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) 1996-2019, 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/pg_enum.h"
  23 #include "catalog/index.h"
  24 #include "catalog/namespace.h"
  25 #include "commands/async.h"
  26 #include "executor/execParallel.h"
  27 #include "libpq/libpq.h"
  28 #include "libpq/pqformat.h"
  29 #include "libpq/pqmq.h"
  30 #include "miscadmin.h"
  31 #include "optimizer/optimizer.h"
  32 #include "pgstat.h"
  33 #include "storage/ipc.h"
  34 #include "storage/predicate.h"
  35 #include "storage/sinval.h"
  36 #include "storage/spin.h"
  37 #include "tcop/tcopprot.h"
  38 #include "utils/combocid.h"
  39 #include "utils/guc.h"
  40 #include "utils/inval.h"
  41 #include "utils/memutils.h"
  42 #include "utils/relmapper.h"
  43 #include "utils/snapmgr.h"
  44 #include "utils/typcache.h"
  45
  46
  47 /*
  48  * We don't want to waste a lot of memory on an error queue which, most of
  49  * the time, will process only a handful of small messages.  However, it is
  50  * desirable to make it large enough that a typical ErrorResponse can be sent
  51  * without blocking.  That way, a worker that errors out can write the whole
  52  * message into the queue and terminate without waiting for the user backend.
  53  */
  54 #define PARALLEL_ERROR_QUEUE_SIZE                       16384
  55
  56 /* Magic number for parallel context TOC. */
  57 #define PARALLEL_MAGIC                                          0x50477c7c
  58
  59 /*
  60  * Magic numbers for per-context parallel state sharing.  Higher-level code
  61  * should use smaller values, leaving these very large ones for use by this
  62  * module.
  63  */
  64 #define PARALLEL_KEY_FIXED                                      UINT64CONST(0xFFFFFFFFFFFF0001)
  65 #define PARALLEL_KEY_ERROR_QUEUE                        UINT64CONST(0xFFFFFFFFFFFF0002)
  66 #define PARALLEL_KEY_LIBRARY                            UINT64CONST(0xFFFFFFFFFFFF0003)
  67 #define PARALLEL_KEY_GUC                                        UINT64CONST(0xFFFFFFFFFFFF0004)
  68 #define PARALLEL_KEY_COMBO_CID                          UINT64CONST(0xFFFFFFFFFFFF0005)
  69 #define PARALLEL_KEY_TRANSACTION_SNAPSHOT       UINT64CONST(0xFFFFFFFFFFFF0006)
  70 #define PARALLEL_KEY_ACTIVE_SNAPSHOT            UINT64CONST(0xFFFFFFFFFFFF0007)
  71 #define PARALLEL_KEY_TRANSACTION_STATE          UINT64CONST(0xFFFFFFFFFFFF0008)
  72 #define PARALLEL_KEY_ENTRYPOINT                         UINT64CONST(0xFFFFFFFFFFFF0009)
  73 #define PARALLEL_KEY_SESSION_DSM                        UINT64CONST(0xFFFFFFFFFFFF000A)
  74 #define PARALLEL_KEY_REINDEX_STATE                      UINT64CONST(0xFFFFFFFFFFFF000B)
  75 #define PARALLEL_KEY_RELMAPPER_STATE            UINT64CONST(0xFFFFFFFFFFFF000C)
  76 #define PARALLEL_KEY_ENUMBLACKLIST                      UINT64CONST(0xFFFFFFFFFFFF000D)
  77
  78 /* Fixed-size parallel state. */
  79 typedef struct FixedParallelState
  80 {
  81         /* Fixed-size state that workers must restore. */
  82         Oid                     database_id;
  83         Oid                     authenticated_user_id;
  84         Oid                     current_user_id;
  85         Oid                     outer_user_id;
  86         Oid                     temp_namespace_id;
  87         Oid                     temp_toast_namespace_id;
  88         int                     sec_context;
  89         bool            is_superuser;
  90         PGPROC     *parallel_master_pgproc;
  91         pid_t           parallel_master_pid;
  92         BackendId       parallel_master_backend_id;
  93         TimestampTz xact_ts;
  94         TimestampTz stmt_ts;
  95         SerializableXactHandle serializable_xact_handle;
  96
  97         /* Mutex protects remaining fields. */
  98         slock_t         mutex;
  99
 100         /* Maximum XactLastRecEnd of any worker. */
 101         XLogRecPtr      last_xlog_end;
 102 } FixedParallelState;
 103
 104 /*
 105  * Our parallel worker number.  We initialize this to -1, meaning that we are
 106  * not a parallel worker.  In parallel workers, it will be set to a value >= 0
 107  * and < the number of workers before any user code is invoked; each parallel
 108  * worker will get a different parallel worker number.
 109  */
 110 int                     ParallelWorkerNumber = -1;
 111
 112 /* Is there a parallel message pending which we need to receive? */
 113 volatile bool ParallelMessagePending = false;
 114
 115 /* Are we initializing a parallel worker? */
 116 bool            InitializingParallelWorker = false;
 117
 118 /* Pointer to our fixed parallel state. */
 119 static FixedParallelState *MyFixedParallelState;
 120
 121 /* List of active parallel contexts. */
 122 static dlist_head pcxt_list = DLIST_STATIC_INIT(pcxt_list);
 123
 124 /* Backend-local copy of data from FixedParallelState. */
 125 static pid_t ParallelMasterPid;
 126
 127 /*
 128  * List of internal parallel worker entry points.  We need this for
 129  * reasons explained in LookupParallelWorkerFunction(), below.
 130  */
 131 static const struct
 132 {
 133         const char *fn_name;
 134         parallel_worker_main_type fn_addr;
 135 }                       InternalParallelWorkers[] =
 136
 137 {
 138         {
 139                 "ParallelQueryMain", ParallelQueryMain
 140         },
 141         {
 142                 "_bt_parallel_build_main", _bt_parallel_build_main
 143         }
 144 };
 145
 146 /* Private functions. */
 147 static void HandleParallelMessage(ParallelContext *pcxt, int i, StringInfo msg);
 148 static void WaitForParallelWorkersToExit(ParallelContext *pcxt);
 149 static parallel_worker_main_type LookupParallelWorkerFunction(const char *libraryname, const char *funcname);
 150 static void ParallelWorkerShutdown(int code, Datum arg);
 151
 152
 153 /*
 154  * Establish a new parallel context.  This should be done after entering
 155  * parallel mode, and (unless there is an error) the context should be
 156  * destroyed before exiting the current subtransaction.
 157  */
 158 ParallelContext *
 159 CreateParallelContext(const char *library_name, const char *function_name,
 160                                           int nworkers)
 161 {
 162         MemoryContext oldcontext;
 163         ParallelContext *pcxt;
 164
 165         /* It is unsafe to create a parallel context if not in parallel mode. */
 166         Assert(IsInParallelMode());
 167
 168         /* Number of workers should be non-negative. */
 169         Assert(nworkers >= 0);
 170
 171         /* We might be running in a short-lived memory context. */
 172         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 173
 174         /* Initialize a new ParallelContext. */
 175         pcxt = palloc0(sizeof(ParallelContext));
 176         pcxt->subid = GetCurrentSubTransactionId();
 177         pcxt->nworkers = nworkers;
 178         pcxt->library_name = pstrdup(library_name);
 179         pcxt->function_name = pstrdup(function_name);
 180         pcxt->error_context_stack = error_context_stack;
 181         shm_toc_initialize_estimator(&pcxt->estimator);
 182         dlist_push_head(&pcxt_list, &pcxt->node);
 183
 184         /* Restore previous memory context. */
 185         MemoryContextSwitchTo(oldcontext);
 186
 187         return pcxt;
 188 }
 189
 190 /*
 191  * Establish the dynamic shared memory segment for a parallel context and
 192  * copy state and other bookkeeping information that will be needed by
 193  * parallel workers into it.
