PostgreSQL Source Code  git master
parallel.c
Go to the documentation of this file.
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/index.h"
23 #include "catalog/namespace.h"
24 #include "catalog/pg_enum.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  * We don't want to waste a lot of memory on an error queue which, most of
48  * the time, will process only a handful of small messages. However, it is
49  * desirable to make it large enough that a typical ErrorResponse can be sent
50  * without blocking. That way, a worker that errors out can write the whole
51  * message into the queue and terminate without waiting for the user backend.
52  */
53 #define PARALLEL_ERROR_QUEUE_SIZE 16384
54 
55 /* Magic number for parallel context TOC. */
56 #define PARALLEL_MAGIC 0x50477c7c
57 
58 /*
59  * Magic numbers for per-context parallel state sharing. Higher-level code
60  * should use smaller values, leaving these very large ones for use by this
61  * module.
62  */
63 #define PARALLEL_KEY_FIXED UINT64CONST(0xFFFFFFFFFFFF0001)
64 #define PARALLEL_KEY_ERROR_QUEUE UINT64CONST(0xFFFFFFFFFFFF0002)
65 #define PARALLEL_KEY_LIBRARY UINT64CONST(0xFFFFFFFFFFFF0003)
66 #define PARALLEL_KEY_GUC UINT64CONST(0xFFFFFFFFFFFF0004)
67 #define PARALLEL_KEY_COMBO_CID UINT64CONST(0xFFFFFFFFFFFF0005)
68 #define PARALLEL_KEY_TRANSACTION_SNAPSHOT UINT64CONST(0xFFFFFFFFFFFF0006)
69 #define PARALLEL_KEY_ACTIVE_SNAPSHOT UINT64CONST(0xFFFFFFFFFFFF0007)
70 #define PARALLEL_KEY_TRANSACTION_STATE UINT64CONST(0xFFFFFFFFFFFF0008)
71 #define PARALLEL_KEY_ENTRYPOINT UINT64CONST(0xFFFFFFFFFFFF0009)
72 #define PARALLEL_KEY_SESSION_DSM UINT64CONST(0xFFFFFFFFFFFF000A)
73 #define PARALLEL_KEY_REINDEX_STATE UINT64CONST(0xFFFFFFFFFFFF000B)
74 #define PARALLEL_KEY_RELMAPPER_STATE UINT64CONST(0xFFFFFFFFFFFF000C)
75 #define PARALLEL_KEY_ENUMBLACKLIST UINT64CONST(0xFFFFFFFFFFFF000D)
76 
77 /* Fixed-size parallel state. */
78 typedef struct FixedParallelState
79 {
80  /* Fixed-size state that workers must restore. */
95 
96  /* Mutex protects remaining fields. */
98 
99  /* Maximum XactLastRecEnd of any worker. */
102 
103 /*
104  * Our parallel worker number. We initialize this to -1, meaning that we are
105  * not a parallel worker. In parallel workers, it will be set to a value >= 0
106  * and < the number of workers before any user code is invoked; each parallel
107  * worker will get a different parallel worker number.
108  */
110 
111 /* Is there a parallel message pending which we need to receive? */
112 volatile bool ParallelMessagePending = false;
113 
114 /* Are we initializing a parallel worker? */
116 
117 /* Pointer to our fixed parallel state. */
119 
120 /* List of active parallel contexts. */
122 
123 /* Backend-local copy of data from FixedParallelState. */
124 static pid_t ParallelMasterPid;
125 
126 /*
127  * List of internal parallel worker entry points. We need this for
128  * reasons explained in LookupParallelWorkerFunction(), below.
129  */
130 static const struct
131 {
132  const char *fn_name;
135 
136 {
137  {
138  "ParallelQueryMain", ParallelQueryMain
139  },
140  {
141  "_bt_parallel_build_main", _bt_parallel_build_main
142  }
143 };
144 
145 /* Private functions. */
146 static void HandleParallelMessage(ParallelContext *pcxt, int i, StringInfo msg);
148 static parallel_worker_main_type LookupParallelWorkerFunction(const char *libraryname, const char *funcname);
149 static void ParallelWorkerShutdown(int code, Datum arg);
150 
151 
152 /*
153  * Establish a new parallel context. This should be done after entering
154  * parallel mode, and (unless there is an error) the context should be
155  * destroyed before exiting the current subtransaction.
156  */
158 CreateParallelContext(const char *library_name, const char *function_name,
159  int nworkers)
160 {
161  MemoryContext oldcontext;
162  ParallelContext *pcxt;
163 
164  /* It is unsafe to create a parallel context if not in parallel mode. */
166 
167  /* Number of workers should be non-negative. */
168  Assert(nworkers >= 0);
169 
170  /* We might be running in a short-lived memory context. */
172 
173  /* Initialize a new ParallelContext. */
174  pcxt = palloc0(sizeof(ParallelContext));
176  pcxt->nworkers = nworkers;
177  pcxt->library_name = pstrdup(library_name);
178  pcxt->function_name = pstrdup(function_name);
181  dlist_push_head(&pcxt_list, &pcxt->node);
182 
183  /* Restore previous memory context. */
184  MemoryContextSwitchTo(oldcontext);
185 
186  return pcxt;
187 }
188 
189 /*
190  * Establish the dynamic shared memory segment for a parallel context and
191  * copy state and other bookkeeping information that will be needed by
192  * parallel workers into it.
193  */
194 void
196 {
197  MemoryContext oldcontext;
198  Size library_len = 0;
199  Size guc_len = 0;
200  Size combocidlen = 0;
201  Size tsnaplen = 0;
202  Size asnaplen = 0;
203  Size tstatelen = 0;
204  Size reindexlen = 0;
205  Size relmapperlen = 0;
206  Size enumblacklistlen = 0;
207  Size segsize = 0;
208  int i;
209  FixedParallelState *fps;
210  dsm_handle session_dsm_handle = DSM_HANDLE_INVALID;
211  Snapshot transaction_snapshot = GetTransactionSnapshot();
212  Snapshot active_snapshot = GetActiveSnapshot();
213 
214  /* We might be running in a very short-lived memory context. */
216 
217  /* Allow space to store the fixed-size parallel state. */
219  shm_toc_estimate_keys(&pcxt->estimator, 1);
220 
221  /*
222  * Normally, the user will have requested at least one worker process, but
223  * if by chance they have not, we can skip a bunch of things here.
224  */
225  if (pcxt->nworkers > 0)
226  {
227  /* Get (or create) the per-session DSM segment's handle. */
228  session_dsm_handle = GetSessionDsmHandle();
229 
230  /*
231  * If we weren't able to create a per-session DSM segment, then we can
232  * continue but we can't safely launch any workers because their
233  * record typmods would be incompatible so they couldn't exchange
234  * tuples.
235  */
236  if (session_dsm_handle == DSM_HANDLE_INVALID)
237  pcxt->nworkers = 0;
238  }
239 
240  if (pcxt->nworkers > 0)
241  {
242  /* Estimate space for various kinds of state sharing. */
243  library_len = EstimateLibraryStateSpace();
244  shm_toc_estimate_chunk(&pcxt->estimator, library_len);
245  guc_len = EstimateGUCStateSpace();
246  shm_toc_estimate_chunk(&pcxt->estimator, guc_len);
247  combocidlen = EstimateComboCIDStateSpace();
248  shm_toc_estimate_chunk(&pcxt->estimator, combocidlen);
249  tsnaplen = EstimateSnapshotSpace(transaction_snapshot);
250  shm_toc_estimate_chunk(&pcxt->estimator, tsnaplen);
251  asnaplen = EstimateSnapshotSpace(active_snapshot);
252  shm_toc_estimate_chunk(&pcxt->estimator, asnaplen);
253  tstatelen = EstimateTransactionStateSpace();
254  shm_toc_estimate_chunk(&pcxt->estimator, tstatelen);
256  reindexlen = EstimateReindexStateSpace();
257  shm_toc_estimate_chunk(&pcxt->estimator, reindexlen);
258  relmapperlen = EstimateRelationMapSpace();
259  shm_toc_estimate_chunk(&pcxt->estimator, relmapperlen);
260  enumblacklistlen = EstimateEnumBlacklistSpace();
261  shm_toc_estimate_chunk(&pcxt->estimator, enumblacklistlen);
262  /* If you add more chunks here, you probably need to add keys. */
263  shm_toc_estimate_keys(&pcxt->estimator, 10);
264 
265  /* Estimate space need for error queues. */
268  "parallel error queue size not buffer-aligned");
271  pcxt->nworkers));
272  shm_toc_estimate_keys(&pcxt->estimator, 1);
273 
274  /* Estimate how much we'll need for the entrypoint info. */
275  shm_toc_estimate_chunk(&pcxt->estimator, strlen(pcxt->library_name) +
276  strlen(pcxt->function_name) + 2);
277  shm_toc_estimate_keys(&pcxt->estimator, 1);
278  }
279 
280  /*
281  * Create DSM and initialize with new table of contents. But if the user
282  * didn't request any workers, then don't bother creating a dynamic shared
283  * memory segment; instead, just use backend-private memory.
