1 /*-------------------------------------------------------------------------
2 *
3 * parallel.c
4 * Infrastructure for launching parallel workers
5 *
6 * Portions Copyright (c) 1996-2018, PostgreSQL Global Development Group
7 * Portions Copyright (c) 1994, Regents of the University of California
8 *
9 * IDENTIFICATION
10 * src/backend/access/transam/parallel.c
11 *
12 *-------------------------------------------------------------------------
13 */
14
15 #include "postgres.h"
16
17 #include "access/nbtree.h"
18 #include "access/parallel.h"
19 #include "access/session.h"
20 #include "access/xact.h"
21 #include "access/xlog.h"
22 #include "catalog/index.h"
23 #include "catalog/namespace.h"
24 #include "commands/async.h"
25 #include "executor/execParallel.h"
26 #include "libpq/libpq.h"
27 #include "libpq/pqformat.h"
28 #include "libpq/pqmq.h"
29 #include "miscadmin.h"
30 #include "optimizer/planmain.h"
31 #include "pgstat.h"
32 #include "storage/ipc.h"
33 #include "storage/sinval.h"
34 #include "storage/spin.h"
35 #include "tcop/tcopprot.h"
36 #include "utils/combocid.h"
37 #include "utils/guc.h"
38 #include "utils/inval.h"
39 #include "utils/memutils.h"
40 #include "utils/resowner.h"
41 #include "utils/snapmgr.h"
42 #include "utils/typcache.h"
43
44
45 /*
46 * We don't want to waste a lot of memory on an error queue which, most of
47 * the time, will process only a handful of small messages. However, it is
48 * desirable to make it large enough that a typical ErrorResponse can be sent
49 * without blocking. That way, a worker that errors out can write the whole
50 * message into the queue and terminate without waiting for the user backend.
51 */
52 #define PARALLEL_ERROR_QUEUE_SIZE 16384
53
54 /* Magic number for parallel context TOC. */
55 #define PARALLEL_MAGIC 0x50477c7c
56
57 /*
58 * Magic numbers for per-context parallel state sharing. Higher-level code
59 * should use smaller values, leaving these very large ones for use by this
60 * module.
61 */
62 #define PARALLEL_KEY_FIXED UINT64CONST(0xFFFFFFFFFFFF0001)
63 #define PARALLEL_KEY_ERROR_QUEUE UINT64CONST(0xFFFFFFFFFFFF0002)
64 #define PARALLEL_KEY_LIBRARY UINT64CONST(0xFFFFFFFFFFFF0003)
65 #define PARALLEL_KEY_GUC UINT64CONST(0xFFFFFFFFFFFF0004)
66 #define PARALLEL_KEY_COMBO_CID UINT64CONST(0xFFFFFFFFFFFF0005)
67 #define PARALLEL_KEY_TRANSACTION_SNAPSHOT UINT64CONST(0xFFFFFFFFFFFF0006)
68 #define PARALLEL_KEY_ACTIVE_SNAPSHOT UINT64CONST(0xFFFFFFFFFFFF0007)
69 #define PARALLEL_KEY_TRANSACTION_STATE UINT64CONST(0xFFFFFFFFFFFF0008)
70 #define PARALLEL_KEY_ENTRYPOINT UINT64CONST(0xFFFFFFFFFFFF0009)
71 #define PARALLEL_KEY_SESSION_DSM UINT64CONST(0xFFFFFFFFFFFF000A)
72 #define PARALLEL_KEY_REINDEX_STATE UINT64CONST(0xFFFFFFFFFFFF000B)
73
74 /* Fixed-size parallel state. */
75 typedef struct FixedParallelState
76 {
77 /* Fixed-size state that workers must restore. */
78 Oid database_id;
79 Oid authenticated_user_id;
80 Oid current_user_id;
81 Oid outer_user_id;
82 Oid temp_namespace_id;
83 Oid temp_toast_namespace_id;
84 int sec_context;
85 bool is_superuser;
86 PGPROC *parallel_master_pgproc;
87 pid_t parallel_master_pid;
88 BackendId parallel_master_backend_id;
89 TimestampTz xact_ts;
90 TimestampTz stmt_ts;
91
92 /* Mutex protects remaining fields. */
93 slock_t mutex;
94
95 /* Maximum XactLastRecEnd of any worker. */
96 XLogRecPtr last_xlog_end;
97 } FixedParallelState;
98
99 /*
100 * Our parallel worker number. We initialize this to -1, meaning that we are
101 * not a parallel worker. In parallel workers, it will be set to a value >= 0
102 * and < the number of workers before any user code is invoked; each parallel
103 * worker will get a different parallel worker number.
104 */
105 int ParallelWorkerNumber = -1;
106
107 /* Is there a parallel message pending which we need to receive? */
108 volatile bool ParallelMessagePending = false;
109
110 /* Are we initializing a parallel worker? */
111 bool InitializingParallelWorker = false;
112
113 /* Pointer to our fixed parallel state. */
114 static FixedParallelState *MyFixedParallelState;
115
116 /* List of active parallel contexts. */
117 static dlist_head pcxt_list = DLIST_STATIC_INIT(pcxt_list);
118
119 /* Backend-local copy of data from FixedParallelState. */
120 static pid_t ParallelMasterPid;
121
122 /*
123 * List of internal parallel worker entry points. We need this for
124 * reasons explained in LookupParallelWorkerFunction(), below.
125 */
126 static const struct
127 {
128 const char *fn_name;
129 parallel_worker_main_type fn_addr;
130 } InternalParallelWorkers[] =
131
132 {
133 {
134 "ParallelQueryMain", ParallelQueryMain
135 },
136 {
137 "_bt_parallel_build_main", _bt_parallel_build_main
138 }
139 };
140
141 /* Private functions. */
142 static void HandleParallelMessage(ParallelContext *pcxt, int i, StringInfo msg);
143 static void WaitForParallelWorkersToExit(ParallelContext *pcxt);
144 static parallel_worker_main_type LookupParallelWorkerFunction(const char *libraryname, const char *funcname);
145 static void ParallelWorkerShutdown(int code, Datum arg);
146
147
148 /*
149 * Establish a new parallel context. This should be done after entering
150 * parallel mode, and (unless there is an error) the context should be
151 * destroyed before exiting the current subtransaction.
152 */
153 ParallelContext *
CreateParallelContext(const char * library_name,const char * function_name,int nworkers,bool serializable_okay)154 CreateParallelContext(const char *library_name, const char *function_name,
155 int nworkers, bool serializable_okay)
156 {
157 MemoryContext oldcontext;
158 ParallelContext *pcxt;
159
160 /* It is unsafe to create a parallel context if not in parallel mode. */
161 Assert(IsInParallelMode());
162
163 /* Number of workers should be non-negative. */
164 Assert(nworkers >= 0);
165
166 /*
167 * If dynamic shared memory is not available, we won't be able to use
168 * background workers.
169 */
170 if (dynamic_shared_memory_type == DSM_IMPL_NONE)
171 nworkers = 0;
172
173 /*
174 * If we are running under serializable isolation, we can't use parallel
175 * workers, at least not until somebody enhances that mechanism to be
176 * parallel-aware. Utility statement callers may ask us to ignore this
177 * restriction because they're always able to safely ignore the fact that
178 * SIREAD locks do not work with parallelism.
