1 /* ---------------------------------------------------------------------------
3 * (c) The GHC Team, 1998-2004
7 * Different GHC ways use this scheduler quite differently (see comments below)
8 * Here is the global picture:
10 * WAY Name CPP flag What's it for
11 * --------------------------------------
12 * mp GUM PARALLEL_HASKELL Parallel execution on a distrib. memory machine
13 * s SMP SMP Parallel execution on a shared memory machine
14 * mg GranSim GRAN Simulation of parallel execution
15 * md GUM/GdH DIST Distributed execution (based on GUM)
17 * --------------------------------------------------------------------------*/
20 * Version with support for distributed memory parallelism aka GUM (WAY=mp):
22 The main scheduling loop in GUM iterates until a finish message is received.
23 In that case a global flag @receivedFinish@ is set and this instance of
24 the RTS shuts down. See ghc/rts/parallel/HLComms.c:processMessages()
25 for the handling of incoming messages, such as PP_FINISH.
26 Note that in the parallel case we have a system manager that coordinates
27 different PEs, each of which are running one instance of the RTS.
28 See ghc/rts/parallel/SysMan.c for the main routine of the parallel program.
29 From this routine processes executing ghc/rts/Main.c are spawned. -- HWL
31 * Version with support for simulating parallel execution aka GranSim (WAY=mg):
33 The main scheduling code in GranSim is quite different from that in std
34 (concurrent) Haskell: while concurrent Haskell just iterates over the
35 threads in the runnable queue, GranSim is event driven, i.e. it iterates
36 over the events in the global event queue. -- HWL
39 #include "PosixSource.h"
44 #include "BlockAlloc.h"
45 #include "OSThreads.h"
49 #define COMPILING_SCHEDULER
51 #include "StgMiscClosures.h"
52 #include "Interpreter.h"
53 #include "Exception.h"
61 #include "ThreadLabels.h"
62 #include "LdvProfile.h"
65 #include "Proftimer.h"
68 #if defined(GRAN) || defined(PARALLEL_HASKELL)
69 # include "GranSimRts.h"
71 # include "ParallelRts.h"
72 # include "Parallel.h"
73 # include "ParallelDebug.h"
78 #include "Capability.h"
81 #ifdef HAVE_SYS_TYPES_H
82 #include <sys/types.h>
96 // Turn off inlining when debugging - it obfuscates things
99 # define STATIC_INLINE static
103 #define USED_IN_THREADED_RTS
105 #define USED_IN_THREADED_RTS STG_UNUSED
108 #ifdef RTS_SUPPORTS_THREADS
109 #define USED_WHEN_RTS_SUPPORTS_THREADS
111 #define USED_WHEN_RTS_SUPPORTS_THREADS STG_UNUSED
114 /* Main thread queue.
115 * Locks required: sched_mutex.
117 StgMainThread *main_threads = NULL;
121 StgTSO* ActiveTSO = NULL; /* for assigning system costs; GranSim-Light only */
122 /* rtsTime TimeOfNextEvent, EndOfTimeSlice; now in GranSim.c */
125 In GranSim we have a runnable and a blocked queue for each processor.
126 In order to minimise code changes new arrays run_queue_hds/tls
127 are created. run_queue_hd is then a short cut (macro) for
128 run_queue_hds[CurrentProc] (see GranSim.h).
131 StgTSO *run_queue_hds[MAX_PROC], *run_queue_tls[MAX_PROC];
132 StgTSO *blocked_queue_hds[MAX_PROC], *blocked_queue_tls[MAX_PROC];
133 StgTSO *ccalling_threadss[MAX_PROC];
134 /* We use the same global list of threads (all_threads) in GranSim as in
135 the std RTS (i.e. we are cheating). However, we don't use this list in
136 the GranSim specific code at the moment (so we are only potentially
142 * Locks required: sched_mutex.
144 StgTSO *run_queue_hd = NULL;
145 StgTSO *run_queue_tl = NULL;
146 StgTSO *blocked_queue_hd = NULL;
147 StgTSO *blocked_queue_tl = NULL;
148 StgTSO *blackhole_queue = NULL;
149 StgTSO *sleeping_queue = NULL; /* perhaps replace with a hash table? */
153 /* The blackhole_queue should be checked for threads to wake up. See
154 * Schedule.h for more thorough comment.
156 rtsBool blackholes_need_checking = rtsFalse;
158 /* Linked list of all threads.
159 * Used for detecting garbage collected threads.
161 StgTSO *all_threads = NULL;
163 /* When a thread performs a safe C call (_ccall_GC, using old
164 * terminology), it gets put on the suspended_ccalling_threads
165 * list. Used by the garbage collector.
167 static StgTSO *suspended_ccalling_threads;
169 /* KH: The following two flags are shared memory locations. There is no need
170 to lock them, since they are only unset at the end of a scheduler
174 /* flag set by signal handler to precipitate a context switch */
175 int context_switch = 0;
177 /* flag that tracks whether we have done any execution in this time slice. */
178 nat recent_activity = ACTIVITY_YES;
180 /* if this flag is set as well, give up execution */
181 rtsBool interrupted = rtsFalse;
183 /* Next thread ID to allocate.
184 * Locks required: thread_id_mutex
186 static StgThreadID next_thread_id = 1;
189 * Pointers to the state of the current thread.
190 * Rule of thumb: if CurrentTSO != NULL, then we're running a Haskell
191 * thread. If CurrentTSO == NULL, then we're at the scheduler level.
194 /* The smallest stack size that makes any sense is:
195 * RESERVED_STACK_WORDS (so we can get back from the stack overflow)
196 * + sizeofW(StgStopFrame) (the stg_stop_thread_info frame)
197 * + 1 (the closure to enter)
199 * + 1 (spare slot req'd by stg_ap_v_ret)
201 * A thread with this stack will bomb immediately with a stack
202 * overflow, which will increase its stack size.
205 #define MIN_STACK_WORDS (RESERVED_STACK_WORDS + sizeofW(StgStopFrame) + 3)
212 /* This is used in `TSO.h' and gcc 2.96 insists that this variable actually
213 * exists - earlier gccs apparently didn't.
219 * Set to TRUE when entering a shutdown state (via shutdownHaskellAndExit()) --
220 * in an MT setting, needed to signal that a worker thread shouldn't hang around
221 * in the scheduler when it is out of work.
223 static rtsBool shutting_down_scheduler = rtsFalse;
225 #if defined(RTS_SUPPORTS_THREADS)
226 /* ToDo: carefully document the invariants that go together
227 * with these synchronisation objects.
229 Mutex sched_mutex = INIT_MUTEX_VAR;
230 Mutex term_mutex = INIT_MUTEX_VAR;
232 #endif /* RTS_SUPPORTS_THREADS */
234 #if defined(PARALLEL_HASKELL)
236 rtsTime TimeOfLastYield;
237 rtsBool emitSchedule = rtsTrue;
241 static char *whatNext_strs[] = {
251 /* -----------------------------------------------------------------------------
252 * static function prototypes
253 * -------------------------------------------------------------------------- */
255 #if defined(RTS_SUPPORTS_THREADS)
256 static void taskStart(void);
259 static void schedule( StgMainThread *mainThread USED_WHEN_RTS_SUPPORTS_THREADS,
260 Capability *initialCapability );
263 // These function all encapsulate parts of the scheduler loop, and are
264 // abstracted only to make the structure and control flow of the
265 // scheduler clearer.
267 static void schedulePreLoop(void);
268 static void scheduleStartSignalHandlers(void);
269 static void scheduleCheckBlockedThreads(void);
270 static void scheduleCheckBlackHoles(void);
271 static void scheduleDetectDeadlock(void);
273 static StgTSO *scheduleProcessEvent(rtsEvent *event);
275 #if defined(PARALLEL_HASKELL)
276 static StgTSO *scheduleSendPendingMessages(void);
277 static void scheduleActivateSpark(void);
278 static rtsBool scheduleGetRemoteWork(rtsBool *receivedFinish);
280 #if defined(PAR) || defined(GRAN)
281 static void scheduleGranParReport(void);
283 static void schedulePostRunThread(void);
284 static rtsBool scheduleHandleHeapOverflow( Capability *cap, StgTSO *t );
285 static void scheduleHandleStackOverflow( StgTSO *t);
286 static rtsBool scheduleHandleYield( StgTSO *t, nat prev_what_next );
287 static void scheduleHandleThreadBlocked( StgTSO *t );
288 static rtsBool scheduleHandleThreadFinished( StgMainThread *mainThread,
289 Capability *cap, StgTSO *t );
290 static rtsBool scheduleDoHeapProfile(rtsBool ready_to_gc);
291 static void scheduleDoGC(Capability *cap);
293 static void unblockThread(StgTSO *tso);
294 static rtsBool checkBlackHoles(void);
295 static SchedulerStatus waitThread_(/*out*/StgMainThread* m,
296 Capability *initialCapability
298 static void scheduleThread_ (StgTSO* tso);
299 static void AllRoots(evac_fn evac);
301 static StgTSO *threadStackOverflow(StgTSO *tso);
303 static void raiseAsync_(StgTSO *tso, StgClosure *exception,
304 rtsBool stop_at_atomically);
306 static void printThreadBlockage(StgTSO *tso);
307 static void printThreadStatus(StgTSO *tso);
309 #if defined(PARALLEL_HASKELL)
310 StgTSO * createSparkThread(rtsSpark spark);
311 StgTSO * activateSpark (rtsSpark spark);
314 /* ----------------------------------------------------------------------------
316 * ------------------------------------------------------------------------- */
318 #if defined(RTS_SUPPORTS_THREADS)
319 static nat startingWorkerThread = 0;
324 ACQUIRE_LOCK(&sched_mutex);
325 startingWorkerThread--;
328 RELEASE_LOCK(&sched_mutex);
332 startSchedulerTaskIfNecessary(void)
334 if ( !EMPTY_RUN_QUEUE()
335 && !shutting_down_scheduler // not if we're shutting down
336 && startingWorkerThread==0)
338 // we don't want to start another worker thread
339 // just because the last one hasn't yet reached the
340 // "waiting for capability" state
341 startingWorkerThread++;
342 if (!maybeStartNewWorker(taskStart)) {
343 startingWorkerThread--;
349 /* -----------------------------------------------------------------------------
350 * Putting a thread on the run queue: different scheduling policies
351 * -------------------------------------------------------------------------- */
354 addToRunQueue( StgTSO *t )
356 #if defined(PARALLEL_HASKELL)
357 if (RtsFlags.ParFlags.doFairScheduling) {
358 // this does round-robin scheduling; good for concurrency
359 APPEND_TO_RUN_QUEUE(t);
361 // this does unfair scheduling; good for parallelism
362 PUSH_ON_RUN_QUEUE(t);
365 // this does round-robin scheduling; good for concurrency
366 APPEND_TO_RUN_QUEUE(t);
370 /* ---------------------------------------------------------------------------
371 Main scheduling loop.
373 We use round-robin scheduling, each thread returning to the
374 scheduler loop when one of these conditions is detected:
377 * timer expires (thread yields)
382 Locking notes: we acquire the scheduler lock once at the beginning
383 of the scheduler loop, and release it when
385 * running a thread, or
386 * waiting for work, or
387 * waiting for a GC to complete.
390 In a GranSim setup this loop iterates over the global event queue.
391 This revolves around the global event queue, which determines what
392 to do next. Therefore, it's more complicated than either the
393 concurrent or the parallel (GUM) setup.
396 GUM iterates over incoming messages.
397 It starts with nothing to do (thus CurrentTSO == END_TSO_QUEUE),
398 and sends out a fish whenever it has nothing to do; in-between
399 doing the actual reductions (shared code below) it processes the
400 incoming messages and deals with delayed operations
401 (see PendingFetches).
402 This is not the ugliest code you could imagine, but it's bloody close.
404 ------------------------------------------------------------------------ */
407 schedule( StgMainThread *mainThread USED_WHEN_RTS_SUPPORTS_THREADS,
408 Capability *initialCapability )
412 StgThreadReturnCode ret;
415 #elif defined(PARALLEL_HASKELL)
418 rtsBool receivedFinish = rtsFalse;
420 nat tp_size, sp_size; // stats only
426 // Pre-condition: sched_mutex is held.
427 // We might have a capability, passed in as initialCapability.
428 cap = initialCapability;
430 #if !defined(RTS_SUPPORTS_THREADS)
431 // simply initialise it in the non-threaded case
432 grabCapability(&cap);
436 sched_belch("### NEW SCHEDULER LOOP (main thr: %p, cap: %p)",
437 mainThread, initialCapability);
442 // -----------------------------------------------------------
443 // Scheduler loop starts here:
445 #if defined(PARALLEL_HASKELL)
446 #define TERMINATION_CONDITION (!receivedFinish)
448 #define TERMINATION_CONDITION ((event = get_next_event()) != (rtsEvent*)NULL)
450 #define TERMINATION_CONDITION rtsTrue
453 while (TERMINATION_CONDITION) {
456 /* Choose the processor with the next event */
457 CurrentProc = event->proc;
458 CurrentTSO = event->tso;
461 IF_DEBUG(scheduler, printAllThreads());
463 #if defined(RTS_SUPPORTS_THREADS)
464 // Yield the capability to higher-priority tasks if necessary.
467 yieldCapability(&cap);
470 // If we do not currently hold a capability, we wait for one
473 waitForCapability(&sched_mutex, &cap,
474 mainThread ? &mainThread->bound_thread_cond : NULL);
477 // We now have a capability...
480 #if 0 /* extra sanity checking */
483 for (m = main_threads; m != NULL; m = m->link) {
484 ASSERT(get_itbl(m->tso)->type == TSO);
489 // Check whether we have re-entered the RTS from Haskell without
490 // going via suspendThread()/resumeThread (i.e. a 'safe' foreign
492 if (cap->r.rInHaskell) {
493 errorBelch("schedule: re-entered unsafely.\n"
494 " Perhaps a 'foreign import unsafe' should be 'safe'?");
499 // Test for interruption. If interrupted==rtsTrue, then either
500 // we received a keyboard interrupt (^C), or the scheduler is
501 // trying to shut down all the tasks (shutting_down_scheduler) in
505 if (shutting_down_scheduler) {
506 IF_DEBUG(scheduler, sched_belch("shutting down"));
507 releaseCapability(cap);
509 mainThread->stat = Interrupted;
510 mainThread->ret = NULL;
514 IF_DEBUG(scheduler, sched_belch("interrupted"));
519 #if defined(not_yet) && defined(SMP)
521 // Top up the run queue from our spark pool. We try to make the
522 // number of threads in the run queue equal to the number of
523 // free capabilities.
527 if (EMPTY_RUN_QUEUE()) {
528 spark = findSpark(rtsFalse);
530 break; /* no more sparks in the pool */
532 createSparkThread(spark);
534 sched_belch("==^^ turning spark of closure %p into a thread",
535 (StgClosure *)spark));
541 scheduleStartSignalHandlers();
543 // Only check the black holes here if we've nothing else to do.
544 // During normal execution, the black hole list only gets checked
545 // at GC time, to avoid repeatedly traversing this possibly long
546 // list each time around the scheduler.
547 if (EMPTY_RUN_QUEUE()) { scheduleCheckBlackHoles(); }
549 scheduleCheckBlockedThreads();
551 scheduleDetectDeadlock();
553 // Normally, the only way we can get here with no threads to
554 // run is if a keyboard interrupt received during
555 // scheduleCheckBlockedThreads() or scheduleDetectDeadlock().
556 // Additionally, it is not fatal for the
557 // threaded RTS to reach here with no threads to run.
559 // win32: might be here due to awaitEvent() being abandoned
560 // as a result of a console event having been delivered.
561 if ( EMPTY_RUN_QUEUE() ) {
562 #if !defined(RTS_SUPPORTS_THREADS) && !defined(mingw32_HOST_OS)
565 continue; // nothing to do
568 #if defined(PARALLEL_HASKELL)
569 scheduleSendPendingMessages();
570 if (EMPTY_RUN_QUEUE() && scheduleActivateSpark())
574 ASSERT(next_fish_to_send_at==0); // i.e. no delayed fishes left!
577 /* If we still have no work we need to send a FISH to get a spark
579 if (EMPTY_RUN_QUEUE()) {
580 if (!scheduleGetRemoteWork(&receivedFinish)) continue;
581 ASSERT(rtsFalse); // should not happen at the moment
583 // from here: non-empty run queue.
