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");
706 // in SMP mode, we might return with a different capability than
707 // we started with, if the Haskell thread made a foreign call. So
708 // let's find out what our current Capability is:
709 cap = myCapability();
712 // We have run some Haskell code: there might be blackhole-blocked
713 // threads to wake up now.
714 if ( blackhole_queue != END_TSO_QUEUE ) {
715 blackholes_need_checking = rtsTrue;
718 cap->r.rInHaskell = rtsFalse;
720 // The TSO might have moved, eg. if it re-entered the RTS and a GC
721 // happened. So find the new location:
722 t = cap->r.rCurrentTSO;
724 // And save the current errno in this thread.
725 t->saved_errno = errno;
727 // ----------------------------------------------------------------------
729 /* Costs for the scheduler are assigned to CCS_SYSTEM */
730 #if defined(PROFILING)
735 ACQUIRE_LOCK(&sched_mutex);
737 #if defined(RTS_SUPPORTS_THREADS)
738 IF_DEBUG(scheduler,debugBelch("sched (task %p): ", osThreadId()););
739 #elif !defined(GRAN) && !defined(PARALLEL_HASKELL)
740 IF_DEBUG(scheduler,debugBelch("sched: "););
743 schedulePostRunThread();
745 ready_to_gc = rtsFalse;
749 ready_to_gc = scheduleHandleHeapOverflow(cap,t);
753 scheduleHandleStackOverflow(t);
757 if (scheduleHandleYield(t, prev_what_next)) {
758 // shortcut for switching between compiler/interpreter:
764 scheduleHandleThreadBlocked(t);
769 if (scheduleHandleThreadFinished(mainThread, cap, t)) return;;
773 barf("schedule: invalid thread return code %d", (int)ret);
776 if (scheduleDoHeapProfile(ready_to_gc)) { ready_to_gc = rtsFalse; }
777 if (ready_to_gc) { scheduleDoGC(cap); }
778 } /* end of while() */
780 IF_PAR_DEBUG(verbose,
781 debugBelch("== Leaving schedule() after having received Finish\n"));
784 /* ----------------------------------------------------------------------------
785 * Setting up the scheduler loop
786 * ASSUMES: sched_mutex
787 * ------------------------------------------------------------------------- */
790 schedulePreLoop(void)
793 /* set up first event to get things going */
794 /* ToDo: assign costs for system setup and init MainTSO ! */
795 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
797 CurrentTSO, (StgClosure*)NULL, (rtsSpark*)NULL);
800 debugBelch("GRAN: Init CurrentTSO (in schedule) = %p\n",
802 G_TSO(CurrentTSO, 5));
804 if (RtsFlags.GranFlags.Light) {
805 /* Save current time; GranSim Light only */
806 CurrentTSO->gran.clock = CurrentTime[CurrentProc];
811 /* ----------------------------------------------------------------------------
812 * Start any pending signal handlers
813 * ASSUMES: sched_mutex
814 * ------------------------------------------------------------------------- */
817 scheduleStartSignalHandlers(void)
819 #if defined(RTS_USER_SIGNALS) && !defined(RTS_SUPPORTS_THREADS)
820 if (signals_pending()) {
821 RELEASE_LOCK(&sched_mutex); /* ToDo: kill */
822 startSignalHandlers();
823 ACQUIRE_LOCK(&sched_mutex);
828 /* ----------------------------------------------------------------------------
829 * Check for blocked threads that can be woken up.
830 * ASSUMES: sched_mutex
831 * ------------------------------------------------------------------------- */
834 scheduleCheckBlockedThreads(void)
837 // Check whether any waiting threads need to be woken up. If the
838 // run queue is empty, and there are no other tasks running, we
839 // can wait indefinitely for something to happen.
841 if ( !EMPTY_QUEUE(blocked_queue_hd) || !EMPTY_QUEUE(sleeping_queue) )
843 #if defined(RTS_SUPPORTS_THREADS)
844 // We shouldn't be here...
845 barf("schedule: awaitEvent() in threaded RTS");
847 awaitEvent( EMPTY_RUN_QUEUE() && !blackholes_need_checking );
852 /* ----------------------------------------------------------------------------
853 * Check for threads blocked on BLACKHOLEs that can be woken up
854 * ASSUMES: sched_mutex
855 * ------------------------------------------------------------------------- */
857 scheduleCheckBlackHoles( void )
859 if ( blackholes_need_checking )
862 blackholes_need_checking = rtsFalse;
866 /* ----------------------------------------------------------------------------
867 * Detect deadlock conditions and attempt to resolve them.
868 * ASSUMES: sched_mutex
869 * ------------------------------------------------------------------------- */
872 scheduleDetectDeadlock(void)
875 #if defined(PARALLEL_HASKELL)
876 // ToDo: add deadlock detection in GUM (similar to SMP) -- HWL
881 * Detect deadlock: when we have no threads to run, there are no
882 * threads blocked, waiting for I/O, or sleeping, and all the
883 * other tasks are waiting for work, we must have a deadlock of
886 if ( EMPTY_THREAD_QUEUES() )
888 #if defined(RTS_SUPPORTS_THREADS)
890 * In the threaded RTS, we only check for deadlock if there
891 * has been no activity in a complete timeslice. This means
892 * we won't eagerly start a full GC just because we don't have
893 * any threads to run currently.
895 if (recent_activity != ACTIVITY_INACTIVE) return;
898 IF_DEBUG(scheduler, sched_belch("deadlocked, forcing major GC..."));
900 // Garbage collection can release some new threads due to
901 // either (a) finalizers or (b) threads resurrected because
902 // they are unreachable and will therefore be sent an
903 // exception. Any threads thus released will be immediately
905 GarbageCollect(GetRoots,rtsTrue);
906 recent_activity = ACTIVITY_DONE_GC;
907 if ( !EMPTY_RUN_QUEUE() ) return;
909 #if defined(RTS_USER_SIGNALS) && !defined(RTS_SUPPORTS_THREADS)
910 /* If we have user-installed signal handlers, then wait
911 * for signals to arrive rather then bombing out with a
914 if ( anyUserHandlers() ) {
916 sched_belch("still deadlocked, waiting for signals..."));
920 if (signals_pending()) {
921 RELEASE_LOCK(&sched_mutex);
922 startSignalHandlers();
923 ACQUIRE_LOCK(&sched_mutex);
926 // either we have threads to run, or we were interrupted:
927 ASSERT(!EMPTY_RUN_QUEUE() || interrupted);
931 #if !defined(RTS_SUPPORTS_THREADS)
932 /* Probably a real deadlock. Send the current main thread the
933 * Deadlock exception (or in the SMP build, send *all* main
934 * threads the deadlock exception, since none of them can make
940 switch (m->tso->why_blocked) {
942 case BlockedOnBlackHole:
943 case BlockedOnException:
945 raiseAsync(m->tso, (StgClosure *)NonTermination_closure);
948 barf("deadlock: main thread blocked in a strange way");
955 /* ----------------------------------------------------------------------------
956 * Process an event (GRAN only)
957 * ------------------------------------------------------------------------- */
961 scheduleProcessEvent(rtsEvent *event)
965 if (RtsFlags.GranFlags.Light)
966 GranSimLight_enter_system(event, &ActiveTSO); // adjust ActiveTSO etc
968 /* adjust time based on time-stamp */
969 if (event->time > CurrentTime[CurrentProc] &&
970 event->evttype != ContinueThread)
971 CurrentTime[CurrentProc] = event->time;
973 /* Deal with the idle PEs (may issue FindWork or MoveSpark events) */
974 if (!RtsFlags.GranFlags.Light)
977 IF_DEBUG(gran, debugBelch("GRAN: switch by event-type\n"));
979 /* main event dispatcher in GranSim */
980 switch (event->evttype) {
981 /* Should just be continuing execution */
983 IF_DEBUG(gran, debugBelch("GRAN: doing ContinueThread\n"));
984 /* ToDo: check assertion
985 ASSERT(run_queue_hd != (StgTSO*)NULL &&
986 run_queue_hd != END_TSO_QUEUE);
988 /* Ignore ContinueThreads for fetching threads (if synchr comm) */
989 if (!RtsFlags.GranFlags.DoAsyncFetch &&
990 procStatus[CurrentProc]==Fetching) {
991 debugBelch("ghuH: Spurious ContinueThread while Fetching ignored; TSO %d (%p) [PE %d]\n",
992 CurrentTSO->id, CurrentTSO, CurrentProc);
995 /* Ignore ContinueThreads for completed threads */
996 if (CurrentTSO->what_next == ThreadComplete) {
997 debugBelch("ghuH: found a ContinueThread event for completed thread %d (%p) [PE %d] (ignoring ContinueThread)\n",
998 CurrentTSO->id, CurrentTSO, CurrentProc);
1001 /* Ignore ContinueThreads for threads that are being migrated */
1002 if (PROCS(CurrentTSO)==Nowhere) {
1003 debugBelch("ghuH: trying to run the migrating TSO %d (%p) [PE %d] (ignoring ContinueThread)\n",
1004 CurrentTSO->id, CurrentTSO, CurrentProc);
1007 /* The thread should be at the beginning of the run queue */
1008 if (CurrentTSO!=run_queue_hds[CurrentProc]) {
1009 debugBelch("ghuH: TSO %d (%p) [PE %d] is not at the start of the run_queue when doing a ContinueThread\n",
1010 CurrentTSO->id, CurrentTSO, CurrentProc);
1011 break; // run the thread anyway
1014 new_event(proc, proc, CurrentTime[proc],
1016 (StgTSO*)NULL, (StgClosure*)NULL, (rtsSpark*)NULL);
1018 */ /* Catches superfluous CONTINUEs -- should be unnecessary */
1019 break; // now actually run the thread; DaH Qu'vam yImuHbej
1022 do_the_fetchnode(event);
1023 goto next_thread; /* handle next event in event queue */
1026 do_the_globalblock(event);
1027 goto next_thread; /* handle next event in event queue */
1030 do_the_fetchreply(event);
1031 goto next_thread; /* handle next event in event queue */
1033 case UnblockThread: /* Move from the blocked queue to the tail of */
1034 do_the_unblock(event);
1035 goto next_thread; /* handle next event in event queue */
1037 case ResumeThread: /* Move from the blocked queue to the tail of */
1038 /* the runnable queue ( i.e. Qu' SImqa'lu') */
1039 event->tso->gran.blocktime +=
1040 CurrentTime[CurrentProc] - event->tso->gran.blockedat;
1041 do_the_startthread(event);
1042 goto next_thread; /* handle next event in event queue */
1045 do_the_startthread(event);
1046 goto next_thread; /* handle next event in event queue */
1049 do_the_movethread(event);
1050 goto next_thread; /* handle next event in event queue */
1053 do_the_movespark(event);
1054 goto next_thread; /* handle next event in event queue */
1057 do_the_findwork(event);
1058 goto next_thread; /* handle next event in event queue */
1061 barf("Illegal event type %u\n", event->evttype);
1064 /* This point was scheduler_loop in the old RTS */
1066 IF_DEBUG(gran, debugBelch("GRAN: after main switch\n"));
1068 TimeOfLastEvent = CurrentTime[CurrentProc];
1069 TimeOfNextEvent = get_time_of_next_event();
1070 IgnoreEvents=(TimeOfNextEvent==0); // HWL HACK
1071 // CurrentTSO = ThreadQueueHd;
1073 IF_DEBUG(gran, debugBelch("GRAN: time of next event is: %ld\n",
1076 if (RtsFlags.GranFlags.Light)
1077 GranSimLight_leave_system(event, &ActiveTSO);
1079 EndOfTimeSlice = CurrentTime[CurrentProc]+RtsFlags.GranFlags.time_slice;
1082 debugBelch("GRAN: end of time-slice is %#lx\n", EndOfTimeSlice));
1084 /* in a GranSim setup the TSO stays on the run queue */
1086 /* Take a thread from the run queue. */
1087 POP_RUN_QUEUE(t); // take_off_run_queue(t);
1090 debugBelch("GRAN: About to run current thread, which is\n");
1093 context_switch = 0; // turned on via GranYield, checking events and time slice
1096 DumpGranEvent(GR_SCHEDULE, t));
1098 procStatus[CurrentProc] = Busy;
1102 /* ----------------------------------------------------------------------------
1103 * Send pending messages (PARALLEL_HASKELL only)
1104 * ------------------------------------------------------------------------- */
1106 #if defined(PARALLEL_HASKELL)
1108 scheduleSendPendingMessages(void)
1114 # if defined(PAR) // global Mem.Mgmt., omit for now
1115 if (PendingFetches != END_BF_QUEUE) {
1120 if (RtsFlags.ParFlags.BufferTime) {
1121 // if we use message buffering, we must send away all message
1122 // packets which have become too old...
1128 /* ----------------------------------------------------------------------------
1129 * Activate spark threads (PARALLEL_HASKELL only)
1130 * ------------------------------------------------------------------------- */
1132 #if defined(PARALLEL_HASKELL)
1134 scheduleActivateSpark(void)
1137 ASSERT(EMPTY_RUN_QUEUE());
1138 /* We get here if the run queue is empty and want some work.
1139 We try to turn a spark into a thread, and add it to the run queue,
1140 from where it will be picked up in the next iteration of the scheduler
1144 /* :-[ no local threads => look out for local sparks */
1145 /* the spark pool for the current PE */
1146 pool = &(cap.r.rSparks); // JB: cap = (old) MainCap
1147 if (advisory_thread_count < RtsFlags.ParFlags.maxThreads &&
1148 pool->hd < pool->tl) {
1150 * ToDo: add GC code check that we really have enough heap afterwards!!
