Expose all EventLog events as DTrace probes
[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         task = cap->run_queue_hd->bound;
399         giveCapabilityToTask(cap,task);
400         return;
401     }
402
403     if (!cap->spare_workers) {
404         // Create a worker thread if we don't have one.  If the system
405         // is interrupted, we only create a worker task if there
406         // are threads that need to be completed.  If the system is
407         // shutting down, we never create a new worker.
408         if (sched_state < SCHED_SHUTTING_DOWN || !emptyRunQueue(cap)) {
409             debugTrace(DEBUG_sched,
410                        "starting new worker on capability %d", cap->no);
411             startWorkerTask(cap, workerStart);
412             return;
413         }
414     }
415
416     // If we have an unbound thread on the run queue, or if there's
417     // anything else to do, give the Capability to a worker thread.
418     if (always_wakeup || 
419         !emptyRunQueue(cap) || !emptyWakeupQueue(cap) ||
420         !emptySparkPoolCap(cap) || globalWorkToDo()) {
421         if (cap->spare_workers) {
422             giveCapabilityToTask(cap,cap->spare_workers);
423             // The worker Task pops itself from the queue;
424             return;
425         }
426     }
427
428     last_free_capability = cap;
429     debugTrace(DEBUG_sched, "freeing capability %d", cap->no);
430 }
431
432 void
433 releaseCapability (Capability* cap USED_IF_THREADS)
434 {
435     ACQUIRE_LOCK(&cap->lock);
436     releaseCapability_(cap, rtsFalse);
437     RELEASE_LOCK(&cap->lock);
438 }
439
440 void
441 releaseAndWakeupCapability (Capability* cap USED_IF_THREADS)
442 {
443     ACQUIRE_LOCK(&cap->lock);
444     releaseCapability_(cap, rtsTrue);
445     RELEASE_LOCK(&cap->lock);
446 }
447
448 static void
449 releaseCapabilityAndQueueWorker (Capability* cap USED_IF_THREADS)
450 {
451     Task *task;
452
453     ACQUIRE_LOCK(&cap->lock);
454
455     task = cap->running_task;
456
457     // If the current task is a worker, save it on the spare_workers
458     // list of this Capability.  A worker can mark itself as stopped,
459     // in which case it is not replaced on the spare_worker queue.
460     // This happens when the system is shutting down (see
461     // Schedule.c:workerStart()).
462     // Also, be careful to check that this task hasn't just exited
463     // Haskell to do a foreign call (task->suspended_tso).
464     if (!isBoundTask(task) && !task->stopped && !task->suspended_tso) {
465         task->next = cap->spare_workers;
466         cap->spare_workers = task;
467     }
468     // Bound tasks just float around attached to their TSOs.
469
470     releaseCapability_(cap,rtsFalse);
471
472     RELEASE_LOCK(&cap->lock);
473 }
474 #endif
475
476 /* ----------------------------------------------------------------------------
477  * waitForReturnCapability( Task *task )
478  *
479  * Purpose:  when an OS thread returns from an external call,
480  * it calls waitForReturnCapability() (via Schedule.resumeThread())
481  * to wait for permission to enter the RTS & communicate the
482  * result of the external call back to the Haskell thread that
483  * made it.
484  *
485  * ------------------------------------------------------------------------- */
486 void
487 waitForReturnCapability (Capability **pCap, Task *task)
488 {
489 #if !defined(THREADED_RTS)
490
491     MainCapability.running_task = task;
492     task->cap = &MainCapability;
493     *pCap = &MainCapability;
494
495 #else
496     Capability *cap = *pCap;
497
498     if (cap == NULL) {
499         // Try last_free_capability first
500         cap = last_free_capability;
501         if (cap->running_task) {
502             nat i;
503             // otherwise, search for a free capability
504             cap = NULL;
505             for (i = 0; i < n_capabilities; i++) {
506                 if (!capabilities[i].running_task) {
507                     cap = &capabilities[i];
508                     break;
509                 }
510             }
511             if (cap == NULL) {
512                 // Can't find a free one, use last_free_capability.
513                 cap = last_free_capability;
514             }
515         }
516
517         // record the Capability as the one this Task is now assocated with.
