Debian Codegraph @

src/backend/access/transam/parallel.c

raw
1 /*-------------------------------------------------------------------------
2 *
3 * parallel.c
4 * Infrastructure for launching parallel workers
5 *
6 * Portions Copyright (c) 1996-2025, 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/brin.h"
18 #include "access/gin.h"
19 #include "access/nbtree.h"
20 #include "access/parallel.h"
21 #include "access/session.h"
22 #include "access/xact.h"
23 #include "access/xlog.h"
24 #include "catalog/index.h"
25 #include "catalog/namespace.h"
26 #include "catalog/pg_enum.h"
27 #include "catalog/storage.h"
28 #include "commands/async.h"
29 #include "commands/vacuum.h"
30 #include "executor/execParallel.h"
31 #include "libpq/libpq.h"
32 #include "libpq/pqformat.h"
33 #include "libpq/pqmq.h"
34 #include "miscadmin.h"
35 #include "optimizer/optimizer.h"
36 #include "pgstat.h"
37 #include "storage/ipc.h"
38 #include "storage/predicate.h"
39 #include "storage/spin.h"
40 #include "tcop/tcopprot.h"
41 #include "utils/combocid.h"
42 #include "utils/guc.h"
43 #include "utils/inval.h"
44 #include "utils/memutils.h"
45 #include "utils/relmapper.h"
46 #include "utils/snapmgr.h"
47
48 /*
49 * We don't want to waste a lot of memory on an error queue which, most of
50 * the time, will process only a handful of small messages. However, it is
51 * desirable to make it large enough that a typical ErrorResponse can be sent
52 * without blocking. That way, a worker that errors out can write the whole
53 * message into the queue and terminate without waiting for the user backend.
54 */
55 #define PARALLEL_ERROR_QUEUE_SIZE 16384
56
57 /* Magic number for parallel context TOC. */
58 #define PARALLEL_MAGIC 0x50477c7c
59
60 /*
61 * Magic numbers for per-context parallel state sharing. Higher-level code
62 * should use smaller values, leaving these very large ones for use by this
63 * module.
64 */
65 #define PARALLEL_KEY_FIXED UINT64CONST(0xFFFFFFFFFFFF0001)
66 #define PARALLEL_KEY_ERROR_QUEUE UINT64CONST(0xFFFFFFFFFFFF0002)
67 #define PARALLEL_KEY_LIBRARY UINT64CONST(0xFFFFFFFFFFFF0003)
68 #define PARALLEL_KEY_GUC UINT64CONST(0xFFFFFFFFFFFF0004)
69 #define PARALLEL_KEY_COMBO_CID UINT64CONST(0xFFFFFFFFFFFF0005)
70 #define PARALLEL_KEY_TRANSACTION_SNAPSHOT UINT64CONST(0xFFFFFFFFFFFF0006)
71 #define PARALLEL_KEY_ACTIVE_SNAPSHOT UINT64CONST(0xFFFFFFFFFFFF0007)
72 #define PARALLEL_KEY_TRANSACTION_STATE UINT64CONST(0xFFFFFFFFFFFF0008)
73 #define PARALLEL_KEY_ENTRYPOINT UINT64CONST(0xFFFFFFFFFFFF0009)
74 #define PARALLEL_KEY_SESSION_DSM UINT64CONST(0xFFFFFFFFFFFF000A)
75 #define PARALLEL_KEY_PENDING_SYNCS UINT64CONST(0xFFFFFFFFFFFF000B)
76 #define PARALLEL_KEY_REINDEX_STATE UINT64CONST(0xFFFFFFFFFFFF000C)
77 #define PARALLEL_KEY_RELMAPPER_STATE UINT64CONST(0xFFFFFFFFFFFF000D)
78 #define PARALLEL_KEY_UNCOMMITTEDENUMS UINT64CONST(0xFFFFFFFFFFFF000E)
79 #define PARALLEL_KEY_CLIENTCONNINFO UINT64CONST(0xFFFFFFFFFFFF000F)
80
81 /* Fixed-size parallel state. */
82 typedef struct FixedParallelState
83 {
84 /* Fixed-size state that workers must restore. */
85 Oid database_id;
86 Oid authenticated_user_id;
87 Oid session_user_id;
88 Oid outer_user_id;
89 Oid current_user_id;
90 Oid temp_namespace_id;
91 Oid temp_toast_namespace_id;
92 int sec_context;
93 bool session_user_is_superuser;
94 bool role_is_superuser;
95 PGPROC *parallel_leader_pgproc;
96 pid_t parallel_leader_pid;
97 ProcNumber parallel_leader_proc_number;
98 TimestampTz xact_ts;
99 TimestampTz stmt_ts;
100 SerializableXactHandle serializable_xact_handle;
101
102 /* Mutex protects remaining fields. */
103 slock_t mutex;
104
105 /* Maximum XactLastRecEnd of any worker. */
106 XLogRecPtr last_xlog_end;
107 } FixedParallelState;
108
109 /*
110 * Our parallel worker number. We initialize this to -1, meaning that we are
111 * not a parallel worker. In parallel workers, it will be set to a value >= 0
112 * and < the number of workers before any user code is invoked; each parallel
113 * worker will get a different parallel worker number.
114 */
115 int ParallelWorkerNumber = -1;
116
117 /* Is there a parallel message pending which we need to receive? */
118 volatile sig_atomic_t ParallelMessagePending = false;
119
120 /* Are we initializing a parallel worker? */
121 bool InitializingParallelWorker = false;
122
123 /* Pointer to our fixed parallel state. */
124 static FixedParallelState *MyFixedParallelState;
125
126 /* List of active parallel contexts. */
127 static dlist_head pcxt_list = DLIST_STATIC_INIT(pcxt_list);
128
129 /* Backend-local copy of data from FixedParallelState. */
130 static pid_t ParallelLeaderPid;
131
132 /*
133 * List of internal parallel worker entry points. We need this for
134 * reasons explained in LookupParallelWorkerFunction(), below.
135 */
136 static const struct
137 {
138 const char *fn_name;
139 parallel_worker_main_type fn_addr;
140 } InternalParallelWorkers[] =
141
142 {
143 {
144 "ParallelQueryMain", ParallelQueryMain
145 },
146 {
147 "_bt_parallel_build_main", _bt_parallel_build_main
148 },
149 {
150 "_brin_parallel_build_main", _brin_parallel_build_main
151 },
152 {
153 "_gin_parallel_build_main", _gin_parallel_build_main
154 },
155 {
156 "parallel_vacuum_main", parallel_vacuum_main
157 }
158 };
159
160 /* Private functions. */
161 static void ProcessParallelMessage(ParallelContext *pcxt, int i, StringInfo msg);
162 static void WaitForParallelWorkersToExit(ParallelContext *pcxt);
163 static parallel_worker_main_type LookupParallelWorkerFunction(const char *libraryname, const char *funcname);
164 static void ParallelWorkerShutdown(int code, Datum arg);
165
166
167 /*
168 * Establish a new parallel context. This should be done after entering
169 * parallel mode, and (unless there is an error) the context should be
170 * destroyed before exiting the current subtransaction.
171 */
172 ParallelContext *
173 CreateParallelContext(const char *library_name, const char *function_name,
174 int nworkers)
175 {
176 MemoryContext oldcontext;
177 ParallelContext *pcxt;
178
179 /* It is unsafe to create a parallel context if not in parallel mode. */
180 Assert(IsInParallelMode());
181
182 /* Number of workers should be non-negative. */
183 Assert(nworkers >= 0);
184
185 /* We might be running in a short-lived memory context. */
186 oldcontext = MemoryContextSwitchTo(TopTransactionContext);
187
188 /* Initialize a new ParallelContext. */
189 pcxt = palloc0(sizeof(ParallelContext));
190 pcxt->subid = GetCurrentSubTransactionId();
191 pcxt->nworkers = nworkers;
192 pcxt->nworkers_to_launch = nworkers;
193 pcxt->library_name = pstrdup(library_name);
194 pcxt->function_name = pstrdup(function_name);
195 pcxt->error_context_stack = error_context_stack;
196 shm_toc_initialize_estimator(&pcxt->estimator);
197 dlist_push_head(&pcxt_list, &pcxt->node);
198
199 /* Restore previous memory context. */
200 MemoryContextSwitchTo(oldcontext);
201
202 return pcxt;
203 }
204
205 /*
206 * Establish the dynamic shared memory segment for a parallel context and
207 * copy state and other bookkeeping information that will be needed by
208 * parallel workers into it.
