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