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"));
411 // If we are a worker, just exit. If we're a bound thread
412 // then we will exit below when we've removed our TSO from
414 if (task->tso == NULL) {
418 IF_DEBUG(scheduler, sched_belch("interrupted"));
422 #if defined(not_yet) && defined(SMP)
424 // Top up the run queue from our spark pool. We try to make the
425 // number of threads in the run queue equal to the number of
426 // free capabilities.
430 if (emptyRunQueue()) {
431 spark = findSpark(rtsFalse);
433 break; /* no more sparks in the pool */
435 createSparkThread(spark);
437 sched_belch("==^^ turning spark of closure %p into a thread",
438 (StgClosure *)spark));
444 scheduleStartSignalHandlers();
446 // Only check the black holes here if we've nothing else to do.
447 // During normal execution, the black hole list only gets checked
448 // at GC time, to avoid repeatedly traversing this possibly long
449 // list each time around the scheduler.
450 if (emptyRunQueue(cap)) { scheduleCheckBlackHoles(cap); }
452 scheduleCheckBlockedThreads(cap);
454 scheduleDetectDeadlock(cap,task);
456 // Normally, the only way we can get here with no threads to
457 // run is if a keyboard interrupt received during
458 // scheduleCheckBlockedThreads() or scheduleDetectDeadlock().
459 // Additionally, it is not fatal for the
460 // threaded RTS to reach here with no threads to run.
462 // win32: might be here due to awaitEvent() being abandoned
463 // as a result of a console event having been delivered.
464 if ( emptyRunQueue(cap) ) {
465 #if !defined(THREADED_RTS) && !defined(mingw32_HOST_OS)
468 continue; // nothing to do
471 #if defined(PARALLEL_HASKELL)
472 scheduleSendPendingMessages();
473 if (emptyRunQueue(cap) && scheduleActivateSpark())
477 ASSERT(next_fish_to_send_at==0); // i.e. no delayed fishes left!
480 /* If we still have no work we need to send a FISH to get a spark
482 if (emptyRunQueue(cap)) {
483 if (!scheduleGetRemoteWork(&receivedFinish)) continue;
484 ASSERT(rtsFalse); // should not happen at the moment
486 // from here: non-empty run queue.
487 // TODO: merge above case with this, only one call processMessages() !
488 if (PacketsWaiting()) { /* process incoming messages, if
489 any pending... only in else
490 because getRemoteWork waits for
492 receivedFinish = processMessages();
497 scheduleProcessEvent(event);
501 // Get a thread to run
503 t = popRunQueue(cap);
505 #if defined(GRAN) || defined(PAR)
506 scheduleGranParReport(); // some kind of debuging output
508 // Sanity check the thread we're about to run. This can be
509 // expensive if there is lots of thread switching going on...
510 IF_DEBUG(sanity,checkTSO(t));
513 #if defined(THREADED_RTS)
514 // Check whether we can run this thread in the current task.
515 // If not, we have to pass our capability to the right task.
517 Task *bound = t->bound;
522 sched_belch("### Running thread %d in bound thread",
524 // yes, the Haskell thread is bound to the current native thread
527 sched_belch("### thread %d bound to another OS thread",
529 // no, bound to a different Haskell thread: pass to that thread
530 pushOnRunQueue(cap,t);
534 // The thread we want to run is unbound.
537 sched_belch("### this OS thread cannot run thread %d", t->id));
538 // no, the current native thread is bound to a different
539 // Haskell thread, so pass it to any worker thread
540 pushOnRunQueue(cap,t);
547 cap->r.rCurrentTSO = t;
549 /* context switches are initiated by the timer signal, unless
550 * the user specified "context switch as often as possible", with
553 if (RtsFlags.ConcFlags.ctxtSwitchTicks == 0
554 && !emptyThreadQueues(cap)) {
560 IF_DEBUG(scheduler, sched_belch("-->> running thread %ld %s ...",
561 (long)t->id, whatNext_strs[t->what_next]));
563 #if defined(PROFILING)
564 startHeapProfTimer();
567 // ----------------------------------------------------------------------
568 // Run the current thread
570 prev_what_next = t->what_next;
572 errno = t->saved_errno;
573 cap->in_haskell = rtsTrue;
575 recent_activity = ACTIVITY_YES;
577 switch (prev_what_next) {
581 /* Thread already finished, return to scheduler. */
582 ret = ThreadFinished;
588 r = StgRun((StgFunPtr) stg_returnToStackTop, &cap->r);
589 cap = regTableToCapability(r);
594 case ThreadInterpret:
595 cap = interpretBCO(cap);
600 barf("schedule: invalid what_next field");
603 cap->in_haskell = rtsFalse;
606 // If ret is ThreadBlocked, and this Task is bound to the TSO that
607 // blocked, we are in limbo - the TSO is now owned by whatever it
608 // is blocked on, and may in fact already have been woken up,
609 // perhaps even on a different Capability. It may be the case
610 // that task->cap != cap. We better yield this Capability
611 // immediately and return to normaility.
612 if (ret == ThreadBlocked) continue;
615 ASSERT(cap->running_task == task);
616 ASSERT(task->cap == cap);
617 ASSERT(myTask() == task);
619 // The TSO might have moved, eg. if it re-entered the RTS and a GC
620 // happened. So find the new location:
621 t = cap->r.rCurrentTSO;
623 // And save the current errno in this thread.
624 t->saved_errno = errno;
626 // ----------------------------------------------------------------------
628 // Costs for the scheduler are assigned to CCS_SYSTEM
629 #if defined(PROFILING)
634 // We have run some Haskell code: there might be blackhole-blocked
635 // threads to wake up now.
636 // Lock-free test here should be ok, we're just setting a flag.
637 if ( blackhole_queue != END_TSO_QUEUE ) {
638 blackholes_need_checking = rtsTrue;
641 #if defined(THREADED_RTS)
642 IF_DEBUG(scheduler,debugBelch("sched (task %p): ", (void *)(unsigned long)(unsigned int)osThreadId()););
643 #elif !defined(GRAN) && !defined(PARALLEL_HASKELL)
644 IF_DEBUG(scheduler,debugBelch("sched: "););
647 schedulePostRunThread();
649 ready_to_gc = rtsFalse;
653 ready_to_gc = scheduleHandleHeapOverflow(cap,t);
657 scheduleHandleStackOverflow(cap,task,t);
661 if (scheduleHandleYield(cap, t, prev_what_next)) {
662 // shortcut for switching between compiler/interpreter:
668 scheduleHandleThreadBlocked(t);
672 if (scheduleHandleThreadFinished(cap, task, t)) return cap;
676 barf("schedule: invalid thread return code %d", (int)ret);
679 if (scheduleDoHeapProfile(ready_to_gc)) { ready_to_gc = rtsFalse; }
680 if (ready_to_gc) { scheduleDoGC(cap,task,rtsFalse); }
681 } /* end of while() */
683 IF_PAR_DEBUG(verbose,
684 debugBelch("== Leaving schedule() after having received Finish\n"));
687 /* ----------------------------------------------------------------------------
688 * Setting up the scheduler loop
689 * ------------------------------------------------------------------------- */
692 schedulePreLoop(void)
695 /* set up first event to get things going */
696 /* ToDo: assign costs for system setup and init MainTSO ! */
697 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
699 CurrentTSO, (StgClosure*)NULL, (rtsSpark*)NULL);
702 debugBelch("GRAN: Init CurrentTSO (in schedule) = %p\n",
704 G_TSO(CurrentTSO, 5));
706 if (RtsFlags.GranFlags.Light) {
707 /* Save current time; GranSim Light only */
708 CurrentTSO->gran.clock = CurrentTime[CurrentProc];
713 /* ----------------------------------------------------------------------------
714 * Start any pending signal handlers
715 * ------------------------------------------------------------------------- */
718 scheduleStartSignalHandlers(void)
720 #if defined(RTS_USER_SIGNALS) && !defined(THREADED_RTS)
721 if (signals_pending()) { // safe outside the lock
722 startSignalHandlers();
727 /* ----------------------------------------------------------------------------
728 * Check for blocked threads that can be woken up.
729 * ------------------------------------------------------------------------- */
732 scheduleCheckBlockedThreads(Capability *cap USED_WHEN_NON_THREADED_RTS)
734 #if !defined(THREADED_RTS)
736 // Check whether any waiting threads need to be woken up. If the
737 // run queue is empty, and there are no other tasks running, we
738 // can wait indefinitely for something to happen.
740 if ( !emptyQueue(blocked_queue_hd) || !emptyQueue(sleeping_queue) )
742 awaitEvent( emptyRunQueue(cap) && !blackholes_need_checking );
748 /* ----------------------------------------------------------------------------
749 * Check for threads blocked on BLACKHOLEs that can be woken up
750 * ------------------------------------------------------------------------- */
752 scheduleCheckBlackHoles (Capability *cap)
754 if ( blackholes_need_checking ) // check without the lock first
756 ACQUIRE_LOCK(&sched_mutex);
757 if ( blackholes_need_checking ) {
758 checkBlackHoles(cap);
759 blackholes_need_checking = rtsFalse;
761 RELEASE_LOCK(&sched_mutex);
765 /* ----------------------------------------------------------------------------
766 * Detect deadlock conditions and attempt to resolve them.
767 * ------------------------------------------------------------------------- */
770 scheduleDetectDeadlock (Capability *cap, Task *task)
773 #if defined(PARALLEL_HASKELL)
774 // ToDo: add deadlock detection in GUM (similar to SMP) -- HWL
779 * Detect deadlock: when we have no threads to run, there are no
780 * threads blocked, waiting for I/O, or sleeping, and all the
781 * other tasks are waiting for work, we must have a deadlock of
784 if ( emptyThreadQueues(cap) )
786 #if defined(THREADED_RTS)
788 * In the threaded RTS, we only check for deadlock if there
789 * has been no activity in a complete timeslice. This means
790 * we won't eagerly start a full GC just because we don't have
791 * any threads to run currently.
793 if (recent_activity != ACTIVITY_INACTIVE) return;
796 IF_DEBUG(scheduler, sched_belch("deadlocked, forcing major GC..."));
798 // Garbage collection can release some new threads due to
799 // either (a) finalizers or (b) threads resurrected because
800 // they are unreachable and will therefore be sent an
801 // exception. Any threads thus released will be immediately
803 scheduleDoGC( cap, task, rtsTrue/*force major GC*/ );
804 recent_activity = ACTIVITY_DONE_GC;
806 if ( !emptyRunQueue(cap) ) return;
808 #if defined(RTS_USER_SIGNALS) && !defined(THREADED_RTS)
809 /* If we have user-installed signal handlers, then wait
810 * for signals to arrive rather then bombing out with a
813 if ( anyUserHandlers() ) {
815 sched_belch("still deadlocked, waiting for signals..."));
819 if (signals_pending()) {
820 startSignalHandlers();
823 // either we have threads to run, or we were interrupted:
824 ASSERT(!emptyRunQueue(cap) || interrupted);
828 #if !defined(THREADED_RTS)
829 /* Probably a real deadlock. Send the current main thread the
830 * Deadlock exception.
833 switch (task->tso->why_blocked) {
835 case BlockedOnBlackHole:
836 case BlockedOnException:
838 raiseAsync(cap, task->tso, (StgClosure *)NonTermination_closure);
841 barf("deadlock: main thread blocked in a strange way");
849 /* ----------------------------------------------------------------------------
850 * Process an event (GRAN only)
851 * ------------------------------------------------------------------------- */
855 scheduleProcessEvent(rtsEvent *event)
859 if (RtsFlags.GranFlags.Light)
860 GranSimLight_enter_system(event, &ActiveTSO); // adjust ActiveTSO etc
862 /* adjust time based on time-stamp */
863 if (event->time > CurrentTime[CurrentProc] &&
864 event->evttype != ContinueThread)
865 CurrentTime[CurrentProc] = event->time;
867 /* Deal with the idle PEs (may issue FindWork or MoveSpark events) */
868 if (!RtsFlags.GranFlags.Light)
871 IF_DEBUG(gran, debugBelch("GRAN: switch by event-type\n"));
873 /* main event dispatcher in GranSim */
874 switch (event->evttype) {
875 /* Should just be continuing execution */
877 IF_DEBUG(gran, debugBelch("GRAN: doing ContinueThread\n"));
878 /* ToDo: check assertion
879 ASSERT(run_queue_hd != (StgTSO*)NULL &&
880 run_queue_hd != END_TSO_QUEUE);
882 /* Ignore ContinueThreads for fetching threads (if synchr comm) */
883 if (!RtsFlags.GranFlags.DoAsyncFetch &&
884 procStatus[CurrentProc]==Fetching) {
885 debugBelch("ghuH: Spurious ContinueThread while Fetching ignored; TSO %d (%p) [PE %d]\n",
886 CurrentTSO->id, CurrentTSO, CurrentProc);
889 /* Ignore ContinueThreads for completed threads */
890 if (CurrentTSO->what_next == ThreadComplete) {
891 debugBelch("ghuH: found a ContinueThread event for completed thread %d (%p) [PE %d] (ignoring ContinueThread)\n",
892 CurrentTSO->id, CurrentTSO, CurrentProc);
895 /* Ignore ContinueThreads for threads that are being migrated */
896 if (PROCS(CurrentTSO)==Nowhere) {
897 debugBelch("ghuH: trying to run the migrating TSO %d (%p) [PE %d] (ignoring ContinueThread)\n",
898 CurrentTSO->id, CurrentTSO, CurrentProc);
901 /* The thread should be at the beginning of the run queue */
902 if (CurrentTSO!=run_queue_hds[CurrentProc]) {
903 debugBelch("ghuH: TSO %d (%p) [PE %d] is not at the start of the run_queue when doing a ContinueThread\n",
904 CurrentTSO->id, CurrentTSO, CurrentProc);
905 break; // run the thread anyway
908 new_event(proc, proc, CurrentTime[proc],
910 (StgTSO*)NULL, (StgClosure*)NULL, (rtsSpark*)NULL);
912 */ /* Catches superfluous CONTINUEs -- should be unnecessary */
913 break; // now actually run the thread; DaH Qu'vam yImuHbej
916 do_the_fetchnode(event);
917 goto next_thread; /* handle next event in event queue */
920 do_the_globalblock(event);
921 goto next_thread; /* handle next event in event queue */
924 do_the_fetchreply(event);
925 goto next_thread; /* handle next event in event queue */
927 case UnblockThread: /* Move from the blocked queue to the tail of */
928 do_the_unblock(event);
929 goto next_thread; /* handle next event in event queue */
931 case ResumeThread: /* Move from the blocked queue to the tail of */
932 /* the runnable queue ( i.e. Qu' SImqa'lu') */
933 event->tso->gran.blocktime +=
934 CurrentTime[CurrentProc] - event->tso->gran.blockedat;
935 do_the_startthread(event);
936 goto next_thread; /* handle next event in event queue */
939 do_the_startthread(event);
940 goto next_thread; /* handle next event in event queue */
943 do_the_movethread(event);
944 goto next_thread; /* handle next event in event queue */
947 do_the_movespark(event);
948 goto next_thread; /* handle next event in event queue */
951 do_the_findwork(event);
952 goto next_thread; /* handle next event in event queue */
955 barf("Illegal event type %u\n", event->evttype);
958 /* This point was scheduler_loop in the old RTS */
960 IF_DEBUG(gran, debugBelch("GRAN: after main switch\n"));
962 TimeOfLastEvent = CurrentTime[CurrentProc];
963 TimeOfNextEvent = get_time_of_next_event();
964 IgnoreEvents=(TimeOfNextEvent==0); // HWL HACK
965 // CurrentTSO = ThreadQueueHd;
967 IF_DEBUG(gran, debugBelch("GRAN: time of next event is: %ld\n",
970 if (RtsFlags.GranFlags.Light)
971 GranSimLight_leave_system(event, &ActiveTSO);
973 EndOfTimeSlice = CurrentTime[CurrentProc]+RtsFlags.GranFlags.time_slice;
976 debugBelch("GRAN: end of time-slice is %#lx\n", EndOfTimeSlice));
978 /* in a GranSim setup the TSO stays on the run queue */
980 /* Take a thread from the run queue. */
981 POP_RUN_QUEUE(t); // take_off_run_queue(t);
984 debugBelch("GRAN: About to run current thread, which is\n");
987 context_switch = 0; // turned on via GranYield, checking events and time slice
990 DumpGranEvent(GR_SCHEDULE, t));
992 procStatus[CurrentProc] = Busy;
996 /* ----------------------------------------------------------------------------
997 * Send pending messages (PARALLEL_HASKELL only)
998 * ------------------------------------------------------------------------- */
1000 #if defined(PARALLEL_HASKELL)
1002 scheduleSendPendingMessages(void)
1008 # if defined(PAR) // global Mem.Mgmt., omit for now
1009 if (PendingFetches != END_BF_QUEUE) {
1014 if (RtsFlags.ParFlags.BufferTime) {
1015 // if we use message buffering, we must send away all message
1016 // packets which have become too old...