284  *
285  * Also, if we can't create a dynamic shared memory segment because the
286  * maximum number of segments have already been created, then fall back to
287  * backend-private memory, and plan not to use any workers. We hope this
288  * won't happen very often, but it's better to abandon the use of
289  * parallelism than to fail outright.
290  */
291  segsize = shm_toc_estimate(&pcxt->estimator);
292  if (pcxt->nworkers > 0)
294  if (pcxt->seg != NULL)
296  dsm_segment_address(pcxt->seg),
297  segsize);
298  else
299  {
300  pcxt->nworkers = 0;
303  segsize);
304  }
305 
306  /* Initialize fixed-size state in shared memory. */
307  fps = (FixedParallelState *)
308  shm_toc_allocate(pcxt->toc, sizeof(FixedParallelState));
309  fps->database_id = MyDatabaseId;
322  SpinLockInit(&fps->mutex);
323  fps->last_xlog_end = 0;
325 
326  /* We can skip the rest of this if we're not budgeting for any workers. */
327  if (pcxt->nworkers > 0)
328  {
329  char *libraryspace;
330  char *gucspace;
331  char *combocidspace;
332  char *tsnapspace;
333  char *asnapspace;
334  char *tstatespace;
335  char *reindexspace;
336  char *relmapperspace;
337  char *error_queue_space;
338  char *session_dsm_handle_space;
339  char *entrypointstate;
340  char *enumblacklistspace;
341  Size lnamelen;
342 
343  /* Serialize shared libraries we have loaded. */
344  libraryspace = shm_toc_allocate(pcxt->toc, library_len);
345  SerializeLibraryState(library_len, libraryspace);
346  shm_toc_insert(pcxt->toc, PARALLEL_KEY_LIBRARY, libraryspace);
347 
348  /* Serialize GUC settings. */
349  gucspace = shm_toc_allocate(pcxt->toc, guc_len);
350  SerializeGUCState(guc_len, gucspace);
351  shm_toc_insert(pcxt->toc, PARALLEL_KEY_GUC, gucspace);
352 
353  /* Serialize combo CID state. */
354  combocidspace = shm_toc_allocate(pcxt->toc, combocidlen);
355  SerializeComboCIDState(combocidlen, combocidspace);
356  shm_toc_insert(pcxt->toc, PARALLEL_KEY_COMBO_CID, combocidspace);
357 
358  /* Serialize transaction snapshot and active snapshot. */
359  tsnapspace = shm_toc_allocate(pcxt->toc, tsnaplen);
360  SerializeSnapshot(transaction_snapshot, tsnapspace);
362  tsnapspace);
363  asnapspace = shm_toc_allocate(pcxt->toc, asnaplen);
364  SerializeSnapshot(active_snapshot, asnapspace);
365  shm_toc_insert(pcxt->toc, PARALLEL_KEY_ACTIVE_SNAPSHOT, asnapspace);
366 
367  /* Provide the handle for per-session segment. */
368  session_dsm_handle_space = shm_toc_allocate(pcxt->toc,
369  sizeof(dsm_handle));
370  *(dsm_handle *) session_dsm_handle_space = session_dsm_handle;
372  session_dsm_handle_space);
373 
374  /* Serialize transaction state. */
375  tstatespace = shm_toc_allocate(pcxt->toc, tstatelen);
376  SerializeTransactionState(tstatelen, tstatespace);
378 
379  /* Serialize reindex state. */
380  reindexspace = shm_toc_allocate(pcxt->toc, reindexlen);
381  SerializeReindexState(reindexlen, reindexspace);
382  shm_toc_insert(pcxt->toc, PARALLEL_KEY_REINDEX_STATE, reindexspace);
383 
384  /* Serialize relmapper state. */
385  relmapperspace = shm_toc_allocate(pcxt->toc, relmapperlen);
386  SerializeRelationMap(relmapperlen, relmapperspace);
388  relmapperspace);
389 
390  /* Serialize enum blacklist state. */
391  enumblacklistspace = shm_toc_allocate(pcxt->toc, enumblacklistlen);
392  SerializeEnumBlacklist(enumblacklistspace, enumblacklistlen);
394  enumblacklistspace);
395 
396  /* Allocate space for worker information. */
397  pcxt->worker = palloc0(sizeof(ParallelWorkerInfo) * pcxt->nworkers);
398 
399  /*
400  * Establish error queues in dynamic shared memory.
401  *
402  * These queues should be used only for transmitting ErrorResponse,
403  * NoticeResponse, and NotifyResponse protocol messages. Tuple data
404  * should be transmitted via separate (possibly larger?) queues.
405  */
406  error_queue_space =
407  shm_toc_allocate(pcxt->toc,
409  pcxt->nworkers));
410  for (i = 0; i < pcxt->nworkers; ++i)
411  {
412  char *start;
413  shm_mq *mq;
414 
415  start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
416  mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
418  pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
419  }
420  shm_toc_insert(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE, error_queue_space);
421 
422  /*
423  * Serialize entrypoint information. It's unsafe to pass function
424  * pointers across processes, as the function pointer may be different
425  * in each process in EXEC_BACKEND builds, so we always pass library
426  * and function name. (We use library name "postgres" for functions
427  * in the core backend.)
428  */
429  lnamelen = strlen(pcxt->library_name);
430  entrypointstate = shm_toc_allocate(pcxt->toc, lnamelen +
431  strlen(pcxt->function_name) + 2);
432  strcpy(entrypointstate, pcxt->library_name);
433  strcpy(entrypointstate + lnamelen + 1, pcxt->function_name);
434  shm_toc_insert(pcxt->toc, PARALLEL_KEY_ENTRYPOINT, entrypointstate);
435  }
436 
437  /* Restore previous memory context. */
438  MemoryContextSwitchTo(oldcontext);
439 }
440 
441 /*
442  * Reinitialize the dynamic shared memory segment for a parallel context such
443  * that we could launch workers for it again.
444  */
445 void
447 {
448  FixedParallelState *fps;
449 
450  /* Wait for any old workers to exit. */
451  if (pcxt->nworkers_launched > 0)
452  {
455  pcxt->nworkers_launched = 0;
456  if (pcxt->known_attached_workers)
457  {
459  pcxt->known_attached_workers = NULL;
460  pcxt->nknown_attached_workers = 0;
461  }
462  }
463 
464  /* Reset a few bits of fixed parallel state to a clean state. */
465  fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED, false);
466  fps->last_xlog_end = 0;
467 
468  /* Recreate error queues (if they exist). */
469  if (pcxt->nworkers > 0)
470  {
471  char *error_queue_space;
472  int i;
473 
474  error_queue_space =
476  for (i = 0; i < pcxt->nworkers; ++i)
477  {
478  char *start;
479  shm_mq *mq;
480 
481  start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
482  mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
484  pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
485  }
486  }
487 }
488 
489 /*
490  * Launch parallel workers.
491  */
492 void
494 {
495  MemoryContext oldcontext;
496  BackgroundWorker worker;
497  int i;
498  bool any_registrations_failed = false;
499 
500  /* Skip this if we have no workers. */
501  if (pcxt->nworkers == 0)
502  return;
503 
504  /* We need to be a lock group leader. */
506 
507  /* If we do have workers, we'd better have a DSM segment. */
508  Assert(pcxt->seg != NULL);
509 
510  /* We might be running in a short-lived memory context. */
512 
513  /* Configure a worker. */
514  memset(&worker, 0, sizeof(worker));
515  snprintf(worker.bgw_name, BGW_MAXLEN, "parallel worker for PID %d",
516  MyProcPid);
517  snprintf(worker.bgw_type, BGW_MAXLEN, "parallel worker");
518  worker.bgw_flags =
523  sprintf(worker.bgw_library_name, "postgres");
524  sprintf(worker.bgw_function_name, "ParallelWorkerMain");
526  worker.bgw_notify_pid = MyProcPid;
527 
528  /*
529  * Start workers.
530  *
531  * The caller must be able to tolerate ending up with fewer workers than
532  * expected, so there is no need to throw an error here if registration
533  * fails. It wouldn't help much anyway, because registering the worker in
534  * no way guarantees that it will start up and initialize successfully.
535  */
536  for (i = 0; i < pcxt->nworkers; ++i)
537  {
538  memcpy(worker.bgw_extra, &i, sizeof(int));
539  if (!any_registrations_failed &&
541  &pcxt->worker[i].bgwhandle))
542  {
544  pcxt->worker[i].bgwhandle);
545  pcxt->nworkers_launched++;
546  }
547  else
548  {
549  /*
550  * If we weren't able to register the worker, then we've bumped up
551  * against the max_worker_processes limit, and future
552  * registrations will probably fail too, so arrange to skip them.
