granicus.if.org Git - postgresql/blob - src/backend/postmaster/bgworker.c

   1 /*--------------------------------------------------------------------
   2  * bgworker.c
   3  *              POSTGRES pluggable background workers implementation
   4  *
   5  * Portions Copyright (c) 1996-2015, PostgreSQL Global Development Group
   6  *
   7  * IDENTIFICATION
   8  *        src/backend/postmaster/bgworker.c
   9  *
  10  *-------------------------------------------------------------------------
  11  */
  12
  13 #include "postgres.h"
  14
  15 #include <unistd.h>
  16
  17 #include "miscadmin.h"
  18 #include "libpq/pqsignal.h"
  19 #include "postmaster/bgworker_internals.h"
  20 #include "postmaster/postmaster.h"
  21 #include "storage/barrier.h"
  22 #include "storage/dsm.h"
  23 #include "storage/ipc.h"
  24 #include "storage/latch.h"
  25 #include "storage/lwlock.h"
  26 #include "storage/pg_shmem.h"
  27 #include "storage/pmsignal.h"
  28 #include "storage/proc.h"
  29 #include "storage/procsignal.h"
  30 #include "storage/shmem.h"
  31 #include "tcop/tcopprot.h"
  32 #include "utils/ascii.h"
  33 #include "utils/ps_status.h"
  34 #include "utils/timeout.h"
  35
  36 /*
  37  * The postmaster's list of registered background workers, in private memory.
  38  */
  39 slist_head      BackgroundWorkerList = SLIST_STATIC_INIT(BackgroundWorkerList);
  40
  41 /*
  42  * BackgroundWorkerSlots exist in shared memory and can be accessed (via
  43  * the BackgroundWorkerArray) by both the postmaster and by regular backends.
  44  * However, the postmaster cannot take locks, even spinlocks, because this
  45  * might allow it to crash or become wedged if shared memory gets corrupted.
  46  * Such an outcome is intolerable.  Therefore, we need a lockless protocol
  47  * for coordinating access to this data.
  48  *
  49  * The 'in_use' flag is used to hand off responsibility for the slot between
  50  * the postmaster and the rest of the system.  When 'in_use' is false,
  51  * the postmaster will ignore the slot entirely, except for the 'in_use' flag
  52  * itself, which it may read.  In this state, regular backends may modify the
  53  * slot.  Once a backend sets 'in_use' to true, the slot becomes the
  54  * responsibility of the postmaster.  Regular backends may no longer modify it,
  55  * but the postmaster may examine it.  Thus, a backend initializing a slot
  56  * must fully initialize the slot - and insert a write memory barrier - before
  57  * marking it as in use.
  58  *
  59  * As an exception, however, even when the slot is in use, regular backends
  60  * may set the 'terminate' flag for a slot, telling the postmaster not
  61  * to restart it.  Once the background worker is no longer running, the slot
  62  * will be released for reuse.
  63  *
  64  * In addition to coordinating with the postmaster, backends modifying this
  65  * data structure must coordinate with each other.  Since they can take locks,
  66  * this is straightforward: any backend wishing to manipulate a slot must
  67  * take BackgroundWorkerLock in exclusive mode.  Backends wishing to read
  68  * data that might get concurrently modified by other backends should take
  69  * this lock in shared mode.  No matter what, backends reading this data
  70  * structure must be able to tolerate concurrent modifications by the
  71  * postmaster.
  72  */
  73 typedef struct BackgroundWorkerSlot
  74 {
  75         bool            in_use;
  76         bool            terminate;
  77         pid_t           pid;                    /* InvalidPid = not started yet; 0 = dead */
  78         uint64          generation;             /* incremented when slot is recycled */
  79         BackgroundWorker worker;
  80 } BackgroundWorkerSlot;
  81
  82 typedef struct BackgroundWorkerArray
  83 {
  84         int                     total_slots;
  85         BackgroundWorkerSlot slot[FLEXIBLE_ARRAY_MEMBER];
  86 } BackgroundWorkerArray;
  87
  88 struct BackgroundWorkerHandle
  89 {
  90         int                     slot;
  91         uint64          generation;
  92 };
  93
  94 static BackgroundWorkerArray *BackgroundWorkerData;
  95
  96 /*
  97  * Calculate shared memory needed.
  98  */
  99 Size
 100 BackgroundWorkerShmemSize(void)
 101 {
 102         Size            size;
 103
 104         /* Array of workers is variably sized. */
 105         size = offsetof(BackgroundWorkerArray, slot);
 106         size = add_size(size, mul_size(max_worker_processes,
 107                                                                    sizeof(BackgroundWorkerSlot)));
 108
 109         return size;
 110 }
 111
 112 /*
 113  * Initialize shared memory.
 114  */
 115 void
 116 BackgroundWorkerShmemInit(void)
 117 {
 118         bool            found;
 119
 120         BackgroundWorkerData = ShmemInitStruct("Background Worker Data",
 121                                                                                    BackgroundWorkerShmemSize(),
 122                                                                                    &found);
 123         if (!IsUnderPostmaster)
 124         {
 125                 slist_iter      siter;
 126                 int                     slotno = 0;
 127
 128                 BackgroundWorkerData->total_slots = max_worker_processes;
 129
 130                 /*
 131                  * Copy contents of worker list into shared memory.  Record the shared
 132                  * memory slot assigned to each worker.  This ensures a 1-to-1
 133                  * correspondence between the postmaster's private list and the array
 134                  * in shared memory.