179 */
180 if (IsolationIsSerializable() && !serializable_okay)
181 nworkers = 0;
182
183 /* We might be running in a short-lived memory context. */
184 oldcontext = MemoryContextSwitchTo(TopTransactionContext);
185
186 /* Initialize a new ParallelContext. */
187 pcxt = palloc0(sizeof(ParallelContext));
188 pcxt->subid = GetCurrentSubTransactionId();
189 pcxt->nworkers = nworkers;
190 pcxt->library_name = pstrdup(library_name);
191 pcxt->function_name = pstrdup(function_name);
192 pcxt->error_context_stack = error_context_stack;
193 shm_toc_initialize_estimator(&pcxt->estimator);
194 dlist_push_head(&pcxt_list, &pcxt->node);
195
196 /* Restore previous memory context. */
197 MemoryContextSwitchTo(oldcontext);
198
199 return pcxt;
200 }
201
202 /*
203 * Establish the dynamic shared memory segment for a parallel context and
204 * copy state and other bookkeeping information that will be needed by
205 * parallel workers into it.
206 */
207 void
InitializeParallelDSM(ParallelContext * pcxt)208 InitializeParallelDSM(ParallelContext *pcxt)
209 {
210 MemoryContext oldcontext;
211 Size library_len = 0;
212 Size guc_len = 0;
213 Size combocidlen = 0;
214 Size tsnaplen = 0;
215 Size asnaplen = 0;
216 Size tstatelen = 0;
217 Size reindexlen = 0;
218 Size segsize = 0;
219 int i;
220 FixedParallelState *fps;
221 dsm_handle session_dsm_handle = DSM_HANDLE_INVALID;
222 Snapshot transaction_snapshot = GetTransactionSnapshot();
223 Snapshot active_snapshot = GetActiveSnapshot();
224
225 /* We might be running in a very short-lived memory context. */
226 oldcontext = MemoryContextSwitchTo(TopTransactionContext);
227
228 /* Allow space to store the fixed-size parallel state. */
229 shm_toc_estimate_chunk(&pcxt->estimator, sizeof(FixedParallelState));
230 shm_toc_estimate_keys(&pcxt->estimator, 1);
231
232 /*
233 * Normally, the user will have requested at least one worker process, but
234 * if by chance they have not, we can skip a bunch of things here.
235 */
236 if (pcxt->nworkers > 0)
237 {
238 /* Get (or create) the per-session DSM segment's handle. */
239 session_dsm_handle = GetSessionDsmHandle();
240
241 /*
242 * If we weren't able to create a per-session DSM segment, then we can
243 * continue but we can't safely launch any workers because their
244 * record typmods would be incompatible so they couldn't exchange
245 * tuples.
246 */
247 if (session_dsm_handle == DSM_HANDLE_INVALID)
248 pcxt->nworkers = 0;
249 }
250
251 if (pcxt->nworkers > 0)
252 {
253 /* Estimate space for various kinds of state sharing. */
254 library_len = EstimateLibraryStateSpace();
255 shm_toc_estimate_chunk(&pcxt->estimator, library_len);
256 guc_len = EstimateGUCStateSpace();
257 shm_toc_estimate_chunk(&pcxt->estimator, guc_len);
258 combocidlen = EstimateComboCIDStateSpace();
259 shm_toc_estimate_chunk(&pcxt->estimator, combocidlen);
260 if (IsolationUsesXactSnapshot())
261 {
262 tsnaplen = EstimateSnapshotSpace(transaction_snapshot);
263 shm_toc_estimate_chunk(&pcxt->estimator, tsnaplen);
264 }
265 asnaplen = EstimateSnapshotSpace(active_snapshot);
266 shm_toc_estimate_chunk(&pcxt->estimator, asnaplen);
267 tstatelen = EstimateTransactionStateSpace();
268 shm_toc_estimate_chunk(&pcxt->estimator, tstatelen);
269 shm_toc_estimate_chunk(&pcxt->estimator, sizeof(dsm_handle));
270 reindexlen = EstimateReindexStateSpace();
271 shm_toc_estimate_chunk(&pcxt->estimator, reindexlen);
272 /* If you add more chunks here, you probably need to add keys. */
273 shm_toc_estimate_keys(&pcxt->estimator, 8);
274
275 /* Estimate space need for error queues. */
276 StaticAssertStmt(BUFFERALIGN(PARALLEL_ERROR_QUEUE_SIZE) ==
277 PARALLEL_ERROR_QUEUE_SIZE,
278 "parallel error queue size not buffer-aligned");
279 shm_toc_estimate_chunk(&pcxt->estimator,
280 mul_size(PARALLEL_ERROR_QUEUE_SIZE,
281 pcxt->nworkers));
282 shm_toc_estimate_keys(&pcxt->estimator, 1);
283
284 /* Estimate how much we'll need for the entrypoint info. */
285 shm_toc_estimate_chunk(&pcxt->estimator, strlen(pcxt->library_name) +
286 strlen(pcxt->function_name) + 2);
287 shm_toc_estimate_keys(&pcxt->estimator, 1);
288 }
289
290 /*
291 * Create DSM and initialize with new table of contents. But if the user
292 * didn't request any workers, then don't bother creating a dynamic shared
293 * memory segment; instead, just use backend-private memory.
294 *
295 * Also, if we can't create a dynamic shared memory segment because the
296 * maximum number of segments have already been created, then fall back to
297 * backend-private memory, and plan not to use any workers. We hope this
298 * won't happen very often, but it's better to abandon the use of
299 * parallelism than to fail outright.
300 */
301 segsize = shm_toc_estimate(&pcxt->estimator);
302 if (pcxt->nworkers > 0)
303 pcxt->seg = dsm_create(segsize, DSM_CREATE_NULL_IF_MAXSEGMENTS);
304 if (pcxt->seg != NULL)
305 pcxt->toc = shm_toc_create(PARALLEL_MAGIC,
306 dsm_segment_address(pcxt->seg),
307 segsize);
308 else
309 {
310 pcxt->nworkers = 0;
311 pcxt->private_memory = MemoryContextAlloc(TopMemoryContext, segsize);
312 pcxt->toc = shm_toc_create(PARALLEL_MAGIC, pcxt->private_memory,
313 segsize);
314 }
315
316 /* Initialize fixed-size state in shared memory. */
317 fps = (FixedParallelState *)
318 shm_toc_allocate(pcxt->toc, sizeof(FixedParallelState));
319 fps->database_id = MyDatabaseId;
320 fps->authenticated_user_id = GetAuthenticatedUserId();
321 fps->outer_user_id = GetCurrentRoleId();
322 fps->is_superuser = session_auth_is_superuser;
323 GetUserIdAndSecContext(&fps->current_user_id, &fps->sec_context);
324 GetTempNamespaceState(&fps->temp_namespace_id,
325 &fps->temp_toast_namespace_id);
326 fps->parallel_master_pgproc = MyProc;
327 fps->parallel_master_pid = MyProcPid;
328 fps->parallel_master_backend_id = MyBackendId;
329 fps->xact_ts = GetCurrentTransactionStartTimestamp();
330 fps->stmt_ts = GetCurrentStatementStartTimestamp();
331 SpinLockInit(&fps->mutex);
332 fps->last_xlog_end = 0;
333 shm_toc_insert(pcxt->toc, PARALLEL_KEY_FIXED, fps);
334
335 /* We can skip the rest of this if we're not budgeting for any workers. */
336 if (pcxt->nworkers > 0)
337 {
338 char *libraryspace;
339 char *gucspace;
340 char *combocidspace;
341 char *tsnapspace;
342 char *asnapspace;
343 char *tstatespace;
344 char *reindexspace;
345 char *error_queue_space;
346 char *session_dsm_handle_space;
347 char *entrypointstate;
348 Size lnamelen;
349
350 /* Serialize shared libraries we have loaded. */
351 libraryspace = shm_toc_allocate(pcxt->toc, library_len);
352 SerializeLibraryState(library_len, libraryspace);
353 shm_toc_insert(pcxt->toc, PARALLEL_KEY_LIBRARY, libraryspace);
354
355 /* Serialize GUC settings. */
356 gucspace = shm_toc_allocate(pcxt->toc, guc_len);
357 SerializeGUCState(guc_len, gucspace);
358 shm_toc_insert(pcxt->toc, PARALLEL_KEY_GUC, gucspace);
359
360 /* Serialize combo CID state. */
361 combocidspace = shm_toc_allocate(pcxt->toc, combocidlen);
362 SerializeComboCIDState(combocidlen, combocidspace);
363 shm_toc_insert(pcxt->toc, PARALLEL_KEY_COMBO_CID, combocidspace);
364
365 /*
366 * Serialize the transaction snapshot if the transaction
367 * isolation-level uses a transaction snapshot.