 194  */
 195 void
 196 InitializeParallelDSM(ParallelContext *pcxt)
 197 {
 198         MemoryContext oldcontext;
 199         Size            library_len = 0;
 200         Size            guc_len = 0;
 201         Size            combocidlen = 0;
 202         Size            tsnaplen = 0;
 203         Size            asnaplen = 0;
 204         Size            tstatelen = 0;
 205         Size            reindexlen = 0;
 206         Size            relmapperlen = 0;
 207         Size            enumblacklistlen = 0;
 208         Size            segsize = 0;
 209         int                     i;
 210         FixedParallelState *fps;
 211         dsm_handle      session_dsm_handle = DSM_HANDLE_INVALID;
 212         Snapshot        transaction_snapshot = GetTransactionSnapshot();
 213         Snapshot        active_snapshot = GetActiveSnapshot();
 214
 215         /* We might be running in a very short-lived memory context. */
 216         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 217
 218         /* Allow space to store the fixed-size parallel state. */
 219         shm_toc_estimate_chunk(&pcxt->estimator, sizeof(FixedParallelState));
 220         shm_toc_estimate_keys(&pcxt->estimator, 1);
 221
 222         /*
 223          * Normally, the user will have requested at least one worker process, but
 224          * if by chance they have not, we can skip a bunch of things here.
 225          */
 226         if (pcxt->nworkers > 0)
 227         {
 228                 /* Get (or create) the per-session DSM segment's handle. */
 229                 session_dsm_handle = GetSessionDsmHandle();
 230
 231                 /*
 232                  * If we weren't able to create a per-session DSM segment, then we can
 233                  * continue but we can't safely launch any workers because their
 234                  * record typmods would be incompatible so they couldn't exchange
 235                  * tuples.
 236                  */
 237                 if (session_dsm_handle == DSM_HANDLE_INVALID)
 238                         pcxt->nworkers = 0;
 239         }
 240
 241         if (pcxt->nworkers > 0)
 242         {
 243                 /* Estimate space for various kinds of state sharing. */
 244                 library_len = EstimateLibraryStateSpace();
 245                 shm_toc_estimate_chunk(&pcxt->estimator, library_len);
 246                 guc_len = EstimateGUCStateSpace();
 247                 shm_toc_estimate_chunk(&pcxt->estimator, guc_len);
 248                 combocidlen = EstimateComboCIDStateSpace();
 249                 shm_toc_estimate_chunk(&pcxt->estimator, combocidlen);
 250                 tsnaplen = EstimateSnapshotSpace(transaction_snapshot);
 251                 shm_toc_estimate_chunk(&pcxt->estimator, tsnaplen);
 252                 asnaplen = EstimateSnapshotSpace(active_snapshot);
 253                 shm_toc_estimate_chunk(&pcxt->estimator, asnaplen);
 254                 tstatelen = EstimateTransactionStateSpace();
 255                 shm_toc_estimate_chunk(&pcxt->estimator, tstatelen);
 256                 shm_toc_estimate_chunk(&pcxt->estimator, sizeof(dsm_handle));
 257                 reindexlen = EstimateReindexStateSpace();
 258                 shm_toc_estimate_chunk(&pcxt->estimator, reindexlen);
 259                 relmapperlen = EstimateRelationMapSpace();
 260                 shm_toc_estimate_chunk(&pcxt->estimator, relmapperlen);
 261                 enumblacklistlen = EstimateEnumBlacklistSpace();
 262                 shm_toc_estimate_chunk(&pcxt->estimator, enumblacklistlen);
 263                 /* If you add more chunks here, you probably need to add keys. */
 264                 shm_toc_estimate_keys(&pcxt->estimator, 10);
 265
 266                 /* Estimate space need for error queues. */
 267                 StaticAssertStmt(BUFFERALIGN(PARALLEL_ERROR_QUEUE_SIZE) ==
 268                                                  PARALLEL_ERROR_QUEUE_SIZE,
 269                                                  "parallel error queue size not buffer-aligned");
 270                 shm_toc_estimate_chunk(&pcxt->estimator,
 271                                                            mul_size(PARALLEL_ERROR_QUEUE_SIZE,
 272                                                                                 pcxt->nworkers));
 273                 shm_toc_estimate_keys(&pcxt->estimator, 1);
 274
 275                 /* Estimate how much we'll need for the entrypoint info. */
 276                 shm_toc_estimate_chunk(&pcxt->estimator, strlen(pcxt->library_name) +
 277                                                            strlen(pcxt->function_name) + 2);
 278                 shm_toc_estimate_keys(&pcxt->estimator, 1);
 279         }
 280
 281         /*
 282          * Create DSM and initialize with new table of contents.  But if the user
 283          * didn't request any workers, then don't bother creating a dynamic shared
 284          * memory segment; instead, just use backend-private memory.
 285          *
 286          * Also, if we can't create a dynamic shared memory segment because the
 287          * maximum number of segments have already been created, then fall back to
 288          * backend-private memory, and plan not to use any workers.  We hope this
 289          * won't happen very often, but it's better to abandon the use of
 290          * parallelism than to fail outright.
 291          */
 292         segsize = shm_toc_estimate(&pcxt->estimator);
 293         if (pcxt->nworkers > 0)
 294                 pcxt->seg = dsm_create(segsize, DSM_CREATE_NULL_IF_MAXSEGMENTS);
 295         if (pcxt->seg != NULL)
 296                 pcxt->toc = shm_toc_create(PARALLEL_MAGIC,
 297                                                                    dsm_segment_address(pcxt->seg),
 298                                                                    segsize);
 299         else
 300         {
 301                 pcxt->nworkers = 0;
 302                 pcxt->private_memory = MemoryContextAlloc(TopMemoryContext, segsize);
 303                 pcxt->toc = shm_toc_create(PARALLEL_MAGIC, pcxt->private_memory,
 304                                                                    segsize);
 305         }
 306
 307         /* Initialize fixed-size state in shared memory. */
 308         fps = (FixedParallelState *)
 309                 shm_toc_allocate(pcxt->toc, sizeof(FixedParallelState));
 310         fps->database_id = MyDatabaseId;
 311         fps->authenticated_user_id = GetAuthenticatedUserId();
 312         fps->outer_user_id = GetCurrentRoleId();
 313         fps->is_superuser = session_auth_is_superuser;
 314         GetUserIdAndSecContext(&fps->current_user_id, &fps->sec_context);
 315         GetTempNamespaceState(&fps->temp_namespace_id,
 316                                                   &fps->temp_toast_namespace_id);
 317         fps->parallel_master_pgproc = MyProc;
 318         fps->parallel_master_pid = MyProcPid;
 319         fps->parallel_master_backend_id = MyBackendId;
 320         fps->xact_ts = GetCurrentTransactionStartTimestamp();
 321         fps->stmt_ts = GetCurrentStatementStartTimestamp();
 322         fps->serializable_xact_handle = ShareSerializableXact();
 323         SpinLockInit(&fps->mutex);
 324         fps->last_xlog_end = 0;
 325         shm_toc_insert(pcxt->toc, PARALLEL_KEY_FIXED, fps);
 326
 327         /* We can skip the rest of this if we're not budgeting for any workers. */
 328         if (pcxt->nworkers > 0)
 329         {
 330                 char       *libraryspace;
 331                 char       *gucspace;
 332                 char       *combocidspace;
 333                 char       *tsnapspace;
 334                 char       *asnapspace;
 335                 char       *tstatespace;
 336                 char       *reindexspace;
 337                 char       *relmapperspace;
 338                 char       *error_queue_space;
 339                 char       *session_dsm_handle_space;
 340                 char       *entrypointstate;
 341                 char       *enumblacklistspace;
 342                 Size            lnamelen;
 343
 344                 /* Serialize shared libraries we have loaded. */
 345                 libraryspace = shm_toc_allocate(pcxt->toc, library_len);
 346                 SerializeLibraryState(library_len, libraryspace);
 347                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_LIBRARY, libraryspace);
 348
 349                 /* Serialize GUC settings. */
 350                 gucspace = shm_toc_allocate(pcxt->toc, guc_len);
 351                 SerializeGUCState(guc_len, gucspace);
 352                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_GUC, gucspace);
 353
 354                 /* Serialize combo CID state. */
 355                 combocidspace = shm_toc_allocate(pcxt->toc, combocidlen);
