1 /*-------------------------------------------------------------------------
2 *
3 * parallel.c
4 * Infrastructure for launching parallel workers
5 *
6 * Portions Copyright (c) 1996-2021, 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/heapam.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 "executor/execParallel.h"
29 #include "libpq/libpq.h"
30 #include "libpq/pqformat.h"
31 #include "libpq/pqmq.h"
32 #include "miscadmin.h"
33 #include "optimizer/optimizer.h"
34 #include "pgstat.h"
35 #include "storage/ipc.h"
36 #include "storage/predicate.h"
37 #include "storage/sinval.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 #include "utils/typcache.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
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 BackendId parallel_leader_backend_id;
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 bool 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 "parallel_vacuum_main", parallel_vacuum_main
148 }
149 };
150
151 /* Private functions. */
152 static void HandleParallelMessage(ParallelContext *pcxt, int i, StringInfo msg);
153 static void WaitForParallelWorkersToExit(ParallelContext *pcxt);
154 static parallel_worker_main_type LookupParallelWorkerFunction(const char *libraryname, const char *funcname);
155 static void ParallelWorkerShutdown(int code, Datum arg);
156
157
158 /*
159 * Establish a new parallel context. This should be done after entering
160 * parallel mode, and (unless there is an error) the context should be
161 * destroyed before exiting the current subtransaction.
162 */
163 ParallelContext *
CreateParallelContext(const char * library_name,const char * function_name,int nworkers)164 CreateParallelContext(const char *library_name, const char *function_name,
165 int nworkers)
166 {
167 MemoryContext oldcontext;
168 ParallelContext *pcxt;
169
170 /* It is unsafe to create a parallel context if not in parallel mode. */
171 Assert(IsInParallelMode());
172
173 /* Number of workers should be non-negative. */
174 Assert(nworkers >= 0);
175
176 /* We might be running in a short-lived memory context. */
177 oldcontext = MemoryContextSwitchTo(TopTransactionContext);
178
179 /* Initialize a new ParallelContext. */
180 pcxt = palloc0(sizeof(ParallelContext));
181 pcxt->subid = GetCurrentSubTransactionId();
182 pcxt->nworkers = nworkers;
183 pcxt->nworkers_to_launch = nworkers;
184 pcxt->library_name = pstrdup(library_name);
185 pcxt->function_name = pstrdup(function_name);
186 pcxt->error_context_stack = error_context_stack;
187 shm_toc_initialize_estimator(&pcxt->estimator);
188 dlist_push_head(&pcxt_list, &pcxt->node);
189
190 /* Restore previous memory context. */
191 MemoryContextSwitchTo(oldcontext);
192
193 return pcxt;
194 }
195
196 /*
197 * Establish the dynamic shared memory segment for a parallel context and
198 * copy state and other bookkeeping information that will be needed by
199 * parallel workers into it.
200 */
201 void
InitializeParallelDSM(ParallelContext * pcxt)202 InitializeParallelDSM(ParallelContext *pcxt)
203 {
204 MemoryContext oldcontext;
205 Size library_len = 0;
206 Size guc_len = 0;
207 Size combocidlen = 0;
208 Size tsnaplen = 0;
209 Size asnaplen = 0;
210 Size tstatelen = 0;
211 Size pendingsyncslen = 0;
212 Size reindexlen = 0;
213 Size relmapperlen = 0;
214 Size uncommittedenumslen = 0;
215 Size segsize = 0;
216 int i;
217 FixedParallelState *fps;
218 dsm_handle session_dsm_handle = DSM_HANDLE_INVALID;
219 Snapshot transaction_snapshot = GetTransactionSnapshot();
220 Snapshot active_snapshot = GetActiveSnapshot();
221
222 /* We might be running in a very short-lived memory context. */
223 oldcontext = MemoryContextSwitchTo(TopTransactionContext);
224
225 /* Allow space to store the fixed-size parallel state. */
226 shm_toc_estimate_chunk(&pcxt->estimator, sizeof(FixedParallelState));
227 shm_toc_estimate_keys(&pcxt->estimator, 1);
228
229 /*
230 * Normally, the user will have requested at least one worker process, but
231 * if by chance they have not, we can skip a bunch of things here.
232 */
233 if (pcxt->nworkers > 0)
234 {
235 /* Get (or create) the per-session DSM segment's handle. */
236 session_dsm_handle = GetSessionDsmHandle();
237
238 /*
239 * If we weren't able to create a per-session DSM segment, then we can
240 * continue but we can't safely launch any workers because their
241 * record typmods would be incompatible so they couldn't exchange
242 * tuples.
243 */
244 if (session_dsm_handle == DSM_HANDLE_INVALID)
245 pcxt->nworkers = 0;
246 }
247
248 if (pcxt->nworkers > 0)
249 {
250 /* Estimate space for various kinds of state sharing. */
251 library_len = EstimateLibraryStateSpace();
252 shm_toc_estimate_chunk(&pcxt->estimator, library_len);
253 guc_len = EstimateGUCStateSpace();
254 shm_toc_estimate_chunk(&pcxt->estimator, guc_len);
255 combocidlen = EstimateComboCIDStateSpace();
256 shm_toc_estimate_chunk(&pcxt->estimator, combocidlen);
257 if (IsolationUsesXactSnapshot())
258 {
259 tsnaplen = EstimateSnapshotSpace(transaction_snapshot);
260 shm_toc_estimate_chunk(&pcxt->estimator, tsnaplen);
261 }
262 asnaplen = EstimateSnapshotSpace(active_snapshot);
263 shm_toc_estimate_chunk(&pcxt->estimator, asnaplen);
264 tstatelen = EstimateTransactionStateSpace();
265 shm_toc_estimate_chunk(&pcxt->estimator, tstatelen);
266 shm_toc_estimate_chunk(&pcxt->estimator, sizeof(dsm_handle));
267 pendingsyncslen = EstimatePendingSyncsSpace();
268 shm_toc_estimate_chunk(&pcxt->estimator, pendingsyncslen);
269 reindexlen = EstimateReindexStateSpace();
270 shm_toc_estimate_chunk(&pcxt->estimator, reindexlen);
271 relmapperlen = EstimateRelationMapSpace();
272 shm_toc_estimate_chunk(&pcxt->estimator, relmapperlen);
273 uncommittedenumslen = EstimateUncommittedEnumsSpace();
274 shm_toc_estimate_chunk(&pcxt->estimator, uncommittedenumslen);
275 /* If you add more chunks here, you probably need to add keys. */
276 shm_toc_estimate_keys(&pcxt->estimator, 11);
277
278 /* Estimate space need for error queues. */
279 StaticAssertStmt(BUFFERALIGN(PARALLEL_ERROR_QUEUE_SIZE) ==
280 PARALLEL_ERROR_QUEUE_SIZE,
281 "parallel error queue size not buffer-aligned");
282 shm_toc_estimate_chunk(&pcxt->estimator,
283 mul_size(PARALLEL_ERROR_QUEUE_SIZE,
284 pcxt->nworkers));
285 shm_toc_estimate_keys(&pcxt->estimator, 1);
286
287 /* Estimate how much we'll need for the entrypoint info. */
288 shm_toc_estimate_chunk(&pcxt->estimator, strlen(pcxt->library_name) +
289 strlen(pcxt->function_name) + 2);
290 shm_toc_estimate_keys(&pcxt->estimator, 1);
291 }
292
293 /*
294 * Create DSM and initialize with new table of contents. But if the user
295 * didn't request any workers, then don't bother creating a dynamic shared
296 * memory segment; instead, just use backend-private memory.
