Use local mut lists in UPD_IND(), also clean up Updates.h
[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
13 #include "sm/Storage.h"
14 #include "RtsUtils.h"
15 #include "StgRun.h"
16 #include "Schedule.h"
17 #include "Interpreter.h"
18 #include "Printer.h"
19 #include "RtsSignals.h"
20 #include "sm/Sanity.h"
21 #include "Stats.h"
22 #include "STM.h"
23 #include "Prelude.h"
24 #include "ThreadLabels.h"
25 #include "Updates.h"
26 #include "Proftimer.h"
27 #include "ProfHeap.h"
28 #include "Weak.h"
29 #include "sm/GC.h" // waitForGcThreads, releaseGCThreads, N
30 #include "Sparks.h"
31 #include "Capability.h"
32 #include "Task.h"
33 #include "AwaitEvent.h"
34 #if defined(mingw32_HOST_OS)
35 #include "win32/IOManager.h"
36 #endif
37 #include "Trace.h"
38 #include "RaiseAsync.h"
39 #include "Threads.h"
40 #include "Timer.h"
41 #include "ThreadPaused.h"
42
43 #ifdef HAVE_SYS_TYPES_H
44 #include <sys/types.h>
45 #endif
46 #ifdef HAVE_UNISTD_H
47 #include <unistd.h>
48 #endif
49
50 #include <string.h>
51 #include <stdlib.h>
52 #include <stdarg.h>
53
54 #ifdef HAVE_ERRNO_H
55 #include <errno.h>
56 #endif
57
58 /* -----------------------------------------------------------------------------
59 * Global variables
60 * -------------------------------------------------------------------------- */
61
62 #if !defined(THREADED_RTS)
63 // Blocked/sleeping thrads
64 StgTSO *blocked_queue_hd = NULL;
65 StgTSO *blocked_queue_tl = NULL;
66 StgTSO *sleeping_queue = NULL; // perhaps replace with a hash table?
67 #endif
68
69 /* Threads blocked on blackholes.
70 * LOCK: sched_mutex+capability, or all capabilities
71 */
72 StgTSO *blackhole_queue = NULL;
73
74 /* The blackhole_queue should be checked for threads to wake up. See
75 * Schedule.h for more thorough comment.
76 * LOCK: none (doesn't matter if we miss an update)
77 */
78 rtsBool blackholes_need_checking = rtsFalse;
79
80 /* Set to true when the latest garbage collection failed to reclaim
81 * enough space, and the runtime should proceed to shut itself down in
82 * an orderly fashion (emitting profiling info etc.)
83 */
84 rtsBool heap_overflow = rtsFalse;
85
86 /* flag that tracks whether we have done any execution in this time slice.
87 * LOCK: currently none, perhaps we should lock (but needs to be
88 * updated in the fast path of the scheduler).
89 *
90 * NB. must be StgWord, we do xchg() on it.
91 */
92 volatile StgWord recent_activity = ACTIVITY_YES;
93
94 /* if this flag is set as well, give up execution
95 * LOCK: none (changes monotonically)
96 */
97 volatile StgWord sched_state = SCHED_RUNNING;
98
99 /* This is used in `TSO.h' and gcc 2.96 insists that this variable actually
100 * exists - earlier gccs apparently didn't.
101 * -= chak
102 */
103 StgTSO dummy_tso;
104
105 /*
106 * Set to TRUE when entering a shutdown state (via shutdownHaskellAndExit()) --
107 * in an MT setting, needed to signal that a worker thread shouldn't hang around
108 * in the scheduler when it is out of work.
109 */
110 rtsBool shutting_down_scheduler = rtsFalse;
111
112 /*
113 * This mutex protects most of the global scheduler data in
114 * the THREADED_RTS runtime.
115 */
116 #if defined(THREADED_RTS)
117 Mutex sched_mutex;
118 #endif
119
120 #if !defined(mingw32_HOST_OS)
121 #define FORKPROCESS_PRIMOP_SUPPORTED
122 #endif
123
124 /* -----------------------------------------------------------------------------
125 * static function prototypes
126 * -------------------------------------------------------------------------- */
127
128 static Capability *schedule (Capability *initialCapability, Task *task);
129
130 //
131 // These function all encapsulate parts of the scheduler loop, and are
132 // abstracted only to make the structure and control flow of the
133 // scheduler clearer.
134 //
135 static void schedulePreLoop (void);
136 static void scheduleFindWork (Capability *cap);
137 #if defined(THREADED_RTS)
138 static void scheduleYield (Capability **pcap, Task *task, rtsBool);
139 #endif
140 static void scheduleStartSignalHandlers (Capability *cap);
141 static void scheduleCheckBlockedThreads (Capability *cap);
142 static void scheduleCheckWakeupThreads(Capability *cap USED_IF_NOT_THREADS);
143 static void scheduleCheckBlackHoles (Capability *cap);
144 static void scheduleDetectDeadlock (Capability *cap, Task *task);
145 static void schedulePushWork(Capability *cap, Task *task);
146 #if defined(THREADED_RTS)
147 static void scheduleActivateSpark(Capability *cap);
148 #endif
149 static void schedulePostRunThread(Capability *cap, StgTSO *t);
150 static rtsBool scheduleHandleHeapOverflow( Capability *cap, StgTSO *t );
151 static void scheduleHandleStackOverflow( Capability *cap, Task *task,
152 StgTSO *t);
153 static rtsBool scheduleHandleYield( Capability *cap, StgTSO *t,
154 nat prev_what_next );
155 static void scheduleHandleThreadBlocked( StgTSO *t );
156 static rtsBool scheduleHandleThreadFinished( Capability *cap, Task *task,
157 StgTSO *t );
158 static rtsBool scheduleNeedHeapProfile(rtsBool ready_to_gc);
159 static Capability *scheduleDoGC(Capability *cap, Task *task,
160 rtsBool force_major);
161
162 static rtsBool checkBlackHoles(Capability *cap);
163
164 static StgTSO *threadStackOverflow(Capability *cap, StgTSO *tso);
165 static StgTSO *threadStackUnderflow(Capability *cap, Task *task, StgTSO *tso);
166
167 static void deleteThread (Capability *cap, StgTSO *tso);
168 static void deleteAllThreads (Capability *cap);
169
170 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
171 static void deleteThread_(Capability *cap, StgTSO *tso);
172 #endif
173
174 /* -----------------------------------------------------------------------------
175 * Putting a thread on the run queue: different scheduling policies
176 * -------------------------------------------------------------------------- */
177
178 STATIC_INLINE void
179 addToRunQueue( Capability *cap, StgTSO *t )
180 {
181 // this does round-robin scheduling; good for concurrency
182 appendToRunQueue(cap,t);
183 }
184
185 /* ---------------------------------------------------------------------------
186 Main scheduling loop.
187
188 We use round-robin scheduling, each thread returning to the
189 scheduler loop when one of these conditions is detected:
190
191 * out of heap space
192 * timer expires (thread yields)
193 * thread blocks
194 * thread ends
195 * stack overflow
196
197 GRAN version:
198 In a GranSim setup this loop iterates over the global event queue.
199 This revolves around the global event queue, which determines what
200 to do next. Therefore, it's more complicated than either the
201 concurrent or the parallel (GUM) setup.
202 This version has been entirely removed (JB 2008/08).
203
204 GUM version:
205 GUM iterates over incoming messages.
206 It starts with nothing to do (thus CurrentTSO == END_TSO_QUEUE),
207 and sends out a fish whenever it has nothing to do; in-between
208 doing the actual reductions (shared code below) it processes the
209 incoming messages and deals with delayed operations
210 (see PendingFetches).
211 This is not the ugliest code you could imagine, but it's bloody close.
212
213 (JB 2008/08) This version was formerly indicated by a PP-Flag PAR,
214 now by PP-flag PARALLEL_HASKELL. The Eden RTS (in GHC-6.x) uses it,
215 as well as future GUM versions. This file has been refurbished to
216 only contain valid code, which is however incomplete, refers to
217 invalid includes etc.
218
219 ------------------------------------------------------------------------ */
220
221 static Capability *
222 schedule (Capability *initialCapability, Task *task)
223 {
224 StgTSO *t;
225 Capability *cap;
226 StgThreadReturnCode ret;
227 nat prev_what_next;
228 rtsBool ready_to_gc;
229 #if defined(THREADED_RTS)
230 rtsBool first = rtsTrue;
231 rtsBool force_yield = rtsFalse;
232 #endif
233
234 cap = initialCapability;
235
236 // Pre-condition: this task owns initialCapability.
237 // The sched_mutex is *NOT* held
238 // NB. on return, we still hold a capability.
239
240 debugTrace (DEBUG_sched, "cap %d: schedule()", initialCapability->no);
241
242 schedulePreLoop();
243
244 // -----------------------------------------------------------
245 // Scheduler loop starts here:
246
247 while (1) {
248
249 // Check whether we have re-entered the RTS from Haskell without
250 // going via suspendThread()/resumeThread (i.e. a 'safe' foreign
251 // call).
252 if (cap->in_haskell) {
253 errorBelch("schedule: re-entered unsafely.\n"
254 " Perhaps a 'foreign import unsafe' should be 'safe'?");
255 stg_exit(EXIT_FAILURE);
256 }
257
258 // The interruption / shutdown sequence.
259 //
260 // In order to cleanly shut down the runtime, we want to:
261 // * make sure that all main threads return to their callers
262 // with the state 'Interrupted'.
263 // * clean up all OS threads assocated with the runtime
264 // * free all memory etc.
265 //
266 // So the sequence for ^C goes like this:
267 //
268 // * ^C handler sets sched_state := SCHED_INTERRUPTING and
269 // arranges for some Capability to wake up
270 //
271 // * all threads in the system are halted, and the zombies are
272 // placed on the run queue for cleaning up. We acquire all
273 // the capabilities in order to delete the threads, this is
274 // done by scheduleDoGC() for convenience (because GC already
275 // needs to acquire all the capabilities). We can't kill
276 // threads involved in foreign calls.
277 //
278 // * somebody calls shutdownHaskell(), which calls exitScheduler()
279 //
280 // * sched_state := SCHED_SHUTTING_DOWN
281 //
282 // * all workers exit when the run queue on their capability
283 // drains. All main threads will also exit when their TSO
284 // reaches the head of the run queue and they can return.
285 //
286 // * eventually all Capabilities will shut down, and the RTS can
287 // exit.
288 //
289 // * We might be left with threads blocked in foreign calls,
290 // we should really attempt to kill these somehow (TODO);
291
292 switch (sched_state) {
293 case SCHED_RUNNING:
294 break;
295 case SCHED_INTERRUPTING:
296 debugTrace(DEBUG_sched, "SCHED_INTERRUPTING");
297 #if defined(THREADED_RTS)
298 discardSparksCap(cap);
299 #endif
300 /* scheduleDoGC() deletes all the threads */
301 cap = scheduleDoGC(cap,task,rtsFalse);
302
303 // after scheduleDoGC(), we must be shutting down. Either some
304 // other Capability did the final GC, or we did it above,
305 // either way we can fall through to the SCHED_SHUTTING_DOWN
306 // case now.
307 ASSERT(sched_state == SCHED_SHUTTING_DOWN);
308 // fall through
309
310 case SCHED_SHUTTING_DOWN:
311 debugTrace(DEBUG_sched, "SCHED_SHUTTING_DOWN");
312 // If we are a worker, just exit. If we're a bound thread
313 // then we will exit below when we've removed our TSO from
314 // the run queue.
315 if (task->tso == NULL && emptyRunQueue(cap)) {
316 return cap;
317 }
318 break;
319 default:
320 barf("sched_state: %d", sched_state);
321 }
322
323 scheduleFindWork(cap);
324
325 /* work pushing, currently relevant only for THREADED_RTS:
326 (pushes threads, wakes up idle capabilities for stealing) */
327 schedulePushWork(cap,task);
328
329 scheduleDetectDeadlock(cap,task);
330
331 #if defined(THREADED_RTS)
332 cap = task->cap; // reload cap, it might have changed
333 #endif
334
335 // Normally, the only way we can get here with no threads to
336 // run is if a keyboard interrupt received during
337 // scheduleCheckBlockedThreads() or scheduleDetectDeadlock().
338 // Additionally, it is not fatal for the
339 // threaded RTS to reach here with no threads to run.
340 //
341 // win32: might be here due to awaitEvent() being abandoned
342 // as a result of a console event having been delivered.
