Unify event logging and debug tracing.
[ghc.git] / rts / Schedule.c
1 /* ---------------------------------------------------------------------------
2 *
3 * (c) The GHC Team, 1998-2006
4 *
5 * The scheduler and thread-related functionality
6 *
7 * --------------------------------------------------------------------------*/
8
9 #include "PosixSource.h"
10 #define KEEP_LOCKCLOSURE
11 #include "Rts.h"
12
13 #include "sm/Storage.h"
14 #include "RtsUtils.h"
15 #include "StgRun.h"
16 #include "Schedule.h"
17 #include "Interpreter.h"
18 #include "Printer.h"
19 #include "RtsSignals.h"
20 #include "Sanity.h"
21 #include "Stats.h"
22 #include "STM.h"
23 #include "Prelude.h"
24 #include "ThreadLabels.h"
25 #include "Updates.h"
26 #include "Proftimer.h"
27 #include "ProfHeap.h"
28 #include "Weak.h"
29 #include "sm/GC.h" // waitForGcThreads, releaseGCThreads, N
30 #include "Sparks.h"
31 #include "Capability.h"
32 #include "Task.h"
33 #include "AwaitEvent.h"
34 #if defined(mingw32_HOST_OS)
35 #include "win32/IOManager.h"
36 #endif
37 #include "Trace.h"
38 #include "RaiseAsync.h"
39 #include "Threads.h"
40 #include "Timer.h"
41 #include "ThreadPaused.h"
42
43 #ifdef HAVE_SYS_TYPES_H
44 #include <sys/types.h>
45 #endif
46 #ifdef HAVE_UNISTD_H
47 #include <unistd.h>
48 #endif
49
50 #include <string.h>
51 #include <stdlib.h>
52 #include <stdarg.h>
53
54 #ifdef HAVE_ERRNO_H
55 #include <errno.h>
56 #endif
57
58 /* -----------------------------------------------------------------------------
59 * Global variables
60 * -------------------------------------------------------------------------- */
61
62 #if !defined(THREADED_RTS)
63 // Blocked/sleeping thrads
64 StgTSO *blocked_queue_hd = NULL;
65 StgTSO *blocked_queue_tl = NULL;
66 StgTSO *sleeping_queue = NULL; // perhaps replace with a hash table?
67 #endif
68
69 /* Threads blocked on blackholes.
70 * LOCK: sched_mutex+capability, or all capabilities
71 */
72 StgTSO *blackhole_queue = NULL;
73
74 /* The blackhole_queue should be checked for threads to wake up. See
75 * Schedule.h for more thorough comment.
76 * LOCK: none (doesn't matter if we miss an update)
77 */
78 rtsBool blackholes_need_checking = rtsFalse;
79
80 /* Set to true when the latest garbage collection failed to reclaim
81 * enough space, and the runtime should proceed to shut itself down in
82 * an orderly fashion (emitting profiling info etc.)
83 */
84 rtsBool heap_overflow = rtsFalse;
85
86 /* flag that tracks whether we have done any execution in this time slice.
87 * LOCK: currently none, perhaps we should lock (but needs to be
88 * updated in the fast path of the scheduler).
89 *
90 * NB. must be StgWord, we do xchg() on it.
91 */
92 volatile StgWord recent_activity = ACTIVITY_YES;
93
94 /* if this flag is set as well, give up execution
95 * LOCK: none (changes monotonically)
96 */
97 volatile StgWord sched_state = SCHED_RUNNING;
98
99 /* This is used in `TSO.h' and gcc 2.96 insists that this variable actually
100 * exists - earlier gccs apparently didn't.
101 * -= chak
102 */
103 StgTSO dummy_tso;
104
105 /*
106 * Set to TRUE when entering a shutdown state (via shutdownHaskellAndExit()) --
107 * in an MT setting, needed to signal that a worker thread shouldn't hang around
108 * in the scheduler when it is out of work.
109 */
110 rtsBool shutting_down_scheduler = rtsFalse;
111
112 /*
113 * This mutex protects most of the global scheduler data in
114 * the THREADED_RTS runtime.
115 */
116 #if defined(THREADED_RTS)
117 Mutex sched_mutex;
118 #endif
119
120 #if !defined(mingw32_HOST_OS)
121 #define FORKPROCESS_PRIMOP_SUPPORTED
122 #endif
123
124 /* -----------------------------------------------------------------------------
125 * static function prototypes
126 * -------------------------------------------------------------------------- */
127
128 static Capability *schedule (Capability *initialCapability, Task *task);
129
130 //
131 // These function all encapsulate parts of the scheduler loop, and are
132 // abstracted only to make the structure and control flow of the
133 // scheduler clearer.
134 //
135 static void schedulePreLoop (void);
136 static void scheduleFindWork (Capability *cap);
137 #if defined(THREADED_RTS)
138 static void scheduleYield (Capability **pcap, Task *task, rtsBool);
139 #endif
140 static void scheduleStartSignalHandlers (Capability *cap);
141 static void scheduleCheckBlockedThreads (Capability *cap);
142 static void scheduleCheckWakeupThreads(Capability *cap USED_IF_NOT_THREADS);
143 static void scheduleCheckBlackHoles (Capability *cap);
144 static void scheduleDetectDeadlock (Capability *cap, Task *task);
145 static void schedulePushWork(Capability *cap, Task *task);
146 #if defined(THREADED_RTS)
147 static void scheduleActivateSpark(Capability *cap);
148 #endif
149 static void schedulePostRunThread(Capability *cap, StgTSO *t);
150 static rtsBool scheduleHandleHeapOverflow( Capability *cap, StgTSO *t );
151 static void scheduleHandleStackOverflow( Capability *cap, Task *task,
152 StgTSO *t);
153 static rtsBool scheduleHandleYield( Capability *cap, StgTSO *t,
154 nat prev_what_next );
155 static void scheduleHandleThreadBlocked( StgTSO *t );
156 static rtsBool scheduleHandleThreadFinished( Capability *cap, Task *task,
157 StgTSO *t );
158 static rtsBool scheduleNeedHeapProfile(rtsBool ready_to_gc);
159 static Capability *scheduleDoGC(Capability *cap, Task *task,
160 rtsBool force_major);
161
162 static rtsBool checkBlackHoles(Capability *cap);
163
164 static StgTSO *threadStackOverflow(Capability *cap, StgTSO *tso);
165 static StgTSO *threadStackUnderflow(Task *task, StgTSO *tso);
166
167 static void deleteThread (Capability *cap, StgTSO *tso);
168 static void deleteAllThreads (Capability *cap);
169
170 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
171 static void deleteThread_(Capability *cap, StgTSO *tso);
172 #endif
173
174 /* -----------------------------------------------------------------------------
175 * Putting a thread on the run queue: different scheduling policies
176 * -------------------------------------------------------------------------- */
177
178 STATIC_INLINE void
179 addToRunQueue( Capability *cap, StgTSO *t )
180 {
181 // this does round-robin scheduling; good for concurrency
182 appendToRunQueue(cap,t);
183 }
184
185 /* ---------------------------------------------------------------------------
186 Main scheduling loop.
187
188 We use round-robin scheduling, each thread returning to the
189 scheduler loop when one of these conditions is detected:
190
191 * out of heap space
192 * timer expires (thread yields)
193 * thread blocks
194 * thread ends
195 * stack overflow
196
197 GRAN version:
198 In a GranSim setup this loop iterates over the global event queue.
199 This revolves around the global event queue, which determines what
200 to do next. Therefore, it's more complicated than either the
201 concurrent or the parallel (GUM) setup.
202 This version has been entirely removed (JB 2008/08).
203
204 GUM version:
205 GUM iterates over incoming messages.
206 It starts with nothing to do (thus CurrentTSO == END_TSO_QUEUE),
207 and sends out a fish whenever it has nothing to do; in-between
208 doing the actual reductions (shared code below) it processes the
209 incoming messages and deals with delayed operations
210 (see PendingFetches).
211 This is not the ugliest code you could imagine, but it's bloody close.
212
213 (JB 2008/08) This version was formerly indicated by a PP-Flag PAR,
214 now by PP-flag PARALLEL_HASKELL. The Eden RTS (in GHC-6.x) uses it,
215 as well as future GUM versions. This file has been refurbished to
216 only contain valid code, which is however incomplete, refers to
217 invalid includes etc.
218
219 ------------------------------------------------------------------------ */
220
221 static Capability *
222 schedule (Capability *initialCapability, Task *task)
223 {
224 StgTSO *t;
225 Capability *cap;
226 StgThreadReturnCode ret;
227 nat prev_what_next;
228 rtsBool ready_to_gc;
229 #if defined(THREADED_RTS)
230 rtsBool first = rtsTrue;
231 rtsBool force_yield = rtsFalse;
232 #endif
233
234 cap = initialCapability;
235
236 // Pre-condition: this task owns initialCapability.
237 // The sched_mutex is *NOT* held
238 // NB. on return, we still hold a capability.
239
240 debugTrace (DEBUG_sched, "cap %d: schedule()", initialCapability->no);
241
242 schedulePreLoop();
243
244 // -----------------------------------------------------------
245 // Scheduler loop starts here:
246
247 while (1) {
248
249 // Check whether we have re-entered the RTS from Haskell without
250 // going via suspendThread()/resumeThread (i.e. a 'safe' foreign
251 // call).
252 if (cap->in_haskell) {
253 errorBelch("schedule: re-entered unsafely.\n"
254 " Perhaps a 'foreign import unsafe' should be 'safe'?");
255 stg_exit(EXIT_FAILURE);
256 }
257
258 // The interruption / shutdown sequence.
259 //
260 // In order to cleanly shut down the runtime, we want to:
261 // * make sure that all main threads return to their callers
262 // with the state 'Interrupted'.
263 // * clean up all OS threads assocated with the runtime
264 // * free all memory etc.
265 //
266 // So the sequence for ^C goes like this:
267 //
268 // * ^C handler sets sched_state := SCHED_INTERRUPTING and
269 // arranges for some Capability to wake up
270 //
271 // * all threads in the system are halted, and the zombies are
272 // placed on the run queue for cleaning up. We acquire all
273 // the capabilities in order to delete the threads, this is
274 // done by scheduleDoGC() for convenience (because GC already
275 // needs to acquire all the capabilities). We can't kill
276 // threads involved in foreign calls.
277 //
278 // * somebody calls shutdownHaskell(), which calls exitScheduler()
279 //
280 // * sched_state := SCHED_SHUTTING_DOWN
281 //
282 // * all workers exit when the run queue on their capability
283 // drains. All main threads will also exit when their TSO
284 // reaches the head of the run queue and they can return.
285 //
286 // * eventually all Capabilities will shut down, and the RTS can
287 // exit.
288 //
289 // * We might be left with threads blocked in foreign calls,
290 // we should really attempt to kill these somehow (TODO);
291
292 switch (sched_state) {
293 case SCHED_RUNNING:
294 break;
295 case SCHED_INTERRUPTING:
296 debugTrace(DEBUG_sched, "SCHED_INTERRUPTING");
297 #if defined(THREADED_RTS)
298 discardSparksCap(cap);
299 #endif
300 /* scheduleDoGC() deletes all the threads */
301 cap = scheduleDoGC(cap,task,rtsFalse);
302
303 // after scheduleDoGC(), we must be shutting down. Either some
304 // other Capability did the final GC, or we did it above,
305 // either way we can fall through to the SCHED_SHUTTING_DOWN
306 // case now.
307 ASSERT(sched_state == SCHED_SHUTTING_DOWN);
308 // fall through
309
310 case SCHED_SHUTTING_DOWN:
311 debugTrace(DEBUG_sched, "SCHED_SHUTTING_DOWN");
312 // If we are a worker, just exit. If we're a bound thread
313 // then we will exit below when we've removed our TSO from
314 // the run queue.
315 if (task->tso == NULL && emptyRunQueue(cap)) {
316 return cap;
317 }
318 break;
319 default:
320 barf("sched_state: %d", sched_state);
321 }
322
323 scheduleFindWork(cap);
324
325 /* work pushing, currently relevant only for THREADED_RTS:
326 (pushes threads, wakes up idle capabilities for stealing) */
327 schedulePushWork(cap,task);
328
329 scheduleDetectDeadlock(cap,task);
330
331 #if defined(THREADED_RTS)
332 cap = task->cap; // reload cap, it might have changed
333 #endif
334
335 // Normally, the only way we can get here with no threads to
336 // run is if a keyboard interrupt received during
337 // scheduleCheckBlockedThreads() or scheduleDetectDeadlock().