297 *
298 * Also, if we can't create a dynamic shared memory segment because the
299 * maximum number of segments have already been created, then fall back to
300 * backend-private memory, and plan not to use any workers. We hope this
301 * won't happen very often, but it's better to abandon the use of
302 * parallelism than to fail outright.
303 */
304 segsize = shm_toc_estimate(&pcxt->estimator);
305 if (pcxt->nworkers > 0)
306 pcxt->seg = dsm_create(segsize, DSM_CREATE_NULL_IF_MAXSEGMENTS);
307 if (pcxt->seg != NULL)
308 pcxt->toc = shm_toc_create(PARALLEL_MAGIC,
309 dsm_segment_address(pcxt->seg),
310 segsize);
311 else
312 {
313 pcxt->nworkers = 0;
314 pcxt->private_memory = MemoryContextAlloc(TopMemoryContext, segsize);
315 pcxt->toc = shm_toc_create(PARALLEL_MAGIC, pcxt->private_memory,
316 segsize);
317 }
318
319 /* Initialize fixed-size state in shared memory. */
320 fps = (FixedParallelState *)
321 shm_toc_allocate(pcxt->toc, sizeof(FixedParallelState));
322 fps->database_id = MyDatabaseId;
323 fps->authenticated_user_id = GetAuthenticatedUserId();
324 fps->outer_user_id = GetCurrentRoleId();
325 fps->is_superuser = session_auth_is_superuser;
326 GetUserIdAndSecContext(&fps->current_user_id, &fps->sec_context);
327 GetTempNamespaceState(&fps->temp_namespace_id,
328 &fps->temp_toast_namespace_id);
329 fps->parallel_leader_pgproc = MyProc;
330 fps->parallel_leader_pid = MyProcPid;
331 fps->parallel_leader_backend_id = MyBackendId;
332 fps->xact_ts = GetCurrentTransactionStartTimestamp();
333 fps->stmt_ts = GetCurrentStatementStartTimestamp();
334 fps->serializable_xact_handle = ShareSerializableXact();
335 SpinLockInit(&fps->mutex);
336 fps->last_xlog_end = 0;
337 shm_toc_insert(pcxt->toc, PARALLEL_KEY_FIXED, fps);
338
339 /* We can skip the rest of this if we're not budgeting for any workers. */
340 if (pcxt->nworkers > 0)
341 {
342 char *libraryspace;
343 char *gucspace;
344 char *combocidspace;
345 char *tsnapspace;
346 char *asnapspace;
347 char *tstatespace;
348 char *pendingsyncsspace;
349 char *reindexspace;
350 char *relmapperspace;
351 char *error_queue_space;
352 char *session_dsm_handle_space;
353 char *entrypointstate;
354 char *uncommittedenumsspace;
355 Size lnamelen;
356
357 /* Serialize shared libraries we have loaded. */
358 libraryspace = shm_toc_allocate(pcxt->toc, library_len);
359 SerializeLibraryState(library_len, libraryspace);
360 shm_toc_insert(pcxt->toc, PARALLEL_KEY_LIBRARY, libraryspace);
361
362 /* Serialize GUC settings. */
363 gucspace = shm_toc_allocate(pcxt->toc, guc_len);
364 SerializeGUCState(guc_len, gucspace);
365 shm_toc_insert(pcxt->toc, PARALLEL_KEY_GUC, gucspace);
366
367 /* Serialize combo CID state. */
368 combocidspace = shm_toc_allocate(pcxt->toc, combocidlen);
369 SerializeComboCIDState(combocidlen, combocidspace);
370 shm_toc_insert(pcxt->toc, PARALLEL_KEY_COMBO_CID, combocidspace);
371
372 /*
373 * Serialize the transaction snapshot if the transaction
374 * isolation-level uses a transaction snapshot.
375 */
376 if (IsolationUsesXactSnapshot())
377 {
378 tsnapspace = shm_toc_allocate(pcxt->toc, tsnaplen);
379 SerializeSnapshot(transaction_snapshot, tsnapspace);
380 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT,
381 tsnapspace);
382 }
383
384 /* Serialize the active snapshot. */
385 asnapspace = shm_toc_allocate(pcxt->toc, asnaplen);
386 SerializeSnapshot(active_snapshot, asnapspace);
387 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ACTIVE_SNAPSHOT, asnapspace);
388
389 /* Provide the handle for per-session segment. */
390 session_dsm_handle_space = shm_toc_allocate(pcxt->toc,
391 sizeof(dsm_handle));
392 *(dsm_handle *) session_dsm_handle_space = session_dsm_handle;
393 shm_toc_insert(pcxt->toc, PARALLEL_KEY_SESSION_DSM,
394 session_dsm_handle_space);
395
396 /* Serialize transaction state. */
397 tstatespace = shm_toc_allocate(pcxt->toc, tstatelen);
398 SerializeTransactionState(tstatelen, tstatespace);
399 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_STATE, tstatespace);
400
401 /* Serialize pending syncs. */
402 pendingsyncsspace = shm_toc_allocate(pcxt->toc, pendingsyncslen);
403 SerializePendingSyncs(pendingsyncslen, pendingsyncsspace);
404 shm_toc_insert(pcxt->toc, PARALLEL_KEY_PENDING_SYNCS,
405 pendingsyncsspace);
406
407 /* Serialize reindex state. */
408 reindexspace = shm_toc_allocate(pcxt->toc, reindexlen);
409 SerializeReindexState(reindexlen, reindexspace);
410 shm_toc_insert(pcxt->toc, PARALLEL_KEY_REINDEX_STATE, reindexspace);
411
412 /* Serialize relmapper state. */
413 relmapperspace = shm_toc_allocate(pcxt->toc, relmapperlen);
414 SerializeRelationMap(relmapperlen, relmapperspace);
415 shm_toc_insert(pcxt->toc, PARALLEL_KEY_RELMAPPER_STATE,
416 relmapperspace);
417
418 /* Serialize uncommitted enum state. */
419 uncommittedenumsspace = shm_toc_allocate(pcxt->toc,
420 uncommittedenumslen);
421 SerializeUncommittedEnums(uncommittedenumsspace, uncommittedenumslen);
422 shm_toc_insert(pcxt->toc, PARALLEL_KEY_UNCOMMITTEDENUMS,
423 uncommittedenumsspace);
424
425 /* Allocate space for worker information. */
426 pcxt->worker = palloc0(sizeof(ParallelWorkerInfo) * pcxt->nworkers);
427
428 /*
429 * Establish error queues in dynamic shared memory.
430 *
431 * These queues should be used only for transmitting ErrorResponse,
432 * NoticeResponse, and NotifyResponse protocol messages. Tuple data
433 * should be transmitted via separate (possibly larger?) queues.