368 */
369 if (IsolationUsesXactSnapshot())
370 {
371 tsnapspace = shm_toc_allocate(pcxt->toc, tsnaplen);
372 SerializeSnapshot(transaction_snapshot, tsnapspace);
373 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT,
374 tsnapspace);
375 }
376
377 /* Serialize the active snapshot. */
378 asnapspace = shm_toc_allocate(pcxt->toc, asnaplen);
379 SerializeSnapshot(active_snapshot, asnapspace);
380 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ACTIVE_SNAPSHOT, asnapspace);
381
382 /* Provide the handle for per-session segment. */
383 session_dsm_handle_space = shm_toc_allocate(pcxt->toc,
384 sizeof(dsm_handle));
385 *(dsm_handle *) session_dsm_handle_space = session_dsm_handle;
386 shm_toc_insert(pcxt->toc, PARALLEL_KEY_SESSION_DSM,
387 session_dsm_handle_space);
388
389 /* Serialize transaction state. */
390 tstatespace = shm_toc_allocate(pcxt->toc, tstatelen);
391 SerializeTransactionState(tstatelen, tstatespace);
392 shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_STATE, tstatespace);
393
394 /* Serialize reindex state. */
395 reindexspace = shm_toc_allocate(pcxt->toc, reindexlen);
396 SerializeReindexState(reindexlen, reindexspace);
397 shm_toc_insert(pcxt->toc, PARALLEL_KEY_REINDEX_STATE, reindexspace);
398
399 /* Allocate space for worker information. */
400 pcxt->worker = palloc0(sizeof(ParallelWorkerInfo) * pcxt->nworkers);
401
402 /*
403 * Establish error queues in dynamic shared memory.
404 *
405 * These queues should be used only for transmitting ErrorResponse,
406 * NoticeResponse, and NotifyResponse protocol messages. Tuple data
407 * should be transmitted via separate (possibly larger?) queues.
408 */
409 error_queue_space =
410 shm_toc_allocate(pcxt->toc,
411 mul_size(PARALLEL_ERROR_QUEUE_SIZE,
412 pcxt->nworkers));
413 for (i = 0; i < pcxt->nworkers; ++i)
414 {
415 char *start;
416 shm_mq *mq;
417
418 start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
419 mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
420 shm_mq_set_receiver(mq, MyProc);
421 pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
422 }
423 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE, error_queue_space);
424
425 /*
426 * Serialize entrypoint information. It's unsafe to pass function
427 * pointers across processes, as the function pointer may be different
428 * in each process in EXEC_BACKEND builds, so we always pass library
429 * and function name. (We use library name "postgres" for functions
430 * in the core backend.)
431 */
432 lnamelen = strlen(pcxt->library_name);
433 entrypointstate = shm_toc_allocate(pcxt->toc, lnamelen +
434 strlen(pcxt->function_name) + 2);
435 strcpy(entrypointstate, pcxt->library_name);
436 strcpy(entrypointstate + lnamelen + 1, pcxt->function_name);
437 shm_toc_insert(pcxt->toc, PARALLEL_KEY_ENTRYPOINT, entrypointstate);
438 }
439
440 /* Restore previous memory context. */
441 MemoryContextSwitchTo(oldcontext);
442 }
443
444 /*
445 * Reinitialize the dynamic shared memory segment for a parallel context such
446 * that we could launch workers for it again.
447 */
448 void
ReinitializeParallelDSM(ParallelContext * pcxt)449 ReinitializeParallelDSM(ParallelContext *pcxt)
450 {
451 FixedParallelState *fps;
452
453 /* Wait for any old workers to exit. */
454 if (pcxt->nworkers_launched > 0)
455 {
456 WaitForParallelWorkersToFinish(pcxt);
457 WaitForParallelWorkersToExit(pcxt);
458 pcxt->nworkers_launched = 0;
459 if (pcxt->known_attached_workers)
460 {
461 pfree(pcxt->known_attached_workers);
462 pcxt->known_attached_workers = NULL;
463 pcxt->nknown_attached_workers = 0;
464 }
465 }
466
467 /* Reset a few bits of fixed parallel state to a clean state. */
468 fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED, false);
469 fps->last_xlog_end = 0;
470
471 /* Recreate error queues (if they exist). */
472 if (pcxt->nworkers > 0)
473 {
474 char *error_queue_space;
475 int i;
476
477 error_queue_space =
478 shm_toc_lookup(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE, false);
479 for (i = 0; i < pcxt->nworkers; ++i)
480 {
481 char *start;
482 shm_mq *mq;
483
484 start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
485 mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
486 shm_mq_set_receiver(mq, MyProc);
487 pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
488 }
489 }
490 }
491
492 /*
493 * Launch parallel workers.
494 */
495 void
LaunchParallelWorkers(ParallelContext * pcxt)496 LaunchParallelWorkers(ParallelContext *pcxt)
497 {
498 MemoryContext oldcontext;
499 BackgroundWorker worker;
500 int i;
501 bool any_registrations_failed = false;
502
503 /* Skip this if we have no workers. */
504 if (pcxt->nworkers == 0)
505 return;
506
507 /* We need to be a lock group leader. */
508 BecomeLockGroupLeader();
509
510 /* If we do have workers, we'd better have a DSM segment. */
511 Assert(pcxt->seg != NULL);
512
513 /* We might be running in a short-lived memory context. */
514 oldcontext = MemoryContextSwitchTo(TopTransactionContext);
515
516 /* Configure a worker. */
517 memset(&worker, 0, sizeof(worker));
518 snprintf(worker.bgw_name, BGW_MAXLEN, "parallel worker for PID %d",
519 MyProcPid);
520 snprintf(worker.bgw_type, BGW_MAXLEN, "parallel worker");
521 worker.bgw_flags =
522 BGWORKER_SHMEM_ACCESS | BGWORKER_BACKEND_DATABASE_CONNECTION
523 | BGWORKER_CLASS_PARALLEL;
524 worker.bgw_start_time = BgWorkerStart_ConsistentState;
525 worker.bgw_restart_time = BGW_NEVER_RESTART;
526 sprintf(worker.bgw_library_name, "postgres");
527 sprintf(worker.bgw_function_name, "ParallelWorkerMain");
528 worker.bgw_main_arg = UInt32GetDatum(dsm_segment_handle(pcxt->seg));
529 worker.bgw_notify_pid = MyProcPid;
530
531 /*
532 * Start workers.
