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) 1996-2017, 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/parallel.h"
  18 #include "access/xact.h"
  19 #include "access/xlog.h"
  20 #include "catalog/namespace.h"
  21 #include "commands/async.h"
  22 #include "executor/execParallel.h"
  23 #include "libpq/libpq.h"
  24 #include "libpq/pqformat.h"
  25 #include "libpq/pqmq.h"
  26 #include "miscadmin.h"
  27 #include "optimizer/planmain.h"
  28 #include "pgstat.h"
  29 #include "storage/ipc.h"
  30 #include "storage/sinval.h"
  31 #include "storage/spin.h"
  32 #include "tcop/tcopprot.h"
  33 #include "utils/combocid.h"
  34 #include "utils/guc.h"
  35 #include "utils/inval.h"
  36 #include "utils/memutils.h"
  37 #include "utils/resowner.h"
  38 #include "utils/snapmgr.h"
  39
  40
  41 /*
  42  * We don't want to waste a lot of memory on an error queue which, most of
  43  * the time, will process only a handful of small messages.  However, it is
  44  * desirable to make it large enough that a typical ErrorResponse can be sent
  45  * without blocking.  That way, a worker that errors out can write the whole
  46  * message into the queue and terminate without waiting for the user backend.
  47  */
  48 #define PARALLEL_ERROR_QUEUE_SIZE                       16384
  49
  50 /* Magic number for parallel context TOC. */
  51 #define PARALLEL_MAGIC                                          0x50477c7c
  52
  53 /*
  54  * Magic numbers for parallel state sharing.  Higher-level code should use
  55  * smaller values, leaving these very large ones for use by this module.
  56  */
  57 #define PARALLEL_KEY_FIXED                                      UINT64CONST(0xFFFFFFFFFFFF0001)
  58 #define PARALLEL_KEY_ERROR_QUEUE                        UINT64CONST(0xFFFFFFFFFFFF0002)
  59 #define PARALLEL_KEY_LIBRARY                            UINT64CONST(0xFFFFFFFFFFFF0003)
  60 #define PARALLEL_KEY_GUC                                        UINT64CONST(0xFFFFFFFFFFFF0004)
  61 #define PARALLEL_KEY_COMBO_CID                          UINT64CONST(0xFFFFFFFFFFFF0005)
  62 #define PARALLEL_KEY_TRANSACTION_SNAPSHOT       UINT64CONST(0xFFFFFFFFFFFF0006)
  63 #define PARALLEL_KEY_ACTIVE_SNAPSHOT            UINT64CONST(0xFFFFFFFFFFFF0007)
  64 #define PARALLEL_KEY_TRANSACTION_STATE          UINT64CONST(0xFFFFFFFFFFFF0008)
  65 #define PARALLEL_KEY_ENTRYPOINT                         UINT64CONST(0xFFFFFFFFFFFF0009)
  66
  67 /* Fixed-size parallel state. */
  68 typedef struct FixedParallelState
  69 {
  70         /* Fixed-size state that workers must restore. */
  71         Oid                     database_id;
  72         Oid                     authenticated_user_id;
  73         Oid                     current_user_id;
  74         Oid                     temp_namespace_id;
  75         Oid                     temp_toast_namespace_id;
  76         int                     sec_context;
  77         PGPROC     *parallel_master_pgproc;
  78         pid_t           parallel_master_pid;
  79         BackendId       parallel_master_backend_id;
  80
  81         /* Mutex protects remaining fields. */
  82         slock_t         mutex;
  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 volatile 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 /*
 109  * List of internal parallel worker entry points.  We need this for
 110  * reasons explained in LookupParallelWorkerFunction(), below.
 111  */
 112 static const struct
 113 {
 114         const char *fn_name;
 115         parallel_worker_main_type fn_addr;
 116 }                       InternalParallelWorkers[] =
 117
 118 {
 119         {
 120                 "ParallelQueryMain", ParallelQueryMain
 121         }
 122 };
 123
 124 /* Private functions. */
 125 static void HandleParallelMessage(ParallelContext *pcxt, int i, StringInfo msg);
 126 static void WaitForParallelWorkersToExit(ParallelContext *pcxt);
 127 static parallel_worker_main_type LookupParallelWorkerFunction(const char *libraryname, const char *funcname);
 128
 129
 130 /*
 131  * Establish a new parallel context.  This should be done after entering
 132  * parallel mode, and (unless there is an error) the context should be
 133  * destroyed before exiting the current subtransaction.
 134  */
 135 ParallelContext *
 136 CreateParallelContext(const char *library_name, const char *function_name,
 137                                           int nworkers)
 138 {
 139         MemoryContext oldcontext;
 140         ParallelContext *pcxt;
 141
 142         /* It is unsafe to create a parallel context if not in parallel mode. */
 143         Assert(IsInParallelMode());
 144
 145         /* Number of workers should be non-negative. */
 146         Assert(nworkers >= 0);
 147
 148         /*
 149          * If dynamic shared memory is not available, we won't be able to use
 150          * background workers.
 151          */
 152         if (dynamic_shared_memory_type == DSM_IMPL_NONE)
 153                 nworkers = 0;
 154
 155         /*
 156          * If we are running under serializable isolation, we can't use parallel
 157          * workers, at least not until somebody enhances that mechanism to be
 158          * parallel-aware.
