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
382 ASSERT_CAPABILITY_INVARIANTS(cap,task);
384 // Yield the capability to higher-priority tasks if necessary.
385 yieldCapability(&cap, task);
389 // Check whether we have re-entered the RTS from Haskell without
390 // going via suspendThread()/resumeThread (i.e. a 'safe' foreign
392 if (cap->in_haskell) {
393 errorBelch("schedule: re-entered unsafely.\n"
394 " Perhaps a 'foreign import unsafe' should be 'safe'?");
395 stg_exit(EXIT_FAILURE);
399 // Test for interruption. If interrupted==rtsTrue, then either
400 // we received a keyboard interrupt (^C), or the scheduler is
401 // trying to shut down all the tasks (shutting_down_scheduler) in
406 if (shutting_down_scheduler) {
407 IF_DEBUG(scheduler, sched_belch("shutting down"));
408 // If we are a worker, just exit. If we're a bound thread
409 // then we will exit below when we've removed our TSO from
411 if (task->tso == NULL) {
415 IF_DEBUG(scheduler, sched_belch("interrupted"));
419 #if defined(not_yet) && defined(SMP)
421 // Top up the run queue from our spark pool. We try to make the
422 // number of threads in the run queue equal to the number of
423 // free capabilities.
427 if (emptyRunQueue()) {
428 spark = findSpark(rtsFalse);
430 break; /* no more sparks in the pool */
432 createSparkThread(spark);
434 sched_belch("==^^ turning spark of closure %p into a thread",
435 (StgClosure *)spark));
441 scheduleStartSignalHandlers();
443 // Only check the black holes here if we've nothing else to do.
444 // During normal execution, the black hole list only gets checked
445 // at GC time, to avoid repeatedly traversing this possibly long
446 // list each time around the scheduler.
447 if (emptyRunQueue(cap)) { scheduleCheckBlackHoles(cap); }
449 scheduleCheckBlockedThreads(cap);
451 scheduleDetectDeadlock(cap,task);
453 // Normally, the only way we can get here with no threads to
454 // run is if a keyboard interrupt received during
455 // scheduleCheckBlockedThreads() or scheduleDetectDeadlock().
456 // Additionally, it is not fatal for the
457 // threaded RTS to reach here with no threads to run.
459 // win32: might be here due to awaitEvent() being abandoned
460 // as a result of a console event having been delivered.
461 if ( emptyRunQueue(cap) ) {
462 #if !defined(THREADED_RTS) && !defined(mingw32_HOST_OS)
465 continue; // nothing to do
468 #if defined(PARALLEL_HASKELL)
469 scheduleSendPendingMessages();
470 if (emptyRunQueue(cap) && scheduleActivateSpark())
474 ASSERT(next_fish_to_send_at==0); // i.e. no delayed fishes left!
477 /* If we still have no work we need to send a FISH to get a spark
479 if (emptyRunQueue(cap)) {
480 if (!scheduleGetRemoteWork(&receivedFinish)) continue;
481 ASSERT(rtsFalse); // should not happen at the moment
483 // from here: non-empty run queue.
484 // TODO: merge above case with this, only one call processMessages() !
485 if (PacketsWaiting()) { /* process incoming messages, if
486 any pending... only in else
487 because getRemoteWork waits for
489 receivedFinish = processMessages();
494 scheduleProcessEvent(event);
498 // Get a thread to run
500 t = popRunQueue(cap);
502 #if defined(GRAN) || defined(PAR)
503 scheduleGranParReport(); // some kind of debuging output
505 // Sanity check the thread we're about to run. This can be
506 // expensive if there is lots of thread switching going on...
507 IF_DEBUG(sanity,checkTSO(t));
510 #if defined(THREADED_RTS)
511 // Check whether we can run this thread in the current task.
512 // If not, we have to pass our capability to the right task.
514 Task *bound = t->bound;
519 sched_belch("### Running thread %d in bound thread",
521 // yes, the Haskell thread is bound to the current native thread
524 sched_belch("### thread %d bound to another OS thread",
526 // no, bound to a different Haskell thread: pass to that thread
527 pushOnRunQueue(cap,t);
531 // The thread we want to run is unbound.
534 sched_belch("### this OS thread cannot run thread %d", t->id));
535 // no, the current native thread is bound to a different
536 // Haskell thread, so pass it to any worker thread
537 pushOnRunQueue(cap,t);
544 cap->r.rCurrentTSO = t;
546 /* context switches are initiated by the timer signal, unless
547 * the user specified "context switch as often as possible", with
550 if (RtsFlags.ConcFlags.ctxtSwitchTicks == 0
551 && !emptyThreadQueues(cap)) {
557 IF_DEBUG(scheduler, sched_belch("-->> running thread %ld %s ...",
558 (long)t->id, whatNext_strs[t->what_next]));
560 #if defined(PROFILING)
561 startHeapProfTimer();
564 // ----------------------------------------------------------------------
565 // Run the current thread
567 prev_what_next = t->what_next;
569 errno = t->saved_errno;
570 cap->in_haskell = rtsTrue;
572 recent_activity = ACTIVITY_YES;
574 switch (prev_what_next) {
578 /* Thread already finished, return to scheduler. */
579 ret = ThreadFinished;
585 r = StgRun((StgFunPtr) stg_returnToStackTop, &cap->r);
586 cap = regTableToCapability(r);
591 case ThreadInterpret:
592 cap = interpretBCO(cap);
597 barf("schedule: invalid what_next field");
600 cap->in_haskell = rtsFalse;
603 // If ret is ThreadBlocked, and this Task is bound to the TSO that
604 // blocked, we are in limbo - the TSO is now owned by whatever it
605 // is blocked on, and may in fact already have been woken up,
606 // perhaps even on a different Capability. It may be the case
607 // that task->cap != cap. We better yield this Capability
608 // immediately and return to normaility.
609 if (ret == ThreadBlocked) continue;
612 ASSERT_CAPABILITY_INVARIANTS(cap,task);
614 // The TSO might have moved, eg. if it re-entered the RTS and a GC
615 // happened. So find the new location:
616 t = cap->r.rCurrentTSO;
618 // And save the current errno in this thread.
619 t->saved_errno = errno;
621 // ----------------------------------------------------------------------
623 // Costs for the scheduler are assigned to CCS_SYSTEM
624 #if defined(PROFILING)
629 // We have run some Haskell code: there might be blackhole-blocked
630 // threads to wake up now.
631 // Lock-free test here should be ok, we're just setting a flag.
632 if ( blackhole_queue != END_TSO_QUEUE ) {
633 blackholes_need_checking = rtsTrue;
636 #if defined(THREADED_RTS)
637 IF_DEBUG(scheduler,debugBelch("sched (task %p): ", (void *)(unsigned long)(unsigned int)osThreadId()););
638 #elif !defined(GRAN) && !defined(PARALLEL_HASKELL)
639 IF_DEBUG(scheduler,debugBelch("sched: "););
642 schedulePostRunThread();
644 ready_to_gc = rtsFalse;
648 ready_to_gc = scheduleHandleHeapOverflow(cap,t);
652 scheduleHandleStackOverflow(cap,task,t);
656 if (scheduleHandleYield(cap, t, prev_what_next)) {
657 // shortcut for switching between compiler/interpreter:
663 scheduleHandleThreadBlocked(t);
667 if (scheduleHandleThreadFinished(cap, task, t)) return cap;
671 barf("schedule: invalid thread return code %d", (int)ret);
674 if (scheduleDoHeapProfile(ready_to_gc)) { ready_to_gc = rtsFalse; }
675 if (ready_to_gc) { scheduleDoGC(cap,task,rtsFalse); }
676 } /* end of while() */
678 IF_PAR_DEBUG(verbose,
679 debugBelch("== Leaving schedule() after having received Finish\n"));
682 /* ----------------------------------------------------------------------------
683 * Setting up the scheduler loop
684 * ------------------------------------------------------------------------- */
687 schedulePreLoop(void)
690 /* set up first event to get things going */
691 /* ToDo: assign costs for system setup and init MainTSO ! */
692 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
694 CurrentTSO, (StgClosure*)NULL, (rtsSpark*)NULL);
697 debugBelch("GRAN: Init CurrentTSO (in schedule) = %p\n",
699 G_TSO(CurrentTSO, 5));
701 if (RtsFlags.GranFlags.Light) {
702 /* Save current time; GranSim Light only */
703 CurrentTSO->gran.clock = CurrentTime[CurrentProc];
708 /* ----------------------------------------------------------------------------
709 * Start any pending signal handlers
710 * ------------------------------------------------------------------------- */
713 scheduleStartSignalHandlers(void)
715 #if defined(RTS_USER_SIGNALS) && !defined(THREADED_RTS)
716 if (signals_pending()) { // safe outside the lock
717 startSignalHandlers();
722 /* ----------------------------------------------------------------------------
723 * Check for blocked threads that can be woken up.
724 * ------------------------------------------------------------------------- */
727 scheduleCheckBlockedThreads(Capability *cap USED_WHEN_NON_THREADED_RTS)
729 #if !defined(THREADED_RTS)
731 // Check whether any waiting threads need to be woken up. If the
732 // run queue is empty, and there are no other tasks running, we
733 // can wait indefinitely for something to happen.
735 if ( !emptyQueue(blocked_queue_hd) || !emptyQueue(sleeping_queue) )
737 awaitEvent( emptyRunQueue(cap) && !blackholes_need_checking );
743 /* ----------------------------------------------------------------------------
744 * Check for threads blocked on BLACKHOLEs that can be woken up
745 * ------------------------------------------------------------------------- */
747 scheduleCheckBlackHoles (Capability *cap)
749 if ( blackholes_need_checking ) // check without the lock first
751 ACQUIRE_LOCK(&sched_mutex);
752 if ( blackholes_need_checking ) {
753 checkBlackHoles(cap);
754 blackholes_need_checking = rtsFalse;
756 RELEASE_LOCK(&sched_mutex);
760 /* ----------------------------------------------------------------------------
761 * Detect deadlock conditions and attempt to resolve them.
762 * ------------------------------------------------------------------------- */
765 scheduleDetectDeadlock (Capability *cap, Task *task)
768 #if defined(PARALLEL_HASKELL)
769 // ToDo: add deadlock detection in GUM (similar to SMP) -- HWL
774 * Detect deadlock: when we have no threads to run, there are no
775 * threads blocked, waiting for I/O, or sleeping, and all the
776 * other tasks are waiting for work, we must have a deadlock of
779 if ( emptyThreadQueues(cap) )
781 #if defined(THREADED_RTS)
783 * In the threaded RTS, we only check for deadlock if there
784 * has been no activity in a complete timeslice. This means
785 * we won't eagerly start a full GC just because we don't have
786 * any threads to run currently.
788 if (recent_activity != ACTIVITY_INACTIVE) return;
791 IF_DEBUG(scheduler, sched_belch("deadlocked, forcing major GC..."));
793 // Garbage collection can release some new threads due to
794 // either (a) finalizers or (b) threads resurrected because
795 // they are unreachable and will therefore be sent an
796 // exception. Any threads thus released will be immediately
798 scheduleDoGC( cap, task, rtsTrue/*force major GC*/ );
799 recent_activity = ACTIVITY_DONE_GC;
801 if ( !emptyRunQueue(cap) ) return;
803 #if defined(RTS_USER_SIGNALS) && !defined(THREADED_RTS)
804 /* If we have user-installed signal handlers, then wait
805 * for signals to arrive rather then bombing out with a
808 if ( anyUserHandlers() ) {
810 sched_belch("still deadlocked, waiting for signals..."));
814 if (signals_pending()) {
815 startSignalHandlers();
818 // either we have threads to run, or we were interrupted:
819 ASSERT(!emptyRunQueue(cap) || interrupted);
823 #if !defined(THREADED_RTS)
824 /* Probably a real deadlock. Send the current main thread the
825 * Deadlock exception.
