1 /* ---------------------------------------------------------------------------
3 * (c) The GHC Team, 1998-2005
5 * The scheduler and thread-related functionality
7 * --------------------------------------------------------------------------*/
9 #include "PosixSource.h"
14 #include "BlockAlloc.h"
15 #include "OSThreads.h"
20 #include "StgMiscClosures.h"
21 #include "Interpreter.h"
22 #include "Exception.h"
24 #include "RtsSignals.h"
30 #include "ThreadLabels.h"
31 #include "LdvProfile.h"
34 #include "Proftimer.h"
37 #if defined(GRAN) || defined(PARALLEL_HASKELL)
38 # include "GranSimRts.h"
40 # include "ParallelRts.h"
41 # include "Parallel.h"
42 # include "ParallelDebug.h"
47 #include "Capability.h"
49 #include "AwaitEvent.h"
51 #ifdef HAVE_SYS_TYPES_H
52 #include <sys/types.h>
66 // Turn off inlining when debugging - it obfuscates things
69 # define STATIC_INLINE static
73 #define USED_WHEN_THREADED_RTS
74 #define USED_WHEN_NON_THREADED_RTS STG_UNUSED
76 #define USED_WHEN_THREADED_RTS STG_UNUSED
77 #define USED_WHEN_NON_THREADED_RTS
83 #define USED_WHEN_SMP STG_UNUSED
86 /* -----------------------------------------------------------------------------
88 * -------------------------------------------------------------------------- */
92 StgTSO* ActiveTSO = NULL; /* for assigning system costs; GranSim-Light only */
93 /* rtsTime TimeOfNextEvent, EndOfTimeSlice; now in GranSim.c */
96 In GranSim we have a runnable and a blocked queue for each processor.
97 In order to minimise code changes new arrays run_queue_hds/tls
98 are created. run_queue_hd is then a short cut (macro) for
99 run_queue_hds[CurrentProc] (see GranSim.h).
102 StgTSO *run_queue_hds[MAX_PROC], *run_queue_tls[MAX_PROC];
103 StgTSO *blocked_queue_hds[MAX_PROC], *blocked_queue_tls[MAX_PROC];
104 StgTSO *ccalling_threadss[MAX_PROC];
105 /* We use the same global list of threads (all_threads) in GranSim as in
106 the std RTS (i.e. we are cheating). However, we don't use this list in
107 the GranSim specific code at the moment (so we are only potentially
112 #if !defined(THREADED_RTS)
113 // Blocked/sleeping thrads
114 StgTSO *blocked_queue_hd = NULL;
115 StgTSO *blocked_queue_tl = NULL;
116 StgTSO *sleeping_queue = NULL; // perhaps replace with a hash table?
119 /* Threads blocked on blackholes.
120 * LOCK: sched_mutex+capability, or all capabilities
122 StgTSO *blackhole_queue = NULL;
125 /* The blackhole_queue should be checked for threads to wake up. See
126 * Schedule.h for more thorough comment.
127 * LOCK: none (doesn't matter if we miss an update)
129 rtsBool blackholes_need_checking = rtsFalse;
131 /* Linked list of all threads.
132 * Used for detecting garbage collected threads.
133 * LOCK: sched_mutex+capability, or all capabilities
135 StgTSO *all_threads = NULL;
137 /* flag set by signal handler to precipitate a context switch
138 * LOCK: none (just an advisory flag)
140 int context_switch = 0;
142 /* flag that tracks whether we have done any execution in this time slice.
143 * LOCK: currently none, perhaps we should lock (but needs to be
144 * updated in the fast path of the scheduler).
146 nat recent_activity = ACTIVITY_YES;
148 /* if this flag is set as well, give up execution
149 * LOCK: none (changes once, from false->true)
151 rtsBool interrupted = rtsFalse;
153 /* Next thread ID to allocate.
156 static StgThreadID next_thread_id = 1;
158 /* The smallest stack size that makes any sense is:
159 * RESERVED_STACK_WORDS (so we can get back from the stack overflow)
160 * + sizeofW(StgStopFrame) (the stg_stop_thread_info frame)
161 * + 1 (the closure to enter)
163 * + 1 (spare slot req'd by stg_ap_v_ret)
165 * A thread with this stack will bomb immediately with a stack
166 * overflow, which will increase its stack size.
168 #define MIN_STACK_WORDS (RESERVED_STACK_WORDS + sizeofW(StgStopFrame) + 3)
174 /* This is used in `TSO.h' and gcc 2.96 insists that this variable actually
175 * exists - earlier gccs apparently didn't.
181 * Set to TRUE when entering a shutdown state (via shutdownHaskellAndExit()) --
182 * in an MT setting, needed to signal that a worker thread shouldn't hang around
183 * in the scheduler when it is out of work.
185 rtsBool shutting_down_scheduler = rtsFalse;
188 * This mutex protects most of the global scheduler data in
189 * the THREADED_RTS and (inc. SMP) runtime.
191 #if defined(THREADED_RTS)
192 Mutex sched_mutex = INIT_MUTEX_VAR;
195 #if defined(PARALLEL_HASKELL)
197 rtsTime TimeOfLastYield;
198 rtsBool emitSchedule = rtsTrue;
201 /* -----------------------------------------------------------------------------
202 * static function prototypes
203 * -------------------------------------------------------------------------- */
205 static Capability *schedule (Capability *initialCapability, Task *task);
208 // These function all encapsulate parts of the scheduler loop, and are
209 // abstracted only to make the structure and control flow of the
210 // scheduler clearer.
212 static void schedulePreLoop (void);
213 static void scheduleStartSignalHandlers (void);
214 static void scheduleCheckBlockedThreads (Capability *cap);
215 static void scheduleCheckBlackHoles (Capability *cap);
216 static void scheduleDetectDeadlock (Capability *cap, Task *task);
218 static StgTSO *scheduleProcessEvent(rtsEvent *event);
220 #if defined(PARALLEL_HASKELL)
221 static StgTSO *scheduleSendPendingMessages(void);
222 static void scheduleActivateSpark(void);
223 static rtsBool scheduleGetRemoteWork(rtsBool *receivedFinish);
225 #if defined(PAR) || defined(GRAN)
226 static void scheduleGranParReport(void);
228 static void schedulePostRunThread(void);
229 static rtsBool scheduleHandleHeapOverflow( Capability *cap, StgTSO *t );
230 static void scheduleHandleStackOverflow( Capability *cap, Task *task,
232 static rtsBool scheduleHandleYield( Capability *cap, StgTSO *t,
233 nat prev_what_next );
234 static void scheduleHandleThreadBlocked( StgTSO *t );
235 static rtsBool scheduleHandleThreadFinished( Capability *cap, Task *task,
237 static rtsBool scheduleDoHeapProfile(rtsBool ready_to_gc);
238 static void scheduleDoGC(Capability *cap, Task *task, rtsBool force_major);
240 static void unblockThread(Capability *cap, StgTSO *tso);
241 static rtsBool checkBlackHoles(Capability *cap);
242 static void AllRoots(evac_fn evac);
244 static StgTSO *threadStackOverflow(Capability *cap, StgTSO *tso);
246 static void raiseAsync_(Capability *cap, StgTSO *tso, StgClosure *exception,
247 rtsBool stop_at_atomically);
249 static void deleteThread (Capability *cap, StgTSO *tso);
250 static void deleteRunQueue (Capability *cap);
253 static void printThreadBlockage(StgTSO *tso);
254 static void printThreadStatus(StgTSO *tso);
255 void printThreadQueue(StgTSO *tso);
258 #if defined(PARALLEL_HASKELL)
259 StgTSO * createSparkThread(rtsSpark spark);
260 StgTSO * activateSpark (rtsSpark spark);
264 static char *whatNext_strs[] = {
274 /* -----------------------------------------------------------------------------
275 * Putting a thread on the run queue: different scheduling policies
276 * -------------------------------------------------------------------------- */
279 addToRunQueue( Capability *cap, StgTSO *t )
281 #if defined(PARALLEL_HASKELL)
282 if (RtsFlags.ParFlags.doFairScheduling) {
283 // this does round-robin scheduling; good for concurrency
284 appendToRunQueue(cap,t);
286 // this does unfair scheduling; good for parallelism
287 pushOnRunQueue(cap,t);
290 // this does round-robin scheduling; good for concurrency
291 appendToRunQueue(cap,t);
295 /* ---------------------------------------------------------------------------
296 Main scheduling loop.
298 We use round-robin scheduling, each thread returning to the
299 scheduler loop when one of these conditions is detected:
302 * timer expires (thread yields)
308 In a GranSim setup this loop iterates over the global event queue.
309 This revolves around the global event queue, which determines what
310 to do next. Therefore, it's more complicated than either the
311 concurrent or the parallel (GUM) setup.
314 GUM iterates over incoming messages.
315 It starts with nothing to do (thus CurrentTSO == END_TSO_QUEUE),
316 and sends out a fish whenever it has nothing to do; in-between
317 doing the actual reductions (shared code below) it processes the
318 incoming messages and deals with delayed operations
319 (see PendingFetches).
320 This is not the ugliest code you could imagine, but it's bloody close.
322 ------------------------------------------------------------------------ */
325 schedule (Capability *initialCapability, Task *task)
329 StgThreadReturnCode ret;
332 #elif defined(PARALLEL_HASKELL)
335 rtsBool receivedFinish = rtsFalse;
337 nat tp_size, sp_size; // stats only
342 rtsBool first = rtsTrue;
344 cap = initialCapability;
346 // Pre-condition: this task owns initialCapability.
347 // The sched_mutex is *NOT* held
348 // NB. on return, we still hold a capability.
351 sched_belch("### NEW SCHEDULER LOOP (task: %p, cap: %p)",
352 task, initialCapability);
357 // -----------------------------------------------------------
358 // Scheduler loop starts here:
360 #if defined(PARALLEL_HASKELL)
361 #define TERMINATION_CONDITION (!receivedFinish)
363 #define TERMINATION_CONDITION ((event = get_next_event()) != (rtsEvent*)NULL)
365 #define TERMINATION_CONDITION rtsTrue
368 while (TERMINATION_CONDITION) {
371 /* Choose the processor with the next event */
372 CurrentProc = event->proc;
373 CurrentTSO = event->tso;
376 #if defined(THREADED_RTS)
378 // don't yield the first time, we want a chance to run this
379 // thread for a bit, even if there are others banging at the
383 // Yield the capability to higher-priority tasks if necessary.
384 yieldCapability(&cap, task);
388 ASSERT(cap->running_task == task);
389 ASSERT(task->cap == cap);
390 ASSERT(myTask() == task);
392 // Check whether we have re-entered the RTS from Haskell without
393 // going via suspendThread()/resumeThread (i.e. a 'safe' foreign
395 if (cap->in_haskell) {
396 errorBelch("schedule: re-entered unsafely.\n"
397 " Perhaps a 'foreign import unsafe' should be 'safe'?");
398 stg_exit(EXIT_FAILURE);
402 // Test for interruption. If interrupted==rtsTrue, then either
403 // we received a keyboard interrupt (^C), or the scheduler is
404 // trying to shut down all the tasks (shutting_down_scheduler) in
409 if (shutting_down_scheduler) {
410 IF_DEBUG(scheduler, sched_belch("shutting down"));
412 IF_DEBUG(scheduler, sched_belch("interrupted"));
416 #if defined(not_yet) && defined(SMP)
418 // Top up the run queue from our spark pool. We try to make the
419 // number of threads in the run queue equal to the number of
420 // free capabilities.
424 if (emptyRunQueue()) {
425 spark = findSpark(rtsFalse);
427 break; /* no more sparks in the pool */
429 createSparkThread(spark);
431 sched_belch("==^^ turning spark of closure %p into a thread",
432 (StgClosure *)spark));
438 scheduleStartSignalHandlers();
440 // Only check the black holes here if we've nothing else to do.
441 // During normal execution, the black hole list only gets checked
442 // at GC time, to avoid repeatedly traversing this possibly long
443 // list each time around the scheduler.
444 if (emptyRunQueue(cap)) { scheduleCheckBlackHoles(cap); }
446 scheduleCheckBlockedThreads(cap);
448 scheduleDetectDeadlock(cap,task);
450 // Normally, the only way we can get here with no threads to
451 // run is if a keyboard interrupt received during
452 // scheduleCheckBlockedThreads() or scheduleDetectDeadlock().
453 // Additionally, it is not fatal for the
454 // threaded RTS to reach here with no threads to run.
456 // win32: might be here due to awaitEvent() being abandoned
457 // as a result of a console event having been delivered.
458 if ( emptyRunQueue(cap) ) {
459 #if !defined(THREADED_RTS) && !defined(mingw32_HOST_OS)
462 continue; // nothing to do
465 #if defined(PARALLEL_HASKELL)
466 scheduleSendPendingMessages();
467 if (emptyRunQueue(cap) && scheduleActivateSpark())
471 ASSERT(next_fish_to_send_at==0); // i.e. no delayed fishes left!
474 /* If we still have no work we need to send a FISH to get a spark
476 if (emptyRunQueue(cap)) {
477 if (!scheduleGetRemoteWork(&receivedFinish)) continue;
478 ASSERT(rtsFalse); // should not happen at the moment
480 // from here: non-empty run queue.
481 // TODO: merge above case with this, only one call processMessages() !
482 if (PacketsWaiting()) { /* process incoming messages, if
483 any pending... only in else
484 because getRemoteWork waits for
486 receivedFinish = processMessages();
491 scheduleProcessEvent(event);
495 // Get a thread to run
497 t = popRunQueue(cap);
499 #if defined(GRAN) || defined(PAR)
500 scheduleGranParReport(); // some kind of debuging output
502 // Sanity check the thread we're about to run. This can be
503 // expensive if there is lots of thread switching going on...
504 IF_DEBUG(sanity,checkTSO(t));
507 #if defined(THREADED_RTS)
508 // Check whether we can run this thread in the current task.
509 // If not, we have to pass our capability to the right task.
511 Task *bound = t->bound;
516 sched_belch("### Running thread %d in bound thread",
518 // yes, the Haskell thread is bound to the current native thread
521 sched_belch("### thread %d bound to another OS thread",
523 // no, bound to a different Haskell thread: pass to that thread
524 pushOnRunQueue(cap,t);
528 // The thread we want to run is unbound.
531 sched_belch("### this OS thread cannot run thread %d", t->id));
532 // no, the current native thread is bound to a different
533 // Haskell thread, so pass it to any worker thread
534 pushOnRunQueue(cap,t);
541 cap->r.rCurrentTSO = t;
543 /* context switches are initiated by the timer signal, unless
544 * the user specified "context switch as often as possible", with
547 if (RtsFlags.ConcFlags.ctxtSwitchTicks == 0
548 && !emptyThreadQueues(cap)) {
554 IF_DEBUG(scheduler, sched_belch("-->> running thread %ld %s ...",
555 (long)t->id, whatNext_strs[t->what_next]));
557 #if defined(PROFILING)
558 startHeapProfTimer();
561 // ----------------------------------------------------------------------
562 // Run the current thread
564 prev_what_next = t->what_next;
566 errno = t->saved_errno;
567 cap->in_haskell = rtsTrue;
569 recent_activity = ACTIVITY_YES;
571 switch (prev_what_next) {
575 /* Thread already finished, return to scheduler. */
576 ret = ThreadFinished;
582 r = StgRun((StgFunPtr) stg_returnToStackTop, &cap->r);
583 cap = regTableToCapability(r);
588 case ThreadInterpret:
589 cap = interpretBCO(cap);
594 barf("schedule: invalid what_next field");
597 cap->in_haskell = rtsFalse;
600 // If ret is ThreadBlocked, and this Task is bound to the TSO that
601 // blocked, we are in limbo - the TSO is now owned by whatever it
602 // is blocked on, and may in fact already have been woken up,
603 // perhaps even on a different Capability. It may be the case
604 // that task->cap != cap. We better yield this Capability
605 // immediately and return to normaility.
