Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * parallel.c
4 : * Infrastructure for launching parallel workers
5 : *
6 : * Portions Copyright (c) 1996-2017, 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/parallel.h"
18 : #include "access/xact.h"
19 : #include "access/xlog.h"
20 : #include "catalog/namespace.h"
21 : #include "commands/async.h"
22 : #include "executor/execParallel.h"
23 : #include "libpq/libpq.h"
24 : #include "libpq/pqformat.h"
25 : #include "libpq/pqmq.h"
26 : #include "miscadmin.h"
27 : #include "optimizer/planmain.h"
28 : #include "pgstat.h"
29 : #include "storage/ipc.h"
30 : #include "storage/sinval.h"
31 : #include "storage/spin.h"
32 : #include "tcop/tcopprot.h"
33 : #include "utils/combocid.h"
34 : #include "utils/guc.h"
35 : #include "utils/inval.h"
36 : #include "utils/memutils.h"
37 : #include "utils/resowner.h"
38 : #include "utils/snapmgr.h"
39 :
40 :
41 : /*
42 : * We don't want to waste a lot of memory on an error queue which, most of
43 : * the time, will process only a handful of small messages. However, it is
44 : * desirable to make it large enough that a typical ErrorResponse can be sent
45 : * without blocking. That way, a worker that errors out can write the whole
46 : * message into the queue and terminate without waiting for the user backend.
47 : */
48 : #define PARALLEL_ERROR_QUEUE_SIZE 16384
49 :
50 : /* Magic number for parallel context TOC. */
51 : #define PARALLEL_MAGIC 0x50477c7c
52 :
53 : /*
54 : * Magic numbers for parallel state sharing. Higher-level code should use
55 : * smaller values, leaving these very large ones for use by this module.
56 : */
57 : #define PARALLEL_KEY_FIXED UINT64CONST(0xFFFFFFFFFFFF0001)
58 : #define PARALLEL_KEY_ERROR_QUEUE UINT64CONST(0xFFFFFFFFFFFF0002)
59 : #define PARALLEL_KEY_LIBRARY UINT64CONST(0xFFFFFFFFFFFF0003)
60 : #define PARALLEL_KEY_GUC UINT64CONST(0xFFFFFFFFFFFF0004)
61 : #define PARALLEL_KEY_COMBO_CID UINT64CONST(0xFFFFFFFFFFFF0005)
62 : #define PARALLEL_KEY_TRANSACTION_SNAPSHOT UINT64CONST(0xFFFFFFFFFFFF0006)
63 : #define PARALLEL_KEY_ACTIVE_SNAPSHOT UINT64CONST(0xFFFFFFFFFFFF0007)
64 : #define PARALLEL_KEY_TRANSACTION_STATE UINT64CONST(0xFFFFFFFFFFFF0008)
65 : #define PARALLEL_KEY_ENTRYPOINT UINT64CONST(0xFFFFFFFFFFFF0009)
66 :
67 : /* Fixed-size parallel state. */
68 : typedef struct FixedParallelState
69 : {
70 : /* Fixed-size state that workers must restore. */
71 : Oid database_id;
72 : Oid authenticated_user_id;
73 : Oid current_user_id;
74 : Oid temp_namespace_id;
75 : Oid temp_toast_namespace_id;
76 : int sec_context;
77 : PGPROC *parallel_master_pgproc;
78 : pid_t parallel_master_pid;
79 : BackendId parallel_master_backend_id;
80 :
81 : /* Mutex protects remaining fields. */
82 : slock_t mutex;
83 :
84 : /* Maximum XactLastRecEnd of any worker. */
85 : XLogRecPtr last_xlog_end;
86 : } FixedParallelState;
87 :
88 : /*
89 : * Our parallel worker number. We initialize this to -1, meaning that we are
90 : * not a parallel worker. In parallel workers, it will be set to a value >= 0
91 : * and < the number of workers before any user code is invoked; each parallel
92 : * worker will get a different parallel worker number.
93 : */
94 : int ParallelWorkerNumber = -1;
95 :
96 : /* Is there a parallel message pending which we need to receive? */
97 : volatile bool ParallelMessagePending = false;
98 :
99 : /* Are we initializing a parallel worker? */
100 : bool InitializingParallelWorker = false;
101 :
102 : /* Pointer to our fixed parallel state. */
103 : static FixedParallelState *MyFixedParallelState;
104 :
105 : /* List of active parallel contexts. */
106 : static dlist_head pcxt_list = DLIST_STATIC_INIT(pcxt_list);
107 :
108 : /*
109 : * List of internal parallel worker entry points. We need this for
110 : * reasons explained in LookupParallelWorkerFunction(), below.
111 : */
112 : static const struct
113 : {
114 : const char *fn_name;
115 : parallel_worker_main_type fn_addr;
116 : } InternalParallelWorkers[] =
117 :
118 : {
119 : {
120 : "ParallelQueryMain", ParallelQueryMain
121 : }
122 : };
123 :
124 : /* Private functions. */
125 : static void HandleParallelMessage(ParallelContext *pcxt, int i, StringInfo msg);
126 : static void WaitForParallelWorkersToExit(ParallelContext *pcxt);
127 : static parallel_worker_main_type LookupParallelWorkerFunction(const char *libraryname, const char *funcname);
128 :
129 :
130 : /*
131 : * Establish a new parallel context. This should be done after entering
132 : * parallel mode, and (unless there is an error) the context should be
133 : * destroyed before exiting the current subtransaction.
134 : */
135 : ParallelContext *
136 17 : CreateParallelContext(const char *library_name, const char *function_name,
137 : int nworkers)
138 : {
139 : MemoryContext oldcontext;
140 : ParallelContext *pcxt;
141 :
142 : /* It is unsafe to create a parallel context if not in parallel mode. */
143 17 : Assert(IsInParallelMode());
144 :
145 : /* Number of workers should be non-negative. */
146 17 : Assert(nworkers >= 0);
147 :
148 : /*
149 : * If dynamic shared memory is not available, we won't be able to use
150 : * background workers.