1022 /* ----------------------------------------------------------------------------
1023 * Activate spark threads (PARALLEL_HASKELL only)
1024 * ------------------------------------------------------------------------- */
1026 #if defined(PARALLEL_HASKELL)
1028 scheduleActivateSpark(void)
1031 ASSERT(emptyRunQueue());
1032 /* We get here if the run queue is empty and want some work.
1033 We try to turn a spark into a thread, and add it to the run queue,
1034 from where it will be picked up in the next iteration of the scheduler
1038 /* :-[ no local threads => look out for local sparks */
1039 /* the spark pool for the current PE */
1040 pool = &(cap.r.rSparks); // JB: cap = (old) MainCap
1041 if (advisory_thread_count < RtsFlags.ParFlags.maxThreads &&
1042 pool->hd < pool->tl) {
1044 * ToDo: add GC code check that we really have enough heap afterwards!!
1046 * If we're here (no runnable threads) and we have pending
1047 * sparks, we must have a space problem. Get enough space
1048 * to turn one of those pending sparks into a
1052 spark = findSpark(rtsFalse); /* get a spark */
1053 if (spark != (rtsSpark) NULL) {
1054 tso = createThreadFromSpark(spark); /* turn the spark into a thread */
1055 IF_PAR_DEBUG(fish, // schedule,
1056 debugBelch("==== schedule: Created TSO %d (%p); %d threads active\n",
1057 tso->id, tso, advisory_thread_count));
1059 if (tso==END_TSO_QUEUE) { /* failed to activate spark->back to loop */
1060 IF_PAR_DEBUG(fish, // schedule,
1061 debugBelch("==^^ failed to create thread from spark @ %lx\n",
1063 return rtsFalse; /* failed to generate a thread */
1064 } /* otherwise fall through & pick-up new tso */
1066 IF_PAR_DEBUG(fish, // schedule,
1067 debugBelch("==^^ no local sparks (spark pool contains only NFs: %d)\n",
1068 spark_queue_len(pool)));
1069 return rtsFalse; /* failed to generate a thread */
1071 return rtsTrue; /* success in generating a thread */
1072 } else { /* no more threads permitted or pool empty */
1073 return rtsFalse; /* failed to generateThread */
1076 tso = NULL; // avoid compiler warning only
1077 return rtsFalse; /* dummy in non-PAR setup */
1080 #endif // PARALLEL_HASKELL
1082 /* ----------------------------------------------------------------------------
1083 * Get work from a remote node (PARALLEL_HASKELL only)
1084 * ------------------------------------------------------------------------- */
1086 #if defined(PARALLEL_HASKELL)
1088 scheduleGetRemoteWork(rtsBool *receivedFinish)
1090 ASSERT(emptyRunQueue());
1092 if (RtsFlags.ParFlags.BufferTime) {
1093 IF_PAR_DEBUG(verbose,
1094 debugBelch("...send all pending data,"));
1097 for (i=1; i<=nPEs; i++)
1098 sendImmediately(i); // send all messages away immediately
1102 //++EDEN++ idle() , i.e. send all buffers, wait for work
1103 // suppress fishing in EDEN... just look for incoming messages
1104 // (blocking receive)
1105 IF_PAR_DEBUG(verbose,
1106 debugBelch("...wait for incoming messages...\n"));
1107 *receivedFinish = processMessages(); // blocking receive...
1109 // and reenter scheduling loop after having received something
1110 // (return rtsFalse below)
1112 # else /* activate SPARKS machinery */
1113 /* We get here, if we have no work, tried to activate a local spark, but still
1114 have no work. We try to get a remote spark, by sending a FISH message.
1115 Thread migration should be added here, and triggered when a sequence of
1116 fishes returns without work. */
1117 delay = (RtsFlags.ParFlags.fishDelay!=0ll ? RtsFlags.ParFlags.fishDelay : 0ll);
1119 /* =8-[ no local sparks => look for work on other PEs */
1121 * We really have absolutely no work. Send out a fish
1122 * (there may be some out there already), and wait for
1123 * something to arrive. We clearly can't run any threads
1124 * until a SCHEDULE or RESUME arrives, and so that's what
1125 * we're hoping to see. (Of course, we still have to
1126 * respond to other types of messages.)
1128 rtsTime now = msTime() /*CURRENT_TIME*/;
1129 IF_PAR_DEBUG(verbose,
1130 debugBelch("-- now=%ld\n", now));
1131 IF_PAR_DEBUG(fish, // verbose,
1132 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
1133 (last_fish_arrived_at!=0 &&
1134 last_fish_arrived_at+delay > now)) {
1135 debugBelch("--$$ <%llu> delaying FISH until %llu (last fish %llu, delay %llu)\n",
1136 now, last_fish_arrived_at+delay,
1137 last_fish_arrived_at,
1141 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
1142 advisory_thread_count < RtsFlags.ParFlags.maxThreads) { // send a FISH, but when?
1143 if (last_fish_arrived_at==0 ||
1144 (last_fish_arrived_at+delay <= now)) { // send FISH now!
1145 /* outstandingFishes is set in sendFish, processFish;
1146 avoid flooding system with fishes via delay */
1147 next_fish_to_send_at = 0;
1149 /* ToDo: this should be done in the main scheduling loop to avoid the
1150 busy wait here; not so bad if fish delay is very small */
1151 int iq = 0; // DEBUGGING -- HWL
1152 next_fish_to_send_at = last_fish_arrived_at+delay; // remember when to send
1153 /* send a fish when ready, but process messages that arrive in the meantime */
1155 if (PacketsWaiting()) {
1157 *receivedFinish = processMessages();
1160 } while (!*receivedFinish || now<next_fish_to_send_at);
1161 // JB: This means the fish could become obsolete, if we receive
1162 // work. Better check for work again?
1163 // last line: while (!receivedFinish || !haveWork || now<...)
1164 // next line: if (receivedFinish || haveWork )
1166 if (*receivedFinish) // no need to send a FISH if we are finishing anyway
1167 return rtsFalse; // NB: this will leave scheduler loop
1168 // immediately after return!
1170 IF_PAR_DEBUG(fish, // verbose,
1171 debugBelch("--$$ <%llu> sent delayed fish (%d processMessages); active/total threads=%d/%d\n",now,iq,run_queue_len(),advisory_thread_count));
1175 // JB: IMHO, this should all be hidden inside sendFish(...)
1177 sendFish(pe, thisPE, NEW_FISH_AGE, NEW_FISH_HISTORY,
1180 // Global statistics: count no. of fishes
1181 if (RtsFlags.ParFlags.ParStats.Global &&
1182 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
1183 globalParStats.tot_fish_mess++;
1187 /* delayed fishes must have been sent by now! */
1188 next_fish_to_send_at = 0;
1191 *receivedFinish = processMessages();
1192 # endif /* SPARKS */
1195 /* NB: this function always returns rtsFalse, meaning the scheduler
1196 loop continues with the next iteration;
1198 return code means success in finding work; we enter this function
1199 if there is no local work, thus have to send a fish which takes
1200 time until it arrives with work; in the meantime we should process
1201 messages in the main loop;
1204 #endif // PARALLEL_HASKELL
1206 /* ----------------------------------------------------------------------------
1207 * PAR/GRAN: Report stats & debugging info(?)
1208 * ------------------------------------------------------------------------- */
1210 #if defined(PAR) || defined(GRAN)
1212 scheduleGranParReport(void)
1214 ASSERT(run_queue_hd != END_TSO_QUEUE);
1216 /* Take a thread from the run queue, if we have work */
1217 POP_RUN_QUEUE(t); // take_off_run_queue(END_TSO_QUEUE);
1219 /* If this TSO has got its outport closed in the meantime,
1220 * it mustn't be run. Instead, we have to clean it up as if it was finished.
1221 * It has to be marked as TH_DEAD for this purpose.
1222 * If it is TH_TERM instead, it is supposed to have finished in the normal way.
1224 JB: TODO: investigate wether state change field could be nuked
1225 entirely and replaced by the normal tso state (whatnext
1226 field). All we want to do is to kill tsos from outside.
1229 /* ToDo: write something to the log-file
1230 if (RTSflags.ParFlags.granSimStats && !sameThread)
1231 DumpGranEvent(GR_SCHEDULE, RunnableThreadsHd);
1235 /* the spark pool for the current PE */
1236 pool = &(cap.r.rSparks); // cap = (old) MainCap
1239 debugBelch("--=^ %d threads, %d sparks on [%#x]\n",
1240 run_queue_len(), spark_queue_len(pool), CURRENT_PROC));
1243 debugBelch("--=^ %d threads, %d sparks on [%#x]\n",
1244 run_queue_len(), spark_queue_len(pool), CURRENT_PROC));
1246 if (RtsFlags.ParFlags.ParStats.Full &&
1247 (t->par.sparkname != (StgInt)0) && // only log spark generated threads
1248 (emitSchedule || // forced emit
1249 (t && LastTSO && t->id != LastTSO->id))) {
1251 we are running a different TSO, so write a schedule event to log file
1252 NB: If we use fair scheduling we also have to write a deschedule
1253 event for LastTSO; with unfair scheduling we know that the
1254 previous tso has blocked whenever we switch to another tso, so
1255 we don't need it in GUM for now
1257 IF_PAR_DEBUG(fish, // schedule,
1258 debugBelch("____ scheduling spark generated thread %d (%lx) (%lx) via a forced emit\n",t->id,t,t->par.sparkname));
1260 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
1261 GR_SCHEDULE, t, (StgClosure *)NULL, 0, 0);
1262 emitSchedule = rtsFalse;
1267 /* ----------------------------------------------------------------------------
1268 * After running a thread...
1269 * ------------------------------------------------------------------------- */
1272 schedulePostRunThread(void)
1275 /* HACK 675: if the last thread didn't yield, make sure to print a
1276 SCHEDULE event to the log file when StgRunning the next thread, even
1277 if it is the same one as before */
1279 TimeOfLastYield = CURRENT_TIME;
1282 /* some statistics gathering in the parallel case */
1284 #if defined(GRAN) || defined(PAR) || defined(EDEN)
1288 IF_DEBUG(gran, DumpGranEvent(GR_DESCHEDULE, t));
1289 globalGranStats.tot_heapover++;
1291 globalParStats.tot_heapover++;
1298 DumpGranEvent(GR_DESCHEDULE, t));
1299 globalGranStats.tot_stackover++;
1302 // DumpGranEvent(GR_DESCHEDULE, t);
1303 globalParStats.tot_stackover++;
1307 case ThreadYielding:
1310 DumpGranEvent(GR_DESCHEDULE, t));
1311 globalGranStats.tot_yields++;
1314 // DumpGranEvent(GR_DESCHEDULE, t);
1315 globalParStats.tot_yields++;
1322 debugBelch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: ",
1323 t->id, t, whatNext_strs[t->what_next], t->block_info.closure,
1324 (t->block_info.closure==(StgClosure*)NULL ? 99 : where_is(t->block_info.closure)));
1325 if (t->block_info.closure!=(StgClosure*)NULL)
1326 print_bq(t->block_info.closure);
1329 // ??? needed; should emit block before
1331 DumpGranEvent(GR_DESCHEDULE, t));
1332 prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
1335 ASSERT(procStatus[CurrentProc]==Busy ||
1336 ((procStatus[CurrentProc]==Fetching) &&
1337 (t->block_info.closure!=(StgClosure*)NULL)));
1338 if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
1339 !(!RtsFlags.GranFlags.DoAsyncFetch &&
1340 procStatus[CurrentProc]==Fetching))
1341 procStatus[CurrentProc] = Idle;
1344 //++PAR++ blockThread() writes the event (change?)
