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 defined(RTS_SUPPORTS_THREADS)
462 // Yield the capability to higher-priority tasks if necessary.
465 yieldCapability(&cap);
468 // If we do not currently hold a capability, we wait for one
471 waitForCapability(&sched_mutex, &cap,
472 mainThread ? &mainThread->bound_thread_cond : NULL);
475 // We now have a capability...
478 #if 0 /* extra sanity checking */
481 for (m = main_threads; m != NULL; m = m->link) {
482 ASSERT(get_itbl(m->tso)->type == TSO);
487 // Check whether we have re-entered the RTS from Haskell without
488 // going via suspendThread()/resumeThread (i.e. a 'safe' foreign
490 if (cap->r.rInHaskell) {
491 errorBelch("schedule: re-entered unsafely.\n"
492 " Perhaps a 'foreign import unsafe' should be 'safe'?");
497 // Test for interruption. If interrupted==rtsTrue, then either
498 // we received a keyboard interrupt (^C), or the scheduler is
499 // trying to shut down all the tasks (shutting_down_scheduler) in
503 if (shutting_down_scheduler) {
504 IF_DEBUG(scheduler, sched_belch("shutting down"));
505 releaseCapability(cap);
507 mainThread->stat = Interrupted;
508 mainThread->ret = NULL;
512 IF_DEBUG(scheduler, sched_belch("interrupted"));
517 #if defined(not_yet) && defined(SMP)
519 // Top up the run queue from our spark pool. We try to make the
520 // number of threads in the run queue equal to the number of
521 // free capabilities.
525 if (EMPTY_RUN_QUEUE()) {
526 spark = findSpark(rtsFalse);
528 break; /* no more sparks in the pool */
530 createSparkThread(spark);
532 sched_belch("==^^ turning spark of closure %p into a thread",
533 (StgClosure *)spark));
539 scheduleStartSignalHandlers();
541 // Only check the black holes here if we've nothing else to do.
542 // During normal execution, the black hole list only gets checked
543 // at GC time, to avoid repeatedly traversing this possibly long
544 // list each time around the scheduler.
545 if (EMPTY_RUN_QUEUE()) { scheduleCheckBlackHoles(); }
547 scheduleCheckBlockedThreads();
549 scheduleDetectDeadlock();
551 // Normally, the only way we can get here with no threads to
552 // run is if a keyboard interrupt received during
553 // scheduleCheckBlockedThreads() or scheduleDetectDeadlock().
554 // Additionally, it is not fatal for the
555 // threaded RTS to reach here with no threads to run.
557 // win32: might be here due to awaitEvent() being abandoned
558 // as a result of a console event having been delivered.
559 if ( EMPTY_RUN_QUEUE() ) {
560 #if !defined(RTS_SUPPORTS_THREADS) && !defined(mingw32_HOST_OS)
563 continue; // nothing to do
566 #if defined(PARALLEL_HASKELL)
567 scheduleSendPendingMessages();
568 if (EMPTY_RUN_QUEUE() && scheduleActivateSpark())
572 ASSERT(next_fish_to_send_at==0); // i.e. no delayed fishes left!
575 /* If we still have no work we need to send a FISH to get a spark
577 if (EMPTY_RUN_QUEUE()) {
578 if (!scheduleGetRemoteWork(&receivedFinish)) continue;
579 ASSERT(rtsFalse); // should not happen at the moment
581 // from here: non-empty run queue.
582 // TODO: merge above case with this, only one call processMessages() !
583 if (PacketsWaiting()) { /* process incoming messages, if
584 any pending... only in else
585 because getRemoteWork waits for
587 receivedFinish = processMessages();
592 scheduleProcessEvent(event);
596 // Get a thread to run
598 ASSERT(run_queue_hd != END_TSO_QUEUE);
601 #if defined(GRAN) || defined(PAR)
602 scheduleGranParReport(); // some kind of debuging output
604 // Sanity check the thread we're about to run. This can be
605 // expensive if there is lots of thread switching going on...
606 IF_DEBUG(sanity,checkTSO(t));
609 #if defined(RTS_SUPPORTS_THREADS)
610 // Check whether we can run this thread in the current task.
611 // If not, we have to pass our capability to the right task.
613 StgMainThread *m = t->main;
620 sched_belch("### Running thread %d in bound thread", t->id));
621 // yes, the Haskell thread is bound to the current native thread
626 sched_belch("### thread %d bound to another OS thread", t->id));
627 // no, bound to a different Haskell thread: pass to that thread
628 PUSH_ON_RUN_QUEUE(t);
629 passCapability(&m->bound_thread_cond);
635 if(mainThread != NULL)
636 // The thread we want to run is bound.
639 sched_belch("### this OS thread cannot run thread %d", t->id));
640 // no, the current native thread is bound to a different
641 // Haskell thread, so pass it to any worker thread
642 PUSH_ON_RUN_QUEUE(t);
643 passCapabilityToWorker();
650 cap->r.rCurrentTSO = t;
652 /* context switches are now initiated by the timer signal, unless
653 * the user specified "context switch as often as possible", with
656 if ((RtsFlags.ConcFlags.ctxtSwitchTicks == 0
657 && (run_queue_hd != END_TSO_QUEUE
658 || blocked_queue_hd != END_TSO_QUEUE
659 || sleeping_queue != END_TSO_QUEUE)))
664 RELEASE_LOCK(&sched_mutex);
666 IF_DEBUG(scheduler, sched_belch("-->> running thread %ld %s ...",
667 (long)t->id, whatNext_strs[t->what_next]));
669 #if defined(PROFILING)
670 startHeapProfTimer();
673 // ----------------------------------------------------------------------
674 // Run the current thread
676 prev_what_next = t->what_next;
678 errno = t->saved_errno;
679 cap->r.rInHaskell = rtsTrue;
681 recent_activity = ACTIVITY_YES;
683 switch (prev_what_next) {
687 /* Thread already finished, return to scheduler. */
688 ret = ThreadFinished;
692 ret = StgRun((StgFunPtr) stg_returnToStackTop, &cap->r);
695 case ThreadInterpret:
696 ret = interpretBCO(cap);
700 barf("schedule: invalid what_next field");
704 // in SMP mode, we might return with a different capability than
705 // we started with, if the Haskell thread made a foreign call. So
706 // let's find out what our current Capability is:
707 cap = myCapability();
710 // We have run some Haskell code: there might be blackhole-blocked
711 // threads to wake up now.
712 if ( blackhole_queue != END_TSO_QUEUE ) {
713 blackholes_need_checking = rtsTrue;
716 cap->r.rInHaskell = rtsFalse;
718 // The TSO might have moved, eg. if it re-entered the RTS and a GC
719 // happened. So find the new location:
720 t = cap->r.rCurrentTSO;
722 // And save the current errno in this thread.
723 t->saved_errno = errno;
725 // ----------------------------------------------------------------------
727 /* Costs for the scheduler are assigned to CCS_SYSTEM */
728 #if defined(PROFILING)
733 ACQUIRE_LOCK(&sched_mutex);
735 #if defined(RTS_SUPPORTS_THREADS)
736 IF_DEBUG(scheduler,debugBelch("sched (task %p): ", osThreadId()););
737 #elif !defined(GRAN) && !defined(PARALLEL_HASKELL)
738 IF_DEBUG(scheduler,debugBelch("sched: "););
741 schedulePostRunThread();
743 ready_to_gc = rtsFalse;
747 ready_to_gc = scheduleHandleHeapOverflow(cap,t);
751 scheduleHandleStackOverflow(t);
755 if (scheduleHandleYield(t, prev_what_next)) {
756 // shortcut for switching between compiler/interpreter:
762 scheduleHandleThreadBlocked(t);
766 if (scheduleHandleThreadFinished(mainThread, cap, t)) return;;
770 barf("schedule: invalid thread return code %d", (int)ret);
773 if (scheduleDoHeapProfile(ready_to_gc)) { ready_to_gc = rtsFalse; }
774 if (ready_to_gc) { scheduleDoGC(cap); }
775 } /* end of while() */
777 IF_PAR_DEBUG(verbose,
778 debugBelch("== Leaving schedule() after having received Finish\n"));
781 /* ----------------------------------------------------------------------------
782 * Setting up the scheduler loop
783 * ASSUMES: sched_mutex
784 * ------------------------------------------------------------------------- */
787 schedulePreLoop(void)
790 /* set up first event to get things going */
791 /* ToDo: assign costs for system setup and init MainTSO ! */
792 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
794 CurrentTSO, (StgClosure*)NULL, (rtsSpark*)NULL);
797 debugBelch("GRAN: Init CurrentTSO (in schedule) = %p\n",
799 G_TSO(CurrentTSO, 5));
801 if (RtsFlags.GranFlags.Light) {
802 /* Save current time; GranSim Light only */
803 CurrentTSO->gran.clock = CurrentTime[CurrentProc];
808 /* ----------------------------------------------------------------------------
809 * Start any pending signal handlers
810 * ASSUMES: sched_mutex
811 * ------------------------------------------------------------------------- */
814 scheduleStartSignalHandlers(void)
816 #if defined(RTS_USER_SIGNALS) && !defined(RTS_SUPPORTS_THREADS)
817 if (signals_pending()) {
818 RELEASE_LOCK(&sched_mutex); /* ToDo: kill */
819 startSignalHandlers();
820 ACQUIRE_LOCK(&sched_mutex);
825 /* ----------------------------------------------------------------------------
826 * Check for blocked threads that can be woken up.
827 * ASSUMES: sched_mutex
828 * ------------------------------------------------------------------------- */
831 scheduleCheckBlockedThreads(void)
834 // Check whether any waiting threads need to be woken up. If the
835 // run queue is empty, and there are no other tasks running, we
836 // can wait indefinitely for something to happen.
838 if ( !EMPTY_QUEUE(blocked_queue_hd) || !EMPTY_QUEUE(sleeping_queue) )
840 #if defined(RTS_SUPPORTS_THREADS)
841 // We shouldn't be here...
842 barf("schedule: awaitEvent() in threaded RTS");
844 awaitEvent( EMPTY_RUN_QUEUE() && !blackholes_need_checking );
849 /* ----------------------------------------------------------------------------
850 * Check for threads blocked on BLACKHOLEs that can be woken up
851 * ASSUMES: sched_mutex
852 * ------------------------------------------------------------------------- */
854 scheduleCheckBlackHoles( void )
856 if ( blackholes_need_checking )
859 blackholes_need_checking = rtsFalse;
863 /* ----------------------------------------------------------------------------
864 * Detect deadlock conditions and attempt to resolve them.
865 * ASSUMES: sched_mutex
866 * ------------------------------------------------------------------------- */
869 scheduleDetectDeadlock(void)
872 #if defined(PARALLEL_HASKELL)
873 // ToDo: add deadlock detection in GUM (similar to SMP) -- HWL
878 * Detect deadlock: when we have no threads to run, there are no
879 * threads blocked, waiting for I/O, or sleeping, and all the
880 * other tasks are waiting for work, we must have a deadlock of
883 if ( EMPTY_THREAD_QUEUES() )
885 #if defined(RTS_SUPPORTS_THREADS)
887 * In the threaded RTS, we only check for deadlock if there
888 * has been no activity in a complete timeslice. This means
889 * we won't eagerly start a full GC just because we don't have
890 * any threads to run currently.
