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(rtsBool force_major);
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);
308 void printThreadQueue(StgTSO *tso);
310 #if defined(PARALLEL_HASKELL)
311 StgTSO * createSparkThread(rtsSpark spark);
312 StgTSO * activateSpark (rtsSpark spark);
315 /* ----------------------------------------------------------------------------
317 * ------------------------------------------------------------------------- */
319 #if defined(RTS_SUPPORTS_THREADS)
320 static nat startingWorkerThread = 0;
325 ACQUIRE_LOCK(&sched_mutex);
326 startingWorkerThread--;
329 RELEASE_LOCK(&sched_mutex);
333 startSchedulerTaskIfNecessary(void)
335 if ( !EMPTY_RUN_QUEUE()
336 && !shutting_down_scheduler // not if we're shutting down
337 && startingWorkerThread==0)
339 // we don't want to start another worker thread
340 // just because the last one hasn't yet reached the
341 // "waiting for capability" state
342 startingWorkerThread++;
343 if (!maybeStartNewWorker(taskStart)) {
344 startingWorkerThread--;
350 /* -----------------------------------------------------------------------------
351 * Putting a thread on the run queue: different scheduling policies
352 * -------------------------------------------------------------------------- */
355 addToRunQueue( StgTSO *t )
357 #if defined(PARALLEL_HASKELL)
358 if (RtsFlags.ParFlags.doFairScheduling) {
359 // this does round-robin scheduling; good for concurrency
360 APPEND_TO_RUN_QUEUE(t);
362 // this does unfair scheduling; good for parallelism
363 PUSH_ON_RUN_QUEUE(t);
366 // this does round-robin scheduling; good for concurrency
367 APPEND_TO_RUN_QUEUE(t);
371 /* ---------------------------------------------------------------------------
372 Main scheduling loop.
374 We use round-robin scheduling, each thread returning to the
375 scheduler loop when one of these conditions is detected:
378 * timer expires (thread yields)
383 Locking notes: we acquire the scheduler lock once at the beginning
384 of the scheduler loop, and release it when
386 * running a thread, or
387 * waiting for work, or
388 * waiting for a GC to complete.
391 In a GranSim setup this loop iterates over the global event queue.
392 This revolves around the global event queue, which determines what
393 to do next. Therefore, it's more complicated than either the
394 concurrent or the parallel (GUM) setup.
397 GUM iterates over incoming messages.
398 It starts with nothing to do (thus CurrentTSO == END_TSO_QUEUE),
399 and sends out a fish whenever it has nothing to do; in-between
400 doing the actual reductions (shared code below) it processes the
401 incoming messages and deals with delayed operations
402 (see PendingFetches).
403 This is not the ugliest code you could imagine, but it's bloody close.
405 ------------------------------------------------------------------------ */
408 schedule( StgMainThread *mainThread USED_WHEN_RTS_SUPPORTS_THREADS,
409 Capability *initialCapability )
413 StgThreadReturnCode ret;
416 #elif defined(PARALLEL_HASKELL)
419 rtsBool receivedFinish = rtsFalse;
421 nat tp_size, sp_size; // stats only
427 // Pre-condition: sched_mutex is held.
428 // We might have a capability, passed in as initialCapability.
429 cap = initialCapability;
431 #if !defined(RTS_SUPPORTS_THREADS)
432 // simply initialise it in the non-threaded case
433 grabCapability(&cap);
437 sched_belch("### NEW SCHEDULER LOOP (main thr: %p, cap: %p)",
438 mainThread, initialCapability);
443 // -----------------------------------------------------------
444 // Scheduler loop starts here:
446 #if defined(PARALLEL_HASKELL)
447 #define TERMINATION_CONDITION (!receivedFinish)
449 #define TERMINATION_CONDITION ((event = get_next_event()) != (rtsEvent*)NULL)
451 #define TERMINATION_CONDITION rtsTrue
454 while (TERMINATION_CONDITION) {
457 /* Choose the processor with the next event */
458 CurrentProc = event->proc;
459 CurrentTSO = event->tso;
462 #if defined(RTS_SUPPORTS_THREADS)
463 // Yield the capability to higher-priority tasks if necessary.
466 yieldCapability(&cap,
467 mainThread ? &mainThread->bound_thread_cond : NULL );
470 // If we do not currently hold a capability, we wait for one
473 waitForCapability(&sched_mutex, &cap,
474 mainThread ? &mainThread->bound_thread_cond : NULL);
477 // We now have a capability...
480 #if 0 /* extra sanity checking */
483 for (m = main_threads; m != NULL; m = m->link) {
484 ASSERT(get_itbl(m->tso)->type == TSO);
489 // Check whether we have re-entered the RTS from Haskell without
490 // going via suspendThread()/resumeThread (i.e. a 'safe' foreign
492 if (cap->r.rInHaskell) {
493 errorBelch("schedule: re-entered unsafely.\n"
494 " Perhaps a 'foreign import unsafe' should be 'safe'?");
499 // Test for interruption. If interrupted==rtsTrue, then either
500 // we received a keyboard interrupt (^C), or the scheduler is
501 // trying to shut down all the tasks (shutting_down_scheduler) in
505 if (shutting_down_scheduler) {
506 IF_DEBUG(scheduler, sched_belch("shutting down"));
507 releaseCapability(cap);
509 mainThread->stat = Interrupted;
510 mainThread->ret = NULL;
514 IF_DEBUG(scheduler, sched_belch("interrupted"));
519 #if defined(not_yet) && defined(SMP)
521 // Top up the run queue from our spark pool. We try to make the
522 // number of threads in the run queue equal to the number of
523 // free capabilities.