338 // Additionally, it is not fatal for the
339 // threaded RTS to reach here with no threads to run.
340 //
341 // win32: might be here due to awaitEvent() being abandoned
342 // as a result of a console event having been delivered.
343
344 #if defined(THREADED_RTS)
345 if (first)
346 {
347 // XXX: ToDo
348 // // don't yield the first time, we want a chance to run this
349 // // thread for a bit, even if there are others banging at the
350 // // door.
351 // first = rtsFalse;
352 // ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
353 }
354
355 yield:
356 scheduleYield(&cap,task,force_yield);
357 force_yield = rtsFalse;
358
359 if (emptyRunQueue(cap)) continue; // look for work again
360 #endif
361
362 #if !defined(THREADED_RTS) && !defined(mingw32_HOST_OS)
363 if ( emptyRunQueue(cap) ) {
364 ASSERT(sched_state >= SCHED_INTERRUPTING);
365 }
366 #endif
367
368 //
369 // Get a thread to run
370 //
371 t = popRunQueue(cap);
372
373 // Sanity check the thread we're about to run. This can be
374 // expensive if there is lots of thread switching going on...
375 IF_DEBUG(sanity,checkTSO(t));
376
377 #if defined(THREADED_RTS)
378 // Check whether we can run this thread in the current task.
379 // If not, we have to pass our capability to the right task.
380 {
381 Task *bound = t->bound;
382
383 if (bound) {
384 if (bound == task) {
385 // yes, the Haskell thread is bound to the current native thread
386 } else {
387 debugTrace(DEBUG_sched,
388 "thread %lu bound to another OS thread",
389 (unsigned long)t->id);
390 // no, bound to a different Haskell thread: pass to that thread
391 pushOnRunQueue(cap,t);
392 continue;
393 }
394 } else {
395 // The thread we want to run is unbound.
396 if (task->tso) {
397 debugTrace(DEBUG_sched,
398 "this OS thread cannot run thread %lu",
399 (unsigned long)t->id);
400 // no, the current native thread is bound to a different
401 // Haskell thread, so pass it to any worker thread
402 pushOnRunQueue(cap,t);
403 continue;
404 }
405 }
406 }
407 #endif
408
409 // If we're shutting down, and this thread has not yet been
410 // killed, kill it now. This sometimes happens when a finalizer
411 // thread is created by the final GC, or a thread previously
412 // in a foreign call returns.
413 if (sched_state >= SCHED_INTERRUPTING &&
414 !(t->what_next == ThreadComplete || t->what_next == ThreadKilled)) {
415 deleteThread(cap,t);
416 }
417
418 /* context switches are initiated by the timer signal, unless
419 * the user specified "context switch as often as possible", with
420 * +RTS -C0
421 */
422 if (RtsFlags.ConcFlags.ctxtSwitchTicks == 0
423 && !emptyThreadQueues(cap)) {
424 cap->context_switch = 1;
425 }
426
427 run_thread:
428
429 // CurrentTSO is the thread to run. t might be different if we
430 // loop back to run_thread, so make sure to set CurrentTSO after
431 // that.
432 cap->r.rCurrentTSO = t;
433
434 startHeapProfTimer();
435
436 // Check for exceptions blocked on this thread
437 maybePerformBlockedException (cap, t);
438
439 // ----------------------------------------------------------------------
440 // Run the current thread
441
442 ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
443 ASSERT(t->cap == cap);
444 ASSERT(t->bound ? t->bound->cap == cap : 1);
445
446 prev_what_next = t->what_next;
447
448 errno = t->saved_errno;
449 #if mingw32_HOST_OS
450 SetLastError(t->saved_winerror);
451 #endif
452
453 cap->in_haskell = rtsTrue;
454
455 dirty_TSO(cap,t);
456
457 #if defined(THREADED_RTS)
458 if (recent_activity == ACTIVITY_DONE_GC) {
459 // ACTIVITY_DONE_GC means we turned off the timer signal to
460 // conserve power (see #1623). Re-enable it here.
461 nat prev;
462 prev = xchg((P_)&recent_activity, ACTIVITY_YES);
463 if (prev == ACTIVITY_DONE_GC) {
464 startTimer();
465 }
466 } else {
467 recent_activity = ACTIVITY_YES;
468 }
469 #endif
470
471 traceSchedEvent(cap, EVENT_RUN_THREAD, t, 0);
472
473 switch (prev_what_next) {
474
475 case ThreadKilled:
476 case ThreadComplete:
477 /* Thread already finished, return to scheduler. */
478 ret = ThreadFinished;
479 break;
480
481 case ThreadRunGHC:
482 {
483 StgRegTable *r;
484 r = StgRun((StgFunPtr) stg_returnToStackTop, &cap->r);
485 cap = regTableToCapability(r);
486 ret = r->rRet;
487 break;
488 }
489
490 case ThreadInterpret:
491 cap = interpretBCO(cap);
492 ret = cap->r.rRet;
493 break;
494
495 default:
496 barf("schedule: invalid what_next field");
497 }
498
499 cap->in_haskell = rtsFalse;
500
501 // The TSO might have moved, eg. if it re-entered the RTS and a GC
502 // happened. So find the new location:
503 t = cap->r.rCurrentTSO;
504
505 // We have run some Haskell code: there might be blackhole-blocked
506 // threads to wake up now.
507 // Lock-free test here should be ok, we're just setting a flag.
508 if ( blackhole_queue != END_TSO_QUEUE ) {
509 blackholes_need_checking = rtsTrue;
510 }
511
512 // And save the current errno in this thread.
513 // XXX: possibly bogus for SMP because this thread might already
514 // be running again, see code below.
515 t->saved_errno = errno;
516 #if mingw32_HOST_OS
517 // Similarly for Windows error code
518 t->saved_winerror = GetLastError();
519 #endif
520
521 traceSchedEvent (cap, EVENT_STOP_THREAD, t, ret);
522
523 #if defined(THREADED_RTS)
524 // If ret is ThreadBlocked, and this Task is bound to the TSO that
525 // blocked, we are in limbo - the TSO is now owned by whatever it
526 // is blocked on, and may in fact already have been woken up,
527 // perhaps even on a different Capability. It may be the case
528 // that task->cap != cap. We better yield this Capability
529 // immediately and return to normaility.
530 if (ret == ThreadBlocked) {
531 force_yield = rtsTrue;
532 goto yield;
533 }
534 #endif
535
536 ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
537 ASSERT(t->cap == cap);
538
539 // ----------------------------------------------------------------------
540
541 // Costs for the scheduler are assigned to CCS_SYSTEM
542 stopHeapProfTimer();
543 #if defined(PROFILING)
544 CCCS = CCS_SYSTEM;
545 #endif
546
547 schedulePostRunThread(cap,t);
548
549 if (ret != StackOverflow) {
550 t = threadStackUnderflow(task,t);
551 }
552
553 ready_to_gc = rtsFalse;
554
555 switch (ret) {
556 case HeapOverflow:
557 ready_to_gc = scheduleHandleHeapOverflow(cap,t);
558 break;
559
560 case StackOverflow:
561 scheduleHandleStackOverflow(cap,task,t);
562 break;
563
564 case ThreadYielding:
565 if (scheduleHandleYield(cap, t, prev_what_next)) {
566 // shortcut for switching between compiler/interpreter:
567 goto run_thread;
568 }
569 break;
570
571 case ThreadBlocked:
572 scheduleHandleThreadBlocked(t);
573 break;
574
575 case ThreadFinished:
576 if (scheduleHandleThreadFinished(cap, task, t)) return cap;
577 ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
578 break;
579
580 default:
581 barf("schedule: invalid thread return code %d", (int)ret);
582 }
583
584 if (ready_to_gc || scheduleNeedHeapProfile(ready_to_gc)) {
585 cap = scheduleDoGC(cap,task,rtsFalse);
586 }
587 } /* end of while() */
588 }
589
590 /* ----------------------------------------------------------------------------
591 * Setting up the scheduler loop
592 * ------------------------------------------------------------------------- */
593
594 static void
595 schedulePreLoop(void)
596 {
597 // initialisation for scheduler - what cannot go into initScheduler()
598 }
599
600 /* -----------------------------------------------------------------------------
601 * scheduleFindWork()
602 *
603 * Search for work to do, and handle messages from elsewhere.
604 * -------------------------------------------------------------------------- */
605
606 static void
607 scheduleFindWork (Capability *cap)
608 {
609 scheduleStartSignalHandlers(cap);
610
611 // Only check the black holes here if we've nothing else to do.
612 // During normal execution, the black hole list only gets checked
613 // at GC time, to avoid repeatedly traversing this possibly long
614 // list each time around the scheduler.
615 if (emptyRunQueue(cap)) { scheduleCheckBlackHoles(cap); }
616
617 scheduleCheckWakeupThreads(cap);
618
619 scheduleCheckBlockedThreads(cap);
620
621 #if defined(THREADED_RTS)
622 if (emptyRunQueue(cap)) { scheduleActivateSpark(cap); }
623 #endif
624 }
625
626 #if defined(THREADED_RTS)
627 STATIC_INLINE rtsBool
628 shouldYieldCapability (Capability *cap, Task *task)
629 {
630 // we need to yield this capability to someone else if..
631 // - another thread is initiating a GC
632 // - another Task is returning from a foreign call
633 // - the thread at the head of the run queue cannot be run
634 // by this Task (it is bound to another Task, or it is unbound
635 // and this task it bound).
636 return (waiting_for_gc ||
637 cap->returning_tasks_hd != NULL ||
638 (!emptyRunQueue(cap) && (task->tso == NULL
639 ? cap->run_queue_hd->bound != NULL
640 : cap->run_queue_hd->bound != task)));
641 }
642
643 // This is the single place where a Task goes to sleep. There are
644 // two reasons it might need to sleep:
645 // - there are no threads to run
646 // - we need to yield this Capability to someone else
647 // (see shouldYieldCapability())
648 //
649 // Careful: the scheduler loop is quite delicate. Make sure you run
650 // the tests in testsuite/concurrent (all ways) after modifying this,
651 // and also check the benchmarks in nofib/parallel for regressions.
652
653 static void
654 scheduleYield (Capability **pcap, Task *task, rtsBool force_yield)
655 {
656 Capability *cap = *pcap;
657
658 // if we have work, and we don't need to give up the Capability, continue.
659 //
660 // The force_yield flag is used when a bound thread blocks. This
661 // is a particularly tricky situation: the current Task does not
662 // own the TSO any more, since it is on some queue somewhere, and
663 // might be woken up or manipulated by another thread at any time.
664 // The TSO and Task might be migrated to another Capability.
665 // Certain invariants might be in doubt, such as task->bound->cap
666 // == cap. We have to yield the current Capability immediately,
667 // no messing around.
668 //
669 if (!force_yield &&
670 !shouldYieldCapability(cap,task) &&
671 (!emptyRunQueue(cap) ||
672 !emptyWakeupQueue(cap) ||
673 blackholes_need_checking ||
674 sched_state >= SCHED_INTERRUPTING))
675 return;
676
677 // otherwise yield (sleep), and keep yielding if necessary.
678 do {
679 yieldCapability(&cap,task);
680 }
681 while (shouldYieldCapability(cap,task));
682
683 // note there may still be no threads on the run queue at this
684 // point, the caller has to check.
685
686 *pcap = cap;
687 return;
688 }
689 #endif
690
691 /* -----------------------------------------------------------------------------
692 * schedulePushWork()
693 *
694 * Push work to other Capabilities if we have some.
695 * -------------------------------------------------------------------------- */
696
697 static void
698 schedulePushWork(Capability *cap USED_IF_THREADS,
699 Task *task USED_IF_THREADS)
700 {
701 /* following code not for PARALLEL_HASKELL. I kept the call general,
702 future GUM versions might use pushing in a distributed setup */
703 #if defined(THREADED_RTS)
704
705 Capability *free_caps[n_capabilities], *cap0;
706 nat i, n_free_caps;
707
708 // migration can be turned off with +RTS -qg
709 if (!RtsFlags.ParFlags.migrate) return;
710
711 // Check whether we have more threads on our run queue, or sparks
712 // in our pool, that we could hand to another Capability.
