?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Allow "internal" subtransactions in parallel mode.
[pgsql.git] / src / backend / access / transam / parallel.c
blob8613fc6fb547c05d30d01a44282dba2b3f2178a3
1 /*-------------------------------------------------------------------------
2 *
3 * parallel.c
4 * Infrastructure for launching parallel workers
5 *
6 * Portions Copyright (c) 1996-2024, 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/nbtree.h"
19 #include "access/parallel.h"
20 #include "access/session.h"
21 #include "access/xact.h"
22 #include "access/xlog.h"
23 #include "catalog/index.h"
24 #include "catalog/namespace.h"
25 #include "catalog/pg_enum.h"
26 #include "catalog/storage.h"
27 #include "commands/async.h"
28 #include "commands/vacuum.h"
29 #include "executor/execParallel.h"
30 #include "libpq/libpq.h"
31 #include "libpq/pqformat.h"
32 #include "libpq/pqmq.h"
33 #include "miscadmin.h"
34 #include "optimizer/optimizer.h"
35 #include "pgstat.h"
36 #include "storage/ipc.h"
37 #include "storage/predicate.h"
38 #include "storage/spin.h"
39 #include "tcop/tcopprot.h"
40 #include "utils/combocid.h"
41 #include "utils/guc.h"
42 #include "utils/inval.h"
43 #include "utils/memutils.h"
44 #include "utils/relmapper.h"
45 #include "utils/snapmgr.h"
46
47 /*
48 * We don't want to waste a lot of memory on an error queue which, most of
49 * the time, will process only a handful of small messages. However, it is
50 * desirable to make it large enough that a typical ErrorResponse can be sent
51 * without blocking. That way, a worker that errors out can write the whole
52 * message into the queue and terminate without waiting for the user backend.
53 */
54 #define PARALLEL_ERROR_QUEUE_SIZE 16384
55
56 /* Magic number for parallel context TOC. */
57 #define PARALLEL_MAGIC 0x50477c7c
58
59 /*
60 * Magic numbers for per-context parallel state sharing. Higher-level code
61 * should use smaller values, leaving these very large ones for use by this
62 * module.
63 */
64 #define PARALLEL_KEY_FIXED UINT64CONST(0xFFFFFFFFFFFF0001)
65 #define PARALLEL_KEY_ERROR_QUEUE UINT64CONST(0xFFFFFFFFFFFF0002)
66 #define PARALLEL_KEY_LIBRARY UINT64CONST(0xFFFFFFFFFFFF0003)
67 #define PARALLEL_KEY_GUC UINT64CONST(0xFFFFFFFFFFFF0004)
68 #define PARALLEL_KEY_COMBO_CID UINT64CONST(0xFFFFFFFFFFFF0005)
69 #define PARALLEL_KEY_TRANSACTION_SNAPSHOT UINT64CONST(0xFFFFFFFFFFFF0006)
70 #define PARALLEL_KEY_ACTIVE_SNAPSHOT UINT64CONST(0xFFFFFFFFFFFF0007)
71 #define PARALLEL_KEY_TRANSACTION_STATE UINT64CONST(0xFFFFFFFFFFFF0008)
72 #define PARALLEL_KEY_ENTRYPOINT UINT64CONST(0xFFFFFFFFFFFF0009)
73 #define PARALLEL_KEY_SESSION_DSM UINT64CONST(0xFFFFFFFFFFFF000A)
74 #define PARALLEL_KEY_PENDING_SYNCS UINT64CONST(0xFFFFFFFFFFFF000B)
75 #define PARALLEL_KEY_REINDEX_STATE UINT64CONST(0xFFFFFFFFFFFF000C)
76 #define PARALLEL_KEY_RELMAPPER_STATE UINT64CONST(0xFFFFFFFFFFFF000D)
77 #define PARALLEL_KEY_UNCOMMITTEDENUMS UINT64CONST(0xFFFFFFFFFFFF000E)
78 #define PARALLEL_KEY_CLIENTCONNINFO UINT64CONST(0xFFFFFFFFFFFF000F)
79
80 /* Fixed-size parallel state. */
81 typedef struct FixedParallelState
82 {
83 /* Fixed-size state that workers must restore. */
84 Oid database_id;
85 Oid authenticated_user_id;
86 Oid current_user_id;
87 Oid outer_user_id;
88 Oid temp_namespace_id;
89 Oid temp_toast_namespace_id;
90 int sec_context;
91 bool is_superuser;
92 PGPROC *parallel_leader_pgproc;
93 pid_t parallel_leader_pid;
94 ProcNumber parallel_leader_proc_number;
95 TimestampTz xact_ts;
96 TimestampTz stmt_ts;
97 SerializableXactHandle serializable_xact_handle;
98
99 /* Mutex protects remaining fields. */
100 slock_t mutex;
101
102 /* Maximum XactLastRecEnd of any worker. */
103 XLogRecPtr last_xlog_end;
104 } FixedParallelState;
105
106 /*
107 * Our parallel worker number. We initialize this to -1, meaning that we are
108 * not a parallel worker. In parallel workers, it will be set to a value >= 0
109 * and < the number of workers before any user code is invoked; each parallel
110 * worker will get a different parallel worker number.
111 */
112 int ParallelWorkerNumber = -1;
113
114 /* Is there a parallel message pending which we need to receive? */
115 volatile sig_atomic_t ParallelMessagePending = false;
116
117 /* Are we initializing a parallel worker? */
118 bool InitializingParallelWorker = false;
119
120 /* Pointer to our fixed parallel state. */
121 static FixedParallelState *MyFixedParallelState;
122
123 /* List of active parallel contexts. */
124 static dlist_head pcxt_list = DLIST_STATIC_INIT(pcxt_list);
125
126 /* Backend-local copy of data from FixedParallelState. */
127 static pid_t ParallelLeaderPid;
128
129 /*
130 * List of internal parallel worker entry points. We need this for
131 * reasons explained in LookupParallelWorkerFunction(), below.
132 */
133 static const struct
134 {
135 const char *fn_name;
136 parallel_worker_main_type fn_addr;
137 } InternalParallelWorkers[] =
138
139 {
140 {
141 "ParallelQueryMain", ParallelQueryMain
142 },
143 {
144 "_bt_parallel_build_main", _bt_parallel_build_main
145 },
146 {
147 "_brin_parallel_build_main", _brin_parallel_build_main
148 },
149 {
150 "parallel_vacuum_main", parallel_vacuum_main
151 }
152 };
153
154 /* Private functions. */
155 static void HandleParallelMessage(ParallelContext *pcxt, int i, StringInfo msg);
156 static void WaitForParallelWorkersToExit(ParallelContext *pcxt);
157 static parallel_worker_main_type LookupParallelWorkerFunction(const char *libraryname, const char *funcname);
158 static void ParallelWorkerShutdown(int code, Datum arg);
159
160
161 /*
162 * Establish a new parallel context. This should be done after entering
163 * parallel mode, and (unless there is an error) the context should be
164 * destroyed before exiting the current subtransaction.
165 */
166 ParallelContext *
167 CreateParallelContext(const char *library_name, const char *function_name,
168 int nworkers)
169 {
170 MemoryContext oldcontext;
171 ParallelContext *pcxt;
172
173 /* It is unsafe to create a parallel context if not in parallel mode. */
174 Assert(IsInParallelMode());
175
176 /* Number of workers should be non-negative. */
177 Assert(nworkers >= 0);
178
179 /* We might be running in a short-lived memory context. */
180 oldcontext = MemoryContextSwitchTo(TopTransactionContext);
181
182 /* Initialize a new ParallelContext. */
183 pcxt = palloc0(sizeof(ParallelContext));
184 pcxt->subid = GetCurrentSubTransactionId();
185 pcxt->nworkers = nworkers;
186 pcxt->nworkers_to_launch = nworkers;
187 pcxt->library_name = pstrdup(library_name);
188 pcxt->function_name = pstrdup(function_name);
189 pcxt->error_context_stack = error_context_stack;
190 shm_toc_initialize_estimator(&pcxt->estimator);
191 dlist_push_head(&pcxt_list, &pcxt->node);
192
193 /* Restore previous memory context. */
194 MemoryContextSwitchTo(oldcontext);
195
196 return pcxt;
197 }
198
199 /*
200 * Establish the dynamic shared memory segment for a parallel context and
201 * copy state and other bookkeeping information that will be needed by
202 * parallel workers into it.