606 if (ret == ThreadBlocked) continue;
609 ASSERT(cap->running_task == task);
610 ASSERT(task->cap == cap);
611 ASSERT(myTask() == task);
613 // The TSO might have moved, eg. if it re-entered the RTS and a GC
614 // happened. So find the new location:
615 t = cap->r.rCurrentTSO;
617 // And save the current errno in this thread.
618 t->saved_errno = errno;
620 // ----------------------------------------------------------------------
622 // Costs for the scheduler are assigned to CCS_SYSTEM
623 #if defined(PROFILING)
628 // We have run some Haskell code: there might be blackhole-blocked
629 // threads to wake up now.
630 // Lock-free test here should be ok, we're just setting a flag.
631 if ( blackhole_queue != END_TSO_QUEUE ) {
632 blackholes_need_checking = rtsTrue;
635 #if defined(THREADED_RTS)
636 IF_DEBUG(scheduler,debugBelch("sched (task %p): ", (void *)(unsigned long)(unsigned int)osThreadId()););
637 #elif !defined(GRAN) && !defined(PARALLEL_HASKELL)
638 IF_DEBUG(scheduler,debugBelch("sched: "););
641 schedulePostRunThread();
643 ready_to_gc = rtsFalse;
647 ready_to_gc = scheduleHandleHeapOverflow(cap,t);
651 scheduleHandleStackOverflow(cap,task,t);
655 if (scheduleHandleYield(cap, t, prev_what_next)) {
656 // shortcut for switching between compiler/interpreter:
662 scheduleHandleThreadBlocked(t);
666 if (scheduleHandleThreadFinished(cap, task, t)) return cap;
670 barf("schedule: invalid thread return code %d", (int)ret);
673 if (scheduleDoHeapProfile(ready_to_gc)) { ready_to_gc = rtsFalse; }
674 if (ready_to_gc) { scheduleDoGC(cap,task,rtsFalse); }
675 } /* end of while() */
677 IF_PAR_DEBUG(verbose,
678 debugBelch("== Leaving schedule() after having received Finish\n"));
681 /* ----------------------------------------------------------------------------
682 * Setting up the scheduler loop
683 * ------------------------------------------------------------------------- */
686 schedulePreLoop(void)
689 /* set up first event to get things going */
690 /* ToDo: assign costs for system setup and init MainTSO ! */
691 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
693 CurrentTSO, (StgClosure*)NULL, (rtsSpark*)NULL);
696 debugBelch("GRAN: Init CurrentTSO (in schedule) = %p\n",
698 G_TSO(CurrentTSO, 5));
700 if (RtsFlags.GranFlags.Light) {
701 /* Save current time; GranSim Light only */
702 CurrentTSO->gran.clock = CurrentTime[CurrentProc];
707 /* ----------------------------------------------------------------------------
708 * Start any pending signal handlers
709 * ------------------------------------------------------------------------- */
712 scheduleStartSignalHandlers(void)
714 #if defined(RTS_USER_SIGNALS) && !defined(THREADED_RTS)
715 if (signals_pending()) { // safe outside the lock
716 startSignalHandlers();
721 /* ----------------------------------------------------------------------------
722 * Check for blocked threads that can be woken up.
723 * ------------------------------------------------------------------------- */
726 scheduleCheckBlockedThreads(Capability *cap USED_WHEN_NON_THREADED_RTS)
728 #if !defined(THREADED_RTS)
730 // Check whether any waiting threads need to be woken up. If the
731 // run queue is empty, and there are no other tasks running, we
732 // can wait indefinitely for something to happen.
734 if ( !emptyQueue(blocked_queue_hd) || !emptyQueue(sleeping_queue) )
736 awaitEvent( emptyRunQueue(cap) && !blackholes_need_checking );
742 /* ----------------------------------------------------------------------------
743 * Check for threads blocked on BLACKHOLEs that can be woken up
744 * ------------------------------------------------------------------------- */
746 scheduleCheckBlackHoles (Capability *cap)
748 if ( blackholes_need_checking ) // check without the lock first
750 ACQUIRE_LOCK(&sched_mutex);
751 if ( blackholes_need_checking ) {
752 checkBlackHoles(cap);
753 blackholes_need_checking = rtsFalse;
755 RELEASE_LOCK(&sched_mutex);
759 /* ----------------------------------------------------------------------------
760 * Detect deadlock conditions and attempt to resolve them.
761 * ------------------------------------------------------------------------- */
764 scheduleDetectDeadlock (Capability *cap, Task *task)
767 #if defined(PARALLEL_HASKELL)
768 // ToDo: add deadlock detection in GUM (similar to SMP) -- HWL
773 * Detect deadlock: when we have no threads to run, there are no
774 * threads blocked, waiting for I/O, or sleeping, and all the
775 * other tasks are waiting for work, we must have a deadlock of
778 if ( emptyThreadQueues(cap) )
780 #if defined(THREADED_RTS)
782 * In the threaded RTS, we only check for deadlock if there
783 * has been no activity in a complete timeslice. This means
784 * we won't eagerly start a full GC just because we don't have
785 * any threads to run currently.
787 if (recent_activity != ACTIVITY_INACTIVE) return;
790 IF_DEBUG(scheduler, sched_belch("deadlocked, forcing major GC..."));
792 // Garbage collection can release some new threads due to
793 // either (a) finalizers or (b) threads resurrected because
794 // they are unreachable and will therefore be sent an
795 // exception. Any threads thus released will be immediately
797 scheduleDoGC( cap, task, rtsTrue/*force major GC*/ );
798 recent_activity = ACTIVITY_DONE_GC;
800 if ( !emptyRunQueue(cap) ) return;
802 #if defined(RTS_USER_SIGNALS) && !defined(THREADED_RTS)
803 /* If we have user-installed signal handlers, then wait
804 * for signals to arrive rather then bombing out with a
807 if ( anyUserHandlers() ) {
809 sched_belch("still deadlocked, waiting for signals..."));
813 if (signals_pending()) {
814 startSignalHandlers();
817 // either we have threads to run, or we were interrupted:
818 ASSERT(!emptyRunQueue(cap) || interrupted);
822 #if !defined(THREADED_RTS)
823 /* Probably a real deadlock. Send the current main thread the
824 * Deadlock exception.
827 switch (task->tso->why_blocked) {
829 case BlockedOnBlackHole:
830 case BlockedOnException:
832 raiseAsync(cap, task->tso, (StgClosure *)NonTermination_closure);
835 barf("deadlock: main thread blocked in a strange way");
843 /* ----------------------------------------------------------------------------
844 * Process an event (GRAN only)
845 * ------------------------------------------------------------------------- */
849 scheduleProcessEvent(rtsEvent *event)
853 if (RtsFlags.GranFlags.Light)
854 GranSimLight_enter_system(event, &ActiveTSO); // adjust ActiveTSO etc
856 /* adjust time based on time-stamp */
857 if (event->time > CurrentTime[CurrentProc] &&
858 event->evttype != ContinueThread)
859 CurrentTime[CurrentProc] = event->time;
861 /* Deal with the idle PEs (may issue FindWork or MoveSpark events) */
862 if (!RtsFlags.GranFlags.Light)
865 IF_DEBUG(gran, debugBelch("GRAN: switch by event-type\n"));
867 /* main event dispatcher in GranSim */
868 switch (event->evttype) {
869 /* Should just be continuing execution */
871 IF_DEBUG(gran, debugBelch("GRAN: doing ContinueThread\n"));
872 /* ToDo: check assertion
873 ASSERT(run_queue_hd != (StgTSO*)NULL &&
874 run_queue_hd != END_TSO_QUEUE);
876 /* Ignore ContinueThreads for fetching threads (if synchr comm) */
877 if (!RtsFlags.GranFlags.DoAsyncFetch &&
878 procStatus[CurrentProc]==Fetching) {
879 debugBelch("ghuH: Spurious ContinueThread while Fetching ignored; TSO %d (%p) [PE %d]\n",
880 CurrentTSO->id, CurrentTSO, CurrentProc);
883 /* Ignore ContinueThreads for completed threads */
884 if (CurrentTSO->what_next == ThreadComplete) {
885 debugBelch("ghuH: found a ContinueThread event for completed thread %d (%p) [PE %d] (ignoring ContinueThread)\n",
886 CurrentTSO->id, CurrentTSO, CurrentProc);
889 /* Ignore ContinueThreads for threads that are being migrated */
890 if (PROCS(CurrentTSO)==Nowhere) {
891 debugBelch("ghuH: trying to run the migrating TSO %d (%p) [PE %d] (ignoring ContinueThread)\n",
892 CurrentTSO->id, CurrentTSO, CurrentProc);
895 /* The thread should be at the beginning of the run queue */
896 if (CurrentTSO!=run_queue_hds[CurrentProc]) {
897 debugBelch("ghuH: TSO %d (%p) [PE %d] is not at the start of the run_queue when doing a ContinueThread\n",
898 CurrentTSO->id, CurrentTSO, CurrentProc);
899 break; // run the thread anyway
902 new_event(proc, proc, CurrentTime[proc],
904 (StgTSO*)NULL, (StgClosure*)NULL, (rtsSpark*)NULL);
906 */ /* Catches superfluous CONTINUEs -- should be unnecessary */
907 break; // now actually run the thread; DaH Qu'vam yImuHbej
910 do_the_fetchnode(event);
911 goto next_thread; /* handle next event in event queue */
914 do_the_globalblock(event);
915 goto next_thread; /* handle next event in event queue */
918 do_the_fetchreply(event);
919 goto next_thread; /* handle next event in event queue */
921 case UnblockThread: /* Move from the blocked queue to the tail of */
922 do_the_unblock(event);
923 goto next_thread; /* handle next event in event queue */
925 case ResumeThread: /* Move from the blocked queue to the tail of */
926 /* the runnable queue ( i.e. Qu' SImqa'lu') */
927 event->tso->gran.blocktime +=
928 CurrentTime[CurrentProc] - event->tso->gran.blockedat;
929 do_the_startthread(event);
930 goto next_thread; /* handle next event in event queue */
933 do_the_startthread(event);
934 goto next_thread; /* handle next event in event queue */
937 do_the_movethread(event);
938 goto next_thread; /* handle next event in event queue */
941 do_the_movespark(event);
942 goto next_thread; /* handle next event in event queue */
945 do_the_findwork(event);
946 goto next_thread; /* handle next event in event queue */
949 barf("Illegal event type %u\n", event->evttype);
952 /* This point was scheduler_loop in the old RTS */
954 IF_DEBUG(gran, debugBelch("GRAN: after main switch\n"));
956 TimeOfLastEvent = CurrentTime[CurrentProc];
957 TimeOfNextEvent = get_time_of_next_event();
958 IgnoreEvents=(TimeOfNextEvent==0); // HWL HACK
959 // CurrentTSO = ThreadQueueHd;
961 IF_DEBUG(gran, debugBelch("GRAN: time of next event is: %ld\n",
964 if (RtsFlags.GranFlags.Light)
965 GranSimLight_leave_system(event, &ActiveTSO);
967 EndOfTimeSlice = CurrentTime[CurrentProc]+RtsFlags.GranFlags.time_slice;
970 debugBelch("GRAN: end of time-slice is %#lx\n", EndOfTimeSlice));
972 /* in a GranSim setup the TSO stays on the run queue */
974 /* Take a thread from the run queue. */
975 POP_RUN_QUEUE(t); // take_off_run_queue(t);
978 debugBelch("GRAN: About to run current thread, which is\n");
981 context_switch = 0; // turned on via GranYield, checking events and time slice
984 DumpGranEvent(GR_SCHEDULE, t));
986 procStatus[CurrentProc] = Busy;
990 /* ----------------------------------------------------------------------------
991 * Send pending messages (PARALLEL_HASKELL only)
992 * ------------------------------------------------------------------------- */
994 #if defined(PARALLEL_HASKELL)
996 scheduleSendPendingMessages(void)
1002 # if defined(PAR) // global Mem.Mgmt., omit for now
1003 if (PendingFetches != END_BF_QUEUE) {
1008 if (RtsFlags.ParFlags.BufferTime) {
1009 // if we use message buffering, we must send away all message
1010 // packets which have become too old...
1016 /* ----------------------------------------------------------------------------
1017 * Activate spark threads (PARALLEL_HASKELL only)
1018 * ------------------------------------------------------------------------- */
1020 #if defined(PARALLEL_HASKELL)
1022 scheduleActivateSpark(void)
1025 ASSERT(emptyRunQueue());
1026 /* We get here if the run queue is empty and want some work.
1027 We try to turn a spark into a thread, and add it to the run queue,
1028 from where it will be picked up in the next iteration of the scheduler
1032 /* :-[ no local threads => look out for local sparks */
1033 /* the spark pool for the current PE */
1034 pool = &(cap.r.rSparks); // JB: cap = (old) MainCap
1035 if (advisory_thread_count < RtsFlags.ParFlags.maxThreads &&
1036 pool->hd < pool->tl) {
1038 * ToDo: add GC code check that we really have enough heap afterwards!!
1040 * If we're here (no runnable threads) and we have pending
1041 * sparks, we must have a space problem. Get enough space
1042 * to turn one of those pending sparks into a
1046 spark = findSpark(rtsFalse); /* get a spark */
1047 if (spark != (rtsSpark) NULL) {
1048 tso = createThreadFromSpark(spark); /* turn the spark into a thread */
1049 IF_PAR_DEBUG(fish, // schedule,
1050 debugBelch("==== schedule: Created TSO %d (%p); %d threads active\n",
1051 tso->id, tso, advisory_thread_count));
1053 if (tso==END_TSO_QUEUE) { /* failed to activate spark->back to loop */
1054 IF_PAR_DEBUG(fish, // schedule,
1055 debugBelch("==^^ failed to create thread from spark @ %lx\n",
1057 return rtsFalse; /* failed to generate a thread */
1058 } /* otherwise fall through & pick-up new tso */
1060 IF_PAR_DEBUG(fish, // schedule,
1061 debugBelch("==^^ no local sparks (spark pool contains only NFs: %d)\n",
1062 spark_queue_len(pool)));
1063 return rtsFalse; /* failed to generate a thread */
1065 return rtsTrue; /* success in generating a thread */
1066 } else { /* no more threads permitted or pool empty */
1067 return rtsFalse; /* failed to generateThread */
1070 tso = NULL; // avoid compiler warning only
1071 return rtsFalse; /* dummy in non-PAR setup */
1074 #endif // PARALLEL_HASKELL
1076 /* ----------------------------------------------------------------------------
1077 * Get work from a remote node (PARALLEL_HASKELL only)
1078 * ------------------------------------------------------------------------- */
1080 #if defined(PARALLEL_HASKELL)
1082 scheduleGetRemoteWork(rtsBool *receivedFinish)
1084 ASSERT(emptyRunQueue());
1086 if (RtsFlags.ParFlags.BufferTime) {
1087 IF_PAR_DEBUG(verbose,
1088 debugBelch("...send all pending data,"));
1091 for (i=1; i<=nPEs; i++)
1092 sendImmediately(i); // send all messages away immediately
1096 //++EDEN++ idle() , i.e. send all buffers, wait for work
1097 // suppress fishing in EDEN... just look for incoming messages
1098 // (blocking receive)
1099 IF_PAR_DEBUG(verbose,
1100 debugBelch("...wait for incoming messages...\n"));
1101 *receivedFinish = processMessages(); // blocking receive...