828 switch (task->tso->why_blocked) {
830 case BlockedOnBlackHole:
831 case BlockedOnException:
833 raiseAsync(cap, task->tso, (StgClosure *)NonTermination_closure);
836 barf("deadlock: main thread blocked in a strange way");
844 /* ----------------------------------------------------------------------------
845 * Process an event (GRAN only)
846 * ------------------------------------------------------------------------- */
850 scheduleProcessEvent(rtsEvent *event)
854 if (RtsFlags.GranFlags.Light)
855 GranSimLight_enter_system(event, &ActiveTSO); // adjust ActiveTSO etc
857 /* adjust time based on time-stamp */
858 if (event->time > CurrentTime[CurrentProc] &&
859 event->evttype != ContinueThread)
860 CurrentTime[CurrentProc] = event->time;
862 /* Deal with the idle PEs (may issue FindWork or MoveSpark events) */
863 if (!RtsFlags.GranFlags.Light)
866 IF_DEBUG(gran, debugBelch("GRAN: switch by event-type\n"));
868 /* main event dispatcher in GranSim */
869 switch (event->evttype) {
870 /* Should just be continuing execution */
872 IF_DEBUG(gran, debugBelch("GRAN: doing ContinueThread\n"));
873 /* ToDo: check assertion
874 ASSERT(run_queue_hd != (StgTSO*)NULL &&
875 run_queue_hd != END_TSO_QUEUE);
877 /* Ignore ContinueThreads for fetching threads (if synchr comm) */
878 if (!RtsFlags.GranFlags.DoAsyncFetch &&
879 procStatus[CurrentProc]==Fetching) {
880 debugBelch("ghuH: Spurious ContinueThread while Fetching ignored; TSO %d (%p) [PE %d]\n",
881 CurrentTSO->id, CurrentTSO, CurrentProc);
884 /* Ignore ContinueThreads for completed threads */
885 if (CurrentTSO->what_next == ThreadComplete) {
886 debugBelch("ghuH: found a ContinueThread event for completed thread %d (%p) [PE %d] (ignoring ContinueThread)\n",
887 CurrentTSO->id, CurrentTSO, CurrentProc);
890 /* Ignore ContinueThreads for threads that are being migrated */
891 if (PROCS(CurrentTSO)==Nowhere) {
892 debugBelch("ghuH: trying to run the migrating TSO %d (%p) [PE %d] (ignoring ContinueThread)\n",
893 CurrentTSO->id, CurrentTSO, CurrentProc);
896 /* The thread should be at the beginning of the run queue */
897 if (CurrentTSO!=run_queue_hds[CurrentProc]) {
898 debugBelch("ghuH: TSO %d (%p) [PE %d] is not at the start of the run_queue when doing a ContinueThread\n",
899 CurrentTSO->id, CurrentTSO, CurrentProc);
900 break; // run the thread anyway
903 new_event(proc, proc, CurrentTime[proc],
905 (StgTSO*)NULL, (StgClosure*)NULL, (rtsSpark*)NULL);
907 */ /* Catches superfluous CONTINUEs -- should be unnecessary */
908 break; // now actually run the thread; DaH Qu'vam yImuHbej
911 do_the_fetchnode(event);
912 goto next_thread; /* handle next event in event queue */
915 do_the_globalblock(event);
916 goto next_thread; /* handle next event in event queue */
919 do_the_fetchreply(event);
920 goto next_thread; /* handle next event in event queue */
922 case UnblockThread: /* Move from the blocked queue to the tail of */
923 do_the_unblock(event);
924 goto next_thread; /* handle next event in event queue */
926 case ResumeThread: /* Move from the blocked queue to the tail of */
927 /* the runnable queue ( i.e. Qu' SImqa'lu') */
928 event->tso->gran.blocktime +=
929 CurrentTime[CurrentProc] - event->tso->gran.blockedat;
930 do_the_startthread(event);
931 goto next_thread; /* handle next event in event queue */
934 do_the_startthread(event);
935 goto next_thread; /* handle next event in event queue */
938 do_the_movethread(event);
939 goto next_thread; /* handle next event in event queue */
942 do_the_movespark(event);
943 goto next_thread; /* handle next event in event queue */
946 do_the_findwork(event);
947 goto next_thread; /* handle next event in event queue */
950 barf("Illegal event type %u\n", event->evttype);
953 /* This point was scheduler_loop in the old RTS */
955 IF_DEBUG(gran, debugBelch("GRAN: after main switch\n"));
957 TimeOfLastEvent = CurrentTime[CurrentProc];
958 TimeOfNextEvent = get_time_of_next_event();
959 IgnoreEvents=(TimeOfNextEvent==0); // HWL HACK
960 // CurrentTSO = ThreadQueueHd;
962 IF_DEBUG(gran, debugBelch("GRAN: time of next event is: %ld\n",
965 if (RtsFlags.GranFlags.Light)
966 GranSimLight_leave_system(event, &ActiveTSO);
968 EndOfTimeSlice = CurrentTime[CurrentProc]+RtsFlags.GranFlags.time_slice;
971 debugBelch("GRAN: end of time-slice is %#lx\n", EndOfTimeSlice));
973 /* in a GranSim setup the TSO stays on the run queue */
975 /* Take a thread from the run queue. */
976 POP_RUN_QUEUE(t); // take_off_run_queue(t);
979 debugBelch("GRAN: About to run current thread, which is\n");
982 context_switch = 0; // turned on via GranYield, checking events and time slice
985 DumpGranEvent(GR_SCHEDULE, t));
987 procStatus[CurrentProc] = Busy;
991 /* ----------------------------------------------------------------------------
992 * Send pending messages (PARALLEL_HASKELL only)
993 * ------------------------------------------------------------------------- */
995 #if defined(PARALLEL_HASKELL)
997 scheduleSendPendingMessages(void)
1003 # if defined(PAR) // global Mem.Mgmt., omit for now
1004 if (PendingFetches != END_BF_QUEUE) {
1009 if (RtsFlags.ParFlags.BufferTime) {
1010 // if we use message buffering, we must send away all message
1011 // packets which have become too old...
1017 /* ----------------------------------------------------------------------------
1018 * Activate spark threads (PARALLEL_HASKELL only)
1019 * ------------------------------------------------------------------------- */
1021 #if defined(PARALLEL_HASKELL)
1023 scheduleActivateSpark(void)
1026 ASSERT(emptyRunQueue());
1027 /* We get here if the run queue is empty and want some work.
1028 We try to turn a spark into a thread, and add it to the run queue,
1029 from where it will be picked up in the next iteration of the scheduler
1033 /* :-[ no local threads => look out for local sparks */
1034 /* the spark pool for the current PE */
1035 pool = &(cap.r.rSparks); // JB: cap = (old) MainCap
1036 if (advisory_thread_count < RtsFlags.ParFlags.maxThreads &&
1037 pool->hd < pool->tl) {
1039 * ToDo: add GC code check that we really have enough heap afterwards!!
1041 * If we're here (no runnable threads) and we have pending
1042 * sparks, we must have a space problem. Get enough space
1043 * to turn one of those pending sparks into a
1047 spark = findSpark(rtsFalse); /* get a spark */
1048 if (spark != (rtsSpark) NULL) {
1049 tso = createThreadFromSpark(spark); /* turn the spark into a thread */
1050 IF_PAR_DEBUG(fish, // schedule,
1051 debugBelch("==== schedule: Created TSO %d (%p); %d threads active\n",
1052 tso->id, tso, advisory_thread_count));
1054 if (tso==END_TSO_QUEUE) { /* failed to activate spark->back to loop */
1055 IF_PAR_DEBUG(fish, // schedule,
1056 debugBelch("==^^ failed to create thread from spark @ %lx\n",
1058 return rtsFalse; /* failed to generate a thread */
1059 } /* otherwise fall through & pick-up new tso */
1061 IF_PAR_DEBUG(fish, // schedule,
1062 debugBelch("==^^ no local sparks (spark pool contains only NFs: %d)\n",
1063 spark_queue_len(pool)));
1064 return rtsFalse; /* failed to generate a thread */
1066 return rtsTrue; /* success in generating a thread */
1067 } else { /* no more threads permitted or pool empty */
1068 return rtsFalse; /* failed to generateThread */
1071 tso = NULL; // avoid compiler warning only
1072 return rtsFalse; /* dummy in non-PAR setup */
1075 #endif // PARALLEL_HASKELL
1077 /* ----------------------------------------------------------------------------
1078 * Get work from a remote node (PARALLEL_HASKELL only)
1079 * ------------------------------------------------------------------------- */
1081 #if defined(PARALLEL_HASKELL)
1083 scheduleGetRemoteWork(rtsBool *receivedFinish)
1085 ASSERT(emptyRunQueue());
1087 if (RtsFlags.ParFlags.BufferTime) {
1088 IF_PAR_DEBUG(verbose,
1089 debugBelch("...send all pending data,"));
1092 for (i=1; i<=nPEs; i++)
1093 sendImmediately(i); // send all messages away immediately
1097 //++EDEN++ idle() , i.e. send all buffers, wait for work
1098 // suppress fishing in EDEN... just look for incoming messages
1099 // (blocking receive)
1100 IF_PAR_DEBUG(verbose,
1101 debugBelch("...wait for incoming messages...\n"));
1102 *receivedFinish = processMessages(); // blocking receive...
1104 // and reenter scheduling loop after having received something
1105 // (return rtsFalse below)
1107 # else /* activate SPARKS machinery */
1108 /* We get here, if we have no work, tried to activate a local spark, but still
1109 have no work. We try to get a remote spark, by sending a FISH message.
1110 Thread migration should be added here, and triggered when a sequence of
1111 fishes returns without work. */
1112 delay = (RtsFlags.ParFlags.fishDelay!=0ll ? RtsFlags.ParFlags.fishDelay : 0ll);
1114 /* =8-[ no local sparks => look for work on other PEs */
1116 * We really have absolutely no work. Send out a fish
1117 * (there may be some out there already), and wait for
1118 * something to arrive. We clearly can't run any threads
1119 * until a SCHEDULE or RESUME arrives, and so that's what
1120 * we're hoping to see. (Of course, we still have to
1121 * respond to other types of messages.)
1123 rtsTime now = msTime() /*CURRENT_TIME*/;
1124 IF_PAR_DEBUG(verbose,
1125 debugBelch("-- now=%ld\n", now));
1126 IF_PAR_DEBUG(fish, // verbose,
1127 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
1128 (last_fish_arrived_at!=0 &&
1129 last_fish_arrived_at+delay > now)) {
1130 debugBelch("--$$ <%llu> delaying FISH until %llu (last fish %llu, delay %llu)\n",
1131 now, last_fish_arrived_at+delay,
1132 last_fish_arrived_at,
1136 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
1137 advisory_thread_count < RtsFlags.ParFlags.maxThreads) { // send a FISH, but when?
1138 if (last_fish_arrived_at==0 ||
1139 (last_fish_arrived_at+delay <= now)) { // send FISH now!
1140 /* outstandingFishes is set in sendFish, processFish;
1141 avoid flooding system with fishes via delay */
1142 next_fish_to_send_at = 0;
1144 /* ToDo: this should be done in the main scheduling loop to avoid the
1145 busy wait here; not so bad if fish delay is very small */
1146 int iq = 0; // DEBUGGING -- HWL
1147 next_fish_to_send_at = last_fish_arrived_at+delay; // remember when to send
1148 /* send a fish when ready, but process messages that arrive in the meantime */
1150 if (PacketsWaiting()) {
1152 *receivedFinish = processMessages();
1155 } while (!*receivedFinish || now<next_fish_to_send_at);
1156 // JB: This means the fish could become obsolete, if we receive
1157 // work. Better check for work again?
1158 // last line: while (!receivedFinish || !haveWork || now<...)
1159 // next line: if (receivedFinish || haveWork )
1161 if (*receivedFinish) // no need to send a FISH if we are finishing anyway
1162 return rtsFalse; // NB: this will leave scheduler loop
1163 // immediately after return!
1165 IF_PAR_DEBUG(fish, // verbose,
1166 debugBelch("--$$ <%llu> sent delayed fish (%d processMessages); active/total threads=%d/%d\n",now,iq,run_queue_len(),advisory_thread_count));
1170 // JB: IMHO, this should all be hidden inside sendFish(...)
1172 sendFish(pe, thisPE, NEW_FISH_AGE, NEW_FISH_HISTORY,
1175 // Global statistics: count no. of fishes
1176 if (RtsFlags.ParFlags.ParStats.Global &&
1177 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
1178 globalParStats.tot_fish_mess++;
1182 /* delayed fishes must have been sent by now! */
1183 next_fish_to_send_at = 0;
1186 *receivedFinish = processMessages();
1187 # endif /* SPARKS */
1190 /* NB: this function always returns rtsFalse, meaning the scheduler
1191 loop continues with the next iteration;
1193 return code means success in finding work; we enter this function
1194 if there is no local work, thus have to send a fish which takes
1195 time until it arrives with work; in the meantime we should process
1196 messages in the main loop;
1199 #endif // PARALLEL_HASKELL
1201 /* ----------------------------------------------------------------------------
1202 * PAR/GRAN: Report stats & debugging info(?)
1203 * ------------------------------------------------------------------------- */
1205 #if defined(PAR) || defined(GRAN)
1207 scheduleGranParReport(void)
1209 ASSERT(run_queue_hd != END_TSO_QUEUE);
1211 /* Take a thread from the run queue, if we have work */
1212 POP_RUN_QUEUE(t); // take_off_run_queue(END_TSO_QUEUE);
1214 /* If this TSO has got its outport closed in the meantime,
1215 * it mustn't be run. Instead, we have to clean it up as if it was finished.
1216 * It has to be marked as TH_DEAD for this purpose.
1217 * If it is TH_TERM instead, it is supposed to have finished in the normal way.
1219 JB: TODO: investigate wether state change field could be nuked
1220 entirely and replaced by the normal tso state (whatnext
1221 field). All we want to do is to kill tsos from outside.
1224 /* ToDo: write something to the log-file
1225 if (RTSflags.ParFlags.granSimStats && !sameThread)
1226 DumpGranEvent(GR_SCHEDULE, RunnableThreadsHd);
1230 /* the spark pool for the current PE */
1231 pool = &(cap.r.rSparks); // cap = (old) MainCap
1234 debugBelch("--=^ %d threads, %d sparks on [%#x]\n",
1235 run_queue_len(), spark_queue_len(pool), CURRENT_PROC));
1238 debugBelch("--=^ %d threads, %d sparks on [%#x]\n",
1239 run_queue_len(), spark_queue_len(pool), CURRENT_PROC));
1241 if (RtsFlags.ParFlags.ParStats.Full &&
1242 (t->par.sparkname != (StgInt)0) && // only log spark generated threads
1243 (emitSchedule || // forced emit
1244 (t && LastTSO && t->id != LastTSO->id))) {
1246 we are running a different TSO, so write a schedule event to log file
1247 NB: If we use fair scheduling we also have to write a deschedule
1248 event for LastTSO; with unfair scheduling we know that the
1249 previous tso has blocked whenever we switch to another tso, so
1250 we don't need it in GUM for now
1252 IF_PAR_DEBUG(fish, // schedule,
1253 debugBelch("____ scheduling spark generated thread %d (%lx) (%lx) via a forced emit\n",t->id,t,t->par.sparkname));
1255 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
1256 GR_SCHEDULE, t, (StgClosure *)NULL, 0, 0);
1257 emitSchedule = rtsFalse;
1262 /* ----------------------------------------------------------------------------
1263 * After running a thread...