892 if (recent_activity != ACTIVITY_INACTIVE) return;
895 IF_DEBUG(scheduler, sched_belch("deadlocked, forcing major GC..."));
897 // Garbage collection can release some new threads due to
898 // either (a) finalizers or (b) threads resurrected because
899 // they are unreachable and will therefore be sent an
900 // exception. Any threads thus released will be immediately
903 GarbageCollect(GetRoots,rtsTrue);
904 recent_activity = ACTIVITY_DONE_GC;
905 if ( !EMPTY_RUN_QUEUE() ) return;
907 #if defined(RTS_USER_SIGNALS) && !defined(RTS_SUPPORTS_THREADS)
908 /* If we have user-installed signal handlers, then wait
909 * for signals to arrive rather then bombing out with a
912 if ( anyUserHandlers() ) {
914 sched_belch("still deadlocked, waiting for signals..."));
918 if (signals_pending()) {
919 RELEASE_LOCK(&sched_mutex);
920 startSignalHandlers();
921 ACQUIRE_LOCK(&sched_mutex);
924 // either we have threads to run, or we were interrupted:
925 ASSERT(!EMPTY_RUN_QUEUE() || interrupted);
929 #if !defined(RTS_SUPPORTS_THREADS)
930 /* Probably a real deadlock. Send the current main thread the
931 * Deadlock exception (or in the SMP build, send *all* main
932 * threads the deadlock exception, since none of them can make
938 switch (m->tso->why_blocked) {
940 case BlockedOnBlackHole:
941 case BlockedOnException:
943 raiseAsync(m->tso, (StgClosure *)NonTermination_closure);
946 barf("deadlock: main thread blocked in a strange way");
953 /* ----------------------------------------------------------------------------
954 * Process an event (GRAN only)
955 * ------------------------------------------------------------------------- */
959 scheduleProcessEvent(rtsEvent *event)
963 if (RtsFlags.GranFlags.Light)
964 GranSimLight_enter_system(event, &ActiveTSO); // adjust ActiveTSO etc
966 /* adjust time based on time-stamp */
967 if (event->time > CurrentTime[CurrentProc] &&
968 event->evttype != ContinueThread)
969 CurrentTime[CurrentProc] = event->time;
971 /* Deal with the idle PEs (may issue FindWork or MoveSpark events) */
972 if (!RtsFlags.GranFlags.Light)
975 IF_DEBUG(gran, debugBelch("GRAN: switch by event-type\n"));
977 /* main event dispatcher in GranSim */
978 switch (event->evttype) {
979 /* Should just be continuing execution */
981 IF_DEBUG(gran, debugBelch("GRAN: doing ContinueThread\n"));
982 /* ToDo: check assertion
983 ASSERT(run_queue_hd != (StgTSO*)NULL &&
984 run_queue_hd != END_TSO_QUEUE);
986 /* Ignore ContinueThreads for fetching threads (if synchr comm) */
987 if (!RtsFlags.GranFlags.DoAsyncFetch &&
988 procStatus[CurrentProc]==Fetching) {
989 debugBelch("ghuH: Spurious ContinueThread while Fetching ignored; TSO %d (%p) [PE %d]\n",
990 CurrentTSO->id, CurrentTSO, CurrentProc);
993 /* Ignore ContinueThreads for completed threads */
994 if (CurrentTSO->what_next == ThreadComplete) {
995 debugBelch("ghuH: found a ContinueThread event for completed thread %d (%p) [PE %d] (ignoring ContinueThread)\n",
996 CurrentTSO->id, CurrentTSO, CurrentProc);
999 /* Ignore ContinueThreads for threads that are being migrated */
1000 if (PROCS(CurrentTSO)==Nowhere) {
1001 debugBelch("ghuH: trying to run the migrating TSO %d (%p) [PE %d] (ignoring ContinueThread)\n",
1002 CurrentTSO->id, CurrentTSO, CurrentProc);
1005 /* The thread should be at the beginning of the run queue */
1006 if (CurrentTSO!=run_queue_hds[CurrentProc]) {
1007 debugBelch("ghuH: TSO %d (%p) [PE %d] is not at the start of the run_queue when doing a ContinueThread\n",
1008 CurrentTSO->id, CurrentTSO, CurrentProc);
1009 break; // run the thread anyway
1012 new_event(proc, proc, CurrentTime[proc],
1014 (StgTSO*)NULL, (StgClosure*)NULL, (rtsSpark*)NULL);
1016 */ /* Catches superfluous CONTINUEs -- should be unnecessary */
1017 break; // now actually run the thread; DaH Qu'vam yImuHbej
1020 do_the_fetchnode(event);
1021 goto next_thread; /* handle next event in event queue */
1024 do_the_globalblock(event);
1025 goto next_thread; /* handle next event in event queue */
1028 do_the_fetchreply(event);
1029 goto next_thread; /* handle next event in event queue */
1031 case UnblockThread: /* Move from the blocked queue to the tail of */
1032 do_the_unblock(event);
1033 goto next_thread; /* handle next event in event queue */
1035 case ResumeThread: /* Move from the blocked queue to the tail of */
1036 /* the runnable queue ( i.e. Qu' SImqa'lu') */
1037 event->tso->gran.blocktime +=
1038 CurrentTime[CurrentProc] - event->tso->gran.blockedat;
1039 do_the_startthread(event);
1040 goto next_thread; /* handle next event in event queue */
1043 do_the_startthread(event);
1044 goto next_thread; /* handle next event in event queue */
1047 do_the_movethread(event);
1048 goto next_thread; /* handle next event in event queue */
1051 do_the_movespark(event);
1052 goto next_thread; /* handle next event in event queue */
1055 do_the_findwork(event);
1056 goto next_thread; /* handle next event in event queue */
1059 barf("Illegal event type %u\n", event->evttype);
1062 /* This point was scheduler_loop in the old RTS */
1064 IF_DEBUG(gran, debugBelch("GRAN: after main switch\n"));
1066 TimeOfLastEvent = CurrentTime[CurrentProc];
1067 TimeOfNextEvent = get_time_of_next_event();
1068 IgnoreEvents=(TimeOfNextEvent==0); // HWL HACK
1069 // CurrentTSO = ThreadQueueHd;
1071 IF_DEBUG(gran, debugBelch("GRAN: time of next event is: %ld\n",
1074 if (RtsFlags.GranFlags.Light)
1075 GranSimLight_leave_system(event, &ActiveTSO);
1077 EndOfTimeSlice = CurrentTime[CurrentProc]+RtsFlags.GranFlags.time_slice;
1080 debugBelch("GRAN: end of time-slice is %#lx\n", EndOfTimeSlice));
1082 /* in a GranSim setup the TSO stays on the run queue */
1084 /* Take a thread from the run queue. */
1085 POP_RUN_QUEUE(t); // take_off_run_queue(t);
1088 debugBelch("GRAN: About to run current thread, which is\n");
1091 context_switch = 0; // turned on via GranYield, checking events and time slice
1094 DumpGranEvent(GR_SCHEDULE, t));
1096 procStatus[CurrentProc] = Busy;
1100 /* ----------------------------------------------------------------------------
1101 * Send pending messages (PARALLEL_HASKELL only)
1102 * ------------------------------------------------------------------------- */
1104 #if defined(PARALLEL_HASKELL)
1106 scheduleSendPendingMessages(void)
1112 # if defined(PAR) // global Mem.Mgmt., omit for now
1113 if (PendingFetches != END_BF_QUEUE) {
1118 if (RtsFlags.ParFlags.BufferTime) {
1119 // if we use message buffering, we must send away all message
1120 // packets which have become too old...
1126 /* ----------------------------------------------------------------------------
1127 * Activate spark threads (PARALLEL_HASKELL only)
1128 * ------------------------------------------------------------------------- */
1130 #if defined(PARALLEL_HASKELL)
1132 scheduleActivateSpark(void)
1135 ASSERT(EMPTY_RUN_QUEUE());
1136 /* We get here if the run queue is empty and want some work.
1137 We try to turn a spark into a thread, and add it to the run queue,
1138 from where it will be picked up in the next iteration of the scheduler
1142 /* :-[ no local threads => look out for local sparks */
1143 /* the spark pool for the current PE */
1144 pool = &(cap.r.rSparks); // JB: cap = (old) MainCap
1145 if (advisory_thread_count < RtsFlags.ParFlags.maxThreads &&
1146 pool->hd < pool->tl) {
1148 * ToDo: add GC code check that we really have enough heap afterwards!!
1150 * If we're here (no runnable threads) and we have pending
1151 * sparks, we must have a space problem. Get enough space
1152 * to turn one of those pending sparks into a
1156 spark = findSpark(rtsFalse); /* get a spark */
1157 if (spark != (rtsSpark) NULL) {
1158 tso = createThreadFromSpark(spark); /* turn the spark into a thread */
1159 IF_PAR_DEBUG(fish, // schedule,
1160 debugBelch("==== schedule: Created TSO %d (%p); %d threads active\n",
1161 tso->id, tso, advisory_thread_count));
1163 if (tso==END_TSO_QUEUE) { /* failed to activate spark->back to loop */
1164 IF_PAR_DEBUG(fish, // schedule,
1165 debugBelch("==^^ failed to create thread from spark @ %lx\n",
1167 return rtsFalse; /* failed to generate a thread */
1168 } /* otherwise fall through & pick-up new tso */
1170 IF_PAR_DEBUG(fish, // schedule,
1171 debugBelch("==^^ no local sparks (spark pool contains only NFs: %d)\n",
1172 spark_queue_len(pool)));
1173 return rtsFalse; /* failed to generate a thread */
1175 return rtsTrue; /* success in generating a thread */
1176 } else { /* no more threads permitted or pool empty */
1177 return rtsFalse; /* failed to generateThread */
1180 tso = NULL; // avoid compiler warning only
1181 return rtsFalse; /* dummy in non-PAR setup */
1184 #endif // PARALLEL_HASKELL
1186 /* ----------------------------------------------------------------------------
1187 * Get work from a remote node (PARALLEL_HASKELL only)
1188 * ------------------------------------------------------------------------- */
1190 #if defined(PARALLEL_HASKELL)
1192 scheduleGetRemoteWork(rtsBool *receivedFinish)
1194 ASSERT(EMPTY_RUN_QUEUE());
1196 if (RtsFlags.ParFlags.BufferTime) {
1197 IF_PAR_DEBUG(verbose,
1198 debugBelch("...send all pending data,"));
1201 for (i=1; i<=nPEs; i++)
1202 sendImmediately(i); // send all messages away immediately
1206 //++EDEN++ idle() , i.e. send all buffers, wait for work
1207 // suppress fishing in EDEN... just look for incoming messages
1208 // (blocking receive)
1209 IF_PAR_DEBUG(verbose,
1210 debugBelch("...wait for incoming messages...\n"));
1211 *receivedFinish = processMessages(); // blocking receive...
1213 // and reenter scheduling loop after having received something
1214 // (return rtsFalse below)
1216 # else /* activate SPARKS machinery */
1217 /* We get here, if we have no work, tried to activate a local spark, but still
1218 have no work. We try to get a remote spark, by sending a FISH message.
1219 Thread migration should be added here, and triggered when a sequence of
1220 fishes returns without work. */
1221 delay = (RtsFlags.ParFlags.fishDelay!=0ll ? RtsFlags.ParFlags.fishDelay : 0ll);
1223 /* =8-[ no local sparks => look for work on other PEs */
1225 * We really have absolutely no work. Send out a fish
1226 * (there may be some out there already), and wait for
1227 * something to arrive. We clearly can't run any threads
1228 * until a SCHEDULE or RESUME arrives, and so that's what
1229 * we're hoping to see. (Of course, we still have to
1230 * respond to other types of messages.)
1232 rtsTime now = msTime() /*CURRENT_TIME*/;
1233 IF_PAR_DEBUG(verbose,
1234 debugBelch("-- now=%ld\n", now));
1235 IF_PAR_DEBUG(fish, // verbose,
1236 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
1237 (last_fish_arrived_at!=0 &&
1238 last_fish_arrived_at+delay > now)) {
1239 debugBelch("--$$ <%llu> delaying FISH until %llu (last fish %llu, delay %llu)\n",
1240 now, last_fish_arrived_at+delay,
1241 last_fish_arrived_at,
1245 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
1246 advisory_thread_count < RtsFlags.ParFlags.maxThreads) { // send a FISH, but when?
1247 if (last_fish_arrived_at==0 ||
1248 (last_fish_arrived_at+delay <= now)) { // send FISH now!
1249 /* outstandingFishes is set in sendFish, processFish;
1250 avoid flooding system with fishes via delay */
1251 next_fish_to_send_at = 0;
1253 /* ToDo: this should be done in the main scheduling loop to avoid the
1254 busy wait here; not so bad if fish delay is very small */
1255 int iq = 0; // DEBUGGING -- HWL
1256 next_fish_to_send_at = last_fish_arrived_at+delay; // remember when to send
1257 /* send a fish when ready, but process messages that arrive in the meantime */
1259 if (PacketsWaiting()) {
1261 *receivedFinish = processMessages();
1264 } while (!*receivedFinish || now<next_fish_to_send_at);
1265 // JB: This means the fish could become obsolete, if we receive
1266 // work. Better check for work again?
1267 // last line: while (!receivedFinish || !haveWork || now<...)
1268 // next line: if (receivedFinish || haveWork )
1270 if (*receivedFinish) // no need to send a FISH if we are finishing anyway
1271 return rtsFalse; // NB: this will leave scheduler loop
1272 // immediately after return!
1274 IF_PAR_DEBUG(fish, // verbose,
1275 debugBelch("--$$ <%llu> sent delayed fish (%d processMessages); active/total threads=%d/%d\n",now,iq,run_queue_len(),advisory_thread_count));
1279 // JB: IMHO, this should all be hidden inside sendFish(...)
1281 sendFish(pe, thisPE, NEW_FISH_AGE, NEW_FISH_HISTORY,
1284 // Global statistics: count no. of fishes
1285 if (RtsFlags.ParFlags.ParStats.Global &&
1286 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
1287 globalParStats.tot_fish_mess++;
1291 /* delayed fishes must have been sent by now! */
1292 next_fish_to_send_at = 0;
1295 *receivedFinish = processMessages();
1296 # endif /* SPARKS */
1299 /* NB: this function always returns rtsFalse, meaning the scheduler
1300 loop continues with the next iteration;
1302 return code means success in finding work; we enter this function
1303 if there is no local work, thus have to send a fish which takes
1304 time until it arrives with work; in the meantime we should process
1305 messages in the main loop;
1308 #endif // PARALLEL_HASKELL
1310 /* ----------------------------------------------------------------------------
1311 * PAR/GRAN: Report stats & debugging info(?)
1312 * ------------------------------------------------------------------------- */
1314 #if defined(PAR) || defined(GRAN)
1316 scheduleGranParReport(void)
1318 ASSERT(run_queue_hd != END_TSO_QUEUE);
1320 /* Take a thread from the run queue, if we have work */
1321 POP_RUN_QUEUE(t); // take_off_run_queue(END_TSO_QUEUE);
1323 /* If this TSO has got its outport closed in the meantime,
1324 * it mustn't be run. Instead, we have to clean it up as if it was finished.
1325 * It has to be marked as TH_DEAD for this purpose.
1326 * If it is TH_TERM instead, it is supposed to have finished in the normal way.
1328 JB: TODO: investigate wether state change field could be nuked
1329 entirely and replaced by the normal tso state (whatnext
1330 field). All we want to do is to kill tsos from outside.
1333 /* ToDo: write something to the log-file
1334 if (RTSflags.ParFlags.granSimStats && !sameThread)
1335 DumpGranEvent(GR_SCHEDULE, RunnableThreadsHd);
1339 /* the spark pool for the current PE */
1340 pool = &(cap.r.rSparks); // cap = (old) MainCap
1343 debugBelch("--=^ %d threads, %d sparks on [%#x]\n",
1344 run_queue_len(), spark_queue_len(pool), CURRENT_PROC));
1347 debugBelch("--=^ %d threads, %d sparks on [%#x]\n",
1348 run_queue_len(), spark_queue_len(pool), CURRENT_PROC));
1350 if (RtsFlags.ParFlags.ParStats.Full &&
1351 (t->par.sparkname != (StgInt)0) && // only log spark generated threads
1352 (emitSchedule || // forced emit
1353 (t && LastTSO && t->id != LastTSO->id))) {
1355 we are running a different TSO, so write a schedule event to log file
1356 NB: If we use fair scheduling we also have to write a deschedule
1357 event for LastTSO; with unfair scheduling we know that the
1358 previous tso has blocked whenever we switch to another tso, so
1359 we don't need it in GUM for now
1361 IF_PAR_DEBUG(fish, // schedule,
1362 debugBelch("____ scheduling spark generated thread %d (%lx) (%lx) via a forced emit\n",t->id,t,t->par.sparkname));
1364 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
1365 GR_SCHEDULE, t, (StgClosure *)NULL, 0, 0);
1366 emitSchedule = rtsFalse;
1371 /* ----------------------------------------------------------------------------
1372 * After running a thread...