343
344 #if defined(THREADED_RTS)
345 if (first)
346 {
347 // XXX: ToDo
348 // // don't yield the first time, we want a chance to run this
349 // // thread for a bit, even if there are others banging at the
350 // // door.
351 // first = rtsFalse;
352 // ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
353 }
354
355 yield:
356 scheduleYield(&cap,task,force_yield);
357 force_yield = rtsFalse;
358
359 if (emptyRunQueue(cap)) continue; // look for work again
360 #endif
361
362 #if !defined(THREADED_RTS) && !defined(mingw32_HOST_OS)
363 if ( emptyRunQueue(cap) ) {
364 ASSERT(sched_state >= SCHED_INTERRUPTING);
365 }
366 #endif
367
368 //
369 // Get a thread to run
370 //
371 t = popRunQueue(cap);
372
373 // Sanity check the thread we're about to run. This can be
374 // expensive if there is lots of thread switching going on...
375 IF_DEBUG(sanity,checkTSO(t));
376
377 #if defined(THREADED_RTS)
378 // Check whether we can run this thread in the current task.
379 // If not, we have to pass our capability to the right task.
380 {
381 Task *bound = t->bound;
382
383 if (bound) {
384 if (bound == task) {
385 // yes, the Haskell thread is bound to the current native thread
386 } else {
387 debugTrace(DEBUG_sched,
388 "thread %lu bound to another OS thread",
389 (unsigned long)t->id);
390 // no, bound to a different Haskell thread: pass to that thread
391 pushOnRunQueue(cap,t);
392 continue;
393 }
394 } else {
395 // The thread we want to run is unbound.
396 if (task->tso) {
397 debugTrace(DEBUG_sched,
398 "this OS thread cannot run thread %lu",
399 (unsigned long)t->id);
400 // no, the current native thread is bound to a different
401 // Haskell thread, so pass it to any worker thread
402 pushOnRunQueue(cap,t);
403 continue;
404 }
405 }
406 }
407 #endif
408
409 // If we're shutting down, and this thread has not yet been
410 // killed, kill it now. This sometimes happens when a finalizer
411 // thread is created by the final GC, or a thread previously
412 // in a foreign call returns.
413 if (sched_state >= SCHED_INTERRUPTING &&
414 !(t->what_next == ThreadComplete || t->what_next == ThreadKilled)) {
415 deleteThread(cap,t);
416 }
417
418 /* context switches are initiated by the timer signal, unless
419 * the user specified "context switch as often as possible", with
420 * +RTS -C0
421 */
422 if (RtsFlags.ConcFlags.ctxtSwitchTicks == 0
423 && !emptyThreadQueues(cap)) {
424 cap->context_switch = 1;
425 }
426
427 run_thread:
428
429 // CurrentTSO is the thread to run. t might be different if we
430 // loop back to run_thread, so make sure to set CurrentTSO after
431 // that.
432 cap->r.rCurrentTSO = t;
433
434 startHeapProfTimer();
435
436 // Check for exceptions blocked on this thread
437 maybePerformBlockedException (cap, t);
438
439 // ----------------------------------------------------------------------
440 // Run the current thread
441
442 ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
443 ASSERT(t->cap == cap);
444 ASSERT(t->bound ? t->bound->cap == cap : 1);
445
446 prev_what_next = t->what_next;
447
448 errno = t->saved_errno;
449 #if mingw32_HOST_OS
450 SetLastError(t->saved_winerror);
451 #endif
452
453 cap->in_haskell = rtsTrue;
454
455 dirty_TSO(cap,t);
456
457 #if defined(THREADED_RTS)
458 if (recent_activity == ACTIVITY_DONE_GC) {
459 // ACTIVITY_DONE_GC means we turned off the timer signal to
460 // conserve power (see #1623). Re-enable it here.
461 nat prev;
462 prev = xchg((P_)&recent_activity, ACTIVITY_YES);
463 if (prev == ACTIVITY_DONE_GC) {
464 startTimer();
465 }
466 } else if (recent_activity != ACTIVITY_INACTIVE) {
467 // If we reached ACTIVITY_INACTIVE, then don't reset it until
468 // we've done the GC. The thread running here might just be
469 // the IO manager thread that handle_tick() woke up via
470 // wakeUpRts().
471 recent_activity = ACTIVITY_YES;
472 }
473 #endif
474
475 traceEventRunThread(cap, t);
476
477 switch (prev_what_next) {
478
479 case ThreadKilled:
480 case ThreadComplete:
481 /* Thread already finished, return to scheduler. */
482 ret = ThreadFinished;
483 break;
484
485 case ThreadRunGHC:
486 {
487 StgRegTable *r;
488 r = StgRun((StgFunPtr) stg_returnToStackTop, &cap->r);
489 cap = regTableToCapability(r);
490 ret = r->rRet;
491 break;
492 }
493
494 case ThreadInterpret:
495 cap = interpretBCO(cap);
496 ret = cap->r.rRet;
497 break;
498
499 default:
500 barf("schedule: invalid what_next field");
501 }
502
503 cap->in_haskell = rtsFalse;
504
505 // The TSO might have moved, eg. if it re-entered the RTS and a GC
506 // happened. So find the new location:
507 t = cap->r.rCurrentTSO;
508
509 // We have run some Haskell code: there might be blackhole-blocked
510 // threads to wake up now.
511 // Lock-free test here should be ok, we're just setting a flag.
512 if ( blackhole_queue != END_TSO_QUEUE ) {
513 blackholes_need_checking = rtsTrue;
514 }
515
516 // And save the current errno in this thread.
517 // XXX: possibly bogus for SMP because this thread might already
518 // be running again, see code below.
519 t->saved_errno = errno;
520 #if mingw32_HOST_OS
521 // Similarly for Windows error code
522 t->saved_winerror = GetLastError();
523 #endif
524
525 traceEventStopThread(cap, t, ret);
526
527 #if defined(THREADED_RTS)
528 // If ret is ThreadBlocked, and this Task is bound to the TSO that
529 // blocked, we are in limbo - the TSO is now owned by whatever it
530 // is blocked on, and may in fact already have been woken up,
531 // perhaps even on a different Capability. It may be the case
532 // that task->cap != cap. We better yield this Capability
533 // immediately and return to normaility.
534 if (ret == ThreadBlocked) {
535 force_yield = rtsTrue;
536 goto yield;
537 }
538 #endif
539
540 ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
541 ASSERT(t->cap == cap);
542
543 // ----------------------------------------------------------------------
544
545 // Costs for the scheduler are assigned to CCS_SYSTEM
546 stopHeapProfTimer();
547 #if defined(PROFILING)
548 CCCS = CCS_SYSTEM;
549 #endif
550
551 schedulePostRunThread(cap,t);
552
553 if (ret != StackOverflow) {
554 t = threadStackUnderflow(cap,task,t);
555 }
556
557 ready_to_gc = rtsFalse;
558
559 switch (ret) {
560 case HeapOverflow:
561 ready_to_gc = scheduleHandleHeapOverflow(cap,t);
562 break;
563
564 case StackOverflow:
565 scheduleHandleStackOverflow(cap,task,t);
566 break;
567
568 case ThreadYielding:
569 if (scheduleHandleYield(cap, t, prev_what_next)) {
570 // shortcut for switching between compiler/interpreter:
571 goto run_thread;
572 }
573 break;
574
575 case ThreadBlocked:
576 scheduleHandleThreadBlocked(t);
577 break;
578
579 case ThreadFinished:
580 if (scheduleHandleThreadFinished(cap, task, t)) return cap;
581 ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
582 break;
583
584 default:
585 barf("schedule: invalid thread return code %d", (int)ret);
586 }
587
588 if (ready_to_gc || scheduleNeedHeapProfile(ready_to_gc)) {
589 cap = scheduleDoGC(cap,task,rtsFalse);
590 }
591 } /* end of while() */
592 }
593
594 /* ----------------------------------------------------------------------------
595 * Setting up the scheduler loop
596 * ------------------------------------------------------------------------- */
597
598 static void
599 schedulePreLoop(void)
600 {
601 // initialisation for scheduler - what cannot go into initScheduler()
602 }
603
604 /* -----------------------------------------------------------------------------
605 * scheduleFindWork()
606 *
607 * Search for work to do, and handle messages from elsewhere.
608 * -------------------------------------------------------------------------- */
609
610 static void
611 scheduleFindWork (Capability *cap)
612 {
613 scheduleStartSignalHandlers(cap);
614
615 // Only check the black holes here if we've nothing else to do.
616 // During normal execution, the black hole list only gets checked
617 // at GC time, to avoid repeatedly traversing this possibly long
618 // list each time around the scheduler.
619 if (emptyRunQueue(cap)) { scheduleCheckBlackHoles(cap); }
620
621 scheduleCheckWakeupThreads(cap);
622
623 scheduleCheckBlockedThreads(cap);
624
625 #if defined(THREADED_RTS)
626 if (emptyRunQueue(cap)) { scheduleActivateSpark(cap); }
627 #endif
628 }
629
630 #if defined(THREADED_RTS)
631 STATIC_INLINE rtsBool
632 shouldYieldCapability (Capability *cap, Task *task)
633 {
634 // we need to yield this capability to someone else if..
635 // - another thread is initiating a GC
636 // - another Task is returning from a foreign call
637 // - the thread at the head of the run queue cannot be run
638 // by this Task (it is bound to another Task, or it is unbound
639 // and this task it bound).
640 return (waiting_for_gc ||
641 cap->returning_tasks_hd != NULL ||
642 (!emptyRunQueue(cap) && (task->tso == NULL
643 ? cap->run_queue_hd->bound != NULL
644 : cap->run_queue_hd->bound != task)));
645 }
646
647 // This is the single place where a Task goes to sleep. There are
648 // two reasons it might need to sleep:
649 // - there are no threads to run
650 // - we need to yield this Capability to someone else
651 // (see shouldYieldCapability())
652 //
653 // Careful: the scheduler loop is quite delicate. Make sure you run
654 // the tests in testsuite/concurrent (all ways) after modifying this,
655 // and also check the benchmarks in nofib/parallel for regressions.
656
657 static void
658 scheduleYield (Capability **pcap, Task *task, rtsBool force_yield)
659 {
660 Capability *cap = *pcap;
661
662 // if we have work, and we don't need to give up the Capability, continue.
663 //
664 // The force_yield flag is used when a bound thread blocks. This
665 // is a particularly tricky situation: the current Task does not
666 // own the TSO any more, since it is on some queue somewhere, and
667 // might be woken up or manipulated by another thread at any time.
668 // The TSO and Task might be migrated to another Capability.
669 // Certain invariants might be in doubt, such as task->bound->cap
670 // == cap. We have to yield the current Capability immediately,
671 // no messing around.
672 //
673 if (!force_yield &&
674 !shouldYieldCapability(cap,task) &&
675 (!emptyRunQueue(cap) ||
676 !emptyWakeupQueue(cap) ||
677 blackholes_need_checking ||
678 sched_state >= SCHED_INTERRUPTING))
679 return;
680
681 // otherwise yield (sleep), and keep yielding if necessary.
682 do {
683 yieldCapability(&cap,task);
684 }
685 while (shouldYieldCapability(cap,task));
686
687 // note there may still be no threads on the run queue at this
688 // point, the caller has to check.
689
690 *pcap = cap;
691 return;
692 }
693 #endif
694
695 /* -----------------------------------------------------------------------------
696 * schedulePushWork()
697 *
698 * Push work to other Capabilities if we have some.
699 * -------------------------------------------------------------------------- */
700
701 static void
702 schedulePushWork(Capability *cap USED_IF_THREADS,
703 Task *task USED_IF_THREADS)
704 {
705 /* following code not for PARALLEL_HASKELL. I kept the call general,
706 future GUM versions might use pushing in a distributed setup */
707 #if defined(THREADED_RTS)
708
709 Capability *free_caps[n_capabilities], *cap0;
710 nat i, n_free_caps;
711
712 // migration can be turned off with +RTS -qg
713 if (!RtsFlags.ParFlags.migrate) return;
714
715 // Check whether we have more threads on our run queue, or sparks
716 // in our pool, that we could hand to another Capability.
717 if (cap->run_queue_hd == END_TSO_QUEUE) {
718 if (sparkPoolSizeCap(cap) < 2) return;
719 } else {
720 if (cap->run_queue_hd->_link == END_TSO_QUEUE &&
721 sparkPoolSizeCap(cap) < 1) return;
722 }
723
724 // First grab as many free Capabilities as we can.