1264 * ------------------------------------------------------------------------- */
1267 schedulePostRunThread(void)
1270 /* HACK 675: if the last thread didn't yield, make sure to print a
1271 SCHEDULE event to the log file when StgRunning the next thread, even
1272 if it is the same one as before */
1274 TimeOfLastYield = CURRENT_TIME;
1277 /* some statistics gathering in the parallel case */
1279 #if defined(GRAN) || defined(PAR) || defined(EDEN)
1283 IF_DEBUG(gran, DumpGranEvent(GR_DESCHEDULE, t));
1284 globalGranStats.tot_heapover++;
1286 globalParStats.tot_heapover++;
1293 DumpGranEvent(GR_DESCHEDULE, t));
1294 globalGranStats.tot_stackover++;
1297 // DumpGranEvent(GR_DESCHEDULE, t);
1298 globalParStats.tot_stackover++;
1302 case ThreadYielding:
1305 DumpGranEvent(GR_DESCHEDULE, t));
1306 globalGranStats.tot_yields++;
1309 // DumpGranEvent(GR_DESCHEDULE, t);
1310 globalParStats.tot_yields++;
1317 debugBelch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: ",
1318 t->id, t, whatNext_strs[t->what_next], t->block_info.closure,
1319 (t->block_info.closure==(StgClosure*)NULL ? 99 : where_is(t->block_info.closure)));
1320 if (t->block_info.closure!=(StgClosure*)NULL)
1321 print_bq(t->block_info.closure);
1324 // ??? needed; should emit block before
1326 DumpGranEvent(GR_DESCHEDULE, t));
1327 prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
1330 ASSERT(procStatus[CurrentProc]==Busy ||
1331 ((procStatus[CurrentProc]==Fetching) &&
1332 (t->block_info.closure!=(StgClosure*)NULL)));
1333 if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
1334 !(!RtsFlags.GranFlags.DoAsyncFetch &&
1335 procStatus[CurrentProc]==Fetching))
1336 procStatus[CurrentProc] = Idle;
1339 //++PAR++ blockThread() writes the event (change?)
1343 case ThreadFinished:
1347 barf("parGlobalStats: unknown return code");
1353 /* -----------------------------------------------------------------------------
1354 * Handle a thread that returned to the scheduler with ThreadHeepOverflow
1355 * -------------------------------------------------------------------------- */
1358 scheduleHandleHeapOverflow( Capability *cap, StgTSO *t )
1360 // did the task ask for a large block?
1361 if (cap->r.rHpAlloc > BLOCK_SIZE) {
1362 // if so, get one and push it on the front of the nursery.
1366 blocks = (lnat)BLOCK_ROUND_UP(cap->r.rHpAlloc) / BLOCK_SIZE;
1369 debugBelch("--<< thread %ld (%s) stopped: requesting a large block (size %ld)\n",
1370 (long)t->id, whatNext_strs[t->what_next], blocks));
1372 // don't do this if the nursery is (nearly) full, we'll GC first.
1373 if (cap->r.rCurrentNursery->link != NULL ||
1374 cap->r.rNursery->n_blocks == 1) { // paranoia to prevent infinite loop
1375 // if the nursery has only one block.
1378 bd = allocGroup( blocks );
1380 cap->r.rNursery->n_blocks += blocks;
1382 // link the new group into the list
1383 bd->link = cap->r.rCurrentNursery;
1384 bd->u.back = cap->r.rCurrentNursery->u.back;
1385 if (cap->r.rCurrentNursery->u.back != NULL) {
1386 cap->r.rCurrentNursery->u.back->link = bd;
1389 ASSERT(g0s0->blocks == cap->r.rCurrentNursery &&
1390 g0s0 == cap->r.rNursery);
1392 cap->r.rNursery->blocks = bd;
1394 cap->r.rCurrentNursery->u.back = bd;
1396 // initialise it as a nursery block. We initialise the
1397 // step, gen_no, and flags field of *every* sub-block in
1398 // this large block, because this is easier than making
1399 // sure that we always find the block head of a large
1400 // block whenever we call Bdescr() (eg. evacuate() and
1401 // isAlive() in the GC would both have to do this, at
1405 for (x = bd; x < bd + blocks; x++) {
1406 x->step = cap->r.rNursery;
1412 // This assert can be a killer if the app is doing lots
1413 // of large block allocations.
1414 IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
1416 // now update the nursery to point to the new block
1417 cap->r.rCurrentNursery = bd;
1419 // we might be unlucky and have another thread get on the
1420 // run queue before us and steal the large block, but in that
1421 // case the thread will just end up requesting another large
1423 pushOnRunQueue(cap,t);
1424 return rtsFalse; /* not actually GC'ing */
1429 debugBelch("--<< thread %ld (%s) stopped: HeapOverflow\n",
1430 (long)t->id, whatNext_strs[t->what_next]));
1432 ASSERT(!is_on_queue(t,CurrentProc));
1433 #elif defined(PARALLEL_HASKELL)
1434 /* Currently we emit a DESCHEDULE event before GC in GUM.
1435 ToDo: either add separate event to distinguish SYSTEM time from rest
1436 or just nuke this DESCHEDULE (and the following SCHEDULE) */
1437 if (0 && RtsFlags.ParFlags.ParStats.Full) {
1438 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
1439 GR_DESCHEDULE, t, (StgClosure *)NULL, 0, 0);
1440 emitSchedule = rtsTrue;
1444 pushOnRunQueue(cap,t);
1446 /* actual GC is done at the end of the while loop in schedule() */
1449 /* -----------------------------------------------------------------------------
1450 * Handle a thread that returned to the scheduler with ThreadStackOverflow
1451 * -------------------------------------------------------------------------- */
1454 scheduleHandleStackOverflow (Capability *cap, Task *task, StgTSO *t)
1456 IF_DEBUG(scheduler,debugBelch("--<< thread %ld (%s) stopped, StackOverflow\n",
1457 (long)t->id, whatNext_strs[t->what_next]));
1458 /* just adjust the stack for this thread, then pop it back
1462 /* enlarge the stack */
1463 StgTSO *new_t = threadStackOverflow(cap, t);
1465 /* This TSO has moved, so update any pointers to it from the
1466 * main thread stack. It better not be on any other queues...
1467 * (it shouldn't be).
1469 if (task->tso != NULL) {
1472 pushOnRunQueue(cap,new_t);
1476 /* -----------------------------------------------------------------------------
1477 * Handle a thread that returned to the scheduler with ThreadYielding
1478 * -------------------------------------------------------------------------- */
1481 scheduleHandleYield( Capability *cap, StgTSO *t, nat prev_what_next )
1483 // Reset the context switch flag. We don't do this just before
1484 // running the thread, because that would mean we would lose ticks
1485 // during GC, which can lead to unfair scheduling (a thread hogs
1486 // the CPU because the tick always arrives during GC). This way
1487 // penalises threads that do a lot of allocation, but that seems
1488 // better than the alternative.
1491 /* put the thread back on the run queue. Then, if we're ready to
1492 * GC, check whether this is the last task to stop. If so, wake
1493 * up the GC thread. getThread will block during a GC until the
1497 if (t->what_next != prev_what_next) {
1498 debugBelch("--<< thread %ld (%s) stopped to switch evaluators\n",
1499 (long)t->id, whatNext_strs[t->what_next]);
1501 debugBelch("--<< thread %ld (%s) stopped, yielding\n",
1502 (long)t->id, whatNext_strs[t->what_next]);
1507 //debugBelch("&& Doing sanity check on yielding TSO %ld.", t->id);
1509 ASSERT(t->link == END_TSO_QUEUE);
1511 // Shortcut if we're just switching evaluators: don't bother
1512 // doing stack squeezing (which can be expensive), just run the
1514 if (t->what_next != prev_what_next) {
1519 ASSERT(!is_on_queue(t,CurrentProc));
1522 //debugBelch("&& Doing sanity check on all ThreadQueues (and their TSOs).");
1523 checkThreadQsSanity(rtsTrue));
1527 addToRunQueue(cap,t);
1530 /* add a ContinueThread event to actually process the thread */
1531 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1533 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1535 debugBelch("GRAN: eventq and runnableq after adding yielded thread to queue again:\n");
1542 /* -----------------------------------------------------------------------------
1543 * Handle a thread that returned to the scheduler with ThreadBlocked
1544 * -------------------------------------------------------------------------- */
1547 scheduleHandleThreadBlocked( StgTSO *t
1548 #if !defined(GRAN) && !defined(DEBUG)
1555 debugBelch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: \n",
1556 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)));
1557 if (t->block_info.closure!=(StgClosure*)NULL) print_bq(t->block_info.closure));
1559 // ??? needed; should emit block before
1561 DumpGranEvent(GR_DESCHEDULE, t));
1562 prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
1565 ASSERT(procStatus[CurrentProc]==Busy ||
1566 ((procStatus[CurrentProc]==Fetching) &&
1567 (t->block_info.closure!=(StgClosure*)NULL)));
1568 if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
1569 !(!RtsFlags.GranFlags.DoAsyncFetch &&
1570 procStatus[CurrentProc]==Fetching))
1571 procStatus[CurrentProc] = Idle;
1575 debugBelch("--<< thread %ld (%p; %s) stopped, blocking on node %p with BQ: \n",
1576 t->id, t, whatNext_strs[t->what_next], t->block_info.closure));
1579 if (t->block_info.closure!=(StgClosure*)NULL)
1580 print_bq(t->block_info.closure));
1582 /* Send a fetch (if BlockedOnGA) and dump event to log file */
1585 /* whatever we schedule next, we must log that schedule */
1586 emitSchedule = rtsTrue;
1590 // We don't need to do anything. The thread is blocked, and it
1591 // has tidied up its stack and placed itself on whatever queue
1592 // it needs to be on.
1595 ASSERT(t->why_blocked != NotBlocked);
1596 // This might not be true under SMP: we don't have
1597 // exclusive access to this TSO, so someone might have
1598 // woken it up by now. This actually happens: try
1599 // conc023 +RTS -N2.
1603 debugBelch("--<< thread %d (%s) stopped: ",
1604 t->id, whatNext_strs[t->what_next]);
1605 printThreadBlockage(t);
1608 /* Only for dumping event to log file
1609 ToDo: do I need this in GranSim, too?
1615 /* -----------------------------------------------------------------------------
1616 * Handle a thread that returned to the scheduler with ThreadFinished
1617 * -------------------------------------------------------------------------- */
1620 scheduleHandleThreadFinished (Capability *cap STG_UNUSED, Task *task, StgTSO *t)
1622 /* Need to check whether this was a main thread, and if so,
1623 * return with the return value.
1625 * We also end up here if the thread kills itself with an
1626 * uncaught exception, see Exception.cmm.
1628 IF_DEBUG(scheduler,debugBelch("--++ thread %d (%s) finished\n",
1629 t->id, whatNext_strs[t->what_next]));
1632 endThread(t, CurrentProc); // clean-up the thread
1633 #elif defined(PARALLEL_HASKELL)
1634 /* For now all are advisory -- HWL */
1635 //if(t->priority==AdvisoryPriority) ??
1636 advisory_thread_count--; // JB: Caution with this counter, buggy!
1639 if(t->dist.priority==RevalPriority)
1643 # if defined(EDENOLD)
1644 // the thread could still have an outport... (BUG)
1645 if (t->eden.outport != -1) {
1646 // delete the outport for the tso which has finished...
1647 IF_PAR_DEBUG(eden_ports,
1648 debugBelch("WARNING: Scheduler removes outport %d for TSO %d.\n",
1649 t->eden.outport, t->id));
1652 // thread still in the process (HEAVY BUG! since outport has just been closed...)
1653 if (t->eden.epid != -1) {
1654 IF_PAR_DEBUG(eden_ports,
1655 debugBelch("WARNING: Scheduler removes TSO %d from process %d .\n",
1656 t->id, t->eden.epid));
1657 removeTSOfromProcess(t);
1662 if (RtsFlags.ParFlags.ParStats.Full &&
1663 !RtsFlags.ParFlags.ParStats.Suppressed)
1664 DumpEndEvent(CURRENT_PROC, t, rtsFalse /* not mandatory */);
1666 // t->par only contains statistics: left out for now...
1668 debugBelch("**** end thread: ended sparked thread %d (%lx); sparkname: %lx\n",
1669 t->id,t,t->par.sparkname));
1671 #endif // PARALLEL_HASKELL
1674 // Check whether the thread that just completed was a bound
1675 // thread, and if so return with the result.
1677 // There is an assumption here that all thread completion goes
1678 // through this point; we need to make sure that if a thread
1679 // ends up in the ThreadKilled state, that it stays on the run
1680 // queue so it can be dealt with here.
1685 if (t->bound != task) {
1686 #if !defined(THREADED_RTS)
1687 // Must be a bound thread that is not the topmost one. Leave
1688 // it on the run queue until the stack has unwound to the
1689 // point where we can deal with this. Leaving it on the run
1690 // queue also ensures that the garbage collector knows about
1691 // this thread and its return value (it gets dropped from the
1692 // all_threads list so there's no other way to find it).
1693 appendToRunQueue(cap,t);
1696 // this cannot happen in the threaded RTS, because a
1697 // bound thread can only be run by the appropriate Task.
1698 barf("finished bound thread that isn't mine");
1702 ASSERT(task->tso == t);
1704 if (t->what_next == ThreadComplete) {
1706 // NOTE: return val is tso->sp[1] (see StgStartup.hc)
1707 *(task->ret) = (StgClosure *)task->tso->sp[1];
1709 task->stat = Success;
1712 *(task->ret) = NULL;
1715 task->stat = Interrupted;
1717 task->stat = Killed;
1721 removeThreadLabel((StgWord)task->tso->id);
1723 return rtsTrue; // tells schedule() to return
1729 /* -----------------------------------------------------------------------------
1730 * Perform a heap census, if PROFILING
1731 * -------------------------------------------------------------------------- */
1734 scheduleDoHeapProfile( rtsBool ready_to_gc STG_UNUSED )
1736 #if defined(PROFILING)
1737 // When we have +RTS -i0 and we're heap profiling, do a census at
1738 // every GC. This lets us get repeatable runs for debugging.
1739 if (performHeapProfile ||
1740 (RtsFlags.ProfFlags.profileInterval==0 &&
1741 RtsFlags.ProfFlags.doHeapProfile && ready_to_gc)) {
1742 GarbageCollect(GetRoots, rtsTrue);
1744 performHeapProfile = rtsFalse;
1745 return rtsTrue; // true <=> we already GC'd
1751 /* -----------------------------------------------------------------------------
1752 * Perform a garbage collection if necessary
1753 * -------------------------------------------------------------------------- */
1756 scheduleDoGC( Capability *cap, Task *task USED_WHEN_SMP, rtsBool force_major )
1760 static volatile StgWord waiting_for_gc;
1761 rtsBool was_waiting;
1766 // In order to GC, there must be no threads running Haskell code.
1767 // Therefore, the GC thread needs to hold *all* the capabilities,
1768 // and release them after the GC has completed.