 159          */
 160         if (IsolationIsSerializable())
 161                 nworkers = 0;
 162
 163         /* We might be running in a short-lived memory context. */
 164         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 165
 166         /* Initialize a new ParallelContext. */
 167         pcxt = palloc0(sizeof(ParallelContext));
 168         pcxt->subid = GetCurrentSubTransactionId();
 169         pcxt->nworkers = nworkers;
 170         pcxt->library_name = pstrdup(library_name);
 171         pcxt->function_name = pstrdup(function_name);
 172         pcxt->error_context_stack = error_context_stack;
 173         shm_toc_initialize_estimator(&pcxt->estimator);
 174         dlist_push_head(&pcxt_list, &pcxt->node);
 175
 176         /* Restore previous memory context. */
 177         MemoryContextSwitchTo(oldcontext);
 178
 179         return pcxt;
 180 }
 181
 182 /*
 183  * Establish the dynamic shared memory segment for a parallel context and
 184  * copy state and other bookkeeping information that will be needed by
 185  * parallel workers into it.
 186  */
 187 void
 188 InitializeParallelDSM(ParallelContext *pcxt)
 189 {
 190         MemoryContext oldcontext;
 191         Size            library_len = 0;
 192         Size            guc_len = 0;
 193         Size            combocidlen = 0;
 194         Size            tsnaplen = 0;
 195         Size            asnaplen = 0;
 196         Size            tstatelen = 0;
 197         Size            segsize = 0;
 198         int                     i;
 199         FixedParallelState *fps;
 200         Snapshot        transaction_snapshot = GetTransactionSnapshot();
 201         Snapshot        active_snapshot = GetActiveSnapshot();
 202
 203         /* We might be running in a very short-lived memory context. */
 204         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 205
 206         /* Allow space to store the fixed-size parallel state. */
 207         shm_toc_estimate_chunk(&pcxt->estimator, sizeof(FixedParallelState));
 208         shm_toc_estimate_keys(&pcxt->estimator, 1);
 209
 210         /*
 211          * Normally, the user will have requested at least one worker process, but
 212          * if by chance they have not, we can skip a bunch of things here.
 213          */
 214         if (pcxt->nworkers > 0)
 215         {
 216                 /* Estimate space for various kinds of state sharing. */
 217                 library_len = EstimateLibraryStateSpace();
 218                 shm_toc_estimate_chunk(&pcxt->estimator, library_len);
 219                 guc_len = EstimateGUCStateSpace();
 220                 shm_toc_estimate_chunk(&pcxt->estimator, guc_len);
 221                 combocidlen = EstimateComboCIDStateSpace();
 222                 shm_toc_estimate_chunk(&pcxt->estimator, combocidlen);
 223                 tsnaplen = EstimateSnapshotSpace(transaction_snapshot);
 224                 shm_toc_estimate_chunk(&pcxt->estimator, tsnaplen);
 225                 asnaplen = EstimateSnapshotSpace(active_snapshot);
 226                 shm_toc_estimate_chunk(&pcxt->estimator, asnaplen);
 227                 tstatelen = EstimateTransactionStateSpace();
 228                 shm_toc_estimate_chunk(&pcxt->estimator, tstatelen);
 229                 /* If you add more chunks here, you probably need to add keys. */
 230                 shm_toc_estimate_keys(&pcxt->estimator, 6);
 231
 232                 /* Estimate space need for error queues. */
 233                 StaticAssertStmt(BUFFERALIGN(PARALLEL_ERROR_QUEUE_SIZE) ==
 234                                                  PARALLEL_ERROR_QUEUE_SIZE,
 235                                                  "parallel error queue size not buffer-aligned");
 236                 shm_toc_estimate_chunk(&pcxt->estimator,
 237                                                            mul_size(PARALLEL_ERROR_QUEUE_SIZE,
 238                                                                                 pcxt->nworkers));
 239                 shm_toc_estimate_keys(&pcxt->estimator, 1);
 240
 241                 /* Estimate how much we'll need for the entrypoint info. */
 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          * Create DSM and initialize with new table of contents.  But if the user
 249          * didn't request any workers, then don't bother creating a dynamic shared
 250          * memory segment; instead, just use backend-private memory.
 251          *
 252          * Also, if we can't create a dynamic shared memory segment because the
 253          * maximum number of segments have already been created, then fall back to
 254          * backend-private memory, and plan not to use any workers.  We hope this
 255          * won't happen very often, but it's better to abandon the use of
 256          * parallelism than to fail outright.