725 for (i=0, n_free_caps=0; i < n_capabilities; i++) {
726 cap0 = &capabilities[i];
727 if (cap != cap0 && tryGrabCapability(cap0,task)) {
728 if (!emptyRunQueue(cap0) || cap->returning_tasks_hd != NULL) {
729 // it already has some work, we just grabbed it at
730 // the wrong moment. Or maybe it's deadlocked!
731 releaseCapability(cap0);
732 } else {
733 free_caps[n_free_caps++] = cap0;
734 }
735 }
736 }
737
738 // we now have n_free_caps free capabilities stashed in
739 // free_caps[]. Share our run queue equally with them. This is
740 // probably the simplest thing we could do; improvements we might
741 // want to do include:
742 //
743 // - giving high priority to moving relatively new threads, on
744 // the gournds that they haven't had time to build up a
745 // working set in the cache on this CPU/Capability.
746 //
747 // - giving low priority to moving long-lived threads
748
749 if (n_free_caps > 0) {
750 StgTSO *prev, *t, *next;
751 rtsBool pushed_to_all;
752
753 debugTrace(DEBUG_sched,
754 "cap %d: %s and %d free capabilities, sharing...",
755 cap->no,
756 (!emptyRunQueue(cap) && cap->run_queue_hd->_link != END_TSO_QUEUE)?
757 "excess threads on run queue":"sparks to share (>=2)",
758 n_free_caps);
759
760 i = 0;
761 pushed_to_all = rtsFalse;
762
763 if (cap->run_queue_hd != END_TSO_QUEUE) {
764 prev = cap->run_queue_hd;
765 t = prev->_link;
766 prev->_link = END_TSO_QUEUE;
767 for (; t != END_TSO_QUEUE; t = next) {
768 next = t->_link;
769 t->_link = END_TSO_QUEUE;
770 if (t->what_next == ThreadRelocated
771 || t->bound == task // don't move my bound thread
772 || tsoLocked(t)) { // don't move a locked thread
773 setTSOLink(cap, prev, t);
774 prev = t;
775 } else if (i == n_free_caps) {
776 pushed_to_all = rtsTrue;
777 i = 0;
778 // keep one for us
779 setTSOLink(cap, prev, t);
780 prev = t;
781 } else {
782 debugTrace(DEBUG_sched, "pushing thread %lu to capability %d", (unsigned long)t->id, free_caps[i]->no);
783 appendToRunQueue(free_caps[i],t);
784
785 traceEventMigrateThread (cap, t, free_caps[i]->no);
786
787 if (t->bound) { t->bound->cap = free_caps[i]; }
788 t->cap = free_caps[i];
789 i++;
790 }
791 }
792 cap->run_queue_tl = prev;
793 }
794
795 #ifdef SPARK_PUSHING
796 /* JB I left this code in place, it would work but is not necessary */
797
798 // If there are some free capabilities that we didn't push any
799 // threads to, then try to push a spark to each one.
800 if (!pushed_to_all) {
801 StgClosure *spark;
802 // i is the next free capability to push to
803 for (; i < n_free_caps; i++) {
804 if (emptySparkPoolCap(free_caps[i])) {
805 spark = tryStealSpark(cap->sparks);
806 if (spark != NULL) {
807 debugTrace(DEBUG_sched, "pushing spark %p to capability %d", spark, free_caps[i]->no);
808
809 traceEventStealSpark(free_caps[i], t, cap->no);
810
811 newSpark(&(free_caps[i]->r), spark);
812 }
813 }
814 }
815 }
816 #endif /* SPARK_PUSHING */
817
818 // release the capabilities
819 for (i = 0; i < n_free_caps; i++) {
820 task->cap = free_caps[i];
821 releaseAndWakeupCapability(free_caps[i]);
822 }
823 }
824 task->cap = cap; // reset to point to our Capability.
825
826 #endif /* THREADED_RTS */
827
828 }
829
830 /* ----------------------------------------------------------------------------
831 * Start any pending signal handlers
832 * ------------------------------------------------------------------------- */
833
834 #if defined(RTS_USER_SIGNALS) && !defined(THREADED_RTS)
835 static void
836 scheduleStartSignalHandlers(Capability *cap)
837 {
838 if (RtsFlags.MiscFlags.install_signal_handlers && signals_pending()) {
839 // safe outside the lock
840 startSignalHandlers(cap);
841 }
842 }
843 #else
844 static void
845 scheduleStartSignalHandlers(Capability *cap STG_UNUSED)
846 {
847 }
848 #endif
849
850 /* ----------------------------------------------------------------------------
851 * Check for blocked threads that can be woken up.
852 * ------------------------------------------------------------------------- */
853
854 static void
855 scheduleCheckBlockedThreads(Capability *cap USED_IF_NOT_THREADS)
856 {
857 #if !defined(THREADED_RTS)
858 //
859 // Check whether any waiting threads need to be woken up. If the
860 // run queue is empty, and there are no other tasks running, we
861 // can wait indefinitely for something to happen.
862 //
863 if ( !emptyQueue(blocked_queue_hd) || !emptyQueue(sleeping_queue) )
864 {
865 awaitEvent( emptyRunQueue(cap) && !blackholes_need_checking );
866 }
867 #endif
868 }
869
870
871 /* ----------------------------------------------------------------------------
872 * Check for threads woken up by other Capabilities
873 * ------------------------------------------------------------------------- */
874
875 static void
876 scheduleCheckWakeupThreads(Capability *cap USED_IF_THREADS)
877 {
878 #if defined(THREADED_RTS)
879 // Any threads that were woken up by other Capabilities get
880 // appended to our run queue.
881 if (!emptyWakeupQueue(cap)) {
882 ACQUIRE_LOCK(&cap->lock);
883 if (emptyRunQueue(cap)) {
884 cap->run_queue_hd = cap->wakeup_queue_hd;
885 cap->run_queue_tl = cap->wakeup_queue_tl;
886 } else {
887 setTSOLink(cap, cap->run_queue_tl, cap->wakeup_queue_hd);
888 cap->run_queue_tl = cap->wakeup_queue_tl;
889 }
890 cap->wakeup_queue_hd = cap->wakeup_queue_tl = END_TSO_QUEUE;
891 RELEASE_LOCK(&cap->lock);
892 }
893 #endif
894 }
895
896 /* ----------------------------------------------------------------------------
897 * Check for threads blocked on BLACKHOLEs that can be woken up
898 * ------------------------------------------------------------------------- */
899 static void
900 scheduleCheckBlackHoles (Capability *cap)
901 {
902 if ( blackholes_need_checking ) // check without the lock first
903 {
904 ACQUIRE_LOCK(&sched_mutex);
905 if ( blackholes_need_checking ) {
906 blackholes_need_checking = rtsFalse;
907 // important that we reset the flag *before* checking the
908 // blackhole queue, otherwise we could get deadlock. This
909 // happens as follows: we wake up a thread that
910 // immediately runs on another Capability, blocks on a
911 // blackhole, and then we reset the blackholes_need_checking flag.
912 checkBlackHoles(cap);
913 }
914 RELEASE_LOCK(&sched_mutex);
915 }
916 }
917
918 /* ----------------------------------------------------------------------------
919 * Detect deadlock conditions and attempt to resolve them.
920 * ------------------------------------------------------------------------- */
921
922 static void
923 scheduleDetectDeadlock (Capability *cap, Task *task)
924 {
925 /*
926 * Detect deadlock: when we have no threads to run, there are no
927 * threads blocked, waiting for I/O, or sleeping, and all the
928 * other tasks are waiting for work, we must have a deadlock of
929 * some description.
930 */
931 if ( emptyThreadQueues(cap) )
932 {
933 #if defined(THREADED_RTS)
934 /*
935 * In the threaded RTS, we only check for deadlock if there
936 * has been no activity in a complete timeslice. This means
937 * we won't eagerly start a full GC just because we don't have
938 * any threads to run currently.
939 */
940 if (recent_activity != ACTIVITY_INACTIVE) return;
941 #endif
942
943 debugTrace(DEBUG_sched, "deadlocked, forcing major GC...");
944
945 // Garbage collection can release some new threads due to
946 // either (a) finalizers or (b) threads resurrected because
947 // they are unreachable and will therefore be sent an
948 // exception. Any threads thus released will be immediately
949 // runnable.
950 cap = scheduleDoGC (cap, task, rtsTrue/*force major GC*/);
951 // when force_major == rtsTrue. scheduleDoGC sets
952 // recent_activity to ACTIVITY_DONE_GC and turns off the timer
953 // signal.
954
955 if ( !emptyRunQueue(cap) ) return;
956
957 #if defined(RTS_USER_SIGNALS) && !defined(THREADED_RTS)
958 /* If we have user-installed signal handlers, then wait
959 * for signals to arrive rather then bombing out with a
960 * deadlock.
961 */
962 if ( RtsFlags.MiscFlags.install_signal_handlers && anyUserHandlers() ) {
963 debugTrace(DEBUG_sched,
964 "still deadlocked, waiting for signals...");
965
966 awaitUserSignals();
967
968 if (signals_pending()) {
969 startSignalHandlers(cap);
970 }
971
972 // either we have threads to run, or we were interrupted:
973 ASSERT(!emptyRunQueue(cap) || sched_state >= SCHED_INTERRUPTING);
974
975 return;
976 }
977 #endif
978
979 #if !defined(THREADED_RTS)
980 /* Probably a real deadlock. Send the current main thread the
981 * Deadlock exception.
982 */
983 if (task->tso) {
984 switch (task->tso->why_blocked) {
985 case BlockedOnSTM:
986 case BlockedOnBlackHole:
987 case BlockedOnException:
988 case BlockedOnMVar:
989 throwToSingleThreaded(cap, task->tso,
990 (StgClosure *)nonTermination_closure);
991 return;
992 default:
993 barf("deadlock: main thread blocked in a strange way");
994 }
995 }
996 return;
997 #endif
998 }
999 }
1000
1001
1002 /* ----------------------------------------------------------------------------
1003 * Send pending messages (PARALLEL_HASKELL only)
1004 * ------------------------------------------------------------------------- */
1005
1006 #if defined(PARALLEL_HASKELL)
1007 static void
1008 scheduleSendPendingMessages(void)
1009 {
1010
1011 # if defined(PAR) // global Mem.Mgmt., omit for now
1012 if (PendingFetches != END_BF_QUEUE) {
1013 processFetches();
1014 }
1015 # endif
1016
1017 if (RtsFlags.ParFlags.BufferTime) {
1018 // if we use message buffering, we must send away all message
1019 // packets which have become too old...
1020 sendOldBuffers();
1021 }
1022 }
1023 #endif
1024
1025 /* ----------------------------------------------------------------------------
1026 * Activate spark threads (PARALLEL_HASKELL and THREADED_RTS)
1027 * ------------------------------------------------------------------------- */
1028
1029 #if defined(THREADED_RTS)
1030 static void
1031 scheduleActivateSpark(Capability *cap)
1032 {
1033 if (anySparks())
1034 {
1035 createSparkThread(cap);
1036 debugTrace(DEBUG_sched, "creating a spark thread");
1037 }
1038 }
1039 #endif // PARALLEL_HASKELL || THREADED_RTS
1040
1041 /* ----------------------------------------------------------------------------
1042 * After running a thread...
1043 * ------------------------------------------------------------------------- */
1044
1045 static void
1046 schedulePostRunThread (Capability *cap, StgTSO *t)
1047 {
1048 // We have to be able to catch transactions that are in an
1049 // infinite loop as a result of seeing an inconsistent view of
1050 // memory, e.g.
1051 //
1052 // atomically $ do
1053 // [a,b] <- mapM readTVar [ta,tb]
1054 // when (a == b) loop
1055 //
1056 // and a is never equal to b given a consistent view of memory.
1057 //
1058 if (t -> trec != NO_TREC && t -> why_blocked == NotBlocked) {
1059 if (!stmValidateNestOfTransactions (t -> trec)) {
1060 debugTrace(DEBUG_sched | DEBUG_stm,
1061 "trec %p found wasting its time", t);
1062
1063 // strip the stack back to the
1064 // ATOMICALLY_FRAME, aborting the (nested)
1065 // transaction, and saving the stack of any
1066 // partially-evaluated thunks on the heap.