1152 * If we're here (no runnable threads) and we have pending
1153 * sparks, we must have a space problem. Get enough space
1154 * to turn one of those pending sparks into a
1158 spark = findSpark(rtsFalse); /* get a spark */
1159 if (spark != (rtsSpark) NULL) {
1160 tso = createThreadFromSpark(spark); /* turn the spark into a thread */
1161 IF_PAR_DEBUG(fish, // schedule,
1162 debugBelch("==== schedule: Created TSO %d (%p); %d threads active\n",
1163 tso->id, tso, advisory_thread_count));
1165 if (tso==END_TSO_QUEUE) { /* failed to activate spark->back to loop */
1166 IF_PAR_DEBUG(fish, // schedule,
1167 debugBelch("==^^ failed to create thread from spark @ %lx\n",
1169 return rtsFalse; /* failed to generate a thread */
1170 } /* otherwise fall through & pick-up new tso */
1172 IF_PAR_DEBUG(fish, // schedule,
1173 debugBelch("==^^ no local sparks (spark pool contains only NFs: %d)\n",
1174 spark_queue_len(pool)));
1175 return rtsFalse; /* failed to generate a thread */
1177 return rtsTrue; /* success in generating a thread */
1178 } else { /* no more threads permitted or pool empty */
1179 return rtsFalse; /* failed to generateThread */
1182 tso = NULL; // avoid compiler warning only
1183 return rtsFalse; /* dummy in non-PAR setup */
1186 #endif // PARALLEL_HASKELL
1188 /* ----------------------------------------------------------------------------
1189 * Get work from a remote node (PARALLEL_HASKELL only)
1190 * ------------------------------------------------------------------------- */
1192 #if defined(PARALLEL_HASKELL)
1194 scheduleGetRemoteWork(rtsBool *receivedFinish)
1196 ASSERT(EMPTY_RUN_QUEUE());
1198 if (RtsFlags.ParFlags.BufferTime) {
1199 IF_PAR_DEBUG(verbose,
1200 debugBelch("...send all pending data,"));
1203 for (i=1; i<=nPEs; i++)
1204 sendImmediately(i); // send all messages away immediately
1208 //++EDEN++ idle() , i.e. send all buffers, wait for work
1209 // suppress fishing in EDEN... just look for incoming messages
1210 // (blocking receive)
1211 IF_PAR_DEBUG(verbose,
1212 debugBelch("...wait for incoming messages...\n"));
1213 *receivedFinish = processMessages(); // blocking receive...
1215 // and reenter scheduling loop after having received something
1216 // (return rtsFalse below)
1218 # else /* activate SPARKS machinery */
1219 /* We get here, if we have no work, tried to activate a local spark, but still
1220 have no work. We try to get a remote spark, by sending a FISH message.
1221 Thread migration should be added here, and triggered when a sequence of
1222 fishes returns without work. */
1223 delay = (RtsFlags.ParFlags.fishDelay!=0ll ? RtsFlags.ParFlags.fishDelay : 0ll);
1225 /* =8-[ no local sparks => look for work on other PEs */
1227 * We really have absolutely no work. Send out a fish
1228 * (there may be some out there already), and wait for
1229 * something to arrive. We clearly can't run any threads
1230 * until a SCHEDULE or RESUME arrives, and so that's what
1231 * we're hoping to see. (Of course, we still have to
1232 * respond to other types of messages.)
1234 rtsTime now = msTime() /*CURRENT_TIME*/;
1235 IF_PAR_DEBUG(verbose,
1236 debugBelch("-- now=%ld\n", now));
1237 IF_PAR_DEBUG(fish, // verbose,
1238 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
1239 (last_fish_arrived_at!=0 &&
1240 last_fish_arrived_at+delay > now)) {
1241 debugBelch("--$$ <%llu> delaying FISH until %llu (last fish %llu, delay %llu)\n",
1242 now, last_fish_arrived_at+delay,
1243 last_fish_arrived_at,
1247 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
1248 advisory_thread_count < RtsFlags.ParFlags.maxThreads) { // send a FISH, but when?
1249 if (last_fish_arrived_at==0 ||
1250 (last_fish_arrived_at+delay <= now)) { // send FISH now!
1251 /* outstandingFishes is set in sendFish, processFish;
1252 avoid flooding system with fishes via delay */
1253 next_fish_to_send_at = 0;
1255 /* ToDo: this should be done in the main scheduling loop to avoid the
1256 busy wait here; not so bad if fish delay is very small */
1257 int iq = 0; // DEBUGGING -- HWL
1258 next_fish_to_send_at = last_fish_arrived_at+delay; // remember when to send
1259 /* send a fish when ready, but process messages that arrive in the meantime */
1261 if (PacketsWaiting()) {
1263 *receivedFinish = processMessages();
1266 } while (!*receivedFinish || now<next_fish_to_send_at);
1267 // JB: This means the fish could become obsolete, if we receive
1268 // work. Better check for work again?
1269 // last line: while (!receivedFinish || !haveWork || now<...)
1270 // next line: if (receivedFinish || haveWork )
1272 if (*receivedFinish) // no need to send a FISH if we are finishing anyway
1273 return rtsFalse; // NB: this will leave scheduler loop
1274 // immediately after return!
1276 IF_PAR_DEBUG(fish, // verbose,
1277 debugBelch("--$$ <%llu> sent delayed fish (%d processMessages); active/total threads=%d/%d\n",now,iq,run_queue_len(),advisory_thread_count));
1281 // JB: IMHO, this should all be hidden inside sendFish(...)
1283 sendFish(pe, thisPE, NEW_FISH_AGE, NEW_FISH_HISTORY,
1286 // Global statistics: count no. of fishes
1287 if (RtsFlags.ParFlags.ParStats.Global &&
1288 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
1289 globalParStats.tot_fish_mess++;
1293 /* delayed fishes must have been sent by now! */
1294 next_fish_to_send_at = 0;
1297 *receivedFinish = processMessages();
1298 # endif /* SPARKS */
1301 /* NB: this function always returns rtsFalse, meaning the scheduler
1302 loop continues with the next iteration;
1304 return code means success in finding work; we enter this function
1305 if there is no local work, thus have to send a fish which takes
1306 time until it arrives with work; in the meantime we should process
1307 messages in the main loop;
1310 #endif // PARALLEL_HASKELL
1312 /* ----------------------------------------------------------------------------
1313 * PAR/GRAN: Report stats & debugging info(?)
1314 * ------------------------------------------------------------------------- */
1316 #if defined(PAR) || defined(GRAN)
1318 scheduleGranParReport(void)
1320 ASSERT(run_queue_hd != END_TSO_QUEUE);
1322 /* Take a thread from the run queue, if we have work */
1323 POP_RUN_QUEUE(t); // take_off_run_queue(END_TSO_QUEUE);
1325 /* If this TSO has got its outport closed in the meantime,
1326 * it mustn't be run. Instead, we have to clean it up as if it was finished.
1327 * It has to be marked as TH_DEAD for this purpose.
1328 * If it is TH_TERM instead, it is supposed to have finished in the normal way.
1330 JB: TODO: investigate wether state change field could be nuked
1331 entirely and replaced by the normal tso state (whatnext
1332 field). All we want to do is to kill tsos from outside.
1335 /* ToDo: write something to the log-file
1336 if (RTSflags.ParFlags.granSimStats && !sameThread)
1337 DumpGranEvent(GR_SCHEDULE, RunnableThreadsHd);
1341 /* the spark pool for the current PE */
1342 pool = &(cap.r.rSparks); // cap = (old) MainCap
1345 debugBelch("--=^ %d threads, %d sparks on [%#x]\n",
1346 run_queue_len(), spark_queue_len(pool), CURRENT_PROC));
1349 debugBelch("--=^ %d threads, %d sparks on [%#x]\n",
1350 run_queue_len(), spark_queue_len(pool), CURRENT_PROC));
1352 if (RtsFlags.ParFlags.ParStats.Full &&
1353 (t->par.sparkname != (StgInt)0) && // only log spark generated threads
1354 (emitSchedule || // forced emit
1355 (t && LastTSO && t->id != LastTSO->id))) {
1357 we are running a different TSO, so write a schedule event to log file
1358 NB: If we use fair scheduling we also have to write a deschedule
1359 event for LastTSO; with unfair scheduling we know that the
1360 previous tso has blocked whenever we switch to another tso, so
1361 we don't need it in GUM for now
1363 IF_PAR_DEBUG(fish, // schedule,
1364 debugBelch("____ scheduling spark generated thread %d (%lx) (%lx) via a forced emit\n",t->id,t,t->par.sparkname));
1366 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
1367 GR_SCHEDULE, t, (StgClosure *)NULL, 0, 0);
1368 emitSchedule = rtsFalse;
1373 /* ----------------------------------------------------------------------------
1374 * After running a thread...
1375 * ASSUMES: sched_mutex
1376 * ------------------------------------------------------------------------- */
1379 schedulePostRunThread(void)
1382 /* HACK 675: if the last thread didn't yield, make sure to print a
1383 SCHEDULE event to the log file when StgRunning the next thread, even
1384 if it is the same one as before */
1386 TimeOfLastYield = CURRENT_TIME;
1389 /* some statistics gathering in the parallel case */
1391 #if defined(GRAN) || defined(PAR) || defined(EDEN)
1395 IF_DEBUG(gran, DumpGranEvent(GR_DESCHEDULE, t));
1396 globalGranStats.tot_heapover++;
1398 globalParStats.tot_heapover++;
1405 DumpGranEvent(GR_DESCHEDULE, t));
1406 globalGranStats.tot_stackover++;
1409 // DumpGranEvent(GR_DESCHEDULE, t);
1410 globalParStats.tot_stackover++;
1414 case ThreadYielding:
1417 DumpGranEvent(GR_DESCHEDULE, t));
1418 globalGranStats.tot_yields++;
1421 // DumpGranEvent(GR_DESCHEDULE, t);
1422 globalParStats.tot_yields++;
1429 debugBelch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: ",
1430 t->id, t, whatNext_strs[t->what_next], t->block_info.closure,
1431 (t->block_info.closure==(StgClosure*)NULL ? 99 : where_is(t->block_info.closure)));
1432 if (t->block_info.closure!=(StgClosure*)NULL)
1433 print_bq(t->block_info.closure);
1436 // ??? needed; should emit block before
1438 DumpGranEvent(GR_DESCHEDULE, t));
1439 prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
1442 ASSERT(procStatus[CurrentProc]==Busy ||
1443 ((procStatus[CurrentProc]==Fetching) &&
1444 (t->block_info.closure!=(StgClosure*)NULL)));
1445 if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
1446 !(!RtsFlags.GranFlags.DoAsyncFetch &&
1447 procStatus[CurrentProc]==Fetching))
1448 procStatus[CurrentProc] = Idle;
1451 //++PAR++ blockThread() writes the event (change?)
1455 case ThreadFinished:
1459 barf("parGlobalStats: unknown return code");
1465 /* -----------------------------------------------------------------------------
1466 * Handle a thread that returned to the scheduler with ThreadHeepOverflow
1467 * ASSUMES: sched_mutex
1468 * -------------------------------------------------------------------------- */
1471 scheduleHandleHeapOverflow( Capability *cap, StgTSO *t )
1473 // did the task ask for a large block?
1474 if (cap->r.rHpAlloc > BLOCK_SIZE) {
1475 // if so, get one and push it on the front of the nursery.
1479 blocks = (lnat)BLOCK_ROUND_UP(cap->r.rHpAlloc) / BLOCK_SIZE;
1482 debugBelch("--<< thread %ld (%s) stopped: requesting a large block (size %ld)\n",
1483 (long)t->id, whatNext_strs[t->what_next], blocks));
1485 // don't do this if the nursery is (nearly) full, we'll GC first.
1486 if (cap->r.rCurrentNursery->link != NULL ||
1487 cap->r.rNursery->n_blocks == 1) { // paranoia to prevent infinite loop
1488 // if the nursery has only one block.
1490 bd = allocGroup( blocks );
1491 cap->r.rNursery->n_blocks += blocks;
1493 // link the new group into the list
1494 bd->link = cap->r.rCurrentNursery;
1495 bd->u.back = cap->r.rCurrentNursery->u.back;
1496 if (cap->r.rCurrentNursery->u.back != NULL) {
1497 cap->r.rCurrentNursery->u.back->link = bd;
1500 ASSERT(g0s0->blocks == cap->r.rCurrentNursery &&
1501 g0s0 == cap->r.rNursery);
1503 cap->r.rNursery->blocks = bd;
1505 cap->r.rCurrentNursery->u.back = bd;
1507 // initialise it as a nursery block. We initialise the
1508 // step, gen_no, and flags field of *every* sub-block in
1509 // this large block, because this is easier than making
1510 // sure that we always find the block head of a large
1511 // block whenever we call Bdescr() (eg. evacuate() and
1512 // isAlive() in the GC would both have to do this, at
1516 for (x = bd; x < bd + blocks; x++) {
1517 x->step = cap->r.rNursery;
1523 // This assert can be a killer if the app is doing lots
1524 // of large block allocations.
1525 IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
1527 // now update the nursery to point to the new block
1528 cap->r.rCurrentNursery = bd;
1530 // we might be unlucky and have another thread get on the
1531 // run queue before us and steal the large block, but in that
1532 // case the thread will just end up requesting another large
1534 PUSH_ON_RUN_QUEUE(t);
1535 return rtsFalse; /* not actually GC'ing */
1540 debugBelch("--<< thread %ld (%s) stopped: HeapOverflow\n",
1541 (long)t->id, whatNext_strs[t->what_next]));
1544 ASSERT(!is_on_queue(t,CurrentProc));
1545 #elif defined(PARALLEL_HASKELL)
1546 /* Currently we emit a DESCHEDULE event before GC in GUM.