713 if (cap->run_queue_hd == END_TSO_QUEUE) {
714 if (sparkPoolSizeCap(cap) < 2) return;
715 } else {
716 if (cap->run_queue_hd->_link == END_TSO_QUEUE &&
717 sparkPoolSizeCap(cap) < 1) return;
718 }
719
720 // First grab as many free Capabilities as we can.
721 for (i=0, n_free_caps=0; i < n_capabilities; i++) {
722 cap0 = &capabilities[i];
723 if (cap != cap0 && tryGrabCapability(cap0,task)) {
724 if (!emptyRunQueue(cap0) || cap->returning_tasks_hd != NULL) {
725 // it already has some work, we just grabbed it at
726 // the wrong moment. Or maybe it's deadlocked!
727 releaseCapability(cap0);
728 } else {
729 free_caps[n_free_caps++] = cap0;
730 }
731 }
732 }
733
734 // we now have n_free_caps free capabilities stashed in
735 // free_caps[]. Share our run queue equally with them. This is
736 // probably the simplest thing we could do; improvements we might
737 // want to do include:
738 //
739 // - giving high priority to moving relatively new threads, on
740 // the gournds that they haven't had time to build up a
741 // working set in the cache on this CPU/Capability.
742 //
743 // - giving low priority to moving long-lived threads
744
745 if (n_free_caps > 0) {
746 StgTSO *prev, *t, *next;
747 rtsBool pushed_to_all;
748
749 debugTrace(DEBUG_sched,
750 "cap %d: %s and %d free capabilities, sharing...",
751 cap->no,
752 (!emptyRunQueue(cap) && cap->run_queue_hd->_link != END_TSO_QUEUE)?
753 "excess threads on run queue":"sparks to share (>=2)",
754 n_free_caps);
755
756 i = 0;
757 pushed_to_all = rtsFalse;
758
759 if (cap->run_queue_hd != END_TSO_QUEUE) {
760 prev = cap->run_queue_hd;
761 t = prev->_link;
762 prev->_link = END_TSO_QUEUE;
763 for (; t != END_TSO_QUEUE; t = next) {
764 next = t->_link;
765 t->_link = END_TSO_QUEUE;
766 if (t->what_next == ThreadRelocated
767 || t->bound == task // don't move my bound thread
768 || tsoLocked(t)) { // don't move a locked thread
769 setTSOLink(cap, prev, t);
770 prev = t;
771 } else if (i == n_free_caps) {
772 pushed_to_all = rtsTrue;
773 i = 0;
774 // keep one for us
775 setTSOLink(cap, prev, t);
776 prev = t;
777 } else {
778 debugTrace(DEBUG_sched, "pushing thread %lu to capability %d", (unsigned long)t->id, free_caps[i]->no);
779 appendToRunQueue(free_caps[i],t);
780
781 traceSchedEvent (cap, EVENT_MIGRATE_THREAD, t, free_caps[i]->no);
782
783 if (t->bound) { t->bound->cap = free_caps[i]; }
784 t->cap = free_caps[i];
785 i++;
786 }
787 }
788 cap->run_queue_tl = prev;
789 }
790
791 #ifdef SPARK_PUSHING
792 /* JB I left this code in place, it would work but is not necessary */
793
794 // If there are some free capabilities that we didn't push any
795 // threads to, then try to push a spark to each one.
796 if (!pushed_to_all) {
797 StgClosure *spark;
798 // i is the next free capability to push to
799 for (; i < n_free_caps; i++) {
800 if (emptySparkPoolCap(free_caps[i])) {
801 spark = tryStealSpark(cap->sparks);
802 if (spark != NULL) {
803 debugTrace(DEBUG_sched, "pushing spark %p to capability %d", spark, free_caps[i]->no);
804
805 traceSchedEvent(free_caps[i], EVENT_STEAL_SPARK, t, cap->no);
806
807 newSpark(&(free_caps[i]->r), spark);
808 }
809 }
810 }
811 }
812 #endif /* SPARK_PUSHING */
813
814 // release the capabilities
815 for (i = 0; i < n_free_caps; i++) {
816 task->cap = free_caps[i];
817 releaseAndWakeupCapability(free_caps[i]);
818 }
819 }
820 task->cap = cap; // reset to point to our Capability.
821
822 #endif /* THREADED_RTS */
823
824 }
825
826 /* ----------------------------------------------------------------------------
827 * Start any pending signal handlers
828 * ------------------------------------------------------------------------- */
829
830 #if defined(RTS_USER_SIGNALS) && !defined(THREADED_RTS)
831 static void
832 scheduleStartSignalHandlers(Capability *cap)
833 {
834 if (RtsFlags.MiscFlags.install_signal_handlers && signals_pending()) {
835 // safe outside the lock
836 startSignalHandlers(cap);
837 }
838 }
839 #else
840 static void
841 scheduleStartSignalHandlers(Capability *cap STG_UNUSED)
842 {
843 }
844 #endif
845
846 /* ----------------------------------------------------------------------------
847 * Check for blocked threads that can be woken up.
848 * ------------------------------------------------------------------------- */
849
850 static void
851 scheduleCheckBlockedThreads(Capability *cap USED_IF_NOT_THREADS)
852 {
853 #if !defined(THREADED_RTS)
854 //
855 // Check whether any waiting threads need to be woken up. If the
856 // run queue is empty, and there are no other tasks running, we
857 // can wait indefinitely for something to happen.
858 //
859 if ( !emptyQueue(blocked_queue_hd) || !emptyQueue(sleeping_queue) )
860 {
861 awaitEvent( emptyRunQueue(cap) && !blackholes_need_checking );
862 }
863 #endif
864 }
865
866
867 /* ----------------------------------------------------------------------------
868 * Check for threads woken up by other Capabilities
869 * ------------------------------------------------------------------------- */
870
871 static void
872 scheduleCheckWakeupThreads(Capability *cap USED_IF_THREADS)
873 {
874 #if defined(THREADED_RTS)
875 // Any threads that were woken up by other Capabilities get
876 // appended to our run queue.
877 if (!emptyWakeupQueue(cap)) {
878 ACQUIRE_LOCK(&cap->lock);
879 if (emptyRunQueue(cap)) {
880 cap->run_queue_hd = cap->wakeup_queue_hd;
881 cap->run_queue_tl = cap->wakeup_queue_tl;
882 } else {
883 setTSOLink(cap, cap->run_queue_tl, cap->wakeup_queue_hd);
884 cap->run_queue_tl = cap->wakeup_queue_tl;
885 }
886 cap->wakeup_queue_hd = cap->wakeup_queue_tl = END_TSO_QUEUE;
887 RELEASE_LOCK(&cap->lock);
888 }
889 #endif
890 }
891
892 /* ----------------------------------------------------------------------------
893 * Check for threads blocked on BLACKHOLEs that can be woken up
894 * ------------------------------------------------------------------------- */
895 static void
896 scheduleCheckBlackHoles (Capability *cap)
897 {
898 if ( blackholes_need_checking ) // check without the lock first
899 {
900 ACQUIRE_LOCK(&sched_mutex);
901 if ( blackholes_need_checking ) {
902 blackholes_need_checking = rtsFalse;
903 // important that we reset the flag *before* checking the
904 // blackhole queue, otherwise we could get deadlock. This
905 // happens as follows: we wake up a thread that
906 // immediately runs on another Capability, blocks on a
907 // blackhole, and then we reset the blackholes_need_checking flag.
908 checkBlackHoles(cap);
909 }
910 RELEASE_LOCK(&sched_mutex);
911 }
912 }
913
914 /* ----------------------------------------------------------------------------
915 * Detect deadlock conditions and attempt to resolve them.
916 * ------------------------------------------------------------------------- */
917
918 static void
919 scheduleDetectDeadlock (Capability *cap, Task *task)
920 {
921 /*
922 * Detect deadlock: when we have no threads to run, there are no
923 * threads blocked, waiting for I/O, or sleeping, and all the
924 * other tasks are waiting for work, we must have a deadlock of
925 * some description.
926 */
927 if ( emptyThreadQueues(cap) )
928 {
929 #if defined(THREADED_RTS)
930 /*
931 * In the threaded RTS, we only check for deadlock if there
932 * has been no activity in a complete timeslice. This means
933 * we won't eagerly start a full GC just because we don't have
934 * any threads to run currently.
935 */
936 if (recent_activity != ACTIVITY_INACTIVE) return;
937 #endif
938
939 debugTrace(DEBUG_sched, "deadlocked, forcing major GC...");
940
941 // Garbage collection can release some new threads due to
942 // either (a) finalizers or (b) threads resurrected because
943 // they are unreachable and will therefore be sent an
944 // exception. Any threads thus released will be immediately
945 // runnable.
946 cap = scheduleDoGC (cap, task, rtsTrue/*force major GC*/);
947 // when force_major == rtsTrue. scheduleDoGC sets
948 // recent_activity to ACTIVITY_DONE_GC and turns off the timer
949 // signal.
950
951 if ( !emptyRunQueue(cap) ) return;
952
953 #if defined(RTS_USER_SIGNALS) && !defined(THREADED_RTS)
954 /* If we have user-installed signal handlers, then wait
955 * for signals to arrive rather then bombing out with a
956 * deadlock.
957 */
958 if ( RtsFlags.MiscFlags.install_signal_handlers && anyUserHandlers() ) {
959 debugTrace(DEBUG_sched,
960 "still deadlocked, waiting for signals...");
961
962 awaitUserSignals();
963
964 if (signals_pending()) {
965 startSignalHandlers(cap);
966 }
967
968 // either we have threads to run, or we were interrupted:
969 ASSERT(!emptyRunQueue(cap) || sched_state >= SCHED_INTERRUPTING);
970
971 return;
972 }
973 #endif
974
975 #if !defined(THREADED_RTS)
976 /* Probably a real deadlock. Send the current main thread the
977 * Deadlock exception.
978 */
979 if (task->tso) {
980 switch (task->tso->why_blocked) {
981 case BlockedOnSTM:
982 case BlockedOnBlackHole:
983 case BlockedOnException:
984 case BlockedOnMVar:
985 throwToSingleThreaded(cap, task->tso,
986 (StgClosure *)nonTermination_closure);
987 return;
988 default:
989 barf("deadlock: main thread blocked in a strange way");
990 }
991 }
992 return;
993 #endif
994 }
995 }
996
997
998 /* ----------------------------------------------------------------------------
999 * Send pending messages (PARALLEL_HASKELL only)
1000 * ------------------------------------------------------------------------- */
1001
1002 #if defined(PARALLEL_HASKELL)
1003 static void
1004 scheduleSendPendingMessages(void)
1005 {
1006
1007 # if defined(PAR) // global Mem.Mgmt., omit for now
1008 if (PendingFetches != END_BF_QUEUE) {
1009 processFetches();
1010 }
1011 # endif
1012
1013 if (RtsFlags.ParFlags.BufferTime) {
1014 // if we use message buffering, we must send away all message
1015 // packets which have become too old...
1016 sendOldBuffers();
1017 }
1018 }
1019 #endif
1020
1021 /* ----------------------------------------------------------------------------
1022 * Activate spark threads (PARALLEL_HASKELL and THREADED_RTS)
1023 * ------------------------------------------------------------------------- */
1024
1025 #if defined(THREADED_RTS)
1026 static void
1027 scheduleActivateSpark(Capability *cap)
1028 {
1029 if (anySparks())
1030 {
1031 createSparkThread(cap);
1032 debugTrace(DEBUG_sched, "creating a spark thread");
1033 }
1034 }
1035 #endif // PARALLEL_HASKELL || THREADED_RTS
1036
1037 /* ----------------------------------------------------------------------------
1038 * After running a thread...
1039 * ------------------------------------------------------------------------- */
1040
1041 static void
1042 schedulePostRunThread (Capability *cap, StgTSO *t)
1043 {
1044 // We have to be able to catch transactions that are in an
1045 // infinite loop as a result of seeing an inconsistent view of
1046 // memory, e.g.