209 */
210 void
211 InitializeParallelDSM(ParallelContext *pcxt)
212 {
213 MemoryContext oldcontext;
214 Size library_len = 0;
215 Size guc_len = 0;
216 Size combocidlen = 0;
217 Size tsnaplen = 0;
218 Size asnaplen = 0;
219 Size tstatelen = 0;
220 Size pendingsyncslen = 0;
221 Size reindexlen = 0;
222 Size relmapperlen = 0;
223 Size uncommittedenumslen = 0;
224 Size clientconninfolen = 0;
225 Size segsize = 0;
226 int i;
227 FixedParallelState *fps;
228 dsm_handle session_dsm_handle = DSM_HANDLE_INVALID;
229 Snapshot transaction_snapshot = GetTransactionSnapshot();
230 Snapshot active_snapshot = GetActiveSnapshot();
231
232 /* We might be running in a very short-lived memory context. */
233 oldcontext = MemoryContextSwitchTo(TopTransactionContext);
234
235 /* Allow space to store the fixed-size parallel state. */
236 shm_toc_estimate_chunk(&pcxt->estimator, sizeof(FixedParallelState));
237 shm_toc_estimate_keys(&pcxt->estimator, 1);
238
239 /*
240 * If we manage to reach here while non-interruptible, it's unsafe to
241 * launch any workers: we would fail to process interrupts sent by them.
242 * We can deal with that edge case by pretending no workers were
243 * requested.
244 */
245 if (!INTERRUPTS_CAN_BE_PROCESSED())
246 pcxt->nworkers = 0;
247
248 /*
249 * Normally, the user will have requested at least one worker process, but
250 * if by chance they have not, we can skip a bunch of things here.
251 */
252 if (pcxt->nworkers > 0)
253 {
254 /* Get (or create) the per-session DSM segment's handle. */
255 session_dsm_handle = GetSessionDsmHandle();
256
257 /*
258 * If we weren't able to create a per-session DSM segment, then we can
259 * continue but we can't safely launch any workers because their
260 * record typmods would be incompatible so they couldn't exchange
261 * tuples.
262 */
263 if (session_dsm_handle == DSM_HANDLE_INVALID)
264 pcxt->nworkers = 0;
265 }
266
267 if (pcxt->nworkers > 0)
268 {
269 /* Estimate space for various kinds of state sharing. */
270 library_len = EstimateLibraryStateSpace();
271 shm_toc_estimate_chunk(&pcxt->estimator, library_len);
272 guc_len = EstimateGUCStateSpace();
273 shm_toc_estimate_chunk(&pcxt->estimator, guc_len);
274 combocidlen = EstimateComboCIDStateSpace();
275 shm_toc_estimate_chunk(&pcxt->estimator, combocidlen);
276 if (IsolationUsesXactSnapshot())
277 {
278 tsnaplen = EstimateSnapshotSpace(transaction_snapshot);
279 shm_toc_estimate_chunk(&pcxt->estimator, tsnaplen);
280 }
281 asnaplen = EstimateSnapshotSpace(active_snapshot);
282 shm_toc_estimate_chunk(&pcxt->estimator, asnaplen);
283 tstatelen = EstimateTransactionStateSpace();
284 shm_toc_estimate_chunk(&pcxt->estimator, tstatelen);
285 shm_toc_estimate_chunk(&pcxt->estimator, sizeof(dsm_handle));
286 pendingsyncslen = EstimatePendingSyncsSpace();
287 shm_toc_estimate_chunk(&pcxt->estimator, pendingsyncslen);
288 reindexlen = EstimateReindexStateSpace();
289 shm_toc_estimate_chunk(&pcxt->estimator, reindexlen);
290 relmapperlen = EstimateRelationMapSpace();
291 shm_toc_estimate_chunk(&pcxt->estimator, relmapperlen);
292 uncommittedenumslen = EstimateUncommittedEnumsSpace();
293 shm_toc_estimate_chunk(&pcxt->estimator, uncommittedenumslen);
294 clientconninfolen = EstimateClientConnectionInfoSpace();
295 shm_toc_estimate_chunk(&pcxt->estimator, clientconninfolen);
296 /* If you add more chunks here, you probably need to add keys. */
297 shm_toc_estimate_keys(&pcxt->estimator, 12);
298
299 /* Estimate space need for error queues. */
300 StaticAssertStmt(BUFFERALIGN(PARALLEL_ERROR_QUEUE_SIZE) ==
301 PARALLEL_ERROR_QUEUE_SIZE,
302 "parallel error queue size not buffer-aligned");
303 shm_toc_estimate_chunk(&pcxt->estimator,
304 mul_size(PARALLEL_ERROR_QUEUE_SIZE,
305 pcxt->nworkers));
306 shm_toc_estimate_keys(&pcxt->estimator, 1);
307
308 /* Estimate how much we'll need for the entrypoint info. */
309 shm_toc_estimate_chunk(&pcxt->estimator, strlen(pcxt->library_name) +
310 strlen(pcxt->function_name) + 2);
311 shm_toc_estimate_keys(&pcxt->estimator, 1);
312 }
313
314 /*
315 * Create DSM and initialize with new table of contents. But if the user
316 * didn't request any workers, then don't bother creating a dynamic shared
317 * memory segment; instead, just use backend-private memory.
318 *
319 * Also, if we can't create a dynamic shared memory segment because the
320 * maximum number of segments have already been created, then fall back to
321 * backend-private memory, and plan not to use any workers. We hope this
322 * won't happen very often, but it's better to abandon the use of
323 * parallelism than to fail outright.
324 */
325 segsize = shm_toc_estimate(&pcxt->estimator);
326 if (pcxt->nworkers > 0)
327 pcxt->seg = dsm_create(segsize, DSM_CREATE_NULL_IF_MAXSEGMENTS);
328 if (pcxt->seg != NULL)
329 pcxt->toc = shm_toc_create(PARALLEL_MAGIC,
330 dsm_segment_address(pcxt->seg),
331 segsize);
332 else
333 {
334 pcxt->nworkers = 0;
335 pcxt->private_memory = MemoryContextAlloc(TopMemoryContext, segsize);
336 pcxt->toc = shm_toc_create(PARALLEL_MAGIC, pcxt->private_memory,
337 segsize);
338 }
339
340 /* Initialize fixed-size state in shared memory. */
341 fps = (FixedParallelState *)
342 shm_toc_allocate(pcxt->toc, sizeof(FixedParallelState));
343 fps->database_id = MyDatabaseId;
344 fps->authenticated_user_id = GetAuthenticatedUserId();
345 fps->session_user_id = GetSessionUserId();
346 fps->outer_user_id = GetCurrentRoleId();
347 GetUserIdAndSecContext(&fps->current_user_id, &fps->sec_context);
348 fps->session_user_is_superuser = GetSessionUserIsSuperuser();
349 fps->role_is_superuser = current_role_is_superuser;
350 GetTempNamespaceState(&fps->temp_namespace_id,
351 &fps->temp_toast_namespace_id);
352 fps->parallel_leader_pgproc = MyProc;
353 fps->parallel_leader_pid = MyProcPid;
354 fps->parallel_leader_proc_number = MyProcNumber;
355 fps->xact_ts = GetCurrentTransactionStartTimestamp();
356 fps->stmt_ts = GetCurrentStatementStartTimestamp();
357 fps->serializable_xact_handle = ShareSerializableXact();
358 SpinLockInit(&fps->mutex);
359 fps->last_xlog_end = 0;
360 shm_toc_insert(pcxt->toc, PARALLEL_KEY_FIXED, fps);
361
362 /* We can skip the rest of this if we're not budgeting for any workers. */
363 if (pcxt->nworkers > 0)
364 {
365 char *libraryspace;
366 char *gucspace;
367 char *combocidspace;
368 char *tsnapspace;
369 char *asnapspace;
370 char *tstatespace;
371 char *pendingsyncsspace;
372 char *reindexspace;
373 char *relmapperspace;
374 char *error_queue_space;
375 char *session_dsm_handle_space;
376 char *entrypointstate;
377 char *uncommittedenumsspace;
378 char *clientconninfospace;
379 Size lnamelen;
380
381 /* Serialize shared libraries we have loaded. */
382 libraryspace = shm_toc_allocate(pcxt->toc, library_len);
383 SerializeLibraryState(library_len, libraryspace);
384 shm_toc_insert(pcxt->toc, PARALLEL_KEY_LIBRARY, libraryspace);
385
386 /* Serialize GUC settings. */
387 gucspace = shm_toc_allocate(pcxt->toc, guc_len);
388 SerializeGUCState(guc_len, gucspace);
389 shm_toc_insert(pcxt->toc, PARALLEL_KEY_GUC, gucspace);
390
391 /* Serialize combo CID state. */
392 combocidspace = shm_toc_allocate(pcxt->toc, combocidlen);
393 SerializeComboCIDState(combocidlen, combocidspace);
394 shm_toc_insert(pcxt->toc, PARALLEL_KEY_COMBO_CID, combocidspace);
395
396 /*
397 * Serialize the transaction snapshot if the transaction isolation
398 * level uses a transaction snapshot.
