Eventlog support for new event type: create spark.
[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
885 postEvent(free_caps[i], EVENT_STEAL_SPARK, t->id, cap->no);
886
887 newSpark(&(free_caps[i]->r), spark);
888 }
889 }
890 }
891 }
892 #endif /* SPARK_PUSHING */
893
894 // release the capabilities
895 for (i = 0; i < n_free_caps; i++) {
896 task->cap = free_caps[i];
897 releaseAndWakeupCapability(free_caps[i]);
898 }
899 }
900 task->cap = cap; // reset to point to our Capability.
901
902 #endif /* THREADED_RTS */
903
904 }
905
906 /* ----------------------------------------------------------------------------
907 * Start any pending signal handlers
908 * ------------------------------------------------------------------------- */
909
910 #if defined(RTS_USER_SIGNALS) && !defined(THREADED_RTS)
911 static void
912 scheduleStartSignalHandlers(Capability *cap)
913 {
914 if (RtsFlags.MiscFlags.install_signal_handlers && signals_pending()) {
915 // safe outside the lock
916 startSignalHandlers(cap);
917 }
918 }
919 #else
920 static void
921 scheduleStartSignalHandlers(Capability *cap STG_UNUSED)
922 {
923 }
924 #endif
925
926 /* ----------------------------------------------------------------------------
927 * Check for blocked threads that can be woken up.
928 * ------------------------------------------------------------------------- */
929
930 static void
931 scheduleCheckBlockedThreads(Capability *cap USED_IF_NOT_THREADS)
932 {
933 #if !defined(THREADED_RTS)
934 //
935 // Check whether any waiting threads need to be woken up. If the
936 // run queue is empty, and there are no other tasks running, we
937 // can wait indefinitely for something to happen.
938 //
939 if ( !emptyQueue(blocked_queue_hd) || !emptyQueue(sleeping_queue) )
940 {
941 awaitEvent( emptyRunQueue(cap) && !blackholes_need_checking );
942 }
943 #endif
944 }
945
946
947 /* ----------------------------------------------------------------------------
948 * Check for threads woken up by other Capabilities
949 * ------------------------------------------------------------------------- */
950
951 static void
952 scheduleCheckWakeupThreads(Capability *cap USED_IF_THREADS)
953 {
954 #if defined(THREADED_RTS)
955 // Any threads that were woken up by other Capabilities get
956 // appended to our run queue.
957 if (!emptyWakeupQueue(cap)) {
958 ACQUIRE_LOCK(&cap->lock);
959 if (emptyRunQueue(cap)) {
960 cap->run_queue_hd = cap->wakeup_queue_hd;
961 cap->run_queue_tl = cap->wakeup_queue_tl;
962 } else {
963 setTSOLink(cap, cap->run_queue_tl, cap->wakeup_queue_hd);
964 cap->run_queue_tl = cap->wakeup_queue_tl;
965 }
966 cap->wakeup_queue_hd = cap->wakeup_queue_tl = END_TSO_QUEUE;
967 RELEASE_LOCK(&cap->lock);
968 }
969 #endif
970 }
971
972 /* ----------------------------------------------------------------------------
973 * Check for threads blocked on BLACKHOLEs that can be woken up
974 * ------------------------------------------------------------------------- */
975 static void
976 scheduleCheckBlackHoles (Capability *cap)
977 {
978 if ( blackholes_need_checking ) // check without the lock first
979 {
980 ACQUIRE_LOCK(&sched_mutex);
981 if ( blackholes_need_checking ) {
982 blackholes_need_checking = rtsFalse;
983 // important that we reset the flag *before* checking the
984 // blackhole queue, otherwise we could get deadlock. This
985 // happens as follows: we wake up a thread that
986 // immediately runs on another Capability, blocks on a
987 // blackhole, and then we reset the blackholes_need_checking flag.
988 checkBlackHoles(cap);
989 }
990 RELEASE_LOCK(&sched_mutex);
991 }
992 }
993
994 /* ----------------------------------------------------------------------------
995 * Detect deadlock conditions and attempt to resolve them.
996 * ------------------------------------------------------------------------- */
997
998 static void
999 scheduleDetectDeadlock (Capability *cap, Task *task)
1000 {
1001
1002 #if defined(PARALLEL_HASKELL)
1003 // ToDo: add deadlock detection in GUM (similar to THREADED_RTS) -- HWL
1004 return;
1005 #endif
1006
1007 /*
1008 * Detect deadlock: when we have no threads to run, there are no
1009 * threads blocked, waiting for I/O, or sleeping, and all the
1010 * other tasks are waiting for work, we must have a deadlock of
1011 * some description.
1012 */
1013 if ( emptyThreadQueues(cap) )
1014 {
1015 #if defined(THREADED_RTS)
1016 /*
1017 * In the threaded RTS, we only check for deadlock if there
1018 * has been no activity in a complete timeslice. This means
1019 * we won't eagerly start a full GC just because we don't have
1020 * any threads to run currently.
1021 */
1022 if (recent_activity != ACTIVITY_INACTIVE) return;
1023 #endif
1024
1025 debugTrace(DEBUG_sched, "deadlocked, forcing major GC...");
1026
1027 // Garbage collection can release some new threads due to
1028 // either (a) finalizers or (b) threads resurrected because
1029 // they are unreachable and will therefore be sent an
1030 // exception. Any threads thus released will be immediately
1031 // runnable.
1032 cap = scheduleDoGC (cap, task, rtsTrue/*force major GC*/);
1033 // when force_major == rtsTrue. scheduleDoGC sets
1034 // recent_activity to ACTIVITY_DONE_GC and turns off the timer
1035 // signal.
1036
1037 if ( !emptyRunQueue(cap) ) return;
1038
1039 #if defined(RTS_USER_SIGNALS) && !defined(THREADED_RTS)
1040 /* If we have user-installed signal handlers, then wait
1041 * for signals to arrive rather then bombing out with a
1042 * deadlock.
1043 */
1044 if ( RtsFlags.MiscFlags.install_signal_handlers && anyUserHandlers() ) {
1045 debugTrace(DEBUG_sched,
1046 "still deadlocked, waiting for signals...");
1047
1048 awaitUserSignals();
1049
1050 if (signals_pending()) {
1051 startSignalHandlers(cap);
1052 }
1053
1054 // either we have threads to run, or we were interrupted:
1055 ASSERT(!emptyRunQueue(cap) || sched_state >= SCHED_INTERRUPTING);
1056
1057 return;
1058 }
1059 #endif
1060
1061 #if !defined(THREADED_RTS)
1062 /* Probably a real deadlock. Send the current main thread the
1063 * Deadlock exception.
1064 */
1065 if (task->tso) {
1066 switch (task->tso->why_blocked) {
1067 case BlockedOnSTM:
1068 case BlockedOnBlackHole:
1069 case BlockedOnException:
1070 case BlockedOnMVar:
1071 throwToSingleThreaded(cap, task->tso,
1072 (StgClosure *)nonTermination_closure);
1073 return;
1074 default:
1075 barf("deadlock: main thread blocked in a strange way");
1076 }
1077 }
1078 return;
1079 #endif
1080 }
1081 }
1082
1083
1084 /* ----------------------------------------------------------------------------
1085 * Send pending messages (PARALLEL_HASKELL only)
1086 * ------------------------------------------------------------------------- */
1087
1088 #if defined(PARALLEL_HASKELL)
1089 static void
1090 scheduleSendPendingMessages(void)
1091 {
1092
1093 # if defined(PAR) // global Mem.Mgmt., omit for now
1094 if (PendingFetches != END_BF_QUEUE) {
1095 processFetches();
1096 }
1097 # endif
1098
1099 if (RtsFlags.ParFlags.BufferTime) {
1100 // if we use message buffering, we must send away all message
1101 // packets which have become too old...
1102 sendOldBuffers();
1103 }
1104 }
1105 #endif
1106
1107 /* ----------------------------------------------------------------------------
1108 * Activate spark threads (PARALLEL_HASKELL and THREADED_RTS)
1109 * ------------------------------------------------------------------------- */
1110
1111 #if defined(PARALLEL_HASKELL) || defined(THREADED_RTS)
1112 static void
1113 scheduleActivateSpark(Capability *cap)
1114 {
1115 if (anySparks())
1116 {
1117 createSparkThread(cap);
1118 debugTrace(DEBUG_sched, "creating a spark thread");
1119 }
1120 }
1121 #endif // PARALLEL_HASKELL || THREADED_RTS
1122
1123 /* ----------------------------------------------------------------------------
1124 * Get work from a remote node (PARALLEL_HASKELL only)
1125 * ------------------------------------------------------------------------- */
1126
1127 #if defined(PARALLEL_HASKELL)
1128 static rtsBool /* return value used in PARALLEL_HASKELL only */
1129 scheduleGetRemoteWork (Capability *cap STG_UNUSED)
1130 {
1131 #if defined(PARALLEL_HASKELL)
1132 rtsBool receivedFinish = rtsFalse;
1133
1134 // idle() , i.e. send all buffers, wait for work
1135 if (RtsFlags.ParFlags.BufferTime) {
1136 IF_PAR_DEBUG(verbose,
1137 debugBelch("...send all pending data,"));
1138 {
1139 nat i;
1140 for (i=1; i<=nPEs; i++)
1141 sendImmediately(i); // send all messages away immediately
1142 }
1143 }
1144
1145 /* this would be the place for fishing in GUM...