 257          */
 258         segsize = shm_toc_estimate(&pcxt->estimator);
 259         if (pcxt->nworkers > 0)
 260                 pcxt->seg = dsm_create(segsize, DSM_CREATE_NULL_IF_MAXSEGMENTS);
 261         if (pcxt->seg != NULL)
 262                 pcxt->toc = shm_toc_create(PARALLEL_MAGIC,
 263                                                                    dsm_segment_address(pcxt->seg),
 264                                                                    segsize);
 265         else
 266         {
 267                 pcxt->nworkers = 0;
 268                 pcxt->private_memory = MemoryContextAlloc(TopMemoryContext, segsize);
 269                 pcxt->toc = shm_toc_create(PARALLEL_MAGIC, pcxt->private_memory,
 270                                                                    segsize);
 271         }
 272
 273         /* Initialize fixed-size state in shared memory. */
 274         fps = (FixedParallelState *)
 275                 shm_toc_allocate(pcxt->toc, sizeof(FixedParallelState));
 276         fps->database_id = MyDatabaseId;
 277         fps->authenticated_user_id = GetAuthenticatedUserId();
 278         GetUserIdAndSecContext(&fps->current_user_id, &fps->sec_context);
 279         GetTempNamespaceState(&fps->temp_namespace_id,
 280                                                   &fps->temp_toast_namespace_id);
 281         fps->parallel_master_pgproc = MyProc;
 282         fps->parallel_master_pid = MyProcPid;
 283         fps->parallel_master_backend_id = MyBackendId;
 284         SpinLockInit(&fps->mutex);
 285         fps->last_xlog_end = 0;
 286         shm_toc_insert(pcxt->toc, PARALLEL_KEY_FIXED, fps);
 287
 288         /* We can skip the rest of this if we're not budgeting for any workers. */
 289         if (pcxt->nworkers > 0)
 290         {
 291                 char       *libraryspace;
 292                 char       *gucspace;
 293                 char       *combocidspace;
 294                 char       *tsnapspace;
 295                 char       *asnapspace;
 296                 char       *tstatespace;
 297                 char       *error_queue_space;
 298                 char       *entrypointstate;
 299                 Size            lnamelen;
 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                                                          mul_size(PARALLEL_ERROR_QUEUE_SIZE,
 343                                                                           pcxt->nworkers));
 344                 for (i = 0; i < pcxt->nworkers; ++i)
 345                 {
 346                         char       *start;
 347                         shm_mq     *mq;
 348
 349                         start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
 350                         mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
 351                         shm_mq_set_receiver(mq, MyProc);
 352                         pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
 353                 }
 354                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE, error_queue_space);
 355
 356                 /*
 357                  * Serialize entrypoint information.  It's unsafe to pass function
 358                  * pointers across processes, as the function pointer may be different
 359                  * in each process in EXEC_BACKEND builds, so we always pass library
 360                  * and function name.  (We use library name "postgres" for functions
 361                  * in the core backend.)
 362                  */
 363                 lnamelen = strlen(pcxt->library_name);
 364                 entrypointstate = shm_toc_allocate(pcxt->toc, lnamelen +
 365                                                                                    strlen(pcxt->function_name) + 2);
 366                 strcpy(entrypointstate, pcxt->library_name);
 367                 strcpy(entrypointstate + lnamelen + 1, pcxt->function_name);
 368                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ENTRYPOINT, entrypointstate);
 369         }
 370
 371         /* Restore previous memory context. */
 372         MemoryContextSwitchTo(oldcontext);
 373 }
 374
 375 /*
 376  * Reinitialize the dynamic shared memory segment for a parallel context such
 377  * that we could launch workers for it again.
 378  */
 379 void
 380 ReinitializeParallelDSM(ParallelContext *pcxt)
 381 {
 382         FixedParallelState *fps;
 383         char       *error_queue_space;
 384         int                     i;
 385
 386         /* Wait for any old workers to exit. */
 387         if (pcxt->nworkers_launched > 0)
 388         {
 389                 WaitForParallelWorkersToFinish(pcxt);
 390                 WaitForParallelWorkersToExit(pcxt);
 391                 pcxt->nworkers_launched = 0;
 392         }
 393
 394         /* Reset a few bits of fixed parallel state to a clean state. */
 395         fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED, false);
 396         fps->last_xlog_end = 0;
 397
 398         /* Recreate error queues. */
 399         error_queue_space =
 400                 shm_toc_lookup(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE, false);
 401         for (i = 0; i < pcxt->nworkers; ++i)
 402         {
 403                 char       *start;
 404                 shm_mq     *mq;
 405
 406                 start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
 407                 mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
 408                 shm_mq_set_receiver(mq, MyProc);
 409                 pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
 410         }
 411 }
 412
 413 /*
 414  * Launch parallel workers.
 415  */
 416 void
 417 LaunchParallelWorkers(ParallelContext *pcxt)
 418 {
 419         MemoryContext oldcontext;
 420         BackgroundWorker worker;
 421         int                     i;
 422         bool            any_registrations_failed = false;
 423
 424         /* Skip this if we have no workers. */
 425         if (pcxt->nworkers == 0)
 426                 return;
 427
 428         /* We need to be a lock group leader. */
 429         BecomeLockGroupLeader();
 430
 431         /* If we do have workers, we'd better have a DSM segment. */
 432         Assert(pcxt->seg != NULL);
 433
 434         /* We might be running in a short-lived memory context. */
 435         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 436
 437         /* Configure a worker. */
 438         memset(&worker, 0, sizeof(worker));
 439         snprintf(worker.bgw_name, BGW_MAXLEN, "parallel worker for PID %d",
 440                          MyProcPid);
 441         worker.bgw_flags =
 442                 BGWORKER_SHMEM_ACCESS | BGWORKER_BACKEND_DATABASE_CONNECTION
 443                 | BGWORKER_CLASS_PARALLEL;
 444         worker.bgw_start_time = BgWorkerStart_ConsistentState;
 445         worker.bgw_restart_time = BGW_NEVER_RESTART;
 446         sprintf(worker.bgw_library_name, "postgres");
 447         sprintf(worker.bgw_function_name, "ParallelWorkerMain");
 448         worker.bgw_main_arg = UInt32GetDatum(dsm_segment_handle(pcxt->seg));
 449         worker.bgw_notify_pid = MyProcPid;
 450
 451         /*
 452          * Start workers.
 453          *
 454          * The caller must be able to tolerate ending up with fewer workers than
 455          * expected, so there is no need to throw an error here if registration
 456          * fails.  It wouldn't help much anyway, because registering the worker in
 457          * no way guarantees that it will start up and initialize successfully.