553  * But we still have to execute this code for the remaining slots
554  * to make sure that we forget about the error queues we budgeted
555  * for those workers. Otherwise, we'll wait for them to start,
556  * but they never will.
557  */
558  any_registrations_failed = true;
559  pcxt->worker[i].bgwhandle = NULL;
560  shm_mq_detach(pcxt->worker[i].error_mqh);
561  pcxt->worker[i].error_mqh = NULL;
562  }
563  }
564 
565  /*
566  * Now that nworkers_launched has taken its final value, we can initialize
567  * known_attached_workers.
568  */
569  if (pcxt->nworkers_launched > 0)
570  {
571  pcxt->known_attached_workers =
572  palloc0(sizeof(bool) * pcxt->nworkers_launched);
573  pcxt->nknown_attached_workers = 0;
574  }
575 
576  /* Restore previous memory context. */
577  MemoryContextSwitchTo(oldcontext);
578 }
579 
580 /*
581  * Wait for all workers to attach to their error queues, and throw an error if
582  * any worker fails to do this.
583  *
584  * Callers can assume that if this function returns successfully, then the
585  * number of workers given by pcxt->nworkers_launched have initialized and
586  * attached to their error queues. Whether or not these workers are guaranteed
587  * to still be running depends on what code the caller asked them to run;
588  * this function does not guarantee that they have not exited. However, it
589  * does guarantee that any workers which exited must have done so cleanly and
590  * after successfully performing the work with which they were tasked.
591  *
592  * If this function is not called, then some of the workers that were launched
593  * may not have been started due to a fork() failure, or may have exited during
594  * early startup prior to attaching to the error queue, so nworkers_launched
595  * cannot be viewed as completely reliable. It will never be less than the
596  * number of workers which actually started, but it might be more. Any workers
597  * that failed to start will still be discovered by
598  * WaitForParallelWorkersToFinish and an error will be thrown at that time,
599  * provided that function is eventually reached.
600  *
601  * In general, the leader process should do as much work as possible before
602  * calling this function. fork() failures and other early-startup failures
603  * are very uncommon, and having the leader sit idle when it could be doing
604  * useful work is undesirable. However, if the leader needs to wait for
605  * all of its workers or for a specific worker, it may want to call this
606  * function before doing so. If not, it must make some other provision for
607  * the failure-to-start case, lest it wait forever. On the other hand, a
608  * leader which never waits for a worker that might not be started yet, or
609  * at least never does so prior to WaitForParallelWorkersToFinish(), need not
610  * call this function at all.
611  */
612 void
614 {
615  int i;
616 
617  /* Skip this if we have no launched workers. */
618  if (pcxt->nworkers_launched == 0)
619  return;
620 
621  for (;;)
622  {
623  /*
624  * This will process any parallel messages that are pending and it may
625  * also throw an error propagated from a worker.
626  */
628 
629  for (i = 0; i < pcxt->nworkers_launched; ++i)
630  {
632  shm_mq *mq;
633  int rc;
634  pid_t pid;
635 
636  if (pcxt->known_attached_workers[i])
637  continue;
638 
639  /*
640  * If error_mqh is NULL, then the worker has already exited
641  * cleanly.
642  */
643  if (pcxt->worker[i].error_mqh == NULL)
644  {
645  pcxt->known_attached_workers[i] = true;
646  ++pcxt->nknown_attached_workers;
647  continue;
648  }
649 
650  status = GetBackgroundWorkerPid(pcxt->worker[i].bgwhandle, &pid);
651  if (status == BGWH_STARTED)
652  {
653  /* Has the worker attached to the error queue? */
654  mq = shm_mq_get_queue(pcxt->worker[i].error_mqh);
655  if (shm_mq_get_sender(mq) != NULL)
656  {
657  /* Yes, so it is known to be attached. */
658  pcxt->known_attached_workers[i] = true;
659  ++pcxt->nknown_attached_workers;
660  }
661  }
662  else if (status == BGWH_STOPPED)
663  {
664  /*
665  * If the worker stopped without attaching to the error queue,
666  * throw an error.
667  */
668  mq = shm_mq_get_queue(pcxt->worker[i].error_mqh);
669  if (shm_mq_get_sender(mq) == NULL)
670  ereport(ERROR,
671  (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
672  errmsg("parallel worker failed to initialize"),
673  errhint("More details may be available in the server log.")));
674 
675  pcxt->known_attached_workers[i] = true;
676  ++pcxt->nknown_attached_workers;
677  }
678  else
679  {
680  /*
681  * Worker not yet started, so we must wait. The postmaster
682  * will notify us if the worker's state changes. Our latch
683  * might also get set for some other reason, but if so we'll
684  * just end up waiting for the same worker again.
685  */
686  rc = WaitLatch(MyLatch,
689 
690  if (rc & WL_LATCH_SET)
692  }
693  }
694 
695  /* If all workers are known to have started, we're done. */
696  if (pcxt->nknown_attached_workers >= pcxt->nworkers_launched)
697  {
699  break;
700  }
701  }
702 }
703 
704 /*
705  * Wait for all workers to finish computing.
706  *
707  * Even if the parallel operation seems to have completed successfully, it's
708  * important to call this function afterwards. We must not miss any errors
709  * the workers may have thrown during the parallel operation, or any that they
710  * may yet throw while shutting down.
711  *
712  * Also, we want to update our notion of XactLastRecEnd based on worker
713  * feedback.
714  */
715 void
717 {
718  for (;;)
719  {
720  bool anyone_alive = false;
721  int nfinished = 0;
722  int i;
723 
724  /*
725  * This will process any parallel messages that are pending, which may
726  * change the outcome of the loop that follows. It may also throw an
727  * error propagated from a worker.
728  */
730 
731  for (i = 0; i < pcxt->nworkers_launched; ++i)
732  {
733  /*
734  * If error_mqh is NULL, then the worker has already exited
735  * cleanly. If we have received a message through error_mqh from
736  * the worker, we know it started up cleanly, and therefore we're
737  * certain to be notified when it exits.
738  */
739  if (pcxt->worker[i].error_mqh == NULL)
740  ++nfinished;
741  else if (pcxt->known_attached_workers[i])
742  {
743  anyone_alive = true;
744  break;
745  }
746  }
747 
748  if (!anyone_alive)
749  {
750  /* If all workers are known to have finished, we're done. */
751  if (nfinished >= pcxt->nworkers_launched)
752  {
753  Assert(nfinished == pcxt->nworkers_launched);
754  break;
755  }
756 
757  /*
758  * We didn't detect any living workers, but not all workers are
759  * known to have exited cleanly. Either not all workers have
760  * launched yet, or maybe some of them failed to start or
761  * terminated abnormally.
762  */
763  for (i = 0; i < pcxt->nworkers_launched; ++i)
764  {
765  pid_t pid;
766  shm_mq *mq;
767 
768  /*
769  * If the worker is BGWH_NOT_YET_STARTED or BGWH_STARTED, we
770  * should just keep waiting. If it is BGWH_STOPPED, then
771  * further investigation is needed.
772  */
773  if (pcxt->worker[i].error_mqh == NULL ||
774  pcxt->worker[i].bgwhandle == NULL ||
776  &pid) != BGWH_STOPPED)
777  continue;
778 
779  /*
780  * Check whether the worker ended up stopped without ever
781  * attaching to the error queue. If so, the postmaster was
782  * unable to fork the worker or it exited without initializing
783  * properly. We must throw an error, since the caller may
784  * have been expecting the worker to do some work before
785  * exiting.
786  */
787  mq = shm_mq_get_queue(pcxt->worker[i].error_mqh);
788  if (shm_mq_get_sender(mq) == NULL)
789  ereport(ERROR,
790  (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
791  errmsg("parallel worker failed to initialize"),
792  errhint("More details may be available in the server log.")));
793 
794  /*
795  * The worker is stopped, but is attached to the error queue.
796  * Unless there's a bug somewhere, this will only happen when
797  * the worker writes messages and terminates after the
798  * CHECK_FOR_INTERRUPTS() near the top of this function and
799  * before the call to GetBackgroundWorkerPid(). In that case,
800  * or latch should have been set as well and the right things
801  * will happen on the next pass through the loop.
802  */
803  }
804  }
805 
809  }
810 
811  if (pcxt->toc != NULL)
812  {
813  FixedParallelState *fps;
814 
815  fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED, false);
816  if (fps->last_xlog_end > XactLastRecEnd)
818  }
819 }
820 
821 /*
822  * Wait for all workers to exit.