1146
1147 if (no-earlier-fish-around)
1148 sendFish(choosePe());
1149 */
1150
1151 // Eden:just look for incoming messages (blocking receive)
1152 IF_PAR_DEBUG(verbose,
1153 debugBelch("...wait for incoming messages...\n"));
1154 processMessages(cap, &receivedFinish); // blocking receive...
1155
1156
1157 return receivedFinish;
1158 // reenter scheduling look after having received something
1159
1160 #else /* !PARALLEL_HASKELL, i.e. THREADED_RTS */
1161
1162 return rtsFalse; /* return value unused in THREADED_RTS */
1163
1164 #endif /* PARALLEL_HASKELL */
1165 }
1166 #endif // PARALLEL_HASKELL || THREADED_RTS
1167
1168 /* ----------------------------------------------------------------------------
1169 * After running a thread...
1170 * ------------------------------------------------------------------------- */
1171
1172 static void
1173 schedulePostRunThread (Capability *cap, StgTSO *t)
1174 {
1175 // We have to be able to catch transactions that are in an
1176 // infinite loop as a result of seeing an inconsistent view of
1177 // memory, e.g.
1178 //
1179 // atomically $ do
1180 // [a,b] <- mapM readTVar [ta,tb]
1181 // when (a == b) loop
1182 //
1183 // and a is never equal to b given a consistent view of memory.
1184 //
1185 if (t -> trec != NO_TREC && t -> why_blocked == NotBlocked) {
1186 if (!stmValidateNestOfTransactions (t -> trec)) {
1187 debugTrace(DEBUG_sched | DEBUG_stm,
1188 "trec %p found wasting its time", t);
1189
1190 // strip the stack back to the
1191 // ATOMICALLY_FRAME, aborting the (nested)
1192 // transaction, and saving the stack of any
1193 // partially-evaluated thunks on the heap.
1194 throwToSingleThreaded_(cap, t, NULL, rtsTrue);
1195
1196 ASSERT(get_itbl((StgClosure *)t->sp)->type == ATOMICALLY_FRAME);
1197 }
1198 }
1199
1200 /* some statistics gathering in the parallel case */
1201 }
1202
1203 /* -----------------------------------------------------------------------------
1204 * Handle a thread that returned to the scheduler with ThreadHeepOverflow
1205 * -------------------------------------------------------------------------- */
1206
1207 static rtsBool
1208 scheduleHandleHeapOverflow( Capability *cap, StgTSO *t )
1209 {
1210 // did the task ask for a large block?
1211 if (cap->r.rHpAlloc > BLOCK_SIZE) {
1212 // if so, get one and push it on the front of the nursery.
1213 bdescr *bd;
1214 lnat blocks;
1215
1216 blocks = (lnat)BLOCK_ROUND_UP(cap->r.rHpAlloc) / BLOCK_SIZE;
1217
1218 debugTrace(DEBUG_sched,
1219 "--<< thread %ld (%s) stopped: requesting a large block (size %ld)\n",
1220 (long)t->id, whatNext_strs[t->what_next], blocks);
1221
1222 // don't do this if the nursery is (nearly) full, we'll GC first.
1223 if (cap->r.rCurrentNursery->link != NULL ||
1224 cap->r.rNursery->n_blocks == 1) { // paranoia to prevent infinite loop
1225 // if the nursery has only one block.
1226
1227 ACQUIRE_SM_LOCK
1228 bd = allocGroup( blocks );
1229 RELEASE_SM_LOCK
1230 cap->r.rNursery->n_blocks += blocks;
1231
1232 // link the new group into the list
1233 bd->link = cap->r.rCurrentNursery;
1234 bd->u.back = cap->r.rCurrentNursery->u.back;
1235 if (cap->r.rCurrentNursery->u.back != NULL) {
1236 cap->r.rCurrentNursery->u.back->link = bd;
1237 } else {
1238 #if !defined(THREADED_RTS)
1239 ASSERT(g0s0->blocks == cap->r.rCurrentNursery &&
1240 g0s0 == cap->r.rNursery);
1241 #endif
1242 cap->r.rNursery->blocks = bd;
1243 }
1244 cap->r.rCurrentNursery->u.back = bd;
1245
1246 // initialise it as a nursery block. We initialise the
1247 // step, gen_no, and flags field of *every* sub-block in
1248 // this large block, because this is easier than making
1249 // sure that we always find the block head of a large
1250 // block whenever we call Bdescr() (eg. evacuate() and
1251 // isAlive() in the GC would both have to do this, at
1252 // least).
1253 {
1254 bdescr *x;
1255 for (x = bd; x < bd + blocks; x++) {
1256 x->step = cap->r.rNursery;
1257 x->gen_no = 0;
1258 x->flags = 0;
1259 }
1260 }
1261
1262 // This assert can be a killer if the app is doing lots
1263 // of large block allocations.
1264 IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
1265
1266 // now update the nursery to point to the new block
1267 cap->r.rCurrentNursery = bd;
1268
1269 // we might be unlucky and have another thread get on the
1270 // run queue before us and steal the large block, but in that
1271 // case the thread will just end up requesting another large
1272 // block.
1273 pushOnRunQueue(cap,t);
1274 return rtsFalse; /* not actually GC'ing */
1275 }
1276 }
1277
1278 debugTrace(DEBUG_sched,
1279 "--<< thread %ld (%s) stopped: HeapOverflow",
1280 (long)t->id, whatNext_strs[t->what_next]);
1281
1282 if (cap->r.rHpLim == NULL || cap->context_switch) {
1283 // Sometimes we miss a context switch, e.g. when calling
1284 // primitives in a tight loop, MAYBE_GC() doesn't check the
1285 // context switch flag, and we end up waiting for a GC.
1286 // See #1984, and concurrent/should_run/1984
1287 cap->context_switch = 0;
1288 addToRunQueue(cap,t);
1289 } else {
1290 pushOnRunQueue(cap,t);
1291 }
1292 return rtsTrue;
1293 /* actual GC is done at the end of the while loop in schedule() */
1294 }
1295
1296 /* -----------------------------------------------------------------------------
1297 * Handle a thread that returned to the scheduler with ThreadStackOverflow
1298 * -------------------------------------------------------------------------- */
1299
1300 static void
1301 scheduleHandleStackOverflow (Capability *cap, Task *task, StgTSO *t)
1302 {
1303 debugTrace (DEBUG_sched,
1304 "--<< thread %ld (%s) stopped, StackOverflow",
1305 (long)t->id, whatNext_strs[t->what_next]);
1306
1307 /* just adjust the stack for this thread, then pop it back
1308 * on the run queue.
1309 */
1310 {
1311 /* enlarge the stack */
1312 StgTSO *new_t = threadStackOverflow(cap, t);
1313
1314 /* The TSO attached to this Task may have moved, so update the
1315 * pointer to it.
1316 */
1317 if (task->tso == t) {
1318 task->tso = new_t;
1319 }
1320 pushOnRunQueue(cap,new_t);
1321 }
1322 }
1323
1324 /* -----------------------------------------------------------------------------
1325 * Handle a thread that returned to the scheduler with ThreadYielding
1326 * -------------------------------------------------------------------------- */
1327
1328 static rtsBool
1329 scheduleHandleYield( Capability *cap, StgTSO *t, nat prev_what_next )
1330 {
1331 // Reset the context switch flag. We don't do this just before
1332 // running the thread, because that would mean we would lose ticks
1333 // during GC, which can lead to unfair scheduling (a thread hogs
1334 // the CPU because the tick always arrives during GC). This way
1335 // penalises threads that do a lot of allocation, but that seems
1336 // better than the alternative.
1337 cap->context_switch = 0;
1338
1339 /* put the thread back on the run queue. Then, if we're ready to
1340 * GC, check whether this is the last task to stop. If so, wake
1341 * up the GC thread. getThread will block during a GC until the
1342 * GC is finished.
1343 */
1344 #ifdef DEBUG
1345 if (t->what_next != prev_what_next) {
1346 debugTrace(DEBUG_sched,
1347 "--<< thread %ld (%s) stopped to switch evaluators",
1348 (long)t->id, whatNext_strs[t->what_next]);
1349 } else {
1350 debugTrace(DEBUG_sched,
1351 "--<< thread %ld (%s) stopped, yielding",
1352 (long)t->id, whatNext_strs[t->what_next]);
1353 }
1354 #endif
1355
1356 IF_DEBUG(sanity,
1357 //debugBelch("&& Doing sanity check on yielding TSO %ld.", t->id);
1358 checkTSO(t));
1359 ASSERT(t->_link == END_TSO_QUEUE);
1360
1361 // Shortcut if we're just switching evaluators: don't bother
1362 // doing stack squeezing (which can be expensive), just run the
1363 // thread.