1373 * ASSUMES: sched_mutex
1374 * ------------------------------------------------------------------------- */
1377 schedulePostRunThread(void)
1380 /* HACK 675: if the last thread didn't yield, make sure to print a
1381 SCHEDULE event to the log file when StgRunning the next thread, even
1382 if it is the same one as before */
1384 TimeOfLastYield = CURRENT_TIME;
1387 /* some statistics gathering in the parallel case */
1389 #if defined(GRAN) || defined(PAR) || defined(EDEN)
1393 IF_DEBUG(gran, DumpGranEvent(GR_DESCHEDULE, t));
1394 globalGranStats.tot_heapover++;
1396 globalParStats.tot_heapover++;
1403 DumpGranEvent(GR_DESCHEDULE, t));
1404 globalGranStats.tot_stackover++;
1407 // DumpGranEvent(GR_DESCHEDULE, t);
1408 globalParStats.tot_stackover++;
1412 case ThreadYielding:
1415 DumpGranEvent(GR_DESCHEDULE, t));
1416 globalGranStats.tot_yields++;
1419 // DumpGranEvent(GR_DESCHEDULE, t);
1420 globalParStats.tot_yields++;
1427 debugBelch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: ",
1428 t->id, t, whatNext_strs[t->what_next], t->block_info.closure,
1429 (t->block_info.closure==(StgClosure*)NULL ? 99 : where_is(t->block_info.closure)));
1430 if (t->block_info.closure!=(StgClosure*)NULL)
1431 print_bq(t->block_info.closure);
1434 // ??? needed; should emit block before
1436 DumpGranEvent(GR_DESCHEDULE, t));
1437 prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
1440 ASSERT(procStatus[CurrentProc]==Busy ||
1441 ((procStatus[CurrentProc]==Fetching) &&
1442 (t->block_info.closure!=(StgClosure*)NULL)));
1443 if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
1444 !(!RtsFlags.GranFlags.DoAsyncFetch &&
1445 procStatus[CurrentProc]==Fetching))
1446 procStatus[CurrentProc] = Idle;
1449 //++PAR++ blockThread() writes the event (change?)
1453 case ThreadFinished:
1457 barf("parGlobalStats: unknown return code");
1463 /* -----------------------------------------------------------------------------
1464 * Handle a thread that returned to the scheduler with ThreadHeepOverflow
1465 * ASSUMES: sched_mutex
1466 * -------------------------------------------------------------------------- */
1469 scheduleHandleHeapOverflow( Capability *cap, StgTSO *t )
1471 // did the task ask for a large block?
1472 if (cap->r.rHpAlloc > BLOCK_SIZE) {
1473 // if so, get one and push it on the front of the nursery.
1477 blocks = (lnat)BLOCK_ROUND_UP(cap->r.rHpAlloc) / BLOCK_SIZE;
1480 debugBelch("--<< thread %ld (%s) stopped: requesting a large block (size %ld)\n",
1481 (long)t->id, whatNext_strs[t->what_next], blocks));
1483 // don't do this if the nursery is (nearly) full, we'll GC first.
1484 if (cap->r.rCurrentNursery->link != NULL ||
1485 cap->r.rNursery->n_blocks == 1) { // paranoia to prevent infinite loop
1486 // if the nursery has only one block.
1488 bd = allocGroup( blocks );
1489 cap->r.rNursery->n_blocks += blocks;
1491 // link the new group into the list
1492 bd->link = cap->r.rCurrentNursery;
1493 bd->u.back = cap->r.rCurrentNursery->u.back;
1494 if (cap->r.rCurrentNursery->u.back != NULL) {
1495 cap->r.rCurrentNursery->u.back->link = bd;
1498 ASSERT(g0s0->blocks == cap->r.rCurrentNursery &&
1499 g0s0 == cap->r.rNursery);
1501 cap->r.rNursery->blocks = bd;
1503 cap->r.rCurrentNursery->u.back = bd;
1505 // initialise it as a nursery block. We initialise the
1506 // step, gen_no, and flags field of *every* sub-block in
1507 // this large block, because this is easier than making
1508 // sure that we always find the block head of a large
1509 // block whenever we call Bdescr() (eg. evacuate() and
1510 // isAlive() in the GC would both have to do this, at
1514 for (x = bd; x < bd + blocks; x++) {
1515 x->step = cap->r.rNursery;
1521 // This assert can be a killer if the app is doing lots
1522 // of large block allocations.
1523 IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
1525 // now update the nursery to point to the new block
1526 cap->r.rCurrentNursery = bd;
1528 // we might be unlucky and have another thread get on the
1529 // run queue before us and steal the large block, but in that
1530 // case the thread will just end up requesting another large
1532 PUSH_ON_RUN_QUEUE(t);
1533 return rtsFalse; /* not actually GC'ing */
1538 debugBelch("--<< thread %ld (%s) stopped: HeapOverflow\n",
1539 (long)t->id, whatNext_strs[t->what_next]));
1541 ASSERT(!is_on_queue(t,CurrentProc));
1542 #elif defined(PARALLEL_HASKELL)
1543 /* Currently we emit a DESCHEDULE event before GC in GUM.
1544 ToDo: either add separate event to distinguish SYSTEM time from rest
1545 or just nuke this DESCHEDULE (and the following SCHEDULE) */
1546 if (0 && RtsFlags.ParFlags.ParStats.Full) {
1547 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
1548 GR_DESCHEDULE, t, (StgClosure *)NULL, 0, 0);
1549 emitSchedule = rtsTrue;
1553 PUSH_ON_RUN_QUEUE(t);
1555 /* actual GC is done at the end of the while loop in schedule() */
1558 /* -----------------------------------------------------------------------------
1559 * Handle a thread that returned to the scheduler with ThreadStackOverflow
1560 * ASSUMES: sched_mutex
1561 * -------------------------------------------------------------------------- */
1564 scheduleHandleStackOverflow( StgTSO *t)
1566 IF_DEBUG(scheduler,debugBelch("--<< thread %ld (%s) stopped, StackOverflow\n",
1567 (long)t->id, whatNext_strs[t->what_next]));
1568 /* just adjust the stack for this thread, then pop it back
1572 /* enlarge the stack */
1573 StgTSO *new_t = threadStackOverflow(t);
1575 /* This TSO has moved, so update any pointers to it from the
1576 * main thread stack. It better not be on any other queues...
1577 * (it shouldn't be).
1579 if (t->main != NULL) {
1580 t->main->tso = new_t;
1582 PUSH_ON_RUN_QUEUE(new_t);
1586 /* -----------------------------------------------------------------------------
1587 * Handle a thread that returned to the scheduler with ThreadYielding
1588 * ASSUMES: sched_mutex
1589 * -------------------------------------------------------------------------- */
1592 scheduleHandleYield( StgTSO *t, nat prev_what_next )
1594 // Reset the context switch flag. We don't do this just before
1595 // running the thread, because that would mean we would lose ticks
1596 // during GC, which can lead to unfair scheduling (a thread hogs
1597 // the CPU because the tick always arrives during GC). This way
1598 // penalises threads that do a lot of allocation, but that seems
1599 // better than the alternative.
1602 /* put the thread back on the run queue. Then, if we're ready to
1603 * GC, check whether this is the last task to stop. If so, wake
1604 * up the GC thread. getThread will block during a GC until the
1608 if (t->what_next != prev_what_next) {
1609 debugBelch("--<< thread %ld (%s) stopped to switch evaluators\n",
1610 (long)t->id, whatNext_strs[t->what_next]);
1612 debugBelch("--<< thread %ld (%s) stopped, yielding\n",
1613 (long)t->id, whatNext_strs[t->what_next]);
1618 //debugBelch("&& Doing sanity check on yielding TSO %ld.", t->id);
1620 ASSERT(t->link == END_TSO_QUEUE);
1622 // Shortcut if we're just switching evaluators: don't bother
1623 // doing stack squeezing (which can be expensive), just run the
1625 if (t->what_next != prev_what_next) {
1630 ASSERT(!is_on_queue(t,CurrentProc));
1633 //debugBelch("&& Doing sanity check on all ThreadQueues (and their TSOs).");
1634 checkThreadQsSanity(rtsTrue));
1641 /* add a ContinueThread event to actually process the thread */
1642 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1644 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1646 debugBelch("GRAN: eventq and runnableq after adding yielded thread to queue again:\n");
1653 /* -----------------------------------------------------------------------------
1654 * Handle a thread that returned to the scheduler with ThreadBlocked
1655 * ASSUMES: sched_mutex
1656 * -------------------------------------------------------------------------- */
1659 scheduleHandleThreadBlocked( StgTSO *t
1660 #if !defined(GRAN) && !defined(DEBUG)
1667 debugBelch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: \n",
1668 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)));
1669 if (t->block_info.closure!=(StgClosure*)NULL) print_bq(t->block_info.closure));
1671 // ??? needed; should emit block before
1673 DumpGranEvent(GR_DESCHEDULE, t));
1674 prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
1677 ASSERT(procStatus[CurrentProc]==Busy ||
1678 ((procStatus[CurrentProc]==Fetching) &&
1679 (t->block_info.closure!=(StgClosure*)NULL)));
1680 if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
1681 !(!RtsFlags.GranFlags.DoAsyncFetch &&
1682 procStatus[CurrentProc]==Fetching))
1683 procStatus[CurrentProc] = Idle;
1687 debugBelch("--<< thread %ld (%p; %s) stopped, blocking on node %p with BQ: \n",
1688 t->id, t, whatNext_strs[t->what_next], t->block_info.closure));
1691 if (t->block_info.closure!=(StgClosure*)NULL)
1692 print_bq(t->block_info.closure));
1694 /* Send a fetch (if BlockedOnGA) and dump event to log file */
1697 /* whatever we schedule next, we must log that schedule */
1698 emitSchedule = rtsTrue;
1702 // We don't need to do anything. The thread is blocked, and it
1703 // has tidied up its stack and placed itself on whatever queue
1704 // it needs to be on.
1707 ASSERT(t->why_blocked != NotBlocked);
1708 // This might not be true under SMP: we don't have
1709 // exclusive access to this TSO, so someone might have
1710 // woken it up by now. This actually happens: try
1711 // conc023 +RTS -N2.
1715 debugBelch("--<< thread %d (%s) stopped: ",
1716 t->id, whatNext_strs[t->what_next]);
1717 printThreadBlockage(t);
1720 /* Only for dumping event to log file
1721 ToDo: do I need this in GranSim, too?
1727 /* -----------------------------------------------------------------------------
1728 * Handle a thread that returned to the scheduler with ThreadFinished
1729 * ASSUMES: sched_mutex
1730 * -------------------------------------------------------------------------- */
1733 scheduleHandleThreadFinished( StgMainThread *mainThread
1734 USED_WHEN_RTS_SUPPORTS_THREADS,
1738 /* Need to check whether this was a main thread, and if so,
1739 * return with the return value.
1741 * We also end up here if the thread kills itself with an
1742 * uncaught exception, see Exception.cmm.
1744 IF_DEBUG(scheduler,debugBelch("--++ thread %d (%s) finished\n",
1745 t->id, whatNext_strs[t->what_next]));
1748 endThread(t, CurrentProc); // clean-up the thread
1749 #elif defined(PARALLEL_HASKELL)
1750 /* For now all are advisory -- HWL */
1751 //if(t->priority==AdvisoryPriority) ??
1752 advisory_thread_count--; // JB: Caution with this counter, buggy!
1755 if(t->dist.priority==RevalPriority)
1759 # if defined(EDENOLD)
1760 // the thread could still have an outport... (BUG)
1761 if (t->eden.outport != -1) {
1762 // delete the outport for the tso which has finished...
1763 IF_PAR_DEBUG(eden_ports,
1764 debugBelch("WARNING: Scheduler removes outport %d for TSO %d.\n",
1765 t->eden.outport, t->id));
1768 // thread still in the process (HEAVY BUG! since outport has just been closed...)
1769 if (t->eden.epid != -1) {
1770 IF_PAR_DEBUG(eden_ports,
1771 debugBelch("WARNING: Scheduler removes TSO %d from process %d .\n",
1772 t->id, t->eden.epid));
1773 removeTSOfromProcess(t);
1778 if (RtsFlags.ParFlags.ParStats.Full &&
1779 !RtsFlags.ParFlags.ParStats.Suppressed)
1780 DumpEndEvent(CURRENT_PROC, t, rtsFalse /* not mandatory */);
1782 // t->par only contains statistics: left out for now...
1784 debugBelch("**** end thread: ended sparked thread %d (%lx); sparkname: %lx\n",
1785 t->id,t,t->par.sparkname));
1787 #endif // PARALLEL_HASKELL
1790 // Check whether the thread that just completed was a main
1791 // thread, and if so return with the result.
1793 // There is an assumption here that all thread completion goes
1794 // through this point; we need to make sure that if a thread
1795 // ends up in the ThreadKilled state, that it stays on the run
1796 // queue so it can be dealt with here.
1799 #if defined(RTS_SUPPORTS_THREADS)
1802 mainThread->tso == t
1806 // We are a bound thread: this must be our thread that just
1808 ASSERT(mainThread->tso == t);
1810 if (t->what_next == ThreadComplete) {
1811 if (mainThread->ret) {
1812 // NOTE: return val is tso->sp[1] (see StgStartup.hc)
1813 *(mainThread->ret) = (StgClosure *)mainThread->tso->sp[1];
1815 mainThread->stat = Success;
1817 if (mainThread->ret) {
1818 *(mainThread->ret) = NULL;
1821 mainThread->stat = Interrupted;
1823 mainThread->stat = Killed;
1827 removeThreadLabel((StgWord)mainThread->tso->id);
1829 if (mainThread->prev == NULL) {
1830 ASSERT(mainThread == main_threads);
1831 main_threads = mainThread->link;
1833 mainThread->prev->link = mainThread->link;
1835 if (mainThread->link != NULL) {
1836 mainThread->link->prev = mainThread->prev;
1838 releaseCapability(cap);
1839 return rtsTrue; // tells schedule() to return
1842 #ifdef RTS_SUPPORTS_THREADS
1843 ASSERT(t->main == NULL);
1845 if (t->main != NULL) {
1846 // Must be a main thread that is not the topmost one. Leave
1847 // it on the run queue until the stack has unwound to the
1848 // point where we can deal with this. Leaving it on the run
1849 // queue also ensures that the garbage collector knows about
1850 // this thread and its return value (it gets dropped from the
1851 // all_threads list so there's no other way to find it).
1852 APPEND_TO_RUN_QUEUE(t);
1858 /* -----------------------------------------------------------------------------
1859 * Perform a heap census, if PROFILING
1860 * -------------------------------------------------------------------------- */
1863 scheduleDoHeapProfile( rtsBool ready_to_gc STG_UNUSED )
1865 #if defined(PROFILING)
1866 // When we have +RTS -i0 and we're heap profiling, do a census at
1867 // every GC. This lets us get repeatable runs for debugging.
1868 if (performHeapProfile ||
1869 (RtsFlags.ProfFlags.profileInterval==0 &&
1870 RtsFlags.ProfFlags.doHeapProfile && ready_to_gc)) {
1871 GarbageCollect(GetRoots, rtsTrue);
1873 performHeapProfile = rtsFalse;
1874 return rtsTrue; // true <=> we already GC'd
1880 /* -----------------------------------------------------------------------------
1881 * Perform a garbage collection if necessary
1882 * ASSUMES: sched_mutex
1883 * -------------------------------------------------------------------------- */
1886 scheduleDoGC( Capability *cap STG_UNUSED )
1890 static rtsBool waiting_for_gc;
1891 int n_capabilities = RtsFlags.ParFlags.nNodes - 1;
1892 // subtract one because we're already holding one.