1103 // and reenter scheduling loop after having received something
1104 // (return rtsFalse below)
1106 # else /* activate SPARKS machinery */
1107 /* We get here, if we have no work, tried to activate a local spark, but still
1108 have no work. We try to get a remote spark, by sending a FISH message.
1109 Thread migration should be added here, and triggered when a sequence of
1110 fishes returns without work. */
1111 delay = (RtsFlags.ParFlags.fishDelay!=0ll ? RtsFlags.ParFlags.fishDelay : 0ll);
1113 /* =8-[ no local sparks => look for work on other PEs */
1115 * We really have absolutely no work. Send out a fish
1116 * (there may be some out there already), and wait for
1117 * something to arrive. We clearly can't run any threads
1118 * until a SCHEDULE or RESUME arrives, and so that's what
1119 * we're hoping to see. (Of course, we still have to
1120 * respond to other types of messages.)
1122 rtsTime now = msTime() /*CURRENT_TIME*/;
1123 IF_PAR_DEBUG(verbose,
1124 debugBelch("-- now=%ld\n", now));
1125 IF_PAR_DEBUG(fish, // verbose,
1126 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
1127 (last_fish_arrived_at!=0 &&
1128 last_fish_arrived_at+delay > now)) {
1129 debugBelch("--$$ <%llu> delaying FISH until %llu (last fish %llu, delay %llu)\n",
1130 now, last_fish_arrived_at+delay,
1131 last_fish_arrived_at,
1135 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
1136 advisory_thread_count < RtsFlags.ParFlags.maxThreads) { // send a FISH, but when?
1137 if (last_fish_arrived_at==0 ||
1138 (last_fish_arrived_at+delay <= now)) { // send FISH now!
1139 /* outstandingFishes is set in sendFish, processFish;
1140 avoid flooding system with fishes via delay */
1141 next_fish_to_send_at = 0;
1143 /* ToDo: this should be done in the main scheduling loop to avoid the
1144 busy wait here; not so bad if fish delay is very small */
1145 int iq = 0; // DEBUGGING -- HWL
1146 next_fish_to_send_at = last_fish_arrived_at+delay; // remember when to send
1147 /* send a fish when ready, but process messages that arrive in the meantime */
1149 if (PacketsWaiting()) {
1151 *receivedFinish = processMessages();
1154 } while (!*receivedFinish || now<next_fish_to_send_at);
1155 // JB: This means the fish could become obsolete, if we receive
1156 // work. Better check for work again?
1157 // last line: while (!receivedFinish || !haveWork || now<...)
1158 // next line: if (receivedFinish || haveWork )
1160 if (*receivedFinish) // no need to send a FISH if we are finishing anyway
1161 return rtsFalse; // NB: this will leave scheduler loop
1162 // immediately after return!
1164 IF_PAR_DEBUG(fish, // verbose,
1165 debugBelch("--$$ <%llu> sent delayed fish (%d processMessages); active/total threads=%d/%d\n",now,iq,run_queue_len(),advisory_thread_count));
1169 // JB: IMHO, this should all be hidden inside sendFish(...)
1171 sendFish(pe, thisPE, NEW_FISH_AGE, NEW_FISH_HISTORY,
1174 // Global statistics: count no. of fishes
1175 if (RtsFlags.ParFlags.ParStats.Global &&
1176 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
1177 globalParStats.tot_fish_mess++;
1181 /* delayed fishes must have been sent by now! */
1182 next_fish_to_send_at = 0;
1185 *receivedFinish = processMessages();
1186 # endif /* SPARKS */
1189 /* NB: this function always returns rtsFalse, meaning the scheduler
1190 loop continues with the next iteration;
1192 return code means success in finding work; we enter this function
1193 if there is no local work, thus have to send a fish which takes
1194 time until it arrives with work; in the meantime we should process
1195 messages in the main loop;
1198 #endif // PARALLEL_HASKELL
1200 /* ----------------------------------------------------------------------------
1201 * PAR/GRAN: Report stats & debugging info(?)
1202 * ------------------------------------------------------------------------- */
1204 #if defined(PAR) || defined(GRAN)
1206 scheduleGranParReport(void)
1208 ASSERT(run_queue_hd != END_TSO_QUEUE);
1210 /* Take a thread from the run queue, if we have work */
1211 POP_RUN_QUEUE(t); // take_off_run_queue(END_TSO_QUEUE);
1213 /* If this TSO has got its outport closed in the meantime,
1214 * it mustn't be run. Instead, we have to clean it up as if it was finished.
1215 * It has to be marked as TH_DEAD for this purpose.
1216 * If it is TH_TERM instead, it is supposed to have finished in the normal way.
1218 JB: TODO: investigate wether state change field could be nuked
1219 entirely and replaced by the normal tso state (whatnext
1220 field). All we want to do is to kill tsos from outside.
1223 /* ToDo: write something to the log-file
1224 if (RTSflags.ParFlags.granSimStats && !sameThread)
1225 DumpGranEvent(GR_SCHEDULE, RunnableThreadsHd);
1229 /* the spark pool for the current PE */
1230 pool = &(cap.r.rSparks); // cap = (old) MainCap
1233 debugBelch("--=^ %d threads, %d sparks on [%#x]\n",
1234 run_queue_len(), spark_queue_len(pool), CURRENT_PROC));
1237 debugBelch("--=^ %d threads, %d sparks on [%#x]\n",
1238 run_queue_len(), spark_queue_len(pool), CURRENT_PROC));
1240 if (RtsFlags.ParFlags.ParStats.Full &&
1241 (t->par.sparkname != (StgInt)0) && // only log spark generated threads
1242 (emitSchedule || // forced emit
1243 (t && LastTSO && t->id != LastTSO->id))) {
1245 we are running a different TSO, so write a schedule event to log file
1246 NB: If we use fair scheduling we also have to write a deschedule
1247 event for LastTSO; with unfair scheduling we know that the
1248 previous tso has blocked whenever we switch to another tso, so
1249 we don't need it in GUM for now
1251 IF_PAR_DEBUG(fish, // schedule,
1252 debugBelch("____ scheduling spark generated thread %d (%lx) (%lx) via a forced emit\n",t->id,t,t->par.sparkname));
1254 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
1255 GR_SCHEDULE, t, (StgClosure *)NULL, 0, 0);
1256 emitSchedule = rtsFalse;
1261 /* ----------------------------------------------------------------------------
1262 * After running a thread...
1263 * ------------------------------------------------------------------------- */
1266 schedulePostRunThread(void)
1269 /* HACK 675: if the last thread didn't yield, make sure to print a
1270 SCHEDULE event to the log file when StgRunning the next thread, even
1271 if it is the same one as before */
1273 TimeOfLastYield = CURRENT_TIME;
1276 /* some statistics gathering in the parallel case */
1278 #if defined(GRAN) || defined(PAR) || defined(EDEN)
1282 IF_DEBUG(gran, DumpGranEvent(GR_DESCHEDULE, t));
1283 globalGranStats.tot_heapover++;
1285 globalParStats.tot_heapover++;
1292 DumpGranEvent(GR_DESCHEDULE, t));
1293 globalGranStats.tot_stackover++;
1296 // DumpGranEvent(GR_DESCHEDULE, t);
1297 globalParStats.tot_stackover++;
1301 case ThreadYielding:
1304 DumpGranEvent(GR_DESCHEDULE, t));
1305 globalGranStats.tot_yields++;
1308 // DumpGranEvent(GR_DESCHEDULE, t);
1309 globalParStats.tot_yields++;
1316 debugBelch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: ",
1317 t->id, t, whatNext_strs[t->what_next], t->block_info.closure,
1318 (t->block_info.closure==(StgClosure*)NULL ? 99 : where_is(t->block_info.closure)));
1319 if (t->block_info.closure!=(StgClosure*)NULL)
1320 print_bq(t->block_info.closure);
1323 // ??? needed; should emit block before
1325 DumpGranEvent(GR_DESCHEDULE, t));
1326 prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
1329 ASSERT(procStatus[CurrentProc]==Busy ||
1330 ((procStatus[CurrentProc]==Fetching) &&
1331 (t->block_info.closure!=(StgClosure*)NULL)));
1332 if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
1333 !(!RtsFlags.GranFlags.DoAsyncFetch &&
1334 procStatus[CurrentProc]==Fetching))
1335 procStatus[CurrentProc] = Idle;
1338 //++PAR++ blockThread() writes the event (change?)
1342 case ThreadFinished:
1346 barf("parGlobalStats: unknown return code");
1352 /* -----------------------------------------------------------------------------
1353 * Handle a thread that returned to the scheduler with ThreadHeepOverflow
1354 * -------------------------------------------------------------------------- */
1357 scheduleHandleHeapOverflow( Capability *cap, StgTSO *t )
1359 // did the task ask for a large block?
1360 if (cap->r.rHpAlloc > BLOCK_SIZE) {
1361 // if so, get one and push it on the front of the nursery.
1365 blocks = (lnat)BLOCK_ROUND_UP(cap->r.rHpAlloc) / BLOCK_SIZE;
1368 debugBelch("--<< thread %ld (%s) stopped: requesting a large block (size %ld)\n",
1369 (long)t->id, whatNext_strs[t->what_next], blocks));
1371 // don't do this if the nursery is (nearly) full, we'll GC first.
1372 if (cap->r.rCurrentNursery->link != NULL ||
1373 cap->r.rNursery->n_blocks == 1) { // paranoia to prevent infinite loop
1374 // if the nursery has only one block.
1377 bd = allocGroup( blocks );
1379 cap->r.rNursery->n_blocks += blocks;
1381 // link the new group into the list
1382 bd->link = cap->r.rCurrentNursery;
1383 bd->u.back = cap->r.rCurrentNursery->u.back;
1384 if (cap->r.rCurrentNursery->u.back != NULL) {
1385 cap->r.rCurrentNursery->u.back->link = bd;
1388 ASSERT(g0s0->blocks == cap->r.rCurrentNursery &&
1389 g0s0 == cap->r.rNursery);
1391 cap->r.rNursery->blocks = bd;
1393 cap->r.rCurrentNursery->u.back = bd;
1395 // initialise it as a nursery block. We initialise the
1396 // step, gen_no, and flags field of *every* sub-block in
1397 // this large block, because this is easier than making
1398 // sure that we always find the block head of a large
1399 // block whenever we call Bdescr() (eg. evacuate() and
1400 // isAlive() in the GC would both have to do this, at
1404 for (x = bd; x < bd + blocks; x++) {
1405 x->step = cap->r.rNursery;
1411 // This assert can be a killer if the app is doing lots
1412 // of large block allocations.
1413 IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
1415 // now update the nursery to point to the new block
1416 cap->r.rCurrentNursery = bd;
1418 // we might be unlucky and have another thread get on the
1419 // run queue before us and steal the large block, but in that
1420 // case the thread will just end up requesting another large
1422 pushOnRunQueue(cap,t);
1423 return rtsFalse; /* not actually GC'ing */
1428 debugBelch("--<< thread %ld (%s) stopped: HeapOverflow\n",
1429 (long)t->id, whatNext_strs[t->what_next]));
1431 ASSERT(!is_on_queue(t,CurrentProc));
1432 #elif defined(PARALLEL_HASKELL)
1433 /* Currently we emit a DESCHEDULE event before GC in GUM.
1434 ToDo: either add separate event to distinguish SYSTEM time from rest
1435 or just nuke this DESCHEDULE (and the following SCHEDULE) */
1436 if (0 && RtsFlags.ParFlags.ParStats.Full) {
1437 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
1438 GR_DESCHEDULE, t, (StgClosure *)NULL, 0, 0);
1439 emitSchedule = rtsTrue;
1443 pushOnRunQueue(cap,t);
1445 /* actual GC is done at the end of the while loop in schedule() */
1448 /* -----------------------------------------------------------------------------
1449 * Handle a thread that returned to the scheduler with ThreadStackOverflow
1450 * -------------------------------------------------------------------------- */
1453 scheduleHandleStackOverflow (Capability *cap, Task *task, StgTSO *t)
1455 IF_DEBUG(scheduler,debugBelch("--<< thread %ld (%s) stopped, StackOverflow\n",
1456 (long)t->id, whatNext_strs[t->what_next]));
1457 /* just adjust the stack for this thread, then pop it back
1461 /* enlarge the stack */
1462 StgTSO *new_t = threadStackOverflow(cap, t);
1464 /* This TSO has moved, so update any pointers to it from the
1465 * main thread stack. It better not be on any other queues...
1466 * (it shouldn't be).
1468 if (task->tso != NULL) {
1471 pushOnRunQueue(cap,new_t);
1475 /* -----------------------------------------------------------------------------
1476 * Handle a thread that returned to the scheduler with ThreadYielding
1477 * -------------------------------------------------------------------------- */
1480 scheduleHandleYield( Capability *cap, StgTSO *t, nat prev_what_next )
1482 // Reset the context switch flag. We don't do this just before
1483 // running the thread, because that would mean we would lose ticks
1484 // during GC, which can lead to unfair scheduling (a thread hogs
1485 // the CPU because the tick always arrives during GC). This way
1486 // penalises threads that do a lot of allocation, but that seems
1487 // better than the alternative.
1490 /* put the thread back on the run queue. Then, if we're ready to
1491 * GC, check whether this is the last task to stop. If so, wake
1492 * up the GC thread. getThread will block during a GC until the
1496 if (t->what_next != prev_what_next) {
1497 debugBelch("--<< thread %ld (%s) stopped to switch evaluators\n",
1498 (long)t->id, whatNext_strs[t->what_next]);
1500 debugBelch("--<< thread %ld (%s) stopped, yielding\n",
1501 (long)t->id, whatNext_strs[t->what_next]);
1506 //debugBelch("&& Doing sanity check on yielding TSO %ld.", t->id);
1508 ASSERT(t->link == END_TSO_QUEUE);
1510 // Shortcut if we're just switching evaluators: don't bother
1511 // doing stack squeezing (which can be expensive), just run the
1513 if (t->what_next != prev_what_next) {
1518 ASSERT(!is_on_queue(t,CurrentProc));
1521 //debugBelch("&& Doing sanity check on all ThreadQueues (and their TSOs).");
1522 checkThreadQsSanity(rtsTrue));
1526 addToRunQueue(cap,t);
1529 /* add a ContinueThread event to actually process the thread */
1530 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1532 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1534 debugBelch("GRAN: eventq and runnableq after adding yielded thread to queue again:\n");
1541 /* -----------------------------------------------------------------------------
1542 * Handle a thread that returned to the scheduler with ThreadBlocked
1543 * -------------------------------------------------------------------------- */
1546 scheduleHandleThreadBlocked( StgTSO *t
1547 #if !defined(GRAN) && !defined(DEBUG)
1554 debugBelch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: \n",
1555 t->id, t, whatNext_strs[t->what_next], t->block_info.closure, (t->block_info.closure==(StgClosure*)NULL ? 99 : where_is(t->block_info.closure)));
1556 if (t->block_info.closure!=(StgClosure*)NULL) print_bq(t->block_info.closure));
1558 // ??? needed; should emit block before
1560 DumpGranEvent(GR_DESCHEDULE, t));
1561 prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
1564 ASSERT(procStatus[CurrentProc]==Busy ||
1565 ((procStatus[CurrentProc]==Fetching) &&
1566 (t->block_info.closure!=(StgClosure*)NULL)));
1567 if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
1568 !(!RtsFlags.GranFlags.DoAsyncFetch &&
1569 procStatus[CurrentProc]==Fetching))
1570 procStatus[CurrentProc] = Idle;
1574 debugBelch("--<< thread %ld (%p; %s) stopped, blocking on node %p with BQ: \n",
1575 t->id, t, whatNext_strs[t->what_next], t->block_info.closure));
1578 if (t->block_info.closure!=(StgClosure*)NULL)
1579 print_bq(t->block_info.closure));
1581 /* Send a fetch (if BlockedOnGA) and dump event to log file */
1584 /* whatever we schedule next, we must log that schedule */
1585 emitSchedule = rtsTrue;
1589 // We don't need to do anything. The thread is blocked, and it
1590 // has tidied up its stack and placed itself on whatever queue
1591 // it needs to be on.