1770 // This seems to be the simplest way: previous attempts involved
1771 // making all the threads with capabilities give up their
1772 // capabilities and sleep except for the *last* one, which
1773 // actually did the GC. But it's quite hard to arrange for all
1774 // the other tasks to sleep and stay asleep.
1777 was_waiting = cas(&waiting_for_gc, 0, 1);
1778 if (was_waiting) return;
1780 for (i=0; i < n_capabilities; i++) {
1781 IF_DEBUG(scheduler, sched_belch("ready_to_gc, grabbing all the capabilies (%d/%d)", i, n_capabilities));
1782 if (cap != &capabilities[i]) {
1783 Capability *pcap = &capabilities[i];
1784 // we better hope this task doesn't get migrated to
1785 // another Capability while we're waiting for this one.
1786 // It won't, because load balancing happens while we have
1787 // all the Capabilities, but even so it's a slightly
1788 // unsavoury invariant.
1790 waitForReturnCapability(&pcap, task);
1791 if (pcap != &capabilities[i]) {
1792 barf("scheduleDoGC: got the wrong capability");
1797 waiting_for_gc = rtsFalse;
1800 /* Kick any transactions which are invalid back to their
1801 * atomically frames. When next scheduled they will try to
1802 * commit, this commit will fail and they will retry.
1807 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
1808 if (t->what_next == ThreadRelocated) {
1811 next = t->global_link;
1812 if (t -> trec != NO_TREC && t -> why_blocked == NotBlocked) {
1813 if (!stmValidateNestOfTransactions (t -> trec)) {
1814 IF_DEBUG(stm, sched_belch("trec %p found wasting its time", t));
1816 // strip the stack back to the
1817 // ATOMICALLY_FRAME, aborting the (nested)
1818 // transaction, and saving the stack of any
1819 // partially-evaluated thunks on the heap.
1820 raiseAsync_(cap, t, NULL, rtsTrue);
1823 ASSERT(get_itbl((StgClosure *)t->sp)->type == ATOMICALLY_FRAME);
1831 // so this happens periodically:
1832 scheduleCheckBlackHoles(cap);
1834 IF_DEBUG(scheduler, printAllThreads());
1836 /* everybody back, start the GC.
1837 * Could do it in this thread, or signal a condition var
1838 * to do it in another thread. Either way, we need to
1839 * broadcast on gc_pending_cond afterward.
1841 #if defined(THREADED_RTS)
1842 IF_DEBUG(scheduler,sched_belch("doing GC"));
1844 GarbageCollect(GetRoots, force_major);
1847 // release our stash of capabilities.
1848 for (i = 0; i < n_capabilities; i++) {
1849 if (cap != &capabilities[i]) {
1850 task->cap = &capabilities[i];
1851 releaseCapability(&capabilities[i]);
1858 /* add a ContinueThread event to continue execution of current thread */
1859 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1861 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1863 debugBelch("GRAN: eventq and runnableq after Garbage collection:\n\n");
1869 /* ---------------------------------------------------------------------------
1870 * rtsSupportsBoundThreads(): is the RTS built to support bound threads?
1871 * used by Control.Concurrent for error checking.
1872 * ------------------------------------------------------------------------- */
1875 rtsSupportsBoundThreads(void)
1877 #if defined(THREADED_RTS)
1884 /* ---------------------------------------------------------------------------
1885 * isThreadBound(tso): check whether tso is bound to an OS thread.
1886 * ------------------------------------------------------------------------- */
1889 isThreadBound(StgTSO* tso USED_WHEN_THREADED_RTS)
1891 #if defined(THREADED_RTS)
1892 return (tso->bound != NULL);
1897 /* ---------------------------------------------------------------------------
1898 * Singleton fork(). Do not copy any running threads.
1899 * ------------------------------------------------------------------------- */
1901 #if !defined(mingw32_HOST_OS) && !defined(SMP)
1902 #define FORKPROCESS_PRIMOP_SUPPORTED
1905 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
1907 deleteThreadImmediately(Capability *cap, StgTSO *tso);
1910 forkProcess(HsStablePtr *entry
1911 #ifndef FORKPROCESS_PRIMOP_SUPPORTED
1916 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
1922 IF_DEBUG(scheduler,sched_belch("forking!"));
1924 // ToDo: for SMP, we should probably acquire *all* the capabilities
1929 if (pid) { // parent
1931 // just return the pid
1937 // delete all threads
1938 cap->run_queue_hd = END_TSO_QUEUE;
1939 cap->run_queue_tl = END_TSO_QUEUE;
1941 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
1944 // don't allow threads to catch the ThreadKilled exception
1945 deleteThreadImmediately(cap,t);
1948 // wipe the main thread list
1949 while ((task = all_tasks) != NULL) {
1950 all_tasks = task->all_link;
1954 cap = rts_evalStableIO(cap, entry, NULL); // run the action
1955 rts_checkSchedStatus("forkProcess",cap);
1958 hs_exit(); // clean up and exit
1959 stg_exit(EXIT_SUCCESS);
1961 #else /* !FORKPROCESS_PRIMOP_SUPPORTED */
1962 barf("forkProcess#: primop not supported on this platform, sorry!\n");
1967 /* ---------------------------------------------------------------------------
1968 * Delete the threads on the run queue of the current capability.
1969 * ------------------------------------------------------------------------- */
1972 deleteRunQueue (Capability *cap)
1975 for (t = cap->run_queue_hd; t != END_TSO_QUEUE; t = next) {
1976 ASSERT(t->what_next != ThreadRelocated);
1978 deleteThread(cap, t);
1982 /* startThread and insertThread are now in GranSim.c -- HWL */
1985 /* -----------------------------------------------------------------------------
1986 Managing the suspended_ccalling_tasks list.
1987 Locks required: sched_mutex
1988 -------------------------------------------------------------------------- */
1991 suspendTask (Capability *cap, Task *task)
1993 ASSERT(task->next == NULL && task->prev == NULL);
1994 task->next = cap->suspended_ccalling_tasks;
1996 if (cap->suspended_ccalling_tasks) {
1997 cap->suspended_ccalling_tasks->prev = task;
1999 cap->suspended_ccalling_tasks = task;
2003 recoverSuspendedTask (Capability *cap, Task *task)
2006 task->prev->next = task->next;
2008 ASSERT(cap->suspended_ccalling_tasks == task);
2009 cap->suspended_ccalling_tasks = task->next;
2012 task->next->prev = task->prev;
2014 task->next = task->prev = NULL;
2017 /* ---------------------------------------------------------------------------
2018 * Suspending & resuming Haskell threads.
2020 * When making a "safe" call to C (aka _ccall_GC), the task gives back
2021 * its capability before calling the C function. This allows another
2022 * task to pick up the capability and carry on running Haskell
2023 * threads. It also means that if the C call blocks, it won't lock
2026 * The Haskell thread making the C call is put to sleep for the
2027 * duration of the call, on the susepended_ccalling_threads queue. We
2028 * give out a token to the task, which it can use to resume the thread
2029 * on return from the C function.
2030 * ------------------------------------------------------------------------- */
2033 suspendThread (StgRegTable *reg)
2036 int saved_errno = errno;
2040 /* assume that *reg is a pointer to the StgRegTable part of a Capability.
2042 cap = regTableToCapability(reg);
2044 task = cap->running_task;
2045 tso = cap->r.rCurrentTSO;
2048 sched_belch("thread %d did a safe foreign call", cap->r.rCurrentTSO->id));
2050 // XXX this might not be necessary --SDM
2051 tso->what_next = ThreadRunGHC;
2055 if(tso->blocked_exceptions == NULL) {
2056 tso->why_blocked = BlockedOnCCall;
2057 tso->blocked_exceptions = END_TSO_QUEUE;
2059 tso->why_blocked = BlockedOnCCall_NoUnblockExc;
2062 // Hand back capability
2063 task->suspended_tso = tso;
2065 ACQUIRE_LOCK(&cap->lock);
2067 suspendTask(cap,task);
2068 cap->in_haskell = rtsFalse;
2069 releaseCapability_(cap);
2071 RELEASE_LOCK(&cap->lock);
2073 #if defined(THREADED_RTS)
2074 /* Preparing to leave the RTS, so ensure there's a native thread/task
2075 waiting to take over.
2077 IF_DEBUG(scheduler, sched_belch("thread %d: leaving RTS", tso->id));
2080 errno = saved_errno;
2085 resumeThread (void *task_)
2089 int saved_errno = errno;
2093 // Wait for permission to re-enter the RTS with the result.
2094 waitForReturnCapability(&cap,task);
2095 // we might be on a different capability now... but if so, our
2096 // entry on the suspended_ccalling_tasks list will also have been
2099 // Remove the thread from the suspended list
2100 recoverSuspendedTask(cap,task);
2102 tso = task->suspended_tso;
2103 task->suspended_tso = NULL;
2104 tso->link = END_TSO_QUEUE;
2105 IF_DEBUG(scheduler, sched_belch("thread %d: re-entering RTS", tso->id));
2107 if (tso->why_blocked == BlockedOnCCall) {
2108 awakenBlockedQueue(cap,tso->blocked_exceptions);
2109 tso->blocked_exceptions = NULL;
2112 /* Reset blocking status */
2113 tso->why_blocked = NotBlocked;
2115 cap->r.rCurrentTSO = tso;
2116 cap->in_haskell = rtsTrue;
2117 errno = saved_errno;
2122 /* ---------------------------------------------------------------------------
2123 * Comparing Thread ids.
2125 * This is used from STG land in the implementation of the
2126 * instances of Eq/Ord for ThreadIds.
2127 * ------------------------------------------------------------------------ */
2130 cmp_thread(StgPtr tso1, StgPtr tso2)
2132 StgThreadID id1 = ((StgTSO *)tso1)->id;
2133 StgThreadID id2 = ((StgTSO *)tso2)->id;
2135 if (id1 < id2) return (-1);
2136 if (id1 > id2) return 1;
2140 /* ---------------------------------------------------------------------------
2141 * Fetching the ThreadID from an StgTSO.
2143 * This is used in the implementation of Show for ThreadIds.
2144 * ------------------------------------------------------------------------ */
2146 rts_getThreadId(StgPtr tso)
2148 return ((StgTSO *)tso)->id;
2153 labelThread(StgPtr tso, char *label)
2158 /* Caveat: Once set, you can only set the thread name to "" */
2159 len = strlen(label)+1;
2160 buf = stgMallocBytes(len * sizeof(char), "Schedule.c:labelThread()");
2161 strncpy(buf,label,len);
2162 /* Update will free the old memory for us */
2163 updateThreadLabel(((StgTSO *)tso)->id,buf);
2167 /* ---------------------------------------------------------------------------
2168 Create a new thread.
2170 The new thread starts with the given stack size. Before the
2171 scheduler can run, however, this thread needs to have a closure
2172 (and possibly some arguments) pushed on its stack. See
2173 pushClosure() in Schedule.h.
2175 createGenThread() and createIOThread() (in SchedAPI.h) are
2176 convenient packaged versions of this function.
2178 currently pri (priority) is only used in a GRAN setup -- HWL
2179 ------------------------------------------------------------------------ */
2181 /* currently pri (priority) is only used in a GRAN setup -- HWL */
2183 createThread(nat size, StgInt pri)
2186 createThread(Capability *cap, nat size)
2192 /* sched_mutex is *not* required */
2194 /* First check whether we should create a thread at all */
2195 #if defined(PARALLEL_HASKELL)
2196 /* check that no more than RtsFlags.ParFlags.maxThreads threads are created */
2197 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads) {
2199 debugBelch("{createThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)\n",
2200 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
2201 return END_TSO_QUEUE;
2207 ASSERT(!RtsFlags.GranFlags.Light || CurrentProc==0);
2210 // ToDo: check whether size = stack_size - TSO_STRUCT_SIZEW
2212 /* catch ridiculously small stack sizes */
2213 if (size < MIN_STACK_WORDS + TSO_STRUCT_SIZEW) {
2214 size = MIN_STACK_WORDS + TSO_STRUCT_SIZEW;
2217 stack_size = size - TSO_STRUCT_SIZEW;
2219 tso = (StgTSO *)allocateLocal(cap, size);
2220 TICK_ALLOC_TSO(stack_size, 0);
2222 SET_HDR(tso, &stg_TSO_info, CCS_SYSTEM);
2224 SET_GRAN_HDR(tso, ThisPE);
2227 // Always start with the compiled code evaluator
2228 tso->what_next = ThreadRunGHC;
2230 tso->why_blocked = NotBlocked;
2231 tso->blocked_exceptions = NULL;
2233 tso->saved_errno = 0;
2236 tso->stack_size = stack_size;
2237 tso->max_stack_size = round_to_mblocks(RtsFlags.GcFlags.maxStkSize)
2239 tso->sp = (P_)&(tso->stack) + stack_size;
2241 tso->trec = NO_TREC;
2244 tso->prof.CCCS = CCS_MAIN;
2247 /* put a stop frame on the stack */
2248 tso->sp -= sizeofW(StgStopFrame);
2249 SET_HDR((StgClosure*)tso->sp,(StgInfoTable *)&stg_stop_thread_info,CCS_SYSTEM);
2250 tso->link = END_TSO_QUEUE;
2254 /* uses more flexible routine in GranSim */
2255 insertThread(tso, CurrentProc);
2257 /* In a non-GranSim setup the pushing of a TSO onto the runq is separated
2263 if (RtsFlags.GranFlags.GranSimStats.Full)
2264 DumpGranEvent(GR_START,tso);
2265 #elif defined(PARALLEL_HASKELL)
2266 if (RtsFlags.ParFlags.ParStats.Full)
2267 DumpGranEvent(GR_STARTQ,tso);
2268 /* HACk to avoid SCHEDULE
2272 /* Link the new thread on the global thread list.
2274 ACQUIRE_LOCK(&sched_mutex);
2275 tso->id = next_thread_id++; // while we have the mutex
2276 tso->global_link = all_threads;
2278 RELEASE_LOCK(&sched_mutex);
2281 tso->dist.priority = MandatoryPriority; //by default that is...