584 // TODO: merge above case with this, only one call processMessages() !
585 if (PacketsWaiting()) { /* process incoming messages, if
586 any pending... only in else
587 because getRemoteWork waits for
589 receivedFinish = processMessages();
594 scheduleProcessEvent(event);
598 // Get a thread to run
600 ASSERT(run_queue_hd != END_TSO_QUEUE);
603 #if defined(GRAN) || defined(PAR)
604 scheduleGranParReport(); // some kind of debuging output
606 // Sanity check the thread we're about to run. This can be
607 // expensive if there is lots of thread switching going on...
608 IF_DEBUG(sanity,checkTSO(t));
611 #if defined(RTS_SUPPORTS_THREADS)
612 // Check whether we can run this thread in the current task.
613 // If not, we have to pass our capability to the right task.
615 StgMainThread *m = t->main;
622 sched_belch("### Running thread %d in bound thread", t->id));
623 // yes, the Haskell thread is bound to the current native thread
628 sched_belch("### thread %d bound to another OS thread", t->id));
629 // no, bound to a different Haskell thread: pass to that thread
630 PUSH_ON_RUN_QUEUE(t);
631 passCapability(&m->bound_thread_cond);
637 if(mainThread != NULL)
638 // The thread we want to run is bound.
641 sched_belch("### this OS thread cannot run thread %d", t->id));
642 // no, the current native thread is bound to a different
643 // Haskell thread, so pass it to any worker thread
644 PUSH_ON_RUN_QUEUE(t);
645 passCapabilityToWorker();
652 cap->r.rCurrentTSO = t;
654 /* context switches are now initiated by the timer signal, unless
655 * the user specified "context switch as often as possible", with
658 if ((RtsFlags.ConcFlags.ctxtSwitchTicks == 0
659 && (run_queue_hd != END_TSO_QUEUE
660 || blocked_queue_hd != END_TSO_QUEUE
661 || sleeping_queue != END_TSO_QUEUE)))
666 RELEASE_LOCK(&sched_mutex);
668 IF_DEBUG(scheduler, sched_belch("-->> running thread %ld %s ...",
669 (long)t->id, whatNext_strs[t->what_next]));
671 #if defined(PROFILING)
672 startHeapProfTimer();
675 // ----------------------------------------------------------------------
676 // Run the current thread
678 prev_what_next = t->what_next;
680 errno = t->saved_errno;
681 cap->r.rInHaskell = rtsTrue;
683 recent_activity = ACTIVITY_YES;
685 switch (prev_what_next) {
689 /* Thread already finished, return to scheduler. */
690 ret = ThreadFinished;
694 ret = StgRun((StgFunPtr) stg_returnToStackTop, &cap->r);
697 case ThreadInterpret:
698 ret = interpretBCO(cap);
702 barf("schedule: invalid what_next field");
705 // We have run some Haskell code: there might be blackhole-blocked
706 // threads to wake up now.
707 if ( blackhole_queue != END_TSO_QUEUE ) {
708 blackholes_need_checking = rtsTrue;
711 cap->r.rInHaskell = rtsFalse;
713 // The TSO might have moved, eg. if it re-entered the RTS and a GC
714 // happened. So find the new location:
715 t = cap->r.rCurrentTSO;
717 // And save the current errno in this thread.
718 t->saved_errno = errno;
720 // ----------------------------------------------------------------------
722 /* Costs for the scheduler are assigned to CCS_SYSTEM */
723 #if defined(PROFILING)
728 ACQUIRE_LOCK(&sched_mutex);
730 #if defined(RTS_SUPPORTS_THREADS)
731 IF_DEBUG(scheduler,debugBelch("sched (task %p): ", osThreadId()););
732 #elif !defined(GRAN) && !defined(PARALLEL_HASKELL)
733 IF_DEBUG(scheduler,debugBelch("sched: "););
736 schedulePostRunThread();
738 ready_to_gc = rtsFalse;
742 ready_to_gc = scheduleHandleHeapOverflow(cap,t);
746 scheduleHandleStackOverflow(t);
750 if (scheduleHandleYield(t, prev_what_next)) {
751 // shortcut for switching between compiler/interpreter:
757 scheduleHandleThreadBlocked(t);
762 if (scheduleHandleThreadFinished(mainThread, cap, t)) return;;
766 barf("schedule: invalid thread return code %d", (int)ret);
769 if (scheduleDoHeapProfile(ready_to_gc)) { ready_to_gc = rtsFalse; }
770 if (ready_to_gc) { scheduleDoGC(cap); }
771 } /* end of while() */
773 IF_PAR_DEBUG(verbose,
774 debugBelch("== Leaving schedule() after having received Finish\n"));
777 /* ----------------------------------------------------------------------------
778 * Setting up the scheduler loop
779 * ASSUMES: sched_mutex
780 * ------------------------------------------------------------------------- */
783 schedulePreLoop(void)
786 /* set up first event to get things going */
787 /* ToDo: assign costs for system setup and init MainTSO ! */
788 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
790 CurrentTSO, (StgClosure*)NULL, (rtsSpark*)NULL);
793 debugBelch("GRAN: Init CurrentTSO (in schedule) = %p\n",
795 G_TSO(CurrentTSO, 5));
797 if (RtsFlags.GranFlags.Light) {
798 /* Save current time; GranSim Light only */
799 CurrentTSO->gran.clock = CurrentTime[CurrentProc];
804 /* ----------------------------------------------------------------------------
805 * Start any pending signal handlers
806 * ASSUMES: sched_mutex
807 * ------------------------------------------------------------------------- */
810 scheduleStartSignalHandlers(void)
812 #if defined(RTS_USER_SIGNALS) && !defined(RTS_SUPPORTS_THREADS)
813 if (signals_pending()) {
814 RELEASE_LOCK(&sched_mutex); /* ToDo: kill */
815 startSignalHandlers();
816 ACQUIRE_LOCK(&sched_mutex);
821 /* ----------------------------------------------------------------------------
822 * Check for blocked threads that can be woken up.
823 * ASSUMES: sched_mutex
824 * ------------------------------------------------------------------------- */
827 scheduleCheckBlockedThreads(void)
830 // Check whether any waiting threads need to be woken up. If the
831 // run queue is empty, and there are no other tasks running, we
832 // can wait indefinitely for something to happen.
834 if ( !EMPTY_QUEUE(blocked_queue_hd) || !EMPTY_QUEUE(sleeping_queue) )
836 #if defined(RTS_SUPPORTS_THREADS)
837 // We shouldn't be here...
838 barf("schedule: awaitEvent() in threaded RTS");
840 awaitEvent( EMPTY_RUN_QUEUE() && !blackholes_need_checking );
845 /* ----------------------------------------------------------------------------
846 * Check for threads blocked on BLACKHOLEs that can be woken up
847 * ASSUMES: sched_mutex
848 * ------------------------------------------------------------------------- */
850 scheduleCheckBlackHoles( void )
852 if ( blackholes_need_checking )
855 blackholes_need_checking = rtsFalse;
859 /* ----------------------------------------------------------------------------
860 * Detect deadlock conditions and attempt to resolve them.
861 * ASSUMES: sched_mutex
862 * ------------------------------------------------------------------------- */
865 scheduleDetectDeadlock(void)
868 #if defined(PARALLEL_HASKELL)
869 // ToDo: add deadlock detection in GUM (similar to SMP) -- HWL
874 * Detect deadlock: when we have no threads to run, there are no
875 * threads blocked, waiting for I/O, or sleeping, and all the
876 * other tasks are waiting for work, we must have a deadlock of
879 if ( EMPTY_THREAD_QUEUES() )
881 #if defined(RTS_SUPPORTS_THREADS)
883 * In the threaded RTS, we only check for deadlock if there
884 * has been no activity in a complete timeslice. This means
885 * we won't eagerly start a full GC just because we don't have
886 * any threads to run currently.
888 if (recent_activity != ACTIVITY_INACTIVE) return;
891 IF_DEBUG(scheduler, sched_belch("deadlocked, forcing major GC..."));
893 // Garbage collection can release some new threads due to
894 // either (a) finalizers or (b) threads resurrected because
895 // they are unreachable and will therefore be sent an
896 // exception. Any threads thus released will be immediately
898 GarbageCollect(GetRoots,rtsTrue);
899 recent_activity = ACTIVITY_DONE_GC;
900 if ( !EMPTY_RUN_QUEUE() ) return;
902 #if defined(RTS_USER_SIGNALS) && !defined(RTS_SUPPORTS_THREADS)
903 /* If we have user-installed signal handlers, then wait
904 * for signals to arrive rather then bombing out with a
907 if ( anyUserHandlers() ) {
909 sched_belch("still deadlocked, waiting for signals..."));
913 if (signals_pending()) {
914 RELEASE_LOCK(&sched_mutex);
915 startSignalHandlers();
916 ACQUIRE_LOCK(&sched_mutex);
919 // either we have threads to run, or we were interrupted:
920 ASSERT(!EMPTY_RUN_QUEUE() || interrupted);
924 #if !defined(RTS_SUPPORTS_THREADS)
925 /* Probably a real deadlock. Send the current main thread the
926 * Deadlock exception (or in the SMP build, send *all* main
927 * threads the deadlock exception, since none of them can make
933 switch (m->tso->why_blocked) {
935 case BlockedOnBlackHole:
936 case BlockedOnException:
938 raiseAsync(m->tso, (StgClosure *)NonTermination_closure);
941 barf("deadlock: main thread blocked in a strange way");
948 /* ----------------------------------------------------------------------------
949 * Process an event (GRAN only)
950 * ------------------------------------------------------------------------- */
954 scheduleProcessEvent(rtsEvent *event)
958 if (RtsFlags.GranFlags.Light)
959 GranSimLight_enter_system(event, &ActiveTSO); // adjust ActiveTSO etc
961 /* adjust time based on time-stamp */
962 if (event->time > CurrentTime[CurrentProc] &&
963 event->evttype != ContinueThread)
964 CurrentTime[CurrentProc] = event->time;
966 /* Deal with the idle PEs (may issue FindWork or MoveSpark events) */
967 if (!RtsFlags.GranFlags.Light)
970 IF_DEBUG(gran, debugBelch("GRAN: switch by event-type\n"));
972 /* main event dispatcher in GranSim */
973 switch (event->evttype) {
974 /* Should just be continuing execution */
976 IF_DEBUG(gran, debugBelch("GRAN: doing ContinueThread\n"));
977 /* ToDo: check assertion
978 ASSERT(run_queue_hd != (StgTSO*)NULL &&
979 run_queue_hd != END_TSO_QUEUE);
981 /* Ignore ContinueThreads for fetching threads (if synchr comm) */
982 if (!RtsFlags.GranFlags.DoAsyncFetch &&
983 procStatus[CurrentProc]==Fetching) {
984 debugBelch("ghuH: Spurious ContinueThread while Fetching ignored; TSO %d (%p) [PE %d]\n",
985 CurrentTSO->id, CurrentTSO, CurrentProc);
988 /* Ignore ContinueThreads for completed threads */
989 if (CurrentTSO->what_next == ThreadComplete) {
990 debugBelch("ghuH: found a ContinueThread event for completed thread %d (%p) [PE %d] (ignoring ContinueThread)\n",
991 CurrentTSO->id, CurrentTSO, CurrentProc);
994 /* Ignore ContinueThreads for threads that are being migrated */
995 if (PROCS(CurrentTSO)==Nowhere) {
996 debugBelch("ghuH: trying to run the migrating TSO %d (%p) [PE %d] (ignoring ContinueThread)\n",
997 CurrentTSO->id, CurrentTSO, CurrentProc);
1000 /* The thread should be at the beginning of the run queue */
1001 if (CurrentTSO!=run_queue_hds[CurrentProc]) {
1002 debugBelch("ghuH: TSO %d (%p) [PE %d] is not at the start of the run_queue when doing a ContinueThread\n",
1003 CurrentTSO->id, CurrentTSO, CurrentProc);
1004 break; // run the thread anyway
1007 new_event(proc, proc, CurrentTime[proc],
1009 (StgTSO*)NULL, (StgClosure*)NULL, (rtsSpark*)NULL);
1011 */ /* Catches superfluous CONTINUEs -- should be unnecessary */
1012 break; // now actually run the thread; DaH Qu'vam yImuHbej
1015 do_the_fetchnode(event);
1016 goto next_thread; /* handle next event in event queue */
1019 do_the_globalblock(event);
1020 goto next_thread; /* handle next event in event queue */
1023 do_the_fetchreply(event);
1024 goto next_thread; /* handle next event in event queue */
1026 case UnblockThread: /* Move from the blocked queue to the tail of */
1027 do_the_unblock(event);
1028 goto next_thread; /* handle next event in event queue */
1030 case ResumeThread: /* Move from the blocked queue to the tail of */
1031 /* the runnable queue ( i.e. Qu' SImqa'lu') */
1032 event->tso->gran.blocktime +=
1033 CurrentTime[CurrentProc] - event->tso->gran.blockedat;
1034 do_the_startthread(event);
1035 goto next_thread; /* handle next event in event queue */
1038 do_the_startthread(event);
1039 goto next_thread; /* handle next event in event queue */
1042 do_the_movethread(event);
1043 goto next_thread; /* handle next event in event queue */
1046 do_the_movespark(event);
1047 goto next_thread; /* handle next event in event queue */
1050 do_the_findwork(event);
1051 goto next_thread; /* handle next event in event queue */
1054 barf("Illegal event type %u\n", event->evttype);
1057 /* This point was scheduler_loop in the old RTS */
1059 IF_DEBUG(gran, debugBelch("GRAN: after main switch\n"));
1061 TimeOfLastEvent = CurrentTime[CurrentProc];
1062 TimeOfNextEvent = get_time_of_next_event();
1063 IgnoreEvents=(TimeOfNextEvent==0); // HWL HACK
1064 // CurrentTSO = ThreadQueueHd;
1066 IF_DEBUG(gran, debugBelch("GRAN: time of next event is: %ld\n",
1069 if (RtsFlags.GranFlags.Light)
1070 GranSimLight_leave_system(event, &ActiveTSO);
1072 EndOfTimeSlice = CurrentTime[CurrentProc]+RtsFlags.GranFlags.time_slice;
1075 debugBelch("GRAN: end of time-slice is %#lx\n", EndOfTimeSlice));
1077 /* in a GranSim setup the TSO stays on the run queue */
1079 /* Take a thread from the run queue. */
1080 POP_RUN_QUEUE(t); // take_off_run_queue(t);
1083 debugBelch("GRAN: About to run current thread, which is\n");
1086 context_switch = 0; // turned on via GranYield, checking events and time slice
1089 DumpGranEvent(GR_SCHEDULE, t));
1091 procStatus[CurrentProc] = Busy;
1095 /* ----------------------------------------------------------------------------
1096 * Send pending messages (PARALLEL_HASKELL only)
1097 * ------------------------------------------------------------------------- */
1099 #if defined(PARALLEL_HASKELL)
1101 scheduleSendPendingMessages(void)
1107 # if defined(PAR) // global Mem.Mgmt., omit for now
1108 if (PendingFetches != END_BF_QUEUE) {
1113 if (RtsFlags.ParFlags.BufferTime) {
1114 // if we use message buffering, we must send away all message
1115 // packets which have become too old...
1121 /* ----------------------------------------------------------------------------
1122 * Activate spark threads (PARALLEL_HASKELL only)
1123 * ------------------------------------------------------------------------- */
1125 #if defined(PARALLEL_HASKELL)
1127 scheduleActivateSpark(void)
1130 ASSERT(EMPTY_RUN_QUEUE());
1131 /* We get here if the run queue is empty and want some work.
1132 We try to turn a spark into a thread, and add it to the run queue,
1133 from where it will be picked up in the next iteration of the scheduler
1137 /* :-[ no local threads => look out for local sparks */
1138 /* the spark pool for the current PE */
1139 pool = &(cap.r.rSparks); // JB: cap = (old) MainCap
1140 if (advisory_thread_count < RtsFlags.ParFlags.maxThreads &&
1141 pool->hd < pool->tl) {
1143 * ToDo: add GC code check that we really have enough heap afterwards!!
