cf85372ce08ab696a17c43950fcb003ba7bc63ca
[ghc-hetmet.git] / rts / Capability.c
1 /* ---------------------------------------------------------------------------
2  *
3  * (c) The GHC Team, 2003-2006
4  *
5  * Capabilities
6  *
7  * A Capability represent the token required to execute STG code,
8  * and all the state an OS thread/task needs to run Haskell code:
9  * its STG registers, a pointer to its TSO, a nursery etc. During
10  * STG execution, a pointer to the capabilitity is kept in a
11  * register (BaseReg; actually it is a pointer to cap->r).
12  *
13  * Only in an THREADED_RTS build will there be multiple capabilities,
14  * for non-threaded builds there is only one global capability, namely
15  * MainCapability.
16  *
17  * --------------------------------------------------------------------------*/
18
19 #include "PosixSource.h"
20 #include "Rts.h"
21
22 #include "Capability.h"
23 #include "Schedule.h"
24 #include "Sparks.h"
25 #include "Trace.h"
26 #include "sm/GC.h" // for gcWorkerThread()
27 #include "STM.h"
28 #include "RtsUtils.h"
29
30 // one global capability, this is the Capability for non-threaded
31 // builds, and for +RTS -N1
32 Capability MainCapability;
33
34 nat n_capabilities = 0;
35 Capability *capabilities = NULL;
36
37 // Holds the Capability which last became free.  This is used so that
38 // an in-call has a chance of quickly finding a free Capability.
39 // Maintaining a global free list of Capabilities would require global
40 // locking, so we don't do that.
41 Capability *last_free_capability = NULL;
42
43 /* GC indicator, in scope for the scheduler, init'ed to false */
44 volatile StgWord waiting_for_gc = 0;
45
46 /* Let foreign code get the current Capability -- assuming there is one!
47  * This is useful for unsafe foreign calls because they are called with
48  * the current Capability held, but they are not passed it. For example,
49  * see see the integer-gmp package which calls allocateLocal() in its
50  * stgAllocForGMP() function (which gets called by gmp functions).
51  * */
52 Capability * rts_unsafeGetMyCapability (void)
53 {
54 #if defined(THREADED_RTS)
55   return myTask()->cap;
56 #else
57   return &MainCapability;
58 #endif
59 }
60
61 #if defined(THREADED_RTS)
62 STATIC_INLINE rtsBool
63 globalWorkToDo (void)
64 {
65     return blackholes_need_checking
66         || sched_state >= SCHED_INTERRUPTING
67         ;
68 }
69 #endif
70
71 #if defined(THREADED_RTS)
72 StgClosure *
73 findSpark (Capability *cap)
74 {
75   Capability *robbed;
76   StgClosurePtr spark;
77   rtsBool retry;
78   nat i = 0;
79
80   if (!emptyRunQueue(cap) || cap->returning_tasks_hd != NULL) {
81       // If there are other threads, don't try to run any new
82       // sparks: sparks might be speculative, we don't want to take
83       // resources away from the main computation.
84       return 0;
85   }
86
87   do {
88       retry = rtsFalse;
89
90       // first try to get a spark from our own pool.
91       // We should be using reclaimSpark(), because it works without
92       // needing any atomic instructions:
93       //   spark = reclaimSpark(cap->sparks);
94       // However, measurements show that this makes at least one benchmark
95       // slower (prsa) and doesn't affect the others.
96       spark = tryStealSpark(cap);
97       if (spark != NULL) {
98           cap->sparks_converted++;
99
100           // Post event for running a spark from capability's own pool.
101           traceEventRunSpark(cap, cap->r.rCurrentTSO);
102
103           return spark;
104       }
105       if (!emptySparkPoolCap(cap)) {
106           retry = rtsTrue;
107       }
108
109       if (n_capabilities == 1) { return NULL; } // makes no sense...