203 */
204 void
205 InitializeParallelDSM(ParallelContext *pcxt)
206 {
207 MemoryContext oldcontext;
208 Size library_len = 0;
209 Size guc_len = 0;
210 Size combocidlen = 0;
211 Size tsnaplen = 0;
212 Size asnaplen = 0;
213 Size tstatelen = 0;
214 Size pendingsyncslen = 0;
215 Size reindexlen = 0;
216 Size relmapperlen = 0;
217 Size uncommittedenumslen = 0;
218 Size clientconninfolen = 0;
219 Size segsize = 0;
220 int i;
221 FixedParallelState *fps;
222 dsm_handle session_dsm_handle = DSM_HANDLE_INVALID;
223 Snapshot transaction_snapshot = GetTransactionSnapshot();
224 Snapshot active_snapshot = GetActiveSnapshot();
225
226 /* We might be running in a very short-lived memory context. */
227 oldcontext = MemoryContextSwitchTo(TopTransactionContext);
228
229 /* Allow space to store the fixed-size parallel state. */
230 shm_toc_estimate_chunk(&pcxt->estimator, sizeof(FixedParallelState));
231 shm_toc_estimate_keys(&pcxt->estimator, 1);
232
233 /*
234 * Normally, the user will have requested at least one worker process, but
235 * if by chance they have not, we can skip a bunch of things here.
236 */
237 if (pcxt->nworkers > 0)
238 {
239 /* Get (or create) the per-session DSM segment's handle. */
240 session_dsm_handle = GetSessionDsmHandle();
241
242 /*
243 * If we weren't able to create a per-session DSM segment, then we can
244 * continue but we can't safely launch any workers because their
245 * record typmods would be incompatible so they couldn't exchange
246 * tuples.
247 */
248 if (session_dsm_handle == DSM_HANDLE_INVALID)
249 pcxt->nworkers = 0;
250 }
251
252 if (pcxt->nworkers > 0)
253 {
254 /* Estimate space for various kinds of state sharing. */
255 library_len = EstimateLibraryStateSpace();
256 shm_toc_estimate_chunk(&pcxt->estimator, library_len);
257 guc_len = EstimateGUCStateSpace();
258 shm_toc_estimate_chunk(&pcxt->estimator, guc_len);
259 combocidlen = EstimateComboCIDStateSpace();
260 shm_toc_estimate_chunk(&pcxt->estimator, combocidlen);
261 if (IsolationUsesXactSnapshot())
262 {
263 tsnaplen = EstimateSnapshotSpace(transaction_snapshot);
264 shm_toc_estimate_chunk(&pcxt->estimator, tsnaplen);
265 }
266 asnaplen = EstimateSnapshotSpace(active_snapshot);
267 shm_toc_estimate_chunk(&pcxt->estimator, asnaplen);
268 tstatelen = EstimateTransactionStateSpace();
269 shm_toc_estimate_chunk(&pcxt->estimator, tstatelen);
270 shm_toc_estimate_chunk(&pcxt->estimator, sizeof(dsm_handle));
271 pendingsyncslen = EstimatePendingSyncsSpace();
272 shm_toc_estimate_chunk(&pcxt->estimator, pendingsyncslen);
273 reindexlen = EstimateReindexStateSpace();
274 shm_toc_estimate_chunk(&pcxt->estimator, reindexlen);
275 relmapperlen = EstimateRelationMapSpace();
276 shm_toc_estimate_chunk(&pcxt->estimator, relmapperlen);
277 uncommittedenumslen = EstimateUncommittedEnumsSpace();
278 shm_toc_estimate_chunk(&pcxt->estimator, uncommittedenumslen);
279 clientconninfolen = EstimateClientConnectionInfoSpace();
280 shm_toc_estimate_chunk(&pcxt->estimator, clientconninfolen);
281 /* If you add more chunks here, you probably need to add keys. */
282 shm_toc_estimate_keys(&pcxt->estimator, 12);
283
284 /* Estimate space need for error queues. */
285 StaticAssertStmt(BUFFERALIGN(PARALLEL_ERROR_QUEUE_SIZE) ==
286 PARALLEL_ERROR_QUEUE_SIZE,
287 "parallel error queue size not buffer-aligned");
288 shm_toc_estimate_chunk(&pcxt->estimator,
289 mul_size(PARALLEL_ERROR_QUEUE_SIZE,
290 pcxt->nworkers));
291 shm_toc_estimate_keys(&pcxt->estimator, 1);
292
293 /* Estimate how much we'll need for the entrypoint info. */
294 shm_toc_estimate_chunk(&pcxt->estimator, strlen(pcxt->library_name) +
295 strlen(pcxt->function_name) + 2);
296 shm_toc_estimate_keys(&pcxt->estimator, 1);
297 }
298
299 /*
300 * Create DSM and initialize with new table of contents. But if the user
301 * didn't request any workers, then don't bother creating a dynamic shared
302 * memory segment; instead, just use backend-private memory.
303 *
304 * Also, if we can't create a dynamic shared memory segment because the
305 * maximum number of segments have already been created, then fall back to
306 * backend-private memory, and plan not to use any workers. We hope this
307 * won't happen very often, but it's better to abandon the use of
308 * parallelism than to fail outright.
309 */
310 segsize = shm_toc_estimate(&pcxt->estimator);
311 if (pcxt->nworkers > 0)
312 pcxt->seg = dsm_create(segsize, DSM_CREATE_NULL_IF_MAXSEGMENTS);
313 if (pcxt->seg != NULL)
314 pcxt->toc = shm_toc_create(PARALLEL_MAGIC,
315 dsm_segment_address(pcxt->seg),
316 segsize);
317 else
318 {
319 pcxt->nworkers = 0;
320 pcxt->private_memory = MemoryContextAlloc(TopMemoryContext, segsize);
321 pcxt->toc = shm_toc_create(PARALLEL_MAGIC, pcxt->private_memory,
322 segsize);
323 }
324
325 /* Initialize fixed-size state in shared memory. */
326 fps = (FixedParallelState *)
327 shm_toc_allocate(pcxt->toc, sizeof(FixedParallelState));
328 fps->database_id = MyDatabaseId;
329 fps->authenticated_user_id = GetAuthenticatedUserId();
330 fps->outer_user_id = GetCurrentRoleId();
331 fps->is_superuser = current_role_is_superuser;
332 GetUserIdAndSecContext(&fps->current_user_id, &fps->sec_context);
333 GetTempNamespaceState(&fps->temp_namespace_id,
334 &fps->temp_toast_namespace_id);
335 fps->parallel_leader_pgproc = MyProc;
336 fps->parallel_leader_pid = MyProcPid;
337 fps->parallel_leader_proc_number = MyProcNumber;
338 fps->xact_ts = GetCurrentTransactionStartTimestamp();
339 fps->stmt_ts = GetCurrentStatementStartTimestamp();
340 fps->serializable_xact_handle = ShareSerializableXact();
341 SpinLockInit(&fps->mutex);
342 fps->last_xlog_end = 0;
343 shm_toc_insert(pcxt->toc, PARALLEL_KEY_FIXED, fps);
344
345 /* We can skip the rest of this if we're not budgeting for any workers. */
346 if (pcxt->nworkers > 0)
347 {
348 char *libraryspace;
349 char *gucspace;
350 char *combocidspace;
351 char *tsnapspace;
352 char *asnapspace;
353 char *tstatespace;
354 char *pendingsyncsspace;
355 char *reindexspace;
356 char *relmapperspace;
357 char *error_queue_space;
358 char *session_dsm_handle_space;
359 char *entrypointstate;
360 char *uncommittedenumsspace;
361 char *clientconninfospace;
362 Size lnamelen;
363
364 /* Serialize shared libraries we have loaded. */
365 libraryspace = shm_toc_allocate(pcxt->toc, library_len);
366 SerializeLibraryState(library_len, libraryspace);
367 shm_toc_insert(pcxt->toc, PARALLEL_KEY_LIBRARY, libraryspace);
368
369 /* Serialize GUC settings. */
370 gucspace = shm_toc_allocate(pcxt->toc, guc_len);
371 SerializeGUCState(guc_len, gucspace);
372 shm_toc_insert(pcxt->toc, PARALLEL_KEY_GUC, gucspace);
373
374 /* Serialize combo CID state. */
375 combocidspace = shm_toc_allocate(pcxt->toc, combocidlen);
376 SerializeComboCIDState(combocidlen, combocidspace);
377 shm_toc_insert(pcxt->toc, PARALLEL_KEY_COMBO_CID, combocidspace);
378
379 /*
380 * Serialize the transaction snapshot if the transaction isolation
381 * level uses a transaction snapshot.