1893 Capability *caps[n_capabilities];
1897 // In order to GC, there must be no threads running Haskell code.
1898 // Therefore, the GC thread needs to hold *all* the capabilities,
1899 // and release them after the GC has completed.
1901 // This seems to be the simplest way: previous attempts involved
1902 // making all the threads with capabilities give up their
1903 // capabilities and sleep except for the *last* one, which
1904 // actually did the GC. But it's quite hard to arrange for all
1905 // the other tasks to sleep and stay asleep.
1908 // Someone else is already trying to GC
1909 if (waiting_for_gc) return;
1910 waiting_for_gc = rtsTrue;
1912 caps[n_capabilities] = cap;
1913 while (n_capabilities > 0) {
1914 IF_DEBUG(scheduler, sched_belch("ready_to_gc, grabbing all the capabilies (%d left)", n_capabilities));
1915 waitForReturnCapability(&sched_mutex, &cap);
1917 caps[n_capabilities] = cap;
1920 waiting_for_gc = rtsFalse;
1923 /* Kick any transactions which are invalid back to their
1924 * atomically frames. When next scheduled they will try to
1925 * commit, this commit will fail and they will retry.
1927 for (t = all_threads; t != END_TSO_QUEUE; t = t -> link) {
1928 if (t -> what_next != ThreadRelocated && t -> trec != NO_TREC && t -> why_blocked == NotBlocked) {
1929 if (!stmValidateTransaction (t -> trec)) {
1930 IF_DEBUG(stm, sched_belch("trec %p found wasting its time", t));
1932 // strip the stack back to the ATOMICALLY_FRAME, aborting
1933 // the (nested) transaction, and saving the stack of any
1934 // partially-evaluated thunks on the heap.
1935 raiseAsync_(t, NULL, rtsTrue);
1938 ASSERT(get_itbl((StgClosure *)t->sp)->type == ATOMICALLY_FRAME);
1944 // so this happens periodically:
1945 scheduleCheckBlackHoles();
1947 IF_DEBUG(scheduler, printAllThreads());
1949 /* everybody back, start the GC.
1950 * Could do it in this thread, or signal a condition var
1951 * to do it in another thread. Either way, we need to
1952 * broadcast on gc_pending_cond afterward.
1954 #if defined(RTS_SUPPORTS_THREADS)
1955 IF_DEBUG(scheduler,sched_belch("doing GC"));
1957 GarbageCollect(GetRoots,rtsFalse);
1961 // release our stash of capabilities.
1963 for (i = 0; i < RtsFlags.ParFlags.nNodes-1; i++) {
1964 releaseCapability(caps[i]);
1970 /* add a ContinueThread event to continue execution of current thread */
1971 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1973 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1975 debugBelch("GRAN: eventq and runnableq after Garbage collection:\n\n");
1981 /* ---------------------------------------------------------------------------
1982 * rtsSupportsBoundThreads(): is the RTS built to support bound threads?
1983 * used by Control.Concurrent for error checking.
1984 * ------------------------------------------------------------------------- */
1987 rtsSupportsBoundThreads(void)
1989 #if defined(RTS_SUPPORTS_THREADS)
1996 /* ---------------------------------------------------------------------------
1997 * isThreadBound(tso): check whether tso is bound to an OS thread.
1998 * ------------------------------------------------------------------------- */
2001 isThreadBound(StgTSO* tso USED_IN_THREADED_RTS)
2003 #if defined(RTS_SUPPORTS_THREADS)
2004 return (tso->main != NULL);
2009 /* ---------------------------------------------------------------------------
2010 * Singleton fork(). Do not copy any running threads.
2011 * ------------------------------------------------------------------------- */
2013 #ifndef mingw32_HOST_OS
2014 #define FORKPROCESS_PRIMOP_SUPPORTED
2017 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
2019 deleteThreadImmediately(StgTSO *tso);
2022 forkProcess(HsStablePtr *entry
2023 #ifndef FORKPROCESS_PRIMOP_SUPPORTED
2028 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
2034 IF_DEBUG(scheduler,sched_belch("forking!"));
2035 rts_lock(); // This not only acquires sched_mutex, it also
2036 // makes sure that no other threads are running
2040 if (pid) { /* parent */
2042 /* just return the pid */
2046 } else { /* child */
2049 // delete all threads
2050 run_queue_hd = run_queue_tl = END_TSO_QUEUE;
2052 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
2055 // don't allow threads to catch the ThreadKilled exception
2056 deleteThreadImmediately(t);
2059 // wipe the main thread list
2060 while((m = main_threads) != NULL) {
2061 main_threads = m->link;
2062 # ifdef THREADED_RTS
2063 closeCondition(&m->bound_thread_cond);
2068 rc = rts_evalStableIO(entry, NULL); // run the action
2069 rts_checkSchedStatus("forkProcess",rc);
2073 hs_exit(); // clean up and exit
2076 #else /* !FORKPROCESS_PRIMOP_SUPPORTED */
2077 barf("forkProcess#: primop not supported, sorry!\n");
2082 /* ---------------------------------------------------------------------------
2083 * deleteAllThreads(): kill all the live threads.
2085 * This is used when we catch a user interrupt (^C), before performing
2086 * any necessary cleanups and running finalizers.
2088 * Locks: sched_mutex held.
2089 * ------------------------------------------------------------------------- */
2092 deleteAllThreads ( void )
2095 IF_DEBUG(scheduler,sched_belch("deleting all threads"));
2096 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
2097 if (t->what_next == ThreadRelocated) {
2100 next = t->global_link;
2105 // The run queue now contains a bunch of ThreadKilled threads. We
2106 // must not throw these away: the main thread(s) will be in there
2107 // somewhere, and the main scheduler loop has to deal with it.
2108 // Also, the run queue is the only thing keeping these threads from
2109 // being GC'd, and we don't want the "main thread has been GC'd" panic.
2111 ASSERT(blocked_queue_hd == END_TSO_QUEUE);
2112 ASSERT(blackhole_queue == END_TSO_QUEUE);
2113 ASSERT(sleeping_queue == END_TSO_QUEUE);
2116 /* startThread and insertThread are now in GranSim.c -- HWL */
2119 /* ---------------------------------------------------------------------------
2120 * Suspending & resuming Haskell threads.
2122 * When making a "safe" call to C (aka _ccall_GC), the task gives back
2123 * its capability before calling the C function. This allows another
2124 * task to pick up the capability and carry on running Haskell
2125 * threads. It also means that if the C call blocks, it won't lock
2128 * The Haskell thread making the C call is put to sleep for the
2129 * duration of the call, on the susepended_ccalling_threads queue. We
2130 * give out a token to the task, which it can use to resume the thread
2131 * on return from the C function.
2132 * ------------------------------------------------------------------------- */
2135 suspendThread( StgRegTable *reg )
2139 int saved_errno = errno;
2141 /* assume that *reg is a pointer to the StgRegTable part
2144 cap = (Capability *)((void *)((unsigned char*)reg - sizeof(StgFunTable)));
2146 ACQUIRE_LOCK(&sched_mutex);
2149 sched_belch("thread %d did a _ccall_gc", cap->r.rCurrentTSO->id));
2151 // XXX this might not be necessary --SDM
2152 cap->r.rCurrentTSO->what_next = ThreadRunGHC;
2154 threadPaused(cap->r.rCurrentTSO);
2155 cap->r.rCurrentTSO->link = suspended_ccalling_threads;
2156 suspended_ccalling_threads = cap->r.rCurrentTSO;
2158 if(cap->r.rCurrentTSO->blocked_exceptions == NULL) {
2159 cap->r.rCurrentTSO->why_blocked = BlockedOnCCall;
2160 cap->r.rCurrentTSO->blocked_exceptions = END_TSO_QUEUE;
2162 cap->r.rCurrentTSO->why_blocked = BlockedOnCCall_NoUnblockExc;
2165 /* Use the thread ID as the token; it should be unique */
2166 tok = cap->r.rCurrentTSO->id;
2168 /* Hand back capability */
2169 cap->r.rInHaskell = rtsFalse;
2170 releaseCapability(cap);
2172 #if defined(RTS_SUPPORTS_THREADS)
2173 /* Preparing to leave the RTS, so ensure there's a native thread/task
2174 waiting to take over.
2176 IF_DEBUG(scheduler, sched_belch("worker (token %d): leaving RTS", tok));
2179 RELEASE_LOCK(&sched_mutex);
2181 errno = saved_errno;
2186 resumeThread( StgInt tok )
2188 StgTSO *tso, **prev;
2190 int saved_errno = errno;
2192 #if defined(RTS_SUPPORTS_THREADS)
2193 /* Wait for permission to re-enter the RTS with the result. */
2194 ACQUIRE_LOCK(&sched_mutex);
2195 waitForReturnCapability(&sched_mutex, &cap);
2197 IF_DEBUG(scheduler, sched_belch("worker (token %d): re-entering RTS", tok));
2199 grabCapability(&cap);
2202 /* Remove the thread off of the suspended list */
2203 prev = &suspended_ccalling_threads;
2204 for (tso = suspended_ccalling_threads;
2205 tso != END_TSO_QUEUE;
2206 prev = &tso->link, tso = tso->link) {
2207 if (tso->id == (StgThreadID)tok) {
2212 if (tso == END_TSO_QUEUE) {
2213 barf("resumeThread: thread not found");
2215 tso->link = END_TSO_QUEUE;
2217 if(tso->why_blocked == BlockedOnCCall) {
2218 awakenBlockedQueueNoLock(tso->blocked_exceptions);
2219 tso->blocked_exceptions = NULL;
2222 /* Reset blocking status */
2223 tso->why_blocked = NotBlocked;
2225 cap->r.rCurrentTSO = tso;
2226 cap->r.rInHaskell = rtsTrue;
2227 RELEASE_LOCK(&sched_mutex);
2228 errno = saved_errno;
2232 /* ---------------------------------------------------------------------------
2233 * Comparing Thread ids.
2235 * This is used from STG land in the implementation of the
2236 * instances of Eq/Ord for ThreadIds.
2237 * ------------------------------------------------------------------------ */
2240 cmp_thread(StgPtr tso1, StgPtr tso2)
2242 StgThreadID id1 = ((StgTSO *)tso1)->id;
2243 StgThreadID id2 = ((StgTSO *)tso2)->id;
2245 if (id1 < id2) return (-1);
2246 if (id1 > id2) return 1;
2250 /* ---------------------------------------------------------------------------
2251 * Fetching the ThreadID from an StgTSO.
2253 * This is used in the implementation of Show for ThreadIds.
2254 * ------------------------------------------------------------------------ */
2256 rts_getThreadId(StgPtr tso)
2258 return ((StgTSO *)tso)->id;
2263 labelThread(StgPtr tso, char *label)
2268 /* Caveat: Once set, you can only set the thread name to "" */
2269 len = strlen(label)+1;
2270 buf = stgMallocBytes(len * sizeof(char), "Schedule.c:labelThread()");
2271 strncpy(buf,label,len);
2272 /* Update will free the old memory for us */
2273 updateThreadLabel(((StgTSO *)tso)->id,buf);
2277 /* ---------------------------------------------------------------------------
2278 Create a new thread.
2280 The new thread starts with the given stack size. Before the
2281 scheduler can run, however, this thread needs to have a closure
2282 (and possibly some arguments) pushed on its stack. See
2283 pushClosure() in Schedule.h.
2285 createGenThread() and createIOThread() (in SchedAPI.h) are
2286 convenient packaged versions of this function.
2288 currently pri (priority) is only used in a GRAN setup -- HWL
2289 ------------------------------------------------------------------------ */
2291 /* currently pri (priority) is only used in a GRAN setup -- HWL */
2293 createThread(nat size, StgInt pri)
2296 createThread(nat size)
2303 /* First check whether we should create a thread at all */
2304 #if defined(PARALLEL_HASKELL)
2305 /* check that no more than RtsFlags.ParFlags.maxThreads threads are created */
2306 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads) {
2308 debugBelch("{createThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)\n",
2309 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
2310 return END_TSO_QUEUE;
2316 ASSERT(!RtsFlags.GranFlags.Light || CurrentProc==0);
2319 // ToDo: check whether size = stack_size - TSO_STRUCT_SIZEW
2321 /* catch ridiculously small stack sizes */
2322 if (size < MIN_STACK_WORDS + TSO_STRUCT_SIZEW) {
2323 size = MIN_STACK_WORDS + TSO_STRUCT_SIZEW;
2326 stack_size = size - TSO_STRUCT_SIZEW;
2328 tso = (StgTSO *)allocate(size);
2329 TICK_ALLOC_TSO(stack_size, 0);
2331 SET_HDR(tso, &stg_TSO_info, CCS_SYSTEM);
2333 SET_GRAN_HDR(tso, ThisPE);
2336 // Always start with the compiled code evaluator
2337 tso->what_next = ThreadRunGHC;
2339 tso->id = next_thread_id++;
2340 tso->why_blocked = NotBlocked;
2341 tso->blocked_exceptions = NULL;
2343 tso->saved_errno = 0;
2346 tso->stack_size = stack_size;
2347 tso->max_stack_size = round_to_mblocks(RtsFlags.GcFlags.maxStkSize)
2349 tso->sp = (P_)&(tso->stack) + stack_size;
2351 tso->trec = NO_TREC;
2354 tso->prof.CCCS = CCS_MAIN;
2357 /* put a stop frame on the stack */
2358 tso->sp -= sizeofW(StgStopFrame);
2359 SET_HDR((StgClosure*)tso->sp,(StgInfoTable *)&stg_stop_thread_info,CCS_SYSTEM);
2360 tso->link = END_TSO_QUEUE;
2364 /* uses more flexible routine in GranSim */
2365 insertThread(tso, CurrentProc);
2367 /* In a non-GranSim setup the pushing of a TSO onto the runq is separated
2373 if (RtsFlags.GranFlags.GranSimStats.Full)
2374 DumpGranEvent(GR_START,tso);
2375 #elif defined(PARALLEL_HASKELL)
2376 if (RtsFlags.ParFlags.ParStats.Full)
2377 DumpGranEvent(GR_STARTQ,tso);
2378 /* HACk to avoid SCHEDULE
2382 /* Link the new thread on the global thread list.