518         task->cap = cap;
519
520     } else {
521         ASSERT(task->cap == cap);
522     }
523
524     ACQUIRE_LOCK(&cap->lock);
525
526     debugTrace(DEBUG_sched, "returning; I want capability %d", cap->no);
527
528     if (!cap->running_task) {
529         // It's free; just grab it
530         cap->running_task = task;
531         RELEASE_LOCK(&cap->lock);
532     } else {
533         newReturningTask(cap,task);
534         RELEASE_LOCK(&cap->lock);
535
536         for (;;) {
537             ACQUIRE_LOCK(&task->lock);
538             // task->lock held, cap->lock not held
539             if (!task->wakeup) waitCondition(&task->cond, &task->lock);
540             cap = task->cap;
541             task->wakeup = rtsFalse;
542             RELEASE_LOCK(&task->lock);
543
544             // now check whether we should wake up...
545             ACQUIRE_LOCK(&cap->lock);
546             if (cap->running_task == NULL) {
547                 if (cap->returning_tasks_hd != task) {
548                     giveCapabilityToTask(cap,cap->returning_tasks_hd);
549                     RELEASE_LOCK(&cap->lock);
550                     continue;
551                 }
552                 cap->running_task = task;
553                 popReturningTask(cap);
554                 RELEASE_LOCK(&cap->lock);
555                 break;
556             }
557             RELEASE_LOCK(&cap->lock);
558         }
559
560     }
561
562     ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
563
564     debugTrace(DEBUG_sched, "resuming capability %d", cap->no);
565
566     *pCap = cap;
567 #endif
568 }
569
570 #if defined(THREADED_RTS)
571 /* ----------------------------------------------------------------------------
572  * yieldCapability
573  * ------------------------------------------------------------------------- */
574
575 void
576 yieldCapability (Capability** pCap, Task *task)
577 {
578     Capability *cap = *pCap;
579
580     if (waiting_for_gc == PENDING_GC_PAR) {
581         traceEventGcStart(cap);
582         gcWorkerThread(cap);
583         traceEventGcEnd(cap);
584         return;
585     }
586
587         debugTrace(DEBUG_sched, "giving up capability %d", cap->no);
588
589         // We must now release the capability and wait to be woken up
590         // again.
591         task->wakeup = rtsFalse;
592         releaseCapabilityAndQueueWorker(cap);
593
594         for (;;) {
595             ACQUIRE_LOCK(&task->lock);
596             // task->lock held, cap->lock not held
597             if (!task->wakeup) waitCondition(&task->cond, &task->lock);
598             cap = task->cap;
599             task->wakeup = rtsFalse;
600             RELEASE_LOCK(&task->lock);
601
602             debugTrace(DEBUG_sched, "woken up on capability %d", cap->no);
603
604             ACQUIRE_LOCK(&cap->lock);
605             if (cap->running_task != NULL) {
606                 debugTrace(DEBUG_sched, 
607                            "capability %d is owned by another task", cap->no);
608                 RELEASE_LOCK(&cap->lock);
609                 continue;
610             }
611
612             if (task->tso == NULL) {
613                 ASSERT(cap->spare_workers != NULL);
614                 // if we're not at the front of the queue, release it
615                 // again.  This is unlikely to happen.
616                 if (cap->spare_workers != task) {
617                     giveCapabilityToTask(cap,cap->spare_workers);
618                     RELEASE_LOCK(&cap->lock);
619                     continue;
620                 }
621                 cap->spare_workers = task->next;
622                 task->next = NULL;
623             }
624             cap->running_task = task;
625             RELEASE_LOCK(&cap->lock);
626             break;
627         }
628
629         debugTrace(DEBUG_sched, "resuming capability %d", cap->no);
630         ASSERT(cap->running_task == task);
631
632     *pCap = cap;
633
634     ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
635
636     return;
637 }
638
639 /* ----------------------------------------------------------------------------
640  * Wake up a thread on a Capability.
641  *
642  * This is used when the current Task is running on a Capability and
643  * wishes to wake up a thread on a different Capability.
644  * ------------------------------------------------------------------------- */
645
646 void
647 wakeupThreadOnCapability (Capability *my_cap, 
648                           Capability *other_cap, 
649                           StgTSO *tso)
650 {
651     ACQUIRE_LOCK(&other_cap->lock);
652
653     // ASSUMES: cap->lock is held (asserted in wakeupThreadOnCapability)
654     if (tso->bound) {
655         ASSERT(tso->bound->cap == tso->cap);
656         tso->bound->cap = other_cap;
657     }
658     tso->cap = other_cap;
659
660     ASSERT(tso->bound ? tso->bound->cap == other_cap : 1);
661
662     if (other_cap->running_task == NULL) {
663         // nobody is running this Capability, we can add our thread
664         // directly onto the run queue and start up a Task to run it.
