141c973f3aa09693f4e94c4a6f3253163469727d
[ghc.git] / rts / Schedule.c
1 /* ---------------------------------------------------------------------------
2 *
3 * (c) The GHC Team, 1998-2006
4 *
5 * The scheduler and thread-related functionality
6 *
7 * --------------------------------------------------------------------------*/
8
9 #include "PosixSource.h"
10 #define KEEP_LOCKCLOSURE
11 #include "Rts.h"
12 #include "SchedAPI.h"
13 #include "RtsUtils.h"
14 #include "RtsFlags.h"
15 #include "OSThreads.h"
16 #include "Storage.h"
17 #include "StgRun.h"
18 #include "Hooks.h"
19 #include "Schedule.h"
20 #include "StgMiscClosures.h"
21 #include "Interpreter.h"
22 #include "Printer.h"
23 #include "RtsSignals.h"
24 #include "Sanity.h"
25 #include "Stats.h"
26 #include "STM.h"
27 #include "Timer.h"
28 #include "Prelude.h"
29 #include "ThreadLabels.h"
30 #include "LdvProfile.h"
31 #include "Updates.h"
32 #include "Proftimer.h"
33 #include "ProfHeap.h"
34 #include "GC.h"
35 #include "Weak.h"
36 #include "EventLog.h"
37
38 /* PARALLEL_HASKELL includes go here */
39
40 #include "Sparks.h"
41 #include "Capability.h"
42 #include "Task.h"
43 #include "AwaitEvent.h"
44 #if defined(mingw32_HOST_OS)
45 #include "win32/IOManager.h"
46 #endif
47 #include "Trace.h"
48 #include "RaiseAsync.h"
49 #include "Threads.h"
50 #include "ThrIOManager.h"
51
52 #ifdef HAVE_SYS_TYPES_H
53 #include <sys/types.h>
54 #endif
55 #ifdef HAVE_UNISTD_H
56 #include <unistd.h>
57 #endif
58
59 #include <string.h>
60 #include <stdlib.h>
61 #include <stdarg.h>
62
63 #ifdef HAVE_ERRNO_H
64 #include <errno.h>
65 #endif
66
67 // Turn off inlining when debugging - it obfuscates things
68 #ifdef DEBUG
69 # undef STATIC_INLINE
70 # define STATIC_INLINE static
71 #endif
72
73 /* -----------------------------------------------------------------------------
74 * Global variables
75 * -------------------------------------------------------------------------- */
76
77 #if !defined(THREADED_RTS)
78 // Blocked/sleeping thrads
79 StgTSO *blocked_queue_hd = NULL;
80 StgTSO *blocked_queue_tl = NULL;
81 StgTSO *sleeping_queue = NULL; // perhaps replace with a hash table?
82 #endif
83
84 /* Threads blocked on blackholes.
85 * LOCK: sched_mutex+capability, or all capabilities
86 */
87 StgTSO *blackhole_queue = NULL;
88
89 /* The blackhole_queue should be checked for threads to wake up. See
90 * Schedule.h for more thorough comment.
91 * LOCK: none (doesn't matter if we miss an update)
92 */
93 rtsBool blackholes_need_checking = rtsFalse;
94
95 /* Set to true when the latest garbage collection failed to reclaim
96 * enough space, and the runtime should proceed to shut itself down in
97 * an orderly fashion (emitting profiling info etc.)
98 */
99 rtsBool heap_overflow = rtsFalse;
100
101 /* flag that tracks whether we have done any execution in this time slice.
102 * LOCK: currently none, perhaps we should lock (but needs to be
103 * updated in the fast path of the scheduler).
104 *
105 * NB. must be StgWord, we do xchg() on it.
106 */
107 volatile StgWord recent_activity = ACTIVITY_YES;
108
109 /* if this flag is set as well, give up execution
110 * LOCK: none (changes monotonically)
111 */
112 volatile StgWord sched_state = SCHED_RUNNING;
113
114 /* This is used in `TSO.h' and gcc 2.96 insists that this variable actually
115 * exists - earlier gccs apparently didn't.
116 * -= chak
117 */
118 StgTSO dummy_tso;
119
120 /*
121 * Set to TRUE when entering a shutdown state (via shutdownHaskellAndExit()) --
122 * in an MT setting, needed to signal that a worker thread shouldn't hang around
123 * in the scheduler when it is out of work.
124 */
125 rtsBool shutting_down_scheduler = rtsFalse;
126
127 /*
128 * This mutex protects most of the global scheduler data in
129 * the THREADED_RTS runtime.
130 */
131 #if defined(THREADED_RTS)
132 Mutex sched_mutex;
133 #endif
134
135 #if !defined(mingw32_HOST_OS)
136 #define FORKPROCESS_PRIMOP_SUPPORTED
137 #endif
138
139 /* -----------------------------------------------------------------------------
140 * static function prototypes
141 * -------------------------------------------------------------------------- */
142
143 static Capability *schedule (Capability *initialCapability, Task *task);
144
145 //
146 // These function all encapsulate parts of the scheduler loop, and are
147 // abstracted only to make the structure and control flow of the
148 // scheduler clearer.
149 //
150 static void schedulePreLoop (void);
151 static void scheduleFindWork (Capability *cap);
152 #if defined(THREADED_RTS)
153 static void scheduleYield (Capability **pcap, Task *task);
154 #endif
155 static void scheduleStartSignalHandlers (Capability *cap);
156 static void scheduleCheckBlockedThreads (Capability *cap);
157 static void scheduleCheckWakeupThreads(Capability *cap USED_IF_NOT_THREADS);
158 static void scheduleCheckBlackHoles (Capability *cap);
159 static void scheduleDetectDeadlock (Capability *cap, Task *task);
160 static void schedulePushWork(Capability *cap, Task *task);
161 #if defined(PARALLEL_HASKELL)
162 static rtsBool scheduleGetRemoteWork(Capability *cap);
163 static void scheduleSendPendingMessages(void);
164 #endif
165 #if defined(PARALLEL_HASKELL) || defined(THREADED_RTS)
166 static void scheduleActivateSpark(Capability *cap);
167 #endif
168 static void schedulePostRunThread(Capability *cap, StgTSO *t);
169 static rtsBool scheduleHandleHeapOverflow( Capability *cap, StgTSO *t );
170 static void scheduleHandleStackOverflow( Capability *cap, Task *task,
171 StgTSO *t);
172 static rtsBool scheduleHandleYield( Capability *cap, StgTSO *t,
173 nat prev_what_next );
174 static void scheduleHandleThreadBlocked( StgTSO *t );
175 static rtsBool scheduleHandleThreadFinished( Capability *cap, Task *task,
176 StgTSO *t );
177 static rtsBool scheduleNeedHeapProfile(rtsBool ready_to_gc);
178 static Capability *scheduleDoGC(Capability *cap, Task *task,
179 rtsBool force_major);
180
181 static rtsBool checkBlackHoles(Capability *cap);
182
183 static StgTSO *threadStackOverflow(Capability *cap, StgTSO *tso);
184 static StgTSO *threadStackUnderflow(Task *task, StgTSO *tso);
185
186 static void deleteThread (Capability *cap, StgTSO *tso);
187 static void deleteAllThreads (Capability *cap);
188
189 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
190 static void deleteThread_(Capability *cap, StgTSO *tso);
191 #endif
192
193 #ifdef DEBUG
194 static char *whatNext_strs[] = {
195 "(unknown)",
196 "ThreadRunGHC",
197 "ThreadInterpret",
198 "ThreadKilled",
199 "ThreadRelocated",
200 "ThreadComplete"
201 };
202 #endif
203
204 /* -----------------------------------------------------------------------------
205 * Putting a thread on the run queue: different scheduling policies
206 * -------------------------------------------------------------------------- */
207
208 STATIC_INLINE void
209 addToRunQueue( Capability *cap, StgTSO *t )
210 {
211 #if defined(PARALLEL_HASKELL)
212 if (RtsFlags.ParFlags.doFairScheduling) {
213 // this does round-robin scheduling; good for concurrency
214 appendToRunQueue(cap,t);
215 } else {
216 // this does unfair scheduling; good for parallelism
217 pushOnRunQueue(cap,t);
218 }
219 #else
220 // this does round-robin scheduling; good for concurrency
221 appendToRunQueue(cap,t);
222 #endif
223 }
224
225 /* ---------------------------------------------------------------------------
226 Main scheduling loop.
227
228 We use round-robin scheduling, each thread returning to the
229 scheduler loop when one of these conditions is detected:
230
231 * out of heap space
232 * timer expires (thread yields)
233 * thread blocks
234 * thread ends
235 * stack overflow
236
237 GRAN version:
238 In a GranSim setup this loop iterates over the global event queue.
239 This revolves around the global event queue, which determines what
240 to do next. Therefore, it's more complicated than either the
241 concurrent or the parallel (GUM) setup.
242 This version has been entirely removed (JB 2008/08).
243
244 GUM version:
245 GUM iterates over incoming messages.
246 It starts with nothing to do (thus CurrentTSO == END_TSO_QUEUE),
247 and sends out a fish whenever it has nothing to do; in-between
248 doing the actual reductions (shared code below) it processes the
249 incoming messages and deals with delayed operations
250 (see PendingFetches).
251 This is not the ugliest code you could imagine, but it's bloody close.
252
253 (JB 2008/08) This version was formerly indicated by a PP-Flag PAR,
254 now by PP-flag PARALLEL_HASKELL. The Eden RTS (in GHC-6.x) uses it,
255 as well as future GUM versions. This file has been refurbished to
256 only contain valid code, which is however incomplete, refers to
257 invalid includes etc.
258
259 ------------------------------------------------------------------------ */
260
261 static Capability *
262 schedule (Capability *initialCapability, Task *task)
263 {
264 StgTSO *t;
265 Capability *cap;
266 StgThreadReturnCode ret;
267 #if defined(PARALLEL_HASKELL)
268 rtsBool receivedFinish = rtsFalse;
269 #endif
270 nat prev_what_next;
271 rtsBool ready_to_gc;
272 #if defined(THREADED_RTS)
273 rtsBool first = rtsTrue;
274 #endif
275
276 cap = initialCapability;
277
278 // Pre-condition: this task owns initialCapability.
279 // The sched_mutex is *NOT* held
280 // NB. on return, we still hold a capability.
281
282 debugTrace (DEBUG_sched,
283 "### NEW SCHEDULER LOOP (task: %p, cap: %p)",
284 task, initialCapability);
285
286 if (running_finalizers) {
287 errorBelch("error: a C finalizer called back into Haskell.\n"
288 " use Foreign.Concurrent.newForeignPtr for Haskell finalizers.");
289 stg_exit(EXIT_FAILURE);
290 }
291
292 schedulePreLoop();
293
294 // -----------------------------------------------------------
295 // Scheduler loop starts here:
296
297 #if defined(PARALLEL_HASKELL)
298 #define TERMINATION_CONDITION (!receivedFinish)
299 #else
300 #define TERMINATION_CONDITION rtsTrue
301 #endif
302
303 while (TERMINATION_CONDITION) {
304
305 // Check whether we have re-entered the RTS from Haskell without
306 // going via suspendThread()/resumeThread (i.e. a 'safe' foreign
307 // call).
308 if (cap->in_haskell) {
309 errorBelch("schedule: re-entered unsafely.\n"
310 " Perhaps a 'foreign import unsafe' should be 'safe'?");
311 stg_exit(EXIT_FAILURE);
312 }
313
314 // The interruption / shutdown sequence.
315 //
316 // In order to cleanly shut down the runtime, we want to:
317 // * make sure that all main threads return to their callers
318 // with the state 'Interrupted'.
319 // * clean up all OS threads assocated with the runtime
320 // * free all memory etc.
321 //
322 // So the sequence for ^C goes like this:
323 //
324 // * ^C handler sets sched_state := SCHED_INTERRUPTING and
325 // arranges for some Capability to wake up
326 //
327 // * all threads in the system are halted, and the zombies are
328 // placed on the run queue for cleaning up. We acquire all
329 // the capabilities in order to delete the threads, this is
330 // done by scheduleDoGC() for convenience (because GC already
331 // needs to acquire all the capabilities). We can't kill
332 // threads involved in foreign calls.
333 //
334 // * somebody calls shutdownHaskell(), which calls exitScheduler()
335 //
336 // * sched_state := SCHED_SHUTTING_DOWN
337 //
338 // * all workers exit when the run queue on their capability
339 // drains. All main threads will also exit when their TSO
340 // reaches the head of the run queue and they can return.
341 //
342 // * eventually all Capabilities will shut down, and the RTS can
343 // exit.