1348 case ThreadFinished:
1352 barf("parGlobalStats: unknown return code");
1358 /* -----------------------------------------------------------------------------
1359 * Handle a thread that returned to the scheduler with ThreadHeepOverflow
1360 * -------------------------------------------------------------------------- */
1363 scheduleHandleHeapOverflow( Capability *cap, StgTSO *t )
1365 // did the task ask for a large block?
1366 if (cap->r.rHpAlloc > BLOCK_SIZE) {
1367 // if so, get one and push it on the front of the nursery.
1371 blocks = (lnat)BLOCK_ROUND_UP(cap->r.rHpAlloc) / BLOCK_SIZE;
1374 debugBelch("--<< thread %ld (%s) stopped: requesting a large block (size %ld)\n",
1375 (long)t->id, whatNext_strs[t->what_next], blocks));
1377 // don't do this if the nursery is (nearly) full, we'll GC first.
1378 if (cap->r.rCurrentNursery->link != NULL ||
1379 cap->r.rNursery->n_blocks == 1) { // paranoia to prevent infinite loop
1380 // if the nursery has only one block.
1383 bd = allocGroup( blocks );
1385 cap->r.rNursery->n_blocks += blocks;
1387 // link the new group into the list
1388 bd->link = cap->r.rCurrentNursery;
1389 bd->u.back = cap->r.rCurrentNursery->u.back;
1390 if (cap->r.rCurrentNursery->u.back != NULL) {
1391 cap->r.rCurrentNursery->u.back->link = bd;
1394 ASSERT(g0s0->blocks == cap->r.rCurrentNursery &&
1395 g0s0 == cap->r.rNursery);
1397 cap->r.rNursery->blocks = bd;
1399 cap->r.rCurrentNursery->u.back = bd;
1401 // initialise it as a nursery block. We initialise the
1402 // step, gen_no, and flags field of *every* sub-block in
1403 // this large block, because this is easier than making
1404 // sure that we always find the block head of a large
1405 // block whenever we call Bdescr() (eg. evacuate() and
1406 // isAlive() in the GC would both have to do this, at
1410 for (x = bd; x < bd + blocks; x++) {
1411 x->step = cap->r.rNursery;
1417 // This assert can be a killer if the app is doing lots
1418 // of large block allocations.
1419 IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
1421 // now update the nursery to point to the new block
1422 cap->r.rCurrentNursery = bd;
1424 // we might be unlucky and have another thread get on the
1425 // run queue before us and steal the large block, but in that
1426 // case the thread will just end up requesting another large
1428 pushOnRunQueue(cap,t);
1429 return rtsFalse; /* not actually GC'ing */
1434 debugBelch("--<< thread %ld (%s) stopped: HeapOverflow\n",
1435 (long)t->id, whatNext_strs[t->what_next]));
1437 ASSERT(!is_on_queue(t,CurrentProc));
1438 #elif defined(PARALLEL_HASKELL)
1439 /* Currently we emit a DESCHEDULE event before GC in GUM.
1440 ToDo: either add separate event to distinguish SYSTEM time from rest
1441 or just nuke this DESCHEDULE (and the following SCHEDULE) */
1442 if (0 && RtsFlags.ParFlags.ParStats.Full) {
1443 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
1444 GR_DESCHEDULE, t, (StgClosure *)NULL, 0, 0);
1445 emitSchedule = rtsTrue;
1449 pushOnRunQueue(cap,t);
1451 /* actual GC is done at the end of the while loop in schedule() */
1454 /* -----------------------------------------------------------------------------
1455 * Handle a thread that returned to the scheduler with ThreadStackOverflow
1456 * -------------------------------------------------------------------------- */
1459 scheduleHandleStackOverflow (Capability *cap, Task *task, StgTSO *t)
1461 IF_DEBUG(scheduler,debugBelch("--<< thread %ld (%s) stopped, StackOverflow\n",
1462 (long)t->id, whatNext_strs[t->what_next]));
1463 /* just adjust the stack for this thread, then pop it back
1467 /* enlarge the stack */
1468 StgTSO *new_t = threadStackOverflow(cap, t);
1470 /* This TSO has moved, so update any pointers to it from the
1471 * main thread stack. It better not be on any other queues...
1472 * (it shouldn't be).
1474 if (task->tso != NULL) {
1477 pushOnRunQueue(cap,new_t);
1481 /* -----------------------------------------------------------------------------
1482 * Handle a thread that returned to the scheduler with ThreadYielding
1483 * -------------------------------------------------------------------------- */
1486 scheduleHandleYield( Capability *cap, StgTSO *t, nat prev_what_next )
1488 // Reset the context switch flag. We don't do this just before
1489 // running the thread, because that would mean we would lose ticks
1490 // during GC, which can lead to unfair scheduling (a thread hogs
1491 // the CPU because the tick always arrives during GC). This way
1492 // penalises threads that do a lot of allocation, but that seems
1493 // better than the alternative.
1496 /* put the thread back on the run queue. Then, if we're ready to
1497 * GC, check whether this is the last task to stop. If so, wake
1498 * up the GC thread. getThread will block during a GC until the
1502 if (t->what_next != prev_what_next) {
1503 debugBelch("--<< thread %ld (%s) stopped to switch evaluators\n",
1504 (long)t->id, whatNext_strs[t->what_next]);
1506 debugBelch("--<< thread %ld (%s) stopped, yielding\n",
1507 (long)t->id, whatNext_strs[t->what_next]);
1512 //debugBelch("&& Doing sanity check on yielding TSO %ld.", t->id);
1514 ASSERT(t->link == END_TSO_QUEUE);
1516 // Shortcut if we're just switching evaluators: don't bother
1517 // doing stack squeezing (which can be expensive), just run the
1519 if (t->what_next != prev_what_next) {
1524 ASSERT(!is_on_queue(t,CurrentProc));
1527 //debugBelch("&& Doing sanity check on all ThreadQueues (and their TSOs).");
1528 checkThreadQsSanity(rtsTrue));
1532 addToRunQueue(cap,t);
1535 /* add a ContinueThread event to actually process the thread */
1536 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1538 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1540 debugBelch("GRAN: eventq and runnableq after adding yielded thread to queue again:\n");
1547 /* -----------------------------------------------------------------------------
1548 * Handle a thread that returned to the scheduler with ThreadBlocked
1549 * -------------------------------------------------------------------------- */
1552 scheduleHandleThreadBlocked( StgTSO *t
1553 #if !defined(GRAN) && !defined(DEBUG)
1560 debugBelch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: \n",
1561 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)));
1562 if (t->block_info.closure!=(StgClosure*)NULL) print_bq(t->block_info.closure));
1564 // ??? needed; should emit block before
1566 DumpGranEvent(GR_DESCHEDULE, t));
1567 prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
1570 ASSERT(procStatus[CurrentProc]==Busy ||
1571 ((procStatus[CurrentProc]==Fetching) &&
1572 (t->block_info.closure!=(StgClosure*)NULL)));
1573 if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
1574 !(!RtsFlags.GranFlags.DoAsyncFetch &&
1575 procStatus[CurrentProc]==Fetching))
1576 procStatus[CurrentProc] = Idle;
1580 debugBelch("--<< thread %ld (%p; %s) stopped, blocking on node %p with BQ: \n",
1581 t->id, t, whatNext_strs[t->what_next], t->block_info.closure));
1584 if (t->block_info.closure!=(StgClosure*)NULL)
1585 print_bq(t->block_info.closure));
1587 /* Send a fetch (if BlockedOnGA) and dump event to log file */
1590 /* whatever we schedule next, we must log that schedule */
1591 emitSchedule = rtsTrue;
1595 // We don't need to do anything. The thread is blocked, and it
1596 // has tidied up its stack and placed itself on whatever queue
1597 // it needs to be on.
1600 ASSERT(t->why_blocked != NotBlocked);
1601 // This might not be true under SMP: we don't have
1602 // exclusive access to this TSO, so someone might have
1603 // woken it up by now. This actually happens: try
1604 // conc023 +RTS -N2.
1608 debugBelch("--<< thread %d (%s) stopped: ",
1609 t->id, whatNext_strs[t->what_next]);
1610 printThreadBlockage(t);
1613 /* Only for dumping event to log file
1614 ToDo: do I need this in GranSim, too?
1620 /* -----------------------------------------------------------------------------
1621 * Handle a thread that returned to the scheduler with ThreadFinished
1622 * -------------------------------------------------------------------------- */
1625 scheduleHandleThreadFinished (Capability *cap STG_UNUSED, Task *task, StgTSO *t)
1627 /* Need to check whether this was a main thread, and if so,
1628 * return with the return value.
1630 * We also end up here if the thread kills itself with an
1631 * uncaught exception, see Exception.cmm.
1633 IF_DEBUG(scheduler,debugBelch("--++ thread %d (%s) finished\n",
1634 t->id, whatNext_strs[t->what_next]));
1637 endThread(t, CurrentProc); // clean-up the thread
1638 #elif defined(PARALLEL_HASKELL)
1639 /* For now all are advisory -- HWL */
1640 //if(t->priority==AdvisoryPriority) ??
1641 advisory_thread_count--; // JB: Caution with this counter, buggy!
1644 if(t->dist.priority==RevalPriority)
1648 # if defined(EDENOLD)
1649 // the thread could still have an outport... (BUG)
1650 if (t->eden.outport != -1) {
1651 // delete the outport for the tso which has finished...
1652 IF_PAR_DEBUG(eden_ports,
1653 debugBelch("WARNING: Scheduler removes outport %d for TSO %d.\n",
1654 t->eden.outport, t->id));
1657 // thread still in the process (HEAVY BUG! since outport has just been closed...)
1658 if (t->eden.epid != -1) {
1659 IF_PAR_DEBUG(eden_ports,
1660 debugBelch("WARNING: Scheduler removes TSO %d from process %d .\n",
1661 t->id, t->eden.epid));
1662 removeTSOfromProcess(t);
1667 if (RtsFlags.ParFlags.ParStats.Full &&
1668 !RtsFlags.ParFlags.ParStats.Suppressed)
1669 DumpEndEvent(CURRENT_PROC, t, rtsFalse /* not mandatory */);
1671 // t->par only contains statistics: left out for now...
1673 debugBelch("**** end thread: ended sparked thread %d (%lx); sparkname: %lx\n",
1674 t->id,t,t->par.sparkname));
1676 #endif // PARALLEL_HASKELL
1679 // Check whether the thread that just completed was a bound
1680 // thread, and if so return with the result.
1682 // There is an assumption here that all thread completion goes
1683 // through this point; we need to make sure that if a thread
1684 // ends up in the ThreadKilled state, that it stays on the run
1685 // queue so it can be dealt with here.
1690 if (t->bound != task) {
1691 #if !defined(THREADED_RTS)
1692 // Must be a bound thread that is not the topmost one. Leave
1693 // it on the run queue until the stack has unwound to the
1694 // point where we can deal with this. Leaving it on the run
1695 // queue also ensures that the garbage collector knows about
1696 // this thread and its return value (it gets dropped from the
1697 // all_threads list so there's no other way to find it).
1698 appendToRunQueue(cap,t);
1701 // this cannot happen in the threaded RTS, because a
1702 // bound thread can only be run by the appropriate Task.
1703 barf("finished bound thread that isn't mine");
1707 ASSERT(task->tso == t);
1709 if (t->what_next == ThreadComplete) {
1711 // NOTE: return val is tso->sp[1] (see StgStartup.hc)
1712 *(task->ret) = (StgClosure *)task->tso->sp[1];
1714 task->stat = Success;
1717 *(task->ret) = NULL;
1720 task->stat = Interrupted;
1722 task->stat = Killed;
1726 removeThreadLabel((StgWord)task->tso->id);
1728 return rtsTrue; // tells schedule() to return
1734 /* -----------------------------------------------------------------------------
1735 * Perform a heap census, if PROFILING
1736 * -------------------------------------------------------------------------- */
1739 scheduleDoHeapProfile( rtsBool ready_to_gc STG_UNUSED )
1741 #if defined(PROFILING)
1742 // When we have +RTS -i0 and we're heap profiling, do a census at
1743 // every GC. This lets us get repeatable runs for debugging.
1744 if (performHeapProfile ||
1745 (RtsFlags.ProfFlags.profileInterval==0 &&
1746 RtsFlags.ProfFlags.doHeapProfile && ready_to_gc)) {
1747 GarbageCollect(GetRoots, rtsTrue);
1749 performHeapProfile = rtsFalse;
1750 return rtsTrue; // true <=> we already GC'd
1756 /* -----------------------------------------------------------------------------
1757 * Perform a garbage collection if necessary
1758 * -------------------------------------------------------------------------- */
1761 scheduleDoGC( Capability *cap, Task *task USED_WHEN_SMP, rtsBool force_major )
1765 static volatile StgWord waiting_for_gc;
1766 rtsBool was_waiting;
1771 // In order to GC, there must be no threads running Haskell code.
1772 // Therefore, the GC thread needs to hold *all* the capabilities,
1773 // and release them after the GC has completed.
1775 // This seems to be the simplest way: previous attempts involved
1776 // making all the threads with capabilities give up their
1777 // capabilities and sleep except for the *last* one, which
1778 // actually did the GC. But it's quite hard to arrange for all
1779 // the other tasks to sleep and stay asleep.
1782 was_waiting = cas(&waiting_for_gc, 0, 1);
1783 if (was_waiting) return;
1785 for (i=0; i < n_capabilities; i++) {
1786 IF_DEBUG(scheduler, sched_belch("ready_to_gc, grabbing all the capabilies (%d/%d)", i, n_capabilities));
1787 if (cap != &capabilities[i]) {
1788 Capability *pcap = &capabilities[i];
1789 // we better hope this task doesn't get migrated to
1790 // another Capability while we're waiting for this one.