110
111       debugTrace(DEBUG_sched,
112                  "cap %d: Trying to steal work from other capabilities", 
113                  cap->no);
114
115       /* visit cap.s 0..n-1 in sequence until a theft succeeds. We could
116       start at a random place instead of 0 as well.  */
117       for ( i=0 ; i < n_capabilities ; i++ ) {
118           robbed = &capabilities[i];
119           if (cap == robbed)  // ourselves...
120               continue;
121
122           if (emptySparkPoolCap(robbed)) // nothing to steal here
123               continue;
124
125           spark = tryStealSpark(robbed);
126           if (spark == NULL && !emptySparkPoolCap(robbed)) {
127               // we conflicted with another thread while trying to steal;
128               // try again later.
129               retry = rtsTrue;
130           }
131
132           if (spark != NULL) {
133               cap->sparks_converted++;
134
135               traceEventStealSpark(cap, cap->r.rCurrentTSO, robbed->no);
136               
137               return spark;
138           }
139           // otherwise: no success, try next one
140       }
141   } while (retry);
142
143   debugTrace(DEBUG_sched, "No sparks stolen");
144   return NULL;
145 }
146
147 // Returns True if any spark pool is non-empty at this moment in time
148 // The result is only valid for an instant, of course, so in a sense
149 // is immediately invalid, and should not be relied upon for
150 // correctness.
151 rtsBool
152 anySparks (void)
153 {
154     nat i;
155
156     for (i=0; i < n_capabilities; i++) {
157         if (!emptySparkPoolCap(&capabilities[i])) {
158             return rtsTrue;
159         }
160     }
161     return rtsFalse;
162 }
163 #endif
164
165 /* -----------------------------------------------------------------------------
166  * Manage the returning_tasks lists.
167  *
168  * These functions require cap->lock
169  * -------------------------------------------------------------------------- */
170
171 #if defined(THREADED_RTS)
172 STATIC_INLINE void
173 newReturningTask (Capability *cap, Task *task)
174 {
175     ASSERT_LOCK_HELD(&cap->lock);
176     ASSERT(task->return_link == NULL);
177     if (cap->returning_tasks_hd) {
178         ASSERT(cap->returning_tasks_tl->return_link == NULL);
179         cap->returning_tasks_tl->return_link = task;
180     } else {
181         cap->returning_tasks_hd = task;
182     }
183     cap->returning_tasks_tl = task;
184 }
185
186 STATIC_INLINE Task *
187 popReturningTask (Capability *cap)
188 {
189     ASSERT_LOCK_HELD(&cap->lock);
190     Task *task;
191     task = cap->returning_tasks_hd;
192     ASSERT(task);
193     cap->returning_tasks_hd = task->return_link;
194     if (!cap->returning_tasks_hd) {
195         cap->returning_tasks_tl = NULL;
196     }
197     task->return_link = NULL;
198     return task;
199 }
200 #endif
201
202 /* ----------------------------------------------------------------------------
203  * Initialisation
204  *
205  * The Capability is initially marked not free.