665
666         other_cap->running_task = myTask(); 
667             // precond for releaseCapability_() and appendToRunQueue()
668
669         appendToRunQueue(other_cap,tso);
670
671         releaseCapability_(other_cap,rtsFalse);
672     } else {
673         appendToWakeupQueue(my_cap,other_cap,tso);
674         other_cap->context_switch = 1;
675         // someone is running on this Capability, so it cannot be
676         // freed without first checking the wakeup queue (see
677         // releaseCapability_).
678     }
679
680     RELEASE_LOCK(&other_cap->lock);
681 }
682
683 /* ----------------------------------------------------------------------------
684  * prodCapability
685  *
686  * If a Capability is currently idle, wake up a Task on it.  Used to 
687  * get every Capability into the GC.
688  * ------------------------------------------------------------------------- */
689
690 void
691 prodCapability (Capability *cap, Task *task)
692 {
693     ACQUIRE_LOCK(&cap->lock);
694     if (!cap->running_task) {
695         cap->running_task = task;
696         releaseCapability_(cap,rtsTrue);
697     }
698     RELEASE_LOCK(&cap->lock);
699 }
700
701 /* ----------------------------------------------------------------------------
702  * shutdownCapability
703  *
704  * At shutdown time, we want to let everything exit as cleanly as
705  * possible.  For each capability, we let its run queue drain, and
706  * allow the workers to stop.
707  *
708  * This function should be called when interrupted and
709  * shutting_down_scheduler = rtsTrue, thus any worker that wakes up
710  * will exit the scheduler and call taskStop(), and any bound thread
711  * that wakes up will return to its caller.  Runnable threads are
712  * killed.
713  *
714  * ------------------------------------------------------------------------- */
715
716 void
717 shutdownCapability (Capability *cap, Task *task, rtsBool safe)
718 {
719     nat i;
720
721     task->cap = cap;
722
723     // Loop indefinitely until all the workers have exited and there
724     // are no Haskell threads left.  We used to bail out after 50
725     // iterations of this loop, but that occasionally left a worker
726     // running which caused problems later (the closeMutex() below
727     // isn't safe, for one thing).
728
729     for (i = 0; /* i < 50 */; i++) {
730         ASSERT(sched_state == SCHED_SHUTTING_DOWN);
731
732         debugTrace(DEBUG_sched, 
733                    "shutting down capability %d, attempt %d", cap->no, i);
734         ACQUIRE_LOCK(&cap->lock);
735         if (cap->running_task) {
736             RELEASE_LOCK(&cap->lock);
737             debugTrace(DEBUG_sched, "not owner, yielding");
738             yieldThread();
739             continue;
740         }
741         cap->running_task = task;
742
743         if (cap->spare_workers) {
744             // Look for workers that have died without removing
745             // themselves from the list; this could happen if the OS
746             // summarily killed the thread, for example.  This
747             // actually happens on Windows when the system is
748             // terminating the program, and the RTS is running in a
749             // DLL.
750             Task *t, *prev;
751             prev = NULL;
752             for (t = cap->spare_workers; t != NULL; t = t->next) {
753                 if (!osThreadIsAlive(t->id)) {
754                     debugTrace(DEBUG_sched, 
755                                "worker thread %p has died unexpectedly", (void *)t->id);
756                         if (!prev) {
757                             cap->spare_workers = t->next;
758                         } else {
759                             prev->next = t->next;
760                         }
761                         prev = t;
762                 }
763             }
764         }
765
766         if (!emptyRunQueue(cap) || cap->spare_workers) {
767             debugTrace(DEBUG_sched, 
768                        "runnable threads or workers still alive, yielding");
769             releaseCapability_(cap,rtsFalse); // this will wake up a worker
770             RELEASE_LOCK(&cap->lock);
771             yieldThread();
772             continue;
773         }
774
775         // If "safe", then busy-wait for any threads currently doing
776         // foreign calls.  If we're about to unload this DLL, for
777         // example, we need to be sure that there are no OS threads
778         // that will try to return to code that has been unloaded.