2384 tso->global_link = all_threads;
2388 tso->dist.priority = MandatoryPriority; //by default that is...
2392 tso->gran.pri = pri;
2394 tso->gran.magic = TSO_MAGIC; // debugging only
2396 tso->gran.sparkname = 0;
2397 tso->gran.startedat = CURRENT_TIME;
2398 tso->gran.exported = 0;
2399 tso->gran.basicblocks = 0;
2400 tso->gran.allocs = 0;
2401 tso->gran.exectime = 0;
2402 tso->gran.fetchtime = 0;
2403 tso->gran.fetchcount = 0;
2404 tso->gran.blocktime = 0;
2405 tso->gran.blockcount = 0;
2406 tso->gran.blockedat = 0;
2407 tso->gran.globalsparks = 0;
2408 tso->gran.localsparks = 0;
2409 if (RtsFlags.GranFlags.Light)
2410 tso->gran.clock = Now; /* local clock */
2412 tso->gran.clock = 0;
2414 IF_DEBUG(gran,printTSO(tso));
2415 #elif defined(PARALLEL_HASKELL)
2417 tso->par.magic = TSO_MAGIC; // debugging only
2419 tso->par.sparkname = 0;
2420 tso->par.startedat = CURRENT_TIME;
2421 tso->par.exported = 0;
2422 tso->par.basicblocks = 0;
2423 tso->par.allocs = 0;
2424 tso->par.exectime = 0;
2425 tso->par.fetchtime = 0;
2426 tso->par.fetchcount = 0;
2427 tso->par.blocktime = 0;
2428 tso->par.blockcount = 0;
2429 tso->par.blockedat = 0;
2430 tso->par.globalsparks = 0;
2431 tso->par.localsparks = 0;
2435 globalGranStats.tot_threads_created++;
2436 globalGranStats.threads_created_on_PE[CurrentProc]++;
2437 globalGranStats.tot_sq_len += spark_queue_len(CurrentProc);
2438 globalGranStats.tot_sq_probes++;
2439 #elif defined(PARALLEL_HASKELL)
2440 // collect parallel global statistics (currently done together with GC stats)
2441 if (RtsFlags.ParFlags.ParStats.Global &&
2442 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
2443 //debugBelch("Creating thread %d @ %11.2f\n", tso->id, usertime());
2444 globalParStats.tot_threads_created++;
2450 sched_belch("==__ schedule: Created TSO %d (%p);",
2451 CurrentProc, tso, tso->id));
2452 #elif defined(PARALLEL_HASKELL)
2453 IF_PAR_DEBUG(verbose,
2454 sched_belch("==__ schedule: Created TSO %d (%p); %d threads active",
2455 (long)tso->id, tso, advisory_thread_count));
2457 IF_DEBUG(scheduler,sched_belch("created thread %ld, stack size = %lx words",
2458 (long)tso->id, (long)tso->stack_size));
2465 all parallel thread creation calls should fall through the following routine.
2468 createThreadFromSpark(rtsSpark spark)
2470 ASSERT(spark != (rtsSpark)NULL);
2471 // JB: TAKE CARE OF THIS COUNTER! BUGGY
2472 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads)
2474 barf("{createSparkThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
2475 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
2476 return END_TSO_QUEUE;
2480 tso = createThread(RtsFlags.GcFlags.initialStkSize);
2481 if (tso==END_TSO_QUEUE)
2482 barf("createSparkThread: Cannot create TSO");
2484 tso->priority = AdvisoryPriority;
2486 pushClosure(tso,spark);
2488 advisory_thread_count++; // JB: TAKE CARE OF THIS COUNTER! BUGGY
2495 Turn a spark into a thread.
2496 ToDo: fix for SMP (needs to acquire SCHED_MUTEX!)
2500 activateSpark (rtsSpark spark)
2504 tso = createSparkThread(spark);
2505 if (RtsFlags.ParFlags.ParStats.Full) {
2506 //ASSERT(run_queue_hd == END_TSO_QUEUE); // I think ...
2507 IF_PAR_DEBUG(verbose,
2508 debugBelch("==^^ activateSpark: turning spark of closure %p (%s) into a thread\n",
2509 (StgClosure *)spark, info_type((StgClosure *)spark)));
2511 // ToDo: fwd info on local/global spark to thread -- HWL
2512 // tso->gran.exported = spark->exported;
2513 // tso->gran.locked = !spark->global;
2514 // tso->gran.sparkname = spark->name;
2520 /* ---------------------------------------------------------------------------
2523 * scheduleThread puts a thread on the head of the runnable queue.
2524 * This will usually be done immediately after a thread is created.
2525 * The caller of scheduleThread must create the thread using e.g.
2526 * createThread and push an appropriate closure
2527 * on this thread's stack before the scheduler is invoked.
2528 * ------------------------------------------------------------------------ */
2531 scheduleThread_(StgTSO *tso)
2533 // The thread goes at the *end* of the run-queue, to avoid possible
2534 // starvation of any threads already on the queue.
2535 APPEND_TO_RUN_QUEUE(tso);
2540 scheduleThread(StgTSO* tso)
2542 ACQUIRE_LOCK(&sched_mutex);
2543 scheduleThread_(tso);
2544 RELEASE_LOCK(&sched_mutex);
2547 #if defined(RTS_SUPPORTS_THREADS)
2548 static Condition bound_cond_cache;
2549 static int bound_cond_cache_full = 0;
2554 scheduleWaitThread(StgTSO* tso, /*[out]*/HaskellObj* ret,
2555 Capability *initialCapability)
2557 // Precondition: sched_mutex must be held
2560 m = stgMallocBytes(sizeof(StgMainThread), "waitThread");
2565 m->link = main_threads;
2567 if (main_threads != NULL) {
2568 main_threads->prev = m;
2572 #if defined(RTS_SUPPORTS_THREADS)
2573 // Allocating a new condition for each thread is expensive, so we
2574 // cache one. This is a pretty feeble hack, but it helps speed up
2575 // consecutive call-ins quite a bit.
2576 if (bound_cond_cache_full) {
2577 m->bound_thread_cond = bound_cond_cache;
2578 bound_cond_cache_full = 0;
2580 initCondition(&m->bound_thread_cond);
2584 /* Put the thread on the main-threads list prior to scheduling the TSO.
2585 Failure to do so introduces a race condition in the MT case (as
2586 identified by Wolfgang Thaller), whereby the new task/OS thread
2587 created by scheduleThread_() would complete prior to the thread
2588 that spawned it managed to put 'itself' on the main-threads list.
2589 The upshot of it all being that the worker thread wouldn't get to
2590 signal the completion of the its work item for the main thread to
2591 see (==> it got stuck waiting.) -- sof 6/02.
2593 IF_DEBUG(scheduler, sched_belch("waiting for thread (%d)", tso->id));
2595 APPEND_TO_RUN_QUEUE(tso);
2596 // NB. Don't call threadRunnable() here, because the thread is
2597 // bound and only runnable by *this* OS thread, so waking up other
2598 // workers will just slow things down.
2600 return waitThread_(m, initialCapability);
2603 /* ---------------------------------------------------------------------------
2606 * Initialise the scheduler. This resets all the queues - if the
2607 * queues contained any threads, they'll be garbage collected at the
2610 * ------------------------------------------------------------------------ */
2618 for (i=0; i<=MAX_PROC; i++) {
2619 run_queue_hds[i] = END_TSO_QUEUE;
2620 run_queue_tls[i] = END_TSO_QUEUE;
2621 blocked_queue_hds[i] = END_TSO_QUEUE;
2622 blocked_queue_tls[i] = END_TSO_QUEUE;
2623 ccalling_threadss[i] = END_TSO_QUEUE;
2624 blackhole_queue[i] = END_TSO_QUEUE;
2625 sleeping_queue = END_TSO_QUEUE;
2628 run_queue_hd = END_TSO_QUEUE;
2629 run_queue_tl = END_TSO_QUEUE;
2630 blocked_queue_hd = END_TSO_QUEUE;
2631 blocked_queue_tl = END_TSO_QUEUE;
2632 blackhole_queue = END_TSO_QUEUE;
2633 sleeping_queue = END_TSO_QUEUE;
2636 suspended_ccalling_threads = END_TSO_QUEUE;
2638 main_threads = NULL;
2639 all_threads = END_TSO_QUEUE;
2644 RtsFlags.ConcFlags.ctxtSwitchTicks =
2645 RtsFlags.ConcFlags.ctxtSwitchTime / TICK_MILLISECS;
2647 #if defined(RTS_SUPPORTS_THREADS)
2648 /* Initialise the mutex and condition variables used by
2650 initMutex(&sched_mutex);
2651 initMutex(&term_mutex);
2654 ACQUIRE_LOCK(&sched_mutex);
2656 /* A capability holds the state a native thread needs in
2657 * order to execute STG code. At least one capability is
2658 * floating around (only SMP builds have more than one).
2662 #if defined(RTS_SUPPORTS_THREADS)
2667 /* eagerly start some extra workers */
2668 startingWorkerThread = RtsFlags.ParFlags.nNodes;
2669 startTasks(RtsFlags.ParFlags.nNodes, taskStart);
2672 #if /* defined(SMP) ||*/ defined(PARALLEL_HASKELL)
2676 RELEASE_LOCK(&sched_mutex);
2680 exitScheduler( void )
2682 interrupted = rtsTrue;
2683 shutting_down_scheduler = rtsTrue;
2684 #if defined(RTS_SUPPORTS_THREADS)
2685 if (threadIsTask(osThreadId())) { taskStop(); }
2690 /* ----------------------------------------------------------------------------
2691 Managing the per-task allocation areas.
2693 Each capability comes with an allocation area. These are
2694 fixed-length block lists into which allocation can be done.
2696 ToDo: no support for two-space collection at the moment???
2697 ------------------------------------------------------------------------- */
2699 static SchedulerStatus
2700 waitThread_(StgMainThread* m, Capability *initialCapability)
2702 SchedulerStatus stat;
2704 // Precondition: sched_mutex must be held.
2705 IF_DEBUG(scheduler, sched_belch("new main thread (%d)", m->tso->id));
2708 /* GranSim specific init */
2709 CurrentTSO = m->tso; // the TSO to run
2710 procStatus[MainProc] = Busy; // status of main PE
2711 CurrentProc = MainProc; // PE to run it on
2712 schedule(m,initialCapability);
2714 schedule(m,initialCapability);
2715 ASSERT(m->stat != NoStatus);
2720 #if defined(RTS_SUPPORTS_THREADS)
2721 // Free the condition variable, returning it to the cache if possible.
2722 if (!bound_cond_cache_full) {
2723 bound_cond_cache = m->bound_thread_cond;
2724 bound_cond_cache_full = 1;
2726 closeCondition(&m->bound_thread_cond);
2730 IF_DEBUG(scheduler, sched_belch("main thread (%d) finished", m->tso->id));
2733 // Postcondition: sched_mutex still held
2737 /* ---------------------------------------------------------------------------
2738 Where are the roots that we know about?
2740 - all the threads on the runnable queue
2741 - all the threads on the blocked queue
2742 - all the threads on the sleeping queue
2743 - all the thread currently executing a _ccall_GC
2744 - all the "main threads"
2746 ------------------------------------------------------------------------ */
2748 /* This has to be protected either by the scheduler monitor, or by the
2749 garbage collection monitor (probably the latter).
2754 GetRoots( evac_fn evac )
2759 for (i=0; i<=RtsFlags.GranFlags.proc; i++) {
2760 if ((run_queue_hds[i] != END_TSO_QUEUE) && ((run_queue_hds[i] != NULL)))
2761 evac((StgClosure **)&run_queue_hds[i]);
2762 if ((run_queue_tls[i] != END_TSO_QUEUE) && ((run_queue_tls[i] != NULL)))
2763 evac((StgClosure **)&run_queue_tls[i]);
2765 if ((blocked_queue_hds[i] != END_TSO_QUEUE) && ((blocked_queue_hds[i] != NULL)))
2766 evac((StgClosure **)&blocked_queue_hds[i]);
2767 if ((blocked_queue_tls[i] != END_TSO_QUEUE) && ((blocked_queue_tls[i] != NULL)))
2768 evac((StgClosure **)&blocked_queue_tls[i]);
2769 if ((ccalling_threadss[i] != END_TSO_QUEUE) && ((ccalling_threadss[i] != NULL)))
2770 evac((StgClosure **)&ccalling_threads[i]);
2777 if (run_queue_hd != END_TSO_QUEUE) {
2778 ASSERT(run_queue_tl != END_TSO_QUEUE);
2779 evac((StgClosure **)&run_queue_hd);
2780 evac((StgClosure **)&run_queue_tl);
2783 if (blocked_queue_hd != END_TSO_QUEUE) {
2784 ASSERT(blocked_queue_tl != END_TSO_QUEUE);
2785 evac((StgClosure **)&blocked_queue_hd);
2786 evac((StgClosure **)&blocked_queue_tl);
2789 if (sleeping_queue != END_TSO_QUEUE) {
2790 evac((StgClosure **)&sleeping_queue);
2794 if (blackhole_queue != END_TSO_QUEUE) {
2795 evac((StgClosure **)&blackhole_queue);
2798 if (suspended_ccalling_threads != END_TSO_QUEUE) {
2799 evac((StgClosure **)&suspended_ccalling_threads);
2802 #if defined(PARALLEL_HASKELL) || defined(GRAN)
2803 markSparkQueue(evac);
2806 #if defined(RTS_USER_SIGNALS)
2807 // mark the signal handlers (signals should be already blocked)
2808 markSignalHandlers(evac);
2812 /* -----------------------------------------------------------------------------
2815 This is the interface to the garbage collector from Haskell land.
2816 We provide this so that external C code can allocate and garbage
2817 collect when called from Haskell via _ccall_GC.
2819 It might be useful to provide an interface whereby the programmer
2820 can specify more roots (ToDo).