 356                 SerializeComboCIDState(combocidlen, combocidspace);
 357                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_COMBO_CID, combocidspace);
 358
 359                 /* Serialize transaction snapshot and active snapshot. */
 360                 tsnapspace = shm_toc_allocate(pcxt->toc, tsnaplen);
 361                 SerializeSnapshot(transaction_snapshot, tsnapspace);
 362                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT,
 363                                            tsnapspace);
 364                 asnapspace = shm_toc_allocate(pcxt->toc, asnaplen);
 365                 SerializeSnapshot(active_snapshot, asnapspace);
 366                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ACTIVE_SNAPSHOT, asnapspace);
 367
 368                 /* Provide the handle for per-session segment. */
 369                 session_dsm_handle_space = shm_toc_allocate(pcxt->toc,
 370                                                                                                         sizeof(dsm_handle));
 371                 *(dsm_handle *) session_dsm_handle_space = session_dsm_handle;
 372                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_SESSION_DSM,
 373                                            session_dsm_handle_space);
 374
 375                 /* Serialize transaction state. */
 376                 tstatespace = shm_toc_allocate(pcxt->toc, tstatelen);
 377                 SerializeTransactionState(tstatelen, tstatespace);
 378                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_STATE, tstatespace);
 379
 380                 /* Serialize reindex state. */
 381                 reindexspace = shm_toc_allocate(pcxt->toc, reindexlen);
 382                 SerializeReindexState(reindexlen, reindexspace);
 383                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_REINDEX_STATE, reindexspace);
 384
 385                 /* Serialize relmapper state. */
 386                 relmapperspace = shm_toc_allocate(pcxt->toc, relmapperlen);
 387                 SerializeRelationMap(relmapperlen, relmapperspace);
 388                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_RELMAPPER_STATE,
 389                                            relmapperspace);
 390
 391                 /* Serialize enum blacklist state. */
 392                 enumblacklistspace = shm_toc_allocate(pcxt->toc, enumblacklistlen);
 393                 SerializeEnumBlacklist(enumblacklistspace, enumblacklistlen);
 394                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ENUMBLACKLIST,
 395                                            enumblacklistspace);
 396
 397                 /* Allocate space for worker information. */
 398                 pcxt->worker = palloc0(sizeof(ParallelWorkerInfo) * pcxt->nworkers);
 399
 400                 /*
 401                  * Establish error queues in dynamic shared memory.
 402                  *
 403                  * These queues should be used only for transmitting ErrorResponse,
 404                  * NoticeResponse, and NotifyResponse protocol messages.  Tuple data
 405                  * should be transmitted via separate (possibly larger?) queues.
 406                  */
 407                 error_queue_space =
 408                         shm_toc_allocate(pcxt->toc,
 409                                                          mul_size(PARALLEL_ERROR_QUEUE_SIZE,
 410                                                                           pcxt->nworkers));
 411                 for (i = 0; i < pcxt->nworkers; ++i)
 412                 {
 413                         char       *start;
 414                         shm_mq     *mq;
 415
 416                         start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
 417                         mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
 418                         shm_mq_set_receiver(mq, MyProc);
 419                         pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
 420                 }
 421                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE, error_queue_space);
 422
 423                 /*
 424                  * Serialize entrypoint information.  It's unsafe to pass function
 425                  * pointers across processes, as the function pointer may be different
 426                  * in each process in EXEC_BACKEND builds, so we always pass library
 427                  * and function name.  (We use library name "postgres" for functions
 428                  * in the core backend.)
 429                  */
 430                 lnamelen = strlen(pcxt->library_name);
 431                 entrypointstate = shm_toc_allocate(pcxt->toc, lnamelen +
 432                                                                                    strlen(pcxt->function_name) + 2);
 433                 strcpy(entrypointstate, pcxt->library_name);
 434                 strcpy(entrypointstate + lnamelen + 1, pcxt->function_name);
 435                 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ENTRYPOINT, entrypointstate);
 436         }
 437
 438         /* Restore previous memory context. */
 439         MemoryContextSwitchTo(oldcontext);
 440 }
 441
 442 /*
 443  * Reinitialize the dynamic shared memory segment for a parallel context such
 444  * that we could launch workers for it again.
 445  */
 446 void
 447 ReinitializeParallelDSM(ParallelContext *pcxt)
 448 {
 449         FixedParallelState *fps;
 450
 451         /* Wait for any old workers to exit. */
 452         if (pcxt->nworkers_launched > 0)
 453         {
 454                 WaitForParallelWorkersToFinish(pcxt);
 455                 WaitForParallelWorkersToExit(pcxt);
 456                 pcxt->nworkers_launched = 0;
 457                 if (pcxt->known_attached_workers)
 458                 {
 459                         pfree(pcxt->known_attached_workers);
 460                         pcxt->known_attached_workers = NULL;
 461                         pcxt->nknown_attached_workers = 0;
 462                 }
 463         }
 464
 465         /* Reset a few bits of fixed parallel state to a clean state. */
 466         fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED, false);
 467         fps->last_xlog_end = 0;
 468
 469         /* Recreate error queues (if they exist). */
 470         if (pcxt->nworkers > 0)
 471         {
 472                 char       *error_queue_space;
 473                 int                     i;
 474
 475                 error_queue_space =
 476                         shm_toc_lookup(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE, false);
 477                 for (i = 0; i < pcxt->nworkers; ++i)
 478                 {
 479                         char       *start;
 480                         shm_mq     *mq;
 481
 482                         start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
 483                         mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
 484                         shm_mq_set_receiver(mq, MyProc);
 485                         pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
 486                 }
 487         }
 488 }
 489
 490 /*
 491  * Launch parallel workers.
 492  */
 493 void
 494 LaunchParallelWorkers(ParallelContext *pcxt)
 495 {
 496         MemoryContext oldcontext;
 497         BackgroundWorker worker;
 498         int                     i;
 499         bool            any_registrations_failed = false;
 500
 501         /* Skip this if we have no workers. */
 502         if (pcxt->nworkers == 0)
 503                 return;
 504
 505         /* We need to be a lock group leader. */
 506         BecomeLockGroupLeader();
 507
 508         /* If we do have workers, we'd better have a DSM segment. */
 509         Assert(pcxt->seg != NULL);
 510
 511         /* We might be running in a short-lived memory context. */
 512         oldcontext = MemoryContextSwitchTo(TopTransactionContext);
 513
 514         /* Configure a worker. */
 515         memset(&worker, 0, sizeof(worker));
 516         snprintf(worker.bgw_name, BGW_MAXLEN, "parallel worker for PID %d",
 517                          MyProcPid);
 518         snprintf(worker.bgw_type, BGW_MAXLEN, "parallel worker");
 519         worker.bgw_flags =
 520                 BGWORKER_SHMEM_ACCESS | BGWORKER_BACKEND_DATABASE_CONNECTION
 521                 | BGWORKER_CLASS_PARALLEL;
 522         worker.bgw_start_time = BgWorkerStart_ConsistentState;
 523         worker.bgw_restart_time = BGW_NEVER_RESTART;
 524         sprintf(worker.bgw_library_name, "postgres");
 525         sprintf(worker.bgw_function_name, "ParallelWorkerMain");
 526         worker.bgw_main_arg = UInt32GetDatum(dsm_segment_handle(pcxt->seg));
 527         worker.bgw_notify_pid = MyProcPid;
 528
 529         /*
 530          * Start workers.