1067 throwToSingleThreaded_(cap, t, NULL, rtsTrue);
1068
1069 // ASSERT(get_itbl((StgClosure *)t->sp)->type == ATOMICALLY_FRAME);
1070 }
1071 }
1072
1073 /* some statistics gathering in the parallel case */
1074 }
1075
1076 /* -----------------------------------------------------------------------------
1077 * Handle a thread that returned to the scheduler with ThreadHeepOverflow
1078 * -------------------------------------------------------------------------- */
1079
1080 static rtsBool
1081 scheduleHandleHeapOverflow( Capability *cap, StgTSO *t )
1082 {
1083 // did the task ask for a large block?
1084 if (cap->r.rHpAlloc > BLOCK_SIZE) {
1085 // if so, get one and push it on the front of the nursery.
1086 bdescr *bd;
1087 lnat blocks;
1088
1089 blocks = (lnat)BLOCK_ROUND_UP(cap->r.rHpAlloc) / BLOCK_SIZE;
1090
1091 debugTrace(DEBUG_sched,
1092 "--<< thread %ld (%s) stopped: requesting a large block (size %ld)\n",
1093 (long)t->id, what_next_strs[t->what_next], blocks);
1094
1095 // don't do this if the nursery is (nearly) full, we'll GC first.
1096 if (cap->r.rCurrentNursery->link != NULL ||
1097 cap->r.rNursery->n_blocks == 1) { // paranoia to prevent infinite loop
1098 // if the nursery has only one block.
1099
1100 ACQUIRE_SM_LOCK
1101 bd = allocGroup( blocks );
1102 RELEASE_SM_LOCK
1103 cap->r.rNursery->n_blocks += blocks;
1104
1105 // link the new group into the list
1106 bd->link = cap->r.rCurrentNursery;
1107 bd->u.back = cap->r.rCurrentNursery->u.back;
1108 if (cap->r.rCurrentNursery->u.back != NULL) {
1109 cap->r.rCurrentNursery->u.back->link = bd;
1110 } else {
1111 cap->r.rNursery->blocks = bd;
1112 }
1113 cap->r.rCurrentNursery->u.back = bd;
1114
1115 // initialise it as a nursery block. We initialise the
1116 // step, gen_no, and flags field of *every* sub-block in
1117 // this large block, because this is easier than making
1118 // sure that we always find the block head of a large
1119 // block whenever we call Bdescr() (eg. evacuate() and
1120 // isAlive() in the GC would both have to do this, at
1121 // least).
1122 {
1123 bdescr *x;
1124 for (x = bd; x < bd + blocks; x++) {
1125 initBdescr(x,g0,g0);
1126 x->free = x->start;
1127 x->flags = 0;
1128 }
1129 }
1130
1131 // This assert can be a killer if the app is doing lots
1132 // of large block allocations.
1133 IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
1134
1135 // now update the nursery to point to the new block
1136 cap->r.rCurrentNursery = bd;
1137
1138 // we might be unlucky and have another thread get on the
1139 // run queue before us and steal the large block, but in that
1140 // case the thread will just end up requesting another large
1141 // block.
1142 pushOnRunQueue(cap,t);
1143 return rtsFalse; /* not actually GC'ing */
1144 }
1145 }
1146
1147 if (cap->r.rHpLim == NULL || cap->context_switch) {
1148 // Sometimes we miss a context switch, e.g. when calling
1149 // primitives in a tight loop, MAYBE_GC() doesn't check the
1150 // context switch flag, and we end up waiting for a GC.
1151 // See #1984, and concurrent/should_run/1984
1152 cap->context_switch = 0;
1153 addToRunQueue(cap,t);
1154 } else {
1155 pushOnRunQueue(cap,t);
1156 }
1157 return rtsTrue;
1158 /* actual GC is done at the end of the while loop in schedule() */
1159 }
1160
1161 /* -----------------------------------------------------------------------------
1162 * Handle a thread that returned to the scheduler with ThreadStackOverflow
1163 * -------------------------------------------------------------------------- */
1164
1165 static void
1166 scheduleHandleStackOverflow (Capability *cap, Task *task, StgTSO *t)
1167 {
1168 /* just adjust the stack for this thread, then pop it back
1169 * on the run queue.
1170 */
1171 {
1172 /* enlarge the stack */
1173 StgTSO *new_t = threadStackOverflow(cap, t);
1174
1175 /* The TSO attached to this Task may have moved, so update the
1176 * pointer to it.
1177 */
1178 if (task->tso == t) {
1179 task->tso = new_t;
1180 }
1181 pushOnRunQueue(cap,new_t);
1182 }
1183 }
1184
1185 /* -----------------------------------------------------------------------------
1186 * Handle a thread that returned to the scheduler with ThreadYielding
1187 * -------------------------------------------------------------------------- */
1188
1189 static rtsBool
1190 scheduleHandleYield( Capability *cap, StgTSO *t, nat prev_what_next )
1191 {
1192 // Reset the context switch flag. We don't do this just before
1193 // running the thread, because that would mean we would lose ticks
1194 // during GC, which can lead to unfair scheduling (a thread hogs
1195 // the CPU because the tick always arrives during GC). This way
1196 // penalises threads that do a lot of allocation, but that seems
1197 // better than the alternative.
1198 cap->context_switch = 0;
1199
1200 /* put the thread back on the run queue. Then, if we're ready to
1201 * GC, check whether this is the last task to stop. If so, wake
1202 * up the GC thread. getThread will block during a GC until the
1203 * GC is finished.
1204 */
1205 #ifdef DEBUG
1206 if (t->what_next != prev_what_next) {
1207 debugTrace(DEBUG_sched,
1208 "--<< thread %ld (%s) stopped to switch evaluators",
1209 (long)t->id, what_next_strs[t->what_next]);
1210 }
1211 #endif
1212
1213 ASSERT(t->_link == END_TSO_QUEUE);
1214
1215 // Shortcut if we're just switching evaluators: don't bother
1216 // doing stack squeezing (which can be expensive), just run the
1217 // thread.
1218 if (t->what_next != prev_what_next) {
1219 return rtsTrue;
1220 }
1221
1222 IF_DEBUG(sanity,
1223 //debugBelch("&& Doing sanity check on yielding TSO %ld.", t->id);
1224 checkTSO(t));
1225
1226 addToRunQueue(cap,t);
1227
1228 return rtsFalse;
1229 }
1230
1231 /* -----------------------------------------------------------------------------
1232 * Handle a thread that returned to the scheduler with ThreadBlocked
1233 * -------------------------------------------------------------------------- */
1234
1235 static void
1236 scheduleHandleThreadBlocked( StgTSO *t
1237 #if !defined(DEBUG)
1238 STG_UNUSED
1239 #endif
1240 )
1241 {
1242
1243 // We don't need to do anything. The thread is blocked, and it
1244 // has tidied up its stack and placed itself on whatever queue
1245 // it needs to be on.
1246
1247 // ASSERT(t->why_blocked != NotBlocked);
1248 // Not true: for example,
1249 // - in THREADED_RTS, the thread may already have been woken
1250 // up by another Capability. This actually happens: try
1251 // conc023 +RTS -N2.
1252 // - the thread may have woken itself up already, because
1253 // threadPaused() might have raised a blocked throwTo
1254 // exception, see maybePerformBlockedException().
1255
1256 #ifdef DEBUG
1257 traceThreadStatus(DEBUG_sched, t);
1258 #endif
1259 }
1260
1261 /* -----------------------------------------------------------------------------
1262 * Handle a thread that returned to the scheduler with ThreadFinished
1263 * -------------------------------------------------------------------------- */
1264
1265 static rtsBool
1266 scheduleHandleThreadFinished (Capability *cap STG_UNUSED, Task *task, StgTSO *t)
1267 {
1268 /* Need to check whether this was a main thread, and if so,
1269 * return with the return value.
1270 *
1271 * We also end up here if the thread kills itself with an
1272 * uncaught exception, see Exception.cmm.
1273 */
1274
1275 // blocked exceptions can now complete, even if the thread was in
1276 // blocked mode (see #2910). This unconditionally calls
1277 // lockTSO(), which ensures that we don't miss any threads that
1278 // are engaged in throwTo() with this thread as a target.
1279 awakenBlockedExceptionQueue (cap, t);
1280
1281 //
1282 // Check whether the thread that just completed was a bound
1283 // thread, and if so return with the result.
1284 //
1285 // There is an assumption here that all thread completion goes
1286 // through this point; we need to make sure that if a thread
1287 // ends up in the ThreadKilled state, that it stays on the run
1288 // queue so it can be dealt with here.
1289 //
1290
1291 if (t->bound) {
1292
1293 if (t->bound != task) {
1294 #if !defined(THREADED_RTS)
1295 // Must be a bound thread that is not the topmost one. Leave
1296 // it on the run queue until the stack has unwound to the
1297 // point where we can deal with this. Leaving it on the run
1298 // queue also ensures that the garbage collector knows about
1299 // this thread and its return value (it gets dropped from the
1300 // step->threads list so there's no other way to find it).
1301 appendToRunQueue(cap,t);
1302 return rtsFalse;
1303 #else
1304 // this cannot happen in the threaded RTS, because a
1305 // bound thread can only be run by the appropriate Task.
1306 barf("finished bound thread that isn't mine");
1307 #endif
1308 }
1309
1310 ASSERT(task->tso == t);
1311
1312 if (t->what_next == ThreadComplete) {
1313 if (task->ret) {
1314 // NOTE: return val is tso->sp[1] (see StgStartup.hc)
1315 *(task->ret) = (StgClosure *)task->tso->sp[1];
1316 }
1317 task->stat = Success;
1318 } else {
1319 if (task->ret) {
1320 *(task->ret) = NULL;
1321 }
1322 if (sched_state >= SCHED_INTERRUPTING) {
1323 if (heap_overflow) {
1324 task->stat = HeapExhausted;
1325 } else {
1326 task->stat = Interrupted;
1327 }
1328 } else {
1329 task->stat = Killed;
1330 }
1331 }
1332 #ifdef DEBUG
1333 removeThreadLabel((StgWord)task->tso->id);
1334 #endif
1335 return rtsTrue; // tells schedule() to return
1336 }
1337
1338 return rtsFalse;
1339 }
1340
1341 /* -----------------------------------------------------------------------------
1342 * Perform a heap census
1343 * -------------------------------------------------------------------------- */
1344
1345 static rtsBool
1346 scheduleNeedHeapProfile( rtsBool ready_to_gc STG_UNUSED )
1347 {
1348 // When we have +RTS -i0 and we're heap profiling, do a census at
1349 // every GC. This lets us get repeatable runs for debugging.
1350 if (performHeapProfile ||
1351 (RtsFlags.ProfFlags.profileInterval==0 &&
1352 RtsFlags.ProfFlags.doHeapProfile && ready_to_gc)) {
1353 return rtsTrue;
1354 } else {
1355 return rtsFalse;
1356 }
1357 }
1358
1359 /* -----------------------------------------------------------------------------
1360 * Perform a garbage collection if necessary
1361 * -------------------------------------------------------------------------- */
1362
1363 static Capability *
1364 scheduleDoGC (Capability *cap, Task *task USED_IF_THREADS, rtsBool force_major)
1365 {
1366 rtsBool heap_census;
1367 #ifdef THREADED_RTS
1368 /* extern static volatile StgWord waiting_for_gc;
1369 lives inside capability.c */
1370 rtsBool gc_type, prev_pending_gc;
1371 nat i;
1372 #endif
1373
1374 if (sched_state == SCHED_SHUTTING_DOWN) {
1375 // The final GC has already been done, and the system is
1376 // shutting down. We'll probably deadlock if we try to GC
1377 // now.
1378 return cap;
1379 }
1380
1381 #ifdef THREADED_RTS
1382 if (sched_state < SCHED_INTERRUPTING
1383 && RtsFlags.ParFlags.parGcEnabled
1384 && N >= RtsFlags.ParFlags.parGcGen
1385 && ! oldest_gen->mark)
1386 {
1387 gc_type = PENDING_GC_PAR;
1388 } else {
1389 gc_type = PENDING_GC_SEQ;
1390 }
1391
1392 // In order to GC, there must be no threads running Haskell code.
1393 // Therefore, the GC thread needs to hold *all* the capabilities,
1394 // and release them after the GC has completed.
1395 //
1396 // This seems to be the simplest way: previous attempts involved
1397 // making all the threads with capabilities give up their
1398 // capabilities and sleep except for the *last* one, which
1399 // actually did the GC. But it's quite hard to arrange for all
1400 // the other tasks to sleep and stay asleep.