151 : */
152 17 : if (dynamic_shared_memory_type == DSM_IMPL_NONE)
153 0 : nworkers = 0;
154 :
155 : /*
156 : * If we are running under serializable isolation, we can't use parallel
157 : * workers, at least not until somebody enhances that mechanism to be
158 : * parallel-aware.
159 : */
160 17 : if (IsolationIsSerializable())
161 0 : nworkers = 0;
162 :
163 : /* We might be running in a short-lived memory context. */
164 17 : oldcontext = MemoryContextSwitchTo(TopTransactionContext);
165 :
166 : /* Initialize a new ParallelContext. */
167 17 : pcxt = palloc0(sizeof(ParallelContext));
168 17 : pcxt->subid = GetCurrentSubTransactionId();
169 17 : pcxt->nworkers = nworkers;
170 17 : pcxt->library_name = pstrdup(library_name);
171 17 : pcxt->function_name = pstrdup(function_name);
172 17 : pcxt->error_context_stack = error_context_stack;
173 17 : shm_toc_initialize_estimator(&pcxt->estimator);
174 17 : dlist_push_head(&pcxt_list, &pcxt->node);
175 :
176 : /* Restore previous memory context. */
177 17 : MemoryContextSwitchTo(oldcontext);
178 :
179 17 : return pcxt;
180 : }
181 :
182 : /*
183 : * Establish the dynamic shared memory segment for a parallel context and
184 : * copy state and other bookkeeping information that will be needed by
185 : * parallel workers into it.
186 : */
187 : void
188 17 : InitializeParallelDSM(ParallelContext *pcxt)
189 : {
190 : MemoryContext oldcontext;
191 17 : Size library_len = 0;
192 17 : Size guc_len = 0;
193 17 : Size combocidlen = 0;
194 17 : Size tsnaplen = 0;
195 17 : Size asnaplen = 0;
196 17 : Size tstatelen = 0;
197 17 : Size segsize = 0;
198 : int i;
199 : FixedParallelState *fps;
200 17 : Snapshot transaction_snapshot = GetTransactionSnapshot();
201 17 : Snapshot active_snapshot = GetActiveSnapshot();
202 :
203 : /* We might be running in a very short-lived memory context. */
204 17 : oldcontext = MemoryContextSwitchTo(TopTransactionContext);
205 :
206 : /* Allow space to store the fixed-size parallel state. */
207 17 : shm_toc_estimate_chunk(&pcxt->estimator, sizeof(FixedParallelState));
208 17 : shm_toc_estimate_keys(&pcxt->estimator, 1);
209 :
210 : /*
211 : * Normally, the user will have requested at least one worker process, but
212 : * if by chance they have not, we can skip a bunch of things here.
213 : */
214 17 : if (pcxt->nworkers > 0)
215 : {
216 : /* Estimate space for various kinds of state sharing. */
217 17 : library_len = EstimateLibraryStateSpace();
218 17 : shm_toc_estimate_chunk(&pcxt->estimator, library_len);
219 17 : guc_len = EstimateGUCStateSpace();
220 17 : shm_toc_estimate_chunk(&pcxt->estimator, guc_len);
221 17 : combocidlen = EstimateComboCIDStateSpace();
222 17 : shm_toc_estimate_chunk(&pcxt->estimator, combocidlen);
223 17 : tsnaplen = EstimateSnapshotSpace(transaction_snapshot);
224 17 : shm_toc_estimate_chunk(&pcxt->estimator, tsnaplen);
225 17 : asnaplen = EstimateSnapshotSpace(active_snapshot);
226 17 : shm_toc_estimate_chunk(&pcxt->estimator, asnaplen);
227 17 : tstatelen = EstimateTransactionStateSpace();
228 17 : shm_toc_estimate_chunk(&pcxt->estimator, tstatelen);
229 : /* If you add more chunks here, you probably need to add keys. */
230 17 : shm_toc_estimate_keys(&pcxt->estimator, 6);
231 :
232 : /* Estimate space need for error queues. */
233 : StaticAssertStmt(BUFFERALIGN(PARALLEL_ERROR_QUEUE_SIZE) ==
234 : PARALLEL_ERROR_QUEUE_SIZE,
235 : "parallel error queue size not buffer-aligned");
236 17 : shm_toc_estimate_chunk(&pcxt->estimator,
237 : mul_size(PARALLEL_ERROR_QUEUE_SIZE,
238 : pcxt->nworkers));
239 17 : shm_toc_estimate_keys(&pcxt->estimator, 1);
240 :
241 : /* Estimate how much we'll need for the entrypoint info. */
242 17 : shm_toc_estimate_chunk(&pcxt->estimator, strlen(pcxt->library_name) +
243 : strlen(pcxt->function_name) + 2);
244 17 : shm_toc_estimate_keys(&pcxt->estimator, 1);
245 : }
246 :
247 : /*
248 : * Create DSM and initialize with new table of contents. But if the user
249 : * didn't request any workers, then don't bother creating a dynamic shared
250 : * memory segment; instead, just use backend-private memory.
251 : *
252 : * Also, if we can't create a dynamic shared memory segment because the
253 : * maximum number of segments have already been created, then fall back to
254 : * backend-private memory, and plan not to use any workers. We hope this
255 : * won't happen very often, but it's better to abandon the use of
256 : * parallelism than to fail outright.