779         // We can be a bit more relaxed when this is a standalone
780         // program that is about to terminate, and let safe=false.
781         if (cap->suspended_ccalling_tasks && safe) {
782             debugTrace(DEBUG_sched, 
783                        "thread(s) are involved in foreign calls, yielding");
784             cap->running_task = NULL;
785             RELEASE_LOCK(&cap->lock);
786             yieldThread();
787             continue;
788         }
789             
790         traceEventShutdown(cap);
791         RELEASE_LOCK(&cap->lock);
792         break;
793     }
794     // we now have the Capability, its run queue and spare workers
795     // list are both empty.
796
797     // ToDo: we can't drop this mutex, because there might still be
798     // threads performing foreign calls that will eventually try to 
799     // return via resumeThread() and attempt to grab cap->lock.
800     // closeMutex(&cap->lock);
801 }
802
803 /* ----------------------------------------------------------------------------
804  * tryGrabCapability
805  *
806  * Attempt to gain control of a Capability if it is free.
807  *
808  * ------------------------------------------------------------------------- */
809
810 rtsBool
811 tryGrabCapability (Capability *cap, Task *task)
812 {
813     if (cap->running_task != NULL) return rtsFalse;
814     ACQUIRE_LOCK(&cap->lock);
815     if (cap->running_task != NULL) {
816         RELEASE_LOCK(&cap->lock);
817         return rtsFalse;
818     }
819     task->cap = cap;
820     cap->running_task = task;
821     RELEASE_LOCK(&cap->lock);
822     return rtsTrue;
823 }
824
825
826 #endif /* THREADED_RTS */
827
828 static void
829 freeCapability (Capability *cap)
830 {
831     stgFree(cap->mut_lists);
832     stgFree(cap->saved_mut_lists);
833 #if defined(THREADED_RTS)
834     freeSparkPool(cap->sparks);
835 #endif
836 }
837
838 void
839 freeCapabilities (void)
840 {
841 #if defined(THREADED_RTS)
842     nat i;
843     for (i=0; i < n_capabilities; i++) {
844         freeCapability(&capabilities[i]);
845     }
846 #else
847     freeCapability(&MainCapability);
848 #endif
849 }
850
851 /* ---------------------------------------------------------------------------
852    Mark everything directly reachable from the Capabilities.  When
853    using multiple GC threads, each GC thread marks all Capabilities
854    for which (c `mod` n == 0), for Capability c and thread n.
855    ------------------------------------------------------------------------ */
856
857 void
858 markSomeCapabilities (evac_fn evac, void *user, nat i0, nat delta, 
859                       rtsBool prune_sparks USED_IF_THREADS)
860 {
861     nat i;
862     Capability *cap;
863     Task *task;
864
865     // Each GC thread is responsible for following roots from the
866     // Capability of the same number.  There will usually be the same
867     // or fewer Capabilities as GC threads, but just in case there
868     // are more, we mark every Capability whose number is the GC
869     // thread's index plus a multiple of the number of GC threads.
870     for (i = i0; i < n_capabilities; i += delta) {
871         cap = &capabilities[i];
872         evac(user, (StgClosure **)(void *)&cap->run_queue_hd);
873         evac(user, (StgClosure **)(void *)&cap->run_queue_tl);
874 #if defined(THREADED_RTS)
875         evac(user, (StgClosure **)(void *)&cap->wakeup_queue_hd);
876         evac(user, (StgClosure **)(void *)&cap->wakeup_queue_tl);
877 #endif
878         for (task = cap->suspended_ccalling_tasks; task != NULL; 
879              task=task->next) {
880             evac(user, (StgClosure **)(void *)&task->suspended_tso);
881         }
882
883 #if defined(THREADED_RTS)
884         if (prune_sparks) {
885             pruneSparkQueue (evac, user, cap);
886         } else {
887             traverseSparkQueue (evac, user, cap);
888         }
889 #endif
890     }
891
892 #if !defined(THREADED_RTS)
893     evac(user, (StgClosure **)(void *)&blocked_queue_hd);
894     evac(user, (StgClosure **)(void *)&blocked_queue_tl);
895     evac(user, (StgClosure **)(void *)&sleeping_queue);
896 #endif 
897 }
898
899 void
900 markCapabilities (evac_fn evac, void *user)
901 {
902     markSomeCapabilities(evac, user, 0, 1, rtsFalse);
903 }