1401 //
1402
1403 /* Other capabilities are prevented from running yet more Haskell
1404 threads if waiting_for_gc is set. Tested inside
1405 yieldCapability() and releaseCapability() in Capability.c */
1406
1407 prev_pending_gc = cas(&waiting_for_gc, 0, gc_type);
1408 if (prev_pending_gc) {
1409 do {
1410 debugTrace(DEBUG_sched, "someone else is trying to GC (%d)...",
1411 prev_pending_gc);
1412 ASSERT(cap);
1413 yieldCapability(&cap,task);
1414 } while (waiting_for_gc);
1415 return cap; // NOTE: task->cap might have changed here
1416 }
1417
1418 setContextSwitches();
1419
1420 // The final shutdown GC is always single-threaded, because it's
1421 // possible that some of the Capabilities have no worker threads.
1422
1423 if (gc_type == PENDING_GC_SEQ)
1424 {
1425 traceEventRequestSeqGc(cap);
1426 }
1427 else
1428 {
1429 traceEventRequestParGc(cap);
1430 debugTrace(DEBUG_sched, "ready_to_gc, grabbing GC threads");
1431 }
1432
1433 // do this while the other Capabilities stop:
1434 if (cap) scheduleCheckBlackHoles(cap);
1435
1436 if (gc_type == PENDING_GC_SEQ)
1437 {
1438 // single-threaded GC: grab all the capabilities
1439 for (i=0; i < n_capabilities; i++) {
1440 debugTrace(DEBUG_sched, "ready_to_gc, grabbing all the capabilies (%d/%d)", i, n_capabilities);
1441 if (cap != &capabilities[i]) {
1442 Capability *pcap = &capabilities[i];
1443 // we better hope this task doesn't get migrated to
1444 // another Capability while we're waiting for this one.
1445 // It won't, because load balancing happens while we have
1446 // all the Capabilities, but even so it's a slightly
1447 // unsavoury invariant.
1448 task->cap = pcap;
1449 waitForReturnCapability(&pcap, task);
1450 if (pcap != &capabilities[i]) {
1451 barf("scheduleDoGC: got the wrong capability");
1452 }
1453 }
1454 }
1455 }
1456 else
1457 {
1458 // multi-threaded GC: make sure all the Capabilities donate one
1459 // GC thread each.
1460 waitForGcThreads(cap);
1461 }
1462
1463 #else /* !THREADED_RTS */
1464
1465 // do this while the other Capabilities stop:
1466 if (cap) scheduleCheckBlackHoles(cap);
1467
1468 #endif
1469
1470 IF_DEBUG(scheduler, printAllThreads());
1471
1472 delete_threads_and_gc:
1473 /*
1474 * We now have all the capabilities; if we're in an interrupting
1475 * state, then we should take the opportunity to delete all the
1476 * threads in the system.
1477 */
1478 if (sched_state == SCHED_INTERRUPTING) {
1479 deleteAllThreads(cap);
1480 sched_state = SCHED_SHUTTING_DOWN;
1481 }
1482
1483 heap_census = scheduleNeedHeapProfile(rtsTrue);
1484
1485 traceEventGcStart(cap);
1486 #if defined(THREADED_RTS)
1487 // reset waiting_for_gc *before* GC, so that when the GC threads
1488 // emerge they don't immediately re-enter the GC.
1489 waiting_for_gc = 0;
1490 GarbageCollect(force_major || heap_census, gc_type, cap);
1491 #else
1492 GarbageCollect(force_major || heap_census, 0, cap);
1493 #endif
1494 traceEventGcEnd(cap);
1495
1496 if (recent_activity == ACTIVITY_INACTIVE && force_major)
1497 {
1498 // We are doing a GC because the system has been idle for a
1499 // timeslice and we need to check for deadlock. Record the
1500 // fact that we've done a GC and turn off the timer signal;
1501 // it will get re-enabled if we run any threads after the GC.
1502 recent_activity = ACTIVITY_DONE_GC;
1503 stopTimer();
1504 }
1505 else
1506 {
1507 // the GC might have taken long enough for the timer to set
1508 // recent_activity = ACTIVITY_INACTIVE, but we aren't
1509 // necessarily deadlocked:
1510 recent_activity = ACTIVITY_YES;
1511 }
1512
1513 #if defined(THREADED_RTS)
1514 if (gc_type == PENDING_GC_PAR)
1515 {
1516 releaseGCThreads(cap);
1517 }
1518 #endif
1519
1520 if (heap_census) {
1521 debugTrace(DEBUG_sched, "performing heap census");
1522 heapCensus();
1523 performHeapProfile = rtsFalse;
1524 }
1525
1526 if (heap_overflow && sched_state < SCHED_INTERRUPTING) {
1527 // GC set the heap_overflow flag, so we should proceed with
1528 // an orderly shutdown now. Ultimately we want the main
1529 // thread to return to its caller with HeapExhausted, at which
1530 // point the caller should call hs_exit(). The first step is
1531 // to delete all the threads.
1532 //
1533 // Another way to do this would be to raise an exception in
1534 // the main thread, which we really should do because it gives
1535 // the program a chance to clean up. But how do we find the
1536 // main thread? It should presumably be the same one that
1537 // gets ^C exceptions, but that's all done on the Haskell side
1538 // (GHC.TopHandler).
1539 sched_state = SCHED_INTERRUPTING;
1540 goto delete_threads_and_gc;
1541 }
1542
1543 #ifdef SPARKBALANCE
1544 /* JB
1545 Once we are all together... this would be the place to balance all
1546 spark pools. No concurrent stealing or adding of new sparks can
1547 occur. Should be defined in Sparks.c. */
1548 balanceSparkPoolsCaps(n_capabilities, capabilities);
1549 #endif
1550
1551 #if defined(THREADED_RTS)
1552 if (gc_type == PENDING_GC_SEQ) {
1553 // release our stash of capabilities.
1554 for (i = 0; i < n_capabilities; i++) {
1555 if (cap != &capabilities[i]) {
1556 task->cap = &capabilities[i];
1557 releaseCapability(&capabilities[i]);
1558 }
1559 }
1560 }
1561 if (cap) {
1562 task->cap = cap;
1563 } else {
1564 task->cap = NULL;
1565 }
1566 #endif
1567
1568 return cap;
1569 }
1570
1571 /* ---------------------------------------------------------------------------
1572 * Singleton fork(). Do not copy any running threads.
1573 * ------------------------------------------------------------------------- */
1574
1575 pid_t
1576 forkProcess(HsStablePtr *entry
1577 #ifndef FORKPROCESS_PRIMOP_SUPPORTED
1578 STG_UNUSED
1579 #endif
1580 )
1581 {
1582 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
1583 Task *task;
1584 pid_t pid;
1585 StgTSO* t,*next;
1586 Capability *cap;
1587 nat g;
1588
1589 #if defined(THREADED_RTS)
1590 if (RtsFlags.ParFlags.nNodes > 1) {
1591 errorBelch("forking not supported with +RTS -N<n> greater than 1");
1592 stg_exit(EXIT_FAILURE);
1593 }
1594 #endif
1595
1596 debugTrace(DEBUG_sched, "forking!");
1597
1598 // ToDo: for SMP, we should probably acquire *all* the capabilities
1599 cap = rts_lock();
1600
1601 // no funny business: hold locks while we fork, otherwise if some
1602 // other thread is holding a lock when the fork happens, the data
1603 // structure protected by the lock will forever be in an
1604 // inconsistent state in the child. See also #1391.
1605 ACQUIRE_LOCK(&sched_mutex);
1606 ACQUIRE_LOCK(&cap->lock);
1607 ACQUIRE_LOCK(&cap->running_task->lock);
1608
1609 pid = fork();
1610
1611 if (pid) { // parent
1612
1613 RELEASE_LOCK(&sched_mutex);
1614 RELEASE_LOCK(&cap->lock);
1615 RELEASE_LOCK(&cap->running_task->lock);
1616
1617 // just return the pid
1618 rts_unlock(cap);
1619 return pid;
1620
1621 } else { // child
1622
1623 #if defined(THREADED_RTS)
1624 initMutex(&sched_mutex);
1625 initMutex(&cap->lock);
1626 initMutex(&cap->running_task->lock);
1627 #endif
1628
1629 // Now, all OS threads except the thread that forked are
1630 // stopped. We need to stop all Haskell threads, including
1631 // those involved in foreign calls. Also we need to delete
1632 // all Tasks, because they correspond to OS threads that are
1633 // now gone.
1634
1635 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1636 for (t = generations[g].threads; t != END_TSO_QUEUE; t = next) {
1637 if (t->what_next == ThreadRelocated) {
1638 next = t->_link;
1639 } else {
1640 next = t->global_link;
1641 // don't allow threads to catch the ThreadKilled
1642 // exception, but we do want to raiseAsync() because these
1643 // threads may be evaluating thunks that we need later.
1644 deleteThread_(cap,t);
1645 }
1646 }
1647 }
1648
1649 // Empty the run queue. It seems tempting to let all the
1650 // killed threads stay on the run queue as zombies to be
1651 // cleaned up later, but some of them correspond to bound
1652 // threads for which the corresponding Task does not exist.
1653 cap->run_queue_hd = END_TSO_QUEUE;
1654 cap->run_queue_tl = END_TSO_QUEUE;
1655
1656 // Any suspended C-calling Tasks are no more, their OS threads
1657 // don't exist now:
1658 cap->suspended_ccalling_tasks = NULL;
1659
1660 // Empty the threads lists. Otherwise, the garbage
1661 // collector may attempt to resurrect some of these threads.
1662 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1663 generations[g].threads = END_TSO_QUEUE;
1664 }
1665
1666 // Wipe the task list, except the current Task.
1667 ACQUIRE_LOCK(&sched_mutex);
1668 for (task = all_tasks; task != NULL; task=task->all_link) {
1669 if (task != cap->running_task) {
1670 #if defined(THREADED_RTS)
1671 initMutex(&task->lock); // see #1391
1672 #endif
1673 discardTask(task);
1674 }
1675 }
1676 RELEASE_LOCK(&sched_mutex);
1677
1678 #if defined(THREADED_RTS)
1679 // Wipe our spare workers list, they no longer exist. New
1680 // workers will be created if necessary.
1681 cap->spare_workers = NULL;
1682 cap->returning_tasks_hd = NULL;
1683 cap->returning_tasks_tl = NULL;
1684 #endif
1685
1686 // On Unix, all timers are reset in the child, so we need to start
1687 // the timer again.
1688 initTimer();
1689 startTimer();
1690
1691 #if defined(THREADED_RTS)
1692 cap = ioManagerStartCap(cap);
1693 #endif
1694
1695 cap = rts_evalStableIO(cap, entry, NULL); // run the action
1696 rts_checkSchedStatus("forkProcess",cap);
1697
1698 rts_unlock(cap);
1699 hs_exit(); // clean up and exit
1700 stg_exit(EXIT_SUCCESS);
1701 }
1702 #else /* !FORKPROCESS_PRIMOP_SUPPORTED */
1703 barf("forkProcess#: primop not supported on this platform, sorry!\n");
1704 #endif
1705 }
1706
1707 /* ---------------------------------------------------------------------------
1708 * Delete all the threads in the system
1709 * ------------------------------------------------------------------------- */
1710
1711 static void
1712 deleteAllThreads ( Capability *cap )
1713 {
1714 // NOTE: only safe to call if we own all capabilities.
1715
1716 StgTSO* t, *next;
1717 nat g;
1718
1719 debugTrace(DEBUG_sched,"deleting all threads");
1720 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1721 for (t = generations[g].threads; t != END_TSO_QUEUE; t = next) {
1722 if (t->what_next == ThreadRelocated) {
1723 next = t->_link;
1724 } else {
1725 next = t->global_link;
1726 deleteThread(cap,t);
1727 }
1728 }
1729 }
1730
1731 // The run queue now contains a bunch of ThreadKilled threads. We
1732 // must not throw these away: the main thread(s) will be in there
1733 // somewhere, and the main scheduler loop has to deal with it.
1734 // Also, the run queue is the only thing keeping these threads from
1735 // being GC'd, and we don't want the "main thread has been GC'd" panic.
