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) 1996-2018, 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/nbtree.h"
  18 #include "access/parallel.h"
  19 #include "access/session.h"
  20 #include "access/xact.h"
  21 #include "access/xlog.h"
  22 #include "catalog/index.h"
  23 #include "catalog/namespace.h"
  24 #include "commands/async.h"
  25 #include "executor/execParallel.h"
  26 #include "libpq/libpq.h"
  27 #include "libpq/pqformat.h"
  28 #include "libpq/pqmq.h"
  29 #include "miscadmin.h"
  30 #include "optimizer/planmain.h"
  31 #include "pgstat.h"
  32 #include "storage/ipc.h"
  33 #include "storage/sinval.h"
  34 #include "storage/spin.h"
  35 #include "tcop/tcopprot.h"
  36 #include "utils/combocid.h"
  37 #include "utils/guc.h"
  38 #include "utils/inval.h"
  39 #include "utils/memutils.h"
  40 #include "utils/relmapper.h"
  41 #include "utils/snapmgr.h"
  42 #include "utils/typcache.h"
  43
  44
  45 /*
  46  * We don't want to waste a lot of memory on an error queue which, most of
  47  * the time, will process only a handful of small messages.  However, it is
  48  * desirable to make it large enough that a typical ErrorResponse can be sent
  49  * without blocking.  That way, a worker that errors out can write the whole
  50  * message into the queue and terminate without waiting for the user backend.
  51  */
  52 #define PARALLEL_ERROR_QUEUE_SIZE                       16384
  53
  54 /* Magic number for parallel context TOC. */
  55 #define PARALLEL_MAGIC                                          0x50477c7c
  56
  57 /*
  58  * Magic numbers for per-context parallel state sharing.  Higher-level code
  59  * should use smaller values, leaving these very large ones for use by this
  60  * module.
  61  */
  62 #define PARALLEL_KEY_FIXED                                      UINT64CONST(0xFFFFFFFFFFFF0001)
  63 #define PARALLEL_KEY_ERROR_QUEUE                        UINT64CONST(0xFFFFFFFFFFFF0002)
  64 #define PARALLEL_KEY_LIBRARY                            UINT64CONST(0xFFFFFFFFFFFF0003)
  65 #define PARALLEL_KEY_GUC                                        UINT64CONST(0xFFFFFFFFFFFF0004)
  66 #define PARALLEL_KEY_COMBO_CID                          UINT64CONST(0xFFFFFFFFFFFF0005)
  67 #define PARALLEL_KEY_TRANSACTION_SNAPSHOT       UINT64CONST(0xFFFFFFFFFFFF0006)
  68 #define PARALLEL_KEY_ACTIVE_SNAPSHOT            UINT64CONST(0xFFFFFFFFFFFF0007)
  69 #define PARALLEL_KEY_TRANSACTION_STATE          UINT64CONST(0xFFFFFFFFFFFF0008)
  70 #define PARALLEL_KEY_ENTRYPOINT                         UINT64CONST(0xFFFFFFFFFFFF0009)
  71 #define PARALLEL_KEY_SESSION_DSM                        UINT64CONST(0xFFFFFFFFFFFF000A)
  72 #define PARALLEL_KEY_REINDEX_STATE                      UINT64CONST(0xFFFFFFFFFFFF000B)
  73 #define PARALLEL_KEY_RELMAPPER_STATE            UINT64CONST(0xFFFFFFFFFFFF000C)
  74
  75 /* Fixed-size parallel state. */
  76 typedef struct FixedParallelState
  77 {
  78         /* Fixed-size state that workers must restore. */
  79         Oid                     database_id;
  80         Oid                     authenticated_user_id;
  81         Oid                     current_user_id;
  82         Oid                     outer_user_id;
  83         Oid                     temp_namespace_id;
  84         Oid                     temp_toast_namespace_id;
  85         int                     sec_context;
  86         bool            is_superuser;
  87         PGPROC     *parallel_master_pgproc;
  88         pid_t           parallel_master_pid;
  89         BackendId       parallel_master_backend_id;
  90         TimestampTz xact_ts;
  91         TimestampTz stmt_ts;
  92
  93         /* Mutex protects remaining fields. */
  94         slock_t         mutex;
  95
  96         /* Maximum XactLastRecEnd of any worker. */
  97         XLogRecPtr      last_xlog_end;
  98 } FixedParallelState;
  99
 100 /*
 101  * Our parallel worker number.  We initialize this to -1, meaning that we are
 102  * not a parallel worker.  In parallel workers, it will be set to a value >= 0
 103  * and < the number of workers before any user code is invoked; each parallel
 104  * worker will get a different parallel worker number.
 105  */
 106 int                     ParallelWorkerNumber = -1;
 107
 108 /* Is there a parallel message pending which we need to receive? */
 109 volatile bool ParallelMessagePending = false;
 110
 111 /* Are we initializing a parallel worker? */
 112 bool            InitializingParallelWorker = false;
 113
 114 /* Pointer to our fixed parallel state. */
 115 static FixedParallelState *MyFixedParallelState;
 116
 117 /* List of active parallel contexts. */
 118 static dlist_head pcxt_list = DLIST_STATIC_INIT(pcxt_list);
 119
 120 /* Backend-local copy of data from FixedParallelState. */
 121 static pid_t ParallelMasterPid;
 122
 123 /*
 124  * List of internal parallel worker entry points.  We need this for
 125  * reasons explained in LookupParallelWorkerFunction(), below.
 126  */
 127 static const struct
 128 {
 129         const char *fn_name;
 130         parallel_worker_main_type fn_addr;
 131 }                       InternalParallelWorkers[] =
 132
 133 {
 134         {
 135                 "ParallelQueryMain", ParallelQueryMain
 136         },
 137         {
 138                 "_bt_parallel_build_main", _bt_parallel_build_main
 139         }
 140 };
 141
 142 /* Private functions. */
 143 static void HandleParallelMessage(ParallelContext *pcxt, int i, StringInfo msg);
 144 static void WaitForParallelWorkersToExit(ParallelContext *pcxt);
 145 static parallel_worker_main_type LookupParallelWorkerFunction(const char *libraryname, const char *funcname);
 146 static void ParallelWorkerShutdown(int code, Datum arg);
 147
 148
 149 /*
 150  * Establish a new parallel context.  This should be done after entering
 151  * parallel mode, and (unless there is an error) the context should be
 152  * destroyed before exiting the current subtransaction.
 153  */
 154 ParallelContext *
 155 CreateParallelContext(const char *library_name, const char *function_name,
 156                                           int nworkers, bool serializable_okay)
 157 {
 158         MemoryContext oldcontext;
 159         ParallelContext *pcxt;
 160
 161         /* It is unsafe to create a parallel context if not in parallel mode. */
 162         Assert(IsInParallelMode());
 163
 164         /* Number of workers should be non-negative. */
 165         Assert(nworkers >= 0);
 166
 167         /*
 168          * If we are running under serializable isolation, we can't use parallel
 169          * workers, at least not until somebody enhances that mechanism to be
 170          * parallel-aware.  Utility statement callers may ask us to ignore this
 171          * restriction because they're always able to safely ignore the fact that
 172          * SIREAD locks do not work with parallelism.
 173          */
 174         if (IsolationIsSerializable() && !serializable_okay)
 175                 nworkers = 0;
 176
 177         /* We might be running in a short-lived memory context. */
 178         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 179
 180         /* Initialize a new ParallelContext. */
 181         pcxt = palloc0(sizeof(ParallelContext));
 182         pcxt->subid = GetCurrentSubTransactionId();
 183         pcxt->nworkers = nworkers;
 184         pcxt->library_name = pstrdup(library_name);
 185         pcxt->function_name = pstrdup(function_name);
 186         pcxt->error_context_stack = error_context_stack;
 187         shm_toc_initialize_estimator(&pcxt->estimator);
 188         dlist_push_head(&pcxt_list, &pcxt->node);
 189
 190         /* Restore previous memory context. */
 191         MemoryContextSwitchTo(oldcontext);
 192
 193         return pcxt;
 194 }
 195
 196 /*
 197  * Establish the dynamic shared memory segment for a parallel context and
 198  * copy state and other bookkeeping information that will be needed by
 199  * parallel workers into it.