1364 if (t->what_next != prev_what_next) {
1365 return rtsTrue;
1366 }
1367
1368 addToRunQueue(cap,t);
1369
1370 return rtsFalse;
1371 }
1372
1373 /* -----------------------------------------------------------------------------
1374 * Handle a thread that returned to the scheduler with ThreadBlocked
1375 * -------------------------------------------------------------------------- */
1376
1377 static void
1378 scheduleHandleThreadBlocked( StgTSO *t
1379 #if !defined(GRAN) && !defined(DEBUG)
1380 STG_UNUSED
1381 #endif
1382 )
1383 {
1384
1385 // We don't need to do anything. The thread is blocked, and it
1386 // has tidied up its stack and placed itself on whatever queue
1387 // it needs to be on.
1388
1389 // ASSERT(t->why_blocked != NotBlocked);
1390 // Not true: for example,
1391 // - in THREADED_RTS, the thread may already have been woken
1392 // up by another Capability. This actually happens: try
1393 // conc023 +RTS -N2.
1394 // - the thread may have woken itself up already, because
1395 // threadPaused() might have raised a blocked throwTo
1396 // exception, see maybePerformBlockedException().
1397
1398 #ifdef DEBUG
1399 if (traceClass(DEBUG_sched)) {
1400 debugTraceBegin("--<< thread %lu (%s) stopped: ",
1401 (unsigned long)t->id, whatNext_strs[t->what_next]);
1402 printThreadBlockage(t);
1403 debugTraceEnd();
1404 }
1405 #endif
1406 }
1407
1408 /* -----------------------------------------------------------------------------
1409 * Handle a thread that returned to the scheduler with ThreadFinished
1410 * -------------------------------------------------------------------------- */
1411
1412 static rtsBool
1413 scheduleHandleThreadFinished (Capability *cap STG_UNUSED, Task *task, StgTSO *t)
1414 {
1415 /* Need to check whether this was a main thread, and if so,
1416 * return with the return value.
1417 *
1418 * We also end up here if the thread kills itself with an
1419 * uncaught exception, see Exception.cmm.
1420 */
1421 debugTrace(DEBUG_sched, "--++ thread %lu (%s) finished",
1422 (unsigned long)t->id, whatNext_strs[t->what_next]);
1423
1424 // blocked exceptions can now complete, even if the thread was in
1425 // blocked mode (see #2910). This unconditionally calls
1426 // lockTSO(), which ensures that we don't miss any threads that
1427 // are engaged in throwTo() with this thread as a target.
1428 awakenBlockedExceptionQueue (cap, t);
1429
1430 //
1431 // Check whether the thread that just completed was a bound
1432 // thread, and if so return with the result.
1433 //
1434 // There is an assumption here that all thread completion goes
1435 // through this point; we need to make sure that if a thread
1436 // ends up in the ThreadKilled state, that it stays on the run
1437 // queue so it can be dealt with here.
1438 //
1439
1440 if (t->bound) {
1441
1442 if (t->bound != task) {
1443 #if !defined(THREADED_RTS)
1444 // Must be a bound thread that is not the topmost one. Leave
1445 // it on the run queue until the stack has unwound to the
1446 // point where we can deal with this. Leaving it on the run
1447 // queue also ensures that the garbage collector knows about
1448 // this thread and its return value (it gets dropped from the
1449 // step->threads list so there's no other way to find it).
1450 appendToRunQueue(cap,t);
1451 return rtsFalse;
1452 #else
1453 // this cannot happen in the threaded RTS, because a
1454 // bound thread can only be run by the appropriate Task.
1455 barf("finished bound thread that isn't mine");
1456 #endif
1457 }
1458
1459 ASSERT(task->tso == t);
1460
1461 if (t->what_next == ThreadComplete) {
1462 if (task->ret) {
1463 // NOTE: return val is tso->sp[1] (see StgStartup.hc)
1464 *(task->ret) = (StgClosure *)task->tso->sp[1];
1465 }
1466 task->stat = Success;
1467 } else {
1468 if (task->ret) {
1469 *(task->ret) = NULL;
1470 }
1471 if (sched_state >= SCHED_INTERRUPTING) {
1472 if (heap_overflow) {
1473 task->stat = HeapExhausted;
1474 } else {
1475 task->stat = Interrupted;
1476 }
1477 } else {
1478 task->stat = Killed;
1479 }
1480 }
1481 #ifdef DEBUG
1482 removeThreadLabel((StgWord)task->tso->id);
1483 #endif
1484 return rtsTrue; // tells schedule() to return
1485 }
1486
1487 return rtsFalse;
1488 }
1489
1490 /* -----------------------------------------------------------------------------
1491 * Perform a heap census
1492 * -------------------------------------------------------------------------- */
1493
1494 static rtsBool
1495 scheduleNeedHeapProfile( rtsBool ready_to_gc STG_UNUSED )
1496 {
1497 // When we have +RTS -i0 and we're heap profiling, do a census at
1498 // every GC. This lets us get repeatable runs for debugging.
1499 if (performHeapProfile ||
1500 (RtsFlags.ProfFlags.profileInterval==0 &&
1501 RtsFlags.ProfFlags.doHeapProfile && ready_to_gc)) {
1502 return rtsTrue;
1503 } else {
1504 return rtsFalse;
1505 }
1506 }
1507
1508 /* -----------------------------------------------------------------------------
1509 * Perform a garbage collection if necessary
1510 * -------------------------------------------------------------------------- */
1511
1512 static Capability *
1513 scheduleDoGC (Capability *cap, Task *task USED_IF_THREADS, rtsBool force_major)
1514 {
1515 rtsBool heap_census;
1516 #ifdef THREADED_RTS
1517 /* extern static volatile StgWord waiting_for_gc;
1518 lives inside capability.c */
1519 rtsBool gc_type, prev_pending_gc;
1520 nat i;
1521 #endif
1522
1523 if (sched_state == SCHED_SHUTTING_DOWN) {
1524 // The final GC has already been done, and the system is
1525 // shutting down. We'll probably deadlock if we try to GC
1526 // now.
1527 return cap;
1528 }
1529
1530 #ifdef THREADED_RTS
1531 if (sched_state < SCHED_INTERRUPTING
1532 && RtsFlags.ParFlags.parGcEnabled
1533 && N >= RtsFlags.ParFlags.parGcGen
1534 && ! oldest_gen->steps[0].mark)
1535 {
1536 gc_type = PENDING_GC_PAR;
1537 } else {
1538 gc_type = PENDING_GC_SEQ;
1539 }
1540
1541 // In order to GC, there must be no threads running Haskell code.
1542 // Therefore, the GC thread needs to hold *all* the capabilities,
1543 // and release them after the GC has completed.
1544 //
1545 // This seems to be the simplest way: previous attempts involved
1546 // making all the threads with capabilities give up their
1547 // capabilities and sleep except for the *last* one, which
1548 // actually did the GC. But it's quite hard to arrange for all
1549 // the other tasks to sleep and stay asleep.
1550 //
1551
1552 /* Other capabilities are prevented from running yet more Haskell
1553 threads if waiting_for_gc is set. Tested inside
1554 yieldCapability() and releaseCapability() in Capability.c */
1555
1556 prev_pending_gc = cas(&waiting_for_gc, 0, gc_type);
1557 if (prev_pending_gc) {
1558 do {
1559 debugTrace(DEBUG_sched, "someone else is trying to GC (%d)...",
1560 prev_pending_gc);
1561 ASSERT(cap);
1562 yieldCapability(&cap,task);
1563 } while (waiting_for_gc);
1564 return cap; // NOTE: task->cap might have changed here
1565 }
1566
1567 setContextSwitches();
1568
1569 // The final shutdown GC is always single-threaded, because it's
1570 // possible that some of the Capabilities have no worker threads.
1571
1572 if (gc_type == PENDING_GC_SEQ)
1573 {
1574 postEvent(cap, EVENT_REQUEST_SEQ_GC, 0, 0);
1575 // single-threaded GC: grab all the capabilities
1576 for (i=0; i < n_capabilities; i++) {
1577 debugTrace(DEBUG_sched, "ready_to_gc, grabbing all the capabilies (%d/%d)", i, n_capabilities);
1578 if (cap != &capabilities[i]) {
1579 Capability *pcap = &capabilities[i];
1580 // we better hope this task doesn't get migrated to
1581 // another Capability while we're waiting for this one.
1582 // It won't, because load balancing happens while we have
1583 // all the Capabilities, but even so it's a slightly
1584 // unsavoury invariant.
1585 task->cap = pcap;
1586 waitForReturnCapability(&pcap, task);
1587 if (pcap != &capabilities[i]) {
1588 barf("scheduleDoGC: got the wrong capability");
1589 }
1590 }
1591 }
1592 }
1593 else
1594 {
1595 // multi-threaded GC: make sure all the Capabilities donate one
1596 // GC thread each.
1597 postEvent(cap, EVENT_REQUEST_PAR_GC, 0, 0);
1598 debugTrace(DEBUG_sched, "ready_to_gc, grabbing GC threads");
1599
1600 waitForGcThreads(cap);
1601 }
1602 #endif
1603
1604 // so this happens periodically:
1605 if (cap) scheduleCheckBlackHoles(cap);
1606
1607 IF_DEBUG(scheduler, printAllThreads());
1608
1609 delete_threads_and_gc:
1610 /*
1611 * We now have all the capabilities; if we're in an interrupting
1612 * state, then we should take the opportunity to delete all the
1613 * threads in the system.