1594 ASSERT(t->why_blocked != NotBlocked);
1595 // This might not be true under SMP: we don't have
1596 // exclusive access to this TSO, so someone might have
1597 // woken it up by now. This actually happens: try
1598 // conc023 +RTS -N2.
1602 debugBelch("--<< thread %d (%s) stopped: ",
1603 t->id, whatNext_strs[t->what_next]);
1604 printThreadBlockage(t);
1607 /* Only for dumping event to log file
1608 ToDo: do I need this in GranSim, too?
1614 /* -----------------------------------------------------------------------------
1615 * Handle a thread that returned to the scheduler with ThreadFinished
1616 * -------------------------------------------------------------------------- */
1619 scheduleHandleThreadFinished (Capability *cap STG_UNUSED, Task *task, StgTSO *t)
1621 /* Need to check whether this was a main thread, and if so,
1622 * return with the return value.
1624 * We also end up here if the thread kills itself with an
1625 * uncaught exception, see Exception.cmm.
1627 IF_DEBUG(scheduler,debugBelch("--++ thread %d (%s) finished\n",
1628 t->id, whatNext_strs[t->what_next]));
1631 endThread(t, CurrentProc); // clean-up the thread
1632 #elif defined(PARALLEL_HASKELL)
1633 /* For now all are advisory -- HWL */
1634 //if(t->priority==AdvisoryPriority) ??
1635 advisory_thread_count--; // JB: Caution with this counter, buggy!
1638 if(t->dist.priority==RevalPriority)
1642 # if defined(EDENOLD)
1643 // the thread could still have an outport... (BUG)
1644 if (t->eden.outport != -1) {
1645 // delete the outport for the tso which has finished...
1646 IF_PAR_DEBUG(eden_ports,
1647 debugBelch("WARNING: Scheduler removes outport %d for TSO %d.\n",
1648 t->eden.outport, t->id));
1651 // thread still in the process (HEAVY BUG! since outport has just been closed...)
1652 if (t->eden.epid != -1) {
1653 IF_PAR_DEBUG(eden_ports,
1654 debugBelch("WARNING: Scheduler removes TSO %d from process %d .\n",
1655 t->id, t->eden.epid));
1656 removeTSOfromProcess(t);
1661 if (RtsFlags.ParFlags.ParStats.Full &&
1662 !RtsFlags.ParFlags.ParStats.Suppressed)
1663 DumpEndEvent(CURRENT_PROC, t, rtsFalse /* not mandatory */);
1665 // t->par only contains statistics: left out for now...
1667 debugBelch("**** end thread: ended sparked thread %d (%lx); sparkname: %lx\n",
1668 t->id,t,t->par.sparkname));
1670 #endif // PARALLEL_HASKELL
1673 // Check whether the thread that just completed was a bound
1674 // thread, and if so return with the result.
1676 // There is an assumption here that all thread completion goes
1677 // through this point; we need to make sure that if a thread
1678 // ends up in the ThreadKilled state, that it stays on the run
1679 // queue so it can be dealt with here.
1684 if (t->bound != task) {
1685 #if !defined(THREADED_RTS)
1686 // Must be a bound thread that is not the topmost one. Leave
1687 // it on the run queue until the stack has unwound to the
1688 // point where we can deal with this. Leaving it on the run
1689 // queue also ensures that the garbage collector knows about
1690 // this thread and its return value (it gets dropped from the
1691 // all_threads list so there's no other way to find it).
1692 appendToRunQueue(cap,t);
1695 // this cannot happen in the threaded RTS, because a
1696 // bound thread can only be run by the appropriate Task.
1697 barf("finished bound thread that isn't mine");
1701 ASSERT(task->tso == t);
1703 if (t->what_next == ThreadComplete) {
1705 // NOTE: return val is tso->sp[1] (see StgStartup.hc)
1706 *(task->ret) = (StgClosure *)task->tso->sp[1];
1708 task->stat = Success;
1711 *(task->ret) = NULL;
1714 task->stat = Interrupted;
1716 task->stat = Killed;
1720 removeThreadLabel((StgWord)task->tso->id);
1722 return rtsTrue; // tells schedule() to return
1728 /* -----------------------------------------------------------------------------
1729 * Perform a heap census, if PROFILING
1730 * -------------------------------------------------------------------------- */
1733 scheduleDoHeapProfile( rtsBool ready_to_gc STG_UNUSED )
1735 #if defined(PROFILING)
1736 // When we have +RTS -i0 and we're heap profiling, do a census at
1737 // every GC. This lets us get repeatable runs for debugging.
1738 if (performHeapProfile ||
1739 (RtsFlags.ProfFlags.profileInterval==0 &&
1740 RtsFlags.ProfFlags.doHeapProfile && ready_to_gc)) {
1741 GarbageCollect(GetRoots, rtsTrue);
1743 performHeapProfile = rtsFalse;
1744 return rtsTrue; // true <=> we already GC'd
1750 /* -----------------------------------------------------------------------------
1751 * Perform a garbage collection if necessary
1752 * -------------------------------------------------------------------------- */
1755 scheduleDoGC( Capability *cap, Task *task USED_WHEN_SMP, rtsBool force_major )
1759 static volatile StgWord waiting_for_gc;
1760 rtsBool was_waiting;
1765 // In order to GC, there must be no threads running Haskell code.
1766 // Therefore, the GC thread needs to hold *all* the capabilities,
1767 // and release them after the GC has completed.
1769 // This seems to be the simplest way: previous attempts involved
1770 // making all the threads with capabilities give up their
1771 // capabilities and sleep except for the *last* one, which
1772 // actually did the GC. But it's quite hard to arrange for all
1773 // the other tasks to sleep and stay asleep.
1776 was_waiting = cas(&waiting_for_gc, 0, 1);
1777 if (was_waiting) return;
1779 for (i=0; i < n_capabilities; i++) {
1780 IF_DEBUG(scheduler, sched_belch("ready_to_gc, grabbing all the capabilies (%d/%d)", i, n_capabilities));
1781 if (cap != &capabilities[i]) {
1782 Capability *pcap = &capabilities[i];
1783 // we better hope this task doesn't get migrated to
1784 // another Capability while we're waiting for this one.
1785 // It won't, because load balancing happens while we have
1786 // all the Capabilities, but even so it's a slightly
1787 // unsavoury invariant.
1789 waitForReturnCapability(&pcap, task);
1790 if (pcap != &capabilities[i]) {
1791 barf("scheduleDoGC: got the wrong capability");
1796 waiting_for_gc = rtsFalse;
1799 /* Kick any transactions which are invalid back to their
1800 * atomically frames. When next scheduled they will try to
1801 * commit, this commit will fail and they will retry.
1806 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
1807 if (t->what_next == ThreadRelocated) {
1810 next = t->global_link;
1811 if (t -> trec != NO_TREC && t -> why_blocked == NotBlocked) {
1812 if (!stmValidateNestOfTransactions (t -> trec)) {
1813 IF_DEBUG(stm, sched_belch("trec %p found wasting its time", t));
1815 // strip the stack back to the
1816 // ATOMICALLY_FRAME, aborting the (nested)
1817 // transaction, and saving the stack of any
1818 // partially-evaluated thunks on the heap.
1819 raiseAsync_(cap, t, NULL, rtsTrue);
1822 ASSERT(get_itbl((StgClosure *)t->sp)->type == ATOMICALLY_FRAME);
1830 // so this happens periodically:
1831 scheduleCheckBlackHoles(cap);
1833 IF_DEBUG(scheduler, printAllThreads());
1835 /* everybody back, start the GC.
1836 * Could do it in this thread, or signal a condition var
1837 * to do it in another thread. Either way, we need to
1838 * broadcast on gc_pending_cond afterward.
1840 #if defined(THREADED_RTS)
1841 IF_DEBUG(scheduler,sched_belch("doing GC"));
1843 GarbageCollect(GetRoots, force_major);
1846 // release our stash of capabilities.
1847 for (i = 0; i < n_capabilities; i++) {
1848 if (cap != &capabilities[i]) {
1849 task->cap = &capabilities[i];
1850 releaseCapability(&capabilities[i]);
1857 /* add a ContinueThread event to continue execution of current thread */
1858 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1860 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1862 debugBelch("GRAN: eventq and runnableq after Garbage collection:\n\n");
1868 /* ---------------------------------------------------------------------------
1869 * rtsSupportsBoundThreads(): is the RTS built to support bound threads?
1870 * used by Control.Concurrent for error checking.
1871 * ------------------------------------------------------------------------- */
1874 rtsSupportsBoundThreads(void)
1876 #if defined(THREADED_RTS)
1883 /* ---------------------------------------------------------------------------
1884 * isThreadBound(tso): check whether tso is bound to an OS thread.
1885 * ------------------------------------------------------------------------- */
1888 isThreadBound(StgTSO* tso USED_WHEN_THREADED_RTS)
1890 #if defined(THREADED_RTS)
1891 return (tso->bound != NULL);
1896 /* ---------------------------------------------------------------------------
1897 * Singleton fork(). Do not copy any running threads.
1898 * ------------------------------------------------------------------------- */
1900 #if !defined(mingw32_HOST_OS) && !defined(SMP)
1901 #define FORKPROCESS_PRIMOP_SUPPORTED
1904 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
1906 deleteThreadImmediately(Capability *cap, StgTSO *tso);
1909 forkProcess(HsStablePtr *entry
1910 #ifndef FORKPROCESS_PRIMOP_SUPPORTED
1915 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
1921 IF_DEBUG(scheduler,sched_belch("forking!"));
1923 // ToDo: for SMP, we should probably acquire *all* the capabilities
1928 if (pid) { // parent
1930 // just return the pid
1935 // delete all threads
1936 cap->run_queue_hd = END_TSO_QUEUE;
1937 cap->run_queue_tl = END_TSO_QUEUE;
1939 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
1942 // don't allow threads to catch the ThreadKilled exception
1943 deleteThreadImmediately(cap,t);
1946 // wipe the main thread list
1947 while ((task = all_tasks) != NULL) {
1948 all_tasks = task->all_link;
1952 cap = rts_evalStableIO(cap, entry, NULL); // run the action
1953 rts_checkSchedStatus("forkProcess",cap);
1956 hs_exit(); // clean up and exit
1957 stg_exit(EXIT_SUCCESS);
1959 #else /* !FORKPROCESS_PRIMOP_SUPPORTED */
1960 barf("forkProcess#: primop not supported on this platform, sorry!\n");
1965 /* ---------------------------------------------------------------------------
1966 * Delete the threads on the run queue of the current capability.
1967 * ------------------------------------------------------------------------- */
1970 deleteRunQueue (Capability *cap)
1973 for (t = cap->run_queue_hd; t != END_TSO_QUEUE; t = next) {
1974 ASSERT(t->what_next != ThreadRelocated);
1976 deleteThread(cap, t);
1980 /* startThread and insertThread are now in GranSim.c -- HWL */
1983 /* -----------------------------------------------------------------------------
1984 Managing the suspended_ccalling_tasks list.
1985 Locks required: sched_mutex
1986 -------------------------------------------------------------------------- */
1989 suspendTask (Capability *cap, Task *task)
1991 ASSERT(task->next == NULL && task->prev == NULL);
1992 task->next = cap->suspended_ccalling_tasks;
1994 if (cap->suspended_ccalling_tasks) {
1995 cap->suspended_ccalling_tasks->prev = task;
1997 cap->suspended_ccalling_tasks = task;
2001 recoverSuspendedTask (Capability *cap, Task *task)
2004 task->prev->next = task->next;
2006 ASSERT(cap->suspended_ccalling_tasks == task);
2007 cap->suspended_ccalling_tasks = task->next;
2010 task->next->prev = task->prev;
2012 task->next = task->prev = NULL;
2015 /* ---------------------------------------------------------------------------
2016 * Suspending & resuming Haskell threads.
2018 * When making a "safe" call to C (aka _ccall_GC), the task gives back
2019 * its capability before calling the C function. This allows another
2020 * task to pick up the capability and carry on running Haskell
2021 * threads. It also means that if the C call blocks, it won't lock
2024 * The Haskell thread making the C call is put to sleep for the
2025 * duration of the call, on the susepended_ccalling_threads queue. We
2026 * give out a token to the task, which it can use to resume the thread
2027 * on return from the C function.
2028 * ------------------------------------------------------------------------- */
2031 suspendThread (StgRegTable *reg)
2034 int saved_errno = errno;
2038 /* assume that *reg is a pointer to the StgRegTable part of a Capability.
2040 cap = regTableToCapability(reg);
2042 task = cap->running_task;
2043 tso = cap->r.rCurrentTSO;
2046 sched_belch("thread %d did a safe foreign call", cap->r.rCurrentTSO->id));
2048 // XXX this might not be necessary --SDM
2049 tso->what_next = ThreadRunGHC;
2053 if(tso->blocked_exceptions == NULL) {
2054 tso->why_blocked = BlockedOnCCall;
2055 tso->blocked_exceptions = END_TSO_QUEUE;
2057 tso->why_blocked = BlockedOnCCall_NoUnblockExc;
2060 // Hand back capability
2061 task->suspended_tso = tso;
2063 ACQUIRE_LOCK(&cap->lock);
2065 suspendTask(cap,task);
2066 cap->in_haskell = rtsFalse;
2067 releaseCapability_(cap);
2069 RELEASE_LOCK(&cap->lock);
2071 #if defined(THREADED_RTS)
2072 /* Preparing to leave the RTS, so ensure there's a native thread/task
2073 waiting to take over.
2075 IF_DEBUG(scheduler, sched_belch("thread %d: leaving RTS", tso->id));
2078 errno = saved_errno;
2083 resumeThread (void *task_)
2087 int saved_errno = errno;
2091 // Wait for permission to re-enter the RTS with the result.
2092 waitForReturnCapability(&cap,task);
2093 // we might be on a different capability now... but if so, our
2094 // entry on the suspended_ccalling_tasks list will also have been
2097 // Remove the thread from the suspended list
2098 recoverSuspendedTask(cap,task);
2100 tso = task->suspended_tso;
2101 task->suspended_tso = NULL;
2102 tso->link = END_TSO_QUEUE;
2103 IF_DEBUG(scheduler, sched_belch("thread %d: re-entering RTS", tso->id));
2105 if (tso->why_blocked == BlockedOnCCall) {
2106 awakenBlockedQueue(cap,tso->blocked_exceptions);
2107 tso->blocked_exceptions = NULL;
2110 /* Reset blocking status */
2111 tso->why_blocked = NotBlocked;
2113 cap->r.rCurrentTSO = tso;
2114 cap->in_haskell = rtsTrue;
2115 errno = saved_errno;
2120 /* ---------------------------------------------------------------------------
2121 * Comparing Thread ids.