 200  */
 201 void
 202 InitializeParallelDSM(ParallelContext *pcxt)
 203 {
 204         MemoryContext oldcontext;
 205         Size            library_len = 0;
 206         Size            guc_len = 0;
 207         Size            combocidlen = 0;
 208         Size            tsnaplen = 0;
 209         Size            asnaplen = 0;
 210         Size            tstatelen = 0;
 211         Size            reindexlen = 0;
 212         Size            relmapperlen = 0;
 213         Size            segsize = 0;
 214         int                     i;
 215         FixedParallelState *fps;
 216         dsm_handle      session_dsm_handle = DSM_HANDLE_INVALID;
 217         Snapshot        transaction_snapshot = GetTransactionSnapshot();
 218         Snapshot        active_snapshot = GetActiveSnapshot();
 219
 220         /* We might be running in a very short-lived memory context. */
 221         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 222
 223         /* Allow space to store the fixed-size parallel state. */
 224         shm_toc_estimate_chunk(&pcxt->estimator, sizeof(FixedParallelState));
 225         shm_toc_estimate_keys(&pcxt->estimator, 1);
 226
 227         /*
 228          * Normally, the user will have requested at least one worker process, but
 229          * if by chance they have not, we can skip a bunch of things here.
 230          */
 231         if (pcxt->nworkers > 0)
 232         {
 233                 /* Get (or create) the per-session DSM segment's handle. */
 234                 session_dsm_handle = GetSessionDsmHandle();
 235
 236                 /*
 237                  * If we weren't able to create a per-session DSM segment, then we can
 238                  * continue but we can't safely launch any workers because their
 239                  * record typmods would be incompatible so they couldn't exchange
 240                  * tuples.
 241                  */
 242                 if (session_dsm_handle == DSM_HANDLE_INVALID)
 243                         pcxt->nworkers = 0;
 244         }
 245
 246         if (pcxt->nworkers > 0)
 247         {
 248                 /* Estimate space for various kinds of state sharing. */
 249                 library_len = EstimateLibraryStateSpace();
 250                 shm_toc_estimate_chunk(&pcxt->estimator, library_len);
 251                 guc_len = EstimateGUCStateSpace();
 252                 shm_toc_estimate_chunk(&pcxt->estimator, guc_len);
 253                 combocidlen = EstimateComboCIDStateSpace();
 254                 shm_toc_estimate_chunk(&pcxt->estimator, combocidlen);
 255                 tsnaplen = EstimateSnapshotSpace(transaction_snapshot);
 256                 shm_toc_estimate_chunk(&pcxt->estimator, tsnaplen);
 257                 asnaplen = EstimateSnapshotSpace(active_snapshot);
 258                 shm_toc_estimate_chunk(&pcxt->estimator, asnaplen);
 259                 tstatelen = EstimateTransactionStateSpace();
 260                 shm_toc_estimate_chunk(&pcxt->estimator, tstatelen);
 261                 shm_toc_estimate_chunk(&pcxt->estimator, sizeof(dsm_handle));
 262                 reindexlen = EstimateReindexStateSpace();
 263                 shm_toc_estimate_chunk(&pcxt->estimator, reindexlen);
 264                 relmapperlen = EstimateRelationMapSpace();
 265                 shm_toc_estimate_chunk(&pcxt->estimator, relmapperlen);
 266                 /* If you add more chunks here, you probably need to add keys. */
 267                 shm_toc_estimate_keys(&pcxt->estimator, 9);
 268
 269                 /* Estimate space need for error queues. */
 270                 StaticAssertStmt(BUFFERALIGN(PARALLEL_ERROR_QUEUE_SIZE) ==
 271                                                  PARALLEL_ERROR_QUEUE_SIZE,
 272                                                  "parallel error queue size not buffer-aligned");
 273                 shm_toc_estimate_chunk(&pcxt->estimator,
 274                                                            mul_size(PARALLEL_ERROR_QUEUE_SIZE,
 275                                                                                 pcxt->nworkers));
 276                 shm_toc_estimate_keys(&pcxt->estimator, 1);
 277
 278                 /* Estimate how much we'll need for the entrypoint info. */
 279                 shm_toc_estimate_chunk(&pcxt->estimator, strlen(pcxt->library_name) +
 280                                                            strlen(pcxt->function_name) + 2);
 281                 shm_toc_estimate_keys(&pcxt->estimator, 1);
 282         }
 283
 284         /*
 285          * Create DSM and initialize with new table of contents.  But if the user
 286          * didn't request any workers, then don't bother creating a dynamic shared
 287          * memory segment; instead, just use backend-private memory.
 288          *
 289          * Also, if we can't create a dynamic shared memory segment because the
 290          * maximum number of segments have already been created, then fall back to
 291          * backend-private memory, and plan not to use any workers.  We hope this
 292          * won't happen very often, but it's better to abandon the use of
 293          * parallelism than to fail outright.
 294          */
 295         segsize = shm_toc_estimate(&pcxt->estimator);
 296         if (pcxt->nworkers > 0)
 297                 pcxt->seg = dsm_create(segsize, DSM_CREATE_NULL_IF_MAXSEGMENTS);
 298         if (pcxt->seg != NULL)
 299                 pcxt->toc = shm_toc_create(PARALLEL_MAGIC,
 300                                                                    dsm_segment_address(pcxt->seg),
 301                                                                    segsize);
 302         else
 303         {
 304                 pcxt->nworkers = 0;
 305                 pcxt->private_memory = MemoryContextAlloc(TopMemoryContext, segsize);
 306                 pcxt->toc = shm_toc_create(PARALLEL_MAGIC, pcxt->private_memory,
 307                                                                    segsize);
 308         }
 309
 310         /* Initialize fixed-size state in shared memory. */
 311         fps = (FixedParallelState *)
 312                 shm_toc_allocate(pcxt->toc, sizeof(FixedParallelState));
 313         fps->database_id = MyDatabaseId;
 314         fps->authenticated_user_id = GetAuthenticatedUserId();
 315         fps->outer_user_id = GetCurrentRoleId();
 316         fps->is_superuser = session_auth_is_superuser;
 317         GetUserIdAndSecContext(&fps->current_user_id, &fps->sec_context);
 318         GetTempNamespaceState(&fps->temp_namespace_id,
 319                                                   &fps->temp_toast_namespace_id);
 320         fps->parallel_master_pgproc = MyProc;
 321         fps->parallel_master_pid = MyProcPid;
 322         fps->parallel_master_backend_id = MyBackendId;
 323         fps->xact_ts = GetCurrentTransactionStartTimestamp();
 324         fps->stmt_ts = GetCurrentStatementStartTimestamp();
 325         SpinLockInit(&fps->mutex);
 326         fps->last_xlog_end = 0;
 327         shm_toc_insert(pcxt->toc, PARALLEL_KEY_FIXED, fps);
 328
 329         /* We can skip the rest of this if we're not budgeting for any workers. */
 330         if (pcxt->nworkers > 0)
 331         {
 332                 char       *libraryspace;
 333                 char       *gucspace;
 334                 char       *combocidspace;
 335                 char       *tsnapspace;
 336                 char       *asnapspace;
 337                 char       *tstatespace;
 338                 char       *reindexspace;
 339                 char       *relmapperspace;
 340                 char       *error_queue_space;
 341                 char       *session_dsm_handle_space;
 342                 char       *entrypointstate;
 343                 Size            lnamelen;
 344
 345                 /* Serialize shared libraries we have loaded. */
 346                 libraryspace = shm_toc_allocate(pcxt->toc, library_len);
 347                 SerializeLibraryState(library_len, libraryspace);
 348                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_LIBRARY, libraryspace);
 349
 350                 /* Serialize GUC settings. */
 351                 gucspace = shm_toc_allocate(pcxt->toc, guc_len);
 352                 SerializeGUCState(guc_len, gucspace);
 353                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_GUC, gucspace);
 354
 355                 /* Serialize combo CID state. */