 135                  */
 136                 slist_foreach(siter, &BackgroundWorkerList)
 137                 {
 138                         BackgroundWorkerSlot *slot = &BackgroundWorkerData->slot[slotno];
 139                         RegisteredBgWorker *rw;
 140
 141                         rw = slist_container(RegisteredBgWorker, rw_lnode, siter.cur);
 142                         Assert(slotno < max_worker_processes);
 143                         slot->in_use = true;
 144                         slot->terminate = false;
 145                         slot->pid = InvalidPid;
 146                         slot->generation = 0;
 147                         rw->rw_shmem_slot = slotno;
 148                         rw->rw_worker.bgw_notify_pid = 0;       /* might be reinit after crash */
 149                         memcpy(&slot->worker, &rw->rw_worker, sizeof(BackgroundWorker));
 150                         ++slotno;
 151                 }
 152
 153                 /*
 154                  * Mark any remaining slots as not in use.
 155                  */
 156                 while (slotno < max_worker_processes)
 157                 {
 158                         BackgroundWorkerSlot *slot = &BackgroundWorkerData->slot[slotno];
 159
 160                         slot->in_use = false;
 161                         ++slotno;
 162                 }
 163         }
 164         else
 165                 Assert(found);
 166 }
 167
 168 /*
 169  * Search the postmaster's backend-private list of RegisteredBgWorker objects
 170  * for the one that maps to the given slot number.
 171  */
 172 static RegisteredBgWorker *
 173 FindRegisteredWorkerBySlotNumber(int slotno)
 174 {
 175         slist_iter      siter;
 176
 177         slist_foreach(siter, &BackgroundWorkerList)
 178         {
 179                 RegisteredBgWorker *rw;
 180
 181                 rw = slist_container(RegisteredBgWorker, rw_lnode, siter.cur);
 182                 if (rw->rw_shmem_slot == slotno)
 183                         return rw;
 184         }
 185
 186         return NULL;
 187 }
 188
 189 /*
 190  * Notice changes to shared memory made by other backends.  This code
 191  * runs in the postmaster, so we must be very careful not to assume that
 192  * shared memory contents are sane.  Otherwise, a rogue backend could take
 193  * out the postmaster.
 194  */
 195 void
 196 BackgroundWorkerStateChange(void)
 197 {
 198         int                     slotno;
 199
 200         /*
 201          * The total number of slots stored in shared memory should match our
 202          * notion of max_worker_processes.  If it does not, something is very
 203          * wrong.  Further down, we always refer to this value as
 204          * max_worker_processes, in case shared memory gets corrupted while we're
 205          * looping.
 206          */
 207         if (max_worker_processes != BackgroundWorkerData->total_slots)
 208         {
 209                 elog(LOG,
 210                          "inconsistent background worker state (max_worker_processes=%d, total_slots=%d",
 211                          max_worker_processes,
 212                          BackgroundWorkerData->total_slots);
 213                 return;
 214         }
 215
 216         /*
 217          * Iterate through slots, looking for newly-registered workers or workers
 218          * who must die.
 219          */
 220         for (slotno = 0; slotno < max_worker_processes; ++slotno)
 221         {
 222                 BackgroundWorkerSlot *slot = &BackgroundWorkerData->slot[slotno];
 223                 RegisteredBgWorker *rw;
 224
 225                 if (!slot->in_use)
 226                         continue;
 227
 228                 /*
 229                  * Make sure we don't see the in_use flag before the updated slot
 230                  * contents.
 231                  */
 232                 pg_read_barrier();
 233
 234                 /* See whether we already know about this worker. */
 235                 rw = FindRegisteredWorkerBySlotNumber(slotno);
 236                 if (rw != NULL)
 237                 {
 238                         /*
 239                          * In general, the worker data can't change after it's initially
 240                          * registered.  However, someone can set the terminate flag.
 241                          */
 242                         if (slot->terminate && !rw->rw_terminate)
 243                         {
 244                                 rw->rw_terminate = true;
 245                                 if (rw->rw_pid != 0)
 246                                         kill(rw->rw_pid, SIGTERM);
 247                                 else
 248                                 {
 249                                         /* Report never-started, now-terminated worker as dead. */
 250                                         ReportBackgroundWorkerPID(rw);
 251                                 }
 252                         }
 253                         continue;
 254                 }
 255
 256                 /*
 257                  * If the worker is marked for termination, we don't need to add it to
 258                  * the registered workers list; we can just free the slot. However, if
 259                  * bgw_notify_pid is set, the process that registered the worker may
 260                  * need to know that we've processed the terminate request, so be sure
 261                  * to signal it.
 262                  */
 263                 if (slot->terminate)
 264                 {
 265                         int                     notify_pid;
 266
 267                         /*
 268                          * We need a memory barrier here to make sure that the load of
 269                          * bgw_notify_pid completes before the store to in_use.