344 //
345 // * We might be left with threads blocked in foreign calls,
346 // we should really attempt to kill these somehow (TODO);
347
348 switch (sched_state) {
349 case SCHED_RUNNING:
350 break;
351 case SCHED_INTERRUPTING:
352 debugTrace(DEBUG_sched, "SCHED_INTERRUPTING");
353 #if defined(THREADED_RTS)
354 discardSparksCap(cap);
355 #endif
356 /* scheduleDoGC() deletes all the threads */
357 cap = scheduleDoGC(cap,task,rtsFalse);
358
359 // after scheduleDoGC(), we must be shutting down. Either some
360 // other Capability did the final GC, or we did it above,
361 // either way we can fall through to the SCHED_SHUTTING_DOWN
362 // case now.
363 ASSERT(sched_state == SCHED_SHUTTING_DOWN);
364 // fall through
365
366 case SCHED_SHUTTING_DOWN:
367 debugTrace(DEBUG_sched, "SCHED_SHUTTING_DOWN");
368 // If we are a worker, just exit. If we're a bound thread
369 // then we will exit below when we've removed our TSO from
370 // the run queue.
371 if (task->tso == NULL && emptyRunQueue(cap)) {
372 return cap;
373 }
374 break;
375 default:
376 barf("sched_state: %d", sched_state);
377 }
378
379 scheduleFindWork(cap);
380
381 /* work pushing, currently relevant only for THREADED_RTS:
382 (pushes threads, wakes up idle capabilities for stealing) */
383 schedulePushWork(cap,task);
384
385 #if defined(PARALLEL_HASKELL)
386 /* since we perform a blocking receive and continue otherwise,
387 either we never reach here or we definitely have work! */
388 // from here: non-empty run queue
389 ASSERT(!emptyRunQueue(cap));
390
391 if (PacketsWaiting()) { /* now process incoming messages, if any
392 pending...
393
394 CAUTION: scheduleGetRemoteWork called
395 above, waits for messages as well! */
396 processMessages(cap, &receivedFinish);
397 }
398 #endif // PARALLEL_HASKELL: non-empty run queue!
399
400 scheduleDetectDeadlock(cap,task);
401
402 #if defined(THREADED_RTS)
403 cap = task->cap; // reload cap, it might have changed
404 #endif
405
406 // Normally, the only way we can get here with no threads to
407 // run is if a keyboard interrupt received during
408 // scheduleCheckBlockedThreads() or scheduleDetectDeadlock().
409 // Additionally, it is not fatal for the
410 // threaded RTS to reach here with no threads to run.
411 //
412 // win32: might be here due to awaitEvent() being abandoned
413 // as a result of a console event having been delivered.
414
415 #if defined(THREADED_RTS)
416 if (first)
417 {
418 // XXX: ToDo
419 // // don't yield the first time, we want a chance to run this
420 // // thread for a bit, even if there are others banging at the
421 // // door.
422 // first = rtsFalse;
423 // ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
424 }
425
426 yield:
427 scheduleYield(&cap,task);
428 if (emptyRunQueue(cap)) continue; // look for work again
429 #endif
430
431 #if !defined(THREADED_RTS) && !defined(mingw32_HOST_OS)
432 if ( emptyRunQueue(cap) ) {
433 ASSERT(sched_state >= SCHED_INTERRUPTING);
434 }
435 #endif
436
437 //
438 // Get a thread to run
439 //
440 t = popRunQueue(cap);
441
442 // Sanity check the thread we're about to run. This can be
443 // expensive if there is lots of thread switching going on...
444 IF_DEBUG(sanity,checkTSO(t));
445
446 #if defined(THREADED_RTS)
447 // Check whether we can run this thread in the current task.
448 // If not, we have to pass our capability to the right task.
449 {
450 Task *bound = t->bound;
451
452 if (bound) {
453 if (bound == task) {
454 debugTrace(DEBUG_sched,
455 "### Running thread %lu in bound thread", (unsigned long)t->id);
456 // yes, the Haskell thread is bound to the current native thread
457 } else {
458 debugTrace(DEBUG_sched,
459 "### thread %lu bound to another OS thread", (unsigned long)t->id);
460 // no, bound to a different Haskell thread: pass to that thread
461 pushOnRunQueue(cap,t);
462 continue;
463 }
464 } else {
465 // The thread we want to run is unbound.
466 if (task->tso) {
467 debugTrace(DEBUG_sched,
468 "### this OS thread cannot run thread %lu", (unsigned long)t->id);
469 // no, the current native thread is bound to a different
470 // Haskell thread, so pass it to any worker thread
471 pushOnRunQueue(cap,t);
472 continue;
473 }
474 }
475 }
476 #endif
477
478 // If we're shutting down, and this thread has not yet been
479 // killed, kill it now. This sometimes happens when a finalizer
480 // thread is created by the final GC, or a thread previously
481 // in a foreign call returns.
482 if (sched_state >= SCHED_INTERRUPTING &&
483 !(t->what_next == ThreadComplete || t->what_next == ThreadKilled)) {
484 deleteThread(cap,t);
485 }
486
487 /* context switches are initiated by the timer signal, unless
488 * the user specified "context switch as often as possible", with
489 * +RTS -C0
490 */
491 if (RtsFlags.ConcFlags.ctxtSwitchTicks == 0
492 && !emptyThreadQueues(cap)) {
493 cap->context_switch = 1;
494 }
495
496 run_thread:
497
498 // CurrentTSO is the thread to run. t might be different if we
499 // loop back to run_thread, so make sure to set CurrentTSO after
500 // that.
501 cap->r.rCurrentTSO = t;
502
503 debugTrace(DEBUG_sched, "-->> running thread %ld %s ...",
504 (long)t->id, whatNext_strs[t->what_next]);
505
506 startHeapProfTimer();
507
508 // Check for exceptions blocked on this thread
509 maybePerformBlockedException (cap, t);
510
511 // ----------------------------------------------------------------------
512 // Run the current thread
513
514 ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
515 ASSERT(t->cap == cap);
516 ASSERT(t->bound ? t->bound->cap == cap : 1);
517
518 prev_what_next = t->what_next;
519
520 errno = t->saved_errno;
521 #if mingw32_HOST_OS
522 SetLastError(t->saved_winerror);
523 #endif
524
525 cap->in_haskell = rtsTrue;
526
527 dirty_TSO(cap,t);
528
529 #if defined(THREADED_RTS)
530 if (recent_activity == ACTIVITY_DONE_GC) {
531 // ACTIVITY_DONE_GC means we turned off the timer signal to
532 // conserve power (see #1623). Re-enable it here.
533 nat prev;
534 prev = xchg((P_)&recent_activity, ACTIVITY_YES);
535 if (prev == ACTIVITY_DONE_GC) {
536 startTimer();
537 }
538 } else {
539 recent_activity = ACTIVITY_YES;
540 }
541 #endif
542
543 postEvent(cap, EVENT_RUN_THREAD, t->id, 0);
544
545 switch (prev_what_next) {
546
547 case ThreadKilled:
548 case ThreadComplete:
549 /* Thread already finished, return to scheduler. */
550 ret = ThreadFinished;
551 break;
552
553 case ThreadRunGHC:
554 {
555 StgRegTable *r;
556 r = StgRun((StgFunPtr) stg_returnToStackTop, &cap->r);
557 cap = regTableToCapability(r);
558 ret = r->rRet;
559 break;
560 }
561
562 case ThreadInterpret:
563 cap = interpretBCO(cap);
564 ret = cap->r.rRet;
565 break;
566
567 default:
568 barf("schedule: invalid what_next field");
569 }
570
571 cap->in_haskell = rtsFalse;
572
573 // The TSO might have moved, eg. if it re-entered the RTS and a GC
574 // happened. So find the new location:
575 t = cap->r.rCurrentTSO;
576
577 // We have run some Haskell code: there might be blackhole-blocked
578 // threads to wake up now.
579 // Lock-free test here should be ok, we're just setting a flag.
580 if ( blackhole_queue != END_TSO_QUEUE ) {
581 blackholes_need_checking = rtsTrue;
582 }
583
584 // And save the current errno in this thread.
585 // XXX: possibly bogus for SMP because this thread might already
586 // be running again, see code below.
587 t->saved_errno = errno;
588 #if mingw32_HOST_OS
589 // Similarly for Windows error code
590 t->saved_winerror = GetLastError();
591 #endif
592
593 postEvent (cap, EVENT_STOP_THREAD, t->id, ret);
594
595 #if defined(THREADED_RTS)
596 // If ret is ThreadBlocked, and this Task is bound to the TSO that
597 // blocked, we are in limbo - the TSO is now owned by whatever it
598 // is blocked on, and may in fact already have been woken up,
599 // perhaps even on a different Capability. It may be the case
600 // that task->cap != cap. We better yield this Capability
601 // immediately and return to normaility.
602 if (ret == ThreadBlocked) {
603 debugTrace(DEBUG_sched,
604 "--<< thread %lu (%s) stopped: blocked",
605 (unsigned long)t->id, whatNext_strs[t->what_next]);
606 goto yield;
607 }
608 #endif
609
610 ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
611 ASSERT(t->cap == cap);
612
613 // ----------------------------------------------------------------------
614
615 // Costs for the scheduler are assigned to CCS_SYSTEM
616 stopHeapProfTimer();
617 #if defined(PROFILING)
618 CCCS = CCS_SYSTEM;
619 #endif
620
621 schedulePostRunThread(cap,t);
622
623 t = threadStackUnderflow(task,t);
624
625 ready_to_gc = rtsFalse;
626
627 switch (ret) {
628 case HeapOverflow:
629 ready_to_gc = scheduleHandleHeapOverflow(cap,t);
630 break;
631
632 case StackOverflow:
633 scheduleHandleStackOverflow(cap,task,t);
634 break;
635
636 case ThreadYielding:
637 if (scheduleHandleYield(cap, t, prev_what_next)) {
638 // shortcut for switching between compiler/interpreter:
639 goto run_thread;
640 }
641 break;
642
643 case ThreadBlocked:
644 scheduleHandleThreadBlocked(t);
645 break;
646
647 case ThreadFinished:
648 if (scheduleHandleThreadFinished(cap, task, t)) return cap;
649 ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
650 break;
651
652 default:
653 barf("schedule: invalid thread return code %d", (int)ret);
654 }
655
656 if (ready_to_gc || scheduleNeedHeapProfile(ready_to_gc)) {
657 cap = scheduleDoGC(cap,task,rtsFalse);
658 }
659 } /* end of while() */
660 }
661
662 /* ----------------------------------------------------------------------------
663 * Setting up the scheduler loop
664 * ------------------------------------------------------------------------- */
665
666 static void
667 schedulePreLoop(void)
668 {
669 // initialisation for scheduler - what cannot go into initScheduler()
670 }
671
672 /* -----------------------------------------------------------------------------
673 * scheduleFindWork()
674 *
675 * Search for work to do, and handle messages from elsewhere.
676 * -------------------------------------------------------------------------- */
677
678 static void
679 scheduleFindWork (Capability *cap)
680 {
681 scheduleStartSignalHandlers(cap);
682
683 // Only check the black holes here if we've nothing else to do.
684 // During normal execution, the black hole list only gets checked
685 // at GC time, to avoid repeatedly traversing this possibly long
686 // list each time around the scheduler.
687 if (emptyRunQueue(cap)) { scheduleCheckBlackHoles(cap); }
688
689 scheduleCheckWakeupThreads(cap);
690
691 scheduleCheckBlockedThreads(cap);
692
693 #if defined(THREADED_RTS) || defined(PARALLEL_HASKELL)
694 if (emptyRunQueue(cap)) { scheduleActivateSpark(cap); }
695 #endif
696
697 #if defined(PARALLEL_HASKELL)
698 // if messages have been buffered...
699 scheduleSendPendingMessages();
700 #endif
701
702 #if defined(PARALLEL_HASKELL)
703 if (emptyRunQueue(cap)) {
704 receivedFinish = scheduleGetRemoteWork(cap);
705 continue; // a new round, (hopefully) with new work
706 /*
707 in GUM, this a) sends out a FISH and returns IF no fish is
708 out already
709 b) (blocking) awaits and receives messages
710
711 in Eden, this is only the blocking receive, as b) in GUM.
712 */
713 }
714 #endif
715 }
716
717 #if defined(THREADED_RTS)
718 STATIC_INLINE rtsBool
719 shouldYieldCapability (Capability *cap, Task *task)
720 {
721 // we need to yield this capability to someone else if..
722 // - another thread is initiating a GC
723 // - another Task is returning from a foreign call
724 // - the thread at the head of the run queue cannot be run
725 // by this Task (it is bound to another Task, or it is unbound
726 // and this task it bound).
727 return (waiting_for_gc ||
728 cap->returning_tasks_hd != NULL ||
729 (!emptyRunQueue(cap) && (task->tso == NULL
730 ? cap->run_queue_hd->bound != NULL
731 : cap->run_queue_hd->bound != task)));
732 }
733
734 // This is the single place where a Task goes to sleep. There are
735 // two reasons it might need to sleep:
736 // - there are no threads to run
737 // - we need to yield this Capability to someone else
738 // (see shouldYieldCapability())
739 //
740 // Careful: the scheduler loop is quite delicate. Make sure you run
741 // the tests in testsuite/concurrent (all ways) after modifying this,
742 // and also check the benchmarks in nofib/parallel for regressions.