 356                 combocidspace = shm_toc_allocate(pcxt->toc, combocidlen);
 357                 SerializeComboCIDState(combocidlen, combocidspace);
 358                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_COMBO_CID, combocidspace);
 359
 360                 /* Serialize transaction snapshot and active snapshot. */
 361                 tsnapspace = shm_toc_allocate(pcxt->toc, tsnaplen);
 362                 SerializeSnapshot(transaction_snapshot, tsnapspace);
 363                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT,
 364                                            tsnapspace);
 365                 asnapspace = shm_toc_allocate(pcxt->toc, asnaplen);
 366                 SerializeSnapshot(active_snapshot, asnapspace);
 367                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ACTIVE_SNAPSHOT, asnapspace);
 368
 369                 /* Provide the handle for per-session segment. */
 370                 session_dsm_handle_space = shm_toc_allocate(pcxt->toc,
 371                                                                                                         sizeof(dsm_handle));
 372                 *(dsm_handle *) session_dsm_handle_space = session_dsm_handle;
 373                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_SESSION_DSM,
 374                                            session_dsm_handle_space);
 375
 376                 /* Serialize transaction state. */
 377                 tstatespace = shm_toc_allocate(pcxt->toc, tstatelen);
 378                 SerializeTransactionState(tstatelen, tstatespace);
 379                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_STATE, tstatespace);
 380
 381                 /* Serialize reindex state. */
 382                 reindexspace = shm_toc_allocate(pcxt->toc, reindexlen);
 383                 SerializeReindexState(reindexlen, reindexspace);
 384                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_REINDEX_STATE, reindexspace);
 385
 386                 /* Serialize relmapper state. */
 387                 relmapperspace = shm_toc_allocate(pcxt->toc, relmapperlen);
 388                 SerializeRelationMap(relmapperlen, relmapperspace);
 389                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_RELMAPPER_STATE,
 390                                            relmapperspace);
 391
 392                 /* Allocate space for worker information. */
 393                 pcxt->worker = palloc0(sizeof(ParallelWorkerInfo) * pcxt->nworkers);
 394
 395                 /*
 396                  * Establish error queues in dynamic shared memory.
 397                  *
 398                  * These queues should be used only for transmitting ErrorResponse,
 399                  * NoticeResponse, and NotifyResponse protocol messages.  Tuple data
 400                  * should be transmitted via separate (possibly larger?) queues.
 401                  */
 402                 error_queue_space =
 403                         shm_toc_allocate(pcxt->toc,
 404                                                          mul_size(PARALLEL_ERROR_QUEUE_SIZE,
 405                                                                           pcxt->nworkers));
 406                 for (i = 0; i < pcxt->nworkers; ++i)
 407                 {
 408                         char       *start;
 409                         shm_mq     *mq;
 410
 411                         start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
 412                         mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
 413                         shm_mq_set_receiver(mq, MyProc);
 414                         pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
 415                 }
 416                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE, error_queue_space);
 417
 418                 /*
 419                  * Serialize entrypoint information.  It's unsafe to pass function
 420                  * pointers across processes, as the function pointer may be different
 421                  * in each process in EXEC_BACKEND builds, so we always pass library
 422                  * and function name.  (We use library name "postgres" for functions
 423                  * in the core backend.)
 424                  */
 425                 lnamelen = strlen(pcxt->library_name);
 426                 entrypointstate = shm_toc_allocate(pcxt->toc, lnamelen +
 427                                                                                    strlen(pcxt->function_name) + 2);
 428                 strcpy(entrypointstate, pcxt->library_name);
 429                 strcpy(entrypointstate + lnamelen + 1, pcxt->function_name);
 430                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ENTRYPOINT, entrypointstate);
 431         }
 432
 433         /* Restore previous memory context. */
 434         MemoryContextSwitchTo(oldcontext);
 435 }
 436
 437 /*
 438  * Reinitialize the dynamic shared memory segment for a parallel context such
 439  * that we could launch workers for it again.
 440  */
 441 void
 442 ReinitializeParallelDSM(ParallelContext *pcxt)
 443 {
 444         FixedParallelState *fps;
 445
 446         /* Wait for any old workers to exit. */
 447         if (pcxt->nworkers_launched > 0)
 448         {
 449                 WaitForParallelWorkersToFinish(pcxt);
 450                 WaitForParallelWorkersToExit(pcxt);
 451                 pcxt->nworkers_launched = 0;
 452                 if (pcxt->known_attached_workers)
 453                 {
 454                         pfree(pcxt->known_attached_workers);
 455                         pcxt->known_attached_workers = NULL;
 456                         pcxt->nknown_attached_workers = 0;
 457                 }
 458         }
 459
 460         /* Reset a few bits of fixed parallel state to a clean state. */
 461         fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED, false);
 462         fps->last_xlog_end = 0;
 463
 464         /* Recreate error queues (if they exist). */
 465         if (pcxt->nworkers > 0)
 466         {
 467                 char       *error_queue_space;
 468                 int                     i;
 469
 470                 error_queue_space =
 471                         shm_toc_lookup(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE, false);
 472                 for (i = 0; i < pcxt->nworkers; ++i)
 473                 {
 474                         char       *start;
 475                         shm_mq     *mq;
 476
 477                         start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
 478                         mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
 479                         shm_mq_set_receiver(mq, MyProc);
 480                         pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
 481                 }
 482         }
 483 }
 484
 485 /*
 486  * Launch parallel workers.
 487  */
 488 void
 489 LaunchParallelWorkers(ParallelContext *pcxt)
 490 {
 491         MemoryContext oldcontext;
 492         BackgroundWorker worker;
 493         int                     i;
 494         bool            any_registrations_failed = false;
 495
 496         /* Skip this if we have no workers. */
 497         if (pcxt->nworkers == 0)
 498                 return;
 499
 500         /* We need to be a lock group leader. */
 501         BecomeLockGroupLeader();
 502
 503         /* If we do have workers, we'd better have a DSM segment. */
 504         Assert(pcxt->seg != NULL);
 505
 506         /* We might be running in a short-lived memory context. */
 507         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 508
 509         /* Configure a worker. */
 510         memset(&worker, 0, sizeof(worker));
 511         snprintf(worker.bgw_name, BGW_MAXLEN, "parallel worker for PID %d",
 512                          MyProcPid);
 513         snprintf(worker.bgw_type, BGW_MAXLEN, "parallel worker");
 514         worker.bgw_flags =
 515                 BGWORKER_SHMEM_ACCESS | BGWORKER_BACKEND_DATABASE_CONNECTION
 516                 | BGWORKER_CLASS_PARALLEL;
 517         worker.bgw_start_time = BgWorkerStart_ConsistentState;
 518         worker.bgw_restart_time = BGW_NEVER_RESTART;
 519         sprintf(worker.bgw_library_name, "postgres");
 520         sprintf(worker.bgw_function_name, "ParallelWorkerMain");
 521         worker.bgw_main_arg = UInt32GetDatum(dsm_segment_handle(pcxt->seg));
 522         worker.bgw_notify_pid = MyProcPid;
 523
 524         /*
 525          * Start workers.