2822 This needs to be protected by the GC condition variable above. KH.
2823 -------------------------------------------------------------------------- */
2825 static void (*extra_roots)(evac_fn);
2830 /* Obligated to hold this lock upon entry */
2831 ACQUIRE_LOCK(&sched_mutex);
2832 GarbageCollect(GetRoots,rtsFalse);
2833 RELEASE_LOCK(&sched_mutex);
2837 performMajorGC(void)
2839 ACQUIRE_LOCK(&sched_mutex);
2840 GarbageCollect(GetRoots,rtsTrue);
2841 RELEASE_LOCK(&sched_mutex);
2845 AllRoots(evac_fn evac)
2847 GetRoots(evac); // the scheduler's roots
2848 extra_roots(evac); // the user's roots
2852 performGCWithRoots(void (*get_roots)(evac_fn))
2854 ACQUIRE_LOCK(&sched_mutex);
2855 extra_roots = get_roots;
2856 GarbageCollect(AllRoots,rtsFalse);
2857 RELEASE_LOCK(&sched_mutex);
2860 /* -----------------------------------------------------------------------------
2863 If the thread has reached its maximum stack size, then raise the
2864 StackOverflow exception in the offending thread. Otherwise
2865 relocate the TSO into a larger chunk of memory and adjust its stack
2867 -------------------------------------------------------------------------- */
2870 threadStackOverflow(StgTSO *tso)
2872 nat new_stack_size, stack_words;
2877 IF_DEBUG(sanity,checkTSO(tso));
2878 if (tso->stack_size >= tso->max_stack_size) {
2881 debugBelch("@@ threadStackOverflow of TSO %ld (%p): stack too large (now %ld; max is %ld)\n",
2882 (long)tso->id, tso, (long)tso->stack_size, (long)tso->max_stack_size);
2883 /* If we're debugging, just print out the top of the stack */
2884 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2887 /* Send this thread the StackOverflow exception */
2888 raiseAsync(tso, (StgClosure *)stackOverflow_closure);
2892 /* Try to double the current stack size. If that takes us over the
2893 * maximum stack size for this thread, then use the maximum instead.
2894 * Finally round up so the TSO ends up as a whole number of blocks.
2896 new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
2897 new_tso_size = (lnat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) +
2898 TSO_STRUCT_SIZE)/sizeof(W_);
2899 new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */
2900 new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
2902 IF_DEBUG(scheduler, debugBelch("== sched: increasing stack size from %d words to %d.\n", tso->stack_size, new_stack_size));
2904 dest = (StgTSO *)allocate(new_tso_size);
2905 TICK_ALLOC_TSO(new_stack_size,0);
2907 /* copy the TSO block and the old stack into the new area */
2908 memcpy(dest,tso,TSO_STRUCT_SIZE);
2909 stack_words = tso->stack + tso->stack_size - tso->sp;
2910 new_sp = (P_)dest + new_tso_size - stack_words;
2911 memcpy(new_sp, tso->sp, stack_words * sizeof(W_));
2913 /* relocate the stack pointers... */
2915 dest->stack_size = new_stack_size;
2917 /* Mark the old TSO as relocated. We have to check for relocated
2918 * TSOs in the garbage collector and any primops that deal with TSOs.
2920 * It's important to set the sp value to just beyond the end
2921 * of the stack, so we don't attempt to scavenge any part of the
2924 tso->what_next = ThreadRelocated;
2926 tso->sp = (P_)&(tso->stack[tso->stack_size]);
2927 tso->why_blocked = NotBlocked;
2929 IF_PAR_DEBUG(verbose,
2930 debugBelch("@@ threadStackOverflow of TSO %d (now at %p): stack size increased to %ld\n",
2931 tso->id, tso, tso->stack_size);
2932 /* If we're debugging, just print out the top of the stack */
2933 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2936 IF_DEBUG(sanity,checkTSO(tso));
2938 IF_DEBUG(scheduler,printTSO(dest));
2944 /* ---------------------------------------------------------------------------
2945 Wake up a queue that was blocked on some resource.
2946 ------------------------------------------------------------------------ */
2950 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2953 #elif defined(PARALLEL_HASKELL)
2955 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2957 /* write RESUME events to log file and
2958 update blocked and fetch time (depending on type of the orig closure) */
2959 if (RtsFlags.ParFlags.ParStats.Full) {
2960 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
2961 GR_RESUMEQ, ((StgTSO *)bqe), ((StgTSO *)bqe)->block_info.closure,
2962 0, 0 /* spark_queue_len(ADVISORY_POOL) */);
2963 if (EMPTY_RUN_QUEUE())
2964 emitSchedule = rtsTrue;
2966 switch (get_itbl(node)->type) {
2968 ((StgTSO *)bqe)->par.fetchtime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2973 ((StgTSO *)bqe)->par.blocktime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2980 barf("{unblockOneLocked}Daq Qagh: unexpected closure in blocking queue");
2987 StgBlockingQueueElement *
2988 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2991 PEs node_loc, tso_loc;
2993 node_loc = where_is(node); // should be lifted out of loop
2994 tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2995 tso_loc = where_is((StgClosure *)tso);
2996 if (IS_LOCAL_TO(PROCS(node),tso_loc)) { // TSO is local
2997 /* !fake_fetch => TSO is on CurrentProc is same as IS_LOCAL_TO */
2998 ASSERT(CurrentProc!=node_loc || tso_loc==CurrentProc);
2999 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.lunblocktime;
3000 // insertThread(tso, node_loc);
3001 new_event(tso_loc, tso_loc, CurrentTime[CurrentProc],
3003 tso, node, (rtsSpark*)NULL);
3004 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
3007 } else { // TSO is remote (actually should be FMBQ)
3008 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.mpacktime +
3009 RtsFlags.GranFlags.Costs.gunblocktime +
3010 RtsFlags.GranFlags.Costs.latency;
3011 new_event(tso_loc, CurrentProc, CurrentTime[CurrentProc],
3013 tso, node, (rtsSpark*)NULL);
3014 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
3017 /* the thread-queue-overhead is accounted for in either Resume or UnblockThread */
3019 debugBelch(" %s TSO %d (%p) [PE %d] (block_info.closure=%p) (next=%p) ,",
3020 (node_loc==tso_loc ? "Local" : "Global"),
3021 tso->id, tso, CurrentProc, tso->block_info.closure, tso->link));
3022 tso->block_info.closure = NULL;
3023 IF_DEBUG(scheduler,debugBelch("-- Waking up thread %ld (%p)\n",
3026 #elif defined(PARALLEL_HASKELL)
3027 StgBlockingQueueElement *
3028 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
3030 StgBlockingQueueElement *next;
3032 switch (get_itbl(bqe)->type) {
3034 ASSERT(((StgTSO *)bqe)->why_blocked != NotBlocked);
3035 /* if it's a TSO just push it onto the run_queue */
3037 ((StgTSO *)bqe)->link = END_TSO_QUEUE; // debugging?
3038 APPEND_TO_RUN_QUEUE((StgTSO *)bqe);
3040 unblockCount(bqe, node);
3041 /* reset blocking status after dumping event */
3042 ((StgTSO *)bqe)->why_blocked = NotBlocked;
3046 /* if it's a BLOCKED_FETCH put it on the PendingFetches list */
3048 bqe->link = (StgBlockingQueueElement *)PendingFetches;
3049 PendingFetches = (StgBlockedFetch *)bqe;
3053 /* can ignore this case in a non-debugging setup;
3054 see comments on RBHSave closures above */
3056 /* check that the closure is an RBHSave closure */
3057 ASSERT(get_itbl((StgClosure *)bqe) == &stg_RBH_Save_0_info ||
3058 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_1_info ||
3059 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_2_info);
3063 barf("{unblockOneLocked}Daq Qagh: Unexpected IP (%#lx; %s) in blocking queue at %#lx\n",
3064 get_itbl((StgClosure *)bqe), info_type((StgClosure *)bqe),
3068 IF_PAR_DEBUG(bq, debugBelch(", %p (%s)\n", bqe, info_type((StgClosure*)bqe)));
3072 #else /* !GRAN && !PARALLEL_HASKELL */
3074 unblockOneLocked(StgTSO *tso)
3078 ASSERT(get_itbl(tso)->type == TSO);
3079 ASSERT(tso->why_blocked != NotBlocked);
3080 tso->why_blocked = NotBlocked;
3082 tso->link = END_TSO_QUEUE;
3083 APPEND_TO_RUN_QUEUE(tso);
3085 IF_DEBUG(scheduler,sched_belch("waking up thread %ld", (long)tso->id));
3090 #if defined(GRAN) || defined(PARALLEL_HASKELL)
3091 INLINE_ME StgBlockingQueueElement *
3092 unblockOne(StgBlockingQueueElement *bqe, StgClosure *node)
3094 ACQUIRE_LOCK(&sched_mutex);
3095 bqe = unblockOneLocked(bqe, node);
3096 RELEASE_LOCK(&sched_mutex);
3101 unblockOne(StgTSO *tso)
3103 ACQUIRE_LOCK(&sched_mutex);
3104 tso = unblockOneLocked(tso);
3105 RELEASE_LOCK(&sched_mutex);
3112 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
3114 StgBlockingQueueElement *bqe;
3119 debugBelch("##-_ AwBQ for node %p on PE %d @ %ld by TSO %d (%p): \n", \
3120 node, CurrentProc, CurrentTime[CurrentProc],
3121 CurrentTSO->id, CurrentTSO));
3123 node_loc = where_is(node);
3125 ASSERT(q == END_BQ_QUEUE ||
3126 get_itbl(q)->type == TSO || // q is either a TSO or an RBHSave
3127 get_itbl(q)->type == CONSTR); // closure (type constructor)
3128 ASSERT(is_unique(node));
3130 /* FAKE FETCH: magically copy the node to the tso's proc;
3131 no Fetch necessary because in reality the node should not have been
3132 moved to the other PE in the first place
3134 if (CurrentProc!=node_loc) {
3136 debugBelch("## node %p is on PE %d but CurrentProc is %d (TSO %d); assuming fake fetch and adjusting bitmask (old: %#x)\n",
3137 node, node_loc, CurrentProc, CurrentTSO->id,
3138 // CurrentTSO, where_is(CurrentTSO),
3139 node->header.gran.procs));
3140 node->header.gran.procs = (node->header.gran.procs) | PE_NUMBER(CurrentProc);
3142 debugBelch("## new bitmask of node %p is %#x\n",
3143 node, node->header.gran.procs));
3144 if (RtsFlags.GranFlags.GranSimStats.Global) {
3145 globalGranStats.tot_fake_fetches++;
3150 // ToDo: check: ASSERT(CurrentProc==node_loc);
3151 while (get_itbl(bqe)->type==TSO) { // q != END_TSO_QUEUE) {
3154 bqe points to the current element in the queue
3155 next points to the next element in the queue
3157 //tso = (StgTSO *)bqe; // wastes an assignment to get the type right
3158 //tso_loc = where_is(tso);
3160 bqe = unblockOneLocked(bqe, node);
3163 /* if this is the BQ of an RBH, we have to put back the info ripped out of
3164 the closure to make room for the anchor of the BQ */
3165 if (bqe!=END_BQ_QUEUE) {
3166 ASSERT(get_itbl(node)->type == RBH && get_itbl(bqe)->type == CONSTR);
3168 ASSERT((info_ptr==&RBH_Save_0_info) ||
3169 (info_ptr==&RBH_Save_1_info) ||
3170 (info_ptr==&RBH_Save_2_info));
3172 /* cf. convertToRBH in RBH.c for writing the RBHSave closure */
3173 ((StgRBH *)node)->blocking_queue = (StgBlockingQueueElement *)((StgRBHSave *)bqe)->payload[0];
3174 ((StgRBH *)node)->mut_link = (StgMutClosure *)((StgRBHSave *)bqe)->payload[1];
3177 debugBelch("## Filled in RBH_Save for %p (%s) at end of AwBQ\n",
3178 node, info_type(node)));
3181 /* statistics gathering */
3182 if (RtsFlags.GranFlags.GranSimStats.Global) {
3183 // globalGranStats.tot_bq_processing_time += bq_processing_time;
3184 globalGranStats.tot_bq_len += len; // total length of all bqs awakened
3185 // globalGranStats.tot_bq_len_local += len_local; // same for local TSOs only
3186 globalGranStats.tot_awbq++; // total no. of bqs awakened
3189 debugBelch("## BQ Stats of %p: [%d entries] %s\n",
3190 node, len, (bqe!=END_BQ_QUEUE) ? "RBH" : ""));
3192 #elif defined(PARALLEL_HASKELL)
3194 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
3196 StgBlockingQueueElement *bqe;
3198 ACQUIRE_LOCK(&sched_mutex);
3200 IF_PAR_DEBUG(verbose,
3201 debugBelch("##-_ AwBQ for node %p on [%x]: \n",
3205 if(get_itbl(q)->type == CONSTR || q==END_BQ_QUEUE) {
3206 IF_PAR_DEBUG(verbose, debugBelch("## ... nothing to unblock so lets just return. RFP (BUG?)\n"));
3211 ASSERT(q == END_BQ_QUEUE ||
3212 get_itbl(q)->type == TSO ||
3213 get_itbl(q)->type == BLOCKED_FETCH ||
3214 get_itbl(q)->type == CONSTR);
3217 while (get_itbl(bqe)->type==TSO ||
3218 get_itbl(bqe)->type==BLOCKED_FETCH) {
3219 bqe = unblockOneLocked(bqe, node);
3221 RELEASE_LOCK(&sched_mutex);
3224 #else /* !GRAN && !PARALLEL_HASKELL */
3227 awakenBlockedQueueNoLock(StgTSO *tso)
3229 while (tso != END_TSO_QUEUE) {
3230 tso = unblockOneLocked(tso);
3235 awakenBlockedQueue(StgTSO *tso)
3237 ACQUIRE_LOCK(&sched_mutex);
3238 while (tso != END_TSO_QUEUE) {
3239 tso = unblockOneLocked(tso);
3241 RELEASE_LOCK(&sched_mutex);
3245 /* ---------------------------------------------------------------------------
3247 - usually called inside a signal handler so it mustn't do anything fancy.
3248 ------------------------------------------------------------------------ */
3251 interruptStgRts(void)
3256 /* ToDo: if invoked from a signal handler, this threadRunnable
3257 * only works if there's another thread (not this one) waiting to
3262 /* -----------------------------------------------------------------------------
3265 This is for use when we raise an exception in another thread, which
3267 This has nothing to do with the UnblockThread event in GranSim. -- HWL
3268 -------------------------------------------------------------------------- */
3270 #if defined(GRAN) || defined(PARALLEL_HASKELL)
3272 NB: only the type of the blocking queue is different in GranSim and GUM
3273 the operations on the queue-elements are the same
3274 long live polymorphism!
3276 Locks: sched_mutex is held upon entry and exit.
3280 unblockThread(StgTSO *tso)
3282 StgBlockingQueueElement *t, **last;
3284 switch (tso->why_blocked) {
3287 return; /* not blocked */
3290 // Be careful: nothing to do here! We tell the scheduler that the thread
3291 // is runnable and we leave it to the stack-walking code to abort the
3292 // transaction while unwinding the stack. We should perhaps have a debugging
3293 // test to make sure that this really happens and that the 'zombie' transaction
3294 // does not get committed.