1736
1737 #if !defined(THREADED_RTS)
1738 ASSERT(blocked_queue_hd == END_TSO_QUEUE);
1739 ASSERT(sleeping_queue == END_TSO_QUEUE);
1740 #endif
1741 }
1742
1743 /* -----------------------------------------------------------------------------
1744 Managing the suspended_ccalling_tasks list.
1745 Locks required: sched_mutex
1746 -------------------------------------------------------------------------- */
1747
1748 STATIC_INLINE void
1749 suspendTask (Capability *cap, Task *task)
1750 {
1751 ASSERT(task->next == NULL && task->prev == NULL);
1752 task->next = cap->suspended_ccalling_tasks;
1753 task->prev = NULL;
1754 if (cap->suspended_ccalling_tasks) {
1755 cap->suspended_ccalling_tasks->prev = task;
1756 }
1757 cap->suspended_ccalling_tasks = task;
1758 }
1759
1760 STATIC_INLINE void
1761 recoverSuspendedTask (Capability *cap, Task *task)
1762 {
1763 if (task->prev) {
1764 task->prev->next = task->next;
1765 } else {
1766 ASSERT(cap->suspended_ccalling_tasks == task);
1767 cap->suspended_ccalling_tasks = task->next;
1768 }
1769 if (task->next) {
1770 task->next->prev = task->prev;
1771 }
1772 task->next = task->prev = NULL;
1773 }
1774
1775 /* ---------------------------------------------------------------------------
1776 * Suspending & resuming Haskell threads.
1777 *
1778 * When making a "safe" call to C (aka _ccall_GC), the task gives back
1779 * its capability before calling the C function. This allows another
1780 * task to pick up the capability and carry on running Haskell
1781 * threads. It also means that if the C call blocks, it won't lock
1782 * the whole system.
1783 *
1784 * The Haskell thread making the C call is put to sleep for the
1785 * duration of the call, on the susepended_ccalling_threads queue. We
1786 * give out a token to the task, which it can use to resume the thread
1787 * on return from the C function.
1788 * ------------------------------------------------------------------------- */
1789
1790 void *
1791 suspendThread (StgRegTable *reg)
1792 {
1793 Capability *cap;
1794 int saved_errno;
1795 StgTSO *tso;
1796 Task *task;
1797 #if mingw32_HOST_OS
1798 StgWord32 saved_winerror;
1799 #endif
1800
1801 saved_errno = errno;
1802 #if mingw32_HOST_OS
1803 saved_winerror = GetLastError();
1804 #endif
1805
1806 /* assume that *reg is a pointer to the StgRegTable part of a Capability.
1807 */
1808 cap = regTableToCapability(reg);
1809
1810 task = cap->running_task;
1811 tso = cap->r.rCurrentTSO;
1812
1813 traceEventStopThread(cap, tso, THREAD_SUSPENDED_FOREIGN_CALL);
1814
1815 // XXX this might not be necessary --SDM
1816 tso->what_next = ThreadRunGHC;
1817
1818 threadPaused(cap,tso);
1819
1820 if ((tso->flags & TSO_BLOCKEX) == 0) {
1821 tso->why_blocked = BlockedOnCCall;
1822 tso->flags |= TSO_BLOCKEX;
1823 tso->flags &= ~TSO_INTERRUPTIBLE;
1824 } else {
1825 tso->why_blocked = BlockedOnCCall_NoUnblockExc;
1826 }
1827
1828 // Hand back capability
1829 task->suspended_tso = tso;
1830
1831 ACQUIRE_LOCK(&cap->lock);
1832
1833 suspendTask(cap,task);
1834 cap->in_haskell = rtsFalse;
1835 releaseCapability_(cap,rtsFalse);
1836
1837 RELEASE_LOCK(&cap->lock);
1838
1839 errno = saved_errno;
1840 #if mingw32_HOST_OS
1841 SetLastError(saved_winerror);
1842 #endif
1843 return task;
1844 }
1845
1846 StgRegTable *
1847 resumeThread (void *task_)
1848 {
1849 StgTSO *tso;
1850 Capability *cap;
1851 Task *task = task_;
1852 int saved_errno;
1853 #if mingw32_HOST_OS
1854 StgWord32 saved_winerror;
1855 #endif
1856
1857 saved_errno = errno;
1858 #if mingw32_HOST_OS
1859 saved_winerror = GetLastError();
1860 #endif
1861
1862 cap = task->cap;
1863 // Wait for permission to re-enter the RTS with the result.
1864 waitForReturnCapability(&cap,task);
1865 // we might be on a different capability now... but if so, our
1866 // entry on the suspended_ccalling_tasks list will also have been
1867 // migrated.
1868
1869 // Remove the thread from the suspended list
1870 recoverSuspendedTask(cap,task);
1871
1872 tso = task->suspended_tso;
1873 task->suspended_tso = NULL;
1874 tso->_link = END_TSO_QUEUE; // no write barrier reqd
1875
1876 traceEventRunThread(cap, tso);
1877
1878 if (tso->why_blocked == BlockedOnCCall) {
1879 // avoid locking the TSO if we don't have to
1880 if (tso->blocked_exceptions != END_TSO_QUEUE) {
1881 awakenBlockedExceptionQueue(cap,tso);
1882 }
1883 tso->flags &= ~(TSO_BLOCKEX | TSO_INTERRUPTIBLE);
1884 }
1885
1886 /* Reset blocking status */
1887 tso->why_blocked = NotBlocked;
1888
1889 cap->r.rCurrentTSO = tso;
1890 cap->in_haskell = rtsTrue;
1891 errno = saved_errno;
1892 #if mingw32_HOST_OS
1893 SetLastError(saved_winerror);
1894 #endif
1895
1896 /* We might have GC'd, mark the TSO dirty again */
1897 dirty_TSO(cap,tso);
1898
1899 IF_DEBUG(sanity, checkTSO(tso));
1900
1901 return &cap->r;
1902 }
1903
1904 /* ---------------------------------------------------------------------------
1905 * scheduleThread()
1906 *
1907 * scheduleThread puts a thread on the end of the runnable queue.
1908 * This will usually be done immediately after a thread is created.
1909 * The caller of scheduleThread must create the thread using e.g.
1910 * createThread and push an appropriate closure
1911 * on this thread's stack before the scheduler is invoked.
1912 * ------------------------------------------------------------------------ */
1913
1914 void
1915 scheduleThread(Capability *cap, StgTSO *tso)
1916 {
1917 // The thread goes at the *end* of the run-queue, to avoid possible
1918 // starvation of any threads already on the queue.
1919 appendToRunQueue(cap,tso);
1920 }
1921
1922 void
1923 scheduleThreadOn(Capability *cap, StgWord cpu USED_IF_THREADS, StgTSO *tso)
1924 {
1925 #if defined(THREADED_RTS)
1926 tso->flags |= TSO_LOCKED; // we requested explicit affinity; don't
1927 // move this thread from now on.
1928 cpu %= RtsFlags.ParFlags.nNodes;
1929 if (cpu == cap->no) {
1930 appendToRunQueue(cap,tso);
1931 } else {
1932 traceEventMigrateThread (cap, tso, capabilities[cpu].no);
1933 wakeupThreadOnCapability(cap, &capabilities[cpu], tso);
1934 }
1935 #else
1936 appendToRunQueue(cap,tso);
1937 #endif
1938 }
1939
1940 Capability *
1941 scheduleWaitThread (StgTSO* tso, /*[out]*/HaskellObj* ret, Capability *cap)
1942 {
1943 Task *task;
1944
1945 // We already created/initialised the Task
1946 task = cap->running_task;
1947
1948 // This TSO is now a bound thread; make the Task and TSO
1949 // point to each other.
1950 tso->bound = task;
1951 tso->cap = cap;
1952
1953 task->tso = tso;
1954 task->ret = ret;
1955 task->stat = NoStatus;
1956
1957 appendToRunQueue(cap,tso);
1958
1959 debugTrace(DEBUG_sched, "new bound thread (%lu)", (unsigned long)tso->id);
1960
1961 cap = schedule(cap,task);
1962
1963 ASSERT(task->stat != NoStatus);
1964 ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
1965
1966 debugTrace(DEBUG_sched, "bound thread (%lu) finished", (unsigned long)task->tso->id);
1967 return cap;
1968 }
1969
1970 /* ----------------------------------------------------------------------------
1971 * Starting Tasks
1972 * ------------------------------------------------------------------------- */
1973
1974 #if defined(THREADED_RTS)
1975 void OSThreadProcAttr
1976 workerStart(Task *task)
1977 {
1978 Capability *cap;
1979
1980 // See startWorkerTask().
1981 ACQUIRE_LOCK(&task->lock);
1982 cap = task->cap;
1983 RELEASE_LOCK(&task->lock);
1984
1985 if (RtsFlags.ParFlags.setAffinity) {
1986 setThreadAffinity(cap->no, n_capabilities);
1987 }
1988
1989 // set the thread-local pointer to the Task:
1990 taskEnter(task);
1991
1992 // schedule() runs without a lock.
1993 cap = schedule(cap,task);
1994
1995 // On exit from schedule(), we have a Capability, but possibly not
1996 // the same one we started with.
1997
1998 // During shutdown, the requirement is that after all the
1999 // Capabilities are shut down, all workers that are shutting down
2000 // have finished workerTaskStop(). This is why we hold on to
2001 // cap->lock until we've finished workerTaskStop() below.
2002 //
2003 // There may be workers still involved in foreign calls; those
2004 // will just block in waitForReturnCapability() because the
2005 // Capability has been shut down.
2006 //
2007 ACQUIRE_LOCK(&cap->lock);
2008 releaseCapability_(cap,rtsFalse);
2009 workerTaskStop(task);
2010 RELEASE_LOCK(&cap->lock);
2011 }
2012 #endif
2013
2014 /* ---------------------------------------------------------------------------
2015 * initScheduler()
2016 *
2017 * Initialise the scheduler. This resets all the queues - if the
2018 * queues contained any threads, they'll be garbage collected at the
2019 * next pass.
2020 *
2021 * ------------------------------------------------------------------------ */
2022
2023 void
2024 initScheduler(void)
2025 {
2026 #if !defined(THREADED_RTS)
2027 blocked_queue_hd = END_TSO_QUEUE;
2028 blocked_queue_tl = END_TSO_QUEUE;
2029 sleeping_queue = END_TSO_QUEUE;
2030 #endif
2031
2032 blackhole_queue = END_TSO_QUEUE;
2033
2034 sched_state = SCHED_RUNNING;
2035 recent_activity = ACTIVITY_YES;
2036
2037 #if defined(THREADED_RTS)
2038 /* Initialise the mutex and condition variables used by
2039 * the scheduler. */
2040 initMutex(&sched_mutex);
2041 #endif
2042
2043 ACQUIRE_LOCK(&sched_mutex);
2044
2045 /* A capability holds the state a native thread needs in
2046 * order to execute STG code. At least one capability is
2047 * floating around (only THREADED_RTS builds have more than one).
2048 */
2049 initCapabilities();
2050
2051 initTaskManager();
2052
2053 #if defined(THREADED_RTS)
2054 initSparkPools();
2055 #endif
2056
2057 #if defined(THREADED_RTS)
2058 /*
2059 * Eagerly start one worker to run each Capability, except for
2060 * Capability 0. The idea is that we're probably going to start a
2061 * bound thread on Capability 0 pretty soon, so we don't want a
2062 * worker task hogging it.
2063 */
2064 {
2065 nat i;
2066 Capability *cap;
2067 for (i = 1; i < n_capabilities; i++) {
2068 cap = &capabilities[i];
2069 ACQUIRE_LOCK(&cap->lock);
2070 startWorkerTask(cap, workerStart);
2071 RELEASE_LOCK(&cap->lock);
2072 }
2073 }
2074 #endif
2075
2076 RELEASE_LOCK(&sched_mutex);
2077 }
2078
2079 void
2080 exitScheduler(
2081 rtsBool wait_foreign
2082 #if !defined(THREADED_RTS)
2083 __attribute__((unused))
2084 #endif
2085 )
2086 /* see Capability.c, shutdownCapability() */
2087 {
2088 Task *task = NULL;
2089
2090 task = newBoundTask();
2091
2092 // If we haven't killed all the threads yet, do it now.