533 *
534 * The caller must be able to tolerate ending up with fewer workers than
535 * expected, so there is no need to throw an error here if registration
536 * fails. It wouldn't help much anyway, because registering the worker in
537 * no way guarantees that it will start up and initialize successfully.
538 */
539 for (i = 0; i < pcxt->nworkers; ++i)
540 {
541 memcpy(worker.bgw_extra, &i, sizeof(int));
542 if (!any_registrations_failed &&
543 RegisterDynamicBackgroundWorker(&worker,
544 &pcxt->worker[i].bgwhandle))
545 {
546 shm_mq_set_handle(pcxt->worker[i].error_mqh,
547 pcxt->worker[i].bgwhandle);
548 pcxt->nworkers_launched++;
549 }
550 else
551 {
552 /*
553 * If we weren't able to register the worker, then we've bumped up
554 * against the max_worker_processes limit, and future
555 * registrations will probably fail too, so arrange to skip them.
556 * But we still have to execute this code for the remaining slots
557 * to make sure that we forget about the error queues we budgeted
558 * for those workers. Otherwise, we'll wait for them to start,
559 * but they never will.
560 */
561 any_registrations_failed = true;
562 pcxt->worker[i].bgwhandle = NULL;
563 shm_mq_detach(pcxt->worker[i].error_mqh);
564 pcxt->worker[i].error_mqh = NULL;
565 }
566 }
567
568 /*
569 * Now that nworkers_launched has taken its final value, we can initialize
570 * known_attached_workers.
571 */
572 if (pcxt->nworkers_launched > 0)
573 {
574 pcxt->known_attached_workers =
575 palloc0(sizeof(bool) * pcxt->nworkers_launched);
576 pcxt->nknown_attached_workers = 0;
577 }
578
579 /* Restore previous memory context. */
580 MemoryContextSwitchTo(oldcontext);
581 }
582
583 /*
584 * Wait for all workers to attach to their error queues, and throw an error if
585 * any worker fails to do this.
586 *
587 * Callers can assume that if this function returns successfully, then the
588 * number of workers given by pcxt->nworkers_launched have initialized and
589 * attached to their error queues. Whether or not these workers are guaranteed
590 * to still be running depends on what code the caller asked them to run;
591 * this function does not guarantee that they have not exited. However, it
592 * does guarantee that any workers which exited must have done so cleanly and
593 * after successfully performing the work with which they were tasked.
594 *
595 * If this function is not called, then some of the workers that were launched
596 * may not have been started due to a fork() failure, or may have exited during
597 * early startup prior to attaching to the error queue, so nworkers_launched
598 * cannot be viewed as completely reliable. It will never be less than the
599 * number of workers which actually started, but it might be more. Any workers
600 * that failed to start will still be discovered by
601 * WaitForParallelWorkersToFinish and an error will be thrown at that time,
602 * provided that function is eventually reached.
603 *
604 * In general, the leader process should do as much work as possible before
605 * calling this function. fork() failures and other early-startup failures
606 * are very uncommon, and having the leader sit idle when it could be doing
607 * useful work is undesirable. However, if the leader needs to wait for
608 * all of its workers or for a specific worker, it may want to call this
609 * function before doing so. If not, it must make some other provision for
610 * the failure-to-start case, lest it wait forever. On the other hand, a
611 * leader which never waits for a worker that might not be started yet, or
612 * at least never does so prior to WaitForParallelWorkersToFinish(), need not
613 * call this function at all.
614 */
615 void
WaitForParallelWorkersToAttach(ParallelContext * pcxt)616 WaitForParallelWorkersToAttach(ParallelContext *pcxt)
617 {
618 int i;
619
620 /* Skip this if we have no launched workers. */
621 if (pcxt->nworkers_launched == 0)
622 return;
623
624 for (;;)
625 {
626 /*
627 * This will process any parallel messages that are pending and it may
628 * also throw an error propagated from a worker.
629 */
630 CHECK_FOR_INTERRUPTS();
631
632 for (i = 0; i < pcxt->nworkers_launched; ++i)
633 {
634 BgwHandleStatus status;
635 shm_mq *mq;
636 int rc;
637 pid_t pid;
638
639 if (pcxt->known_attached_workers[i])
640 continue;
641
642 /*
643 * If error_mqh is NULL, then the worker has already exited
644 * cleanly.
645 */
646 if (pcxt->worker[i].error_mqh == NULL)
647 {
648 pcxt->known_attached_workers[i] = true;
649 ++pcxt->nknown_attached_workers;
650 continue;
651 }
652
653 status = GetBackgroundWorkerPid(pcxt->worker[i].bgwhandle, &pid);
654 if (status == BGWH_STARTED)
655 {
656 /* Has the worker attached to the error queue? */
657 mq = shm_mq_get_queue(pcxt->worker[i].error_mqh);
658 if (shm_mq_get_sender(mq) != NULL)
659 {
660 /* Yes, so it is known to be attached. */
661 pcxt->known_attached_workers[i] = true;
662 ++pcxt->nknown_attached_workers;
663 }
664 }
665 else if (status == BGWH_STOPPED)
666 {
667 /*
668 * If the worker stopped without attaching to the error queue,
669 * throw an error.
670 */
671 mq = shm_mq_get_queue(pcxt->worker[i].error_mqh);
672 if (shm_mq_get_sender(mq) == NULL)
673 ereport(ERROR,
674 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
675 errmsg("parallel worker failed to initialize"),
676 errhint("More details may be available in the server log.")));
677
678 pcxt->known_attached_workers[i] = true;
679 ++pcxt->nknown_attached_workers;
680 }
681 else
682 {
683 /*
684 * Worker not yet started, so we must wait. The postmaster
685 * will notify us if the worker's state changes. Our latch
686 * might also get set for some other reason, but if so we'll
687 * just end up waiting for the same worker again.
688 */
689 rc = WaitLatch(MyLatch,
690 WL_LATCH_SET | WL_POSTMASTER_DEATH,
691 -1, WAIT_EVENT_BGWORKER_STARTUP);
692
693 /* emergency bailout if postmaster has died */
694 if (rc & WL_POSTMASTER_DEATH)
695 proc_exit(1);
696
697 if (rc & WL_LATCH_SET)
698 ResetLatch(MyLatch);
699 }
700 }
701
702 /* If all workers are known to have started, we're done. */
703 if (pcxt->nknown_attached_workers >= pcxt->nworkers_launched)
704 {
705 Assert(pcxt->nknown_attached_workers == pcxt->nworkers_launched);
706 break;
707 }
708 }
709 }
710
711 /*
712 * Wait for all workers to finish computing.
713 *
714 * Even if the parallel operation seems to have completed successfully, it's
715 * important to call this function afterwards. We must not miss any errors
716 * the workers may have thrown during the parallel operation, or any that they
717 * may yet throw while shutting down.
718 *
719 * Also, we want to update our notion of XactLastRecEnd based on worker
720 * feedback.
721 */
722 void
WaitForParallelWorkersToFinish(ParallelContext * pcxt)723 WaitForParallelWorkersToFinish(ParallelContext *pcxt)
724 {
725 for (;;)
726 {
727 bool anyone_alive = false;
728 int nfinished = 0;
729 int i;
730
731 /*
732 * This will process any parallel messages that are pending, which may
733 * change the outcome of the loop that follows. It may also throw an
734 * error propagated from a worker.