 270                          */
 271                         notify_pid = slot->worker.bgw_notify_pid;
 272                         pg_memory_barrier();
 273                         slot->pid = 0;
 274                         slot->in_use = false;
 275                         if (notify_pid != 0)
 276                                 kill(notify_pid, SIGUSR1);
 277
 278                         continue;
 279                 }
 280
 281                 /*
 282                  * Copy the registration data into the registered workers list.
 283                  */
 284                 rw = malloc(sizeof(RegisteredBgWorker));
 285                 if (rw == NULL)
 286                 {
 287                         ereport(LOG,
 288                                         (errcode(ERRCODE_OUT_OF_MEMORY),
 289                                          errmsg("out of memory")));
 290                         return;
 291                 }
 292
 293                 /*
 294                  * Copy strings in a paranoid way.  If shared memory is corrupted, the
 295                  * source data might not even be NUL-terminated.
 296                  */
 297                 ascii_safe_strlcpy(rw->rw_worker.bgw_name,
 298                                                    slot->worker.bgw_name, BGW_MAXLEN);
 299                 ascii_safe_strlcpy(rw->rw_worker.bgw_library_name,
 300                                                    slot->worker.bgw_library_name, BGW_MAXLEN);
 301                 ascii_safe_strlcpy(rw->rw_worker.bgw_function_name,
 302                                                    slot->worker.bgw_function_name, BGW_MAXLEN);
 303
 304                 /*
 305                  * Copy various fixed-size fields.
 306                  *
 307                  * flags, start_time, and restart_time are examined by the postmaster,
 308                  * but nothing too bad will happen if they are corrupted.  The
 309                  * remaining fields will only be examined by the child process.  It
 310                  * might crash, but we won't.
 311                  */
 312                 rw->rw_worker.bgw_flags = slot->worker.bgw_flags;
 313                 rw->rw_worker.bgw_start_time = slot->worker.bgw_start_time;
 314                 rw->rw_worker.bgw_restart_time = slot->worker.bgw_restart_time;
 315                 rw->rw_worker.bgw_main = slot->worker.bgw_main;
 316                 rw->rw_worker.bgw_main_arg = slot->worker.bgw_main_arg;
 317
 318                 /*
 319                  * Copy the PID to be notified about state changes, but only if the
 320                  * postmaster knows about a backend with that PID.  It isn't an error
 321                  * if the postmaster doesn't know about the PID, because the backend
 322                  * that requested the worker could have died (or been killed) just
 323                  * after doing so.  Nonetheless, at least until we get some experience
 324                  * with how this plays out in the wild, log a message at a relative
 325                  * high debug level.
 326                  */
 327                 rw->rw_worker.bgw_notify_pid = slot->worker.bgw_notify_pid;
 328                 if (!PostmasterMarkPIDForWorkerNotify(rw->rw_worker.bgw_notify_pid))
 329                 {
 330                         elog(DEBUG1, "worker notification PID %lu is not valid",
 331                                  (long) rw->rw_worker.bgw_notify_pid);
 332                         rw->rw_worker.bgw_notify_pid = 0;
 333                 }
 334
 335                 /* Initialize postmaster bookkeeping. */
 336                 rw->rw_backend = NULL;
 337                 rw->rw_pid = 0;
 338                 rw->rw_child_slot = 0;
 339                 rw->rw_crashed_at = 0;
 340                 rw->rw_shmem_slot = slotno;
 341                 rw->rw_terminate = false;
 342
 343                 /* Log it! */
 344                 ereport(LOG,
 345                                 (errmsg("registering background worker \"%s\"",
 346                                                 rw->rw_worker.bgw_name)));
 347
 348                 slist_push_head(&BackgroundWorkerList, &rw->rw_lnode);
 349         }
 350 }
 351
 352 /*
 353  * Forget about a background worker that's no longer needed.
 354  *
 355  * The worker must be identified by passing an slist_mutable_iter that
 356  * points to it.  This convention allows deletion of workers during
 357  * searches of the worker list, and saves having to search the list again.
 358  *
 359  * This function must be invoked only in the postmaster.
 360  */
 361 void
 362 ForgetBackgroundWorker(slist_mutable_iter *cur)
 363 {
 364         RegisteredBgWorker *rw;
 365         BackgroundWorkerSlot *slot;
 366
 367         rw = slist_container(RegisteredBgWorker, rw_lnode, cur->cur);
 368
 369         Assert(rw->rw_shmem_slot < max_worker_processes);
 370         slot = &BackgroundWorkerData->slot[rw->rw_shmem_slot];
 371         slot->in_use = false;
 372
 373         ereport(LOG,
 374                         (errmsg("unregistering background worker \"%s\"",
 375                                         rw->rw_worker.bgw_name)));
 376
 377         slist_delete_current(cur);
 378         free(rw);
 379 }
 380
 381 /*
 382  * Report the PID of a newly-launched background worker in shared memory.
 383  *
 384  * This function should only be called from the postmaster.