1547 ToDo: either add separate event to distinguish SYSTEM time from rest
1548 or just nuke this DESCHEDULE (and the following SCHEDULE) */
1549 if (0 && RtsFlags.ParFlags.ParStats.Full) {
1550 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
1551 GR_DESCHEDULE, t, (StgClosure *)NULL, 0, 0);
1552 emitSchedule = rtsTrue;
1556 PUSH_ON_RUN_QUEUE(t);
1558 /* actual GC is done at the end of the while loop in schedule() */
1561 /* -----------------------------------------------------------------------------
1562 * Handle a thread that returned to the scheduler with ThreadStackOverflow
1563 * ASSUMES: sched_mutex
1564 * -------------------------------------------------------------------------- */
1567 scheduleHandleStackOverflow( StgTSO *t)
1569 IF_DEBUG(scheduler,debugBelch("--<< thread %ld (%s) stopped, StackOverflow\n",
1570 (long)t->id, whatNext_strs[t->what_next]));
1571 /* just adjust the stack for this thread, then pop it back
1576 /* enlarge the stack */
1577 StgTSO *new_t = threadStackOverflow(t);
1579 /* This TSO has moved, so update any pointers to it from the
1580 * main thread stack. It better not be on any other queues...
1581 * (it shouldn't be).
1583 if (t->main != NULL) {
1584 t->main->tso = new_t;
1586 PUSH_ON_RUN_QUEUE(new_t);
1590 /* -----------------------------------------------------------------------------
1591 * Handle a thread that returned to the scheduler with ThreadYielding
1592 * ASSUMES: sched_mutex
1593 * -------------------------------------------------------------------------- */
1596 scheduleHandleYield( StgTSO *t, nat prev_what_next )
1598 // Reset the context switch flag. We don't do this just before
1599 // running the thread, because that would mean we would lose ticks
1600 // during GC, which can lead to unfair scheduling (a thread hogs
1601 // the CPU because the tick always arrives during GC). This way
1602 // penalises threads that do a lot of allocation, but that seems
1603 // better than the alternative.
1606 /* put the thread back on the run queue. Then, if we're ready to
1607 * GC, check whether this is the last task to stop. If so, wake
1608 * up the GC thread. getThread will block during a GC until the
1612 if (t->what_next != prev_what_next) {
1613 debugBelch("--<< thread %ld (%s) stopped to switch evaluators\n",
1614 (long)t->id, whatNext_strs[t->what_next]);
1616 debugBelch("--<< thread %ld (%s) stopped, yielding\n",
1617 (long)t->id, whatNext_strs[t->what_next]);
1622 //debugBelch("&& Doing sanity check on yielding TSO %ld.", t->id);
1624 ASSERT(t->link == END_TSO_QUEUE);
1626 // Shortcut if we're just switching evaluators: don't bother
1627 // doing stack squeezing (which can be expensive), just run the
1629 if (t->what_next != prev_what_next) {
1636 ASSERT(!is_on_queue(t,CurrentProc));
1639 //debugBelch("&& Doing sanity check on all ThreadQueues (and their TSOs).");
1640 checkThreadQsSanity(rtsTrue));
1647 /* add a ContinueThread event to actually process the thread */
1648 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1650 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1652 debugBelch("GRAN: eventq and runnableq after adding yielded thread to queue again:\n");
1659 /* -----------------------------------------------------------------------------
1660 * Handle a thread that returned to the scheduler with ThreadBlocked
1661 * ASSUMES: sched_mutex
1662 * -------------------------------------------------------------------------- */
1665 scheduleHandleThreadBlocked( StgTSO *t
1666 #if !defined(GRAN) && !defined(DEBUG)
1673 debugBelch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: \n",
1674 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)));
1675 if (t->block_info.closure!=(StgClosure*)NULL) print_bq(t->block_info.closure));
1677 // ??? needed; should emit block before
1679 DumpGranEvent(GR_DESCHEDULE, t));
1680 prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
1683 ASSERT(procStatus[CurrentProc]==Busy ||
1684 ((procStatus[CurrentProc]==Fetching) &&
1685 (t->block_info.closure!=(StgClosure*)NULL)));
1686 if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
1687 !(!RtsFlags.GranFlags.DoAsyncFetch &&
1688 procStatus[CurrentProc]==Fetching))
1689 procStatus[CurrentProc] = Idle;
1693 debugBelch("--<< thread %ld (%p; %s) stopped, blocking on node %p with BQ: \n",
1694 t->id, t, whatNext_strs[t->what_next], t->block_info.closure));
1697 if (t->block_info.closure!=(StgClosure*)NULL)
1698 print_bq(t->block_info.closure));
1700 /* Send a fetch (if BlockedOnGA) and dump event to log file */
1703 /* whatever we schedule next, we must log that schedule */
1704 emitSchedule = rtsTrue;
1707 /* don't need to do anything. Either the thread is blocked on
1708 * I/O, in which case we'll have called addToBlockedQueue
1709 * previously, or it's blocked on an MVar or Blackhole, in which
1710 * case it'll be on the relevant queue already.
1712 ASSERT(t->why_blocked != NotBlocked);
1714 debugBelch("--<< thread %d (%s) stopped: ",
1715 t->id, whatNext_strs[t->what_next]);
1716 printThreadBlockage(t);
1719 /* Only for dumping event to log file
1720 ToDo: do I need this in GranSim, too?
1726 /* -----------------------------------------------------------------------------
1727 * Handle a thread that returned to the scheduler with ThreadFinished
1728 * ASSUMES: sched_mutex
1729 * -------------------------------------------------------------------------- */
1732 scheduleHandleThreadFinished( StgMainThread *mainThread
1733 USED_WHEN_RTS_SUPPORTS_THREADS,
1737 /* Need to check whether this was a main thread, and if so,
1738 * return with the return value.
1740 * We also end up here if the thread kills itself with an
1741 * uncaught exception, see Exception.cmm.
1743 IF_DEBUG(scheduler,debugBelch("--++ thread %d (%s) finished\n",
1744 t->id, whatNext_strs[t->what_next]));
1747 endThread(t, CurrentProc); // clean-up the thread
1748 #elif defined(PARALLEL_HASKELL)
1749 /* For now all are advisory -- HWL */
1750 //if(t->priority==AdvisoryPriority) ??
1751 advisory_thread_count--; // JB: Caution with this counter, buggy!
1754 if(t->dist.priority==RevalPriority)
1758 # if defined(EDENOLD)
1759 // the thread could still have an outport... (BUG)
1760 if (t->eden.outport != -1) {
1761 // delete the outport for the tso which has finished...
1762 IF_PAR_DEBUG(eden_ports,
1763 debugBelch("WARNING: Scheduler removes outport %d for TSO %d.\n",
1764 t->eden.outport, t->id));
1767 // thread still in the process (HEAVY BUG! since outport has just been closed...)
1768 if (t->eden.epid != -1) {
1769 IF_PAR_DEBUG(eden_ports,
1770 debugBelch("WARNING: Scheduler removes TSO %d from process %d .\n",
1771 t->id, t->eden.epid));
1772 removeTSOfromProcess(t);
1777 if (RtsFlags.ParFlags.ParStats.Full &&
1778 !RtsFlags.ParFlags.ParStats.Suppressed)
1779 DumpEndEvent(CURRENT_PROC, t, rtsFalse /* not mandatory */);
1781 // t->par only contains statistics: left out for now...
1783 debugBelch("**** end thread: ended sparked thread %d (%lx); sparkname: %lx\n",
1784 t->id,t,t->par.sparkname));
1786 #endif // PARALLEL_HASKELL
1789 // Check whether the thread that just completed was a main
1790 // thread, and if so return with the result.
1792 // There is an assumption here that all thread completion goes
1793 // through this point; we need to make sure that if a thread
1794 // ends up in the ThreadKilled state, that it stays on the run
1795 // queue so it can be dealt with here.
1798 #if defined(RTS_SUPPORTS_THREADS)
1801 mainThread->tso == t
1805 // We are a bound thread: this must be our thread that just
1807 ASSERT(mainThread->tso == t);
1809 if (t->what_next == ThreadComplete) {
1810 if (mainThread->ret) {
1811 // NOTE: return val is tso->sp[1] (see StgStartup.hc)
1812 *(mainThread->ret) = (StgClosure *)mainThread->tso->sp[1];
1814 mainThread->stat = Success;
1816 if (mainThread->ret) {
1817 *(mainThread->ret) = NULL;
1820 mainThread->stat = Interrupted;
1822 mainThread->stat = Killed;
1826 removeThreadLabel((StgWord)mainThread->tso->id);
1828 if (mainThread->prev == NULL) {
1829 ASSERT(mainThread == main_threads);
1830 main_threads = mainThread->link;
1832 mainThread->prev->link = mainThread->link;
1834 if (mainThread->link != NULL) {
1835 mainThread->link->prev = mainThread->prev;
1837 releaseCapability(cap);
1838 return rtsTrue; // tells schedule() to return
1841 #ifdef RTS_SUPPORTS_THREADS
1842 ASSERT(t->main == NULL);
1844 if (t->main != NULL) {
1845 // Must be a main thread that is not the topmost one. Leave
1846 // it on the run queue until the stack has unwound to the
1847 // point where we can deal with this. Leaving it on the run
1848 // queue also ensures that the garbage collector knows about
1849 // this thread and its return value (it gets dropped from the
1850 // all_threads list so there's no other way to find it).
1851 APPEND_TO_RUN_QUEUE(t);
1857 /* -----------------------------------------------------------------------------
1858 * Perform a heap census, if PROFILING
1859 * -------------------------------------------------------------------------- */
1862 scheduleDoHeapProfile( rtsBool ready_to_gc STG_UNUSED )
1864 #if defined(PROFILING)
1865 // When we have +RTS -i0 and we're heap profiling, do a census at
1866 // every GC. This lets us get repeatable runs for debugging.
1867 if (performHeapProfile ||
1868 (RtsFlags.ProfFlags.profileInterval==0 &&
1869 RtsFlags.ProfFlags.doHeapProfile && ready_to_gc)) {
1870 GarbageCollect(GetRoots, rtsTrue);
1872 performHeapProfile = rtsFalse;
1873 return rtsTrue; // true <=> we already GC'd
1879 /* -----------------------------------------------------------------------------
1880 * Perform a garbage collection if necessary
1881 * ASSUMES: sched_mutex
1882 * -------------------------------------------------------------------------- */
1885 scheduleDoGC( Capability *cap STG_UNUSED )
1889 static rtsBool waiting_for_gc;
1890 int n_capabilities = RtsFlags.ParFlags.nNodes - 1;
1891 // subtract one because we're already holding one.
1892 Capability *caps[n_capabilities];
1896 // In order to GC, there must be no threads running Haskell code.
1897 // Therefore, the GC thread needs to hold *all* the capabilities,
1898 // and release them after the GC has completed.
1900 // This seems to be the simplest way: previous attempts involved
1901 // making all the threads with capabilities give up their
1902 // capabilities and sleep except for the *last* one, which
1903 // actually did the GC. But it's quite hard to arrange for all
1904 // the other tasks to sleep and stay asleep.
1907 // Someone else is already trying to GC
1908 if (waiting_for_gc) return;
1909 waiting_for_gc = rtsTrue;
1911 caps[n_capabilities] = cap;
1912 while (n_capabilities > 0) {
1913 IF_DEBUG(scheduler, sched_belch("ready_to_gc, grabbing all the capabilies (%d left)", n_capabilities));
1914 waitForReturnCapability(&sched_mutex, &cap);
1916 caps[n_capabilities] = cap;
1919 waiting_for_gc = rtsFalse;
1922 /* Kick any transactions which are invalid back to their
1923 * atomically frames. When next scheduled they will try to
1924 * commit, this commit will fail and they will retry.
1926 for (t = all_threads; t != END_TSO_QUEUE; t = t -> link) {
1927 if (t -> what_next != ThreadRelocated && t -> trec != NO_TREC && t -> why_blocked == NotBlocked) {
1928 if (!stmValidateTransaction (t -> trec)) {
1929 IF_DEBUG(stm, sched_belch("trec %p found wasting its time", t));
1931 // strip the stack back to the ATOMICALLY_FRAME, aborting
1932 // the (nested) transaction, and saving the stack of any
1933 // partially-evaluated thunks on the heap.
1934 raiseAsync_(t, NULL, rtsTrue);
1937 ASSERT(get_itbl((StgClosure *)t->sp)->type == ATOMICALLY_FRAME);
1943 // so this happens periodically:
1944 scheduleCheckBlackHoles();
1946 /* everybody back, start the GC.
1947 * Could do it in this thread, or signal a condition var
1948 * to do it in another thread. Either way, we need to
1949 * broadcast on gc_pending_cond afterward.
1951 #if defined(RTS_SUPPORTS_THREADS)
1952 IF_DEBUG(scheduler,sched_belch("doing GC"));
1954 GarbageCollect(GetRoots,rtsFalse);
1958 // release our stash of capabilities.
1960 for (i = 0; i < RtsFlags.ParFlags.nNodes-1; i++) {
1961 releaseCapability(caps[i]);
1967 /* add a ContinueThread event to continue execution of current thread */
1968 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1970 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1972 debugBelch("GRAN: eventq and runnableq after Garbage collection:\n\n");
1978 /* ---------------------------------------------------------------------------
1979 * rtsSupportsBoundThreads(): is the RTS built to support bound threads?
1980 * used by Control.Concurrent for error checking.
1981 * ------------------------------------------------------------------------- */
1984 rtsSupportsBoundThreads(void)
1986 #if defined(RTS_SUPPORTS_THREADS)
1993 /* ---------------------------------------------------------------------------
1994 * isThreadBound(tso): check whether tso is bound to an OS thread.
1995 * ------------------------------------------------------------------------- */
1998 isThreadBound(StgTSO* tso USED_IN_THREADED_RTS)
2000 #if defined(RTS_SUPPORTS_THREADS)
2001 return (tso->main != NULL);
2006 /* ---------------------------------------------------------------------------
2007 * Singleton fork(). Do not copy any running threads.