1145 * If we're here (no runnable threads) and we have pending
1146 * sparks, we must have a space problem. Get enough space
1147 * to turn one of those pending sparks into a
1151 spark = findSpark(rtsFalse); /* get a spark */
1152 if (spark != (rtsSpark) NULL) {
1153 tso = createThreadFromSpark(spark); /* turn the spark into a thread */
1154 IF_PAR_DEBUG(fish, // schedule,
1155 debugBelch("==== schedule: Created TSO %d (%p); %d threads active\n",
1156 tso->id, tso, advisory_thread_count));
1158 if (tso==END_TSO_QUEUE) { /* failed to activate spark->back to loop */
1159 IF_PAR_DEBUG(fish, // schedule,
1160 debugBelch("==^^ failed to create thread from spark @ %lx\n",
1162 return rtsFalse; /* failed to generate a thread */
1163 } /* otherwise fall through & pick-up new tso */
1165 IF_PAR_DEBUG(fish, // schedule,
1166 debugBelch("==^^ no local sparks (spark pool contains only NFs: %d)\n",
1167 spark_queue_len(pool)));
1168 return rtsFalse; /* failed to generate a thread */
1170 return rtsTrue; /* success in generating a thread */
1171 } else { /* no more threads permitted or pool empty */
1172 return rtsFalse; /* failed to generateThread */
1175 tso = NULL; // avoid compiler warning only
1176 return rtsFalse; /* dummy in non-PAR setup */
1179 #endif // PARALLEL_HASKELL
1181 /* ----------------------------------------------------------------------------
1182 * Get work from a remote node (PARALLEL_HASKELL only)
1183 * ------------------------------------------------------------------------- */
1185 #if defined(PARALLEL_HASKELL)
1187 scheduleGetRemoteWork(rtsBool *receivedFinish)
1189 ASSERT(EMPTY_RUN_QUEUE());
1191 if (RtsFlags.ParFlags.BufferTime) {
1192 IF_PAR_DEBUG(verbose,
1193 debugBelch("...send all pending data,"));
1196 for (i=1; i<=nPEs; i++)
1197 sendImmediately(i); // send all messages away immediately
1201 //++EDEN++ idle() , i.e. send all buffers, wait for work
1202 // suppress fishing in EDEN... just look for incoming messages
1203 // (blocking receive)
1204 IF_PAR_DEBUG(verbose,
1205 debugBelch("...wait for incoming messages...\n"));
1206 *receivedFinish = processMessages(); // blocking receive...
1208 // and reenter scheduling loop after having received something
1209 // (return rtsFalse below)
1211 # else /* activate SPARKS machinery */
1212 /* We get here, if we have no work, tried to activate a local spark, but still
1213 have no work. We try to get a remote spark, by sending a FISH message.
1214 Thread migration should be added here, and triggered when a sequence of
1215 fishes returns without work. */
1216 delay = (RtsFlags.ParFlags.fishDelay!=0ll ? RtsFlags.ParFlags.fishDelay : 0ll);
1218 /* =8-[ no local sparks => look for work on other PEs */
1220 * We really have absolutely no work. Send out a fish
1221 * (there may be some out there already), and wait for
1222 * something to arrive. We clearly can't run any threads
1223 * until a SCHEDULE or RESUME arrives, and so that's what
1224 * we're hoping to see. (Of course, we still have to
1225 * respond to other types of messages.)
1227 rtsTime now = msTime() /*CURRENT_TIME*/;
1228 IF_PAR_DEBUG(verbose,
1229 debugBelch("-- now=%ld\n", now));
1230 IF_PAR_DEBUG(fish, // verbose,
1231 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
1232 (last_fish_arrived_at!=0 &&
1233 last_fish_arrived_at+delay > now)) {
1234 debugBelch("--$$ <%llu> delaying FISH until %llu (last fish %llu, delay %llu)\n",
1235 now, last_fish_arrived_at+delay,
1236 last_fish_arrived_at,
1240 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
1241 advisory_thread_count < RtsFlags.ParFlags.maxThreads) { // send a FISH, but when?
1242 if (last_fish_arrived_at==0 ||
1243 (last_fish_arrived_at+delay <= now)) { // send FISH now!
1244 /* outstandingFishes is set in sendFish, processFish;
1245 avoid flooding system with fishes via delay */
1246 next_fish_to_send_at = 0;
1248 /* ToDo: this should be done in the main scheduling loop to avoid the
1249 busy wait here; not so bad if fish delay is very small */
1250 int iq = 0; // DEBUGGING -- HWL
1251 next_fish_to_send_at = last_fish_arrived_at+delay; // remember when to send
1252 /* send a fish when ready, but process messages that arrive in the meantime */
1254 if (PacketsWaiting()) {
1256 *receivedFinish = processMessages();
1259 } while (!*receivedFinish || now<next_fish_to_send_at);
1260 // JB: This means the fish could become obsolete, if we receive
1261 // work. Better check for work again?
1262 // last line: while (!receivedFinish || !haveWork || now<...)
1263 // next line: if (receivedFinish || haveWork )
1265 if (*receivedFinish) // no need to send a FISH if we are finishing anyway
1266 return rtsFalse; // NB: this will leave scheduler loop
1267 // immediately after return!
1269 IF_PAR_DEBUG(fish, // verbose,
1270 debugBelch("--$$ <%llu> sent delayed fish (%d processMessages); active/total threads=%d/%d\n",now,iq,run_queue_len(),advisory_thread_count));
1274 // JB: IMHO, this should all be hidden inside sendFish(...)
1276 sendFish(pe, thisPE, NEW_FISH_AGE, NEW_FISH_HISTORY,
1279 // Global statistics: count no. of fishes
1280 if (RtsFlags.ParFlags.ParStats.Global &&
1281 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
1282 globalParStats.tot_fish_mess++;
1286 /* delayed fishes must have been sent by now! */
1287 next_fish_to_send_at = 0;
1290 *receivedFinish = processMessages();
1291 # endif /* SPARKS */
1294 /* NB: this function always returns rtsFalse, meaning the scheduler
1295 loop continues with the next iteration;
1297 return code means success in finding work; we enter this function
1298 if there is no local work, thus have to send a fish which takes
1299 time until it arrives with work; in the meantime we should process
1300 messages in the main loop;
1303 #endif // PARALLEL_HASKELL
1305 /* ----------------------------------------------------------------------------
1306 * PAR/GRAN: Report stats & debugging info(?)
1307 * ------------------------------------------------------------------------- */
1309 #if defined(PAR) || defined(GRAN)
1311 scheduleGranParReport(void)
1313 ASSERT(run_queue_hd != END_TSO_QUEUE);
1315 /* Take a thread from the run queue, if we have work */
1316 POP_RUN_QUEUE(t); // take_off_run_queue(END_TSO_QUEUE);
1318 /* If this TSO has got its outport closed in the meantime,
1319 * it mustn't be run. Instead, we have to clean it up as if it was finished.
1320 * It has to be marked as TH_DEAD for this purpose.
1321 * If it is TH_TERM instead, it is supposed to have finished in the normal way.
1323 JB: TODO: investigate wether state change field could be nuked
1324 entirely and replaced by the normal tso state (whatnext
1325 field). All we want to do is to kill tsos from outside.
1328 /* ToDo: write something to the log-file
1329 if (RTSflags.ParFlags.granSimStats && !sameThread)
1330 DumpGranEvent(GR_SCHEDULE, RunnableThreadsHd);
1334 /* the spark pool for the current PE */
1335 pool = &(cap.r.rSparks); // cap = (old) MainCap
1338 debugBelch("--=^ %d threads, %d sparks on [%#x]\n",
1339 run_queue_len(), spark_queue_len(pool), CURRENT_PROC));
1342 debugBelch("--=^ %d threads, %d sparks on [%#x]\n",
1343 run_queue_len(), spark_queue_len(pool), CURRENT_PROC));
1345 if (RtsFlags.ParFlags.ParStats.Full &&
1346 (t->par.sparkname != (StgInt)0) && // only log spark generated threads
1347 (emitSchedule || // forced emit
1348 (t && LastTSO && t->id != LastTSO->id))) {
1350 we are running a different TSO, so write a schedule event to log file
1351 NB: If we use fair scheduling we also have to write a deschedule
1352 event for LastTSO; with unfair scheduling we know that the
1353 previous tso has blocked whenever we switch to another tso, so
1354 we don't need it in GUM for now
1356 IF_PAR_DEBUG(fish, // schedule,
1357 debugBelch("____ scheduling spark generated thread %d (%lx) (%lx) via a forced emit\n",t->id,t,t->par.sparkname));
1359 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
1360 GR_SCHEDULE, t, (StgClosure *)NULL, 0, 0);
1361 emitSchedule = rtsFalse;
1366 /* ----------------------------------------------------------------------------
1367 * After running a thread...
1368 * ASSUMES: sched_mutex
1369 * ------------------------------------------------------------------------- */
1372 schedulePostRunThread(void)
1375 /* HACK 675: if the last thread didn't yield, make sure to print a
1376 SCHEDULE event to the log file when StgRunning the next thread, even
1377 if it is the same one as before */
1379 TimeOfLastYield = CURRENT_TIME;
1382 /* some statistics gathering in the parallel case */
1384 #if defined(GRAN) || defined(PAR) || defined(EDEN)
1388 IF_DEBUG(gran, DumpGranEvent(GR_DESCHEDULE, t));
1389 globalGranStats.tot_heapover++;
1391 globalParStats.tot_heapover++;
1398 DumpGranEvent(GR_DESCHEDULE, t));
1399 globalGranStats.tot_stackover++;
1402 // DumpGranEvent(GR_DESCHEDULE, t);
1403 globalParStats.tot_stackover++;
1407 case ThreadYielding:
1410 DumpGranEvent(GR_DESCHEDULE, t));
1411 globalGranStats.tot_yields++;
1414 // DumpGranEvent(GR_DESCHEDULE, t);
1415 globalParStats.tot_yields++;
1422 debugBelch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: ",
1423 t->id, t, whatNext_strs[t->what_next], t->block_info.closure,
1424 (t->block_info.closure==(StgClosure*)NULL ? 99 : where_is(t->block_info.closure)));
1425 if (t->block_info.closure!=(StgClosure*)NULL)
1426 print_bq(t->block_info.closure);
1429 // ??? needed; should emit block before
1431 DumpGranEvent(GR_DESCHEDULE, t));
1432 prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
1435 ASSERT(procStatus[CurrentProc]==Busy ||
1436 ((procStatus[CurrentProc]==Fetching) &&
1437 (t->block_info.closure!=(StgClosure*)NULL)));
1438 if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
1439 !(!RtsFlags.GranFlags.DoAsyncFetch &&
1440 procStatus[CurrentProc]==Fetching))
1441 procStatus[CurrentProc] = Idle;
1444 //++PAR++ blockThread() writes the event (change?)
1448 case ThreadFinished:
1452 barf("parGlobalStats: unknown return code");
1458 /* -----------------------------------------------------------------------------
1459 * Handle a thread that returned to the scheduler with ThreadHeepOverflow
1460 * ASSUMES: sched_mutex
1461 * -------------------------------------------------------------------------- */
1464 scheduleHandleHeapOverflow( Capability *cap, StgTSO *t )
1466 // did the task ask for a large block?
1467 if (cap->r.rHpAlloc > BLOCK_SIZE) {
1468 // if so, get one and push it on the front of the nursery.
1472 blocks = (lnat)BLOCK_ROUND_UP(cap->r.rHpAlloc) / BLOCK_SIZE;
1475 debugBelch("--<< thread %ld (%s) stopped: requesting a large block (size %ld)\n",
1476 (long)t->id, whatNext_strs[t->what_next], blocks));
1478 // don't do this if the nursery is (nearly) full, we'll GC first.
1479 if (cap->r.rCurrentNursery->link != NULL ||
1480 cap->r.rNursery->n_blocks == 1) { // paranoia to prevent infinite loop
1481 // if the nursery has only one block.
1483 bd = allocGroup( blocks );
1484 cap->r.rNursery->n_blocks += blocks;
1486 // link the new group into the list
1487 bd->link = cap->r.rCurrentNursery;
1488 bd->u.back = cap->r.rCurrentNursery->u.back;
1489 if (cap->r.rCurrentNursery->u.back != NULL) {
1490 cap->r.rCurrentNursery->u.back->link = bd;
1493 ASSERT(g0s0->blocks == cap->r.rCurrentNursery &&
1494 g0s0 == cap->r.rNursery);
1496 cap->r.rNursery->blocks = bd;
1498 cap->r.rCurrentNursery->u.back = bd;
1500 // initialise it as a nursery block. We initialise the
1501 // step, gen_no, and flags field of *every* sub-block in
1502 // this large block, because this is easier than making
1503 // sure that we always find the block head of a large
1504 // block whenever we call Bdescr() (eg. evacuate() and
1505 // isAlive() in the GC would both have to do this, at
1509 for (x = bd; x < bd + blocks; x++) {
1510 x->step = cap->r.rNursery;
1516 // This assert can be a killer if the app is doing lots
1517 // of large block allocations.
1518 IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
1520 // now update the nursery to point to the new block
1521 cap->r.rCurrentNursery = bd;
1523 // we might be unlucky and have another thread get on the
1524 // run queue before us and steal the large block, but in that
1525 // case the thread will just end up requesting another large
1527 PUSH_ON_RUN_QUEUE(t);
1528 return rtsFalse; /* not actually GC'ing */
1532 /* make all the running tasks block on a condition variable,
1533 * maybe set context_switch and wait till they all pile in,
1534 * then have them wait on a GC condition variable.
1537 debugBelch("--<< thread %ld (%s) stopped: HeapOverflow\n",
1538 (long)t->id, whatNext_strs[t->what_next]));
1541 ASSERT(!is_on_queue(t,CurrentProc));
1542 #elif defined(PARALLEL_HASKELL)
1543 /* Currently we emit a DESCHEDULE event before GC in GUM.
1544 ToDo: either add separate event to distinguish SYSTEM time from rest
1545 or just nuke this DESCHEDULE (and the following SCHEDULE) */
1546 if (0 && RtsFlags.ParFlags.ParStats.Full) {
1547 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
1548 GR_DESCHEDULE, t, (StgClosure *)NULL, 0, 0);
1549 emitSchedule = rtsTrue;
1553 PUSH_ON_RUN_QUEUE(t);
1555 /* actual GC is done at the end of the while loop in schedule() */
1558 /* -----------------------------------------------------------------------------
1559 * Handle a thread that returned to the scheduler with ThreadStackOverflow
1560 * ASSUMES: sched_mutex
1561 * -------------------------------------------------------------------------- */
1564 scheduleHandleStackOverflow( StgTSO *t)
1566 IF_DEBUG(scheduler,debugBelch("--<< thread %ld (%s) stopped, StackOverflow\n",
1567 (long)t->id, whatNext_strs[t->what_next]));
1568 /* just adjust the stack for this thread, then pop it back
1573 /* enlarge the stack */
1574 StgTSO *new_t = threadStackOverflow(t);
1576 /* This TSO has moved, so update any pointers to it from the
1577 * main thread stack. It better not be on any other queues...
1578 * (it shouldn't be).
1580 if (t->main != NULL) {
1581 t->main->tso = new_t;
1583 PUSH_ON_RUN_QUEUE(new_t);
1587 /* -----------------------------------------------------------------------------
1588 * Handle a thread that returned to the scheduler with ThreadYielding
1589 * ASSUMES: sched_mutex
1590 * -------------------------------------------------------------------------- */
1593 scheduleHandleYield( StgTSO *t, nat prev_what_next )
1595 // Reset the context switch flag. We don't do this just before
1596 // running the thread, because that would mean we would lose ticks
1597 // during GC, which can lead to unfair scheduling (a thread hogs
1598 // the CPU because the tick always arrives during GC). This way
1599 // penalises threads that do a lot of allocation, but that seems
1600 // better than the alternative.