2123 * This is used from STG land in the implementation of the
2124 * instances of Eq/Ord for ThreadIds.
2125 * ------------------------------------------------------------------------ */
2128 cmp_thread(StgPtr tso1, StgPtr tso2)
2130 StgThreadID id1 = ((StgTSO *)tso1)->id;
2131 StgThreadID id2 = ((StgTSO *)tso2)->id;
2133 if (id1 < id2) return (-1);
2134 if (id1 > id2) return 1;
2138 /* ---------------------------------------------------------------------------
2139 * Fetching the ThreadID from an StgTSO.
2141 * This is used in the implementation of Show for ThreadIds.
2142 * ------------------------------------------------------------------------ */
2144 rts_getThreadId(StgPtr tso)
2146 return ((StgTSO *)tso)->id;
2151 labelThread(StgPtr tso, char *label)
2156 /* Caveat: Once set, you can only set the thread name to "" */
2157 len = strlen(label)+1;
2158 buf = stgMallocBytes(len * sizeof(char), "Schedule.c:labelThread()");
2159 strncpy(buf,label,len);
2160 /* Update will free the old memory for us */
2161 updateThreadLabel(((StgTSO *)tso)->id,buf);
2165 /* ---------------------------------------------------------------------------
2166 Create a new thread.
2168 The new thread starts with the given stack size. Before the
2169 scheduler can run, however, this thread needs to have a closure
2170 (and possibly some arguments) pushed on its stack. See
2171 pushClosure() in Schedule.h.
2173 createGenThread() and createIOThread() (in SchedAPI.h) are
2174 convenient packaged versions of this function.
2176 currently pri (priority) is only used in a GRAN setup -- HWL
2177 ------------------------------------------------------------------------ */
2179 /* currently pri (priority) is only used in a GRAN setup -- HWL */
2181 createThread(nat size, StgInt pri)
2184 createThread(Capability *cap, nat size)
2190 /* sched_mutex is *not* required */
2192 /* First check whether we should create a thread at all */
2193 #if defined(PARALLEL_HASKELL)
2194 /* check that no more than RtsFlags.ParFlags.maxThreads threads are created */
2195 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads) {
2197 debugBelch("{createThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)\n",
2198 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
2199 return END_TSO_QUEUE;
2205 ASSERT(!RtsFlags.GranFlags.Light || CurrentProc==0);
2208 // ToDo: check whether size = stack_size - TSO_STRUCT_SIZEW
2210 /* catch ridiculously small stack sizes */
2211 if (size < MIN_STACK_WORDS + TSO_STRUCT_SIZEW) {
2212 size = MIN_STACK_WORDS + TSO_STRUCT_SIZEW;
2215 stack_size = size - TSO_STRUCT_SIZEW;
2217 tso = (StgTSO *)allocateLocal(cap, size);
2218 TICK_ALLOC_TSO(stack_size, 0);
2220 SET_HDR(tso, &stg_TSO_info, CCS_SYSTEM);
2222 SET_GRAN_HDR(tso, ThisPE);
2225 // Always start with the compiled code evaluator
2226 tso->what_next = ThreadRunGHC;
2228 tso->why_blocked = NotBlocked;
2229 tso->blocked_exceptions = NULL;
2231 tso->saved_errno = 0;
2234 tso->stack_size = stack_size;
2235 tso->max_stack_size = round_to_mblocks(RtsFlags.GcFlags.maxStkSize)
2237 tso->sp = (P_)&(tso->stack) + stack_size;
2239 tso->trec = NO_TREC;
2242 tso->prof.CCCS = CCS_MAIN;
2245 /* put a stop frame on the stack */
2246 tso->sp -= sizeofW(StgStopFrame);
2247 SET_HDR((StgClosure*)tso->sp,(StgInfoTable *)&stg_stop_thread_info,CCS_SYSTEM);
2248 tso->link = END_TSO_QUEUE;
2252 /* uses more flexible routine in GranSim */
2253 insertThread(tso, CurrentProc);
2255 /* In a non-GranSim setup the pushing of a TSO onto the runq is separated
2261 if (RtsFlags.GranFlags.GranSimStats.Full)
2262 DumpGranEvent(GR_START,tso);
2263 #elif defined(PARALLEL_HASKELL)
2264 if (RtsFlags.ParFlags.ParStats.Full)
2265 DumpGranEvent(GR_STARTQ,tso);
2266 /* HACk to avoid SCHEDULE
2270 /* Link the new thread on the global thread list.
2272 ACQUIRE_LOCK(&sched_mutex);
2273 tso->id = next_thread_id++; // while we have the mutex
2274 tso->global_link = all_threads;
2276 RELEASE_LOCK(&sched_mutex);
2279 tso->dist.priority = MandatoryPriority; //by default that is...
2283 tso->gran.pri = pri;
2285 tso->gran.magic = TSO_MAGIC; // debugging only
2287 tso->gran.sparkname = 0;
2288 tso->gran.startedat = CURRENT_TIME;
2289 tso->gran.exported = 0;
2290 tso->gran.basicblocks = 0;
2291 tso->gran.allocs = 0;
2292 tso->gran.exectime = 0;
2293 tso->gran.fetchtime = 0;
2294 tso->gran.fetchcount = 0;
2295 tso->gran.blocktime = 0;
2296 tso->gran.blockcount = 0;
2297 tso->gran.blockedat = 0;
2298 tso->gran.globalsparks = 0;
2299 tso->gran.localsparks = 0;
2300 if (RtsFlags.GranFlags.Light)
2301 tso->gran.clock = Now; /* local clock */
2303 tso->gran.clock = 0;
2305 IF_DEBUG(gran,printTSO(tso));
2306 #elif defined(PARALLEL_HASKELL)
2308 tso->par.magic = TSO_MAGIC; // debugging only
2310 tso->par.sparkname = 0;
2311 tso->par.startedat = CURRENT_TIME;
2312 tso->par.exported = 0;
2313 tso->par.basicblocks = 0;
2314 tso->par.allocs = 0;
2315 tso->par.exectime = 0;
2316 tso->par.fetchtime = 0;
2317 tso->par.fetchcount = 0;
2318 tso->par.blocktime = 0;
2319 tso->par.blockcount = 0;
2320 tso->par.blockedat = 0;
2321 tso->par.globalsparks = 0;
2322 tso->par.localsparks = 0;
2326 globalGranStats.tot_threads_created++;
2327 globalGranStats.threads_created_on_PE[CurrentProc]++;
2328 globalGranStats.tot_sq_len += spark_queue_len(CurrentProc);
2329 globalGranStats.tot_sq_probes++;
2330 #elif defined(PARALLEL_HASKELL)
2331 // collect parallel global statistics (currently done together with GC stats)
2332 if (RtsFlags.ParFlags.ParStats.Global &&
2333 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
2334 //debugBelch("Creating thread %d @ %11.2f\n", tso->id, usertime());
2335 globalParStats.tot_threads_created++;
2341 sched_belch("==__ schedule: Created TSO %d (%p);",
2342 CurrentProc, tso, tso->id));
2343 #elif defined(PARALLEL_HASKELL)
2344 IF_PAR_DEBUG(verbose,
2345 sched_belch("==__ schedule: Created TSO %d (%p); %d threads active",
2346 (long)tso->id, tso, advisory_thread_count));
2348 IF_DEBUG(scheduler,sched_belch("created thread %ld, stack size = %lx words",
2349 (long)tso->id, (long)tso->stack_size));
2356 all parallel thread creation calls should fall through the following routine.
2359 createThreadFromSpark(rtsSpark spark)
2361 ASSERT(spark != (rtsSpark)NULL);
2362 // JB: TAKE CARE OF THIS COUNTER! BUGGY
2363 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads)
2365 barf("{createSparkThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
2366 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
2367 return END_TSO_QUEUE;
2371 tso = createThread(RtsFlags.GcFlags.initialStkSize);
2372 if (tso==END_TSO_QUEUE)
2373 barf("createSparkThread: Cannot create TSO");
2375 tso->priority = AdvisoryPriority;
2377 pushClosure(tso,spark);
2379 advisory_thread_count++; // JB: TAKE CARE OF THIS COUNTER! BUGGY
2386 Turn a spark into a thread.
2387 ToDo: fix for SMP (needs to acquire SCHED_MUTEX!)
2391 activateSpark (rtsSpark spark)
2395 tso = createSparkThread(spark);
2396 if (RtsFlags.ParFlags.ParStats.Full) {
2397 //ASSERT(run_queue_hd == END_TSO_QUEUE); // I think ...
2398 IF_PAR_DEBUG(verbose,
2399 debugBelch("==^^ activateSpark: turning spark of closure %p (%s) into a thread\n",
2400 (StgClosure *)spark, info_type((StgClosure *)spark)));
2402 // ToDo: fwd info on local/global spark to thread -- HWL
2403 // tso->gran.exported = spark->exported;
2404 // tso->gran.locked = !spark->global;
2405 // tso->gran.sparkname = spark->name;
2411 /* ---------------------------------------------------------------------------
2414 * scheduleThread puts a thread on the end of the runnable queue.
2415 * This will usually be done immediately after a thread is created.
2416 * The caller of scheduleThread must create the thread using e.g.
2417 * createThread and push an appropriate closure
2418 * on this thread's stack before the scheduler is invoked.
2419 * ------------------------------------------------------------------------ */
2422 scheduleThread(Capability *cap, StgTSO *tso)
2424 // The thread goes at the *end* of the run-queue, to avoid possible
2425 // starvation of any threads already on the queue.
2426 appendToRunQueue(cap,tso);
2430 scheduleWaitThread (StgTSO* tso, /*[out]*/HaskellObj* ret, Capability *cap)
2434 // We already created/initialised the Task
2435 task = cap->running_task;
2437 // This TSO is now a bound thread; make the Task and TSO
2438 // point to each other.
2443 task->stat = NoStatus;
2445 appendToRunQueue(cap,tso);
2447 IF_DEBUG(scheduler, sched_belch("new bound thread (%d)", tso->id));
2450 /* GranSim specific init */
2451 CurrentTSO = m->tso; // the TSO to run
2452 procStatus[MainProc] = Busy; // status of main PE
2453 CurrentProc = MainProc; // PE to run it on
2456 cap = schedule(cap,task);
2458 ASSERT(task->stat != NoStatus);
2460 IF_DEBUG(scheduler, sched_belch("bound thread (%d) finished", task->tso->id));
2464 /* ----------------------------------------------------------------------------
2466 * ------------------------------------------------------------------------- */
2468 #if defined(THREADED_RTS)
2470 workerStart(Task *task)
2474 // See startWorkerTask().
2475 ACQUIRE_LOCK(&task->lock);
2477 RELEASE_LOCK(&task->lock);
2479 // set the thread-local pointer to the Task:
2482 // schedule() runs without a lock.
2483 cap = schedule(cap,task);
2485 // On exit from schedule(), we have a Capability.
2486 releaseCapability(cap);
2491 /* ---------------------------------------------------------------------------
2494 * Initialise the scheduler. This resets all the queues - if the
2495 * queues contained any threads, they'll be garbage collected at the
2498 * ------------------------------------------------------------------------ */
2505 for (i=0; i<=MAX_PROC; i++) {
2506 run_queue_hds[i] = END_TSO_QUEUE;
2507 run_queue_tls[i] = END_TSO_QUEUE;
2508 blocked_queue_hds[i] = END_TSO_QUEUE;
2509 blocked_queue_tls[i] = END_TSO_QUEUE;
2510 ccalling_threadss[i] = END_TSO_QUEUE;
2511 blackhole_queue[i] = END_TSO_QUEUE;
2512 sleeping_queue = END_TSO_QUEUE;
2514 #elif !defined(THREADED_RTS)
2515 blocked_queue_hd = END_TSO_QUEUE;
2516 blocked_queue_tl = END_TSO_QUEUE;
2517 sleeping_queue = END_TSO_QUEUE;
2520 blackhole_queue = END_TSO_QUEUE;
2521 all_threads = END_TSO_QUEUE;
2526 RtsFlags.ConcFlags.ctxtSwitchTicks =
2527 RtsFlags.ConcFlags.ctxtSwitchTime / TICK_MILLISECS;
2529 #if defined(THREADED_RTS)
2530 /* Initialise the mutex and condition variables used by
2532 initMutex(&sched_mutex);
2535 ACQUIRE_LOCK(&sched_mutex);
2537 /* A capability holds the state a native thread needs in
2538 * order to execute STG code. At least one capability is
2539 * floating around (only SMP builds have more than one).
2547 * Eagerly start one worker to run each Capability, except for
2548 * Capability 0. The idea is that we're probably going to start a
2549 * bound thread on Capability 0 pretty soon, so we don't want a
2550 * worker task hogging it.
2555 for (i = 1; i < n_capabilities; i++) {
2556 cap = &capabilities[i];
2557 ACQUIRE_LOCK(&cap->lock);
2558 startWorkerTask(cap, workerStart);
2559 RELEASE_LOCK(&cap->lock);
2564 #if /* defined(SMP) ||*/ defined(PARALLEL_HASKELL)
2568 RELEASE_LOCK(&sched_mutex);
2572 exitScheduler( void )
2574 interrupted = rtsTrue;
2575 shutting_down_scheduler = rtsTrue;
2577 #if defined(THREADED_RTS)
2582 ACQUIRE_LOCK(&sched_mutex);
2583 task = newBoundTask();
2584 RELEASE_LOCK(&sched_mutex);
2586 for (i = 0; i < n_capabilities; i++) {
2587 shutdownCapability(&capabilities[i], task);
2589 boundTaskExiting(task);
2595 /* ---------------------------------------------------------------------------
2596 Where are the roots that we know about?
2598 - all the threads on the runnable queue
2599 - all the threads on the blocked queue
2600 - all the threads on the sleeping queue
2601 - all the thread currently executing a _ccall_GC
2602 - all the "main threads"
2604 ------------------------------------------------------------------------ */
2606 /* This has to be protected either by the scheduler monitor, or by the
2607 garbage collection monitor (probably the latter).