399 */
400 if (IsolationUsesXactSnapshot())
401 {
402 tsnapspace = shm_toc_allocate(pcxt->toc, tsnaplen);
403 SerializeSnapshot(transaction_snapshot, tsnapspace);
404 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT,
405 tsnapspace);
406 }
407
408 /* Serialize the active snapshot. */
409 asnapspace = shm_toc_allocate(pcxt->toc, asnaplen);
410 SerializeSnapshot(active_snapshot, asnapspace);
411 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ACTIVE_SNAPSHOT, asnapspace);
412
413 /* Provide the handle for per-session segment. */
414 session_dsm_handle_space = shm_toc_allocate(pcxt->toc,
415 sizeof(dsm_handle));
416 *(dsm_handle *) session_dsm_handle_space = session_dsm_handle;
417 shm_toc_insert(pcxt->toc, PARALLEL_KEY_SESSION_DSM,
418 session_dsm_handle_space);
419
420 /* Serialize transaction state. */
421 tstatespace = shm_toc_allocate(pcxt->toc, tstatelen);
422 SerializeTransactionState(tstatelen, tstatespace);
423 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_STATE, tstatespace);
424
425 /* Serialize pending syncs. */
426 pendingsyncsspace = shm_toc_allocate(pcxt->toc, pendingsyncslen);
427 SerializePendingSyncs(pendingsyncslen, pendingsyncsspace);
428 shm_toc_insert(pcxt->toc, PARALLEL_KEY_PENDING_SYNCS,
429 pendingsyncsspace);
430
431 /* Serialize reindex state. */
432 reindexspace = shm_toc_allocate(pcxt->toc, reindexlen);
433 SerializeReindexState(reindexlen, reindexspace);
434 shm_toc_insert(pcxt->toc, PARALLEL_KEY_REINDEX_STATE, reindexspace);
435
436 /* Serialize relmapper state. */
437 relmapperspace = shm_toc_allocate(pcxt->toc, relmapperlen);
438 SerializeRelationMap(relmapperlen, relmapperspace);
439 shm_toc_insert(pcxt->toc, PARALLEL_KEY_RELMAPPER_STATE,
440 relmapperspace);
441
442 /* Serialize uncommitted enum state. */
443 uncommittedenumsspace = shm_toc_allocate(pcxt->toc,
444 uncommittedenumslen);
445 SerializeUncommittedEnums(uncommittedenumsspace, uncommittedenumslen);
446 shm_toc_insert(pcxt->toc, PARALLEL_KEY_UNCOMMITTEDENUMS,
447 uncommittedenumsspace);
448
449 /* Serialize our ClientConnectionInfo. */
450 clientconninfospace = shm_toc_allocate(pcxt->toc, clientconninfolen);
451 SerializeClientConnectionInfo(clientconninfolen, clientconninfospace);
452 shm_toc_insert(pcxt->toc, PARALLEL_KEY_CLIENTCONNINFO,
453 clientconninfospace);
454
455 /* Allocate space for worker information. */
456 pcxt->worker = palloc0(sizeof(ParallelWorkerInfo) * pcxt->nworkers);
457
458 /*
459 * Establish error queues in dynamic shared memory.
460 *
461 * These queues should be used only for transmitting ErrorResponse,
462 * NoticeResponse, and NotifyResponse protocol messages. Tuple data
463 * should be transmitted via separate (possibly larger?) queues.
464 */
465 error_queue_space =
466 shm_toc_allocate(pcxt->toc,
467 mul_size(PARALLEL_ERROR_QUEUE_SIZE,
468 pcxt->nworkers));
469 for (i = 0; i < pcxt->nworkers; ++i)
470 {
471 char *start;
472 shm_mq *mq;
473
474 start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
475 mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
476 shm_mq_set_receiver(mq, MyProc);
477 pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
478 }
479 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE, error_queue_space);
480
481 /*
482 * Serialize entrypoint information. It's unsafe to pass function
483 * pointers across processes, as the function pointer may be different
484 * in each process in EXEC_BACKEND builds, so we always pass library
485 * and function name. (We use library name "postgres" for functions
486 * in the core backend.)
487 */
488 lnamelen = strlen(pcxt->library_name);
489 entrypointstate = shm_toc_allocate(pcxt->toc, lnamelen +
490 strlen(pcxt->function_name) + 2);
491 strcpy(entrypointstate, pcxt->library_name);
492 strcpy(entrypointstate + lnamelen + 1, pcxt->function_name);
493 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ENTRYPOINT, entrypointstate);
494 }
495
496 /* Update nworkers_to_launch, in case we changed nworkers above. */
497 pcxt->nworkers_to_launch = pcxt->nworkers;
498
499 /* Restore previous memory context. */
500 MemoryContextSwitchTo(oldcontext);
501 }
502
503 /*
504 * Reinitialize the dynamic shared memory segment for a parallel context such
505 * that we could launch workers for it again.
506 */
507 void
508 ReinitializeParallelDSM(ParallelContext *pcxt)
509 {
510 MemoryContext oldcontext;
511 FixedParallelState *fps;
512
513 /* We might be running in a very short-lived memory context. */
514 oldcontext = MemoryContextSwitchTo(TopTransactionContext);
515
516 /* Wait for any old workers to exit. */
517 if (pcxt->nworkers_launched > 0)
518 {
519 WaitForParallelWorkersToFinish(pcxt);
520 WaitForParallelWorkersToExit(pcxt);
521 pcxt->nworkers_launched = 0;
522 if (pcxt->known_attached_workers)
523 {
524 pfree(pcxt->known_attached_workers);
525 pcxt->known_attached_workers = NULL;
526 pcxt->nknown_attached_workers = 0;
527 }
528 }
529
530 /* Reset a few bits of fixed parallel state to a clean state. */
531 fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED, false);
532 fps->last_xlog_end = 0;
533
534 /* Recreate error queues (if they exist). */
535 if (pcxt->nworkers > 0)
536 {
537 char *error_queue_space;
538 int i;
539
540 error_queue_space =
541 shm_toc_lookup(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE, false);
542 for (i = 0; i < pcxt->nworkers; ++i)
543 {
544 char *start;
545 shm_mq *mq;
546
547 start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
548 mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
549 shm_mq_set_receiver(mq, MyProc);
550 pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
551 }
552 }
553
554 /* Restore previous memory context. */
555 MemoryContextSwitchTo(oldcontext);
556 }
557
558 /*
559 * Reinitialize parallel workers for a parallel context such that we could
560 * launch a different number of workers. This is required for cases where
561 * we need to reuse the same DSM segment, but the number of workers can
562 * vary from run-to-run.
563 */
564 void
565 ReinitializeParallelWorkers(ParallelContext *pcxt, int nworkers_to_launch)
566 {
567 /*
568 * The number of workers that need to be launched must be less than the
569 * number of workers with which the parallel context is initialized. But
570 * the caller might not know that InitializeParallelDSM reduced nworkers,
571 * so just silently trim the request.
572 */
573 pcxt->nworkers_to_launch = Min(pcxt->nworkers, nworkers_to_launch);
574 }
575
576 /*
577 * Launch parallel workers.