 385  */
 386 void
 387 ReportBackgroundWorkerPID(RegisteredBgWorker *rw)
 388 {
 389         BackgroundWorkerSlot *slot;
 390
 391         Assert(rw->rw_shmem_slot < max_worker_processes);
 392         slot = &BackgroundWorkerData->slot[rw->rw_shmem_slot];
 393         slot->pid = rw->rw_pid;
 394
 395         if (rw->rw_worker.bgw_notify_pid != 0)
 396                 kill(rw->rw_worker.bgw_notify_pid, SIGUSR1);
 397 }
 398
 399 /*
 400  * Cancel SIGUSR1 notifications for a PID belonging to an exiting backend.
 401  *
 402  * This function should only be called from the postmaster.
 403  */
 404 void
 405 BackgroundWorkerStopNotifications(pid_t pid)
 406 {
 407         slist_iter      siter;
 408
 409         slist_foreach(siter, &BackgroundWorkerList)
 410         {
 411                 RegisteredBgWorker *rw;
 412
 413                 rw = slist_container(RegisteredBgWorker, rw_lnode, siter.cur);
 414                 if (rw->rw_worker.bgw_notify_pid == pid)
 415                         rw->rw_worker.bgw_notify_pid = 0;
 416         }
 417 }
 418
 419 /*
 420  * Reset background worker crash state.
 421  *
 422  * We assume that, after a crash-and-restart cycle, background workers without
 423  * the never-restart flag should be restarted immediately, instead of waiting
 424  * for bgw_restart_time to elapse.
 425  */
 426 void
 427 ResetBackgroundWorkerCrashTimes(void)
 428 {
 429         slist_mutable_iter iter;
 430
 431         slist_foreach_modify(iter, &BackgroundWorkerList)
 432         {
 433                 RegisteredBgWorker *rw;
 434
 435                 rw = slist_container(RegisteredBgWorker, rw_lnode, iter.cur);
 436
 437                 /*
 438                  * For workers that should not be restarted, we don't want to lose the
 439                  * information that they have crashed; otherwise, they would be
 440                  * restarted, which is wrong.
 441                  */
 442                 if (rw->rw_worker.bgw_restart_time != BGW_NEVER_RESTART)
 443                         rw->rw_crashed_at = 0;
 444         }
 445 }
 446
 447 #ifdef EXEC_BACKEND
 448 /*
 449  * In EXEC_BACKEND mode, workers use this to retrieve their details from
 450  * shared memory.
 451  */
 452 BackgroundWorker *
 453 BackgroundWorkerEntry(int slotno)
 454 {
 455         static BackgroundWorker myEntry;
 456         BackgroundWorkerSlot *slot;
 457
 458         Assert(slotno < BackgroundWorkerData->total_slots);
 459         slot = &BackgroundWorkerData->slot[slotno];
 460         Assert(slot->in_use);
 461
 462         /* must copy this in case we don't intend to retain shmem access */
 463         memcpy(&myEntry, &slot->worker, sizeof myEntry);
 464         return &myEntry;
 465 }
 466 #endif
 467
 468 /*
 469  * Complain about the BackgroundWorker definition using error level elevel.
 470  * Return true if it looks ok, false if not (unless elevel >= ERROR, in
 471  * which case we won't return at all in the not-OK case).
 472  */
 473 static bool
 474 SanityCheckBackgroundWorker(BackgroundWorker *worker, int elevel)
 475 {
 476         /* sanity check for flags */
 477         if (worker->bgw_flags & BGWORKER_BACKEND_DATABASE_CONNECTION)
 478         {
 479                 if (!(worker->bgw_flags & BGWORKER_SHMEM_ACCESS))
 480                 {
 481                         ereport(elevel,
 482                                         (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
 483                                          errmsg("background worker \"%s\": must attach to shared memory in order to request a database connection",
 484                                                         worker->bgw_name)));
 485                         return false;
 486                 }
 487
 488                 if (worker->bgw_start_time == BgWorkerStart_PostmasterStart)
 489                 {
 490                         ereport(elevel,
 491                                         (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
 492                                          errmsg("background worker \"%s\": cannot request database access if starting at postmaster start",
 493                                                         worker->bgw_name)));
 494                         return false;
 495                 }
 496
 497                 /* XXX other checks? */
 498         }
 499
 500         if ((worker->bgw_restart_time < 0 &&
 501                  worker->bgw_restart_time != BGW_NEVER_RESTART) ||
 502                 (worker->bgw_restart_time > USECS_PER_DAY / 1000))
 503         {
 504                 ereport(elevel,
 505                                 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
 506                                  errmsg("background worker \"%s\": invalid restart interval",
 507                                                 worker->bgw_name)));
 508                 return false;
 509         }
 510
 511         return true;
 512 }
 513
 514 static void
 515 bgworker_quickdie(SIGNAL_ARGS)
 516 {
 517         sigaddset(&BlockSig, SIGQUIT);          /* prevent nested calls */
 518         PG_SETMASK(&BlockSig);
 519
 520         /*
 521          * We DO NOT want to run proc_exit() callbacks -- we're here because
 522          * shared memory may be corrupted, so we don't want to try to clean up our
 523          * transaction.  Just nail the windows shut and get out of town.  Now that
 524          * there's an atexit callback to prevent third-party code from breaking
 525          * things by calling exit() directly, we have to reset the callbacks
 526          * explicitly to make this work as intended.