1047 //
1048 // atomically $ do
1049 // [a,b] <- mapM readTVar [ta,tb]
1050 // when (a == b) loop
1051 //
1052 // and a is never equal to b given a consistent view of memory.
1053 //
1054 if (t -> trec != NO_TREC && t -> why_blocked == NotBlocked) {
1055 if (!stmValidateNestOfTransactions (t -> trec)) {
1056 debugTrace(DEBUG_sched | DEBUG_stm,
1057 "trec %p found wasting its time", t);
1058
1059 // strip the stack back to the
1060 // ATOMICALLY_FRAME, aborting the (nested)
1061 // transaction, and saving the stack of any
1062 // partially-evaluated thunks on the heap.
1063 throwToSingleThreaded_(cap, t, NULL, rtsTrue);
1064
1065 ASSERT(get_itbl((StgClosure *)t->sp)->type == ATOMICALLY_FRAME);
1066 }
1067 }
1068
1069 /* some statistics gathering in the parallel case */
1070 }
1071
1072 /* -----------------------------------------------------------------------------
1073 * Handle a thread that returned to the scheduler with ThreadHeepOverflow
1074 * -------------------------------------------------------------------------- */
1075
1076 static rtsBool
1077 scheduleHandleHeapOverflow( Capability *cap, StgTSO *t )
1078 {
1079 // did the task ask for a large block?
1080 if (cap->r.rHpAlloc > BLOCK_SIZE) {
1081 // if so, get one and push it on the front of the nursery.
1082 bdescr *bd;
1083 lnat blocks;
1084
1085 blocks = (lnat)BLOCK_ROUND_UP(cap->r.rHpAlloc) / BLOCK_SIZE;
1086
1087 debugTrace(DEBUG_sched,
1088 "--<< thread %ld (%s) stopped: requesting a large block (size %ld)\n",
1089 (long)t->id, what_next_strs[t->what_next], blocks);
1090
1091 // don't do this if the nursery is (nearly) full, we'll GC first.
1092 if (cap->r.rCurrentNursery->link != NULL ||
1093 cap->r.rNursery->n_blocks == 1) { // paranoia to prevent infinite loop
1094 // if the nursery has only one block.
1095
1096 ACQUIRE_SM_LOCK
1097 bd = allocGroup( blocks );
1098 RELEASE_SM_LOCK
1099 cap->r.rNursery->n_blocks += blocks;
1100
1101 // link the new group into the list
1102 bd->link = cap->r.rCurrentNursery;
1103 bd->u.back = cap->r.rCurrentNursery->u.back;
1104 if (cap->r.rCurrentNursery->u.back != NULL) {
1105 cap->r.rCurrentNursery->u.back->link = bd;
1106 } else {
1107 cap->r.rNursery->blocks = bd;
1108 }
1109 cap->r.rCurrentNursery->u.back = bd;
1110
1111 // initialise it as a nursery block. We initialise the
1112 // step, gen_no, and flags field of *every* sub-block in
1113 // this large block, because this is easier than making
1114 // sure that we always find the block head of a large
1115 // block whenever we call Bdescr() (eg. evacuate() and
1116 // isAlive() in the GC would both have to do this, at
1117 // least).
1118 {
1119 bdescr *x;
1120 for (x = bd; x < bd + blocks; x++) {
1121 x->step = cap->r.rNursery;
1122 x->gen_no = 0;
1123 x->flags = 0;
1124 }
1125 }
1126
1127 // This assert can be a killer if the app is doing lots
1128 // of large block allocations.
1129 IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
1130
1131 // now update the nursery to point to the new block
1132 cap->r.rCurrentNursery = bd;
1133
1134 // we might be unlucky and have another thread get on the
1135 // run queue before us and steal the large block, but in that
1136 // case the thread will just end up requesting another large
1137 // block.
1138 pushOnRunQueue(cap,t);
1139 return rtsFalse; /* not actually GC'ing */
1140 }
1141 }
1142
1143 if (cap->r.rHpLim == NULL || cap->context_switch) {
1144 // Sometimes we miss a context switch, e.g. when calling
1145 // primitives in a tight loop, MAYBE_GC() doesn't check the
1146 // context switch flag, and we end up waiting for a GC.
1147 // See #1984, and concurrent/should_run/1984
1148 cap->context_switch = 0;
1149 addToRunQueue(cap,t);
1150 } else {
1151 pushOnRunQueue(cap,t);
1152 }
1153 return rtsTrue;
1154 /* actual GC is done at the end of the while loop in schedule() */
1155 }
1156
1157 /* -----------------------------------------------------------------------------
1158 * Handle a thread that returned to the scheduler with ThreadStackOverflow
1159 * -------------------------------------------------------------------------- */
1160
1161 static void
1162 scheduleHandleStackOverflow (Capability *cap, Task *task, StgTSO *t)
1163 {
1164 /* just adjust the stack for this thread, then pop it back
1165 * on the run queue.
1166 */
1167 {
1168 /* enlarge the stack */
1169 StgTSO *new_t = threadStackOverflow(cap, t);
1170
1171 /* The TSO attached to this Task may have moved, so update the
1172 * pointer to it.
1173 */
1174 if (task->tso == t) {
1175 task->tso = new_t;
1176 }
1177 pushOnRunQueue(cap,new_t);
1178 }
1179 }
1180
1181 /* -----------------------------------------------------------------------------
1182 * Handle a thread that returned to the scheduler with ThreadYielding
1183 * -------------------------------------------------------------------------- */
1184
1185 static rtsBool
1186 scheduleHandleYield( Capability *cap, StgTSO *t, nat prev_what_next )
1187 {
1188 // Reset the context switch flag. We don't do this just before
1189 // running the thread, because that would mean we would lose ticks
1190 // during GC, which can lead to unfair scheduling (a thread hogs
1191 // the CPU because the tick always arrives during GC). This way
1192 // penalises threads that do a lot of allocation, but that seems
1193 // better than the alternative.
1194 cap->context_switch = 0;
1195
1196 /* put the thread back on the run queue. Then, if we're ready to
1197 * GC, check whether this is the last task to stop. If so, wake
1198 * up the GC thread. getThread will block during a GC until the
1199 * GC is finished.
1200 */
1201 #ifdef DEBUG
1202 if (t->what_next != prev_what_next) {
1203 debugTrace(DEBUG_sched,
1204 "--<< thread %ld (%s) stopped to switch evaluators",
1205 (long)t->id, what_next_strs[t->what_next]);
1206 }
1207 #endif
1208
1209 IF_DEBUG(sanity,
1210 //debugBelch("&& Doing sanity check on yielding TSO %ld.", t->id);
1211 checkTSO(t));
1212 ASSERT(t->_link == END_TSO_QUEUE);
1213
1214 // Shortcut if we're just switching evaluators: don't bother
1215 // doing stack squeezing (which can be expensive), just run the
1216 // thread.
1217 if (t->what_next != prev_what_next) {
1218 return rtsTrue;
1219 }
1220
1221 addToRunQueue(cap,t);
1222
1223 return rtsFalse;
1224 }
1225
1226 /* -----------------------------------------------------------------------------
1227 * Handle a thread that returned to the scheduler with ThreadBlocked
1228 * -------------------------------------------------------------------------- */
1229
1230 static void
1231 scheduleHandleThreadBlocked( StgTSO *t
1232 #if !defined(DEBUG)
1233 STG_UNUSED
1234 #endif
1235 )
1236 {
1237
1238 // We don't need to do anything. The thread is blocked, and it
1239 // has tidied up its stack and placed itself on whatever queue
1240 // it needs to be on.
1241
1242 // ASSERT(t->why_blocked != NotBlocked);
1243 // Not true: for example,
1244 // - in THREADED_RTS, the thread may already have been woken
1245 // up by another Capability. This actually happens: try
1246 // conc023 +RTS -N2.
1247 // - the thread may have woken itself up already, because
1248 // threadPaused() might have raised a blocked throwTo
1249 // exception, see maybePerformBlockedException().
1250
1251 #ifdef DEBUG
1252 traceThreadStatus(DEBUG_sched, t);
1253 #endif
1254 }
1255
1256 /* -----------------------------------------------------------------------------
1257 * Handle a thread that returned to the scheduler with ThreadFinished
1258 * -------------------------------------------------------------------------- */
1259
1260 static rtsBool
1261 scheduleHandleThreadFinished (Capability *cap STG_UNUSED, Task *task, StgTSO *t)
1262 {
1263 /* Need to check whether this was a main thread, and if so,
1264 * return with the return value.
1265 *
1266 * We also end up here if the thread kills itself with an
1267 * uncaught exception, see Exception.cmm.
1268 */
1269
1270 // blocked exceptions can now complete, even if the thread was in
1271 // blocked mode (see #2910). This unconditionally calls
1272 // lockTSO(), which ensures that we don't miss any threads that
1273 // are engaged in throwTo() with this thread as a target.
1274 awakenBlockedExceptionQueue (cap, t);
1275
1276 //
1277 // Check whether the thread that just completed was a bound
1278 // thread, and if so return with the result.
1279 //
1280 // There is an assumption here that all thread completion goes
1281 // through this point; we need to make sure that if a thread
1282 // ends up in the ThreadKilled state, that it stays on the run
1283 // queue so it can be dealt with here.
1284 //
1285
1286 if (t->bound) {
1287
1288 if (t->bound != task) {
1289 #if !defined(THREADED_RTS)
1290 // Must be a bound thread that is not the topmost one. Leave
1291 // it on the run queue until the stack has unwound to the
1292 // point where we can deal with this. Leaving it on the run
1293 // queue also ensures that the garbage collector knows about
1294 // this thread and its return value (it gets dropped from the
1295 // step->threads list so there's no other way to find it).
1296 appendToRunQueue(cap,t);
1297 return rtsFalse;
1298 #else
1299 // this cannot happen in the threaded RTS, because a
1300 // bound thread can only be run by the appropriate Task.
1301 barf("finished bound thread that isn't mine");
1302 #endif
1303 }
1304
1305 ASSERT(task->tso == t);
1306
1307 if (t->what_next == ThreadComplete) {
1308 if (task->ret) {
1309 // NOTE: return val is tso->sp[1] (see StgStartup.hc)
1310 *(task->ret) = (StgClosure *)task->tso->sp[1];
1311 }
1312 task->stat = Success;
1313 } else {
1314 if (task->ret) {
1315 *(task->ret) = NULL;
1316 }
1317 if (sched_state >= SCHED_INTERRUPTING) {
1318 if (heap_overflow) {
1319 task->stat = HeapExhausted;
1320 } else {
1321 task->stat = Interrupted;
1322 }
1323 } else {
1324 task->stat = Killed;
1325 }
1326 }
1327 #ifdef DEBUG
1328 removeThreadLabel((StgWord)task->tso->id);
1329 #endif
1330 return rtsTrue; // tells schedule() to return
1331 }
1332
1333 return rtsFalse;
1334 }
1335
1336 /* -----------------------------------------------------------------------------
1337 * Perform a heap census
1338 * -------------------------------------------------------------------------- */
1339
1340 static rtsBool
1341 scheduleNeedHeapProfile( rtsBool ready_to_gc STG_UNUSED )
1342 {
1343 // When we have +RTS -i0 and we're heap profiling, do a census at
1344 // every GC. This lets us get repeatable runs for debugging.
1345 if (performHeapProfile ||
1346 (RtsFlags.ProfFlags.profileInterval==0 &&
1347 RtsFlags.ProfFlags.doHeapProfile && ready_to_gc)) {
1348 return rtsTrue;
1349 } else {
1350 return rtsFalse;
1351 }
1352 }
1353
1354 /* -----------------------------------------------------------------------------
1355 * Perform a garbage collection if necessary
1356 * -------------------------------------------------------------------------- */
1357
1358 static Capability *
1359 scheduleDoGC (Capability *cap, Task *task USED_IF_THREADS, rtsBool force_major)
1360 {
1361 rtsBool heap_census;
1362 #ifdef THREADED_RTS
1363 /* extern static volatile StgWord waiting_for_gc;
1364 lives inside capability.c */
1365 rtsBool gc_type, prev_pending_gc;
1366 nat i;
1367 #endif
1368
1369 if (sched_state == SCHED_SHUTTING_DOWN) {
1370 // The final GC has already been done, and the system is
1371 // shutting down. We'll probably deadlock if we try to GC
1372 // now.