578 */
579 void
580 LaunchParallelWorkers(ParallelContext *pcxt)
581 {
582 MemoryContext oldcontext;
583 BackgroundWorker worker;
584 int i;
585 bool any_registrations_failed = false;
586
587 /* Skip this if we have no workers. */
588 if (pcxt->nworkers == 0 || pcxt->nworkers_to_launch == 0)
589 return;
590
591 /* We need to be a lock group leader. */
592 BecomeLockGroupLeader();
593
594 /* If we do have workers, we'd better have a DSM segment. */
595 Assert(pcxt->seg != NULL);
596
597 /* We might be running in a short-lived memory context. */
598 oldcontext = MemoryContextSwitchTo(TopTransactionContext);
599
600 /* Configure a worker. */
601 memset(&worker, 0, sizeof(worker));
602 snprintf(worker.bgw_name, BGW_MAXLEN, "parallel worker for PID %d",
603 MyProcPid);
604 snprintf(worker.bgw_type, BGW_MAXLEN, "parallel worker");
605 worker.bgw_flags =
606 BGWORKER_SHMEM_ACCESS | BGWORKER_BACKEND_DATABASE_CONNECTION
607 | BGWORKER_CLASS_PARALLEL;
608 worker.bgw_start_time = BgWorkerStart_ConsistentState;
609 worker.bgw_restart_time = BGW_NEVER_RESTART;
610 sprintf(worker.bgw_library_name, "postgres");
611 sprintf(worker.bgw_function_name, "ParallelWorkerMain");
612 worker.bgw_main_arg = UInt32GetDatum(dsm_segment_handle(pcxt->seg));
613 worker.bgw_notify_pid = MyProcPid;
614
615 /*
616 * Start workers.
617 *
618 * The caller must be able to tolerate ending up with fewer workers than
619 * expected, so there is no need to throw an error here if registration
620 * fails. It wouldn't help much anyway, because registering the worker in
621 * no way guarantees that it will start up and initialize successfully.
622 */
623 for (i = 0; i < pcxt->nworkers_to_launch; ++i)
624 {
625 memcpy(worker.bgw_extra, &i, sizeof(int));
626 if (!any_registrations_failed &&
627 RegisterDynamicBackgroundWorker(&worker,
628 &pcxt->worker[i].bgwhandle))
629 {
630 shm_mq_set_handle(pcxt->worker[i].error_mqh,
631 pcxt->worker[i].bgwhandle);
632 pcxt->nworkers_launched++;
633 }
634 else
635 {
636 /*
637 * If we weren't able to register the worker, then we've bumped up
638 * against the max_worker_processes limit, and future
639 * registrations will probably fail too, so arrange to skip them.
640 * But we still have to execute this code for the remaining slots
641 * to make sure that we forget about the error queues we budgeted
642 * for those workers. Otherwise, we'll wait for them to start,
643 * but they never will.
644 */
645 any_registrations_failed = true;
646 pcxt->worker[i].bgwhandle = NULL;
647 shm_mq_detach(pcxt->worker[i].error_mqh);
648 pcxt->worker[i].error_mqh = NULL;
649 }
650 }
651
652 /*
653 * Now that nworkers_launched has taken its final value, we can initialize
654 * known_attached_workers.
655 */
656 if (pcxt->nworkers_launched > 0)
657 {
658 pcxt->known_attached_workers =
659 palloc0(sizeof(bool) * pcxt->nworkers_launched);
660 pcxt->nknown_attached_workers = 0;
661 }
662
663 /* Restore previous memory context. */
664 MemoryContextSwitchTo(oldcontext);
665 }
666
667 /*
668 * Wait for all workers to attach to their error queues, and throw an error if
669 * any worker fails to do this.
670 *
671 * Callers can assume that if this function returns successfully, then the
672 * number of workers given by pcxt->nworkers_launched have initialized and
673 * attached to their error queues. Whether or not these workers are guaranteed
674 * to still be running depends on what code the caller asked them to run;
675 * this function does not guarantee that they have not exited. However, it
676 * does guarantee that any workers which exited must have done so cleanly and
677 * after successfully performing the work with which they were tasked.
678 *
679 * If this function is not called, then some of the workers that were launched
680 * may not have been started due to a fork() failure, or may have exited during
681 * early startup prior to attaching to the error queue, so nworkers_launched
682 * cannot be viewed as completely reliable. It will never be less than the
683 * number of workers which actually started, but it might be more. Any workers
684 * that failed to start will still be discovered by
685 * WaitForParallelWorkersToFinish and an error will be thrown at that time,
686 * provided that function is eventually reached.
687 *
688 * In general, the leader process should do as much work as possible before
689 * calling this function. fork() failures and other early-startup failures
690 * are very uncommon, and having the leader sit idle when it could be doing
691 * useful work is undesirable. However, if the leader needs to wait for
692 * all of its workers or for a specific worker, it may want to call this
693 * function before doing so. If not, it must make some other provision for
694 * the failure-to-start case, lest it wait forever. On the other hand, a
695 * leader which never waits for a worker that might not be started yet, or
696 * at least never does so prior to WaitForParallelWorkersToFinish(), need not
697 * call this function at all.
698 */
699 void
700 WaitForParallelWorkersToAttach(ParallelContext *pcxt)
701 {
702 int i;
703
704 /* Skip this if we have no launched workers. */
705 if (pcxt->nworkers_launched == 0)
706 return;
707
708 for (;;)
709 {
710 /*
711 * This will process any parallel messages that are pending and it may
712 * also throw an error propagated from a worker.
713 */
714 CHECK_FOR_INTERRUPTS();
715
716 for (i = 0; i < pcxt->nworkers_launched; ++i)
717 {
718 BgwHandleStatus status;
719 shm_mq *mq;
720 int rc;
721 pid_t pid;
722
723 if (pcxt->known_attached_workers[i])
724 continue;
725
726 /*
727 * If error_mqh is NULL, then the worker has already exited
728 * cleanly.
729 */
730 if (pcxt->worker[i].error_mqh == NULL)
731 {
732 pcxt->known_attached_workers[i] = true;
733 ++pcxt->nknown_attached_workers;
734 continue;
735 }
736
737 status = GetBackgroundWorkerPid(pcxt->worker[i].bgwhandle, &pid);
738 if (status == BGWH_STARTED)
739 {
740 /* Has the worker attached to the error queue? */
741 mq = shm_mq_get_queue(pcxt->worker[i].error_mqh);
742 if (shm_mq_get_sender(mq) != NULL)
743 {
744 /* Yes, so it is known to be attached. */
745 pcxt->known_attached_workers[i] = true;
746 ++pcxt->nknown_attached_workers;
747 }
748 }
749 else if (status == BGWH_STOPPED)
750 {
751 /*
752 * If the worker stopped without attaching to the error queue,
753 * throw an error.
754 */
755 mq = shm_mq_get_queue(pcxt->worker[i].error_mqh);
756 if (shm_mq_get_sender(mq) == NULL)
757 ereport(ERROR,
758 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
759 errmsg("parallel worker failed to initialize"),
760 errhint("More details may be available in the server log.")));
761
762 pcxt->known_attached_workers[i] = true;
763 ++pcxt->nknown_attached_workers;
764 }
765 else
766 {
767 /*
768 * Worker not yet started, so we must wait. The postmaster
769 * will notify us if the worker's state changes. Our latch
770 * might also get set for some other reason, but if so we'll
771 * just end up waiting for the same worker again.
772 */
773 rc = WaitLatch(MyLatch,
774 WL_LATCH_SET | WL_EXIT_ON_PM_DEATH,
775 -1, WAIT_EVENT_BGWORKER_STARTUP);
776
777 if (rc & WL_LATCH_SET)
778 ResetLatch(MyLatch);
779 }
780 }
781
782 /* If all workers are known to have started, we're done. */
783 if (pcxt->nknown_attached_workers >= pcxt->nworkers_launched)
784 {
785 Assert(pcxt->nknown_attached_workers == pcxt->nworkers_launched);
786 break;
787 }
788 }
789 }
790
791 /*
792 * Wait for all workers to finish computing.
793 *
794 * Even if the parallel operation seems to have completed successfully, it's
795 * important to call this function afterwards. We must not miss any errors
796 * the workers may have thrown during the parallel operation, or any that they
797 * may yet throw while shutting down.