2008 * ------------------------------------------------------------------------- */
2010 #ifndef mingw32_HOST_OS
2011 #define FORKPROCESS_PRIMOP_SUPPORTED
2014 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
2016 deleteThreadImmediately(StgTSO *tso);
2019 forkProcess(HsStablePtr *entry
2020 #ifndef FORKPROCESS_PRIMOP_SUPPORTED
2025 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
2031 IF_DEBUG(scheduler,sched_belch("forking!"));
2032 rts_lock(); // This not only acquires sched_mutex, it also
2033 // makes sure that no other threads are running
2037 if (pid) { /* parent */
2039 /* just return the pid */
2043 } else { /* child */
2046 // delete all threads
2047 run_queue_hd = run_queue_tl = END_TSO_QUEUE;
2049 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
2052 // don't allow threads to catch the ThreadKilled exception
2053 deleteThreadImmediately(t);
2056 // wipe the main thread list
2057 while((m = main_threads) != NULL) {
2058 main_threads = m->link;
2059 # ifdef THREADED_RTS
2060 closeCondition(&m->bound_thread_cond);
2065 rc = rts_evalStableIO(entry, NULL); // run the action
2066 rts_checkSchedStatus("forkProcess",rc);
2070 hs_exit(); // clean up and exit
2073 #else /* !FORKPROCESS_PRIMOP_SUPPORTED */
2074 barf("forkProcess#: primop not supported, sorry!\n");
2079 /* ---------------------------------------------------------------------------
2080 * deleteAllThreads(): kill all the live threads.
2082 * This is used when we catch a user interrupt (^C), before performing
2083 * any necessary cleanups and running finalizers.
2085 * Locks: sched_mutex held.
2086 * ------------------------------------------------------------------------- */
2089 deleteAllThreads ( void )
2092 IF_DEBUG(scheduler,sched_belch("deleting all threads"));
2093 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
2094 if (t->what_next == ThreadRelocated) {
2097 next = t->global_link;
2102 // The run queue now contains a bunch of ThreadKilled threads. We
2103 // must not throw these away: the main thread(s) will be in there
2104 // somewhere, and the main scheduler loop has to deal with it.
2105 // Also, the run queue is the only thing keeping these threads from
2106 // being GC'd, and we don't want the "main thread has been GC'd" panic.
2108 ASSERT(blocked_queue_hd == END_TSO_QUEUE);
2109 ASSERT(blackhole_queue == END_TSO_QUEUE);
2110 ASSERT(sleeping_queue == END_TSO_QUEUE);
2113 /* startThread and insertThread are now in GranSim.c -- HWL */
2116 /* ---------------------------------------------------------------------------
2117 * Suspending & resuming Haskell threads.
2119 * When making a "safe" call to C (aka _ccall_GC), the task gives back
2120 * its capability before calling the C function. This allows another
2121 * task to pick up the capability and carry on running Haskell
2122 * threads. It also means that if the C call blocks, it won't lock
2125 * The Haskell thread making the C call is put to sleep for the
2126 * duration of the call, on the susepended_ccalling_threads queue. We
2127 * give out a token to the task, which it can use to resume the thread
2128 * on return from the C function.
2129 * ------------------------------------------------------------------------- */
2132 suspendThread( StgRegTable *reg )
2136 int saved_errno = errno;
2138 /* assume that *reg is a pointer to the StgRegTable part
2141 cap = (Capability *)((void *)((unsigned char*)reg - sizeof(StgFunTable)));
2143 ACQUIRE_LOCK(&sched_mutex);
2146 sched_belch("thread %d did a _ccall_gc", cap->r.rCurrentTSO->id));
2148 // XXX this might not be necessary --SDM
2149 cap->r.rCurrentTSO->what_next = ThreadRunGHC;
2151 threadPaused(cap->r.rCurrentTSO);
2152 cap->r.rCurrentTSO->link = suspended_ccalling_threads;
2153 suspended_ccalling_threads = cap->r.rCurrentTSO;
2155 if(cap->r.rCurrentTSO->blocked_exceptions == NULL) {
2156 cap->r.rCurrentTSO->why_blocked = BlockedOnCCall;
2157 cap->r.rCurrentTSO->blocked_exceptions = END_TSO_QUEUE;
2159 cap->r.rCurrentTSO->why_blocked = BlockedOnCCall_NoUnblockExc;
2162 /* Use the thread ID as the token; it should be unique */
2163 tok = cap->r.rCurrentTSO->id;
2165 /* Hand back capability */
2166 cap->r.rInHaskell = rtsFalse;
2167 releaseCapability(cap);
2169 #if defined(RTS_SUPPORTS_THREADS)
2170 /* Preparing to leave the RTS, so ensure there's a native thread/task
2171 waiting to take over.
2173 IF_DEBUG(scheduler, sched_belch("worker (token %d): leaving RTS", tok));
2176 RELEASE_LOCK(&sched_mutex);
2178 errno = saved_errno;
2183 resumeThread( StgInt tok )
2185 StgTSO *tso, **prev;
2187 int saved_errno = errno;
2189 #if defined(RTS_SUPPORTS_THREADS)
2190 /* Wait for permission to re-enter the RTS with the result. */
2191 ACQUIRE_LOCK(&sched_mutex);
2192 waitForReturnCapability(&sched_mutex, &cap);
2194 IF_DEBUG(scheduler, sched_belch("worker (token %d): re-entering RTS", tok));
2196 grabCapability(&cap);
2199 /* Remove the thread off of the suspended list */
2200 prev = &suspended_ccalling_threads;
2201 for (tso = suspended_ccalling_threads;
2202 tso != END_TSO_QUEUE;
2203 prev = &tso->link, tso = tso->link) {
2204 if (tso->id == (StgThreadID)tok) {
2209 if (tso == END_TSO_QUEUE) {
2210 barf("resumeThread: thread not found");
2212 tso->link = END_TSO_QUEUE;
2214 if(tso->why_blocked == BlockedOnCCall) {
2215 awakenBlockedQueueNoLock(tso->blocked_exceptions);
2216 tso->blocked_exceptions = NULL;
2219 /* Reset blocking status */
2220 tso->why_blocked = NotBlocked;
2222 cap->r.rCurrentTSO = tso;
2223 cap->r.rInHaskell = rtsTrue;
2224 RELEASE_LOCK(&sched_mutex);
2225 errno = saved_errno;
2229 /* ---------------------------------------------------------------------------
2230 * Comparing Thread ids.
2232 * This is used from STG land in the implementation of the
2233 * instances of Eq/Ord for ThreadIds.
2234 * ------------------------------------------------------------------------ */
2237 cmp_thread(StgPtr tso1, StgPtr tso2)
2239 StgThreadID id1 = ((StgTSO *)tso1)->id;
2240 StgThreadID id2 = ((StgTSO *)tso2)->id;
2242 if (id1 < id2) return (-1);
2243 if (id1 > id2) return 1;
2247 /* ---------------------------------------------------------------------------
2248 * Fetching the ThreadID from an StgTSO.
2250 * This is used in the implementation of Show for ThreadIds.
2251 * ------------------------------------------------------------------------ */
2253 rts_getThreadId(StgPtr tso)
2255 return ((StgTSO *)tso)->id;
2260 labelThread(StgPtr tso, char *label)
2265 /* Caveat: Once set, you can only set the thread name to "" */
2266 len = strlen(label)+1;
2267 buf = stgMallocBytes(len * sizeof(char), "Schedule.c:labelThread()");
2268 strncpy(buf,label,len);
2269 /* Update will free the old memory for us */
2270 updateThreadLabel(((StgTSO *)tso)->id,buf);
2274 /* ---------------------------------------------------------------------------
2275 Create a new thread.
2277 The new thread starts with the given stack size. Before the
2278 scheduler can run, however, this thread needs to have a closure
2279 (and possibly some arguments) pushed on its stack. See
2280 pushClosure() in Schedule.h.
2282 createGenThread() and createIOThread() (in SchedAPI.h) are
2283 convenient packaged versions of this function.
2285 currently pri (priority) is only used in a GRAN setup -- HWL
2286 ------------------------------------------------------------------------ */
2288 /* currently pri (priority) is only used in a GRAN setup -- HWL */
2290 createThread(nat size, StgInt pri)
2293 createThread(nat size)
2300 /* First check whether we should create a thread at all */
2301 #if defined(PARALLEL_HASKELL)
2302 /* check that no more than RtsFlags.ParFlags.maxThreads threads are created */
2303 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads) {
2305 debugBelch("{createThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)\n",
2306 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
2307 return END_TSO_QUEUE;
2313 ASSERT(!RtsFlags.GranFlags.Light || CurrentProc==0);
2316 // ToDo: check whether size = stack_size - TSO_STRUCT_SIZEW
2318 /* catch ridiculously small stack sizes */
2319 if (size < MIN_STACK_WORDS + TSO_STRUCT_SIZEW) {
2320 size = MIN_STACK_WORDS + TSO_STRUCT_SIZEW;
2323 stack_size = size - TSO_STRUCT_SIZEW;
2325 tso = (StgTSO *)allocate(size);
2326 TICK_ALLOC_TSO(stack_size, 0);
2328 SET_HDR(tso, &stg_TSO_info, CCS_SYSTEM);
2330 SET_GRAN_HDR(tso, ThisPE);
2333 // Always start with the compiled code evaluator
2334 tso->what_next = ThreadRunGHC;
2336 tso->id = next_thread_id++;
2337 tso->why_blocked = NotBlocked;
2338 tso->blocked_exceptions = NULL;
2340 tso->saved_errno = 0;
2343 tso->stack_size = stack_size;
2344 tso->max_stack_size = round_to_mblocks(RtsFlags.GcFlags.maxStkSize)
2346 tso->sp = (P_)&(tso->stack) + stack_size;
2348 tso->trec = NO_TREC;
2351 tso->prof.CCCS = CCS_MAIN;
2354 /* put a stop frame on the stack */
2355 tso->sp -= sizeofW(StgStopFrame);
2356 SET_HDR((StgClosure*)tso->sp,(StgInfoTable *)&stg_stop_thread_info,CCS_SYSTEM);
2357 tso->link = END_TSO_QUEUE;
2361 /* uses more flexible routine in GranSim */
2362 insertThread(tso, CurrentProc);
2364 /* In a non-GranSim setup the pushing of a TSO onto the runq is separated
2370 if (RtsFlags.GranFlags.GranSimStats.Full)
2371 DumpGranEvent(GR_START,tso);
2372 #elif defined(PARALLEL_HASKELL)
2373 if (RtsFlags.ParFlags.ParStats.Full)
2374 DumpGranEvent(GR_STARTQ,tso);
2375 /* HACk to avoid SCHEDULE
2379 /* Link the new thread on the global thread list.
2381 tso->global_link = all_threads;
2385 tso->dist.priority = MandatoryPriority; //by default that is...
2389 tso->gran.pri = pri;
2391 tso->gran.magic = TSO_MAGIC; // debugging only
2393 tso->gran.sparkname = 0;
2394 tso->gran.startedat = CURRENT_TIME;
2395 tso->gran.exported = 0;
2396 tso->gran.basicblocks = 0;
2397 tso->gran.allocs = 0;
2398 tso->gran.exectime = 0;
2399 tso->gran.fetchtime = 0;
2400 tso->gran.fetchcount = 0;
2401 tso->gran.blocktime = 0;
2402 tso->gran.blockcount = 0;
2403 tso->gran.blockedat = 0;
2404 tso->gran.globalsparks = 0;
2405 tso->gran.localsparks = 0;
2406 if (RtsFlags.GranFlags.Light)
2407 tso->gran.clock = Now; /* local clock */
2409 tso->gran.clock = 0;
2411 IF_DEBUG(gran,printTSO(tso));
2412 #elif defined(PARALLEL_HASKELL)
2414 tso->par.magic = TSO_MAGIC; // debugging only
2416 tso->par.sparkname = 0;
2417 tso->par.startedat = CURRENT_TIME;
2418 tso->par.exported = 0;
2419 tso->par.basicblocks = 0;
2420 tso->par.allocs = 0;
2421 tso->par.exectime = 0;
2422 tso->par.fetchtime = 0;
2423 tso->par.fetchcount = 0;
2424 tso->par.blocktime = 0;
2425 tso->par.blockcount = 0;
2426 tso->par.blockedat = 0;
2427 tso->par.globalsparks = 0;
2428 tso->par.localsparks = 0;
2432 globalGranStats.tot_threads_created++;
2433 globalGranStats.threads_created_on_PE[CurrentProc]++;
2434 globalGranStats.tot_sq_len += spark_queue_len(CurrentProc);
2435 globalGranStats.tot_sq_probes++;
2436 #elif defined(PARALLEL_HASKELL)
2437 // collect parallel global statistics (currently done together with GC stats)
2438 if (RtsFlags.ParFlags.ParStats.Global &&
2439 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
2440 //debugBelch("Creating thread %d @ %11.2f\n", tso->id, usertime());
2441 globalParStats.tot_threads_created++;
2447 sched_belch("==__ schedule: Created TSO %d (%p);",
2448 CurrentProc, tso, tso->id));
2449 #elif defined(PARALLEL_HASKELL)
2450 IF_PAR_DEBUG(verbose,
2451 sched_belch("==__ schedule: Created TSO %d (%p); %d threads active",
2452 (long)tso->id, tso, advisory_thread_count));
2454 IF_DEBUG(scheduler,sched_belch("created thread %ld, stack size = %lx words",
2455 (long)tso->id, (long)tso->stack_size));
2462 all parallel thread creation calls should fall through the following routine.