2285 tso->gran.pri = pri;
2287 tso->gran.magic = TSO_MAGIC; // debugging only
2289 tso->gran.sparkname = 0;
2290 tso->gran.startedat = CURRENT_TIME;
2291 tso->gran.exported = 0;
2292 tso->gran.basicblocks = 0;
2293 tso->gran.allocs = 0;
2294 tso->gran.exectime = 0;
2295 tso->gran.fetchtime = 0;
2296 tso->gran.fetchcount = 0;
2297 tso->gran.blocktime = 0;
2298 tso->gran.blockcount = 0;
2299 tso->gran.blockedat = 0;
2300 tso->gran.globalsparks = 0;
2301 tso->gran.localsparks = 0;
2302 if (RtsFlags.GranFlags.Light)
2303 tso->gran.clock = Now; /* local clock */
2305 tso->gran.clock = 0;
2307 IF_DEBUG(gran,printTSO(tso));
2308 #elif defined(PARALLEL_HASKELL)
2310 tso->par.magic = TSO_MAGIC; // debugging only
2312 tso->par.sparkname = 0;
2313 tso->par.startedat = CURRENT_TIME;
2314 tso->par.exported = 0;
2315 tso->par.basicblocks = 0;
2316 tso->par.allocs = 0;
2317 tso->par.exectime = 0;
2318 tso->par.fetchtime = 0;
2319 tso->par.fetchcount = 0;
2320 tso->par.blocktime = 0;
2321 tso->par.blockcount = 0;
2322 tso->par.blockedat = 0;
2323 tso->par.globalsparks = 0;
2324 tso->par.localsparks = 0;
2328 globalGranStats.tot_threads_created++;
2329 globalGranStats.threads_created_on_PE[CurrentProc]++;
2330 globalGranStats.tot_sq_len += spark_queue_len(CurrentProc);
2331 globalGranStats.tot_sq_probes++;
2332 #elif defined(PARALLEL_HASKELL)
2333 // collect parallel global statistics (currently done together with GC stats)
2334 if (RtsFlags.ParFlags.ParStats.Global &&
2335 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
2336 //debugBelch("Creating thread %d @ %11.2f\n", tso->id, usertime());
2337 globalParStats.tot_threads_created++;
2343 sched_belch("==__ schedule: Created TSO %d (%p);",
2344 CurrentProc, tso, tso->id));
2345 #elif defined(PARALLEL_HASKELL)
2346 IF_PAR_DEBUG(verbose,
2347 sched_belch("==__ schedule: Created TSO %d (%p); %d threads active",
2348 (long)tso->id, tso, advisory_thread_count));
2350 IF_DEBUG(scheduler,sched_belch("created thread %ld, stack size = %lx words",
2351 (long)tso->id, (long)tso->stack_size));
2358 all parallel thread creation calls should fall through the following routine.
2361 createThreadFromSpark(rtsSpark spark)
2363 ASSERT(spark != (rtsSpark)NULL);
2364 // JB: TAKE CARE OF THIS COUNTER! BUGGY
2365 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads)
2367 barf("{createSparkThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
2368 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
2369 return END_TSO_QUEUE;
2373 tso = createThread(RtsFlags.GcFlags.initialStkSize);
2374 if (tso==END_TSO_QUEUE)
2375 barf("createSparkThread: Cannot create TSO");
2377 tso->priority = AdvisoryPriority;
2379 pushClosure(tso,spark);
2381 advisory_thread_count++; // JB: TAKE CARE OF THIS COUNTER! BUGGY
2388 Turn a spark into a thread.
2389 ToDo: fix for SMP (needs to acquire SCHED_MUTEX!)
2393 activateSpark (rtsSpark spark)
2397 tso = createSparkThread(spark);
2398 if (RtsFlags.ParFlags.ParStats.Full) {
2399 //ASSERT(run_queue_hd == END_TSO_QUEUE); // I think ...
2400 IF_PAR_DEBUG(verbose,
2401 debugBelch("==^^ activateSpark: turning spark of closure %p (%s) into a thread\n",
2402 (StgClosure *)spark, info_type((StgClosure *)spark)));
2404 // ToDo: fwd info on local/global spark to thread -- HWL
2405 // tso->gran.exported = spark->exported;
2406 // tso->gran.locked = !spark->global;
2407 // tso->gran.sparkname = spark->name;
2413 /* ---------------------------------------------------------------------------
2416 * scheduleThread puts a thread on the end of the runnable queue.
2417 * This will usually be done immediately after a thread is created.
2418 * The caller of scheduleThread must create the thread using e.g.
2419 * createThread and push an appropriate closure
2420 * on this thread's stack before the scheduler is invoked.
2421 * ------------------------------------------------------------------------ */
2424 scheduleThread(Capability *cap, StgTSO *tso)
2426 // The thread goes at the *end* of the run-queue, to avoid possible
2427 // starvation of any threads already on the queue.
2428 appendToRunQueue(cap,tso);
2432 scheduleWaitThread (StgTSO* tso, /*[out]*/HaskellObj* ret, Capability *cap)
2436 // We already created/initialised the Task
2437 task = cap->running_task;
2439 // This TSO is now a bound thread; make the Task and TSO
2440 // point to each other.
2445 task->stat = NoStatus;
2447 appendToRunQueue(cap,tso);
2449 IF_DEBUG(scheduler, sched_belch("new bound thread (%d)", tso->id));
2452 /* GranSim specific init */
2453 CurrentTSO = m->tso; // the TSO to run
2454 procStatus[MainProc] = Busy; // status of main PE
2455 CurrentProc = MainProc; // PE to run it on
2458 cap = schedule(cap,task);
2460 ASSERT(task->stat != NoStatus);
2461 ASSERT_CAPABILITY_INVARIANTS(cap,task);
2463 IF_DEBUG(scheduler, sched_belch("bound thread (%d) finished", task->tso->id));
2467 /* ----------------------------------------------------------------------------
2469 * ------------------------------------------------------------------------- */
2471 #if defined(THREADED_RTS)
2473 workerStart(Task *task)
2477 // See startWorkerTask().
2478 ACQUIRE_LOCK(&task->lock);
2480 RELEASE_LOCK(&task->lock);
2482 // set the thread-local pointer to the Task:
2485 // schedule() runs without a lock.
2486 cap = schedule(cap,task);
2488 // On exit from schedule(), we have a Capability.
2489 releaseCapability(cap);
2494 /* ---------------------------------------------------------------------------
2497 * Initialise the scheduler. This resets all the queues - if the
2498 * queues contained any threads, they'll be garbage collected at the
2501 * ------------------------------------------------------------------------ */
2508 for (i=0; i<=MAX_PROC; i++) {
2509 run_queue_hds[i] = END_TSO_QUEUE;
2510 run_queue_tls[i] = END_TSO_QUEUE;
2511 blocked_queue_hds[i] = END_TSO_QUEUE;
2512 blocked_queue_tls[i] = END_TSO_QUEUE;
2513 ccalling_threadss[i] = END_TSO_QUEUE;
2514 blackhole_queue[i] = END_TSO_QUEUE;
2515 sleeping_queue = END_TSO_QUEUE;
2517 #elif !defined(THREADED_RTS)
2518 blocked_queue_hd = END_TSO_QUEUE;
2519 blocked_queue_tl = END_TSO_QUEUE;
2520 sleeping_queue = END_TSO_QUEUE;
2523 blackhole_queue = END_TSO_QUEUE;
2524 all_threads = END_TSO_QUEUE;
2529 RtsFlags.ConcFlags.ctxtSwitchTicks =
2530 RtsFlags.ConcFlags.ctxtSwitchTime / TICK_MILLISECS;
2532 #if defined(THREADED_RTS)
2533 /* Initialise the mutex and condition variables used by
2535 initMutex(&sched_mutex);
2538 ACQUIRE_LOCK(&sched_mutex);
2540 /* A capability holds the state a native thread needs in
2541 * order to execute STG code. At least one capability is
2542 * floating around (only SMP builds have more than one).
2550 * Eagerly start one worker to run each Capability, except for
2551 * Capability 0. The idea is that we're probably going to start a
2552 * bound thread on Capability 0 pretty soon, so we don't want a
2553 * worker task hogging it.
2558 for (i = 1; i < n_capabilities; i++) {
2559 cap = &capabilities[i];
2560 ACQUIRE_LOCK(&cap->lock);
2561 startWorkerTask(cap, workerStart);
2562 RELEASE_LOCK(&cap->lock);
2567 #if /* defined(SMP) ||*/ defined(PARALLEL_HASKELL)
2571 RELEASE_LOCK(&sched_mutex);
2575 exitScheduler( void )
2577 interrupted = rtsTrue;
2578 shutting_down_scheduler = rtsTrue;
2580 #if defined(THREADED_RTS)
2585 ACQUIRE_LOCK(&sched_mutex);
2586 task = newBoundTask();
2587 RELEASE_LOCK(&sched_mutex);
2589 for (i = 0; i < n_capabilities; i++) {
2590 shutdownCapability(&capabilities[i], task);
2592 boundTaskExiting(task);
2598 /* ---------------------------------------------------------------------------
2599 Where are the roots that we know about?
2601 - all the threads on the runnable queue
2602 - all the threads on the blocked queue
2603 - all the threads on the sleeping queue
2604 - all the thread currently executing a _ccall_GC
2605 - all the "main threads"
2607 ------------------------------------------------------------------------ */
2609 /* This has to be protected either by the scheduler monitor, or by the
2610 garbage collection monitor (probably the latter).
2615 GetRoots( evac_fn evac )
2622 for (i=0; i<=RtsFlags.GranFlags.proc; i++) {
2623 if ((run_queue_hds[i] != END_TSO_QUEUE) && ((run_queue_hds[i] != NULL)))
2624 evac((StgClosure **)&run_queue_hds[i]);
2625 if ((run_queue_tls[i] != END_TSO_QUEUE) && ((run_queue_tls[i] != NULL)))
2626 evac((StgClosure **)&run_queue_tls[i]);
2628 if ((blocked_queue_hds[i] != END_TSO_QUEUE) && ((blocked_queue_hds[i] != NULL)))
2629 evac((StgClosure **)&blocked_queue_hds[i]);
2630 if ((blocked_queue_tls[i] != END_TSO_QUEUE) && ((blocked_queue_tls[i] != NULL)))
2631 evac((StgClosure **)&blocked_queue_tls[i]);
2632 if ((ccalling_threadss[i] != END_TSO_QUEUE) && ((ccalling_threadss[i] != NULL)))
2633 evac((StgClosure **)&ccalling_threads[i]);
2640 for (i = 0; i < n_capabilities; i++) {
2641 cap = &capabilities[i];
2642 evac((StgClosure **)&cap->run_queue_hd);
2643 evac((StgClosure **)&cap->run_queue_tl);
2645 for (task = cap->suspended_ccalling_tasks; task != NULL;
2647 evac((StgClosure **)&task->suspended_tso);
2651 #if !defined(THREADED_RTS)
2652 evac((StgClosure **)&blocked_queue_hd);
2653 evac((StgClosure **)&blocked_queue_tl);
2654 evac((StgClosure **)&sleeping_queue);
2658 evac((StgClosure **)&blackhole_queue);
2660 #if defined(PARALLEL_HASKELL) || defined(GRAN)
2661 markSparkQueue(evac);
2664 #if defined(RTS_USER_SIGNALS)
2665 // mark the signal handlers (signals should be already blocked)
2666 markSignalHandlers(evac);
2670 /* -----------------------------------------------------------------------------
2673 This is the interface to the garbage collector from Haskell land.
2674 We provide this so that external C code can allocate and garbage
2675 collect when called from Haskell via _ccall_GC.
2677 It might be useful to provide an interface whereby the programmer
2678 can specify more roots (ToDo).
2680 This needs to be protected by the GC condition variable above. KH.
2681 -------------------------------------------------------------------------- */
2683 static void (*extra_roots)(evac_fn);
2689 // ToDo: we have to grab all the capabilities here.
2690 errorBelch("performGC not supported in threaded RTS (yet)");
2691 stg_exit(EXIT_FAILURE);
2693 /* Obligated to hold this lock upon entry */
2694 GarbageCollect(GetRoots,rtsFalse);
2698 performMajorGC(void)
2701 errorBelch("performMayjorGC not supported in threaded RTS (yet)");
2702 stg_exit(EXIT_FAILURE);
2704 GarbageCollect(GetRoots,rtsTrue);
2708 AllRoots(evac_fn evac)
2710 GetRoots(evac); // the scheduler's roots
2711 extra_roots(evac); // the user's roots
2715 performGCWithRoots(void (*get_roots)(evac_fn))
2718 errorBelch("performGCWithRoots not supported in threaded RTS (yet)");
2719 stg_exit(EXIT_FAILURE);
2721 extra_roots = get_roots;
2722 GarbageCollect(AllRoots,rtsFalse);
2725 /* -----------------------------------------------------------------------------
2728 If the thread has reached its maximum stack size, then raise the
2729 StackOverflow exception in the offending thread. Otherwise
2730 relocate the TSO into a larger chunk of memory and adjust its stack
2732 -------------------------------------------------------------------------- */
2735 threadStackOverflow(Capability *cap, StgTSO *tso)
2737 nat new_stack_size, stack_words;
2742 IF_DEBUG(sanity,checkTSO(tso));
2743 if (tso->stack_size >= tso->max_stack_size) {
2746 debugBelch("@@ threadStackOverflow of TSO %ld (%p): stack too large (now %ld; max is %ld)\n",
2747 (long)tso->id, tso, (long)tso->stack_size, (long)tso->max_stack_size);
2748 /* If we're debugging, just print out the top of the stack */
2749 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2752 /* Send this thread the StackOverflow exception */
2753 raiseAsync(cap, tso, (StgClosure *)stackOverflow_closure);
2757 /* Try to double the current stack size. If that takes us over the
2758 * maximum stack size for this thread, then use the maximum instead.
2759 * Finally round up so the TSO ends up as a whole number of blocks.