743
744 static void
745 scheduleYield (Capability **pcap, Task *task)
746 {
747 Capability *cap = *pcap;
748
749 // if we have work, and we don't need to give up the Capability, continue.
750 if (!shouldYieldCapability(cap,task) &&
751 (!emptyRunQueue(cap) ||
752 !emptyWakeupQueue(cap) ||
753 blackholes_need_checking ||
754 sched_state >= SCHED_INTERRUPTING))
755 return;
756
757 // otherwise yield (sleep), and keep yielding if necessary.
758 do {
759 yieldCapability(&cap,task);
760 }
761 while (shouldYieldCapability(cap,task));
762
763 // note there may still be no threads on the run queue at this
764 // point, the caller has to check.
765
766 *pcap = cap;
767 return;
768 }
769 #endif
770
771 /* -----------------------------------------------------------------------------
772 * schedulePushWork()
773 *
774 * Push work to other Capabilities if we have some.
775 * -------------------------------------------------------------------------- */
776
777 static void
778 schedulePushWork(Capability *cap USED_IF_THREADS,
779 Task *task USED_IF_THREADS)
780 {
781 /* following code not for PARALLEL_HASKELL. I kept the call general,
782 future GUM versions might use pushing in a distributed setup */
783 #if defined(THREADED_RTS)
784
785 Capability *free_caps[n_capabilities], *cap0;
786 nat i, n_free_caps;
787
788 // migration can be turned off with +RTS -qg
789 if (!RtsFlags.ParFlags.migrate) return;
790
791 // Check whether we have more threads on our run queue, or sparks
792 // in our pool, that we could hand to another Capability.
793 if (cap->run_queue_hd == END_TSO_QUEUE) {
794 if (sparkPoolSizeCap(cap) < 2) return;
795 } else {
796 if (cap->run_queue_hd->_link == END_TSO_QUEUE &&
797 sparkPoolSizeCap(cap) < 1) return;
798 }
799
800 // First grab as many free Capabilities as we can.
801 for (i=0, n_free_caps=0; i < n_capabilities; i++) {
802 cap0 = &capabilities[i];
803 if (cap != cap0 && tryGrabCapability(cap0,task)) {
804 if (!emptyRunQueue(cap0) || cap->returning_tasks_hd != NULL) {
805 // it already has some work, we just grabbed it at
806 // the wrong moment. Or maybe it's deadlocked!
807 releaseCapability(cap0);
808 } else {
809 free_caps[n_free_caps++] = cap0;
810 }
811 }
812 }
813
814 // we now have n_free_caps free capabilities stashed in
815 // free_caps[]. Share our run queue equally with them. This is
816 // probably the simplest thing we could do; improvements we might
817 // want to do include:
818 //
819 // - giving high priority to moving relatively new threads, on
820 // the gournds that they haven't had time to build up a
821 // working set in the cache on this CPU/Capability.
822 //
823 // - giving low priority to moving long-lived threads
824
825 if (n_free_caps > 0) {
826 StgTSO *prev, *t, *next;
827 rtsBool pushed_to_all;
828
829 debugTrace(DEBUG_sched,
830 "cap %d: %s and %d free capabilities, sharing...",
831 cap->no,
832 (!emptyRunQueue(cap) && cap->run_queue_hd->_link != END_TSO_QUEUE)?
833 "excess threads on run queue":"sparks to share (>=2)",
834 n_free_caps);
835
836 i = 0;
837 pushed_to_all = rtsFalse;
838
839 if (cap->run_queue_hd != END_TSO_QUEUE) {
840 prev = cap->run_queue_hd;
841 t = prev->_link;
842 prev->_link = END_TSO_QUEUE;
843 for (; t != END_TSO_QUEUE; t = next) {
844 next = t->_link;
845 t->_link = END_TSO_QUEUE;
846 if (t->what_next == ThreadRelocated
847 || t->bound == task // don't move my bound thread
848 || tsoLocked(t)) { // don't move a locked thread
849 setTSOLink(cap, prev, t);
850 prev = t;
851 } else if (i == n_free_caps) {
852 pushed_to_all = rtsTrue;
853 i = 0;
854 // keep one for us
855 setTSOLink(cap, prev, t);
856 prev = t;
857 } else {
858 debugTrace(DEBUG_sched, "pushing thread %lu to capability %d", (unsigned long)t->id, free_caps[i]->no);
859 appendToRunQueue(free_caps[i],t);
860
861 postEvent (cap, EVENT_MIGRATE_THREAD, t->id, free_caps[i]->no);
862
863 if (t->bound) { t->bound->cap = free_caps[i]; }
864 t->cap = free_caps[i];
865 i++;
866 }
867 }
868 cap->run_queue_tl = prev;
869 }
870
871 #ifdef SPARK_PUSHING
872 /* JB I left this code in place, it would work but is not necessary */
873
874 // If there are some free capabilities that we didn't push any
875 // threads to, then try to push a spark to each one.
876 if (!pushed_to_all) {
877 StgClosure *spark;
878 // i is the next free capability to push to
879 for (; i < n_free_caps; i++) {
880 if (emptySparkPoolCap(free_caps[i])) {
881 spark = tryStealSpark(cap->sparks);
882 if (spark != NULL) {
883 debugTrace(DEBUG_sched, "pushing spark %p to capability %d", spark, free_caps[i]->no);
884 newSpark(&(free_caps[i]->r), spark);
885 }
886 }
887 }
888 }
889 #endif /* SPARK_PUSHING */
890
891 // release the capabilities
892 for (i = 0; i < n_free_caps; i++) {
893 task->cap = free_caps[i];
894 releaseAndWakeupCapability(free_caps[i]);
895 }
896 }
897 task->cap = cap; // reset to point to our Capability.
898
899 #endif /* THREADED_RTS */
900
901 }
902
903 /* ----------------------------------------------------------------------------
904 * Start any pending signal handlers
905 * ------------------------------------------------------------------------- */
906
907 #if defined(RTS_USER_SIGNALS) && !defined(THREADED_RTS)
908 static void
909 scheduleStartSignalHandlers(Capability *cap)
910 {
911 if (RtsFlags.MiscFlags.install_signal_handlers && signals_pending()) {
912 // safe outside the lock
913 startSignalHandlers(cap);
914 }
915 }
916 #else
917 static void
918 scheduleStartSignalHandlers(Capability *cap STG_UNUSED)
919 {
920 }
921 #endif
922
923 /* ----------------------------------------------------------------------------
924 * Check for blocked threads that can be woken up.
925 * ------------------------------------------------------------------------- */
926
927 static void
928 scheduleCheckBlockedThreads(Capability *cap USED_IF_NOT_THREADS)
929 {
930 #if !defined(THREADED_RTS)
931 //
932 // Check whether any waiting threads need to be woken up. If the
933 // run queue is empty, and there are no other tasks running, we
934 // can wait indefinitely for something to happen.
935 //
936 if ( !emptyQueue(blocked_queue_hd) || !emptyQueue(sleeping_queue) )
937 {
938 awaitEvent( emptyRunQueue(cap) && !blackholes_need_checking );
939 }
940 #endif
941 }
942
943
944 /* ----------------------------------------------------------------------------
945 * Check for threads woken up by other Capabilities
946 * ------------------------------------------------------------------------- */
947
948 static void
949 scheduleCheckWakeupThreads(Capability *cap USED_IF_THREADS)
950 {
951 #if defined(THREADED_RTS)
952 // Any threads that were woken up by other Capabilities get
953 // appended to our run queue.
954 if (!emptyWakeupQueue(cap)) {
955 ACQUIRE_LOCK(&cap->lock);
956 if (emptyRunQueue(cap)) {
957 cap->run_queue_hd = cap->wakeup_queue_hd;
958 cap->run_queue_tl = cap->wakeup_queue_tl;
959 } else {
960 setTSOLink(cap, cap->run_queue_tl, cap->wakeup_queue_hd);
961 cap->run_queue_tl = cap->wakeup_queue_tl;
962 }
963 cap->wakeup_queue_hd = cap->wakeup_queue_tl = END_TSO_QUEUE;
964 RELEASE_LOCK(&cap->lock);
965 }
966 #endif
967 }
968
969 /* ----------------------------------------------------------------------------
970 * Check for threads blocked on BLACKHOLEs that can be woken up
971 * ------------------------------------------------------------------------- */
972 static void
973 scheduleCheckBlackHoles (Capability *cap)
974 {
975 if ( blackholes_need_checking ) // check without the lock first
976 {
977 ACQUIRE_LOCK(&sched_mutex);
978 if ( blackholes_need_checking ) {
979 blackholes_need_checking = rtsFalse;
980 // important that we reset the flag *before* checking the
981 // blackhole queue, otherwise we could get deadlock. This
982 // happens as follows: we wake up a thread that
983 // immediately runs on another Capability, blocks on a
984 // blackhole, and then we reset the blackholes_need_checking flag.
985 checkBlackHoles(cap);
986 }
987 RELEASE_LOCK(&sched_mutex);
988 }
989 }
990
991 /* ----------------------------------------------------------------------------
992 * Detect deadlock conditions and attempt to resolve them.
993 * ------------------------------------------------------------------------- */
994
995 static void
996 scheduleDetectDeadlock (Capability *cap, Task *task)
997 {
998
999 #if defined(PARALLEL_HASKELL)
1000 // ToDo: add deadlock detection in GUM (similar to THREADED_RTS) -- HWL
1001 return;
1002 #endif
1003
1004 /*
1005 * Detect deadlock: when we have no threads to run, there are no
1006 * threads blocked, waiting for I/O, or sleeping, and all the
1007 * other tasks are waiting for work, we must have a deadlock of
1008 * some description.
1009 */
1010 if ( emptyThreadQueues(cap) )
1011 {
1012 #if defined(THREADED_RTS)
1013 /*
1014 * In the threaded RTS, we only check for deadlock if there
1015 * has been no activity in a complete timeslice. This means
1016 * we won't eagerly start a full GC just because we don't have
1017 * any threads to run currently.
1018 */
1019 if (recent_activity != ACTIVITY_INACTIVE) return;
1020 #endif
1021
1022 debugTrace(DEBUG_sched, "deadlocked, forcing major GC...");
1023
1024 // Garbage collection can release some new threads due to
1025 // either (a) finalizers or (b) threads resurrected because
1026 // they are unreachable and will therefore be sent an
1027 // exception. Any threads thus released will be immediately
1028 // runnable.
1029 cap = scheduleDoGC (cap, task, rtsTrue/*force major GC*/);
1030 // when force_major == rtsTrue. scheduleDoGC sets
1031 // recent_activity to ACTIVITY_DONE_GC and turns off the timer
1032 // signal.
1033
1034 if ( !emptyRunQueue(cap) ) return;
1035
1036 #if defined(RTS_USER_SIGNALS) && !defined(THREADED_RTS)
1037 /* If we have user-installed signal handlers, then wait
1038 * for signals to arrive rather then bombing out with a
1039 * deadlock.
1040 */
1041 if ( RtsFlags.MiscFlags.install_signal_handlers && anyUserHandlers() ) {
1042 debugTrace(DEBUG_sched,
1043 "still deadlocked, waiting for signals...");
1044
1045 awaitUserSignals();
1046
1047 if (signals_pending()) {
1048 startSignalHandlers(cap);
1049 }
1050
1051 // either we have threads to run, or we were interrupted:
1052 ASSERT(!emptyRunQueue(cap) || sched_state >= SCHED_INTERRUPTING);
1053
1054 return;
1055 }
1056 #endif
1057
1058 #if !defined(THREADED_RTS)
1059 /* Probably a real deadlock. Send the current main thread the
1060 * Deadlock exception.
1061 */
1062 if (task->tso) {
1063 switch (task->tso->why_blocked) {
1064 case BlockedOnSTM:
1065 case BlockedOnBlackHole:
1066 case BlockedOnException:
1067 case BlockedOnMVar:
1068 throwToSingleThreaded(cap, task->tso,
1069 (StgClosure *)nonTermination_closure);
1070 return;
1071 default:
1072 barf("deadlock: main thread blocked in a strange way");
1073 }
1074 }
1075 return;
1076 #endif
1077 }
1078 }
1079
1080
1081 /* ----------------------------------------------------------------------------
1082 * Send pending messages (PARALLEL_HASKELL only)
1083 * ------------------------------------------------------------------------- */
1084
1085 #if defined(PARALLEL_HASKELL)
1086 static void
1087 scheduleSendPendingMessages(void)
1088 {
1089
1090 # if defined(PAR) // global Mem.Mgmt., omit for now
1091 if (PendingFetches != END_BF_QUEUE) {
1092 processFetches();
1093 }
1094 # endif
1095
1096 if (RtsFlags.ParFlags.BufferTime) {
1097 // if we use message buffering, we must send away all message
1098 // packets which have become too old...