823  *
824  * This function ensures that workers have been completely shutdown. The
825  * difference between WaitForParallelWorkersToFinish and this function is
826  * that former just ensures that last message sent by worker backend is
827  * received by master backend whereas this ensures the complete shutdown.
828  */
829 static void
831 {
832  int i;
833 
834  /* Wait until the workers actually die. */
835  for (i = 0; i < pcxt->nworkers_launched; ++i)
836  {
838 
839  if (pcxt->worker == NULL || pcxt->worker[i].bgwhandle == NULL)
840  continue;
841 
843 
844  /*
845  * If the postmaster kicked the bucket, we have no chance of cleaning
846  * up safely -- we won't be able to tell when our workers are actually
847  * dead. This doesn't necessitate a PANIC since they will all abort
848  * eventually, but we can't safely continue this session.
849  */
850  if (status == BGWH_POSTMASTER_DIED)
851  ereport(FATAL,
852  (errcode(ERRCODE_ADMIN_SHUTDOWN),
853  errmsg("postmaster exited during a parallel transaction")));
854 
855  /* Release memory. */
856  pfree(pcxt->worker[i].bgwhandle);
857  pcxt->worker[i].bgwhandle = NULL;
858  }
859 }
860 
861 /*
862  * Destroy a parallel context.
863  *
864  * If expecting a clean exit, you should use WaitForParallelWorkersToFinish()
865  * first, before calling this function. When this function is invoked, any
866  * remaining workers are forcibly killed; the dynamic shared memory segment
867  * is unmapped; and we then wait (uninterruptibly) for the workers to exit.
868  */
869 void
871 {
872  int i;
873 
874  /*
875  * Be careful about order of operations here! We remove the parallel
876  * context from the list before we do anything else; otherwise, if an
877  * error occurs during a subsequent step, we might try to nuke it again
878  * from AtEOXact_Parallel or AtEOSubXact_Parallel.
879  */
880  dlist_delete(&pcxt->node);
881 
882  /* Kill each worker in turn, and forget their error queues. */
883  if (pcxt->worker != NULL)
884  {
885  for (i = 0; i < pcxt->nworkers_launched; ++i)
886  {
887  if (pcxt->worker[i].error_mqh != NULL)
888  {
890 
891  shm_mq_detach(pcxt->worker[i].error_mqh);
892  pcxt->worker[i].error_mqh = NULL;
893  }
894  }
895  }
896 
897  /*
898  * If we have allocated a shared memory segment, detach it. This will
899  * implicitly detach the error queues, and any other shared memory queues,
900  * stored there.
901  */
902  if (pcxt->seg != NULL)
903  {
904  dsm_detach(pcxt->seg);
905  pcxt->seg = NULL;
906  }
907 
908  /*
909  * If this parallel context is actually in backend-private memory rather
910  * than shared memory, free that memory instead.
911  */
912  if (pcxt->private_memory != NULL)
913  {
914  pfree(pcxt->private_memory);
915  pcxt->private_memory = NULL;
916  }
917 
918  /*
919  * We can't finish transaction commit or abort until all of the workers
920  * have exited. This means, in particular, that we can't respond to
921  * interrupts at this stage.
922  */
923  HOLD_INTERRUPTS();
926 
927  /* Free the worker array itself. */
928  if (pcxt->worker != NULL)
929  {
930  pfree(pcxt->worker);
931  pcxt->worker = NULL;
932  }
933 
934  /* Free memory. */
935  pfree(pcxt->library_name);
936  pfree(pcxt->function_name);
937  pfree(pcxt);
938 }
939 
940 /*
941  * Are there any parallel contexts currently active?
942  */
943 bool
945 {
946  return !dlist_is_empty(&pcxt_list);
947 }
948 
949 /*
950  * Handle receipt of an interrupt indicating a parallel worker message.
951  *
952  * Note: this is called within a signal handler! All we can do is set
953  * a flag that will cause the next CHECK_FOR_INTERRUPTS() to invoke
954  * HandleParallelMessages().
955  */
956 void
958 {
959  InterruptPending = true;
960  ParallelMessagePending = true;
961  SetLatch(MyLatch);
962 }
963 
964 /*
965  * Handle any queued protocol messages received from parallel workers.
966  */
967 void
969 {
970  dlist_iter iter;
971  MemoryContext oldcontext;
972 
973  static MemoryContext hpm_context = NULL;
974 
975  /*
976  * This is invoked from ProcessInterrupts(), and since some of the
977  * functions it calls contain CHECK_FOR_INTERRUPTS(), there is a potential
978  * for recursive calls if more signals are received while this runs. It's
979  * unclear that recursive entry would be safe, and it doesn't seem useful
980  * even if it is safe, so let's block interrupts until done.
981  */
982  HOLD_INTERRUPTS();
983 
984  /*
985  * Moreover, CurrentMemoryContext might be pointing almost anywhere. We
986  * don't want to risk leaking data into long-lived contexts, so let's do
987  * our work here in a private context that we can reset on each use.
988  */
989  if (hpm_context == NULL) /* first time through? */
991  "HandleParallelMessages",
993  else
994  MemoryContextReset(hpm_context);
995 
996  oldcontext = MemoryContextSwitchTo(hpm_context);
997 
998  /* OK to process messages. Reset the flag saying there are more to do. */
999  ParallelMessagePending = false;
1000 
1001  dlist_foreach(iter, &pcxt_list)
1002  {
1003  ParallelContext *pcxt;
1004  int i;
1005 
1006  pcxt = dlist_container(ParallelContext, node, iter.cur);
1007  if (pcxt->worker == NULL)
1008  continue;
1009 
1010  for (i = 0; i < pcxt->nworkers_launched; ++i)
1011  {
1012  /*
1013  * Read as many messages as we can from each worker, but stop when
1014  * either (1) the worker's error queue goes away, which can happen
1015  * if we receive a Terminate message from the worker; or (2) no
1016  * more messages can be read from the worker without blocking.
1017  */
1018  while (pcxt->worker[i].error_mqh != NULL)
1019  {
1020  shm_mq_result res;
1021  Size nbytes;
1022  void *data;
1023 
1024  res = shm_mq_receive(pcxt->worker[i].error_mqh, &nbytes,
1025  &data, true);
1026  if (res == SHM_MQ_WOULD_BLOCK)
1027  break;
1028  else if (res == SHM_MQ_SUCCESS)
1029  {
1030  StringInfoData msg;
1031 
1032  initStringInfo(&msg);
1033  appendBinaryStringInfo(&msg, data, nbytes);
1034  HandleParallelMessage(pcxt, i, &msg);
1035  pfree(msg.data);
1036  }
1037  else
1038  ereport(ERROR,
1039  (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1040  errmsg("lost connection to parallel worker")));
1041  }
1042  }
1043  }
1044 
1045  MemoryContextSwitchTo(oldcontext);
1046 
1047  /* Might as well clear the context on our way out */
1048  MemoryContextReset(hpm_context);
1049 
1051 }
1052 
1053 /*
1054  * Handle a single protocol message received from a single parallel worker.
1055  */
1056 static void
1058 {
1059  char msgtype;
1060 
1061  if (pcxt->known_attached_workers != NULL &&
1062  !pcxt->known_attached_workers[i])
1063  {
1064  pcxt->known_attached_workers[i] = true;
1065  pcxt->nknown_attached_workers++;
1066  }
1067 
1068  msgtype = pq_getmsgbyte(msg);
1069 
1070  switch (msgtype)
1071  {
1072  case 'K': /* BackendKeyData */
1073  {
1074  int32 pid = pq_getmsgint(msg, 4);
1075 
1076  (void) pq_getmsgint(msg, 4); /* discard cancel key */
1077  (void) pq_getmsgend(msg);
1078  pcxt->worker[i].pid = pid;
1079  break;
1080  }
1081 
1082  case 'E': /* ErrorResponse */
1083  case 'N': /* NoticeResponse */
1084  {
1085  ErrorData edata;
1086  ErrorContextCallback *save_error_context_stack;
1087 
1088  /* Parse ErrorResponse or NoticeResponse. */
1089  pq_parse_errornotice(msg, &edata);
1090 
1091  /* Death of a worker isn't enough justification for suicide. */
1092  edata.elevel = Min(edata.elevel, ERROR);
1093 
1094  /*
1095  * If desired, add a context line to show that this is a
1096  * message propagated from a parallel worker. Otherwise, it
1097  * can sometimes be confusing to understand what actually
1098  * happened. (We don't do this in FORCE_PARALLEL_REGRESS mode
1099  * because it causes test-result instability depending on
1100  * whether a parallel worker is actually used or not.)
1101  */
1103  {
1104  if (edata.context)
1105  edata.context = psprintf("%s\n%s", edata.context,
1106  _("parallel worker"));
1107  else
1108  edata.context = pstrdup(_("parallel worker"));
1109  }
1110 
1111  /*
1112  * Context beyond that should use the error context callbacks
1113  * that were in effect when the ParallelContext was created,
1114  * not the current ones.