434 */
435 error_queue_space =
436 shm_toc_allocate(pcxt->toc,
437 mul_size(PARALLEL_ERROR_QUEUE_SIZE,
438 pcxt->nworkers));
439 for (i = 0; i < pcxt->nworkers; ++i)
440 {
441 char *start;
442 shm_mq *mq;
443
444 start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
445 mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
446 shm_mq_set_receiver(mq, MyProc);
447 pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
448 }
449 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE, error_queue_space);
450
451 /*
452 * Serialize entrypoint information. It's unsafe to pass function
453 * pointers across processes, as the function pointer may be different
454 * in each process in EXEC_BACKEND builds, so we always pass library
455 * and function name. (We use library name "postgres" for functions
456 * in the core backend.)
457 */
458 lnamelen = strlen(pcxt->library_name);
459 entrypointstate = shm_toc_allocate(pcxt->toc, lnamelen +
460 strlen(pcxt->function_name) + 2);
461 strcpy(entrypointstate, pcxt->library_name);
462 strcpy(entrypointstate + lnamelen + 1, pcxt->function_name);
463 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ENTRYPOINT, entrypointstate);
464 }
465
466 /* Restore previous memory context. */
467 MemoryContextSwitchTo(oldcontext);
468 }
469
470 /*
471 * Reinitialize the dynamic shared memory segment for a parallel context such
472 * that we could launch workers for it again.
473 */
474 void
ReinitializeParallelDSM(ParallelContext * pcxt)475 ReinitializeParallelDSM(ParallelContext *pcxt)
476 {
477 FixedParallelState *fps;
478
479 /* Wait for any old workers to exit. */
480 if (pcxt->nworkers_launched > 0)
481 {
482 WaitForParallelWorkersToFinish(pcxt);
483 WaitForParallelWorkersToExit(pcxt);
484 pcxt->nworkers_launched = 0;
485 if (pcxt->known_attached_workers)
486 {
487 pfree(pcxt->known_attached_workers);
488 pcxt->known_attached_workers = NULL;
489 pcxt->nknown_attached_workers = 0;
490 }
491 }
492
493 /* Reset a few bits of fixed parallel state to a clean state. */
494 fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED, false);
495 fps->last_xlog_end = 0;
496
497 /* Recreate error queues (if they exist). */
498 if (pcxt->nworkers > 0)
499 {
500 char *error_queue_space;
501 int i;
502
503 error_queue_space =
504 shm_toc_lookup(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE, false);
505 for (i = 0; i < pcxt->nworkers; ++i)
506 {
507 char *start;
508 shm_mq *mq;
509
510 start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
511 mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
512 shm_mq_set_receiver(mq, MyProc);
513 pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
514 }
515 }
516 }
517
518 /*
519 * Reinitialize parallel workers for a parallel context such that we could
520 * launch a different number of workers. This is required for cases where
521 * we need to reuse the same DSM segment, but the number of workers can
522 * vary from run-to-run.
523 */
524 void
ReinitializeParallelWorkers(ParallelContext * pcxt,int nworkers_to_launch)525 ReinitializeParallelWorkers(ParallelContext *pcxt, int nworkers_to_launch)
526 {
527 /*
528 * The number of workers that need to be launched must be less than the
529 * number of workers with which the parallel context is initialized.
530 */
531 Assert(pcxt->nworkers >= nworkers_to_launch);
532 pcxt->nworkers_to_launch = nworkers_to_launch;
533 }
534
535 /*
536 * Launch parallel workers.
537 */
538 void
LaunchParallelWorkers(ParallelContext * pcxt)539 LaunchParallelWorkers(ParallelContext *pcxt)
540 {
541 MemoryContext oldcontext;
542 BackgroundWorker worker;
543 int i;
544 bool any_registrations_failed = false;
545
546 /* Skip this if we have no workers. */
547 if (pcxt->nworkers == 0 || pcxt->nworkers_to_launch == 0)
548 return;
549
550 /* We need to be a lock group leader. */
551 BecomeLockGroupLeader();
552
553 /* If we do have workers, we'd better have a DSM segment. */
554 Assert(pcxt->seg != NULL);
555
556 /* We might be running in a short-lived memory context. */
557 oldcontext = MemoryContextSwitchTo(TopTransactionContext);
558
559 /* Configure a worker. */
560 memset(&worker, 0, sizeof(worker));
561 snprintf(worker.bgw_name, BGW_MAXLEN, "parallel worker for PID %d",
562 MyProcPid);
563 snprintf(worker.bgw_type, BGW_MAXLEN, "parallel worker");
564 worker.bgw_flags =
565 BGWORKER_SHMEM_ACCESS | BGWORKER_BACKEND_DATABASE_CONNECTION
566 | BGWORKER_CLASS_PARALLEL;
567 worker.bgw_start_time = BgWorkerStart_ConsistentState;
568 worker.bgw_restart_time = BGW_NEVER_RESTART;
569 sprintf(worker.bgw_library_name, "postgres");
570 sprintf(worker.bgw_function_name, "ParallelWorkerMain");
571 worker.bgw_main_arg = UInt32GetDatum(dsm_segment_handle(pcxt->seg));
572 worker.bgw_notify_pid = MyProcPid;
573
574 /*
575 * Start workers.
576 *
577 * The caller must be able to tolerate ending up with fewer workers than
578 * expected, so there is no need to throw an error here if registration
579 * fails. It wouldn't help much anyway, because registering the worker in
580 * no way guarantees that it will start up and initialize successfully.
581 */
582 for (i = 0; i < pcxt->nworkers_to_launch; ++i)
583 {
584 memcpy(worker.bgw_extra, &i, sizeof(int));
585 if (!any_registrations_failed &&
586 RegisterDynamicBackgroundWorker(&worker,
587 &pcxt->worker[i].bgwhandle))
588 {
589 shm_mq_set_handle(pcxt->worker[i].error_mqh,
590 pcxt->worker[i].bgwhandle);
591 pcxt->nworkers_launched++;
592 }
593 else
594 {
595 /*
596 * If we weren't able to register the worker, then we've bumped up
597 * against the max_worker_processes limit, and future
598 * registrations will probably fail too, so arrange to skip them.
599 * But we still have to execute this code for the remaining slots
600 * to make sure that we forget about the error queues we budgeted
601 * for those workers. Otherwise, we'll wait for them to start,
602 * but they never will.
603 */
604 any_registrations_failed = true;
605 pcxt->worker[i].bgwhandle = NULL;
606 shm_mq_detach(pcxt->worker[i].error_mqh);
607 pcxt->worker[i].error_mqh = NULL;
608 }
609 }
610
611 /*
612 * Now that nworkers_launched has taken its final value, we can initialize
613 * known_attached_workers.
614 */
615 if (pcxt->nworkers_launched > 0)
616 {
617 pcxt->known_attached_workers =
618 palloc0(sizeof(bool) * pcxt->nworkers_launched);
619 pcxt->nknown_attached_workers = 0;
620 }
621
622 /* Restore previous memory context. */
623 MemoryContextSwitchTo(oldcontext);
624 }
625
626 /*
627 * Wait for all workers to attach to their error queues, and throw an error if
628 * any worker fails to do this.