1614 */
1615 if (sched_state == SCHED_INTERRUPTING) {
1616 deleteAllThreads(cap);
1617 sched_state = SCHED_SHUTTING_DOWN;
1618 }
1619
1620 heap_census = scheduleNeedHeapProfile(rtsTrue);
1621
1622 #if defined(THREADED_RTS)
1623 postEvent(cap, EVENT_GC_START, 0, 0);
1624 debugTrace(DEBUG_sched, "doing GC");
1625 // reset waiting_for_gc *before* GC, so that when the GC threads
1626 // emerge they don't immediately re-enter the GC.
1627 waiting_for_gc = 0;
1628 GarbageCollect(force_major || heap_census, gc_type, cap);
1629 #else
1630 GarbageCollect(force_major || heap_census, 0, cap);
1631 #endif
1632 postEvent(cap, EVENT_GC_END, 0, 0);
1633
1634 if (recent_activity == ACTIVITY_INACTIVE && force_major)
1635 {
1636 // We are doing a GC because the system has been idle for a
1637 // timeslice and we need to check for deadlock. Record the
1638 // fact that we've done a GC and turn off the timer signal;
1639 // it will get re-enabled if we run any threads after the GC.
1640 recent_activity = ACTIVITY_DONE_GC;
1641 stopTimer();
1642 }
1643 else
1644 {
1645 // the GC might have taken long enough for the timer to set
1646 // recent_activity = ACTIVITY_INACTIVE, but we aren't
1647 // necessarily deadlocked:
1648 recent_activity = ACTIVITY_YES;
1649 }
1650
1651 #if defined(THREADED_RTS)
1652 if (gc_type == PENDING_GC_PAR)
1653 {
1654 releaseGCThreads(cap);
1655 }
1656 #endif
1657
1658 if (heap_census) {
1659 debugTrace(DEBUG_sched, "performing heap census");
1660 heapCensus();
1661 performHeapProfile = rtsFalse;
1662 }
1663
1664 if (heap_overflow && sched_state < SCHED_INTERRUPTING) {
1665 // GC set the heap_overflow flag, so we should proceed with
1666 // an orderly shutdown now. Ultimately we want the main
1667 // thread to return to its caller with HeapExhausted, at which
1668 // point the caller should call hs_exit(). The first step is
1669 // to delete all the threads.
1670 //
1671 // Another way to do this would be to raise an exception in
1672 // the main thread, which we really should do because it gives
1673 // the program a chance to clean up. But how do we find the
1674 // main thread? It should presumably be the same one that
1675 // gets ^C exceptions, but that's all done on the Haskell side
1676 // (GHC.TopHandler).
1677 sched_state = SCHED_INTERRUPTING;
1678 goto delete_threads_and_gc;
1679 }
1680
1681 #ifdef SPARKBALANCE
1682 /* JB
1683 Once we are all together... this would be the place to balance all
1684 spark pools. No concurrent stealing or adding of new sparks can
1685 occur. Should be defined in Sparks.c. */
1686 balanceSparkPoolsCaps(n_capabilities, capabilities);
1687 #endif
1688
1689 #if defined(THREADED_RTS)
1690 if (gc_type == PENDING_GC_SEQ) {
1691 // release our stash of capabilities.
1692 for (i = 0; i < n_capabilities; i++) {
1693 if (cap != &capabilities[i]) {
1694 task->cap = &capabilities[i];
1695 releaseCapability(&capabilities[i]);
1696 }
1697 }
1698 }
1699 if (cap) {
1700 task->cap = cap;
1701 } else {
1702 task->cap = NULL;
1703 }
1704 #endif
1705
1706 return cap;
1707 }
1708
1709 /* ---------------------------------------------------------------------------
1710 * Singleton fork(). Do not copy any running threads.
1711 * ------------------------------------------------------------------------- */
1712
1713 pid_t
1714 forkProcess(HsStablePtr *entry
1715 #ifndef FORKPROCESS_PRIMOP_SUPPORTED
1716 STG_UNUSED
1717 #endif
1718 )
1719 {
1720 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
1721 Task *task;
1722 pid_t pid;
1723 StgTSO* t,*next;
1724 Capability *cap;
1725 nat s;
1726
1727 #if defined(THREADED_RTS)
1728 if (RtsFlags.ParFlags.nNodes > 1) {
1729 errorBelch("forking not supported with +RTS -N<n> greater than 1");
1730 stg_exit(EXIT_FAILURE);
1731 }
1732 #endif
1733
1734 debugTrace(DEBUG_sched, "forking!");
1735
1736 // ToDo: for SMP, we should probably acquire *all* the capabilities
1737 cap = rts_lock();
1738
1739 // no funny business: hold locks while we fork, otherwise if some
1740 // other thread is holding a lock when the fork happens, the data
1741 // structure protected by the lock will forever be in an
1742 // inconsistent state in the child. See also #1391.
1743 ACQUIRE_LOCK(&sched_mutex);
1744 ACQUIRE_LOCK(&cap->lock);
1745 ACQUIRE_LOCK(&cap->running_task->lock);
1746
1747 pid = fork();
1748
1749 if (pid) { // parent
1750
1751 RELEASE_LOCK(&sched_mutex);
1752 RELEASE_LOCK(&cap->lock);
1753 RELEASE_LOCK(&cap->running_task->lock);
1754
1755 // just return the pid
1756 rts_unlock(cap);
1757 return pid;
1758
1759 } else { // child
1760
1761 #if defined(THREADED_RTS)
1762 initMutex(&sched_mutex);
1763 initMutex(&cap->lock);
1764 initMutex(&cap->running_task->lock);
1765 #endif
1766
1767 // Now, all OS threads except the thread that forked are
1768 // stopped. We need to stop all Haskell threads, including
1769 // those involved in foreign calls. Also we need to delete
1770 // all Tasks, because they correspond to OS threads that are
1771 // now gone.
1772
1773 for (s = 0; s < total_steps; s++) {
1774 for (t = all_steps[s].threads; t != END_TSO_QUEUE; t = next) {
1775 if (t->what_next == ThreadRelocated) {
1776 next = t->_link;
1777 } else {
1778 next = t->global_link;
1779 // don't allow threads to catch the ThreadKilled
1780 // exception, but we do want to raiseAsync() because these
1781 // threads may be evaluating thunks that we need later.
1782 deleteThread_(cap,t);
1783 }
1784 }
1785 }
1786
1787 // Empty the run queue. It seems tempting to let all the
1788 // killed threads stay on the run queue as zombies to be
1789 // cleaned up later, but some of them correspond to bound
1790 // threads for which the corresponding Task does not exist.
1791 cap->run_queue_hd = END_TSO_QUEUE;
1792 cap->run_queue_tl = END_TSO_QUEUE;
1793
1794 // Any suspended C-calling Tasks are no more, their OS threads
1795 // don't exist now:
1796 cap->suspended_ccalling_tasks = NULL;
1797
1798 // Empty the threads lists. Otherwise, the garbage
1799 // collector may attempt to resurrect some of these threads.
1800 for (s = 0; s < total_steps; s++) {
1801 all_steps[s].threads = END_TSO_QUEUE;
1802 }
1803
1804 // Wipe the task list, except the current Task.
1805 ACQUIRE_LOCK(&sched_mutex);
1806 for (task = all_tasks; task != NULL; task=task->all_link) {
1807 if (task != cap->running_task) {
1808 #if defined(THREADED_RTS)
1809 initMutex(&task->lock); // see #1391
1810 #endif
1811 discardTask(task);
1812 }
1813 }
1814 RELEASE_LOCK(&sched_mutex);
1815
1816 #if defined(THREADED_RTS)
1817 // Wipe our spare workers list, they no longer exist. New
1818 // workers will be created if necessary.
1819 cap->spare_workers = NULL;
1820 cap->returning_tasks_hd = NULL;
1821 cap->returning_tasks_tl = NULL;
1822 #endif
1823
1824 // On Unix, all timers are reset in the child, so we need to start
1825 // the timer again.
1826 initTimer();
1827 startTimer();
1828
1829 cap = rts_evalStableIO(cap, entry, NULL); // run the action
1830 rts_checkSchedStatus("forkProcess",cap);
1831
1832 rts_unlock(cap);
1833 hs_exit(); // clean up and exit
1834 stg_exit(EXIT_SUCCESS);
1835 }
1836 #else /* !FORKPROCESS_PRIMOP_SUPPORTED */
1837 barf("forkProcess#: primop not supported on this platform, sorry!\n");
1838 return -1;
1839 #endif
1840 }
1841
1842 /* ---------------------------------------------------------------------------
1843 * Delete all the threads in the system
1844 * ------------------------------------------------------------------------- */
1845
1846 static void
1847 deleteAllThreads ( Capability *cap )
1848 {
1849 // NOTE: only safe to call if we own all capabilities.