1115  */
1116  save_error_context_stack = error_context_stack;
1118 
1119  /* Rethrow error or print notice. */
1120  ThrowErrorData(&edata);
1121 
1122  /* Not an error, so restore previous context stack. */
1123  error_context_stack = save_error_context_stack;
1124 
1125  break;
1126  }
1127 
1128  case 'A': /* NotifyResponse */
1129  {
1130  /* Propagate NotifyResponse. */
1131  int32 pid;
1132  const char *channel;
1133  const char *payload;
1134 
1135  pid = pq_getmsgint(msg, 4);
1136  channel = pq_getmsgrawstring(msg);
1137  payload = pq_getmsgrawstring(msg);
1138  pq_endmessage(msg);
1139 
1140  NotifyMyFrontEnd(channel, payload, pid);
1141 
1142  break;
1143  }
1144 
1145  case 'X': /* Terminate, indicating clean exit */
1146  {
1147  shm_mq_detach(pcxt->worker[i].error_mqh);
1148  pcxt->worker[i].error_mqh = NULL;
1149  break;
1150  }
1151 
1152  default:
1153  {
1154  elog(ERROR, "unrecognized message type received from parallel worker: %c (message length %d bytes)",
1155  msgtype, msg->len);
1156  }
1157  }
1158 }
1159 
1160 /*
1161  * End-of-subtransaction cleanup for parallel contexts.
1162  *
1163  * Currently, it's forbidden to enter or leave a subtransaction while
1164  * parallel mode is in effect, so we could just blow away everything. But
1165  * we may want to relax that restriction in the future, so this code
1166  * contemplates that there may be multiple subtransaction IDs in pcxt_list.
1167  */
1168 void
1170 {
1171  while (!dlist_is_empty(&pcxt_list))
1172  {
1173  ParallelContext *pcxt;
1174 
1175  pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
1176  if (pcxt->subid != mySubId)
1177  break;
1178  if (isCommit)
1179  elog(WARNING, "leaked parallel context");
1180  DestroyParallelContext(pcxt);
1181  }
1182 }
1183 
1184 /*
1185  * End-of-transaction cleanup for parallel contexts.
1186  */
1187 void
1188 AtEOXact_Parallel(bool isCommit)
1189 {
1190  while (!dlist_is_empty(&pcxt_list))
1191  {
1192  ParallelContext *pcxt;
1193 
1194  pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
1195  if (isCommit)
1196  elog(WARNING, "leaked parallel context");
1197  DestroyParallelContext(pcxt);
1198  }
1199 }
1200 
1201 /*
1202  * Main entrypoint for parallel workers.
1203  */
1204 void
1206 {
1207  dsm_segment *seg;
1208  shm_toc *toc;
1209  FixedParallelState *fps;
1210  char *error_queue_space;
1211  shm_mq *mq;
1212  shm_mq_handle *mqh;
1213  char *libraryspace;
1214  char *entrypointstate;
1215  char *library_name;
1216  char *function_name;
1217  parallel_worker_main_type entrypt;
1218  char *gucspace;
1219  char *combocidspace;
1220  char *tsnapspace;
1221  char *asnapspace;
1222  char *tstatespace;
1223  char *reindexspace;
1224  char *relmapperspace;
1225  char *enumblacklistspace;
1226  StringInfoData msgbuf;
1227  char *session_dsm_handle_space;
1228 
1229  /* Set flag to indicate that we're initializing a parallel worker. */
1231 
1232  /* Establish signal handlers. */
1233  pqsignal(SIGTERM, die);
1235 
1236  /* Determine and set our parallel worker number. */
1238  memcpy(&ParallelWorkerNumber, MyBgworkerEntry->bgw_extra, sizeof(int));
1239 
1240  /* Set up a memory context to work in, just for cleanliness. */
1242  "Parallel worker",
1244 
1245  /*
1246  * Attach to the dynamic shared memory segment for the parallel query, and
1247  * find its table of contents.
1248  *
1249  * Note: at this point, we have not created any ResourceOwner in this
1250  * process. This will result in our DSM mapping surviving until process
1251  * exit, which is fine. If there were a ResourceOwner, it would acquire
1252  * ownership of the mapping, but we have no need for that.
1253  */
1254  seg = dsm_attach(DatumGetUInt32(main_arg));
1255  if (seg == NULL)
1256  ereport(ERROR,
1257  (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1258  errmsg("could not map dynamic shared memory segment")));
1260  if (toc == NULL)
1261  ereport(ERROR,
1262  (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1263  errmsg("invalid magic number in dynamic shared memory segment")));
1264 
1265  /* Look up fixed parallel state. */
1266  fps = shm_toc_lookup(toc, PARALLEL_KEY_FIXED, false);
1267  MyFixedParallelState = fps;
1268 
1269  /* Arrange to signal the leader if we exit. */
1273 
1274  /*
1275  * Now we can find and attach to the error queue provided for us. That's
1276  * good, because until we do that, any errors that happen here will not be
1277  * reported back to the process that requested that this worker be
1278  * launched.
1279  */
1280  error_queue_space = shm_toc_lookup(toc, PARALLEL_KEY_ERROR_QUEUE, false);
1281  mq = (shm_mq *) (error_queue_space +
1284  mqh = shm_mq_attach(mq, seg, NULL);
1285  pq_redirect_to_shm_mq(seg, mqh);
1288 
1289  /*
1290  * Send a BackendKeyData message to the process that initiated parallelism
1291  * so that it has access to our PID before it receives any other messages
1292  * from us. Our cancel key is sent, too, since that's the way the
1293  * protocol message is defined, but it won't actually be used for anything
1294  * in this case.
1295  */
1296  pq_beginmessage(&msgbuf, 'K');
1297  pq_sendint32(&msgbuf, (int32) MyProcPid);
1298  pq_sendint32(&msgbuf, (int32) MyCancelKey);
1299  pq_endmessage(&msgbuf);
1300 
1301  /*
1302  * Hooray! Primary initialization is complete. Now, we need to set up our
1303  * backend-local state to match the original backend.
1304  */
1305 
1306  /*
1307  * Join locking group. We must do this before anything that could try to
1308  * acquire a heavyweight lock, because any heavyweight locks acquired to
1309  * this point could block either directly against the parallel group
1310  * leader or against some process which in turn waits for a lock that
1311  * conflicts with the parallel group leader, causing an undetected
1312  * deadlock. (If we can't join the lock group, the leader has gone away,
1313  * so just exit quietly.)
1314  */
1316  fps->parallel_master_pid))
1317  return;
1318 
1319  /*
1320  * Restore transaction and statement start-time timestamps. This must
1321  * happen before anything that would start a transaction, else asserts in
1322  * xact.c will fire.
1323  */
1325 
1326  /*
1327  * Identify the entry point to be called. In theory this could result in
1328  * loading an additional library, though most likely the entry point is in
1329  * the core backend or in a library we just loaded.
1330  */
1331  entrypointstate = shm_toc_lookup(toc, PARALLEL_KEY_ENTRYPOINT, false);
1332  library_name = entrypointstate;
1333  function_name = entrypointstate + strlen(library_name) + 1;
1334 
1335  entrypt = LookupParallelWorkerFunction(library_name, function_name);
1336 
1337  /* Restore database connection. */
1339  fps->authenticated_user_id,
1340  0);
1341 
1342  /*
1343  * Set the client encoding to the database encoding, since that is what
1344  * the leader will expect.
1345  */
1347 
1348  /*
1349  * Load libraries that were loaded by original backend. We want to do
1350  * this before restoring GUCs, because the libraries might define custom
1351  * variables.
1352  */
1353  libraryspace = shm_toc_lookup(toc, PARALLEL_KEY_LIBRARY, false);
1355  RestoreLibraryState(libraryspace);
1356 
1357  /* Restore GUC values from launching backend. */
1358  gucspace = shm_toc_lookup(toc, PARALLEL_KEY_GUC, false);
1359  RestoreGUCState(gucspace);
1361 
1362  /* Crank up a transaction state appropriate to a parallel worker. */
1363  tstatespace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_STATE, false);
1364  StartParallelWorkerTransaction(tstatespace);
1365 
1366  /* Restore combo CID state. */
1367  combocidspace = shm_toc_lookup(toc, PARALLEL_KEY_COMBO_CID, false);
1368  RestoreComboCIDState(combocidspace);
1369 
1370  /* Attach to the per-session DSM segment and contained objects. */
1371  session_dsm_handle_space =
1373  AttachSession(*(dsm_handle *) session_dsm_handle_space);
1374 
1375  /* Restore transaction snapshot. */
1376  tsnapspace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT, false);
1378  fps->parallel_master_pgproc);
1379 
1380  /* Restore active snapshot. */
1381  asnapspace = shm_toc_lookup(toc, PARALLEL_KEY_ACTIVE_SNAPSHOT, false);
1382  PushActiveSnapshot(RestoreSnapshot(asnapspace));
1383 
1384  /*
1385  * We've changed which tuples we can see, and must therefore invalidate
1386  * system caches.