629 *
630 * Callers can assume that if this function returns successfully, then the
631 * number of workers given by pcxt->nworkers_launched have initialized and
632 * attached to their error queues. Whether or not these workers are guaranteed
633 * to still be running depends on what code the caller asked them to run;
634 * this function does not guarantee that they have not exited. However, it
635 * does guarantee that any workers which exited must have done so cleanly and
636 * after successfully performing the work with which they were tasked.
637 *
638 * If this function is not called, then some of the workers that were launched
639 * may not have been started due to a fork() failure, or may have exited during
640 * early startup prior to attaching to the error queue, so nworkers_launched
641 * cannot be viewed as completely reliable. It will never be less than the
642 * number of workers which actually started, but it might be more. Any workers
643 * that failed to start will still be discovered by
644 * WaitForParallelWorkersToFinish and an error will be thrown at that time,
645 * provided that function is eventually reached.
646 *
647 * In general, the leader process should do as much work as possible before
648 * calling this function. fork() failures and other early-startup failures
649 * are very uncommon, and having the leader sit idle when it could be doing
650 * useful work is undesirable. However, if the leader needs to wait for
651 * all of its workers or for a specific worker, it may want to call this
652 * function before doing so. If not, it must make some other provision for
653 * the failure-to-start case, lest it wait forever. On the other hand, a
654 * leader which never waits for a worker that might not be started yet, or
655 * at least never does so prior to WaitForParallelWorkersToFinish(), need not
656 * call this function at all.
657 */
658 void
WaitForParallelWorkersToAttach(ParallelContext * pcxt)659 WaitForParallelWorkersToAttach(ParallelContext *pcxt)
660 {
661 int i;
662
663 /* Skip this if we have no launched workers. */
664 if (pcxt->nworkers_launched == 0)
665 return;
666
667 for (;;)
668 {
669 /*
670 * This will process any parallel messages that are pending and it may
671 * also throw an error propagated from a worker.
672 */
673 CHECK_FOR_INTERRUPTS();
674
675 for (i = 0; i < pcxt->nworkers_launched; ++i)
676 {
677 BgwHandleStatus status;
678 shm_mq *mq;
679 int rc;
680 pid_t pid;
681
682 if (pcxt->known_attached_workers[i])
683 continue;
684
685 /*
686 * If error_mqh is NULL, then the worker has already exited
687 * cleanly.
688 */
689 if (pcxt->worker[i].error_mqh == NULL)
690 {
691 pcxt->known_attached_workers[i] = true;
692 ++pcxt->nknown_attached_workers;
693 continue;
694 }
695
696 status = GetBackgroundWorkerPid(pcxt->worker[i].bgwhandle, &pid);
697 if (status == BGWH_STARTED)
698 {
699 /* Has the worker attached to the error queue? */
700 mq = shm_mq_get_queue(pcxt->worker[i].error_mqh);
701 if (shm_mq_get_sender(mq) != NULL)
702 {
703 /* Yes, so it is known to be attached. */
704 pcxt->known_attached_workers[i] = true;
705 ++pcxt->nknown_attached_workers;
706 }
707 }
708 else if (status == BGWH_STOPPED)
709 {
710 /*
711 * If the worker stopped without attaching to the error queue,
712 * throw an error.
713 */
714 mq = shm_mq_get_queue(pcxt->worker[i].error_mqh);
715 if (shm_mq_get_sender(mq) == NULL)
716 ereport(ERROR,
717 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
718 errmsg("parallel worker failed to initialize"),
719 errhint("More details may be available in the server log.")));
720
721 pcxt->known_attached_workers[i] = true;
722 ++pcxt->nknown_attached_workers;
723 }
724 else
725 {
726 /*
727 * Worker not yet started, so we must wait. The postmaster
728 * will notify us if the worker's state changes. Our latch
729 * might also get set for some other reason, but if so we'll
730 * just end up waiting for the same worker again.
731 */
732 rc = WaitLatch(MyLatch,
733 WL_LATCH_SET | WL_EXIT_ON_PM_DEATH,
734 -1, WAIT_EVENT_BGWORKER_STARTUP);
735
736 if (rc & WL_LATCH_SET)
737 ResetLatch(MyLatch);
738 }
739 }
740
741 /* If all workers are known to have started, we're done. */
742 if (pcxt->nknown_attached_workers >= pcxt->nworkers_launched)
743 {
744 Assert(pcxt->nknown_attached_workers == pcxt->nworkers_launched);
745 break;
746 }
747 }
748 }
749
750 /*
751 * Wait for all workers to finish computing.
752 *
753 * Even if the parallel operation seems to have completed successfully, it's
754 * important to call this function afterwards. We must not miss any errors
755 * the workers may have thrown during the parallel operation, or any that they
756 * may yet throw while shutting down.
757 *
758 * Also, we want to update our notion of XactLastRecEnd based on worker
759 * feedback.
760 */
761 void
WaitForParallelWorkersToFinish(ParallelContext * pcxt)762 WaitForParallelWorkersToFinish(ParallelContext *pcxt)
763 {
764 for (;;)
765 {
766 bool anyone_alive = false;
767 int nfinished = 0;
768 int i;
769
770 /*
771 * This will process any parallel messages that are pending, which may
772 * change the outcome of the loop that follows. It may also throw an
773 * error propagated from a worker.
774 */
775 CHECK_FOR_INTERRUPTS();
776
777 for (i = 0; i < pcxt->nworkers_launched; ++i)
778 {
779 /*
780 * If error_mqh is NULL, then the worker has already exited
781 * cleanly. If we have received a message through error_mqh from
782 * the worker, we know it started up cleanly, and therefore we're
783 * certain to be notified when it exits.
784 */
785 if (pcxt->worker[i].error_mqh == NULL)
786 ++nfinished;
787 else if (pcxt->known_attached_workers[i])
788 {
789 anyone_alive = true;
790 break;
791 }
792 }
793
794 if (!anyone_alive)
795 {
796 /* If all workers are known to have finished, we're done. */
797 if (nfinished >= pcxt->nworkers_launched)
798 {
799 Assert(nfinished == pcxt->nworkers_launched);
800 break;
801 }
802
803 /*
804 * We didn't detect any living workers, but not all workers are
805 * known to have exited cleanly. Either not all workers have
806 * launched yet, or maybe some of them failed to start or
807 * terminated abnormally.
808 */
809 for (i = 0; i < pcxt->nworkers_launched; ++i)
810 {
811 pid_t pid;
812 shm_mq *mq;
813
814 /*
815 * If the worker is BGWH_NOT_YET_STARTED or BGWH_STARTED, we
816 * should just keep waiting. If it is BGWH_STOPPED, then
817 * further investigation is needed.
818 */
819 if (pcxt->worker[i].error_mqh == NULL ||
820 pcxt->worker[i].bgwhandle == NULL ||
821 GetBackgroundWorkerPid(pcxt->worker[i].bgwhandle,
822 &pid) != BGWH_STOPPED)
823 continue;
824
825 /*
826 * Check whether the worker ended up stopped without ever
827 * attaching to the error queue. If so, the postmaster was
828 * unable to fork the worker or it exited without initializing
829 * properly. We must throw an error, since the caller may
830 * have been expecting the worker to do some work before
831 * exiting.