382 */
383 if (IsolationUsesXactSnapshot())
384 {
385 tsnapspace = shm_toc_allocate(pcxt->toc, tsnaplen);
386 SerializeSnapshot(transaction_snapshot, tsnapspace);
387 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT,
388 tsnapspace);
389 }
390
391 /* Serialize the active snapshot. */
392 asnapspace = shm_toc_allocate(pcxt->toc, asnaplen);
393 SerializeSnapshot(active_snapshot, asnapspace);
394 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ACTIVE_SNAPSHOT, asnapspace);
395
396 /* Provide the handle for per-session segment. */
397 session_dsm_handle_space = shm_toc_allocate(pcxt->toc,
398 sizeof(dsm_handle));
399 *(dsm_handle *) session_dsm_handle_space = session_dsm_handle;
400 shm_toc_insert(pcxt->toc, PARALLEL_KEY_SESSION_DSM,
401 session_dsm_handle_space);
402
403 /* Serialize transaction state. */
404 tstatespace = shm_toc_allocate(pcxt->toc, tstatelen);
405 SerializeTransactionState(tstatelen, tstatespace);
406 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_STATE, tstatespace);
407
408 /* Serialize pending syncs. */
409 pendingsyncsspace = shm_toc_allocate(pcxt->toc, pendingsyncslen);
410 SerializePendingSyncs(pendingsyncslen, pendingsyncsspace);
411 shm_toc_insert(pcxt->toc, PARALLEL_KEY_PENDING_SYNCS,
412 pendingsyncsspace);
413
414 /* Serialize reindex state. */
415 reindexspace = shm_toc_allocate(pcxt->toc, reindexlen);
416 SerializeReindexState(reindexlen, reindexspace);
417 shm_toc_insert(pcxt->toc, PARALLEL_KEY_REINDEX_STATE, reindexspace);
418
419 /* Serialize relmapper state. */
420 relmapperspace = shm_toc_allocate(pcxt->toc, relmapperlen);
421 SerializeRelationMap(relmapperlen, relmapperspace);
422 shm_toc_insert(pcxt->toc, PARALLEL_KEY_RELMAPPER_STATE,
423 relmapperspace);
424
425 /* Serialize uncommitted enum state. */
426 uncommittedenumsspace = shm_toc_allocate(pcxt->toc,
427 uncommittedenumslen);
428 SerializeUncommittedEnums(uncommittedenumsspace, uncommittedenumslen);
429 shm_toc_insert(pcxt->toc, PARALLEL_KEY_UNCOMMITTEDENUMS,
430 uncommittedenumsspace);
431
432 /* Serialize our ClientConnectionInfo. */
433 clientconninfospace = shm_toc_allocate(pcxt->toc, clientconninfolen);
434 SerializeClientConnectionInfo(clientconninfolen, clientconninfospace);
435 shm_toc_insert(pcxt->toc, PARALLEL_KEY_CLIENTCONNINFO,
436 clientconninfospace);
437
438 /* Allocate space for worker information. */
439 pcxt->worker = palloc0(sizeof(ParallelWorkerInfo) * pcxt->nworkers);
440
441 /*
442 * Establish error queues in dynamic shared memory.
443 *
444 * These queues should be used only for transmitting ErrorResponse,
445 * NoticeResponse, and NotifyResponse protocol messages. Tuple data
446 * should be transmitted via separate (possibly larger?) queues.
447 */
448 error_queue_space =
449 shm_toc_allocate(pcxt->toc,
450 mul_size(PARALLEL_ERROR_QUEUE_SIZE,
451 pcxt->nworkers));
452 for (i = 0; i < pcxt->nworkers; ++i)
453 {
454 char *start;
455 shm_mq *mq;
456
457 start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
458 mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
459 shm_mq_set_receiver(mq, MyProc);
460 pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
461 }
462 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE, error_queue_space);
463
464 /*
465 * Serialize entrypoint information. It's unsafe to pass function
466 * pointers across processes, as the function pointer may be different
467 * in each process in EXEC_BACKEND builds, so we always pass library
468 * and function name. (We use library name "postgres" for functions
469 * in the core backend.)
470 */
471 lnamelen = strlen(pcxt->library_name);
472 entrypointstate = shm_toc_allocate(pcxt->toc, lnamelen +
473 strlen(pcxt->function_name) + 2);
474 strcpy(entrypointstate, pcxt->library_name);
475 strcpy(entrypointstate + lnamelen + 1, pcxt->function_name);
476 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ENTRYPOINT, entrypointstate);
477 }
478
479 /* Restore previous memory context. */
480 MemoryContextSwitchTo(oldcontext);
481 }
482
483 /*
484 * Reinitialize the dynamic shared memory segment for a parallel context such
485 * that we could launch workers for it again.
486 */
487 void
488 ReinitializeParallelDSM(ParallelContext *pcxt)
489 {
490 FixedParallelState *fps;
491
492 /* Wait for any old workers to exit. */
493 if (pcxt->nworkers_launched > 0)
494 {
495 WaitForParallelWorkersToFinish(pcxt);
496 WaitForParallelWorkersToExit(pcxt);
497 pcxt->nworkers_launched = 0;
498 if (pcxt->known_attached_workers)
499 {
500 pfree(pcxt->known_attached_workers);
501 pcxt->known_attached_workers = NULL;
502 pcxt->nknown_attached_workers = 0;
503 }
504 }
505
506 /* Reset a few bits of fixed parallel state to a clean state. */
507 fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED, false);
508 fps->last_xlog_end = 0;
509
510 /* Recreate error queues (if they exist). */
511 if (pcxt->nworkers > 0)
512 {
513 char *error_queue_space;
514 int i;
515
516 error_queue_space =
517 shm_toc_lookup(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE, false);
518 for (i = 0; i < pcxt->nworkers; ++i)
519 {
520 char *start;
521 shm_mq *mq;
522
523 start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
524 mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
525 shm_mq_set_receiver(mq, MyProc);
526 pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
527 }
528 }
529 }
530
531 /*
532 * Reinitialize parallel workers for a parallel context such that we could
533 * launch a different number of workers. This is required for cases where
534 * we need to reuse the same DSM segment, but the number of workers can
535 * vary from run-to-run.
536 */
537 void
538 ReinitializeParallelWorkers(ParallelContext *pcxt, int nworkers_to_launch)
539 {
540 /*
541 * The number of workers that need to be launched must be less than the
542 * number of workers with which the parallel context is initialized.
543 */
544 Assert(pcxt->nworkers >= nworkers_to_launch);
545 pcxt->nworkers_to_launch = nworkers_to_launch;
546 }
547
548 /*
549 * Launch parallel workers.