2761 new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
2762 new_tso_size = (lnat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) +
2763 TSO_STRUCT_SIZE)/sizeof(W_);
2764 new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */
2765 new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
2767 IF_DEBUG(scheduler, sched_belch("increasing stack size from %ld words to %d.\n", tso->stack_size, new_stack_size));
2769 dest = (StgTSO *)allocate(new_tso_size);
2770 TICK_ALLOC_TSO(new_stack_size,0);
2772 /* copy the TSO block and the old stack into the new area */
2773 memcpy(dest,tso,TSO_STRUCT_SIZE);
2774 stack_words = tso->stack + tso->stack_size - tso->sp;
2775 new_sp = (P_)dest + new_tso_size - stack_words;
2776 memcpy(new_sp, tso->sp, stack_words * sizeof(W_));
2778 /* relocate the stack pointers... */
2780 dest->stack_size = new_stack_size;
2782 /* Mark the old TSO as relocated. We have to check for relocated
2783 * TSOs in the garbage collector and any primops that deal with TSOs.
2785 * It's important to set the sp value to just beyond the end
2786 * of the stack, so we don't attempt to scavenge any part of the
2789 tso->what_next = ThreadRelocated;
2791 tso->sp = (P_)&(tso->stack[tso->stack_size]);
2792 tso->why_blocked = NotBlocked;
2794 IF_PAR_DEBUG(verbose,
2795 debugBelch("@@ threadStackOverflow of TSO %d (now at %p): stack size increased to %ld\n",
2796 tso->id, tso, tso->stack_size);
2797 /* If we're debugging, just print out the top of the stack */
2798 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2801 IF_DEBUG(sanity,checkTSO(tso));
2803 IF_DEBUG(scheduler,printTSO(dest));
2809 /* ---------------------------------------------------------------------------
2810 Wake up a queue that was blocked on some resource.
2811 ------------------------------------------------------------------------ */
2815 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2818 #elif defined(PARALLEL_HASKELL)
2820 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2822 /* write RESUME events to log file and
2823 update blocked and fetch time (depending on type of the orig closure) */
2824 if (RtsFlags.ParFlags.ParStats.Full) {
2825 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
2826 GR_RESUMEQ, ((StgTSO *)bqe), ((StgTSO *)bqe)->block_info.closure,
2827 0, 0 /* spark_queue_len(ADVISORY_POOL) */);
2828 if (emptyRunQueue())
2829 emitSchedule = rtsTrue;
2831 switch (get_itbl(node)->type) {
2833 ((StgTSO *)bqe)->par.fetchtime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2838 ((StgTSO *)bqe)->par.blocktime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2845 barf("{unblockOne}Daq Qagh: unexpected closure in blocking queue");
2852 StgBlockingQueueElement *
2853 unblockOne(StgBlockingQueueElement *bqe, StgClosure *node)
2856 PEs node_loc, tso_loc;
2858 node_loc = where_is(node); // should be lifted out of loop
2859 tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2860 tso_loc = where_is((StgClosure *)tso);
2861 if (IS_LOCAL_TO(PROCS(node),tso_loc)) { // TSO is local
2862 /* !fake_fetch => TSO is on CurrentProc is same as IS_LOCAL_TO */
2863 ASSERT(CurrentProc!=node_loc || tso_loc==CurrentProc);
2864 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.lunblocktime;
2865 // insertThread(tso, node_loc);
2866 new_event(tso_loc, tso_loc, CurrentTime[CurrentProc],
2868 tso, node, (rtsSpark*)NULL);
2869 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2872 } else { // TSO is remote (actually should be FMBQ)
2873 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.mpacktime +
2874 RtsFlags.GranFlags.Costs.gunblocktime +
2875 RtsFlags.GranFlags.Costs.latency;
2876 new_event(tso_loc, CurrentProc, CurrentTime[CurrentProc],
2878 tso, node, (rtsSpark*)NULL);
2879 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2882 /* the thread-queue-overhead is accounted for in either Resume or UnblockThread */
2884 debugBelch(" %s TSO %d (%p) [PE %d] (block_info.closure=%p) (next=%p) ,",
2885 (node_loc==tso_loc ? "Local" : "Global"),
2886 tso->id, tso, CurrentProc, tso->block_info.closure, tso->link));
2887 tso->block_info.closure = NULL;
2888 IF_DEBUG(scheduler,debugBelch("-- Waking up thread %ld (%p)\n",
2891 #elif defined(PARALLEL_HASKELL)
2892 StgBlockingQueueElement *
2893 unblockOne(StgBlockingQueueElement *bqe, StgClosure *node)
2895 StgBlockingQueueElement *next;
2897 switch (get_itbl(bqe)->type) {
2899 ASSERT(((StgTSO *)bqe)->why_blocked != NotBlocked);
2900 /* if it's a TSO just push it onto the run_queue */
2902 ((StgTSO *)bqe)->link = END_TSO_QUEUE; // debugging?
2903 APPEND_TO_RUN_QUEUE((StgTSO *)bqe);
2905 unblockCount(bqe, node);
2906 /* reset blocking status after dumping event */
2907 ((StgTSO *)bqe)->why_blocked = NotBlocked;
2911 /* if it's a BLOCKED_FETCH put it on the PendingFetches list */
2913 bqe->link = (StgBlockingQueueElement *)PendingFetches;
2914 PendingFetches = (StgBlockedFetch *)bqe;
2918 /* can ignore this case in a non-debugging setup;
2919 see comments on RBHSave closures above */
2921 /* check that the closure is an RBHSave closure */
2922 ASSERT(get_itbl((StgClosure *)bqe) == &stg_RBH_Save_0_info ||
2923 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_1_info ||
2924 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_2_info);
2928 barf("{unblockOne}Daq Qagh: Unexpected IP (%#lx; %s) in blocking queue at %#lx\n",
2929 get_itbl((StgClosure *)bqe), info_type((StgClosure *)bqe),
2933 IF_PAR_DEBUG(bq, debugBelch(", %p (%s)\n", bqe, info_type((StgClosure*)bqe)));
2939 unblockOne(Capability *cap, StgTSO *tso)
2943 ASSERT(get_itbl(tso)->type == TSO);
2944 ASSERT(tso->why_blocked != NotBlocked);
2945 tso->why_blocked = NotBlocked;
2947 tso->link = END_TSO_QUEUE;
2949 // We might have just migrated this TSO to our Capability:
2951 tso->bound->cap = cap;
2954 appendToRunQueue(cap,tso);
2956 // we're holding a newly woken thread, make sure we context switch
2957 // quickly so we can migrate it if necessary.
2959 IF_DEBUG(scheduler,sched_belch("waking up thread %ld", (long)tso->id));
2966 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2968 StgBlockingQueueElement *bqe;
2973 debugBelch("##-_ AwBQ for node %p on PE %d @ %ld by TSO %d (%p): \n", \
2974 node, CurrentProc, CurrentTime[CurrentProc],
2975 CurrentTSO->id, CurrentTSO));
2977 node_loc = where_is(node);
2979 ASSERT(q == END_BQ_QUEUE ||
2980 get_itbl(q)->type == TSO || // q is either a TSO or an RBHSave
2981 get_itbl(q)->type == CONSTR); // closure (type constructor)
2982 ASSERT(is_unique(node));
2984 /* FAKE FETCH: magically copy the node to the tso's proc;
2985 no Fetch necessary because in reality the node should not have been
2986 moved to the other PE in the first place
2988 if (CurrentProc!=node_loc) {
2990 debugBelch("## node %p is on PE %d but CurrentProc is %d (TSO %d); assuming fake fetch and adjusting bitmask (old: %#x)\n",
2991 node, node_loc, CurrentProc, CurrentTSO->id,
2992 // CurrentTSO, where_is(CurrentTSO),
2993 node->header.gran.procs));
2994 node->header.gran.procs = (node->header.gran.procs) | PE_NUMBER(CurrentProc);
2996 debugBelch("## new bitmask of node %p is %#x\n",
2997 node, node->header.gran.procs));
2998 if (RtsFlags.GranFlags.GranSimStats.Global) {
2999 globalGranStats.tot_fake_fetches++;
3004 // ToDo: check: ASSERT(CurrentProc==node_loc);
3005 while (get_itbl(bqe)->type==TSO) { // q != END_TSO_QUEUE) {
3008 bqe points to the current element in the queue
3009 next points to the next element in the queue
3011 //tso = (StgTSO *)bqe; // wastes an assignment to get the type right
3012 //tso_loc = where_is(tso);
3014 bqe = unblockOne(bqe, node);
3017 /* if this is the BQ of an RBH, we have to put back the info ripped out of
3018 the closure to make room for the anchor of the BQ */
3019 if (bqe!=END_BQ_QUEUE) {
3020 ASSERT(get_itbl(node)->type == RBH && get_itbl(bqe)->type == CONSTR);
3022 ASSERT((info_ptr==&RBH_Save_0_info) ||
3023 (info_ptr==&RBH_Save_1_info) ||
3024 (info_ptr==&RBH_Save_2_info));
3026 /* cf. convertToRBH in RBH.c for writing the RBHSave closure */
3027 ((StgRBH *)node)->blocking_queue = (StgBlockingQueueElement *)((StgRBHSave *)bqe)->payload[0];
3028 ((StgRBH *)node)->mut_link = (StgMutClosure *)((StgRBHSave *)bqe)->payload[1];
3031 debugBelch("## Filled in RBH_Save for %p (%s) at end of AwBQ\n",
3032 node, info_type(node)));
3035 /* statistics gathering */
3036 if (RtsFlags.GranFlags.GranSimStats.Global) {
3037 // globalGranStats.tot_bq_processing_time += bq_processing_time;
3038 globalGranStats.tot_bq_len += len; // total length of all bqs awakened
3039 // globalGranStats.tot_bq_len_local += len_local; // same for local TSOs only
3040 globalGranStats.tot_awbq++; // total no. of bqs awakened
3043 debugBelch("## BQ Stats of %p: [%d entries] %s\n",
3044 node, len, (bqe!=END_BQ_QUEUE) ? "RBH" : ""));
3046 #elif defined(PARALLEL_HASKELL)
3048 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
3050 StgBlockingQueueElement *bqe;
3052 IF_PAR_DEBUG(verbose,
3053 debugBelch("##-_ AwBQ for node %p on [%x]: \n",
3057 if(get_itbl(q)->type == CONSTR || q==END_BQ_QUEUE) {
3058 IF_PAR_DEBUG(verbose, debugBelch("## ... nothing to unblock so lets just return. RFP (BUG?)\n"));
3063 ASSERT(q == END_BQ_QUEUE ||
3064 get_itbl(q)->type == TSO ||
3065 get_itbl(q)->type == BLOCKED_FETCH ||
3066 get_itbl(q)->type == CONSTR);
3069 while (get_itbl(bqe)->type==TSO ||
3070 get_itbl(bqe)->type==BLOCKED_FETCH) {
3071 bqe = unblockOne(bqe, node);
3075 #else /* !GRAN && !PARALLEL_HASKELL */
3078 awakenBlockedQueue(Capability *cap, StgTSO *tso)
3080 if (tso == NULL) return; // hack; see bug #1235728, and comments in
3082 while (tso != END_TSO_QUEUE) {
3083 tso = unblockOne(cap,tso);
3088 /* ---------------------------------------------------------------------------
3090 - usually called inside a signal handler so it mustn't do anything fancy.
3091 ------------------------------------------------------------------------ */
3094 interruptStgRts(void)
3098 #if defined(THREADED_RTS)
3099 prodAllCapabilities();
3103 /* -----------------------------------------------------------------------------
3106 This is for use when we raise an exception in another thread, which
3108 This has nothing to do with the UnblockThread event in GranSim. -- HWL
3109 -------------------------------------------------------------------------- */
3111 #if defined(GRAN) || defined(PARALLEL_HASKELL)
3113 NB: only the type of the blocking queue is different in GranSim and GUM
3114 the operations on the queue-elements are the same
3115 long live polymorphism!
3117 Locks: sched_mutex is held upon entry and exit.
3121 unblockThread(Capability *cap, StgTSO *tso)
3123 StgBlockingQueueElement *t, **last;
3125 switch (tso->why_blocked) {
3128 return; /* not blocked */
3131 // Be careful: nothing to do here! We tell the scheduler that the thread
3132 // is runnable and we leave it to the stack-walking code to abort the
3133 // transaction while unwinding the stack. We should perhaps have a debugging
3134 // test to make sure that this really happens and that the 'zombie' transaction
3135 // does not get committed.
3139 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
3141 StgBlockingQueueElement *last_tso = END_BQ_QUEUE;
3142 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
3144 last = (StgBlockingQueueElement **)&mvar->head;
3145 for (t = (StgBlockingQueueElement *)mvar->head;
3147 last = &t->link, last_tso = t, t = t->link) {
3148 if (t == (StgBlockingQueueElement *)tso) {
3149 *last = (StgBlockingQueueElement *)tso->link;
3150 if (mvar->tail == tso) {
3151 mvar->tail = (StgTSO *)last_tso;
3156 barf("unblockThread (MVAR): TSO not found");
3159 case BlockedOnBlackHole:
3160 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
3162 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
3164 last = &bq->blocking_queue;
3165 for (t = bq->blocking_queue;
3167 last = &t->link, t = t->link) {
3168 if (t == (StgBlockingQueueElement *)tso) {
3169 *last = (StgBlockingQueueElement *)tso->link;
3173 barf("unblockThread (BLACKHOLE): TSO not found");
3176 case BlockedOnException:
3178 StgTSO *target = tso->block_info.tso;
3180 ASSERT(get_itbl(target)->type == TSO);
3182 if (target->what_next == ThreadRelocated) {
3183 target = target->link;
3184 ASSERT(get_itbl(target)->type == TSO);
3187 ASSERT(target->blocked_exceptions != NULL);
3189 last = (StgBlockingQueueElement **)&target->blocked_exceptions;
3190 for (t = (StgBlockingQueueElement *)target->blocked_exceptions;
3192 last = &t->link, t = t->link) {
3193 ASSERT(get_itbl(t)->type == TSO);
3194 if (t == (StgBlockingQueueElement *)tso) {
3195 *last = (StgBlockingQueueElement *)tso->link;
3199 barf("unblockThread (Exception): TSO not found");
3203 case BlockedOnWrite:
3204 #if defined(mingw32_HOST_OS)
3205 case BlockedOnDoProc:
3208 /* take TSO off blocked_queue */
3209 StgBlockingQueueElement *prev = NULL;
3210 for (t = (StgBlockingQueueElement *)blocked_queue_hd; t != END_BQ_QUEUE;
3211 prev = t, t = t->link) {
3212 if (t == (StgBlockingQueueElement *)tso) {
3214 blocked_queue_hd = (StgTSO *)t->link;
3215 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
3216 blocked_queue_tl = END_TSO_QUEUE;
3219 prev->link = t->link;
3220 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
3221 blocked_queue_tl = (StgTSO *)prev;
3224 #if defined(mingw32_HOST_OS)
3225 /* (Cooperatively) signal that the worker thread should abort
3228 abandonWorkRequest(tso->block_info.async_result->reqID);
3233 barf("unblockThread (I/O): TSO not found");
3236 case BlockedOnDelay:
3238 /* take TSO off sleeping_queue */
3239 StgBlockingQueueElement *prev = NULL;
3240 for (t = (StgBlockingQueueElement *)sleeping_queue; t != END_BQ_QUEUE;
3241 prev = t, t = t->link) {
3242 if (t == (StgBlockingQueueElement *)tso) {
3244 sleeping_queue = (StgTSO *)t->link;
3246 prev->link = t->link;
3251 barf("unblockThread (delay): TSO not found");
3255 barf("unblockThread");
3259 tso->link = END_TSO_QUEUE;
3260 tso->why_blocked = NotBlocked;
3261 tso->block_info.closure = NULL;
3262 pushOnRunQueue(cap,tso);
3266 unblockThread(Capability *cap, StgTSO *tso)
3270 /* To avoid locking unnecessarily. */
3271 if (tso->why_blocked == NotBlocked) {
3275 switch (tso->why_blocked) {
3278 // Be careful: nothing to do here! We tell the scheduler that the thread
3279 // is runnable and we leave it to the stack-walking code to abort the
3280 // transaction while unwinding the stack. We should perhaps have a debugging
3281 // test to make sure that this really happens and that the 'zombie' transaction
3282 // does not get committed.