798 *
799 * Also, we want to update our notion of XactLastRecEnd based on worker
800 * feedback.
801 */
802 void
803 WaitForParallelWorkersToFinish(ParallelContext *pcxt)
804 {
805 for (;;)
806 {
807 bool anyone_alive = false;
808 int nfinished = 0;
809 int i;
810
811 /*
812 * This will process any parallel messages that are pending, which may
813 * change the outcome of the loop that follows. It may also throw an
814 * error propagated from a worker.
815 */
816 CHECK_FOR_INTERRUPTS();
817
818 for (i = 0; i < pcxt->nworkers_launched; ++i)
819 {
820 /*
821 * If error_mqh is NULL, then the worker has already exited
822 * cleanly. If we have received a message through error_mqh from
823 * the worker, we know it started up cleanly, and therefore we're
824 * certain to be notified when it exits.
825 */
826 if (pcxt->worker[i].error_mqh == NULL)
827 ++nfinished;
828 else if (pcxt->known_attached_workers[i])
829 {
830 anyone_alive = true;
831 break;
832 }
833 }
834
835 if (!anyone_alive)
836 {
837 /* If all workers are known to have finished, we're done. */
838 if (nfinished >= pcxt->nworkers_launched)
839 {
840 Assert(nfinished == pcxt->nworkers_launched);
841 break;
842 }
843
844 /*
845 * We didn't detect any living workers, but not all workers are
846 * known to have exited cleanly. Either not all workers have
847 * launched yet, or maybe some of them failed to start or
848 * terminated abnormally.
849 */
850 for (i = 0; i < pcxt->nworkers_launched; ++i)
851 {
852 pid_t pid;
853 shm_mq *mq;
854
855 /*
856 * If the worker is BGWH_NOT_YET_STARTED or BGWH_STARTED, we
857 * should just keep waiting. If it is BGWH_STOPPED, then
858 * further investigation is needed.
859 */
860 if (pcxt->worker[i].error_mqh == NULL ||
861 pcxt->worker[i].bgwhandle == NULL ||
862 GetBackgroundWorkerPid(pcxt->worker[i].bgwhandle,
863 &pid) != BGWH_STOPPED)
864 continue;
865
866 /*
867 * Check whether the worker ended up stopped without ever
868 * attaching to the error queue. If so, the postmaster was
869 * unable to fork the worker or it exited without initializing
870 * properly. We must throw an error, since the caller may
871 * have been expecting the worker to do some work before
872 * exiting.
873 */
874 mq = shm_mq_get_queue(pcxt->worker[i].error_mqh);
875 if (shm_mq_get_sender(mq) == NULL)
876 ereport(ERROR,
877 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
878 errmsg("parallel worker failed to initialize"),
879 errhint("More details may be available in the server log.")));
880
881 /*
882 * The worker is stopped, but is attached to the error queue.
883 * Unless there's a bug somewhere, this will only happen when
884 * the worker writes messages and terminates after the
885 * CHECK_FOR_INTERRUPTS() near the top of this function and
886 * before the call to GetBackgroundWorkerPid(). In that case,
887 * or latch should have been set as well and the right things
888 * will happen on the next pass through the loop.
889 */
890 }
891 }
892
893 (void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, -1,
894 WAIT_EVENT_PARALLEL_FINISH);
895 ResetLatch(MyLatch);
896 }
897
898 if (pcxt->toc != NULL)
899 {
900 FixedParallelState *fps;
901
902 fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED, false);
903 if (fps->last_xlog_end > XactLastRecEnd)
904 XactLastRecEnd = fps->last_xlog_end;
905 }
906 }
907
908 /*
909 * Wait for all workers to exit.
910 *
911 * This function ensures that workers have been completely shutdown. The
912 * difference between WaitForParallelWorkersToFinish and this function is
913 * that the former just ensures that last message sent by a worker backend is
914 * received by the leader backend whereas this ensures the complete shutdown.
915 */
916 static void
917 WaitForParallelWorkersToExit(ParallelContext *pcxt)
918 {
919 int i;
920
921 /* Wait until the workers actually die. */
922 for (i = 0; i < pcxt->nworkers_launched; ++i)
923 {
924 BgwHandleStatus status;
925
926 if (pcxt->worker == NULL || pcxt->worker[i].bgwhandle == NULL)
927 continue;
928
929 status = WaitForBackgroundWorkerShutdown(pcxt->worker[i].bgwhandle);
930
931 /*
932 * If the postmaster kicked the bucket, we have no chance of cleaning
933 * up safely -- we won't be able to tell when our workers are actually
934 * dead. This doesn't necessitate a PANIC since they will all abort
935 * eventually, but we can't safely continue this session.
936 */
937 if (status == BGWH_POSTMASTER_DIED)
938 ereport(FATAL,
939 (errcode(ERRCODE_ADMIN_SHUTDOWN),
940 errmsg("postmaster exited during a parallel transaction")));
941
942 /* Release memory. */
943 pfree(pcxt->worker[i].bgwhandle);
944 pcxt->worker[i].bgwhandle = NULL;
945 }
946 }
947
948 /*
949 * Destroy a parallel context.
950 *
951 * If expecting a clean exit, you should use WaitForParallelWorkersToFinish()
952 * first, before calling this function. When this function is invoked, any
953 * remaining workers are forcibly killed; the dynamic shared memory segment
954 * is unmapped; and we then wait (uninterruptibly) for the workers to exit.
955 */
956 void
957 DestroyParallelContext(ParallelContext *pcxt)
958 {
959 int i;
960
961 /*
962 * Be careful about order of operations here! We remove the parallel
963 * context from the list before we do anything else; otherwise, if an
964 * error occurs during a subsequent step, we might try to nuke it again
965 * from AtEOXact_Parallel or AtEOSubXact_Parallel.
966 */
967 dlist_delete(&pcxt->node);
968
969 /* Kill each worker in turn, and forget their error queues. */
970 if (pcxt->worker != NULL)
971 {
972 for (i = 0; i < pcxt->nworkers_launched; ++i)
973 {
974 if (pcxt->worker[i].error_mqh != NULL)
975 {
976 TerminateBackgroundWorker(pcxt->worker[i].bgwhandle);
977
978 shm_mq_detach(pcxt->worker[i].error_mqh);
979 pcxt->worker[i].error_mqh = NULL;
980 }
981 }
982 }
983
984 /*
985 * If we have allocated a shared memory segment, detach it. This will
986 * implicitly detach the error queues, and any other shared memory queues,
987 * stored there.
988 */
989 if (pcxt->seg != NULL)
990 {
991 dsm_detach(pcxt->seg);
992 pcxt->seg = NULL;
993 }
994
995 /*
996 * If this parallel context is actually in backend-private memory rather
997 * than shared memory, free that memory instead.
998 */
999 if (pcxt->private_memory != NULL)
1000 {
1001 pfree(pcxt->private_memory);
1002 pcxt->private_memory = NULL;
1003 }
1004
1005 /*
1006 * We can't finish transaction commit or abort until all of the workers
1007 * have exited. This means, in particular, that we can't respond to
1008 * interrupts at this stage.
1009 */
1010 HOLD_INTERRUPTS();
1011 WaitForParallelWorkersToExit(pcxt);
1012 RESUME_INTERRUPTS();
1013
1014 /* Free the worker array itself. */
1015 if (pcxt->worker != NULL)
1016 {
1017 pfree(pcxt->worker);
1018 pcxt->worker = NULL;
1019 }
1020
1021 /* Free memory. */
1022 pfree(pcxt->library_name);
1023 pfree(pcxt->function_name);
1024 pfree(pcxt);
1025 }
1026
1027 /*
1028 * Are there any parallel contexts currently active?
1029 */
1030 bool
1031 ParallelContextActive(void)
1032 {
1033 return !dlist_is_empty(&pcxt_list);
1034 }
1035
1036 /*
1037 * Handle receipt of an interrupt indicating a parallel worker message.
1038 *
1039 * Note: this is called within a signal handler! All we can do is set
1040 * a flag that will cause the next CHECK_FOR_INTERRUPTS() to invoke
1041 * ProcessParallelMessages().