 527          */
 528         on_exit_reset();
 529
 530         /*
 531          * Note we do exit(2) not exit(0).  This is to force the postmaster into a
 532          * system reset cycle if some idiot DBA sends a manual SIGQUIT to a random
 533          * backend.  This is necessary precisely because we don't clean up our
 534          * shared memory state.  (The "dead man switch" mechanism in pmsignal.c
 535          * should ensure the postmaster sees this as a crash, too, but no harm in
 536          * being doubly sure.)
 537          */
 538         exit(2);
 539 }
 540
 541 /*
 542  * Standard SIGTERM handler for background workers
 543  */
 544 static void
 545 bgworker_die(SIGNAL_ARGS)
 546 {
 547         PG_SETMASK(&BlockSig);
 548
 549         ereport(FATAL,
 550                         (errcode(ERRCODE_ADMIN_SHUTDOWN),
 551                          errmsg("terminating background worker \"%s\" due to administrator command",
 552                                         MyBgworkerEntry->bgw_name)));
 553 }
 554
 555 /*
 556  * Standard SIGUSR1 handler for unconnected workers
 557  *
 558  * Here, we want to make sure an unconnected worker will at least heed
 559  * latch activity.
 560  */
 561 static void
 562 bgworker_sigusr1_handler(SIGNAL_ARGS)
 563 {
 564         int                     save_errno = errno;
 565
 566         latch_sigusr1_handler();
 567
 568         errno = save_errno;
 569 }
 570
 571 /*
 572  * Start a new background worker
 573  *
 574  * This is the main entry point for background worker, to be called from
 575  * postmaster.
 576  */
 577 void
 578 StartBackgroundWorker(void)
 579 {
 580         sigjmp_buf      local_sigjmp_buf;
 581         char            buf[MAXPGPATH];
 582         BackgroundWorker *worker = MyBgworkerEntry;
 583         bgworker_main_type entrypt;
 584
 585         if (worker == NULL)
 586                 elog(FATAL, "unable to find bgworker entry");
 587
 588         IsBackgroundWorker = true;
 589
 590         /* Identify myself via ps */
 591         snprintf(buf, MAXPGPATH, "bgworker: %s", worker->bgw_name);
 592         init_ps_display(buf, "", "", "");
 593
 594         /*
 595          * If we're not supposed to have shared memory access, then detach from
 596          * shared memory.  If we didn't request shared memory access, the
 597          * postmaster won't force a cluster-wide restart if we exit unexpectedly,
 598          * so we'd better make sure that we don't mess anything up that would
 599          * require that sort of cleanup.
 600          */
 601         if ((worker->bgw_flags & BGWORKER_SHMEM_ACCESS) == 0)
 602         {
 603                 on_exit_reset();
 604                 dsm_detach_all();
 605                 PGSharedMemoryDetach();
 606         }
 607
 608         SetProcessingMode(InitProcessing);
 609
 610         /* Apply PostAuthDelay */
 611         if (PostAuthDelay > 0)
 612                 pg_usleep(PostAuthDelay * 1000000L);
 613
 614         /*
 615          * Set up signal handlers.
 616          */
 617         if (worker->bgw_flags & BGWORKER_BACKEND_DATABASE_CONNECTION)
 618         {
 619                 /*
 620                  * SIGINT is used to signal canceling the current action
 621                  */
 622                 pqsignal(SIGINT, StatementCancelHandler);
 623                 pqsignal(SIGUSR1, procsignal_sigusr1_handler);
 624                 pqsignal(SIGFPE, FloatExceptionHandler);
 625
 626                 /* XXX Any other handlers needed here? */
 627         }
 628         else
 629         {
 630                 pqsignal(SIGINT, SIG_IGN);
 631                 pqsignal(SIGUSR1, bgworker_sigusr1_handler);
 632                 pqsignal(SIGFPE, SIG_IGN);
 633         }
 634         pqsignal(SIGTERM, bgworker_die);
 635         pqsignal(SIGHUP, SIG_IGN);
 636
 637         pqsignal(SIGQUIT, bgworker_quickdie);
 638         InitializeTimeouts();           /* establishes SIGALRM handler */
 639
 640         pqsignal(SIGPIPE, SIG_IGN);
 641         pqsignal(SIGUSR2, SIG_IGN);
 642         pqsignal(SIGCHLD, SIG_DFL);
 643
 644         /*
 645          * If an exception is encountered, processing resumes here.
 646          *
 647          * See notes in postgres.c about the design of this coding.
 648          */
 649         if (sigsetjmp(local_sigjmp_buf, 1) != 0)
 650         {
 651                 /* Since not using PG_TRY, must reset error stack by hand */
 652                 error_context_stack = NULL;
 653
 654                 /* Prevent interrupts while cleaning up */
 655                 HOLD_INTERRUPTS();
 656
 657                 /* Report the error to the server log */
 658                 EmitErrorReport();
 659
 660                 /*
 661                  * Do we need more cleanup here?  For shmem-connected bgworkers, we
 662                  * will call InitProcess below, which will install ProcKill as exit
 663                  * callback.  That will take care of releasing locks, etc.