257 : */
258 17 : segsize = shm_toc_estimate(&pcxt->estimator);
259 17 : if (pcxt->nworkers > 0)
260 17 : pcxt->seg = dsm_create(segsize, DSM_CREATE_NULL_IF_MAXSEGMENTS);
261 17 : if (pcxt->seg != NULL)
262 17 : pcxt->toc = shm_toc_create(PARALLEL_MAGIC,
263 : dsm_segment_address(pcxt->seg),
264 : segsize);
265 : else
266 : {
267 0 : pcxt->nworkers = 0;
268 0 : pcxt->private_memory = MemoryContextAlloc(TopMemoryContext, segsize);
269 0 : pcxt->toc = shm_toc_create(PARALLEL_MAGIC, pcxt->private_memory,
270 : segsize);
271 : }
272 :
273 : /* Initialize fixed-size state in shared memory. */
274 17 : fps = (FixedParallelState *)
275 17 : shm_toc_allocate(pcxt->toc, sizeof(FixedParallelState));
276 17 : fps->database_id = MyDatabaseId;
277 17 : fps->authenticated_user_id = GetAuthenticatedUserId();
278 17 : GetUserIdAndSecContext(&fps->current_user_id, &fps->sec_context);
279 17 : GetTempNamespaceState(&fps->temp_namespace_id,
280 : &fps->temp_toast_namespace_id);
281 17 : fps->parallel_master_pgproc = MyProc;
282 17 : fps->parallel_master_pid = MyProcPid;
283 17 : fps->parallel_master_backend_id = MyBackendId;
284 17 : SpinLockInit(&fps->mutex);
285 17 : fps->last_xlog_end = 0;
286 17 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_FIXED, fps);
287 :
288 : /* We can skip the rest of this if we're not budgeting for any workers. */
289 17 : if (pcxt->nworkers > 0)
290 : {
291 : char *libraryspace;
292 : char *gucspace;
293 : char *combocidspace;
294 : char *tsnapspace;
295 : char *asnapspace;
296 : char *tstatespace;
297 : char *error_queue_space;
298 : char *entrypointstate;
299 : Size lnamelen;
300 :
301 : /* Serialize shared libraries we have loaded. */
302 17 : libraryspace = shm_toc_allocate(pcxt->toc, library_len);
303 17 : SerializeLibraryState(library_len, libraryspace);
304 17 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_LIBRARY, libraryspace);
305 :
306 : /* Serialize GUC settings. */
307 17 : gucspace = shm_toc_allocate(pcxt->toc, guc_len);
308 17 : SerializeGUCState(guc_len, gucspace);
309 17 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_GUC, gucspace);
310 :
311 : /* Serialize combo CID state. */
312 17 : combocidspace = shm_toc_allocate(pcxt->toc, combocidlen);
313 17 : SerializeComboCIDState(combocidlen, combocidspace);
314 17 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_COMBO_CID, combocidspace);
315 :
316 : /* Serialize transaction snapshot and active snapshot. */
317 17 : tsnapspace = shm_toc_allocate(pcxt->toc, tsnaplen);
318 17 : SerializeSnapshot(transaction_snapshot, tsnapspace);
319 17 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT,
320 : tsnapspace);
321 17 : asnapspace = shm_toc_allocate(pcxt->toc, asnaplen);
322 17 : SerializeSnapshot(active_snapshot, asnapspace);
323 17 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_ACTIVE_SNAPSHOT, asnapspace);
324 :
325 : /* Serialize transaction state. */
326 17 : tstatespace = shm_toc_allocate(pcxt->toc, tstatelen);
327 17 : SerializeTransactionState(tstatelen, tstatespace);
328 17 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_TRANSACTION_STATE, tstatespace);
329 :
330 : /* Allocate space for worker information. */
331 17 : pcxt->worker = palloc0(sizeof(ParallelWorkerInfo) * pcxt->nworkers);
332 :
333 : /*
334 : * Establish error queues in dynamic shared memory.
335 : *
336 : * These queues should be used only for transmitting ErrorResponse,
337 : * NoticeResponse, and NotifyResponse protocol messages. Tuple data
338 : * should be transmitted via separate (possibly larger?) queues.
339 : */
340 17 : error_queue_space =
341 17 : shm_toc_allocate(pcxt->toc,
342 : mul_size(PARALLEL_ERROR_QUEUE_SIZE,
343 17 : pcxt->nworkers));
344 76 : for (i = 0; i < pcxt->nworkers; ++i)
345 : {
346 : char *start;
347 : shm_mq *mq;
348 :
349 59 : start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
350 59 : mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
351 59 : shm_mq_set_receiver(mq, MyProc);
352 59 : pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
353 : }
354 17 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE, error_queue_space);
355 :
356 : /*
357 : * Serialize entrypoint information. It's unsafe to pass function
358 : * pointers across processes, as the function pointer may be different
359 : * in each process in EXEC_BACKEND builds, so we always pass library
360 : * and function name. (We use library name "postgres" for functions
361 : * in the core backend.)
362 : */
363 17 : lnamelen = strlen(pcxt->library_name);
364 17 : entrypointstate = shm_toc_allocate(pcxt->toc, lnamelen +
365 17 : strlen(pcxt->function_name) + 2);
366 17 : strcpy(entrypointstate, pcxt->library_name);
367 17 : strcpy(entrypointstate + lnamelen + 1, pcxt->function_name);
368 17 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_ENTRYPOINT, entrypointstate);
369 : }
370 :
371 : /* Restore previous memory context. */
372 17 : MemoryContextSwitchTo(oldcontext);
373 17 : }
374 :
375 : /*
376 : * Reinitialize the dynamic shared memory segment for a parallel context such
377 : * that we could launch workers for it again.
378 : */
379 : void
380 15 : ReinitializeParallelDSM(ParallelContext *pcxt)
381 : {
382 : FixedParallelState *fps;
383 : char *error_queue_space;
384 : int i;
385 :
386 : /* Wait for any old workers to exit. */
387 15 : if (pcxt->nworkers_launched > 0)
388 : {
389 15 : WaitForParallelWorkersToFinish(pcxt);
390 15 : WaitForParallelWorkersToExit(pcxt);
391 15 : pcxt->nworkers_launched = 0;
392 : }
393 :
394 : /* Reset a few bits of fixed parallel state to a clean state. */
395 15 : fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED, false);
396 15 : fps->last_xlog_end = 0;
397 :
398 : /* Recreate error queues. */
399 15 : error_queue_space =
400 15 : shm_toc_lookup(pcxt->toc, PARALLEL_KEY_ERROR_QUEUE, false);
401 75 : for (i = 0; i < pcxt->nworkers; ++i)
402 : {
403 : char *start;
404 : shm_mq *mq;
405 :
406 60 : start = error_queue_space + i * PARALLEL_ERROR_QUEUE_SIZE;
407 60 : mq = shm_mq_create(start, PARALLEL_ERROR_QUEUE_SIZE);
408 60 : shm_mq_set_receiver(mq, MyProc);
409 60 : pcxt->worker[i].error_mqh = shm_mq_attach(mq, pcxt->seg, NULL);
410 : }
411 15 : }
412 :
413 : /*
414 : * Launch parallel workers.