206  * ------------------------------------------------------------------------- */
207
208 static void
209 initCapability( Capability *cap, nat i )
210 {
211     nat g;
212
213     cap->no = i;
214     cap->in_haskell        = rtsFalse;
215
216     cap->run_queue_hd      = END_TSO_QUEUE;
217     cap->run_queue_tl      = END_TSO_QUEUE;
218
219 #if defined(THREADED_RTS)
220     initMutex(&cap->lock);
221     cap->running_task      = NULL; // indicates cap is free
222     cap->spare_workers     = NULL;
223     cap->suspended_ccalling_tasks = NULL;
224     cap->returning_tasks_hd = NULL;
225     cap->returning_tasks_tl = NULL;
226     cap->wakeup_queue_hd    = END_TSO_QUEUE;
227     cap->wakeup_queue_tl    = END_TSO_QUEUE;
228     cap->sparks_created     = 0;
229     cap->sparks_converted   = 0;
230     cap->sparks_pruned      = 0;
231 #endif
232
233     cap->f.stgEagerBlackholeInfo = (W_)&__stg_EAGER_BLACKHOLE_info;
234     cap->f.stgGCEnter1     = (StgFunPtr)__stg_gc_enter_1;
235     cap->f.stgGCFun        = (StgFunPtr)__stg_gc_fun;
236
237     cap->mut_lists  = stgMallocBytes(sizeof(bdescr *) *
238                                      RtsFlags.GcFlags.generations,
239                                      "initCapability");
240     cap->saved_mut_lists = stgMallocBytes(sizeof(bdescr *) *
241                                           RtsFlags.GcFlags.generations,
242                                           "initCapability");
243
244     for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
245         cap->mut_lists[g] = NULL;
246     }
247
248     cap->free_tvar_watch_queues = END_STM_WATCH_QUEUE;
249     cap->free_invariant_check_queues = END_INVARIANT_CHECK_QUEUE;
250     cap->free_trec_chunks = END_STM_CHUNK_LIST;
251     cap->free_trec_headers = NO_TREC;
252     cap->transaction_tokens = 0;
253     cap->context_switch = 0;
254     cap->pinned_object_block = NULL;
255 }
256
257 /* ---------------------------------------------------------------------------
258  * Function:  initCapabilities()
259  *
260  * Purpose:   set up the Capability handling. For the THREADED_RTS build,
261  *            we keep a table of them, the size of which is
262  *            controlled by the user via the RTS flag -N.
263  *
264  * ------------------------------------------------------------------------- */
265 void
266 initCapabilities( void )
267 {
268 #if defined(THREADED_RTS)
269     nat i;
270
271 #ifndef REG_Base
272     // We can't support multiple CPUs if BaseReg is not a register
273     if (RtsFlags.ParFlags.nNodes > 1) {
274         errorBelch("warning: multiple CPUs not supported in this build, reverting to 1");
275         RtsFlags.ParFlags.nNodes = 1;
276     }
277 #endif
278
279     n_capabilities = RtsFlags.ParFlags.nNodes;
280
281     if (n_capabilities == 1) {
282         capabilities = &MainCapability;
283         // THREADED_RTS must work on builds that don't have a mutable
284         // BaseReg (eg. unregisterised), so in this case
285         // capabilities[0] must coincide with &MainCapability.
286     } else {
287         capabilities = stgMallocBytes(n_capabilities * sizeof(Capability),
288                                       "initCapabilities");
289     }
290
291     for (i = 0; i < n_capabilities; i++) {
292         initCapability(&capabilities[i], i);
293     }
294
295     debugTrace(DEBUG_sched, "allocated %d capabilities", n_capabilities);
296
297 #else /* !THREADED_RTS */
298
299     n_capabilities = 1;
300     capabilities = &MainCapability;
301     initCapability(&MainCapability, 0);
302
303 #endif
304
305     // There are no free capabilities to begin with.  We will start
306     // a worker Task to each Capability, which will quickly put the
307     // Capability on the free list when it finds nothing to do.
308     last_free_capability = &capabilities[0];
309 }
310
311 /* ----------------------------------------------------------------------------
312  * setContextSwitches: cause all capabilities to context switch as
313  * soon as possible.
314  * ------------------------------------------------------------------------- */
315
316 void setContextSwitches(void)
317 {
318     nat i;
319     for (i=0; i < n_capabilities; i++) {
320         contextSwitchCapability(&capabilities[i]);
321     }
322 }
323
324 /* ----------------------------------------------------------------------------
325  * Give a Capability to a Task.  The task must currently be sleeping
326  * on its condition variable.
327  *
328  * Requires cap->lock (modifies cap->running_task).
329  *
330  * When migrating a Task, the migrater must take task->lock before
331  * modifying task->cap, to synchronise with the waking up Task.
332  * Additionally, the migrater should own the Capability (when
333  * migrating the run queue), or cap->lock (when migrating
334  * returning_workers).