832 */
833 mq = shm_mq_get_queue(pcxt->worker[i].error_mqh);
834 if (shm_mq_get_sender(mq) == NULL)
835 ereport(ERROR,
836 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
837 errmsg("parallel worker failed to initialize"),
838 errhint("More details may be available in the server log.")));
839
840 /*
841 * The worker is stopped, but is attached to the error queue.
842 * Unless there's a bug somewhere, this will only happen when
843 * the worker writes messages and terminates after the
844 * CHECK_FOR_INTERRUPTS() near the top of this function and
845 * before the call to GetBackgroundWorkerPid(). In that case,
846 * or latch should have been set as well and the right things
847 * will happen on the next pass through the loop.
848 */
849 }
850 }
851
852 (void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, -1,
853 WAIT_EVENT_PARALLEL_FINISH);
854 ResetLatch(MyLatch);
855 }
856
857 if (pcxt->toc != NULL)
858 {
859 FixedParallelState *fps;
860
861 fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED, false);
862 if (fps->last_xlog_end > XactLastRecEnd)
863 XactLastRecEnd = fps->last_xlog_end;
864 }
865 }
866
867 /*
868 * Wait for all workers to exit.
869 *
870 * This function ensures that workers have been completely shutdown. The
871 * difference between WaitForParallelWorkersToFinish and this function is
872 * that the former just ensures that last message sent by a worker backend is
873 * received by the leader backend whereas this ensures the complete shutdown.
874 */
875 static void
WaitForParallelWorkersToExit(ParallelContext * pcxt)876 WaitForParallelWorkersToExit(ParallelContext *pcxt)
877 {
878 int i;
879
880 /* Wait until the workers actually die. */
881 for (i = 0; i < pcxt->nworkers_launched; ++i)
882 {
883 BgwHandleStatus status;
884
885 if (pcxt->worker == NULL || pcxt->worker[i].bgwhandle == NULL)
886 continue;
887
888 status = WaitForBackgroundWorkerShutdown(pcxt->worker[i].bgwhandle);
889
890 /*
891 * If the postmaster kicked the bucket, we have no chance of cleaning
892 * up safely -- we won't be able to tell when our workers are actually
893 * dead. This doesn't necessitate a PANIC since they will all abort
894 * eventually, but we can't safely continue this session.
895 */
896 if (status == BGWH_POSTMASTER_DIED)
897 ereport(FATAL,
898 (errcode(ERRCODE_ADMIN_SHUTDOWN),
899 errmsg("postmaster exited during a parallel transaction")));
900
901 /* Release memory. */
902 pfree(pcxt->worker[i].bgwhandle);
903 pcxt->worker[i].bgwhandle = NULL;
904 }
905 }
906
907 /*
908 * Destroy a parallel context.
909 *
910 * If expecting a clean exit, you should use WaitForParallelWorkersToFinish()
911 * first, before calling this function. When this function is invoked, any
912 * remaining workers are forcibly killed; the dynamic shared memory segment
913 * is unmapped; and we then wait (uninterruptibly) for the workers to exit.
914 */
915 void
DestroyParallelContext(ParallelContext * pcxt)916 DestroyParallelContext(ParallelContext *pcxt)
917 {
918 int i;
919
920 /*
921 * Be careful about order of operations here! We remove the parallel
922 * context from the list before we do anything else; otherwise, if an
923 * error occurs during a subsequent step, we might try to nuke it again
924 * from AtEOXact_Parallel or AtEOSubXact_Parallel.
925 */
926 dlist_delete(&pcxt->node);
927
928 /* Kill each worker in turn, and forget their error queues. */
929 if (pcxt->worker != NULL)
930 {
931 for (i = 0; i < pcxt->nworkers_launched; ++i)
932 {
933 if (pcxt->worker[i].error_mqh != NULL)
934 {
935 TerminateBackgroundWorker(pcxt->worker[i].bgwhandle);
936
937 shm_mq_detach(pcxt->worker[i].error_mqh);
938 pcxt->worker[i].error_mqh = NULL;
939 }
940 }
941 }
942
943 /*
944 * If we have allocated a shared memory segment, detach it. This will
945 * implicitly detach the error queues, and any other shared memory queues,
946 * stored there.
947 */
948 if (pcxt->seg != NULL)
949 {
950 dsm_detach(pcxt->seg);
951 pcxt->seg = NULL;
952 }
953
954 /*
955 * If this parallel context is actually in backend-private memory rather
956 * than shared memory, free that memory instead.
957 */
958 if (pcxt->private_memory != NULL)
959 {
960 pfree(pcxt->private_memory);
961 pcxt->private_memory = NULL;
962 }
963
964 /*
965 * We can't finish transaction commit or abort until all of the workers
966 * have exited. This means, in particular, that we can't respond to
967 * interrupts at this stage.
968 */
969 HOLD_INTERRUPTS();
970 WaitForParallelWorkersToExit(pcxt);
971 RESUME_INTERRUPTS();
972
973 /* Free the worker array itself. */
974 if (pcxt->worker != NULL)
975 {
976 pfree(pcxt->worker);
977 pcxt->worker = NULL;
978 }
979
980 /* Free memory. */
981 pfree(pcxt->library_name);
982 pfree(pcxt->function_name);
983 pfree(pcxt);
984 }
985
986 /*
987 * Are there any parallel contexts currently active?
988 */
989 bool
ParallelContextActive(void)990 ParallelContextActive(void)
991 {
992 return !dlist_is_empty(&pcxt_list);
993 }
994
995 /*
996 * Handle receipt of an interrupt indicating a parallel worker message.
997 *
998 * Note: this is called within a signal handler! All we can do is set
999 * a flag that will cause the next CHECK_FOR_INTERRUPTS() to invoke
1000 * HandleParallelMessages().
1001 */
1002 void
HandleParallelMessageInterrupt(void)1003 HandleParallelMessageInterrupt(void)
1004 {
1005 InterruptPending = true;
1006 ParallelMessagePending = true;
1007 SetLatch(MyLatch);
1008 }
1009
1010 /*
1011 * Handle any queued protocol messages received from parallel workers.
1012 */
1013 void
HandleParallelMessages(void)1014 HandleParallelMessages(void)
1015 {
1016 dlist_iter iter;
1017 MemoryContext oldcontext;
1018
1019 static MemoryContext hpm_context = NULL;
1020
1021 /*
1022 * This is invoked from ProcessInterrupts(), and since some of the
1023 * functions it calls contain CHECK_FOR_INTERRUPTS(), there is a potential
1024 * for recursive calls if more signals are received while this runs. It's
1025 * unclear that recursive entry would be safe, and it doesn't seem useful
1026 * even if it is safe, so let's block interrupts until done.
1027 */
1028 HOLD_INTERRUPTS();
1029
1030 /*
1031 * Moreover, CurrentMemoryContext might be pointing almost anywhere. We
1032 * don't want to risk leaking data into long-lived contexts, so let's do
1033 * our work here in a private context that we can reset on each use.