 526          *
 527          * The caller must be able to tolerate ending up with fewer workers than
 528          * expected, so there is no need to throw an error here if registration
 529          * fails.  It wouldn't help much anyway, because registering the worker in
 530          * no way guarantees that it will start up and initialize successfully.
 531          */
 532         for (i = 0; i < pcxt->nworkers; ++i)
 533         {
 534                 memcpy(worker.bgw_extra, &i, sizeof(int));
 535                 if (!any_registrations_failed &&
 536                         RegisterDynamicBackgroundWorker(&worker,
 537                                                                                         &pcxt->worker[i].bgwhandle))
 538                 {
 539                         shm_mq_set_handle(pcxt->worker[i].error_mqh,
 540                                                           pcxt->worker[i].bgwhandle);
 541                         pcxt->nworkers_launched++;
 542                 }
 543                 else
 544                 {
 545                         /*
 546                          * If we weren't able to register the worker, then we've bumped up
 547                          * against the max_worker_processes limit, and future
 548                          * registrations will probably fail too, so arrange to skip them.
 549                          * But we still have to execute this code for the remaining slots
 550                          * to make sure that we forget about the error queues we budgeted
 551                          * for those workers.  Otherwise, we'll wait for them to start,
 552                          * but they never will.
 553                          */
 554                         any_registrations_failed = true;
 555                         pcxt->worker[i].bgwhandle = NULL;
 556                         shm_mq_detach(pcxt->worker[i].error_mqh);
 557                         pcxt->worker[i].error_mqh = NULL;
 558                 }
 559         }
 560
 561         /*
 562          * Now that nworkers_launched has taken its final value, we can initialize
 563          * known_attached_workers.
 564          */
 565         if (pcxt->nworkers_launched > 0)
 566         {
 567                 pcxt->known_attached_workers =
 568                         palloc0(sizeof(bool) * pcxt->nworkers_launched);
 569                 pcxt->nknown_attached_workers = 0;
 570         }
 571
 572         /* Restore previous memory context. */
 573         MemoryContextSwitchTo(oldcontext);
 574 }
 575
 576 /*
 577  * Wait for all workers to attach to their error queues, and throw an error if
 578  * any worker fails to do this.
 579  *
 580  * Callers can assume that if this function returns successfully, then the
 581  * number of workers given by pcxt->nworkers_launched have initialized and
 582  * attached to their error queues.  Whether or not these workers are guaranteed
 583  * to still be running depends on what code the caller asked them to run;
 584  * this function does not guarantee that they have not exited.  However, it
 585  * does guarantee that any workers which exited must have done so cleanly and
 586  * after successfully performing the work with which they were tasked.
 587  *
 588  * If this function is not called, then some of the workers that were launched
 589  * may not have been started due to a fork() failure, or may have exited during
 590  * early startup prior to attaching to the error queue, so nworkers_launched
 591  * cannot be viewed as completely reliable.  It will never be less than the
 592  * number of workers which actually started, but it might be more.  Any workers
 593  * that failed to start will still be discovered by
 594  * WaitForParallelWorkersToFinish and an error will be thrown at that time,
 595  * provided that function is eventually reached.
 596  *
 597  * In general, the leader process should do as much work as possible before
 598  * calling this function.  fork() failures and other early-startup failures
 599  * are very uncommon, and having the leader sit idle when it could be doing
 600  * useful work is undesirable.  However, if the leader needs to wait for
 601  * all of its workers or for a specific worker, it may want to call this
 602  * function before doing so.  If not, it must make some other provision for
 603  * the failure-to-start case, lest it wait forever.  On the other hand, a
 604  * leader which never waits for a worker that might not be started yet, or
 605  * at least never does so prior to WaitForParallelWorkersToFinish(), need not
 606  * call this function at all.
 607  */
 608 void
 609 WaitForParallelWorkersToAttach(ParallelContext *pcxt)
 610 {
 611         int                     i;
 612
 613         /* Skip this if we have no launched workers. */
 614         if (pcxt->nworkers_launched == 0)
 615                 return;
 616
 617         for (;;)
 618         {
 619                 /*
 620                  * This will process any parallel messages that are pending and it may
 621                  * also throw an error propagated from a worker.
 622                  */
 623                 CHECK_FOR_INTERRUPTS();
 624
 625                 for (i = 0; i < pcxt->nworkers_launched; ++i)
 626                 {
 627                         BgwHandleStatus status;
 628                         shm_mq     *mq;
 629                         int                     rc;
 630                         pid_t           pid;
 631
 632                         if (pcxt->known_attached_workers[i])
 633                                 continue;
 634
 635                         /*
 636                          * If error_mqh is NULL, then the worker has already exited
 637                          * cleanly.
 638                          */
 639                         if (pcxt->worker[i].error_mqh == NULL)
 640                         {
 641                                 pcxt->known_attached_workers[i] = true;
 642                                 ++pcxt->nknown_attached_workers;
 643                                 continue;
 644                         }
 645
 646                         status = GetBackgroundWorkerPid(pcxt->worker[i].bgwhandle, &pid);
 647                         if (status == BGWH_STARTED)
 648                         {
 649                                 /* Has the worker attached to the error queue? */
 650                                 mq = shm_mq_get_queue(pcxt->worker[i].error_mqh);
 651                                 if (shm_mq_get_sender(mq) != NULL)
 652                                 {
 653                                         /* Yes, so it is known to be attached. */
 654                                         pcxt->known_attached_workers[i] = true;
 655                                         ++pcxt->nknown_attached_workers;
 656                                 }
 657                         }
 658                         else if (status == BGWH_STOPPED)
 659                         {
 660                                 /*
 661                                  * If the worker stopped without attaching to the error queue,
 662                                  * throw an error.
 663                                  */
 664                                 mq = shm_mq_get_queue(pcxt->worker[i].error_mqh);
 665                                 if (shm_mq_get_sender(mq) == NULL)
 666                                         ereport(ERROR,
 667                                                         (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 668                                                          errmsg("parallel worker failed to initialize"),
 669                                                          errhint("More details may be available in the server log.")));
 670
 671                                 pcxt->known_attached_workers[i] = true;
 672                                 ++pcxt->nknown_attached_workers;
 673                         }
 674                         else
 675                         {
 676                                 /*
 677                                  * Worker not yet started, so we must wait.  The postmaster
 678                                  * will notify us if the worker's state changes.  Our latch
 679                                  * might also get set for some other reason, but if so we'll
 680                                  * just end up waiting for the same worker again.
 681                                  */
 682                                 rc = WaitLatch(MyLatch,
 683                                                            WL_LATCH_SET | WL_POSTMASTER_DEATH,
 684                                                            -1, WAIT_EVENT_BGWORKER_STARTUP);
 685
 686                                 /* emergency bailout if postmaster has died */
 687                                 if (rc & WL_POSTMASTER_DEATH)
 688                                         proc_exit(1);
 689
 690                                 if (rc & WL_LATCH_SET)
 691                                         ResetLatch(MyLatch);
 692                         }
 693                 }
 694
 695                 /* If all workers are known to have started, we're done. */
 696                 if (pcxt->nknown_attached_workers >= pcxt->nworkers_launched)
 697                 {
 698                         Assert(pcxt->nknown_attached_workers == pcxt->nworkers_launched);
 699                         break;
 700                 }
 701         }
 702 }
 703
 704 /*
 705  * Wait for all workers to finish computing.
 706  *
 707  * Even if the parallel operation seems to have completed successfully, it's
 708  * important to call this function afterwards.  We must not miss any errors
 709  * the workers may have thrown during the parallel operation, or any that they
 710  * may yet throw while shutting down.