550 */
551 void
552 LaunchParallelWorkers(ParallelContext *pcxt)
553 {
554 MemoryContext oldcontext;
555 BackgroundWorker worker;
556 int i;
557 bool any_registrations_failed = false;
558
559 /* Skip this if we have no workers. */
560 if (pcxt->nworkers == 0 || pcxt->nworkers_to_launch == 0)
561 return;
562
563 /* We need to be a lock group leader. */
564 BecomeLockGroupLeader();
565
566 /* If we do have workers, we'd better have a DSM segment. */
567 Assert(pcxt->seg != NULL);
568
569 /* We might be running in a short-lived memory context. */
570 oldcontext = MemoryContextSwitchTo(TopTransactionContext);
571
572 /* Configure a worker. */
573 memset(&worker, 0, sizeof(worker));
574 snprintf(worker.bgw_name, BGW_MAXLEN, "parallel worker for PID %d",
575 MyProcPid);
576 snprintf(worker.bgw_type, BGW_MAXLEN, "parallel worker");
577 worker.bgw_flags =
578 BGWORKER_SHMEM_ACCESS | BGWORKER_BACKEND_DATABASE_CONNECTION
579 | BGWORKER_CLASS_PARALLEL;
580 worker.bgw_start_time = BgWorkerStart_ConsistentState;
581 worker.bgw_restart_time = BGW_NEVER_RESTART;
582 sprintf(worker.bgw_library_name, "postgres");
583 sprintf(worker.bgw_function_name, "ParallelWorkerMain");
584 worker.bgw_main_arg = UInt32GetDatum(dsm_segment_handle(pcxt->seg));
585 worker.bgw_notify_pid = MyProcPid;
586
587 /*
588 * Start workers.
589 *
590 * The caller must be able to tolerate ending up with fewer workers than
591 * expected, so there is no need to throw an error here if registration
592 * fails. It wouldn't help much anyway, because registering the worker in
593 * no way guarantees that it will start up and initialize successfully.
594 */
595 for (i = 0; i < pcxt->nworkers_to_launch; ++i)
596 {
597 memcpy(worker.bgw_extra, &i, sizeof(int));
598 if (!any_registrations_failed &&
599 RegisterDynamicBackgroundWorker(&worker,
600 &pcxt->worker[i].bgwhandle))
601 {
602 shm_mq_set_handle(pcxt->worker[i].error_mqh,
603 pcxt->worker[i].bgwhandle);
604 pcxt->nworkers_launched++;
605 }
606 else
607 {
608 /*
609 * If we weren't able to register the worker, then we've bumped up
610 * against the max_worker_processes limit, and future
611 * registrations will probably fail too, so arrange to skip them.
612 * But we still have to execute this code for the remaining slots
613 * to make sure that we forget about the error queues we budgeted
614 * for those workers. Otherwise, we'll wait for them to start,
615 * but they never will.
616 */
617 any_registrations_failed = true;
618 pcxt->worker[i].bgwhandle = NULL;
619 shm_mq_detach(pcxt->worker[i].error_mqh);
620 pcxt->worker[i].error_mqh = NULL;
621 }
622 }
623
624 /*
625 * Now that nworkers_launched has taken its final value, we can initialize
626 * known_attached_workers.
627 */
628 if (pcxt->nworkers_launched > 0)
629 {
630 pcxt->known_attached_workers =
631 palloc0(sizeof(bool) * pcxt->nworkers_launched);
632 pcxt->nknown_attached_workers = 0;
633 }
634
635 /* Restore previous memory context. */
636 MemoryContextSwitchTo(oldcontext);
637 }
638
639 /*
640 * Wait for all workers to attach to their error queues, and throw an error if
641 * any worker fails to do this.
642 *
643 * Callers can assume that if this function returns successfully, then the
644 * number of workers given by pcxt->nworkers_launched have initialized and
645 * attached to their error queues. Whether or not these workers are guaranteed
646 * to still be running depends on what code the caller asked them to run;
647 * this function does not guarantee that they have not exited. However, it
648 * does guarantee that any workers which exited must have done so cleanly and
649 * after successfully performing the work with which they were tasked.
650 *
651 * If this function is not called, then some of the workers that were launched
652 * may not have been started due to a fork() failure, or may have exited during
653 * early startup prior to attaching to the error queue, so nworkers_launched
654 * cannot be viewed as completely reliable. It will never be less than the
655 * number of workers which actually started, but it might be more. Any workers
656 * that failed to start will still be discovered by
657 * WaitForParallelWorkersToFinish and an error will be thrown at that time,
658 * provided that function is eventually reached.
659 *
660 * In general, the leader process should do as much work as possible before
661 * calling this function. fork() failures and other early-startup failures
662 * are very uncommon, and having the leader sit idle when it could be doing
663 * useful work is undesirable. However, if the leader needs to wait for
664 * all of its workers or for a specific worker, it may want to call this
665 * function before doing so. If not, it must make some other provision for
666 * the failure-to-start case, lest it wait forever. On the other hand, a
667 * leader which never waits for a worker that might not be started yet, or
668 * at least never does so prior to WaitForParallelWorkersToFinish(), need not
669 * call this function at all.
670 */
671 void
672 WaitForParallelWorkersToAttach(ParallelContext *pcxt)
673 {
674 int i;
675
676 /* Skip this if we have no launched workers. */
677 if (pcxt->nworkers_launched == 0)
678 return;
679
680 for (;;)
681 {
682 /*
683 * This will process any parallel messages that are pending and it may
684 * also throw an error propagated from a worker.
685 */
686 CHECK_FOR_INTERRUPTS();
687
688 for (i = 0; i < pcxt->nworkers_launched; ++i)
689 {
690 BgwHandleStatus status;
691 shm_mq *mq;
692 int rc;
693 pid_t pid;
694
695 if (pcxt->known_attached_workers[i])
696 continue;
697
698 /*
699 * If error_mqh is NULL, then the worker has already exited
700 * cleanly.
701 */
702 if (pcxt->worker[i].error_mqh == NULL)
703 {
704 pcxt->known_attached_workers[i] = true;
705 ++pcxt->nknown_attached_workers;
706 continue;
707 }
708
709 status = GetBackgroundWorkerPid(pcxt->worker[i].bgwhandle, &pid);
710 if (status == BGWH_STARTED)
711 {
712 /* Has the worker attached to the error queue? */
713 mq = shm_mq_get_queue(pcxt->worker[i].error_mqh);
714 if (shm_mq_get_sender(mq) != NULL)
715 {
716 /* Yes, so it is known to be attached. */
717 pcxt->known_attached_workers[i] = true;
718 ++pcxt->nknown_attached_workers;
719 }
720 }
721 else if (status == BGWH_STOPPED)
722 {
723 /*
724 * If the worker stopped without attaching to the error queue,
725 * throw an error.
726 */
727 mq = shm_mq_get_queue(pcxt->worker[i].error_mqh);
728 if (shm_mq_get_sender(mq) == NULL)
729 ereport(ERROR,
730 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
731 errmsg("parallel worker failed to initialize"),
732 errhint("More details may be available in the server log.")));
733
734 pcxt->known_attached_workers[i] = true;
735 ++pcxt->nknown_attached_workers;
736 }
737 else
738 {
739 /*
740 * Worker not yet started, so we must wait. The postmaster
741 * will notify us if the worker's state changes. Our latch
742 * might also get set for some other reason, but if so we'll
743 * just end up waiting for the same worker again.
744 */
745 rc = WaitLatch(MyLatch,
746 WL_LATCH_SET | WL_EXIT_ON_PM_DEATH,
747 -1, WAIT_EVENT_BGWORKER_STARTUP);
748
749 if (rc & WL_LATCH_SET)
750 ResetLatch(MyLatch);
751 }
752 }
753
754 /* If all workers are known to have started, we're done. */
755 if (pcxt->nknown_attached_workers >= pcxt->nworkers_launched)
756 {
757 Assert(pcxt->nknown_attached_workers == pcxt->nworkers_launched);
758 break;
759 }
760 }
761 }
762
763 /*
764 * Wait for all workers to finish computing.