1373 return cap;
1374 }
1375
1376 #ifdef THREADED_RTS
1377 if (sched_state < SCHED_INTERRUPTING
1378 && RtsFlags.ParFlags.parGcEnabled
1379 && N >= RtsFlags.ParFlags.parGcGen
1380 && ! oldest_gen->steps[0].mark)
1381 {
1382 gc_type = PENDING_GC_PAR;
1383 } else {
1384 gc_type = PENDING_GC_SEQ;
1385 }
1386
1387 // In order to GC, there must be no threads running Haskell code.
1388 // Therefore, the GC thread needs to hold *all* the capabilities,
1389 // and release them after the GC has completed.
1390 //
1391 // This seems to be the simplest way: previous attempts involved
1392 // making all the threads with capabilities give up their
1393 // capabilities and sleep except for the *last* one, which
1394 // actually did the GC. But it's quite hard to arrange for all
1395 // the other tasks to sleep and stay asleep.
1396 //
1397
1398 /* Other capabilities are prevented from running yet more Haskell
1399 threads if waiting_for_gc is set. Tested inside
1400 yieldCapability() and releaseCapability() in Capability.c */
1401
1402 prev_pending_gc = cas(&waiting_for_gc, 0, gc_type);
1403 if (prev_pending_gc) {
1404 do {
1405 debugTrace(DEBUG_sched, "someone else is trying to GC (%d)...",
1406 prev_pending_gc);
1407 ASSERT(cap);
1408 yieldCapability(&cap,task);
1409 } while (waiting_for_gc);
1410 return cap; // NOTE: task->cap might have changed here
1411 }
1412
1413 setContextSwitches();
1414
1415 // The final shutdown GC is always single-threaded, because it's
1416 // possible that some of the Capabilities have no worker threads.
1417
1418 if (gc_type == PENDING_GC_SEQ)
1419 {
1420 traceSchedEvent(cap, EVENT_REQUEST_SEQ_GC, 0, 0);
1421 // single-threaded GC: grab all the capabilities
1422 for (i=0; i < n_capabilities; i++) {
1423 debugTrace(DEBUG_sched, "ready_to_gc, grabbing all the capabilies (%d/%d)", i, n_capabilities);
1424 if (cap != &capabilities[i]) {
1425 Capability *pcap = &capabilities[i];
1426 // we better hope this task doesn't get migrated to
1427 // another Capability while we're waiting for this one.
1428 // It won't, because load balancing happens while we have
1429 // all the Capabilities, but even so it's a slightly
1430 // unsavoury invariant.
1431 task->cap = pcap;
1432 waitForReturnCapability(&pcap, task);
1433 if (pcap != &capabilities[i]) {
1434 barf("scheduleDoGC: got the wrong capability");
1435 }
1436 }
1437 }
1438 }
1439 else
1440 {
1441 // multi-threaded GC: make sure all the Capabilities donate one
1442 // GC thread each.
1443 traceSchedEvent(cap, EVENT_REQUEST_PAR_GC, 0, 0);
1444 debugTrace(DEBUG_sched, "ready_to_gc, grabbing GC threads");
1445
1446 waitForGcThreads(cap);
1447 }
1448 #endif
1449
1450 // so this happens periodically:
1451 if (cap) scheduleCheckBlackHoles(cap);
1452
1453 IF_DEBUG(scheduler, printAllThreads());
1454
1455 delete_threads_and_gc:
1456 /*
1457 * We now have all the capabilities; if we're in an interrupting
1458 * state, then we should take the opportunity to delete all the
1459 * threads in the system.
1460 */
1461 if (sched_state == SCHED_INTERRUPTING) {
1462 deleteAllThreads(cap);
1463 sched_state = SCHED_SHUTTING_DOWN;
1464 }
1465
1466 heap_census = scheduleNeedHeapProfile(rtsTrue);
1467
1468 #if defined(THREADED_RTS)
1469 traceSchedEvent(cap, EVENT_GC_START, 0, 0);
1470 // reset waiting_for_gc *before* GC, so that when the GC threads
1471 // emerge they don't immediately re-enter the GC.
1472 waiting_for_gc = 0;
1473 GarbageCollect(force_major || heap_census, gc_type, cap);
1474 #else
1475 GarbageCollect(force_major || heap_census, 0, cap);
1476 #endif
1477 traceSchedEvent(cap, EVENT_GC_END, 0, 0);
1478
1479 if (recent_activity == ACTIVITY_INACTIVE && force_major)
1480 {
1481 // We are doing a GC because the system has been idle for a
1482 // timeslice and we need to check for deadlock. Record the
1483 // fact that we've done a GC and turn off the timer signal;
1484 // it will get re-enabled if we run any threads after the GC.
1485 recent_activity = ACTIVITY_DONE_GC;
1486 stopTimer();
1487 }
1488 else
1489 {
1490 // the GC might have taken long enough for the timer to set
1491 // recent_activity = ACTIVITY_INACTIVE, but we aren't
1492 // necessarily deadlocked:
1493 recent_activity = ACTIVITY_YES;
1494 }
1495
1496 #if defined(THREADED_RTS)
1497 if (gc_type == PENDING_GC_PAR)
1498 {
1499 releaseGCThreads(cap);
1500 }
1501 #endif
1502
1503 if (heap_census) {
1504 debugTrace(DEBUG_sched, "performing heap census");
1505 heapCensus();
1506 performHeapProfile = rtsFalse;
1507 }
1508
1509 if (heap_overflow && sched_state < SCHED_INTERRUPTING) {
1510 // GC set the heap_overflow flag, so we should proceed with
1511 // an orderly shutdown now. Ultimately we want the main
1512 // thread to return to its caller with HeapExhausted, at which
1513 // point the caller should call hs_exit(). The first step is
1514 // to delete all the threads.
1515 //
1516 // Another way to do this would be to raise an exception in
1517 // the main thread, which we really should do because it gives
1518 // the program a chance to clean up. But how do we find the
1519 // main thread? It should presumably be the same one that
1520 // gets ^C exceptions, but that's all done on the Haskell side
1521 // (GHC.TopHandler).
1522 sched_state = SCHED_INTERRUPTING;
1523 goto delete_threads_and_gc;
1524 }
1525
1526 #ifdef SPARKBALANCE
1527 /* JB
1528 Once we are all together... this would be the place to balance all
1529 spark pools. No concurrent stealing or adding of new sparks can
1530 occur. Should be defined in Sparks.c. */
1531 balanceSparkPoolsCaps(n_capabilities, capabilities);
1532 #endif
1533
1534 #if defined(THREADED_RTS)
1535 if (gc_type == PENDING_GC_SEQ) {
1536 // release our stash of capabilities.
1537 for (i = 0; i < n_capabilities; i++) {
1538 if (cap != &capabilities[i]) {
1539 task->cap = &capabilities[i];
1540 releaseCapability(&capabilities[i]);
1541 }
1542 }
1543 }
1544 if (cap) {
1545 task->cap = cap;
1546 } else {
1547 task->cap = NULL;
1548 }
1549 #endif
1550
1551 return cap;
1552 }
1553
1554 /* ---------------------------------------------------------------------------
1555 * Singleton fork(). Do not copy any running threads.
1556 * ------------------------------------------------------------------------- */
1557
1558 pid_t
1559 forkProcess(HsStablePtr *entry
1560 #ifndef FORKPROCESS_PRIMOP_SUPPORTED
1561 STG_UNUSED
1562 #endif
1563 )
1564 {
1565 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
1566 Task *task;
1567 pid_t pid;
1568 StgTSO* t,*next;
1569 Capability *cap;
1570 nat s;
1571
1572 #if defined(THREADED_RTS)
1573 if (RtsFlags.ParFlags.nNodes > 1) {
1574 errorBelch("forking not supported with +RTS -N<n> greater than 1");
1575 stg_exit(EXIT_FAILURE);
1576 }
1577 #endif
1578
1579 debugTrace(DEBUG_sched, "forking!");
1580
1581 // ToDo: for SMP, we should probably acquire *all* the capabilities
1582 cap = rts_lock();
1583
1584 // no funny business: hold locks while we fork, otherwise if some
1585 // other thread is holding a lock when the fork happens, the data
1586 // structure protected by the lock will forever be in an
1587 // inconsistent state in the child. See also #1391.
1588 ACQUIRE_LOCK(&sched_mutex);
1589 ACQUIRE_LOCK(&cap->lock);
1590 ACQUIRE_LOCK(&cap->running_task->lock);
1591
1592 pid = fork();
1593
1594 if (pid) { // parent
1595
1596 RELEASE_LOCK(&sched_mutex);
1597 RELEASE_LOCK(&cap->lock);
1598 RELEASE_LOCK(&cap->running_task->lock);
1599
1600 // just return the pid
1601 rts_unlock(cap);
1602 return pid;
1603
1604 } else { // child
1605
1606 #if defined(THREADED_RTS)
1607 initMutex(&sched_mutex);
1608 initMutex(&cap->lock);
1609 initMutex(&cap->running_task->lock);
1610 #endif
1611
1612 // Now, all OS threads except the thread that forked are
1613 // stopped. We need to stop all Haskell threads, including
1614 // those involved in foreign calls. Also we need to delete
1615 // all Tasks, because they correspond to OS threads that are
1616 // now gone.
1617
1618 for (s = 0; s < total_steps; s++) {
1619 for (t = all_steps[s].threads; t != END_TSO_QUEUE; t = next) {
1620 if (t->what_next == ThreadRelocated) {
1621 next = t->_link;
1622 } else {
1623 next = t->global_link;
1624 // don't allow threads to catch the ThreadKilled
1625 // exception, but we do want to raiseAsync() because these
1626 // threads may be evaluating thunks that we need later.
1627 deleteThread_(cap,t);
1628 }
1629 }
1630 }
1631
1632 // Empty the run queue. It seems tempting to let all the
1633 // killed threads stay on the run queue as zombies to be
1634 // cleaned up later, but some of them correspond to bound
1635 // threads for which the corresponding Task does not exist.
1636 cap->run_queue_hd = END_TSO_QUEUE;
1637 cap->run_queue_tl = END_TSO_QUEUE;
1638
1639 // Any suspended C-calling Tasks are no more, their OS threads
1640 // don't exist now:
1641 cap->suspended_ccalling_tasks = NULL;
1642
1643 // Empty the threads lists. Otherwise, the garbage
1644 // collector may attempt to resurrect some of these threads.
1645 for (s = 0; s < total_steps; s++) {
1646 all_steps[s].threads = END_TSO_QUEUE;
1647 }
1648
1649 // Wipe the task list, except the current Task.
1650 ACQUIRE_LOCK(&sched_mutex);
1651 for (task = all_tasks; task != NULL; task=task->all_link) {
1652 if (task != cap->running_task) {
1653 #if defined(THREADED_RTS)
1654 initMutex(&task->lock); // see #1391
1655 #endif
1656 discardTask(task);
1657 }
1658 }
1659 RELEASE_LOCK(&sched_mutex);
1660
1661 #if defined(THREADED_RTS)
1662 // Wipe our spare workers list, they no longer exist. New
1663 // workers will be created if necessary.
1664 cap->spare_workers = NULL;
1665 cap->returning_tasks_hd = NULL;
1666 cap->returning_tasks_tl = NULL;
1667 #endif
1668
1669 // On Unix, all timers are reset in the child, so we need to start
1670 // the timer again.
1671 initTimer();
1672 startTimer();
1673
1674 cap = rts_evalStableIO(cap, entry, NULL); // run the action
1675 rts_checkSchedStatus("forkProcess",cap);
1676
1677 rts_unlock(cap);
1678 hs_exit(); // clean up and exit
1679 stg_exit(EXIT_SUCCESS);
1680 }
1681 #else /* !FORKPROCESS_PRIMOP_SUPPORTED */
1682 barf("forkProcess#: primop not supported on this platform, sorry!\n");
1683 #endif
1684 }
1685
1686 /* ---------------------------------------------------------------------------
1687 * Delete all the threads in the system
1688 * ------------------------------------------------------------------------- */
1689
1690 static void
1691 deleteAllThreads ( Capability *cap )
1692 {
1693 // NOTE: only safe to call if we own all capabilities.