1034 */
1035 if (hpm_context == NULL) /* first time through? */
1036 hpm_context = AllocSetContextCreate(TopMemoryContext,
1037 "HandleParallelMessages",
1038 ALLOCSET_DEFAULT_SIZES);
1039 else
1040 MemoryContextReset(hpm_context);
1041
1042 oldcontext = MemoryContextSwitchTo(hpm_context);
1043
1044 /* OK to process messages. Reset the flag saying there are more to do. */
1045 ParallelMessagePending = false;
1046
1047 dlist_foreach(iter, &pcxt_list)
1048 {
1049 ParallelContext *pcxt;
1050 int i;
1051
1052 pcxt = dlist_container(ParallelContext, node, iter.cur);
1053 if (pcxt->worker == NULL)
1054 continue;
1055
1056 for (i = 0; i < pcxt->nworkers_launched; ++i)
1057 {
1058 /*
1059 * Read as many messages as we can from each worker, but stop when
1060 * either (1) the worker's error queue goes away, which can happen
1061 * if we receive a Terminate message from the worker; or (2) no
1062 * more messages can be read from the worker without blocking.
1063 */
1064 while (pcxt->worker[i].error_mqh != NULL)
1065 {
1066 shm_mq_result res;
1067 Size nbytes;
1068 void *data;
1069
1070 res = shm_mq_receive(pcxt->worker[i].error_mqh, &nbytes,
1071 &data, true);
1072 if (res == SHM_MQ_WOULD_BLOCK)
1073 break;
1074 else if (res == SHM_MQ_SUCCESS)
1075 {
1076 StringInfoData msg;
1077
1078 initStringInfo(&msg);
1079 appendBinaryStringInfo(&msg, data, nbytes);
1080 HandleParallelMessage(pcxt, i, &msg);
1081 pfree(msg.data);
1082 }
1083 else
1084 ereport(ERROR,
1085 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1086 errmsg("lost connection to parallel worker")));
1087 }
1088 }
1089 }
1090
1091 MemoryContextSwitchTo(oldcontext);
1092
1093 /* Might as well clear the context on our way out */
1094 MemoryContextReset(hpm_context);
1095
1096 RESUME_INTERRUPTS();
1097 }
1098
1099 /*
1100 * Handle a single protocol message received from a single parallel worker.
1101 */
1102 static void
HandleParallelMessage(ParallelContext * pcxt,int i,StringInfo msg)1103 HandleParallelMessage(ParallelContext *pcxt, int i, StringInfo msg)
1104 {
1105 char msgtype;
1106
1107 if (pcxt->known_attached_workers != NULL &&
1108 !pcxt->known_attached_workers[i])
1109 {
1110 pcxt->known_attached_workers[i] = true;
1111 pcxt->nknown_attached_workers++;
1112 }
1113
1114 msgtype = pq_getmsgbyte(msg);
1115
1116 switch (msgtype)
1117 {
1118 case 'K': /* BackendKeyData */
1119 {
1120 int32 pid = pq_getmsgint(msg, 4);
1121
1122 (void) pq_getmsgint(msg, 4); /* discard cancel key */
1123 (void) pq_getmsgend(msg);
1124 pcxt->worker[i].pid = pid;
1125 break;
1126 }
1127
1128 case 'E': /* ErrorResponse */
1129 case 'N': /* NoticeResponse */
1130 {
1131 ErrorData edata;
1132 ErrorContextCallback *save_error_context_stack;
1133
1134 /* Parse ErrorResponse or NoticeResponse. */
1135 pq_parse_errornotice(msg, &edata);
1136
1137 /* Death of a worker isn't enough justification for suicide. */
1138 edata.elevel = Min(edata.elevel, ERROR);
1139
1140 /*
1141 * If desired, add a context line to show that this is a
1142 * message propagated from a parallel worker. Otherwise, it
1143 * can sometimes be confusing to understand what actually
1144 * happened. (We don't do this in FORCE_PARALLEL_REGRESS mode
1145 * because it causes test-result instability depending on
1146 * whether a parallel worker is actually used or not.)
1147 */
1148 if (force_parallel_mode != FORCE_PARALLEL_REGRESS)
1149 {
1150 if (edata.context)
1151 edata.context = psprintf("%s\n%s", edata.context,
1152 _("parallel worker"));
1153 else
1154 edata.context = pstrdup(_("parallel worker"));
1155 }
1156
1157 /*
1158 * Context beyond that should use the error context callbacks
1159 * that were in effect when the ParallelContext was created,
1160 * not the current ones.
1161 */
1162 save_error_context_stack = error_context_stack;
1163 error_context_stack = pcxt->error_context_stack;
1164
1165 /* Rethrow error or print notice. */
1166 ThrowErrorData(&edata);
1167
1168 /* Not an error, so restore previous context stack. */
1169 error_context_stack = save_error_context_stack;
1170
1171 break;
1172 }
1173
1174 case 'A': /* NotifyResponse */
1175 {
1176 /* Propagate NotifyResponse. */
1177 int32 pid;
1178 const char *channel;
1179 const char *payload;
1180
1181 pid = pq_getmsgint(msg, 4);
1182 channel = pq_getmsgrawstring(msg);
1183 payload = pq_getmsgrawstring(msg);
1184 pq_endmessage(msg);
1185
1186 NotifyMyFrontEnd(channel, payload, pid);
1187
1188 break;
1189 }
1190
1191 case 'X': /* Terminate, indicating clean exit */
1192 {
1193 shm_mq_detach(pcxt->worker[i].error_mqh);
1194 pcxt->worker[i].error_mqh = NULL;
1195 break;
1196 }
1197
1198 default:
1199 {
1200 elog(ERROR, "unrecognized message type received from parallel worker: %c (message length %d bytes)",
1201 msgtype, msg->len);
1202 }
1203 }
1204 }
1205
1206 /*
1207 * End-of-subtransaction cleanup for parallel contexts.
1208 *
1209 * Currently, it's forbidden to enter or leave a subtransaction while
1210 * parallel mode is in effect, so we could just blow away everything. But
1211 * we may want to relax that restriction in the future, so this code
1212 * contemplates that there may be multiple subtransaction IDs in pcxt_list.
1213 */
1214 void
AtEOSubXact_Parallel(bool isCommit,SubTransactionId mySubId)1215 AtEOSubXact_Parallel(bool isCommit, SubTransactionId mySubId)
1216 {
1217 while (!dlist_is_empty(&pcxt_list))
1218 {
1219 ParallelContext *pcxt;
1220
1221 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
1222 if (pcxt->subid != mySubId)
1223 break;
1224 if (isCommit)
1225 elog(WARNING, "leaked parallel context");
1226 DestroyParallelContext(pcxt);
1227 }
1228 }
1229
1230 /*
1231 * End-of-transaction cleanup for parallel contexts.
1232 */
1233 void
AtEOXact_Parallel(bool isCommit)1234 AtEOXact_Parallel(bool isCommit)
1235 {
1236 while (!dlist_is_empty(&pcxt_list))
1237 {
1238 ParallelContext *pcxt;
1239
1240 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
1241 if (isCommit)
1242 elog(WARNING, "leaked parallel context");
1243 DestroyParallelContext(pcxt);
1244 }
1245 }
1246
1247 /*
1248 * Main entrypoint for parallel workers.