1791 // It won't, because load balancing happens while we have
1792 // all the Capabilities, but even so it's a slightly
1793 // unsavoury invariant.
1795 waitForReturnCapability(&pcap, task);
1796 if (pcap != &capabilities[i]) {
1797 barf("scheduleDoGC: got the wrong capability");
1802 waiting_for_gc = rtsFalse;
1805 /* Kick any transactions which are invalid back to their
1806 * atomically frames. When next scheduled they will try to
1807 * commit, this commit will fail and they will retry.
1812 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
1813 if (t->what_next == ThreadRelocated) {
1816 next = t->global_link;
1817 if (t -> trec != NO_TREC && t -> why_blocked == NotBlocked) {
1818 if (!stmValidateNestOfTransactions (t -> trec)) {
1819 IF_DEBUG(stm, sched_belch("trec %p found wasting its time", t));
1821 // strip the stack back to the
1822 // ATOMICALLY_FRAME, aborting the (nested)
1823 // transaction, and saving the stack of any
1824 // partially-evaluated thunks on the heap.
1825 raiseAsync_(cap, t, NULL, rtsTrue);
1828 ASSERT(get_itbl((StgClosure *)t->sp)->type == ATOMICALLY_FRAME);
1836 // so this happens periodically:
1837 scheduleCheckBlackHoles(cap);
1839 IF_DEBUG(scheduler, printAllThreads());
1841 /* everybody back, start the GC.
1842 * Could do it in this thread, or signal a condition var
1843 * to do it in another thread. Either way, we need to
1844 * broadcast on gc_pending_cond afterward.
1846 #if defined(THREADED_RTS)
1847 IF_DEBUG(scheduler,sched_belch("doing GC"));
1849 GarbageCollect(GetRoots, force_major);
1852 // release our stash of capabilities.
1853 for (i = 0; i < n_capabilities; i++) {
1854 if (cap != &capabilities[i]) {
1855 task->cap = &capabilities[i];
1856 releaseCapability(&capabilities[i]);
1863 /* add a ContinueThread event to continue execution of current thread */
1864 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1866 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1868 debugBelch("GRAN: eventq and runnableq after Garbage collection:\n\n");
1874 /* ---------------------------------------------------------------------------
1875 * rtsSupportsBoundThreads(): is the RTS built to support bound threads?
1876 * used by Control.Concurrent for error checking.
1877 * ------------------------------------------------------------------------- */
1880 rtsSupportsBoundThreads(void)
1882 #if defined(THREADED_RTS)
1889 /* ---------------------------------------------------------------------------
1890 * isThreadBound(tso): check whether tso is bound to an OS thread.
1891 * ------------------------------------------------------------------------- */
1894 isThreadBound(StgTSO* tso USED_WHEN_THREADED_RTS)
1896 #if defined(THREADED_RTS)
1897 return (tso->bound != NULL);
1902 /* ---------------------------------------------------------------------------
1903 * Singleton fork(). Do not copy any running threads.
1904 * ------------------------------------------------------------------------- */
1906 #if !defined(mingw32_HOST_OS) && !defined(SMP)
1907 #define FORKPROCESS_PRIMOP_SUPPORTED
1910 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
1912 deleteThreadImmediately(Capability *cap, StgTSO *tso);
1915 forkProcess(HsStablePtr *entry
1916 #ifndef FORKPROCESS_PRIMOP_SUPPORTED
1921 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
1927 IF_DEBUG(scheduler,sched_belch("forking!"));
1929 // ToDo: for SMP, we should probably acquire *all* the capabilities
1934 if (pid) { // parent
1936 // just return the pid
1942 // delete all threads
1943 cap->run_queue_hd = END_TSO_QUEUE;
1944 cap->run_queue_tl = END_TSO_QUEUE;
1946 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
1949 // don't allow threads to catch the ThreadKilled exception
1950 deleteThreadImmediately(cap,t);
1953 // wipe the main thread list
1954 while ((task = all_tasks) != NULL) {
1955 all_tasks = task->all_link;
1959 cap = rts_evalStableIO(cap, entry, NULL); // run the action
1960 rts_checkSchedStatus("forkProcess",cap);
1963 hs_exit(); // clean up and exit
1964 stg_exit(EXIT_SUCCESS);
1966 #else /* !FORKPROCESS_PRIMOP_SUPPORTED */
1967 barf("forkProcess#: primop not supported on this platform, sorry!\n");
1972 /* ---------------------------------------------------------------------------
1973 * Delete the threads on the run queue of the current capability.
1974 * ------------------------------------------------------------------------- */
1977 deleteRunQueue (Capability *cap)
1980 for (t = cap->run_queue_hd; t != END_TSO_QUEUE; t = next) {
1981 ASSERT(t->what_next != ThreadRelocated);
1983 deleteThread(cap, t);
1987 /* startThread and insertThread are now in GranSim.c -- HWL */
1990 /* -----------------------------------------------------------------------------
1991 Managing the suspended_ccalling_tasks list.
1992 Locks required: sched_mutex
1993 -------------------------------------------------------------------------- */
1996 suspendTask (Capability *cap, Task *task)
1998 ASSERT(task->next == NULL && task->prev == NULL);
1999 task->next = cap->suspended_ccalling_tasks;
2001 if (cap->suspended_ccalling_tasks) {
2002 cap->suspended_ccalling_tasks->prev = task;
2004 cap->suspended_ccalling_tasks = task;
2008 recoverSuspendedTask (Capability *cap, Task *task)
2011 task->prev->next = task->next;
2013 ASSERT(cap->suspended_ccalling_tasks == task);
2014 cap->suspended_ccalling_tasks = task->next;
2017 task->next->prev = task->prev;
2019 task->next = task->prev = NULL;
2022 /* ---------------------------------------------------------------------------
2023 * Suspending & resuming Haskell threads.
2025 * When making a "safe" call to C (aka _ccall_GC), the task gives back
2026 * its capability before calling the C function. This allows another
2027 * task to pick up the capability and carry on running Haskell
2028 * threads. It also means that if the C call blocks, it won't lock
2031 * The Haskell thread making the C call is put to sleep for the
2032 * duration of the call, on the susepended_ccalling_threads queue. We
2033 * give out a token to the task, which it can use to resume the thread
2034 * on return from the C function.
2035 * ------------------------------------------------------------------------- */
2038 suspendThread (StgRegTable *reg)
2041 int saved_errno = errno;
2045 /* assume that *reg is a pointer to the StgRegTable part of a Capability.
2047 cap = regTableToCapability(reg);
2049 task = cap->running_task;
2050 tso = cap->r.rCurrentTSO;
2053 sched_belch("thread %d did a safe foreign call", cap->r.rCurrentTSO->id));
2055 // XXX this might not be necessary --SDM
2056 tso->what_next = ThreadRunGHC;
2060 if(tso->blocked_exceptions == NULL) {
2061 tso->why_blocked = BlockedOnCCall;
2062 tso->blocked_exceptions = END_TSO_QUEUE;
2064 tso->why_blocked = BlockedOnCCall_NoUnblockExc;
2067 // Hand back capability
2068 task->suspended_tso = tso;
2070 ACQUIRE_LOCK(&cap->lock);
2072 suspendTask(cap,task);
2073 cap->in_haskell = rtsFalse;
2074 releaseCapability_(cap);
2076 RELEASE_LOCK(&cap->lock);
2078 #if defined(THREADED_RTS)
2079 /* Preparing to leave the RTS, so ensure there's a native thread/task
2080 waiting to take over.
2082 IF_DEBUG(scheduler, sched_belch("thread %d: leaving RTS", tso->id));
2085 errno = saved_errno;
2090 resumeThread (void *task_)
2094 int saved_errno = errno;
2098 // Wait for permission to re-enter the RTS with the result.
2099 waitForReturnCapability(&cap,task);
2100 // we might be on a different capability now... but if so, our
2101 // entry on the suspended_ccalling_tasks list will also have been
2104 // Remove the thread from the suspended list
2105 recoverSuspendedTask(cap,task);
2107 tso = task->suspended_tso;
2108 task->suspended_tso = NULL;
2109 tso->link = END_TSO_QUEUE;
2110 IF_DEBUG(scheduler, sched_belch("thread %d: re-entering RTS", tso->id));
2112 if (tso->why_blocked == BlockedOnCCall) {
2113 awakenBlockedQueue(cap,tso->blocked_exceptions);
2114 tso->blocked_exceptions = NULL;
2117 /* Reset blocking status */
2118 tso->why_blocked = NotBlocked;
2120 cap->r.rCurrentTSO = tso;
2121 cap->in_haskell = rtsTrue;
2122 errno = saved_errno;
2127 /* ---------------------------------------------------------------------------
2128 * Comparing Thread ids.
2130 * This is used from STG land in the implementation of the
2131 * instances of Eq/Ord for ThreadIds.
2132 * ------------------------------------------------------------------------ */
2135 cmp_thread(StgPtr tso1, StgPtr tso2)
2137 StgThreadID id1 = ((StgTSO *)tso1)->id;
2138 StgThreadID id2 = ((StgTSO *)tso2)->id;
2140 if (id1 < id2) return (-1);
2141 if (id1 > id2) return 1;
2145 /* ---------------------------------------------------------------------------
2146 * Fetching the ThreadID from an StgTSO.
2148 * This is used in the implementation of Show for ThreadIds.
2149 * ------------------------------------------------------------------------ */
2151 rts_getThreadId(StgPtr tso)
2153 return ((StgTSO *)tso)->id;
2158 labelThread(StgPtr tso, char *label)
2163 /* Caveat: Once set, you can only set the thread name to "" */
2164 len = strlen(label)+1;
2165 buf = stgMallocBytes(len * sizeof(char), "Schedule.c:labelThread()");
2166 strncpy(buf,label,len);
2167 /* Update will free the old memory for us */
2168 updateThreadLabel(((StgTSO *)tso)->id,buf);
2172 /* ---------------------------------------------------------------------------
2173 Create a new thread.
2175 The new thread starts with the given stack size. Before the
2176 scheduler can run, however, this thread needs to have a closure
2177 (and possibly some arguments) pushed on its stack. See
2178 pushClosure() in Schedule.h.
2180 createGenThread() and createIOThread() (in SchedAPI.h) are
2181 convenient packaged versions of this function.
2183 currently pri (priority) is only used in a GRAN setup -- HWL
2184 ------------------------------------------------------------------------ */
2186 /* currently pri (priority) is only used in a GRAN setup -- HWL */
2188 createThread(nat size, StgInt pri)
2191 createThread(Capability *cap, nat size)
2197 /* sched_mutex is *not* required */
2199 /* First check whether we should create a thread at all */
2200 #if defined(PARALLEL_HASKELL)
2201 /* check that no more than RtsFlags.ParFlags.maxThreads threads are created */
2202 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads) {
2204 debugBelch("{createThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)\n",
2205 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
2206 return END_TSO_QUEUE;
2212 ASSERT(!RtsFlags.GranFlags.Light || CurrentProc==0);
2215 // ToDo: check whether size = stack_size - TSO_STRUCT_SIZEW
2217 /* catch ridiculously small stack sizes */
2218 if (size < MIN_STACK_WORDS + TSO_STRUCT_SIZEW) {
2219 size = MIN_STACK_WORDS + TSO_STRUCT_SIZEW;
2222 stack_size = size - TSO_STRUCT_SIZEW;
2224 tso = (StgTSO *)allocateLocal(cap, size);
2225 TICK_ALLOC_TSO(stack_size, 0);
2227 SET_HDR(tso, &stg_TSO_info, CCS_SYSTEM);
2229 SET_GRAN_HDR(tso, ThisPE);
2232 // Always start with the compiled code evaluator
2233 tso->what_next = ThreadRunGHC;
2235 tso->why_blocked = NotBlocked;
2236 tso->blocked_exceptions = NULL;
2238 tso->saved_errno = 0;
2241 tso->stack_size = stack_size;
2242 tso->max_stack_size = round_to_mblocks(RtsFlags.GcFlags.maxStkSize)
2244 tso->sp = (P_)&(tso->stack) + stack_size;
2246 tso->trec = NO_TREC;
2249 tso->prof.CCCS = CCS_MAIN;
2252 /* put a stop frame on the stack */
2253 tso->sp -= sizeofW(StgStopFrame);
2254 SET_HDR((StgClosure*)tso->sp,(StgInfoTable *)&stg_stop_thread_info,CCS_SYSTEM);
2255 tso->link = END_TSO_QUEUE;
2259 /* uses more flexible routine in GranSim */
2260 insertThread(tso, CurrentProc);
2262 /* In a non-GranSim setup the pushing of a TSO onto the runq is separated
2268 if (RtsFlags.GranFlags.GranSimStats.Full)
2269 DumpGranEvent(GR_START,tso);
2270 #elif defined(PARALLEL_HASKELL)
2271 if (RtsFlags.ParFlags.ParStats.Full)
2272 DumpGranEvent(GR_STARTQ,tso);
2273 /* HACk to avoid SCHEDULE
2277 /* Link the new thread on the global thread list.
2279 ACQUIRE_LOCK(&sched_mutex);
2280 tso->id = next_thread_id++; // while we have the mutex
2281 tso->global_link = all_threads;
2283 RELEASE_LOCK(&sched_mutex);
2286 tso->dist.priority = MandatoryPriority; //by default that is...