2612 GetRoots( evac_fn evac )
2619 for (i=0; i<=RtsFlags.GranFlags.proc; i++) {
2620 if ((run_queue_hds[i] != END_TSO_QUEUE) && ((run_queue_hds[i] != NULL)))
2621 evac((StgClosure **)&run_queue_hds[i]);
2622 if ((run_queue_tls[i] != END_TSO_QUEUE) && ((run_queue_tls[i] != NULL)))
2623 evac((StgClosure **)&run_queue_tls[i]);
2625 if ((blocked_queue_hds[i] != END_TSO_QUEUE) && ((blocked_queue_hds[i] != NULL)))
2626 evac((StgClosure **)&blocked_queue_hds[i]);
2627 if ((blocked_queue_tls[i] != END_TSO_QUEUE) && ((blocked_queue_tls[i] != NULL)))
2628 evac((StgClosure **)&blocked_queue_tls[i]);
2629 if ((ccalling_threadss[i] != END_TSO_QUEUE) && ((ccalling_threadss[i] != NULL)))
2630 evac((StgClosure **)&ccalling_threads[i]);
2637 for (i = 0; i < n_capabilities; i++) {
2638 cap = &capabilities[i];
2639 evac((StgClosure **)&cap->run_queue_hd);
2640 evac((StgClosure **)&cap->run_queue_tl);
2642 for (task = cap->suspended_ccalling_tasks; task != NULL;
2644 evac((StgClosure **)&task->suspended_tso);
2648 #if !defined(THREADED_RTS)
2649 evac((StgClosure **)&blocked_queue_hd);
2650 evac((StgClosure **)&blocked_queue_tl);
2651 evac((StgClosure **)&sleeping_queue);
2655 evac((StgClosure **)&blackhole_queue);
2657 #if defined(PARALLEL_HASKELL) || defined(GRAN)
2658 markSparkQueue(evac);
2661 #if defined(RTS_USER_SIGNALS)
2662 // mark the signal handlers (signals should be already blocked)
2663 markSignalHandlers(evac);
2667 /* -----------------------------------------------------------------------------
2670 This is the interface to the garbage collector from Haskell land.
2671 We provide this so that external C code can allocate and garbage
2672 collect when called from Haskell via _ccall_GC.
2674 It might be useful to provide an interface whereby the programmer
2675 can specify more roots (ToDo).
2677 This needs to be protected by the GC condition variable above. KH.
2678 -------------------------------------------------------------------------- */
2680 static void (*extra_roots)(evac_fn);
2686 // ToDo: we have to grab all the capabilities here.
2687 errorBelch("performGC not supported in threaded RTS (yet)");
2688 stg_exit(EXIT_FAILURE);
2690 /* Obligated to hold this lock upon entry */
2691 GarbageCollect(GetRoots,rtsFalse);
2695 performMajorGC(void)
2698 errorBelch("performMayjorGC not supported in threaded RTS (yet)");
2699 stg_exit(EXIT_FAILURE);
2701 GarbageCollect(GetRoots,rtsTrue);
2705 AllRoots(evac_fn evac)
2707 GetRoots(evac); // the scheduler's roots
2708 extra_roots(evac); // the user's roots
2712 performGCWithRoots(void (*get_roots)(evac_fn))
2715 errorBelch("performGCWithRoots not supported in threaded RTS (yet)");
2716 stg_exit(EXIT_FAILURE);
2718 extra_roots = get_roots;
2719 GarbageCollect(AllRoots,rtsFalse);
2722 /* -----------------------------------------------------------------------------
2725 If the thread has reached its maximum stack size, then raise the
2726 StackOverflow exception in the offending thread. Otherwise
2727 relocate the TSO into a larger chunk of memory and adjust its stack
2729 -------------------------------------------------------------------------- */
2732 threadStackOverflow(Capability *cap, StgTSO *tso)
2734 nat new_stack_size, stack_words;
2739 IF_DEBUG(sanity,checkTSO(tso));
2740 if (tso->stack_size >= tso->max_stack_size) {
2743 debugBelch("@@ threadStackOverflow of TSO %ld (%p): stack too large (now %ld; max is %ld)\n",
2744 (long)tso->id, tso, (long)tso->stack_size, (long)tso->max_stack_size);
2745 /* If we're debugging, just print out the top of the stack */
2746 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2749 /* Send this thread the StackOverflow exception */
2750 raiseAsync(cap, tso, (StgClosure *)stackOverflow_closure);
2754 /* Try to double the current stack size. If that takes us over the
2755 * maximum stack size for this thread, then use the maximum instead.
2756 * Finally round up so the TSO ends up as a whole number of blocks.
2758 new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
2759 new_tso_size = (lnat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) +
2760 TSO_STRUCT_SIZE)/sizeof(W_);
2761 new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */
2762 new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
2764 IF_DEBUG(scheduler, sched_belch("increasing stack size from %ld words to %d.\n", tso->stack_size, new_stack_size));
2766 dest = (StgTSO *)allocate(new_tso_size);
2767 TICK_ALLOC_TSO(new_stack_size,0);
2769 /* copy the TSO block and the old stack into the new area */
2770 memcpy(dest,tso,TSO_STRUCT_SIZE);
2771 stack_words = tso->stack + tso->stack_size - tso->sp;
2772 new_sp = (P_)dest + new_tso_size - stack_words;
2773 memcpy(new_sp, tso->sp, stack_words * sizeof(W_));
2775 /* relocate the stack pointers... */
2777 dest->stack_size = new_stack_size;
2779 /* Mark the old TSO as relocated. We have to check for relocated
2780 * TSOs in the garbage collector and any primops that deal with TSOs.
2782 * It's important to set the sp value to just beyond the end
2783 * of the stack, so we don't attempt to scavenge any part of the
2786 tso->what_next = ThreadRelocated;
2788 tso->sp = (P_)&(tso->stack[tso->stack_size]);
2789 tso->why_blocked = NotBlocked;
2791 IF_PAR_DEBUG(verbose,
2792 debugBelch("@@ threadStackOverflow of TSO %d (now at %p): stack size increased to %ld\n",
2793 tso->id, tso, tso->stack_size);
2794 /* If we're debugging, just print out the top of the stack */
2795 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2798 IF_DEBUG(sanity,checkTSO(tso));
2800 IF_DEBUG(scheduler,printTSO(dest));
2806 /* ---------------------------------------------------------------------------
2807 Wake up a queue that was blocked on some resource.
2808 ------------------------------------------------------------------------ */
2812 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2815 #elif defined(PARALLEL_HASKELL)
2817 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2819 /* write RESUME events to log file and
2820 update blocked and fetch time (depending on type of the orig closure) */
2821 if (RtsFlags.ParFlags.ParStats.Full) {
2822 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
2823 GR_RESUMEQ, ((StgTSO *)bqe), ((StgTSO *)bqe)->block_info.closure,
2824 0, 0 /* spark_queue_len(ADVISORY_POOL) */);
2825 if (emptyRunQueue())
2826 emitSchedule = rtsTrue;
2828 switch (get_itbl(node)->type) {
2830 ((StgTSO *)bqe)->par.fetchtime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2835 ((StgTSO *)bqe)->par.blocktime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2842 barf("{unblockOne}Daq Qagh: unexpected closure in blocking queue");
2849 StgBlockingQueueElement *
2850 unblockOne(StgBlockingQueueElement *bqe, StgClosure *node)
2853 PEs node_loc, tso_loc;
2855 node_loc = where_is(node); // should be lifted out of loop
2856 tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2857 tso_loc = where_is((StgClosure *)tso);
2858 if (IS_LOCAL_TO(PROCS(node),tso_loc)) { // TSO is local
2859 /* !fake_fetch => TSO is on CurrentProc is same as IS_LOCAL_TO */
2860 ASSERT(CurrentProc!=node_loc || tso_loc==CurrentProc);
2861 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.lunblocktime;
2862 // insertThread(tso, node_loc);
2863 new_event(tso_loc, tso_loc, CurrentTime[CurrentProc],
2865 tso, node, (rtsSpark*)NULL);
2866 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2869 } else { // TSO is remote (actually should be FMBQ)
2870 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.mpacktime +
2871 RtsFlags.GranFlags.Costs.gunblocktime +
2872 RtsFlags.GranFlags.Costs.latency;
2873 new_event(tso_loc, CurrentProc, CurrentTime[CurrentProc],
2875 tso, node, (rtsSpark*)NULL);
2876 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2879 /* the thread-queue-overhead is accounted for in either Resume or UnblockThread */
2881 debugBelch(" %s TSO %d (%p) [PE %d] (block_info.closure=%p) (next=%p) ,",
2882 (node_loc==tso_loc ? "Local" : "Global"),
2883 tso->id, tso, CurrentProc, tso->block_info.closure, tso->link));
2884 tso->block_info.closure = NULL;
2885 IF_DEBUG(scheduler,debugBelch("-- Waking up thread %ld (%p)\n",
2888 #elif defined(PARALLEL_HASKELL)
2889 StgBlockingQueueElement *
2890 unblockOne(StgBlockingQueueElement *bqe, StgClosure *node)
2892 StgBlockingQueueElement *next;
2894 switch (get_itbl(bqe)->type) {
2896 ASSERT(((StgTSO *)bqe)->why_blocked != NotBlocked);
2897 /* if it's a TSO just push it onto the run_queue */
2899 ((StgTSO *)bqe)->link = END_TSO_QUEUE; // debugging?
2900 APPEND_TO_RUN_QUEUE((StgTSO *)bqe);
2902 unblockCount(bqe, node);
2903 /* reset blocking status after dumping event */
2904 ((StgTSO *)bqe)->why_blocked = NotBlocked;
2908 /* if it's a BLOCKED_FETCH put it on the PendingFetches list */
2910 bqe->link = (StgBlockingQueueElement *)PendingFetches;
2911 PendingFetches = (StgBlockedFetch *)bqe;
2915 /* can ignore this case in a non-debugging setup;
2916 see comments on RBHSave closures above */
2918 /* check that the closure is an RBHSave closure */
2919 ASSERT(get_itbl((StgClosure *)bqe) == &stg_RBH_Save_0_info ||
2920 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_1_info ||
2921 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_2_info);
2925 barf("{unblockOne}Daq Qagh: Unexpected IP (%#lx; %s) in blocking queue at %#lx\n",
2926 get_itbl((StgClosure *)bqe), info_type((StgClosure *)bqe),
2930 IF_PAR_DEBUG(bq, debugBelch(", %p (%s)\n", bqe, info_type((StgClosure*)bqe)));
2936 unblockOne(Capability *cap, StgTSO *tso)
2940 ASSERT(get_itbl(tso)->type == TSO);
2941 ASSERT(tso->why_blocked != NotBlocked);
2942 tso->why_blocked = NotBlocked;
2944 tso->link = END_TSO_QUEUE;
2946 // We might have just migrated this TSO to our Capability:
2948 tso->bound->cap = cap;
2951 appendToRunQueue(cap,tso);
2953 // we're holding a newly woken thread, make sure we context switch
2954 // quickly so we can migrate it if necessary.
2956 IF_DEBUG(scheduler,sched_belch("waking up thread %ld", (long)tso->id));
2963 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2965 StgBlockingQueueElement *bqe;
2970 debugBelch("##-_ AwBQ for node %p on PE %d @ %ld by TSO %d (%p): \n", \
2971 node, CurrentProc, CurrentTime[CurrentProc],
2972 CurrentTSO->id, CurrentTSO));
2974 node_loc = where_is(node);
2976 ASSERT(q == END_BQ_QUEUE ||
2977 get_itbl(q)->type == TSO || // q is either a TSO or an RBHSave
2978 get_itbl(q)->type == CONSTR); // closure (type constructor)
2979 ASSERT(is_unique(node));
2981 /* FAKE FETCH: magically copy the node to the tso's proc;
2982 no Fetch necessary because in reality the node should not have been
2983 moved to the other PE in the first place
2985 if (CurrentProc!=node_loc) {
2987 debugBelch("## node %p is on PE %d but CurrentProc is %d (TSO %d); assuming fake fetch and adjusting bitmask (old: %#x)\n",
2988 node, node_loc, CurrentProc, CurrentTSO->id,
2989 // CurrentTSO, where_is(CurrentTSO),
2990 node->header.gran.procs));
2991 node->header.gran.procs = (node->header.gran.procs) | PE_NUMBER(CurrentProc);
2993 debugBelch("## new bitmask of node %p is %#x\n",
2994 node, node->header.gran.procs));
2995 if (RtsFlags.GranFlags.GranSimStats.Global) {
2996 globalGranStats.tot_fake_fetches++;
3001 // ToDo: check: ASSERT(CurrentProc==node_loc);
3002 while (get_itbl(bqe)->type==TSO) { // q != END_TSO_QUEUE) {
3005 bqe points to the current element in the queue
3006 next points to the next element in the queue
3008 //tso = (StgTSO *)bqe; // wastes an assignment to get the type right
3009 //tso_loc = where_is(tso);
3011 bqe = unblockOne(bqe, node);
3014 /* if this is the BQ of an RBH, we have to put back the info ripped out of
3015 the closure to make room for the anchor of the BQ */
3016 if (bqe!=END_BQ_QUEUE) {
3017 ASSERT(get_itbl(node)->type == RBH && get_itbl(bqe)->type == CONSTR);
3019 ASSERT((info_ptr==&RBH_Save_0_info) ||
3020 (info_ptr==&RBH_Save_1_info) ||
3021 (info_ptr==&RBH_Save_2_info));
3023 /* cf. convertToRBH in RBH.c for writing the RBHSave closure */
3024 ((StgRBH *)node)->blocking_queue = (StgBlockingQueueElement *)((StgRBHSave *)bqe)->payload[0];
3025 ((StgRBH *)node)->mut_link = (StgMutClosure *)((StgRBHSave *)bqe)->payload[1];
3028 debugBelch("## Filled in RBH_Save for %p (%s) at end of AwBQ\n",
3029 node, info_type(node)));
3032 /* statistics gathering */
3033 if (RtsFlags.GranFlags.GranSimStats.Global) {
3034 // globalGranStats.tot_bq_processing_time += bq_processing_time;
3035 globalGranStats.tot_bq_len += len; // total length of all bqs awakened
3036 // globalGranStats.tot_bq_len_local += len_local; // same for local TSOs only
3037 globalGranStats.tot_awbq++; // total no. of bqs awakened
3040 debugBelch("## BQ Stats of %p: [%d entries] %s\n",
3041 node, len, (bqe!=END_BQ_QUEUE) ? "RBH" : ""));
3043 #elif defined(PARALLEL_HASKELL)
3045 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
3047 StgBlockingQueueElement *bqe;
3049 IF_PAR_DEBUG(verbose,
3050 debugBelch("##-_ AwBQ for node %p on [%x]: \n",
3054 if(get_itbl(q)->type == CONSTR || q==END_BQ_QUEUE) {
3055 IF_PAR_DEBUG(verbose, debugBelch("## ... nothing to unblock so lets just return. RFP (BUG?)\n"));
3060 ASSERT(q == END_BQ_QUEUE ||
3061 get_itbl(q)->type == TSO ||
3062 get_itbl(q)->type == BLOCKED_FETCH ||
3063 get_itbl(q)->type == CONSTR);
3066 while (get_itbl(bqe)->type==TSO ||
3067 get_itbl(bqe)->type==BLOCKED_FETCH) {
3068 bqe = unblockOne(bqe, node);
3072 #else /* !GRAN && !PARALLEL_HASKELL */
3075 awakenBlockedQueue(Capability *cap, StgTSO *tso)
3077 if (tso == NULL) return; // hack; see bug #1235728, and comments in
3079 while (tso != END_TSO_QUEUE) {
3080 tso = unblockOne(cap,tso);
3085 /* ---------------------------------------------------------------------------
3087 - usually called inside a signal handler so it mustn't do anything fancy.
3088 ------------------------------------------------------------------------ */
3091 interruptStgRts(void)
3095 #if defined(THREADED_RTS)
3096 prodAllCapabilities();
3100 /* -----------------------------------------------------------------------------
3103 This is for use when we raise an exception in another thread, which
3105 This has nothing to do with the UnblockThread event in GranSim. -- HWL
3106 -------------------------------------------------------------------------- */
3108 #if defined(GRAN) || defined(PARALLEL_HASKELL)
3110 NB: only the type of the blocking queue is different in GranSim and GUM
3111 the operations on the queue-elements are the same
3112 long live polymorphism!