1042 */
1043 void
1044 HandleParallelMessageInterrupt(void)
1045 {
1046 InterruptPending = true;
1047 ParallelMessagePending = true;
1048 SetLatch(MyLatch);
1049 }
1050
1051 /*
1052 * Process any queued protocol messages received from parallel workers.
1053 */
1054 void
1055 ProcessParallelMessages(void)
1056 {
1057 dlist_iter iter;
1058 MemoryContext oldcontext;
1059
1060 static MemoryContext hpm_context = NULL;
1061
1062 /*
1063 * This is invoked from ProcessInterrupts(), and since some of the
1064 * functions it calls contain CHECK_FOR_INTERRUPTS(), there is a potential
1065 * for recursive calls if more signals are received while this runs. It's
1066 * unclear that recursive entry would be safe, and it doesn't seem useful
1067 * even if it is safe, so let's block interrupts until done.
1068 */
1069 HOLD_INTERRUPTS();
1070
1071 /*
1072 * Moreover, CurrentMemoryContext might be pointing almost anywhere. We
1073 * don't want to risk leaking data into long-lived contexts, so let's do
1074 * our work here in a private context that we can reset on each use.
1075 */
1076 if (hpm_context == NULL) /* first time through? */
1077 hpm_context = AllocSetContextCreate(TopMemoryContext,
1078 "ProcessParallelMessages",
1079 ALLOCSET_DEFAULT_SIZES);
1080 else
1081 MemoryContextReset(hpm_context);
1082
1083 oldcontext = MemoryContextSwitchTo(hpm_context);
1084
1085 /* OK to process messages. Reset the flag saying there are more to do. */
1086 ParallelMessagePending = false;
1087
1088 dlist_foreach(iter, &pcxt_list)
1089 {
1090 ParallelContext *pcxt;
1091 int i;
1092
1093 pcxt = dlist_container(ParallelContext, node, iter.cur);
1094 if (pcxt->worker == NULL)
1095 continue;
1096
1097 for (i = 0; i < pcxt->nworkers_launched; ++i)
1098 {
1099 /*
1100 * Read as many messages as we can from each worker, but stop when
1101 * either (1) the worker's error queue goes away, which can happen
1102 * if we receive a Terminate message from the worker; or (2) no
1103 * more messages can be read from the worker without blocking.
1104 */
1105 while (pcxt->worker[i].error_mqh != NULL)
1106 {
1107 shm_mq_result res;
1108 Size nbytes;
1109 void *data;
1110
1111 res = shm_mq_receive(pcxt->worker[i].error_mqh, &nbytes,
1112 &data, true);
1113 if (res == SHM_MQ_WOULD_BLOCK)
1114 break;
1115 else if (res == SHM_MQ_SUCCESS)
1116 {
1117 StringInfoData msg;
1118
1119 initStringInfo(&msg);
1120 appendBinaryStringInfo(&msg, data, nbytes);
1121 ProcessParallelMessage(pcxt, i, &msg);
1122 pfree(msg.data);
1123 }
1124 else
1125 ereport(ERROR,
1126 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1127 errmsg("lost connection to parallel worker")));
1128 }
1129 }
1130 }
1131
1132 MemoryContextSwitchTo(oldcontext);
1133
1134 /* Might as well clear the context on our way out */
1135 MemoryContextReset(hpm_context);
1136
1137 RESUME_INTERRUPTS();
1138 }
1139
1140 /*
1141 * Process a single protocol message received from a single parallel worker.
1142 */
1143 static void
1144 ProcessParallelMessage(ParallelContext *pcxt, int i, StringInfo msg)
1145 {
1146 char msgtype;
1147
1148 if (pcxt->known_attached_workers != NULL &&
1149 !pcxt->known_attached_workers[i])
1150 {
1151 pcxt->known_attached_workers[i] = true;
1152 pcxt->nknown_attached_workers++;
1153 }
1154
1155 msgtype = pq_getmsgbyte(msg);
1156
1157 switch (msgtype)
1158 {
1159 case PqMsg_ErrorResponse:
1160 case PqMsg_NoticeResponse:
1161 {
1162 ErrorData edata;
1163 ErrorContextCallback *save_error_context_stack;
1164
1165 /* Parse ErrorResponse or NoticeResponse. */
1166 pq_parse_errornotice(msg, &edata);
1167
1168 /* Death of a worker isn't enough justification for suicide. */
1169 edata.elevel = Min(edata.elevel, ERROR);
1170
1171 /*
1172 * If desired, add a context line to show that this is a
1173 * message propagated from a parallel worker. Otherwise, it
1174 * can sometimes be confusing to understand what actually
1175 * happened. (We don't do this in DEBUG_PARALLEL_REGRESS mode
1176 * because it causes test-result instability depending on
1177 * whether a parallel worker is actually used or not.)
1178 */
1179 if (debug_parallel_query != DEBUG_PARALLEL_REGRESS)
1180 {
1181 if (edata.context)
1182 edata.context = psprintf("%s\n%s", edata.context,
1183 _("parallel worker"));
1184 else
1185 edata.context = pstrdup(_("parallel worker"));
1186 }
1187
1188 /*
1189 * Context beyond that should use the error context callbacks
1190 * that were in effect when the ParallelContext was created,
1191 * not the current ones.
1192 */
1193 save_error_context_stack = error_context_stack;
1194 error_context_stack = pcxt->error_context_stack;
1195
1196 /* Rethrow error or print notice. */
1197 ThrowErrorData(&edata);
1198
1199 /* Not an error, so restore previous context stack. */
1200 error_context_stack = save_error_context_stack;
1201
1202 break;
1203 }
1204
1205 case PqMsg_NotificationResponse:
1206 {
1207 /* Propagate NotifyResponse. */
1208 int32 pid;
1209 const char *channel;
1210 const char *payload;
1211
1212 pid = pq_getmsgint(msg, 4);
1213 channel = pq_getmsgrawstring(msg);
1214 payload = pq_getmsgrawstring(msg);
1215 pq_endmessage(msg);
1216
1217 NotifyMyFrontEnd(channel, payload, pid);
1218
1219 break;
1220 }
1221
1222 case PqMsg_Progress:
1223 {
1224 /*
1225 * Only incremental progress reporting is currently supported.
1226 * However, it's possible to add more fields to the message to
1227 * allow for handling of other backend progress APIs.
1228 */
1229 int index = pq_getmsgint(msg, 4);
1230 int64 incr = pq_getmsgint64(msg);
1231
1232 pq_getmsgend(msg);
1233
1234 pgstat_progress_incr_param(index, incr);
1235
1236 break;
1237 }
1238
1239 case PqMsg_Terminate:
1240 {
1241 shm_mq_detach(pcxt->worker[i].error_mqh);
1242 pcxt->worker[i].error_mqh = NULL;
1243 break;
1244 }
1245
1246 default:
1247 {
1248 elog(ERROR, "unrecognized message type received from parallel worker: %c (message length %d bytes)",
1249 msgtype, msg->len);
1250 }
1251 }
1252 }
1253
1254 /*
1255 * End-of-subtransaction cleanup for parallel contexts.
1256 *
1257 * Here we remove only parallel contexts initiated within the current
1258 * subtransaction.
1259 */
1260 void
1261 AtEOSubXact_Parallel(bool isCommit, SubTransactionId mySubId)
1262 {
1263 while (!dlist_is_empty(&pcxt_list))
1264 {
1265 ParallelContext *pcxt;
1266
1267 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
1268 if (pcxt->subid != mySubId)
1269 break;
1270 if (isCommit)
1271 elog(WARNING, "leaked parallel context");
1272 DestroyParallelContext(pcxt);
1273 }
1274 }
1275
1276 /*
1277 * End-of-transaction cleanup for parallel contexts.
1278 *
1279 * We nuke all remaining parallel contexts.
1280 */
1281 void
1282 AtEOXact_Parallel(bool isCommit)
1283 {
1284 while (!dlist_is_empty(&pcxt_list))
1285 {
1286 ParallelContext *pcxt;
1287
1288 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
1289 if (isCommit)
1290 elog(WARNING, "leaked parallel context");
1291 DestroyParallelContext(pcxt);
1292 }
1293 }
1294
1295 /*
1296 * Main entrypoint for parallel workers.