2290 tso->gran.pri = pri;
2292 tso->gran.magic = TSO_MAGIC; // debugging only
2294 tso->gran.sparkname = 0;
2295 tso->gran.startedat = CURRENT_TIME;
2296 tso->gran.exported = 0;
2297 tso->gran.basicblocks = 0;
2298 tso->gran.allocs = 0;
2299 tso->gran.exectime = 0;
2300 tso->gran.fetchtime = 0;
2301 tso->gran.fetchcount = 0;
2302 tso->gran.blocktime = 0;
2303 tso->gran.blockcount = 0;
2304 tso->gran.blockedat = 0;
2305 tso->gran.globalsparks = 0;
2306 tso->gran.localsparks = 0;
2307 if (RtsFlags.GranFlags.Light)
2308 tso->gran.clock = Now; /* local clock */
2310 tso->gran.clock = 0;
2312 IF_DEBUG(gran,printTSO(tso));
2313 #elif defined(PARALLEL_HASKELL)
2315 tso->par.magic = TSO_MAGIC; // debugging only
2317 tso->par.sparkname = 0;
2318 tso->par.startedat = CURRENT_TIME;
2319 tso->par.exported = 0;
2320 tso->par.basicblocks = 0;
2321 tso->par.allocs = 0;
2322 tso->par.exectime = 0;
2323 tso->par.fetchtime = 0;
2324 tso->par.fetchcount = 0;
2325 tso->par.blocktime = 0;
2326 tso->par.blockcount = 0;
2327 tso->par.blockedat = 0;
2328 tso->par.globalsparks = 0;
2329 tso->par.localsparks = 0;
2333 globalGranStats.tot_threads_created++;
2334 globalGranStats.threads_created_on_PE[CurrentProc]++;
2335 globalGranStats.tot_sq_len += spark_queue_len(CurrentProc);
2336 globalGranStats.tot_sq_probes++;
2337 #elif defined(PARALLEL_HASKELL)
2338 // collect parallel global statistics (currently done together with GC stats)
2339 if (RtsFlags.ParFlags.ParStats.Global &&
2340 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
2341 //debugBelch("Creating thread %d @ %11.2f\n", tso->id, usertime());
2342 globalParStats.tot_threads_created++;
2348 sched_belch("==__ schedule: Created TSO %d (%p);",
2349 CurrentProc, tso, tso->id));
2350 #elif defined(PARALLEL_HASKELL)
2351 IF_PAR_DEBUG(verbose,
2352 sched_belch("==__ schedule: Created TSO %d (%p); %d threads active",
2353 (long)tso->id, tso, advisory_thread_count));
2355 IF_DEBUG(scheduler,sched_belch("created thread %ld, stack size = %lx words",
2356 (long)tso->id, (long)tso->stack_size));
2363 all parallel thread creation calls should fall through the following routine.
2366 createThreadFromSpark(rtsSpark spark)
2368 ASSERT(spark != (rtsSpark)NULL);
2369 // JB: TAKE CARE OF THIS COUNTER! BUGGY
2370 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads)
2372 barf("{createSparkThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
2373 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
2374 return END_TSO_QUEUE;
2378 tso = createThread(RtsFlags.GcFlags.initialStkSize);
2379 if (tso==END_TSO_QUEUE)
2380 barf("createSparkThread: Cannot create TSO");
2382 tso->priority = AdvisoryPriority;
2384 pushClosure(tso,spark);
2386 advisory_thread_count++; // JB: TAKE CARE OF THIS COUNTER! BUGGY
2393 Turn a spark into a thread.
2394 ToDo: fix for SMP (needs to acquire SCHED_MUTEX!)
2398 activateSpark (rtsSpark spark)
2402 tso = createSparkThread(spark);
2403 if (RtsFlags.ParFlags.ParStats.Full) {
2404 //ASSERT(run_queue_hd == END_TSO_QUEUE); // I think ...
2405 IF_PAR_DEBUG(verbose,
2406 debugBelch("==^^ activateSpark: turning spark of closure %p (%s) into a thread\n",
2407 (StgClosure *)spark, info_type((StgClosure *)spark)));
2409 // ToDo: fwd info on local/global spark to thread -- HWL
2410 // tso->gran.exported = spark->exported;
2411 // tso->gran.locked = !spark->global;
2412 // tso->gran.sparkname = spark->name;
2418 /* ---------------------------------------------------------------------------
2421 * scheduleThread puts a thread on the end of the runnable queue.
2422 * This will usually be done immediately after a thread is created.
2423 * The caller of scheduleThread must create the thread using e.g.
2424 * createThread and push an appropriate closure
2425 * on this thread's stack before the scheduler is invoked.
2426 * ------------------------------------------------------------------------ */
2429 scheduleThread(Capability *cap, StgTSO *tso)
2431 // The thread goes at the *end* of the run-queue, to avoid possible
2432 // starvation of any threads already on the queue.
2433 appendToRunQueue(cap,tso);
2437 scheduleWaitThread (StgTSO* tso, /*[out]*/HaskellObj* ret, Capability *cap)
2441 // We already created/initialised the Task
2442 task = cap->running_task;
2444 // This TSO is now a bound thread; make the Task and TSO
2445 // point to each other.
2450 task->stat = NoStatus;
2452 appendToRunQueue(cap,tso);
2454 IF_DEBUG(scheduler, sched_belch("new bound thread (%d)", tso->id));
2457 /* GranSim specific init */
2458 CurrentTSO = m->tso; // the TSO to run
2459 procStatus[MainProc] = Busy; // status of main PE
2460 CurrentProc = MainProc; // PE to run it on
2463 cap = schedule(cap,task);
2465 ASSERT(task->stat != NoStatus);
2466 ASSERT(cap->running_task == task);
2467 ASSERT(task->cap == cap);
2469 IF_DEBUG(scheduler, sched_belch("bound thread (%d) finished", task->tso->id));
2473 /* ----------------------------------------------------------------------------
2475 * ------------------------------------------------------------------------- */
2477 #if defined(THREADED_RTS)
2479 workerStart(Task *task)
2483 // See startWorkerTask().
2484 ACQUIRE_LOCK(&task->lock);
2486 RELEASE_LOCK(&task->lock);
2488 // set the thread-local pointer to the Task:
2491 // schedule() runs without a lock.
2492 cap = schedule(cap,task);
2494 // On exit from schedule(), we have a Capability.
2495 releaseCapability(cap);
2500 /* ---------------------------------------------------------------------------
2503 * Initialise the scheduler. This resets all the queues - if the
2504 * queues contained any threads, they'll be garbage collected at the
2507 * ------------------------------------------------------------------------ */
2514 for (i=0; i<=MAX_PROC; i++) {
2515 run_queue_hds[i] = END_TSO_QUEUE;
2516 run_queue_tls[i] = END_TSO_QUEUE;
2517 blocked_queue_hds[i] = END_TSO_QUEUE;
2518 blocked_queue_tls[i] = END_TSO_QUEUE;
2519 ccalling_threadss[i] = END_TSO_QUEUE;
2520 blackhole_queue[i] = END_TSO_QUEUE;
2521 sleeping_queue = END_TSO_QUEUE;
2523 #elif !defined(THREADED_RTS)
2524 blocked_queue_hd = END_TSO_QUEUE;
2525 blocked_queue_tl = END_TSO_QUEUE;
2526 sleeping_queue = END_TSO_QUEUE;
2529 blackhole_queue = END_TSO_QUEUE;
2530 all_threads = END_TSO_QUEUE;
2535 RtsFlags.ConcFlags.ctxtSwitchTicks =
2536 RtsFlags.ConcFlags.ctxtSwitchTime / TICK_MILLISECS;
2538 #if defined(THREADED_RTS)
2539 /* Initialise the mutex and condition variables used by
2541 initMutex(&sched_mutex);
2544 ACQUIRE_LOCK(&sched_mutex);
2546 /* A capability holds the state a native thread needs in
2547 * order to execute STG code. At least one capability is
2548 * floating around (only SMP builds have more than one).
2556 * Eagerly start one worker to run each Capability, except for
2557 * Capability 0. The idea is that we're probably going to start a
2558 * bound thread on Capability 0 pretty soon, so we don't want a
2559 * worker task hogging it.
2564 for (i = 1; i < n_capabilities; i++) {
2565 cap = &capabilities[i];
2566 ACQUIRE_LOCK(&cap->lock);
2567 startWorkerTask(cap, workerStart);
2568 RELEASE_LOCK(&cap->lock);
2573 #if /* defined(SMP) ||*/ defined(PARALLEL_HASKELL)
2577 RELEASE_LOCK(&sched_mutex);
2581 exitScheduler( void )
2583 interrupted = rtsTrue;
2584 shutting_down_scheduler = rtsTrue;
2586 #if defined(THREADED_RTS)
2591 ACQUIRE_LOCK(&sched_mutex);
2592 task = newBoundTask();
2593 RELEASE_LOCK(&sched_mutex);
2595 for (i = 0; i < n_capabilities; i++) {
2596 shutdownCapability(&capabilities[i], task);
2598 boundTaskExiting(task);
2604 /* ---------------------------------------------------------------------------
2605 Where are the roots that we know about?
2607 - all the threads on the runnable queue
2608 - all the threads on the blocked queue
2609 - all the threads on the sleeping queue
2610 - all the thread currently executing a _ccall_GC
2611 - all the "main threads"
2613 ------------------------------------------------------------------------ */
2615 /* This has to be protected either by the scheduler monitor, or by the
2616 garbage collection monitor (probably the latter).
2621 GetRoots( evac_fn evac )
2628 for (i=0; i<=RtsFlags.GranFlags.proc; i++) {
2629 if ((run_queue_hds[i] != END_TSO_QUEUE) && ((run_queue_hds[i] != NULL)))
2630 evac((StgClosure **)&run_queue_hds[i]);
2631 if ((run_queue_tls[i] != END_TSO_QUEUE) && ((run_queue_tls[i] != NULL)))
2632 evac((StgClosure **)&run_queue_tls[i]);
2634 if ((blocked_queue_hds[i] != END_TSO_QUEUE) && ((blocked_queue_hds[i] != NULL)))
2635 evac((StgClosure **)&blocked_queue_hds[i]);
2636 if ((blocked_queue_tls[i] != END_TSO_QUEUE) && ((blocked_queue_tls[i] != NULL)))
2637 evac((StgClosure **)&blocked_queue_tls[i]);
2638 if ((ccalling_threadss[i] != END_TSO_QUEUE) && ((ccalling_threadss[i] != NULL)))
2639 evac((StgClosure **)&ccalling_threads[i]);
2646 for (i = 0; i < n_capabilities; i++) {
2647 cap = &capabilities[i];
2648 evac((StgClosure **)&cap->run_queue_hd);
2649 evac((StgClosure **)&cap->run_queue_tl);
2651 for (task = cap->suspended_ccalling_tasks; task != NULL;
2653 evac((StgClosure **)&task->suspended_tso);
2657 #if !defined(THREADED_RTS)
2658 evac((StgClosure **)&blocked_queue_hd);
2659 evac((StgClosure **)&blocked_queue_tl);
2660 evac((StgClosure **)&sleeping_queue);
2664 evac((StgClosure **)&blackhole_queue);
2666 #if defined(PARALLEL_HASKELL) || defined(GRAN)
2667 markSparkQueue(evac);
2670 #if defined(RTS_USER_SIGNALS)
2671 // mark the signal handlers (signals should be already blocked)
2672 markSignalHandlers(evac);
2676 /* -----------------------------------------------------------------------------
2679 This is the interface to the garbage collector from Haskell land.
2680 We provide this so that external C code can allocate and garbage
2681 collect when called from Haskell via _ccall_GC.
2683 It might be useful to provide an interface whereby the programmer
2684 can specify more roots (ToDo).
2686 This needs to be protected by the GC condition variable above. KH.
2687 -------------------------------------------------------------------------- */
2689 static void (*extra_roots)(evac_fn);
2695 // ToDo: we have to grab all the capabilities here.
2696 errorBelch("performGC not supported in threaded RTS (yet)");
2697 stg_exit(EXIT_FAILURE);
2699 /* Obligated to hold this lock upon entry */
2700 GarbageCollect(GetRoots,rtsFalse);
2704 performMajorGC(void)
2707 errorBelch("performMayjorGC not supported in threaded RTS (yet)");
2708 stg_exit(EXIT_FAILURE);
2710 GarbageCollect(GetRoots,rtsTrue);
2714 AllRoots(evac_fn evac)
2716 GetRoots(evac); // the scheduler's roots
2717 extra_roots(evac); // the user's roots
2721 performGCWithRoots(void (*get_roots)(evac_fn))
2724 errorBelch("performGCWithRoots not supported in threaded RTS (yet)");
2725 stg_exit(EXIT_FAILURE);
2727 extra_roots = get_roots;
2728 GarbageCollect(AllRoots,rtsFalse);
2731 /* -----------------------------------------------------------------------------
2734 If the thread has reached its maximum stack size, then raise the
2735 StackOverflow exception in the offending thread. Otherwise
2736 relocate the TSO into a larger chunk of memory and adjust its stack
2738 -------------------------------------------------------------------------- */
2741 threadStackOverflow(Capability *cap, StgTSO *tso)
2743 nat new_stack_size, stack_words;
2748 IF_DEBUG(sanity,checkTSO(tso));
2749 if (tso->stack_size >= tso->max_stack_size) {
2752 debugBelch("@@ threadStackOverflow of TSO %ld (%p): stack too large (now %ld; max is %ld)\n",
2753 (long)tso->id, tso, (long)tso->stack_size, (long)tso->max_stack_size);
2754 /* If we're debugging, just print out the top of the stack */
2755 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2758 /* Send this thread the StackOverflow exception */
2759 raiseAsync(cap, tso, (StgClosure *)stackOverflow_closure);
2763 /* Try to double the current stack size. If that takes us over the
2764 * maximum stack size for this thread, then use the maximum instead.
2765 * Finally round up so the TSO ends up as a whole number of blocks.
2767 new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
2768 new_tso_size = (lnat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) +
2769 TSO_STRUCT_SIZE)/sizeof(W_);
2770 new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */
2771 new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
2773 IF_DEBUG(scheduler, sched_belch("increasing stack size from %ld words to %d.\n", tso->stack_size, new_stack_size));
2775 dest = (StgTSO *)allocate(new_tso_size);
2776 TICK_ALLOC_TSO(new_stack_size,0);
2778 /* copy the TSO block and the old stack into the new area */
2779 memcpy(dest,tso,TSO_STRUCT_SIZE);
2780 stack_words = tso->stack + tso->stack_size - tso->sp;
2781 new_sp = (P_)dest + new_tso_size - stack_words;
2782 memcpy(new_sp, tso->sp, stack_words * sizeof(W_));
2784 /* relocate the stack pointers... */
2786 dest->stack_size = new_stack_size;
2788 /* Mark the old TSO as relocated. We have to check for relocated
2789 * TSOs in the garbage collector and any primops that deal with TSOs.