527 if (EMPTY_RUN_QUEUE()) {
528 spark = findSpark(rtsFalse);
530 break; /* no more sparks in the pool */
532 createSparkThread(spark);
534 sched_belch("==^^ turning spark of closure %p into a thread",
535 (StgClosure *)spark));
541 scheduleStartSignalHandlers();
543 // Only check the black holes here if we've nothing else to do.
544 // During normal execution, the black hole list only gets checked
545 // at GC time, to avoid repeatedly traversing this possibly long
546 // list each time around the scheduler.
547 if (EMPTY_RUN_QUEUE()) { scheduleCheckBlackHoles(); }
549 scheduleCheckBlockedThreads();
551 scheduleDetectDeadlock();
553 // Normally, the only way we can get here with no threads to
554 // run is if a keyboard interrupt received during
555 // scheduleCheckBlockedThreads() or scheduleDetectDeadlock().
556 // Additionally, it is not fatal for the
557 // threaded RTS to reach here with no threads to run.
559 // win32: might be here due to awaitEvent() being abandoned
560 // as a result of a console event having been delivered.
561 if ( EMPTY_RUN_QUEUE() ) {
562 #if !defined(RTS_SUPPORTS_THREADS) && !defined(mingw32_HOST_OS)
565 continue; // nothing to do
568 #if defined(PARALLEL_HASKELL)
569 scheduleSendPendingMessages();
570 if (EMPTY_RUN_QUEUE() && scheduleActivateSpark())
574 ASSERT(next_fish_to_send_at==0); // i.e. no delayed fishes left!
577 /* If we still have no work we need to send a FISH to get a spark
579 if (EMPTY_RUN_QUEUE()) {
580 if (!scheduleGetRemoteWork(&receivedFinish)) continue;
581 ASSERT(rtsFalse); // should not happen at the moment
583 // from here: non-empty run queue.
584 // TODO: merge above case with this, only one call processMessages() !
585 if (PacketsWaiting()) { /* process incoming messages, if
586 any pending... only in else
587 because getRemoteWork waits for
589 receivedFinish = processMessages();
594 scheduleProcessEvent(event);
598 // Get a thread to run
600 ASSERT(run_queue_hd != END_TSO_QUEUE);
603 #if defined(GRAN) || defined(PAR)
604 scheduleGranParReport(); // some kind of debuging output
606 // Sanity check the thread we're about to run. This can be
607 // expensive if there is lots of thread switching going on...
608 IF_DEBUG(sanity,checkTSO(t));
611 #if defined(RTS_SUPPORTS_THREADS)
612 // Check whether we can run this thread in the current task.
613 // If not, we have to pass our capability to the right task.
615 StgMainThread *m = t->main;
622 sched_belch("### Running thread %d in bound thread", t->id));
623 // yes, the Haskell thread is bound to the current native thread
628 sched_belch("### thread %d bound to another OS thread", t->id));
629 // no, bound to a different Haskell thread: pass to that thread
630 PUSH_ON_RUN_QUEUE(t);
636 if(mainThread != NULL)
637 // The thread we want to run is unbound.
640 sched_belch("### this OS thread cannot run thread %d", t->id));
641 // no, the current native thread is bound to a different
642 // Haskell thread, so pass it to any worker thread
643 PUSH_ON_RUN_QUEUE(t);
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 cap->r.rInHaskell = rtsFalse;
712 // The TSO might have moved, eg. if it re-entered the RTS and a GC
713 // happened. So find the new location:
714 t = cap->r.rCurrentTSO;
716 // And save the current errno in this thread.
717 t->saved_errno = errno;
719 // ----------------------------------------------------------------------
721 /* Costs for the scheduler are assigned to CCS_SYSTEM */
722 #if defined(PROFILING)
727 ACQUIRE_LOCK(&sched_mutex);
729 // We have run some Haskell code: there might be blackhole-blocked
730 // threads to wake up now.
731 if ( blackhole_queue != END_TSO_QUEUE ) {
732 blackholes_need_checking = rtsTrue;
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(rtsFalse); }
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 );
850 /* ----------------------------------------------------------------------------
851 * Check for threads blocked on BLACKHOLEs that can be woken up
852 * ASSUMES: sched_mutex
853 * ------------------------------------------------------------------------- */
855 scheduleCheckBlackHoles( void )
857 if ( blackholes_need_checking )
860 blackholes_need_checking = rtsFalse;
864 /* ----------------------------------------------------------------------------
865 * Detect deadlock conditions and attempt to resolve them.
866 * ASSUMES: sched_mutex
867 * ------------------------------------------------------------------------- */
870 scheduleDetectDeadlock()
873 #if defined(PARALLEL_HASKELL)
874 // ToDo: add deadlock detection in GUM (similar to SMP) -- HWL
879 * Detect deadlock: when we have no threads to run, there are no
880 * threads blocked, waiting for I/O, or sleeping, and all the
881 * other tasks are waiting for work, we must have a deadlock of
884 if ( EMPTY_THREAD_QUEUES() )
886 #if defined(RTS_SUPPORTS_THREADS)
888 * In the threaded RTS, we only check for deadlock if there
889 * has been no activity in a complete timeslice. This means
890 * we won't eagerly start a full GC just because we don't have
891 * any threads to run currently.