765 *
766 * Even if the parallel operation seems to have completed successfully, it's
767 * important to call this function afterwards. We must not miss any errors
768 * the workers may have thrown during the parallel operation, or any that they
769 * may yet throw while shutting down.
770 *
771 * Also, we want to update our notion of XactLastRecEnd based on worker
772 * feedback.
773 */
774 void
775 WaitForParallelWorkersToFinish(ParallelContext *pcxt)
776 {
777 for (;;)
778 {
779 bool anyone_alive = false;
780 int nfinished = 0;
781 int i;
782
783 /*
784 * This will process any parallel messages that are pending, which may
785 * change the outcome of the loop that follows. It may also throw an
786 * error propagated from a worker.
787 */
788 CHECK_FOR_INTERRUPTS();
789
790 for (i = 0; i < pcxt->nworkers_launched; ++i)
791 {
792 /*
793 * If error_mqh is NULL, then the worker has already exited
794 * cleanly. If we have received a message through error_mqh from
795 * the worker, we know it started up cleanly, and therefore we're
796 * certain to be notified when it exits.
797 */
798 if (pcxt->worker[i].error_mqh == NULL)
799 ++nfinished;
800 else if (pcxt->known_attached_workers[i])
801 {
802 anyone_alive = true;
803 break;
804 }
805 }
806
807 if (!anyone_alive)
808 {
809 /* If all workers are known to have finished, we're done. */
810 if (nfinished >= pcxt->nworkers_launched)
811 {
812 Assert(nfinished == pcxt->nworkers_launched);
813 break;
814 }
815
816 /*
817 * We didn't detect any living workers, but not all workers are
818 * known to have exited cleanly. Either not all workers have
819 * launched yet, or maybe some of them failed to start or
820 * terminated abnormally.
821 */
822 for (i = 0; i < pcxt->nworkers_launched; ++i)
823 {
824 pid_t pid;
825 shm_mq *mq;
826
827 /*
828 * If the worker is BGWH_NOT_YET_STARTED or BGWH_STARTED, we
829 * should just keep waiting. If it is BGWH_STOPPED, then
830 * further investigation is needed.
831 */
832 if (pcxt->worker[i].error_mqh == NULL ||
833 pcxt->worker[i].bgwhandle == NULL ||
834 GetBackgroundWorkerPid(pcxt->worker[i].bgwhandle,
835 &pid) != BGWH_STOPPED)
836 continue;
837
838 /*
839 * Check whether the worker ended up stopped without ever
840 * attaching to the error queue. If so, the postmaster was
841 * unable to fork the worker or it exited without initializing
842 * properly. We must throw an error, since the caller may
843 * have been expecting the worker to do some work before
844 * exiting.
845 */
846 mq = shm_mq_get_queue(pcxt->worker[i].error_mqh);
847 if (shm_mq_get_sender(mq) == NULL)
848 ereport(ERROR,
849 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
850 errmsg("parallel worker failed to initialize"),
851 errhint("More details may be available in the server log.")));
852
853 /*
854 * The worker is stopped, but is attached to the error queue.
855 * Unless there's a bug somewhere, this will only happen when
856 * the worker writes messages and terminates after the
857 * CHECK_FOR_INTERRUPTS() near the top of this function and
858 * before the call to GetBackgroundWorkerPid(). In that case,
859 * or latch should have been set as well and the right things
860 * will happen on the next pass through the loop.
861 */
862 }
863 }
864
865 (void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, -1,
866 WAIT_EVENT_PARALLEL_FINISH);
867 ResetLatch(MyLatch);
868 }
869
870 if (pcxt->toc != NULL)
871 {
872 FixedParallelState *fps;
873
874 fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED, false);
875 if (fps->last_xlog_end > XactLastRecEnd)
876 XactLastRecEnd = fps->last_xlog_end;
877 }
878 }
879
880 /*
881 * Wait for all workers to exit.
882 *
883 * This function ensures that workers have been completely shutdown. The
884 * difference between WaitForParallelWorkersToFinish and this function is
885 * that the former just ensures that last message sent by a worker backend is
886 * received by the leader backend whereas this ensures the complete shutdown.
887 */
888 static void
889 WaitForParallelWorkersToExit(ParallelContext *pcxt)
890 {
891 int i;
892
893 /* Wait until the workers actually die. */
894 for (i = 0; i < pcxt->nworkers_launched; ++i)
895 {
896 BgwHandleStatus status;
897
898 if (pcxt->worker == NULL || pcxt->worker[i].bgwhandle == NULL)
899 continue;
900
901 status = WaitForBackgroundWorkerShutdown(pcxt->worker[i].bgwhandle);
902
903 /*
904 * If the postmaster kicked the bucket, we have no chance of cleaning
905 * up safely -- we won't be able to tell when our workers are actually
906 * dead. This doesn't necessitate a PANIC since they will all abort
907 * eventually, but we can't safely continue this session.
908 */
909 if (status == BGWH_POSTMASTER_DIED)
910 ereport(FATAL,
911 (errcode(ERRCODE_ADMIN_SHUTDOWN),
912 errmsg("postmaster exited during a parallel transaction")));
913
914 /* Release memory. */
915 pfree(pcxt->worker[i].bgwhandle);
916 pcxt->worker[i].bgwhandle = NULL;
917 }
918 }
919
920 /*
921 * Destroy a parallel context.
922 *
923 * If expecting a clean exit, you should use WaitForParallelWorkersToFinish()
924 * first, before calling this function. When this function is invoked, any
925 * remaining workers are forcibly killed; the dynamic shared memory segment
926 * is unmapped; and we then wait (uninterruptibly) for the workers to exit.
927 */
928 void
929 DestroyParallelContext(ParallelContext *pcxt)
930 {
931 int i;
932
933 /*
934 * Be careful about order of operations here! We remove the parallel
935 * context from the list before we do anything else; otherwise, if an
936 * error occurs during a subsequent step, we might try to nuke it again
937 * from AtEOXact_Parallel or AtEOSubXact_Parallel.
938 */
939 dlist_delete(&pcxt->node);
940
941 /* Kill each worker in turn, and forget their error queues. */
942 if (pcxt->worker != NULL)
943 {
944 for (i = 0; i < pcxt->nworkers_launched; ++i)
945 {
946 if (pcxt->worker[i].error_mqh != NULL)
947 {
948 TerminateBackgroundWorker(pcxt->worker[i].bgwhandle);
949
950 shm_mq_detach(pcxt->worker[i].error_mqh);
951 pcxt->worker[i].error_mqh = NULL;
952 }
953 }
954 }
955
956 /*
957 * If we have allocated a shared memory segment, detach it. This will
958 * implicitly detach the error queues, and any other shared memory queues,
959 * stored there.
960 */
961 if (pcxt->seg != NULL)
962 {
963 dsm_detach(pcxt->seg);
964 pcxt->seg = NULL;
965 }
966
967 /*
968 * If this parallel context is actually in backend-private memory rather
969 * than shared memory, free that memory instead.
970 */
971 if (pcxt->private_memory != NULL)
972 {
973 pfree(pcxt->private_memory);
974 pcxt->private_memory = NULL;
975 }
976
977 /*
978 * We can't finish transaction commit or abort until all of the workers
979 * have exited. This means, in particular, that we can't respond to
980 * interrupts at this stage.
981 */
982 HOLD_INTERRUPTS();
983 WaitForParallelWorkersToExit(pcxt);
984 RESUME_INTERRUPTS();
985
986 /* Free the worker array itself. */
987 if (pcxt->worker != NULL)
988 {
989 pfree(pcxt->worker);
990 pcxt->worker = NULL;
991 }
992
993 /* Free memory. */
994 pfree(pcxt->library_name);
995 pfree(pcxt->function_name);
996 pfree(pcxt);
997 }
998
999 /*
1000 * Are there any parallel contexts currently active?