1694
1695 StgTSO* t, *next;
1696 nat s;
1697
1698 debugTrace(DEBUG_sched,"deleting all threads");
1699 for (s = 0; s < total_steps; s++) {
1700 for (t = all_steps[s].threads; t != END_TSO_QUEUE; t = next) {
1701 if (t->what_next == ThreadRelocated) {
1702 next = t->_link;
1703 } else {
1704 next = t->global_link;
1705 deleteThread(cap,t);
1706 }
1707 }
1708 }
1709
1710 // The run queue now contains a bunch of ThreadKilled threads. We
1711 // must not throw these away: the main thread(s) will be in there
1712 // somewhere, and the main scheduler loop has to deal with it.
1713 // Also, the run queue is the only thing keeping these threads from
1714 // being GC'd, and we don't want the "main thread has been GC'd" panic.
1715
1716 #if !defined(THREADED_RTS)
1717 ASSERT(blocked_queue_hd == END_TSO_QUEUE);
1718 ASSERT(sleeping_queue == END_TSO_QUEUE);
1719 #endif
1720 }
1721
1722 /* -----------------------------------------------------------------------------
1723 Managing the suspended_ccalling_tasks list.
1724 Locks required: sched_mutex
1725 -------------------------------------------------------------------------- */
1726
1727 STATIC_INLINE void
1728 suspendTask (Capability *cap, Task *task)
1729 {
1730 ASSERT(task->next == NULL && task->prev == NULL);
1731 task->next = cap->suspended_ccalling_tasks;
1732 task->prev = NULL;
1733 if (cap->suspended_ccalling_tasks) {
1734 cap->suspended_ccalling_tasks->prev = task;
1735 }
1736 cap->suspended_ccalling_tasks = task;
1737 }
1738
1739 STATIC_INLINE void
1740 recoverSuspendedTask (Capability *cap, Task *task)
1741 {
1742 if (task->prev) {
1743 task->prev->next = task->next;
1744 } else {
1745 ASSERT(cap->suspended_ccalling_tasks == task);
1746 cap->suspended_ccalling_tasks = task->next;
1747 }
1748 if (task->next) {
1749 task->next->prev = task->prev;
1750 }
1751 task->next = task->prev = NULL;
1752 }
1753
1754 /* ---------------------------------------------------------------------------
1755 * Suspending & resuming Haskell threads.
1756 *
1757 * When making a "safe" call to C (aka _ccall_GC), the task gives back
1758 * its capability before calling the C function. This allows another
1759 * task to pick up the capability and carry on running Haskell
1760 * threads. It also means that if the C call blocks, it won't lock
1761 * the whole system.
1762 *
1763 * The Haskell thread making the C call is put to sleep for the
1764 * duration of the call, on the susepended_ccalling_threads queue. We
1765 * give out a token to the task, which it can use to resume the thread
1766 * on return from the C function.
1767 * ------------------------------------------------------------------------- */
1768
1769 void *
1770 suspendThread (StgRegTable *reg)
1771 {
1772 Capability *cap;
1773 int saved_errno;
1774 StgTSO *tso;
1775 Task *task;
1776 #if mingw32_HOST_OS
1777 StgWord32 saved_winerror;
1778 #endif
1779
1780 saved_errno = errno;
1781 #if mingw32_HOST_OS
1782 saved_winerror = GetLastError();
1783 #endif
1784
1785 /* assume that *reg is a pointer to the StgRegTable part of a Capability.
1786 */
1787 cap = regTableToCapability(reg);
1788
1789 task = cap->running_task;
1790 tso = cap->r.rCurrentTSO;
1791
1792 traceSchedEvent(cap, EVENT_STOP_THREAD, tso, THREAD_SUSPENDED_FOREIGN_CALL);
1793
1794 // XXX this might not be necessary --SDM
1795 tso->what_next = ThreadRunGHC;
1796
1797 threadPaused(cap,tso);
1798
1799 if ((tso->flags & TSO_BLOCKEX) == 0) {
1800 tso->why_blocked = BlockedOnCCall;
1801 tso->flags |= TSO_BLOCKEX;
1802 tso->flags &= ~TSO_INTERRUPTIBLE;
1803 } else {
1804 tso->why_blocked = BlockedOnCCall_NoUnblockExc;
1805 }
1806
1807 // Hand back capability
1808 task->suspended_tso = tso;
1809
1810 ACQUIRE_LOCK(&cap->lock);
1811
1812 suspendTask(cap,task);
1813 cap->in_haskell = rtsFalse;
1814 releaseCapability_(cap,rtsFalse);
1815
1816 RELEASE_LOCK(&cap->lock);
1817
1818 errno = saved_errno;
1819 #if mingw32_HOST_OS
1820 SetLastError(saved_winerror);
1821 #endif
1822 return task;
1823 }
1824
1825 StgRegTable *
1826 resumeThread (void *task_)
1827 {
1828 StgTSO *tso;
1829 Capability *cap;
1830 Task *task = task_;
1831 int saved_errno;
1832 #if mingw32_HOST_OS
1833 StgWord32 saved_winerror;
1834 #endif
1835
1836 saved_errno = errno;
1837 #if mingw32_HOST_OS
1838 saved_winerror = GetLastError();
1839 #endif
1840
1841 cap = task->cap;
1842 // Wait for permission to re-enter the RTS with the result.
1843 waitForReturnCapability(&cap,task);
1844 // we might be on a different capability now... but if so, our
1845 // entry on the suspended_ccalling_tasks list will also have been
1846 // migrated.
1847
1848 // Remove the thread from the suspended list
1849 recoverSuspendedTask(cap,task);
1850
1851 tso = task->suspended_tso;
1852 task->suspended_tso = NULL;
1853 tso->_link = END_TSO_QUEUE; // no write barrier reqd
1854
1855 traceSchedEvent(cap, EVENT_RUN_THREAD, tso, tso->what_next);
1856
1857 if (tso->why_blocked == BlockedOnCCall) {
1858 // avoid locking the TSO if we don't have to
1859 if (tso->blocked_exceptions != END_TSO_QUEUE) {
1860 awakenBlockedExceptionQueue(cap,tso);
1861 }
1862 tso->flags &= ~(TSO_BLOCKEX | TSO_INTERRUPTIBLE);
1863 }
1864
1865 /* Reset blocking status */
1866 tso->why_blocked = NotBlocked;
1867
1868 cap->r.rCurrentTSO = tso;
1869 cap->in_haskell = rtsTrue;
1870 errno = saved_errno;
1871 #if mingw32_HOST_OS
1872 SetLastError(saved_winerror);
1873 #endif
1874
1875 /* We might have GC'd, mark the TSO dirty again */
1876 dirty_TSO(cap,tso);
1877
1878 IF_DEBUG(sanity, checkTSO(tso));
1879
1880 return &cap->r;
1881 }
1882
1883 /* ---------------------------------------------------------------------------
1884 * scheduleThread()
1885 *
1886 * scheduleThread puts a thread on the end of the runnable queue.
1887 * This will usually be done immediately after a thread is created.
1888 * The caller of scheduleThread must create the thread using e.g.
1889 * createThread and push an appropriate closure
1890 * on this thread's stack before the scheduler is invoked.
1891 * ------------------------------------------------------------------------ */
1892
1893 void
1894 scheduleThread(Capability *cap, StgTSO *tso)
1895 {
1896 // The thread goes at the *end* of the run-queue, to avoid possible
1897 // starvation of any threads already on the queue.
1898 appendToRunQueue(cap,tso);
1899 }
1900
1901 void
1902 scheduleThreadOn(Capability *cap, StgWord cpu USED_IF_THREADS, StgTSO *tso)
1903 {
1904 #if defined(THREADED_RTS)
1905 tso->flags |= TSO_LOCKED; // we requested explicit affinity; don't
1906 // move this thread from now on.
1907 cpu %= RtsFlags.ParFlags.nNodes;
1908 if (cpu == cap->no) {
1909 appendToRunQueue(cap,tso);
1910 } else {
1911 traceSchedEvent (cap, EVENT_MIGRATE_THREAD, tso, capabilities[cpu].no);
1912 wakeupThreadOnCapability(cap, &capabilities[cpu], tso);
1913 }
1914 #else
1915 appendToRunQueue(cap,tso);
1916 #endif
1917 }
1918
1919 Capability *
1920 scheduleWaitThread (StgTSO* tso, /*[out]*/HaskellObj* ret, Capability *cap)
1921 {
1922 Task *task;
1923
1924 // We already created/initialised the Task
1925 task = cap->running_task;
1926
1927 // This TSO is now a bound thread; make the Task and TSO
1928 // point to each other.
1929 tso->bound = task;
1930 tso->cap = cap;
1931
1932 task->tso = tso;
1933 task->ret = ret;
1934 task->stat = NoStatus;
1935
1936 appendToRunQueue(cap,tso);
1937
1938 debugTrace(DEBUG_sched, "new bound thread (%lu)", (unsigned long)tso->id);
1939
1940 cap = schedule(cap,task);
1941
1942 ASSERT(task->stat != NoStatus);
1943 ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
1944
1945 debugTrace(DEBUG_sched, "bound thread (%lu) finished", (unsigned long)task->tso->id);
1946 return cap;
1947 }
1948
1949 /* ----------------------------------------------------------------------------
1950 * Starting Tasks
1951 * ------------------------------------------------------------------------- */
1952
1953 #if defined(THREADED_RTS)
1954 void OSThreadProcAttr
1955 workerStart(Task *task)
1956 {
1957 Capability *cap;
1958
1959 // See startWorkerTask().
1960 ACQUIRE_LOCK(&task->lock);
1961 cap = task->cap;
1962 RELEASE_LOCK(&task->lock);
1963
1964 if (RtsFlags.ParFlags.setAffinity) {
1965 setThreadAffinity(cap->no, n_capabilities);
1966 }
1967
1968 // set the thread-local pointer to the Task:
1969 taskEnter(task);
1970
1971 // schedule() runs without a lock.
1972 cap = schedule(cap,task);
1973
1974 // On exit from schedule(), we have a Capability, but possibly not
1975 // the same one we started with.
1976
1977 // During shutdown, the requirement is that after all the
1978 // Capabilities are shut down, all workers that are shutting down
1979 // have finished workerTaskStop(). This is why we hold on to
1980 // cap->lock until we've finished workerTaskStop() below.
1981 //
1982 // There may be workers still involved in foreign calls; those
1983 // will just block in waitForReturnCapability() because the
1984 // Capability has been shut down.
1985 //
1986 ACQUIRE_LOCK(&cap->lock);
1987 releaseCapability_(cap,rtsFalse);
1988 workerTaskStop(task);
1989 RELEASE_LOCK(&cap->lock);
1990 }
1991 #endif
1992
1993 /* ---------------------------------------------------------------------------
1994 * initScheduler()
1995 *
1996 * Initialise the scheduler. This resets all the queues - if the
1997 * queues contained any threads, they'll be garbage collected at the
1998 * next pass.
1999 *
2000 * ------------------------------------------------------------------------ */
2001
2002 void
2003 initScheduler(void)
2004 {
2005 #if !defined(THREADED_RTS)
2006 blocked_queue_hd = END_TSO_QUEUE;
2007 blocked_queue_tl = END_TSO_QUEUE;
2008 sleeping_queue = END_TSO_QUEUE;
2009 #endif
2010
2011 blackhole_queue = END_TSO_QUEUE;
2012
2013 sched_state = SCHED_RUNNING;
2014 recent_activity = ACTIVITY_YES;
2015
2016 #if defined(THREADED_RTS)
2017 /* Initialise the mutex and condition variables used by
2018 * the scheduler. */
2019 initMutex(&sched_mutex);
2020 #endif
2021
2022 ACQUIRE_LOCK(&sched_mutex);
2023
2024 /* A capability holds the state a native thread needs in
2025 * order to execute STG code. At least one capability is
2026 * floating around (only THREADED_RTS builds have more than one).
2027 */
2028 initCapabilities();
2029
2030 initTaskManager();
2031
2032 #if defined(THREADED_RTS)
2033 initSparkPools();
2034 #endif
2035
2036 #if defined(THREADED_RTS)
2037 /*
2038 * Eagerly start one worker to run each Capability, except for
2039 * Capability 0. The idea is that we're probably going to start a
2040 * bound thread on Capability 0 pretty soon, so we don't want a
2041 * worker task hogging it.