 531          *
 532          * The caller must be able to tolerate ending up with fewer workers than
 533          * expected, so there is no need to throw an error here if registration
 534          * fails.  It wouldn't help much anyway, because registering the worker in
 535          * no way guarantees that it will start up and initialize successfully.
 536          */
 537         for (i = 0; i < pcxt->nworkers; ++i)
 538         {
 539                 memcpy(worker.bgw_extra, &i, sizeof(int));
 540                 if (!any_registrations_failed &&
 541                         RegisterDynamicBackgroundWorker(&worker,
 542                                                                                         &pcxt->worker[i].bgwhandle))
 543                 {
 544                         shm_mq_set_handle(pcxt->worker[i].error_mqh,
 545                                                           pcxt->worker[i].bgwhandle);
 546                         pcxt->nworkers_launched++;
 547                 }
 548                 else
 549                 {
 550                         /*
 551                          * If we weren't able to register the worker, then we've bumped up
 552                          * against the max_worker_processes limit, and future
 553                          * registrations will probably fail too, so arrange to skip them.
 554                          * But we still have to execute this code for the remaining slots
 555                          * to make sure that we forget about the error queues we budgeted
 556                          * for those workers.  Otherwise, we'll wait for them to start,
 557                          * but they never will.
 558                          */
 559                         any_registrations_failed = true;
 560                         pcxt->worker[i].bgwhandle = NULL;
 561                         shm_mq_detach(pcxt->worker[i].error_mqh);
 562                         pcxt->worker[i].error_mqh = NULL;
 563                 }
 564         }
 565
 566         /*
 567          * Now that nworkers_launched has taken its final value, we can initialize
 568          * known_attached_workers.
 569          */
 570         if (pcxt->nworkers_launched > 0)
 571         {
 572                 pcxt->known_attached_workers =
 573                         palloc0(sizeof(bool) * pcxt->nworkers_launched);
 574                 pcxt->nknown_attached_workers = 0;
 575         }
 576
 577         /* Restore previous memory context. */
 578         MemoryContextSwitchTo(oldcontext);
 579 }
 580
 581 /*
 582  * Wait for all workers to attach to their error queues, and throw an error if
 583  * any worker fails to do this.
 584  *
 585  * Callers can assume that if this function returns successfully, then the
 586  * number of workers given by pcxt->nworkers_launched have initialized and
 587  * attached to their error queues.  Whether or not these workers are guaranteed
 588  * to still be running depends on what code the caller asked them to run;
 589  * this function does not guarantee that they have not exited.  However, it
 590  * does guarantee that any workers which exited must have done so cleanly and
 591  * after successfully performing the work with which they were tasked.
 592  *
 593  * If this function is not called, then some of the workers that were launched
 594  * may not have been started due to a fork() failure, or may have exited during
 595  * early startup prior to attaching to the error queue, so nworkers_launched
 596  * cannot be viewed as completely reliable.  It will never be less than the
 597  * number of workers which actually started, but it might be more.  Any workers
 598  * that failed to start will still be discovered by
 599  * WaitForParallelWorkersToFinish and an error will be thrown at that time,
 600  * provided that function is eventually reached.
 601  *
 602  * In general, the leader process should do as much work as possible before
 603  * calling this function.  fork() failures and other early-startup failures
 604  * are very uncommon, and having the leader sit idle when it could be doing
 605  * useful work is undesirable.  However, if the leader needs to wait for
 606  * all of its workers or for a specific worker, it may want to call this
 607  * function before doing so.  If not, it must make some other provision for
 608  * the failure-to-start case, lest it wait forever.  On the other hand, a
 609  * leader which never waits for a worker that might not be started yet, or
 610  * at least never does so prior to WaitForParallelWorkersToFinish(), need not
 611  * call this function at all.
 612  */
 613 void
 614 WaitForParallelWorkersToAttach(ParallelContext *pcxt)
 615 {
 616         int                     i;
 617
 618         /* Skip this if we have no launched workers. */
 619         if (pcxt->nworkers_launched == 0)
 620                 return;
 621
 622         for (;;)
 623         {
 624                 /*
 625                  * This will process any parallel messages that are pending and it may
 626                  * also throw an error propagated from a worker.
 627                  */
 628                 CHECK_FOR_INTERRUPTS();
 629
 630                 for (i = 0; i < pcxt->nworkers_launched; ++i)
 631                 {
 632                         BgwHandleStatus status;
 633                         shm_mq     *mq;
 634                         int                     rc;
 635                         pid_t           pid;
 636
 637                         if (pcxt->known_attached_workers[i])
 638                                 continue;
 639
 640                         /*
 641                          * If error_mqh is NULL, then the worker has already exited
 642                          * cleanly.
 643                          */
 644                         if (pcxt->worker[i].error_mqh == NULL)
 645                         {
 646                                 pcxt->known_attached_workers[i] = true;
 647                                 ++pcxt->nknown_attached_workers;
 648                                 continue;
 649                         }
 650
 651                         status = GetBackgroundWorkerPid(pcxt->worker[i].bgwhandle, &pid);
 652                         if (status == BGWH_STARTED)
 653                         {
 654                                 /* Has the worker attached to the error queue? */
 655                                 mq = shm_mq_get_queue(pcxt->worker[i].error_mqh);
 656                                 if (shm_mq_get_sender(mq) != NULL)
 657                                 {
 658                                         /* Yes, so it is known to be attached. */
 659                                         pcxt->known_attached_workers[i] = true;
 660                                         ++pcxt->nknown_attached_workers;
 661                                 }
 662                         }
 663                         else if (status == BGWH_STOPPED)
 664                         {
 665                                 /*
 666                                  * If the worker stopped without attaching to the error queue,
 667                                  * throw an error.
 668                                  */
 669                                 mq = shm_mq_get_queue(pcxt->worker[i].error_mqh);
 670                                 if (shm_mq_get_sender(mq) == NULL)
 671                                         ereport(ERROR,
 672                                                         (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 673                                                          errmsg("parallel worker failed to initialize"),
 674                                                          errhint("More details may be available in the server log.")));
 675
 676                                 pcxt->known_attached_workers[i] = true;
 677                                 ++pcxt->nknown_attached_workers;
 678                         }
 679                         else
 680                         {
 681                                 /*
 682                                  * Worker not yet started, so we must wait.  The postmaster
 683                                  * will notify us if the worker's state changes.  Our latch
 684                                  * might also get set for some other reason, but if so we'll
 685                                  * just end up waiting for the same worker again.
 686                                  */
 687                                 rc = WaitLatch(MyLatch,
 688                                                            WL_LATCH_SET | WL_EXIT_ON_PM_DEATH,
 689                                                            -1, WAIT_EVENT_BGWORKER_STARTUP);
 690
 691                                 if (rc & WL_LATCH_SET)
 692                                         ResetLatch(MyLatch);
 693                         }
 694                 }
 695
 696                 /* If all workers are known to have started, we're done. */
 697                 if (pcxt->nknown_attached_workers >= pcxt->nworkers_launched)
 698                 {
 699                         Assert(pcxt->nknown_attached_workers == pcxt->nworkers_launched);
 700                         break;
 701                 }
 702         }
 703 }
 704
 705 /*
 706  * Wait for all workers to finish computing.
 707  *
 708  * Even if the parallel operation seems to have completed successfully, it's
 709  * important to call this function afterwards.  We must not miss any errors
 710  * the workers may have thrown during the parallel operation, or any that they
 711  * may yet throw while shutting down.
 712  *
 713  * Also, we want to update our notion of XactLastRecEnd based on worker
 714  * feedback.