1603 /* put the thread back on the run queue. Then, if we're ready to
1604 * GC, check whether this is the last task to stop. If so, wake
1605 * up the GC thread. getThread will block during a GC until the
1609 if (t->what_next != prev_what_next) {
1610 debugBelch("--<< thread %ld (%s) stopped to switch evaluators\n",
1611 (long)t->id, whatNext_strs[t->what_next]);
1613 debugBelch("--<< thread %ld (%s) stopped, yielding\n",
1614 (long)t->id, whatNext_strs[t->what_next]);
1619 //debugBelch("&& Doing sanity check on yielding TSO %ld.", t->id);
1621 ASSERT(t->link == END_TSO_QUEUE);
1623 // Shortcut if we're just switching evaluators: don't bother
1624 // doing stack squeezing (which can be expensive), just run the
1626 if (t->what_next != prev_what_next) {
1633 ASSERT(!is_on_queue(t,CurrentProc));
1636 //debugBelch("&& Doing sanity check on all ThreadQueues (and their TSOs).");
1637 checkThreadQsSanity(rtsTrue));
1644 /* add a ContinueThread event to actually process the thread */
1645 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1647 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1649 debugBelch("GRAN: eventq and runnableq after adding yielded thread to queue again:\n");
1656 /* -----------------------------------------------------------------------------
1657 * Handle a thread that returned to the scheduler with ThreadBlocked
1658 * ASSUMES: sched_mutex
1659 * -------------------------------------------------------------------------- */
1662 scheduleHandleThreadBlocked( StgTSO *t
1663 #if !defined(GRAN) && !defined(DEBUG)
1670 debugBelch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: \n",
1671 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)));
1672 if (t->block_info.closure!=(StgClosure*)NULL) print_bq(t->block_info.closure));
1674 // ??? needed; should emit block before
1676 DumpGranEvent(GR_DESCHEDULE, t));
1677 prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
1680 ASSERT(procStatus[CurrentProc]==Busy ||
1681 ((procStatus[CurrentProc]==Fetching) &&
1682 (t->block_info.closure!=(StgClosure*)NULL)));
1683 if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
1684 !(!RtsFlags.GranFlags.DoAsyncFetch &&
1685 procStatus[CurrentProc]==Fetching))
1686 procStatus[CurrentProc] = Idle;
1690 debugBelch("--<< thread %ld (%p; %s) stopped, blocking on node %p with BQ: \n",
1691 t->id, t, whatNext_strs[t->what_next], t->block_info.closure));
1694 if (t->block_info.closure!=(StgClosure*)NULL)
1695 print_bq(t->block_info.closure));
1697 /* Send a fetch (if BlockedOnGA) and dump event to log file */
1700 /* whatever we schedule next, we must log that schedule */
1701 emitSchedule = rtsTrue;
1704 /* don't need to do anything. Either the thread is blocked on
1705 * I/O, in which case we'll have called addToBlockedQueue
1706 * previously, or it's blocked on an MVar or Blackhole, in which
1707 * case it'll be on the relevant queue already.
1709 ASSERT(t->why_blocked != NotBlocked);
1711 debugBelch("--<< thread %d (%s) stopped: ",
1712 t->id, whatNext_strs[t->what_next]);
1713 printThreadBlockage(t);
1716 /* Only for dumping event to log file
1717 ToDo: do I need this in GranSim, too?
1723 /* -----------------------------------------------------------------------------
1724 * Handle a thread that returned to the scheduler with ThreadFinished
1725 * ASSUMES: sched_mutex
1726 * -------------------------------------------------------------------------- */
1729 scheduleHandleThreadFinished( StgMainThread *mainThread
1730 USED_WHEN_RTS_SUPPORTS_THREADS,
1734 /* Need to check whether this was a main thread, and if so,
1735 * return with the return value.
1737 * We also end up here if the thread kills itself with an
1738 * uncaught exception, see Exception.cmm.
1740 IF_DEBUG(scheduler,debugBelch("--++ thread %d (%s) finished\n",
1741 t->id, whatNext_strs[t->what_next]));
1744 endThread(t, CurrentProc); // clean-up the thread
1745 #elif defined(PARALLEL_HASKELL)
1746 /* For now all are advisory -- HWL */
1747 //if(t->priority==AdvisoryPriority) ??
1748 advisory_thread_count--; // JB: Caution with this counter, buggy!
1751 if(t->dist.priority==RevalPriority)
1755 # if defined(EDENOLD)
1756 // the thread could still have an outport... (BUG)
1757 if (t->eden.outport != -1) {
1758 // delete the outport for the tso which has finished...
1759 IF_PAR_DEBUG(eden_ports,
1760 debugBelch("WARNING: Scheduler removes outport %d for TSO %d.\n",
1761 t->eden.outport, t->id));
1764 // thread still in the process (HEAVY BUG! since outport has just been closed...)
1765 if (t->eden.epid != -1) {
1766 IF_PAR_DEBUG(eden_ports,
1767 debugBelch("WARNING: Scheduler removes TSO %d from process %d .\n",
1768 t->id, t->eden.epid));
1769 removeTSOfromProcess(t);
1774 if (RtsFlags.ParFlags.ParStats.Full &&
1775 !RtsFlags.ParFlags.ParStats.Suppressed)
1776 DumpEndEvent(CURRENT_PROC, t, rtsFalse /* not mandatory */);
1778 // t->par only contains statistics: left out for now...
1780 debugBelch("**** end thread: ended sparked thread %d (%lx); sparkname: %lx\n",
1781 t->id,t,t->par.sparkname));
1783 #endif // PARALLEL_HASKELL
1786 // Check whether the thread that just completed was a main
1787 // thread, and if so return with the result.
1789 // There is an assumption here that all thread completion goes
1790 // through this point; we need to make sure that if a thread
1791 // ends up in the ThreadKilled state, that it stays on the run
1792 // queue so it can be dealt with here.
1795 #if defined(RTS_SUPPORTS_THREADS)
1798 mainThread->tso == t
1802 // We are a bound thread: this must be our thread that just
1804 ASSERT(mainThread->tso == t);
1806 if (t->what_next == ThreadComplete) {
1807 if (mainThread->ret) {
1808 // NOTE: return val is tso->sp[1] (see StgStartup.hc)
1809 *(mainThread->ret) = (StgClosure *)mainThread->tso->sp[1];
1811 mainThread->stat = Success;
1813 if (mainThread->ret) {
1814 *(mainThread->ret) = NULL;
1817 mainThread->stat = Interrupted;
1819 mainThread->stat = Killed;
1823 removeThreadLabel((StgWord)mainThread->tso->id);
1825 if (mainThread->prev == NULL) {
1826 ASSERT(mainThread == main_threads);
1827 main_threads = mainThread->link;
1829 mainThread->prev->link = mainThread->link;
1831 if (mainThread->link != NULL) {
1832 mainThread->link->prev = mainThread->prev;
1834 releaseCapability(cap);
1835 return rtsTrue; // tells schedule() to return
1838 #ifdef RTS_SUPPORTS_THREADS
1839 ASSERT(t->main == NULL);
1841 if (t->main != NULL) {
1842 // Must be a main thread that is not the topmost one. Leave
1843 // it on the run queue until the stack has unwound to the
1844 // point where we can deal with this. Leaving it on the run
1845 // queue also ensures that the garbage collector knows about
1846 // this thread and its return value (it gets dropped from the
1847 // all_threads list so there's no other way to find it).
1848 APPEND_TO_RUN_QUEUE(t);
1854 /* -----------------------------------------------------------------------------
1855 * Perform a heap census, if PROFILING
1856 * -------------------------------------------------------------------------- */
1859 scheduleDoHeapProfile( rtsBool ready_to_gc STG_UNUSED )
1861 #if defined(PROFILING)
1862 // When we have +RTS -i0 and we're heap profiling, do a census at
1863 // every GC. This lets us get repeatable runs for debugging.
1864 if (performHeapProfile ||
1865 (RtsFlags.ProfFlags.profileInterval==0 &&
1866 RtsFlags.ProfFlags.doHeapProfile && ready_to_gc)) {
1867 GarbageCollect(GetRoots, rtsTrue);
1869 performHeapProfile = rtsFalse;
1870 return rtsTrue; // true <=> we already GC'd
1876 /* -----------------------------------------------------------------------------
1877 * Perform a garbage collection if necessary
1878 * ASSUMES: sched_mutex
1879 * -------------------------------------------------------------------------- */
1882 scheduleDoGC( Capability *cap STG_UNUSED )
1886 static rtsBool waiting_for_gc;
1887 int n_capabilities = RtsFlags.ParFlags.nNodes - 1;
1888 // subtract one because we're already holding one.
1889 Capability *caps[n_capabilities];
1893 // In order to GC, there must be no threads running Haskell code.
1894 // Therefore, the GC thread needs to hold *all* the capabilities,
1895 // and release them after the GC has completed.
1897 // This seems to be the simplest way: previous attempts involved
1898 // making all the threads with capabilities give up their
1899 // capabilities and sleep except for the *last* one, which
1900 // actually did the GC. But it's quite hard to arrange for all
1901 // the other tasks to sleep and stay asleep.
1904 // Someone else is already trying to GC
1905 if (waiting_for_gc) return;
1906 waiting_for_gc = rtsTrue;
1908 caps[n_capabilities] = cap;
1909 while (n_capabilities > 0) {
1910 IF_DEBUG(scheduler, sched_belch("ready_to_gc, grabbing all the capabilies (%d left)", n_capabilities));
1911 waitForReturnCapability(&sched_mutex, &cap);
1913 caps[n_capabilities] = cap;
1916 waiting_for_gc = rtsFalse;
1919 /* Kick any transactions which are invalid back to their
1920 * atomically frames. When next scheduled they will try to
1921 * commit, this commit will fail and they will retry.
1923 for (t = all_threads; t != END_TSO_QUEUE; t = t -> link) {
1924 if (t -> what_next != ThreadRelocated && t -> trec != NO_TREC && t -> why_blocked == NotBlocked) {
1925 if (!stmValidateTransaction (t -> trec)) {
1926 IF_DEBUG(stm, sched_belch("trec %p found wasting its time", t));
1928 // strip the stack back to the ATOMICALLY_FRAME, aborting
1929 // the (nested) transaction, and saving the stack of any
1930 // partially-evaluated thunks on the heap.
1931 raiseAsync_(t, NULL, rtsTrue);
1934 ASSERT(get_itbl((StgClosure *)t->sp)->type == ATOMICALLY_FRAME);
1940 // so this happens periodically:
1941 scheduleCheckBlackHoles();
1943 /* everybody back, start the GC.
1944 * Could do it in this thread, or signal a condition var
1945 * to do it in another thread. Either way, we need to
1946 * broadcast on gc_pending_cond afterward.
1948 #if defined(RTS_SUPPORTS_THREADS)
1949 IF_DEBUG(scheduler,sched_belch("doing GC"));
1951 GarbageCollect(GetRoots,rtsFalse);
1955 // release our stash of capabilities.
1957 for (i = 0; i < RtsFlags.ParFlags.nNodes-1; i++) {
1958 releaseCapability(caps[i]);
1964 /* add a ContinueThread event to continue execution of current thread */
1965 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1967 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1969 debugBelch("GRAN: eventq and runnableq after Garbage collection:\n\n");
1975 /* ---------------------------------------------------------------------------
1976 * rtsSupportsBoundThreads(): is the RTS built to support bound threads?
1977 * used by Control.Concurrent for error checking.
1978 * ------------------------------------------------------------------------- */
1981 rtsSupportsBoundThreads(void)
1983 #if defined(RTS_SUPPORTS_THREADS)
1990 /* ---------------------------------------------------------------------------
1991 * isThreadBound(tso): check whether tso is bound to an OS thread.
1992 * ------------------------------------------------------------------------- */
1995 isThreadBound(StgTSO* tso USED_IN_THREADED_RTS)
1997 #if defined(RTS_SUPPORTS_THREADS)
1998 return (tso->main != NULL);
2003 /* ---------------------------------------------------------------------------
2004 * Singleton fork(). Do not copy any running threads.
2005 * ------------------------------------------------------------------------- */
2007 #ifndef mingw32_HOST_OS
2008 #define FORKPROCESS_PRIMOP_SUPPORTED
2011 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
2013 deleteThreadImmediately(StgTSO *tso);
2016 forkProcess(HsStablePtr *entry
2017 #ifndef FORKPROCESS_PRIMOP_SUPPORTED
2022 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
2028 IF_DEBUG(scheduler,sched_belch("forking!"));
2029 rts_lock(); // This not only acquires sched_mutex, it also
2030 // makes sure that no other threads are running
2034 if (pid) { /* parent */
2036 /* just return the pid */
2040 } else { /* child */
2043 // delete all threads
2044 run_queue_hd = run_queue_tl = END_TSO_QUEUE;
2046 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
2049 // don't allow threads to catch the ThreadKilled exception
2050 deleteThreadImmediately(t);
2053 // wipe the main thread list
2054 while((m = main_threads) != NULL) {
2055 main_threads = m->link;
2056 # ifdef THREADED_RTS
2057 closeCondition(&m->bound_thread_cond);
2062 rc = rts_evalStableIO(entry, NULL); // run the action
2063 rts_checkSchedStatus("forkProcess",rc);
2067 hs_exit(); // clean up and exit
2070 #else /* !FORKPROCESS_PRIMOP_SUPPORTED */
2071 barf("forkProcess#: primop not supported, sorry!\n");
2076 /* ---------------------------------------------------------------------------
2077 * deleteAllThreads(): kill all the live threads.
2079 * This is used when we catch a user interrupt (^C), before performing
2080 * any necessary cleanups and running finalizers.
2082 * Locks: sched_mutex held.
2083 * ------------------------------------------------------------------------- */
2086 deleteAllThreads ( void )
2089 IF_DEBUG(scheduler,sched_belch("deleting all threads"));
2090 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
2091 if (t->what_next == ThreadRelocated) {
2094 next = t->global_link;
2099 // The run queue now contains a bunch of ThreadKilled threads. We
2100 // must not throw these away: the main thread(s) will be in there
2101 // somewhere, and the main scheduler loop has to deal with it.
2102 // Also, the run queue is the only thing keeping these threads from
2103 // being GC'd, and we don't want the "main thread has been GC'd" panic.
2105 ASSERT(blocked_queue_hd == END_TSO_QUEUE);
2106 ASSERT(blackhole_queue == END_TSO_QUEUE);
2107 ASSERT(sleeping_queue == END_TSO_QUEUE);
2110 /* startThread and insertThread are now in GranSim.c -- HWL */
2113 /* ---------------------------------------------------------------------------
2114 * Suspending & resuming Haskell threads.
2116 * When making a "safe" call to C (aka _ccall_GC), the task gives back
2117 * its capability before calling the C function. This allows another
2118 * task to pick up the capability and carry on running Haskell
2119 * threads. It also means that if the C call blocks, it won't lock
2122 * The Haskell thread making the C call is put to sleep for the
2123 * duration of the call, on the susepended_ccalling_threads queue. We
2124 * give out a token to the task, which it can use to resume the thread
2125 * on return from the C function.
2126 * ------------------------------------------------------------------------- */
2129 suspendThread( StgRegTable *reg )
2133 int saved_errno = errno;
2135 /* assume that *reg is a pointer to the StgRegTable part
2138 cap = (Capability *)((void *)((unsigned char*)reg - sizeof(StgFunTable)));
2140 ACQUIRE_LOCK(&sched_mutex);
2143 sched_belch("thread %d did a _ccall_gc", cap->r.rCurrentTSO->id));
2145 // XXX this might not be necessary --SDM
2146 cap->r.rCurrentTSO->what_next = ThreadRunGHC;
2148 threadPaused(cap->r.rCurrentTSO);
2149 cap->r.rCurrentTSO->link = suspended_ccalling_threads;
2150 suspended_ccalling_threads = cap->r.rCurrentTSO;
2152 if(cap->r.rCurrentTSO->blocked_exceptions == NULL) {
2153 cap->r.rCurrentTSO->why_blocked = BlockedOnCCall;
2154 cap->r.rCurrentTSO->blocked_exceptions = END_TSO_QUEUE;
2156 cap->r.rCurrentTSO->why_blocked = BlockedOnCCall_NoUnblockExc;
2159 /* Use the thread ID as the token; it should be unique */
2160 tok = cap->r.rCurrentTSO->id;
2162 /* Hand back capability */
2163 cap->r.rInHaskell = rtsFalse;
2164 releaseCapability(cap);
2166 #if defined(RTS_SUPPORTS_THREADS)
2167 /* Preparing to leave the RTS, so ensure there's a native thread/task
2168 waiting to take over.