2093 if (sched_state < SCHED_SHUTTING_DOWN) {
2094 sched_state = SCHED_INTERRUPTING;
2095 waitForReturnCapability(&task->cap,task);
2096 scheduleDoGC(task->cap,task,rtsFalse);
2097 releaseCapability(task->cap);
2098 }
2099 sched_state = SCHED_SHUTTING_DOWN;
2100
2101 #if defined(THREADED_RTS)
2102 {
2103 nat i;
2104
2105 for (i = 0; i < n_capabilities; i++) {
2106 shutdownCapability(&capabilities[i], task, wait_foreign);
2107 }
2108 }
2109 #endif
2110
2111 boundTaskExiting(task);
2112 }
2113
2114 void
2115 freeScheduler( void )
2116 {
2117 nat still_running;
2118
2119 ACQUIRE_LOCK(&sched_mutex);
2120 still_running = freeTaskManager();
2121 // We can only free the Capabilities if there are no Tasks still
2122 // running. We might have a Task about to return from a foreign
2123 // call into waitForReturnCapability(), for example (actually,
2124 // this should be the *only* thing that a still-running Task can
2125 // do at this point, and it will block waiting for the
2126 // Capability).
2127 if (still_running == 0) {
2128 freeCapabilities();
2129 if (n_capabilities != 1) {
2130 stgFree(capabilities);
2131 }
2132 }
2133 RELEASE_LOCK(&sched_mutex);
2134 #if defined(THREADED_RTS)
2135 closeMutex(&sched_mutex);
2136 #endif
2137 }
2138
2139 /* -----------------------------------------------------------------------------
2140 performGC
2141
2142 This is the interface to the garbage collector from Haskell land.
2143 We provide this so that external C code can allocate and garbage
2144 collect when called from Haskell via _ccall_GC.
2145 -------------------------------------------------------------------------- */
2146
2147 static void
2148 performGC_(rtsBool force_major)
2149 {
2150 Task *task;
2151
2152 // We must grab a new Task here, because the existing Task may be
2153 // associated with a particular Capability, and chained onto the
2154 // suspended_ccalling_tasks queue.
2155 task = newBoundTask();
2156
2157 waitForReturnCapability(&task->cap,task);
2158 scheduleDoGC(task->cap,task,force_major);
2159 releaseCapability(task->cap);
2160 boundTaskExiting(task);
2161 }
2162
2163 void
2164 performGC(void)
2165 {
2166 performGC_(rtsFalse);
2167 }
2168
2169 void
2170 performMajorGC(void)
2171 {
2172 performGC_(rtsTrue);
2173 }
2174
2175 /* -----------------------------------------------------------------------------
2176 Stack overflow
2177
2178 If the thread has reached its maximum stack size, then raise the
2179 StackOverflow exception in the offending thread. Otherwise
2180 relocate the TSO into a larger chunk of memory and adjust its stack
2181 size appropriately.
2182 -------------------------------------------------------------------------- */
2183
2184 static StgTSO *
2185 threadStackOverflow(Capability *cap, StgTSO *tso)
2186 {
2187 nat new_stack_size, stack_words;
2188 lnat new_tso_size;
2189 StgPtr new_sp;
2190 StgTSO *dest;
2191
2192 IF_DEBUG(sanity,checkTSO(tso));
2193
2194 // don't allow throwTo() to modify the blocked_exceptions queue
2195 // while we are moving the TSO:
2196 lockClosure((StgClosure *)tso);
2197
2198 if (tso->stack_size >= tso->max_stack_size
2199 && !(tso->flags & TSO_BLOCKEX)) {
2200 // NB. never raise a StackOverflow exception if the thread is
2201 // inside Control.Exceptino.block. It is impractical to protect
2202 // against stack overflow exceptions, since virtually anything
2203 // can raise one (even 'catch'), so this is the only sensible
2204 // thing to do here. See bug #767.
2205 //
2206
2207 if (tso->flags & TSO_SQUEEZED) {
2208 unlockTSO(tso);
2209 return tso;
2210 }
2211 // #3677: In a stack overflow situation, stack squeezing may
2212 // reduce the stack size, but we don't know whether it has been
2213 // reduced enough for the stack check to succeed if we try
2214 // again. Fortunately stack squeezing is idempotent, so all we
2215 // need to do is record whether *any* squeezing happened. If we
2216 // are at the stack's absolute -K limit, and stack squeezing
2217 // happened, then we try running the thread again. The
2218 // TSO_SQUEEZED flag is set by threadPaused() to tell us whether
2219 // squeezing happened or not.
2220
2221 debugTrace(DEBUG_gc,
2222 "threadStackOverflow of TSO %ld (%p): stack too large (now %ld; max is %ld)",
2223 (long)tso->id, tso, (long)tso->stack_size, (long)tso->max_stack_size);
2224 IF_DEBUG(gc,
2225 /* If we're debugging, just print out the top of the stack */
2226 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2227 tso->sp+64)));
2228
2229 // Send this thread the StackOverflow exception
2230 unlockTSO(tso);
2231 throwToSingleThreaded(cap, tso, (StgClosure *)stackOverflow_closure);
2232 return tso;
2233 }
2234
2235
2236 // We also want to avoid enlarging the stack if squeezing has
2237 // already released some of it. However, we don't want to get into
2238 // a pathalogical situation where a thread has a nearly full stack
2239 // (near its current limit, but not near the absolute -K limit),
2240 // keeps allocating a little bit, squeezing removes a little bit,
2241 // and then it runs again. So to avoid this, if we squeezed *and*
2242 // there is still less than BLOCK_SIZE_W words free, then we enlarge
2243 // the stack anyway.
2244 if ((tso->flags & TSO_SQUEEZED) &&
2245 ((W_)(tso->sp - tso->stack) >= BLOCK_SIZE_W)) {
2246 unlockTSO(tso);
2247 return tso;
2248 }
2249
2250 /* Try to double the current stack size. If that takes us over the
2251 * maximum stack size for this thread, then use the maximum instead
2252 * (that is, unless we're already at or over the max size and we
2253 * can't raise the StackOverflow exception (see above), in which
2254 * case just double the size). Finally round up so the TSO ends up as
2255 * a whole number of blocks.
2256 */
2257 if (tso->stack_size >= tso->max_stack_size) {
2258 new_stack_size = tso->stack_size * 2;
2259 } else {
2260 new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
2261 }
2262 new_tso_size = (lnat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) +
2263 TSO_STRUCT_SIZE)/sizeof(W_);
2264 new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */
2265 new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
2266
2267 debugTrace(DEBUG_sched,
2268 "increasing stack size from %ld words to %d.",
2269 (long)tso->stack_size, new_stack_size);
2270
2271 dest = (StgTSO *)allocate(cap,new_tso_size);
2272 TICK_ALLOC_TSO(new_stack_size,0);
2273
2274 /* copy the TSO block and the old stack into the new area */
2275 memcpy(dest,tso,TSO_STRUCT_SIZE);
2276 stack_words = tso->stack + tso->stack_size - tso->sp;
2277 new_sp = (P_)dest + new_tso_size - stack_words;
2278 memcpy(new_sp, tso->sp, stack_words * sizeof(W_));
2279
2280 /* relocate the stack pointers... */
2281 dest->sp = new_sp;
2282 dest->stack_size = new_stack_size;
2283
2284 /* Mark the old TSO as relocated. We have to check for relocated
2285 * TSOs in the garbage collector and any primops that deal with TSOs.
2286 *
2287 * It's important to set the sp value to just beyond the end
2288 * of the stack, so we don't attempt to scavenge any part of the
2289 * dead TSO's stack.
2290 */
2291 tso->what_next = ThreadRelocated;
2292 setTSOLink(cap,tso,dest);
2293 tso->sp = (P_)&(tso->stack[tso->stack_size]);
2294 tso->why_blocked = NotBlocked;
2295
2296 unlockTSO(dest);
2297 unlockTSO(tso);
2298
2299 IF_DEBUG(sanity,checkTSO(dest));
2300 #if 0
2301 IF_DEBUG(scheduler,printTSO(dest));
2302 #endif
2303
2304 return dest;
2305 }
2306
2307 static StgTSO *
2308 threadStackUnderflow (Capability *cap, Task *task, StgTSO *tso)
2309 {
2310 bdescr *bd, *new_bd;
2311 lnat free_w, tso_size_w;
2312 StgTSO *new_tso;
2313
2314 tso_size_w = tso_sizeW(tso);
2315
2316 if (tso_size_w < MBLOCK_SIZE_W ||
2317 // TSO is less than 2 mblocks (since the first mblock is
2318 // shorter than MBLOCK_SIZE_W)
2319 (tso_size_w - BLOCKS_PER_MBLOCK*BLOCK_SIZE_W) % MBLOCK_SIZE_W != 0 ||
2320 // or TSO is not a whole number of megablocks (ensuring
2321 // precondition of splitLargeBlock() below)
2322 (tso_size_w <= round_up_to_mblocks(RtsFlags.GcFlags.initialStkSize)) ||
2323 // or TSO is smaller than the minimum stack size (rounded up)
2324 (nat)(tso->stack + tso->stack_size - tso->sp) > tso->stack_size / 4)
2325 // or stack is using more than 1/4 of the available space
2326 {
2327 // then do nothing
2328 return tso;
2329 }
2330
2331 // don't allow throwTo() to modify the blocked_exceptions queue
2332 // while we are moving the TSO:
2333 lockClosure((StgClosure *)tso);
2334
2335 // this is the number of words we'll free
2336 free_w = round_to_mblocks(tso_size_w/2);
2337
2338 bd = Bdescr((StgPtr)tso);
2339 new_bd = splitLargeBlock(bd, free_w / BLOCK_SIZE_W);
2340 bd->free = bd->start + TSO_STRUCT_SIZEW;
2341
2342 new_tso = (StgTSO *)new_bd->start;
2343 memcpy(new_tso,tso,TSO_STRUCT_SIZE);
2344 new_tso->stack_size = new_bd->free - new_tso->stack;
2345
2346 // The original TSO was dirty and probably on the mutable
2347 // list. The new TSO is not yet on the mutable list, so we better
2348 // put it there.
2349 new_tso->dirty = 0;
2350 new_tso->flags &= ~TSO_LINK_DIRTY;
2351 dirty_TSO(cap, new_tso);
2352
2353 debugTrace(DEBUG_sched, "thread %ld: reducing TSO size from %lu words to %lu",
2354 (long)tso->id, tso_size_w, tso_sizeW(new_tso));
2355
2356 tso->what_next = ThreadRelocated;
2357 tso->_link = new_tso; // no write barrier reqd: same generation
2358
2359 // The TSO attached to this Task may have moved, so update the
2360 // pointer to it.
2361 if (task->tso == tso) {
2362 task->tso = new_tso;
2363 }
2364
2365 unlockTSO(new_tso);
2366 unlockTSO(tso);
2367
2368 IF_DEBUG(sanity,checkTSO(new_tso));
2369
2370 return new_tso;
2371 }
2372
2373 /* ---------------------------------------------------------------------------
2374 Interrupt execution
2375 - usually called inside a signal handler so it mustn't do anything fancy.
2376 ------------------------------------------------------------------------ */
2377
2378 void
2379 interruptStgRts(void)
2380 {
2381 sched_state = SCHED_INTERRUPTING;
2382 setContextSwitches();
2383 #if defined(THREADED_RTS)
2384 wakeUpRts();
2385 #endif
2386 }
2387
2388 /* -----------------------------------------------------------------------------
2389 Wake up the RTS
2390
2391 This function causes at least one OS thread to wake up and run the
2392 scheduler loop. It is invoked when the RTS might be deadlocked, or
2393 an external event has arrived that may need servicing (eg. a
2394 keyboard interrupt).
2395
2396 In the single-threaded RTS we don't do anything here; we only have
2397 one thread anyway, and the event that caused us to want to wake up
2398 will have interrupted any blocking system call in progress anyway.
2399 -------------------------------------------------------------------------- */
2400
2401 #if defined(THREADED_RTS)
2402 void wakeUpRts(void)
2403 {
2404 // This forces the IO Manager thread to wakeup, which will
2405 // in turn ensure that some OS thread wakes up and runs the
2406 // scheduler loop, which will cause a GC and deadlock check.
2407 ioManagerWakeup();
2408 }
2409 #endif
2410
2411 /* -----------------------------------------------------------------------------
2412 * checkBlackHoles()
2413 *
2414 * Check the blackhole_queue for threads that can be woken up. We do
2415 * this periodically: before every GC, and whenever the run queue is
2416 * empty.
2417 *
2418 * An elegant solution might be to just wake up all the blocked
2419 * threads with awakenBlockedQueue occasionally: they'll go back to
2420 * sleep again if the object is still a BLACKHOLE. Unfortunately this
2421 * doesn't give us a way to tell whether we've actually managed to
2422 * wake up any threads, so we would be busy-waiting.