1297 */
1298 void
1299 ParallelWorkerMain(Datum main_arg)
1300 {
1301 dsm_segment *seg;
1302 shm_toc *toc;
1303 FixedParallelState *fps;
1304 char *error_queue_space;
1305 shm_mq *mq;
1306 shm_mq_handle *mqh;
1307 char *libraryspace;
1308 char *entrypointstate;
1309 char *library_name;
1310 char *function_name;
1311 parallel_worker_main_type entrypt;
1312 char *gucspace;
1313 char *combocidspace;
1314 char *tsnapspace;
1315 char *asnapspace;
1316 char *tstatespace;
1317 char *pendingsyncsspace;
1318 char *reindexspace;
1319 char *relmapperspace;
1320 char *uncommittedenumsspace;
1321 char *clientconninfospace;
1322 char *session_dsm_handle_space;
1323 Snapshot tsnapshot;
1324 Snapshot asnapshot;
1325
1326 /* Set flag to indicate that we're initializing a parallel worker. */
1327 InitializingParallelWorker = true;
1328
1329 /* Establish signal handlers. */
1330 pqsignal(SIGTERM, die);
1331 BackgroundWorkerUnblockSignals();
1332
1333 /* Determine and set our parallel worker number. */
1334 Assert(ParallelWorkerNumber == -1);
1335 memcpy(&ParallelWorkerNumber, MyBgworkerEntry->bgw_extra, sizeof(int));
1336
1337 /* Set up a memory context to work in, just for cleanliness. */
1338 CurrentMemoryContext = AllocSetContextCreate(TopMemoryContext,
1339 "Parallel worker",
1340 ALLOCSET_DEFAULT_SIZES);
1341
1342 /*
1343 * Attach to the dynamic shared memory segment for the parallel query, and
1344 * find its table of contents.
1345 *
1346 * Note: at this point, we have not created any ResourceOwner in this
1347 * process. This will result in our DSM mapping surviving until process
1348 * exit, which is fine. If there were a ResourceOwner, it would acquire
1349 * ownership of the mapping, but we have no need for that.
1350 */
1351 seg = dsm_attach(DatumGetUInt32(main_arg));
1352 if (seg == NULL)
1353 ereport(ERROR,
1354 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1355 errmsg("could not map dynamic shared memory segment")));
1356 toc = shm_toc_attach(PARALLEL_MAGIC, dsm_segment_address(seg));
1357 if (toc == NULL)
1358 ereport(ERROR,
1359 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1360 errmsg("invalid magic number in dynamic shared memory segment")));
1361
1362 /* Look up fixed parallel state. */
1363 fps = shm_toc_lookup(toc, PARALLEL_KEY_FIXED, false);
1364 MyFixedParallelState = fps;
1365
1366 /* Arrange to signal the leader if we exit. */
1367 ParallelLeaderPid = fps->parallel_leader_pid;
1368 ParallelLeaderProcNumber = fps->parallel_leader_proc_number;
1369 before_shmem_exit(ParallelWorkerShutdown, PointerGetDatum(seg));
1370
1371 /*
1372 * Now we can find and attach to the error queue provided for us. That's
1373 * good, because until we do that, any errors that happen here will not be
1374 * reported back to the process that requested that this worker be
1375 * launched.
1376 */
1377 error_queue_space = shm_toc_lookup(toc, PARALLEL_KEY_ERROR_QUEUE, false);
1378 mq = (shm_mq *) (error_queue_space +
1379 ParallelWorkerNumber * PARALLEL_ERROR_QUEUE_SIZE);
1380 shm_mq_set_sender(mq, MyProc);
1381 mqh = shm_mq_attach(mq, seg, NULL);
1382 pq_redirect_to_shm_mq(seg, mqh);
1383 pq_set_parallel_leader(fps->parallel_leader_pid,
1384 fps->parallel_leader_proc_number);
1385
1386 /*
1387 * Hooray! Primary initialization is complete. Now, we need to set up our
1388 * backend-local state to match the original backend.
1389 */
1390
1391 /*
1392 * Join locking group. We must do this before anything that could try to
1393 * acquire a heavyweight lock, because any heavyweight locks acquired to
1394 * this point could block either directly against the parallel group
1395 * leader or against some process which in turn waits for a lock that
1396 * conflicts with the parallel group leader, causing an undetected
1397 * deadlock. (If we can't join the lock group, the leader has gone away,
1398 * so just exit quietly.)
1399 */
1400 if (!BecomeLockGroupMember(fps->parallel_leader_pgproc,
1401 fps->parallel_leader_pid))
1402 return;
1403
1404 /*
1405 * Restore transaction and statement start-time timestamps. This must
1406 * happen before anything that would start a transaction, else asserts in
1407 * xact.c will fire.
1408 */
1409 SetParallelStartTimestamps(fps->xact_ts, fps->stmt_ts);
1410
1411 /*
1412 * Identify the entry point to be called. In theory this could result in
1413 * loading an additional library, though most likely the entry point is in
1414 * the core backend or in a library we just loaded.
1415 */
1416 entrypointstate = shm_toc_lookup(toc, PARALLEL_KEY_ENTRYPOINT, false);
1417 library_name = entrypointstate;
1418 function_name = entrypointstate + strlen(library_name) + 1;
1419
1420 entrypt = LookupParallelWorkerFunction(library_name, function_name);
1421
1422 /*
1423 * Restore current session authorization and role id. No verification
1424 * happens here, we just blindly adopt the leader's state. Note that this
1425 * has to happen before InitPostgres, since InitializeSessionUserId will
1426 * not set these variables.
1427 */
1428 SetAuthenticatedUserId(fps->authenticated_user_id);
1429 SetSessionAuthorization(fps->session_user_id,
1430 fps->session_user_is_superuser);
1431 SetCurrentRoleId(fps->outer_user_id, fps->role_is_superuser);
1432
1433 /*
1434 * Restore database connection. We skip connection authorization checks,
1435 * reasoning that (a) the leader checked these things when it started, and
1436 * (b) we do not want parallel mode to cause these failures, because that
1437 * would make use of parallel query plans not transparent to applications.
1438 */
1439 BackgroundWorkerInitializeConnectionByOid(fps->database_id,
1440 fps->authenticated_user_id,
1441 BGWORKER_BYPASS_ALLOWCONN |
1442 BGWORKER_BYPASS_ROLELOGINCHECK);
1443
1444 /*
1445 * Set the client encoding to the database encoding, since that is what
1446 * the leader will expect. (We're cheating a bit by not calling
1447 * PrepareClientEncoding first. It's okay because this call will always
1448 * result in installing a no-op conversion. No error should be possible,
1449 * but check anyway.)
1450 */
1451 if (SetClientEncoding(GetDatabaseEncoding()) < 0)
1452 elog(ERROR, "SetClientEncoding(%d) failed", GetDatabaseEncoding());
1453
1454 /*
1455 * Load libraries that were loaded by original backend. We want to do
1456 * this before restoring GUCs, because the libraries might define custom
1457 * variables.
1458 */
1459 libraryspace = shm_toc_lookup(toc, PARALLEL_KEY_LIBRARY, false);
1460 StartTransactionCommand();
1461 RestoreLibraryState(libraryspace);
1462 CommitTransactionCommand();
1463
1464 /* Crank up a transaction state appropriate to a parallel worker. */
1465 tstatespace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_STATE, false);
1466 StartParallelWorkerTransaction(tstatespace);
1467
1468 /*
1469 * Restore state that affects catalog access. Ideally we'd do this even
1470 * before calling InitPostgres, but that has order-of-initialization
1471 * problems, and also the relmapper would get confused during the
1472 * CommitTransactionCommand call above.