1249 */
1250 void
ParallelWorkerMain(Datum main_arg)1251 ParallelWorkerMain(Datum main_arg)
1252 {
1253 dsm_segment *seg;
1254 shm_toc *toc;
1255 FixedParallelState *fps;
1256 char *error_queue_space;
1257 shm_mq *mq;
1258 shm_mq_handle *mqh;
1259 char *libraryspace;
1260 char *entrypointstate;
1261 char *library_name;
1262 char *function_name;
1263 parallel_worker_main_type entrypt;
1264 char *gucspace;
1265 char *combocidspace;
1266 char *tsnapspace;
1267 char *asnapspace;
1268 char *tstatespace;
1269 char *pendingsyncsspace;
1270 char *reindexspace;
1271 char *relmapperspace;
1272 char *uncommittedenumsspace;
1273 StringInfoData msgbuf;
1274 char *session_dsm_handle_space;
1275 Snapshot tsnapshot;
1276 Snapshot asnapshot;
1277
1278 /* Set flag to indicate that we're initializing a parallel worker. */
1279 InitializingParallelWorker = true;
1280
1281 /* Establish signal handlers. */
1282 pqsignal(SIGTERM, die);
1283 BackgroundWorkerUnblockSignals();
1284
1285 /* Determine and set our parallel worker number. */
1286 Assert(ParallelWorkerNumber == -1);
1287 memcpy(&ParallelWorkerNumber, MyBgworkerEntry->bgw_extra, sizeof(int));
1288
1289 /* Set up a memory context to work in, just for cleanliness. */
1290 CurrentMemoryContext = AllocSetContextCreate(TopMemoryContext,
1291 "Parallel worker",
1292 ALLOCSET_DEFAULT_SIZES);
1293
1294 /*
1295 * Attach to the dynamic shared memory segment for the parallel query, and
1296 * find its table of contents.
1297 *
1298 * Note: at this point, we have not created any ResourceOwner in this
1299 * process. This will result in our DSM mapping surviving until process
1300 * exit, which is fine. If there were a ResourceOwner, it would acquire
1301 * ownership of the mapping, but we have no need for that.
1302 */
1303 seg = dsm_attach(DatumGetUInt32(main_arg));
1304 if (seg == NULL)
1305 ereport(ERROR,
1306 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1307 errmsg("could not map dynamic shared memory segment")));
1308 toc = shm_toc_attach(PARALLEL_MAGIC, dsm_segment_address(seg));
1309 if (toc == NULL)
1310 ereport(ERROR,
1311 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1312 errmsg("invalid magic number in dynamic shared memory segment")));
1313
1314 /* Look up fixed parallel state. */
1315 fps = shm_toc_lookup(toc, PARALLEL_KEY_FIXED, false);
1316 MyFixedParallelState = fps;
1317
1318 /* Arrange to signal the leader if we exit. */
1319 ParallelLeaderPid = fps->parallel_leader_pid;
1320 ParallelLeaderBackendId = fps->parallel_leader_backend_id;
1321 on_shmem_exit(ParallelWorkerShutdown, (Datum) 0);
1322
1323 /*
1324 * Now we can find and attach to the error queue provided for us. That's
1325 * good, because until we do that, any errors that happen here will not be
1326 * reported back to the process that requested that this worker be
1327 * launched.
1328 */
1329 error_queue_space = shm_toc_lookup(toc, PARALLEL_KEY_ERROR_QUEUE, false);
1330 mq = (shm_mq *) (error_queue_space +
1331 ParallelWorkerNumber * PARALLEL_ERROR_QUEUE_SIZE);
1332 shm_mq_set_sender(mq, MyProc);
1333 mqh = shm_mq_attach(mq, seg, NULL);
1334 pq_redirect_to_shm_mq(seg, mqh);
1335 pq_set_parallel_leader(fps->parallel_leader_pid,
1336 fps->parallel_leader_backend_id);
1337
1338 /*
1339 * Send a BackendKeyData message to the process that initiated parallelism
1340 * so that it has access to our PID before it receives any other messages
1341 * from us. Our cancel key is sent, too, since that's the way the
1342 * protocol message is defined, but it won't actually be used for anything
1343 * in this case.
1344 */
1345 pq_beginmessage(&msgbuf, 'K');
1346 pq_sendint32(&msgbuf, (int32) MyProcPid);
1347 pq_sendint32(&msgbuf, (int32) MyCancelKey);
1348 pq_endmessage(&msgbuf);
1349
1350 /*
1351 * Hooray! Primary initialization is complete. Now, we need to set up our
1352 * backend-local state to match the original backend.
1353 */
1354
1355 /*
1356 * Join locking group. We must do this before anything that could try to
1357 * acquire a heavyweight lock, because any heavyweight locks acquired to
1358 * this point could block either directly against the parallel group
1359 * leader or against some process which in turn waits for a lock that
1360 * conflicts with the parallel group leader, causing an undetected
1361 * deadlock. (If we can't join the lock group, the leader has gone away,
1362 * so just exit quietly.)
1363 */
1364 if (!BecomeLockGroupMember(fps->parallel_leader_pgproc,
1365 fps->parallel_leader_pid))
1366 return;
1367
1368 /*
1369 * Restore transaction and statement start-time timestamps. This must
1370 * happen before anything that would start a transaction, else asserts in
1371 * xact.c will fire.
1372 */
1373 SetParallelStartTimestamps(fps->xact_ts, fps->stmt_ts);
1374
1375 /*
1376 * Identify the entry point to be called. In theory this could result in
1377 * loading an additional library, though most likely the entry point is in
1378 * the core backend or in a library we just loaded.
1379 */
1380 entrypointstate = shm_toc_lookup(toc, PARALLEL_KEY_ENTRYPOINT, false);
1381 library_name = entrypointstate;
1382 function_name = entrypointstate + strlen(library_name) + 1;
1383
1384 entrypt = LookupParallelWorkerFunction(library_name, function_name);
1385
1386 /* Restore database connection. */
1387 BackgroundWorkerInitializeConnectionByOid(fps->database_id,
1388 fps->authenticated_user_id,
1389 0);
1390
1391 /*
1392 * Set the client encoding to the database encoding, since that is what
1393 * the leader will expect.
1394 */
1395 SetClientEncoding(GetDatabaseEncoding());
1396
1397 /*
1398 * Load libraries that were loaded by original backend. We want to do
1399 * this before restoring GUCs, because the libraries might define custom
1400 * variables.
1401 */
1402 libraryspace = shm_toc_lookup(toc, PARALLEL_KEY_LIBRARY, false);
1403 StartTransactionCommand();
1404 RestoreLibraryState(libraryspace);
1405
1406 /* Restore GUC values from launching backend. */
1407 gucspace = shm_toc_lookup(toc, PARALLEL_KEY_GUC, false);
1408 RestoreGUCState(gucspace);
1409 CommitTransactionCommand();
1410
1411 /* Crank up a transaction state appropriate to a parallel worker. */
1412 tstatespace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_STATE, false);
1413 StartParallelWorkerTransaction(tstatespace);
1414
1415 /* Restore combo CID state. */
1416 combocidspace = shm_toc_lookup(toc, PARALLEL_KEY_COMBO_CID, false);
1417 RestoreComboCIDState(combocidspace);
1418
1419 /* Attach to the per-session DSM segment and contained objects. */
1420 session_dsm_handle_space =
1421 shm_toc_lookup(toc, PARALLEL_KEY_SESSION_DSM, false);
1422 AttachSession(*(dsm_handle *) session_dsm_handle_space);
1423
1424 /*
1425 * If the transaction isolation level is REPEATABLE READ or SERIALIZABLE,
1426 * the leader has serialized the transaction snapshot and we must restore
1427 * it. At lower isolation levels, there is no transaction-lifetime
1428 * snapshot, but we need TransactionXmin to get set to a value which is
1429 * less than or equal to the xmin of every snapshot that will be used by
1430 * this worker. The easiest way to accomplish that is to install the
1431 * active snapshot as the transaction snapshot. Code running in this
1432 * parallel worker might take new snapshots via GetTransactionSnapshot()
1433 * or GetLatestSnapshot(), but it shouldn't have any way of acquiring a
1434 * snapshot older than the active snapshot.