1099 sendOldBuffers();
1100 }
1101 }
1102 #endif
1103
1104 /* ----------------------------------------------------------------------------
1105 * Activate spark threads (PARALLEL_HASKELL and THREADED_RTS)
1106 * ------------------------------------------------------------------------- */
1107
1108 #if defined(PARALLEL_HASKELL) || defined(THREADED_RTS)
1109 static void
1110 scheduleActivateSpark(Capability *cap)
1111 {
1112 if (anySparks())
1113 {
1114 createSparkThread(cap);
1115 debugTrace(DEBUG_sched, "creating a spark thread");
1116 }
1117 }
1118 #endif // PARALLEL_HASKELL || THREADED_RTS
1119
1120 /* ----------------------------------------------------------------------------
1121 * Get work from a remote node (PARALLEL_HASKELL only)
1122 * ------------------------------------------------------------------------- */
1123
1124 #if defined(PARALLEL_HASKELL)
1125 static rtsBool /* return value used in PARALLEL_HASKELL only */
1126 scheduleGetRemoteWork (Capability *cap STG_UNUSED)
1127 {
1128 #if defined(PARALLEL_HASKELL)
1129 rtsBool receivedFinish = rtsFalse;
1130
1131 // idle() , i.e. send all buffers, wait for work
1132 if (RtsFlags.ParFlags.BufferTime) {
1133 IF_PAR_DEBUG(verbose,
1134 debugBelch("...send all pending data,"));
1135 {
1136 nat i;
1137 for (i=1; i<=nPEs; i++)
1138 sendImmediately(i); // send all messages away immediately
1139 }
1140 }
1141
1142 /* this would be the place for fishing in GUM...
1143
1144 if (no-earlier-fish-around)
1145 sendFish(choosePe());
1146 */
1147
1148 // Eden:just look for incoming messages (blocking receive)
1149 IF_PAR_DEBUG(verbose,
1150 debugBelch("...wait for incoming messages...\n"));
1151 processMessages(cap, &receivedFinish); // blocking receive...
1152
1153
1154 return receivedFinish;
1155 // reenter scheduling look after having received something
1156
1157 #else /* !PARALLEL_HASKELL, i.e. THREADED_RTS */
1158
1159 return rtsFalse; /* return value unused in THREADED_RTS */
1160
1161 #endif /* PARALLEL_HASKELL */
1162 }
1163 #endif // PARALLEL_HASKELL || THREADED_RTS
1164
1165 /* ----------------------------------------------------------------------------
1166 * After running a thread...
1167 * ------------------------------------------------------------------------- */
1168
1169 static void
1170 schedulePostRunThread (Capability *cap, StgTSO *t)
1171 {
1172 // We have to be able to catch transactions that are in an
1173 // infinite loop as a result of seeing an inconsistent view of
1174 // memory, e.g.
1175 //
1176 // atomically $ do
1177 // [a,b] <- mapM readTVar [ta,tb]
1178 // when (a == b) loop
1179 //
1180 // and a is never equal to b given a consistent view of memory.
1181 //
1182 if (t -> trec != NO_TREC && t -> why_blocked == NotBlocked) {
1183 if (!stmValidateNestOfTransactions (t -> trec)) {
1184 debugTrace(DEBUG_sched | DEBUG_stm,
1185 "trec %p found wasting its time", t);
1186
1187 // strip the stack back to the
1188 // ATOMICALLY_FRAME, aborting the (nested)
1189 // transaction, and saving the stack of any
1190 // partially-evaluated thunks on the heap.
1191 throwToSingleThreaded_(cap, t, NULL, rtsTrue);
1192
1193 ASSERT(get_itbl((StgClosure *)t->sp)->type == ATOMICALLY_FRAME);
1194 }
1195 }
1196
1197 /* some statistics gathering in the parallel case */
1198 }
1199
1200 /* -----------------------------------------------------------------------------
1201 * Handle a thread that returned to the scheduler with ThreadHeepOverflow
1202 * -------------------------------------------------------------------------- */
1203
1204 static rtsBool
1205 scheduleHandleHeapOverflow( Capability *cap, StgTSO *t )
1206 {
1207 // did the task ask for a large block?
1208 if (cap->r.rHpAlloc > BLOCK_SIZE) {
1209 // if so, get one and push it on the front of the nursery.
1210 bdescr *bd;
1211 lnat blocks;
1212
1213 blocks = (lnat)BLOCK_ROUND_UP(cap->r.rHpAlloc) / BLOCK_SIZE;
1214
1215 debugTrace(DEBUG_sched,
1216 "--<< thread %ld (%s) stopped: requesting a large block (size %ld)\n",
1217 (long)t->id, whatNext_strs[t->what_next], blocks);
1218
1219 // don't do this if the nursery is (nearly) full, we'll GC first.
1220 if (cap->r.rCurrentNursery->link != NULL ||
1221 cap->r.rNursery->n_blocks == 1) { // paranoia to prevent infinite loop
1222 // if the nursery has only one block.
1223
1224 ACQUIRE_SM_LOCK
1225 bd = allocGroup( blocks );
1226 RELEASE_SM_LOCK
1227 cap->r.rNursery->n_blocks += blocks;
1228
1229 // link the new group into the list
1230 bd->link = cap->r.rCurrentNursery;
1231 bd->u.back = cap->r.rCurrentNursery->u.back;
1232 if (cap->r.rCurrentNursery->u.back != NULL) {
1233 cap->r.rCurrentNursery->u.back->link = bd;
1234 } else {
1235 #if !defined(THREADED_RTS)
1236 ASSERT(g0s0->blocks == cap->r.rCurrentNursery &&
1237 g0s0 == cap->r.rNursery);
1238 #endif
1239 cap->r.rNursery->blocks = bd;
1240 }
1241 cap->r.rCurrentNursery->u.back = bd;
1242
1243 // initialise it as a nursery block. We initialise the
1244 // step, gen_no, and flags field of *every* sub-block in
1245 // this large block, because this is easier than making
1246 // sure that we always find the block head of a large
1247 // block whenever we call Bdescr() (eg. evacuate() and
1248 // isAlive() in the GC would both have to do this, at
1249 // least).
1250 {
1251 bdescr *x;
1252 for (x = bd; x < bd + blocks; x++) {
1253 x->step = cap->r.rNursery;
1254 x->gen_no = 0;
1255 x->flags = 0;
1256 }
1257 }
1258
1259 // This assert can be a killer if the app is doing lots
1260 // of large block allocations.
1261 IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
1262
1263 // now update the nursery to point to the new block
1264 cap->r.rCurrentNursery = bd;
1265
1266 // we might be unlucky and have another thread get on the
1267 // run queue before us and steal the large block, but in that
1268 // case the thread will just end up requesting another large
1269 // block.
1270 pushOnRunQueue(cap,t);
1271 return rtsFalse; /* not actually GC'ing */
1272 }
1273 }
1274
1275 debugTrace(DEBUG_sched,
1276 "--<< thread %ld (%s) stopped: HeapOverflow",
1277 (long)t->id, whatNext_strs[t->what_next]);
1278
1279 if (cap->r.rHpLim == NULL || cap->context_switch) {
1280 // Sometimes we miss a context switch, e.g. when calling
1281 // primitives in a tight loop, MAYBE_GC() doesn't check the
1282 // context switch flag, and we end up waiting for a GC.
1283 // See #1984, and concurrent/should_run/1984
1284 cap->context_switch = 0;
1285 addToRunQueue(cap,t);
1286 } else {
1287 pushOnRunQueue(cap,t);
1288 }
1289 return rtsTrue;
1290 /* actual GC is done at the end of the while loop in schedule() */
1291 }
1292
1293 /* -----------------------------------------------------------------------------
1294 * Handle a thread that returned to the scheduler with ThreadStackOverflow
1295 * -------------------------------------------------------------------------- */
1296
1297 static void
1298 scheduleHandleStackOverflow (Capability *cap, Task *task, StgTSO *t)
1299 {
1300 debugTrace (DEBUG_sched,
1301 "--<< thread %ld (%s) stopped, StackOverflow",
1302 (long)t->id, whatNext_strs[t->what_next]);
1303
1304 /* just adjust the stack for this thread, then pop it back
1305 * on the run queue.
1306 */
1307 {
1308 /* enlarge the stack */
1309 StgTSO *new_t = threadStackOverflow(cap, t);
1310
1311 /* The TSO attached to this Task may have moved, so update the
1312 * pointer to it.
1313 */
1314 if (task->tso == t) {
1315 task->tso = new_t;
1316 }
1317 pushOnRunQueue(cap,new_t);
1318 }
1319 }
1320
1321 /* -----------------------------------------------------------------------------
1322 * Handle a thread that returned to the scheduler with ThreadYielding
1323 * -------------------------------------------------------------------------- */
1324
1325 static rtsBool
1326 scheduleHandleYield( Capability *cap, StgTSO *t, nat prev_what_next )
1327 {
1328 // Reset the context switch flag. We don't do this just before
1329 // running the thread, because that would mean we would lose ticks
1330 // during GC, which can lead to unfair scheduling (a thread hogs
1331 // the CPU because the tick always arrives during GC). This way
1332 // penalises threads that do a lot of allocation, but that seems
1333 // better than the alternative.
1334 cap->context_switch = 0;
1335
1336 /* put the thread back on the run queue. Then, if we're ready to
1337 * GC, check whether this is the last task to stop. If so, wake
1338 * up the GC thread. getThread will block during a GC until the
1339 * GC is finished.
1340 */
1341 #ifdef DEBUG
1342 if (t->what_next != prev_what_next) {
1343 debugTrace(DEBUG_sched,
1344 "--<< thread %ld (%s) stopped to switch evaluators",
1345 (long)t->id, whatNext_strs[t->what_next]);
1346 } else {
1347 debugTrace(DEBUG_sched,
1348 "--<< thread %ld (%s) stopped, yielding",
1349 (long)t->id, whatNext_strs[t->what_next]);
1350 }
1351 #endif
1352
1353 IF_DEBUG(sanity,
1354 //debugBelch("&& Doing sanity check on yielding TSO %ld.", t->id);
1355 checkTSO(t));
1356 ASSERT(t->_link == END_TSO_QUEUE);
1357
1358 // Shortcut if we're just switching evaluators: don't bother
1359 // doing stack squeezing (which can be expensive), just run the
1360 // thread.
1361 if (t->what_next != prev_what_next) {
1362 return rtsTrue;
1363 }
1364
1365 addToRunQueue(cap,t);
1366
1367 return rtsFalse;
1368 }
1369
1370 /* -----------------------------------------------------------------------------
1371 * Handle a thread that returned to the scheduler with ThreadBlocked
1372 * -------------------------------------------------------------------------- */
1373
1374 static void
1375 scheduleHandleThreadBlocked( StgTSO *t
1376 #if !defined(GRAN) && !defined(DEBUG)
1377 STG_UNUSED
1378 #endif
1379 )
1380 {
1381
1382 // We don't need to do anything. The thread is blocked, and it
1383 // has tidied up its stack and placed itself on whatever queue
1384 // it needs to be on.
1385
1386 // ASSERT(t->why_blocked != NotBlocked);
1387 // Not true: for example,
1388 // - in THREADED_RTS, the thread may already have been woken
1389 // up by another Capability. This actually happens: try
1390 // conc023 +RTS -N2.
1391 // - the thread may have woken itself up already, because
1392 // threadPaused() might have raised a blocked throwTo
1393 // exception, see maybePerformBlockedException().
1394
1395 #ifdef DEBUG
1396 if (traceClass(DEBUG_sched)) {
1397 debugTraceBegin("--<< thread %lu (%s) stopped: ",
1398 (unsigned long)t->id, whatNext_strs[t->what_next]);
1399 printThreadBlockage(t);
1400 debugTraceEnd();
1401 }
1402 #endif
1403 }
1404
1405 /* -----------------------------------------------------------------------------
1406 * Handle a thread that returned to the scheduler with ThreadFinished
1407 * -------------------------------------------------------------------------- */
1408
1409 static rtsBool
1410 scheduleHandleThreadFinished (Capability *cap STG_UNUSED, Task *task, StgTSO *t)
1411 {
1412 /* Need to check whether this was a main thread, and if so,
1413 * return with the return value.
1414 *
1415 * We also end up here if the thread kills itself with an
1416 * uncaught exception, see Exception.cmm.
1417 */
1418 debugTrace(DEBUG_sched, "--++ thread %lu (%s) finished",
1419 (unsigned long)t->id, whatNext_strs[t->what_next]);
1420
1421 // blocked exceptions can now complete, even if the thread was in
1422 // blocked mode (see #2910). This unconditionally calls
1423 // lockTSO(), which ensures that we don't miss any threads that
1424 // are engaged in throwTo() with this thread as a target.
1425 awakenBlockedExceptionQueue (cap, t);
1426
1427 //
1428 // Check whether the thread that just completed was a bound
1429 // thread, and if so return with the result.
1430 //
1431 // There is an assumption here that all thread completion goes
1432 // through this point; we need to make sure that if a thread
1433 // ends up in the ThreadKilled state, that it stays on the run
1434 // queue so it can be dealt with here.