 458          */
 459         for (i = 0; i < pcxt->nworkers; ++i)
 460         {
 461                 memcpy(worker.bgw_extra, &i, sizeof(int));
 462                 if (!any_registrations_failed &&
 463                         RegisterDynamicBackgroundWorker(&worker,
 464                                                                                         &pcxt->worker[i].bgwhandle))
 465                 {
 466                         shm_mq_set_handle(pcxt->worker[i].error_mqh,
 467                                                           pcxt->worker[i].bgwhandle);
 468                         pcxt->nworkers_launched++;
 469                 }
 470                 else
 471                 {
 472                         /*
 473                          * If we weren't able to register the worker, then we've bumped up
 474                          * against the max_worker_processes limit, and future
 475                          * registrations will probably fail too, so arrange to skip them.
 476                          * But we still have to execute this code for the remaining slots
 477                          * to make sure that we forget about the error queues we budgeted
 478                          * for those workers.  Otherwise, we'll wait for them to start,
 479                          * but they never will.
 480                          */
 481                         any_registrations_failed = true;
 482                         pcxt->worker[i].bgwhandle = NULL;
 483                         shm_mq_detach(pcxt->worker[i].error_mqh);
 484                         pcxt->worker[i].error_mqh = NULL;
 485                 }
 486         }
 487
 488         /* Restore previous memory context. */
 489         MemoryContextSwitchTo(oldcontext);
 490 }
 491
 492 /*
 493  * Wait for all workers to finish computing.
 494  *
 495  * Even if the parallel operation seems to have completed successfully, it's
 496  * important to call this function afterwards.  We must not miss any errors
 497  * the workers may have thrown during the parallel operation, or any that they
 498  * may yet throw while shutting down.
 499  *
 500  * Also, we want to update our notion of XactLastRecEnd based on worker
 501  * feedback.
 502  */
 503 void
 504 WaitForParallelWorkersToFinish(ParallelContext *pcxt)
 505 {
 506         for (;;)
 507         {
 508                 bool            anyone_alive = false;
 509                 int                     i;
 510
 511                 /*
 512                  * This will process any parallel messages that are pending, which may
 513                  * change the outcome of the loop that follows.  It may also throw an
 514                  * error propagated from a worker.
 515                  */
 516                 CHECK_FOR_INTERRUPTS();
 517
 518                 for (i = 0; i < pcxt->nworkers_launched; ++i)
 519                 {
 520                         if (pcxt->worker[i].error_mqh != NULL)
 521                         {
 522                                 anyone_alive = true;
 523                                 break;
 524                         }
 525                 }
 526
 527                 if (!anyone_alive)
 528                         break;
 529
 530                 WaitLatch(MyLatch, WL_LATCH_SET, -1,
 531                                   WAIT_EVENT_PARALLEL_FINISH);
 532                 ResetLatch(MyLatch);
 533         }
 534
 535         if (pcxt->toc != NULL)
 536         {
 537                 FixedParallelState *fps;
 538
 539                 fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED, false);
 540                 if (fps->last_xlog_end > XactLastRecEnd)
 541                         XactLastRecEnd = fps->last_xlog_end;
 542         }
 543 }
 544
 545 /*
 546  * Wait for all workers to exit.
 547  *
 548  * This function ensures that workers have been completely shutdown.  The
 549  * difference between WaitForParallelWorkersToFinish and this function is
 550  * that former just ensures that last message sent by worker backend is
 551  * received by master backend whereas this ensures the complete shutdown.
 552  */
 553 static void
 554 WaitForParallelWorkersToExit(ParallelContext *pcxt)
 555 {
 556         int                     i;
 557
 558         /* Wait until the workers actually die. */
 559         for (i = 0; i < pcxt->nworkers_launched; ++i)
 560         {
 561                 BgwHandleStatus status;
 562
 563                 if (pcxt->worker == NULL || pcxt->worker[i].bgwhandle == NULL)
 564                         continue;
 565
 566                 status = WaitForBackgroundWorkerShutdown(pcxt->worker[i].bgwhandle);
 567
 568                 /*
 569                  * If the postmaster kicked the bucket, we have no chance of cleaning
 570                  * up safely -- we won't be able to tell when our workers are actually
 571                  * dead.  This doesn't necessitate a PANIC since they will all abort
 572                  * eventually, but we can't safely continue this session.
 573                  */
 574                 if (status == BGWH_POSTMASTER_DIED)
 575                         ereport(FATAL,
 576                                         (errcode(ERRCODE_ADMIN_SHUTDOWN),
 577                                          errmsg("postmaster exited during a parallel transaction")));
 578
 579                 /* Release memory. */
 580                 pfree(pcxt->worker[i].bgwhandle);
 581                 pcxt->worker[i].bgwhandle = NULL;
 582         }
 583 }
 584
 585 /*
 586  * Destroy a parallel context.
 587  *
 588  * If expecting a clean exit, you should use WaitForParallelWorkersToFinish()
 589  * first, before calling this function.  When this function is invoked, any
 590  * remaining workers are forcibly killed; the dynamic shared memory segment
 591  * is unmapped; and we then wait (uninterruptibly) for the workers to exit.
 592  */
 593 void
 594 DestroyParallelContext(ParallelContext *pcxt)
 595 {
 596         int                     i;
 597
 598         /*
 599          * Be careful about order of operations here!  We remove the parallel
 600          * context from the list before we do anything else; otherwise, if an
 601          * error occurs during a subsequent step, we might try to nuke it again
 602          * from AtEOXact_Parallel or AtEOSubXact_Parallel.