2042 */
2043 {
2044 nat i;
2045 Capability *cap;
2046 for (i = 1; i < n_capabilities; i++) {
2047 cap = &capabilities[i];
2048 ACQUIRE_LOCK(&cap->lock);
2049 startWorkerTask(cap, workerStart);
2050 RELEASE_LOCK(&cap->lock);
2051 }
2052 }
2053 #endif
2054
2055 RELEASE_LOCK(&sched_mutex);
2056 }
2057
2058 void
2059 exitScheduler(
2060 rtsBool wait_foreign
2061 #if !defined(THREADED_RTS)
2062 __attribute__((unused))
2063 #endif
2064 )
2065 /* see Capability.c, shutdownCapability() */
2066 {
2067 Task *task = NULL;
2068
2069 task = newBoundTask();
2070
2071 // If we haven't killed all the threads yet, do it now.
2072 if (sched_state < SCHED_SHUTTING_DOWN) {
2073 sched_state = SCHED_INTERRUPTING;
2074 waitForReturnCapability(&task->cap,task);
2075 scheduleDoGC(task->cap,task,rtsFalse);
2076 releaseCapability(task->cap);
2077 }
2078 sched_state = SCHED_SHUTTING_DOWN;
2079
2080 #if defined(THREADED_RTS)
2081 {
2082 nat i;
2083
2084 for (i = 0; i < n_capabilities; i++) {
2085 shutdownCapability(&capabilities[i], task, wait_foreign);
2086 }
2087 boundTaskExiting(task);
2088 }
2089 #endif
2090 }
2091
2092 void
2093 freeScheduler( void )
2094 {
2095 nat still_running;
2096
2097 ACQUIRE_LOCK(&sched_mutex);
2098 still_running = freeTaskManager();
2099 // We can only free the Capabilities if there are no Tasks still
2100 // running. We might have a Task about to return from a foreign
2101 // call into waitForReturnCapability(), for example (actually,
2102 // this should be the *only* thing that a still-running Task can
2103 // do at this point, and it will block waiting for the
2104 // Capability).
2105 if (still_running == 0) {
2106 freeCapabilities();
2107 if (n_capabilities != 1) {
2108 stgFree(capabilities);
2109 }
2110 }
2111 RELEASE_LOCK(&sched_mutex);
2112 #if defined(THREADED_RTS)
2113 closeMutex(&sched_mutex);
2114 #endif
2115 }
2116
2117 /* -----------------------------------------------------------------------------
2118 performGC
2119
2120 This is the interface to the garbage collector from Haskell land.
2121 We provide this so that external C code can allocate and garbage
2122 collect when called from Haskell via _ccall_GC.
2123 -------------------------------------------------------------------------- */
2124
2125 static void
2126 performGC_(rtsBool force_major)
2127 {
2128 Task *task;
2129
2130 // We must grab a new Task here, because the existing Task may be
2131 // associated with a particular Capability, and chained onto the
2132 // suspended_ccalling_tasks queue.
2133 task = newBoundTask();
2134
2135 waitForReturnCapability(&task->cap,task);
2136 scheduleDoGC(task->cap,task,force_major);
2137 releaseCapability(task->cap);
2138 boundTaskExiting(task);
2139 }
2140
2141 void
2142 performGC(void)
2143 {
2144 performGC_(rtsFalse);
2145 }
2146
2147 void
2148 performMajorGC(void)
2149 {
2150 performGC_(rtsTrue);
2151 }
2152
2153 /* -----------------------------------------------------------------------------
2154 Stack overflow
2155
2156 If the thread has reached its maximum stack size, then raise the
2157 StackOverflow exception in the offending thread. Otherwise
2158 relocate the TSO into a larger chunk of memory and adjust its stack
2159 size appropriately.
2160 -------------------------------------------------------------------------- */
2161
2162 static StgTSO *
2163 threadStackOverflow(Capability *cap, StgTSO *tso)
2164 {
2165 nat new_stack_size, stack_words;
2166 lnat new_tso_size;
2167 StgPtr new_sp;
2168 StgTSO *dest;
2169
2170 IF_DEBUG(sanity,checkTSO(tso));
2171
2172 // don't allow throwTo() to modify the blocked_exceptions queue
2173 // while we are moving the TSO:
2174 lockClosure((StgClosure *)tso);
2175
2176 if (tso->stack_size >= tso->max_stack_size && !(tso->flags & TSO_BLOCKEX)) {
2177 // NB. never raise a StackOverflow exception if the thread is
2178 // inside Control.Exceptino.block. It is impractical to protect
2179 // against stack overflow exceptions, since virtually anything
2180 // can raise one (even 'catch'), so this is the only sensible
2181 // thing to do here. See bug #767.
2182
2183 debugTrace(DEBUG_gc,
2184 "threadStackOverflow of TSO %ld (%p): stack too large (now %ld; max is %ld)",
2185 (long)tso->id, tso, (long)tso->stack_size, (long)tso->max_stack_size);
2186 IF_DEBUG(gc,
2187 /* If we're debugging, just print out the top of the stack */
2188 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2189 tso->sp+64)));
2190
2191 // Send this thread the StackOverflow exception
2192 unlockTSO(tso);
2193 throwToSingleThreaded(cap, tso, (StgClosure *)stackOverflow_closure);
2194 return tso;
2195 }
2196
2197 /* Try to double the current stack size. If that takes us over the
2198 * maximum stack size for this thread, then use the maximum instead
2199 * (that is, unless we're already at or over the max size and we
2200 * can't raise the StackOverflow exception (see above), in which
2201 * case just double the size). Finally round up so the TSO ends up as
2202 * a whole number of blocks.
2203 */
2204 if (tso->stack_size >= tso->max_stack_size) {
2205 new_stack_size = tso->stack_size * 2;
2206 } else {
2207 new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
2208 }
2209 new_tso_size = (lnat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) +
2210 TSO_STRUCT_SIZE)/sizeof(W_);
2211 new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */
2212 new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
2213
2214 debugTrace(DEBUG_sched,
2215 "increasing stack size from %ld words to %d.",
2216 (long)tso->stack_size, new_stack_size);
2217
2218 dest = (StgTSO *)allocateLocal(cap,new_tso_size);
2219 TICK_ALLOC_TSO(new_stack_size,0);
2220
2221 /* copy the TSO block and the old stack into the new area */
2222 memcpy(dest,tso,TSO_STRUCT_SIZE);
2223 stack_words = tso->stack + tso->stack_size - tso->sp;
2224 new_sp = (P_)dest + new_tso_size - stack_words;
2225 memcpy(new_sp, tso->sp, stack_words * sizeof(W_));
2226
2227 /* relocate the stack pointers... */
2228 dest->sp = new_sp;
2229 dest->stack_size = new_stack_size;
2230
2231 /* Mark the old TSO as relocated. We have to check for relocated
2232 * TSOs in the garbage collector and any primops that deal with TSOs.
2233 *
2234 * It's important to set the sp value to just beyond the end
2235 * of the stack, so we don't attempt to scavenge any part of the
2236 * dead TSO's stack.
2237 */
2238 tso->what_next = ThreadRelocated;
2239 setTSOLink(cap,tso,dest);
2240 tso->sp = (P_)&(tso->stack[tso->stack_size]);
2241 tso->why_blocked = NotBlocked;
2242
2243 unlockTSO(dest);
2244 unlockTSO(tso);
2245
2246 IF_DEBUG(sanity,checkTSO(dest));
2247 #if 0
2248 IF_DEBUG(scheduler,printTSO(dest));
2249 #endif
2250
2251 return dest;
2252 }
2253
2254 static StgTSO *
2255 threadStackUnderflow (Task *task STG_UNUSED, StgTSO *tso)
2256 {
2257 bdescr *bd, *new_bd;
2258 lnat free_w, tso_size_w;
2259 StgTSO *new_tso;
2260
2261 tso_size_w = tso_sizeW(tso);
2262
2263 if (tso_size_w < MBLOCK_SIZE_W ||
2264 // TSO is less than 2 mblocks (since the first mblock is
2265 // shorter than MBLOCK_SIZE_W)
2266 (tso_size_w - BLOCKS_PER_MBLOCK*BLOCK_SIZE_W) % MBLOCK_SIZE_W != 0 ||
2267 // or TSO is not a whole number of megablocks (ensuring
2268 // precondition of splitLargeBlock() below)
2269 (tso_size_w <= round_up_to_mblocks(RtsFlags.GcFlags.initialStkSize)) ||
2270 // or TSO is smaller than the minimum stack size (rounded up)
2271 (nat)(tso->stack + tso->stack_size - tso->sp) > tso->stack_size / 4)
2272 // or stack is using more than 1/4 of the available space
2273 {
2274 // then do nothing
2275 return tso;
2276 }
2277
2278 // don't allow throwTo() to modify the blocked_exceptions queue
2279 // while we are moving the TSO:
2280 lockClosure((StgClosure *)tso);
2281
2282 // this is the number of words we'll free
2283 free_w = round_to_mblocks(tso_size_w/2);
2284
2285 bd = Bdescr((StgPtr)tso);
2286 new_bd = splitLargeBlock(bd, free_w / BLOCK_SIZE_W);
2287 bd->free = bd->start + TSO_STRUCT_SIZEW;
2288
2289 new_tso = (StgTSO *)new_bd->start;
2290 memcpy(new_tso,tso,TSO_STRUCT_SIZE);
2291 new_tso->stack_size = new_bd->free - new_tso->stack;
2292
2293 debugTrace(DEBUG_sched, "thread %ld: reducing TSO size from %lu words to %lu",
2294 (long)tso->id, tso_size_w, tso_sizeW(new_tso));
2295
2296 tso->what_next = ThreadRelocated;
2297 tso->_link = new_tso; // no write barrier reqd: same generation
2298
2299 // The TSO attached to this Task may have moved, so update the
2300 // pointer to it.
2301 if (task->tso == tso) {
2302 task->tso = new_tso;
2303 }
2304
2305 unlockTSO(new_tso);
2306 unlockTSO(tso);
2307
2308 IF_DEBUG(sanity,checkTSO(new_tso));
2309
2310 return new_tso;
2311 }
2312
2313 /* ---------------------------------------------------------------------------
2314 Interrupt execution
2315 - usually called inside a signal handler so it mustn't do anything fancy.
2316 ------------------------------------------------------------------------ */
2317
2318 void
2319 interruptStgRts(void)
2320 {
2321 sched_state = SCHED_INTERRUPTING;
2322 setContextSwitches();
2323 #if defined(THREADED_RTS)
2324 wakeUpRts();
2325 #endif
2326 }
2327
2328 /* -----------------------------------------------------------------------------
2329 Wake up the RTS
2330
2331 This function causes at least one OS thread to wake up and run the
2332 scheduler loop. It is invoked when the RTS might be deadlocked, or
2333 an external event has arrived that may need servicing (eg. a
2334 keyboard interrupt).
2335
2336 In the single-threaded RTS we don't do anything here; we only have
2337 one thread anyway, and the event that caused us to want to wake up
2338 will have interrupted any blocking system call in progress anyway.
2339 -------------------------------------------------------------------------- */
2340
2341 #if defined(THREADED_RTS)
2342 void wakeUpRts(void)
2343 {
2344 // This forces the IO Manager thread to wakeup, which will
2345 // in turn ensure that some OS thread wakes up and runs the
2346 // scheduler loop, which will cause a GC and deadlock check.
2347 ioManagerWakeup();
2348 }
2349 #endif
2350
2351 /* -----------------------------------------------------------------------------
2352 * checkBlackHoles()
2353 *
2354 * Check the blackhole_queue for threads that can be woken up. We do
2355 * this periodically: before every GC, and whenever the run queue is
2356 * empty.
2357 *
2358 * An elegant solution might be to just wake up all the blocked
2359 * threads with awakenBlockedQueue occasionally: they'll go back to
2360 * sleep again if the object is still a BLACKHOLE. Unfortunately this
2361 * doesn't give us a way to tell whether we've actually managed to
2362 * wake up any threads, so we would be busy-waiting.