735 */
736 CHECK_FOR_INTERRUPTS();
737
738 for (i = 0; i < pcxt->nworkers_launched; ++i)
739 {
740 /*
741 * If error_mqh is NULL, then the worker has already exited
742 * cleanly. If we have received a message through error_mqh from
743 * the worker, we know it started up cleanly, and therefore we're
744 * certain to be notified when it exits.
745 */
746 if (pcxt->worker[i].error_mqh == NULL)
747 ++nfinished;
748 else if (pcxt->known_attached_workers[i])
749 {
750 anyone_alive = true;
751 break;
752 }
753 }
754
755 if (!anyone_alive)
756 {
757 /* If all workers are known to have finished, we're done. */
758 if (nfinished >= pcxt->nworkers_launched)
759 {
760 Assert(nfinished == pcxt->nworkers_launched);
761 break;
762 }
763
764 /*
765 * We didn't detect any living workers, but not all workers are
766 * known to have exited cleanly. Either not all workers have
767 * launched yet, or maybe some of them failed to start or
768 * terminated abnormally.
769 */
770 for (i = 0; i < pcxt->nworkers_launched; ++i)
771 {
772 pid_t pid;
773 shm_mq *mq;
774
775 /*
776 * If the worker is BGWH_NOT_YET_STARTED or BGWH_STARTED, we
777 * should just keep waiting. If it is BGWH_STOPPED, then
778 * further investigation is needed.
779 */
780 if (pcxt->worker[i].error_mqh == NULL ||
781 pcxt->worker[i].bgwhandle == NULL ||
782 GetBackgroundWorkerPid(pcxt->worker[i].bgwhandle,
783 &pid) != BGWH_STOPPED)
784 continue;
785
786 /*
787 * Check whether the worker ended up stopped without ever
788 * attaching to the error queue. If so, the postmaster was
789 * unable to fork the worker or it exited without initializing
790 * properly. We must throw an error, since the caller may
791 * have been expecting the worker to do some work before
792 * exiting.
793 */
794 mq = shm_mq_get_queue(pcxt->worker[i].error_mqh);
795 if (shm_mq_get_sender(mq) == NULL)
796 ereport(ERROR,
797 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
798 errmsg("parallel worker failed to initialize"),
799 errhint("More details may be available in the server log.")));
800
801 /*
802 * The worker is stopped, but is attached to the error queue.
803 * Unless there's a bug somewhere, this will only happen when
804 * the worker writes messages and terminates after the
805 * CHECK_FOR_INTERRUPTS() near the top of this function and
806 * before the call to GetBackgroundWorkerPid(). In that case,
807 * or latch should have been set as well and the right things
808 * will happen on the next pass through the loop.
809 */
810 }
811 }
812
813 WaitLatch(MyLatch, WL_LATCH_SET, -1,
814 WAIT_EVENT_PARALLEL_FINISH);
815 ResetLatch(MyLatch);
816 }
817
818 if (pcxt->toc != NULL)
819 {
820 FixedParallelState *fps;
821
822 fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED, false);
823 if (fps->last_xlog_end > XactLastRecEnd)
824 XactLastRecEnd = fps->last_xlog_end;
825 }
826 }
827
828 /*
829 * Wait for all workers to exit.
830 *
831 * This function ensures that workers have been completely shutdown. The
832 * difference between WaitForParallelWorkersToFinish and this function is
833 * that former just ensures that last message sent by worker backend is
834 * received by master backend whereas this ensures the complete shutdown.
835 */
836 static void
WaitForParallelWorkersToExit(ParallelContext * pcxt)837 WaitForParallelWorkersToExit(ParallelContext *pcxt)
838 {
839 int i;
840
841 /* Wait until the workers actually die. */
842 for (i = 0; i < pcxt->nworkers_launched; ++i)
843 {
844 BgwHandleStatus status;
845
846 if (pcxt->worker == NULL || pcxt->worker[i].bgwhandle == NULL)
847 continue;
848
849 status = WaitForBackgroundWorkerShutdown(pcxt->worker[i].bgwhandle);
850
851 /*
852 * If the postmaster kicked the bucket, we have no chance of cleaning
853 * up safely -- we won't be able to tell when our workers are actually
854 * dead. This doesn't necessitate a PANIC since they will all abort
855 * eventually, but we can't safely continue this session.
856 */
857 if (status == BGWH_POSTMASTER_DIED)
858 ereport(FATAL,
859 (errcode(ERRCODE_ADMIN_SHUTDOWN),
860 errmsg("postmaster exited during a parallel transaction")));
861
862 /* Release memory. */
863 pfree(pcxt->worker[i].bgwhandle);
864 pcxt->worker[i].bgwhandle = NULL;
865 }
866 }
867
868 /*
869 * Destroy a parallel context.
870 *
871 * If expecting a clean exit, you should use WaitForParallelWorkersToFinish()
872 * first, before calling this function. When this function is invoked, any
873 * remaining workers are forcibly killed; the dynamic shared memory segment
874 * is unmapped; and we then wait (uninterruptibly) for the workers to exit.
875 */
876 void
DestroyParallelContext(ParallelContext * pcxt)877 DestroyParallelContext(ParallelContext *pcxt)
878 {
879 int i;
880
881 /*
882 * Be careful about order of operations here! We remove the parallel
883 * context from the list before we do anything else; otherwise, if an
884 * error occurs during a subsequent step, we might try to nuke it again
885 * from AtEOXact_Parallel or AtEOSubXact_Parallel.
886 */
887 dlist_delete(&pcxt->node);
888
889 /* Kill each worker in turn, and forget their error queues. */
890 if (pcxt->worker != NULL)
891 {
892 for (i = 0; i < pcxt->nworkers_launched; ++i)
893 {
894 if (pcxt->worker[i].error_mqh != NULL)
895 {
896 TerminateBackgroundWorker(pcxt->worker[i].bgwhandle);
897
898 shm_mq_detach(pcxt->worker[i].error_mqh);
899 pcxt->worker[i].error_mqh = NULL;
900 }
901 }
902 }
903
904 /*
905 * If we have allocated a shared memory segment, detach it. This will
906 * implicitly detach the error queues, and any other shared memory queues,
907 * stored there.
908 */
909 if (pcxt->seg != NULL)
910 {
911 dsm_detach(pcxt->seg);
912 pcxt->seg = NULL;
913 }
914
915 /*
916 * If this parallel context is actually in backend-private memory rather
917 * than shared memory, free that memory instead.
918 */
919 if (pcxt->private_memory != NULL)
920 {
921 pfree(pcxt->private_memory);
922 pcxt->private_memory = NULL;
923 }
924
925 /*
926 * We can't finish transaction commit or abort until all of the workers
927 * have exited. This means, in particular, that we can't respond to
928 * interrupts at this stage.
929 */
930 HOLD_INTERRUPTS();
931 WaitForParallelWorkersToExit(pcxt);
932 RESUME_INTERRUPTS();
933
934 /* Free the worker array itself. */
935 if (pcxt->worker != NULL)
936 {
937 pfree(pcxt->worker);
938 pcxt->worker = NULL;
939 }
940
941 /* Free memory. */
942 pfree(pcxt->library_name);
943 pfree(pcxt->function_name);
944 pfree(pcxt);
945 }
946
947 /*
948 * Are there any parallel contexts currently active?