2791 * It's important to set the sp value to just beyond the end
2792 * of the stack, so we don't attempt to scavenge any part of the
2795 tso->what_next = ThreadRelocated;
2797 tso->sp = (P_)&(tso->stack[tso->stack_size]);
2798 tso->why_blocked = NotBlocked;
2800 IF_PAR_DEBUG(verbose,
2801 debugBelch("@@ threadStackOverflow of TSO %d (now at %p): stack size increased to %ld\n",
2802 tso->id, tso, tso->stack_size);
2803 /* If we're debugging, just print out the top of the stack */
2804 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2807 IF_DEBUG(sanity,checkTSO(tso));
2809 IF_DEBUG(scheduler,printTSO(dest));
2815 /* ---------------------------------------------------------------------------
2816 Wake up a queue that was blocked on some resource.
2817 ------------------------------------------------------------------------ */
2821 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2824 #elif defined(PARALLEL_HASKELL)
2826 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2828 /* write RESUME events to log file and
2829 update blocked and fetch time (depending on type of the orig closure) */
2830 if (RtsFlags.ParFlags.ParStats.Full) {
2831 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
2832 GR_RESUMEQ, ((StgTSO *)bqe), ((StgTSO *)bqe)->block_info.closure,
2833 0, 0 /* spark_queue_len(ADVISORY_POOL) */);
2834 if (emptyRunQueue())
2835 emitSchedule = rtsTrue;
2837 switch (get_itbl(node)->type) {
2839 ((StgTSO *)bqe)->par.fetchtime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2844 ((StgTSO *)bqe)->par.blocktime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2851 barf("{unblockOne}Daq Qagh: unexpected closure in blocking queue");
2858 StgBlockingQueueElement *
2859 unblockOne(StgBlockingQueueElement *bqe, StgClosure *node)
2862 PEs node_loc, tso_loc;
2864 node_loc = where_is(node); // should be lifted out of loop
2865 tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2866 tso_loc = where_is((StgClosure *)tso);
2867 if (IS_LOCAL_TO(PROCS(node),tso_loc)) { // TSO is local
2868 /* !fake_fetch => TSO is on CurrentProc is same as IS_LOCAL_TO */
2869 ASSERT(CurrentProc!=node_loc || tso_loc==CurrentProc);
2870 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.lunblocktime;
2871 // insertThread(tso, node_loc);
2872 new_event(tso_loc, tso_loc, CurrentTime[CurrentProc],
2874 tso, node, (rtsSpark*)NULL);
2875 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2878 } else { // TSO is remote (actually should be FMBQ)
2879 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.mpacktime +
2880 RtsFlags.GranFlags.Costs.gunblocktime +
2881 RtsFlags.GranFlags.Costs.latency;
2882 new_event(tso_loc, CurrentProc, CurrentTime[CurrentProc],
2884 tso, node, (rtsSpark*)NULL);
2885 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2888 /* the thread-queue-overhead is accounted for in either Resume or UnblockThread */
2890 debugBelch(" %s TSO %d (%p) [PE %d] (block_info.closure=%p) (next=%p) ,",
2891 (node_loc==tso_loc ? "Local" : "Global"),
2892 tso->id, tso, CurrentProc, tso->block_info.closure, tso->link));
2893 tso->block_info.closure = NULL;
2894 IF_DEBUG(scheduler,debugBelch("-- Waking up thread %ld (%p)\n",
2897 #elif defined(PARALLEL_HASKELL)
2898 StgBlockingQueueElement *
2899 unblockOne(StgBlockingQueueElement *bqe, StgClosure *node)
2901 StgBlockingQueueElement *next;
2903 switch (get_itbl(bqe)->type) {
2905 ASSERT(((StgTSO *)bqe)->why_blocked != NotBlocked);
2906 /* if it's a TSO just push it onto the run_queue */
2908 ((StgTSO *)bqe)->link = END_TSO_QUEUE; // debugging?
2909 APPEND_TO_RUN_QUEUE((StgTSO *)bqe);
2911 unblockCount(bqe, node);
2912 /* reset blocking status after dumping event */
2913 ((StgTSO *)bqe)->why_blocked = NotBlocked;
2917 /* if it's a BLOCKED_FETCH put it on the PendingFetches list */
2919 bqe->link = (StgBlockingQueueElement *)PendingFetches;
2920 PendingFetches = (StgBlockedFetch *)bqe;
2924 /* can ignore this case in a non-debugging setup;
2925 see comments on RBHSave closures above */
2927 /* check that the closure is an RBHSave closure */
2928 ASSERT(get_itbl((StgClosure *)bqe) == &stg_RBH_Save_0_info ||
2929 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_1_info ||
2930 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_2_info);
2934 barf("{unblockOne}Daq Qagh: Unexpected IP (%#lx; %s) in blocking queue at %#lx\n",
2935 get_itbl((StgClosure *)bqe), info_type((StgClosure *)bqe),
2939 IF_PAR_DEBUG(bq, debugBelch(", %p (%s)\n", bqe, info_type((StgClosure*)bqe)));
2945 unblockOne(Capability *cap, StgTSO *tso)
2949 ASSERT(get_itbl(tso)->type == TSO);
2950 ASSERT(tso->why_blocked != NotBlocked);
2951 tso->why_blocked = NotBlocked;
2953 tso->link = END_TSO_QUEUE;
2955 // We might have just migrated this TSO to our Capability:
2957 tso->bound->cap = cap;
2960 appendToRunQueue(cap,tso);
2962 // we're holding a newly woken thread, make sure we context switch
2963 // quickly so we can migrate it if necessary.
2965 IF_DEBUG(scheduler,sched_belch("waking up thread %ld", (long)tso->id));
2972 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2974 StgBlockingQueueElement *bqe;
2979 debugBelch("##-_ AwBQ for node %p on PE %d @ %ld by TSO %d (%p): \n", \
2980 node, CurrentProc, CurrentTime[CurrentProc],
2981 CurrentTSO->id, CurrentTSO));
2983 node_loc = where_is(node);
2985 ASSERT(q == END_BQ_QUEUE ||
2986 get_itbl(q)->type == TSO || // q is either a TSO or an RBHSave
2987 get_itbl(q)->type == CONSTR); // closure (type constructor)
2988 ASSERT(is_unique(node));
2990 /* FAKE FETCH: magically copy the node to the tso's proc;
2991 no Fetch necessary because in reality the node should not have been
2992 moved to the other PE in the first place
2994 if (CurrentProc!=node_loc) {
2996 debugBelch("## node %p is on PE %d but CurrentProc is %d (TSO %d); assuming fake fetch and adjusting bitmask (old: %#x)\n",
2997 node, node_loc, CurrentProc, CurrentTSO->id,
2998 // CurrentTSO, where_is(CurrentTSO),
2999 node->header.gran.procs));
3000 node->header.gran.procs = (node->header.gran.procs) | PE_NUMBER(CurrentProc);
3002 debugBelch("## new bitmask of node %p is %#x\n",
3003 node, node->header.gran.procs));
3004 if (RtsFlags.GranFlags.GranSimStats.Global) {
3005 globalGranStats.tot_fake_fetches++;
3010 // ToDo: check: ASSERT(CurrentProc==node_loc);
3011 while (get_itbl(bqe)->type==TSO) { // q != END_TSO_QUEUE) {
3014 bqe points to the current element in the queue
3015 next points to the next element in the queue
3017 //tso = (StgTSO *)bqe; // wastes an assignment to get the type right
3018 //tso_loc = where_is(tso);
3020 bqe = unblockOne(bqe, node);
3023 /* if this is the BQ of an RBH, we have to put back the info ripped out of
3024 the closure to make room for the anchor of the BQ */
3025 if (bqe!=END_BQ_QUEUE) {
3026 ASSERT(get_itbl(node)->type == RBH && get_itbl(bqe)->type == CONSTR);
3028 ASSERT((info_ptr==&RBH_Save_0_info) ||
3029 (info_ptr==&RBH_Save_1_info) ||
3030 (info_ptr==&RBH_Save_2_info));
3032 /* cf. convertToRBH in RBH.c for writing the RBHSave closure */
3033 ((StgRBH *)node)->blocking_queue = (StgBlockingQueueElement *)((StgRBHSave *)bqe)->payload[0];
3034 ((StgRBH *)node)->mut_link = (StgMutClosure *)((StgRBHSave *)bqe)->payload[1];
3037 debugBelch("## Filled in RBH_Save for %p (%s) at end of AwBQ\n",
3038 node, info_type(node)));
3041 /* statistics gathering */
3042 if (RtsFlags.GranFlags.GranSimStats.Global) {
3043 // globalGranStats.tot_bq_processing_time += bq_processing_time;
3044 globalGranStats.tot_bq_len += len; // total length of all bqs awakened
3045 // globalGranStats.tot_bq_len_local += len_local; // same for local TSOs only
3046 globalGranStats.tot_awbq++; // total no. of bqs awakened
3049 debugBelch("## BQ Stats of %p: [%d entries] %s\n",
3050 node, len, (bqe!=END_BQ_QUEUE) ? "RBH" : ""));
3052 #elif defined(PARALLEL_HASKELL)
3054 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
3056 StgBlockingQueueElement *bqe;
3058 IF_PAR_DEBUG(verbose,
3059 debugBelch("##-_ AwBQ for node %p on [%x]: \n",
3063 if(get_itbl(q)->type == CONSTR || q==END_BQ_QUEUE) {
3064 IF_PAR_DEBUG(verbose, debugBelch("## ... nothing to unblock so lets just return. RFP (BUG?)\n"));
3069 ASSERT(q == END_BQ_QUEUE ||
3070 get_itbl(q)->type == TSO ||
3071 get_itbl(q)->type == BLOCKED_FETCH ||
3072 get_itbl(q)->type == CONSTR);
3075 while (get_itbl(bqe)->type==TSO ||
3076 get_itbl(bqe)->type==BLOCKED_FETCH) {
3077 bqe = unblockOne(bqe, node);
3081 #else /* !GRAN && !PARALLEL_HASKELL */
3084 awakenBlockedQueue(Capability *cap, StgTSO *tso)
3086 if (tso == NULL) return; // hack; see bug #1235728, and comments in
3088 while (tso != END_TSO_QUEUE) {
3089 tso = unblockOne(cap,tso);
3094 /* ---------------------------------------------------------------------------
3096 - usually called inside a signal handler so it mustn't do anything fancy.
3097 ------------------------------------------------------------------------ */
3100 interruptStgRts(void)
3104 #if defined(THREADED_RTS)
3105 prodAllCapabilities();
3109 /* -----------------------------------------------------------------------------
3112 This is for use when we raise an exception in another thread, which
3114 This has nothing to do with the UnblockThread event in GranSim. -- HWL
3115 -------------------------------------------------------------------------- */
3117 #if defined(GRAN) || defined(PARALLEL_HASKELL)
3119 NB: only the type of the blocking queue is different in GranSim and GUM
3120 the operations on the queue-elements are the same
3121 long live polymorphism!
3123 Locks: sched_mutex is held upon entry and exit.
3127 unblockThread(Capability *cap, StgTSO *tso)
3129 StgBlockingQueueElement *t, **last;
3131 switch (tso->why_blocked) {
3134 return; /* not blocked */
3137 // Be careful: nothing to do here! We tell the scheduler that the thread
3138 // is runnable and we leave it to the stack-walking code to abort the
3139 // transaction while unwinding the stack. We should perhaps have a debugging
3140 // test to make sure that this really happens and that the 'zombie' transaction
3141 // does not get committed.
3145 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
3147 StgBlockingQueueElement *last_tso = END_BQ_QUEUE;
3148 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
3150 last = (StgBlockingQueueElement **)&mvar->head;
3151 for (t = (StgBlockingQueueElement *)mvar->head;
3153 last = &t->link, last_tso = t, t = t->link) {
3154 if (t == (StgBlockingQueueElement *)tso) {
3155 *last = (StgBlockingQueueElement *)tso->link;
3156 if (mvar->tail == tso) {
3157 mvar->tail = (StgTSO *)last_tso;
3162 barf("unblockThread (MVAR): TSO not found");
3165 case BlockedOnBlackHole:
3166 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
3168 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
3170 last = &bq->blocking_queue;
3171 for (t = bq->blocking_queue;
3173 last = &t->link, t = t->link) {
3174 if (t == (StgBlockingQueueElement *)tso) {
3175 *last = (StgBlockingQueueElement *)tso->link;
3179 barf("unblockThread (BLACKHOLE): TSO not found");
3182 case BlockedOnException:
3184 StgTSO *target = tso->block_info.tso;
3186 ASSERT(get_itbl(target)->type == TSO);
3188 if (target->what_next == ThreadRelocated) {
3189 target = target->link;
3190 ASSERT(get_itbl(target)->type == TSO);
3193 ASSERT(target->blocked_exceptions != NULL);
3195 last = (StgBlockingQueueElement **)&target->blocked_exceptions;
3196 for (t = (StgBlockingQueueElement *)target->blocked_exceptions;
3198 last = &t->link, t = t->link) {
3199 ASSERT(get_itbl(t)->type == TSO);
3200 if (t == (StgBlockingQueueElement *)tso) {
3201 *last = (StgBlockingQueueElement *)tso->link;
3205 barf("unblockThread (Exception): TSO not found");
3209 case BlockedOnWrite:
3210 #if defined(mingw32_HOST_OS)
3211 case BlockedOnDoProc:
3214 /* take TSO off blocked_queue */
3215 StgBlockingQueueElement *prev = NULL;
3216 for (t = (StgBlockingQueueElement *)blocked_queue_hd; t != END_BQ_QUEUE;
3217 prev = t, t = t->link) {
3218 if (t == (StgBlockingQueueElement *)tso) {
3220 blocked_queue_hd = (StgTSO *)t->link;
3221 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
3222 blocked_queue_tl = END_TSO_QUEUE;
3225 prev->link = t->link;
3226 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
3227 blocked_queue_tl = (StgTSO *)prev;
3230 #if defined(mingw32_HOST_OS)
3231 /* (Cooperatively) signal that the worker thread should abort
3234 abandonWorkRequest(tso->block_info.async_result->reqID);
3239 barf("unblockThread (I/O): TSO not found");
3242 case BlockedOnDelay:
3244 /* take TSO off sleeping_queue */
3245 StgBlockingQueueElement *prev = NULL;
3246 for (t = (StgBlockingQueueElement *)sleeping_queue; t != END_BQ_QUEUE;
3247 prev = t, t = t->link) {
3248 if (t == (StgBlockingQueueElement *)tso) {
3250 sleeping_queue = (StgTSO *)t->link;
3252 prev->link = t->link;
3257 barf("unblockThread (delay): TSO not found");
3261 barf("unblockThread");
3265 tso->link = END_TSO_QUEUE;
3266 tso->why_blocked = NotBlocked;
3267 tso->block_info.closure = NULL;
3268 pushOnRunQueue(cap,tso);
3272 unblockThread(Capability *cap, StgTSO *tso)
3276 /* To avoid locking unnecessarily. */
3277 if (tso->why_blocked == NotBlocked) {
3281 switch (tso->why_blocked) {
3284 // Be careful: nothing to do here! We tell the scheduler that the thread
3285 // is runnable and we leave it to the stack-walking code to abort the
3286 // transaction while unwinding the stack. We should perhaps have a debugging
3287 // test to make sure that this really happens and that the 'zombie' transaction
3288 // does not get committed.