335  *
336  * ------------------------------------------------------------------------- */
337
338 #if defined(THREADED_RTS)
339 STATIC_INLINE void
340 giveCapabilityToTask (Capability *cap USED_IF_DEBUG, Task *task)
341 {
342     ASSERT_LOCK_HELD(&cap->lock);
343     ASSERT(task->cap == cap);
344     debugTrace(DEBUG_sched, "passing capability %d to %s %p",
345                cap->no, task->tso ? "bound task" : "worker",
346                (void *)task->id);
347     ACQUIRE_LOCK(&task->lock);
348     task->wakeup = rtsTrue;
349     // the wakeup flag is needed because signalCondition() doesn't
350     // flag the condition if the thread is already runniing, but we want
351     // it to be sticky.
352     signalCondition(&task->cond);
353     RELEASE_LOCK(&task->lock);
354 }
355 #endif
356
357 /* ----------------------------------------------------------------------------
358  * Function:  releaseCapability(Capability*)
359  *
360  * Purpose:   Letting go of a capability. Causes a
361  *            'returning worker' thread or a 'waiting worker'
362  *            to wake up, in that order.
363  * ------------------------------------------------------------------------- */
364
365 #if defined(THREADED_RTS)
366 void
367 releaseCapability_ (Capability* cap, 
368                     rtsBool always_wakeup)
369 {
370     Task *task;
371
372     task = cap->running_task;
373
374     ASSERT_PARTIAL_CAPABILITY_INVARIANTS(cap,task);
375
376     cap->running_task = NULL;
377
378     // Check to see whether a worker thread can be given
379     // the go-ahead to return the result of an external call..
380     if (cap->returning_tasks_hd != NULL) {
381         giveCapabilityToTask(cap,cap->returning_tasks_hd);
382         // The Task pops itself from the queue (see waitForReturnCapability())
383         return;
384     }
385
386     if (waiting_for_gc == PENDING_GC_SEQ) {
387       last_free_capability = cap; // needed?
388       debugTrace(DEBUG_sched, "GC pending, set capability %d free", cap->no);
389       return;
390     } 
391
392
393     // If the next thread on the run queue is a bound thread,
394     // give this Capability to the appropriate Task.
395     if (!emptyRunQueue(cap) && cap->run_queue_hd->bound) {
396         // Make sure we're not about to try to wake ourselves up
397         // ASSERT(task != cap->run_queue_hd->bound);
398         // assertion is false: in schedule() we force a yield after
399         // ThreadBlocked, but the thread may be back on the run queue
400         // by now.
401         task = cap->run_queue_hd->bound;
402         giveCapabilityToTask(cap,task);
403         return;
404     }
405
406     if (!cap->spare_workers) {
407         // Create a worker thread if we don't have one.  If the system
408         // is interrupted, we only create a worker task if there
409         // are threads that need to be completed.  If the system is
410         // shutting down, we never create a new worker.
411         if (sched_state < SCHED_SHUTTING_DOWN || !emptyRunQueue(cap)) {
412             debugTrace(DEBUG_sched,
413                        "starting new worker on capability %d", cap->no);
414             startWorkerTask(cap, workerStart);
415             return;
416         }
417     }
418
419     // If we have an unbound thread on the run queue, or if there's
420     // anything else to do, give the Capability to a worker thread.
421     if (always_wakeup || 
422         !emptyRunQueue(cap) || !emptyWakeupQueue(cap) ||
423         !emptySparkPoolCap(cap) || globalWorkToDo()) {
424         if (cap->spare_workers) {
425             giveCapabilityToTask(cap,cap->spare_workers);
426             // The worker Task pops itself from the queue;
427             return;
428         }
429     }
430
431     last_free_capability = cap;
432     debugTrace(DEBUG_sched, "freeing capability %d", cap->no);
433 }
434
435 void
436 releaseCapability (Capability* cap USED_IF_THREADS)
437 {
438     ACQUIRE_LOCK(&cap->lock);
439     releaseCapability_(cap, rtsFalse);
440     RELEASE_LOCK(&cap->lock);
441 }
442
443 void
444 releaseAndWakeupCapability (Capability* cap USED_IF_THREADS)
445 {
446     ACQUIRE_LOCK(&cap->lock);
447     releaseCapability_(cap, rtsTrue);
448     RELEASE_LOCK(&cap->lock);
449 }
450
451 static void
452 releaseCapabilityAndQueueWorker (Capability* cap USED_IF_THREADS)
453 {
454     Task *task;
455
456     ACQUIRE_LOCK(&cap->lock);
457
458     task = cap->running_task;
459
460     // If the current task is a worker, save it on the spare_workers
461     // list of this Capability.  A worker can mark itself as stopped,
462     // in which case it is not replaced on the spare_worker queue.