949 */
950 bool
ParallelContextActive(void)951 ParallelContextActive(void)
952 {
953 return !dlist_is_empty(&pcxt_list);
954 }
955
956 /*
957 * Handle receipt of an interrupt indicating a parallel worker message.
958 *
959 * Note: this is called within a signal handler! All we can do is set
960 * a flag that will cause the next CHECK_FOR_INTERRUPTS() to invoke
961 * HandleParallelMessages().
962 */
963 void
HandleParallelMessageInterrupt(void)964 HandleParallelMessageInterrupt(void)
965 {
966 InterruptPending = true;
967 ParallelMessagePending = true;
968 SetLatch(MyLatch);
969 }
970
971 /*
972 * Handle any queued protocol messages received from parallel workers.
973 */
974 void
HandleParallelMessages(void)975 HandleParallelMessages(void)
976 {
977 dlist_iter iter;
978 MemoryContext oldcontext;
979
980 static MemoryContext hpm_context = NULL;
981
982 /*
983 * This is invoked from ProcessInterrupts(), and since some of the
984 * functions it calls contain CHECK_FOR_INTERRUPTS(), there is a potential
985 * for recursive calls if more signals are received while this runs. It's
986 * unclear that recursive entry would be safe, and it doesn't seem useful
987 * even if it is safe, so let's block interrupts until done.
988 */
989 HOLD_INTERRUPTS();
990
991 /*
992 * Moreover, CurrentMemoryContext might be pointing almost anywhere. We
993 * don't want to risk leaking data into long-lived contexts, so let's do
994 * our work here in a private context that we can reset on each use.
995 */
996 if (hpm_context == NULL) /* first time through? */
997 hpm_context = AllocSetContextCreate(TopMemoryContext,
998 "HandleParallelMessages",
999 ALLOCSET_DEFAULT_SIZES);
1000 else
1001 MemoryContextReset(hpm_context);
1002
1003 oldcontext = MemoryContextSwitchTo(hpm_context);
1004
1005 /* OK to process messages. Reset the flag saying there are more to do. */
1006 ParallelMessagePending = false;
1007
1008 dlist_foreach(iter, &pcxt_list)
1009 {
1010 ParallelContext *pcxt;
1011 int i;
1012
1013 pcxt = dlist_container(ParallelContext, node, iter.cur);
1014 if (pcxt->worker == NULL)
1015 continue;
1016
1017 for (i = 0; i < pcxt->nworkers_launched; ++i)
1018 {
1019 /*
1020 * Read as many messages as we can from each worker, but stop when
1021 * either (1) the worker's error queue goes away, which can happen
1022 * if we receive a Terminate message from the worker; or (2) no
1023 * more messages can be read from the worker without blocking.
1024 */
1025 while (pcxt->worker[i].error_mqh != NULL)
1026 {
1027 shm_mq_result res;
1028 Size nbytes;
1029 void *data;
1030
1031 res = shm_mq_receive(pcxt->worker[i].error_mqh, &nbytes,
1032 &data, true);
1033 if (res == SHM_MQ_WOULD_BLOCK)
1034 break;
1035 else if (res == SHM_MQ_SUCCESS)
1036 {
1037 StringInfoData msg;
1038
1039 initStringInfo(&msg);
1040 appendBinaryStringInfo(&msg, data, nbytes);
1041 HandleParallelMessage(pcxt, i, &msg);
1042 pfree(msg.data);
1043 }
1044 else
1045 ereport(ERROR,
1046 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1047 errmsg("lost connection to parallel worker")));
1048 }
1049 }
1050 }
1051
1052 MemoryContextSwitchTo(oldcontext);
1053
1054 /* Might as well clear the context on our way out */
1055 MemoryContextReset(hpm_context);
1056
1057 RESUME_INTERRUPTS();
1058 }
1059
1060 /*
1061 * Handle a single protocol message received from a single parallel worker.
1062 */
1063 static void
HandleParallelMessage(ParallelContext * pcxt,int i,StringInfo msg)1064 HandleParallelMessage(ParallelContext *pcxt, int i, StringInfo msg)
1065 {
1066 char msgtype;
1067
1068 if (pcxt->known_attached_workers != NULL &&
1069 !pcxt->known_attached_workers[i])
1070 {
1071 pcxt->known_attached_workers[i] = true;
1072 pcxt->nknown_attached_workers++;
1073 }
1074
1075 msgtype = pq_getmsgbyte(msg);
1076
1077 switch (msgtype)
1078 {
1079 case 'K': /* BackendKeyData */
1080 {
1081 int32 pid = pq_getmsgint(msg, 4);
1082
1083 (void) pq_getmsgint(msg, 4); /* discard cancel key */
1084 (void) pq_getmsgend(msg);
1085 pcxt->worker[i].pid = pid;
1086 break;
1087 }
1088
1089 case 'E': /* ErrorResponse */
1090 case 'N': /* NoticeResponse */
1091 {
1092 ErrorData edata;
1093 ErrorContextCallback *save_error_context_stack;
1094
1095 /* Parse ErrorResponse or NoticeResponse. */
1096 pq_parse_errornotice(msg, &edata);
1097
1098 /* Death of a worker isn't enough justification for suicide. */
1099 edata.elevel = Min(edata.elevel, ERROR);
1100
1101 /*
1102 * If desired, add a context line to show that this is a
1103 * message propagated from a parallel worker. Otherwise, it
1104 * can sometimes be confusing to understand what actually
1105 * happened. (We don't do this in FORCE_PARALLEL_REGRESS mode
1106 * because it causes test-result instability depending on
1107 * whether a parallel worker is actually used or not.)
1108 */
1109 if (force_parallel_mode != FORCE_PARALLEL_REGRESS)
1110 {
1111 if (edata.context)
1112 edata.context = psprintf("%s\n%s", edata.context,
1113 _("parallel worker"));
1114 else
1115 edata.context = pstrdup(_("parallel worker"));
1116 }
1117
1118 /*
1119 * Context beyond that should use the error context callbacks
1120 * that were in effect when the ParallelContext was created,
1121 * not the current ones.
1122 */
1123 save_error_context_stack = error_context_stack;
1124 error_context_stack = pcxt->error_context_stack;
1125
1126 /* Rethrow error or print notice. */
1127 ThrowErrorData(&edata);
1128
1129 /* Not an error, so restore previous context stack. */
1130 error_context_stack = save_error_context_stack;
1131
1132 break;
1133 }
1134
1135 case 'A': /* NotifyResponse */
1136 {
1137 /* Propagate NotifyResponse. */
1138 int32 pid;
1139 const char *channel;
1140 const char *payload;
1141
1142 pid = pq_getmsgint(msg, 4);
1143 channel = pq_getmsgrawstring(msg);
1144 payload = pq_getmsgrawstring(msg);
1145 pq_endmessage(msg);
1146
1147 NotifyMyFrontEnd(channel, payload, pid);
1148
1149 break;
1150 }
1151
1152 case 'X': /* Terminate, indicating clean exit */
1153 {
1154 shm_mq_detach(pcxt->worker[i].error_mqh);
1155 pcxt->worker[i].error_mqh = NULL;
1156 break;
1157 }
1158
1159 default:
1160 {
1161 elog(ERROR, "unrecognized message type received from parallel worker: %c (message length %d bytes)",
1162 msgtype, msg->len);
1163 }
1164 }
1165 }
1166
1167 /*
1168 * End-of-subtransaction cleanup for parallel contexts.