1850
1851 StgTSO* t, *next;
1852 nat s;
1853
1854 debugTrace(DEBUG_sched,"deleting all threads");
1855 for (s = 0; s < total_steps; s++) {
1856 for (t = all_steps[s].threads; t != END_TSO_QUEUE; t = next) {
1857 if (t->what_next == ThreadRelocated) {
1858 next = t->_link;
1859 } else {
1860 next = t->global_link;
1861 deleteThread(cap,t);
1862 }
1863 }
1864 }
1865
1866 // The run queue now contains a bunch of ThreadKilled threads. We
1867 // must not throw these away: the main thread(s) will be in there
1868 // somewhere, and the main scheduler loop has to deal with it.
1869 // Also, the run queue is the only thing keeping these threads from
1870 // being GC'd, and we don't want the "main thread has been GC'd" panic.
1871
1872 #if !defined(THREADED_RTS)
1873 ASSERT(blocked_queue_hd == END_TSO_QUEUE);
1874 ASSERT(sleeping_queue == END_TSO_QUEUE);
1875 #endif
1876 }
1877
1878 /* -----------------------------------------------------------------------------
1879 Managing the suspended_ccalling_tasks list.
1880 Locks required: sched_mutex
1881 -------------------------------------------------------------------------- */
1882
1883 STATIC_INLINE void
1884 suspendTask (Capability *cap, Task *task)
1885 {
1886 ASSERT(task->next == NULL && task->prev == NULL);
1887 task->next = cap->suspended_ccalling_tasks;
1888 task->prev = NULL;
1889 if (cap->suspended_ccalling_tasks) {
1890 cap->suspended_ccalling_tasks->prev = task;
1891 }
1892 cap->suspended_ccalling_tasks = task;
1893 }
1894
1895 STATIC_INLINE void
1896 recoverSuspendedTask (Capability *cap, Task *task)
1897 {
1898 if (task->prev) {
1899 task->prev->next = task->next;
1900 } else {
1901 ASSERT(cap->suspended_ccalling_tasks == task);
1902 cap->suspended_ccalling_tasks = task->next;
1903 }
1904 if (task->next) {
1905 task->next->prev = task->prev;
1906 }
1907 task->next = task->prev = NULL;
1908 }
1909
1910 /* ---------------------------------------------------------------------------
1911 * Suspending & resuming Haskell threads.
1912 *
1913 * When making a "safe" call to C (aka _ccall_GC), the task gives back
1914 * its capability before calling the C function. This allows another
1915 * task to pick up the capability and carry on running Haskell
1916 * threads. It also means that if the C call blocks, it won't lock
1917 * the whole system.
1918 *
1919 * The Haskell thread making the C call is put to sleep for the
1920 * duration of the call, on the susepended_ccalling_threads queue. We
1921 * give out a token to the task, which it can use to resume the thread
1922 * on return from the C function.
1923 * ------------------------------------------------------------------------- */
1924
1925 void *
1926 suspendThread (StgRegTable *reg)
1927 {
1928 Capability *cap;
1929 int saved_errno;
1930 StgTSO *tso;
1931 Task *task;
1932 #if mingw32_HOST_OS
1933 StgWord32 saved_winerror;
1934 #endif
1935
1936 saved_errno = errno;
1937 #if mingw32_HOST_OS
1938 saved_winerror = GetLastError();
1939 #endif
1940
1941 /* assume that *reg is a pointer to the StgRegTable part of a Capability.
1942 */
1943 cap = regTableToCapability(reg);
1944
1945 task = cap->running_task;
1946 tso = cap->r.rCurrentTSO;
1947
1948 postEvent(cap, EVENT_STOP_THREAD, tso->id, THREAD_SUSPENDED_FOREIGN_CALL);
1949 debugTrace(DEBUG_sched,
1950 "thread %lu did a safe foreign call",
1951 (unsigned long)cap->r.rCurrentTSO->id);
1952
1953 // XXX this might not be necessary --SDM
1954 tso->what_next = ThreadRunGHC;
1955
1956 threadPaused(cap,tso);
1957
1958 if ((tso->flags & TSO_BLOCKEX) == 0) {
1959 tso->why_blocked = BlockedOnCCall;
1960 tso->flags |= TSO_BLOCKEX;
1961 tso->flags &= ~TSO_INTERRUPTIBLE;
1962 } else {
1963 tso->why_blocked = BlockedOnCCall_NoUnblockExc;
1964 }
1965
1966 // Hand back capability
1967 task->suspended_tso = tso;
1968
1969 ACQUIRE_LOCK(&cap->lock);
1970
1971 suspendTask(cap,task);
1972 cap->in_haskell = rtsFalse;
1973 releaseCapability_(cap,rtsFalse);
1974
1975 RELEASE_LOCK(&cap->lock);
1976
1977 #if defined(THREADED_RTS)
1978 /* Preparing to leave the RTS, so ensure there's a native thread/task
1979 waiting to take over.
1980 */
1981 debugTrace(DEBUG_sched, "thread %lu: leaving RTS", (unsigned long)tso->id);
1982 #endif
1983
1984 errno = saved_errno;
1985 #if mingw32_HOST_OS
1986 SetLastError(saved_winerror);
1987 #endif
1988 return task;
1989 }
1990
1991 StgRegTable *
1992 resumeThread (void *task_)
1993 {
1994 StgTSO *tso;
1995 Capability *cap;
1996 Task *task = task_;
1997 int saved_errno;
1998 #if mingw32_HOST_OS
1999 StgWord32 saved_winerror;
2000 #endif
2001
2002 saved_errno = errno;
2003 #if mingw32_HOST_OS
2004 saved_winerror = GetLastError();
2005 #endif
2006
2007 cap = task->cap;
2008 // Wait for permission to re-enter the RTS with the result.
2009 waitForReturnCapability(&cap,task);
2010 // we might be on a different capability now... but if so, our
2011 // entry on the suspended_ccalling_tasks list will also have been
2012 // migrated.
2013
2014 // Remove the thread from the suspended list
2015 recoverSuspendedTask(cap,task);
2016
2017 tso = task->suspended_tso;
2018 task->suspended_tso = NULL;
2019 tso->_link = END_TSO_QUEUE; // no write barrier reqd
2020
2021 postEvent(cap, EVENT_RUN_THREAD, tso->id, 0);
2022 debugTrace(DEBUG_sched, "thread %lu: re-entering RTS", (unsigned long)tso->id);
2023
2024 if (tso->why_blocked == BlockedOnCCall) {
2025 // avoid locking the TSO if we don't have to
2026 if (tso->blocked_exceptions != END_TSO_QUEUE) {
2027 awakenBlockedExceptionQueue(cap,tso);
2028 }
2029 tso->flags &= ~(TSO_BLOCKEX | TSO_INTERRUPTIBLE);
2030 }
2031
2032 /* Reset blocking status */
2033 tso->why_blocked = NotBlocked;
2034
2035 cap->r.rCurrentTSO = tso;
2036 cap->in_haskell = rtsTrue;
2037 errno = saved_errno;
2038 #if mingw32_HOST_OS
2039 SetLastError(saved_winerror);
2040 #endif
2041
2042 /* We might have GC'd, mark the TSO dirty again */
2043 dirty_TSO(cap,tso);
2044
2045 IF_DEBUG(sanity, checkTSO(tso));
2046
2047 return &cap->r;
2048 }
2049
2050 /* ---------------------------------------------------------------------------
2051 * scheduleThread()
2052 *
2053 * scheduleThread puts a thread on the end of the runnable queue.
2054 * This will usually be done immediately after a thread is created.
2055 * The caller of scheduleThread must create the thread using e.g.
2056 * createThread and push an appropriate closure
2057 * on this thread's stack before the scheduler is invoked.
2058 * ------------------------------------------------------------------------ */
2059
2060 void
2061 scheduleThread(Capability *cap, StgTSO *tso)
2062 {
2063 // The thread goes at the *end* of the run-queue, to avoid possible
2064 // starvation of any threads already on the queue.
2065 appendToRunQueue(cap,tso);
2066 }
2067
2068 void
2069 scheduleThreadOn(Capability *cap, StgWord cpu USED_IF_THREADS, StgTSO *tso)
2070 {
2071 #if defined(THREADED_RTS)
2072 tso->flags |= TSO_LOCKED; // we requested explicit affinity; don't
2073 // move this thread from now on.
2074 cpu %= RtsFlags.ParFlags.nNodes;
2075 if (cpu == cap->no) {
2076 appendToRunQueue(cap,tso);
2077 } else {
2078 postEvent (cap, EVENT_MIGRATE_THREAD, tso->id, capabilities[cpu].no);
2079 wakeupThreadOnCapability(cap, &capabilities[cpu], tso);
2080 }
2081 #else
2082 appendToRunQueue(cap,tso);
2083 #endif
2084 }
2085
2086 Capability *
2087 scheduleWaitThread (StgTSO* tso, /*[out]*/HaskellObj* ret, Capability *cap)
2088 {
2089 Task *task;
2090
2091 // We already created/initialised the Task
2092 task = cap->running_task;
2093
2094 // This TSO is now a bound thread; make the Task and TSO
2095 // point to each other.