 711  *
 712  * Also, we want to update our notion of XactLastRecEnd based on worker
 713  * feedback.
 714  */
 715 void
 716 WaitForParallelWorkersToFinish(ParallelContext *pcxt)
 717 {
 718         for (;;)
 719         {
 720                 bool            anyone_alive = false;
 721                 int                     nfinished = 0;
 722                 int                     i;
 723
 724                 /*
 725                  * This will process any parallel messages that are pending, which may
 726                  * change the outcome of the loop that follows.  It may also throw an
 727                  * error propagated from a worker.
 728                  */
 729                 CHECK_FOR_INTERRUPTS();
 730
 731                 for (i = 0; i < pcxt->nworkers_launched; ++i)
 732                 {
 733                         /*
 734                          * If error_mqh is NULL, then the worker has already exited
 735                          * cleanly.  If we have received a message through error_mqh from
 736                          * the worker, we know it started up cleanly, and therefore we're
 737                          * certain to be notified when it exits.
 738                          */
 739                         if (pcxt->worker[i].error_mqh == NULL)
 740                                 ++nfinished;
 741                         else if (pcxt->known_attached_workers[i])
 742                         {
 743                                 anyone_alive = true;
 744                                 break;
 745                         }
 746                 }
 747
 748                 if (!anyone_alive)
 749                 {
 750                         /* If all workers are known to have finished, we're done. */
 751                         if (nfinished >= pcxt->nworkers_launched)
 752                         {
 753                                 Assert(nfinished == pcxt->nworkers_launched);
 754                                 break;
 755                         }
 756
 757                         /*
 758                          * We didn't detect any living workers, but not all workers are
 759                          * known to have exited cleanly.  Either not all workers have
 760                          * launched yet, or maybe some of them failed to start or
 761                          * terminated abnormally.
 762                          */
 763                         for (i = 0; i < pcxt->nworkers_launched; ++i)
 764                         {
 765                                 pid_t           pid;
 766                                 shm_mq     *mq;
 767
 768                                 /*
 769                                  * If the worker is BGWH_NOT_YET_STARTED or BGWH_STARTED, we
 770                                  * should just keep waiting.  If it is BGWH_STOPPED, then
 771                                  * further investigation is needed.
 772                                  */
 773                                 if (pcxt->worker[i].error_mqh == NULL ||
 774                                         pcxt->worker[i].bgwhandle == NULL ||
 775                                         GetBackgroundWorkerPid(pcxt->worker[i].bgwhandle,
 776                                                                                    &pid) != BGWH_STOPPED)
 777                                         continue;
 778
 779                                 /*
 780                                  * Check whether the worker ended up stopped without ever
 781                                  * attaching to the error queue.  If so, the postmaster was
 782                                  * unable to fork the worker or it exited without initializing
 783                                  * properly.  We must throw an error, since the caller may
 784                                  * have been expecting the worker to do some work before
 785                                  * exiting.
 786                                  */
 787                                 mq = shm_mq_get_queue(pcxt->worker[i].error_mqh);
 788                                 if (shm_mq_get_sender(mq) == NULL)
 789                                         ereport(ERROR,
 790                                                         (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 791                                                          errmsg("parallel worker failed to initialize"),
 792                                                          errhint("More details may be available in the server log.")));
 793
 794                                 /*
 795                                  * The worker is stopped, but is attached to the error queue.
 796                                  * Unless there's a bug somewhere, this will only happen when
 797                                  * the worker writes messages and terminates after the
 798                                  * CHECK_FOR_INTERRUPTS() near the top of this function and
 799                                  * before the call to GetBackgroundWorkerPid().  In that case,
 800                                  * or latch should have been set as well and the right things
 801                                  * will happen on the next pass through the loop.
 802                                  */
 803                         }
 804                 }
 805
 806                 WaitLatch(MyLatch, WL_LATCH_SET, -1,
 807                                   WAIT_EVENT_PARALLEL_FINISH);
 808                 ResetLatch(MyLatch);
 809         }
 810
 811         if (pcxt->toc != NULL)
 812         {
 813                 FixedParallelState *fps;
 814
 815                 fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED, false);
 816                 if (fps->last_xlog_end > XactLastRecEnd)
 817                         XactLastRecEnd = fps->last_xlog_end;
 818         }
 819 }
 820
 821 /*
 822  * Wait for all workers to exit.
 823  *
 824  * This function ensures that workers have been completely shutdown.  The
 825  * difference between WaitForParallelWorkersToFinish and this function is
 826  * that former just ensures that last message sent by worker backend is
 827  * received by master backend whereas this ensures the complete shutdown.
 828  */
 829 static void
 830 WaitForParallelWorkersToExit(ParallelContext *pcxt)
 831 {
 832         int                     i;
 833
 834         /* Wait until the workers actually die. */
 835         for (i = 0; i < pcxt->nworkers_launched; ++i)
 836         {
 837                 BgwHandleStatus status;
 838
 839                 if (pcxt->worker == NULL || pcxt->worker[i].bgwhandle == NULL)
 840                         continue;
 841
 842                 status = WaitForBackgroundWorkerShutdown(pcxt->worker[i].bgwhandle);
 843
 844                 /*
 845                  * If the postmaster kicked the bucket, we have no chance of cleaning
 846                  * up safely -- we won't be able to tell when our workers are actually
 847                  * dead.  This doesn't necessitate a PANIC since they will all abort
 848                  * eventually, but we can't safely continue this session.
 849                  */
 850                 if (status == BGWH_POSTMASTER_DIED)
 851                         ereport(FATAL,
 852                                         (errcode(ERRCODE_ADMIN_SHUTDOWN),
 853                                          errmsg("postmaster exited during a parallel transaction")));
 854
 855                 /* Release memory. */
 856                 pfree(pcxt->worker[i].bgwhandle);
 857                 pcxt->worker[i].bgwhandle = NULL;
 858         }
 859 }
 860
 861 /*
 862  * Destroy a parallel context.
 863  *
 864  * If expecting a clean exit, you should use WaitForParallelWorkersToFinish()
 865  * first, before calling this function.  When this function is invoked, any
 866  * remaining workers are forcibly killed; the dynamic shared memory segment
 867  * is unmapped; and we then wait (uninterruptibly) for the workers to exit.
 868  */
 869 void
 870 DestroyParallelContext(ParallelContext *pcxt)
 871 {
 872         int                     i;
 873
 874         /*
 875          * Be careful about order of operations here!  We remove the parallel
 876          * context from the list before we do anything else; otherwise, if an
 877          * error occurs during a subsequent step, we might try to nuke it again
 878          * from AtEOXact_Parallel or AtEOSubXact_Parallel.
 879          */
 880         dlist_delete(&pcxt->node);
 881
 882         /* Kill each worker in turn, and forget their error queues. */
 883         if (pcxt->worker != NULL)
 884         {
 885                 for (i = 0; i < pcxt->nworkers_launched; ++i)
 886                 {
 887                         if (pcxt->worker[i].error_mqh != NULL)
 888                         {
 889                                 TerminateBackgroundWorker(pcxt->worker[i].bgwhandle);
 890
 891                                 shm_mq_detach(pcxt->worker[i].error_mqh);
 892                                 pcxt->worker[i].error_mqh = NULL;
 893                         }
 894                 }
 895         }
 896
 897         /*
 898          * If we have allocated a shared memory segment, detach it.  This will
 899          * implicitly detach the error queues, and any other shared memory queues,
 900          * stored there.
 901          */
 902         if (pcxt->seg != NULL)
 903         {
 904                 dsm_detach(pcxt->seg);
 905                 pcxt->seg = NULL;
 906         }
 907
 908         /*
 909          * If this parallel context is actually in backend-private memory rather
 910          * than shared memory, free that memory instead.