1001 */
1002 bool
1003 ParallelContextActive(void)
1004 {
1005 return !dlist_is_empty(&pcxt_list);
1006 }
1007
1008 /*
1009 * Handle receipt of an interrupt indicating a parallel worker message.
1010 *
1011 * Note: this is called within a signal handler! All we can do is set
1012 * a flag that will cause the next CHECK_FOR_INTERRUPTS() to invoke
1013 * HandleParallelMessages().
1014 */
1015 void
1016 HandleParallelMessageInterrupt(void)
1017 {
1018 InterruptPending = true;
1019 ParallelMessagePending = true;
1020 SetLatch(MyLatch);
1021 }
1022
1023 /*
1024 * Handle any queued protocol messages received from parallel workers.
1025 */
1026 void
1027 HandleParallelMessages(void)
1028 {
1029 dlist_iter iter;
1030 MemoryContext oldcontext;
1031
1032 static MemoryContext hpm_context = NULL;
1033
1034 /*
1035 * This is invoked from ProcessInterrupts(), and since some of the
1036 * functions it calls contain CHECK_FOR_INTERRUPTS(), there is a potential
1037 * for recursive calls if more signals are received while this runs. It's
1038 * unclear that recursive entry would be safe, and it doesn't seem useful
1039 * even if it is safe, so let's block interrupts until done.
1040 */
1041 HOLD_INTERRUPTS();
1042
1043 /*
1044 * Moreover, CurrentMemoryContext might be pointing almost anywhere. We
1045 * don't want to risk leaking data into long-lived contexts, so let's do
1046 * our work here in a private context that we can reset on each use.
1047 */
1048 if (hpm_context == NULL) /* first time through? */
1049 hpm_context = AllocSetContextCreate(TopMemoryContext,
1050 "HandleParallelMessages",
1051 ALLOCSET_DEFAULT_SIZES);
1052 else
1053 MemoryContextReset(hpm_context);
1054
1055 oldcontext = MemoryContextSwitchTo(hpm_context);
1056
1057 /* OK to process messages. Reset the flag saying there are more to do. */
1058 ParallelMessagePending = false;
1059
1060 dlist_foreach(iter, &pcxt_list)
1061 {
1062 ParallelContext *pcxt;
1063 int i;
1064
1065 pcxt = dlist_container(ParallelContext, node, iter.cur);
1066 if (pcxt->worker == NULL)
1067 continue;
1068
1069 for (i = 0; i < pcxt->nworkers_launched; ++i)
1070 {
1071 /*
1072 * Read as many messages as we can from each worker, but stop when
1073 * either (1) the worker's error queue goes away, which can happen
1074 * if we receive a Terminate message from the worker; or (2) no
1075 * more messages can be read from the worker without blocking.
1076 */
1077 while (pcxt->worker[i].error_mqh != NULL)
1078 {
1079 shm_mq_result res;
1080 Size nbytes;
1081 void *data;
1082
1083 res = shm_mq_receive(pcxt->worker[i].error_mqh, &nbytes,
1084 &data, true);
1085 if (res == SHM_MQ_WOULD_BLOCK)
1086 break;
1087 else if (res == SHM_MQ_SUCCESS)
1088 {
1089 StringInfoData msg;
1090
1091 initStringInfo(&msg);
1092 appendBinaryStringInfo(&msg, data, nbytes);
1093 HandleParallelMessage(pcxt, i, &msg);
1094 pfree(msg.data);
1095 }
1096 else
1097 ereport(ERROR,
1098 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1099 errmsg("lost connection to parallel worker")));
1100 }
1101 }
1102 }
1103
1104 MemoryContextSwitchTo(oldcontext);
1105
1106 /* Might as well clear the context on our way out */
1107 MemoryContextReset(hpm_context);
1108
1109 RESUME_INTERRUPTS();
1110 }
1111
1112 /*
1113 * Handle a single protocol message received from a single parallel worker.
1114 */
1115 static void
1116 HandleParallelMessage(ParallelContext *pcxt, int i, StringInfo msg)
1117 {
1118 char msgtype;
1119
1120 if (pcxt->known_attached_workers != NULL &&
1121 !pcxt->known_attached_workers[i])
1122 {
1123 pcxt->known_attached_workers[i] = true;
1124 pcxt->nknown_attached_workers++;
1125 }
1126
1127 msgtype = pq_getmsgbyte(msg);
1128
1129 switch (msgtype)
1130 {
1131 case PqMsg_ErrorResponse:
1132 case PqMsg_NoticeResponse:
1133 {
1134 ErrorData edata;
1135 ErrorContextCallback *save_error_context_stack;
1136
1137 /* Parse ErrorResponse or NoticeResponse. */
1138 pq_parse_errornotice(msg, &edata);
1139
1140 /* Death of a worker isn't enough justification for suicide. */
1141 edata.elevel = Min(edata.elevel, ERROR);
1142
1143 /*
1144 * If desired, add a context line to show that this is a
1145 * message propagated from a parallel worker. Otherwise, it
1146 * can sometimes be confusing to understand what actually
1147 * happened. (We don't do this in DEBUG_PARALLEL_REGRESS mode
1148 * because it causes test-result instability depending on
1149 * whether a parallel worker is actually used or not.)
1150 */
1151 if (debug_parallel_query != DEBUG_PARALLEL_REGRESS)
1152 {
1153 if (edata.context)
1154 edata.context = psprintf("%s\n%s", edata.context,
1155 _("parallel worker"));
1156 else
1157 edata.context = pstrdup(_("parallel worker"));
1158 }
1159
1160 /*
1161 * Context beyond that should use the error context callbacks
1162 * that were in effect when the ParallelContext was created,
1163 * not the current ones.
1164 */
1165 save_error_context_stack = error_context_stack;
1166 error_context_stack = pcxt->error_context_stack;
1167
1168 /* Rethrow error or print notice. */
1169 ThrowErrorData(&edata);
1170
1171 /* Not an error, so restore previous context stack. */
1172 error_context_stack = save_error_context_stack;
1173
1174 break;
1175 }
1176
1177 case PqMsg_NotificationResponse:
1178 {
1179 /* Propagate NotifyResponse. */
1180 int32 pid;
1181 const char *channel;
1182 const char *payload;
1183
1184 pid = pq_getmsgint(msg, 4);
1185 channel = pq_getmsgrawstring(msg);
1186 payload = pq_getmsgrawstring(msg);
1187 pq_endmessage(msg);
1188
1189 NotifyMyFrontEnd(channel, payload, pid);
1190
1191 break;
1192 }
1193
1194 case 'P': /* Parallel progress reporting */
1195 {
1196 /*
1197 * Only incremental progress reporting is currently supported.
1198 * However, it's possible to add more fields to the message to
1199 * allow for handling of other backend progress APIs.
1200 */
1201 int index = pq_getmsgint(msg, 4);
1202 int64 incr = pq_getmsgint64(msg);
1203
1204 pq_getmsgend(msg);
1205
1206 pgstat_progress_incr_param(index, incr);
1207
1208 break;
1209 }
1210
1211 case PqMsg_Terminate:
1212 {
1213 shm_mq_detach(pcxt->worker[i].error_mqh);
1214 pcxt->worker[i].error_mqh = NULL;
1215 break;
1216 }
1217
1218 default:
1219 {
1220 elog(ERROR, "unrecognized message type received from parallel worker: %c (message length %d bytes)",
1221 msgtype, msg->len);
1222 }
1223 }
1224 }
1225
1226 /*
1227 * End-of-subtransaction cleanup for parallel contexts.
1228 *
1229 * Here we remove only parallel contexts initiated within the current
1230 * subtransaction.
1231 */
1232 void
1233 AtEOSubXact_Parallel(bool isCommit, SubTransactionId mySubId)
1234 {
1235 while (!dlist_is_empty(&pcxt_list))
1236 {
1237 ParallelContext *pcxt;
1238
1239 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
1240 if (pcxt->subid != mySubId)
1241 break;
1242 if (isCommit)
1243 elog(WARNING, "leaked parallel context");
1244 DestroyParallelContext(pcxt);
1245 }
1246 }
1247
1248 /*
1249 * End-of-transaction cleanup for parallel contexts.