893 if (recent_activity != ACTIVITY_INACTIVE) return;
896 IF_DEBUG(scheduler, sched_belch("deadlocked, forcing major GC..."));
898 // Garbage collection can release some new threads due to
899 // either (a) finalizers or (b) threads resurrected because
900 // they are unreachable and will therefore be sent an
901 // exception. Any threads thus released will be immediately
904 scheduleDoGC( rtsTrue/*force major GC*/ );
905 recent_activity = ACTIVITY_DONE_GC;
906 if ( !EMPTY_RUN_QUEUE() ) return;
908 #if defined(RTS_USER_SIGNALS) && !defined(RTS_SUPPORTS_THREADS)
909 /* If we have user-installed signal handlers, then wait
910 * for signals to arrive rather then bombing out with a
913 if ( anyUserHandlers() ) {
915 sched_belch("still deadlocked, waiting for signals..."));
919 if (signals_pending()) {
920 RELEASE_LOCK(&sched_mutex);
921 startSignalHandlers();
922 ACQUIRE_LOCK(&sched_mutex);
925 // either we have threads to run, or we were interrupted:
926 ASSERT(!EMPTY_RUN_QUEUE() || interrupted);
930 #if !defined(RTS_SUPPORTS_THREADS)
931 /* Probably a real deadlock. Send the current main thread the
932 * Deadlock exception (or in the SMP build, send *all* main
933 * threads the deadlock exception, since none of them can make
939 switch (m->tso->why_blocked) {
941 case BlockedOnBlackHole:
942 case BlockedOnException:
944 raiseAsync(m->tso, (StgClosure *)NonTermination_closure);
947 barf("deadlock: main thread blocked in a strange way");
954 /* ----------------------------------------------------------------------------
955 * Process an event (GRAN only)
956 * ------------------------------------------------------------------------- */
960 scheduleProcessEvent(rtsEvent *event)
964 if (RtsFlags.GranFlags.Light)
965 GranSimLight_enter_system(event, &ActiveTSO); // adjust ActiveTSO etc
967 /* adjust time based on time-stamp */
968 if (event->time > CurrentTime[CurrentProc] &&
969 event->evttype != ContinueThread)
970 CurrentTime[CurrentProc] = event->time;
972 /* Deal with the idle PEs (may issue FindWork or MoveSpark events) */
973 if (!RtsFlags.GranFlags.Light)
976 IF_DEBUG(gran, debugBelch("GRAN: switch by event-type\n"));
978 /* main event dispatcher in GranSim */
979 switch (event->evttype) {
980 /* Should just be continuing execution */
982 IF_DEBUG(gran, debugBelch("GRAN: doing ContinueThread\n"));
983 /* ToDo: check assertion
984 ASSERT(run_queue_hd != (StgTSO*)NULL &&
985 run_queue_hd != END_TSO_QUEUE);
987 /* Ignore ContinueThreads for fetching threads (if synchr comm) */
988 if (!RtsFlags.GranFlags.DoAsyncFetch &&
989 procStatus[CurrentProc]==Fetching) {
990 debugBelch("ghuH: Spurious ContinueThread while Fetching ignored; TSO %d (%p) [PE %d]\n",
991 CurrentTSO->id, CurrentTSO, CurrentProc);
994 /* Ignore ContinueThreads for completed threads */
995 if (CurrentTSO->what_next == ThreadComplete) {
996 debugBelch("ghuH: found a ContinueThread event for completed thread %d (%p) [PE %d] (ignoring ContinueThread)\n",
997 CurrentTSO->id, CurrentTSO, CurrentProc);
1000 /* Ignore ContinueThreads for threads that are being migrated */
1001 if (PROCS(CurrentTSO)==Nowhere) {
1002 debugBelch("ghuH: trying to run the migrating TSO %d (%p) [PE %d] (ignoring ContinueThread)\n",
1003 CurrentTSO->id, CurrentTSO, CurrentProc);
1006 /* The thread should be at the beginning of the run queue */
1007 if (CurrentTSO!=run_queue_hds[CurrentProc]) {
1008 debugBelch("ghuH: TSO %d (%p) [PE %d] is not at the start of the run_queue when doing a ContinueThread\n",
1009 CurrentTSO->id, CurrentTSO, CurrentProc);
1010 break; // run the thread anyway
1013 new_event(proc, proc, CurrentTime[proc],
1015 (StgTSO*)NULL, (StgClosure*)NULL, (rtsSpark*)NULL);
1017 */ /* Catches superfluous CONTINUEs -- should be unnecessary */
1018 break; // now actually run the thread; DaH Qu'vam yImuHbej
1021 do_the_fetchnode(event);
1022 goto next_thread; /* handle next event in event queue */
1025 do_the_globalblock(event);
1026 goto next_thread; /* handle next event in event queue */
1029 do_the_fetchreply(event);
1030 goto next_thread; /* handle next event in event queue */
1032 case UnblockThread: /* Move from the blocked queue to the tail of */
1033 do_the_unblock(event);
1034 goto next_thread; /* handle next event in event queue */
1036 case ResumeThread: /* Move from the blocked queue to the tail of */
1037 /* the runnable queue ( i.e. Qu' SImqa'lu') */
1038 event->tso->gran.blocktime +=
1039 CurrentTime[CurrentProc] - event->tso->gran.blockedat;
1040 do_the_startthread(event);
1041 goto next_thread; /* handle next event in event queue */
1044 do_the_startthread(event);
1045 goto next_thread; /* handle next event in event queue */
1048 do_the_movethread(event);
1049 goto next_thread; /* handle next event in event queue */