463     // This happens when the system is shutting down (see
464     // Schedule.c:workerStart()).
465     // Also, be careful to check that this task hasn't just exited
466     // Haskell to do a foreign call (task->suspended_tso).
467     if (!isBoundTask(task) && !task->stopped && !task->suspended_tso) {
468         task->next = cap->spare_workers;
469         cap->spare_workers = task;
470     }
471     // Bound tasks just float around attached to their TSOs.
472
473     releaseCapability_(cap,rtsFalse);
474
475     RELEASE_LOCK(&cap->lock);
476 }
477 #endif
478
479 /* ----------------------------------------------------------------------------
480  * waitForReturnCapability( Task *task )
481  *
482  * Purpose:  when an OS thread returns from an external call,
483  * it calls waitForReturnCapability() (via Schedule.resumeThread())
484  * to wait for permission to enter the RTS & communicate the
485  * result of the external call back to the Haskell thread that
486  * made it.
487  *
488  * ------------------------------------------------------------------------- */
489 void
490 waitForReturnCapability (Capability **pCap, Task *task)
491 {
492 #if !defined(THREADED_RTS)
493
494     MainCapability.running_task = task;
495     task->cap = &MainCapability;
496     *pCap = &MainCapability;
497
498 #else
499     Capability *cap = *pCap;
500
501     if (cap == NULL) {
502         // Try last_free_capability first
503         cap = last_free_capability;
504         if (cap->running_task) {
505             nat i;
506             // otherwise, search for a free capability
507             cap = NULL;
508             for (i = 0; i < n_capabilities; i++) {
509                 if (!capabilities[i].running_task) {
510                     cap = &capabilities[i];
511                     break;
512                 }
513             }
514             if (cap == NULL) {
515                 // Can't find a free one, use last_free_capability.
516                 cap = last_free_capability;
517             }
518         }
519
520         // record the Capability as the one this Task is now assocated with.
521         task->cap = cap;
522
523     } else {
524         ASSERT(task->cap == cap);
525     }
526
527     ACQUIRE_LOCK(&cap->lock);
528
529     debugTrace(DEBUG_sched, "returning; I want capability %d", cap->no);
530
531     if (!cap->running_task) {
532         // It's free; just grab it
533         cap->running_task = task;
534         RELEASE_LOCK(&cap->lock);
535     } else {
536         newReturningTask(cap,task);
537         RELEASE_LOCK(&cap->lock);
538
539         for (;;) {
540             ACQUIRE_LOCK(&task->lock);
541             // task->lock held, cap->lock not held
542             if (!task->wakeup) waitCondition(&task->cond, &task->lock);
543             cap = task->cap;
544             task->wakeup = rtsFalse;
545             RELEASE_LOCK(&task->lock);
546
547             // now check whether we should wake up...