 715  */
 716 void
 717 WaitForParallelWorkersToFinish(ParallelContext *pcxt)
 718 {
 719         for (;;)
 720         {
 721                 bool            anyone_alive = false;
 722                 int                     nfinished = 0;
 723                 int                     i;
 724
 725                 /*
 726                  * This will process any parallel messages that are pending, which may
 727                  * change the outcome of the loop that follows.  It may also throw an
 728                  * error propagated from a worker.
 729                  */
 730                 CHECK_FOR_INTERRUPTS();
 731
 732                 for (i = 0; i < pcxt->nworkers_launched; ++i)
 733                 {
 734                         /*
 735                          * If error_mqh is NULL, then the worker has already exited
 736                          * cleanly.  If we have received a message through error_mqh from
 737                          * the worker, we know it started up cleanly, and therefore we're
 738                          * certain to be notified when it exits.
 739                          */
 740                         if (pcxt->worker[i].error_mqh == NULL)
 741                                 ++nfinished;
 742                         else if (pcxt->known_attached_workers[i])
 743                         {
 744                                 anyone_alive = true;
 745                                 break;
 746                         }
 747                 }
 748
 749                 if (!anyone_alive)
 750                 {
 751                         /* If all workers are known to have finished, we're done. */
 752                         if (nfinished >= pcxt->nworkers_launched)
 753                         {
 754                                 Assert(nfinished == pcxt->nworkers_launched);
 755                                 break;
 756                         }
 757
 758                         /*
 759                          * We didn't detect any living workers, but not all workers are
 760                          * known to have exited cleanly.  Either not all workers have
 761                          * launched yet, or maybe some of them failed to start or
 762                          * terminated abnormally.
 763                          */
 764                         for (i = 0; i < pcxt->nworkers_launched; ++i)
 765                         {
 766                                 pid_t           pid;
 767                                 shm_mq     *mq;
 768
 769                                 /*
 770                                  * If the worker is BGWH_NOT_YET_STARTED or BGWH_STARTED, we
 771                                  * should just keep waiting.  If it is BGWH_STOPPED, then
 772                                  * further investigation is needed.
 773                                  */
 774                                 if (pcxt->worker[i].error_mqh == NULL ||
 775                                         pcxt->worker[i].bgwhandle == NULL ||
 776                                         GetBackgroundWorkerPid(pcxt->worker[i].bgwhandle,
 777                                                                                    &pid) != BGWH_STOPPED)
 778                                         continue;
 779
 780                                 /*
 781                                  * Check whether the worker ended up stopped without ever
 782                                  * attaching to the error queue.  If so, the postmaster was
 783                                  * unable to fork the worker or it exited without initializing
 784                                  * properly.  We must throw an error, since the caller may
 785                                  * have been expecting the worker to do some work before
 786                                  * exiting.
 787                                  */
 788                                 mq = shm_mq_get_queue(pcxt->worker[i].error_mqh);
 789                                 if (shm_mq_get_sender(mq) == NULL)
 790                                         ereport(ERROR,
 791                                                         (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 792                                                          errmsg("parallel worker failed to initialize"),
 793                                                          errhint("More details may be available in the server log.")));
 794
 795                                 /*
 796                                  * The worker is stopped, but is attached to the error queue.
 797                                  * Unless there's a bug somewhere, this will only happen when
 798                                  * the worker writes messages and terminates after the
 799                                  * CHECK_FOR_INTERRUPTS() near the top of this function and
 800                                  * before the call to GetBackgroundWorkerPid().  In that case,
 801                                  * or latch should have been set as well and the right things
 802                                  * will happen on the next pass through the loop.
 803                                  */
 804                         }
 805                 }
 806
 807                 (void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, -1,
 808                                                  WAIT_EVENT_PARALLEL_FINISH);
 809                 ResetLatch(MyLatch);
 810         }
 811
 812         if (pcxt->toc != NULL)
 813         {
 814                 FixedParallelState *fps;
 815
 816                 fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED, false);
 817                 if (fps->last_xlog_end > XactLastRecEnd)
 818                         XactLastRecEnd = fps->last_xlog_end;
 819         }
 820 }
 821
 822 /*
 823  * Wait for all workers to exit.
 824  *
 825  * This function ensures that workers have been completely shutdown.  The
 826  * difference between WaitForParallelWorkersToFinish and this function is
 827  * that former just ensures that last message sent by worker backend is
 828  * received by master backend whereas this ensures the complete shutdown.
 829  */
 830 static void
 831 WaitForParallelWorkersToExit(ParallelContext *pcxt)
 832 {
 833         int                     i;
 834
 835         /* Wait until the workers actually die. */
 836         for (i = 0; i < pcxt->nworkers_launched; ++i)
 837         {
 838                 BgwHandleStatus status;
 839
 840                 if (pcxt->worker == NULL || pcxt->worker[i].bgwhandle == NULL)
 841                         continue;
 842
 843                 status = WaitForBackgroundWorkerShutdown(pcxt->worker[i].bgwhandle);
 844
 845                 /*
 846                  * If the postmaster kicked the bucket, we have no chance of cleaning
 847                  * up safely -- we won't be able to tell when our workers are actually
 848                  * dead.  This doesn't necessitate a PANIC since they will all abort
 849                  * eventually, but we can't safely continue this session.
 850                  */
 851                 if (status == BGWH_POSTMASTER_DIED)
 852                         ereport(FATAL,
 853                                         (errcode(ERRCODE_ADMIN_SHUTDOWN),
 854                                          errmsg("postmaster exited during a parallel transaction")));
 855
 856                 /* Release memory. */
 857                 pfree(pcxt->worker[i].bgwhandle);
 858                 pcxt->worker[i].bgwhandle = NULL;
 859         }
 860 }
 861
 862 /*
 863  * Destroy a parallel context.
 864  *
 865  * If expecting a clean exit, you should use WaitForParallelWorkersToFinish()
 866  * first, before calling this function.  When this function is invoked, any
 867  * remaining workers are forcibly killed; the dynamic shared memory segment
 868  * is unmapped; and we then wait (uninterruptibly) for the workers to exit.
 869  */
 870 void
 871 DestroyParallelContext(ParallelContext *pcxt)
 872 {
 873         int                     i;
 874
 875         /*
 876          * Be careful about order of operations here!  We remove the parallel
 877          * context from the list before we do anything else; otherwise, if an
 878          * error occurs during a subsequent step, we might try to nuke it again
 879          * from AtEOXact_Parallel or AtEOSubXact_Parallel.
 880          */
 881         dlist_delete(&pcxt->node);
 882
 883         /* Kill each worker in turn, and forget their error queues. */
 884         if (pcxt->worker != NULL)
 885         {
 886                 for (i = 0; i < pcxt->nworkers_launched; ++i)
 887                 {
 888                         if (pcxt->worker[i].error_mqh != NULL)
 889                         {
 890                                 TerminateBackgroundWorker(pcxt->worker[i].bgwhandle);
 891
 892                                 shm_mq_detach(pcxt->worker[i].error_mqh);
 893                                 pcxt->worker[i].error_mqh = NULL;
 894                         }
 895                 }
 896         }
 897
 898         /*
 899          * If we have allocated a shared memory segment, detach it.  This will
 900          * implicitly detach the error queues, and any other shared memory queues,
 901          * stored there.
 902          */
 903         if (pcxt->seg != NULL)
 904         {
 905                 dsm_detach(pcxt->seg);
 906                 pcxt->seg = NULL;
 907         }
 908
 909         /*
 910          * If this parallel context is actually in backend-private memory rather
 911          * than shared memory, free that memory instead.
 912          */
 913         if (pcxt->private_memory != NULL)
 914         {
 915                 pfree(pcxt->private_memory);
 916                 pcxt->private_memory = NULL;
 917         }
 918
 919         /*
 920          * We can't finish transaction commit or abort until all of the workers
 921          * have exited.  This means, in particular, that we can't respond to
 922          * interrupts at this stage.