 664                  */
 665
 666                 /* and go away */
 667                 proc_exit(1);
 668         }
 669
 670         /* We can now handle ereport(ERROR) */
 671         PG_exception_stack = &local_sigjmp_buf;
 672
 673         /*
 674          * If the background worker request shared memory access, set that up now;
 675          * else, detach all shared memory segments.
 676          */
 677         if (worker->bgw_flags & BGWORKER_SHMEM_ACCESS)
 678         {
 679                 /*
 680                  * Early initialization.  Some of this could be useful even for
 681                  * background workers that aren't using shared memory, but they can
 682                  * call the individual startup routines for those subsystems if
 683                  * needed.
 684                  */
 685                 BaseInit();
 686
 687                 /*
 688                  * Create a per-backend PGPROC struct in shared memory, except in the
 689                  * EXEC_BACKEND case where this was done in SubPostmasterMain. We must
 690                  * do this before we can use LWLocks (and in the EXEC_BACKEND case we
 691                  * already had to do some stuff with LWLocks).
 692                  */
 693 #ifndef EXEC_BACKEND
 694                 InitProcess();
 695 #endif
 696         }
 697
 698         /*
 699          * If bgw_main is set, we use that value as the initial entrypoint.
 700          * However, if the library containing the entrypoint wasn't loaded at
 701          * postmaster startup time, passing it as a direct function pointer is not
 702          * possible.  To work around that, we allow callers for whom a function
 703          * pointer is not available to pass a library name (which will be loaded,
 704          * if necessary) and a function name (which will be looked up in the named
 705          * library).
 706          */
 707         if (worker->bgw_main != NULL)
 708                 entrypt = worker->bgw_main;
 709         else
 710                 entrypt = (bgworker_main_type)
 711                         load_external_function(worker->bgw_library_name,
 712                                                                    worker->bgw_function_name,
 713                                                                    true, NULL);
 714
 715         /*
 716          * Note that in normal processes, we would call InitPostgres here.  For a
 717          * worker, however, we don't know what database to connect to, yet; so we
 718          * need to wait until the user code does it via
 719          * BackgroundWorkerInitializeConnection().
 720          */
 721
 722         /*
 723          * Now invoke the user-defined worker code
 724          */
 725         entrypt(worker->bgw_main_arg);
 726
 727         /* ... and if it returns, we're done */
 728         proc_exit(0);
 729 }
 730
 731 /*
 732  * Register a new background worker while processing shared_preload_libraries.
 733  *
 734  * This can only be called in the _PG_init function of a module library
 735  * that's loaded by shared_preload_libraries; otherwise it has no effect.
 736  */
 737 void
 738 RegisterBackgroundWorker(BackgroundWorker *worker)
 739 {
 740         RegisteredBgWorker *rw;
 741         static int      numworkers = 0;
 742
 743         if (!IsUnderPostmaster)
 744                 ereport(LOG,
 745                  (errmsg("registering background worker \"%s\"", worker->bgw_name)));
 746
 747         if (!process_shared_preload_libraries_in_progress)
 748         {
 749                 if (!IsUnderPostmaster)
 750                         ereport(LOG,
 751                                         (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
 752                                          errmsg("background worker \"%s\": must be registered in shared_preload_libraries",
 753                                                         worker->bgw_name)));
 754                 return;
 755         }
 756
 757         if (!SanityCheckBackgroundWorker(worker, LOG))
 758                 return;
 759
 760         if (worker->bgw_notify_pid != 0)
 761         {
 762                 ereport(LOG,
 763                                 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
 764                                  errmsg("background worker \"%s\": only dynamic background workers can request notification",
 765                                                 worker->bgw_name)));
 766                 return;
 767         }
 768
 769         /*
 770          * Enforce maximum number of workers.  Note this is overly restrictive: we
 771          * could allow more non-shmem-connected workers, because these don't count
 772          * towards the MAX_BACKENDS limit elsewhere.  For now, it doesn't seem
 773          * important to relax this restriction.
 774          */
 775         if (++numworkers > max_worker_processes)
 776         {
 777                 ereport(LOG,
 778                                 (errcode(ERRCODE_CONFIGURATION_LIMIT_EXCEEDED),
 779                                  errmsg("too many background workers"),
 780                                  errdetail_plural("Up to %d background worker can be registered with the current settings.",
 781                                                                   "Up to %d background workers can be registered with the current settings.",
 782                                                                   max_worker_processes,
 783                                                                   max_worker_processes),
 784                                  errhint("Consider increasing the configuration parameter \"max_worker_processes\".")));
 785                 return;
 786         }
 787
 788         /*
 789          * Copy the registration data into the registered workers list.
 790          */
 791         rw = malloc(sizeof(RegisteredBgWorker));
 792         if (rw == NULL)
 793         {
 794                 ereport(LOG,
 795                                 (errcode(ERRCODE_OUT_OF_MEMORY),
 796                                  errmsg("out of memory")));
 797                 return;
 798         }
 799
 800         rw->rw_worker = *worker;
 801         rw->rw_backend = NULL;
 802         rw->rw_pid = 0;
 803         rw->rw_child_slot = 0;
 804         rw->rw_crashed_at = 0;
 805         rw->rw_terminate = false;
 806
 807         slist_push_head(&BackgroundWorkerList, &rw->rw_lnode);
 808 }
 809
 810 /*
 811  * Register a new background worker from a regular backend.