 911          */
 912         if (pcxt->private_memory != NULL)
 913         {
 914                 pfree(pcxt->private_memory);
 915                 pcxt->private_memory = NULL;
 916         }
 917
 918         /*
 919          * We can't finish transaction commit or abort until all of the workers
 920          * have exited.  This means, in particular, that we can't respond to
 921          * interrupts at this stage.
 922          */
 923         HOLD_INTERRUPTS();
 924         WaitForParallelWorkersToExit(pcxt);
 925         RESUME_INTERRUPTS();
 926
 927         /* Free the worker array itself. */
 928         if (pcxt->worker != NULL)
 929         {
 930                 pfree(pcxt->worker);
 931                 pcxt->worker = NULL;
 932         }
 933
 934         /* Free memory. */
 935         pfree(pcxt->library_name);
 936         pfree(pcxt->function_name);
 937         pfree(pcxt);
 938 }
 939
 940 /*
 941  * Are there any parallel contexts currently active?
 942  */
 943 bool
 944 ParallelContextActive(void)
 945 {
 946         return !dlist_is_empty(&pcxt_list);
 947 }
 948
 949 /*
 950  * Handle receipt of an interrupt indicating a parallel worker message.
 951  *
 952  * Note: this is called within a signal handler!  All we can do is set
 953  * a flag that will cause the next CHECK_FOR_INTERRUPTS() to invoke
 954  * HandleParallelMessages().
 955  */
 956 void
 957 HandleParallelMessageInterrupt(void)
 958 {
 959         InterruptPending = true;
 960         ParallelMessagePending = true;
 961         SetLatch(MyLatch);
 962 }
 963
 964 /*
 965  * Handle any queued protocol messages received from parallel workers.
 966  */
 967 void
 968 HandleParallelMessages(void)
 969 {
 970         dlist_iter      iter;
 971         MemoryContext oldcontext;
 972
 973         static MemoryContext hpm_context = NULL;
 974
 975         /*
 976          * This is invoked from ProcessInterrupts(), and since some of the
 977          * functions it calls contain CHECK_FOR_INTERRUPTS(), there is a potential
 978          * for recursive calls if more signals are received while this runs.  It's
 979          * unclear that recursive entry would be safe, and it doesn't seem useful
 980          * even if it is safe, so let's block interrupts until done.
 981          */
 982         HOLD_INTERRUPTS();
 983
 984         /*
 985          * Moreover, CurrentMemoryContext might be pointing almost anywhere.  We
 986          * don't want to risk leaking data into long-lived contexts, so let's do
 987          * our work here in a private context that we can reset on each use.
 988          */
 989         if (hpm_context == NULL)        /* first time through? */
 990                 hpm_context = AllocSetContextCreate(TopMemoryContext,
 991                                                                                         "HandleParallelMessages",
 992                                                                                         ALLOCSET_DEFAULT_SIZES);
 993         else
 994                 MemoryContextReset(hpm_context);
 995
 996         oldcontext = MemoryContextSwitchTo(hpm_context);
 997
 998         /* OK to process messages.  Reset the flag saying there are more to do. */
 999         ParallelMessagePending = false;
1000
1001         dlist_foreach(iter, &pcxt_list)
1002         {
1003                 ParallelContext *pcxt;
1004                 int                     i;
1005
1006                 pcxt = dlist_container(ParallelContext, node, iter.cur);
1007                 if (pcxt->worker == NULL)
1008                         continue;
1009
1010                 for (i = 0; i < pcxt->nworkers_launched; ++i)
1011                 {
1012                         /*
1013                          * Read as many messages as we can from each worker, but stop when
1014                          * either (1) the worker's error queue goes away, which can happen
1015                          * if we receive a Terminate message from the worker; or (2) no
1016                          * more messages can be read from the worker without blocking.
1017                          */
1018                         while (pcxt->worker[i].error_mqh != NULL)
1019                         {
1020                                 shm_mq_result res;
1021                                 Size            nbytes;
1022                                 void       *data;
1023
1024                                 res = shm_mq_receive(pcxt->worker[i].error_mqh, &nbytes,
1025                                                                          &data, true);
1026                                 if (res == SHM_MQ_WOULD_BLOCK)
1027                                         break;
1028                                 else if (res == SHM_MQ_SUCCESS)
1029                                 {
1030                                         StringInfoData msg;
1031
1032                                         initStringInfo(&msg);
1033                                         appendBinaryStringInfo(&msg, data, nbytes);
1034                                         HandleParallelMessage(pcxt, i, &msg);
1035                                         pfree(msg.data);
1036                                 }
1037                                 else
1038                                         ereport(ERROR,
1039                                                         (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1040                                                          errmsg("lost connection to parallel worker")));
1041                         }
1042                 }
1043         }
1044
1045         MemoryContextSwitchTo(oldcontext);
1046
1047         /* Might as well clear the context on our way out */
1048         MemoryContextReset(hpm_context);
1049
1050         RESUME_INTERRUPTS();
1051 }
1052
1053 /*
1054  * Handle a single protocol message received from a single parallel worker.
1055  */
1056 static void
1057 HandleParallelMessage(ParallelContext *pcxt, int i, StringInfo msg)
1058 {
1059         char            msgtype;
1060
1061         if (pcxt->known_attached_workers != NULL &&
1062                 !pcxt->known_attached_workers[i])
1063         {
1064                 pcxt->known_attached_workers[i] = true;
1065                 pcxt->nknown_attached_workers++;
1066         }
1067
1068         msgtype = pq_getmsgbyte(msg);
1069
1070         switch (msgtype)
1071         {
1072                 case 'K':                               /* BackendKeyData */
1073                         {
1074                                 int32           pid = pq_getmsgint(msg, 4);
1075
1076                                 (void) pq_getmsgint(msg, 4);    /* discard cancel key */
1077                                 (void) pq_getmsgend(msg);
1078                                 pcxt->worker[i].pid = pid;
1079                                 break;
1080                         }
1081
1082                 case 'E':                               /* ErrorResponse */
1083                 case 'N':                               /* NoticeResponse */
1084                         {
1085                                 ErrorData       edata;
1086                                 ErrorContextCallback *save_error_context_stack;
1087
1088                                 /* Parse ErrorResponse or NoticeResponse. */
1089                                 pq_parse_errornotice(msg, &edata);
1090
1091                                 /* Death of a worker isn't enough justification for suicide. */
1092                                 edata.elevel = Min(edata.elevel, ERROR);
1093
1094                                 /*
1095                                  * If desired, add a context line to show that this is a
1096                                  * message propagated from a parallel worker.  Otherwise, it
1097                                  * can sometimes be confusing to understand what actually
1098                                  * happened.  (We don't do this in FORCE_PARALLEL_REGRESS mode
1099                                  * because it causes test-result instability depending on
1100                                  * whether a parallel worker is actually used or not.)
1101                                  */
1102                                 if (force_parallel_mode != FORCE_PARALLEL_REGRESS)
1103                                 {
1104                                         if (edata.context)
1105                                                 edata.context = psprintf("%s\n%s", edata.context,
1106                                                                                                  _("parallel worker"));
1107                                         else
1108                                                 edata.context = pstrdup(_("parallel worker"));
1109                                 }
1110
1111                                 /*
1112                                  * Context beyond that should use the error context callbacks
1113                                  * that were in effect when the ParallelContext was created,
1114                                  * not the current ones.