2170 IF_DEBUG(scheduler, sched_belch("worker (token %d): leaving RTS", tok));
2173 RELEASE_LOCK(&sched_mutex);
2175 errno = saved_errno;
2180 resumeThread( StgInt tok )
2182 StgTSO *tso, **prev;
2184 int saved_errno = errno;
2186 #if defined(RTS_SUPPORTS_THREADS)
2187 /* Wait for permission to re-enter the RTS with the result. */
2188 ACQUIRE_LOCK(&sched_mutex);
2189 waitForReturnCapability(&sched_mutex, &cap);
2191 IF_DEBUG(scheduler, sched_belch("worker (token %d): re-entering RTS", tok));
2193 grabCapability(&cap);
2196 /* Remove the thread off of the suspended list */
2197 prev = &suspended_ccalling_threads;
2198 for (tso = suspended_ccalling_threads;
2199 tso != END_TSO_QUEUE;
2200 prev = &tso->link, tso = tso->link) {
2201 if (tso->id == (StgThreadID)tok) {
2206 if (tso == END_TSO_QUEUE) {
2207 barf("resumeThread: thread not found");
2209 tso->link = END_TSO_QUEUE;
2211 if(tso->why_blocked == BlockedOnCCall) {
2212 awakenBlockedQueueNoLock(tso->blocked_exceptions);
2213 tso->blocked_exceptions = NULL;
2216 /* Reset blocking status */
2217 tso->why_blocked = NotBlocked;
2219 cap->r.rCurrentTSO = tso;
2220 cap->r.rInHaskell = rtsTrue;
2221 RELEASE_LOCK(&sched_mutex);
2222 errno = saved_errno;
2226 /* ---------------------------------------------------------------------------
2227 * Comparing Thread ids.
2229 * This is used from STG land in the implementation of the
2230 * instances of Eq/Ord for ThreadIds.
2231 * ------------------------------------------------------------------------ */
2234 cmp_thread(StgPtr tso1, StgPtr tso2)
2236 StgThreadID id1 = ((StgTSO *)tso1)->id;
2237 StgThreadID id2 = ((StgTSO *)tso2)->id;
2239 if (id1 < id2) return (-1);
2240 if (id1 > id2) return 1;
2244 /* ---------------------------------------------------------------------------
2245 * Fetching the ThreadID from an StgTSO.
2247 * This is used in the implementation of Show for ThreadIds.
2248 * ------------------------------------------------------------------------ */
2250 rts_getThreadId(StgPtr tso)
2252 return ((StgTSO *)tso)->id;
2257 labelThread(StgPtr tso, char *label)
2262 /* Caveat: Once set, you can only set the thread name to "" */
2263 len = strlen(label)+1;
2264 buf = stgMallocBytes(len * sizeof(char), "Schedule.c:labelThread()");
2265 strncpy(buf,label,len);
2266 /* Update will free the old memory for us */
2267 updateThreadLabel(((StgTSO *)tso)->id,buf);
2271 /* ---------------------------------------------------------------------------
2272 Create a new thread.
2274 The new thread starts with the given stack size. Before the
2275 scheduler can run, however, this thread needs to have a closure
2276 (and possibly some arguments) pushed on its stack. See
2277 pushClosure() in Schedule.h.
2279 createGenThread() and createIOThread() (in SchedAPI.h) are
2280 convenient packaged versions of this function.
2282 currently pri (priority) is only used in a GRAN setup -- HWL
2283 ------------------------------------------------------------------------ */
2285 /* currently pri (priority) is only used in a GRAN setup -- HWL */
2287 createThread(nat size, StgInt pri)
2290 createThread(nat size)
2297 /* First check whether we should create a thread at all */
2298 #if defined(PARALLEL_HASKELL)
2299 /* check that no more than RtsFlags.ParFlags.maxThreads threads are created */
2300 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads) {
2302 debugBelch("{createThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)\n",
2303 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
2304 return END_TSO_QUEUE;
2310 ASSERT(!RtsFlags.GranFlags.Light || CurrentProc==0);
2313 // ToDo: check whether size = stack_size - TSO_STRUCT_SIZEW
2315 /* catch ridiculously small stack sizes */
2316 if (size < MIN_STACK_WORDS + TSO_STRUCT_SIZEW) {
2317 size = MIN_STACK_WORDS + TSO_STRUCT_SIZEW;
2320 stack_size = size - TSO_STRUCT_SIZEW;
2322 tso = (StgTSO *)allocate(size);
2323 TICK_ALLOC_TSO(stack_size, 0);
2325 SET_HDR(tso, &stg_TSO_info, CCS_SYSTEM);
2327 SET_GRAN_HDR(tso, ThisPE);
2330 // Always start with the compiled code evaluator
2331 tso->what_next = ThreadRunGHC;
2333 tso->id = next_thread_id++;
2334 tso->why_blocked = NotBlocked;
2335 tso->blocked_exceptions = NULL;
2337 tso->saved_errno = 0;
2340 tso->stack_size = stack_size;
2341 tso->max_stack_size = round_to_mblocks(RtsFlags.GcFlags.maxStkSize)
2343 tso->sp = (P_)&(tso->stack) + stack_size;
2345 tso->trec = NO_TREC;
2348 tso->prof.CCCS = CCS_MAIN;
2351 /* put a stop frame on the stack */
2352 tso->sp -= sizeofW(StgStopFrame);
2353 SET_HDR((StgClosure*)tso->sp,(StgInfoTable *)&stg_stop_thread_info,CCS_SYSTEM);
2354 tso->link = END_TSO_QUEUE;
2358 /* uses more flexible routine in GranSim */
2359 insertThread(tso, CurrentProc);
2361 /* In a non-GranSim setup the pushing of a TSO onto the runq is separated
2367 if (RtsFlags.GranFlags.GranSimStats.Full)
2368 DumpGranEvent(GR_START,tso);
2369 #elif defined(PARALLEL_HASKELL)
2370 if (RtsFlags.ParFlags.ParStats.Full)
2371 DumpGranEvent(GR_STARTQ,tso);
2372 /* HACk to avoid SCHEDULE
2376 /* Link the new thread on the global thread list.
2378 tso->global_link = all_threads;
2382 tso->dist.priority = MandatoryPriority; //by default that is...
2386 tso->gran.pri = pri;
2388 tso->gran.magic = TSO_MAGIC; // debugging only
2390 tso->gran.sparkname = 0;
2391 tso->gran.startedat = CURRENT_TIME;
2392 tso->gran.exported = 0;
2393 tso->gran.basicblocks = 0;
2394 tso->gran.allocs = 0;
2395 tso->gran.exectime = 0;
2396 tso->gran.fetchtime = 0;
2397 tso->gran.fetchcount = 0;
2398 tso->gran.blocktime = 0;
2399 tso->gran.blockcount = 0;
2400 tso->gran.blockedat = 0;
2401 tso->gran.globalsparks = 0;
2402 tso->gran.localsparks = 0;
2403 if (RtsFlags.GranFlags.Light)
2404 tso->gran.clock = Now; /* local clock */
2406 tso->gran.clock = 0;
2408 IF_DEBUG(gran,printTSO(tso));
2409 #elif defined(PARALLEL_HASKELL)
2411 tso->par.magic = TSO_MAGIC; // debugging only
2413 tso->par.sparkname = 0;
2414 tso->par.startedat = CURRENT_TIME;
2415 tso->par.exported = 0;
2416 tso->par.basicblocks = 0;
2417 tso->par.allocs = 0;
2418 tso->par.exectime = 0;
2419 tso->par.fetchtime = 0;
2420 tso->par.fetchcount = 0;
2421 tso->par.blocktime = 0;
2422 tso->par.blockcount = 0;
2423 tso->par.blockedat = 0;
2424 tso->par.globalsparks = 0;
2425 tso->par.localsparks = 0;
2429 globalGranStats.tot_threads_created++;
2430 globalGranStats.threads_created_on_PE[CurrentProc]++;
2431 globalGranStats.tot_sq_len += spark_queue_len(CurrentProc);
2432 globalGranStats.tot_sq_probes++;
2433 #elif defined(PARALLEL_HASKELL)
2434 // collect parallel global statistics (currently done together with GC stats)
2435 if (RtsFlags.ParFlags.ParStats.Global &&
2436 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
2437 //debugBelch("Creating thread %d @ %11.2f\n", tso->id, usertime());
2438 globalParStats.tot_threads_created++;
2444 sched_belch("==__ schedule: Created TSO %d (%p);",
2445 CurrentProc, tso, tso->id));
2446 #elif defined(PARALLEL_HASKELL)
2447 IF_PAR_DEBUG(verbose,
2448 sched_belch("==__ schedule: Created TSO %d (%p); %d threads active",
2449 (long)tso->id, tso, advisory_thread_count));
2451 IF_DEBUG(scheduler,sched_belch("created thread %ld, stack size = %lx words",
2452 (long)tso->id, (long)tso->stack_size));
2459 all parallel thread creation calls should fall through the following routine.
2462 createThreadFromSpark(rtsSpark spark)
2464 ASSERT(spark != (rtsSpark)NULL);
2465 // JB: TAKE CARE OF THIS COUNTER! BUGGY
2466 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads)
2468 barf("{createSparkThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
2469 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
2470 return END_TSO_QUEUE;
2474 tso = createThread(RtsFlags.GcFlags.initialStkSize);
2475 if (tso==END_TSO_QUEUE)
2476 barf("createSparkThread: Cannot create TSO");
2478 tso->priority = AdvisoryPriority;
2480 pushClosure(tso,spark);
2482 advisory_thread_count++; // JB: TAKE CARE OF THIS COUNTER! BUGGY
2489 Turn a spark into a thread.
2490 ToDo: fix for SMP (needs to acquire SCHED_MUTEX!)
2494 activateSpark (rtsSpark spark)
2498 tso = createSparkThread(spark);
2499 if (RtsFlags.ParFlags.ParStats.Full) {
2500 //ASSERT(run_queue_hd == END_TSO_QUEUE); // I think ...
2501 IF_PAR_DEBUG(verbose,
2502 debugBelch("==^^ activateSpark: turning spark of closure %p (%s) into a thread\n",
2503 (StgClosure *)spark, info_type((StgClosure *)spark)));
2505 // ToDo: fwd info on local/global spark to thread -- HWL
2506 // tso->gran.exported = spark->exported;
2507 // tso->gran.locked = !spark->global;
2508 // tso->gran.sparkname = spark->name;
2514 /* ---------------------------------------------------------------------------
2517 * scheduleThread puts a thread on the head of the runnable queue.
2518 * This will usually be done immediately after a thread is created.
2519 * The caller of scheduleThread must create the thread using e.g.
2520 * createThread and push an appropriate closure
2521 * on this thread's stack before the scheduler is invoked.
2522 * ------------------------------------------------------------------------ */
2525 scheduleThread_(StgTSO *tso)
2527 // The thread goes at the *end* of the run-queue, to avoid possible
2528 // starvation of any threads already on the queue.
2529 APPEND_TO_RUN_QUEUE(tso);
2534 scheduleThread(StgTSO* tso)
2536 ACQUIRE_LOCK(&sched_mutex);
2537 scheduleThread_(tso);
2538 RELEASE_LOCK(&sched_mutex);
2541 #if defined(RTS_SUPPORTS_THREADS)
2542 static Condition bound_cond_cache;
2543 static int bound_cond_cache_full = 0;
2548 scheduleWaitThread(StgTSO* tso, /*[out]*/HaskellObj* ret,
2549 Capability *initialCapability)
2551 // Precondition: sched_mutex must be held
2554 m = stgMallocBytes(sizeof(StgMainThread), "waitThread");
2559 m->link = main_threads;
2561 if (main_threads != NULL) {
2562 main_threads->prev = m;
2566 #if defined(RTS_SUPPORTS_THREADS)
2567 // Allocating a new condition for each thread is expensive, so we
2568 // cache one. This is a pretty feeble hack, but it helps speed up
2569 // consecutive call-ins quite a bit.
2570 if (bound_cond_cache_full) {
2571 m->bound_thread_cond = bound_cond_cache;
2572 bound_cond_cache_full = 0;
2574 initCondition(&m->bound_thread_cond);
2578 /* Put the thread on the main-threads list prior to scheduling the TSO.
2579 Failure to do so introduces a race condition in the MT case (as
2580 identified by Wolfgang Thaller), whereby the new task/OS thread
2581 created by scheduleThread_() would complete prior to the thread
2582 that spawned it managed to put 'itself' on the main-threads list.
2583 The upshot of it all being that the worker thread wouldn't get to
2584 signal the completion of the its work item for the main thread to
2585 see (==> it got stuck waiting.) -- sof 6/02.
2587 IF_DEBUG(scheduler, sched_belch("waiting for thread (%d)", tso->id));
2589 APPEND_TO_RUN_QUEUE(tso);
2590 // NB. Don't call threadRunnable() here, because the thread is
2591 // bound and only runnable by *this* OS thread, so waking up other
2592 // workers will just slow things down.
2594 return waitThread_(m, initialCapability);
2597 /* ---------------------------------------------------------------------------
2600 * Initialise the scheduler. This resets all the queues - if the
2601 * queues contained any threads, they'll be garbage collected at the
2604 * ------------------------------------------------------------------------ */
2612 for (i=0; i<=MAX_PROC; i++) {
2613 run_queue_hds[i] = END_TSO_QUEUE;
2614 run_queue_tls[i] = END_TSO_QUEUE;
2615 blocked_queue_hds[i] = END_TSO_QUEUE;
2616 blocked_queue_tls[i] = END_TSO_QUEUE;
2617 ccalling_threadss[i] = END_TSO_QUEUE;
2618 blackhole_queue[i] = END_TSO_QUEUE;
2619 sleeping_queue = END_TSO_QUEUE;
2622 run_queue_hd = END_TSO_QUEUE;
2623 run_queue_tl = END_TSO_QUEUE;
2624 blocked_queue_hd = END_TSO_QUEUE;
2625 blocked_queue_tl = END_TSO_QUEUE;
2626 blackhole_queue = END_TSO_QUEUE;
2627 sleeping_queue = END_TSO_QUEUE;
2630 suspended_ccalling_threads = END_TSO_QUEUE;
2632 main_threads = NULL;
2633 all_threads = END_TSO_QUEUE;
2638 RtsFlags.ConcFlags.ctxtSwitchTicks =
2639 RtsFlags.ConcFlags.ctxtSwitchTime / TICK_MILLISECS;
2641 #if defined(RTS_SUPPORTS_THREADS)
2642 /* Initialise the mutex and condition variables used by
2644 initMutex(&sched_mutex);
2645 initMutex(&term_mutex);
2648 ACQUIRE_LOCK(&sched_mutex);
2650 /* A capability holds the state a native thread needs in
2651 * order to execute STG code. At least one capability is
2652 * floating around (only SMP builds have more than one).
2656 #if defined(RTS_SUPPORTS_THREADS)
2661 /* eagerly start some extra workers */
2662 startingWorkerThread = RtsFlags.ParFlags.nNodes;
2663 startTasks(RtsFlags.ParFlags.nNodes, taskStart);
2666 #if /* defined(SMP) ||*/ defined(PARALLEL_HASKELL)
2670 RELEASE_LOCK(&sched_mutex);
2674 exitScheduler( void )
2676 interrupted = rtsTrue;
2677 shutting_down_scheduler = rtsTrue;
2678 #if defined(RTS_SUPPORTS_THREADS)
2679 if (threadIsTask(osThreadId())) { taskStop(); }
2684 /* ----------------------------------------------------------------------------
2685 Managing the per-task allocation areas.
2687 Each capability comes with an allocation area. These are
2688 fixed-length block lists into which allocation can be done.