2096 tso->bound = task;
2097 tso->cap = cap;
2098
2099 task->tso = tso;
2100 task->ret = ret;
2101 task->stat = NoStatus;
2102
2103 appendToRunQueue(cap,tso);
2104
2105 debugTrace(DEBUG_sched, "new bound thread (%lu)", (unsigned long)tso->id);
2106
2107 cap = schedule(cap,task);
2108
2109 ASSERT(task->stat != NoStatus);
2110 ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
2111
2112 debugTrace(DEBUG_sched, "bound thread (%lu) finished", (unsigned long)task->tso->id);
2113 return cap;
2114 }
2115
2116 /* ----------------------------------------------------------------------------
2117 * Starting Tasks
2118 * ------------------------------------------------------------------------- */
2119
2120 #if defined(THREADED_RTS)
2121 void OSThreadProcAttr
2122 workerStart(Task *task)
2123 {
2124 Capability *cap;
2125
2126 // See startWorkerTask().
2127 ACQUIRE_LOCK(&task->lock);
2128 cap = task->cap;
2129 RELEASE_LOCK(&task->lock);
2130
2131 if (RtsFlags.ParFlags.setAffinity) {
2132 setThreadAffinity(cap->no, n_capabilities);
2133 }
2134
2135 // set the thread-local pointer to the Task:
2136 taskEnter(task);
2137
2138 // schedule() runs without a lock.
2139 cap = schedule(cap,task);
2140
2141 // On exit from schedule(), we have a Capability, but possibly not
2142 // the same one we started with.
2143
2144 // During shutdown, the requirement is that after all the
2145 // Capabilities are shut down, all workers that are shutting down
2146 // have finished workerTaskStop(). This is why we hold on to
2147 // cap->lock until we've finished workerTaskStop() below.
2148 //
2149 // There may be workers still involved in foreign calls; those
2150 // will just block in waitForReturnCapability() because the
2151 // Capability has been shut down.
2152 //
2153 ACQUIRE_LOCK(&cap->lock);
2154 releaseCapability_(cap,rtsFalse);
2155 workerTaskStop(task);
2156 RELEASE_LOCK(&cap->lock);
2157 }
2158 #endif
2159
2160 /* ---------------------------------------------------------------------------
2161 * initScheduler()
2162 *
2163 * Initialise the scheduler. This resets all the queues - if the
2164 * queues contained any threads, they'll be garbage collected at the
2165 * next pass.
2166 *
2167 * ------------------------------------------------------------------------ */
2168
2169 void
2170 initScheduler(void)
2171 {
2172 #if !defined(THREADED_RTS)
2173 blocked_queue_hd = END_TSO_QUEUE;
2174 blocked_queue_tl = END_TSO_QUEUE;
2175 sleeping_queue = END_TSO_QUEUE;
2176 #endif
2177
2178 blackhole_queue = END_TSO_QUEUE;
2179
2180 sched_state = SCHED_RUNNING;
2181 recent_activity = ACTIVITY_YES;
2182
2183 #if defined(THREADED_RTS)
2184 /* Initialise the mutex and condition variables used by
2185 * the scheduler. */
2186 initMutex(&sched_mutex);
2187 #endif
2188
2189 ACQUIRE_LOCK(&sched_mutex);
2190
2191 /* A capability holds the state a native thread needs in
2192 * order to execute STG code. At least one capability is
2193 * floating around (only THREADED_RTS builds have more than one).
2194 */
2195 initCapabilities();
2196
2197 initTaskManager();
2198
2199 #if defined(THREADED_RTS) || defined(PARALLEL_HASKELL)
2200 initSparkPools();
2201 #endif
2202
2203 #if defined(THREADED_RTS)
2204 /*
2205 * Eagerly start one worker to run each Capability, except for
2206 * Capability 0. The idea is that we're probably going to start a
2207 * bound thread on Capability 0 pretty soon, so we don't want a
2208 * worker task hogging it.
2209 */
2210 {
2211 nat i;
2212 Capability *cap;
2213 for (i = 1; i < n_capabilities; i++) {
2214 cap = &capabilities[i];
2215 ACQUIRE_LOCK(&cap->lock);
2216 startWorkerTask(cap, workerStart);
2217 RELEASE_LOCK(&cap->lock);
2218 }
2219 }
2220 #endif
2221
2222 RELEASE_LOCK(&sched_mutex);
2223 }
2224
2225 void
2226 exitScheduler(
2227 rtsBool wait_foreign
2228 #if !defined(THREADED_RTS)
2229 __attribute__((unused))
2230 #endif
2231 )
2232 /* see Capability.c, shutdownCapability() */
2233 {
2234 Task *task = NULL;
2235
2236 ACQUIRE_LOCK(&sched_mutex);
2237 task = newBoundTask();
2238 RELEASE_LOCK(&sched_mutex);
2239
2240 // If we haven't killed all the threads yet, do it now.
2241 if (sched_state < SCHED_SHUTTING_DOWN) {
2242 sched_state = SCHED_INTERRUPTING;
2243 waitForReturnCapability(&task->cap,task);
2244 scheduleDoGC(task->cap,task,rtsFalse);
2245 releaseCapability(task->cap);
2246 }
2247 sched_state = SCHED_SHUTTING_DOWN;
2248
2249 #if defined(THREADED_RTS)
2250 {
2251 nat i;
2252
2253 for (i = 0; i < n_capabilities; i++) {
2254 shutdownCapability(&capabilities[i], task, wait_foreign);
2255 }
2256 boundTaskExiting(task);
2257 }
2258 #endif
2259 }
2260
2261 void
2262 freeScheduler( void )
2263 {
2264 nat still_running;
2265
2266 ACQUIRE_LOCK(&sched_mutex);
2267 still_running = freeTaskManager();
2268 // We can only free the Capabilities if there are no Tasks still
2269 // running. We might have a Task about to return from a foreign
2270 // call into waitForReturnCapability(), for example (actually,
2271 // this should be the *only* thing that a still-running Task can
2272 // do at this point, and it will block waiting for the
2273 // Capability).
2274 if (still_running == 0) {
2275 freeCapabilities();
2276 if (n_capabilities != 1) {
2277 stgFree(capabilities);
2278 }
2279 }
2280 RELEASE_LOCK(&sched_mutex);
2281 #if defined(THREADED_RTS)
2282 closeMutex(&sched_mutex);
2283 #endif
2284 }
2285
2286 /* -----------------------------------------------------------------------------
2287 performGC
2288
2289 This is the interface to the garbage collector from Haskell land.
2290 We provide this so that external C code can allocate and garbage
2291 collect when called from Haskell via _ccall_GC.
2292 -------------------------------------------------------------------------- */
2293
2294 static void
2295 performGC_(rtsBool force_major)
2296 {
2297 Task *task;
2298
2299 // We must grab a new Task here, because the existing Task may be
2300 // associated with a particular Capability, and chained onto the
2301 // suspended_ccalling_tasks queue.
2302 ACQUIRE_LOCK(&sched_mutex);
2303 task = newBoundTask();
2304 RELEASE_LOCK(&sched_mutex);
2305
2306 waitForReturnCapability(&task->cap,task);
2307 scheduleDoGC(task->cap,task,force_major);
2308 releaseCapability(task->cap);
2309 boundTaskExiting(task);
2310 }
2311
2312 void
2313 performGC(void)
2314 {
2315 performGC_(rtsFalse);
2316 }
2317
2318 void
2319 performMajorGC(void)
2320 {
2321 performGC_(rtsTrue);
2322 }
2323
2324 /* -----------------------------------------------------------------------------
2325 Stack overflow
2326
2327 If the thread has reached its maximum stack size, then raise the
2328 StackOverflow exception in the offending thread. Otherwise
2329 relocate the TSO into a larger chunk of memory and adjust its stack
2330 size appropriately.
2331 -------------------------------------------------------------------------- */
2332
2333 static StgTSO *
2334 threadStackOverflow(Capability *cap, StgTSO *tso)
2335 {
2336 nat new_stack_size, stack_words;
2337 lnat new_tso_size;
2338 StgPtr new_sp;
2339 StgTSO *dest;
2340
2341 IF_DEBUG(sanity,checkTSO(tso));
2342
2343 // don't allow throwTo() to modify the blocked_exceptions queue
2344 // while we are moving the TSO:
2345 lockClosure((StgClosure *)tso);
2346
2347 if (tso->stack_size >= tso->max_stack_size && !(tso->flags & TSO_BLOCKEX)) {
2348 // NB. never raise a StackOverflow exception if the thread is
2349 // inside Control.Exceptino.block. It is impractical to protect
2350 // against stack overflow exceptions, since virtually anything
2351 // can raise one (even 'catch'), so this is the only sensible
2352 // thing to do here. See bug #767.