1387  */
1389 
1390  /*
1391  * Restore current role id. Skip verifying whether session user is
1392  * allowed to become this role and blindly restore the leader's state for
1393  * current role.
1394  */
1396 
1397  /* Restore user ID and security context. */
1399 
1400  /* Restore temp-namespace state to ensure search path matches leader's. */
1403 
1404  /* Restore reindex state. */
1405  reindexspace = shm_toc_lookup(toc, PARALLEL_KEY_REINDEX_STATE, false);
1406  RestoreReindexState(reindexspace);
1407 
1408  /* Restore relmapper state. */
1409  relmapperspace = shm_toc_lookup(toc, PARALLEL_KEY_RELMAPPER_STATE, false);
1410  RestoreRelationMap(relmapperspace);
1411 
1412  /* Restore enum blacklist. */
1413  enumblacklistspace = shm_toc_lookup(toc, PARALLEL_KEY_ENUMBLACKLIST,
1414  false);
1415  RestoreEnumBlacklist(enumblacklistspace);
1416 
1417  /* Attach to the leader's serializable transaction, if SERIALIZABLE. */
1419 
1420  /*
1421  * We've initialized all of our state now; nothing should change
1422  * hereafter.
1423  */
1426 
1427  /*
1428  * Time to do the real work: invoke the caller-supplied code.
1429  */
1430  entrypt(seg, toc);
1431 
1432  /* Must exit parallel mode to pop active snapshot. */
1433  ExitParallelMode();
1434 
1435  /* Must pop active snapshot so snapmgr.c doesn't complain. */
1437 
1438  /* Shut down the parallel-worker transaction. */
1440 
1441  /* Detach from the per-session DSM segment. */
1442  DetachSession();
1443 
1444  /* Report success. */
1445  pq_putmessage('X', NULL, 0);
1446 }
1447 
1448 /*
1449  * Update shared memory with the ending location of the last WAL record we
1450  * wrote, if it's greater than the value already stored there.
1451  */
1452 void
1454 {
1456 
1457  Assert(fps != NULL);
1458  SpinLockAcquire(&fps->mutex);
1459  if (fps->last_xlog_end < last_xlog_end)
1461  SpinLockRelease(&fps->mutex);
1462 }
1463 
1464 /*
1465  * Make sure the leader tries to read from our error queue one more time.
1466  * This guards against the case where we exit uncleanly without sending an
1467  * ErrorResponse to the leader, for example because some code calls proc_exit
1468  * directly.
1469  */
1470 static void
1472 {
1476 }
1477 
1478 /*
1479  * Look up (and possibly load) a parallel worker entry point function.
1480  *
1481  * For functions contained in the core code, we use library name "postgres"
1482  * and consult the InternalParallelWorkers array. External functions are
1483  * looked up, and loaded if necessary, using load_external_function().
1484  *
1485  * The point of this is to pass function names as strings across process
1486  * boundaries. We can't pass actual function addresses because of the
1487  * possibility that the function has been loaded at a different address
1488  * in a different process. This is obviously a hazard for functions in
1489  * loadable libraries, but it can happen even for functions in the core code
1490  * on platforms using EXEC_BACKEND (e.g., Windows).
1491  *
1492  * At some point it might be worthwhile to get rid of InternalParallelWorkers[]
1493  * in favor of applying load_external_function() for core functions too;
1494  * but that raises portability issues that are not worth addressing now.
1495  */
1497 LookupParallelWorkerFunction(const char *libraryname, const char *funcname)
1498 {
1499  /*
1500  * If the function is to be loaded from postgres itself, search the
1501  * InternalParallelWorkers array.
1502  */
1503  if (strcmp(libraryname, "postgres") == 0)
1504  {
1505  int i;
1506 
1507  for (i = 0; i < lengthof(InternalParallelWorkers); i++)
1508  {
1509  if (strcmp(InternalParallelWorkers[i].fn_name, funcname) == 0)
1510  return InternalParallelWorkers[i].fn_addr;
1511  }
1512 
1513  /* We can only reach this by programming error. */
1514  elog(ERROR, "internal function \"%s\" not found", funcname);
1515  }
1516 
1517  /* Otherwise load from external library. */
1518  return (parallel_worker_main_type)
1519  load_external_function(libraryname, funcname, true, NULL);
1520 }
char bgw_extra[BGW_EXTRALEN]
Definition: bgworker.h:98
#define DatumGetUInt32(X)
Definition: postgres.h:486
int slock_t
Definition: s_lock.h:934
void SerializeEnumBlacklist(void *space, Size size)
Definition: pg_enum.c:709
#define PARALLEL_ERROR_QUEUE_SIZE
Definition: parallel.c:53
static const struct @21 InternalParallelWorkers[]
#define AllocSetContextCreate
Definition: memutils.h:170
shm_toc * shm_toc_create(uint64 magic, void *address, Size nbytes)
Definition: shm_toc.c:40
int MyProcPid
Definition: globals.c:40
int errhint(const char *fmt,...)
Definition: elog.c:1069
BackendId MyBackendId
Definition: globals.c:81
Snapshot RestoreSnapshot(char *start_address)
Definition: snapmgr.c:2161
MemoryContext TopTransactionContext
Definition: mcxt.c:49
uint32 dsm_handle
Definition: dsm_impl.h:54
ParallelContext * CreateParallelContext(const char *library_name, const char *function_name, int nworkers)
Definition: parallel.c:158
void SetUserIdAndSecContext(Oid userid, int sec_context)
Definition: miscinit.c:492
static void dlist_push_head(dlist_head *head, dlist_node *node)
Definition: ilist.h:300
XLogRecPtr XactLastRecEnd
Definition: xlog.c:352
void AttachSerializableXact(SerializableXactHandle handle)
Definition: predicate.c:5176
void shm_mq_detach(shm_mq_handle *mqh)
Definition: shm_mq.c:793
PGPROC * MyProc
Definition: proc.c:67
int64 TimestampTz
Definition: timestamp.h:39
dsm_segment * seg
Definition: parallel.h:42
static void WaitForParallelWorkersToExit(ParallelContext *pcxt)
Definition: parallel.c:830
#define dlist_foreach(iter, lhead)
Definition: ilist.h:507
void SerializeReindexState(Size maxsize, char *start_address)
Definition: index.c:3887
char * pstrdup(const char *in)
Definition: mcxt.c:1186
void CommitTransactionCommand(void)
Definition: xact.c:2898
shm_toc_estimator estimator
Definition: parallel.h:41
char * psprintf(const char *fmt,...)