 603          */
 604         dlist_delete(&pcxt->node);
 605
 606         /* Kill each worker in turn, and forget their error queues. */
 607         if (pcxt->worker != NULL)
 608         {
 609                 for (i = 0; i < pcxt->nworkers_launched; ++i)
 610                 {
 611                         if (pcxt->worker[i].error_mqh != NULL)
 612                         {
 613                                 TerminateBackgroundWorker(pcxt->worker[i].bgwhandle);
 614
 615                                 shm_mq_detach(pcxt->worker[i].error_mqh);
 616                                 pcxt->worker[i].error_mqh = NULL;
 617                         }
 618                 }
 619         }
 620
 621         /*
 622          * If we have allocated a shared memory segment, detach it.  This will
 623          * implicitly detach the error queues, and any other shared memory queues,
 624          * stored there.
 625          */
 626         if (pcxt->seg != NULL)
 627         {
 628                 dsm_detach(pcxt->seg);
 629                 pcxt->seg = NULL;
 630         }
 631
 632         /*
 633          * If this parallel context is actually in backend-private memory rather
 634          * than shared memory, free that memory instead.
 635          */
 636         if (pcxt->private_memory != NULL)
 637         {
 638                 pfree(pcxt->private_memory);
 639                 pcxt->private_memory = NULL;
 640         }
 641
 642         /*
 643          * We can't finish transaction commit or abort until all of the workers
 644          * have exited.  This means, in particular, that we can't respond to
 645          * interrupts at this stage.
 646          */
 647         HOLD_INTERRUPTS();
 648         WaitForParallelWorkersToExit(pcxt);
 649         RESUME_INTERRUPTS();
 650
 651         /* Free the worker array itself. */
 652         if (pcxt->worker != NULL)
 653         {
 654                 pfree(pcxt->worker);
 655                 pcxt->worker = NULL;
 656         }
 657
 658         /* Free memory. */
 659         pfree(pcxt->library_name);
 660         pfree(pcxt->function_name);
 661         pfree(pcxt);
 662 }
 663
 664 /*
 665  * Are there any parallel contexts currently active?
 666  */
 667 bool
 668 ParallelContextActive(void)
 669 {
 670         return !dlist_is_empty(&pcxt_list);
 671 }
 672
 673 /*
 674  * Handle receipt of an interrupt indicating a parallel worker message.
 675  *
 676  * Note: this is called within a signal handler!  All we can do is set
 677  * a flag that will cause the next CHECK_FOR_INTERRUPTS() to invoke
 678  * HandleParallelMessages().
 679  */
 680 void
 681 HandleParallelMessageInterrupt(void)
 682 {
 683         InterruptPending = true;
 684         ParallelMessagePending = true;
 685         SetLatch(MyLatch);
 686 }
 687
 688 /*
 689  * Handle any queued protocol messages received from parallel workers.
 690  */
 691 void
 692 HandleParallelMessages(void)
 693 {
 694         dlist_iter      iter;
 695         MemoryContext oldcontext;
 696
 697         static MemoryContext hpm_context = NULL;
 698
 699         /*
 700          * This is invoked from ProcessInterrupts(), and since some of the
 701          * functions it calls contain CHECK_FOR_INTERRUPTS(), there is a potential
 702          * for recursive calls if more signals are received while this runs.  It's
 703          * unclear that recursive entry would be safe, and it doesn't seem useful
 704          * even if it is safe, so let's block interrupts until done.
 705          */
 706         HOLD_INTERRUPTS();
 707
 708         /*
 709          * Moreover, CurrentMemoryContext might be pointing almost anywhere.  We
 710          * don't want to risk leaking data into long-lived contexts, so let's do
 711          * our work here in a private context that we can reset on each use.
 712          */
 713         if (hpm_context == NULL)        /* first time through? */
 714                 hpm_context = AllocSetContextCreate(TopMemoryContext,
 715                                                                                         "HandleParallelMessages",
 716                                                                                         ALLOCSET_DEFAULT_SIZES);
 717         else
 718                 MemoryContextReset(hpm_context);
 719
 720         oldcontext = MemoryContextSwitchTo(hpm_context);
 721
 722         /* OK to process messages.  Reset the flag saying there are more to do. */
 723         ParallelMessagePending = false;
 724
 725         dlist_foreach(iter, &pcxt_list)
 726         {
 727                 ParallelContext *pcxt;
 728                 int                     i;
 729
 730                 pcxt = dlist_container(ParallelContext, node, iter.cur);
 731                 if (pcxt->worker == NULL)
 732                         continue;
 733
 734                 for (i = 0; i < pcxt->nworkers_launched; ++i)
 735                 {
 736                         /*
 737                          * Read as many messages as we can from each worker, but stop when
 738                          * either (1) the worker's error queue goes away, which can happen
 739                          * if we receive a Terminate message from the worker; or (2) no
 740                          * more messages can be read from the worker without blocking.