2353
2354 debugTrace(DEBUG_gc,
2355 "threadStackOverflow of TSO %ld (%p): stack too large (now %ld; max is %ld)",
2356 (long)tso->id, tso, (long)tso->stack_size, (long)tso->max_stack_size);
2357 IF_DEBUG(gc,
2358 /* If we're debugging, just print out the top of the stack */
2359 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2360 tso->sp+64)));
2361
2362 // Send this thread the StackOverflow exception
2363 unlockTSO(tso);
2364 throwToSingleThreaded(cap, tso, (StgClosure *)stackOverflow_closure);
2365 return tso;
2366 }
2367
2368 /* Try to double the current stack size. If that takes us over the
2369 * maximum stack size for this thread, then use the maximum instead
2370 * (that is, unless we're already at or over the max size and we
2371 * can't raise the StackOverflow exception (see above), in which
2372 * case just double the size). Finally round up so the TSO ends up as
2373 * a whole number of blocks.
2374 */
2375 if (tso->stack_size >= tso->max_stack_size) {
2376 new_stack_size = tso->stack_size * 2;
2377 } else {
2378 new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
2379 }
2380 new_tso_size = (lnat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) +
2381 TSO_STRUCT_SIZE)/sizeof(W_);
2382 new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */
2383 new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
2384
2385 debugTrace(DEBUG_sched,
2386 "increasing stack size from %ld words to %d.",
2387 (long)tso->stack_size, new_stack_size);
2388
2389 dest = (StgTSO *)allocateLocal(cap,new_tso_size);
2390 TICK_ALLOC_TSO(new_stack_size,0);
2391
2392 /* copy the TSO block and the old stack into the new area */
2393 memcpy(dest,tso,TSO_STRUCT_SIZE);
2394 stack_words = tso->stack + tso->stack_size - tso->sp;
2395 new_sp = (P_)dest + new_tso_size - stack_words;
2396 memcpy(new_sp, tso->sp, stack_words * sizeof(W_));
2397
2398 /* relocate the stack pointers... */
2399 dest->sp = new_sp;
2400 dest->stack_size = new_stack_size;
2401
2402 /* Mark the old TSO as relocated. We have to check for relocated
2403 * TSOs in the garbage collector and any primops that deal with TSOs.
2404 *
2405 * It's important to set the sp value to just beyond the end
2406 * of the stack, so we don't attempt to scavenge any part of the
2407 * dead TSO's stack.
2408 */
2409 tso->what_next = ThreadRelocated;
2410 setTSOLink(cap,tso,dest);
2411 tso->sp = (P_)&(tso->stack[tso->stack_size]);
2412 tso->why_blocked = NotBlocked;
2413
2414 IF_PAR_DEBUG(verbose,
2415 debugBelch("@@ threadStackOverflow of TSO %d (now at %p): stack size increased to %ld\n",
2416 tso->id, tso, tso->stack_size);
2417 /* If we're debugging, just print out the top of the stack */
2418 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2419 tso->sp+64)));
2420
2421 unlockTSO(dest);
2422 unlockTSO(tso);
2423
2424 IF_DEBUG(sanity,checkTSO(dest));
2425 #if 0
2426 IF_DEBUG(scheduler,printTSO(dest));
2427 #endif
2428
2429 return dest;
2430 }
2431
2432 static StgTSO *
2433 threadStackUnderflow (Task *task STG_UNUSED, StgTSO *tso)
2434 {
2435 bdescr *bd, *new_bd;
2436 lnat free_w, tso_size_w;
2437 StgTSO *new_tso;
2438
2439 tso_size_w = tso_sizeW(tso);
2440
2441 if (tso_size_w < MBLOCK_SIZE_W ||
2442 (nat)(tso->stack + tso->stack_size - tso->sp) > tso->stack_size / 4)
2443 {
2444 return tso;
2445 }
2446
2447 // don't allow throwTo() to modify the blocked_exceptions queue
2448 // while we are moving the TSO:
2449 lockClosure((StgClosure *)tso);
2450
2451 // this is the number of words we'll free
2452 free_w = round_to_mblocks(tso_size_w/2);
2453
2454 bd = Bdescr((StgPtr)tso);
2455 new_bd = splitLargeBlock(bd, free_w / BLOCK_SIZE_W);
2456 bd->free = bd->start + TSO_STRUCT_SIZEW;
2457
2458 new_tso = (StgTSO *)new_bd->start;
2459 memcpy(new_tso,tso,TSO_STRUCT_SIZE);
2460 new_tso->stack_size = new_bd->free - new_tso->stack;
2461
2462 debugTrace(DEBUG_sched, "thread %ld: reducing TSO size from %lu words to %lu",
2463 (long)tso->id, tso_size_w, tso_sizeW(new_tso));
2464
2465 tso->what_next = ThreadRelocated;
2466 tso->_link = new_tso; // no write barrier reqd: same generation
2467
2468 // The TSO attached to this Task may have moved, so update the
2469 // pointer to it.
2470 if (task->tso == tso) {
2471 task->tso = new_tso;
2472 }
2473
2474 unlockTSO(new_tso);
2475 unlockTSO(tso);
2476
2477 IF_DEBUG(sanity,checkTSO(new_tso));
2478
2479 return new_tso;
2480 }
2481
2482 /* ---------------------------------------------------------------------------
2483 Interrupt execution
2484 - usually called inside a signal handler so it mustn't do anything fancy.
2485 ------------------------------------------------------------------------ */
2486
2487 void
2488 interruptStgRts(void)
2489 {
2490 sched_state = SCHED_INTERRUPTING;
2491 setContextSwitches();
2492 wakeUpRts();
2493 }
2494
2495 /* -----------------------------------------------------------------------------
2496 Wake up the RTS
2497
2498 This function causes at least one OS thread to wake up and run the
2499 scheduler loop. It is invoked when the RTS might be deadlocked, or
2500 an external event has arrived that may need servicing (eg. a
2501 keyboard interrupt).
2502
2503 In the single-threaded RTS we don't do anything here; we only have
2504 one thread anyway, and the event that caused us to want to wake up
2505 will have interrupted any blocking system call in progress anyway.
2506 -------------------------------------------------------------------------- */
2507
2508 void
2509 wakeUpRts(void)
2510 {
2511 #if defined(THREADED_RTS)
2512 // This forces the IO Manager thread to wakeup, which will
2513 // in turn ensure that some OS thread wakes up and runs the
2514 // scheduler loop, which will cause a GC and deadlock check.
2515 ioManagerWakeup();
2516 #endif
2517 }
2518
2519 /* -----------------------------------------------------------------------------
2520 * checkBlackHoles()
2521 *
2522 * Check the blackhole_queue for threads that can be woken up. We do
2523 * this periodically: before every GC, and whenever the run queue is
2524 * empty.
2525 *
2526 * An elegant solution might be to just wake up all the blocked
2527 * threads with awakenBlockedQueue occasionally: they'll go back to
2528 * sleep again if the object is still a BLACKHOLE. Unfortunately this
2529 * doesn't give us a way to tell whether we've actually managed to
2530 * wake up any threads, so we would be busy-waiting.
2531 *
2532 * -------------------------------------------------------------------------- */
2533
2534 static rtsBool
2535 checkBlackHoles (Capability *cap)
2536 {
2537 StgTSO **prev, *t;
2538 rtsBool any_woke_up = rtsFalse;
2539 StgHalfWord type;
2540
2541 // blackhole_queue is global:
2542 ASSERT_LOCK_HELD(&sched_mutex);
2543
2544 debugTrace(DEBUG_sched, "checking threads blocked on black holes");
2545
2546 // ASSUMES: sched_mutex
2547 prev = &blackhole_queue;
2548 t = blackhole_queue;
2549 while (t != END_TSO_QUEUE) {
2550 if (t->what_next == ThreadRelocated) {
2551 t = t->_link;
2552 continue;
2553 }
2554 ASSERT(t->why_blocked == BlockedOnBlackHole);
2555 type = get_itbl(UNTAG_CLOSURE(t->block_info.closure))->type;
2556 if (type != BLACKHOLE && type != CAF_BLACKHOLE) {
2557 IF_DEBUG(sanity,checkTSO(t));
2558 t = unblockOne(cap, t);
2559 *prev = t;
2560 any_woke_up = rtsTrue;
2561 } else {
2562 prev = &t->_link;
2563 t = t->_link;
2564 }
2565 }
2566
2567 return any_woke_up;
2568 }
2569
2570 /* -----------------------------------------------------------------------------
2571 Deleting threads
2572
2573 This is used for interruption (^C) and forking, and corresponds to
2574 raising an exception but without letting the thread catch the
2575 exception.
2576 -------------------------------------------------------------------------- */
2577
2578 static void
2579 deleteThread (Capability *cap, StgTSO *tso)
2580 {
2581 // NOTE: must only be called on a TSO that we have exclusive
2582 // access to, because we will call throwToSingleThreaded() below.
2583 // The TSO must be on the run queue of the Capability we own, or
2584 // we must own all Capabilities.
2585
2586 if (tso->why_blocked != BlockedOnCCall &&
2587 tso->why_blocked != BlockedOnCCall_NoUnblockExc) {
2588 throwToSingleThreaded(cap,tso,NULL);
2589 }
2590 }
2591
2592 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
2593 static void
2594 deleteThread_(Capability *cap, StgTSO *tso)
2595 { // for forkProcess only:
2596 // like deleteThread(), but we delete threads in foreign calls, too.