1435 //
1436
1437 if (t->bound) {
1438
1439 if (t->bound != task) {
1440 #if !defined(THREADED_RTS)
1441 // Must be a bound thread that is not the topmost one. Leave
1442 // it on the run queue until the stack has unwound to the
1443 // point where we can deal with this. Leaving it on the run
1444 // queue also ensures that the garbage collector knows about
1445 // this thread and its return value (it gets dropped from the
1446 // step->threads list so there's no other way to find it).
1447 appendToRunQueue(cap,t);
1448 return rtsFalse;
1449 #else
1450 // this cannot happen in the threaded RTS, because a
1451 // bound thread can only be run by the appropriate Task.
1452 barf("finished bound thread that isn't mine");
1453 #endif
1454 }
1455
1456 ASSERT(task->tso == t);
1457
1458 if (t->what_next == ThreadComplete) {
1459 if (task->ret) {
1460 // NOTE: return val is tso->sp[1] (see StgStartup.hc)
1461 *(task->ret) = (StgClosure *)task->tso->sp[1];
1462 }
1463 task->stat = Success;
1464 } else {
1465 if (task->ret) {
1466 *(task->ret) = NULL;
1467 }
1468 if (sched_state >= SCHED_INTERRUPTING) {
1469 if (heap_overflow) {
1470 task->stat = HeapExhausted;
1471 } else {
1472 task->stat = Interrupted;
1473 }
1474 } else {
1475 task->stat = Killed;
1476 }
1477 }
1478 #ifdef DEBUG
1479 removeThreadLabel((StgWord)task->tso->id);
1480 #endif
1481 return rtsTrue; // tells schedule() to return
1482 }
1483
1484 return rtsFalse;
1485 }
1486
1487 /* -----------------------------------------------------------------------------
1488 * Perform a heap census
1489 * -------------------------------------------------------------------------- */
1490
1491 static rtsBool
1492 scheduleNeedHeapProfile( rtsBool ready_to_gc STG_UNUSED )
1493 {
1494 // When we have +RTS -i0 and we're heap profiling, do a census at
1495 // every GC. This lets us get repeatable runs for debugging.
1496 if (performHeapProfile ||
1497 (RtsFlags.ProfFlags.profileInterval==0 &&
1498 RtsFlags.ProfFlags.doHeapProfile && ready_to_gc)) {
1499 return rtsTrue;
1500 } else {
1501 return rtsFalse;
1502 }
1503 }
1504
1505 /* -----------------------------------------------------------------------------
1506 * Perform a garbage collection if necessary
1507 * -------------------------------------------------------------------------- */
1508
1509 static Capability *
1510 scheduleDoGC (Capability *cap, Task *task USED_IF_THREADS, rtsBool force_major)
1511 {
1512 rtsBool heap_census;
1513 #ifdef THREADED_RTS
1514 /* extern static volatile StgWord waiting_for_gc;
1515 lives inside capability.c */
1516 rtsBool gc_type, prev_pending_gc;
1517 nat i;
1518 #endif
1519
1520 if (sched_state == SCHED_SHUTTING_DOWN) {
1521 // The final GC has already been done, and the system is
1522 // shutting down. We'll probably deadlock if we try to GC
1523 // now.
1524 return cap;
1525 }
1526
1527 #ifdef THREADED_RTS
1528 if (sched_state < SCHED_INTERRUPTING
1529 && RtsFlags.ParFlags.parGcEnabled
1530 && N >= RtsFlags.ParFlags.parGcGen
1531 && ! oldest_gen->steps[0].mark)
1532 {
1533 gc_type = PENDING_GC_PAR;
1534 } else {
1535 gc_type = PENDING_GC_SEQ;
1536 }
1537
1538 // In order to GC, there must be no threads running Haskell code.
1539 // Therefore, the GC thread needs to hold *all* the capabilities,
1540 // and release them after the GC has completed.
1541 //
1542 // This seems to be the simplest way: previous attempts involved
1543 // making all the threads with capabilities give up their
1544 // capabilities and sleep except for the *last* one, which
1545 // actually did the GC. But it's quite hard to arrange for all
1546 // the other tasks to sleep and stay asleep.
1547 //
1548
1549 /* Other capabilities are prevented from running yet more Haskell
1550 threads if waiting_for_gc is set. Tested inside
1551 yieldCapability() and releaseCapability() in Capability.c */
1552
1553 prev_pending_gc = cas(&waiting_for_gc, 0, gc_type);
1554 if (prev_pending_gc) {
1555 do {
1556 debugTrace(DEBUG_sched, "someone else is trying to GC (%d)...",
1557 prev_pending_gc);
1558 ASSERT(cap);
1559 yieldCapability(&cap,task);
1560 } while (waiting_for_gc);
1561 return cap; // NOTE: task->cap might have changed here
1562 }
1563
1564 setContextSwitches();
1565
1566 // The final shutdown GC is always single-threaded, because it's
1567 // possible that some of the Capabilities have no worker threads.
1568
1569 if (gc_type == PENDING_GC_SEQ)
1570 {
1571 postEvent(cap, EVENT_REQUEST_SEQ_GC, 0, 0);
1572 // single-threaded GC: grab all the capabilities
1573 for (i=0; i < n_capabilities; i++) {
1574 debugTrace(DEBUG_sched, "ready_to_gc, grabbing all the capabilies (%d/%d)", i, n_capabilities);
1575 if (cap != &capabilities[i]) {
1576 Capability *pcap = &capabilities[i];
1577 // we better hope this task doesn't get migrated to
1578 // another Capability while we're waiting for this one.
1579 // It won't, because load balancing happens while we have
1580 // all the Capabilities, but even so it's a slightly
1581 // unsavoury invariant.
1582 task->cap = pcap;
1583 waitForReturnCapability(&pcap, task);
1584 if (pcap != &capabilities[i]) {
1585 barf("scheduleDoGC: got the wrong capability");
1586 }
1587 }
1588 }
1589 }
1590 else
1591 {
1592 // multi-threaded GC: make sure all the Capabilities donate one
1593 // GC thread each.
1594 postEvent(cap, EVENT_REQUEST_PAR_GC, 0, 0);
1595 debugTrace(DEBUG_sched, "ready_to_gc, grabbing GC threads");
1596
1597 waitForGcThreads(cap);
1598 }
1599 #endif
1600
1601 // so this happens periodically:
1602 if (cap) scheduleCheckBlackHoles(cap);
1603
1604 IF_DEBUG(scheduler, printAllThreads());
1605
1606 delete_threads_and_gc:
1607 /*
1608 * We now have all the capabilities; if we're in an interrupting
1609 * state, then we should take the opportunity to delete all the
1610 * threads in the system.
1611 */
1612 if (sched_state == SCHED_INTERRUPTING) {
1613 deleteAllThreads(cap);
1614 sched_state = SCHED_SHUTTING_DOWN;
1615 }
1616
1617 heap_census = scheduleNeedHeapProfile(rtsTrue);
1618
1619 #if defined(THREADED_RTS)
1620 postEvent(cap, EVENT_GC_START, 0, 0);
1621 debugTrace(DEBUG_sched, "doing GC");
1622 // reset waiting_for_gc *before* GC, so that when the GC threads
1623 // emerge they don't immediately re-enter the GC.
1624 waiting_for_gc = 0;
1625 GarbageCollect(force_major || heap_census, gc_type, cap);
1626 #else
1627 GarbageCollect(force_major || heap_census, 0, cap);
1628 #endif
1629 postEvent(cap, EVENT_GC_END, 0, 0);
1630
1631 if (recent_activity == ACTIVITY_INACTIVE && force_major)
1632 {
1633 // We are doing a GC because the system has been idle for a
1634 // timeslice and we need to check for deadlock. Record the
1635 // fact that we've done a GC and turn off the timer signal;
1636 // it will get re-enabled if we run any threads after the GC.
1637 recent_activity = ACTIVITY_DONE_GC;
1638 stopTimer();
1639 }
1640 else
1641 {
1642 // the GC might have taken long enough for the timer to set
1643 // recent_activity = ACTIVITY_INACTIVE, but we aren't
1644 // necessarily deadlocked:
1645 recent_activity = ACTIVITY_YES;
1646 }
1647
1648 #if defined(THREADED_RTS)
1649 if (gc_type == PENDING_GC_PAR)
1650 {
1651 releaseGCThreads(cap);
1652 }
1653 #endif
1654
1655 if (heap_census) {
1656 debugTrace(DEBUG_sched, "performing heap census");
1657 heapCensus();
1658 performHeapProfile = rtsFalse;
1659 }
1660
1661 if (heap_overflow && sched_state < SCHED_INTERRUPTING) {
1662 // GC set the heap_overflow flag, so we should proceed with
1663 // an orderly shutdown now. Ultimately we want the main
1664 // thread to return to its caller with HeapExhausted, at which
1665 // point the caller should call hs_exit(). The first step is
1666 // to delete all the threads.
1667 //
1668 // Another way to do this would be to raise an exception in
1669 // the main thread, which we really should do because it gives
1670 // the program a chance to clean up. But how do we find the
1671 // main thread? It should presumably be the same one that
1672 // gets ^C exceptions, but that's all done on the Haskell side
1673 // (GHC.TopHandler).
1674 sched_state = SCHED_INTERRUPTING;
1675 goto delete_threads_and_gc;
1676 }
1677
1678 #ifdef SPARKBALANCE
1679 /* JB
1680 Once we are all together... this would be the place to balance all
1681 spark pools. No concurrent stealing or adding of new sparks can
1682 occur. Should be defined in Sparks.c. */
1683 balanceSparkPoolsCaps(n_capabilities, capabilities);
1684 #endif
1685
1686 #if defined(THREADED_RTS)
1687 if (gc_type == PENDING_GC_SEQ) {
1688 // release our stash of capabilities.
1689 for (i = 0; i < n_capabilities; i++) {
1690 if (cap != &capabilities[i]) {
1691 task->cap = &capabilities[i];
1692 releaseCapability(&capabilities[i]);
1693 }
1694 }
1695 }
1696 if (cap) {
1697 task->cap = cap;
1698 } else {
1699 task->cap = NULL;
1700 }
1701 #endif
1702
1703 return cap;
1704 }
1705
1706 /* ---------------------------------------------------------------------------
1707 * Singleton fork(). Do not copy any running threads.
1708 * ------------------------------------------------------------------------- */
1709
1710 pid_t
1711 forkProcess(HsStablePtr *entry
1712 #ifndef FORKPROCESS_PRIMOP_SUPPORTED
1713 STG_UNUSED
1714 #endif
1715 )
1716 {
1717 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
1718 Task *task;
1719 pid_t pid;
1720 StgTSO* t,*next;
1721 Capability *cap;
1722 nat s;
1723
1724 #if defined(THREADED_RTS)
1725 if (RtsFlags.ParFlags.nNodes > 1) {
1726 errorBelch("forking not supported with +RTS -N<n> greater than 1");
1727 stg_exit(EXIT_FAILURE);
1728 }
1729 #endif
1730
1731 debugTrace(DEBUG_sched, "forking!");
1732
1733 // ToDo: for SMP, we should probably acquire *all* the capabilities
1734 cap = rts_lock();
1735
1736 // no funny business: hold locks while we fork, otherwise if some
1737 // other thread is holding a lock when the fork happens, the data
1738 // structure protected by the lock will forever be in an
1739 // inconsistent state in the child. See also #1391.
1740 ACQUIRE_LOCK(&sched_mutex);
1741 ACQUIRE_LOCK(&cap->lock);
1742 ACQUIRE_LOCK(&cap->running_task->lock);
1743
1744 pid = fork();
1745
1746 if (pid) { // parent
1747
1748 RELEASE_LOCK(&sched_mutex);
1749 RELEASE_LOCK(&cap->lock);
1750 RELEASE_LOCK(&cap->running_task->lock);
1751
1752 // just return the pid
1753 rts_unlock(cap);
1754 return pid;
1755
1756 } else { // child
1757
1758 #if defined(THREADED_RTS)
1759 initMutex(&sched_mutex);
1760 initMutex(&cap->lock);
1761 initMutex(&cap->running_task->lock);
1762 #endif
1763
1764 // Now, all OS threads except the thread that forked are
1765 // stopped. We need to stop all Haskell threads, including
1766 // those involved in foreign calls. Also we need to delete
1767 // all Tasks, because they correspond to OS threads that are
1768 // now gone.
1769
1770 for (s = 0; s < total_steps; s++) {
1771 for (t = all_steps[s].threads; t != END_TSO_QUEUE; t = next) {
1772 if (t->what_next == ThreadRelocated) {
1773 next = t->_link;
1774 } else {
1775 next = t->global_link;
1776 // don't allow threads to catch the ThreadKilled
1777 // exception, but we do want to raiseAsync() because these
1778 // threads may be evaluating thunks that we need later.