1250 *
1251 * We nuke all remaining parallel contexts.
1252 */
1253 void
1254 AtEOXact_Parallel(bool isCommit)
1255 {
1256 while (!dlist_is_empty(&pcxt_list))
1257 {
1258 ParallelContext *pcxt;
1259
1260 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
1261 if (isCommit)
1262 elog(WARNING, "leaked parallel context");
1263 DestroyParallelContext(pcxt);
1264 }
1265 }
1266
1267 /*
1268 * Main entrypoint for parallel workers.
1269 */
1270 void
1271 ParallelWorkerMain(Datum main_arg)
1272 {
1273 dsm_segment *seg;
1274 shm_toc *toc;
1275 FixedParallelState *fps;
1276 char *error_queue_space;
1277 shm_mq *mq;
1278 shm_mq_handle *mqh;
1279 char *libraryspace;
1280 char *entrypointstate;
1281 char *library_name;
1282 char *function_name;
1283 parallel_worker_main_type entrypt;
1284 char *gucspace;
1285 char *combocidspace;
1286 char *tsnapspace;
1287 char *asnapspace;
1288 char *tstatespace;
1289 char *pendingsyncsspace;
1290 char *reindexspace;
1291 char *relmapperspace;
1292 char *uncommittedenumsspace;
1293 char *clientconninfospace;
1294 char *session_dsm_handle_space;
1295 Snapshot tsnapshot;
1296 Snapshot asnapshot;
1297
1298 /* Set flag to indicate that we're initializing a parallel worker. */
1299 InitializingParallelWorker = true;
1300
1301 /* Establish signal handlers. */
1302 pqsignal(SIGTERM, die);
1303 BackgroundWorkerUnblockSignals();
1304
1305 /* Determine and set our parallel worker number. */
1306 Assert(ParallelWorkerNumber == -1);
1307 memcpy(&ParallelWorkerNumber, MyBgworkerEntry->bgw_extra, sizeof(int));
1308
1309 /* Set up a memory context to work in, just for cleanliness. */
1310 CurrentMemoryContext = AllocSetContextCreate(TopMemoryContext,
1311 "Parallel worker",
1312 ALLOCSET_DEFAULT_SIZES);
1313
1314 /*
1315 * Attach to the dynamic shared memory segment for the parallel query, and
1316 * find its table of contents.
1317 *
1318 * Note: at this point, we have not created any ResourceOwner in this
1319 * process. This will result in our DSM mapping surviving until process
1320 * exit, which is fine. If there were a ResourceOwner, it would acquire
1321 * ownership of the mapping, but we have no need for that.
1322 */
1323 seg = dsm_attach(DatumGetUInt32(main_arg));
1324 if (seg == NULL)
1325 ereport(ERROR,
1326 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1327 errmsg("could not map dynamic shared memory segment")));
1328 toc = shm_toc_attach(PARALLEL_MAGIC, dsm_segment_address(seg));
1329 if (toc == NULL)
1330 ereport(ERROR,
1331 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1332 errmsg("invalid magic number in dynamic shared memory segment")));
1333
1334 /* Look up fixed parallel state. */
1335 fps = shm_toc_lookup(toc, PARALLEL_KEY_FIXED, false);
1336 MyFixedParallelState = fps;
1337
1338 /* Arrange to signal the leader if we exit. */
1339 ParallelLeaderPid = fps->parallel_leader_pid;
1340 ParallelLeaderProcNumber = fps->parallel_leader_proc_number;
1341 before_shmem_exit(ParallelWorkerShutdown, PointerGetDatum(seg));
1342
1343 /*
1344 * Now we can find and attach to the error queue provided for us. That's
1345 * good, because until we do that, any errors that happen here will not be
1346 * reported back to the process that requested that this worker be
1347 * launched.
1348 */
1349 error_queue_space = shm_toc_lookup(toc, PARALLEL_KEY_ERROR_QUEUE, false);
1350 mq = (shm_mq *) (error_queue_space +
1351 ParallelWorkerNumber * PARALLEL_ERROR_QUEUE_SIZE);
1352 shm_mq_set_sender(mq, MyProc);
1353 mqh = shm_mq_attach(mq, seg, NULL);
1354 pq_redirect_to_shm_mq(seg, mqh);
1355 pq_set_parallel_leader(fps->parallel_leader_pid,
1356 fps->parallel_leader_proc_number);
1357
1358 /*
1359 * Hooray! Primary initialization is complete. Now, we need to set up our
1360 * backend-local state to match the original backend.
1361 */
1362
1363 /*
1364 * Join locking group. We must do this before anything that could try to
1365 * acquire a heavyweight lock, because any heavyweight locks acquired to
1366 * this point could block either directly against the parallel group
1367 * leader or against some process which in turn waits for a lock that
1368 * conflicts with the parallel group leader, causing an undetected
1369 * deadlock. (If we can't join the lock group, the leader has gone away,
1370 * so just exit quietly.)
1371 */
1372 if (!BecomeLockGroupMember(fps->parallel_leader_pgproc,
1373 fps->parallel_leader_pid))
1374 return;
1375
1376 /*
1377 * Restore transaction and statement start-time timestamps. This must
1378 * happen before anything that would start a transaction, else asserts in
1379 * xact.c will fire.
1380 */
1381 SetParallelStartTimestamps(fps->xact_ts, fps->stmt_ts);
1382
1383 /*
1384 * Identify the entry point to be called. In theory this could result in
1385 * loading an additional library, though most likely the entry point is in
1386 * the core backend or in a library we just loaded.
1387 */
1388 entrypointstate = shm_toc_lookup(toc, PARALLEL_KEY_ENTRYPOINT, false);
1389 library_name = entrypointstate;
1390 function_name = entrypointstate + strlen(library_name) + 1;
1391
1392 entrypt = LookupParallelWorkerFunction(library_name, function_name);
1393
1394 /* Restore database connection. */
1395 BackgroundWorkerInitializeConnectionByOid(fps->database_id,
1396 fps->authenticated_user_id,
1397 0);
1398
1399 /*
1400 * Set the client encoding to the database encoding, since that is what
1401 * the leader will expect.
1402 */
1403 SetClientEncoding(GetDatabaseEncoding());
1404
1405 /*
1406 * Load libraries that were loaded by original backend. We want to do
1407 * this before restoring GUCs, because the libraries might define custom
1408 * variables.
1409 */
1410 libraryspace = shm_toc_lookup(toc, PARALLEL_KEY_LIBRARY, false);
1411 StartTransactionCommand();
1412 RestoreLibraryState(libraryspace);
1413
1414 /* Restore GUC values from launching backend. */
1415 gucspace = shm_toc_lookup(toc, PARALLEL_KEY_GUC, false);
1416 RestoreGUCState(gucspace);
1417 CommitTransactionCommand();
1418
1419 /* Crank up a transaction state appropriate to a parallel worker. */
1420 tstatespace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_STATE, false);
1421 StartParallelWorkerTransaction(tstatespace);
1422
1423 /* Restore combo CID state. */
1424 combocidspace = shm_toc_lookup(toc, PARALLEL_KEY_COMBO_CID, false);
1425 RestoreComboCIDState(combocidspace);
1426
1427 /* Attach to the per-session DSM segment and contained objects. */
1428 session_dsm_handle_space =
1429 shm_toc_lookup(toc, PARALLEL_KEY_SESSION_DSM, false);
1430 AttachSession(*(dsm_handle *) session_dsm_handle_space);
1431
1432 /*
1433 * If the transaction isolation level is REPEATABLE READ or SERIALIZABLE,
1434 * the leader has serialized the transaction snapshot and we must restore
1435 * it. At lower isolation levels, there is no transaction-lifetime
1436 * snapshot, but we need TransactionXmin to get set to a value which is
1437 * less than or equal to the xmin of every snapshot that will be used by
1438 * this worker. The easiest way to accomplish that is to install the
1439 * active snapshot as the transaction snapshot. Code running in this
1440 * parallel worker might take new snapshots via GetTransactionSnapshot()
1441 * or GetLatestSnapshot(), but it shouldn't have any way of acquiring a
1442 * snapshot older than the active snapshot.