1473 */
1474 pendingsyncsspace = shm_toc_lookup(toc, PARALLEL_KEY_PENDING_SYNCS,
1475 false);
1476 RestorePendingSyncs(pendingsyncsspace);
1477 relmapperspace = shm_toc_lookup(toc, PARALLEL_KEY_RELMAPPER_STATE, false);
1478 RestoreRelationMap(relmapperspace);
1479 reindexspace = shm_toc_lookup(toc, PARALLEL_KEY_REINDEX_STATE, false);
1480 RestoreReindexState(reindexspace);
1481 combocidspace = shm_toc_lookup(toc, PARALLEL_KEY_COMBO_CID, false);
1482 RestoreComboCIDState(combocidspace);
1483
1484 /* Attach to the per-session DSM segment and contained objects. */
1485 session_dsm_handle_space =
1486 shm_toc_lookup(toc, PARALLEL_KEY_SESSION_DSM, false);
1487 AttachSession(*(dsm_handle *) session_dsm_handle_space);
1488
1489 /*
1490 * If the transaction isolation level is REPEATABLE READ or SERIALIZABLE,
1491 * the leader has serialized the transaction snapshot and we must restore
1492 * it. At lower isolation levels, there is no transaction-lifetime
1493 * snapshot, but we need TransactionXmin to get set to a value which is
1494 * less than or equal to the xmin of every snapshot that will be used by
1495 * this worker. The easiest way to accomplish that is to install the
1496 * active snapshot as the transaction snapshot. Code running in this
1497 * parallel worker might take new snapshots via GetTransactionSnapshot()
1498 * or GetLatestSnapshot(), but it shouldn't have any way of acquiring a
1499 * snapshot older than the active snapshot.
1500 */
1501 asnapspace = shm_toc_lookup(toc, PARALLEL_KEY_ACTIVE_SNAPSHOT, false);
1502 tsnapspace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT, true);
1503 asnapshot = RestoreSnapshot(asnapspace);
1504 tsnapshot = tsnapspace ? RestoreSnapshot(tsnapspace) : asnapshot;
1505 RestoreTransactionSnapshot(tsnapshot,
1506 fps->parallel_leader_pgproc);
1507 PushActiveSnapshot(asnapshot);
1508
1509 /*
1510 * We've changed which tuples we can see, and must therefore invalidate
1511 * system caches.
1512 */
1513 InvalidateSystemCaches();
1514
1515 /*
1516 * Restore GUC values from launching backend. We can't do this earlier,
1517 * because GUC check hooks that do catalog lookups need to see the same
1518 * database state as the leader. Also, the check hooks for
1519 * session_authorization and role assume we already set the correct role
1520 * OIDs.
1521 */
1522 gucspace = shm_toc_lookup(toc, PARALLEL_KEY_GUC, false);
1523 RestoreGUCState(gucspace);
1524
1525 /*
1526 * Restore current user ID and security context. No verification happens
1527 * here, we just blindly adopt the leader's state. We can't do this till
1528 * after restoring GUCs, else we'll get complaints about restoring
1529 * session_authorization and role. (In effect, we're assuming that all
1530 * the restored values are okay to set, even if we are now inside a
1531 * restricted context.)
1532 */
1533 SetUserIdAndSecContext(fps->current_user_id, fps->sec_context);
1534
1535 /* Restore temp-namespace state to ensure search path matches leader's. */
1536 SetTempNamespaceState(fps->temp_namespace_id,
1537 fps->temp_toast_namespace_id);
1538
1539 /* Restore uncommitted enums. */
1540 uncommittedenumsspace = shm_toc_lookup(toc, PARALLEL_KEY_UNCOMMITTEDENUMS,
1541 false);
1542 RestoreUncommittedEnums(uncommittedenumsspace);
1543
1544 /* Restore the ClientConnectionInfo. */
1545 clientconninfospace = shm_toc_lookup(toc, PARALLEL_KEY_CLIENTCONNINFO,
1546 false);
1547 RestoreClientConnectionInfo(clientconninfospace);
1548
1549 /*
1550 * Initialize SystemUser now that MyClientConnectionInfo is restored. Also
1551 * ensure that auth_method is actually valid, aka authn_id is not NULL.
1552 */
1553 if (MyClientConnectionInfo.authn_id)
1554 InitializeSystemUser(MyClientConnectionInfo.authn_id,
1555 hba_authname(MyClientConnectionInfo.auth_method));
1556
1557 /* Attach to the leader's serializable transaction, if SERIALIZABLE. */
1558 AttachSerializableXact(fps->serializable_xact_handle);
1559
1560 /*
1561 * We've initialized all of our state now; nothing should change
1562 * hereafter.
1563 */
1564 InitializingParallelWorker = false;
1565 EnterParallelMode();
1566
1567 /*
1568 * Time to do the real work: invoke the caller-supplied code.
1569 */
1570 entrypt(seg, toc);
1571
1572 /* Must exit parallel mode to pop active snapshot. */
1573 ExitParallelMode();
1574
1575 /* Must pop active snapshot so snapmgr.c doesn't complain. */
1576 PopActiveSnapshot();
1577
1578 /* Shut down the parallel-worker transaction. */
1579 EndParallelWorkerTransaction();
1580
1581 /* Detach from the per-session DSM segment. */
1582 DetachSession();
1583
1584 /* Report success. */
1585 pq_putmessage(PqMsg_Terminate, NULL, 0);
1586 }
1587
1588 /*
1589 * Update shared memory with the ending location of the last WAL record we
1590 * wrote, if it's greater than the value already stored there.
1591 */
1592 void
1593 ParallelWorkerReportLastRecEnd(XLogRecPtr last_xlog_end)
1594 {
1595 FixedParallelState *fps = MyFixedParallelState;
1596
1597 Assert(fps != NULL);
1598 SpinLockAcquire(&fps->mutex);
1599 if (fps->last_xlog_end < last_xlog_end)
1600 fps->last_xlog_end = last_xlog_end;
1601 SpinLockRelease(&fps->mutex);
1602 }
1603
1604 /*
1605 * Make sure the leader tries to read from our error queue one more time.
1606 * This guards against the case where we exit uncleanly without sending an
1607 * ErrorResponse to the leader, for example because some code calls proc_exit
1608 * directly.
1609 *
1610 * Also explicitly detach from dsm segment so that subsystems using
1611 * on_dsm_detach() have a chance to send stats before the stats subsystem is
1612 * shut down as part of a before_shmem_exit() hook.
1613 *
1614 * One might think this could instead be solved by carefully ordering the
1615 * attaching to dsm segments, so that the pgstats segments get detached from
1616 * later than the parallel query one. That turns out to not work because the
1617 * stats hash might need to grow which can cause new segments to be allocated,
1618 * which then will be detached from earlier.
1619 */
1620 static void
1621 ParallelWorkerShutdown(int code, Datum arg)
1622 {
1623 SendProcSignal(ParallelLeaderPid,
1624 PROCSIG_PARALLEL_MESSAGE,
1625 ParallelLeaderProcNumber);
1626
1627 dsm_detach((dsm_segment *) DatumGetPointer(arg));
1628 }
1629
1630 /*
1631 * Look up (and possibly load) a parallel worker entry point function.
1632 *
1633 * For functions contained in the core code, we use library name "postgres"
1634 * and consult the InternalParallelWorkers array. External functions are
1635 * looked up, and loaded if necessary, using load_external_function().
1636 *
1637 * The point of this is to pass function names as strings across process
1638 * boundaries. We can't pass actual function addresses because of the
1639 * possibility that the function has been loaded at a different address
1640 * in a different process. This is obviously a hazard for functions in
1641 * loadable libraries, but it can happen even for functions in the core code
1642 * on platforms using EXEC_BACKEND (e.g., Windows).
1643 *
1644 * At some point it might be worthwhile to get rid of InternalParallelWorkers[]
1645 * in favor of applying load_external_function() for core functions too;
1646 * but that raises portability issues that are not worth addressing now.
1647 */
1648 static parallel_worker_main_type
1649 LookupParallelWorkerFunction(const char *libraryname, const char *funcname)
1650 {
1651 /*
1652 * If the function is to be loaded from postgres itself, search the
1653 * InternalParallelWorkers array.
1654 */
1655 if (strcmp(libraryname, "postgres") == 0)
1656 {
1657 int i;
1658
1659 for (i = 0; i < lengthof(InternalParallelWorkers); i++)
1660 {
1661 if (strcmp(InternalParallelWorkers[i].fn_name, funcname) == 0)
1662 return InternalParallelWorkers[i].fn_addr;
1663 }
1664
1665 /* We can only reach this by programming error. */
1666 elog(ERROR, "internal function \"%s\" not found", funcname);
1667 }
1668
1669 /* Otherwise load from external library. */
1670 return (parallel_worker_main_type)
1671 load_external_function(libraryname, funcname, true, NULL);
1672 }