1115                                  */
1116                                 save_error_context_stack = error_context_stack;
1117                                 error_context_stack = pcxt->error_context_stack;
1118
1119                                 /* Rethrow error or print notice. */
1120                                 ThrowErrorData(&edata);
1121
1122                                 /* Not an error, so restore previous context stack. */
1123                                 error_context_stack = save_error_context_stack;
1124
1125                                 break;
1126                         }
1127
1128                 case 'A':                               /* NotifyResponse */
1129                         {
1130                                 /* Propagate NotifyResponse. */
1131                                 int32           pid;
1132                                 const char *channel;
1133                                 const char *payload;
1134
1135                                 pid = pq_getmsgint(msg, 4);
1136                                 channel = pq_getmsgrawstring(msg);
1137                                 payload = pq_getmsgrawstring(msg);
1138                                 pq_endmessage(msg);
1139
1140                                 NotifyMyFrontEnd(channel, payload, pid);
1141
1142                                 break;
1143                         }
1144
1145                 case 'X':                               /* Terminate, indicating clean exit */
1146                         {
1147                                 shm_mq_detach(pcxt->worker[i].error_mqh);
1148                                 pcxt->worker[i].error_mqh = NULL;
1149                                 break;
1150                         }
1151
1152                 default:
1153                         {
1154                                 elog(ERROR, "unrecognized message type received from parallel worker: %c (message length %d bytes)",
1155                                          msgtype, msg->len);
1156                         }
1157         }
1158 }
1159
1160 /*
1161  * End-of-subtransaction cleanup for parallel contexts.
1162  *
1163  * Currently, it's forbidden to enter or leave a subtransaction while
1164  * parallel mode is in effect, so we could just blow away everything.  But
1165  * we may want to relax that restriction in the future, so this code
1166  * contemplates that there may be multiple subtransaction IDs in pcxt_list.
1167  */
1168 void
1169 AtEOSubXact_Parallel(bool isCommit, SubTransactionId mySubId)
1170 {
1171         while (!dlist_is_empty(&pcxt_list))
1172         {
1173                 ParallelContext *pcxt;
1174
1175                 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
1176                 if (pcxt->subid != mySubId)
1177                         break;
1178                 if (isCommit)
1179                         elog(WARNING, "leaked parallel context");
1180                 DestroyParallelContext(pcxt);
1181         }
1182 }
1183
1184 /*
1185  * End-of-transaction cleanup for parallel contexts.
1186  */
1187 void
1188 AtEOXact_Parallel(bool isCommit)
1189 {
1190         while (!dlist_is_empty(&pcxt_list))
1191         {
1192                 ParallelContext *pcxt;
1193
1194                 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
1195                 if (isCommit)
1196                         elog(WARNING, "leaked parallel context");
1197                 DestroyParallelContext(pcxt);
1198         }
1199 }
1200
1201 /*
1202  * Main entrypoint for parallel workers.
1203  */
1204 void
1205 ParallelWorkerMain(Datum main_arg)
1206 {
1207         dsm_segment *seg;
1208         shm_toc    *toc;
1209         FixedParallelState *fps;
1210         char       *error_queue_space;
1211         shm_mq     *mq;
1212         shm_mq_handle *mqh;
1213         char       *libraryspace;
1214         char       *entrypointstate;
1215         char       *library_name;
1216         char       *function_name;
1217         parallel_worker_main_type entrypt;
1218         char       *gucspace;
1219         char       *combocidspace;
1220         char       *tsnapspace;
1221         char       *asnapspace;
1222         char       *tstatespace;
1223         char       *reindexspace;
1224         char       *relmapperspace;
1225         StringInfoData msgbuf;
1226         char       *session_dsm_handle_space;
1227
1228         /* Set flag to indicate that we're initializing a parallel worker. */
1229         InitializingParallelWorker = true;
1230
1231         /* Establish signal handlers. */
1232         pqsignal(SIGTERM, die);
1233         BackgroundWorkerUnblockSignals();
1234
1235         /* Determine and set our parallel worker number. */
1236         Assert(ParallelWorkerNumber == -1);
1237         memcpy(&ParallelWorkerNumber, MyBgworkerEntry->bgw_extra, sizeof(int));
1238
1239         /* Set up a memory context to work in, just for cleanliness. */
1240         CurrentMemoryContext = AllocSetContextCreate(TopMemoryContext,
1241                                                                                                  "Parallel worker",
1242                                                                                                  ALLOCSET_DEFAULT_SIZES);
1243
1244         /*
1245          * Attach to the dynamic shared memory segment for the parallel query, and
1246          * find its table of contents.
1247          *
1248          * Note: at this point, we have not created any ResourceOwner in this
1249          * process.  This will result in our DSM mapping surviving until process
1250          * exit, which is fine.  If there were a ResourceOwner, it would acquire
1251          * ownership of the mapping, but we have no need for that.
1252          */
1253         seg = dsm_attach(DatumGetUInt32(main_arg));
1254         if (seg == NULL)
1255                 ereport(ERROR,
1256                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1257                                  errmsg("could not map dynamic shared memory segment")));
1258         toc = shm_toc_attach(PARALLEL_MAGIC, dsm_segment_address(seg));
1259         if (toc == NULL)
1260                 ereport(ERROR,
1261                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1262                                  errmsg("invalid magic number in dynamic shared memory segment")));
1263
1264         /* Look up fixed parallel state. */
1265         fps = shm_toc_lookup(toc, PARALLEL_KEY_FIXED, false);
1266         MyFixedParallelState = fps;
1267
1268         /* Arrange to signal the leader if we exit. */
1269         ParallelMasterPid = fps->parallel_master_pid;
1270         ParallelMasterBackendId = fps->parallel_master_backend_id;
1271         on_shmem_exit(ParallelWorkerShutdown, (Datum) 0);
1272
1273         /*
1274          * Now we can find and attach to the error queue provided for us.  That's
1275          * good, because until we do that, any errors that happen here will not be
1276          * reported back to the process that requested that this worker be
1277          * launched.
1278          */
1279         error_queue_space = shm_toc_lookup(toc, PARALLEL_KEY_ERROR_QUEUE, false);
1280         mq = (shm_mq *) (error_queue_space +
1281                                          ParallelWorkerNumber * PARALLEL_ERROR_QUEUE_SIZE);
1282         shm_mq_set_sender(mq, MyProc);
1283         mqh = shm_mq_attach(mq, seg, NULL);
1284         pq_redirect_to_shm_mq(seg, mqh);
1285         pq_set_parallel_master(fps->parallel_master_pid,
1286                                                    fps->parallel_master_backend_id);
1287
1288         /*
1289          * Send a BackendKeyData message to the process that initiated parallelism
1290          * so that it has access to our PID before it receives any other messages
1291          * from us.  Our cancel key is sent, too, since that's the way the
1292          * protocol message is defined, but it won't actually be used for anything
1293          * in this case.
1294          */
1295         pq_beginmessage(&msgbuf, 'K');
1296         pq_sendint32(&msgbuf, (int32) MyProcPid);
1297         pq_sendint32(&msgbuf, (int32) MyCancelKey);
1298         pq_endmessage(&msgbuf);
1299
1300         /*
1301          * Hooray! Primary initialization is complete.  Now, we need to set up our
1302          * backend-local state to match the original backend.
1303          */
1304
1305         /*
1306          * Join locking group.  We must do this before anything that could try to
1307          * acquire a heavyweight lock, because any heavyweight locks acquired to
1308          * this point could block either directly against the parallel group
1309          * leader or against some process which in turn waits for a lock that
1310          * conflicts with the parallel group leader, causing an undetected
1311          * deadlock.  (If we can't join the lock group, the leader has gone away,
1312          * so just exit quietly.)
1313          */
1314         if (!BecomeLockGroupMember(fps->parallel_master_pgproc,
1315                                                            fps->parallel_master_pid))
1316                 return;
1317
1318         /*
1319          * Restore transaction and statement start-time timestamps.  This must
1320          * happen before anything that would start a transaction, else asserts in
1321          * xact.c will fire.