1779 deleteThread_(cap,t);
1780 }
1781 }
1782 }
1783
1784 // Empty the run queue. It seems tempting to let all the
1785 // killed threads stay on the run queue as zombies to be
1786 // cleaned up later, but some of them correspond to bound
1787 // threads for which the corresponding Task does not exist.
1788 cap->run_queue_hd = END_TSO_QUEUE;
1789 cap->run_queue_tl = END_TSO_QUEUE;
1790
1791 // Any suspended C-calling Tasks are no more, their OS threads
1792 // don't exist now:
1793 cap->suspended_ccalling_tasks = NULL;
1794
1795 // Empty the threads lists. Otherwise, the garbage
1796 // collector may attempt to resurrect some of these threads.
1797 for (s = 0; s < total_steps; s++) {
1798 all_steps[s].threads = END_TSO_QUEUE;
1799 }
1800
1801 // Wipe the task list, except the current Task.
1802 ACQUIRE_LOCK(&sched_mutex);
1803 for (task = all_tasks; task != NULL; task=task->all_link) {
1804 if (task != cap->running_task) {
1805 #if defined(THREADED_RTS)
1806 initMutex(&task->lock); // see #1391
1807 #endif
1808 discardTask(task);
1809 }
1810 }
1811 RELEASE_LOCK(&sched_mutex);
1812
1813 #if defined(THREADED_RTS)
1814 // Wipe our spare workers list, they no longer exist. New
1815 // workers will be created if necessary.
1816 cap->spare_workers = NULL;
1817 cap->returning_tasks_hd = NULL;
1818 cap->returning_tasks_tl = NULL;
1819 #endif
1820
1821 // On Unix, all timers are reset in the child, so we need to start
1822 // the timer again.
1823 initTimer();
1824 startTimer();
1825
1826 cap = rts_evalStableIO(cap, entry, NULL); // run the action
1827 rts_checkSchedStatus("forkProcess",cap);
1828
1829 rts_unlock(cap);
1830 hs_exit(); // clean up and exit
1831 stg_exit(EXIT_SUCCESS);
1832 }
1833 #else /* !FORKPROCESS_PRIMOP_SUPPORTED */
1834 barf("forkProcess#: primop not supported on this platform, sorry!\n");
1835 return -1;
1836 #endif
1837 }
1838
1839 /* ---------------------------------------------------------------------------
1840 * Delete all the threads in the system
1841 * ------------------------------------------------------------------------- */
1842
1843 static void
1844 deleteAllThreads ( Capability *cap )
1845 {
1846 // NOTE: only safe to call if we own all capabilities.
1847
1848 StgTSO* t, *next;
1849 nat s;
1850
1851 debugTrace(DEBUG_sched,"deleting all threads");
1852 for (s = 0; s < total_steps; s++) {
1853 for (t = all_steps[s].threads; t != END_TSO_QUEUE; t = next) {
1854 if (t->what_next == ThreadRelocated) {
1855 next = t->_link;
1856 } else {
1857 next = t->global_link;
1858 deleteThread(cap,t);
1859 }
1860 }
1861 }
1862
1863 // The run queue now contains a bunch of ThreadKilled threads. We
1864 // must not throw these away: the main thread(s) will be in there
1865 // somewhere, and the main scheduler loop has to deal with it.
1866 // Also, the run queue is the only thing keeping these threads from
1867 // being GC'd, and we don't want the "main thread has been GC'd" panic.
1868
1869 #if !defined(THREADED_RTS)
1870 ASSERT(blocked_queue_hd == END_TSO_QUEUE);
1871 ASSERT(sleeping_queue == END_TSO_QUEUE);
1872 #endif
1873 }
1874
1875 /* -----------------------------------------------------------------------------
1876 Managing the suspended_ccalling_tasks list.
1877 Locks required: sched_mutex
1878 -------------------------------------------------------------------------- */
1879
1880 STATIC_INLINE void
1881 suspendTask (Capability *cap, Task *task)
1882 {
1883 ASSERT(task->next == NULL && task->prev == NULL);
1884 task->next = cap->suspended_ccalling_tasks;
1885 task->prev = NULL;
1886 if (cap->suspended_ccalling_tasks) {
1887 cap->suspended_ccalling_tasks->prev = task;
1888 }
1889 cap->suspended_ccalling_tasks = task;
1890 }
1891
1892 STATIC_INLINE void
1893 recoverSuspendedTask (Capability *cap, Task *task)
1894 {
1895 if (task->prev) {
1896 task->prev->next = task->next;
1897 } else {
1898 ASSERT(cap->suspended_ccalling_tasks == task);
1899 cap->suspended_ccalling_tasks = task->next;
1900 }
1901 if (task->next) {
1902 task->next->prev = task->prev;
1903 }
1904 task->next = task->prev = NULL;
1905 }
1906
1907 /* ---------------------------------------------------------------------------
1908 * Suspending & resuming Haskell threads.
1909 *
1910 * When making a "safe" call to C (aka _ccall_GC), the task gives back
1911 * its capability before calling the C function. This allows another
1912 * task to pick up the capability and carry on running Haskell
1913 * threads. It also means that if the C call blocks, it won't lock
1914 * the whole system.
1915 *
1916 * The Haskell thread making the C call is put to sleep for the
1917 * duration of the call, on the susepended_ccalling_threads queue. We
1918 * give out a token to the task, which it can use to resume the thread
1919 * on return from the C function.
1920 * ------------------------------------------------------------------------- */
1921
1922 void *
1923 suspendThread (StgRegTable *reg)
1924 {
1925 Capability *cap;
1926 int saved_errno;
1927 StgTSO *tso;
1928 Task *task;
1929 #if mingw32_HOST_OS
1930 StgWord32 saved_winerror;
1931 #endif
1932
1933 saved_errno = errno;
1934 #if mingw32_HOST_OS
1935 saved_winerror = GetLastError();
1936 #endif
1937
1938 /* assume that *reg is a pointer to the StgRegTable part of a Capability.
1939 */
1940 cap = regTableToCapability(reg);
1941
1942 task = cap->running_task;
1943 tso = cap->r.rCurrentTSO;
1944
1945 postEvent(cap, EVENT_STOP_THREAD, tso->id, THREAD_SUSPENDED_FOREIGN_CALL);
1946 debugTrace(DEBUG_sched,
1947 "thread %lu did a safe foreign call",
1948 (unsigned long)cap->r.rCurrentTSO->id);
1949
1950 // XXX this might not be necessary --SDM
1951 tso->what_next = ThreadRunGHC;
1952
1953 threadPaused(cap,tso);
1954
1955 if ((tso->flags & TSO_BLOCKEX) == 0) {
1956 tso->why_blocked = BlockedOnCCall;
1957 tso->flags |= TSO_BLOCKEX;
1958 tso->flags &= ~TSO_INTERRUPTIBLE;
1959 } else {
1960 tso->why_blocked = BlockedOnCCall_NoUnblockExc;
1961 }
1962
1963 // Hand back capability
1964 task->suspended_tso = tso;
1965
1966 ACQUIRE_LOCK(&cap->lock);
1967
1968 suspendTask(cap,task);
1969 cap->in_haskell = rtsFalse;
1970 releaseCapability_(cap,rtsFalse);
1971
1972 RELEASE_LOCK(&cap->lock);
1973
1974 #if defined(THREADED_RTS)
1975 /* Preparing to leave the RTS, so ensure there's a native thread/task
1976 waiting to take over.
1977 */
1978 debugTrace(DEBUG_sched, "thread %lu: leaving RTS", (unsigned long)tso->id);
1979 #endif
1980
1981 errno = saved_errno;
1982 #if mingw32_HOST_OS
1983 SetLastError(saved_winerror);
1984 #endif
1985 return task;
1986 }
1987
1988 StgRegTable *
1989 resumeThread (void *task_)
1990 {
1991 StgTSO *tso;
1992 Capability *cap;
1993 Task *task = task_;
1994 int saved_errno;
1995 #if mingw32_HOST_OS
1996 StgWord32 saved_winerror;
1997 #endif
1998
1999 saved_errno = errno;
2000 #if mingw32_HOST_OS
2001 saved_winerror = GetLastError();
2002 #endif
2003
2004 cap = task->cap;
2005 // Wait for permission to re-enter the RTS with the result.
2006 waitForReturnCapability(&cap,task);
2007 // we might be on a different capability now... but if so, our
2008 // entry on the suspended_ccalling_tasks list will also have been
2009 // migrated.
2010
2011 // Remove the thread from the suspended list
2012 recoverSuspendedTask(cap,task);
2013
2014 tso = task->suspended_tso;
2015 task->suspended_tso = NULL;
2016 tso->_link = END_TSO_QUEUE; // no write barrier reqd
2017
2018 postEvent(cap, EVENT_RUN_THREAD, tso->id, 0);
2019 debugTrace(DEBUG_sched, "thread %lu: re-entering RTS", (unsigned long)tso->id);
2020
2021 if (tso->why_blocked == BlockedOnCCall) {
2022 // avoid locking the TSO if we don't have to
2023 if (tso->blocked_exceptions != END_TSO_QUEUE) {
2024 awakenBlockedExceptionQueue(cap,tso);
2025 }
2026 tso->flags &= ~(TSO_BLOCKEX | TSO_INTERRUPTIBLE);
2027 }
2028
2029 /* Reset blocking status */
2030 tso->why_blocked = NotBlocked;
2031
2032 cap->r.rCurrentTSO = tso;
2033 cap->in_haskell = rtsTrue;
2034 errno = saved_errno;
2035 #if mingw32_HOST_OS
2036 SetLastError(saved_winerror);
2037 #endif
2038
2039 /* We might have GC'd, mark the TSO dirty again */
2040 dirty_TSO(cap,tso);
2041
2042 IF_DEBUG(sanity, checkTSO(tso));
2043
2044 return &cap->r;
2045 }
2046
2047 /* ---------------------------------------------------------------------------
2048 * scheduleThread()
2049 *
2050 * scheduleThread puts a thread on the end of the runnable queue.
2051 * This will usually be done immediately after a thread is created.
2052 * The caller of scheduleThread must create the thread using e.g.
2053 * createThread and push an appropriate closure
2054 * on this thread's stack before the scheduler is invoked.
2055 * ------------------------------------------------------------------------ */
2056
2057 void
2058 scheduleThread(Capability *cap, StgTSO *tso)
2059 {
2060 // The thread goes at the *end* of the run-queue, to avoid possible
2061 // starvation of any threads already on the queue.
2062 appendToRunQueue(cap,tso);
2063 }
2064
2065 void
2066 scheduleThreadOn(Capability *cap, StgWord cpu USED_IF_THREADS, StgTSO *tso)
2067 {
2068 #if defined(THREADED_RTS)
2069 tso->flags |= TSO_LOCKED; // we requested explicit affinity; don't
2070 // move this thread from now on.
2071 cpu %= RtsFlags.ParFlags.nNodes;
2072 if (cpu == cap->no) {
2073 appendToRunQueue(cap,tso);
2074 } else {
2075 postEvent (cap, EVENT_MIGRATE_THREAD, tso->id, capabilities[cpu].no);
2076 wakeupThreadOnCapability(cap, &capabilities[cpu], tso);
2077 }
2078 #else
2079 appendToRunQueue(cap,tso);
2080 #endif
2081 }
2082
2083 Capability *
2084 scheduleWaitThread (StgTSO* tso, /*[out]*/HaskellObj* ret, Capability *cap)
2085 {
2086 Task *task;
2087
2088 // We already created/initialised the Task
2089 task = cap->running_task;
2090
2091 // This TSO is now a bound thread; make the Task and TSO
2092 // point to each other.
2093 tso->bound = task;
2094 tso->cap = cap;
2095
2096 task->tso = tso;
2097 task->ret = ret;
2098 task->stat = NoStatus;
2099
2100 appendToRunQueue(cap,tso);
2101
2102 debugTrace(DEBUG_sched, "new bound thread (%lu)", (unsigned long)tso->id);
2103
2104 cap = schedule(cap,task);
2105
2106 ASSERT(task->stat != NoStatus);
2107 ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
2108
2109 debugTrace(DEBUG_sched, "bound thread (%lu) finished", (unsigned long)task->tso->id);
2110 return cap;
2111 }
2112
2113 /* ----------------------------------------------------------------------------
2114 * Starting Tasks
2115 * ------------------------------------------------------------------------- */
2116
2117 #if defined(THREADED_RTS)
2118 void OSThreadProcAttr
2119 workerStart(Task *task)
2120 {
2121 Capability *cap;
2122
2123 // See startWorkerTask().
2124 ACQUIRE_LOCK(&task->lock);
2125 cap = task->cap;
2126 RELEASE_LOCK(&task->lock);
2127
2128 if (RtsFlags.ParFlags.setAffinity) {
2129 setThreadAffinity(cap->no, n_capabilities);
2130 }
2131
2132 // set the thread-local pointer to the Task:
2133 taskEnter(task);
2134
2135 // schedule() runs without a lock.