3286 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
3288 StgTSO *last_tso = END_TSO_QUEUE;
3289 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
3292 for (t = mvar->head; t != END_TSO_QUEUE;
3293 last = &t->link, last_tso = t, t = t->link) {
3296 if (mvar->tail == tso) {
3297 mvar->tail = last_tso;
3302 barf("unblockThread (MVAR): TSO not found");
3305 case BlockedOnBlackHole:
3307 last = &blackhole_queue;
3308 for (t = blackhole_queue; t != END_TSO_QUEUE;
3309 last = &t->link, t = t->link) {
3315 barf("unblockThread (BLACKHOLE): TSO not found");
3318 case BlockedOnException:
3320 StgTSO *target = tso->block_info.tso;
3322 ASSERT(get_itbl(target)->type == TSO);
3324 while (target->what_next == ThreadRelocated) {
3325 target = target->link;
3326 ASSERT(get_itbl(target)->type == TSO);
3329 ASSERT(target->blocked_exceptions != NULL);
3331 last = &target->blocked_exceptions;
3332 for (t = target->blocked_exceptions; t != END_TSO_QUEUE;
3333 last = &t->link, t = t->link) {
3334 ASSERT(get_itbl(t)->type == TSO);
3340 barf("unblockThread (Exception): TSO not found");
3343 #if !defined(THREADED_RTS)
3345 case BlockedOnWrite:
3346 #if defined(mingw32_HOST_OS)
3347 case BlockedOnDoProc:
3350 StgTSO *prev = NULL;
3351 for (t = blocked_queue_hd; t != END_TSO_QUEUE;
3352 prev = t, t = t->link) {
3355 blocked_queue_hd = t->link;
3356 if (blocked_queue_tl == t) {
3357 blocked_queue_tl = END_TSO_QUEUE;
3360 prev->link = t->link;
3361 if (blocked_queue_tl == t) {
3362 blocked_queue_tl = prev;
3365 #if defined(mingw32_HOST_OS)
3366 /* (Cooperatively) signal that the worker thread should abort
3369 abandonWorkRequest(tso->block_info.async_result->reqID);
3374 barf("unblockThread (I/O): TSO not found");
3377 case BlockedOnDelay:
3379 StgTSO *prev = NULL;
3380 for (t = sleeping_queue; t != END_TSO_QUEUE;
3381 prev = t, t = t->link) {
3384 sleeping_queue = t->link;
3386 prev->link = t->link;
3391 barf("unblockThread (delay): TSO not found");
3396 barf("unblockThread");
3400 tso->link = END_TSO_QUEUE;
3401 tso->why_blocked = NotBlocked;
3402 tso->block_info.closure = NULL;
3403 appendToRunQueue(cap,tso);
3407 /* -----------------------------------------------------------------------------
3410 * Check the blackhole_queue for threads that can be woken up. We do
3411 * this periodically: before every GC, and whenever the run queue is
3414 * An elegant solution might be to just wake up all the blocked
3415 * threads with awakenBlockedQueue occasionally: they'll go back to
3416 * sleep again if the object is still a BLACKHOLE. Unfortunately this
3417 * doesn't give us a way to tell whether we've actually managed to
3418 * wake up any threads, so we would be busy-waiting.
3420 * -------------------------------------------------------------------------- */
3423 checkBlackHoles (Capability *cap)
3426 rtsBool any_woke_up = rtsFalse;
3429 // blackhole_queue is global:
3430 ASSERT_LOCK_HELD(&sched_mutex);
3432 IF_DEBUG(scheduler, sched_belch("checking threads blocked on black holes"));
3434 // ASSUMES: sched_mutex
3435 prev = &blackhole_queue;
3436 t = blackhole_queue;
3437 while (t != END_TSO_QUEUE) {
3438 ASSERT(t->why_blocked == BlockedOnBlackHole);
3439 type = get_itbl(t->block_info.closure)->type;
3440 if (type != BLACKHOLE && type != CAF_BLACKHOLE) {
3441 IF_DEBUG(sanity,checkTSO(t));
3442 t = unblockOne(cap, t);
3443 // urk, the threads migrate to the current capability
3444 // here, but we'd like to keep them on the original one.
3446 any_woke_up = rtsTrue;
3456 /* -----------------------------------------------------------------------------
3459 * The following function implements the magic for raising an
3460 * asynchronous exception in an existing thread.
3462 * We first remove the thread from any queue on which it might be
3463 * blocked. The possible blockages are MVARs and BLACKHOLE_BQs.
3465 * We strip the stack down to the innermost CATCH_FRAME, building
3466 * thunks in the heap for all the active computations, so they can
3467 * be restarted if necessary. When we reach a CATCH_FRAME, we build
3468 * an application of the handler to the exception, and push it on
3469 * the top of the stack.
3471 * How exactly do we save all the active computations? We create an
3472 * AP_STACK for every UpdateFrame on the stack. Entering one of these
3473 * AP_STACKs pushes everything from the corresponding update frame
3474 * upwards onto the stack. (Actually, it pushes everything up to the
3475 * next update frame plus a pointer to the next AP_STACK object.
3476 * Entering the next AP_STACK object pushes more onto the stack until we
3477 * reach the last AP_STACK object - at which point the stack should look
3478 * exactly as it did when we killed the TSO and we can continue
3479 * execution by entering the closure on top of the stack.
3481 * We can also kill a thread entirely - this happens if either (a) the
3482 * exception passed to raiseAsync is NULL, or (b) there's no
3483 * CATCH_FRAME on the stack. In either case, we strip the entire
3484 * stack and replace the thread with a zombie.
3486 * ToDo: in SMP mode, this function is only safe if either (a) we hold
3487 * all the Capabilities (eg. in GC), or (b) we own the Capability that
3488 * the TSO is currently blocked on or on the run queue of.
3490 * -------------------------------------------------------------------------- */
3493 raiseAsync(Capability *cap, StgTSO *tso, StgClosure *exception)
3495 raiseAsync_(cap, tso, exception, rtsFalse);
3499 raiseAsync_(Capability *cap, StgTSO *tso, StgClosure *exception,
3500 rtsBool stop_at_atomically)
3502 StgRetInfoTable *info;
3505 // Thread already dead?
3506 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
3511 sched_belch("raising exception in thread %ld.", (long)tso->id));
3513 // Remove it from any blocking queues
3514 unblockThread(cap,tso);
3518 // The stack freezing code assumes there's a closure pointer on
3519 // the top of the stack, so we have to arrange that this is the case...
3521 if (sp[0] == (W_)&stg_enter_info) {
3525 sp[0] = (W_)&stg_dummy_ret_closure;
3531 // 1. Let the top of the stack be the "current closure"
3533 // 2. Walk up the stack until we find either an UPDATE_FRAME or a
3536 // 3. If it's an UPDATE_FRAME, then make an AP_STACK containing the
3537 // current closure applied to the chunk of stack up to (but not
3538 // including) the update frame. This closure becomes the "current
3539 // closure". Go back to step 2.
3541 // 4. If it's a CATCH_FRAME, then leave the exception handler on
3542 // top of the stack applied to the exception.
3544 // 5. If it's a STOP_FRAME, then kill the thread.
3546 // NB: if we pass an ATOMICALLY_FRAME then abort the associated
3553 info = get_ret_itbl((StgClosure *)frame);
3555 while (info->i.type != UPDATE_FRAME
3556 && (info->i.type != CATCH_FRAME || exception == NULL)
3557 && info->i.type != STOP_FRAME
3558 && (info->i.type != ATOMICALLY_FRAME || stop_at_atomically == rtsFalse))
3560 if (info->i.type == CATCH_RETRY_FRAME || info->i.type == ATOMICALLY_FRAME) {
3561 // IF we find an ATOMICALLY_FRAME then we abort the
3562 // current transaction and propagate the exception. In
3563 // this case (unlike ordinary exceptions) we do not care
3564 // whether the transaction is valid or not because its
3565 // possible validity cannot have caused the exception
3566 // and will not be visible after the abort.
3568 debugBelch("Found atomically block delivering async exception\n"));
3569 stmAbortTransaction(tso -> trec);
3570 tso -> trec = stmGetEnclosingTRec(tso -> trec);
3572 frame += stack_frame_sizeW((StgClosure *)frame);
3573 info = get_ret_itbl((StgClosure *)frame);
3576 switch (info->i.type) {
3578 case ATOMICALLY_FRAME:
3579 ASSERT(stop_at_atomically);
3580 ASSERT(stmGetEnclosingTRec(tso->trec) == NO_TREC);
3581 stmCondemnTransaction(tso -> trec);
3585 // R1 is not a register: the return convention for IO in
3586 // this case puts the return value on the stack, so we
3587 // need to set up the stack to return to the atomically
3588 // frame properly...
3589 tso->sp = frame - 2;
3590 tso->sp[1] = (StgWord) &stg_NO_FINALIZER_closure; // why not?
3591 tso->sp[0] = (StgWord) &stg_ut_1_0_unreg_info;
3593 tso->what_next = ThreadRunGHC;
3597 // If we find a CATCH_FRAME, and we've got an exception to raise,
3598 // then build the THUNK raise(exception), and leave it on
3599 // top of the CATCH_FRAME ready to enter.
3603 StgCatchFrame *cf = (StgCatchFrame *)frame;
3607 // we've got an exception to raise, so let's pass it to the
3608 // handler in this frame.
3610 raise = (StgThunk *)allocateLocal(cap,sizeofW(StgThunk)+MIN_UPD_SIZE);
3611 TICK_ALLOC_SE_THK(1,0);
3612 SET_HDR(raise,&stg_raise_info,cf->header.prof.ccs);
3613 raise->payload[0] = exception;
3615 // throw away the stack from Sp up to the CATCH_FRAME.
3619 /* Ensure that async excpetions are blocked now, so we don't get
3620 * a surprise exception before we get around to executing the
3623 if (tso->blocked_exceptions == NULL) {
3624 tso->blocked_exceptions = END_TSO_QUEUE;
3627 /* Put the newly-built THUNK on top of the stack, ready to execute
3628 * when the thread restarts.
3631 sp[-1] = (W_)&stg_enter_info;
3633 tso->what_next = ThreadRunGHC;
3634 IF_DEBUG(sanity, checkTSO(tso));
3643 // First build an AP_STACK consisting of the stack chunk above the
3644 // current update frame, with the top word on the stack as the
3647 words = frame - sp - 1;
3648 ap = (StgAP_STACK *)allocateLocal(cap,AP_STACK_sizeW(words));
3651 ap->fun = (StgClosure *)sp[0];
3653 for(i=0; i < (nat)words; ++i) {
3654 ap->payload[i] = (StgClosure *)*sp++;
3657 SET_HDR(ap,&stg_AP_STACK_info,
3658 ((StgClosure *)frame)->header.prof.ccs /* ToDo */);
3659 TICK_ALLOC_UP_THK(words+1,0);
3662 debugBelch("sched: Updating ");
3663 printPtr((P_)((StgUpdateFrame *)frame)->updatee);
3664 debugBelch(" with ");
3665 printObj((StgClosure *)ap);
3668 // Replace the updatee with an indirection - happily
3669 // this will also wake up any threads currently
3670 // waiting on the result.
3672 // Warning: if we're in a loop, more than one update frame on
3673 // the stack may point to the same object. Be careful not to
3674 // overwrite an IND_OLDGEN in this case, because we'll screw
3675 // up the mutable lists. To be on the safe side, don't
3676 // overwrite any kind of indirection at all. See also
3677 // threadSqueezeStack in GC.c, where we have to make a similar
3680 if (!closure_IND(((StgUpdateFrame *)frame)->updatee)) {
3681 // revert the black hole
3682 UPD_IND_NOLOCK(((StgUpdateFrame *)frame)->updatee,
3685 sp += sizeofW(StgUpdateFrame) - 1;
3686 sp[0] = (W_)ap; // push onto stack
3691 // We've stripped the entire stack, the thread is now dead.