2465 createThreadFromSpark(rtsSpark spark)
2467 ASSERT(spark != (rtsSpark)NULL);
2468 // JB: TAKE CARE OF THIS COUNTER! BUGGY
2469 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads)
2471 barf("{createSparkThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
2472 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
2473 return END_TSO_QUEUE;
2477 tso = createThread(RtsFlags.GcFlags.initialStkSize);
2478 if (tso==END_TSO_QUEUE)
2479 barf("createSparkThread: Cannot create TSO");
2481 tso->priority = AdvisoryPriority;
2483 pushClosure(tso,spark);
2485 advisory_thread_count++; // JB: TAKE CARE OF THIS COUNTER! BUGGY
2492 Turn a spark into a thread.
2493 ToDo: fix for SMP (needs to acquire SCHED_MUTEX!)
2497 activateSpark (rtsSpark spark)
2501 tso = createSparkThread(spark);
2502 if (RtsFlags.ParFlags.ParStats.Full) {
2503 //ASSERT(run_queue_hd == END_TSO_QUEUE); // I think ...
2504 IF_PAR_DEBUG(verbose,
2505 debugBelch("==^^ activateSpark: turning spark of closure %p (%s) into a thread\n",
2506 (StgClosure *)spark, info_type((StgClosure *)spark)));
2508 // ToDo: fwd info on local/global spark to thread -- HWL
2509 // tso->gran.exported = spark->exported;
2510 // tso->gran.locked = !spark->global;
2511 // tso->gran.sparkname = spark->name;
2517 /* ---------------------------------------------------------------------------
2520 * scheduleThread puts a thread on the head of the runnable queue.
2521 * This will usually be done immediately after a thread is created.
2522 * The caller of scheduleThread must create the thread using e.g.
2523 * createThread and push an appropriate closure
2524 * on this thread's stack before the scheduler is invoked.
2525 * ------------------------------------------------------------------------ */
2528 scheduleThread_(StgTSO *tso)
2530 // The thread goes at the *end* of the run-queue, to avoid possible
2531 // starvation of any threads already on the queue.
2532 APPEND_TO_RUN_QUEUE(tso);
2537 scheduleThread(StgTSO* tso)
2539 ACQUIRE_LOCK(&sched_mutex);
2540 scheduleThread_(tso);
2541 RELEASE_LOCK(&sched_mutex);
2544 #if defined(RTS_SUPPORTS_THREADS)
2545 static Condition bound_cond_cache;
2546 static int bound_cond_cache_full = 0;
2551 scheduleWaitThread(StgTSO* tso, /*[out]*/HaskellObj* ret,
2552 Capability *initialCapability)
2554 // Precondition: sched_mutex must be held
2557 m = stgMallocBytes(sizeof(StgMainThread), "waitThread");
2562 m->link = main_threads;
2564 if (main_threads != NULL) {
2565 main_threads->prev = m;
2569 #if defined(RTS_SUPPORTS_THREADS)
2570 // Allocating a new condition for each thread is expensive, so we
2571 // cache one. This is a pretty feeble hack, but it helps speed up
2572 // consecutive call-ins quite a bit.
2573 if (bound_cond_cache_full) {
2574 m->bound_thread_cond = bound_cond_cache;
2575 bound_cond_cache_full = 0;
2577 initCondition(&m->bound_thread_cond);
2581 /* Put the thread on the main-threads list prior to scheduling the TSO.
2582 Failure to do so introduces a race condition in the MT case (as
2583 identified by Wolfgang Thaller), whereby the new task/OS thread
2584 created by scheduleThread_() would complete prior to the thread
2585 that spawned it managed to put 'itself' on the main-threads list.
2586 The upshot of it all being that the worker thread wouldn't get to
2587 signal the completion of the its work item for the main thread to
2588 see (==> it got stuck waiting.) -- sof 6/02.
2590 IF_DEBUG(scheduler, sched_belch("waiting for thread (%d)", tso->id));
2592 APPEND_TO_RUN_QUEUE(tso);
2593 // NB. Don't call threadRunnable() here, because the thread is
2594 // bound and only runnable by *this* OS thread, so waking up other
2595 // workers will just slow things down.
2597 return waitThread_(m, initialCapability);
2600 /* ---------------------------------------------------------------------------
2603 * Initialise the scheduler. This resets all the queues - if the
2604 * queues contained any threads, they'll be garbage collected at the
2607 * ------------------------------------------------------------------------ */
2615 for (i=0; i<=MAX_PROC; i++) {
2616 run_queue_hds[i] = END_TSO_QUEUE;
2617 run_queue_tls[i] = END_TSO_QUEUE;
2618 blocked_queue_hds[i] = END_TSO_QUEUE;
2619 blocked_queue_tls[i] = END_TSO_QUEUE;
2620 ccalling_threadss[i] = END_TSO_QUEUE;
2621 blackhole_queue[i] = END_TSO_QUEUE;
2622 sleeping_queue = END_TSO_QUEUE;
2625 run_queue_hd = END_TSO_QUEUE;
2626 run_queue_tl = END_TSO_QUEUE;
2627 blocked_queue_hd = END_TSO_QUEUE;
2628 blocked_queue_tl = END_TSO_QUEUE;
2629 blackhole_queue = END_TSO_QUEUE;
2630 sleeping_queue = END_TSO_QUEUE;
2633 suspended_ccalling_threads = END_TSO_QUEUE;
2635 main_threads = NULL;
2636 all_threads = END_TSO_QUEUE;
2641 RtsFlags.ConcFlags.ctxtSwitchTicks =
2642 RtsFlags.ConcFlags.ctxtSwitchTime / TICK_MILLISECS;
2644 #if defined(RTS_SUPPORTS_THREADS)
2645 /* Initialise the mutex and condition variables used by
2647 initMutex(&sched_mutex);
2648 initMutex(&term_mutex);
2651 ACQUIRE_LOCK(&sched_mutex);
2653 /* A capability holds the state a native thread needs in
2654 * order to execute STG code. At least one capability is
2655 * floating around (only SMP builds have more than one).
2659 #if defined(RTS_SUPPORTS_THREADS)
2664 /* eagerly start some extra workers */
2665 startingWorkerThread = RtsFlags.ParFlags.nNodes;
2666 startTasks(RtsFlags.ParFlags.nNodes, taskStart);
2669 #if /* defined(SMP) ||*/ defined(PARALLEL_HASKELL)
2673 RELEASE_LOCK(&sched_mutex);
2677 exitScheduler( void )
2679 interrupted = rtsTrue;
2680 shutting_down_scheduler = rtsTrue;
2681 #if defined(RTS_SUPPORTS_THREADS)
2682 if (threadIsTask(osThreadId())) { taskStop(); }
2687 /* ----------------------------------------------------------------------------
2688 Managing the per-task allocation areas.
2690 Each capability comes with an allocation area. These are
2691 fixed-length block lists into which allocation can be done.
2693 ToDo: no support for two-space collection at the moment???
2694 ------------------------------------------------------------------------- */
2696 static SchedulerStatus
2697 waitThread_(StgMainThread* m, Capability *initialCapability)
2699 SchedulerStatus stat;
2701 // Precondition: sched_mutex must be held.
2702 IF_DEBUG(scheduler, sched_belch("new main thread (%d)", m->tso->id));
2705 /* GranSim specific init */
2706 CurrentTSO = m->tso; // the TSO to run
2707 procStatus[MainProc] = Busy; // status of main PE
2708 CurrentProc = MainProc; // PE to run it on
2709 schedule(m,initialCapability);
2711 schedule(m,initialCapability);
2712 ASSERT(m->stat != NoStatus);
2717 #if defined(RTS_SUPPORTS_THREADS)
2718 // Free the condition variable, returning it to the cache if possible.
2719 if (!bound_cond_cache_full) {
2720 bound_cond_cache = m->bound_thread_cond;
2721 bound_cond_cache_full = 1;
2723 closeCondition(&m->bound_thread_cond);
2727 IF_DEBUG(scheduler, sched_belch("main thread (%d) finished", m->tso->id));
2730 // Postcondition: sched_mutex still held
2734 /* ---------------------------------------------------------------------------
2735 Where are the roots that we know about?
2737 - all the threads on the runnable queue
2738 - all the threads on the blocked queue
2739 - all the threads on the sleeping queue
2740 - all the thread currently executing a _ccall_GC
2741 - all the "main threads"
2743 ------------------------------------------------------------------------ */
2745 /* This has to be protected either by the scheduler monitor, or by the
2746 garbage collection monitor (probably the latter).
2751 GetRoots( evac_fn evac )
2756 for (i=0; i<=RtsFlags.GranFlags.proc; i++) {
2757 if ((run_queue_hds[i] != END_TSO_QUEUE) && ((run_queue_hds[i] != NULL)))
2758 evac((StgClosure **)&run_queue_hds[i]);
2759 if ((run_queue_tls[i] != END_TSO_QUEUE) && ((run_queue_tls[i] != NULL)))
2760 evac((StgClosure **)&run_queue_tls[i]);
2762 if ((blocked_queue_hds[i] != END_TSO_QUEUE) && ((blocked_queue_hds[i] != NULL)))
2763 evac((StgClosure **)&blocked_queue_hds[i]);
2764 if ((blocked_queue_tls[i] != END_TSO_QUEUE) && ((blocked_queue_tls[i] != NULL)))
2765 evac((StgClosure **)&blocked_queue_tls[i]);
2766 if ((ccalling_threadss[i] != END_TSO_QUEUE) && ((ccalling_threadss[i] != NULL)))
2767 evac((StgClosure **)&ccalling_threads[i]);
2774 if (run_queue_hd != END_TSO_QUEUE) {
2775 ASSERT(run_queue_tl != END_TSO_QUEUE);
2776 evac((StgClosure **)&run_queue_hd);
2777 evac((StgClosure **)&run_queue_tl);
2780 if (blocked_queue_hd != END_TSO_QUEUE) {
2781 ASSERT(blocked_queue_tl != END_TSO_QUEUE);
2782 evac((StgClosure **)&blocked_queue_hd);
2783 evac((StgClosure **)&blocked_queue_tl);
2786 if (sleeping_queue != END_TSO_QUEUE) {
2787 evac((StgClosure **)&sleeping_queue);
2791 if (blackhole_queue != END_TSO_QUEUE) {
2792 evac((StgClosure **)&blackhole_queue);
2795 if (suspended_ccalling_threads != END_TSO_QUEUE) {
2796 evac((StgClosure **)&suspended_ccalling_threads);
2799 #if defined(PARALLEL_HASKELL) || defined(GRAN)
2800 markSparkQueue(evac);
2803 #if defined(RTS_USER_SIGNALS)
2804 // mark the signal handlers (signals should be already blocked)
2805 markSignalHandlers(evac);
2809 /* -----------------------------------------------------------------------------
2812 This is the interface to the garbage collector from Haskell land.
2813 We provide this so that external C code can allocate and garbage
2814 collect when called from Haskell via _ccall_GC.
2816 It might be useful to provide an interface whereby the programmer
2817 can specify more roots (ToDo).
2819 This needs to be protected by the GC condition variable above. KH.
2820 -------------------------------------------------------------------------- */
2822 static void (*extra_roots)(evac_fn);
2827 /* Obligated to hold this lock upon entry */
2828 ACQUIRE_LOCK(&sched_mutex);
2829 GarbageCollect(GetRoots,rtsFalse);
2830 RELEASE_LOCK(&sched_mutex);
2834 performMajorGC(void)
2836 ACQUIRE_LOCK(&sched_mutex);
2837 GarbageCollect(GetRoots,rtsTrue);
2838 RELEASE_LOCK(&sched_mutex);
2842 AllRoots(evac_fn evac)
2844 GetRoots(evac); // the scheduler's roots
2845 extra_roots(evac); // the user's roots
2849 performGCWithRoots(void (*get_roots)(evac_fn))
2851 ACQUIRE_LOCK(&sched_mutex);
2852 extra_roots = get_roots;
2853 GarbageCollect(AllRoots,rtsFalse);
2854 RELEASE_LOCK(&sched_mutex);
2857 /* -----------------------------------------------------------------------------
2860 If the thread has reached its maximum stack size, then raise the
2861 StackOverflow exception in the offending thread. Otherwise
2862 relocate the TSO into a larger chunk of memory and adjust its stack
2864 -------------------------------------------------------------------------- */
2867 threadStackOverflow(StgTSO *tso)
2869 nat new_stack_size, stack_words;
2874 IF_DEBUG(sanity,checkTSO(tso));
2875 if (tso->stack_size >= tso->max_stack_size) {
2878 debugBelch("@@ threadStackOverflow of TSO %ld (%p): stack too large (now %ld; max is %ld)\n",
2879 (long)tso->id, tso, (long)tso->stack_size, (long)tso->max_stack_size);
2880 /* If we're debugging, just print out the top of the stack */
2881 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2884 /* Send this thread the StackOverflow exception */
2885 raiseAsync(tso, (StgClosure *)stackOverflow_closure);
2889 /* Try to double the current stack size. If that takes us over the
2890 * maximum stack size for this thread, then use the maximum instead.
2891 * Finally round up so the TSO ends up as a whole number of blocks.
2893 new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
2894 new_tso_size = (lnat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) +
2895 TSO_STRUCT_SIZE)/sizeof(W_);
2896 new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */
2897 new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
2899 IF_DEBUG(scheduler, debugBelch("== sched: increasing stack size from %d words to %d.\n", tso->stack_size, new_stack_size));
2901 dest = (StgTSO *)allocate(new_tso_size);
2902 TICK_ALLOC_TSO(new_stack_size,0);
2904 /* copy the TSO block and the old stack into the new area */
2905 memcpy(dest,tso,TSO_STRUCT_SIZE);
2906 stack_words = tso->stack + tso->stack_size - tso->sp;
2907 new_sp = (P_)dest + new_tso_size - stack_words;
2908 memcpy(new_sp, tso->sp, stack_words * sizeof(W_));
2910 /* relocate the stack pointers... */
2912 dest->stack_size = new_stack_size;
2914 /* Mark the old TSO as relocated. We have to check for relocated
2915 * TSOs in the garbage collector and any primops that deal with TSOs.