 812  *
 813  * Returns true on success and false on failure.  Failure typically indicates
 814  * that no background worker slots are currently available.
 815  *
 816  * If handle != NULL, we'll set *handle to a pointer that can subsequently
 817  * be used as an argument to GetBackgroundWorkerPid().  The caller can
 818  * free this pointer using pfree(), if desired.
 819  */
 820 bool
 821 RegisterDynamicBackgroundWorker(BackgroundWorker *worker,
 822                                                                 BackgroundWorkerHandle **handle)
 823 {
 824         int                     slotno;
 825         bool            success = false;
 826         uint64          generation = 0;
 827
 828         /*
 829          * We can't register dynamic background workers from the postmaster. If
 830          * this is a standalone backend, we're the only process and can't start
 831          * any more.  In a multi-process environement, it might be theoretically
 832          * possible, but we don't currently support it due to locking
 833          * considerations; see comments on the BackgroundWorkerSlot data
 834          * structure.
 835          */
 836         if (!IsUnderPostmaster)
 837                 return false;
 838
 839         if (!SanityCheckBackgroundWorker(worker, ERROR))
 840                 return false;
 841
 842         LWLockAcquire(BackgroundWorkerLock, LW_EXCLUSIVE);
 843
 844         /*
 845          * Look for an unused slot.  If we find one, grab it.
 846          */
 847         for (slotno = 0; slotno < BackgroundWorkerData->total_slots; ++slotno)
 848         {
 849                 BackgroundWorkerSlot *slot = &BackgroundWorkerData->slot[slotno];
 850
 851                 if (!slot->in_use)
 852                 {
 853                         memcpy(&slot->worker, worker, sizeof(BackgroundWorker));
 854                         slot->pid = InvalidPid;         /* indicates not started yet */
 855                         slot->generation++;
 856                         slot->terminate = false;
 857                         generation = slot->generation;
 858
 859                         /*
 860                          * Make sure postmaster doesn't see the slot as in use before it
 861                          * sees the new contents.
 862                          */
 863                         pg_write_barrier();
 864
 865                         slot->in_use = true;
 866                         success = true;
 867                         break;
 868                 }
 869         }
 870
 871         LWLockRelease(BackgroundWorkerLock);
 872
 873         /* If we found a slot, tell the postmaster to notice the change. */
 874         if (success)
 875                 SendPostmasterSignal(PMSIGNAL_BACKGROUND_WORKER_CHANGE);
 876
 877         /*
 878          * If we found a slot and the user has provided a handle, initialize it.
 879          */
 880         if (success && handle)
 881         {
 882                 *handle = palloc(sizeof(BackgroundWorkerHandle));
 883                 (*handle)->slot = slotno;
 884                 (*handle)->generation = generation;
 885         }
 886
 887         return success;
 888 }
 889
 890 /*
 891  * Get the PID of a dynamically-registered background worker.
 892  *
 893  * If the worker is determined to be running, the return value will be
 894  * BGWH_STARTED and *pidp will get the PID of the worker process.
 895  * Otherwise, the return value will be BGWH_NOT_YET_STARTED if the worker
 896  * hasn't been started yet, and BGWH_STOPPED if the worker was previously
 897  * running but is no longer.
 898  *
 899  * In the latter case, the worker may be stopped temporarily (if it is
 900  * configured for automatic restart and exited non-zero) or gone for
 901  * good (if it exited with code 0 or if it is configured not to restart).
 902  */
 903 BgwHandleStatus
 904 GetBackgroundWorkerPid(BackgroundWorkerHandle *handle, pid_t *pidp)
 905 {
 906         BackgroundWorkerSlot *slot;
 907         pid_t           pid;
 908
 909         Assert(handle->slot < max_worker_processes);
 910         slot = &BackgroundWorkerData->slot[handle->slot];
 911
 912         /*
 913          * We could probably arrange to synchronize access to data using memory
 914          * barriers only, but for now, let's just keep it simple and grab the
 915          * lock.  It seems unlikely that there will be enough traffic here to
 916          * result in meaningful contention.
 917          */
 918         LWLockAcquire(BackgroundWorkerLock, LW_SHARED);
 919
 920         /*
 921          * The generation number can't be concurrently changed while we hold the
 922          * lock.  The pid, which is updated by the postmaster, can change at any
 923          * time, but we assume such changes are atomic.  So the value we read
 924          * won't be garbage, but it might be out of date by the time the caller
 925          * examines it (but that's unavoidable anyway).
 926          */
 927         if (handle->generation != slot->generation)
 928                 pid = 0;
 929         else
 930                 pid = slot->pid;
 931
 932         /* All done. */
 933         LWLockRelease(BackgroundWorkerLock);
 934
 935         if (pid == 0)
 936                 return BGWH_STOPPED;
 937         else if (pid == InvalidPid)
 938                 return BGWH_NOT_YET_STARTED;
 939         *pidp = pid;
 940         return BGWH_STARTED;
 941 }
 942
 943 /*
 944  * Wait for a background worker to start up.