1435 */
1436 asnapspace = shm_toc_lookup(toc, PARALLEL_KEY_ACTIVE_SNAPSHOT, false);
1437 tsnapspace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT, true);
1438 asnapshot = RestoreSnapshot(asnapspace);
1439 tsnapshot = tsnapspace ? RestoreSnapshot(tsnapspace) : asnapshot;
1440 RestoreTransactionSnapshot(tsnapshot,
1441 fps->parallel_leader_pgproc);
1442 PushActiveSnapshot(asnapshot);
1443
1444 /*
1445 * We've changed which tuples we can see, and must therefore invalidate
1446 * system caches.
1447 */
1448 InvalidateSystemCaches();
1449
1450 /*
1451 * Restore current role id. Skip verifying whether session user is
1452 * allowed to become this role and blindly restore the leader's state for
1453 * current role.
1454 */
1455 SetCurrentRoleId(fps->outer_user_id, fps->is_superuser);
1456
1457 /* Restore user ID and security context. */
1458 SetUserIdAndSecContext(fps->current_user_id, fps->sec_context);
1459
1460 /* Restore temp-namespace state to ensure search path matches leader's. */
1461 SetTempNamespaceState(fps->temp_namespace_id,
1462 fps->temp_toast_namespace_id);
1463
1464 /* Restore pending syncs. */
1465 pendingsyncsspace = shm_toc_lookup(toc, PARALLEL_KEY_PENDING_SYNCS,
1466 false);
1467 RestorePendingSyncs(pendingsyncsspace);
1468
1469 /* Restore reindex state. */
1470 reindexspace = shm_toc_lookup(toc, PARALLEL_KEY_REINDEX_STATE, false);
1471 RestoreReindexState(reindexspace);
1472
1473 /* Restore relmapper state. */
1474 relmapperspace = shm_toc_lookup(toc, PARALLEL_KEY_RELMAPPER_STATE, false);
1475 RestoreRelationMap(relmapperspace);
1476
1477 /* Restore uncommitted enums. */
1478 uncommittedenumsspace = shm_toc_lookup(toc, PARALLEL_KEY_UNCOMMITTEDENUMS,
1479 false);
1480 RestoreUncommittedEnums(uncommittedenumsspace);
1481
1482 /* Attach to the leader's serializable transaction, if SERIALIZABLE. */
1483 AttachSerializableXact(fps->serializable_xact_handle);
1484
1485 /*
1486 * We've initialized all of our state now; nothing should change
1487 * hereafter.
1488 */
1489 InitializingParallelWorker = false;
1490 EnterParallelMode();
1491
1492 /*
1493 * Time to do the real work: invoke the caller-supplied code.
1494 */
1495 entrypt(seg, toc);
1496
1497 /* Must exit parallel mode to pop active snapshot. */
1498 ExitParallelMode();
1499
1500 /* Must pop active snapshot so snapmgr.c doesn't complain. */
1501 PopActiveSnapshot();
1502
1503 /* Shut down the parallel-worker transaction. */
1504 EndParallelWorkerTransaction();
1505
1506 /* Detach from the per-session DSM segment. */
1507 DetachSession();
1508
1509 /* Report success. */
1510 pq_putmessage('X', NULL, 0);
1511 }
1512
1513 /*
1514 * Update shared memory with the ending location of the last WAL record we
1515 * wrote, if it's greater than the value already stored there.
1516 */
1517 void
ParallelWorkerReportLastRecEnd(XLogRecPtr last_xlog_end)1518 ParallelWorkerReportLastRecEnd(XLogRecPtr last_xlog_end)
1519 {
1520 FixedParallelState *fps = MyFixedParallelState;
1521
1522 Assert(fps != NULL);
1523 SpinLockAcquire(&fps->mutex);
1524 if (fps->last_xlog_end < last_xlog_end)
1525 fps->last_xlog_end = last_xlog_end;
1526 SpinLockRelease(&fps->mutex);
1527 }
1528
1529 /*
1530 * Make sure the leader tries to read from our error queue one more time.
1531 * This guards against the case where we exit uncleanly without sending an
1532 * ErrorResponse to the leader, for example because some code calls proc_exit
1533 * directly.
1534 */
1535 static void
ParallelWorkerShutdown(int code,Datum arg)1536 ParallelWorkerShutdown(int code, Datum arg)
1537 {
1538 SendProcSignal(ParallelLeaderPid,
1539 PROCSIG_PARALLEL_MESSAGE,
1540 ParallelLeaderBackendId);
1541 }
1542
1543 /*
1544 * Look up (and possibly load) a parallel worker entry point function.
1545 *
1546 * For functions contained in the core code, we use library name "postgres"
1547 * and consult the InternalParallelWorkers array. External functions are
1548 * looked up, and loaded if necessary, using load_external_function().
1549 *
1550 * The point of this is to pass function names as strings across process
1551 * boundaries. We can't pass actual function addresses because of the
1552 * possibility that the function has been loaded at a different address
1553 * in a different process. This is obviously a hazard for functions in
1554 * loadable libraries, but it can happen even for functions in the core code
1555 * on platforms using EXEC_BACKEND (e.g., Windows).
1556 *
1557 * At some point it might be worthwhile to get rid of InternalParallelWorkers[]
1558 * in favor of applying load_external_function() for core functions too;
1559 * but that raises portability issues that are not worth addressing now.
1560 */
1561 static parallel_worker_main_type
LookupParallelWorkerFunction(const char * libraryname,const char * funcname)1562 LookupParallelWorkerFunction(const char *libraryname, const char *funcname)
1563 {
1564 /*
1565 * If the function is to be loaded from postgres itself, search the
1566 * InternalParallelWorkers array.
1567 */
1568 if (strcmp(libraryname, "postgres") == 0)
1569 {
1570 int i;
1571
1572 for (i = 0; i < lengthof(InternalParallelWorkers); i++)
1573 {
1574 if (strcmp(InternalParallelWorkers[i].fn_name, funcname) == 0)
1575 return InternalParallelWorkers[i].fn_addr;
1576 }
1577
1578 /* We can only reach this by programming error. */
1579 elog(ERROR, "internal function \"%s\" not found", funcname);
1580 }
1581
1582 /* Otherwise load from external library. */
1583 return (parallel_worker_main_type)
1584 load_external_function(libraryname, funcname, true, NULL);
1585 }
1586