2917 * It's important to set the sp value to just beyond the end
2918 * of the stack, so we don't attempt to scavenge any part of the
2921 tso->what_next = ThreadRelocated;
2923 tso->sp = (P_)&(tso->stack[tso->stack_size]);
2924 tso->why_blocked = NotBlocked;
2926 IF_PAR_DEBUG(verbose,
2927 debugBelch("@@ threadStackOverflow of TSO %d (now at %p): stack size increased to %ld\n",
2928 tso->id, tso, tso->stack_size);
2929 /* If we're debugging, just print out the top of the stack */
2930 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2933 IF_DEBUG(sanity,checkTSO(tso));
2935 IF_DEBUG(scheduler,printTSO(dest));
2941 /* ---------------------------------------------------------------------------
2942 Wake up a queue that was blocked on some resource.
2943 ------------------------------------------------------------------------ */
2947 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2950 #elif defined(PARALLEL_HASKELL)
2952 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2954 /* write RESUME events to log file and
2955 update blocked and fetch time (depending on type of the orig closure) */
2956 if (RtsFlags.ParFlags.ParStats.Full) {
2957 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
2958 GR_RESUMEQ, ((StgTSO *)bqe), ((StgTSO *)bqe)->block_info.closure,
2959 0, 0 /* spark_queue_len(ADVISORY_POOL) */);
2960 if (EMPTY_RUN_QUEUE())
2961 emitSchedule = rtsTrue;
2963 switch (get_itbl(node)->type) {
2965 ((StgTSO *)bqe)->par.fetchtime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2970 ((StgTSO *)bqe)->par.blocktime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2977 barf("{unblockOneLocked}Daq Qagh: unexpected closure in blocking queue");
2984 StgBlockingQueueElement *
2985 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2988 PEs node_loc, tso_loc;
2990 node_loc = where_is(node); // should be lifted out of loop
2991 tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2992 tso_loc = where_is((StgClosure *)tso);
2993 if (IS_LOCAL_TO(PROCS(node),tso_loc)) { // TSO is local
2994 /* !fake_fetch => TSO is on CurrentProc is same as IS_LOCAL_TO */
2995 ASSERT(CurrentProc!=node_loc || tso_loc==CurrentProc);
2996 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.lunblocktime;
2997 // insertThread(tso, node_loc);
2998 new_event(tso_loc, tso_loc, CurrentTime[CurrentProc],
3000 tso, node, (rtsSpark*)NULL);
3001 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
3004 } else { // TSO is remote (actually should be FMBQ)
3005 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.mpacktime +
3006 RtsFlags.GranFlags.Costs.gunblocktime +
3007 RtsFlags.GranFlags.Costs.latency;
3008 new_event(tso_loc, CurrentProc, CurrentTime[CurrentProc],
3010 tso, node, (rtsSpark*)NULL);
3011 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
3014 /* the thread-queue-overhead is accounted for in either Resume or UnblockThread */
3016 debugBelch(" %s TSO %d (%p) [PE %d] (block_info.closure=%p) (next=%p) ,",
3017 (node_loc==tso_loc ? "Local" : "Global"),
3018 tso->id, tso, CurrentProc, tso->block_info.closure, tso->link));
3019 tso->block_info.closure = NULL;
3020 IF_DEBUG(scheduler,debugBelch("-- Waking up thread %ld (%p)\n",
3023 #elif defined(PARALLEL_HASKELL)
3024 StgBlockingQueueElement *
3025 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
3027 StgBlockingQueueElement *next;
3029 switch (get_itbl(bqe)->type) {
3031 ASSERT(((StgTSO *)bqe)->why_blocked != NotBlocked);
3032 /* if it's a TSO just push it onto the run_queue */
3034 ((StgTSO *)bqe)->link = END_TSO_QUEUE; // debugging?
3035 APPEND_TO_RUN_QUEUE((StgTSO *)bqe);
3037 unblockCount(bqe, node);
3038 /* reset blocking status after dumping event */
3039 ((StgTSO *)bqe)->why_blocked = NotBlocked;
3043 /* if it's a BLOCKED_FETCH put it on the PendingFetches list */
3045 bqe->link = (StgBlockingQueueElement *)PendingFetches;
3046 PendingFetches = (StgBlockedFetch *)bqe;
3050 /* can ignore this case in a non-debugging setup;
3051 see comments on RBHSave closures above */
3053 /* check that the closure is an RBHSave closure */
3054 ASSERT(get_itbl((StgClosure *)bqe) == &stg_RBH_Save_0_info ||
3055 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_1_info ||
3056 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_2_info);
3060 barf("{unblockOneLocked}Daq Qagh: Unexpected IP (%#lx; %s) in blocking queue at %#lx\n",
3061 get_itbl((StgClosure *)bqe), info_type((StgClosure *)bqe),
3065 IF_PAR_DEBUG(bq, debugBelch(", %p (%s)\n", bqe, info_type((StgClosure*)bqe)));
3069 #else /* !GRAN && !PARALLEL_HASKELL */
3071 unblockOneLocked(StgTSO *tso)
3075 ASSERT(get_itbl(tso)->type == TSO);
3076 ASSERT(tso->why_blocked != NotBlocked);
3077 tso->why_blocked = NotBlocked;
3079 tso->link = END_TSO_QUEUE;
3080 APPEND_TO_RUN_QUEUE(tso);
3082 IF_DEBUG(scheduler,sched_belch("waking up thread %ld", (long)tso->id));
3087 #if defined(GRAN) || defined(PARALLEL_HASKELL)
3088 INLINE_ME StgBlockingQueueElement *
3089 unblockOne(StgBlockingQueueElement *bqe, StgClosure *node)
3091 ACQUIRE_LOCK(&sched_mutex);
3092 bqe = unblockOneLocked(bqe, node);
3093 RELEASE_LOCK(&sched_mutex);
3098 unblockOne(StgTSO *tso)
3100 ACQUIRE_LOCK(&sched_mutex);
3101 tso = unblockOneLocked(tso);
3102 RELEASE_LOCK(&sched_mutex);
3109 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
3111 StgBlockingQueueElement *bqe;
3116 debugBelch("##-_ AwBQ for node %p on PE %d @ %ld by TSO %d (%p): \n", \
3117 node, CurrentProc, CurrentTime[CurrentProc],
3118 CurrentTSO->id, CurrentTSO));
3120 node_loc = where_is(node);
3122 ASSERT(q == END_BQ_QUEUE ||
3123 get_itbl(q)->type == TSO || // q is either a TSO or an RBHSave
3124 get_itbl(q)->type == CONSTR); // closure (type constructor)
3125 ASSERT(is_unique(node));
3127 /* FAKE FETCH: magically copy the node to the tso's proc;
3128 no Fetch necessary because in reality the node should not have been
3129 moved to the other PE in the first place
3131 if (CurrentProc!=node_loc) {
3133 debugBelch("## node %p is on PE %d but CurrentProc is %d (TSO %d); assuming fake fetch and adjusting bitmask (old: %#x)\n",
3134 node, node_loc, CurrentProc, CurrentTSO->id,
3135 // CurrentTSO, where_is(CurrentTSO),
3136 node->header.gran.procs));
3137 node->header.gran.procs = (node->header.gran.procs) | PE_NUMBER(CurrentProc);
3139 debugBelch("## new bitmask of node %p is %#x\n",
3140 node, node->header.gran.procs));
3141 if (RtsFlags.GranFlags.GranSimStats.Global) {
3142 globalGranStats.tot_fake_fetches++;
3147 // ToDo: check: ASSERT(CurrentProc==node_loc);
3148 while (get_itbl(bqe)->type==TSO) { // q != END_TSO_QUEUE) {
3151 bqe points to the current element in the queue
3152 next points to the next element in the queue
3154 //tso = (StgTSO *)bqe; // wastes an assignment to get the type right
3155 //tso_loc = where_is(tso);
3157 bqe = unblockOneLocked(bqe, node);
3160 /* if this is the BQ of an RBH, we have to put back the info ripped out of
3161 the closure to make room for the anchor of the BQ */
3162 if (bqe!=END_BQ_QUEUE) {
3163 ASSERT(get_itbl(node)->type == RBH && get_itbl(bqe)->type == CONSTR);
3165 ASSERT((info_ptr==&RBH_Save_0_info) ||
3166 (info_ptr==&RBH_Save_1_info) ||
3167 (info_ptr==&RBH_Save_2_info));
3169 /* cf. convertToRBH in RBH.c for writing the RBHSave closure */
3170 ((StgRBH *)node)->blocking_queue = (StgBlockingQueueElement *)((StgRBHSave *)bqe)->payload[0];
3171 ((StgRBH *)node)->mut_link = (StgMutClosure *)((StgRBHSave *)bqe)->payload[1];
3174 debugBelch("## Filled in RBH_Save for %p (%s) at end of AwBQ\n",
3175 node, info_type(node)));
3178 /* statistics gathering */
3179 if (RtsFlags.GranFlags.GranSimStats.Global) {
3180 // globalGranStats.tot_bq_processing_time += bq_processing_time;
3181 globalGranStats.tot_bq_len += len; // total length of all bqs awakened
3182 // globalGranStats.tot_bq_len_local += len_local; // same for local TSOs only
3183 globalGranStats.tot_awbq++; // total no. of bqs awakened
3186 debugBelch("## BQ Stats of %p: [%d entries] %s\n",
3187 node, len, (bqe!=END_BQ_QUEUE) ? "RBH" : ""));
3189 #elif defined(PARALLEL_HASKELL)
3191 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
3193 StgBlockingQueueElement *bqe;
3195 ACQUIRE_LOCK(&sched_mutex);
3197 IF_PAR_DEBUG(verbose,
3198 debugBelch("##-_ AwBQ for node %p on [%x]: \n",
3202 if(get_itbl(q)->type == CONSTR || q==END_BQ_QUEUE) {
3203 IF_PAR_DEBUG(verbose, debugBelch("## ... nothing to unblock so lets just return. RFP (BUG?)\n"));
3208 ASSERT(q == END_BQ_QUEUE ||
3209 get_itbl(q)->type == TSO ||
3210 get_itbl(q)->type == BLOCKED_FETCH ||
3211 get_itbl(q)->type == CONSTR);
3214 while (get_itbl(bqe)->type==TSO ||
3215 get_itbl(bqe)->type==BLOCKED_FETCH) {
3216 bqe = unblockOneLocked(bqe, node);
3218 RELEASE_LOCK(&sched_mutex);
3221 #else /* !GRAN && !PARALLEL_HASKELL */
3224 awakenBlockedQueueNoLock(StgTSO *tso)
3226 while (tso != END_TSO_QUEUE) {
3227 tso = unblockOneLocked(tso);
3232 awakenBlockedQueue(StgTSO *tso)
3234 ACQUIRE_LOCK(&sched_mutex);
3235 while (tso != END_TSO_QUEUE) {
3236 tso = unblockOneLocked(tso);
3238 RELEASE_LOCK(&sched_mutex);
3242 /* ---------------------------------------------------------------------------
3244 - usually called inside a signal handler so it mustn't do anything fancy.
3245 ------------------------------------------------------------------------ */
3248 interruptStgRts(void)
3253 /* ToDo: if invoked from a signal handler, this threadRunnable
3254 * only works if there's another thread (not this one) waiting to
3259 /* -----------------------------------------------------------------------------
3262 This is for use when we raise an exception in another thread, which
3264 This has nothing to do with the UnblockThread event in GranSim. -- HWL
3265 -------------------------------------------------------------------------- */
3267 #if defined(GRAN) || defined(PARALLEL_HASKELL)
3269 NB: only the type of the blocking queue is different in GranSim and GUM
3270 the operations on the queue-elements are the same
3271 long live polymorphism!
3273 Locks: sched_mutex is held upon entry and exit.
3277 unblockThread(StgTSO *tso)
3279 StgBlockingQueueElement *t, **last;
3281 switch (tso->why_blocked) {
3284 return; /* not blocked */
3287 // Be careful: nothing to do here! We tell the scheduler that the thread
3288 // is runnable and we leave it to the stack-walking code to abort the
3289 // transaction while unwinding the stack. We should perhaps have a debugging
3290 // test to make sure that this really happens and that the 'zombie' transaction
3291 // does not get committed.