415 : */
416 : void
417 32 : LaunchParallelWorkers(ParallelContext *pcxt)
418 : {
419 : MemoryContext oldcontext;
420 : BackgroundWorker worker;
421 : int i;
422 32 : bool any_registrations_failed = false;
423 :
424 : /* Skip this if we have no workers. */
425 32 : if (pcxt->nworkers == 0)
426 32 : return;
427 :
428 : /* We need to be a lock group leader. */
429 32 : BecomeLockGroupLeader();
430 :
431 : /* If we do have workers, we'd better have a DSM segment. */
432 32 : Assert(pcxt->seg != NULL);
433 :
434 : /* We might be running in a short-lived memory context. */
435 32 : oldcontext = MemoryContextSwitchTo(TopTransactionContext);
436 :
437 : /* Configure a worker. */
438 32 : memset(&worker, 0, sizeof(worker));
439 32 : snprintf(worker.bgw_name, BGW_MAXLEN, "parallel worker for PID %d",
440 : MyProcPid);
441 32 : worker.bgw_flags =
442 : BGWORKER_SHMEM_ACCESS | BGWORKER_BACKEND_DATABASE_CONNECTION
443 : | BGWORKER_CLASS_PARALLEL;
444 32 : worker.bgw_start_time = BgWorkerStart_ConsistentState;
445 32 : worker.bgw_restart_time = BGW_NEVER_RESTART;
446 32 : sprintf(worker.bgw_library_name, "postgres");
447 32 : sprintf(worker.bgw_function_name, "ParallelWorkerMain");
448 32 : worker.bgw_main_arg = UInt32GetDatum(dsm_segment_handle(pcxt->seg));
449 32 : worker.bgw_notify_pid = MyProcPid;
450 :
451 : /*
452 : * Start workers.
453 : *
454 : * The caller must be able to tolerate ending up with fewer workers than
455 : * expected, so there is no need to throw an error here if registration
456 : * fails. It wouldn't help much anyway, because registering the worker in
457 : * no way guarantees that it will start up and initialize successfully.
458 : */
459 151 : for (i = 0; i < pcxt->nworkers; ++i)
460 : {
461 119 : memcpy(worker.bgw_extra, &i, sizeof(int));
462 235 : if (!any_registrations_failed &&
463 116 : RegisterDynamicBackgroundWorker(&worker,
464 116 : &pcxt->worker[i].bgwhandle))
465 : {
466 115 : shm_mq_set_handle(pcxt->worker[i].error_mqh,
467 115 : pcxt->worker[i].bgwhandle);
468 115 : pcxt->nworkers_launched++;
469 : }
470 : else
471 : {
472 : /*
473 : * If we weren't able to register the worker, then we've bumped up
474 : * against the max_worker_processes limit, and future
475 : * registrations will probably fail too, so arrange to skip them.
476 : * But we still have to execute this code for the remaining slots
477 : * to make sure that we forget about the error queues we budgeted
478 : * for those workers. Otherwise, we'll wait for them to start,
479 : * but they never will.
480 : */
481 4 : any_registrations_failed = true;
482 4 : pcxt->worker[i].bgwhandle = NULL;
483 4 : shm_mq_detach(pcxt->worker[i].error_mqh);
484 4 : pcxt->worker[i].error_mqh = NULL;
485 : }
486 : }
487 :
488 : /* Restore previous memory context. */
489 32 : MemoryContextSwitchTo(oldcontext);
490 : }
491 :
492 : /*
493 : * Wait for all workers to finish computing.
494 : *
495 : * Even if the parallel operation seems to have completed successfully, it's
496 : * important to call this function afterwards. We must not miss any errors
497 : * the workers may have thrown during the parallel operation, or any that they
498 : * may yet throw while shutting down.
499 : *
500 : * Also, we want to update our notion of XactLastRecEnd based on worker
501 : * feedback.
502 : */
503 : void
504 68 : WaitForParallelWorkersToFinish(ParallelContext *pcxt)
505 : {
506 : for (;;)
507 : {
508 68 : bool anyone_alive = false;
509 : int i;
510 :
511 : /*
512 : * This will process any parallel messages that are pending, which may
513 : * change the outcome of the loop that follows. It may also throw an
514 : * error propagated from a worker.
515 : */
516 68 : CHECK_FOR_INTERRUPTS();
517 :
518 242 : for (i = 0; i < pcxt->nworkers_launched; ++i)
519 : {
520 196 : if (pcxt->worker[i].error_mqh != NULL)
521 : {
522 22 : anyone_alive = true;
523 22 : break;
524 : }
525 : }
526 :
527 68 : if (!anyone_alive)
528 46 : break;
529 :
530 22 : WaitLatch(MyLatch, WL_LATCH_SET, -1,
531 : WAIT_EVENT_PARALLEL_FINISH);
532 22 : ResetLatch(MyLatch);
533 22 : }
534 :
535 46 : if (pcxt->toc != NULL)
536 : {
537 : FixedParallelState *fps;
538 :
539 46 : fps = shm_toc_lookup(pcxt->toc, PARALLEL_KEY_FIXED, false);
540 46 : if (fps->last_xlog_end > XactLastRecEnd)
541 0 : XactLastRecEnd = fps->last_xlog_end;
542 : }
543 46 : }
544 :
545 : /*
546 : * Wait for all workers to exit.
547 : *
548 : * This function ensures that workers have been completely shutdown. The
549 : * difference between WaitForParallelWorkersToFinish and this function is
550 : * that former just ensures that last message sent by worker backend is
551 : * received by master backend whereas this ensures the complete shutdown.