 741                          */
 742                         while (pcxt->worker[i].error_mqh != NULL)
 743                         {
 744                                 shm_mq_result res;
 745                                 Size            nbytes;
 746                                 void       *data;
 747
 748                                 res = shm_mq_receive(pcxt->worker[i].error_mqh, &nbytes,
 749                                                                          &data, true);
 750                                 if (res == SHM_MQ_WOULD_BLOCK)
 751                                         break;
 752                                 else if (res == SHM_MQ_SUCCESS)
 753                                 {
 754                                         StringInfoData msg;
 755
 756                                         initStringInfo(&msg);
 757                                         appendBinaryStringInfo(&msg, data, nbytes);
 758                                         HandleParallelMessage(pcxt, i, &msg);
 759                                         pfree(msg.data);
 760                                 }
 761                                 else
 762                                         ereport(ERROR,
 763                                                         (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 764                                                          errmsg("lost connection to parallel worker")));
 765                         }
 766                 }
 767         }
 768
 769         MemoryContextSwitchTo(oldcontext);
 770
 771         /* Might as well clear the context on our way out */
 772         MemoryContextReset(hpm_context);
 773
 774         RESUME_INTERRUPTS();
 775 }
 776
 777 /*
 778  * Handle a single protocol message received from a single parallel worker.
 779  */
 780 static void
 781 HandleParallelMessage(ParallelContext *pcxt, int i, StringInfo msg)
 782 {
 783         char            msgtype;
 784
 785         msgtype = pq_getmsgbyte(msg);
 786
 787         switch (msgtype)
 788         {
 789                 case 'K':                               /* BackendKeyData */
 790                         {
 791                                 int32           pid = pq_getmsgint(msg, 4);
 792
 793                                 (void) pq_getmsgint(msg, 4);    /* discard cancel key */
 794                                 (void) pq_getmsgend(msg);
 795                                 pcxt->worker[i].pid = pid;
 796                                 break;
 797                         }
 798
 799                 case 'E':                               /* ErrorResponse */
 800                 case 'N':                               /* NoticeResponse */
 801                         {
 802                                 ErrorData       edata;
 803                                 ErrorContextCallback *save_error_context_stack;
 804
 805                                 /* Parse ErrorResponse or NoticeResponse. */
 806                                 pq_parse_errornotice(msg, &edata);
 807
 808                                 /* Death of a worker isn't enough justification for suicide. */
 809                                 edata.elevel = Min(edata.elevel, ERROR);
 810
 811                                 /*
 812                                  * If desired, add a context line to show that this is a
 813                                  * message propagated from a parallel worker.  Otherwise, it
 814                                  * can sometimes be confusing to understand what actually
 815                                  * happened.  (We don't do this in FORCE_PARALLEL_REGRESS mode
 816                                  * because it causes test-result instability depending on
 817                                  * whether a parallel worker is actually used or not.)
 818                                  */
 819                                 if (force_parallel_mode != FORCE_PARALLEL_REGRESS)
 820                                 {
 821                                         if (edata.context)
 822                                                 edata.context = psprintf("%s\n%s", edata.context,
 823                                                                                                  _("parallel worker"));
 824                                         else
 825                                                 edata.context = pstrdup(_("parallel worker"));
 826                                 }
 827
 828                                 /*
 829                                  * Context beyond that should use the error context callbacks
 830                                  * that were in effect when the ParallelContext was created,
 831                                  * not the current ones.
 832                                  */
 833                                 save_error_context_stack = error_context_stack;
 834                                 error_context_stack = pcxt->error_context_stack;
 835
 836                                 /* Rethrow error or print notice. */
 837                                 ThrowErrorData(&edata);
 838
 839                                 /* Not an error, so restore previous context stack. */
 840                                 error_context_stack = save_error_context_stack;
 841
 842                                 break;
 843                         }
 844
 845                 case 'A':                               /* NotifyResponse */
 846                         {
 847                                 /* Propagate NotifyResponse. */
 848                                 int32           pid;
 849                                 const char *channel;
 850                                 const char *payload;
 851
 852                                 pid = pq_getmsgint(msg, 4);
 853                                 channel = pq_getmsgrawstring(msg);
 854                                 payload = pq_getmsgrawstring(msg);
 855                                 pq_endmessage(msg);
 856
 857                                 NotifyMyFrontEnd(channel, payload, pid);
 858
 859                                 break;
 860                         }
 861
 862                 case 'X':                               /* Terminate, indicating clean exit */
 863                         {
 864                                 shm_mq_detach(pcxt->worker[i].error_mqh);
 865                                 pcxt->worker[i].error_mqh = NULL;
 866                                 break;
 867                         }
 868
 869                 default:
 870                         {
 871                                 elog(ERROR, "unrecognized message type received from parallel worker: %c (message length %d bytes)",
 872                                          msgtype, msg->len);
 873                         }
 874         }
 875 }
 876
 877 /*
 878  * End-of-subtransaction cleanup for parallel contexts.
 879  *
 880  * Currently, it's forbidden to enter or leave a subtransaction while
 881  * parallel mode is in effect, so we could just blow away everything.  But
 882  * we may want to relax that restriction in the future, so this code
 883  * contemplates that there may be multiple subtransaction IDs in pcxt_list.
 884  */
 885 void
 886 AtEOSubXact_Parallel(bool isCommit, SubTransactionId mySubId)
 887 {
 888         while (!dlist_is_empty(&pcxt_list))
 889         {
 890                 ParallelContext *pcxt;
 891
 892                 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
 893                 if (pcxt->subid != mySubId)
 894                         break;
 895                 if (isCommit)
 896                         elog(WARNING, "leaked parallel context");
 897                 DestroyParallelContext(pcxt);
 898         }
 899 }
 900
 901 /*
 902  * End-of-transaction cleanup for parallel contexts.
 903  */
 904 void
 905 AtEOXact_Parallel(bool isCommit)
 906 {
 907         while (!dlist_is_empty(&pcxt_list))
 908         {
 909                 ParallelContext *pcxt;
 910
 911                 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
 912                 if (isCommit)
 913                         elog(WARNING, "leaked parallel context");
 914                 DestroyParallelContext(pcxt);
 915         }
 916 }
 917
 918 /*
 919  * Main entrypoint for parallel workers.