3295 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
3297 StgBlockingQueueElement *last_tso = END_BQ_QUEUE;
3298 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
3300 last = (StgBlockingQueueElement **)&mvar->head;
3301 for (t = (StgBlockingQueueElement *)mvar->head;
3303 last = &t->link, last_tso = t, t = t->link) {
3304 if (t == (StgBlockingQueueElement *)tso) {
3305 *last = (StgBlockingQueueElement *)tso->link;
3306 if (mvar->tail == tso) {
3307 mvar->tail = (StgTSO *)last_tso;
3312 barf("unblockThread (MVAR): TSO not found");
3315 case BlockedOnBlackHole:
3316 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
3318 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
3320 last = &bq->blocking_queue;
3321 for (t = bq->blocking_queue;
3323 last = &t->link, t = t->link) {
3324 if (t == (StgBlockingQueueElement *)tso) {
3325 *last = (StgBlockingQueueElement *)tso->link;
3329 barf("unblockThread (BLACKHOLE): TSO not found");
3332 case BlockedOnException:
3334 StgTSO *target = tso->block_info.tso;
3336 ASSERT(get_itbl(target)->type == TSO);
3338 if (target->what_next == ThreadRelocated) {
3339 target = target->link;
3340 ASSERT(get_itbl(target)->type == TSO);
3343 ASSERT(target->blocked_exceptions != NULL);
3345 last = (StgBlockingQueueElement **)&target->blocked_exceptions;
3346 for (t = (StgBlockingQueueElement *)target->blocked_exceptions;
3348 last = &t->link, t = t->link) {
3349 ASSERT(get_itbl(t)->type == TSO);
3350 if (t == (StgBlockingQueueElement *)tso) {
3351 *last = (StgBlockingQueueElement *)tso->link;
3355 barf("unblockThread (Exception): TSO not found");
3359 case BlockedOnWrite:
3360 #if defined(mingw32_HOST_OS)
3361 case BlockedOnDoProc:
3364 /* take TSO off blocked_queue */
3365 StgBlockingQueueElement *prev = NULL;
3366 for (t = (StgBlockingQueueElement *)blocked_queue_hd; t != END_BQ_QUEUE;
3367 prev = t, t = t->link) {
3368 if (t == (StgBlockingQueueElement *)tso) {
3370 blocked_queue_hd = (StgTSO *)t->link;
3371 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
3372 blocked_queue_tl = END_TSO_QUEUE;
3375 prev->link = t->link;
3376 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
3377 blocked_queue_tl = (StgTSO *)prev;
3380 #if defined(mingw32_HOST_OS)
3381 /* (Cooperatively) signal that the worker thread should abort
3384 abandonWorkRequest(tso->block_info.async_result->reqID);
3389 barf("unblockThread (I/O): TSO not found");
3392 case BlockedOnDelay:
3394 /* take TSO off sleeping_queue */
3395 StgBlockingQueueElement *prev = NULL;
3396 for (t = (StgBlockingQueueElement *)sleeping_queue; t != END_BQ_QUEUE;
3397 prev = t, t = t->link) {
3398 if (t == (StgBlockingQueueElement *)tso) {
3400 sleeping_queue = (StgTSO *)t->link;
3402 prev->link = t->link;
3407 barf("unblockThread (delay): TSO not found");
3411 barf("unblockThread");
3415 tso->link = END_TSO_QUEUE;
3416 tso->why_blocked = NotBlocked;
3417 tso->block_info.closure = NULL;
3418 PUSH_ON_RUN_QUEUE(tso);
3422 unblockThread(StgTSO *tso)
3426 /* To avoid locking unnecessarily. */
3427 if (tso->why_blocked == NotBlocked) {
3431 switch (tso->why_blocked) {
3434 // Be careful: nothing to do here! We tell the scheduler that the thread
3435 // is runnable and we leave it to the stack-walking code to abort the
3436 // transaction while unwinding the stack. We should perhaps have a debugging
3437 // test to make sure that this really happens and that the 'zombie' transaction
3438 // does not get committed.
3442 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
3444 StgTSO *last_tso = END_TSO_QUEUE;
3445 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
3448 for (t = mvar->head; t != END_TSO_QUEUE;
3449 last = &t->link, last_tso = t, t = t->link) {
3452 if (mvar->tail == tso) {
3453 mvar->tail = last_tso;
3458 barf("unblockThread (MVAR): TSO not found");
3461 case BlockedOnBlackHole:
3463 last = &blackhole_queue;
3464 for (t = blackhole_queue; t != END_TSO_QUEUE;
3465 last = &t->link, t = t->link) {
3471 barf("unblockThread (BLACKHOLE): TSO not found");
3474 case BlockedOnException:
3476 StgTSO *target = tso->block_info.tso;
3478 ASSERT(get_itbl(target)->type == TSO);
3480 while (target->what_next == ThreadRelocated) {
3481 target = target->link;
3482 ASSERT(get_itbl(target)->type == TSO);
3485 ASSERT(target->blocked_exceptions != NULL);
3487 last = &target->blocked_exceptions;
3488 for (t = target->blocked_exceptions; t != END_TSO_QUEUE;
3489 last = &t->link, t = t->link) {
3490 ASSERT(get_itbl(t)->type == TSO);
3496 barf("unblockThread (Exception): TSO not found");
3500 case BlockedOnWrite:
3501 #if defined(mingw32_HOST_OS)
3502 case BlockedOnDoProc:
3505 StgTSO *prev = NULL;
3506 for (t = blocked_queue_hd; t != END_TSO_QUEUE;
3507 prev = t, t = t->link) {
3510 blocked_queue_hd = t->link;
3511 if (blocked_queue_tl == t) {
3512 blocked_queue_tl = END_TSO_QUEUE;
3515 prev->link = t->link;
3516 if (blocked_queue_tl == t) {
3517 blocked_queue_tl = prev;
3520 #if defined(mingw32_HOST_OS)
3521 /* (Cooperatively) signal that the worker thread should abort
3524 abandonWorkRequest(tso->block_info.async_result->reqID);
3529 barf("unblockThread (I/O): TSO not found");
3532 case BlockedOnDelay:
3534 StgTSO *prev = NULL;
3535 for (t = sleeping_queue; t != END_TSO_QUEUE;
3536 prev = t, t = t->link) {
3539 sleeping_queue = t->link;
3541 prev->link = t->link;
3546 barf("unblockThread (delay): TSO not found");
3550 barf("unblockThread");
3554 tso->link = END_TSO_QUEUE;
3555 tso->why_blocked = NotBlocked;
3556 tso->block_info.closure = NULL;
3557 APPEND_TO_RUN_QUEUE(tso);
3561 /* -----------------------------------------------------------------------------
3564 * Check the blackhole_queue for threads that can be woken up. We do
3565 * this periodically: before every GC, and whenever the run queue is
3568 * An elegant solution might be to just wake up all the blocked
3569 * threads with awakenBlockedQueue occasionally: they'll go back to
3570 * sleep again if the object is still a BLACKHOLE. Unfortunately this
3571 * doesn't give us a way to tell whether we've actually managed to
3572 * wake up any threads, so we would be busy-waiting.
3574 * -------------------------------------------------------------------------- */
3577 checkBlackHoles( void )
3580 rtsBool any_woke_up = rtsFalse;
3583 IF_DEBUG(scheduler, sched_belch("checking threads blocked on black holes"));
3585 // ASSUMES: sched_mutex
3586 prev = &blackhole_queue;
3587 t = blackhole_queue;
3588 while (t != END_TSO_QUEUE) {
3589 ASSERT(t->why_blocked == BlockedOnBlackHole);
3590 type = get_itbl(t->block_info.closure)->type;
3591 if (type != BLACKHOLE && type != CAF_BLACKHOLE) {
3592 t = unblockOneLocked(t);
3594 any_woke_up = rtsTrue;
3604 /* -----------------------------------------------------------------------------
3607 * The following function implements the magic for raising an
3608 * asynchronous exception in an existing thread.
3610 * We first remove the thread from any queue on which it might be
3611 * blocked. The possible blockages are MVARs and BLACKHOLE_BQs.
3613 * We strip the stack down to the innermost CATCH_FRAME, building
3614 * thunks in the heap for all the active computations, so they can
3615 * be restarted if necessary. When we reach a CATCH_FRAME, we build
3616 * an application of the handler to the exception, and push it on
3617 * the top of the stack.
3619 * How exactly do we save all the active computations? We create an
3620 * AP_STACK for every UpdateFrame on the stack. Entering one of these
3621 * AP_STACKs pushes everything from the corresponding update frame
3622 * upwards onto the stack. (Actually, it pushes everything up to the
3623 * next update frame plus a pointer to the next AP_STACK object.
3624 * Entering the next AP_STACK object pushes more onto the stack until we
3625 * reach the last AP_STACK object - at which point the stack should look
3626 * exactly as it did when we killed the TSO and we can continue
3627 * execution by entering the closure on top of the stack.
3629 * We can also kill a thread entirely - this happens if either (a) the
3630 * exception passed to raiseAsync is NULL, or (b) there's no
3631 * CATCH_FRAME on the stack. In either case, we strip the entire
3632 * stack and replace the thread with a zombie.
3634 * Locks: sched_mutex held upon entry nor exit.
3636 * -------------------------------------------------------------------------- */
3639 deleteThread(StgTSO *tso)
3641 if (tso->why_blocked != BlockedOnCCall &&
3642 tso->why_blocked != BlockedOnCCall_NoUnblockExc) {
3643 raiseAsync(tso,NULL);
3647 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
3649 deleteThreadImmediately(StgTSO *tso)
3650 { // for forkProcess only:
3651 // delete thread without giving it a chance to catch the KillThread exception
3653 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
3657 if (tso->why_blocked != BlockedOnCCall &&
3658 tso->why_blocked != BlockedOnCCall_NoUnblockExc) {
3662 tso->what_next = ThreadKilled;
3667 raiseAsyncWithLock(StgTSO *tso, StgClosure *exception)
3669 /* When raising async exs from contexts where sched_mutex isn't held;
3670 use raiseAsyncWithLock(). */
3671 ACQUIRE_LOCK(&sched_mutex);
3672 raiseAsync(tso,exception);
3673 RELEASE_LOCK(&sched_mutex);
3677 raiseAsync(StgTSO *tso, StgClosure *exception)
3679 raiseAsync_(tso, exception, rtsFalse);
3683 raiseAsync_(StgTSO *tso, StgClosure *exception, rtsBool stop_at_atomically)
3685 StgRetInfoTable *info;
3688 // Thread already dead?
3689 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
3694 sched_belch("raising exception in thread %ld.", (long)tso->id));
3696 // Remove it from any blocking queues
3701 // The stack freezing code assumes there's a closure pointer on
3702 // the top of the stack, so we have to arrange that this is the case...
3704 if (sp[0] == (W_)&stg_enter_info) {
3708 sp[0] = (W_)&stg_dummy_ret_closure;
3714 // 1. Let the top of the stack be the "current closure"
3716 // 2. Walk up the stack until we find either an UPDATE_FRAME or a
3719 // 3. If it's an UPDATE_FRAME, then make an AP_STACK containing the
3720 // current closure applied to the chunk of stack up to (but not
3721 // including) the update frame. This closure becomes the "current
3722 // closure". Go back to step 2.
3724 // 4. If it's a CATCH_FRAME, then leave the exception handler on
3725 // top of the stack applied to the exception.
3727 // 5. If it's a STOP_FRAME, then kill the thread.
3729 // NB: if we pass an ATOMICALLY_FRAME then abort the associated
3736 info = get_ret_itbl((StgClosure *)frame);
3738 while (info->i.type != UPDATE_FRAME
3739 && (info->i.type != CATCH_FRAME || exception == NULL)
3740 && info->i.type != STOP_FRAME
3741 && (info->i.type != ATOMICALLY_FRAME || stop_at_atomically == rtsFalse))
3743 if (info->i.type == CATCH_RETRY_FRAME || info->i.type == ATOMICALLY_FRAME) {
3744 // IF we find an ATOMICALLY_FRAME then we abort the
3745 // current transaction and propagate the exception. In
3746 // this case (unlike ordinary exceptions) we do not care
3747 // whether the transaction is valid or not because its
3748 // possible validity cannot have caused the exception
3749 // and will not be visible after the abort.
3751 debugBelch("Found atomically block delivering async exception\n"));
3752 stmAbortTransaction(tso -> trec);
3753 tso -> trec = stmGetEnclosingTRec(tso -> trec);
3755 frame += stack_frame_sizeW((StgClosure *)frame);
3756 info = get_ret_itbl((StgClosure *)frame);
3759 switch (info->i.type) {
3761 case ATOMICALLY_FRAME:
3762 ASSERT(stop_at_atomically);
3763 ASSERT(stmGetEnclosingTRec(tso->trec) == NO_TREC);
3764 stmCondemnTransaction(tso -> trec);
3768 // R1 is not a register: the return convention for IO in
3769 // this case puts the return value on the stack, so we
3770 // need to set up the stack to return to the atomically
3771 // frame properly...
3772 tso->sp = frame - 2;
3773 tso->sp[1] = (StgWord) &stg_NO_FINALIZER_closure; // why not?
3774 tso->sp[0] = (StgWord) &stg_ut_1_0_unreg_info;
3776 tso->what_next = ThreadRunGHC;
3780 // If we find a CATCH_FRAME, and we've got an exception to raise,
3781 // then build the THUNK raise(exception), and leave it on
3782 // top of the CATCH_FRAME ready to enter.
3786 StgCatchFrame *cf = (StgCatchFrame *)frame;
3790 // we've got an exception to raise, so let's pass it to the
3791 // handler in this frame.
3793 raise = (StgThunk *)allocate(sizeofW(StgThunk)+1);
3794 TICK_ALLOC_SE_THK(1,0);
3795 SET_HDR(raise,&stg_raise_info,cf->header.prof.ccs);
3796 raise->payload[0] = exception;
3798 // throw away the stack from Sp up to the CATCH_FRAME.
3802 /* Ensure that async excpetions are blocked now, so we don't get
3803 * a surprise exception before we get around to executing the
3806 if (tso->blocked_exceptions == NULL) {
3807 tso->blocked_exceptions = END_TSO_QUEUE;
3810 /* Put the newly-built THUNK on top of the stack, ready to execute
3811 * when the thread restarts.
3814 sp[-1] = (W_)&stg_enter_info;
3816 tso->what_next = ThreadRunGHC;
3817 IF_DEBUG(sanity, checkTSO(tso));
3826 // First build an AP_STACK consisting of the stack chunk above the
3827 // current update frame, with the top word on the stack as the
3830 words = frame - sp - 1;
3831 ap = (StgAP_STACK *)allocate(AP_STACK_sizeW(words));
3834 ap->fun = (StgClosure *)sp[0];
3836 for(i=0; i < (nat)words; ++i) {
3837 ap->payload[i] = (StgClosure *)*sp++;
3840 SET_HDR(ap,&stg_AP_STACK_info,
3841 ((StgClosure *)frame)->header.prof.ccs /* ToDo */);
3842 TICK_ALLOC_UP_THK(words+1,0);
3845 debugBelch("sched: Updating ");
3846 printPtr((P_)((StgUpdateFrame *)frame)->updatee);
3847 debugBelch(" with ");
3848 printObj((StgClosure *)ap);
3851 // Replace the updatee with an indirection - happily
3852 // this will also wake up any threads currently
3853 // waiting on the result.