3692 sp += sizeofW(StgStopFrame);
3693 tso->what_next = ThreadKilled;
3704 /* -----------------------------------------------------------------------------
3707 This is used for interruption (^C) and forking, and corresponds to
3708 raising an exception but without letting the thread catch the
3710 -------------------------------------------------------------------------- */
3713 deleteThread (Capability *cap, StgTSO *tso)
3715 if (tso->why_blocked != BlockedOnCCall &&
3716 tso->why_blocked != BlockedOnCCall_NoUnblockExc) {
3717 raiseAsync(cap,tso,NULL);
3721 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
3723 deleteThreadImmediately(Capability *cap, StgTSO *tso)
3724 { // for forkProcess only:
3725 // delete thread without giving it a chance to catch the KillThread exception
3727 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
3731 if (tso->why_blocked != BlockedOnCCall &&
3732 tso->why_blocked != BlockedOnCCall_NoUnblockExc) {
3733 unblockThread(cap,tso);
3736 tso->what_next = ThreadKilled;
3740 /* -----------------------------------------------------------------------------
3741 raiseExceptionHelper
3743 This function is called by the raise# primitve, just so that we can
3744 move some of the tricky bits of raising an exception from C-- into
3745 C. Who knows, it might be a useful re-useable thing here too.
3746 -------------------------------------------------------------------------- */
3749 raiseExceptionHelper (StgRegTable *reg, StgTSO *tso, StgClosure *exception)
3751 Capability *cap = regTableToCapability(reg);
3752 StgThunk *raise_closure = NULL;
3754 StgRetInfoTable *info;
3756 // This closure represents the expression 'raise# E' where E
3757 // is the exception raise. It is used to overwrite all the
3758 // thunks which are currently under evaluataion.
3762 // LDV profiling: stg_raise_info has THUNK as its closure
3763 // type. Since a THUNK takes at least MIN_UPD_SIZE words in its
3764 // payload, MIN_UPD_SIZE is more approprate than 1. It seems that
3765 // 1 does not cause any problem unless profiling is performed.
3766 // However, when LDV profiling goes on, we need to linearly scan
3767 // small object pool, where raise_closure is stored, so we should
3768 // use MIN_UPD_SIZE.
3770 // raise_closure = (StgClosure *)RET_STGCALL1(P_,allocate,
3771 // sizeofW(StgClosure)+1);
3775 // Walk up the stack, looking for the catch frame. On the way,
3776 // we update any closures pointed to from update frames with the
3777 // raise closure that we just built.
3781 info = get_ret_itbl((StgClosure *)p);
3782 next = p + stack_frame_sizeW((StgClosure *)p);
3783 switch (info->i.type) {
3786 // Only create raise_closure if we need to.
3787 if (raise_closure == NULL) {
3789 (StgThunk *)allocateLocal(cap,sizeofW(StgThunk)+MIN_UPD_SIZE);
3790 SET_HDR(raise_closure, &stg_raise_info, CCCS);
3791 raise_closure->payload[0] = exception;
3793 UPD_IND(((StgUpdateFrame *)p)->updatee,(StgClosure *)raise_closure);
3797 case ATOMICALLY_FRAME:
3798 IF_DEBUG(stm, debugBelch("Found ATOMICALLY_FRAME at %p\n", p));
3800 return ATOMICALLY_FRAME;
3806 case CATCH_STM_FRAME:
3807 IF_DEBUG(stm, debugBelch("Found CATCH_STM_FRAME at %p\n", p));
3809 return CATCH_STM_FRAME;
3815 case CATCH_RETRY_FRAME:
3824 /* -----------------------------------------------------------------------------
3825 findRetryFrameHelper
3827 This function is called by the retry# primitive. It traverses the stack
3828 leaving tso->sp referring to the frame which should handle the retry.
3830 This should either be a CATCH_RETRY_FRAME (if the retry# is within an orElse#)
3831 or should be a ATOMICALLY_FRAME (if the retry# reaches the top level).
3833 We skip CATCH_STM_FRAMEs because retries are not considered to be exceptions,
3834 despite the similar implementation.
3836 We should not expect to see CATCH_FRAME or STOP_FRAME because those should
3837 not be created within memory transactions.
3838 -------------------------------------------------------------------------- */
3841 findRetryFrameHelper (StgTSO *tso)
3844 StgRetInfoTable *info;
3848 info = get_ret_itbl((StgClosure *)p);
3849 next = p + stack_frame_sizeW((StgClosure *)p);
3850 switch (info->i.type) {
3852 case ATOMICALLY_FRAME:
3853 IF_DEBUG(stm, debugBelch("Found ATOMICALLY_FRAME at %p during retrry\n", p));
3855 return ATOMICALLY_FRAME;
3857 case CATCH_RETRY_FRAME:
3858 IF_DEBUG(stm, debugBelch("Found CATCH_RETRY_FRAME at %p during retrry\n", p));
3860 return CATCH_RETRY_FRAME;
3862 case CATCH_STM_FRAME:
3864 ASSERT(info->i.type != CATCH_FRAME);
3865 ASSERT(info->i.type != STOP_FRAME);
3872 /* -----------------------------------------------------------------------------
3873 resurrectThreads is called after garbage collection on the list of
3874 threads found to be garbage. Each of these threads will be woken
3875 up and sent a signal: BlockedOnDeadMVar if the thread was blocked
3876 on an MVar, or NonTermination if the thread was blocked on a Black
3879 Locks: assumes we hold *all* the capabilities.
3880 -------------------------------------------------------------------------- */
3883 resurrectThreads (StgTSO *threads)
3888 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
3889 next = tso->global_link;
3890 tso->global_link = all_threads;
3892 IF_DEBUG(scheduler, sched_belch("resurrecting thread %d", tso->id));
3894 // Wake up the thread on the Capability it was last on for a
3895 // bound thread, or last_free_capability otherwise.
3897 cap = tso->bound->cap;
3899 cap = last_free_capability;
3902 switch (tso->why_blocked) {
3904 case BlockedOnException:
3905 /* Called by GC - sched_mutex lock is currently held. */
3906 raiseAsync(cap, tso,(StgClosure *)BlockedOnDeadMVar_closure);
3908 case BlockedOnBlackHole:
3909 raiseAsync(cap, tso,(StgClosure *)NonTermination_closure);
3912 raiseAsync(cap, tso,(StgClosure *)BlockedIndefinitely_closure);
3915 /* This might happen if the thread was blocked on a black hole
3916 * belonging to a thread that we've just woken up (raiseAsync
3917 * can wake up threads, remember...).
3921 barf("resurrectThreads: thread blocked in a strange way");
3926 /* ----------------------------------------------------------------------------
3927 * Debugging: why is a thread blocked
3928 * [Also provides useful information when debugging threaded programs
3929 * at the Haskell source code level, so enable outside of DEBUG. --sof 7/02]
3930 ------------------------------------------------------------------------- */
3934 printThreadBlockage(StgTSO *tso)
3936 switch (tso->why_blocked) {
3938 debugBelch("is blocked on read from fd %d", (int)(tso->block_info.fd));
3940 case BlockedOnWrite:
3941 debugBelch("is blocked on write to fd %d", (int)(tso->block_info.fd));
3943 #if defined(mingw32_HOST_OS)
3944 case BlockedOnDoProc:
3945 debugBelch("is blocked on proc (request: %ld)", tso->block_info.async_result->reqID);
3948 case BlockedOnDelay:
3949 debugBelch("is blocked until %ld", (long)(tso->block_info.target));
3952 debugBelch("is blocked on an MVar @ %p", tso->block_info.closure);
3954 case BlockedOnException:
3955 debugBelch("is blocked on delivering an exception to thread %d",
3956 tso->block_info.tso->id);
3958 case BlockedOnBlackHole:
3959 debugBelch("is blocked on a black hole");
3962 debugBelch("is not blocked");
3964 #if defined(PARALLEL_HASKELL)
3966 debugBelch("is blocked on global address; local FM_BQ is %p (%s)",
3967 tso->block_info.closure, info_type(tso->block_info.closure));
3969 case BlockedOnGA_NoSend:
3970 debugBelch("is blocked on global address (no send); local FM_BQ is %p (%s)",
3971 tso->block_info.closure, info_type(tso->block_info.closure));
3974 case BlockedOnCCall:
3975 debugBelch("is blocked on an external call");
3977 case BlockedOnCCall_NoUnblockExc:
3978 debugBelch("is blocked on an external call (exceptions were already blocked)");
3981 debugBelch("is blocked on an STM operation");
3984 barf("printThreadBlockage: strange tso->why_blocked: %d for TSO %d (%d)",
3985 tso->why_blocked, tso->id, tso);
3990 printThreadStatus(StgTSO *tso)
3992 switch (tso->what_next) {
3994 debugBelch("has been killed");
3996 case ThreadComplete:
3997 debugBelch("has completed");
4000 printThreadBlockage(tso);
4005 printAllThreads(void)
4010 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
4011 ullong_format_string(TIME_ON_PROC(CurrentProc),
4012 time_string, rtsFalse/*no commas!*/);
4014 debugBelch("all threads at [%s]:\n", time_string);
4015 # elif defined(PARALLEL_HASKELL)
4016 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
4017 ullong_format_string(CURRENT_TIME,
4018 time_string, rtsFalse/*no commas!*/);
4020 debugBelch("all threads at [%s]:\n", time_string);
4022 debugBelch("all threads:\n");
4025 for (t = all_threads; t != END_TSO_QUEUE; ) {
4026 debugBelch("\tthread %4d @ %p ", t->id, (void *)t);
4028 void *label = lookupThreadLabel(t->id);
4029 if (label) debugBelch("[\"%s\"] ",(char *)label);
4031 if (t->what_next == ThreadRelocated) {
4032 debugBelch("has been relocated...\n");
4035 printThreadStatus(t);
4044 printThreadQueue(StgTSO *t)
4047 for (; t != END_TSO_QUEUE; t = t->link) {
4048 debugBelch("\tthread %d @ %p ", t->id, (void *)t);
4049 if (t->what_next == ThreadRelocated) {
4050 debugBelch("has been relocated...\n");
4052 printThreadStatus(t);
4057 debugBelch("%d threads on queue\n", i);
4061 Print a whole blocking queue attached to node (debugging only).
4063 # if defined(PARALLEL_HASKELL)
4065 print_bq (StgClosure *node)
4067 StgBlockingQueueElement *bqe;
4071 debugBelch("## BQ of closure %p (%s): ",
4072 node, info_type(node));
4074 /* should cover all closures that may have a blocking queue */
4075 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
4076 get_itbl(node)->type == FETCH_ME_BQ ||
4077 get_itbl(node)->type == RBH ||
4078 get_itbl(node)->type == MVAR);
4080 ASSERT(node!=(StgClosure*)NULL); // sanity check
4082 print_bqe(((StgBlockingQueue*)node)->blocking_queue);
4086 Print a whole blocking queue starting with the element bqe.
4089 print_bqe (StgBlockingQueueElement *bqe)
4094 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
4096 for (end = (bqe==END_BQ_QUEUE);
4097 !end; // iterate until bqe points to a CONSTR
4098 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE),
4099 bqe = end ? END_BQ_QUEUE : bqe->link) {
4100 ASSERT(bqe != END_BQ_QUEUE); // sanity check
4101 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
4102 /* types of closures that may appear in a blocking queue */
4103 ASSERT(get_itbl(bqe)->type == TSO ||
4104 get_itbl(bqe)->type == BLOCKED_FETCH ||
4105 get_itbl(bqe)->type == CONSTR);
4106 /* only BQs of an RBH end with an RBH_Save closure */
4107 //ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
4109 switch (get_itbl(bqe)->type) {
4111 debugBelch(" TSO %u (%x),",
4112 ((StgTSO *)bqe)->id, ((StgTSO *)bqe));
4115 debugBelch(" BF (node=%p, ga=((%x, %d, %x)),",
4116 ((StgBlockedFetch *)bqe)->node,
4117 ((StgBlockedFetch *)bqe)->ga.payload.gc.gtid,
4118 ((StgBlockedFetch *)bqe)->ga.payload.gc.slot,
4119 ((StgBlockedFetch *)bqe)->ga.weight);
4122 debugBelch(" %s (IP %p),",
4123 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
4124 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
4125 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
4126 "RBH_Save_?"), get_itbl(bqe));
4129 barf("Unexpected closure type %s in blocking queue", // of %p (%s)",
4130 info_type((StgClosure *)bqe)); // , node, info_type(node));
4136 # elif defined(GRAN)
4138 print_bq (StgClosure *node)
4140 StgBlockingQueueElement *bqe;
4141 PEs node_loc, tso_loc;
4144 /* should cover all closures that may have a blocking queue */
4145 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
4146 get_itbl(node)->type == FETCH_ME_BQ ||
4147 get_itbl(node)->type == RBH);
4149 ASSERT(node!=(StgClosure*)NULL); // sanity check
4150 node_loc = where_is(node);
4152 debugBelch("## BQ of closure %p (%s) on [PE %d]: ",
4153 node, info_type(node), node_loc);
4156 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
4158 for (bqe = ((StgBlockingQueue*)node)->blocking_queue, end = (bqe==END_BQ_QUEUE);
4159 !end; // iterate until bqe points to a CONSTR
4160 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) {
4161 ASSERT(bqe != END_BQ_QUEUE); // sanity check
4162 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
4163 /* types of closures that may appear in a blocking queue */
4164 ASSERT(get_itbl(bqe)->type == TSO ||
4165 get_itbl(bqe)->type == CONSTR);
4166 /* only BQs of an RBH end with an RBH_Save closure */
4167 ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
4169 tso_loc = where_is((StgClosure *)bqe);
4170 switch (get_itbl(bqe)->type) {
4172 debugBelch(" TSO %d (%p) on [PE %d],",
4173 ((StgTSO *)bqe)->id, (StgTSO *)bqe, tso_loc);
4176 debugBelch(" %s (IP %p),",
4177 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
4178 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
4179 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
4180 "RBH_Save_?"), get_itbl(bqe));
4183 barf("Unexpected closure type %s in blocking queue of %p (%s)",
4184 info_type((StgClosure *)bqe), node, info_type(node));
4192 #if defined(PARALLEL_HASKELL)
4199 for (i=0, tso=run_queue_hd;
4200 tso != END_TSO_QUEUE;
4201 i++, tso=tso->link) {
4210 sched_belch(char *s, ...)
4215 debugBelch("sched (task %p): ", (void *)(unsigned long)(unsigned int)osThreadId());
4216 #elif defined(PARALLEL_HASKELL)
4219 debugBelch("sched: ");