1052 do_the_movespark(event);
1053 goto next_thread; /* handle next event in event queue */
1056 do_the_findwork(event);
1057 goto next_thread; /* handle next event in event queue */
1060 barf("Illegal event type %u\n", event->evttype);
1063 /* This point was scheduler_loop in the old RTS */
1065 IF_DEBUG(gran, debugBelch("GRAN: after main switch\n"));
1067 TimeOfLastEvent = CurrentTime[CurrentProc];
1068 TimeOfNextEvent = get_time_of_next_event();
1069 IgnoreEvents=(TimeOfNextEvent==0); // HWL HACK
1070 // CurrentTSO = ThreadQueueHd;
1072 IF_DEBUG(gran, debugBelch("GRAN: time of next event is: %ld\n",
1075 if (RtsFlags.GranFlags.Light)
1076 GranSimLight_leave_system(event, &ActiveTSO);
1078 EndOfTimeSlice = CurrentTime[CurrentProc]+RtsFlags.GranFlags.time_slice;
1081 debugBelch("GRAN: end of time-slice is %#lx\n", EndOfTimeSlice));
1083 /* in a GranSim setup the TSO stays on the run queue */
1085 /* Take a thread from the run queue. */
1086 POP_RUN_QUEUE(t); // take_off_run_queue(t);
1089 debugBelch("GRAN: About to run current thread, which is\n");
1092 context_switch = 0; // turned on via GranYield, checking events and time slice
1095 DumpGranEvent(GR_SCHEDULE, t));
1097 procStatus[CurrentProc] = Busy;
1101 /* ----------------------------------------------------------------------------
1102 * Send pending messages (PARALLEL_HASKELL only)
1103 * ------------------------------------------------------------------------- */
1105 #if defined(PARALLEL_HASKELL)
1107 scheduleSendPendingMessages(void)
1113 # if defined(PAR) // global Mem.Mgmt., omit for now
1114 if (PendingFetches != END_BF_QUEUE) {
1119 if (RtsFlags.ParFlags.BufferTime) {
1120 // if we use message buffering, we must send away all message
1121 // packets which have become too old...
1127 /* ----------------------------------------------------------------------------
1128 * Activate spark threads (PARALLEL_HASKELL only)
1129 * ------------------------------------------------------------------------- */
1131 #if defined(PARALLEL_HASKELL)
1133 scheduleActivateSpark(void)
1136 ASSERT(EMPTY_RUN_QUEUE());
1137 /* We get here if the run queue is empty and want some work.
1138 We try to turn a spark into a thread, and add it to the run queue,
1139 from where it will be picked up in the next iteration of the scheduler
1143 /* :-[ no local threads => look out for local sparks */
1144 /* the spark pool for the current PE */
1145 pool = &(cap.r.rSparks); // JB: cap = (old) MainCap
1146 if (advisory_thread_count < RtsFlags.ParFlags.maxThreads &&
1147 pool->hd < pool->tl) {
1149 * ToDo: add GC code check that we really have enough heap afterwards!!
1151 * If we're here (no runnable threads) and we have pending
1152 * sparks, we must have a space problem. Get enough space
1153 * to turn one of those pending sparks into a
1157 spark = findSpark(rtsFalse); /* get a spark */
1158 if (spark != (rtsSpark) NULL) {
1159 tso = createThreadFromSpark(spark); /* turn the spark into a thread */
1160 IF_PAR_DEBUG(fish, // schedule,
1161 debugBelch("==== schedule: Created TSO %d (%p); %d threads active\n",
1162 tso->id, tso, advisory_thread_count));
1164 if (tso==END_TSO_QUEUE) { /* failed to activate spark->back to loop */
1165 IF_PAR_DEBUG(fish, // schedule,
1166 debugBelch("==^^ failed to create thread from spark @ %lx\n",
1168 return rtsFalse; /* failed to generate a thread */
1169 } /* otherwise fall through & pick-up new tso */
1171 IF_PAR_DEBUG(fish, // schedule,
1172 debugBelch("==^^ no local sparks (spark pool contains only NFs: %d)\n",
1173 spark_queue_len(pool)));
1174 return rtsFalse; /* failed to generate a thread */
1176 return rtsTrue; /* success in generating a thread */
1177 } else { /* no more threads permitted or pool empty */
1178 return rtsFalse; /* failed to generateThread */
1181 tso = NULL; // avoid compiler warning only
1182 return rtsFalse; /* dummy in non-PAR setup */
1185 #endif // PARALLEL_HASKELL
1187 /* ----------------------------------------------------------------------------
1188 * Get work from a remote node (PARALLEL_HASKELL only)
1189 * ------------------------------------------------------------------------- */
1191 #if defined(PARALLEL_HASKELL)
1193 scheduleGetRemoteWork(rtsBool *receivedFinish)
1195 ASSERT(EMPTY_RUN_QUEUE());
1197 if (RtsFlags.ParFlags.BufferTime) {
1198 IF_PAR_DEBUG(verbose,
1199 debugBelch("...send all pending data,"));
1202 for (i=1; i<=nPEs; i++)
1203 sendImmediately(i); // send all messages away immediately
1207 //++EDEN++ idle() , i.e. send all buffers, wait for work
1208 // suppress fishing in EDEN... just look for incoming messages
1209 // (blocking receive)
1210 IF_PAR_DEBUG(verbose,
1211 debugBelch("...wait for incoming messages...\n"));
1212 *receivedFinish = processMessages(); // blocking receive...