2423 *
2424 * -------------------------------------------------------------------------- */
2425
2426 static rtsBool
2427 checkBlackHoles (Capability *cap)
2428 {
2429 StgTSO **prev, *t;
2430 rtsBool any_woke_up = rtsFalse;
2431 StgHalfWord type;
2432
2433 // blackhole_queue is global:
2434 ASSERT_LOCK_HELD(&sched_mutex);
2435
2436 debugTrace(DEBUG_sched, "checking threads blocked on black holes");
2437
2438 // ASSUMES: sched_mutex
2439 prev = &blackhole_queue;
2440 t = blackhole_queue;
2441 while (t != END_TSO_QUEUE) {
2442 if (t->what_next == ThreadRelocated) {
2443 t = t->_link;
2444 continue;
2445 }
2446 ASSERT(t->why_blocked == BlockedOnBlackHole);
2447 type = get_itbl(UNTAG_CLOSURE(t->block_info.closure))->type;
2448 if (type != BLACKHOLE && type != CAF_BLACKHOLE) {
2449 IF_DEBUG(sanity,checkTSO(t));
2450 t = unblockOne(cap, t);
2451 *prev = t;
2452 any_woke_up = rtsTrue;
2453 } else {
2454 prev = &t->_link;
2455 t = t->_link;
2456 }
2457 }
2458
2459 return any_woke_up;
2460 }
2461
2462 /* -----------------------------------------------------------------------------
2463 Deleting threads
2464
2465 This is used for interruption (^C) and forking, and corresponds to
2466 raising an exception but without letting the thread catch the
2467 exception.
2468 -------------------------------------------------------------------------- */
2469
2470 static void
2471 deleteThread (Capability *cap, StgTSO *tso)
2472 {
2473 // NOTE: must only be called on a TSO that we have exclusive
2474 // access to, because we will call throwToSingleThreaded() below.
2475 // The TSO must be on the run queue of the Capability we own, or
2476 // we must own all Capabilities.
2477
2478 if (tso->why_blocked != BlockedOnCCall &&
2479 tso->why_blocked != BlockedOnCCall_NoUnblockExc) {
2480 throwToSingleThreaded(cap,tso,NULL);
2481 }
2482 }
2483
2484 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
2485 static void
2486 deleteThread_(Capability *cap, StgTSO *tso)
2487 { // for forkProcess only:
2488 // like deleteThread(), but we delete threads in foreign calls, too.
2489
2490 if (tso->why_blocked == BlockedOnCCall ||
2491 tso->why_blocked == BlockedOnCCall_NoUnblockExc) {
2492 unblockOne(cap,tso);
2493 tso->what_next = ThreadKilled;
2494 } else {
2495 deleteThread(cap,tso);
2496 }
2497 }
2498 #endif
2499
2500 /* -----------------------------------------------------------------------------
2501 raiseExceptionHelper
2502
2503 This function is called by the raise# primitve, just so that we can
2504 move some of the tricky bits of raising an exception from C-- into
2505 C. Who knows, it might be a useful re-useable thing here too.
2506 -------------------------------------------------------------------------- */
2507
2508 StgWord
2509 raiseExceptionHelper (StgRegTable *reg, StgTSO *tso, StgClosure *exception)
2510 {
2511 Capability *cap = regTableToCapability(reg);
2512 StgThunk *raise_closure = NULL;
2513 StgPtr p, next;
2514 StgRetInfoTable *info;
2515 //
2516 // This closure represents the expression 'raise# E' where E
2517 // is the exception raise. It is used to overwrite all the
2518 // thunks which are currently under evaluataion.
2519 //
2520
2521 // OLD COMMENT (we don't have MIN_UPD_SIZE now):
2522 // LDV profiling: stg_raise_info has THUNK as its closure
2523 // type. Since a THUNK takes at least MIN_UPD_SIZE words in its
2524 // payload, MIN_UPD_SIZE is more approprate than 1. It seems that
2525 // 1 does not cause any problem unless profiling is performed.
2526 // However, when LDV profiling goes on, we need to linearly scan
2527 // small object pool, where raise_closure is stored, so we should
2528 // use MIN_UPD_SIZE.
2529 //
2530 // raise_closure = (StgClosure *)RET_STGCALL1(P_,allocate,
2531 // sizeofW(StgClosure)+1);
2532 //
2533
2534 //
2535 // Walk up the stack, looking for the catch frame. On the way,
2536 // we update any closures pointed to from update frames with the
2537 // raise closure that we just built.
2538 //
2539 p = tso->sp;
2540 while(1) {
2541 info = get_ret_itbl((StgClosure *)p);
2542 next = p + stack_frame_sizeW((StgClosure *)p);
2543 switch (info->i.type) {
2544
2545 case UPDATE_FRAME:
2546 // Only create raise_closure if we need to.
2547 if (raise_closure == NULL) {
2548 raise_closure =
2549 (StgThunk *)allocate(cap,sizeofW(StgThunk)+1);
2550 SET_HDR(raise_closure, &stg_raise_info, CCCS);
2551 raise_closure->payload[0] = exception;
2552 }
2553 UPD_IND(cap, ((StgUpdateFrame *)p)->updatee,
2554 (StgClosure *)raise_closure);
2555 p = next;
2556 continue;
2557
2558 case ATOMICALLY_FRAME:
2559 debugTrace(DEBUG_stm, "found ATOMICALLY_FRAME at %p", p);
2560 tso->sp = p;
2561 return ATOMICALLY_FRAME;
2562
2563 case CATCH_FRAME:
2564 tso->sp = p;
2565 return CATCH_FRAME;
2566
2567 case CATCH_STM_FRAME:
2568 debugTrace(DEBUG_stm, "found CATCH_STM_FRAME at %p", p);
2569 tso->sp = p;
2570 return CATCH_STM_FRAME;
2571
2572 case STOP_FRAME:
2573 tso->sp = p;
2574 return STOP_FRAME;
2575
2576 case CATCH_RETRY_FRAME:
2577 default:
2578 p = next;
2579 continue;
2580 }
2581 }
2582 }
2583
2584
2585 /* -----------------------------------------------------------------------------
2586 findRetryFrameHelper
2587
2588 This function is called by the retry# primitive. It traverses the stack
2589 leaving tso->sp referring to the frame which should handle the retry.
2590
2591 This should either be a CATCH_RETRY_FRAME (if the retry# is within an orElse#)
2592 or should be a ATOMICALLY_FRAME (if the retry# reaches the top level).
2593
2594 We skip CATCH_STM_FRAMEs (aborting and rolling back the nested tx that they
2595 create) because retries are not considered to be exceptions, despite the
2596 similar implementation.
2597
2598 We should not expect to see CATCH_FRAME or STOP_FRAME because those should
2599 not be created within memory transactions.
2600 -------------------------------------------------------------------------- */
2601
2602 StgWord
2603 findRetryFrameHelper (StgTSO *tso)
2604 {
2605 StgPtr p, next;
2606 StgRetInfoTable *info;
2607
2608 p = tso -> sp;
2609 while (1) {
2610 info = get_ret_itbl((StgClosure *)p);
2611 next = p + stack_frame_sizeW((StgClosure *)p);
2612 switch (info->i.type) {
2613
2614 case ATOMICALLY_FRAME:
2615 debugTrace(DEBUG_stm,
2616 "found ATOMICALLY_FRAME at %p during retry", p);
2617 tso->sp = p;
2618 return ATOMICALLY_FRAME;
2619
2620 case CATCH_RETRY_FRAME:
2621 debugTrace(DEBUG_stm,
2622 "found CATCH_RETRY_FRAME at %p during retrry", p);
2623 tso->sp = p;
2624 return CATCH_RETRY_FRAME;
2625
2626 case CATCH_STM_FRAME: {
2627 StgTRecHeader *trec = tso -> trec;
2628 StgTRecHeader *outer = trec -> enclosing_trec;
2629 debugTrace(DEBUG_stm,
2630 "found CATCH_STM_FRAME at %p during retry", p);
2631 debugTrace(DEBUG_stm, "trec=%p outer=%p", trec, outer);
2632 stmAbortTransaction(tso -> cap, trec);
2633 stmFreeAbortedTRec(tso -> cap, trec);
2634 tso -> trec = outer;
2635 p = next;
2636 continue;
2637 }
2638
2639
2640 default:
2641 ASSERT(info->i.type != CATCH_FRAME);
2642 ASSERT(info->i.type != STOP_FRAME);
2643 p = next;
2644 continue;
2645 }
2646 }
2647 }
2648
2649 /* -----------------------------------------------------------------------------
2650 resurrectThreads is called after garbage collection on the list of
2651 threads found to be garbage. Each of these threads will be woken
2652 up and sent a signal: BlockedOnDeadMVar if the thread was blocked
2653 on an MVar, or NonTermination if the thread was blocked on a Black
2654 Hole.
2655
2656 Locks: assumes we hold *all* the capabilities.
2657 -------------------------------------------------------------------------- */
2658
2659 void
2660 resurrectThreads (StgTSO *threads)
2661 {
2662 StgTSO *tso, *next;
2663 Capability *cap;
2664 generation *gen;
2665
2666 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
2667 next = tso->global_link;
2668
2669 gen = Bdescr((P_)tso)->gen;
2670 tso->global_link = gen->threads;
2671 gen->threads = tso;
2672
2673 debugTrace(DEBUG_sched, "resurrecting thread %lu", (unsigned long)tso->id);
2674
2675 // Wake up the thread on the Capability it was last on
2676 cap = tso->cap;
2677
2678 switch (tso->why_blocked) {
2679 case BlockedOnMVar:
2680 /* Called by GC - sched_mutex lock is currently held. */
2681 throwToSingleThreaded(cap, tso,
2682 (StgClosure *)blockedIndefinitelyOnMVar_closure);
2683 break;
2684 case BlockedOnBlackHole:
2685 throwToSingleThreaded(cap, tso,
2686 (StgClosure *)nonTermination_closure);
2687 break;
2688 case BlockedOnSTM:
2689 throwToSingleThreaded(cap, tso,
2690 (StgClosure *)blockedIndefinitelyOnSTM_closure);
2691 break;
2692 case NotBlocked:
2693 /* This might happen if the thread was blocked on a black hole
2694 * belonging to a thread that we've just woken up (raiseAsync
2695 * can wake up threads, remember...).
2696 */
2697 continue;
2698 case BlockedOnException:
2699 // throwTo should never block indefinitely: if the target
2700 // thread dies or completes, throwTo returns.
2701 barf("resurrectThreads: thread BlockedOnException");
2702 break;
2703 default:
2704 barf("resurrectThreads: thread blocked in a strange way");
2705 }
2706 }
2707 }
2708
2709 /* -----------------------------------------------------------------------------
2710 performPendingThrowTos is called after garbage collection, and
2711 passed a list of threads that were found to have pending throwTos
2712 (tso->blocked_exceptions was not empty), and were blocked.
2713 Normally this doesn't happen, because we would deliver the
2714 exception directly if the target thread is blocked, but there are
2715 small windows where it might occur on a multiprocessor (see
2716 throwTo()).
2717
2718 NB. we must be holding all the capabilities at this point, just
2719 like resurrectThreads().
2720 -------------------------------------------------------------------------- */
2721
2722 void
2723 performPendingThrowTos (StgTSO *threads)
2724 {
2725 StgTSO *tso, *next;
2726 Capability *cap;
2727 Task *task, *saved_task;;
2728 generation *gen;
2729
2730 task = myTask();
2731 cap = task->cap;
2732
2733 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
2734 next = tso->global_link;
2735
2736 gen = Bdescr((P_)tso)->gen;
2737 tso->global_link = gen->threads;
2738 gen->threads = tso;
2739
2740 debugTrace(DEBUG_sched, "performing blocked throwTo to thread %lu", (unsigned long)tso->id);
2741
2742 // We must pretend this Capability belongs to the current Task
2743 // for the time being, as invariants will be broken otherwise.
2744 // In fact the current Task has exclusive access to the systme
2745 // at this point, so this is just bookkeeping:
2746 task->cap = tso->cap;
2747 saved_task = tso->cap->running_task;
2748 tso->cap->running_task = task;
2749 maybePerformBlockedException(tso->cap, tso);
2750 tso->cap->running_task = saved_task;
2751 }
2752
2753 // Restore our original Capability:
2754 task->cap = cap;
2755 }