552 : */
553 : static void
554 32 : WaitForParallelWorkersToExit(ParallelContext *pcxt)
555 : {
556 : int i;
557 :
558 : /* Wait until the workers actually die. */
559 147 : for (i = 0; i < pcxt->nworkers_launched; ++i)
560 : {
561 : BgwHandleStatus status;
562 :
563 115 : if (pcxt->worker == NULL || pcxt->worker[i].bgwhandle == NULL)
564 0 : continue;
565 :
566 115 : status = WaitForBackgroundWorkerShutdown(pcxt->worker[i].bgwhandle);
567 :
568 : /*
569 : * If the postmaster kicked the bucket, we have no chance of cleaning
570 : * up safely -- we won't be able to tell when our workers are actually
571 : * dead. This doesn't necessitate a PANIC since they will all abort
572 : * eventually, but we can't safely continue this session.
573 : */
574 115 : if (status == BGWH_POSTMASTER_DIED)
575 0 : ereport(FATAL,
576 : (errcode(ERRCODE_ADMIN_SHUTDOWN),
577 : errmsg("postmaster exited during a parallel transaction")));
578 :
579 : /* Release memory. */
580 115 : pfree(pcxt->worker[i].bgwhandle);
581 115 : pcxt->worker[i].bgwhandle = NULL;
582 : }
583 32 : }
584 :
585 : /*
586 : * Destroy a parallel context.
587 : *
588 : * If expecting a clean exit, you should use WaitForParallelWorkersToFinish()
589 : * first, before calling this function. When this function is invoked, any
590 : * remaining workers are forcibly killed; the dynamic shared memory segment
591 : * is unmapped; and we then wait (uninterruptibly) for the workers to exit.
592 : */
593 : void
594 17 : DestroyParallelContext(ParallelContext *pcxt)
595 : {
596 : int i;
597 :
598 : /*
599 : * Be careful about order of operations here! We remove the parallel
600 : * context from the list before we do anything else; otherwise, if an
601 : * error occurs during a subsequent step, we might try to nuke it again
602 : * from AtEOXact_Parallel or AtEOSubXact_Parallel.
603 : */
604 17 : dlist_delete(&pcxt->node);
605 :
606 : /* Kill each worker in turn, and forget their error queues. */
607 17 : if (pcxt->worker != NULL)
608 : {
609 72 : for (i = 0; i < pcxt->nworkers_launched; ++i)
610 : {
611 55 : if (pcxt->worker[i].error_mqh != NULL)
612 : {
613 1 : TerminateBackgroundWorker(pcxt->worker[i].bgwhandle);
614 :
615 1 : shm_mq_detach(pcxt->worker[i].error_mqh);
616 1 : pcxt->worker[i].error_mqh = NULL;
617 : }
618 : }
619 : }
620 :
621 : /*
622 : * If we have allocated a shared memory segment, detach it. This will
623 : * implicitly detach the error queues, and any other shared memory queues,
624 : * stored there.
625 : */
626 17 : if (pcxt->seg != NULL)
627 : {
628 17 : dsm_detach(pcxt->seg);
629 17 : pcxt->seg = NULL;
630 : }
631 :
632 : /*
633 : * If this parallel context is actually in backend-private memory rather
634 : * than shared memory, free that memory instead.
635 : */
636 17 : if (pcxt->private_memory != NULL)
637 : {
638 0 : pfree(pcxt->private_memory);
639 0 : pcxt->private_memory = NULL;
640 : }
641 :
642 : /*
643 : * We can't finish transaction commit or abort until all of the workers
644 : * have exited. This means, in particular, that we can't respond to
645 : * interrupts at this stage.
646 : */
647 17 : HOLD_INTERRUPTS();
648 17 : WaitForParallelWorkersToExit(pcxt);
649 17 : RESUME_INTERRUPTS();
650 :
651 : /* Free the worker array itself. */
652 17 : if (pcxt->worker != NULL)
653 : {
654 17 : pfree(pcxt->worker);
655 17 : pcxt->worker = NULL;
656 : }
657 :
658 : /* Free memory. */
659 17 : pfree(pcxt->library_name);
660 17 : pfree(pcxt->function_name);
661 17 : pfree(pcxt);
662 17 : }
663 :
664 : /*
665 : * Are there any parallel contexts currently active?
666 : */
667 : bool
668 130 : ParallelContextActive(void)
669 : {
670 130 : return !dlist_is_empty(&pcxt_list);
671 : }
672 :
673 : /*
674 : * Handle receipt of an interrupt indicating a parallel worker message.
675 : *
676 : * Note: this is called within a signal handler! All we can do is set
677 : * a flag that will cause the next CHECK_FOR_INTERRUPTS() to invoke
678 : * HandleParallelMessages().
679 : */
680 : void
681 230 : HandleParallelMessageInterrupt(void)
682 : {
683 230 : InterruptPending = true;
684 230 : ParallelMessagePending = true;
685 230 : SetLatch(MyLatch);
686 230 : }
687 :
688 : /*
689 : * Handle any queued protocol messages received from parallel workers.
690 : */
691 : void
692 230 : HandleParallelMessages(void)
693 : {
694 : dlist_iter iter;
695 : MemoryContext oldcontext;
696 :
697 : static MemoryContext hpm_context = NULL;
698 :
699 : /*
700 : * This is invoked from ProcessInterrupts(), and since some of the
701 : * functions it calls contain CHECK_FOR_INTERRUPTS(), there is a potential
702 : * for recursive calls if more signals are received while this runs. It's
703 : * unclear that recursive entry would be safe, and it doesn't seem useful
704 : * even if it is safe, so let's block interrupts until done.
705 : */
706 230 : HOLD_INTERRUPTS();
707 :
708 : /*
709 : * Moreover, CurrentMemoryContext might be pointing almost anywhere. We
710 : * don't want to risk leaking data into long-lived contexts, so let's do
711 : * our work here in a private context that we can reset on each use.