3114 Locks: sched_mutex is held upon entry and exit.
3118 unblockThread(Capability *cap, StgTSO *tso)
3120 StgBlockingQueueElement *t, **last;
3122 switch (tso->why_blocked) {
3125 return; /* not blocked */
3128 // Be careful: nothing to do here! We tell the scheduler that the thread
3129 // is runnable and we leave it to the stack-walking code to abort the
3130 // transaction while unwinding the stack. We should perhaps have a debugging
3131 // test to make sure that this really happens and that the 'zombie' transaction
3132 // does not get committed.
3136 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
3138 StgBlockingQueueElement *last_tso = END_BQ_QUEUE;
3139 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
3141 last = (StgBlockingQueueElement **)&mvar->head;
3142 for (t = (StgBlockingQueueElement *)mvar->head;
3144 last = &t->link, last_tso = t, t = t->link) {
3145 if (t == (StgBlockingQueueElement *)tso) {
3146 *last = (StgBlockingQueueElement *)tso->link;
3147 if (mvar->tail == tso) {
3148 mvar->tail = (StgTSO *)last_tso;
3153 barf("unblockThread (MVAR): TSO not found");
3156 case BlockedOnBlackHole:
3157 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
3159 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
3161 last = &bq->blocking_queue;
3162 for (t = bq->blocking_queue;
3164 last = &t->link, t = t->link) {
3165 if (t == (StgBlockingQueueElement *)tso) {
3166 *last = (StgBlockingQueueElement *)tso->link;
3170 barf("unblockThread (BLACKHOLE): TSO not found");
3173 case BlockedOnException:
3175 StgTSO *target = tso->block_info.tso;
3177 ASSERT(get_itbl(target)->type == TSO);
3179 if (target->what_next == ThreadRelocated) {
3180 target = target->link;
3181 ASSERT(get_itbl(target)->type == TSO);
3184 ASSERT(target->blocked_exceptions != NULL);
3186 last = (StgBlockingQueueElement **)&target->blocked_exceptions;
3187 for (t = (StgBlockingQueueElement *)target->blocked_exceptions;
3189 last = &t->link, t = t->link) {
3190 ASSERT(get_itbl(t)->type == TSO);
3191 if (t == (StgBlockingQueueElement *)tso) {
3192 *last = (StgBlockingQueueElement *)tso->link;
3196 barf("unblockThread (Exception): TSO not found");
3200 case BlockedOnWrite:
3201 #if defined(mingw32_HOST_OS)
3202 case BlockedOnDoProc:
3205 /* take TSO off blocked_queue */
3206 StgBlockingQueueElement *prev = NULL;
3207 for (t = (StgBlockingQueueElement *)blocked_queue_hd; t != END_BQ_QUEUE;
3208 prev = t, t = t->link) {
3209 if (t == (StgBlockingQueueElement *)tso) {
3211 blocked_queue_hd = (StgTSO *)t->link;
3212 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
3213 blocked_queue_tl = END_TSO_QUEUE;
3216 prev->link = t->link;
3217 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
3218 blocked_queue_tl = (StgTSO *)prev;
3221 #if defined(mingw32_HOST_OS)
3222 /* (Cooperatively) signal that the worker thread should abort
3225 abandonWorkRequest(tso->block_info.async_result->reqID);
3230 barf("unblockThread (I/O): TSO not found");
3233 case BlockedOnDelay:
3235 /* take TSO off sleeping_queue */
3236 StgBlockingQueueElement *prev = NULL;
3237 for (t = (StgBlockingQueueElement *)sleeping_queue; t != END_BQ_QUEUE;
3238 prev = t, t = t->link) {
3239 if (t == (StgBlockingQueueElement *)tso) {
3241 sleeping_queue = (StgTSO *)t->link;
3243 prev->link = t->link;
3248 barf("unblockThread (delay): TSO not found");
3252 barf("unblockThread");
3256 tso->link = END_TSO_QUEUE;
3257 tso->why_blocked = NotBlocked;
3258 tso->block_info.closure = NULL;
3259 pushOnRunQueue(cap,tso);
3263 unblockThread(Capability *cap, StgTSO *tso)
3267 /* To avoid locking unnecessarily. */
3268 if (tso->why_blocked == NotBlocked) {
3272 switch (tso->why_blocked) {
3275 // Be careful: nothing to do here! We tell the scheduler that the thread
3276 // is runnable and we leave it to the stack-walking code to abort the
3277 // transaction while unwinding the stack. We should perhaps have a debugging
3278 // test to make sure that this really happens and that the 'zombie' transaction
3279 // does not get committed.
3283 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
3285 StgTSO *last_tso = END_TSO_QUEUE;
3286 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
3289 for (t = mvar->head; t != END_TSO_QUEUE;
3290 last = &t->link, last_tso = t, t = t->link) {
3293 if (mvar->tail == tso) {
3294 mvar->tail = last_tso;
3299 barf("unblockThread (MVAR): TSO not found");
3302 case BlockedOnBlackHole:
3304 last = &blackhole_queue;
3305 for (t = blackhole_queue; t != END_TSO_QUEUE;
3306 last = &t->link, t = t->link) {
3312 barf("unblockThread (BLACKHOLE): TSO not found");
3315 case BlockedOnException:
3317 StgTSO *target = tso->block_info.tso;
3319 ASSERT(get_itbl(target)->type == TSO);
3321 while (target->what_next == ThreadRelocated) {
3322 target = target->link;
3323 ASSERT(get_itbl(target)->type == TSO);
3326 ASSERT(target->blocked_exceptions != NULL);
3328 last = &target->blocked_exceptions;
3329 for (t = target->blocked_exceptions; t != END_TSO_QUEUE;
3330 last = &t->link, t = t->link) {
3331 ASSERT(get_itbl(t)->type == TSO);
3337 barf("unblockThread (Exception): TSO not found");
3340 #if !defined(THREADED_RTS)
3342 case BlockedOnWrite:
3343 #if defined(mingw32_HOST_OS)
3344 case BlockedOnDoProc:
3347 StgTSO *prev = NULL;
3348 for (t = blocked_queue_hd; t != END_TSO_QUEUE;
3349 prev = t, t = t->link) {
3352 blocked_queue_hd = t->link;
3353 if (blocked_queue_tl == t) {
3354 blocked_queue_tl = END_TSO_QUEUE;
3357 prev->link = t->link;
3358 if (blocked_queue_tl == t) {
3359 blocked_queue_tl = prev;
3362 #if defined(mingw32_HOST_OS)
3363 /* (Cooperatively) signal that the worker thread should abort
3366 abandonWorkRequest(tso->block_info.async_result->reqID);
3371 barf("unblockThread (I/O): TSO not found");
3374 case BlockedOnDelay:
3376 StgTSO *prev = NULL;
3377 for (t = sleeping_queue; t != END_TSO_QUEUE;
3378 prev = t, t = t->link) {
3381 sleeping_queue = t->link;
3383 prev->link = t->link;
3388 barf("unblockThread (delay): TSO not found");
3393 barf("unblockThread");
3397 tso->link = END_TSO_QUEUE;
3398 tso->why_blocked = NotBlocked;
3399 tso->block_info.closure = NULL;
3400 appendToRunQueue(cap,tso);
3404 /* -----------------------------------------------------------------------------
3407 * Check the blackhole_queue for threads that can be woken up. We do
3408 * this periodically: before every GC, and whenever the run queue is
3411 * An elegant solution might be to just wake up all the blocked
3412 * threads with awakenBlockedQueue occasionally: they'll go back to
3413 * sleep again if the object is still a BLACKHOLE. Unfortunately this
3414 * doesn't give us a way to tell whether we've actually managed to
3415 * wake up any threads, so we would be busy-waiting.
3417 * -------------------------------------------------------------------------- */
3420 checkBlackHoles (Capability *cap)
3423 rtsBool any_woke_up = rtsFalse;
3426 // blackhole_queue is global:
3427 ASSERT_LOCK_HELD(&sched_mutex);
3429 IF_DEBUG(scheduler, sched_belch("checking threads blocked on black holes"));
3431 // ASSUMES: sched_mutex
3432 prev = &blackhole_queue;
3433 t = blackhole_queue;
3434 while (t != END_TSO_QUEUE) {
3435 ASSERT(t->why_blocked == BlockedOnBlackHole);
3436 type = get_itbl(t->block_info.closure)->type;
3437 if (type != BLACKHOLE && type != CAF_BLACKHOLE) {
3438 IF_DEBUG(sanity,checkTSO(t));
3439 t = unblockOne(cap, t);
3440 // urk, the threads migrate to the current capability
3441 // here, but we'd like to keep them on the original one.
3443 any_woke_up = rtsTrue;
3453 /* -----------------------------------------------------------------------------
3456 * The following function implements the magic for raising an
3457 * asynchronous exception in an existing thread.
3459 * We first remove the thread from any queue on which it might be
3460 * blocked. The possible blockages are MVARs and BLACKHOLE_BQs.
3462 * We strip the stack down to the innermost CATCH_FRAME, building
3463 * thunks in the heap for all the active computations, so they can
3464 * be restarted if necessary. When we reach a CATCH_FRAME, we build
3465 * an application of the handler to the exception, and push it on
3466 * the top of the stack.
3468 * How exactly do we save all the active computations? We create an
3469 * AP_STACK for every UpdateFrame on the stack. Entering one of these
3470 * AP_STACKs pushes everything from the corresponding update frame
3471 * upwards onto the stack. (Actually, it pushes everything up to the
3472 * next update frame plus a pointer to the next AP_STACK object.
3473 * Entering the next AP_STACK object pushes more onto the stack until we
3474 * reach the last AP_STACK object - at which point the stack should look
3475 * exactly as it did when we killed the TSO and we can continue
3476 * execution by entering the closure on top of the stack.
3478 * We can also kill a thread entirely - this happens if either (a) the
3479 * exception passed to raiseAsync is NULL, or (b) there's no
3480 * CATCH_FRAME on the stack. In either case, we strip the entire
3481 * stack and replace the thread with a zombie.
3483 * ToDo: in SMP mode, this function is only safe if either (a) we hold
3484 * all the Capabilities (eg. in GC), or (b) we own the Capability that
3485 * the TSO is currently blocked on or on the run queue of.
3487 * -------------------------------------------------------------------------- */
3490 raiseAsync(Capability *cap, StgTSO *tso, StgClosure *exception)
3492 raiseAsync_(cap, tso, exception, rtsFalse);
3496 raiseAsync_(Capability *cap, StgTSO *tso, StgClosure *exception,
3497 rtsBool stop_at_atomically)
3499 StgRetInfoTable *info;
3502 // Thread already dead?
3503 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
3508 sched_belch("raising exception in thread %ld.", (long)tso->id));
3510 // Remove it from any blocking queues
3511 unblockThread(cap,tso);
3515 // The stack freezing code assumes there's a closure pointer on
3516 // the top of the stack, so we have to arrange that this is the case...
3518 if (sp[0] == (W_)&stg_enter_info) {
3522 sp[0] = (W_)&stg_dummy_ret_closure;
3528 // 1. Let the top of the stack be the "current closure"
3530 // 2. Walk up the stack until we find either an UPDATE_FRAME or a
3533 // 3. If it's an UPDATE_FRAME, then make an AP_STACK containing the
3534 // current closure applied to the chunk of stack up to (but not
3535 // including) the update frame. This closure becomes the "current
3536 // closure". Go back to step 2.
3538 // 4. If it's a CATCH_FRAME, then leave the exception handler on
3539 // top of the stack applied to the exception.
3541 // 5. If it's a STOP_FRAME, then kill the thread.
3543 // NB: if we pass an ATOMICALLY_FRAME then abort the associated
3550 info = get_ret_itbl((StgClosure *)frame);
3552 while (info->i.type != UPDATE_FRAME
3553 && (info->i.type != CATCH_FRAME || exception == NULL)
3554 && info->i.type != STOP_FRAME
3555 && (info->i.type != ATOMICALLY_FRAME || stop_at_atomically == rtsFalse))
3557 if (info->i.type == CATCH_RETRY_FRAME || info->i.type == ATOMICALLY_FRAME) {
3558 // IF we find an ATOMICALLY_FRAME then we abort the
3559 // current transaction and propagate the exception. In
3560 // this case (unlike ordinary exceptions) we do not care
3561 // whether the transaction is valid or not because its
3562 // possible validity cannot have caused the exception
3563 // and will not be visible after the abort.
3565 debugBelch("Found atomically block delivering async exception\n"));
3566 stmAbortTransaction(tso -> trec);
3567 tso -> trec = stmGetEnclosingTRec(tso -> trec);
3569 frame += stack_frame_sizeW((StgClosure *)frame);
3570 info = get_ret_itbl((StgClosure *)frame);
3573 switch (info->i.type) {
3575 case ATOMICALLY_FRAME:
3576 ASSERT(stop_at_atomically);
3577 ASSERT(stmGetEnclosingTRec(tso->trec) == NO_TREC);
3578 stmCondemnTransaction(tso -> trec);
3582 // R1 is not a register: the return convention for IO in
3583 // this case puts the return value on the stack, so we
3584 // need to set up the stack to return to the atomically
3585 // frame properly...
3586 tso->sp = frame - 2;
3587 tso->sp[1] = (StgWord) &stg_NO_FINALIZER_closure; // why not?
3588 tso->sp[0] = (StgWord) &stg_ut_1_0_unreg_info;
3590 tso->what_next = ThreadRunGHC;
3594 // If we find a CATCH_FRAME, and we've got an exception to raise,
3595 // then build the THUNK raise(exception), and leave it on
3596 // top of the CATCH_FRAME ready to enter.
3600 StgCatchFrame *cf = (StgCatchFrame *)frame;
3604 // we've got an exception to raise, so let's pass it to the
3605 // handler in this frame.
3607 raise = (StgThunk *)allocateLocal(cap,sizeofW(StgThunk)+MIN_UPD_SIZE);
3608 TICK_ALLOC_SE_THK(1,0);
3609 SET_HDR(raise,&stg_raise_info,cf->header.prof.ccs);
3610 raise->payload[0] = exception;
3612 // throw away the stack from Sp up to the CATCH_FRAME.
3616 /* Ensure that async excpetions are blocked now, so we don't get
3617 * a surprise exception before we get around to executing the
3620 if (tso->blocked_exceptions == NULL) {
3621 tso->blocked_exceptions = END_TSO_QUEUE;
3624 /* Put the newly-built THUNK on top of the stack, ready to execute
3625 * when the thread restarts.
3628 sp[-1] = (W_)&stg_enter_info;
3630 tso->what_next = ThreadRunGHC;
3631 IF_DEBUG(sanity, checkTSO(tso));
3640 // First build an AP_STACK consisting of the stack chunk above the
3641 // current update frame, with the top word on the stack as the
3644 words = frame - sp - 1;
3645 ap = (StgAP_STACK *)allocateLocal(cap,AP_STACK_sizeW(words));
3648 ap->fun = (StgClosure *)sp[0];
3650 for(i=0; i < (nat)words; ++i) {
3651 ap->payload[i] = (StgClosure *)*sp++;
3654 SET_HDR(ap,&stg_AP_STACK_info,
3655 ((StgClosure *)frame)->header.prof.ccs /* ToDo */);
3656 TICK_ALLOC_UP_THK(words+1,0);
3659 debugBelch("sched: Updating ");
3660 printPtr((P_)((StgUpdateFrame *)frame)->updatee);
3661 debugBelch(" with ");
3662 printObj((StgClosure *)ap);
3665 // Replace the updatee with an indirection - happily
3666 // this will also wake up any threads currently
3667 // waiting on the result.