2690 ToDo: no support for two-space collection at the moment???
2691 ------------------------------------------------------------------------- */
2693 static SchedulerStatus
2694 waitThread_(StgMainThread* m, Capability *initialCapability)
2696 SchedulerStatus stat;
2698 // Precondition: sched_mutex must be held.
2699 IF_DEBUG(scheduler, sched_belch("new main thread (%d)", m->tso->id));
2702 /* GranSim specific init */
2703 CurrentTSO = m->tso; // the TSO to run
2704 procStatus[MainProc] = Busy; // status of main PE
2705 CurrentProc = MainProc; // PE to run it on
2706 schedule(m,initialCapability);
2708 schedule(m,initialCapability);
2709 ASSERT(m->stat != NoStatus);
2714 #if defined(RTS_SUPPORTS_THREADS)
2715 // Free the condition variable, returning it to the cache if possible.
2716 if (!bound_cond_cache_full) {
2717 bound_cond_cache = m->bound_thread_cond;
2718 bound_cond_cache_full = 1;
2720 closeCondition(&m->bound_thread_cond);
2724 IF_DEBUG(scheduler, sched_belch("main thread (%d) finished", m->tso->id));
2727 // Postcondition: sched_mutex still held
2731 /* ---------------------------------------------------------------------------
2732 Where are the roots that we know about?
2734 - all the threads on the runnable queue
2735 - all the threads on the blocked queue
2736 - all the threads on the sleeping queue
2737 - all the thread currently executing a _ccall_GC
2738 - all the "main threads"
2740 ------------------------------------------------------------------------ */
2742 /* This has to be protected either by the scheduler monitor, or by the
2743 garbage collection monitor (probably the latter).
2748 GetRoots( evac_fn evac )
2753 for (i=0; i<=RtsFlags.GranFlags.proc; i++) {
2754 if ((run_queue_hds[i] != END_TSO_QUEUE) && ((run_queue_hds[i] != NULL)))
2755 evac((StgClosure **)&run_queue_hds[i]);
2756 if ((run_queue_tls[i] != END_TSO_QUEUE) && ((run_queue_tls[i] != NULL)))
2757 evac((StgClosure **)&run_queue_tls[i]);
2759 if ((blocked_queue_hds[i] != END_TSO_QUEUE) && ((blocked_queue_hds[i] != NULL)))
2760 evac((StgClosure **)&blocked_queue_hds[i]);
2761 if ((blocked_queue_tls[i] != END_TSO_QUEUE) && ((blocked_queue_tls[i] != NULL)))
2762 evac((StgClosure **)&blocked_queue_tls[i]);
2763 if ((ccalling_threadss[i] != END_TSO_QUEUE) && ((ccalling_threadss[i] != NULL)))
2764 evac((StgClosure **)&ccalling_threads[i]);
2771 if (run_queue_hd != END_TSO_QUEUE) {
2772 ASSERT(run_queue_tl != END_TSO_QUEUE);
2773 evac((StgClosure **)&run_queue_hd);
2774 evac((StgClosure **)&run_queue_tl);
2777 if (blocked_queue_hd != END_TSO_QUEUE) {
2778 ASSERT(blocked_queue_tl != END_TSO_QUEUE);
2779 evac((StgClosure **)&blocked_queue_hd);
2780 evac((StgClosure **)&blocked_queue_tl);
2783 if (sleeping_queue != END_TSO_QUEUE) {
2784 evac((StgClosure **)&sleeping_queue);
2788 if (blackhole_queue != END_TSO_QUEUE) {
2789 evac((StgClosure **)&blackhole_queue);
2792 if (suspended_ccalling_threads != END_TSO_QUEUE) {
2793 evac((StgClosure **)&suspended_ccalling_threads);
2796 #if defined(PARALLEL_HASKELL) || defined(GRAN)
2797 markSparkQueue(evac);
2800 #if defined(RTS_USER_SIGNALS)
2801 // mark the signal handlers (signals should be already blocked)
2802 markSignalHandlers(evac);
2806 /* -----------------------------------------------------------------------------
2809 This is the interface to the garbage collector from Haskell land.
2810 We provide this so that external C code can allocate and garbage
2811 collect when called from Haskell via _ccall_GC.
2813 It might be useful to provide an interface whereby the programmer
2814 can specify more roots (ToDo).
2816 This needs to be protected by the GC condition variable above. KH.
2817 -------------------------------------------------------------------------- */
2819 static void (*extra_roots)(evac_fn);
2824 /* Obligated to hold this lock upon entry */
2825 ACQUIRE_LOCK(&sched_mutex);
2826 GarbageCollect(GetRoots,rtsFalse);
2827 RELEASE_LOCK(&sched_mutex);
2831 performMajorGC(void)
2833 ACQUIRE_LOCK(&sched_mutex);
2834 GarbageCollect(GetRoots,rtsTrue);
2835 RELEASE_LOCK(&sched_mutex);
2839 AllRoots(evac_fn evac)
2841 GetRoots(evac); // the scheduler's roots
2842 extra_roots(evac); // the user's roots
2846 performGCWithRoots(void (*get_roots)(evac_fn))
2848 ACQUIRE_LOCK(&sched_mutex);
2849 extra_roots = get_roots;
2850 GarbageCollect(AllRoots,rtsFalse);
2851 RELEASE_LOCK(&sched_mutex);
2854 /* -----------------------------------------------------------------------------
2857 If the thread has reached its maximum stack size, then raise the
2858 StackOverflow exception in the offending thread. Otherwise
2859 relocate the TSO into a larger chunk of memory and adjust its stack
2861 -------------------------------------------------------------------------- */
2864 threadStackOverflow(StgTSO *tso)
2866 nat new_stack_size, stack_words;
2871 IF_DEBUG(sanity,checkTSO(tso));
2872 if (tso->stack_size >= tso->max_stack_size) {
2875 debugBelch("@@ threadStackOverflow of TSO %ld (%p): stack too large (now %ld; max is %ld)\n",
2876 (long)tso->id, tso, (long)tso->stack_size, (long)tso->max_stack_size);
2877 /* If we're debugging, just print out the top of the stack */
2878 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2881 /* Send this thread the StackOverflow exception */
2882 raiseAsync(tso, (StgClosure *)stackOverflow_closure);
2886 /* Try to double the current stack size. If that takes us over the
2887 * maximum stack size for this thread, then use the maximum instead.
2888 * Finally round up so the TSO ends up as a whole number of blocks.
2890 new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
2891 new_tso_size = (lnat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) +
2892 TSO_STRUCT_SIZE)/sizeof(W_);
2893 new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */
2894 new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
2896 IF_DEBUG(scheduler, debugBelch("== sched: increasing stack size from %d words to %d.\n", tso->stack_size, new_stack_size));
2898 dest = (StgTSO *)allocate(new_tso_size);
2899 TICK_ALLOC_TSO(new_stack_size,0);
2901 /* copy the TSO block and the old stack into the new area */
2902 memcpy(dest,tso,TSO_STRUCT_SIZE);
2903 stack_words = tso->stack + tso->stack_size - tso->sp;
2904 new_sp = (P_)dest + new_tso_size - stack_words;
2905 memcpy(new_sp, tso->sp, stack_words * sizeof(W_));
2907 /* relocate the stack pointers... */
2909 dest->stack_size = new_stack_size;
2911 /* Mark the old TSO as relocated. We have to check for relocated
2912 * TSOs in the garbage collector and any primops that deal with TSOs.
2914 * It's important to set the sp value to just beyond the end
2915 * of the stack, so we don't attempt to scavenge any part of the
2918 tso->what_next = ThreadRelocated;
2920 tso->sp = (P_)&(tso->stack[tso->stack_size]);
2921 tso->why_blocked = NotBlocked;
2923 IF_PAR_DEBUG(verbose,
2924 debugBelch("@@ threadStackOverflow of TSO %d (now at %p): stack size increased to %ld\n",
2925 tso->id, tso, tso->stack_size);
2926 /* If we're debugging, just print out the top of the stack */
2927 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2930 IF_DEBUG(sanity,checkTSO(tso));
2932 IF_DEBUG(scheduler,printTSO(dest));
2938 /* ---------------------------------------------------------------------------
2939 Wake up a queue that was blocked on some resource.
2940 ------------------------------------------------------------------------ */
2944 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2947 #elif defined(PARALLEL_HASKELL)
2949 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2951 /* write RESUME events to log file and
2952 update blocked and fetch time (depending on type of the orig closure) */
2953 if (RtsFlags.ParFlags.ParStats.Full) {
2954 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
2955 GR_RESUMEQ, ((StgTSO *)bqe), ((StgTSO *)bqe)->block_info.closure,
2956 0, 0 /* spark_queue_len(ADVISORY_POOL) */);
2957 if (EMPTY_RUN_QUEUE())
2958 emitSchedule = rtsTrue;
2960 switch (get_itbl(node)->type) {
2962 ((StgTSO *)bqe)->par.fetchtime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2967 ((StgTSO *)bqe)->par.blocktime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2974 barf("{unblockOneLocked}Daq Qagh: unexpected closure in blocking queue");
2981 static StgBlockingQueueElement *
2982 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2985 PEs node_loc, tso_loc;
2987 node_loc = where_is(node); // should be lifted out of loop
2988 tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2989 tso_loc = where_is((StgClosure *)tso);
2990 if (IS_LOCAL_TO(PROCS(node),tso_loc)) { // TSO is local
2991 /* !fake_fetch => TSO is on CurrentProc is same as IS_LOCAL_TO */
2992 ASSERT(CurrentProc!=node_loc || tso_loc==CurrentProc);
2993 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.lunblocktime;
2994 // insertThread(tso, node_loc);
2995 new_event(tso_loc, tso_loc, CurrentTime[CurrentProc],
2997 tso, node, (rtsSpark*)NULL);
2998 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
3001 } else { // TSO is remote (actually should be FMBQ)
3002 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.mpacktime +
3003 RtsFlags.GranFlags.Costs.gunblocktime +
3004 RtsFlags.GranFlags.Costs.latency;
3005 new_event(tso_loc, CurrentProc, CurrentTime[CurrentProc],
3007 tso, node, (rtsSpark*)NULL);
3008 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
3011 /* the thread-queue-overhead is accounted for in either Resume or UnblockThread */
3013 debugBelch(" %s TSO %d (%p) [PE %d] (block_info.closure=%p) (next=%p) ,",
3014 (node_loc==tso_loc ? "Local" : "Global"),
3015 tso->id, tso, CurrentProc, tso->block_info.closure, tso->link));
3016 tso->block_info.closure = NULL;
3017 IF_DEBUG(scheduler,debugBelch("-- Waking up thread %ld (%p)\n",
3020 #elif defined(PARALLEL_HASKELL)
3021 static StgBlockingQueueElement *
3022 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
3024 StgBlockingQueueElement *next;
3026 switch (get_itbl(bqe)->type) {
3028 ASSERT(((StgTSO *)bqe)->why_blocked != NotBlocked);
3029 /* if it's a TSO just push it onto the run_queue */
3031 ((StgTSO *)bqe)->link = END_TSO_QUEUE; // debugging?
3032 APPEND_TO_RUN_QUEUE((StgTSO *)bqe);
3034 unblockCount(bqe, node);
3035 /* reset blocking status after dumping event */
3036 ((StgTSO *)bqe)->why_blocked = NotBlocked;
3040 /* if it's a BLOCKED_FETCH put it on the PendingFetches list */
3042 bqe->link = (StgBlockingQueueElement *)PendingFetches;
3043 PendingFetches = (StgBlockedFetch *)bqe;
3047 /* can ignore this case in a non-debugging setup;
3048 see comments on RBHSave closures above */
3050 /* check that the closure is an RBHSave closure */
3051 ASSERT(get_itbl((StgClosure *)bqe) == &stg_RBH_Save_0_info ||
3052 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_1_info ||
3053 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_2_info);
3057 barf("{unblockOneLocked}Daq Qagh: Unexpected IP (%#lx; %s) in blocking queue at %#lx\n",
3058 get_itbl((StgClosure *)bqe), info_type((StgClosure *)bqe),
3062 IF_PAR_DEBUG(bq, debugBelch(", %p (%s)\n", bqe, info_type((StgClosure*)bqe)));
3066 #else /* !GRAN && !PARALLEL_HASKELL */
3068 unblockOneLocked(StgTSO *tso)
3072 ASSERT(get_itbl(tso)->type == TSO);
3073 ASSERT(tso->why_blocked != NotBlocked);
3074 tso->why_blocked = NotBlocked;
3076 tso->link = END_TSO_QUEUE;
3077 APPEND_TO_RUN_QUEUE(tso);
3079 IF_DEBUG(scheduler,sched_belch("waking up thread %ld", (long)tso->id));
3084 #if defined(GRAN) || defined(PARALLEL_HASKELL)
3085 INLINE_ME StgBlockingQueueElement *
3086 unblockOne(StgBlockingQueueElement *bqe, StgClosure *node)
3088 ACQUIRE_LOCK(&sched_mutex);
3089 bqe = unblockOneLocked(bqe, node);
3090 RELEASE_LOCK(&sched_mutex);
3095 unblockOne(StgTSO *tso)
3097 ACQUIRE_LOCK(&sched_mutex);
3098 tso = unblockOneLocked(tso);
3099 RELEASE_LOCK(&sched_mutex);
3106 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
3108 StgBlockingQueueElement *bqe;
3113 debugBelch("##-_ AwBQ for node %p on PE %d @ %ld by TSO %d (%p): \n", \
3114 node, CurrentProc, CurrentTime[CurrentProc],
3115 CurrentTSO->id, CurrentTSO));
3117 node_loc = where_is(node);
3119 ASSERT(q == END_BQ_QUEUE ||
3120 get_itbl(q)->type == TSO || // q is either a TSO or an RBHSave
3121 get_itbl(q)->type == CONSTR); // closure (type constructor)
3122 ASSERT(is_unique(node));
3124 /* FAKE FETCH: magically copy the node to the tso's proc;
3125 no Fetch necessary because in reality the node should not have been
3126 moved to the other PE in the first place
3128 if (CurrentProc!=node_loc) {
3130 debugBelch("## node %p is on PE %d but CurrentProc is %d (TSO %d); assuming fake fetch and adjusting bitmask (old: %#x)\n",
3131 node, node_loc, CurrentProc, CurrentTSO->id,
3132 // CurrentTSO, where_is(CurrentTSO),
3133 node->header.gran.procs));
3134 node->header.gran.procs = (node->header.gran.procs) | PE_NUMBER(CurrentProc);
3136 debugBelch("## new bitmask of node %p is %#x\n",
3137 node, node->header.gran.procs));
3138 if (RtsFlags.GranFlags.GranSimStats.Global) {
3139 globalGranStats.tot_fake_fetches++;
3144 // ToDo: check: ASSERT(CurrentProc==node_loc);
3145 while (get_itbl(bqe)->type==TSO) { // q != END_TSO_QUEUE) {
3148 bqe points to the current element in the queue
3149 next points to the next element in the queue
3151 //tso = (StgTSO *)bqe; // wastes an assignment to get the type right
3152 //tso_loc = where_is(tso);
3154 bqe = unblockOneLocked(bqe, node);
3157 /* if this is the BQ of an RBH, we have to put back the info ripped out of
3158 the closure to make room for the anchor of the BQ */
3159 if (bqe!=END_BQ_QUEUE) {
3160 ASSERT(get_itbl(node)->type == RBH && get_itbl(bqe)->type == CONSTR);
3162 ASSERT((info_ptr==&RBH_Save_0_info) ||
3163 (info_ptr==&RBH_Save_1_info) ||
3164 (info_ptr==&RBH_Save_2_info));
3166 /* cf. convertToRBH in RBH.c for writing the RBHSave closure */
3167 ((StgRBH *)node)->blocking_queue = (StgBlockingQueueElement *)((StgRBHSave *)bqe)->payload[0];
3168 ((StgRBH *)node)->mut_link = (StgMutClosure *)((StgRBHSave *)bqe)->payload[1];
3171 debugBelch("## Filled in RBH_Save for %p (%s) at end of AwBQ\n",
3172 node, info_type(node)));
3175 /* statistics gathering */
3176 if (RtsFlags.GranFlags.GranSimStats.Global) {
3177 // globalGranStats.tot_bq_processing_time += bq_processing_time;
3178 globalGranStats.tot_bq_len += len; // total length of all bqs awakened
3179 // globalGranStats.tot_bq_len_local += len_local; // same for local TSOs only
3180 globalGranStats.tot_awbq++; // total no. of bqs awakened
3183 debugBelch("## BQ Stats of %p: [%d entries] %s\n",
3184 node, len, (bqe!=END_BQ_QUEUE) ? "RBH" : ""));
3186 #elif defined(PARALLEL_HASKELL)
3188 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
3190 StgBlockingQueueElement *bqe;
3192 ACQUIRE_LOCK(&sched_mutex);
3194 IF_PAR_DEBUG(verbose,
3195 debugBelch("##-_ AwBQ for node %p on [%x]: \n",
3199 if(get_itbl(q)->type == CONSTR || q==END_BQ_QUEUE) {
3200 IF_PAR_DEBUG(verbose, debugBelch("## ... nothing to unblock so lets just return. RFP (BUG?)\n"));
3205 ASSERT(q == END_BQ_QUEUE ||
3206 get_itbl(q)->type == TSO ||
3207 get_itbl(q)->type == BLOCKED_FETCH ||
3208 get_itbl(q)->type == CONSTR);
3211 while (get_itbl(bqe)->type==TSO ||
3212 get_itbl(bqe)->type==BLOCKED_FETCH) {
3213 bqe = unblockOneLocked(bqe, node);
3215 RELEASE_LOCK(&sched_mutex);
3218 #else /* !GRAN && !PARALLEL_HASKELL */
3221 awakenBlockedQueueNoLock(StgTSO *tso)
3223 while (tso != END_TSO_QUEUE) {
3224 tso = unblockOneLocked(tso);
3229 awakenBlockedQueue(StgTSO *tso)
3231 ACQUIRE_LOCK(&sched_mutex);
3232 while (tso != END_TSO_QUEUE) {
3233 tso = unblockOneLocked(tso);
3235 RELEASE_LOCK(&sched_mutex);
3239 /* ---------------------------------------------------------------------------
3241 - usually called inside a signal handler so it mustn't do anything fancy.