2363 *
2364 * -------------------------------------------------------------------------- */
2365
2366 static rtsBool
2367 checkBlackHoles (Capability *cap)
2368 {
2369 StgTSO **prev, *t;
2370 rtsBool any_woke_up = rtsFalse;
2371 StgHalfWord type;
2372
2373 // blackhole_queue is global:
2374 ASSERT_LOCK_HELD(&sched_mutex);
2375
2376 debugTrace(DEBUG_sched, "checking threads blocked on black holes");
2377
2378 // ASSUMES: sched_mutex
2379 prev = &blackhole_queue;
2380 t = blackhole_queue;
2381 while (t != END_TSO_QUEUE) {
2382 if (t->what_next == ThreadRelocated) {
2383 t = t->_link;
2384 continue;
2385 }
2386 ASSERT(t->why_blocked == BlockedOnBlackHole);
2387 type = get_itbl(UNTAG_CLOSURE(t->block_info.closure))->type;
2388 if (type != BLACKHOLE && type != CAF_BLACKHOLE) {
2389 IF_DEBUG(sanity,checkTSO(t));
2390 t = unblockOne(cap, t);
2391 *prev = t;
2392 any_woke_up = rtsTrue;
2393 } else {
2394 prev = &t->_link;
2395 t = t->_link;
2396 }
2397 }
2398
2399 return any_woke_up;
2400 }
2401
2402 /* -----------------------------------------------------------------------------
2403 Deleting threads
2404
2405 This is used for interruption (^C) and forking, and corresponds to
2406 raising an exception but without letting the thread catch the
2407 exception.
2408 -------------------------------------------------------------------------- */
2409
2410 static void
2411 deleteThread (Capability *cap, StgTSO *tso)
2412 {
2413 // NOTE: must only be called on a TSO that we have exclusive
2414 // access to, because we will call throwToSingleThreaded() below.
2415 // The TSO must be on the run queue of the Capability we own, or
2416 // we must own all Capabilities.
2417
2418 if (tso->why_blocked != BlockedOnCCall &&
2419 tso->why_blocked != BlockedOnCCall_NoUnblockExc) {
2420 throwToSingleThreaded(cap,tso,NULL);
2421 }
2422 }
2423
2424 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
2425 static void
2426 deleteThread_(Capability *cap, StgTSO *tso)
2427 { // for forkProcess only:
2428 // like deleteThread(), but we delete threads in foreign calls, too.
2429
2430 if (tso->why_blocked == BlockedOnCCall ||
2431 tso->why_blocked == BlockedOnCCall_NoUnblockExc) {
2432 unblockOne(cap,tso);
2433 tso->what_next = ThreadKilled;
2434 } else {
2435 deleteThread(cap,tso);
2436 }
2437 }
2438 #endif
2439
2440 /* -----------------------------------------------------------------------------
2441 raiseExceptionHelper
2442
2443 This function is called by the raise# primitve, just so that we can
2444 move some of the tricky bits of raising an exception from C-- into
2445 C. Who knows, it might be a useful re-useable thing here too.
2446 -------------------------------------------------------------------------- */
2447
2448 StgWord
2449 raiseExceptionHelper (StgRegTable *reg, StgTSO *tso, StgClosure *exception)
2450 {
2451 Capability *cap = regTableToCapability(reg);
2452 StgThunk *raise_closure = NULL;
2453 StgPtr p, next;
2454 StgRetInfoTable *info;
2455 //
2456 // This closure represents the expression 'raise# E' where E
2457 // is the exception raise. It is used to overwrite all the
2458 // thunks which are currently under evaluataion.
2459 //
2460
2461 // OLD COMMENT (we don't have MIN_UPD_SIZE now):
2462 // LDV profiling: stg_raise_info has THUNK as its closure
2463 // type. Since a THUNK takes at least MIN_UPD_SIZE words in its
2464 // payload, MIN_UPD_SIZE is more approprate than 1. It seems that
2465 // 1 does not cause any problem unless profiling is performed.
2466 // However, when LDV profiling goes on, we need to linearly scan
2467 // small object pool, where raise_closure is stored, so we should
2468 // use MIN_UPD_SIZE.
2469 //
2470 // raise_closure = (StgClosure *)RET_STGCALL1(P_,allocate,
2471 // sizeofW(StgClosure)+1);
2472 //
2473
2474 //
2475 // Walk up the stack, looking for the catch frame. On the way,
2476 // we update any closures pointed to from update frames with the
2477 // raise closure that we just built.
2478 //
2479 p = tso->sp;
2480 while(1) {
2481 info = get_ret_itbl((StgClosure *)p);
2482 next = p + stack_frame_sizeW((StgClosure *)p);
2483 switch (info->i.type) {
2484
2485 case UPDATE_FRAME:
2486 // Only create raise_closure if we need to.
2487 if (raise_closure == NULL) {
2488 raise_closure =
2489 (StgThunk *)allocateLocal(cap,sizeofW(StgThunk)+1);
2490 SET_HDR(raise_closure, &stg_raise_info, CCCS);
2491 raise_closure->payload[0] = exception;
2492 }
2493 UPD_IND(((StgUpdateFrame *)p)->updatee,(StgClosure *)raise_closure);
2494 p = next;
2495 continue;
2496
2497 case ATOMICALLY_FRAME:
2498 debugTrace(DEBUG_stm, "found ATOMICALLY_FRAME at %p", p);
2499 tso->sp = p;
2500 return ATOMICALLY_FRAME;
2501
2502 case CATCH_FRAME:
2503 tso->sp = p;
2504 return CATCH_FRAME;
2505
2506 case CATCH_STM_FRAME:
2507 debugTrace(DEBUG_stm, "found CATCH_STM_FRAME at %p", p);
2508 tso->sp = p;
2509 return CATCH_STM_FRAME;
2510
2511 case STOP_FRAME:
2512 tso->sp = p;
2513 return STOP_FRAME;
2514
2515 case CATCH_RETRY_FRAME:
2516 default:
2517 p = next;
2518 continue;
2519 }
2520 }
2521 }
2522
2523
2524 /* -----------------------------------------------------------------------------
2525 findRetryFrameHelper
2526
2527 This function is called by the retry# primitive. It traverses the stack
2528 leaving tso->sp referring to the frame which should handle the retry.
2529
2530 This should either be a CATCH_RETRY_FRAME (if the retry# is within an orElse#)
2531 or should be a ATOMICALLY_FRAME (if the retry# reaches the top level).
2532
2533 We skip CATCH_STM_FRAMEs (aborting and rolling back the nested tx that they
2534 create) because retries are not considered to be exceptions, despite the
2535 similar implementation.
2536
2537 We should not expect to see CATCH_FRAME or STOP_FRAME because those should
2538 not be created within memory transactions.
2539 -------------------------------------------------------------------------- */
2540
2541 StgWord
2542 findRetryFrameHelper (StgTSO *tso)
2543 {
2544 StgPtr p, next;
2545 StgRetInfoTable *info;
2546
2547 p = tso -> sp;
2548 while (1) {
2549 info = get_ret_itbl((StgClosure *)p);
2550 next = p + stack_frame_sizeW((StgClosure *)p);
2551 switch (info->i.type) {
2552
2553 case ATOMICALLY_FRAME:
2554 debugTrace(DEBUG_stm,
2555 "found ATOMICALLY_FRAME at %p during retry", p);
2556 tso->sp = p;
2557 return ATOMICALLY_FRAME;
2558
2559 case CATCH_RETRY_FRAME:
2560 debugTrace(DEBUG_stm,
2561 "found CATCH_RETRY_FRAME at %p during retrry", p);
2562 tso->sp = p;
2563 return CATCH_RETRY_FRAME;
2564
2565 case CATCH_STM_FRAME: {
2566 StgTRecHeader *trec = tso -> trec;
2567 StgTRecHeader *outer = stmGetEnclosingTRec(trec);
2568 debugTrace(DEBUG_stm,
2569 "found CATCH_STM_FRAME at %p during retry", p);
2570 debugTrace(DEBUG_stm, "trec=%p outer=%p", trec, outer);
2571 stmAbortTransaction(tso -> cap, trec);
2572 stmFreeAbortedTRec(tso -> cap, trec);
2573 tso -> trec = outer;
2574 p = next;
2575 continue;
2576 }
2577
2578
2579 default:
2580 ASSERT(info->i.type != CATCH_FRAME);
2581 ASSERT(info->i.type != STOP_FRAME);
2582 p = next;
2583 continue;
2584 }
2585 }
2586 }
2587
2588 /* -----------------------------------------------------------------------------
2589 resurrectThreads is called after garbage collection on the list of
2590 threads found to be garbage. Each of these threads will be woken
2591 up and sent a signal: BlockedOnDeadMVar if the thread was blocked
2592 on an MVar, or NonTermination if the thread was blocked on a Black
2593 Hole.
2594
2595 Locks: assumes we hold *all* the capabilities.
2596 -------------------------------------------------------------------------- */
2597
2598 void
2599 resurrectThreads (StgTSO *threads)
2600 {
2601 StgTSO *tso, *next;
2602 Capability *cap;
2603 step *step;
2604
2605 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
2606 next = tso->global_link;
2607
2608 step = Bdescr((P_)tso)->step;
2609 tso->global_link = step->threads;
2610 step->threads = tso;
2611
2612 debugTrace(DEBUG_sched, "resurrecting thread %lu", (unsigned long)tso->id);
2613
2614 // Wake up the thread on the Capability it was last on
2615 cap = tso->cap;
2616
2617 switch (tso->why_blocked) {
2618 case BlockedOnMVar:
2619 /* Called by GC - sched_mutex lock is currently held. */
2620 throwToSingleThreaded(cap, tso,
2621 (StgClosure *)blockedIndefinitelyOnMVar_closure);
2622 break;
2623 case BlockedOnBlackHole:
2624 throwToSingleThreaded(cap, tso,
2625 (StgClosure *)nonTermination_closure);
2626 break;
2627 case BlockedOnSTM:
2628 throwToSingleThreaded(cap, tso,
2629 (StgClosure *)blockedIndefinitelyOnSTM_closure);
2630 break;
2631 case NotBlocked:
2632 /* This might happen if the thread was blocked on a black hole
2633 * belonging to a thread that we've just woken up (raiseAsync
2634 * can wake up threads, remember...).
2635 */
2636 continue;
2637 case BlockedOnException:
2638 // throwTo should never block indefinitely: if the target
2639 // thread dies or completes, throwTo returns.
2640 barf("resurrectThreads: thread BlockedOnException");
2641 break;
2642 default:
2643 barf("resurrectThreads: thread blocked in a strange way");
2644 }
2645 }
2646 }
2647
2648 /* -----------------------------------------------------------------------------
2649 performPendingThrowTos is called after garbage collection, and
2650 passed a list of threads that were found to have pending throwTos
2651 (tso->blocked_exceptions was not empty), and were blocked.
2652 Normally this doesn't happen, because we would deliver the
2653 exception directly if the target thread is blocked, but there are
2654 small windows where it might occur on a multiprocessor (see
2655 throwTo()).
2656
2657 NB. we must be holding all the capabilities at this point, just
2658 like resurrectThreads().
2659 -------------------------------------------------------------------------- */
2660
2661 void
2662 performPendingThrowTos (StgTSO *threads)
2663 {
2664 StgTSO *tso, *next;
2665 Capability *cap;
2666 Task *task, *saved_task;;
2667 step *step;
2668
2669 task = myTask();
2670 cap = task->cap;
2671
2672 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
2673 next = tso->global_link;
2674
2675 step = Bdescr((P_)tso)->step;
2676 tso->global_link = step->threads;
2677 step->threads = tso;
2678
2679 debugTrace(DEBUG_sched, "performing blocked throwTo to thread %lu", (unsigned long)tso->id);
2680
2681 // We must pretend this Capability belongs to the current Task
2682 // for the time being, as invariants will be broken otherwise.
2683 // In fact the current Task has exclusive access to the systme
2684 // at this point, so this is just bookkeeping:
2685 task->cap = tso->cap;
2686 saved_task = tso->cap->running_task;
2687 tso->cap->running_task = task;
2688 maybePerformBlockedException(tso->cap, tso);
2689 tso->cap->running_task = saved_task;
2690 }
2691
2692 // Restore our original Capability:
2693 task->cap = cap;
2694 }