3855 // Warning: if we're in a loop, more than one update frame on
3856 // the stack may point to the same object. Be careful not to
3857 // overwrite an IND_OLDGEN in this case, because we'll screw
3858 // up the mutable lists. To be on the safe side, don't
3859 // overwrite any kind of indirection at all. See also
3860 // threadSqueezeStack in GC.c, where we have to make a similar
3863 if (!closure_IND(((StgUpdateFrame *)frame)->updatee)) {
3864 // revert the black hole
3865 UPD_IND_NOLOCK(((StgUpdateFrame *)frame)->updatee,
3868 sp += sizeofW(StgUpdateFrame) - 1;
3869 sp[0] = (W_)ap; // push onto stack
3874 // We've stripped the entire stack, the thread is now dead.
3875 sp += sizeofW(StgStopFrame);
3876 tso->what_next = ThreadKilled;
3887 /* -----------------------------------------------------------------------------
3888 raiseExceptionHelper
3890 This function is called by the raise# primitve, just so that we can
3891 move some of the tricky bits of raising an exception from C-- into
3892 C. Who knows, it might be a useful re-useable thing here too.
3893 -------------------------------------------------------------------------- */
3896 raiseExceptionHelper (StgTSO *tso, StgClosure *exception)
3898 StgThunk *raise_closure = NULL;
3900 StgRetInfoTable *info;
3902 // This closure represents the expression 'raise# E' where E
3903 // is the exception raise. It is used to overwrite all the
3904 // thunks which are currently under evaluataion.
3908 // LDV profiling: stg_raise_info has THUNK as its closure
3909 // type. Since a THUNK takes at least MIN_UPD_SIZE words in its
3910 // payload, MIN_UPD_SIZE is more approprate than 1. It seems that
3911 // 1 does not cause any problem unless profiling is performed.
3912 // However, when LDV profiling goes on, we need to linearly scan
3913 // small object pool, where raise_closure is stored, so we should
3914 // use MIN_UPD_SIZE.
3916 // raise_closure = (StgClosure *)RET_STGCALL1(P_,allocate,
3917 // sizeofW(StgClosure)+1);
3921 // Walk up the stack, looking for the catch frame. On the way,
3922 // we update any closures pointed to from update frames with the
3923 // raise closure that we just built.
3927 info = get_ret_itbl((StgClosure *)p);
3928 next = p + stack_frame_sizeW((StgClosure *)p);
3929 switch (info->i.type) {
3932 // Only create raise_closure if we need to.
3933 if (raise_closure == NULL) {
3935 (StgThunk *)allocate(sizeofW(StgThunk)+MIN_UPD_SIZE);
3936 SET_HDR(raise_closure, &stg_raise_info, CCCS);
3937 raise_closure->payload[0] = exception;
3939 UPD_IND(((StgUpdateFrame *)p)->updatee,(StgClosure *)raise_closure);
3943 case ATOMICALLY_FRAME:
3944 IF_DEBUG(stm, debugBelch("Found ATOMICALLY_FRAME at %p\n", p));
3946 return ATOMICALLY_FRAME;
3952 case CATCH_STM_FRAME:
3953 IF_DEBUG(stm, debugBelch("Found CATCH_STM_FRAME at %p\n", p));
3955 return CATCH_STM_FRAME;
3961 case CATCH_RETRY_FRAME:
3970 /* -----------------------------------------------------------------------------
3971 findRetryFrameHelper
3973 This function is called by the retry# primitive. It traverses the stack
3974 leaving tso->sp referring to the frame which should handle the retry.
3976 This should either be a CATCH_RETRY_FRAME (if the retry# is within an orElse#)
3977 or should be a ATOMICALLY_FRAME (if the retry# reaches the top level).
3979 We skip CATCH_STM_FRAMEs because retries are not considered to be exceptions,
3980 despite the similar implementation.
3982 We should not expect to see CATCH_FRAME or STOP_FRAME because those should
3983 not be created within memory transactions.
3984 -------------------------------------------------------------------------- */
3987 findRetryFrameHelper (StgTSO *tso)
3990 StgRetInfoTable *info;
3994 info = get_ret_itbl((StgClosure *)p);
3995 next = p + stack_frame_sizeW((StgClosure *)p);
3996 switch (info->i.type) {
3998 case ATOMICALLY_FRAME:
3999 IF_DEBUG(stm, debugBelch("Found ATOMICALLY_FRAME at %p during retrry\n", p));
4001 return ATOMICALLY_FRAME;
4003 case CATCH_RETRY_FRAME:
4004 IF_DEBUG(stm, debugBelch("Found CATCH_RETRY_FRAME at %p during retrry\n", p));
4006 return CATCH_RETRY_FRAME;
4008 case CATCH_STM_FRAME:
4010 ASSERT(info->i.type != CATCH_FRAME);
4011 ASSERT(info->i.type != STOP_FRAME);
4018 /* -----------------------------------------------------------------------------
4019 resurrectThreads is called after garbage collection on the list of
4020 threads found to be garbage. Each of these threads will be woken
4021 up and sent a signal: BlockedOnDeadMVar if the thread was blocked
4022 on an MVar, or NonTermination if the thread was blocked on a Black
4025 Locks: sched_mutex isn't held upon entry nor exit.
4026 -------------------------------------------------------------------------- */
4029 resurrectThreads( StgTSO *threads )
4033 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
4034 next = tso->global_link;
4035 tso->global_link = all_threads;
4037 IF_DEBUG(scheduler, sched_belch("resurrecting thread %d", tso->id));
4039 switch (tso->why_blocked) {
4041 case BlockedOnException:
4042 /* Called by GC - sched_mutex lock is currently held. */
4043 raiseAsync(tso,(StgClosure *)BlockedOnDeadMVar_closure);
4045 case BlockedOnBlackHole:
4046 raiseAsync(tso,(StgClosure *)NonTermination_closure);
4049 raiseAsync(tso,(StgClosure *)BlockedIndefinitely_closure);
4052 /* This might happen if the thread was blocked on a black hole
4053 * belonging to a thread that we've just woken up (raiseAsync
4054 * can wake up threads, remember...).
4058 barf("resurrectThreads: thread blocked in a strange way");
4063 /* ----------------------------------------------------------------------------
4064 * Debugging: why is a thread blocked
4065 * [Also provides useful information when debugging threaded programs
4066 * at the Haskell source code level, so enable outside of DEBUG. --sof 7/02]
4067 ------------------------------------------------------------------------- */
4070 printThreadBlockage(StgTSO *tso)
4072 switch (tso->why_blocked) {
4074 debugBelch("is blocked on read from fd %ld", tso->block_info.fd);
4076 case BlockedOnWrite:
4077 debugBelch("is blocked on write to fd %ld", tso->block_info.fd);
4079 #if defined(mingw32_HOST_OS)
4080 case BlockedOnDoProc:
4081 debugBelch("is blocked on proc (request: %ld)", tso->block_info.async_result->reqID);
4084 case BlockedOnDelay:
4085 debugBelch("is blocked until %ld", tso->block_info.target);
4088 debugBelch("is blocked on an MVar");
4090 case BlockedOnException:
4091 debugBelch("is blocked on delivering an exception to thread %d",
4092 tso->block_info.tso->id);
4094 case BlockedOnBlackHole:
4095 debugBelch("is blocked on a black hole");
4098 debugBelch("is not blocked");
4100 #if defined(PARALLEL_HASKELL)
4102 debugBelch("is blocked on global address; local FM_BQ is %p (%s)",
4103 tso->block_info.closure, info_type(tso->block_info.closure));
4105 case BlockedOnGA_NoSend:
4106 debugBelch("is blocked on global address (no send); local FM_BQ is %p (%s)",
4107 tso->block_info.closure, info_type(tso->block_info.closure));
4110 case BlockedOnCCall:
4111 debugBelch("is blocked on an external call");
4113 case BlockedOnCCall_NoUnblockExc:
4114 debugBelch("is blocked on an external call (exceptions were already blocked)");
4117 debugBelch("is blocked on an STM operation");
4120 barf("printThreadBlockage: strange tso->why_blocked: %d for TSO %d (%d)",
4121 tso->why_blocked, tso->id, tso);
4126 printThreadStatus(StgTSO *tso)
4128 switch (tso->what_next) {
4130 debugBelch("has been killed");
4132 case ThreadComplete:
4133 debugBelch("has completed");
4136 printThreadBlockage(tso);
4141 printAllThreads(void)
4146 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
4147 ullong_format_string(TIME_ON_PROC(CurrentProc),
4148 time_string, rtsFalse/*no commas!*/);
4150 debugBelch("all threads at [%s]:\n", time_string);
4151 # elif defined(PARALLEL_HASKELL)
4152 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
4153 ullong_format_string(CURRENT_TIME,
4154 time_string, rtsFalse/*no commas!*/);
4156 debugBelch("all threads at [%s]:\n", time_string);
4158 debugBelch("all threads:\n");
4161 for (t = all_threads; t != END_TSO_QUEUE; ) {
4162 debugBelch("\tthread %d @ %p ", t->id, (void *)t);
4165 void *label = lookupThreadLabel(t->id);
4166 if (label) debugBelch("[\"%s\"] ",(char *)label);
4169 if (t->what_next == ThreadRelocated) {
4170 debugBelch("has been relocated...\n");
4173 printThreadStatus(t);
4183 Print a whole blocking queue attached to node (debugging only).
4185 # if defined(PARALLEL_HASKELL)
4187 print_bq (StgClosure *node)
4189 StgBlockingQueueElement *bqe;
4193 debugBelch("## BQ of closure %p (%s): ",
4194 node, info_type(node));
4196 /* should cover all closures that may have a blocking queue */
4197 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
4198 get_itbl(node)->type == FETCH_ME_BQ ||
4199 get_itbl(node)->type == RBH ||
4200 get_itbl(node)->type == MVAR);
4202 ASSERT(node!=(StgClosure*)NULL); // sanity check
4204 print_bqe(((StgBlockingQueue*)node)->blocking_queue);
4208 Print a whole blocking queue starting with the element bqe.
4211 print_bqe (StgBlockingQueueElement *bqe)
4216 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
4218 for (end = (bqe==END_BQ_QUEUE);
4219 !end; // iterate until bqe points to a CONSTR
4220 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE),
4221 bqe = end ? END_BQ_QUEUE : bqe->link) {
4222 ASSERT(bqe != END_BQ_QUEUE); // sanity check
4223 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
4224 /* types of closures that may appear in a blocking queue */
4225 ASSERT(get_itbl(bqe)->type == TSO ||
4226 get_itbl(bqe)->type == BLOCKED_FETCH ||
4227 get_itbl(bqe)->type == CONSTR);
4228 /* only BQs of an RBH end with an RBH_Save closure */
4229 //ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
4231 switch (get_itbl(bqe)->type) {
4233 debugBelch(" TSO %u (%x),",
4234 ((StgTSO *)bqe)->id, ((StgTSO *)bqe));
4237 debugBelch(" BF (node=%p, ga=((%x, %d, %x)),",
4238 ((StgBlockedFetch *)bqe)->node,
4239 ((StgBlockedFetch *)bqe)->ga.payload.gc.gtid,
4240 ((StgBlockedFetch *)bqe)->ga.payload.gc.slot,
4241 ((StgBlockedFetch *)bqe)->ga.weight);
4244 debugBelch(" %s (IP %p),",
4245 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
4246 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
4247 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
4248 "RBH_Save_?"), get_itbl(bqe));
4251 barf("Unexpected closure type %s in blocking queue", // of %p (%s)",
4252 info_type((StgClosure *)bqe)); // , node, info_type(node));
4258 # elif defined(GRAN)
4260 print_bq (StgClosure *node)
4262 StgBlockingQueueElement *bqe;
4263 PEs node_loc, tso_loc;
4266 /* should cover all closures that may have a blocking queue */
4267 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
4268 get_itbl(node)->type == FETCH_ME_BQ ||
4269 get_itbl(node)->type == RBH);
4271 ASSERT(node!=(StgClosure*)NULL); // sanity check
4272 node_loc = where_is(node);
4274 debugBelch("## BQ of closure %p (%s) on [PE %d]: ",
4275 node, info_type(node), node_loc);
4278 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
4280 for (bqe = ((StgBlockingQueue*)node)->blocking_queue, end = (bqe==END_BQ_QUEUE);
4281 !end; // iterate until bqe points to a CONSTR
4282 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) {
4283 ASSERT(bqe != END_BQ_QUEUE); // sanity check
4284 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
4285 /* types of closures that may appear in a blocking queue */
4286 ASSERT(get_itbl(bqe)->type == TSO ||
4287 get_itbl(bqe)->type == CONSTR);
4288 /* only BQs of an RBH end with an RBH_Save closure */
4289 ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
4291 tso_loc = where_is((StgClosure *)bqe);
4292 switch (get_itbl(bqe)->type) {
4294 debugBelch(" TSO %d (%p) on [PE %d],",
4295 ((StgTSO *)bqe)->id, (StgTSO *)bqe, tso_loc);
4298 debugBelch(" %s (IP %p),",
4299 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
4300 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
4301 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
4302 "RBH_Save_?"), get_itbl(bqe));
4305 barf("Unexpected closure type %s in blocking queue of %p (%s)",
4306 info_type((StgClosure *)bqe), node, info_type(node));
4314 #if defined(PARALLEL_HASKELL)
4321 for (i=0, tso=run_queue_hd;
4322 tso != END_TSO_QUEUE;
4331 sched_belch(char *s, ...)
4335 #ifdef RTS_SUPPORTS_THREADS
4336 debugBelch("sched (task %p): ", osThreadId());
4337 #elif defined(PARALLEL_HASKELL)
4340 debugBelch("sched: ");