1169 *
1170 * Currently, it's forbidden to enter or leave a subtransaction while
1171 * parallel mode is in effect, so we could just blow away everything. But
1172 * we may want to relax that restriction in the future, so this code
1173 * contemplates that there may be multiple subtransaction IDs in pcxt_list.
1174 */
1175 void
AtEOSubXact_Parallel(bool isCommit,SubTransactionId mySubId)1176 AtEOSubXact_Parallel(bool isCommit, SubTransactionId mySubId)
1177 {
1178 while (!dlist_is_empty(&pcxt_list))
1179 {
1180 ParallelContext *pcxt;
1181
1182 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
1183 if (pcxt->subid != mySubId)
1184 break;
1185 if (isCommit)
1186 elog(WARNING, "leaked parallel context");
1187 DestroyParallelContext(pcxt);
1188 }
1189 }
1190
1191 /*
1192 * End-of-transaction cleanup for parallel contexts.
1193 */
1194 void
AtEOXact_Parallel(bool isCommit)1195 AtEOXact_Parallel(bool isCommit)
1196 {
1197 while (!dlist_is_empty(&pcxt_list))
1198 {
1199 ParallelContext *pcxt;
1200
1201 pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
1202 if (isCommit)
1203 elog(WARNING, "leaked parallel context");
1204 DestroyParallelContext(pcxt);
1205 }
1206 }
1207
1208 /*
1209 * Main entrypoint for parallel workers.
1210 */
1211 void
ParallelWorkerMain(Datum main_arg)1212 ParallelWorkerMain(Datum main_arg)
1213 {
1214 dsm_segment *seg;
1215 shm_toc *toc;
1216 FixedParallelState *fps;
1217 char *error_queue_space;
1218 shm_mq *mq;
1219 shm_mq_handle *mqh;
1220 char *libraryspace;
1221 char *entrypointstate;
1222 char *library_name;
1223 char *function_name;
1224 parallel_worker_main_type entrypt;
1225 char *gucspace;
1226 char *combocidspace;
1227 char *tsnapspace;
1228 char *asnapspace;
1229 char *tstatespace;
1230 char *reindexspace;
1231 StringInfoData msgbuf;
1232 char *session_dsm_handle_space;
1233 Snapshot tsnapshot;
1234 Snapshot asnapshot;
1235
1236 /* Set flag to indicate that we're initializing a parallel worker. */
1237 InitializingParallelWorker = true;
1238
1239 /* Establish signal handlers. */
1240 pqsignal(SIGTERM, die);
1241 BackgroundWorkerUnblockSignals();
1242
1243 /* Determine and set our parallel worker number. */
1244 Assert(ParallelWorkerNumber == -1);
1245 memcpy(&ParallelWorkerNumber, MyBgworkerEntry->bgw_extra, sizeof(int));
1246
1247 /* Set up a memory context and resource owner. */
1248 Assert(CurrentResourceOwner == NULL);
1249 CurrentResourceOwner = ResourceOwnerCreate(NULL, "parallel toplevel");
1250 CurrentMemoryContext = AllocSetContextCreate(TopMemoryContext,
1251 "Parallel worker",
1252 ALLOCSET_DEFAULT_SIZES);
1253
1254 /*
1255 * Now that we have a resource owner, we can attach to the dynamic shared
1256 * memory segment and read the table of contents.
1257 */
1258 seg = dsm_attach(DatumGetUInt32(main_arg));
1259 if (seg == NULL)
1260 ereport(ERROR,
1261 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1262 errmsg("could not map dynamic shared memory segment")));
1263 toc = shm_toc_attach(PARALLEL_MAGIC, dsm_segment_address(seg));
1264 if (toc == NULL)
1265 ereport(ERROR,
1266 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1267 errmsg("invalid magic number in dynamic shared memory segment")));
1268
1269 /* Look up fixed parallel state. */
1270 fps = shm_toc_lookup(toc, PARALLEL_KEY_FIXED, false);
1271 MyFixedParallelState = fps;
1272
1273 /* Arrange to signal the leader if we exit. */
1274 ParallelMasterPid = fps->parallel_master_pid;
1275 ParallelMasterBackendId = fps->parallel_master_backend_id;
1276 on_shmem_exit(ParallelWorkerShutdown, (Datum) 0);
1277
1278 /*
1279 * Now we can find and attach to the error queue provided for us. That's
1280 * good, because until we do that, any errors that happen here will not be
1281 * reported back to the process that requested that this worker be
1282 * launched.
1283 */
1284 error_queue_space = shm_toc_lookup(toc, PARALLEL_KEY_ERROR_QUEUE, false);
1285 mq = (shm_mq *) (error_queue_space +
1286 ParallelWorkerNumber * PARALLEL_ERROR_QUEUE_SIZE);
1287 shm_mq_set_sender(mq, MyProc);
1288 mqh = shm_mq_attach(mq, seg, NULL);
1289 pq_redirect_to_shm_mq(seg, mqh);
1290 pq_set_parallel_master(fps->parallel_master_pid,
1291 fps->parallel_master_backend_id);
1292
1293 /*
1294 * Send a BackendKeyData message to the process that initiated parallelism
1295 * so that it has access to our PID before it receives any other messages
1296 * from us. Our cancel key is sent, too, since that's the way the
1297 * protocol message is defined, but it won't actually be used for anything
1298 * in this case.
1299 */
1300 pq_beginmessage(&msgbuf, 'K');
1301 pq_sendint32(&msgbuf, (int32) MyProcPid);
1302 pq_sendint32(&msgbuf, (int32) MyCancelKey);
1303 pq_endmessage(&msgbuf);
1304
1305 /*
1306 * Hooray! Primary initialization is complete. Now, we need to set up our
1307 * backend-local state to match the original backend.
1308 */
1309
1310 /*
1311 * Join locking group. We must do this before anything that could try to
1312 * acquire a heavyweight lock, because any heavyweight locks acquired to
1313 * this point could block either directly against the parallel group
1314 * leader or against some process which in turn waits for a lock that
1315 * conflicts with the parallel group leader, causing an undetected
1316 * deadlock. (If we can't join the lock group, the leader has gone away,
1317 * so just exit quietly.)
1318 */
1319 if (!BecomeLockGroupMember(fps->parallel_master_pgproc,
1320 fps->parallel_master_pid))
1321 return;
1322
1323 /*
1324 * Restore transaction and statement start-time timestamps. This must
1325 * happen before anything that would start a transaction, else asserts in
1326 * xact.c will fire.
1327 */
1328 SetParallelStartTimestamps(fps->xact_ts, fps->stmt_ts);
1329
1330 /*
1331 * Identify the entry point to be called. In theory this could result in
1332 * loading an additional library, though most likely the entry point is in
1333 * the core backend or in a library we just loaded.
1334 */
1335 entrypointstate = shm_toc_lookup(toc, PARALLEL_KEY_ENTRYPOINT, false);
1336 library_name = entrypointstate;
1337 function_name = entrypointstate + strlen(library_name) + 1;
1338
1339 entrypt = LookupParallelWorkerFunction(library_name, function_name);
1340
1341 /* Restore database connection. */
1342 BackgroundWorkerInitializeConnectionByOid(fps->database_id,
1343 fps->authenticated_user_id,
1344 0);
1345
1346 /*
1347 * Set the client encoding to the database encoding, since that is what
1348 * the leader will expect.