1214 // and reenter scheduling loop after having received something
1215 // (return rtsFalse below)
1217 # else /* activate SPARKS machinery */
1218 /* We get here, if we have no work, tried to activate a local spark, but still
1219 have no work. We try to get a remote spark, by sending a FISH message.
1220 Thread migration should be added here, and triggered when a sequence of
1221 fishes returns without work. */
1222 delay = (RtsFlags.ParFlags.fishDelay!=0ll ? RtsFlags.ParFlags.fishDelay : 0ll);
1224 /* =8-[ no local sparks => look for work on other PEs */
1226 * We really have absolutely no work. Send out a fish
1227 * (there may be some out there already), and wait for
1228 * something to arrive. We clearly can't run any threads
1229 * until a SCHEDULE or RESUME arrives, and so that's what
1230 * we're hoping to see. (Of course, we still have to
1231 * respond to other types of messages.)
1233 rtsTime now = msTime() /*CURRENT_TIME*/;
1234 IF_PAR_DEBUG(verbose,
1235 debugBelch("-- now=%ld\n", now));
1236 IF_PAR_DEBUG(fish, // verbose,
1237 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
1238 (last_fish_arrived_at!=0 &&
1239 last_fish_arrived_at+delay > now)) {
1240 debugBelch("--$$ <%llu> delaying FISH until %llu (last fish %llu, delay %llu)\n",
1241 now, last_fish_arrived_at+delay,
1242 last_fish_arrived_at,
1246 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
1247 advisory_thread_count < RtsFlags.ParFlags.maxThreads) { // send a FISH, but when?
1248 if (last_fish_arrived_at==0 ||
1249 (last_fish_arrived_at+delay <= now)) { // send FISH now!
1250 /* outstandingFishes is set in sendFish, processFish;
1251 avoid flooding system with fishes via delay */
1252 next_fish_to_send_at = 0;
1254 /* ToDo: this should be done in the main scheduling loop to avoid the
1255 busy wait here; not so bad if fish delay is very small */
1256 int iq = 0; // DEBUGGING -- HWL
1257 next_fish_to_send_at = last_fish_arrived_at+delay; // remember when to send
1258 /* send a fish when ready, but process messages that arrive in the meantime */
1260 if (PacketsWaiting()) {
1262 *receivedFinish = processMessages();
1265 } while (!*receivedFinish || now<next_fish_to_send_at);
1266 // JB: This means the fish could become obsolete, if we receive
1267 // work. Better check for work again?
1268 // last line: while (!receivedFinish || !haveWork || now<...)
1269 // next line: if (receivedFinish || haveWork )
1271 if (*receivedFinish) // no need to send a FISH if we are finishing anyway
1272 return rtsFalse; // NB: this will leave scheduler loop
1273 // immediately after return!
1275 IF_PAR_DEBUG(fish, // verbose,
1276 debugBelch("--$$ <%llu> sent delayed fish (%d processMessages); active/total threads=%d/%d\n",now,iq,run_queue_len(),advisory_thread_count));
1280 // JB: IMHO, this should all be hidden inside sendFish(...)
1282 sendFish(pe, thisPE, NEW_FISH_AGE, NEW_FISH_HISTORY,
1285 // Global statistics: count no. of fishes
1286 if (RtsFlags.ParFlags.ParStats.Global &&
1287 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
1288 globalParStats.tot_fish_mess++;
1292 /* delayed fishes must have been sent by now! */
1293 next_fish_to_send_at = 0;
1296 *receivedFinish = processMessages();
1297 # endif /* SPARKS */
1300 /* NB: this function always returns rtsFalse, meaning the scheduler
1301 loop continues with the next iteration;
1303 return code means success in finding work; we enter this function
1304 if there is no local work, thus have to send a fish which takes
1305 time until it arrives with work; in the meantime we should process
1306 messages in the main loop;
1309 #endif // PARALLEL_HASKELL
1311 /* ----------------------------------------------------------------------------
1312 * PAR/GRAN: Report stats & debugging info(?)
1313 * ------------------------------------------------------------------------- */
1315 #if defined(PAR) || defined(GRAN)
1317 scheduleGranParReport(void)
1319 ASSERT(run_queue_hd != END_TSO_QUEUE);
1321 /* Take a thread from the run queue, if we have work */
1322 POP_RUN_QUEUE(t); // take_off_run_queue(END_TSO_QUEUE);
1324 /* If this TSO has got its outport closed in the meantime,
1325 * it mustn't be run. Instead, we have to clean it up as if it was finished.
1326 * It has to be marked as TH_DEAD for this purpose.
1327 * If it is TH_TERM instead, it is supposed to have finished in the normal way.