 923          */
 924         HOLD_INTERRUPTS();
 925         WaitForParallelWorkersToExit(pcxt);
 926         RESUME_INTERRUPTS();
 927
 928         /* Free the worker array itself. */
 929         if (pcxt->worker != NULL)
 930         {
 931                 pfree(pcxt->worker);
 932                 pcxt->worker = NULL;
 933         }
 934
 935         /* Free memory. */
 936         pfree(pcxt->library_name);
 937         pfree(pcxt->function_name);
 938         pfree(pcxt);
 939 }
 940
 941 /*
 942  * Are there any parallel contexts currently active?
 943  */
 944 bool
 945 ParallelContextActive(void)
 946 {
 947         return !dlist_is_empty(&pcxt_list);
 948 }
 949
 950 /*
 951  * Handle receipt of an interrupt indicating a parallel worker message.
 952  *
 953  * Note: this is called within a signal handler!  All we can do is set
 954  * a flag that will cause the next CHECK_FOR_INTERRUPTS() to invoke
 955  * HandleParallelMessages().
 956  */
 957 void
 958 HandleParallelMessageInterrupt(void)
 959 {
 960         InterruptPending = true;
 961         ParallelMessagePending = true;
 962         SetLatch(MyLatch);
 963 }
 964
 965 /*
 966  * Handle any queued protocol messages received from parallel workers.
 967  */
 968 void
 969 HandleParallelMessages(void)
 970 {
 971         dlist_iter      iter;
 972         MemoryContext oldcontext;
 973
 974         static MemoryContext hpm_context = NULL;
 975
 976         /*
 977          * This is invoked from ProcessInterrupts(), and since some of the
 978          * functions it calls contain CHECK_FOR_INTERRUPTS(), there is a potential
 979          * for recursive calls if more signals are received while this runs.  It's
 980          * unclear that recursive entry would be safe, and it doesn't seem useful
 981          * even if it is safe, so let's block interrupts until done.
 982          */
 983         HOLD_INTERRUPTS();
 984
 985         /*
 986          * Moreover, CurrentMemoryContext might be pointing almost anywhere.  We
 987          * don't want to risk leaking data into long-lived contexts, so let's do
 988          * our work here in a private context that we can reset on each use.
 989          */
 990         if (hpm_context == NULL)        /* first time through? */
 991                 hpm_context = AllocSetContextCreate(TopMemoryContext,
 992                                                                                         "HandleParallelMessages",
 993                                                                                         ALLOCSET_DEFAULT_SIZES);
 994         else
 995                 MemoryContextReset(hpm_context);
 996
 997         oldcontext = MemoryContextSwitchTo(hpm_context);
 998
 999         /* OK to process messages.  Reset the flag saying there are more to do. */
1000         ParallelMessagePending = false;
1001
1002         dlist_foreach(iter, &pcxt_list)
1003         {
1004                 ParallelContext *pcxt;
1005                 int                     i;
1006
1007                 pcxt = dlist_container(ParallelContext, node, iter.cur);
1008                 if (pcxt->worker == NULL)
1009                         continue;
1010
1011                 for (i = 0; i < pcxt->nworkers_launched; ++i)
1012                 {
1013                         /*
1014                          * Read as many messages as we can from each worker, but stop when
1015                          * either (1) the worker's error queue goes away, which can happen
1016                          * if we receive a Terminate message from the worker; or (2) no
1017                          * more messages can be read from the worker without blocking.
1018                          */
1019                         while (pcxt->worker[i].error_mqh != NULL)
1020                         {
1021                                 shm_mq_result res;
1022                                 Size            nbytes;
1023                                 void       *data;
1024
1025                                 res = shm_mq_receive(pcxt->worker[i].error_mqh, &nbytes,
1026                                                                          &data, true);
1027                                 if (res == SHM_MQ_WOULD_BLOCK)
1028                                         break;
1029                                 else if (res == SHM_MQ_SUCCESS)
1030                                 {
1031                                         StringInfoData msg;
1032
1033                                         initStringInfo(&msg);
1034                                         appendBinaryStringInfo(&msg, data, nbytes);
1035                                         HandleParallelMessage(pcxt, i, &msg);
1036                                         pfree(msg.data);
1037                                 }
1038                                 else
1039                                         ereport(ERROR,
1040                                                         (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1041                                                          errmsg("lost connection to parallel worker")));
1042                         }
1043                 }
1044         }
1045
1046         MemoryContextSwitchTo(oldcontext);
1047
1048         /* Might as well clear the context on our way out */
1049         MemoryContextReset(hpm_context);
1050
1051         RESUME_INTERRUPTS();
1052 }
1053
1054 /*
1055  * Handle a single protocol message received from a single parallel worker.
1056  */
1057 static void
1058 HandleParallelMessage(ParallelContext *pcxt, int i, StringInfo msg)
1059 {
1060         char            msgtype;
1061
1062         if (pcxt->known_attached_workers != NULL &&
1063                 !pcxt->known_attached_workers[i])
1064         {
1065                 pcxt->known_attached_workers[i] = true;
1066                 pcxt->nknown_attached_workers++;
1067         }
1068
1069         msgtype = pq_getmsgbyte(msg);
1070
1071         switch (msgtype)
1072         {
1073                 case 'K':                               /* BackendKeyData */
1074                         {
1075                                 int32           pid = pq_getmsgint(msg, 4);
1076
1077                                 (void) pq_getmsgint(msg, 4);    /* discard cancel key */
1078                                 (void) pq_getmsgend(msg);
1079                                 pcxt->worker[i].pid = pid;
1080                                 break;
1081                         }
1082
1083                 case 'E':                               /* ErrorResponse */
1084                 case 'N':                               /* NoticeResponse */
1085                         {
1086                                 ErrorData       edata;
1087                                 ErrorContextCallback *save_error_context_stack;
1088
1089                                 /* Parse ErrorResponse or NoticeResponse. */
1090                                 pq_parse_errornotice(msg, &edata);
1091
1092                                 /* Death of a worker isn't enough justification for suicide. */
1093                                 edata.elevel = Min(edata.elevel, ERROR);
1094
1095                                 /*
1096                                  * If desired, add a context line to show that this is a
1097                                  * message propagated from a parallel worker.  Otherwise, it
1098                                  * can sometimes be confusing to understand what actually
1099                                  * happened.  (We don't do this in FORCE_PARALLEL_REGRESS mode
1100                                  * because it causes test-result instability depending on
1101                                  * whether a parallel worker is actually used or not.)
1102                                  */
1103                                 if (force_parallel_mode != FORCE_PARALLEL_REGRESS)
1104                                 {
1105                                         if (edata.context)
1106                                                 edata.context = psprintf("%s\n%s", edata.context,
1107                                                                                                  _("parallel worker"));
1108                                         else
1109                                                 edata.context = pstrdup(_("parallel worker"));
1110                                 }
1111
1112                                 /*
1113                                  * Context beyond that should use the error context callbacks
1114                                  * that were in effect when the ParallelContext was created,
1115                                  * not the current ones.