1349 */
1350 SetClientEncoding(GetDatabaseEncoding());
1351
1352 /*
1353 * Load libraries that were loaded by original backend. We want to do
1354 * this before restoring GUCs, because the libraries might define custom
1355 * variables.
1356 */
1357 libraryspace = shm_toc_lookup(toc, PARALLEL_KEY_LIBRARY, false);
1358 StartTransactionCommand();
1359 RestoreLibraryState(libraryspace);
1360
1361 /* Restore GUC values from launching backend. */
1362 gucspace = shm_toc_lookup(toc, PARALLEL_KEY_GUC, false);
1363 RestoreGUCState(gucspace);
1364 CommitTransactionCommand();
1365
1366 /* Crank up a transaction state appropriate to a parallel worker. */
1367 tstatespace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_STATE, false);
1368 StartParallelWorkerTransaction(tstatespace);
1369
1370 /* Restore combo CID state. */
1371 combocidspace = shm_toc_lookup(toc, PARALLEL_KEY_COMBO_CID, false);
1372 RestoreComboCIDState(combocidspace);
1373
1374 /* Attach to the per-session DSM segment and contained objects. */
1375 session_dsm_handle_space =
1376 shm_toc_lookup(toc, PARALLEL_KEY_SESSION_DSM, false);
1377 AttachSession(*(dsm_handle *) session_dsm_handle_space);
1378
1379 /*
1380 * If the transaction isolation level is REPEATABLE READ or SERIALIZABLE,
1381 * the leader has serialized the transaction snapshot and we must restore
1382 * it. At lower isolation levels, there is no transaction-lifetime
1383 * snapshot, but we need TransactionXmin to get set to a value which is
1384 * less than or equal to the xmin of every snapshot that will be used by
1385 * this worker. The easiest way to accomplish that is to install the
1386 * active snapshot as the transaction snapshot. Code running in this
1387 * parallel worker might take new snapshots via GetTransactionSnapshot()
1388 * or GetLatestSnapshot(), but it shouldn't have any way of acquiring a
1389 * snapshot older than the active snapshot.
1390 */
1391 asnapspace = shm_toc_lookup(toc, PARALLEL_KEY_ACTIVE_SNAPSHOT, false);
1392 tsnapspace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT, true);
1393 asnapshot = RestoreSnapshot(asnapspace);
1394 tsnapshot = tsnapspace ? RestoreSnapshot(tsnapspace) : asnapshot;
1395 RestoreTransactionSnapshot(tsnapshot,
1396 fps->parallel_master_pgproc);
1397 PushActiveSnapshot(asnapshot);
1398
1399 /*
1400 * We've changed which tuples we can see, and must therefore invalidate
1401 * system caches.
1402 */
1403 InvalidateSystemCaches();
1404
1405 /*
1406 * Restore current role id. Skip verifying whether session user is
1407 * allowed to become this role and blindly restore the leader's state for
1408 * current role.
1409 */
1410 SetCurrentRoleId(fps->outer_user_id, fps->is_superuser);
1411
1412 /* Restore user ID and security context. */
1413 SetUserIdAndSecContext(fps->current_user_id, fps->sec_context);
1414
1415 /* Restore temp-namespace state to ensure search path matches leader's. */
1416 SetTempNamespaceState(fps->temp_namespace_id,
1417 fps->temp_toast_namespace_id);
1418
1419 /* Restore reindex state. */
1420 reindexspace = shm_toc_lookup(toc, PARALLEL_KEY_REINDEX_STATE, false);
1421 RestoreReindexState(reindexspace);
1422
1423 /*
1424 * We've initialized all of our state now; nothing should change
1425 * hereafter.
1426 */
1427 InitializingParallelWorker = false;
1428 EnterParallelMode();
1429
1430 /*
1431 * Time to do the real work: invoke the caller-supplied code.
1432 */
1433 entrypt(seg, toc);
1434
1435 /* Must exit parallel mode to pop active snapshot. */
1436 ExitParallelMode();
1437
1438 /* Must pop active snapshot so resowner.c doesn't complain. */
1439 PopActiveSnapshot();
1440
1441 /* Shut down the parallel-worker transaction. */
1442 EndParallelWorkerTransaction();
1443
1444 /* Detach from the per-session DSM segment. */
1445 DetachSession();
1446
1447 /* Report success. */
1448 pq_putmessage('X', NULL, 0);
1449 }
1450
1451 /*
1452 * Update shared memory with the ending location of the last WAL record we
1453 * wrote, if it's greater than the value already stored there.
1454 */
1455 void
ParallelWorkerReportLastRecEnd(XLogRecPtr last_xlog_end)1456 ParallelWorkerReportLastRecEnd(XLogRecPtr last_xlog_end)
1457 {
1458 FixedParallelState *fps = MyFixedParallelState;
1459
1460 Assert(fps != NULL);
1461 SpinLockAcquire(&fps->mutex);
1462 if (fps->last_xlog_end < last_xlog_end)
1463 fps->last_xlog_end = last_xlog_end;
1464 SpinLockRelease(&fps->mutex);
1465 }
1466
1467 /*
1468 * Make sure the leader tries to read from our error queue one more time.
1469 * This guards against the case where we exit uncleanly without sending an
1470 * ErrorResponse to the leader, for example because some code calls proc_exit
1471 * directly.
1472 */
1473 static void
ParallelWorkerShutdown(int code,Datum arg)1474 ParallelWorkerShutdown(int code, Datum arg)
1475 {
1476 SendProcSignal(ParallelMasterPid,
1477 PROCSIG_PARALLEL_MESSAGE,
1478 ParallelMasterBackendId);
1479 }
1480
1481 /*
1482 * Look up (and possibly load) a parallel worker entry point function.
1483 *
1484 * For functions contained in the core code, we use library name "postgres"
1485 * and consult the InternalParallelWorkers array. External functions are
1486 * looked up, and loaded if necessary, using load_external_function().
1487 *
1488 * The point of this is to pass function names as strings across process
1489 * boundaries. We can't pass actual function addresses because of the
1490 * possibility that the function has been loaded at a different address
1491 * in a different process. This is obviously a hazard for functions in
1492 * loadable libraries, but it can happen even for functions in the core code
1493 * on platforms using EXEC_BACKEND (e.g., Windows).
1494 *
1495 * At some point it might be worthwhile to get rid of InternalParallelWorkers[]
1496 * in favor of applying load_external_function() for core functions too;
1497 * but that raises portability issues that are not worth addressing now.
1498 */
1499 static parallel_worker_main_type
LookupParallelWorkerFunction(const char * libraryname,const char * funcname)1500 LookupParallelWorkerFunction(const char *libraryname, const char *funcname)
1501 {
1502 /*
1503 * If the function is to be loaded from postgres itself, search the
1504 * InternalParallelWorkers array.
1505 */
1506 if (strcmp(libraryname, "postgres") == 0)
1507 {
1508 int i;
1509
1510 for (i = 0; i < lengthof(InternalParallelWorkers); i++)
1511 {
1512 if (strcmp(InternalParallelWorkers[i].fn_name, funcname) == 0)
1513 return InternalParallelWorkers[i].fn_addr;
1514 }
1515
1516 /* We can only reach this by programming error. */
1517 elog(ERROR, "internal function \"%s\" not found", funcname);
1518 }
1519
1520 /* Otherwise load from external library. */
1521 return (parallel_worker_main_type)
1522 load_external_function(libraryname, funcname, true, NULL);
1523 }
1524