3298 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
3300 StgBlockingQueueElement *last_tso = END_BQ_QUEUE;
3301 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
3303 last = (StgBlockingQueueElement **)&mvar->head;
3304 for (t = (StgBlockingQueueElement *)mvar->head;
3306 last = &t->link, last_tso = t, t = t->link) {
3307 if (t == (StgBlockingQueueElement *)tso) {
3308 *last = (StgBlockingQueueElement *)tso->link;
3309 if (mvar->tail == tso) {
3310 mvar->tail = (StgTSO *)last_tso;
3315 barf("unblockThread (MVAR): TSO not found");
3318 case BlockedOnBlackHole:
3319 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
3321 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
3323 last = &bq->blocking_queue;
3324 for (t = bq->blocking_queue;
3326 last = &t->link, t = t->link) {
3327 if (t == (StgBlockingQueueElement *)tso) {
3328 *last = (StgBlockingQueueElement *)tso->link;
3332 barf("unblockThread (BLACKHOLE): TSO not found");
3335 case BlockedOnException:
3337 StgTSO *target = tso->block_info.tso;
3339 ASSERT(get_itbl(target)->type == TSO);
3341 if (target->what_next == ThreadRelocated) {
3342 target = target->link;
3343 ASSERT(get_itbl(target)->type == TSO);
3346 ASSERT(target->blocked_exceptions != NULL);
3348 last = (StgBlockingQueueElement **)&target->blocked_exceptions;
3349 for (t = (StgBlockingQueueElement *)target->blocked_exceptions;
3351 last = &t->link, t = t->link) {
3352 ASSERT(get_itbl(t)->type == TSO);
3353 if (t == (StgBlockingQueueElement *)tso) {
3354 *last = (StgBlockingQueueElement *)tso->link;
3358 barf("unblockThread (Exception): TSO not found");
3362 case BlockedOnWrite:
3363 #if defined(mingw32_HOST_OS)
3364 case BlockedOnDoProc:
3367 /* take TSO off blocked_queue */
3368 StgBlockingQueueElement *prev = NULL;
3369 for (t = (StgBlockingQueueElement *)blocked_queue_hd; t != END_BQ_QUEUE;
3370 prev = t, t = t->link) {
3371 if (t == (StgBlockingQueueElement *)tso) {
3373 blocked_queue_hd = (StgTSO *)t->link;
3374 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
3375 blocked_queue_tl = END_TSO_QUEUE;
3378 prev->link = t->link;
3379 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
3380 blocked_queue_tl = (StgTSO *)prev;
3383 #if defined(mingw32_HOST_OS)
3384 /* (Cooperatively) signal that the worker thread should abort
3387 abandonWorkRequest(tso->block_info.async_result->reqID);
3392 barf("unblockThread (I/O): TSO not found");
3395 case BlockedOnDelay:
3397 /* take TSO off sleeping_queue */
3398 StgBlockingQueueElement *prev = NULL;
3399 for (t = (StgBlockingQueueElement *)sleeping_queue; t != END_BQ_QUEUE;
3400 prev = t, t = t->link) {
3401 if (t == (StgBlockingQueueElement *)tso) {
3403 sleeping_queue = (StgTSO *)t->link;
3405 prev->link = t->link;
3410 barf("unblockThread (delay): TSO not found");
3414 barf("unblockThread");
3418 tso->link = END_TSO_QUEUE;
3419 tso->why_blocked = NotBlocked;
3420 tso->block_info.closure = NULL;
3421 PUSH_ON_RUN_QUEUE(tso);
3425 unblockThread(StgTSO *tso)
3429 /* To avoid locking unnecessarily. */
3430 if (tso->why_blocked == NotBlocked) {
3434 switch (tso->why_blocked) {
3437 // Be careful: nothing to do here! We tell the scheduler that the thread
3438 // is runnable and we leave it to the stack-walking code to abort the
3439 // transaction while unwinding the stack. We should perhaps have a debugging
3440 // test to make sure that this really happens and that the 'zombie' transaction
3441 // does not get committed.
3445 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
3447 StgTSO *last_tso = END_TSO_QUEUE;
3448 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
3451 for (t = mvar->head; t != END_TSO_QUEUE;
3452 last = &t->link, last_tso = t, t = t->link) {
3455 if (mvar->tail == tso) {
3456 mvar->tail = last_tso;
3461 barf("unblockThread (MVAR): TSO not found");
3464 case BlockedOnBlackHole:
3466 last = &blackhole_queue;
3467 for (t = blackhole_queue; t != END_TSO_QUEUE;
3468 last = &t->link, t = t->link) {
3474 barf("unblockThread (BLACKHOLE): TSO not found");
3477 case BlockedOnException:
3479 StgTSO *target = tso->block_info.tso;
3481 ASSERT(get_itbl(target)->type == TSO);
3483 while (target->what_next == ThreadRelocated) {
3484 target = target->link;
3485 ASSERT(get_itbl(target)->type == TSO);
3488 ASSERT(target->blocked_exceptions != NULL);
3490 last = &target->blocked_exceptions;
3491 for (t = target->blocked_exceptions; t != END_TSO_QUEUE;
3492 last = &t->link, t = t->link) {
3493 ASSERT(get_itbl(t)->type == TSO);
3499 barf("unblockThread (Exception): TSO not found");
3503 case BlockedOnWrite:
3504 #if defined(mingw32_HOST_OS)
3505 case BlockedOnDoProc:
3508 StgTSO *prev = NULL;
3509 for (t = blocked_queue_hd; t != END_TSO_QUEUE;
3510 prev = t, t = t->link) {
3513 blocked_queue_hd = t->link;
3514 if (blocked_queue_tl == t) {
3515 blocked_queue_tl = END_TSO_QUEUE;
3518 prev->link = t->link;
3519 if (blocked_queue_tl == t) {
3520 blocked_queue_tl = prev;
3523 #if defined(mingw32_HOST_OS)
3524 /* (Cooperatively) signal that the worker thread should abort
3527 abandonWorkRequest(tso->block_info.async_result->reqID);
3532 barf("unblockThread (I/O): TSO not found");
3535 case BlockedOnDelay:
3537 StgTSO *prev = NULL;
3538 for (t = sleeping_queue; t != END_TSO_QUEUE;
3539 prev = t, t = t->link) {
3542 sleeping_queue = t->link;
3544 prev->link = t->link;
3549 barf("unblockThread (delay): TSO not found");
3553 barf("unblockThread");
3557 tso->link = END_TSO_QUEUE;
3558 tso->why_blocked = NotBlocked;
3559 tso->block_info.closure = NULL;
3560 APPEND_TO_RUN_QUEUE(tso);
3564 /* -----------------------------------------------------------------------------
3567 * Check the blackhole_queue for threads that can be woken up. We do
3568 * this periodically: before every GC, and whenever the run queue is
3571 * An elegant solution might be to just wake up all the blocked
3572 * threads with awakenBlockedQueue occasionally: they'll go back to
3573 * sleep again if the object is still a BLACKHOLE. Unfortunately this
3574 * doesn't give us a way to tell whether we've actually managed to
3575 * wake up any threads, so we would be busy-waiting.
3577 * -------------------------------------------------------------------------- */
3580 checkBlackHoles( void )
3583 rtsBool any_woke_up = rtsFalse;
3586 IF_DEBUG(scheduler, sched_belch("checking threads blocked on black holes"));
3588 // ASSUMES: sched_mutex
3589 prev = &blackhole_queue;
3590 t = blackhole_queue;
3591 while (t != END_TSO_QUEUE) {
3592 ASSERT(t->why_blocked == BlockedOnBlackHole);
3593 type = get_itbl(t->block_info.closure)->type;
3594 if (type != BLACKHOLE && type != CAF_BLACKHOLE) {
3595 t = unblockOneLocked(t);
3597 any_woke_up = rtsTrue;
3607 /* -----------------------------------------------------------------------------
3610 * The following function implements the magic for raising an
3611 * asynchronous exception in an existing thread.
3613 * We first remove the thread from any queue on which it might be
3614 * blocked. The possible blockages are MVARs and BLACKHOLE_BQs.
3616 * We strip the stack down to the innermost CATCH_FRAME, building
3617 * thunks in the heap for all the active computations, so they can
3618 * be restarted if necessary. When we reach a CATCH_FRAME, we build
3619 * an application of the handler to the exception, and push it on
3620 * the top of the stack.
3622 * How exactly do we save all the active computations? We create an
3623 * AP_STACK for every UpdateFrame on the stack. Entering one of these
3624 * AP_STACKs pushes everything from the corresponding update frame
3625 * upwards onto the stack. (Actually, it pushes everything up to the
3626 * next update frame plus a pointer to the next AP_STACK object.
3627 * Entering the next AP_STACK object pushes more onto the stack until we
3628 * reach the last AP_STACK object - at which point the stack should look
3629 * exactly as it did when we killed the TSO and we can continue
3630 * execution by entering the closure on top of the stack.
3632 * We can also kill a thread entirely - this happens if either (a) the
3633 * exception passed to raiseAsync is NULL, or (b) there's no
3634 * CATCH_FRAME on the stack. In either case, we strip the entire
3635 * stack and replace the thread with a zombie.
3637 * Locks: sched_mutex held upon entry nor exit.
3639 * -------------------------------------------------------------------------- */
3642 deleteThread(StgTSO *tso)
3644 if (tso->why_blocked != BlockedOnCCall &&
3645 tso->why_blocked != BlockedOnCCall_NoUnblockExc) {
3646 raiseAsync(tso,NULL);
3650 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
3652 deleteThreadImmediately(StgTSO *tso)
3653 { // for forkProcess only:
3654 // delete thread without giving it a chance to catch the KillThread exception
3656 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
3660 if (tso->why_blocked != BlockedOnCCall &&
3661 tso->why_blocked != BlockedOnCCall_NoUnblockExc) {
3665 tso->what_next = ThreadKilled;
3670 raiseAsyncWithLock(StgTSO *tso, StgClosure *exception)
3672 /* When raising async exs from contexts where sched_mutex isn't held;
3673 use raiseAsyncWithLock(). */
3674 ACQUIRE_LOCK(&sched_mutex);
3675 raiseAsync(tso,exception);
3676 RELEASE_LOCK(&sched_mutex);
3680 raiseAsync(StgTSO *tso, StgClosure *exception)
3682 raiseAsync_(tso, exception, rtsFalse);
3686 raiseAsync_(StgTSO *tso, StgClosure *exception, rtsBool stop_at_atomically)
3688 StgRetInfoTable *info;
3691 // Thread already dead?
3692 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
3697 sched_belch("raising exception in thread %ld.", (long)tso->id));
3699 // Remove it from any blocking queues
3704 // The stack freezing code assumes there's a closure pointer on
3705 // the top of the stack, so we have to arrange that this is the case...
3707 if (sp[0] == (W_)&stg_enter_info) {
3711 sp[0] = (W_)&stg_dummy_ret_closure;
3717 // 1. Let the top of the stack be the "current closure"
3719 // 2. Walk up the stack until we find either an UPDATE_FRAME or a
3722 // 3. If it's an UPDATE_FRAME, then make an AP_STACK containing the
3723 // current closure applied to the chunk of stack up to (but not
3724 // including) the update frame. This closure becomes the "current
3725 // closure". Go back to step 2.
3727 // 4. If it's a CATCH_FRAME, then leave the exception handler on
3728 // top of the stack applied to the exception.
3730 // 5. If it's a STOP_FRAME, then kill the thread.
3732 // NB: if we pass an ATOMICALLY_FRAME then abort the associated
3739 info = get_ret_itbl((StgClosure *)frame);
3741 while (info->i.type != UPDATE_FRAME
3742 && (info->i.type != CATCH_FRAME || exception == NULL)
3743 && info->i.type != STOP_FRAME
3744 && (info->i.type != ATOMICALLY_FRAME || stop_at_atomically == rtsFalse))
3746 if (info->i.type == CATCH_RETRY_FRAME || info->i.type == ATOMICALLY_FRAME) {
3747 // IF we find an ATOMICALLY_FRAME then we abort the
3748 // current transaction and propagate the exception. In
3749 // this case (unlike ordinary exceptions) we do not care
3750 // whether the transaction is valid or not because its
3751 // possible validity cannot have caused the exception
3752 // and will not be visible after the abort.
3754 debugBelch("Found atomically block delivering async exception\n"));
3755 stmAbortTransaction(tso -> trec);
3756 tso -> trec = stmGetEnclosingTRec(tso -> trec);
3758 frame += stack_frame_sizeW((StgClosure *)frame);
3759 info = get_ret_itbl((StgClosure *)frame);
3762 switch (info->i.type) {
3764 case ATOMICALLY_FRAME:
3765 ASSERT(stop_at_atomically);
3766 ASSERT(stmGetEnclosingTRec(tso->trec) == NO_TREC);
3767 stmCondemnTransaction(tso -> trec);
3771 // R1 is not a register: the return convention for IO in
3772 // this case puts the return value on the stack, so we
3773 // need to set up the stack to return to the atomically
3774 // frame properly...
3775 tso->sp = frame - 2;
3776 tso->sp[1] = (StgWord) &stg_NO_FINALIZER_closure; // why not?
3777 tso->sp[0] = (StgWord) &stg_ut_1_0_unreg_info;
3779 tso->what_next = ThreadRunGHC;
3783 // If we find a CATCH_FRAME, and we've got an exception to raise,
3784 // then build the THUNK raise(exception), and leave it on
3785 // top of the CATCH_FRAME ready to enter.
3789 StgCatchFrame *cf = (StgCatchFrame *)frame;
3793 // we've got an exception to raise, so let's pass it to the
3794 // handler in this frame.
3796 raise = (StgThunk *)allocate(sizeofW(StgThunk)+1);
3797 TICK_ALLOC_SE_THK(1,0);
3798 SET_HDR(raise,&stg_raise_info,cf->header.prof.ccs);
3799 raise->payload[0] = exception;
3801 // throw away the stack from Sp up to the CATCH_FRAME.
3805 /* Ensure that async excpetions are blocked now, so we don't get
3806 * a surprise exception before we get around to executing the
3809 if (tso->blocked_exceptions == NULL) {
3810 tso->blocked_exceptions = END_TSO_QUEUE;
3813 /* Put the newly-built THUNK on top of the stack, ready to execute
3814 * when the thread restarts.