2136 cap = schedule(cap,task);
2137
2138 // On exit from schedule(), we have a Capability, but possibly not
2139 // the same one we started with.
2140
2141 // During shutdown, the requirement is that after all the
2142 // Capabilities are shut down, all workers that are shutting down
2143 // have finished workerTaskStop(). This is why we hold on to
2144 // cap->lock until we've finished workerTaskStop() below.
2145 //
2146 // There may be workers still involved in foreign calls; those
2147 // will just block in waitForReturnCapability() because the
2148 // Capability has been shut down.
2149 //
2150 ACQUIRE_LOCK(&cap->lock);
2151 releaseCapability_(cap,rtsFalse);
2152 workerTaskStop(task);
2153 RELEASE_LOCK(&cap->lock);
2154 }
2155 #endif
2156
2157 /* ---------------------------------------------------------------------------
2158 * initScheduler()
2159 *
2160 * Initialise the scheduler. This resets all the queues - if the
2161 * queues contained any threads, they'll be garbage collected at the
2162 * next pass.
2163 *
2164 * ------------------------------------------------------------------------ */
2165
2166 void
2167 initScheduler(void)
2168 {
2169 #if !defined(THREADED_RTS)
2170 blocked_queue_hd = END_TSO_QUEUE;
2171 blocked_queue_tl = END_TSO_QUEUE;
2172 sleeping_queue = END_TSO_QUEUE;
2173 #endif
2174
2175 blackhole_queue = END_TSO_QUEUE;
2176
2177 sched_state = SCHED_RUNNING;
2178 recent_activity = ACTIVITY_YES;
2179
2180 #if defined(THREADED_RTS)
2181 /* Initialise the mutex and condition variables used by
2182 * the scheduler. */
2183 initMutex(&sched_mutex);
2184 #endif
2185
2186 ACQUIRE_LOCK(&sched_mutex);
2187
2188 /* A capability holds the state a native thread needs in
2189 * order to execute STG code. At least one capability is
2190 * floating around (only THREADED_RTS builds have more than one).
2191 */
2192 initCapabilities();
2193
2194 initTaskManager();
2195
2196 #if defined(THREADED_RTS) || defined(PARALLEL_HASKELL)
2197 initSparkPools();
2198 #endif
2199
2200 #if defined(THREADED_RTS)
2201 /*
2202 * Eagerly start one worker to run each Capability, except for
2203 * Capability 0. The idea is that we're probably going to start a
2204 * bound thread on Capability 0 pretty soon, so we don't want a
2205 * worker task hogging it.
2206 */
2207 {
2208 nat i;
2209 Capability *cap;
2210 for (i = 1; i < n_capabilities; i++) {
2211 cap = &capabilities[i];
2212 ACQUIRE_LOCK(&cap->lock);
2213 startWorkerTask(cap, workerStart);
2214 RELEASE_LOCK(&cap->lock);
2215 }
2216 }
2217 #endif
2218
2219 RELEASE_LOCK(&sched_mutex);
2220 }
2221
2222 void
2223 exitScheduler(
2224 rtsBool wait_foreign
2225 #if !defined(THREADED_RTS)
2226 __attribute__((unused))
2227 #endif
2228 )
2229 /* see Capability.c, shutdownCapability() */
2230 {
2231 Task *task = NULL;
2232
2233 ACQUIRE_LOCK(&sched_mutex);
2234 task = newBoundTask();
2235 RELEASE_LOCK(&sched_mutex);
2236
2237 // If we haven't killed all the threads yet, do it now.
2238 if (sched_state < SCHED_SHUTTING_DOWN) {
2239 sched_state = SCHED_INTERRUPTING;
2240 waitForReturnCapability(&task->cap,task);
2241 scheduleDoGC(task->cap,task,rtsFalse);
2242 releaseCapability(task->cap);
2243 }
2244 sched_state = SCHED_SHUTTING_DOWN;
2245
2246 #if defined(THREADED_RTS)
2247 {
2248 nat i;
2249
2250 for (i = 0; i < n_capabilities; i++) {
2251 shutdownCapability(&capabilities[i], task, wait_foreign);
2252 }
2253 boundTaskExiting(task);
2254 }
2255 #endif
2256 }
2257
2258 void
2259 freeScheduler( void )
2260 {
2261 nat still_running;
2262
2263 ACQUIRE_LOCK(&sched_mutex);
2264 still_running = freeTaskManager();
2265 // We can only free the Capabilities if there are no Tasks still
2266 // running. We might have a Task about to return from a foreign
2267 // call into waitForReturnCapability(), for example (actually,
2268 // this should be the *only* thing that a still-running Task can
2269 // do at this point, and it will block waiting for the
2270 // Capability).
2271 if (still_running == 0) {
2272 freeCapabilities();
2273 if (n_capabilities != 1) {
2274 stgFree(capabilities);
2275 }
2276 }
2277 RELEASE_LOCK(&sched_mutex);
2278 #if defined(THREADED_RTS)
2279 closeMutex(&sched_mutex);
2280 #endif
2281 }
2282
2283 /* -----------------------------------------------------------------------------
2284 performGC
2285
2286 This is the interface to the garbage collector from Haskell land.
2287 We provide this so that external C code can allocate and garbage
2288 collect when called from Haskell via _ccall_GC.
2289 -------------------------------------------------------------------------- */
2290
2291 static void
2292 performGC_(rtsBool force_major)
2293 {
2294 Task *task;
2295
2296 // We must grab a new Task here, because the existing Task may be
2297 // associated with a particular Capability, and chained onto the
2298 // suspended_ccalling_tasks queue.
2299 ACQUIRE_LOCK(&sched_mutex);
2300 task = newBoundTask();
2301 RELEASE_LOCK(&sched_mutex);
2302
2303 waitForReturnCapability(&task->cap,task);
2304 scheduleDoGC(task->cap,task,force_major);
2305 releaseCapability(task->cap);
2306 boundTaskExiting(task);
2307 }
2308
2309 void
2310 performGC(void)
2311 {
2312 performGC_(rtsFalse);
2313 }
2314
2315 void
2316 performMajorGC(void)
2317 {
2318 performGC_(rtsTrue);
2319 }
2320
2321 /* -----------------------------------------------------------------------------
2322 Stack overflow
2323
2324 If the thread has reached its maximum stack size, then raise the
2325 StackOverflow exception in the offending thread. Otherwise
2326 relocate the TSO into a larger chunk of memory and adjust its stack
2327 size appropriately.
2328 -------------------------------------------------------------------------- */
2329
2330 static StgTSO *
2331 threadStackOverflow(Capability *cap, StgTSO *tso)
2332 {
2333 nat new_stack_size, stack_words;
2334 lnat new_tso_size;
2335 StgPtr new_sp;
2336 StgTSO *dest;
2337
2338 IF_DEBUG(sanity,checkTSO(tso));
2339
2340 // don't allow throwTo() to modify the blocked_exceptions queue
2341 // while we are moving the TSO:
2342 lockClosure((StgClosure *)tso);
2343
2344 if (tso->stack_size >= tso->max_stack_size && !(tso->flags & TSO_BLOCKEX)) {
2345 // NB. never raise a StackOverflow exception if the thread is
2346 // inside Control.Exceptino.block. It is impractical to protect
2347 // against stack overflow exceptions, since virtually anything
2348 // can raise one (even 'catch'), so this is the only sensible
2349 // thing to do here. See bug #767.
2350
2351 debugTrace(DEBUG_gc,
2352 "threadStackOverflow of TSO %ld (%p): stack too large (now %ld; max is %ld)",
2353 (long)tso->id, tso, (long)tso->stack_size, (long)tso->max_stack_size);
2354 IF_DEBUG(gc,
2355 /* If we're debugging, just print out the top of the stack */
2356 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2357 tso->sp+64)));
2358
2359 // Send this thread the StackOverflow exception
2360 unlockTSO(tso);
2361 throwToSingleThreaded(cap, tso, (StgClosure *)stackOverflow_closure);
2362 return tso;
2363 }
2364
2365 /* Try to double the current stack size. If that takes us over the
2366 * maximum stack size for this thread, then use the maximum instead
2367 * (that is, unless we're already at or over the max size and we
2368 * can't raise the StackOverflow exception (see above), in which
2369 * case just double the size). Finally round up so the TSO ends up as
2370 * a whole number of blocks.
2371 */
2372 if (tso->stack_size >= tso->max_stack_size) {
2373 new_stack_size = tso->stack_size * 2;
2374 } else {
2375 new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
2376 }
2377 new_tso_size = (lnat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) +
2378 TSO_STRUCT_SIZE)/sizeof(W_);
2379 new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */
2380 new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
2381
2382 debugTrace(DEBUG_sched,
2383 "increasing stack size from %ld words to %d.",
2384 (long)tso->stack_size, new_stack_size);
2385
2386 dest = (StgTSO *)allocateLocal(cap,new_tso_size);
2387 TICK_ALLOC_TSO(new_stack_size,0);
2388
2389 /* copy the TSO block and the old stack into the new area */
2390 memcpy(dest,tso,TSO_STRUCT_SIZE);
2391 stack_words = tso->stack + tso->stack_size - tso->sp;
2392 new_sp = (P_)dest + new_tso_size - stack_words;
2393 memcpy(new_sp, tso->sp, stack_words * sizeof(W_));
2394
2395 /* relocate the stack pointers... */
2396 dest->sp = new_sp;
2397 dest->stack_size = new_stack_size;
2398
2399 /* Mark the old TSO as relocated. We have to check for relocated
2400 * TSOs in the garbage collector and any primops that deal with TSOs.
2401 *
2402 * It's important to set the sp value to just beyond the end
2403 * of the stack, so we don't attempt to scavenge any part of the
2404 * dead TSO's stack.
2405 */
2406 tso->what_next = ThreadRelocated;
2407 setTSOLink(cap,tso,dest);
2408 tso->sp = (P_)&(tso->stack[tso->stack_size]);
2409 tso->why_blocked = NotBlocked;
2410
2411 IF_PAR_DEBUG(verbose,
2412 debugBelch("@@ threadStackOverflow of TSO %d (now at %p): stack size increased to %ld\n",
2413 tso->id, tso, tso->stack_size);
2414 /* If we're debugging, just print out the top of the stack */
2415 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2416 tso->sp+64)));
2417
2418 unlockTSO(dest);
2419 unlockTSO(tso);
2420
2421 IF_DEBUG(sanity,checkTSO(dest));
2422 #if 0
2423 IF_DEBUG(scheduler,printTSO(dest));
2424 #endif
2425
2426 return dest;
2427 }
2428
2429 static StgTSO *
2430 threadStackUnderflow (Task *task STG_UNUSED, StgTSO *tso)
2431 {
2432 bdescr *bd, *new_bd;
2433 lnat free_w, tso_size_w;
2434 StgTSO *new_tso;
2435
2436 tso_size_w = tso_sizeW(tso);
2437
2438 if (tso_size_w < MBLOCK_SIZE_W ||
2439 (nat)(tso->stack + tso->stack_size - tso->sp) > tso->stack_size / 4)
2440 {
2441 return tso;
2442 }
2443
2444 // don't allow throwTo() to modify the blocked_exceptions queue
2445 // while we are moving the TSO:
2446 lockClosure((StgClosure *)tso);
2447
2448 // this is the number of words we'll free
2449 free_w = round_to_mblocks(tso_size_w/2);
2450
2451 bd = Bdescr((StgPtr)tso);
2452 new_bd = splitLargeBlock(bd, free_w / BLOCK_SIZE_W);
2453 bd->free = bd->start + TSO_STRUCT_SIZEW;
2454
2455 new_tso = (StgTSO *)new_bd->start;
2456 memcpy(new_tso,tso,TSO_STRUCT_SIZE);
2457 new_tso->stack_size = new_bd->free - new_tso->stack;
2458
2459 debugTrace(DEBUG_sched, "thread %ld: reducing TSO size from %lu words to %lu",
2460 (long)tso->id, tso_size_w, tso_sizeW(new_tso));
2461
2462 tso->what_next = ThreadRelocated;
2463 tso->_link = new_tso; // no write barrier reqd: same generation
2464
2465 // The TSO attached to this Task may have moved, so update the
2466 // pointer to it.
2467 if (task->tso == tso) {
2468 task->tso = new_tso;
2469 }
2470
2471 unlockTSO(new_tso);
2472 unlockTSO(tso);
2473
2474 IF_DEBUG(sanity,checkTSO(new_tso));
2475
2476 return new_tso;
2477 }
2478
2479 /* ---------------------------------------------------------------------------
2480 Interrupt execution
2481 - usually called inside a signal handler so it mustn't do anything fancy.
2482 ------------------------------------------------------------------------ */
2483
2484 void
2485 interruptStgRts(void)
2486 {
2487 sched_state = SCHED_INTERRUPTING;
2488 setContextSwitches();
2489 wakeUpRts();
2490 }
2491
2492 /* -----------------------------------------------------------------------------
2493 Wake up the RTS
2494
2495 This function causes at least one OS thread to wake up and run the
2496 scheduler loop. It is invoked when the RTS might be deadlocked, or
2497 an external event has arrived that may need servicing (eg. a
2498 keyboard interrupt).