Definition: psprintf.c:46
void EndParallelWorkerTransaction(void)
Definition: xact.c:5275
#define SpinLockInit(lock)
Definition: spin.h:60
void GetTempNamespaceState(Oid *tempNamespaceId, Oid *tempToastNamespaceId)
Definition: namespace.c:3305
void RestoreTransactionSnapshot(Snapshot snapshot, void *master_pgproc)
Definition: snapmgr.c:2225
dsm_segment * dsm_attach(dsm_handle h)
Definition: dsm.c:533
void _bt_parallel_build_main(dsm_segment *seg, shm_toc *toc)
Definition: nbtsort.c:1588
#define Min(x, y)
Definition: c.h:911
PGPROC * shm_mq_get_sender(shm_mq *mq)
Definition: shm_mq.c:249
Oid authenticated_user_id
Definition: parallel.c:82
static MemoryContext MemoryContextSwitchTo(MemoryContext context)
Definition: palloc.h:109
Snapshot GetActiveSnapshot(void)
Definition: snapmgr.c:841
#define PARALLEL_KEY_ENUMBLACKLIST
Definition: parallel.c:75
dsm_handle dsm_segment_handle(dsm_segment *seg)
Definition: dsm.c:966
int bgw_restart_time
Definition: bgworker.h:94
int errcode(int sqlerrcode)
Definition: elog.c:608
PGPROC * parallel_master_pgproc
Definition: parallel.c:89
Oid temp_toast_namespace_id
Definition: parallel.c:86
#define BGWORKER_CLASS_PARALLEL
Definition: bgworker.h:67
void DetachSession(void)
Definition: session.c:201
BackgroundWorker * MyBgworkerEntry
Definition: postmaster.c:192
void MemoryContextReset(MemoryContext context)
Definition: mcxt.c:136
bool BecomeLockGroupMember(PGPROC *leader, int pid)
Definition: proc.c:1878
void PopActiveSnapshot(void)
Definition: snapmgr.c:814
int nknown_attached_workers
Definition: parallel.h:46
uint32 SubTransactionId
Definition: c.h:518
#define PARALLEL_KEY_RELMAPPER_STATE
Definition: parallel.c:74
Size shm_toc_estimate(shm_toc_estimator *e)
Definition: shm_toc.c:263
#define lengthof(array)
Definition: c.h:669
void SerializeTransactionState(Size maxsize, char *start_address)
Definition: xact.c:5180
parallel_worker_main_type fn_addr
Definition: parallel.c:133
const char * pq_getmsgrawstring(StringInfo msg)
Definition: pqformat.c:610
unsigned int Oid
Definition: postgres_ext.h:31
#define shm_toc_estimate_chunk(e, sz)
Definition: shm_toc.h:51
void SetLatch(Latch *latch)
Definition: latch.c:436
#define BGWORKER_SHMEM_ACCESS
Definition: bgworker.h:52
Snapshot GetTransactionSnapshot(void)
Definition: snapmgr.c:306
void InvalidateSystemCaches(void)
Definition: inval.c:643
char bgw_function_name[BGW_MAXLEN]
Definition: bgworker.h:96
static pid_t ParallelMasterPid
Definition: parallel.c:124
void ResetLatch(Latch *latch)
Definition: latch.c:519
void pq_beginmessage(StringInfo buf, char msgtype)
Definition: pqformat.c:87
void RestoreComboCIDState(char *comboCIDstate)
Definition: combocid.c:343
signed int int32
Definition: c.h:347
int WaitLatch(Latch *latch, int wakeEvents, long timeout, uint32 wait_event_info)
Definition: latch.c:344
SubTransactionId subid
Definition: parallel.h:35
Oid GetCurrentRoleId(void)
Definition: miscinit.c:739
#define RESUME_INTERRUPTS()
Definition: miscadmin.h:118
ErrorContextCallback * error_context_stack
Definition: elog.c:91
SerializableXactHandle serializable_xact_handle
Definition: parallel.c:94
volatile bool ParallelMessagePending
Definition: parallel.c:112
static void pq_sendint32(StringInfo buf, uint32 i)
Definition: pqformat.h:145
#define DSM_HANDLE_INVALID
Definition: dsm.h:23
#define sprintf
Definition: port.h:194
#define StaticAssertStmt(condition, errmessage)
Definition: c.h:849
void WaitForParallelWorkersToFinish(ParallelContext *pcxt)
Definition: parallel.c:716
#define SpinLockAcquire(lock)
Definition: spin.h:62
void DestroyParallelContext(ParallelContext *pcxt)
Definition: parallel.c:870
int SendProcSignal(pid_t pid, ProcSignalReason reason, BackendId backendId)
Definition: procsignal.c:179
#define dlist_container(type, membername, ptr)
Definition: ilist.h:477
ParallelWorkerInfo * worker
Definition: parallel.h:45
Datum bgw_main_arg
Definition: bgworker.h:97
void pfree(void *pointer)
Definition: mcxt.c:1056
bool IsInParallelMode(void)
Definition: xact.c:996
void SerializeLibraryState(Size maxsize, char *start_address)
Definition: dfmgr.c:727
#define ERROR
Definition: elog.h:43
BgwHandleStatus WaitForBackgroundWorkerShutdown(BackgroundWorkerHandle *handle)
Definition: bgworker.c:1138
void ParallelQueryMain(dsm_segment *seg, shm_toc *toc)
Oid GetAuthenticatedUserId(void)
Definition: miscinit.c:439
#define PARALLEL_KEY_TRANSACTION_SNAPSHOT
Definition: parallel.c:68
char * function_name
Definition: parallel.h:39
void SerializeSnapshot(Snapshot snapshot, char *start_address)
Definition: snapmgr.c:2102
int32 MyCancelKey
Definition: globals.c:44
void pq_parse_errornotice(StringInfo msg, ErrorData *edata)
Definition: pqmq.c:215
#define FATAL
Definition: elog.h:52
shm_mq * shm_mq_create(void *address, Size size)
Definition: shm_mq.c:169
void ExitParallelMode(void)
Definition: xact.c:976
#define ALLOCSET_DEFAULT_SIZES
Definition: memutils.h:192
#define PARALLEL_KEY_FIXED
Definition: parallel.c:63
void HandleParallelMessages(void)
Definition: parallel.c:968
#define PARALLEL_KEY_ERROR_QUEUE
Definition: parallel.c:64
void on_shmem_exit(pg_on_exit_callback function, Datum arg)
Definition: ipc.c:361
void SetTempNamespaceState(Oid tempNamespaceId, Oid tempToastNamespaceId)
Definition: namespace.c:3321
static void HandleParallelMessage(ParallelContext *pcxt, int i, StringInfo msg)
Definition: parallel.c:1057
#define DSM_CREATE_NULL_IF_MAXSEGMENTS
Definition: dsm.h:20
void PushActiveSnapshot(Snapshot snap)
Definition: snapmgr.c:735
void GetUserIdAndSecContext(Oid *userid, int *sec_context)
Definition: miscinit.c:485
shm_mq_handle * error_mqh
Definition: parallel.h:28
#define PARALLEL_KEY_GUC
Definition: parallel.c:66
int SetClientEncoding(int encoding)
Definition: mbutils.c:201
int ParallelWorkerNumber
Definition: parallel.c:109
BackgroundWorkerHandle * bgwhandle
Definition: parallel.h:27
void SerializeRelationMap(Size maxSize, char *startAddress)
Definition: relmapper.c:657
Size EstimateGUCStateSpace(void)
Definition: guc.c:10074
void AttachSession(dsm_handle handle)
Definition: session.c:155
#define BGW_NEVER_RESTART
Definition: bgworker.h:84
#define shm_toc_initialize_estimator(e)
Definition: shm_toc.h:49
Size EstimateComboCIDStateSpace(void)
Definition: combocid.c:298
#define UInt32GetDatum(X)
Definition: postgres.h:493
MemoryContext CurrentMemoryContext
Definition: mcxt.c:38
static void ParallelWorkerShutdown(int code, Datum arg)
Definition: parallel.c:1471
static void dlist_delete(dlist_node *node)
Definition: ilist.h:358
int nworkers_launched
Definition: parallel.h:37
XLogRecPtr last_xlog_end
Definition: parallel.c:100
BgwHandleStatus
Definition: bgworker.h:102
Size EstimateEnumBlacklistSpace(void)
Definition: pg_enum.c:695
PGFunction load_external_function(const char *filename, const char *funcname, bool signalNotFound, void **filehandle)
Definition: dfmgr.c:107
void LaunchParallelWorkers(ParallelContext *pcxt)
Definition: parallel.c:493
Size EstimateReindexStateSpace(void)
Definition: index.c:3876
void shm_mq_set_sender(shm_mq *mq, PGPROC *proc)
Definition: shm_mq.c:216
void BecomeLockGroupLeader(void)
Definition: proc.c:1848
#define ereport(elevel, rest)
Definition: elog.h:141
MemoryContext TopMemoryContext
Definition: mcxt.c:44
TimestampTz GetCurrentTransactionStartTimestamp(void)
Definition: xact.c:782
void ThrowErrorData(ErrorData *edata)
Definition: elog.c:1703
void initStringInfo(StringInfo str)
Definition: stringinfo.c:59
#define DLIST_STATIC_INIT(name)
Definition: ilist.h:248
#define WARNING
Definition: elog.h:40
void InitializeParallelDSM(ParallelContext *pcxt)
Definition: parallel.c:195
int elevel
Definition: elog.h:380
bool * known_attached_workers
Definition: parallel.h:47
bool ParallelContextActive(void)
Definition: parallel.c:944
#define SpinLockRelease(lock)
Definition: spin.h:64
#define dlist_head_element(type, membername, lhead)