3242 ------------------------------------------------------------------------ */
3245 interruptStgRts(void)
3250 /* ToDo: if invoked from a signal handler, this threadRunnable
3251 * only works if there's another thread (not this one) waiting to
3256 /* -----------------------------------------------------------------------------
3259 This is for use when we raise an exception in another thread, which
3261 This has nothing to do with the UnblockThread event in GranSim. -- HWL
3262 -------------------------------------------------------------------------- */
3264 #if defined(GRAN) || defined(PARALLEL_HASKELL)
3266 NB: only the type of the blocking queue is different in GranSim and GUM
3267 the operations on the queue-elements are the same
3268 long live polymorphism!
3270 Locks: sched_mutex is held upon entry and exit.
3274 unblockThread(StgTSO *tso)
3276 StgBlockingQueueElement *t, **last;
3278 switch (tso->why_blocked) {
3281 return; /* not blocked */
3284 // Be careful: nothing to do here! We tell the scheduler that the thread
3285 // is runnable and we leave it to the stack-walking code to abort the
3286 // transaction while unwinding the stack. We should perhaps have a debugging
3287 // test to make sure that this really happens and that the 'zombie' transaction
3288 // does not get committed.
3292 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
3294 StgBlockingQueueElement *last_tso = END_BQ_QUEUE;
3295 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
3297 last = (StgBlockingQueueElement **)&mvar->head;
3298 for (t = (StgBlockingQueueElement *)mvar->head;
3300 last = &t->link, last_tso = t, t = t->link) {
3301 if (t == (StgBlockingQueueElement *)tso) {
3302 *last = (StgBlockingQueueElement *)tso->link;
3303 if (mvar->tail == tso) {
3304 mvar->tail = (StgTSO *)last_tso;
3309 barf("unblockThread (MVAR): TSO not found");
3312 case BlockedOnBlackHole:
3313 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
3315 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
3317 last = &bq->blocking_queue;
3318 for (t = bq->blocking_queue;
3320 last = &t->link, t = t->link) {
3321 if (t == (StgBlockingQueueElement *)tso) {
3322 *last = (StgBlockingQueueElement *)tso->link;
3326 barf("unblockThread (BLACKHOLE): TSO not found");
3329 case BlockedOnException:
3331 StgTSO *target = tso->block_info.tso;
3333 ASSERT(get_itbl(target)->type == TSO);
3335 if (target->what_next == ThreadRelocated) {
3336 target = target->link;
3337 ASSERT(get_itbl(target)->type == TSO);
3340 ASSERT(target->blocked_exceptions != NULL);
3342 last = (StgBlockingQueueElement **)&target->blocked_exceptions;
3343 for (t = (StgBlockingQueueElement *)target->blocked_exceptions;
3345 last = &t->link, t = t->link) {
3346 ASSERT(get_itbl(t)->type == TSO);
3347 if (t == (StgBlockingQueueElement *)tso) {
3348 *last = (StgBlockingQueueElement *)tso->link;
3352 barf("unblockThread (Exception): TSO not found");
3356 case BlockedOnWrite:
3357 #if defined(mingw32_HOST_OS)
3358 case BlockedOnDoProc:
3361 /* take TSO off blocked_queue */
3362 StgBlockingQueueElement *prev = NULL;
3363 for (t = (StgBlockingQueueElement *)blocked_queue_hd; t != END_BQ_QUEUE;
3364 prev = t, t = t->link) {
3365 if (t == (StgBlockingQueueElement *)tso) {
3367 blocked_queue_hd = (StgTSO *)t->link;
3368 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
3369 blocked_queue_tl = END_TSO_QUEUE;
3372 prev->link = t->link;
3373 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
3374 blocked_queue_tl = (StgTSO *)prev;
3377 #if defined(mingw32_HOST_OS)
3378 /* (Cooperatively) signal that the worker thread should abort
3381 abandonWorkRequest(tso->block_info.async_result->reqID);
3386 barf("unblockThread (I/O): TSO not found");
3389 case BlockedOnDelay:
3391 /* take TSO off sleeping_queue */
3392 StgBlockingQueueElement *prev = NULL;
3393 for (t = (StgBlockingQueueElement *)sleeping_queue; t != END_BQ_QUEUE;
3394 prev = t, t = t->link) {
3395 if (t == (StgBlockingQueueElement *)tso) {
3397 sleeping_queue = (StgTSO *)t->link;
3399 prev->link = t->link;
3404 barf("unblockThread (delay): TSO not found");
3408 barf("unblockThread");
3412 tso->link = END_TSO_QUEUE;
3413 tso->why_blocked = NotBlocked;
3414 tso->block_info.closure = NULL;
3415 PUSH_ON_RUN_QUEUE(tso);
3419 unblockThread(StgTSO *tso)
3423 /* To avoid locking unnecessarily. */
3424 if (tso->why_blocked == NotBlocked) {
3428 switch (tso->why_blocked) {
3431 // Be careful: nothing to do here! We tell the scheduler that the thread
3432 // is runnable and we leave it to the stack-walking code to abort the
3433 // transaction while unwinding the stack. We should perhaps have a debugging
3434 // test to make sure that this really happens and that the 'zombie' transaction
3435 // does not get committed.
3439 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
3441 StgTSO *last_tso = END_TSO_QUEUE;
3442 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
3445 for (t = mvar->head; t != END_TSO_QUEUE;
3446 last = &t->link, last_tso = t, t = t->link) {
3449 if (mvar->tail == tso) {
3450 mvar->tail = last_tso;
3455 barf("unblockThread (MVAR): TSO not found");
3458 case BlockedOnBlackHole:
3460 last = &blackhole_queue;
3461 for (t = blackhole_queue; t != END_TSO_QUEUE;
3462 last = &t->link, t = t->link) {
3468 barf("unblockThread (BLACKHOLE): TSO not found");
3471 case BlockedOnException:
3473 StgTSO *target = tso->block_info.tso;
3475 ASSERT(get_itbl(target)->type == TSO);
3477 while (target->what_next == ThreadRelocated) {
3478 target = target->link;
3479 ASSERT(get_itbl(target)->type == TSO);
3482 ASSERT(target->blocked_exceptions != NULL);
3484 last = &target->blocked_exceptions;
3485 for (t = target->blocked_exceptions; t != END_TSO_QUEUE;
3486 last = &t->link, t = t->link) {
3487 ASSERT(get_itbl(t)->type == TSO);
3493 barf("unblockThread (Exception): TSO not found");
3497 case BlockedOnWrite:
3498 #if defined(mingw32_HOST_OS)
3499 case BlockedOnDoProc:
3502 StgTSO *prev = NULL;
3503 for (t = blocked_queue_hd; t != END_TSO_QUEUE;
3504 prev = t, t = t->link) {
3507 blocked_queue_hd = t->link;
3508 if (blocked_queue_tl == t) {
3509 blocked_queue_tl = END_TSO_QUEUE;
3512 prev->link = t->link;
3513 if (blocked_queue_tl == t) {
3514 blocked_queue_tl = prev;
3517 #if defined(mingw32_HOST_OS)
3518 /* (Cooperatively) signal that the worker thread should abort
3521 abandonWorkRequest(tso->block_info.async_result->reqID);
3526 barf("unblockThread (I/O): TSO not found");
3529 case BlockedOnDelay:
3531 StgTSO *prev = NULL;
3532 for (t = sleeping_queue; t != END_TSO_QUEUE;
3533 prev = t, t = t->link) {
3536 sleeping_queue = t->link;
3538 prev->link = t->link;
3543 barf("unblockThread (delay): TSO not found");
3547 barf("unblockThread");
3551 tso->link = END_TSO_QUEUE;
3552 tso->why_blocked = NotBlocked;
3553 tso->block_info.closure = NULL;
3554 APPEND_TO_RUN_QUEUE(tso);
3558 /* -----------------------------------------------------------------------------
3561 * Check the blackhole_queue for threads that can be woken up. We do
3562 * this periodically: before every GC, and whenever the run queue is
3565 * An elegant solution might be to just wake up all the blocked
3566 * threads with awakenBlockedQueue occasionally: they'll go back to
3567 * sleep again if the object is still a BLACKHOLE. Unfortunately this
3568 * doesn't give us a way to tell whether we've actually managed to
3569 * wake up any threads, so we would be busy-waiting.
3571 * -------------------------------------------------------------------------- */
3574 checkBlackHoles( void )
3577 rtsBool any_woke_up = rtsFalse;
3580 IF_DEBUG(scheduler, sched_belch("checking threads blocked on black holes"));
3582 // ASSUMES: sched_mutex
3583 prev = &blackhole_queue;
3584 t = blackhole_queue;
3585 while (t != END_TSO_QUEUE) {
3586 ASSERT(t->why_blocked == BlockedOnBlackHole);
3587 type = get_itbl(t->block_info.closure)->type;
3588 if (type != BLACKHOLE && type != CAF_BLACKHOLE) {
3589 t = unblockOneLocked(t);
3591 any_woke_up = rtsTrue;
3601 /* -----------------------------------------------------------------------------
3604 * The following function implements the magic for raising an
3605 * asynchronous exception in an existing thread.
3607 * We first remove the thread from any queue on which it might be
3608 * blocked. The possible blockages are MVARs and BLACKHOLE_BQs.
3610 * We strip the stack down to the innermost CATCH_FRAME, building
3611 * thunks in the heap for all the active computations, so they can
3612 * be restarted if necessary. When we reach a CATCH_FRAME, we build
3613 * an application of the handler to the exception, and push it on
3614 * the top of the stack.
3616 * How exactly do we save all the active computations? We create an
3617 * AP_STACK for every UpdateFrame on the stack. Entering one of these
3618 * AP_STACKs pushes everything from the corresponding update frame
3619 * upwards onto the stack. (Actually, it pushes everything up to the
3620 * next update frame plus a pointer to the next AP_STACK object.
3621 * Entering the next AP_STACK object pushes more onto the stack until we
3622 * reach the last AP_STACK object - at which point the stack should look
3623 * exactly as it did when we killed the TSO and we can continue
3624 * execution by entering the closure on top of the stack.
3626 * We can also kill a thread entirely - this happens if either (a) the
3627 * exception passed to raiseAsync is NULL, or (b) there's no
3628 * CATCH_FRAME on the stack. In either case, we strip the entire
3629 * stack and replace the thread with a zombie.
3631 * Locks: sched_mutex held upon entry nor exit.
3633 * -------------------------------------------------------------------------- */
3636 deleteThread(StgTSO *tso)
3638 if (tso->why_blocked != BlockedOnCCall &&
3639 tso->why_blocked != BlockedOnCCall_NoUnblockExc) {
3640 raiseAsync(tso,NULL);
3644 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
3646 deleteThreadImmediately(StgTSO *tso)
3647 { // for forkProcess only:
3648 // delete thread without giving it a chance to catch the KillThread exception
3650 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
3654 if (tso->why_blocked != BlockedOnCCall &&
3655 tso->why_blocked != BlockedOnCCall_NoUnblockExc) {
3659 tso->what_next = ThreadKilled;
3664 raiseAsyncWithLock(StgTSO *tso, StgClosure *exception)
3666 /* When raising async exs from contexts where sched_mutex isn't held;
3667 use raiseAsyncWithLock(). */
3668 ACQUIRE_LOCK(&sched_mutex);
3669 raiseAsync(tso,exception);
3670 RELEASE_LOCK(&sched_mutex);
3674 raiseAsync(StgTSO *tso, StgClosure *exception)
3676 raiseAsync_(tso, exception, rtsFalse);
3680 raiseAsync_(StgTSO *tso, StgClosure *exception, rtsBool stop_at_atomically)
3682 StgRetInfoTable *info;
3685 // Thread already dead?
3686 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
3691 sched_belch("raising exception in thread %ld.", (long)tso->id));
3693 // Remove it from any blocking queues
3698 // The stack freezing code assumes there's a closure pointer on
3699 // the top of the stack, so we have to arrange that this is the case...
3701 if (sp[0] == (W_)&stg_enter_info) {
3705 sp[0] = (W_)&stg_dummy_ret_closure;
3711 // 1. Let the top of the stack be the "current closure"
3713 // 2. Walk up the stack until we find either an UPDATE_FRAME or a
3716 // 3. If it's an UPDATE_FRAME, then make an AP_STACK containing the
3717 // current closure applied to the chunk of stack up to (but not
3718 // including) the update frame. This closure becomes the "current
3719 // closure". Go back to step 2.
3721 // 4. If it's a CATCH_FRAME, then leave the exception handler on
3722 // top of the stack applied to the exception.
3724 // 5. If it's a STOP_FRAME, then kill the thread.
3726 // NB: if we pass an ATOMICALLY_FRAME then abort the associated
3733 info = get_ret_itbl((StgClosure *)frame);
3735 while (info->i.type != UPDATE_FRAME
3736 && (info->i.type != CATCH_FRAME || exception == NULL)
3737 && info->i.type != STOP_FRAME
3738 && (info->i.type != ATOMICALLY_FRAME || stop_at_atomically == rtsFalse))
3740 if (info->i.type == CATCH_RETRY_FRAME || info->i.type == ATOMICALLY_FRAME) {
3741 // IF we find an ATOMICALLY_FRAME then we abort the
3742 // current transaction and propagate the exception. In
3743 // this case (unlike ordinary exceptions) we do not care
3744 // whether the transaction is valid or not because its
3745 // possible validity cannot have caused the exception
3746 // and will not be visible after the abort.
3748 debugBelch("Found atomically block delivering async exception\n"));
3749 stmAbortTransaction(tso -> trec);
3750 tso -> trec = stmGetEnclosingTRec(tso -> trec);
3752 frame += stack_frame_sizeW((StgClosure *)frame);
3753 info = get_ret_itbl((StgClosure *)frame);
3756 switch (info->i.type) {
3758 case ATOMICALLY_FRAME:
3759 ASSERT(stop_at_atomically);
3760 ASSERT(stmGetEnclosingTRec(tso->trec) == NO_TREC);
3761 stmCondemnTransaction(tso -> trec);
3765 // R1 is not a register: the return convention for IO in
3766 // this case puts the return value on the stack, so we
3767 // need to set up the stack to return to the atomically
3768 // frame properly...