712 : */
713 230 : if (hpm_context == NULL) /* first time through? */
714 1 : hpm_context = AllocSetContextCreate(TopMemoryContext,
715 : "HandleParallelMessages",
716 : ALLOCSET_DEFAULT_SIZES);
717 : else
718 229 : MemoryContextReset(hpm_context);
719 :
720 230 : oldcontext = MemoryContextSwitchTo(hpm_context);
721 :
722 : /* OK to process messages. Reset the flag saying there are more to do. */
723 230 : ParallelMessagePending = false;
724 :
725 459 : dlist_foreach(iter, &pcxt_list)
726 : {
727 : ParallelContext *pcxt;
728 : int i;
729 :
730 230 : pcxt = dlist_container(ParallelContext, node, iter.cur);
731 230 : if (pcxt->worker == NULL)
732 0 : continue;
733 :
734 1131 : for (i = 0; i < pcxt->nworkers_launched; ++i)
735 : {
736 : /*
737 : * Read as many messages as we can from each worker, but stop when
738 : * either (1) the worker's error queue goes away, which can happen
739 : * if we receive a Terminate message from the worker; or (2) no
740 : * more messages can be read from the worker without blocking.
741 : */
742 2033 : while (pcxt->worker[i].error_mqh != NULL)
743 : {
744 : shm_mq_result res;
745 : Size nbytes;
746 : void *data;
747 :
748 682 : res = shm_mq_receive(pcxt->worker[i].error_mqh, &nbytes,
749 : &data, true);
750 682 : if (res == SHM_MQ_WOULD_BLOCK)
751 452 : break;
752 230 : else if (res == SHM_MQ_SUCCESS)
753 : {
754 : StringInfoData msg;
755 :
756 230 : initStringInfo(&msg);
757 230 : appendBinaryStringInfo(&msg, data, nbytes);
758 230 : HandleParallelMessage(pcxt, i, &msg);
759 229 : pfree(msg.data);
760 : }
761 : else
762 0 : ereport(ERROR,
763 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
764 : errmsg("lost connection to parallel worker")));
765 : }
766 : }
767 : }
768 :
769 229 : MemoryContextSwitchTo(oldcontext);
770 :
771 : /* Might as well clear the context on our way out */
772 229 : MemoryContextReset(hpm_context);
773 :
774 229 : RESUME_INTERRUPTS();
775 229 : }
776 :
777 : /*
778 : * Handle a single protocol message received from a single parallel worker.
779 : */
780 : static void
781 230 : HandleParallelMessage(ParallelContext *pcxt, int i, StringInfo msg)
782 : {
783 : char msgtype;
784 :
785 230 : msgtype = pq_getmsgbyte(msg);
786 :
787 230 : switch (msgtype)
788 : {
789 : case 'K': /* BackendKeyData */
790 : {
791 115 : int32 pid = pq_getmsgint(msg, 4);
792 :
793 115 : (void) pq_getmsgint(msg, 4); /* discard cancel key */
794 115 : (void) pq_getmsgend(msg);
795 115 : pcxt->worker[i].pid = pid;
796 115 : break;
797 : }
798 :
799 : case 'E': /* ErrorResponse */
800 : case 'N': /* NoticeResponse */
801 : {
802 : ErrorData edata;
803 : ErrorContextCallback *save_error_context_stack;
804 :
805 : /* Parse ErrorResponse or NoticeResponse. */
806 1 : pq_parse_errornotice(msg, &edata);
807 :
808 : /* Death of a worker isn't enough justification for suicide. */
809 1 : edata.elevel = Min(edata.elevel, ERROR);
810 :
811 : /*
812 : * If desired, add a context line to show that this is a
813 : * message propagated from a parallel worker. Otherwise, it
814 : * can sometimes be confusing to understand what actually
815 : * happened. (We don't do this in FORCE_PARALLEL_REGRESS mode
816 : * because it causes test-result instability depending on
817 : * whether a parallel worker is actually used or not.)
818 : */
819 1 : if (force_parallel_mode != FORCE_PARALLEL_REGRESS)
820 : {
821 1 : if (edata.context)
822 0 : edata.context = psprintf("%s\n%s", edata.context,
823 : _("parallel worker"));
824 : else
825 1 : edata.context = pstrdup(_("parallel worker"));
826 : }
827 :
828 : /*
829 : * Context beyond that should use the error context callbacks
830 : * that were in effect when the ParallelContext was created,
831 : * not the current ones.
832 : */
833 1 : save_error_context_stack = error_context_stack;
834 1 : error_context_stack = pcxt->error_context_stack;
835 :
836 : /* Rethrow error or print notice. */
837 1 : ThrowErrorData(&edata);
838 :
839 : /* Not an error, so restore previous context stack. */
840 0 : error_context_stack = save_error_context_stack;
841 :
842 0 : break;
843 : }
844 :
845 : case 'A': /* NotifyResponse */
846 : {
847 : /* Propagate NotifyResponse. */
848 : int32 pid;
849 : const char *channel;
850 : const char *payload;
851 :
852 0 : pid = pq_getmsgint(msg, 4);
853 0 : channel = pq_getmsgrawstring(msg);
854 0 : payload = pq_getmsgrawstring(msg);
855 0 : pq_endmessage(msg);
856 :
857 0 : NotifyMyFrontEnd(channel, payload, pid);
858 :
859 0 : break;
860 : }
861 :
862 : case 'X': /* Terminate, indicating clean exit */
863 : {
864 114 : shm_mq_detach(pcxt->worker[i].error_mqh);
865 114 : pcxt->worker[i].error_mqh = NULL;
866 114 : break;
867 : }
868 :
869 : default:
870 : {
871 0 : elog(ERROR, "unrecognized message type received from parallel worker: %c (message length %d bytes)",
872 : msgtype, msg->len);
873 : }
874 : }
875 229 : }
876 :
877 : /*
878 : * End-of-subtransaction cleanup for parallel contexts.
879 : *
880 : * Currently, it's forbidden to enter or leave a subtransaction while
881 : * parallel mode is in effect, so we could just blow away everything. But
882 : * we may want to relax that restriction in the future, so this code
883 : * contemplates that there may be multiple subtransaction IDs in pcxt_list.
884 : */
885 : void
886 1 : AtEOSubXact_Parallel(bool isCommit, SubTransactionId mySubId)
887 : {
888 3 : while (!dlist_is_empty(&pcxt_list))
889 : {
890 : ParallelContext *pcxt;
891 :
892 1 : pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
893 1 : if (pcxt->subid != mySubId)
894 0 : break;
895 1 : if (isCommit)
896 0 : elog(WARNING, "leaked parallel context");
897 1 : DestroyParallelContext(pcxt);
898 : }
899 1 : }
900 :
901 : /*
902 : * End-of-transaction cleanup for parallel contexts.