1329 JB: TODO: investigate wether state change field could be nuked
1330 entirely and replaced by the normal tso state (whatnext
1331 field). All we want to do is to kill tsos from outside.
1334 /* ToDo: write something to the log-file
1335 if (RTSflags.ParFlags.granSimStats && !sameThread)
1336 DumpGranEvent(GR_SCHEDULE, RunnableThreadsHd);
1340 /* the spark pool for the current PE */
1341 pool = &(cap.r.rSparks); // cap = (old) MainCap
1344 debugBelch("--=^ %d threads, %d sparks on [%#x]\n",
1345 run_queue_len(), spark_queue_len(pool), CURRENT_PROC));
1348 debugBelch("--=^ %d threads, %d sparks on [%#x]\n",
1349 run_queue_len(), spark_queue_len(pool), CURRENT_PROC));
1351 if (RtsFlags.ParFlags.ParStats.Full &&
1352 (t->par.sparkname != (StgInt)0) && // only log spark generated threads
1353 (emitSchedule || // forced emit
1354 (t && LastTSO && t->id != LastTSO->id))) {
1356 we are running a different TSO, so write a schedule event to log file
1357 NB: If we use fair scheduling we also have to write a deschedule
1358 event for LastTSO; with unfair scheduling we know that the
1359 previous tso has blocked whenever we switch to another tso, so
1360 we don't need it in GUM for now
1362 IF_PAR_DEBUG(fish, // schedule,
1363 debugBelch("____ scheduling spark generated thread %d (%lx) (%lx) via a forced emit\n",t->id,t,t->par.sparkname));
1365 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
1366 GR_SCHEDULE, t, (StgClosure *)NULL, 0, 0);
1367 emitSchedule = rtsFalse;
1372 /* ----------------------------------------------------------------------------
1373 * After running a thread...
1374 * ASSUMES: sched_mutex
1375 * ------------------------------------------------------------------------- */
1378 schedulePostRunThread(void)
1381 /* HACK 675: if the last thread didn't yield, make sure to print a
1382 SCHEDULE event to the log file when StgRunning the next thread, even
1383 if it is the same one as before */
1385 TimeOfLastYield = CURRENT_TIME;
1388 /* some statistics gathering in the parallel case */
1390 #if defined(GRAN) || defined(PAR) || defined(EDEN)
1394 IF_DEBUG(gran, DumpGranEvent(GR_DESCHEDULE, t));
1395 globalGranStats.tot_heapover++;
1397 globalParStats.tot_heapover++;
1404 DumpGranEvent(GR_DESCHEDULE, t));
1405 globalGranStats.tot_stackover++;
1408 // DumpGranEvent(GR_DESCHEDULE, t);
1409 globalParStats.tot_stackover++;
1413 case ThreadYielding:
1416 DumpGranEvent(GR_DESCHEDULE, t));
1417 globalGranStats.tot_yields++;
1420 // DumpGranEvent(GR_DESCHEDULE, t);
1421 globalParStats.tot_yields++;
1428 debugBelch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: ",
1429 t->id, t, whatNext_strs[t->what_next], t->block_info.closure,
1430 (t->block_info.closure==(StgClosure*)NULL ? 99 : where_is(t->block_info.closure)));
1431 if (t->block_info.closure!=(StgClosure*)NULL)
1432 print_bq(t->block_info.closure);
1435 // ??? needed; should emit block before
1437 DumpGranEvent(GR_DESCHEDULE, t));
1438 prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
1441 ASSERT(procStatus[CurrentProc]==Busy ||
1442 ((procStatus[CurrentProc]==Fetching) &&
1443 (t->block_info.closure!=(StgClosure*)NULL)));
1444 if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
1445 !(!RtsFlags.GranFlags.DoAsyncFetch &&
1446 procStatus[CurrentProc]==Fetching))
1447 procStatus[CurrentProc] = Idle;
1450 //++PAR++ blockThread() writes the event (change?)
1454 case ThreadFinished:
1458 barf("parGlobalStats: unknown return code");
1464 /* -----------------------------------------------------------------------------
1465 * Handle a thread that returned to the scheduler with ThreadHeepOverflow
1466 * ASSUMES: sched_mutex
1467 * -------------------------------------------------------------------------- */
1470 scheduleHandleHeapOverflow( Capability *cap, StgTSO *t )
1472 // did the task ask for a large block?
1473 if (cap->r.rHpAlloc > BLOCK_SIZE) {
1474 // if so, get one and push it on the front of the nursery.
1478 blocks = (lnat)BLOCK_ROUND_UP(cap->r.rHpAlloc) / BLOCK_SIZE;
1481 debugBelch("--<< thread %ld (%s) stopped: requesting a large block (size %ld)\n",
1482 (long)t->id, whatNext_strs[t->what_next], blocks));
1484 // don't do this if the nursery is (nearly) full, we'll GC first.
1485 if (cap->r.rCurrentNursery->link != NULL ||
1486 cap->r.rNursery->n_blocks == 1) { // paranoia to prevent infinite loop
1487 // if the nursery has only one block.