1116                                  */
1117                                 save_error_context_stack = error_context_stack;
1118                                 error_context_stack = pcxt->error_context_stack;
1119
1120                                 /* Rethrow error or print notice. */
1121                                 ThrowErrorData(&edata);
1122
1123                                 /* Not an error, so restore previous context stack. */
1124                                 error_context_stack = save_error_context_stack;
1125
1126                                 break;
1127                         }
1128
1129                 case 'A':                               /* NotifyResponse */
1130                         {
1131                                 /* Propagate NotifyResponse. */
1132                                 int32           pid;
1133                                 const char *channel;
1134                                 const char *payload;
1135
1136                                 pid = pq_getmsgint(msg, 4);
1137                                 channel = pq_getmsgrawstring(msg);
1138                                 payload = pq_getmsgrawstring(msg);
1139                                 pq_endmessage(msg);
1140
1141                                 NotifyMyFrontEnd(channel, payload, pid);
1142
1143                                 break;
1144                         }
1145
1146                 case 'X':                               /* Terminate, indicating clean exit */
1147                         {
1148                                 shm_mq_detach(pcxt->worker[i].error_mqh);
1149                                 pcxt->worker[i].error_mqh = NULL;
1150                                 break;
1151                         }
1152
1153                 default:
1154                         {
1155                                 elog(ERROR, "unrecognized message type received from parallel worker: %c (message length %d bytes)",
1156                                          msgtype, msg->len);
1157                         }
1158         }
1159 }
1160
1161 /*
1162  * End-of-subtransaction cleanup for parallel contexts.
1163  *
1164  * Currently, it's forbidden to enter or leave a subtransaction while
1165  * parallel mode is in effect, so we could just blow away everything.  But
1166  * we may want to relax that restriction in the future, so this code
1167  * contemplates that there may be multiple subtransaction IDs in pcxt_list.
1168  */
1169 void
1170 AtEOSubXact_Parallel(bool isCommit, SubTransactionId mySubId)
1171 {
1172         while (!dlist_is_empty(&pcxt_list))
1173         {
1174                 ParallelContext *pcxt;
1175
1176                 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
1177                 if (pcxt->subid != mySubId)
1178                         break;
1179                 if (isCommit)
1180                         elog(WARNING, "leaked parallel context");
1181                 DestroyParallelContext(pcxt);
1182         }
1183 }
1184
1185 /*
1186  * End-of-transaction cleanup for parallel contexts.
1187  */
1188 void
1189 AtEOXact_Parallel(bool isCommit)
1190 {
1191         while (!dlist_is_empty(&pcxt_list))
1192         {
1193                 ParallelContext *pcxt;
1194
1195                 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
1196                 if (isCommit)
1197                         elog(WARNING, "leaked parallel context");
1198                 DestroyParallelContext(pcxt);
1199         }
1200 }
1201
1202 /*
1203  * Main entrypoint for parallel workers.
1204  */
1205 void
1206 ParallelWorkerMain(Datum main_arg)
1207 {
1208         dsm_segment *seg;
1209         shm_toc    *toc;
1210         FixedParallelState *fps;
1211         char       *error_queue_space;
1212         shm_mq     *mq;
1213         shm_mq_handle *mqh;
1214         char       *libraryspace;
1215         char       *entrypointstate;
1216         char       *library_name;
1217         char       *function_name;
1218         parallel_worker_main_type entrypt;
1219         char       *gucspace;
1220         char       *combocidspace;
1221         char       *tsnapspace;
1222         char       *asnapspace;
1223         char       *tstatespace;
1224         char       *reindexspace;
1225         char       *relmapperspace;
1226         char       *enumblacklistspace;
1227         StringInfoData msgbuf;
1228         char       *session_dsm_handle_space;
1229
1230         /* Set flag to indicate that we're initializing a parallel worker. */
1231         InitializingParallelWorker = true;
1232
1233         /* Establish signal handlers. */
1234         pqsignal(SIGTERM, die);
1235         BackgroundWorkerUnblockSignals();
1236
1237         /* Determine and set our parallel worker number. */
1238         Assert(ParallelWorkerNumber == -1);
1239         memcpy(&ParallelWorkerNumber, MyBgworkerEntry->bgw_extra, sizeof(int));
1240
1241         /* Set up a memory context to work in, just for cleanliness. */
1242         CurrentMemoryContext = AllocSetContextCreate(TopMemoryContext,
1243                                                                                                  "Parallel worker",
1244                                                                                                  ALLOCSET_DEFAULT_SIZES);
1245
1246         /*
1247          * Attach to the dynamic shared memory segment for the parallel query, and
1248          * find its table of contents.
1249          *
1250          * Note: at this point, we have not created any ResourceOwner in this
1251          * process.  This will result in our DSM mapping surviving until process
1252          * exit, which is fine.  If there were a ResourceOwner, it would acquire
1253          * ownership of the mapping, but we have no need for that.
1254          */
1255         seg = dsm_attach(DatumGetUInt32(main_arg));
1256         if (seg == NULL)
1257                 ereport(ERROR,
1258                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1259                                  errmsg("could not map dynamic shared memory segment")));
1260         toc = shm_toc_attach(PARALLEL_MAGIC, dsm_segment_address(seg));
1261         if (toc == NULL)
1262                 ereport(ERROR,
1263                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1264                                  errmsg("invalid magic number in dynamic shared memory segment")));
1265
1266         /* Look up fixed parallel state. */
1267         fps = shm_toc_lookup(toc, PARALLEL_KEY_FIXED, false);
1268         MyFixedParallelState = fps;
1269
1270         /* Arrange to signal the leader if we exit. */
1271         ParallelMasterPid = fps->parallel_master_pid;
1272         ParallelMasterBackendId = fps->parallel_master_backend_id;
1273         on_shmem_exit(ParallelWorkerShutdown, (Datum) 0);
1274
1275         /*
1276          * Now we can find and attach to the error queue provided for us.  That's
1277          * good, because until we do that, any errors that happen here will not be
1278          * reported back to the process that requested that this worker be
1279          * launched.
1280          */
1281         error_queue_space = shm_toc_lookup(toc, PARALLEL_KEY_ERROR_QUEUE, false);
1282         mq = (shm_mq *) (error_queue_space +
1283                                          ParallelWorkerNumber * PARALLEL_ERROR_QUEUE_SIZE);
1284         shm_mq_set_sender(mq, MyProc);
1285         mqh = shm_mq_attach(mq, seg, NULL);
1286         pq_redirect_to_shm_mq(seg, mqh);
1287         pq_set_parallel_master(fps->parallel_master_pid,
1288                                                    fps->parallel_master_backend_id);
1289
1290         /*
1291          * Send a BackendKeyData message to the process that initiated parallelism
1292          * so that it has access to our PID before it receives any other messages
1293          * from us.  Our cancel key is sent, too, since that's the way the
1294          * protocol message is defined, but it won't actually be used for anything
1295          * in this case.
1296          */
1297         pq_beginmessage(&msgbuf, 'K');
1298         pq_sendint32(&msgbuf, (int32) MyProcPid);
1299         pq_sendint32(&msgbuf, (int32) MyCancelKey);
1300         pq_endmessage(&msgbuf);
1301
1302         /*
1303          * Hooray! Primary initialization is complete.  Now, we need to set up our
1304          * backend-local state to match the original backend.
1305          */
1306
1307         /*
1308          * Join locking group.  We must do this before anything that could try to
1309          * acquire a heavyweight lock, because any heavyweight locks acquired to
1310          * this point could block either directly against the parallel group
1311          * leader or against some process which in turn waits for a lock that
1312          * conflicts with the parallel group leader, causing an undetected
1313          * deadlock.  (If we can't join the lock group, the leader has gone away,
1314          * so just exit quietly.)
1315          */
1316         if (!BecomeLockGroupMember(fps->parallel_master_pgproc,
1317                                                            fps->parallel_master_pid))
1318                 return;
1319
1320         /*
1321          * Restore transaction and statement start-time timestamps.  This must
1322          * happen before anything that would start a transaction, else asserts in
1323          * xact.c will fire.
1324          */
1325         SetParallelStartTimestamps(fps->xact_ts, fps->stmt_ts);
1326
1327         /*
1328          * Identify the entry point to be called.  In theory this could result in
1329          * loading an additional library, though most likely the entry point is in
1330          * the core backend or in a library we just loaded.