3292 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
3294 StgTSO *last_tso = END_TSO_QUEUE;
3295 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
3298 for (t = mvar->head; t != END_TSO_QUEUE;
3299 last = &t->link, last_tso = t, t = t->link) {
3302 if (mvar->tail == tso) {
3303 mvar->tail = last_tso;
3308 barf("unblockThread (MVAR): TSO not found");
3311 case BlockedOnBlackHole:
3313 last = &blackhole_queue;
3314 for (t = blackhole_queue; t != END_TSO_QUEUE;
3315 last = &t->link, t = t->link) {
3321 barf("unblockThread (BLACKHOLE): TSO not found");
3324 case BlockedOnException:
3326 StgTSO *target = tso->block_info.tso;
3328 ASSERT(get_itbl(target)->type == TSO);
3330 while (target->what_next == ThreadRelocated) {
3331 target = target->link;
3332 ASSERT(get_itbl(target)->type == TSO);
3335 ASSERT(target->blocked_exceptions != NULL);
3337 last = &target->blocked_exceptions;
3338 for (t = target->blocked_exceptions; t != END_TSO_QUEUE;
3339 last = &t->link, t = t->link) {
3340 ASSERT(get_itbl(t)->type == TSO);
3346 barf("unblockThread (Exception): TSO not found");
3349 #if !defined(THREADED_RTS)
3351 case BlockedOnWrite:
3352 #if defined(mingw32_HOST_OS)
3353 case BlockedOnDoProc:
3356 StgTSO *prev = NULL;
3357 for (t = blocked_queue_hd; t != END_TSO_QUEUE;
3358 prev = t, t = t->link) {
3361 blocked_queue_hd = t->link;
3362 if (blocked_queue_tl == t) {
3363 blocked_queue_tl = END_TSO_QUEUE;
3366 prev->link = t->link;
3367 if (blocked_queue_tl == t) {
3368 blocked_queue_tl = prev;
3371 #if defined(mingw32_HOST_OS)
3372 /* (Cooperatively) signal that the worker thread should abort
3375 abandonWorkRequest(tso->block_info.async_result->reqID);
3380 barf("unblockThread (I/O): TSO not found");
3383 case BlockedOnDelay:
3385 StgTSO *prev = NULL;
3386 for (t = sleeping_queue; t != END_TSO_QUEUE;
3387 prev = t, t = t->link) {
3390 sleeping_queue = t->link;
3392 prev->link = t->link;
3397 barf("unblockThread (delay): TSO not found");
3402 barf("unblockThread");
3406 tso->link = END_TSO_QUEUE;
3407 tso->why_blocked = NotBlocked;
3408 tso->block_info.closure = NULL;
3409 appendToRunQueue(cap,tso);
3413 /* -----------------------------------------------------------------------------
3416 * Check the blackhole_queue for threads that can be woken up. We do
3417 * this periodically: before every GC, and whenever the run queue is
3420 * An elegant solution might be to just wake up all the blocked
3421 * threads with awakenBlockedQueue occasionally: they'll go back to
3422 * sleep again if the object is still a BLACKHOLE. Unfortunately this
3423 * doesn't give us a way to tell whether we've actually managed to
3424 * wake up any threads, so we would be busy-waiting.
3426 * -------------------------------------------------------------------------- */
3429 checkBlackHoles (Capability *cap)
3432 rtsBool any_woke_up = rtsFalse;
3435 // blackhole_queue is global:
3436 ASSERT_LOCK_HELD(&sched_mutex);
3438 IF_DEBUG(scheduler, sched_belch("checking threads blocked on black holes"));
3440 // ASSUMES: sched_mutex
3441 prev = &blackhole_queue;
3442 t = blackhole_queue;
3443 while (t != END_TSO_QUEUE) {
3444 ASSERT(t->why_blocked == BlockedOnBlackHole);
3445 type = get_itbl(t->block_info.closure)->type;
3446 if (type != BLACKHOLE && type != CAF_BLACKHOLE) {
3447 IF_DEBUG(sanity,checkTSO(t));
3448 t = unblockOne(cap, t);
3449 // urk, the threads migrate to the current capability
3450 // here, but we'd like to keep them on the original one.
3452 any_woke_up = rtsTrue;
3462 /* -----------------------------------------------------------------------------
3465 * The following function implements the magic for raising an
3466 * asynchronous exception in an existing thread.
3468 * We first remove the thread from any queue on which it might be
3469 * blocked. The possible blockages are MVARs and BLACKHOLE_BQs.
3471 * We strip the stack down to the innermost CATCH_FRAME, building
3472 * thunks in the heap for all the active computations, so they can
3473 * be restarted if necessary. When we reach a CATCH_FRAME, we build
3474 * an application of the handler to the exception, and push it on
3475 * the top of the stack.
3477 * How exactly do we save all the active computations? We create an
3478 * AP_STACK for every UpdateFrame on the stack. Entering one of these
3479 * AP_STACKs pushes everything from the corresponding update frame
3480 * upwards onto the stack. (Actually, it pushes everything up to the
3481 * next update frame plus a pointer to the next AP_STACK object.
3482 * Entering the next AP_STACK object pushes more onto the stack until we
3483 * reach the last AP_STACK object - at which point the stack should look
3484 * exactly as it did when we killed the TSO and we can continue
3485 * execution by entering the closure on top of the stack.
3487 * We can also kill a thread entirely - this happens if either (a) the
3488 * exception passed to raiseAsync is NULL, or (b) there's no
3489 * CATCH_FRAME on the stack. In either case, we strip the entire
3490 * stack and replace the thread with a zombie.
3492 * ToDo: in SMP mode, this function is only safe if either (a) we hold
3493 * all the Capabilities (eg. in GC), or (b) we own the Capability that
3494 * the TSO is currently blocked on or on the run queue of.
3496 * -------------------------------------------------------------------------- */
3499 raiseAsync(Capability *cap, StgTSO *tso, StgClosure *exception)
3501 raiseAsync_(cap, tso, exception, rtsFalse);
3505 raiseAsync_(Capability *cap, StgTSO *tso, StgClosure *exception,
3506 rtsBool stop_at_atomically)
3508 StgRetInfoTable *info;
3511 // Thread already dead?
3512 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
3517 sched_belch("raising exception in thread %ld.", (long)tso->id));
3519 // Remove it from any blocking queues
3520 unblockThread(cap,tso);
3524 // The stack freezing code assumes there's a closure pointer on
3525 // the top of the stack, so we have to arrange that this is the case...
3527 if (sp[0] == (W_)&stg_enter_info) {
3531 sp[0] = (W_)&stg_dummy_ret_closure;
3537 // 1. Let the top of the stack be the "current closure"
3539 // 2. Walk up the stack until we find either an UPDATE_FRAME or a
3542 // 3. If it's an UPDATE_FRAME, then make an AP_STACK containing the
3543 // current closure applied to the chunk of stack up to (but not
3544 // including) the update frame. This closure becomes the "current
3545 // closure". Go back to step 2.
3547 // 4. If it's a CATCH_FRAME, then leave the exception handler on
3548 // top of the stack applied to the exception.
3550 // 5. If it's a STOP_FRAME, then kill the thread.
3552 // NB: if we pass an ATOMICALLY_FRAME then abort the associated
3559 info = get_ret_itbl((StgClosure *)frame);
3561 while (info->i.type != UPDATE_FRAME
3562 && (info->i.type != CATCH_FRAME || exception == NULL)
3563 && info->i.type != STOP_FRAME
3564 && (info->i.type != ATOMICALLY_FRAME || stop_at_atomically == rtsFalse))
3566 if (info->i.type == CATCH_RETRY_FRAME || info->i.type == ATOMICALLY_FRAME) {
3567 // IF we find an ATOMICALLY_FRAME then we abort the
3568 // current transaction and propagate the exception. In
3569 // this case (unlike ordinary exceptions) we do not care
3570 // whether the transaction is valid or not because its
3571 // possible validity cannot have caused the exception
3572 // and will not be visible after the abort.
3574 debugBelch("Found atomically block delivering async exception\n"));
3575 stmAbortTransaction(tso -> trec);
3576 tso -> trec = stmGetEnclosingTRec(tso -> trec);
3578 frame += stack_frame_sizeW((StgClosure *)frame);
3579 info = get_ret_itbl((StgClosure *)frame);
3582 switch (info->i.type) {
3584 case ATOMICALLY_FRAME:
3585 ASSERT(stop_at_atomically);
3586 ASSERT(stmGetEnclosingTRec(tso->trec) == NO_TREC);
3587 stmCondemnTransaction(tso -> trec);
3591 // R1 is not a register: the return convention for IO in
3592 // this case puts the return value on the stack, so we
3593 // need to set up the stack to return to the atomically
3594 // frame properly...
3595 tso->sp = frame - 2;
3596 tso->sp[1] = (StgWord) &stg_NO_FINALIZER_closure; // why not?
3597 tso->sp[0] = (StgWord) &stg_ut_1_0_unreg_info;
3599 tso->what_next = ThreadRunGHC;
3603 // If we find a CATCH_FRAME, and we've got an exception to raise,
3604 // then build the THUNK raise(exception), and leave it on
3605 // top of the CATCH_FRAME ready to enter.
3609 StgCatchFrame *cf = (StgCatchFrame *)frame;
3613 // we've got an exception to raise, so let's pass it to the
3614 // handler in this frame.
3616 raise = (StgThunk *)allocateLocal(cap,sizeofW(StgThunk)+MIN_UPD_SIZE);
3617 TICK_ALLOC_SE_THK(1,0);
3618 SET_HDR(raise,&stg_raise_info,cf->header.prof.ccs);
3619 raise->payload[0] = exception;
3621 // throw away the stack from Sp up to the CATCH_FRAME.
3625 /* Ensure that async excpetions are blocked now, so we don't get
3626 * a surprise exception before we get around to executing the
3629 if (tso->blocked_exceptions == NULL) {
3630 tso->blocked_exceptions = END_TSO_QUEUE;
3633 /* Put the newly-built THUNK on top of the stack, ready to execute
3634 * when the thread restarts.
3637 sp[-1] = (W_)&stg_enter_info;
3639 tso->what_next = ThreadRunGHC;
3640 IF_DEBUG(sanity, checkTSO(tso));
3649 // First build an AP_STACK consisting of the stack chunk above the
3650 // current update frame, with the top word on the stack as the
3653 words = frame - sp - 1;
3654 ap = (StgAP_STACK *)allocateLocal(cap,AP_STACK_sizeW(words));
3657 ap->fun = (StgClosure *)sp[0];
3659 for(i=0; i < (nat)words; ++i) {
3660 ap->payload[i] = (StgClosure *)*sp++;
3663 SET_HDR(ap,&stg_AP_STACK_info,
3664 ((StgClosure *)frame)->header.prof.ccs /* ToDo */);
3665 TICK_ALLOC_UP_THK(words+1,0);
3668 debugBelch("sched: Updating ");
3669 printPtr((P_)((StgUpdateFrame *)frame)->updatee);
3670 debugBelch(" with ");
3671 printObj((StgClosure *)ap);
3674 // Replace the updatee with an indirection - happily
3675 // this will also wake up any threads currently
3676 // waiting on the result.
3678 // Warning: if we're in a loop, more than one update frame on
3679 // the stack may point to the same object. Be careful not to
3680 // overwrite an IND_OLDGEN in this case, because we'll screw
3681 // up the mutable lists. To be on the safe side, don't
3682 // overwrite any kind of indirection at all. See also
3683 // threadSqueezeStack in GC.c, where we have to make a similar
3686 if (!closure_IND(((StgUpdateFrame *)frame)->updatee)) {
3687 // revert the black hole
3688 UPD_IND_NOLOCK(((StgUpdateFrame *)frame)->updatee,
3691 sp += sizeofW(StgUpdateFrame) - 1;
3692 sp[0] = (W_)ap; // push onto stack
3697 // We've stripped the entire stack, the thread is now dead.