548             ACQUIRE_LOCK(&cap->lock);
549             if (cap->running_task == NULL) {
550                 if (cap->returning_tasks_hd != task) {
551                     giveCapabilityToTask(cap,cap->returning_tasks_hd);
552                     RELEASE_LOCK(&cap->lock);
553                     continue;
554                 }
555                 cap->running_task = task;
556                 popReturningTask(cap);
557                 RELEASE_LOCK(&cap->lock);
558                 break;
559             }
560             RELEASE_LOCK(&cap->lock);
561         }
562
563     }
564
565     ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
566
567     debugTrace(DEBUG_sched, "resuming capability %d", cap->no);
568
569     *pCap = cap;
570 #endif
571 }
572
573 #if defined(THREADED_RTS)
574 /* ----------------------------------------------------------------------------
575  * yieldCapability
576  * ------------------------------------------------------------------------- */
577
578 void
579 yieldCapability (Capability** pCap, Task *task)
580 {
581     Capability *cap = *pCap;
582
583     if (waiting_for_gc == PENDING_GC_PAR) {
584         traceEventGcStart(cap);
585         gcWorkerThread(cap);
586         traceEventGcEnd(cap);
587         return;
588     }
589
590         debugTrace(DEBUG_sched, "giving up capability %d", cap->no);
591
592         // We must now release the capability and wait to be woken up
593         // again.
594         task->wakeup = rtsFalse;
595         releaseCapabilityAndQueueWorker(cap);
596
597         for (;;) {
598             ACQUIRE_LOCK(&task->lock);
599             // task->lock held, cap->lock not held
600             if (!task->wakeup) waitCondition(&task->cond, &task->lock);
601             cap = task->cap;
602             task->wakeup = rtsFalse;
603             RELEASE_LOCK(&task->lock);
604
605             debugTrace(DEBUG_sched, "woken up on capability %d", cap->no);
606
607             ACQUIRE_LOCK(&cap->lock);
608             if (cap->running_task != NULL) {
609                 debugTrace(DEBUG_sched, 
610                            "capability %d is owned by another task", cap->no);
611                 RELEASE_LOCK(&cap->lock);
612                 continue;
613             }
614
615             if (task->tso == NULL) {
616                 ASSERT(cap->spare_workers != NULL);
617                 // if we're not at the front of the queue, release it
618                 // again.  This is unlikely to happen.
619                 if (cap->spare_workers != task) {
620                     giveCapabilityToTask(cap,cap->spare_workers);
621                     RELEASE_LOCK(&cap->lock);
622                     continue;
623                 }
624                 cap->spare_workers = task->next;
625                 task->next = NULL;
626             }
627             cap->running_task = task;
628             RELEASE_LOCK(&cap->lock);
629             break;
630         }
631
632         debugTrace(DEBUG_sched, "resuming capability %d", cap->no);
633         ASSERT(cap->running_task == task);
634
635     *pCap = cap;
636
637     ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
638
639     return;
640 }
641
642 /* ----------------------------------------------------------------------------
643  * Wake up a thread on a Capability.
644  *
645  * This is used when the current Task is running on a Capability and
646  * wishes to wake up a thread on a different Capability.
647  * ------------------------------------------------------------------------- */
648
649 void
650 wakeupThreadOnCapability (Capability *my_cap, 
651                           Capability *other_cap, 
652                           StgTSO *tso)
653 {
654     ACQUIRE_LOCK(&other_cap->lock);
655
656     // ASSUMES: cap->lock is held (asserted in wakeupThreadOnCapability)
657     if (tso->bound) {
658         ASSERT(tso->bound->cap == tso->cap);
659         tso->bound->cap = other_cap;
660     }
661     tso->cap = other_cap;
662
663     ASSERT(tso->bound ? tso->bound->cap == other_cap : 1);
664
665     if (other_cap->running_task == NULL) {
666         // nobody is running this Capability, we can add our thread
667         // directly onto the run queue and start up a Task to run it.
668
669         other_cap->running_task = myTask(); 
670             // precond for releaseCapability_() and appendToRunQueue()
671
672         appendToRunQueue(other_cap,tso);
673
674         releaseCapability_(other_cap,rtsFalse);
675     } else {
676         appendToWakeupQueue(my_cap,other_cap,tso);
677         other_cap->context_switch = 1;
678         // someone is running on this Capability, so it cannot be
679         // freed without first checking the wakeup queue (see
680         // releaseCapability_).
681     }
682
683     RELEASE_LOCK(&other_cap->lock);
684 }
685
686 /* ----------------------------------------------------------------------------
687  * prodCapability
688  *
689  * If a Capability is currently idle, wake up a Task on it.  Used to 
690  * get every Capability into the GC.