1331          */
1332         entrypointstate = shm_toc_lookup(toc, PARALLEL_KEY_ENTRYPOINT, false);
1333         library_name = entrypointstate;
1334         function_name = entrypointstate + strlen(library_name) + 1;
1335
1336         entrypt = LookupParallelWorkerFunction(library_name, function_name);
1337
1338         /* Restore database connection. */
1339         BackgroundWorkerInitializeConnectionByOid(fps->database_id,
1340                                                                                           fps->authenticated_user_id,
1341                                                                                           0);
1342
1343         /*
1344          * Set the client encoding to the database encoding, since that is what
1345          * the leader will expect.
1346          */
1347         SetClientEncoding(GetDatabaseEncoding());
1348
1349         /*
1350          * Load libraries that were loaded by original backend.  We want to do
1351          * this before restoring GUCs, because the libraries might define custom
1352          * variables.
1353          */
1354         libraryspace = shm_toc_lookup(toc, PARALLEL_KEY_LIBRARY, false);
1355         StartTransactionCommand();
1356         RestoreLibraryState(libraryspace);
1357
1358         /* Restore GUC values from launching backend. */
1359         gucspace = shm_toc_lookup(toc, PARALLEL_KEY_GUC, false);
1360         RestoreGUCState(gucspace);
1361         CommitTransactionCommand();
1362
1363         /* Crank up a transaction state appropriate to a parallel worker. */
1364         tstatespace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_STATE, false);
1365         StartParallelWorkerTransaction(tstatespace);
1366
1367         /* Restore combo CID state. */
1368         combocidspace = shm_toc_lookup(toc, PARALLEL_KEY_COMBO_CID, false);
1369         RestoreComboCIDState(combocidspace);
1370
1371         /* Attach to the per-session DSM segment and contained objects. */
1372         session_dsm_handle_space =
1373                 shm_toc_lookup(toc, PARALLEL_KEY_SESSION_DSM, false);
1374         AttachSession(*(dsm_handle *) session_dsm_handle_space);
1375
1376         /* Restore transaction snapshot. */
1377         tsnapspace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT, false);
1378         RestoreTransactionSnapshot(RestoreSnapshot(tsnapspace),
1379                                                            fps->parallel_master_pgproc);
1380
1381         /* Restore active snapshot. */
1382         asnapspace = shm_toc_lookup(toc, PARALLEL_KEY_ACTIVE_SNAPSHOT, false);
1383         PushActiveSnapshot(RestoreSnapshot(asnapspace));
1384
1385         /*
1386          * We've changed which tuples we can see, and must therefore invalidate
1387          * system caches.
1388          */
1389         InvalidateSystemCaches();
1390
1391         /*
1392          * Restore current role id.  Skip verifying whether session user is
1393          * allowed to become this role and blindly restore the leader's state for
1394          * current role.
1395          */
1396         SetCurrentRoleId(fps->outer_user_id, fps->is_superuser);
1397
1398         /* Restore user ID and security context. */
1399         SetUserIdAndSecContext(fps->current_user_id, fps->sec_context);
1400
1401         /* Restore temp-namespace state to ensure search path matches leader's. */
1402         SetTempNamespaceState(fps->temp_namespace_id,
1403                                                   fps->temp_toast_namespace_id);
1404
1405         /* Restore reindex state. */
1406         reindexspace = shm_toc_lookup(toc, PARALLEL_KEY_REINDEX_STATE, false);
1407         RestoreReindexState(reindexspace);
1408
1409         /* Restore relmapper state. */
1410         relmapperspace = shm_toc_lookup(toc, PARALLEL_KEY_RELMAPPER_STATE, false);
1411         RestoreRelationMap(relmapperspace);
1412
1413         /* Restore enum blacklist. */
1414         enumblacklistspace = shm_toc_lookup(toc, PARALLEL_KEY_ENUMBLACKLIST,
1415                                                                                 false);
1416         RestoreEnumBlacklist(enumblacklistspace);
1417
1418         /* Attach to the leader's serializable transaction, if SERIALIZABLE. */
1419         AttachSerializableXact(fps->serializable_xact_handle);
1420
1421         /*
1422          * We've initialized all of our state now; nothing should change
1423          * hereafter.
1424          */
1425         InitializingParallelWorker = false;
1426         EnterParallelMode();
1427
1428         /*
1429          * Time to do the real work: invoke the caller-supplied code.
1430          */
1431         entrypt(seg, toc);
1432
1433         /* Must exit parallel mode to pop active snapshot. */
1434         ExitParallelMode();
1435
1436         /* Must pop active snapshot so snapmgr.c doesn't complain. */
1437         PopActiveSnapshot();
1438
1439         /* Shut down the parallel-worker transaction. */
1440         EndParallelWorkerTransaction();
1441
1442         /* Detach from the per-session DSM segment. */
1443         DetachSession();
1444
1445         /* Report success. */
1446         pq_putmessage('X', NULL, 0);
1447 }
1448
1449 /*
1450  * Update shared memory with the ending location of the last WAL record we
1451  * wrote, if it's greater than the value already stored there.
1452  */
1453 void
1454 ParallelWorkerReportLastRecEnd(XLogRecPtr last_xlog_end)
1455 {
1456         FixedParallelState *fps = MyFixedParallelState;
1457
1458         Assert(fps != NULL);
1459         SpinLockAcquire(&fps->mutex);
1460         if (fps->last_xlog_end < last_xlog_end)
1461                 fps->last_xlog_end = last_xlog_end;
1462         SpinLockRelease(&fps->mutex);
1463 }
1464
1465 /*
1466  * Make sure the leader tries to read from our error queue one more time.
1467  * This guards against the case where we exit uncleanly without sending an
1468  * ErrorResponse to the leader, for example because some code calls proc_exit
1469  * directly.
1470  */
1471 static void
1472 ParallelWorkerShutdown(int code, Datum arg)
1473 {
1474         SendProcSignal(ParallelMasterPid,
1475                                    PROCSIG_PARALLEL_MESSAGE,
1476                                    ParallelMasterBackendId);
1477 }
1478
1479 /*
1480  * Look up (and possibly load) a parallel worker entry point function.
1481  *
1482  * For functions contained in the core code, we use library name "postgres"
1483  * and consult the InternalParallelWorkers array.  External functions are
1484  * looked up, and loaded if necessary, using load_external_function().
1485  *
1486  * The point of this is to pass function names as strings across process
1487  * boundaries.  We can't pass actual function addresses because of the
1488  * possibility that the function has been loaded at a different address
1489  * in a different process.  This is obviously a hazard for functions in
1490  * loadable libraries, but it can happen even for functions in the core code
1491  * on platforms using EXEC_BACKEND (e.g., Windows).
1492  *
1493  * At some point it might be worthwhile to get rid of InternalParallelWorkers[]
1494  * in favor of applying load_external_function() for core functions too;
1495  * but that raises portability issues that are not worth addressing now.
1496  */
1497 static parallel_worker_main_type
1498 LookupParallelWorkerFunction(const char *libraryname, const char *funcname)
1499 {
1500         /*
1501          * If the function is to be loaded from postgres itself, search the
1502          * InternalParallelWorkers array.
1503          */
1504         if (strcmp(libraryname, "postgres") == 0)
1505         {
1506                 int                     i;
1507
1508                 for (i = 0; i < lengthof(InternalParallelWorkers); i++)
1509                 {
1510                         if (strcmp(InternalParallelWorkers[i].fn_name, funcname) == 0)
1511                                 return InternalParallelWorkers[i].fn_addr;
1512                 }
1513
1514                 /* We can only reach this by programming error. */
1515                 elog(ERROR, "internal function \"%s\" not found", funcname);
1516         }
1517
1518         /* Otherwise load from external library. */
1519         return (parallel_worker_main_type)
1520                 load_external_function(libraryname, funcname, true, NULL);
1521 }