2499
2500 In the single-threaded RTS we don't do anything here; we only have
2501 one thread anyway, and the event that caused us to want to wake up
2502 will have interrupted any blocking system call in progress anyway.
2503 -------------------------------------------------------------------------- */
2504
2505 void
2506 wakeUpRts(void)
2507 {
2508 #if defined(THREADED_RTS)
2509 // This forces the IO Manager thread to wakeup, which will
2510 // in turn ensure that some OS thread wakes up and runs the
2511 // scheduler loop, which will cause a GC and deadlock check.
2512 ioManagerWakeup();
2513 #endif
2514 }
2515
2516 /* -----------------------------------------------------------------------------
2517 * checkBlackHoles()
2518 *
2519 * Check the blackhole_queue for threads that can be woken up. We do
2520 * this periodically: before every GC, and whenever the run queue is
2521 * empty.
2522 *
2523 * An elegant solution might be to just wake up all the blocked
2524 * threads with awakenBlockedQueue occasionally: they'll go back to
2525 * sleep again if the object is still a BLACKHOLE. Unfortunately this
2526 * doesn't give us a way to tell whether we've actually managed to
2527 * wake up any threads, so we would be busy-waiting.
2528 *
2529 * -------------------------------------------------------------------------- */
2530
2531 static rtsBool
2532 checkBlackHoles (Capability *cap)
2533 {
2534 StgTSO **prev, *t;
2535 rtsBool any_woke_up = rtsFalse;
2536 StgHalfWord type;
2537
2538 // blackhole_queue is global:
2539 ASSERT_LOCK_HELD(&sched_mutex);
2540
2541 debugTrace(DEBUG_sched, "checking threads blocked on black holes");
2542
2543 // ASSUMES: sched_mutex
2544 prev = &blackhole_queue;
2545 t = blackhole_queue;
2546 while (t != END_TSO_QUEUE) {
2547 if (t->what_next == ThreadRelocated) {
2548 t = t->_link;
2549 continue;
2550 }
2551 ASSERT(t->why_blocked == BlockedOnBlackHole);
2552 type = get_itbl(UNTAG_CLOSURE(t->block_info.closure))->type;
2553 if (type != BLACKHOLE && type != CAF_BLACKHOLE) {
2554 IF_DEBUG(sanity,checkTSO(t));
2555 t = unblockOne(cap, t);
2556 *prev = t;
2557 any_woke_up = rtsTrue;
2558 } else {
2559 prev = &t->_link;
2560 t = t->_link;
2561 }
2562 }
2563
2564 return any_woke_up;
2565 }
2566
2567 /* -----------------------------------------------------------------------------
2568 Deleting threads
2569
2570 This is used for interruption (^C) and forking, and corresponds to
2571 raising an exception but without letting the thread catch the
2572 exception.
2573 -------------------------------------------------------------------------- */
2574
2575 static void
2576 deleteThread (Capability *cap, StgTSO *tso)
2577 {
2578 // NOTE: must only be called on a TSO that we have exclusive
2579 // access to, because we will call throwToSingleThreaded() below.
2580 // The TSO must be on the run queue of the Capability we own, or
2581 // we must own all Capabilities.
2582
2583 if (tso->why_blocked != BlockedOnCCall &&
2584 tso->why_blocked != BlockedOnCCall_NoUnblockExc) {
2585 throwToSingleThreaded(cap,tso,NULL);
2586 }
2587 }
2588
2589 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
2590 static void
2591 deleteThread_(Capability *cap, StgTSO *tso)
2592 { // for forkProcess only:
2593 // like deleteThread(), but we delete threads in foreign calls, too.
2594
2595 if (tso->why_blocked == BlockedOnCCall ||
2596 tso->why_blocked == BlockedOnCCall_NoUnblockExc) {
2597 unblockOne(cap,tso);
2598 tso->what_next = ThreadKilled;
2599 } else {
2600 deleteThread(cap,tso);
2601 }
2602 }
2603 #endif
2604
2605 /* -----------------------------------------------------------------------------
2606 raiseExceptionHelper
2607
2608 This function is called by the raise# primitve, just so that we can
2609 move some of the tricky bits of raising an exception from C-- into
2610 C. Who knows, it might be a useful re-useable thing here too.
2611 -------------------------------------------------------------------------- */
2612
2613 StgWord
2614 raiseExceptionHelper (StgRegTable *reg, StgTSO *tso, StgClosure *exception)
2615 {
2616 Capability *cap = regTableToCapability(reg);
2617 StgThunk *raise_closure = NULL;
2618 StgPtr p, next;
2619 StgRetInfoTable *info;
2620 //
2621 // This closure represents the expression 'raise# E' where E
2622 // is the exception raise. It is used to overwrite all the
2623 // thunks which are currently under evaluataion.
2624 //
2625
2626 // OLD COMMENT (we don't have MIN_UPD_SIZE now):
2627 // LDV profiling: stg_raise_info has THUNK as its closure
2628 // type. Since a THUNK takes at least MIN_UPD_SIZE words in its
2629 // payload, MIN_UPD_SIZE is more approprate than 1. It seems that
2630 // 1 does not cause any problem unless profiling is performed.
2631 // However, when LDV profiling goes on, we need to linearly scan
2632 // small object pool, where raise_closure is stored, so we should
2633 // use MIN_UPD_SIZE.
2634 //
2635 // raise_closure = (StgClosure *)RET_STGCALL1(P_,allocate,
2636 // sizeofW(StgClosure)+1);
2637 //
2638
2639 //
2640 // Walk up the stack, looking for the catch frame. On the way,
2641 // we update any closures pointed to from update frames with the
2642 // raise closure that we just built.
2643 //
2644 p = tso->sp;
2645 while(1) {
2646 info = get_ret_itbl((StgClosure *)p);
2647 next = p + stack_frame_sizeW((StgClosure *)p);
2648 switch (info->i.type) {
2649
2650 case UPDATE_FRAME:
2651 // Only create raise_closure if we need to.
2652 if (raise_closure == NULL) {
2653 raise_closure =
2654 (StgThunk *)allocateLocal(cap,sizeofW(StgThunk)+1);
2655 SET_HDR(raise_closure, &stg_raise_info, CCCS);
2656 raise_closure->payload[0] = exception;
2657 }
2658 UPD_IND(((StgUpdateFrame *)p)->updatee,(StgClosure *)raise_closure);
2659 p = next;
2660 continue;
2661
2662 case ATOMICALLY_FRAME:
2663 debugTrace(DEBUG_stm, "found ATOMICALLY_FRAME at %p", p);
2664 tso->sp = p;
2665 return ATOMICALLY_FRAME;
2666
2667 case CATCH_FRAME:
2668 tso->sp = p;
2669 return CATCH_FRAME;
2670
2671 case CATCH_STM_FRAME:
2672 debugTrace(DEBUG_stm, "found CATCH_STM_FRAME at %p", p);
2673 tso->sp = p;
2674 return CATCH_STM_FRAME;
2675
2676 case STOP_FRAME:
2677 tso->sp = p;
2678 return STOP_FRAME;
2679
2680 case CATCH_RETRY_FRAME:
2681 default:
2682 p = next;
2683 continue;
2684 }
2685 }
2686 }
2687
2688
2689 /* -----------------------------------------------------------------------------
2690 findRetryFrameHelper
2691
2692 This function is called by the retry# primitive. It traverses the stack
2693 leaving tso->sp referring to the frame which should handle the retry.
2694
2695 This should either be a CATCH_RETRY_FRAME (if the retry# is within an orElse#)
2696 or should be a ATOMICALLY_FRAME (if the retry# reaches the top level).
2697
2698 We skip CATCH_STM_FRAMEs (aborting and rolling back the nested tx that they
2699 create) because retries are not considered to be exceptions, despite the
2700 similar implementation.
2701
2702 We should not expect to see CATCH_FRAME or STOP_FRAME because those should
2703 not be created within memory transactions.
2704 -------------------------------------------------------------------------- */
2705
2706 StgWord
2707 findRetryFrameHelper (StgTSO *tso)
2708 {
2709 StgPtr p, next;
2710 StgRetInfoTable *info;
2711
2712 p = tso -> sp;
2713 while (1) {
2714 info = get_ret_itbl((StgClosure *)p);
2715 next = p + stack_frame_sizeW((StgClosure *)p);
2716 switch (info->i.type) {
2717
2718 case ATOMICALLY_FRAME:
2719 debugTrace(DEBUG_stm,
2720 "found ATOMICALLY_FRAME at %p during retry", p);
2721 tso->sp = p;
2722 return ATOMICALLY_FRAME;
2723
2724 case CATCH_RETRY_FRAME:
2725 debugTrace(DEBUG_stm,
2726 "found CATCH_RETRY_FRAME at %p during retrry", p);
2727 tso->sp = p;
2728 return CATCH_RETRY_FRAME;
2729
2730 case CATCH_STM_FRAME: {
2731 StgTRecHeader *trec = tso -> trec;
2732 StgTRecHeader *outer = stmGetEnclosingTRec(trec);
2733 debugTrace(DEBUG_stm,
2734 "found CATCH_STM_FRAME at %p during retry", p);
2735 debugTrace(DEBUG_stm, "trec=%p outer=%p", trec, outer);
2736 stmAbortTransaction(tso -> cap, trec);
2737 stmFreeAbortedTRec(tso -> cap, trec);
2738 tso -> trec = outer;
2739 p = next;
2740 continue;
2741 }
2742
2743
2744 default:
2745 ASSERT(info->i.type != CATCH_FRAME);
2746 ASSERT(info->i.type != STOP_FRAME);
2747 p = next;
2748 continue;
2749 }
2750 }
2751 }
2752
2753 /* -----------------------------------------------------------------------------
2754 resurrectThreads is called after garbage collection on the list of
2755 threads found to be garbage. Each of these threads will be woken
2756 up and sent a signal: BlockedOnDeadMVar if the thread was blocked
2757 on an MVar, or NonTermination if the thread was blocked on a Black
2758 Hole.
2759
2760 Locks: assumes we hold *all* the capabilities.
2761 -------------------------------------------------------------------------- */
2762
2763 void
2764 resurrectThreads (StgTSO *threads)
2765 {
2766 StgTSO *tso, *next;
2767 Capability *cap;
2768 step *step;
2769
2770 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
2771 next = tso->global_link;
2772
2773 step = Bdescr((P_)tso)->step;
2774 tso->global_link = step->threads;
2775 step->threads = tso;
2776
2777 debugTrace(DEBUG_sched, "resurrecting thread %lu", (unsigned long)tso->id);
2778
2779 // Wake up the thread on the Capability it was last on
2780 cap = tso->cap;
2781
2782 switch (tso->why_blocked) {
2783 case BlockedOnMVar:
2784 case BlockedOnException:
2785 /* Called by GC - sched_mutex lock is currently held. */
2786 throwToSingleThreaded(cap, tso,
2787 (StgClosure *)blockedOnDeadMVar_closure);
2788 break;
2789 case BlockedOnBlackHole:
2790 throwToSingleThreaded(cap, tso,
2791 (StgClosure *)nonTermination_closure);
2792 break;
2793 case BlockedOnSTM:
2794 throwToSingleThreaded(cap, tso,
2795 (StgClosure *)blockedIndefinitely_closure);
2796 break;
2797 case NotBlocked:
2798 /* This might happen if the thread was blocked on a black hole
2799 * belonging to a thread that we've just woken up (raiseAsync
2800 * can wake up threads, remember...).
2801 */
2802 continue;
2803 default:
2804 barf("resurrectThreads: thread blocked in a strange way");
2805 }
2806 }
2807 }
2808
2809 /* -----------------------------------------------------------------------------
2810 performPendingThrowTos is called after garbage collection, and
2811 passed a list of threads that were found to have pending throwTos
2812 (tso->blocked_exceptions was not empty), and were blocked.
2813 Normally this doesn't happen, because we would deliver the
2814 exception directly if the target thread is blocked, but there are
2815 small windows where it might occur on a multiprocessor (see
2816 throwTo()).
2817
2818 NB. we must be holding all the capabilities at this point, just
2819 like resurrectThreads().
2820 -------------------------------------------------------------------------- */
2821
2822 void
2823 performPendingThrowTos (StgTSO *threads)
2824 {
2825 StgTSO *tso, *next;
2826 Capability *cap;
2827 step *step;
2828
2829 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
2830 next = tso->global_link;
2831
2832 step = Bdescr((P_)tso)->step;
2833 tso->global_link = step->threads;
2834 step->threads = tso;
2835
2836 debugTrace(DEBUG_sched, "performing blocked throwTo to thread %lu", (unsigned long)tso->id);
2837
2838 cap = tso->cap;
2839 maybePerformBlockedException(cap, tso);
2840 }
2841 }