 920  */
 921 void
 922 ParallelWorkerMain(Datum main_arg)
 923 {
 924         dsm_segment *seg;
 925         shm_toc    *toc;
 926         FixedParallelState *fps;
 927         char       *error_queue_space;
 928         shm_mq     *mq;
 929         shm_mq_handle *mqh;
 930         char       *libraryspace;
 931         char       *entrypointstate;
 932         char       *library_name;
 933         char       *function_name;
 934         parallel_worker_main_type entrypt;
 935         char       *gucspace;
 936         char       *combocidspace;
 937         char       *tsnapspace;
 938         char       *asnapspace;
 939         char       *tstatespace;
 940         StringInfoData msgbuf;
 941
 942         /* Set flag to indicate that we're initializing a parallel worker. */
 943         InitializingParallelWorker = true;
 944
 945         /* Establish signal handlers. */
 946         pqsignal(SIGTERM, die);
 947         BackgroundWorkerUnblockSignals();
 948
 949         /* Determine and set our parallel worker number. */
 950         Assert(ParallelWorkerNumber == -1);
 951         memcpy(&ParallelWorkerNumber, MyBgworkerEntry->bgw_extra, sizeof(int));
 952
 953         /* Set up a memory context and resource owner. */
 954         Assert(CurrentResourceOwner == NULL);
 955         CurrentResourceOwner = ResourceOwnerCreate(NULL, "parallel toplevel");
 956         CurrentMemoryContext = AllocSetContextCreate(TopMemoryContext,
 957                                                                                                  "Parallel worker",
 958                                                                                                  ALLOCSET_DEFAULT_SIZES);
 959
 960         /*
 961          * Now that we have a resource owner, we can attach to the dynamic shared
 962          * memory segment and read the table of contents.
 963          */
 964         seg = dsm_attach(DatumGetUInt32(main_arg));
 965         if (seg == NULL)
 966                 ereport(ERROR,
 967                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 968                                  errmsg("could not map dynamic shared memory segment")));
 969         toc = shm_toc_attach(PARALLEL_MAGIC, dsm_segment_address(seg));
 970         if (toc == NULL)
 971                 ereport(ERROR,
 972                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 973                                  errmsg("invalid magic number in dynamic shared memory segment")));
 974
 975         /* Look up fixed parallel state. */
 976         fps = shm_toc_lookup(toc, PARALLEL_KEY_FIXED, false);
 977         MyFixedParallelState = fps;
 978
 979         /*
 980          * Now that we have a worker number, we can find and attach to the error
 981          * queue provided for us.  That's good, because until we do that, any
 982          * errors that happen here will not be reported back to the process that
 983          * requested that this worker be launched.
 984          */
 985         error_queue_space = shm_toc_lookup(toc, PARALLEL_KEY_ERROR_QUEUE, false);
 986         mq = (shm_mq *) (error_queue_space +
 987                                          ParallelWorkerNumber * PARALLEL_ERROR_QUEUE_SIZE);
 988         shm_mq_set_sender(mq, MyProc);
 989         mqh = shm_mq_attach(mq, seg, NULL);
 990         pq_redirect_to_shm_mq(seg, mqh);
 991         pq_set_parallel_master(fps->parallel_master_pid,
 992                                                    fps->parallel_master_backend_id);
 993
 994         /*
 995          * Send a BackendKeyData message to the process that initiated parallelism
 996          * so that it has access to our PID before it receives any other messages
 997          * from us.  Our cancel key is sent, too, since that's the way the
 998          * protocol message is defined, but it won't actually be used for anything
 999          * in this case.
1000          */
1001         pq_beginmessage(&msgbuf, 'K');
1002         pq_sendint(&msgbuf, (int32) MyProcPid, sizeof(int32));
1003         pq_sendint(&msgbuf, (int32) MyCancelKey, sizeof(int32));
1004         pq_endmessage(&msgbuf);
1005
1006         /*
1007          * Hooray! Primary initialization is complete.  Now, we need to set up our
1008          * backend-local state to match the original backend.
1009          */
1010
1011         /*
1012          * Join locking group.  We must do this before anything that could try to
1013          * acquire a heavyweight lock, because any heavyweight locks acquired to
1014          * this point could block either directly against the parallel group
1015          * leader or against some process which in turn waits for a lock that
1016          * conflicts with the parallel group leader, causing an undetected
1017          * deadlock.  (If we can't join the lock group, the leader has gone away,
1018          * so just exit quietly.)
1019          */
1020         if (!BecomeLockGroupMember(fps->parallel_master_pgproc,
1021                                                            fps->parallel_master_pid))
1022                 return;
1023
1024         /*
1025          * Load libraries that were loaded by original backend.  We want to do
1026          * this before restoring GUCs, because the libraries might define custom
1027          * variables.
1028          */
1029         libraryspace = shm_toc_lookup(toc, PARALLEL_KEY_LIBRARY, false);
1030         RestoreLibraryState(libraryspace);
1031
1032         /*
1033          * Identify the entry point to be called.  In theory this could result in
1034          * loading an additional library, though most likely the entry point is in
1035          * the core backend or in a library we just loaded.