1490 bd = allocGroup( blocks );
1492 cap->r.rNursery->n_blocks += blocks;
1494 // link the new group into the list
1495 bd->link = cap->r.rCurrentNursery;
1496 bd->u.back = cap->r.rCurrentNursery->u.back;
1497 if (cap->r.rCurrentNursery->u.back != NULL) {
1498 cap->r.rCurrentNursery->u.back->link = bd;
1501 ASSERT(g0s0->blocks == cap->r.rCurrentNursery &&
1502 g0s0 == cap->r.rNursery);
1504 cap->r.rNursery->blocks = bd;
1506 cap->r.rCurrentNursery->u.back = bd;
1508 // initialise it as a nursery block. We initialise the
1509 // step, gen_no, and flags field of *every* sub-block in
1510 // this large block, because this is easier than making
1511 // sure that we always find the block head of a large
1512 // block whenever we call Bdescr() (eg. evacuate() and
1513 // isAlive() in the GC would both have to do this, at
1517 for (x = bd; x < bd + blocks; x++) {
1518 x->step = cap->r.rNursery;
1524 // This assert can be a killer if the app is doing lots
1525 // of large block allocations.
1526 IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
1528 // now update the nursery to point to the new block
1529 cap->r.rCurrentNursery = bd;
1531 // we might be unlucky and have another thread get on the
1532 // run queue before us and steal the large block, but in that
1533 // case the thread will just end up requesting another large
1535 PUSH_ON_RUN_QUEUE(t);
1536 return rtsFalse; /* not actually GC'ing */
1541 debugBelch("--<< thread %ld (%s) stopped: HeapOverflow\n",
1542 (long)t->id, whatNext_strs[t->what_next]));
1544 ASSERT(!is_on_queue(t,CurrentProc));
1545 #elif defined(PARALLEL_HASKELL)
1546 /* Currently we emit a DESCHEDULE event before GC in GUM.
1547 ToDo: either add separate event to distinguish SYSTEM time from rest
1548 or just nuke this DESCHEDULE (and the following SCHEDULE) */
1549 if (0 && RtsFlags.ParFlags.ParStats.Full) {
1550 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
1551 GR_DESCHEDULE, t, (StgClosure *)NULL, 0, 0);
1552 emitSchedule = rtsTrue;
1556 PUSH_ON_RUN_QUEUE(t);
1558 /* actual GC is done at the end of the while loop in schedule() */
1561 /* -----------------------------------------------------------------------------
1562 * Handle a thread that returned to the scheduler with ThreadStackOverflow
1563 * ASSUMES: sched_mutex
1564 * -------------------------------------------------------------------------- */
1567 scheduleHandleStackOverflow( StgTSO *t)
1569 IF_DEBUG(scheduler,debugBelch("--<< thread %ld (%s) stopped, StackOverflow\n",
1570 (long)t->id, whatNext_strs[t->what_next]));
1571 /* just adjust the stack for this thread, then pop it back
1575 /* enlarge the stack */
1576 StgTSO *new_t = threadStackOverflow(t);
1578 /* This TSO has moved, so update any pointers to it from the
1579 * main thread stack. It better not be on any other queues...
1580 * (it shouldn't be).
1582 if (t->main != NULL) {
1583 t->main->tso = new_t;
1585 PUSH_ON_RUN_QUEUE(new_t);
1589 /* -----------------------------------------------------------------------------
1590 * Handle a thread that returned to the scheduler with ThreadYielding
1591 * ASSUMES: sched_mutex
1592 * -------------------------------------------------------------------------- */
1595 scheduleHandleYield( StgTSO *t, nat prev_what_next )
1597 // Reset the context switch flag. We don't do this just before
1598 // running the thread, because that would mean we would lose ticks
1599 // during GC, which can lead to unfair scheduling (a thread hogs
1600 // the CPU because the tick always arrives during GC). This way
1601 // penalises threads that do a lot of allocation, but that seems
1602 // better than the alternative.
1605 /* put the thread back on the run queue. Then, if we're ready to
1606 * GC, check whether this is the last task to stop. If so, wake
1607 * up the GC thread. getThread will block during a GC until the
1611 if (t->what_next != prev_what_next) {
1612 debugBelch("--<< thread %ld (%s) stopped to switch evaluators\n",
1613 (long)t->id, whatNext_strs[t->what_next]);
1615 debugBelch("--<< thread %ld (%s) stopped, yielding\n",
1616 (long)t->id, whatNext_strs[t->what_next]);
1621 //debugBelch("&& Doing sanity check on yielding TSO %ld.", t->id);
1623 ASSERT(t->link == END_TSO_QUEUE);
1625 // Shortcut if we're just switching evaluators: don't bother
1626 // doing stack squeezing (which can be expensive), just run the
1628 if (t->what_next != prev_what_next) {
1633 ASSERT(!is_on_queue(t,CurrentProc));
1636 //debugBelch("&& Doing sanity check on all ThreadQueues (and their TSOs).");
1637 checkThreadQsSanity(rtsTrue));
1644 /* add a ContinueThread event to actually process the thread */
1645 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1647 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1649 debugBelch("GRAN: eventq and runnableq after adding yielded thread to queue again:\n");
1656 /* -----------------------------------------------------------------------------
1657 * Handle a thread that returned to the scheduler with ThreadBlocked
1658 * ASSUMES: sched_mutex
1659 * -------------------------------------------------------------------------- */
1662 scheduleHandleThreadBlocked( StgTSO *t
1663 #if !defined(GRAN) && !defined(DEBUG)
1670 debugBelch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: \n",
1671 t->id, t, whatNext_strs[t->what_next], t->block_info.closure, (t->block_info.closure==(StgClosure*)NULL ? 99 : where_is(t->block_info.closure)));
1672 if (t->block_info.closure!=(StgClosure*)NULL) print_bq(t->block_info.closure));
1674 // ??? needed; should emit block before
1676 DumpGranEvent(GR_DESCHEDULE, t));
1677 prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
1680 ASSERT(procStatus[CurrentProc]==Busy ||
1681 ((procStatus[CurrentProc]==Fetching) &&
1682 (t->block_info.closure!=(StgClosure*)NULL)));
1683 if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
1684 !(!RtsFlags.GranFlags.DoAsyncFetch &&
1685 procStatus[CurrentProc]==Fetching))
1686 procStatus[CurrentProc] = Idle;
1690 debugBelch("--<< thread %ld (%p; %s) stopped, blocking on node %p with BQ: \n",
1691 t->id, t, whatNext_strs[t->what_next], t->block_info.closure));
1694 if (t->block_info.closure!=(StgClosure*)NULL)
1695 print_bq(t->block_info.closure));
1697 /* Send a fetch (if BlockedOnGA) and dump event to log file */
1700 /* whatever we schedule next, we must log that schedule */
1701 emitSchedule = rtsTrue;
1705 // We don't need to do anything. The thread is blocked, and it
1706 // has tidied up its stack and placed itself on whatever queue
1707 // it needs to be on.