3669 // Warning: if we're in a loop, more than one update frame on
3670 // the stack may point to the same object. Be careful not to
3671 // overwrite an IND_OLDGEN in this case, because we'll screw
3672 // up the mutable lists. To be on the safe side, don't
3673 // overwrite any kind of indirection at all. See also
3674 // threadSqueezeStack in GC.c, where we have to make a similar
3677 if (!closure_IND(((StgUpdateFrame *)frame)->updatee)) {
3678 // revert the black hole
3679 UPD_IND_NOLOCK(((StgUpdateFrame *)frame)->updatee,
3682 sp += sizeofW(StgUpdateFrame) - 1;
3683 sp[0] = (W_)ap; // push onto stack
3688 // We've stripped the entire stack, the thread is now dead.
3689 sp += sizeofW(StgStopFrame);
3690 tso->what_next = ThreadKilled;
3701 /* -----------------------------------------------------------------------------
3704 This is used for interruption (^C) and forking, and corresponds to
3705 raising an exception but without letting the thread catch the
3707 -------------------------------------------------------------------------- */
3710 deleteThread (Capability *cap, StgTSO *tso)
3712 if (tso->why_blocked != BlockedOnCCall &&
3713 tso->why_blocked != BlockedOnCCall_NoUnblockExc) {
3714 raiseAsync(cap,tso,NULL);
3718 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
3720 deleteThreadImmediately(Capability *cap, StgTSO *tso)
3721 { // for forkProcess only:
3722 // delete thread without giving it a chance to catch the KillThread exception
3724 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
3728 if (tso->why_blocked != BlockedOnCCall &&
3729 tso->why_blocked != BlockedOnCCall_NoUnblockExc) {
3730 unblockThread(cap,tso);
3733 tso->what_next = ThreadKilled;
3737 /* -----------------------------------------------------------------------------
3738 raiseExceptionHelper
3740 This function is called by the raise# primitve, just so that we can
3741 move some of the tricky bits of raising an exception from C-- into
3742 C. Who knows, it might be a useful re-useable thing here too.
3743 -------------------------------------------------------------------------- */
3746 raiseExceptionHelper (StgRegTable *reg, StgTSO *tso, StgClosure *exception)
3748 Capability *cap = regTableToCapability(reg);
3749 StgThunk *raise_closure = NULL;
3751 StgRetInfoTable *info;
3753 // This closure represents the expression 'raise# E' where E
3754 // is the exception raise. It is used to overwrite all the
3755 // thunks which are currently under evaluataion.
3759 // LDV profiling: stg_raise_info has THUNK as its closure
3760 // type. Since a THUNK takes at least MIN_UPD_SIZE words in its
3761 // payload, MIN_UPD_SIZE is more approprate than 1. It seems that
3762 // 1 does not cause any problem unless profiling is performed.
3763 // However, when LDV profiling goes on, we need to linearly scan
3764 // small object pool, where raise_closure is stored, so we should
3765 // use MIN_UPD_SIZE.
3767 // raise_closure = (StgClosure *)RET_STGCALL1(P_,allocate,
3768 // sizeofW(StgClosure)+1);
3772 // Walk up the stack, looking for the catch frame. On the way,
3773 // we update any closures pointed to from update frames with the
3774 // raise closure that we just built.
3778 info = get_ret_itbl((StgClosure *)p);
3779 next = p + stack_frame_sizeW((StgClosure *)p);
3780 switch (info->i.type) {
3783 // Only create raise_closure if we need to.
3784 if (raise_closure == NULL) {
3786 (StgThunk *)allocateLocal(cap,sizeofW(StgThunk)+MIN_UPD_SIZE);
3787 SET_HDR(raise_closure, &stg_raise_info, CCCS);
3788 raise_closure->payload[0] = exception;
3790 UPD_IND(((StgUpdateFrame *)p)->updatee,(StgClosure *)raise_closure);
3794 case ATOMICALLY_FRAME:
3795 IF_DEBUG(stm, debugBelch("Found ATOMICALLY_FRAME at %p\n", p));
3797 return ATOMICALLY_FRAME;
3803 case CATCH_STM_FRAME:
3804 IF_DEBUG(stm, debugBelch("Found CATCH_STM_FRAME at %p\n", p));
3806 return CATCH_STM_FRAME;
3812 case CATCH_RETRY_FRAME:
3821 /* -----------------------------------------------------------------------------
3822 findRetryFrameHelper
3824 This function is called by the retry# primitive. It traverses the stack
3825 leaving tso->sp referring to the frame which should handle the retry.
3827 This should either be a CATCH_RETRY_FRAME (if the retry# is within an orElse#)
3828 or should be a ATOMICALLY_FRAME (if the retry# reaches the top level).
3830 We skip CATCH_STM_FRAMEs because retries are not considered to be exceptions,
3831 despite the similar implementation.
3833 We should not expect to see CATCH_FRAME or STOP_FRAME because those should
3834 not be created within memory transactions.
3835 -------------------------------------------------------------------------- */
3838 findRetryFrameHelper (StgTSO *tso)
3841 StgRetInfoTable *info;
3845 info = get_ret_itbl((StgClosure *)p);
3846 next = p + stack_frame_sizeW((StgClosure *)p);
3847 switch (info->i.type) {
3849 case ATOMICALLY_FRAME:
3850 IF_DEBUG(stm, debugBelch("Found ATOMICALLY_FRAME at %p during retrry\n", p));
3852 return ATOMICALLY_FRAME;
3854 case CATCH_RETRY_FRAME:
3855 IF_DEBUG(stm, debugBelch("Found CATCH_RETRY_FRAME at %p during retrry\n", p));
3857 return CATCH_RETRY_FRAME;
3859 case CATCH_STM_FRAME:
3861 ASSERT(info->i.type != CATCH_FRAME);
3862 ASSERT(info->i.type != STOP_FRAME);
3869 /* -----------------------------------------------------------------------------
3870 resurrectThreads is called after garbage collection on the list of
3871 threads found to be garbage. Each of these threads will be woken
3872 up and sent a signal: BlockedOnDeadMVar if the thread was blocked
3873 on an MVar, or NonTermination if the thread was blocked on a Black
3876 Locks: assumes we hold *all* the capabilities.
3877 -------------------------------------------------------------------------- */
3880 resurrectThreads (StgTSO *threads)
3885 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
3886 next = tso->global_link;
3887 tso->global_link = all_threads;
3889 IF_DEBUG(scheduler, sched_belch("resurrecting thread %d", tso->id));
3891 // Wake up the thread on the Capability it was last on for a
3892 // bound thread, or last_free_capability otherwise.
3894 cap = tso->bound->cap;
3896 cap = last_free_capability;
3899 switch (tso->why_blocked) {
3901 case BlockedOnException:
3902 /* Called by GC - sched_mutex lock is currently held. */
3903 raiseAsync(cap, tso,(StgClosure *)BlockedOnDeadMVar_closure);
3905 case BlockedOnBlackHole:
3906 raiseAsync(cap, tso,(StgClosure *)NonTermination_closure);
3909 raiseAsync(cap, tso,(StgClosure *)BlockedIndefinitely_closure);
3912 /* This might happen if the thread was blocked on a black hole
3913 * belonging to a thread that we've just woken up (raiseAsync
3914 * can wake up threads, remember...).
3918 barf("resurrectThreads: thread blocked in a strange way");
3923 /* ----------------------------------------------------------------------------
3924 * Debugging: why is a thread blocked
3925 * [Also provides useful information when debugging threaded programs
3926 * at the Haskell source code level, so enable outside of DEBUG. --sof 7/02]
3927 ------------------------------------------------------------------------- */
3931 printThreadBlockage(StgTSO *tso)
3933 switch (tso->why_blocked) {
3935 debugBelch("is blocked on read from fd %d", (int)(tso->block_info.fd));
3937 case BlockedOnWrite:
3938 debugBelch("is blocked on write to fd %d", (int)(tso->block_info.fd));
3940 #if defined(mingw32_HOST_OS)
3941 case BlockedOnDoProc:
3942 debugBelch("is blocked on proc (request: %ld)", tso->block_info.async_result->reqID);
3945 case BlockedOnDelay:
3946 debugBelch("is blocked until %ld", (long)(tso->block_info.target));
3949 debugBelch("is blocked on an MVar @ %p", tso->block_info.closure);
3951 case BlockedOnException:
3952 debugBelch("is blocked on delivering an exception to thread %d",
3953 tso->block_info.tso->id);
3955 case BlockedOnBlackHole:
3956 debugBelch("is blocked on a black hole");
3959 debugBelch("is not blocked");
3961 #if defined(PARALLEL_HASKELL)
3963 debugBelch("is blocked on global address; local FM_BQ is %p (%s)",
3964 tso->block_info.closure, info_type(tso->block_info.closure));
3966 case BlockedOnGA_NoSend:
3967 debugBelch("is blocked on global address (no send); local FM_BQ is %p (%s)",
3968 tso->block_info.closure, info_type(tso->block_info.closure));
3971 case BlockedOnCCall:
3972 debugBelch("is blocked on an external call");
3974 case BlockedOnCCall_NoUnblockExc:
3975 debugBelch("is blocked on an external call (exceptions were already blocked)");
3978 debugBelch("is blocked on an STM operation");
3981 barf("printThreadBlockage: strange tso->why_blocked: %d for TSO %d (%d)",
3982 tso->why_blocked, tso->id, tso);
3987 printThreadStatus(StgTSO *tso)
3989 switch (tso->what_next) {
3991 debugBelch("has been killed");
3993 case ThreadComplete:
3994 debugBelch("has completed");
3997 printThreadBlockage(tso);
4002 printAllThreads(void)
4007 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
4008 ullong_format_string(TIME_ON_PROC(CurrentProc),
4009 time_string, rtsFalse/*no commas!*/);
4011 debugBelch("all threads at [%s]:\n", time_string);
4012 # elif defined(PARALLEL_HASKELL)
4013 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
4014 ullong_format_string(CURRENT_TIME,
4015 time_string, rtsFalse/*no commas!*/);
4017 debugBelch("all threads at [%s]:\n", time_string);
4019 debugBelch("all threads:\n");
4022 for (t = all_threads; t != END_TSO_QUEUE; ) {
4023 debugBelch("\tthread %4d @ %p ", t->id, (void *)t);
4025 void *label = lookupThreadLabel(t->id);
4026 if (label) debugBelch("[\"%s\"] ",(char *)label);
4028 if (t->what_next == ThreadRelocated) {
4029 debugBelch("has been relocated...\n");
4032 printThreadStatus(t);
4041 printThreadQueue(StgTSO *t)
4044 for (; t != END_TSO_QUEUE; t = t->link) {
4045 debugBelch("\tthread %d @ %p ", t->id, (void *)t);
4046 if (t->what_next == ThreadRelocated) {
4047 debugBelch("has been relocated...\n");
4049 printThreadStatus(t);
4054 debugBelch("%d threads on queue\n", i);
4058 Print a whole blocking queue attached to node (debugging only).
4060 # if defined(PARALLEL_HASKELL)
4062 print_bq (StgClosure *node)
4064 StgBlockingQueueElement *bqe;
4068 debugBelch("## BQ of closure %p (%s): ",
4069 node, info_type(node));
4071 /* should cover all closures that may have a blocking queue */
4072 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
4073 get_itbl(node)->type == FETCH_ME_BQ ||
4074 get_itbl(node)->type == RBH ||
4075 get_itbl(node)->type == MVAR);
4077 ASSERT(node!=(StgClosure*)NULL); // sanity check
4079 print_bqe(((StgBlockingQueue*)node)->blocking_queue);
4083 Print a whole blocking queue starting with the element bqe.
4086 print_bqe (StgBlockingQueueElement *bqe)
4091 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
4093 for (end = (bqe==END_BQ_QUEUE);
4094 !end; // iterate until bqe points to a CONSTR
4095 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE),
4096 bqe = end ? END_BQ_QUEUE : bqe->link) {
4097 ASSERT(bqe != END_BQ_QUEUE); // sanity check
4098 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
4099 /* types of closures that may appear in a blocking queue */
4100 ASSERT(get_itbl(bqe)->type == TSO ||
4101 get_itbl(bqe)->type == BLOCKED_FETCH ||
4102 get_itbl(bqe)->type == CONSTR);
4103 /* only BQs of an RBH end with an RBH_Save closure */
4104 //ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
4106 switch (get_itbl(bqe)->type) {
4108 debugBelch(" TSO %u (%x),",
4109 ((StgTSO *)bqe)->id, ((StgTSO *)bqe));
4112 debugBelch(" BF (node=%p, ga=((%x, %d, %x)),",
4113 ((StgBlockedFetch *)bqe)->node,
4114 ((StgBlockedFetch *)bqe)->ga.payload.gc.gtid,
4115 ((StgBlockedFetch *)bqe)->ga.payload.gc.slot,
4116 ((StgBlockedFetch *)bqe)->ga.weight);
4119 debugBelch(" %s (IP %p),",
4120 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
4121 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
4122 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
4123 "RBH_Save_?"), get_itbl(bqe));
4126 barf("Unexpected closure type %s in blocking queue", // of %p (%s)",
4127 info_type((StgClosure *)bqe)); // , node, info_type(node));
4133 # elif defined(GRAN)
4135 print_bq (StgClosure *node)
4137 StgBlockingQueueElement *bqe;
4138 PEs node_loc, tso_loc;
4141 /* should cover all closures that may have a blocking queue */
4142 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
4143 get_itbl(node)->type == FETCH_ME_BQ ||
4144 get_itbl(node)->type == RBH);
4146 ASSERT(node!=(StgClosure*)NULL); // sanity check
4147 node_loc = where_is(node);
4149 debugBelch("## BQ of closure %p (%s) on [PE %d]: ",
4150 node, info_type(node), node_loc);
4153 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
4155 for (bqe = ((StgBlockingQueue*)node)->blocking_queue, end = (bqe==END_BQ_QUEUE);
4156 !end; // iterate until bqe points to a CONSTR
4157 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) {
4158 ASSERT(bqe != END_BQ_QUEUE); // sanity check
4159 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
4160 /* types of closures that may appear in a blocking queue */
4161 ASSERT(get_itbl(bqe)->type == TSO ||
4162 get_itbl(bqe)->type == CONSTR);
4163 /* only BQs of an RBH end with an RBH_Save closure */
4164 ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
4166 tso_loc = where_is((StgClosure *)bqe);
4167 switch (get_itbl(bqe)->type) {
4169 debugBelch(" TSO %d (%p) on [PE %d],",
4170 ((StgTSO *)bqe)->id, (StgTSO *)bqe, tso_loc);
4173 debugBelch(" %s (IP %p),",
4174 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
4175 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
4176 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
4177 "RBH_Save_?"), get_itbl(bqe));
4180 barf("Unexpected closure type %s in blocking queue of %p (%s)",
4181 info_type((StgClosure *)bqe), node, info_type(node));
4189 #if defined(PARALLEL_HASKELL)
4196 for (i=0, tso=run_queue_hd;
4197 tso != END_TSO_QUEUE;
4198 i++, tso=tso->link) {
4207 sched_belch(char *s, ...)
4212 debugBelch("sched (task %p): ", (void *)(unsigned long)(unsigned int)osThreadId());
4213 #elif defined(PARALLEL_HASKELL)
4216 debugBelch("sched: ");