2597
2598 if (tso->why_blocked == BlockedOnCCall ||
2599 tso->why_blocked == BlockedOnCCall_NoUnblockExc) {
2600 unblockOne(cap,tso);
2601 tso->what_next = ThreadKilled;
2602 } else {
2603 deleteThread(cap,tso);
2604 }
2605 }
2606 #endif
2607
2608 /* -----------------------------------------------------------------------------
2609 raiseExceptionHelper
2610
2611 This function is called by the raise# primitve, just so that we can
2612 move some of the tricky bits of raising an exception from C-- into
2613 C. Who knows, it might be a useful re-useable thing here too.
2614 -------------------------------------------------------------------------- */
2615
2616 StgWord
2617 raiseExceptionHelper (StgRegTable *reg, StgTSO *tso, StgClosure *exception)
2618 {
2619 Capability *cap = regTableToCapability(reg);
2620 StgThunk *raise_closure = NULL;
2621 StgPtr p, next;
2622 StgRetInfoTable *info;
2623 //
2624 // This closure represents the expression 'raise# E' where E
2625 // is the exception raise. It is used to overwrite all the
2626 // thunks which are currently under evaluataion.
2627 //
2628
2629 // OLD COMMENT (we don't have MIN_UPD_SIZE now):
2630 // LDV profiling: stg_raise_info has THUNK as its closure
2631 // type. Since a THUNK takes at least MIN_UPD_SIZE words in its
2632 // payload, MIN_UPD_SIZE is more approprate than 1. It seems that
2633 // 1 does not cause any problem unless profiling is performed.
2634 // However, when LDV profiling goes on, we need to linearly scan
2635 // small object pool, where raise_closure is stored, so we should
2636 // use MIN_UPD_SIZE.
2637 //
2638 // raise_closure = (StgClosure *)RET_STGCALL1(P_,allocate,
2639 // sizeofW(StgClosure)+1);
2640 //
2641
2642 //
2643 // Walk up the stack, looking for the catch frame. On the way,
2644 // we update any closures pointed to from update frames with the
2645 // raise closure that we just built.
2646 //
2647 p = tso->sp;
2648 while(1) {
2649 info = get_ret_itbl((StgClosure *)p);
2650 next = p + stack_frame_sizeW((StgClosure *)p);
2651 switch (info->i.type) {
2652
2653 case UPDATE_FRAME:
2654 // Only create raise_closure if we need to.
2655 if (raise_closure == NULL) {
2656 raise_closure =
2657 (StgThunk *)allocateLocal(cap,sizeofW(StgThunk)+1);
2658 SET_HDR(raise_closure, &stg_raise_info, CCCS);
2659 raise_closure->payload[0] = exception;
2660 }
2661 UPD_IND(((StgUpdateFrame *)p)->updatee,(StgClosure *)raise_closure);
2662 p = next;
2663 continue;
2664
2665 case ATOMICALLY_FRAME:
2666 debugTrace(DEBUG_stm, "found ATOMICALLY_FRAME at %p", p);
2667 tso->sp = p;
2668 return ATOMICALLY_FRAME;
2669
2670 case CATCH_FRAME:
2671 tso->sp = p;
2672 return CATCH_FRAME;
2673
2674 case CATCH_STM_FRAME:
2675 debugTrace(DEBUG_stm, "found CATCH_STM_FRAME at %p", p);
2676 tso->sp = p;
2677 return CATCH_STM_FRAME;
2678
2679 case STOP_FRAME:
2680 tso->sp = p;
2681 return STOP_FRAME;
2682
2683 case CATCH_RETRY_FRAME:
2684 default:
2685 p = next;
2686 continue;
2687 }
2688 }
2689 }
2690
2691
2692 /* -----------------------------------------------------------------------------
2693 findRetryFrameHelper
2694
2695 This function is called by the retry# primitive. It traverses the stack
2696 leaving tso->sp referring to the frame which should handle the retry.
2697
2698 This should either be a CATCH_RETRY_FRAME (if the retry# is within an orElse#)
2699 or should be a ATOMICALLY_FRAME (if the retry# reaches the top level).
2700
2701 We skip CATCH_STM_FRAMEs (aborting and rolling back the nested tx that they
2702 create) because retries are not considered to be exceptions, despite the
2703 similar implementation.
2704
2705 We should not expect to see CATCH_FRAME or STOP_FRAME because those should
2706 not be created within memory transactions.
2707 -------------------------------------------------------------------------- */
2708
2709 StgWord
2710 findRetryFrameHelper (StgTSO *tso)
2711 {
2712 StgPtr p, next;
2713 StgRetInfoTable *info;
2714
2715 p = tso -> sp;
2716 while (1) {
2717 info = get_ret_itbl((StgClosure *)p);
2718 next = p + stack_frame_sizeW((StgClosure *)p);
2719 switch (info->i.type) {
2720
2721 case ATOMICALLY_FRAME:
2722 debugTrace(DEBUG_stm,
2723 "found ATOMICALLY_FRAME at %p during retry", p);
2724 tso->sp = p;
2725 return ATOMICALLY_FRAME;
2726
2727 case CATCH_RETRY_FRAME:
2728 debugTrace(DEBUG_stm,
2729 "found CATCH_RETRY_FRAME at %p during retrry", p);
2730 tso->sp = p;
2731 return CATCH_RETRY_FRAME;
2732
2733 case CATCH_STM_FRAME: {
2734 StgTRecHeader *trec = tso -> trec;
2735 StgTRecHeader *outer = stmGetEnclosingTRec(trec);
2736 debugTrace(DEBUG_stm,
2737 "found CATCH_STM_FRAME at %p during retry", p);
2738 debugTrace(DEBUG_stm, "trec=%p outer=%p", trec, outer);
2739 stmAbortTransaction(tso -> cap, trec);
2740 stmFreeAbortedTRec(tso -> cap, trec);
2741 tso -> trec = outer;
2742 p = next;
2743 continue;
2744 }
2745
2746
2747 default:
2748 ASSERT(info->i.type != CATCH_FRAME);
2749 ASSERT(info->i.type != STOP_FRAME);
2750 p = next;
2751 continue;
2752 }
2753 }
2754 }
2755
2756 /* -----------------------------------------------------------------------------
2757 resurrectThreads is called after garbage collection on the list of
2758 threads found to be garbage. Each of these threads will be woken
2759 up and sent a signal: BlockedOnDeadMVar if the thread was blocked
2760 on an MVar, or NonTermination if the thread was blocked on a Black
2761 Hole.
2762
2763 Locks: assumes we hold *all* the capabilities.
2764 -------------------------------------------------------------------------- */
2765
2766 void
2767 resurrectThreads (StgTSO *threads)
2768 {
2769 StgTSO *tso, *next;
2770 Capability *cap;
2771 step *step;
2772
2773 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
2774 next = tso->global_link;
2775
2776 step = Bdescr((P_)tso)->step;
2777 tso->global_link = step->threads;
2778 step->threads = tso;
2779
2780 debugTrace(DEBUG_sched, "resurrecting thread %lu", (unsigned long)tso->id);
2781
2782 // Wake up the thread on the Capability it was last on
2783 cap = tso->cap;
2784
2785 switch (tso->why_blocked) {
2786 case BlockedOnMVar:
2787 case BlockedOnException:
2788 /* Called by GC - sched_mutex lock is currently held. */
2789 throwToSingleThreaded(cap, tso,
2790 (StgClosure *)blockedOnDeadMVar_closure);
2791 break;
2792 case BlockedOnBlackHole:
2793 throwToSingleThreaded(cap, tso,
2794 (StgClosure *)nonTermination_closure);
2795 break;
2796 case BlockedOnSTM:
2797 throwToSingleThreaded(cap, tso,
2798 (StgClosure *)blockedIndefinitely_closure);
2799 break;
2800 case NotBlocked:
2801 /* This might happen if the thread was blocked on a black hole
2802 * belonging to a thread that we've just woken up (raiseAsync
2803 * can wake up threads, remember...).
2804 */
2805 continue;
2806 default:
2807 barf("resurrectThreads: thread blocked in a strange way");
2808 }
2809 }
2810 }
2811
2812 /* -----------------------------------------------------------------------------
2813 performPendingThrowTos is called after garbage collection, and
2814 passed a list of threads that were found to have pending throwTos
2815 (tso->blocked_exceptions was not empty), and were blocked.
2816 Normally this doesn't happen, because we would deliver the
2817 exception directly if the target thread is blocked, but there are
2818 small windows where it might occur on a multiprocessor (see
2819 throwTo()).
2820
2821 NB. we must be holding all the capabilities at this point, just
2822 like resurrectThreads().
2823 -------------------------------------------------------------------------- */
2824
2825 void
2826 performPendingThrowTos (StgTSO *threads)
2827 {
2828 StgTSO *tso, *next;
2829 Capability *cap;
2830 step *step;
2831
2832 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
2833 next = tso->global_link;
2834
2835 step = Bdescr((P_)tso)->step;
2836 tso->global_link = step->threads;
2837 step->threads = tso;
2838
2839 debugTrace(DEBUG_sched, "performing blocked throwTo to thread %lu", (unsigned long)tso->id);
2840
2841 cap = tso->cap;
2842 maybePerformBlockedException(cap, tso);
2843 }
2844 }