1322          */
1323         SetParallelStartTimestamps(fps->xact_ts, fps->stmt_ts);
1324
1325         /*
1326          * Identify the entry point to be called.  In theory this could result in
1327          * loading an additional library, though most likely the entry point is in
1328          * the core backend or in a library we just loaded.
1329          */
1330         entrypointstate = shm_toc_lookup(toc, PARALLEL_KEY_ENTRYPOINT, false);
1331         library_name = entrypointstate;
1332         function_name = entrypointstate + strlen(library_name) + 1;
1333
1334         entrypt = LookupParallelWorkerFunction(library_name, function_name);
1335
1336         /* Restore database connection. */
1337         BackgroundWorkerInitializeConnectionByOid(fps->database_id,
1338                                                                                           fps->authenticated_user_id,
1339                                                                                           0);
1340
1341         /*
1342          * Set the client encoding to the database encoding, since that is what
1343          * the leader will expect.
1344          */
1345         SetClientEncoding(GetDatabaseEncoding());
1346
1347         /*
1348          * Load libraries that were loaded by original backend.  We want to do
1349          * this before restoring GUCs, because the libraries might define custom
1350          * variables.
1351          */
1352         libraryspace = shm_toc_lookup(toc, PARALLEL_KEY_LIBRARY, false);
1353         StartTransactionCommand();
1354         RestoreLibraryState(libraryspace);
1355
1356         /* Restore GUC values from launching backend. */
1357         gucspace = shm_toc_lookup(toc, PARALLEL_KEY_GUC, false);
1358         RestoreGUCState(gucspace);
1359         CommitTransactionCommand();
1360
1361         /* Crank up a transaction state appropriate to a parallel worker. */
1362         tstatespace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_STATE, false);
1363         StartParallelWorkerTransaction(tstatespace);
1364
1365         /* Restore combo CID state. */
1366         combocidspace = shm_toc_lookup(toc, PARALLEL_KEY_COMBO_CID, false);
1367         RestoreComboCIDState(combocidspace);
1368
1369         /* Attach to the per-session DSM segment and contained objects. */
1370         session_dsm_handle_space =
1371                 shm_toc_lookup(toc, PARALLEL_KEY_SESSION_DSM, false);
1372         AttachSession(*(dsm_handle *) session_dsm_handle_space);
1373
1374         /* Restore transaction snapshot. */
1375         tsnapspace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT, false);
1376         RestoreTransactionSnapshot(RestoreSnapshot(tsnapspace),
1377                                                            fps->parallel_master_pgproc);
1378
1379         /* Restore active snapshot. */
1380         asnapspace = shm_toc_lookup(toc, PARALLEL_KEY_ACTIVE_SNAPSHOT, false);
1381         PushActiveSnapshot(RestoreSnapshot(asnapspace));
1382
1383         /*
1384          * We've changed which tuples we can see, and must therefore invalidate
1385          * system caches.
1386          */
1387         InvalidateSystemCaches();
1388
1389         /*
1390          * Restore current role id.  Skip verifying whether session user is
1391          * allowed to become this role and blindly restore the leader's state for
1392          * current role.
1393          */
1394         SetCurrentRoleId(fps->outer_user_id, fps->is_superuser);
1395
1396         /* Restore user ID and security context. */
1397         SetUserIdAndSecContext(fps->current_user_id, fps->sec_context);
1398
1399         /* Restore temp-namespace state to ensure search path matches leader's. */
1400         SetTempNamespaceState(fps->temp_namespace_id,
1401                                                   fps->temp_toast_namespace_id);
1402
1403         /* Restore reindex state. */
1404         reindexspace = shm_toc_lookup(toc, PARALLEL_KEY_REINDEX_STATE, false);
1405         RestoreReindexState(reindexspace);
1406
1407         /* Restore relmapper state. */
1408         relmapperspace = shm_toc_lookup(toc, PARALLEL_KEY_RELMAPPER_STATE, false);
1409         RestoreRelationMap(relmapperspace);
1410
1411         /*
1412          * We've initialized all of our state now; nothing should change
1413          * hereafter.
1414          */
1415         InitializingParallelWorker = false;
1416         EnterParallelMode();
1417
1418         /*
1419          * Time to do the real work: invoke the caller-supplied code.
1420          */
1421         entrypt(seg, toc);
1422
1423         /* Must exit parallel mode to pop active snapshot. */
1424         ExitParallelMode();
1425
1426         /* Must pop active snapshot so snapmgr.c doesn't complain. */
1427         PopActiveSnapshot();
1428
1429         /* Shut down the parallel-worker transaction. */
1430         EndParallelWorkerTransaction();
1431
1432         /* Detach from the per-session DSM segment. */
1433         DetachSession();
1434
1435         /* Report success. */
1436         pq_putmessage('X', NULL, 0);
1437 }
1438
1439 /*
1440  * Update shared memory with the ending location of the last WAL record we
1441  * wrote, if it's greater than the value already stored there.
1442  */
1443 void
1444 ParallelWorkerReportLastRecEnd(XLogRecPtr last_xlog_end)
1445 {
1446         FixedParallelState *fps = MyFixedParallelState;
1447
1448         Assert(fps != NULL);
1449         SpinLockAcquire(&fps->mutex);
1450         if (fps->last_xlog_end < last_xlog_end)
1451                 fps->last_xlog_end = last_xlog_end;
1452         SpinLockRelease(&fps->mutex);
1453 }
1454
1455 /*
1456  * Make sure the leader tries to read from our error queue one more time.
1457  * This guards against the case where we exit uncleanly without sending an
1458  * ErrorResponse to the leader, for example because some code calls proc_exit
1459  * directly.
1460  */
1461 static void
1462 ParallelWorkerShutdown(int code, Datum arg)
1463 {
1464         SendProcSignal(ParallelMasterPid,
1465                                    PROCSIG_PARALLEL_MESSAGE,
1466                                    ParallelMasterBackendId);
1467 }
1468
1469 /*
1470  * Look up (and possibly load) a parallel worker entry point function.
1471  *
1472  * For functions contained in the core code, we use library name "postgres"
1473  * and consult the InternalParallelWorkers array.  External functions are
1474  * looked up, and loaded if necessary, using load_external_function().
1475  *
1476  * The point of this is to pass function names as strings across process
1477  * boundaries.  We can't pass actual function addresses because of the
1478  * possibility that the function has been loaded at a different address
1479  * in a different process.  This is obviously a hazard for functions in
1480  * loadable libraries, but it can happen even for functions in the core code
1481  * on platforms using EXEC_BACKEND (e.g., Windows).
1482  *
1483  * At some point it might be worthwhile to get rid of InternalParallelWorkers[]
1484  * in favor of applying load_external_function() for core functions too;
1485  * but that raises portability issues that are not worth addressing now.
1486  */
1487 static parallel_worker_main_type
1488 LookupParallelWorkerFunction(const char *libraryname, const char *funcname)
1489 {
1490         /*
1491          * If the function is to be loaded from postgres itself, search the
1492          * InternalParallelWorkers array.
1493          */
1494         if (strcmp(libraryname, "postgres") == 0)
1495         {
1496                 int                     i;
1497
1498                 for (i = 0; i < lengthof(InternalParallelWorkers); i++)
1499                 {
1500                         if (strcmp(InternalParallelWorkers[i].fn_name, funcname) == 0)
1501                                 return InternalParallelWorkers[i].fn_addr;
1502                 }
1503
1504                 /* We can only reach this by programming error. */
1505                 elog(ERROR, "internal function \"%s\" not found", funcname);
1506         }
1507
1508         /* Otherwise load from external library. */
1509         return (parallel_worker_main_type)
1510                 load_external_function(libraryname, funcname, true, NULL);
1511 }