3698 sp += sizeofW(StgStopFrame);
3699 tso->what_next = ThreadKilled;
3710 /* -----------------------------------------------------------------------------
3713 This is used for interruption (^C) and forking, and corresponds to
3714 raising an exception but without letting the thread catch the
3716 -------------------------------------------------------------------------- */
3719 deleteThread (Capability *cap, StgTSO *tso)
3721 if (tso->why_blocked != BlockedOnCCall &&
3722 tso->why_blocked != BlockedOnCCall_NoUnblockExc) {
3723 raiseAsync(cap,tso,NULL);
3727 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
3729 deleteThreadImmediately(Capability *cap, StgTSO *tso)
3730 { // for forkProcess only:
3731 // delete thread without giving it a chance to catch the KillThread exception
3733 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
3737 if (tso->why_blocked != BlockedOnCCall &&
3738 tso->why_blocked != BlockedOnCCall_NoUnblockExc) {
3739 unblockThread(cap,tso);
3742 tso->what_next = ThreadKilled;
3746 /* -----------------------------------------------------------------------------
3747 raiseExceptionHelper
3749 This function is called by the raise# primitve, just so that we can
3750 move some of the tricky bits of raising an exception from C-- into
3751 C. Who knows, it might be a useful re-useable thing here too.
3752 -------------------------------------------------------------------------- */
3755 raiseExceptionHelper (StgRegTable *reg, StgTSO *tso, StgClosure *exception)
3757 Capability *cap = regTableToCapability(reg);
3758 StgThunk *raise_closure = NULL;
3760 StgRetInfoTable *info;
3762 // This closure represents the expression 'raise# E' where E
3763 // is the exception raise. It is used to overwrite all the
3764 // thunks which are currently under evaluataion.
3768 // LDV profiling: stg_raise_info has THUNK as its closure
3769 // type. Since a THUNK takes at least MIN_UPD_SIZE words in its
3770 // payload, MIN_UPD_SIZE is more approprate than 1. It seems that
3771 // 1 does not cause any problem unless profiling is performed.
3772 // However, when LDV profiling goes on, we need to linearly scan
3773 // small object pool, where raise_closure is stored, so we should
3774 // use MIN_UPD_SIZE.
3776 // raise_closure = (StgClosure *)RET_STGCALL1(P_,allocate,
3777 // sizeofW(StgClosure)+1);
3781 // Walk up the stack, looking for the catch frame. On the way,
3782 // we update any closures pointed to from update frames with the
3783 // raise closure that we just built.
3787 info = get_ret_itbl((StgClosure *)p);
3788 next = p + stack_frame_sizeW((StgClosure *)p);
3789 switch (info->i.type) {
3792 // Only create raise_closure if we need to.
3793 if (raise_closure == NULL) {
3795 (StgThunk *)allocateLocal(cap,sizeofW(StgThunk)+MIN_UPD_SIZE);
3796 SET_HDR(raise_closure, &stg_raise_info, CCCS);
3797 raise_closure->payload[0] = exception;
3799 UPD_IND(((StgUpdateFrame *)p)->updatee,(StgClosure *)raise_closure);
3803 case ATOMICALLY_FRAME:
3804 IF_DEBUG(stm, debugBelch("Found ATOMICALLY_FRAME at %p\n", p));
3806 return ATOMICALLY_FRAME;
3812 case CATCH_STM_FRAME:
3813 IF_DEBUG(stm, debugBelch("Found CATCH_STM_FRAME at %p\n", p));
3815 return CATCH_STM_FRAME;
3821 case CATCH_RETRY_FRAME:
3830 /* -----------------------------------------------------------------------------
3831 findRetryFrameHelper
3833 This function is called by the retry# primitive. It traverses the stack
3834 leaving tso->sp referring to the frame which should handle the retry.
3836 This should either be a CATCH_RETRY_FRAME (if the retry# is within an orElse#)
3837 or should be a ATOMICALLY_FRAME (if the retry# reaches the top level).
3839 We skip CATCH_STM_FRAMEs because retries are not considered to be exceptions,
3840 despite the similar implementation.
3842 We should not expect to see CATCH_FRAME or STOP_FRAME because those should
3843 not be created within memory transactions.
3844 -------------------------------------------------------------------------- */
3847 findRetryFrameHelper (StgTSO *tso)
3850 StgRetInfoTable *info;
3854 info = get_ret_itbl((StgClosure *)p);
3855 next = p + stack_frame_sizeW((StgClosure *)p);
3856 switch (info->i.type) {
3858 case ATOMICALLY_FRAME:
3859 IF_DEBUG(stm, debugBelch("Found ATOMICALLY_FRAME at %p during retrry\n", p));
3861 return ATOMICALLY_FRAME;
3863 case CATCH_RETRY_FRAME:
3864 IF_DEBUG(stm, debugBelch("Found CATCH_RETRY_FRAME at %p during retrry\n", p));
3866 return CATCH_RETRY_FRAME;
3868 case CATCH_STM_FRAME:
3870 ASSERT(info->i.type != CATCH_FRAME);
3871 ASSERT(info->i.type != STOP_FRAME);
3878 /* -----------------------------------------------------------------------------
3879 resurrectThreads is called after garbage collection on the list of
3880 threads found to be garbage. Each of these threads will be woken
3881 up and sent a signal: BlockedOnDeadMVar if the thread was blocked
3882 on an MVar, or NonTermination if the thread was blocked on a Black
3885 Locks: assumes we hold *all* the capabilities.
3886 -------------------------------------------------------------------------- */
3889 resurrectThreads (StgTSO *threads)
3894 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
3895 next = tso->global_link;
3896 tso->global_link = all_threads;
3898 IF_DEBUG(scheduler, sched_belch("resurrecting thread %d", tso->id));
3900 // Wake up the thread on the Capability it was last on for a
3901 // bound thread, or last_free_capability otherwise.
3903 cap = tso->bound->cap;
3905 cap = last_free_capability;
3908 switch (tso->why_blocked) {
3910 case BlockedOnException:
3911 /* Called by GC - sched_mutex lock is currently held. */
3912 raiseAsync(cap, tso,(StgClosure *)BlockedOnDeadMVar_closure);
3914 case BlockedOnBlackHole:
3915 raiseAsync(cap, tso,(StgClosure *)NonTermination_closure);
3918 raiseAsync(cap, tso,(StgClosure *)BlockedIndefinitely_closure);
3921 /* This might happen if the thread was blocked on a black hole
3922 * belonging to a thread that we've just woken up (raiseAsync
3923 * can wake up threads, remember...).
3927 barf("resurrectThreads: thread blocked in a strange way");
3932 /* ----------------------------------------------------------------------------
3933 * Debugging: why is a thread blocked
3934 * [Also provides useful information when debugging threaded programs
3935 * at the Haskell source code level, so enable outside of DEBUG. --sof 7/02]
3936 ------------------------------------------------------------------------- */
3940 printThreadBlockage(StgTSO *tso)
3942 switch (tso->why_blocked) {
3944 debugBelch("is blocked on read from fd %d", (int)(tso->block_info.fd));
3946 case BlockedOnWrite:
3947 debugBelch("is blocked on write to fd %d", (int)(tso->block_info.fd));
3949 #if defined(mingw32_HOST_OS)
3950 case BlockedOnDoProc:
3951 debugBelch("is blocked on proc (request: %ld)", tso->block_info.async_result->reqID);
3954 case BlockedOnDelay:
3955 debugBelch("is blocked until %ld", (long)(tso->block_info.target));
3958 debugBelch("is blocked on an MVar @ %p", tso->block_info.closure);
3960 case BlockedOnException:
3961 debugBelch("is blocked on delivering an exception to thread %d",
3962 tso->block_info.tso->id);
3964 case BlockedOnBlackHole:
3965 debugBelch("is blocked on a black hole");
3968 debugBelch("is not blocked");
3970 #if defined(PARALLEL_HASKELL)
3972 debugBelch("is blocked on global address; local FM_BQ is %p (%s)",
3973 tso->block_info.closure, info_type(tso->block_info.closure));
3975 case BlockedOnGA_NoSend:
3976 debugBelch("is blocked on global address (no send); local FM_BQ is %p (%s)",
3977 tso->block_info.closure, info_type(tso->block_info.closure));
3980 case BlockedOnCCall:
3981 debugBelch("is blocked on an external call");
3983 case BlockedOnCCall_NoUnblockExc:
3984 debugBelch("is blocked on an external call (exceptions were already blocked)");
3987 debugBelch("is blocked on an STM operation");
3990 barf("printThreadBlockage: strange tso->why_blocked: %d for TSO %d (%d)",
3991 tso->why_blocked, tso->id, tso);
3996 printThreadStatus(StgTSO *tso)
3998 switch (tso->what_next) {
4000 debugBelch("has been killed");
4002 case ThreadComplete:
4003 debugBelch("has completed");
4006 printThreadBlockage(tso);
4011 printAllThreads(void)
4016 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
4017 ullong_format_string(TIME_ON_PROC(CurrentProc),
4018 time_string, rtsFalse/*no commas!*/);
4020 debugBelch("all threads at [%s]:\n", time_string);
4021 # elif defined(PARALLEL_HASKELL)
4022 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
4023 ullong_format_string(CURRENT_TIME,
4024 time_string, rtsFalse/*no commas!*/);
4026 debugBelch("all threads at [%s]:\n", time_string);
4028 debugBelch("all threads:\n");
4031 for (t = all_threads; t != END_TSO_QUEUE; ) {
4032 debugBelch("\tthread %4d @ %p ", t->id, (void *)t);
4034 void *label = lookupThreadLabel(t->id);
4035 if (label) debugBelch("[\"%s\"] ",(char *)label);
4037 if (t->what_next == ThreadRelocated) {
4038 debugBelch("has been relocated...\n");
4041 printThreadStatus(t);
4050 printThreadQueue(StgTSO *t)
4053 for (; t != END_TSO_QUEUE; t = t->link) {
4054 debugBelch("\tthread %d @ %p ", t->id, (void *)t);
4055 if (t->what_next == ThreadRelocated) {
4056 debugBelch("has been relocated...\n");
4058 printThreadStatus(t);
4063 debugBelch("%d threads on queue\n", i);
4067 Print a whole blocking queue attached to node (debugging only).
4069 # if defined(PARALLEL_HASKELL)
4071 print_bq (StgClosure *node)
4073 StgBlockingQueueElement *bqe;
4077 debugBelch("## BQ of closure %p (%s): ",
4078 node, info_type(node));
4080 /* should cover all closures that may have a blocking queue */
4081 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
4082 get_itbl(node)->type == FETCH_ME_BQ ||
4083 get_itbl(node)->type == RBH ||
4084 get_itbl(node)->type == MVAR);
4086 ASSERT(node!=(StgClosure*)NULL); // sanity check
4088 print_bqe(((StgBlockingQueue*)node)->blocking_queue);
4092 Print a whole blocking queue starting with the element bqe.
4095 print_bqe (StgBlockingQueueElement *bqe)
4100 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
4102 for (end = (bqe==END_BQ_QUEUE);
4103 !end; // iterate until bqe points to a CONSTR
4104 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE),
4105 bqe = end ? END_BQ_QUEUE : bqe->link) {
4106 ASSERT(bqe != END_BQ_QUEUE); // sanity check
4107 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
4108 /* types of closures that may appear in a blocking queue */
4109 ASSERT(get_itbl(bqe)->type == TSO ||
4110 get_itbl(bqe)->type == BLOCKED_FETCH ||
4111 get_itbl(bqe)->type == CONSTR);
4112 /* only BQs of an RBH end with an RBH_Save closure */
4113 //ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
4115 switch (get_itbl(bqe)->type) {
4117 debugBelch(" TSO %u (%x),",
4118 ((StgTSO *)bqe)->id, ((StgTSO *)bqe));
4121 debugBelch(" BF (node=%p, ga=((%x, %d, %x)),",
4122 ((StgBlockedFetch *)bqe)->node,
4123 ((StgBlockedFetch *)bqe)->ga.payload.gc.gtid,
4124 ((StgBlockedFetch *)bqe)->ga.payload.gc.slot,
4125 ((StgBlockedFetch *)bqe)->ga.weight);
4128 debugBelch(" %s (IP %p),",
4129 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
4130 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
4131 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
4132 "RBH_Save_?"), get_itbl(bqe));
4135 barf("Unexpected closure type %s in blocking queue", // of %p (%s)",
4136 info_type((StgClosure *)bqe)); // , node, info_type(node));
4142 # elif defined(GRAN)
4144 print_bq (StgClosure *node)
4146 StgBlockingQueueElement *bqe;
4147 PEs node_loc, tso_loc;
4150 /* should cover all closures that may have a blocking queue */
4151 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
4152 get_itbl(node)->type == FETCH_ME_BQ ||
4153 get_itbl(node)->type == RBH);
4155 ASSERT(node!=(StgClosure*)NULL); // sanity check
4156 node_loc = where_is(node);
4158 debugBelch("## BQ of closure %p (%s) on [PE %d]: ",
4159 node, info_type(node), node_loc);
4162 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
4164 for (bqe = ((StgBlockingQueue*)node)->blocking_queue, end = (bqe==END_BQ_QUEUE);
4165 !end; // iterate until bqe points to a CONSTR
4166 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) {
4167 ASSERT(bqe != END_BQ_QUEUE); // sanity check
4168 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
4169 /* types of closures that may appear in a blocking queue */
4170 ASSERT(get_itbl(bqe)->type == TSO ||
4171 get_itbl(bqe)->type == CONSTR);
4172 /* only BQs of an RBH end with an RBH_Save closure */
4173 ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
4175 tso_loc = where_is((StgClosure *)bqe);
4176 switch (get_itbl(bqe)->type) {
4178 debugBelch(" TSO %d (%p) on [PE %d],",
4179 ((StgTSO *)bqe)->id, (StgTSO *)bqe, tso_loc);
4182 debugBelch(" %s (IP %p),",
4183 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
4184 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
4185 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
4186 "RBH_Save_?"), get_itbl(bqe));
4189 barf("Unexpected closure type %s in blocking queue of %p (%s)",
4190 info_type((StgClosure *)bqe), node, info_type(node));
4198 #if defined(PARALLEL_HASKELL)
4205 for (i=0, tso=run_queue_hd;
4206 tso != END_TSO_QUEUE;
4207 i++, tso=tso->link) {
4216 sched_belch(char *s, ...)
4221 debugBelch("sched (task %p): ", (void *)(unsigned long)(unsigned int)osThreadId());
4222 #elif defined(PARALLEL_HASKELL)
4225 debugBelch("sched: ");