1443 */
1444 asnapspace = shm_toc_lookup(toc, PARALLEL_KEY_ACTIVE_SNAPSHOT, false);
1445 tsnapspace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT, true);
1446 asnapshot = RestoreSnapshot(asnapspace);
1447 tsnapshot = tsnapspace ? RestoreSnapshot(tsnapspace) : asnapshot;
1448 RestoreTransactionSnapshot(tsnapshot,
1449 fps->parallel_leader_pgproc);
1450 PushActiveSnapshot(asnapshot);
1451
1452 /*
1453 * We've changed which tuples we can see, and must therefore invalidate
1454 * system caches.
1455 */
1456 InvalidateSystemCaches();
1457
1458 /*
1459 * Restore current role id. Skip verifying whether session user is
1460 * allowed to become this role and blindly restore the leader's state for
1461 * current role.
1462 */
1463 SetCurrentRoleId(fps->outer_user_id, fps->is_superuser);
1464
1465 /* Restore user ID and security context. */
1466 SetUserIdAndSecContext(fps->current_user_id, fps->sec_context);
1467
1468 /* Restore temp-namespace state to ensure search path matches leader's. */
1469 SetTempNamespaceState(fps->temp_namespace_id,
1470 fps->temp_toast_namespace_id);
1471
1472 /* Restore pending syncs. */
1473 pendingsyncsspace = shm_toc_lookup(toc, PARALLEL_KEY_PENDING_SYNCS,
1474 false);
1475 RestorePendingSyncs(pendingsyncsspace);
1476
1477 /* Restore reindex state. */
1478 reindexspace = shm_toc_lookup(toc, PARALLEL_KEY_REINDEX_STATE, false);
1479 RestoreReindexState(reindexspace);
1480
1481 /* Restore relmapper state. */
1482 relmapperspace = shm_toc_lookup(toc, PARALLEL_KEY_RELMAPPER_STATE, false);
1483 RestoreRelationMap(relmapperspace);
1484
1485 /* Restore uncommitted enums. */
1486 uncommittedenumsspace = shm_toc_lookup(toc, PARALLEL_KEY_UNCOMMITTEDENUMS,
1487 false);
1488 RestoreUncommittedEnums(uncommittedenumsspace);
1489
1490 /* Restore the ClientConnectionInfo. */
1491 clientconninfospace = shm_toc_lookup(toc, PARALLEL_KEY_CLIENTCONNINFO,
1492 false);
1493 RestoreClientConnectionInfo(clientconninfospace);
1494
1495 /*
1496 * Initialize SystemUser now that MyClientConnectionInfo is restored. Also
1497 * ensure that auth_method is actually valid, aka authn_id is not NULL.
1498 */
1499 if (MyClientConnectionInfo.authn_id)
1500 InitializeSystemUser(MyClientConnectionInfo.authn_id,
1501 hba_authname(MyClientConnectionInfo.auth_method));
1502
1503 /* Attach to the leader's serializable transaction, if SERIALIZABLE. */
1504 AttachSerializableXact(fps->serializable_xact_handle);
1505
1506 /*
1507 * We've initialized all of our state now; nothing should change
1508 * hereafter.
1509 */
1510 InitializingParallelWorker = false;
1511 EnterParallelMode();
1512
1513 /*
1514 * Time to do the real work: invoke the caller-supplied code.
1515 */
1516 entrypt(seg, toc);
1517
1518 /* Must exit parallel mode to pop active snapshot. */
1519 ExitParallelMode();
1520
1521 /* Must pop active snapshot so snapmgr.c doesn't complain. */
1522 PopActiveSnapshot();
1523
1524 /* Shut down the parallel-worker transaction. */
1525 EndParallelWorkerTransaction();
1526
1527 /* Detach from the per-session DSM segment. */
1528 DetachSession();
1529
1530 /* Report success. */
1531 pq_putmessage(PqMsg_Terminate, NULL, 0);
1532 }
1533
1534 /*
1535 * Update shared memory with the ending location of the last WAL record we
1536 * wrote, if it's greater than the value already stored there.
1537 */
1538 void
1539 ParallelWorkerReportLastRecEnd(XLogRecPtr last_xlog_end)
1540 {
1541 FixedParallelState *fps = MyFixedParallelState;
1542
1543 Assert(fps != NULL);
1544 SpinLockAcquire(&fps->mutex);
1545 if (fps->last_xlog_end < last_xlog_end)
1546 fps->last_xlog_end = last_xlog_end;
1547 SpinLockRelease(&fps->mutex);
1548 }
1549
1550 /*
1551 * Make sure the leader tries to read from our error queue one more time.
1552 * This guards against the case where we exit uncleanly without sending an
1553 * ErrorResponse to the leader, for example because some code calls proc_exit
1554 * directly.
1555 *
1556 * Also explicitly detach from dsm segment so that subsystems using
1557 * on_dsm_detach() have a chance to send stats before the stats subsystem is
1558 * shut down as part of a before_shmem_exit() hook.
1559 *
1560 * One might think this could instead be solved by carefully ordering the
1561 * attaching to dsm segments, so that the pgstats segments get detached from
1562 * later than the parallel query one. That turns out to not work because the
1563 * stats hash might need to grow which can cause new segments to be allocated,
1564 * which then will be detached from earlier.
1565 */
1566 static void
1567 ParallelWorkerShutdown(int code, Datum arg)
1568 {
1569 SendProcSignal(ParallelLeaderPid,
1570 PROCSIG_PARALLEL_MESSAGE,
1571 ParallelLeaderProcNumber);
1572
1573 dsm_detach((dsm_segment *) DatumGetPointer(arg));
1574 }
1575
1576 /*
1577 * Look up (and possibly load) a parallel worker entry point function.
1578 *
1579 * For functions contained in the core code, we use library name "postgres"
1580 * and consult the InternalParallelWorkers array. External functions are
1581 * looked up, and loaded if necessary, using load_external_function().
1582 *
1583 * The point of this is to pass function names as strings across process
1584 * boundaries. We can't pass actual function addresses because of the
1585 * possibility that the function has been loaded at a different address
1586 * in a different process. This is obviously a hazard for functions in
1587 * loadable libraries, but it can happen even for functions in the core code
1588 * on platforms using EXEC_BACKEND (e.g., Windows).
1589 *
1590 * At some point it might be worthwhile to get rid of InternalParallelWorkers[]
1591 * in favor of applying load_external_function() for core functions too;
1592 * but that raises portability issues that are not worth addressing now.
1593 */
1594 static parallel_worker_main_type
1595 LookupParallelWorkerFunction(const char *libraryname, const char *funcname)
1596 {
1597 /*
1598 * If the function is to be loaded from postgres itself, search the
1599 * InternalParallelWorkers array.
1600 */
1601 if (strcmp(libraryname, "postgres") == 0)
1602 {
1603 int i;
1604
1605 for (i = 0; i < lengthof(InternalParallelWorkers); i++)
1606 {
1607 if (strcmp(InternalParallelWorkers[i].fn_name, funcname) == 0)
1608 return InternalParallelWorkers[i].fn_addr;
1609 }
1610
1611 /* We can only reach this by programming error. */
1612 elog(ERROR, "internal function \"%s\" not found", funcname);
1613 }
1614
1615 /* Otherwise load from external library. */
1616 return (parallel_worker_main_type)
1617 load_external_function(libraryname, funcname, true, NULL);
1618 }
Postgresql Clone from git://git.postgresql.org/git/postgresql.git