691  * ------------------------------------------------------------------------- */
692
693 void
694 prodCapability (Capability *cap, Task *task)
695 {
696     ACQUIRE_LOCK(&cap->lock);
697     if (!cap->running_task) {
698         cap->running_task = task;
699         releaseCapability_(cap,rtsTrue);
700     }
701     RELEASE_LOCK(&cap->lock);
702 }
703
704 /* ----------------------------------------------------------------------------
705  * shutdownCapability
706  *
707  * At shutdown time, we want to let everything exit as cleanly as
708  * possible.  For each capability, we let its run queue drain, and
709  * allow the workers to stop.
710  *
711  * This function should be called when interrupted and
712  * shutting_down_scheduler = rtsTrue, thus any worker that wakes up
713  * will exit the scheduler and call taskStop(), and any bound thread
714  * that wakes up will return to its caller.  Runnable threads are
715  * killed.
716  *
717  * ------------------------------------------------------------------------- */
718
719 void
720 shutdownCapability (Capability *cap, Task *task, rtsBool safe)
721 {
722     nat i;
723
724     task->cap = cap;
725
726     // Loop indefinitely until all the workers have exited and there
727     // are no Haskell threads left.  We used to bail out after 50
728     // iterations of this loop, but that occasionally left a worker
729     // running which caused problems later (the closeMutex() below
730     // isn't safe, for one thing).
731
732     for (i = 0; /* i < 50 */; i++) {
733         ASSERT(sched_state == SCHED_SHUTTING_DOWN);
734
735         debugTrace(DEBUG_sched, 
736                    "shutting down capability %d, attempt %d", cap->no, i);
737         ACQUIRE_LOCK(&cap->lock);
738         if (cap->running_task) {
739             RELEASE_LOCK(&cap->lock);
740             debugTrace(DEBUG_sched, "not owner, yielding");
741             yieldThread();
742             continue;
743         }
744         cap->running_task = task;
745
746         if (cap->spare_workers) {
747             // Look for workers that have died without removing
748             // themselves from the list; this could happen if the OS
749             // summarily killed the thread, for example.  This
750             // actually happens on Windows when the system is
751             // terminating the program, and the RTS is running in a
752             // DLL.
753             Task *t, *prev;
754             prev = NULL;
755             for (t = cap->spare_workers; t != NULL; t = t->next) {
756                 if (!osThreadIsAlive(t->id)) {
757                     debugTrace(DEBUG_sched, 
758                                "worker thread %p has died unexpectedly", (void *)t->id);
759                         if (!prev) {
760                             cap->spare_workers = t->next;
761                         } else {
762                             prev->next = t->next;
763                         }
764                         prev = t;
765                 }
766             }
767         }
768
769         if (!emptyRunQueue(cap) || cap->spare_workers) {
770             debugTrace(DEBUG_sched, 
771                        "runnable threads or workers still alive, yielding");
772             releaseCapability_(cap,rtsFalse); // this will wake up a worker
773             RELEASE_LOCK(&cap->lock);
774             yieldThread();
775             continue;
776         }
777
778         // If "safe", then busy-wait for any threads currently doing
779         // foreign calls.  If we're about to unload this DLL, for
780         // example, we need to be sure that there are no OS threads
781         // that will try to return to code that has been unloaded.
782         // We can be a bit more relaxed when this is a standalone
783         // program that is about to terminate, and let safe=false.
784         if (cap->suspended_ccalling_tasks && safe) {
785             debugTrace(DEBUG_sched, 
786                        "thread(s) are involved in foreign calls, yielding");
787             cap->running_task = NULL;
788             RELEASE_LOCK(&cap->lock);
789             // The IO manager thread might have been slow to start up,
790             // so the first attempt to kill it might not have
791             // succeeded.  Just in case, try again - the kill message
792             // will only be sent once.
793             //
794             // To reproduce this deadlock: run ffi002(threaded1)
795             // repeatedly on a loaded machine.