903 : */
904 : void
905 115 : AtEOXact_Parallel(bool isCommit)
906 : {
907 230 : while (!dlist_is_empty(&pcxt_list))
908 : {
909 : ParallelContext *pcxt;
910 :
911 0 : pcxt = dlist_head_element(ParallelContext, node, &pcxt_list);
912 0 : if (isCommit)
913 0 : elog(WARNING, "leaked parallel context");
914 0 : DestroyParallelContext(pcxt);
915 : }
916 115 : }
917 :
918 : /*
919 : * Main entrypoint for parallel workers.
920 : */
921 : void
922 115 : ParallelWorkerMain(Datum main_arg)
923 : {
924 : dsm_segment *seg;
925 : shm_toc *toc;
926 : FixedParallelState *fps;
927 : char *error_queue_space;
928 : shm_mq *mq;
929 : shm_mq_handle *mqh;
930 : char *libraryspace;
931 : char *entrypointstate;
932 : char *library_name;
933 : char *function_name;
934 : parallel_worker_main_type entrypt;
935 : char *gucspace;
936 : char *combocidspace;
937 : char *tsnapspace;
938 : char *asnapspace;
939 : char *tstatespace;
940 : StringInfoData msgbuf;
941 :
942 : /* Set flag to indicate that we're initializing a parallel worker. */
943 115 : InitializingParallelWorker = true;
944 :
945 : /* Establish signal handlers. */
946 115 : pqsignal(SIGTERM, die);
947 115 : BackgroundWorkerUnblockSignals();
948 :
949 : /* Determine and set our parallel worker number. */
950 115 : Assert(ParallelWorkerNumber == -1);
951 115 : memcpy(&ParallelWorkerNumber, MyBgworkerEntry->bgw_extra, sizeof(int));
952 :
953 : /* Set up a memory context and resource owner. */
954 115 : Assert(CurrentResourceOwner == NULL);
955 115 : CurrentResourceOwner = ResourceOwnerCreate(NULL, "parallel toplevel");
956 115 : CurrentMemoryContext = AllocSetContextCreate(TopMemoryContext,
957 : "Parallel worker",
958 : ALLOCSET_DEFAULT_SIZES);
959 :
960 : /*
961 : * Now that we have a resource owner, we can attach to the dynamic shared
962 : * memory segment and read the table of contents.
963 : */
964 115 : seg = dsm_attach(DatumGetUInt32(main_arg));
965 115 : if (seg == NULL)
966 0 : ereport(ERROR,
967 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
968 : errmsg("could not map dynamic shared memory segment")));
969 115 : toc = shm_toc_attach(PARALLEL_MAGIC, dsm_segment_address(seg));
970 115 : if (toc == NULL)
971 0 : ereport(ERROR,
972 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
973 : errmsg("invalid magic number in dynamic shared memory segment")));
974 :
975 : /* Look up fixed parallel state. */
976 115 : fps = shm_toc_lookup(toc, PARALLEL_KEY_FIXED, false);
977 115 : MyFixedParallelState = fps;
978 :
979 : /*
980 : * Now that we have a worker number, we can find and attach to the error
981 : * queue provided for us. That's good, because until we do that, any
982 : * errors that happen here will not be reported back to the process that
983 : * requested that this worker be launched.
984 : */
985 115 : error_queue_space = shm_toc_lookup(toc, PARALLEL_KEY_ERROR_QUEUE, false);
986 115 : mq = (shm_mq *) (error_queue_space +
987 115 : ParallelWorkerNumber * PARALLEL_ERROR_QUEUE_SIZE);
988 115 : shm_mq_set_sender(mq, MyProc);
989 115 : mqh = shm_mq_attach(mq, seg, NULL);
990 115 : pq_redirect_to_shm_mq(seg, mqh);
991 115 : pq_set_parallel_master(fps->parallel_master_pid,
992 : fps->parallel_master_backend_id);
993 :
994 : /*
995 : * Send a BackendKeyData message to the process that initiated parallelism
996 : * so that it has access to our PID before it receives any other messages
997 : * from us. Our cancel key is sent, too, since that's the way the
998 : * protocol message is defined, but it won't actually be used for anything
999 : * in this case.
1000 : */
1001 115 : pq_beginmessage(&msgbuf, 'K');
1002 115 : pq_sendint(&msgbuf, (int32) MyProcPid, sizeof(int32));
1003 115 : pq_sendint(&msgbuf, (int32) MyCancelKey, sizeof(int32));
1004 115 : pq_endmessage(&msgbuf);
1005 :
1006 : /*
1007 : * Hooray! Primary initialization is complete. Now, we need to set up our
1008 : * backend-local state to match the original backend.
1009 : */
1010 :
1011 : /*
1012 : * Join locking group. We must do this before anything that could try to
1013 : * acquire a heavyweight lock, because any heavyweight locks acquired to
1014 : * this point could block either directly against the parallel group
1015 : * leader or against some process which in turn waits for a lock that
1016 : * conflicts with the parallel group leader, causing an undetected
1017 : * deadlock. (If we can't join the lock group, the leader has gone away,
1018 : * so just exit quietly.)
1019 : */
1020 115 : if (!BecomeLockGroupMember(fps->parallel_master_pgproc,
1021 : fps->parallel_master_pid))
1022 114 : return;
1023 :
1024 : /*
1025 : * Load libraries that were loaded by original backend. We want to do
1026 : * this before restoring GUCs, because the libraries might define custom
1027 : * variables.
1028 : */
1029 115 : libraryspace = shm_toc_lookup(toc, PARALLEL_KEY_LIBRARY, false);
1030 115 : RestoreLibraryState(libraryspace);
1031 :
1032 : /*
1033 : * Identify the entry point to be called. In theory this could result in
1034 : * loading an additional library, though most likely the entry point is in
1035 : * the core backend or in a library we just loaded.