 945  *
 946  * This is like GetBackgroundWorkerPid(), except that if the worker has not
 947  * yet started, we wait for it to do so; thus, BGWH_NOT_YET_STARTED is never
 948  * returned.  However, if the postmaster has died, we give up and return
 949  * BGWH_POSTMASTER_DIED, since it that case we know that startup will not
 950  * take place.
 951  */
 952 BgwHandleStatus
 953 WaitForBackgroundWorkerStartup(BackgroundWorkerHandle *handle, pid_t *pidp)
 954 {
 955         BgwHandleStatus status;
 956         int                     rc;
 957         bool            save_set_latch_on_sigusr1;
 958
 959         save_set_latch_on_sigusr1 = set_latch_on_sigusr1;
 960         set_latch_on_sigusr1 = true;
 961
 962         PG_TRY();
 963         {
 964                 for (;;)
 965                 {
 966                         pid_t           pid;
 967
 968                         CHECK_FOR_INTERRUPTS();
 969
 970                         status = GetBackgroundWorkerPid(handle, &pid);
 971                         if (status == BGWH_STARTED)
 972                                 *pidp = pid;
 973                         if (status != BGWH_NOT_YET_STARTED)
 974                                 break;
 975
 976                         rc = WaitLatch(MyLatch,
 977                                                    WL_LATCH_SET | WL_POSTMASTER_DEATH, 0);
 978
 979                         if (rc & WL_POSTMASTER_DEATH)
 980                         {
 981                                 status = BGWH_POSTMASTER_DIED;
 982                                 break;
 983                         }
 984
 985                         ResetLatch(MyLatch);
 986                 }
 987         }
 988         PG_CATCH();
 989         {
 990                 set_latch_on_sigusr1 = save_set_latch_on_sigusr1;
 991                 PG_RE_THROW();
 992         }
 993         PG_END_TRY();
 994
 995         set_latch_on_sigusr1 = save_set_latch_on_sigusr1;
 996         return status;
 997 }
 998
 999 /*
1000  * Wait for a background worker to stop.
1001  *
1002  * If the worker hasn't yet started, or is running, we wait for it to stop
1003  * and then return BGWH_STOPPED.  However, if the postmaster has died, we give
1004  * up and return BGWH_POSTMASTER_DIED, because it's the postmaster that
1005  * notifies us when a worker's state changes.
1006  */
1007 BgwHandleStatus
1008 WaitForBackgroundWorkerShutdown(BackgroundWorkerHandle *handle)
1009 {
1010         BgwHandleStatus status;
1011         int                     rc;
1012         bool            save_set_latch_on_sigusr1;
1013
1014         save_set_latch_on_sigusr1 = set_latch_on_sigusr1;
1015         set_latch_on_sigusr1 = true;
1016
1017         PG_TRY();
1018         {
1019                 for (;;)
1020                 {
1021                         pid_t           pid;
1022
1023                         CHECK_FOR_INTERRUPTS();
1024
1025                         status = GetBackgroundWorkerPid(handle, &pid);
1026                         if (status == BGWH_STOPPED)
1027                                 return status;
1028
1029                         rc = WaitLatch(&MyProc->procLatch,
1030                                                    WL_LATCH_SET | WL_POSTMASTER_DEATH, 0);
1031
1032                         if (rc & WL_POSTMASTER_DEATH)
1033                                 return BGWH_POSTMASTER_DIED;
1034
1035                         ResetLatch(&MyProc->procLatch);
1036                 }
1037         }
1038         PG_CATCH();
1039         {
1040                 set_latch_on_sigusr1 = save_set_latch_on_sigusr1;
1041                 PG_RE_THROW();
1042         }
1043         PG_END_TRY();
1044
1045         set_latch_on_sigusr1 = save_set_latch_on_sigusr1;
1046         return status;
1047 }
1048
1049 /*
1050  * Instruct the postmaster to terminate a background worker.
1051  *
1052  * Note that it's safe to do this without regard to whether the worker is
1053  * still running, or even if the worker may already have existed and been
1054  * unregistered.
1055  */
1056 void
1057 TerminateBackgroundWorker(BackgroundWorkerHandle *handle)
1058 {
1059         BackgroundWorkerSlot *slot;
1060         bool            signal_postmaster = false;
1061
1062         Assert(handle->slot < max_worker_processes);
1063         slot = &BackgroundWorkerData->slot[handle->slot];
1064
1065         /* Set terminate flag in shared memory, unless slot has been reused. */
1066         LWLockAcquire(BackgroundWorkerLock, LW_EXCLUSIVE);
1067         if (handle->generation == slot->generation)
1068         {
1069                 slot->terminate = true;
1070                 signal_postmaster = true;
1071         }
1072         LWLockRelease(BackgroundWorkerLock);
1073
1074         /* Make sure the postmaster notices the change to shared memory. */
1075         if (signal_postmaster)
1076                 SendPostmasterSignal(PMSIGNAL_BACKGROUND_WORKER_CHANGE);
1077 }