1036          */
1037         entrypointstate = shm_toc_lookup(toc, PARALLEL_KEY_ENTRYPOINT, false);
1038         library_name = entrypointstate;
1039         function_name = entrypointstate + strlen(library_name) + 1;
1040
1041         entrypt = LookupParallelWorkerFunction(library_name, function_name);
1042
1043         /* Restore database connection. */
1044         BackgroundWorkerInitializeConnectionByOid(fps->database_id,
1045                                                                                           fps->authenticated_user_id);
1046
1047         /*
1048          * Set the client encoding to the database encoding, since that is what
1049          * the leader will expect.
1050          */
1051         SetClientEncoding(GetDatabaseEncoding());
1052
1053         /* Restore GUC values from launching backend. */
1054         gucspace = shm_toc_lookup(toc, PARALLEL_KEY_GUC, false);
1055         StartTransactionCommand();
1056         RestoreGUCState(gucspace);
1057         CommitTransactionCommand();
1058
1059         /* Crank up a transaction state appropriate to a parallel worker. */
1060         tstatespace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_STATE, false);
1061         StartParallelWorkerTransaction(tstatespace);
1062
1063         /* Restore combo CID state. */
1064         combocidspace = shm_toc_lookup(toc, PARALLEL_KEY_COMBO_CID, false);
1065         RestoreComboCIDState(combocidspace);
1066
1067         /* Restore transaction snapshot. */
1068         tsnapspace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT, false);
1069         RestoreTransactionSnapshot(RestoreSnapshot(tsnapspace),
1070                                                            fps->parallel_master_pgproc);
1071
1072         /* Restore active snapshot. */
1073         asnapspace = shm_toc_lookup(toc, PARALLEL_KEY_ACTIVE_SNAPSHOT, false);
1074         PushActiveSnapshot(RestoreSnapshot(asnapspace));
1075
1076         /*
1077          * We've changed which tuples we can see, and must therefore invalidate
1078          * system caches.
1079          */
1080         InvalidateSystemCaches();
1081
1082         /* Restore user ID and security context. */
1083         SetUserIdAndSecContext(fps->current_user_id, fps->sec_context);
1084
1085         /* Restore temp-namespace state to ensure search path matches leader's. */
1086         SetTempNamespaceState(fps->temp_namespace_id,
1087                                                   fps->temp_toast_namespace_id);
1088
1089         /* Set ParallelMasterBackendId so we know how to address temp relations. */
1090         ParallelMasterBackendId = fps->parallel_master_backend_id;
1091
1092         /*
1093          * We've initialized all of our state now; nothing should change
1094          * hereafter.
1095          */
1096         InitializingParallelWorker = false;
1097         EnterParallelMode();
1098
1099         /*
1100          * Time to do the real work: invoke the caller-supplied code.
1101          */
1102         entrypt(seg, toc);
1103
1104         /* Must exit parallel mode to pop active snapshot. */
1105         ExitParallelMode();
1106
1107         /* Must pop active snapshot so resowner.c doesn't complain. */
1108         PopActiveSnapshot();
1109
1110         /* Shut down the parallel-worker transaction. */
1111         EndParallelWorkerTransaction();
1112
1113         /* Report success. */
1114         pq_putmessage('X', NULL, 0);
1115 }
1116
1117 /*
1118  * Update shared memory with the ending location of the last WAL record we
1119  * wrote, if it's greater than the value already stored there.
1120  */
1121 void
1122 ParallelWorkerReportLastRecEnd(XLogRecPtr last_xlog_end)
1123 {
1124         FixedParallelState *fps = MyFixedParallelState;
1125
1126         Assert(fps != NULL);
1127         SpinLockAcquire(&fps->mutex);
1128         if (fps->last_xlog_end < last_xlog_end)
1129                 fps->last_xlog_end = last_xlog_end;
1130         SpinLockRelease(&fps->mutex);
1131 }
1132
1133 /*
1134  * Look up (and possibly load) a parallel worker entry point function.
1135  *
1136  * For functions contained in the core code, we use library name "postgres"
1137  * and consult the InternalParallelWorkers array.  External functions are
1138  * looked up, and loaded if necessary, using load_external_function().
1139  *
1140  * The point of this is to pass function names as strings across process
1141  * boundaries.  We can't pass actual function addresses because of the
1142  * possibility that the function has been loaded at a different address
1143  * in a different process.  This is obviously a hazard for functions in
1144  * loadable libraries, but it can happen even for functions in the core code
1145  * on platforms using EXEC_BACKEND (e.g., Windows).
1146  *
1147  * At some point it might be worthwhile to get rid of InternalParallelWorkers[]
1148  * in favor of applying load_external_function() for core functions too;
1149  * but that raises portability issues that are not worth addressing now.
1150  */
1151 static parallel_worker_main_type
1152 LookupParallelWorkerFunction(const char *libraryname, const char *funcname)
1153 {
1154         /*
1155          * If the function is to be loaded from postgres itself, search the
1156          * InternalParallelWorkers array.
1157          */
1158         if (strcmp(libraryname, "postgres") == 0)
1159         {
1160                 int                     i;
1161
1162                 for (i = 0; i < lengthof(InternalParallelWorkers); i++)
1163                 {
1164                         if (strcmp(InternalParallelWorkers[i].fn_name, funcname) == 0)
1165                                 return InternalParallelWorkers[i].fn_addr;
1166                 }
1167
1168                 /* We can only reach this by programming error. */
1169                 elog(ERROR, "internal function \"%s\" not found", funcname);
1170         }
1171
1172         /* Otherwise load from external library. */
1173         return (parallel_worker_main_type)
1174                 load_external_function(libraryname, funcname, true, NULL);
1175 }