg("unable to map dynamic shared memory segment")));
 836         toc = shm_toc_attach(PARALLEL_MAGIC, dsm_segment_address(seg));
 837         if (toc == NULL)
 838                 ereport(ERROR,
 839                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 840                            errmsg("bad magic number in dynamic shared memory segment")));
 841
 842         /* Determine and set our worker number. */
 843         fps = shm_toc_lookup(toc, PARALLEL_KEY_FIXED);
 844         Assert(fps != NULL);
 845         Assert(ParallelWorkerNumber == -1);
 846         SpinLockAcquire(&fps->mutex);
 847         if (fps->workers_attached < fps->workers_expected)
 848                 ParallelWorkerNumber = fps->workers_attached++;
 849         SpinLockRelease(&fps->mutex);
 850         if (ParallelWorkerNumber < 0)
 851                 ereport(ERROR,
 852                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 853                                  errmsg("too many parallel workers already attached")));
 854         MyFixedParallelState = fps;
 855
 856         /*
 857          * Now that we have a worker number, we can find and attach to the error
 858          * queue provided for us.  That's good, because until we do that, any
 859          * errors that happen here will not be reported back to the process that
 860          * requested that this worker be launched.
 861          */
 862         error_queue_space = shm_toc_lookup(toc, PARALLEL_KEY_ERROR_QUEUE);
 863         mq = (shm_mq *) (error_queue_space +
 864                                          ParallelWorkerNumber * PARALLEL_ERROR_QUEUE_SIZE);
 865         shm_mq_set_sender(mq, MyProc);
 866         mqh = shm_mq_attach(mq, seg, NULL);
 867         pq_redirect_to_shm_mq(mq, mqh);
 868         pq_set_parallel_master(fps->parallel_master_pid,
 869                                                    fps->parallel_master_backend_id);
 870
 871         /*
 872          * Send a BackendKeyData message to the process that initiated parallelism
 873          * so that it has access to our PID before it receives any other messages
 874          * from us.  Our cancel key is sent, too, since that's the way the
 875          * protocol message is defined, but it won't actually be used for anything
 876          * in this case.
 877          */
 878         pq_beginmessage(&msgbuf, 'K');
 879         pq_sendint(&msgbuf, (int32) MyProcPid, sizeof(int32));
 880         pq_sendint(&msgbuf, (int32) MyCancelKey, sizeof(int32));
 881         pq_endmessage(&msgbuf);
 882
 883         /*
 884          * Hooray! Primary initialization is complete.  Now, we need to set up our
 885          * backend-local state to match the original backend.
 886          */
 887
 888         /*
 889          * Load libraries that were loaded by original backend.  We want to do
 890          * this before restoring GUCs, because the libraries might define custom
 891          * variables.
 892          */
 893         libraryspace = shm_toc_lookup(toc, PARALLEL_KEY_LIBRARY);
 894         Assert(libraryspace != NULL);
 895         RestoreLibraryState(libraryspace);
 896
 897         /* Restore database connection. */
 898         BackgroundWorkerInitializeConnectionByOid(fps->database_id,
 899                                                                                           fps->authenticated_user_id);
 900
 901         /* Restore GUC values from launching backend. */
 902         gucspace = shm_toc_lookup(toc, PARALLEL_KEY_GUC);
 903         Assert(gucspace != NULL);
 904         StartTransactionCommand();
 905         RestoreGUCState(gucspace);
 906         CommitTransactionCommand();
 907
 908         /* Crank up a transaction state appropriate to a parallel worker. */
 909         tstatespace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_STATE);
 910         StartParallelWorkerTransaction(tstatespace);
 911
 912         /* Restore combo CID state. */
 913         combocidspace = shm_toc_lookup(toc, PARALLEL_KEY_COMBO_CID);
 914         Assert(combocidspace != NULL);
 915         RestoreComboCIDState(combocidspace);
 916
 917         /* Restore transaction snapshot. */
 918         tsnapspace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT);
 919         Assert(tsnapspace != NULL);
 920         RestoreTransactionSnapshot(RestoreSnapshot(tsnapspace),
 921                                                            fps->parallel_master_pgproc);
 922
 923         /* Restore active snapshot. */
 924         asnapspace = shm_toc_lookup(toc, PARALLEL_KEY_ACTIVE_SNAPSHOT);
 925         Assert(asnapspace != NULL);
 926         PushActiveSnapshot(RestoreSnapshot(asnapspace));
 927
 928         /* Restore user ID and security context. */
 929         SetUserIdAndSecContext(fps->current_user_id, fps->sec_context);
 930
 931         /*
 932          * We've initialized all of our state now; nothing should change
 933          * hereafter.
 934          */
 935         EnterParallelMode();
 936
 937         /*
 938          * Time to do the real work: invoke the caller-supplied code.
 939          *
 940          * If you get a crash at this line, see the comments for
 941          * ParallelExtensionTrampoline.
 942          */
 943         fps->entrypoint(seg, toc);
 944
 945         /* Must exit parallel mode to pop active snapshot. */
 946         ExitParallelMode();
 947
 948         /* Must pop active snapshot so resowner.c doesn't complain. */
 949         PopActiveSnapshot();
 950
 951         /* Shut down the parallel-worker transaction. */
 952         EndParallelWorkerTransaction();
 953
 954         /* Report success. */
 955         pq_putmessage('X', NULL, 0);
 956 }
 957
 958 /*
 959  * It's unsafe for the entrypoint invoked by ParallelWorkerMain to be a
 960  * function living in a dynamically loaded module, because the module might
 961  * not be loaded in every process, or might be loaded but not at the same
 962  * address.  To work around that problem, CreateParallelContextForExtension()
 963  * arranges to call this function rather than calling the extension-provided
 964  * function directly; and this function then looks up the real entrypoint and
 965  * calls it.
 966  */
 967 static void
 968 ParallelExtensionTrampoline(dsm_segment *seg, shm_toc *toc)
 969 {
 970         char       *extensionstate;
 971         char       *library_name;
 972         char       *function_name;
 973         parallel_worker_main_type entrypt;
 974
 975         extensionstate = shm_toc_lookup(toc, PARALLEL_KEY_EXTENSION_TRAMPOLINE);
 976         Assert(extensionstate != NULL);
 977         library_name = extensionstate;
 978         function_name = extensionstate + strlen(library_name) + 1;
 979
 980         entrypt = (parallel_worker_main_type)
 981                 load_external_function(library_name, function_name, true, NULL);
 982         entrypt(seg, toc);
 983 }
 984
 985 /*
 986  * Give the user a hint that this is a message propagated from a parallel
 987  * worker.  Otherwise, it can sometimes be confusing to understand what
 988  * actually happened.
 989  */
 990 static void
 991 ParallelErrorContext(void *arg)
 992 {
 993         errcontext("parallel worker, pid %d", *(int32 *) arg);
 994 }
 995
 996 /*
 997  * Update shared memory with the ending location of the last WAL record we
 998  * wrote, if it's greater than the value already stored there.
 999  */
1000 void
1001 ParallelWorkerReportLastRecEnd(XLogRecPtr last_xlog_end)
1002 {
1003         FixedParallelState *fps = MyFixedParallelState;
1004
1005         Assert(fps != NULL);
1006         SpinLockAcquire(&fps->mutex);
1007         if (fps->last_xlog_end < last_xlog_end)
1008                 fps->last_xlog_end = last_xlog_end;
1009         SpinLockRelease(&fps->mutex);
1010 }