Definition: ilist.h:487
Size EstimateSnapshotSpace(Snapshot snap)
Definition: snapmgr.c:2078
Size mul_size(Size s1, Size s2)
Definition: shmem.c:492
void * palloc0(Size size)
Definition: mcxt.c:980
static parallel_worker_main_type LookupParallelWorkerFunction(const char *libraryname, const char *funcname)
Definition: parallel.c:1497
void RestoreLibraryState(char *start_address)
Definition: dfmgr.c:749
uintptr_t Datum
Definition: postgres.h:367
void RestoreEnumBlacklist(void *space)
Definition: pg_enum.c:741
dsm_segment * dsm_create(Size size, int flags)
Definition: dsm.c:437
shm_toc * shm_toc_attach(uint64 magic, void *address)
Definition: shm_toc.c:64
int GetDatabaseEncoding(void)
Definition: mbutils.c:1046
int BackendId
Definition: backendid.h:21
Oid MyDatabaseId
Definition: globals.c:85
pid_t parallel_master_pid
Definition: parallel.c:90
Size EstimateLibraryStateSpace(void)
Definition: dfmgr.c:710
void ReinitializeParallelDSM(ParallelContext *pcxt)
Definition: parallel.c:446
void shm_mq_set_handle(shm_mq_handle *mqh, BackgroundWorkerHandle *handle)
Definition: shm_mq.c:310
dlist_node * cur
Definition: ilist.h:161
void BackgroundWorkerInitializeConnectionByOid(Oid dboid, Oid useroid, uint32 flags)
Definition: postmaster.c:5708
#define PARALLEL_MAGIC
Definition: parallel.c:56
void pq_redirect_to_shm_mq(dsm_segment *seg, shm_mq_handle *mqh)
Definition: pqmq.c:55
void * SerializableXactHandle
Definition: predicate.h:37
void ParallelWorkerReportLastRecEnd(XLogRecPtr last_xlog_end)
Definition: parallel.c:1453
void TerminateBackgroundWorker(BackgroundWorkerHandle *handle)
Definition: bgworker.c:1177
pqsigfunc pqsignal(int signum, pqsigfunc handler)
Definition: signal.c:170
int pq_getmsgbyte(StringInfo msg)
Definition: pqformat.c:401
shm_mq_result
Definition: shm_mq.h:36
TimestampTz xact_ts
Definition: parallel.c:92
char * library_name
Definition: parallel.h:38
BackendId parallel_master_backend_id
Definition: parallel.c:91
int force_parallel_mode
Definition: planner.c:70
void * dsm_segment_address(dsm_segment *seg)
Definition: dsm.c:938
uint64 XLogRecPtr
Definition: xlogdefs.h:21
char bgw_name[BGW_MAXLEN]
Definition: bgworker.h:90
#define Assert(condition)
Definition: c.h:739
BackendId ParallelMasterBackendId
Definition: globals.c:83
void StartParallelWorkerTransaction(char *tstatespace)
Definition: xact.c:5250
#define BGWORKER_BACKEND_DATABASE_CONNECTION
Definition: bgworker.h:59
SubTransactionId GetCurrentSubTransactionId(void)
Definition: xact.c:707
Size EstimateTransactionStateSpace(void)
Definition: xact.c:5152
void StartTransactionCommand(void)
Definition: xact.c:2797
#define PARALLEL_KEY_REINDEX_STATE
Definition: parallel.c:73
const char * fn_name
Definition: parallel.c:132
static bool dlist_is_empty(dlist_head *head)
Definition: ilist.h:289
#define BGW_MAXLEN
Definition: bgworker.h:85
size_t Size
Definition: c.h:467
BgWorkerStartTime bgw_start_time
Definition: bgworker.h:93
dlist_node node
Definition: parallel.h:34
#define shm_toc_estimate_keys(e, cnt)
Definition: shm_toc.h:53
shm_mq * shm_mq_get_queue(shm_mq_handle *mqh)
Definition: shm_mq.c:848
bool RegisterDynamicBackgroundWorker(BackgroundWorker *worker, BackgroundWorkerHandle **handle)
Definition: bgworker.c:932
void EnterParallelMode(void)
Definition: xact.c:963
volatile sig_atomic_t InterruptPending
Definition: globals.c:30
Size EstimateRelationMapSpace(void)
Definition: relmapper.c:646
void * shm_toc_allocate(shm_toc *toc, Size nbytes)
Definition: shm_toc.c:88
char * context
Definition: elog.h:396
shm_mq_handle * shm_mq_attach(shm_mq *mq, dsm_segment *seg, BackgroundWorkerHandle *handle)
Definition: shm_mq.c:282
ErrorContextCallback * error_context_stack
Definition: parallel.h:40
void pq_set_parallel_master(pid_t pid, BackendId backend_id)
Definition: pqmq.c:80
void SetParallelStartTimestamps(TimestampTz xact_ts, TimestampTz stmt_ts)
Definition: xact.c:771
#define PARALLEL_KEY_TRANSACTION_STATE
Definition: parallel.c:70
char bgw_type[BGW_MAXLEN]
Definition: bgworker.h:91
void dsm_detach(dsm_segment *seg)
Definition: dsm.c:664
void shm_toc_insert(shm_toc *toc, uint64 key, void *address)
Definition: shm_toc.c:171
void RestoreReindexState(void *reindexstate)
Definition: index.c:3905
int errmsg(const char *fmt,...)
Definition: elog.c:822
void(* parallel_worker_main_type)(dsm_segment *seg, shm_toc *toc)
Definition: parallel.h:23
void pq_endmessage(StringInfo buf)
Definition: pqformat.c:298
void ParallelWorkerMain(Datum main_arg)
Definition: parallel.c:1205
pid_t bgw_notify_pid
Definition: bgworker.h:99
static FixedParallelState * MyFixedParallelState
Definition: parallel.c:118
void * MemoryContextAlloc(MemoryContext context, Size size)
Definition: mcxt.c:796
void SetCurrentRoleId(Oid roleid, bool is_superuser)
Definition: miscinit.c:760
#define HOLD_INTERRUPTS()
Definition: miscadmin.h:116
#define elog(elevel,...)
Definition: elog.h:228
bool InitializingParallelWorker
Definition: parallel.c:115
int i
TimestampTz stmt_ts
Definition: parallel.c:93
Definition: shm_mq.c:70
void RestoreRelationMap(char *startAddress)
Definition: relmapper.c:674
#define PARALLEL_KEY_SESSION_DSM
Definition: parallel.c:72
#define BUFFERALIGN(LEN)
Definition: c.h:694
void * arg
struct Latch * MyLatch
Definition: globals.c:54
void HandleParallelMessageInterrupt(void)
Definition: parallel.c:957
unsigned int pq_getmsgint(StringInfo msg, int b)
Definition: pqformat.c:417
#define CHECK_FOR_INTERRUPTS()
Definition: miscadmin.h:99
void shm_mq_set_receiver(shm_mq *mq, PGPROC *proc)
Definition: shm_mq.c:198
static dlist_head pcxt_list
Definition: parallel.c:121
void pq_getmsgend(StringInfo msg)
Definition: pqformat.c:637
#define pq_putmessage(msgtype, s, len)
Definition: libpq.h:42
static void static void status(const char *fmt,...) pg_attribute_printf(1
Definition: pg_regress.c:226
struct FixedParallelState FixedParallelState
dsm_handle GetSessionDsmHandle(void)
Definition: session.c:70
shm_mq_result shm_mq_receive(shm_mq_handle *mqh, Size *nbytesp, void **datap, bool nowait)
Definition: shm_mq.c:540
char bgw_library_name[BGW_MAXLEN]
Definition: bgworker.h:95
void WaitForParallelWorkersToAttach(ParallelContext *pcxt)
Definition: parallel.c:613
bool session_auth_is_superuser
Definition: guc.c:511
SerializableXactHandle ShareSerializableXact(void)
Definition: predicate.c:5167
void AtEOXact_Parallel(bool isCommit)
Definition: parallel.c:1188
Definition: proc.h:95
#define PARALLEL_KEY_ENTRYPOINT
Definition: parallel.c:71
#define snprintf
Definition: port.h:192
#define PARALLEL_KEY_COMBO_CID
Definition: parallel.c:67
#define WL_LATCH_SET
Definition: latch.h:124
#define _(x)
Definition: elog.c:87
void AtEOSubXact_Parallel(bool isCommit, SubTransactionId mySubId)
Definition: parallel.c:1169
void SerializeGUCState(Size maxsize, char *start_address)
Definition: guc.c:10216
void appendBinaryStringInfo(StringInfo str, const char *data, int datalen)
Definition: stringinfo.c:227
#define PARALLEL_KEY_ACTIVE_SNAPSHOT
Definition: parallel.c:69
void * shm_toc_lookup(shm_toc *toc, uint64 key, bool noError)
Definition: shm_toc.c:232
#define die(msg)
Definition: pg_test_fsync.c:96
void SerializeComboCIDState(Size maxsize, char *start_address)
Definition: combocid.c:317
#define PARALLEL_KEY_LIBRARY
Definition: parallel.c:65
BgwHandleStatus GetBackgroundWorkerPid(BackgroundWorkerHandle *handle, pid_t *pidp)
Definition: bgworker.c:1044
void RestoreGUCState(void *gucstate)
Definition: guc.c:10296
shm_toc * toc
Definition: parallel.h:44
#define WL_EXIT_ON_PM_DEATH
Definition: latch.h:129
void NotifyMyFrontEnd(const char *channel, const char *payload, int32 srcPid)
Definition: async.c:2255
TimestampTz GetCurrentStatementStartTimestamp(void)
Definition: xact.c:791
void * private_memory
Definition: parallel.h:43
void BackgroundWorkerUnblockSignals(void)
Definition: postmaster.c:5737