3817 sp[-1] = (W_)&stg_enter_info;
3819 tso->what_next = ThreadRunGHC;
3820 IF_DEBUG(sanity, checkTSO(tso));
3829 // First build an AP_STACK consisting of the stack chunk above the
3830 // current update frame, with the top word on the stack as the
3833 words = frame - sp - 1;
3834 ap = (StgAP_STACK *)allocate(AP_STACK_sizeW(words));
3837 ap->fun = (StgClosure *)sp[0];
3839 for(i=0; i < (nat)words; ++i) {
3840 ap->payload[i] = (StgClosure *)*sp++;
3843 SET_HDR(ap,&stg_AP_STACK_info,
3844 ((StgClosure *)frame)->header.prof.ccs /* ToDo */);
3845 TICK_ALLOC_UP_THK(words+1,0);
3848 debugBelch("sched: Updating ");
3849 printPtr((P_)((StgUpdateFrame *)frame)->updatee);
3850 debugBelch(" with ");
3851 printObj((StgClosure *)ap);
3854 // Replace the updatee with an indirection - happily
3855 // this will also wake up any threads currently
3856 // waiting on the result.
3858 // Warning: if we're in a loop, more than one update frame on
3859 // the stack may point to the same object. Be careful not to
3860 // overwrite an IND_OLDGEN in this case, because we'll screw
3861 // up the mutable lists. To be on the safe side, don't
3862 // overwrite any kind of indirection at all. See also
3863 // threadSqueezeStack in GC.c, where we have to make a similar
3866 if (!closure_IND(((StgUpdateFrame *)frame)->updatee)) {
3867 // revert the black hole
3868 UPD_IND_NOLOCK(((StgUpdateFrame *)frame)->updatee,
3871 sp += sizeofW(StgUpdateFrame) - 1;
3872 sp[0] = (W_)ap; // push onto stack
3877 // We've stripped the entire stack, the thread is now dead.
3878 sp += sizeofW(StgStopFrame);
3879 tso->what_next = ThreadKilled;
3890 /* -----------------------------------------------------------------------------
3891 raiseExceptionHelper
3893 This function is called by the raise# primitve, just so that we can
3894 move some of the tricky bits of raising an exception from C-- into
3895 C. Who knows, it might be a useful re-useable thing here too.
3896 -------------------------------------------------------------------------- */
3899 raiseExceptionHelper (StgTSO *tso, StgClosure *exception)
3901 StgThunk *raise_closure = NULL;
3903 StgRetInfoTable *info;
3905 // This closure represents the expression 'raise# E' where E
3906 // is the exception raise. It is used to overwrite all the
3907 // thunks which are currently under evaluataion.
3911 // LDV profiling: stg_raise_info has THUNK as its closure
3912 // type. Since a THUNK takes at least MIN_UPD_SIZE words in its
3913 // payload, MIN_UPD_SIZE is more approprate than 1. It seems that
3914 // 1 does not cause any problem unless profiling is performed.
3915 // However, when LDV profiling goes on, we need to linearly scan
3916 // small object pool, where raise_closure is stored, so we should
3917 // use MIN_UPD_SIZE.
3919 // raise_closure = (StgClosure *)RET_STGCALL1(P_,allocate,
3920 // sizeofW(StgClosure)+1);
3924 // Walk up the stack, looking for the catch frame. On the way,
3925 // we update any closures pointed to from update frames with the
3926 // raise closure that we just built.
3930 info = get_ret_itbl((StgClosure *)p);
3931 next = p + stack_frame_sizeW((StgClosure *)p);
3932 switch (info->i.type) {
3935 // Only create raise_closure if we need to.
3936 if (raise_closure == NULL) {
3938 (StgThunk *)allocate(sizeofW(StgThunk)+MIN_UPD_SIZE);
3939 SET_HDR(raise_closure, &stg_raise_info, CCCS);
3940 raise_closure->payload[0] = exception;
3942 UPD_IND(((StgUpdateFrame *)p)->updatee,(StgClosure *)raise_closure);
3946 case ATOMICALLY_FRAME:
3947 IF_DEBUG(stm, debugBelch("Found ATOMICALLY_FRAME at %p\n", p));
3949 return ATOMICALLY_FRAME;
3955 case CATCH_STM_FRAME:
3956 IF_DEBUG(stm, debugBelch("Found CATCH_STM_FRAME at %p\n", p));
3958 return CATCH_STM_FRAME;
3964 case CATCH_RETRY_FRAME:
3973 /* -----------------------------------------------------------------------------
3974 findRetryFrameHelper
3976 This function is called by the retry# primitive. It traverses the stack
3977 leaving tso->sp referring to the frame which should handle the retry.
3979 This should either be a CATCH_RETRY_FRAME (if the retry# is within an orElse#)
3980 or should be a ATOMICALLY_FRAME (if the retry# reaches the top level).
3982 We skip CATCH_STM_FRAMEs because retries are not considered to be exceptions,
3983 despite the similar implementation.
3985 We should not expect to see CATCH_FRAME or STOP_FRAME because those should
3986 not be created within memory transactions.
3987 -------------------------------------------------------------------------- */
3990 findRetryFrameHelper (StgTSO *tso)
3993 StgRetInfoTable *info;
3997 info = get_ret_itbl((StgClosure *)p);
3998 next = p + stack_frame_sizeW((StgClosure *)p);
3999 switch (info->i.type) {
4001 case ATOMICALLY_FRAME:
4002 IF_DEBUG(stm, debugBelch("Found ATOMICALLY_FRAME at %p during retrry\n", p));
4004 return ATOMICALLY_FRAME;
4006 case CATCH_RETRY_FRAME:
4007 IF_DEBUG(stm, debugBelch("Found CATCH_RETRY_FRAME at %p during retrry\n", p));
4009 return CATCH_RETRY_FRAME;
4011 case CATCH_STM_FRAME:
4013 ASSERT(info->i.type != CATCH_FRAME);
4014 ASSERT(info->i.type != STOP_FRAME);
4021 /* -----------------------------------------------------------------------------
4022 resurrectThreads is called after garbage collection on the list of
4023 threads found to be garbage. Each of these threads will be woken
4024 up and sent a signal: BlockedOnDeadMVar if the thread was blocked
4025 on an MVar, or NonTermination if the thread was blocked on a Black
4028 Locks: sched_mutex isn't held upon entry nor exit.
4029 -------------------------------------------------------------------------- */
4032 resurrectThreads( StgTSO *threads )
4036 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
4037 next = tso->global_link;
4038 tso->global_link = all_threads;
4040 IF_DEBUG(scheduler, sched_belch("resurrecting thread %d", tso->id));
4042 switch (tso->why_blocked) {
4044 case BlockedOnException:
4045 /* Called by GC - sched_mutex lock is currently held. */
4046 raiseAsync(tso,(StgClosure *)BlockedOnDeadMVar_closure);
4048 case BlockedOnBlackHole:
4049 raiseAsync(tso,(StgClosure *)NonTermination_closure);
4052 raiseAsync(tso,(StgClosure *)BlockedIndefinitely_closure);
4055 /* This might happen if the thread was blocked on a black hole
4056 * belonging to a thread that we've just woken up (raiseAsync
4057 * can wake up threads, remember...).
4061 barf("resurrectThreads: thread blocked in a strange way");
4066 /* ----------------------------------------------------------------------------
4067 * Debugging: why is a thread blocked
4068 * [Also provides useful information when debugging threaded programs
4069 * at the Haskell source code level, so enable outside of DEBUG. --sof 7/02]
4070 ------------------------------------------------------------------------- */
4073 printThreadBlockage(StgTSO *tso)
4075 switch (tso->why_blocked) {
4077 debugBelch("is blocked on read from fd %d", (int)(tso->block_info.fd));
4079 case BlockedOnWrite:
4080 debugBelch("is blocked on write to fd %d", (int)(tso->block_info.fd));
4082 #if defined(mingw32_HOST_OS)
4083 case BlockedOnDoProc:
4084 debugBelch("is blocked on proc (request: %ld)", tso->block_info.async_result->reqID);
4087 case BlockedOnDelay:
4088 debugBelch("is blocked until %ld", (long)(tso->block_info.target));
4091 debugBelch("is blocked on an MVar");
4093 case BlockedOnException:
4094 debugBelch("is blocked on delivering an exception to thread %d",
4095 tso->block_info.tso->id);
4097 case BlockedOnBlackHole:
4098 debugBelch("is blocked on a black hole");
4101 debugBelch("is not blocked");
4103 #if defined(PARALLEL_HASKELL)
4105 debugBelch("is blocked on global address; local FM_BQ is %p (%s)",
4106 tso->block_info.closure, info_type(tso->block_info.closure));
4108 case BlockedOnGA_NoSend:
4109 debugBelch("is blocked on global address (no send); local FM_BQ is %p (%s)",
4110 tso->block_info.closure, info_type(tso->block_info.closure));
4113 case BlockedOnCCall:
4114 debugBelch("is blocked on an external call");
4116 case BlockedOnCCall_NoUnblockExc:
4117 debugBelch("is blocked on an external call (exceptions were already blocked)");
4120 debugBelch("is blocked on an STM operation");
4123 barf("printThreadBlockage: strange tso->why_blocked: %d for TSO %d (%d)",
4124 tso->why_blocked, tso->id, tso);
4129 printThreadStatus(StgTSO *tso)
4131 switch (tso->what_next) {
4133 debugBelch("has been killed");
4135 case ThreadComplete:
4136 debugBelch("has completed");
4139 printThreadBlockage(tso);
4144 printAllThreads(void)
4149 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
4150 ullong_format_string(TIME_ON_PROC(CurrentProc),
4151 time_string, rtsFalse/*no commas!*/);
4153 debugBelch("all threads at [%s]:\n", time_string);
4154 # elif defined(PARALLEL_HASKELL)
4155 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
4156 ullong_format_string(CURRENT_TIME,
4157 time_string, rtsFalse/*no commas!*/);
4159 debugBelch("all threads at [%s]:\n", time_string);
4161 debugBelch("all threads:\n");
4164 for (t = all_threads; t != END_TSO_QUEUE; ) {
4165 debugBelch("\tthread %d @ %p ", t->id, (void *)t);
4168 void *label = lookupThreadLabel(t->id);
4169 if (label) debugBelch("[\"%s\"] ",(char *)label);
4172 if (t->what_next == ThreadRelocated) {
4173 debugBelch("has been relocated...\n");
4176 printThreadStatus(t);
4186 Print a whole blocking queue attached to node (debugging only).
4188 # if defined(PARALLEL_HASKELL)
4190 print_bq (StgClosure *node)
4192 StgBlockingQueueElement *bqe;
4196 debugBelch("## BQ of closure %p (%s): ",
4197 node, info_type(node));
4199 /* should cover all closures that may have a blocking queue */
4200 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
4201 get_itbl(node)->type == FETCH_ME_BQ ||
4202 get_itbl(node)->type == RBH ||
4203 get_itbl(node)->type == MVAR);
4205 ASSERT(node!=(StgClosure*)NULL); // sanity check
4207 print_bqe(((StgBlockingQueue*)node)->blocking_queue);
4211 Print a whole blocking queue starting with the element bqe.
4214 print_bqe (StgBlockingQueueElement *bqe)
4219 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
4221 for (end = (bqe==END_BQ_QUEUE);
4222 !end; // iterate until bqe points to a CONSTR
4223 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE),
4224 bqe = end ? END_BQ_QUEUE : bqe->link) {
4225 ASSERT(bqe != END_BQ_QUEUE); // sanity check
4226 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
4227 /* types of closures that may appear in a blocking queue */
4228 ASSERT(get_itbl(bqe)->type == TSO ||
4229 get_itbl(bqe)->type == BLOCKED_FETCH ||
4230 get_itbl(bqe)->type == CONSTR);
4231 /* only BQs of an RBH end with an RBH_Save closure */
4232 //ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
4234 switch (get_itbl(bqe)->type) {
4236 debugBelch(" TSO %u (%x),",
4237 ((StgTSO *)bqe)->id, ((StgTSO *)bqe));
4240 debugBelch(" BF (node=%p, ga=((%x, %d, %x)),",
4241 ((StgBlockedFetch *)bqe)->node,
4242 ((StgBlockedFetch *)bqe)->ga.payload.gc.gtid,
4243 ((StgBlockedFetch *)bqe)->ga.payload.gc.slot,
4244 ((StgBlockedFetch *)bqe)->ga.weight);
4247 debugBelch(" %s (IP %p),",
4248 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
4249 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
4250 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
4251 "RBH_Save_?"), get_itbl(bqe));
4254 barf("Unexpected closure type %s in blocking queue", // of %p (%s)",
4255 info_type((StgClosure *)bqe)); // , node, info_type(node));
4261 # elif defined(GRAN)
4263 print_bq (StgClosure *node)
4265 StgBlockingQueueElement *bqe;
4266 PEs node_loc, tso_loc;
4269 /* should cover all closures that may have a blocking queue */
4270 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
4271 get_itbl(node)->type == FETCH_ME_BQ ||
4272 get_itbl(node)->type == RBH);
4274 ASSERT(node!=(StgClosure*)NULL); // sanity check
4275 node_loc = where_is(node);
4277 debugBelch("## BQ of closure %p (%s) on [PE %d]: ",
4278 node, info_type(node), node_loc);
4281 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
4283 for (bqe = ((StgBlockingQueue*)node)->blocking_queue, end = (bqe==END_BQ_QUEUE);
4284 !end; // iterate until bqe points to a CONSTR
4285 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) {
4286 ASSERT(bqe != END_BQ_QUEUE); // sanity check
4287 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
4288 /* types of closures that may appear in a blocking queue */
4289 ASSERT(get_itbl(bqe)->type == TSO ||
4290 get_itbl(bqe)->type == CONSTR);
4291 /* only BQs of an RBH end with an RBH_Save closure */
4292 ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
4294 tso_loc = where_is((StgClosure *)bqe);
4295 switch (get_itbl(bqe)->type) {
4297 debugBelch(" TSO %d (%p) on [PE %d],",
4298 ((StgTSO *)bqe)->id, (StgTSO *)bqe, tso_loc);
4301 debugBelch(" %s (IP %p),",
4302 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
4303 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
4304 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
4305 "RBH_Save_?"), get_itbl(bqe));
4308 barf("Unexpected closure type %s in blocking queue of %p (%s)",
4309 info_type((StgClosure *)bqe), node, info_type(node));
4317 #if defined(PARALLEL_HASKELL)
4324 for (i=0, tso=run_queue_hd;
4325 tso != END_TSO_QUEUE;
4334 sched_belch(char *s, ...)
4338 #ifdef RTS_SUPPORTS_THREADS
4339 debugBelch("sched (task %p): ", osThreadId());
4340 #elif defined(PARALLEL_HASKELL)
4343 debugBelch("sched: ");