1036 : */
1037 115 : entrypointstate = shm_toc_lookup(toc, PARALLEL_KEY_ENTRYPOINT, false);
1038 115 : library_name = entrypointstate;
1039 115 : function_name = entrypointstate + strlen(library_name) + 1;
1040 :
1041 115 : entrypt = LookupParallelWorkerFunction(library_name, function_name);
1042 :
1043 : /* Restore database connection. */
1044 115 : BackgroundWorkerInitializeConnectionByOid(fps->database_id,
1045 : fps->authenticated_user_id);
1046 :
1047 : /*
1048 : * Set the client encoding to the database encoding, since that is what
1049 : * the leader will expect.
1050 : */
1051 115 : SetClientEncoding(GetDatabaseEncoding());
1052 :
1053 : /* Restore GUC values from launching backend. */
1054 115 : gucspace = shm_toc_lookup(toc, PARALLEL_KEY_GUC, false);
1055 115 : StartTransactionCommand();
1056 115 : RestoreGUCState(gucspace);
1057 115 : CommitTransactionCommand();
1058 :
1059 : /* Crank up a transaction state appropriate to a parallel worker. */
1060 115 : tstatespace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_STATE, false);
1061 115 : StartParallelWorkerTransaction(tstatespace);
1062 :
1063 : /* Restore combo CID state. */
1064 115 : combocidspace = shm_toc_lookup(toc, PARALLEL_KEY_COMBO_CID, false);
1065 115 : RestoreComboCIDState(combocidspace);
1066 :
1067 : /* Restore transaction snapshot. */
1068 115 : tsnapspace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT, false);
1069 115 : RestoreTransactionSnapshot(RestoreSnapshot(tsnapspace),
1070 115 : fps->parallel_master_pgproc);
1071 :
1072 : /* Restore active snapshot. */
1073 115 : asnapspace = shm_toc_lookup(toc, PARALLEL_KEY_ACTIVE_SNAPSHOT, false);
1074 115 : PushActiveSnapshot(RestoreSnapshot(asnapspace));
1075 :
1076 : /*
1077 : * We've changed which tuples we can see, and must therefore invalidate
1078 : * system caches.
1079 : */
1080 115 : InvalidateSystemCaches();
1081 :
1082 : /* Restore user ID and security context. */
1083 115 : SetUserIdAndSecContext(fps->current_user_id, fps->sec_context);
1084 :
1085 : /* Restore temp-namespace state to ensure search path matches leader's. */
1086 115 : SetTempNamespaceState(fps->temp_namespace_id,
1087 : fps->temp_toast_namespace_id);
1088 :
1089 : /* Set ParallelMasterBackendId so we know how to address temp relations. */
1090 115 : ParallelMasterBackendId = fps->parallel_master_backend_id;
1091 :
1092 : /*
1093 : * We've initialized all of our state now; nothing should change
1094 : * hereafter.
1095 : */
1096 115 : InitializingParallelWorker = false;
1097 115 : EnterParallelMode();
1098 :
1099 : /*
1100 : * Time to do the real work: invoke the caller-supplied code.
1101 : */
1102 115 : entrypt(seg, toc);
1103 :
1104 : /* Must exit parallel mode to pop active snapshot. */
1105 114 : ExitParallelMode();
1106 :
1107 : /* Must pop active snapshot so resowner.c doesn't complain. */
1108 114 : PopActiveSnapshot();
1109 :
1110 : /* Shut down the parallel-worker transaction. */
1111 114 : EndParallelWorkerTransaction();
1112 :
1113 : /* Report success. */
1114 114 : pq_putmessage('X', NULL, 0);
1115 : }
1116 :
1117 : /*
1118 : * Update shared memory with the ending location of the last WAL record we
1119 : * wrote, if it's greater than the value already stored there.
1120 : */
1121 : void
1122 114 : ParallelWorkerReportLastRecEnd(XLogRecPtr last_xlog_end)
1123 : {
1124 114 : FixedParallelState *fps = MyFixedParallelState;
1125 :
1126 114 : Assert(fps != NULL);
1127 114 : SpinLockAcquire(&fps->mutex);
1128 114 : if (fps->last_xlog_end < last_xlog_end)
1129 0 : fps->last_xlog_end = last_xlog_end;
1130 114 : SpinLockRelease(&fps->mutex);
1131 114 : }
1132 :
1133 : /*
1134 : * Look up (and possibly load) a parallel worker entry point function.
1135 : *
1136 : * For functions contained in the core code, we use library name "postgres"
1137 : * and consult the InternalParallelWorkers array. External functions are
1138 : * looked up, and loaded if necessary, using load_external_function().
1139 : *
1140 : * The point of this is to pass function names as strings across process
1141 : * boundaries. We can't pass actual function addresses because of the
1142 : * possibility that the function has been loaded at a different address
1143 : * in a different process. This is obviously a hazard for functions in
1144 : * loadable libraries, but it can happen even for functions in the core code
1145 : * on platforms using EXEC_BACKEND (e.g., Windows).
1146 : *
1147 : * At some point it might be worthwhile to get rid of InternalParallelWorkers[]
1148 : * in favor of applying load_external_function() for core functions too;
1149 : * but that raises portability issues that are not worth addressing now.
1150 : */
1151 : static parallel_worker_main_type
1152 115 : LookupParallelWorkerFunction(const char *libraryname, const char *funcname)
1153 : {
1154 : /*
1155 : * If the function is to be loaded from postgres itself, search the
1156 : * InternalParallelWorkers array.
1157 : */
1158 115 : if (strcmp(libraryname, "postgres") == 0)
1159 : {
1160 : int i;
1161 :
1162 115 : for (i = 0; i < lengthof(InternalParallelWorkers); i++)
1163 : {
1164 115 : if (strcmp(InternalParallelWorkers[i].fn_name, funcname) == 0)
1165 115 : return InternalParallelWorkers[i].fn_addr;
1166 : }
1167 :
1168 : /* We can only reach this by programming error. */
1169 0 : elog(ERROR, "internal function \"%s\" not found", funcname);
1170 : }
1171 :
1172 : /* Otherwise load from external library. */
1173 0 : return (parallel_worker_main_type)
1174 : load_external_function(libraryname, funcname, true, NULL);
1175 : }
|