796             ioManagerDie();
797             yieldThread();
798             continue;
799         }
800
801         traceEventShutdown(cap);
802         RELEASE_LOCK(&cap->lock);
803         break;
804     }
805     // we now have the Capability, its run queue and spare workers
806     // list are both empty.
807
808     // ToDo: we can't drop this mutex, because there might still be
809     // threads performing foreign calls that will eventually try to 
810     // return via resumeThread() and attempt to grab cap->lock.
811     // closeMutex(&cap->lock);
812 }
813
814 /* ----------------------------------------------------------------------------
815  * tryGrabCapability
816  *
817  * Attempt to gain control of a Capability if it is free.
818  *
819  * ------------------------------------------------------------------------- */
820
821 rtsBool
822 tryGrabCapability (Capability *cap, Task *task)
823 {
824     if (cap->running_task != NULL) return rtsFalse;
825     ACQUIRE_LOCK(&cap->lock);
826     if (cap->running_task != NULL) {
827         RELEASE_LOCK(&cap->lock);
828         return rtsFalse;
829     }
830     task->cap = cap;
831     cap->running_task = task;
832     RELEASE_LOCK(&cap->lock);
833     return rtsTrue;
834 }
835
836
837 #endif /* THREADED_RTS */
838
839 static void
840 freeCapability (Capability *cap)
841 {
842     stgFree(cap->mut_lists);
843     stgFree(cap->saved_mut_lists);
844 #if defined(THREADED_RTS)
845     freeSparkPool(cap->sparks);
846 #endif
847 }
848
849 void
850 freeCapabilities (void)
851 {
852 #if defined(THREADED_RTS)
853     nat i;
854     for (i=0; i < n_capabilities; i++) {
855         freeCapability(&capabilities[i]);
856     }
857 #else
858     freeCapability(&MainCapability);
859 #endif
860 }
861
862 /* ---------------------------------------------------------------------------
863    Mark everything directly reachable from the Capabilities.  When
864    using multiple GC threads, each GC thread marks all Capabilities
865    for which (c `mod` n == 0), for Capability c and thread n.
866    ------------------------------------------------------------------------ */
867
868 void
869 markSomeCapabilities (evac_fn evac, void *user, nat i0, nat delta, 
870                       rtsBool prune_sparks USED_IF_THREADS)
871 {
872     nat i;
873     Capability *cap;
874     Task *task;
875
876     // Each GC thread is responsible for following roots from the
877     // Capability of the same number.  There will usually be the same
878     // or fewer Capabilities as GC threads, but just in case there
879     // are more, we mark every Capability whose number is the GC
880     // thread's index plus a multiple of the number of GC threads.
881     for (i = i0; i < n_capabilities; i += delta) {
882         cap = &capabilities[i];
883         evac(user, (StgClosure **)(void *)&cap->run_queue_hd);
884         evac(user, (StgClosure **)(void *)&cap->run_queue_tl);
885 #if defined(THREADED_RTS)
886         evac(user, (StgClosure **)(void *)&cap->wakeup_queue_hd);
887         evac(user, (StgClosure **)(void *)&cap->wakeup_queue_tl);
888 #endif
889         for (task = cap->suspended_ccalling_tasks; task != NULL; 
890              task=task->next) {
891             evac(user, (StgClosure **)(void *)&task->suspended_tso);
892         }
893
894 #if defined(THREADED_RTS)
895         if (prune_sparks) {
896             pruneSparkQueue (evac, user, cap);
897         } else {
898             traverseSparkQueue (evac, user, cap);
899         }
900 #endif
901     }
902
903 #if !defined(THREADED_RTS)
904     evac(user, (StgClosure **)(void *)&blocked_queue_hd);
905     evac(user, (StgClosure **)(void *)&blocked_queue_tl);
906     evac(user, (StgClosure **)(void *)&sleeping_queue);
907 #endif 
908 }
909
910 void
911 markCapabilities (evac_fn evac, void *user)
912 {
913     markSomeCapabilities(evac, user, 0, 1, rtsFalse);
914 }