3769 tso->sp = frame - 2;
3770 tso->sp[1] = (StgWord) &stg_NO_FINALIZER_closure; // why not?
3771 tso->sp[0] = (StgWord) &stg_ut_1_0_unreg_info;
3773 tso->what_next = ThreadRunGHC;
3777 // If we find a CATCH_FRAME, and we've got an exception to raise,
3778 // then build the THUNK raise(exception), and leave it on
3779 // top of the CATCH_FRAME ready to enter.
3783 StgCatchFrame *cf = (StgCatchFrame *)frame;
3787 // we've got an exception to raise, so let's pass it to the
3788 // handler in this frame.
3790 raise = (StgThunk *)allocate(sizeofW(StgThunk)+1);
3791 TICK_ALLOC_SE_THK(1,0);
3792 SET_HDR(raise,&stg_raise_info,cf->header.prof.ccs);
3793 raise->payload[0] = exception;
3795 // throw away the stack from Sp up to the CATCH_FRAME.
3799 /* Ensure that async excpetions are blocked now, so we don't get
3800 * a surprise exception before we get around to executing the
3803 if (tso->blocked_exceptions == NULL) {
3804 tso->blocked_exceptions = END_TSO_QUEUE;
3807 /* Put the newly-built THUNK on top of the stack, ready to execute
3808 * when the thread restarts.
3811 sp[-1] = (W_)&stg_enter_info;
3813 tso->what_next = ThreadRunGHC;
3814 IF_DEBUG(sanity, checkTSO(tso));
3823 // First build an AP_STACK consisting of the stack chunk above the
3824 // current update frame, with the top word on the stack as the
3827 words = frame - sp - 1;
3828 ap = (StgAP_STACK *)allocate(AP_STACK_sizeW(words));
3831 ap->fun = (StgClosure *)sp[0];
3833 for(i=0; i < (nat)words; ++i) {
3834 ap->payload[i] = (StgClosure *)*sp++;
3837 SET_HDR(ap,&stg_AP_STACK_info,
3838 ((StgClosure *)frame)->header.prof.ccs /* ToDo */);
3839 TICK_ALLOC_UP_THK(words+1,0);
3842 debugBelch("sched: Updating ");
3843 printPtr((P_)((StgUpdateFrame *)frame)->updatee);
3844 debugBelch(" with ");
3845 printObj((StgClosure *)ap);
3848 // Replace the updatee with an indirection - happily
3849 // this will also wake up any threads currently
3850 // waiting on the result.
3852 // Warning: if we're in a loop, more than one update frame on
3853 // the stack may point to the same object. Be careful not to
3854 // overwrite an IND_OLDGEN in this case, because we'll screw
3855 // up the mutable lists. To be on the safe side, don't
3856 // overwrite any kind of indirection at all. See also
3857 // threadSqueezeStack in GC.c, where we have to make a similar
3860 if (!closure_IND(((StgUpdateFrame *)frame)->updatee)) {
3861 // revert the black hole
3862 UPD_IND_NOLOCK(((StgUpdateFrame *)frame)->updatee,
3865 sp += sizeofW(StgUpdateFrame) - 1;
3866 sp[0] = (W_)ap; // push onto stack
3871 // We've stripped the entire stack, the thread is now dead.
3872 sp += sizeofW(StgStopFrame);
3873 tso->what_next = ThreadKilled;
3884 /* -----------------------------------------------------------------------------
3885 raiseExceptionHelper
3887 This function is called by the raise# primitve, just so that we can
3888 move some of the tricky bits of raising an exception from C-- into
3889 C. Who knows, it might be a useful re-useable thing here too.
3890 -------------------------------------------------------------------------- */
3893 raiseExceptionHelper (StgTSO *tso, StgClosure *exception)
3895 StgThunk *raise_closure = NULL;
3897 StgRetInfoTable *info;
3899 // This closure represents the expression 'raise# E' where E
3900 // is the exception raise. It is used to overwrite all the
3901 // thunks which are currently under evaluataion.
3905 // LDV profiling: stg_raise_info has THUNK as its closure
3906 // type. Since a THUNK takes at least MIN_UPD_SIZE words in its
3907 // payload, MIN_UPD_SIZE is more approprate than 1. It seems that
3908 // 1 does not cause any problem unless profiling is performed.
3909 // However, when LDV profiling goes on, we need to linearly scan
3910 // small object pool, where raise_closure is stored, so we should
3911 // use MIN_UPD_SIZE.
3913 // raise_closure = (StgClosure *)RET_STGCALL1(P_,allocate,
3914 // sizeofW(StgClosure)+1);
3918 // Walk up the stack, looking for the catch frame. On the way,
3919 // we update any closures pointed to from update frames with the
3920 // raise closure that we just built.
3924 info = get_ret_itbl((StgClosure *)p);
3925 next = p + stack_frame_sizeW((StgClosure *)p);
3926 switch (info->i.type) {
3929 // Only create raise_closure if we need to.
3930 if (raise_closure == NULL) {
3932 (StgThunk *)allocate(sizeofW(StgThunk)+MIN_UPD_SIZE);
3933 SET_HDR(raise_closure, &stg_raise_info, CCCS);
3934 raise_closure->payload[0] = exception;
3936 UPD_IND(((StgUpdateFrame *)p)->updatee,(StgClosure *)raise_closure);
3940 case ATOMICALLY_FRAME:
3941 IF_DEBUG(stm, debugBelch("Found ATOMICALLY_FRAME at %p\n", p));
3943 return ATOMICALLY_FRAME;
3949 case CATCH_STM_FRAME:
3950 IF_DEBUG(stm, debugBelch("Found CATCH_STM_FRAME at %p\n", p));
3952 return CATCH_STM_FRAME;
3958 case CATCH_RETRY_FRAME:
3967 /* -----------------------------------------------------------------------------
3968 findRetryFrameHelper
3970 This function is called by the retry# primitive. It traverses the stack
3971 leaving tso->sp referring to the frame which should handle the retry.
3973 This should either be a CATCH_RETRY_FRAME (if the retry# is within an orElse#)
3974 or should be a ATOMICALLY_FRAME (if the retry# reaches the top level).
3976 We skip CATCH_STM_FRAMEs because retries are not considered to be exceptions,
3977 despite the similar implementation.
3979 We should not expect to see CATCH_FRAME or STOP_FRAME because those should
3980 not be created within memory transactions.
3981 -------------------------------------------------------------------------- */
3984 findRetryFrameHelper (StgTSO *tso)
3987 StgRetInfoTable *info;
3991 info = get_ret_itbl((StgClosure *)p);
3992 next = p + stack_frame_sizeW((StgClosure *)p);
3993 switch (info->i.type) {
3995 case ATOMICALLY_FRAME:
3996 IF_DEBUG(stm, debugBelch("Found ATOMICALLY_FRAME at %p during retrry\n", p));
3998 return ATOMICALLY_FRAME;
4000 case CATCH_RETRY_FRAME:
4001 IF_DEBUG(stm, debugBelch("Found CATCH_RETRY_FRAME at %p during retrry\n", p));
4003 return CATCH_RETRY_FRAME;
4005 case CATCH_STM_FRAME:
4007 ASSERT(info->i.type != CATCH_FRAME);
4008 ASSERT(info->i.type != STOP_FRAME);
4015 /* -----------------------------------------------------------------------------
4016 resurrectThreads is called after garbage collection on the list of
4017 threads found to be garbage. Each of these threads will be woken
4018 up and sent a signal: BlockedOnDeadMVar if the thread was blocked
4019 on an MVar, or NonTermination if the thread was blocked on a Black
4022 Locks: sched_mutex isn't held upon entry nor exit.
4023 -------------------------------------------------------------------------- */
4026 resurrectThreads( StgTSO *threads )
4030 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
4031 next = tso->global_link;
4032 tso->global_link = all_threads;
4034 IF_DEBUG(scheduler, sched_belch("resurrecting thread %d", tso->id));
4036 switch (tso->why_blocked) {
4038 case BlockedOnException:
4039 /* Called by GC - sched_mutex lock is currently held. */
4040 raiseAsync(tso,(StgClosure *)BlockedOnDeadMVar_closure);
4042 case BlockedOnBlackHole:
4043 raiseAsync(tso,(StgClosure *)NonTermination_closure);
4046 raiseAsync(tso,(StgClosure *)BlockedIndefinitely_closure);
4049 /* This might happen if the thread was blocked on a black hole
4050 * belonging to a thread that we've just woken up (raiseAsync
4051 * can wake up threads, remember...).
4055 barf("resurrectThreads: thread blocked in a strange way");
4060 /* ----------------------------------------------------------------------------
4061 * Debugging: why is a thread blocked
4062 * [Also provides useful information when debugging threaded programs
4063 * at the Haskell source code level, so enable outside of DEBUG. --sof 7/02]
4064 ------------------------------------------------------------------------- */
4067 printThreadBlockage(StgTSO *tso)
4069 switch (tso->why_blocked) {
4071 debugBelch("is blocked on read from fd %ld", tso->block_info.fd);
4073 case BlockedOnWrite:
4074 debugBelch("is blocked on write to fd %ld", tso->block_info.fd);
4076 #if defined(mingw32_HOST_OS)
4077 case BlockedOnDoProc:
4078 debugBelch("is blocked on proc (request: %ld)", tso->block_info.async_result->reqID);
4081 case BlockedOnDelay:
4082 debugBelch("is blocked until %ld", tso->block_info.target);
4085 debugBelch("is blocked on an MVar");
4087 case BlockedOnException:
4088 debugBelch("is blocked on delivering an exception to thread %d",
4089 tso->block_info.tso->id);
4091 case BlockedOnBlackHole:
4092 debugBelch("is blocked on a black hole");
4095 debugBelch("is not blocked");
4097 #if defined(PARALLEL_HASKELL)
4099 debugBelch("is blocked on global address; local FM_BQ is %p (%s)",
4100 tso->block_info.closure, info_type(tso->block_info.closure));
4102 case BlockedOnGA_NoSend:
4103 debugBelch("is blocked on global address (no send); local FM_BQ is %p (%s)",
4104 tso->block_info.closure, info_type(tso->block_info.closure));
4107 case BlockedOnCCall:
4108 debugBelch("is blocked on an external call");
4110 case BlockedOnCCall_NoUnblockExc:
4111 debugBelch("is blocked on an external call (exceptions were already blocked)");
4114 debugBelch("is blocked on an STM operation");
4117 barf("printThreadBlockage: strange tso->why_blocked: %d for TSO %d (%d)",
4118 tso->why_blocked, tso->id, tso);
4123 printThreadStatus(StgTSO *tso)
4125 switch (tso->what_next) {
4127 debugBelch("has been killed");
4129 case ThreadComplete:
4130 debugBelch("has completed");
4133 printThreadBlockage(tso);
4138 printAllThreads(void)
4143 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
4144 ullong_format_string(TIME_ON_PROC(CurrentProc),
4145 time_string, rtsFalse/*no commas!*/);
4147 debugBelch("all threads at [%s]:\n", time_string);
4148 # elif defined(PARALLEL_HASKELL)
4149 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
4150 ullong_format_string(CURRENT_TIME,
4151 time_string, rtsFalse/*no commas!*/);
4153 debugBelch("all threads at [%s]:\n", time_string);
4155 debugBelch("all threads:\n");
4158 for (t = all_threads; t != END_TSO_QUEUE; ) {
4159 debugBelch("\tthread %d @ %p ", t->id, (void *)t);
4162 void *label = lookupThreadLabel(t->id);
4163 if (label) debugBelch("[\"%s\"] ",(char *)label);
4166 if (t->what_next == ThreadRelocated) {
4167 debugBelch("has been relocated...\n");
4170 printThreadStatus(t);
4180 Print a whole blocking queue attached to node (debugging only).
4182 # if defined(PARALLEL_HASKELL)
4184 print_bq (StgClosure *node)
4186 StgBlockingQueueElement *bqe;
4190 debugBelch("## BQ of closure %p (%s): ",
4191 node, info_type(node));
4193 /* should cover all closures that may have a blocking queue */
4194 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
4195 get_itbl(node)->type == FETCH_ME_BQ ||
4196 get_itbl(node)->type == RBH ||
4197 get_itbl(node)->type == MVAR);
4199 ASSERT(node!=(StgClosure*)NULL); // sanity check
4201 print_bqe(((StgBlockingQueue*)node)->blocking_queue);
4205 Print a whole blocking queue starting with the element bqe.
4208 print_bqe (StgBlockingQueueElement *bqe)
4213 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
4215 for (end = (bqe==END_BQ_QUEUE);
4216 !end; // iterate until bqe points to a CONSTR
4217 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE),
4218 bqe = end ? END_BQ_QUEUE : bqe->link) {
4219 ASSERT(bqe != END_BQ_QUEUE); // sanity check
4220 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
4221 /* types of closures that may appear in a blocking queue */
4222 ASSERT(get_itbl(bqe)->type == TSO ||
4223 get_itbl(bqe)->type == BLOCKED_FETCH ||
4224 get_itbl(bqe)->type == CONSTR);
4225 /* only BQs of an RBH end with an RBH_Save closure */
4226 //ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
4228 switch (get_itbl(bqe)->type) {
4230 debugBelch(" TSO %u (%x),",
4231 ((StgTSO *)bqe)->id, ((StgTSO *)bqe));
4234 debugBelch(" BF (node=%p, ga=((%x, %d, %x)),",
4235 ((StgBlockedFetch *)bqe)->node,
4236 ((StgBlockedFetch *)bqe)->ga.payload.gc.gtid,
4237 ((StgBlockedFetch *)bqe)->ga.payload.gc.slot,
4238 ((StgBlockedFetch *)bqe)->ga.weight);
4241 debugBelch(" %s (IP %p),",
4242 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
4243 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
4244 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
4245 "RBH_Save_?"), get_itbl(bqe));
4248 barf("Unexpected closure type %s in blocking queue", // of %p (%s)",
4249 info_type((StgClosure *)bqe)); // , node, info_type(node));
4255 # elif defined(GRAN)
4257 print_bq (StgClosure *node)
4259 StgBlockingQueueElement *bqe;
4260 PEs node_loc, tso_loc;
4263 /* should cover all closures that may have a blocking queue */
4264 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
4265 get_itbl(node)->type == FETCH_ME_BQ ||
4266 get_itbl(node)->type == RBH);
4268 ASSERT(node!=(StgClosure*)NULL); // sanity check
4269 node_loc = where_is(node);
4271 debugBelch("## BQ of closure %p (%s) on [PE %d]: ",
4272 node, info_type(node), node_loc);
4275 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
4277 for (bqe = ((StgBlockingQueue*)node)->blocking_queue, end = (bqe==END_BQ_QUEUE);
4278 !end; // iterate until bqe points to a CONSTR
4279 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) {
4280 ASSERT(bqe != END_BQ_QUEUE); // sanity check
4281 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
4282 /* types of closures that may appear in a blocking queue */
4283 ASSERT(get_itbl(bqe)->type == TSO ||
4284 get_itbl(bqe)->type == CONSTR);
4285 /* only BQs of an RBH end with an RBH_Save closure */
4286 ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
4288 tso_loc = where_is((StgClosure *)bqe);
4289 switch (get_itbl(bqe)->type) {
4291 debugBelch(" TSO %d (%p) on [PE %d],",
4292 ((StgTSO *)bqe)->id, (StgTSO *)bqe, tso_loc);
4295 debugBelch(" %s (IP %p),",
4296 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
4297 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
4298 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
4299 "RBH_Save_?"), get_itbl(bqe));
4302 barf("Unexpected closure type %s in blocking queue of %p (%s)",
4303 info_type((StgClosure *)bqe), node, info_type(node));
4311 #if defined(PARALLEL_HASKELL)
4318 for (i=0, tso=run_queue_hd;
4319 tso != END_TSO_QUEUE;
4328 sched_belch(char *s, ...)
4332 #ifdef RTS_SUPPORTS_THREADS
4333 debugBelch("sched (task %p): ", osThreadId());
4334 #elif defined(PARALLEL_HASKELL)
4337 debugBelch("sched: ");