1710 ASSERT(t->why_blocked != NotBlocked);
1711 // This might not be true under SMP: we don't have
1712 // exclusive access to this TSO, so someone might have
1713 // woken it up by now. This actually happens: try
1714 // conc023 +RTS -N2.
1718 debugBelch("--<< thread %d (%s) stopped: ",
1719 t->id, whatNext_strs[t->what_next]);
1720 printThreadBlockage(t);
1723 /* Only for dumping event to log file
1724 ToDo: do I need this in GranSim, too?
1730 /* -----------------------------------------------------------------------------
1731 * Handle a thread that returned to the scheduler with ThreadFinished
1732 * ASSUMES: sched_mutex
1733 * -------------------------------------------------------------------------- */
1736 scheduleHandleThreadFinished( StgMainThread *mainThread
1737 USED_WHEN_RTS_SUPPORTS_THREADS,
1741 /* Need to check whether this was a main thread, and if so,
1742 * return with the return value.
1744 * We also end up here if the thread kills itself with an
1745 * uncaught exception, see Exception.cmm.
1747 IF_DEBUG(scheduler,debugBelch("--++ thread %d (%s) finished\n",
1748 t->id, whatNext_strs[t->what_next]));
1751 endThread(t, CurrentProc); // clean-up the thread
1752 #elif defined(PARALLEL_HASKELL)
1753 /* For now all are advisory -- HWL */
1754 //if(t->priority==AdvisoryPriority) ??
1755 advisory_thread_count--; // JB: Caution with this counter, buggy!
1758 if(t->dist.priority==RevalPriority)
1762 # if defined(EDENOLD)
1763 // the thread could still have an outport... (BUG)
1764 if (t->eden.outport != -1) {
1765 // delete the outport for the tso which has finished...
1766 IF_PAR_DEBUG(eden_ports,
1767 debugBelch("WARNING: Scheduler removes outport %d for TSO %d.\n",
1768 t->eden.outport, t->id));
1771 // thread still in the process (HEAVY BUG! since outport has just been closed...)
1772 if (t->eden.epid != -1) {
1773 IF_PAR_DEBUG(eden_ports,
1774 debugBelch("WARNING: Scheduler removes TSO %d from process %d .\n",
1775 t->id, t->eden.epid));
1776 removeTSOfromProcess(t);
1781 if (RtsFlags.ParFlags.ParStats.Full &&
1782 !RtsFlags.ParFlags.ParStats.Suppressed)
1783 DumpEndEvent(CURRENT_PROC, t, rtsFalse /* not mandatory */);
1785 // t->par only contains statistics: left out for now...
1787 debugBelch("**** end thread: ended sparked thread %d (%lx); sparkname: %lx\n",
1788 t->id,t,t->par.sparkname));
1790 #endif // PARALLEL_HASKELL
1793 // Check whether the thread that just completed was a main
1794 // thread, and if so return with the result.
1796 // There is an assumption here that all thread completion goes
1797 // through this point; we need to make sure that if a thread
1798 // ends up in the ThreadKilled state, that it stays on the run
1799 // queue so it can be dealt with here.
1802 #if defined(RTS_SUPPORTS_THREADS)
1805 mainThread->tso == t
1809 // We are a bound thread: this must be our thread that just
1811 ASSERT(mainThread->tso == t);
1813 if (t->what_next == ThreadComplete) {
1814 if (mainThread->ret) {
1815 // NOTE: return val is tso->sp[1] (see StgStartup.hc)
1816 *(mainThread->ret) = (StgClosure *)mainThread->tso->sp[1];
1818 mainThread->stat = Success;
1820 if (mainThread->ret) {
1821 *(mainThread->ret) = NULL;
1824 mainThread->stat = Interrupted;
1826 mainThread->stat = Killed;
1830 removeThreadLabel((StgWord)mainThread->tso->id);
1832 if (mainThread->prev == NULL) {
1833 ASSERT(mainThread == main_threads);
1834 main_threads = mainThread->link;
1836 mainThread->prev->link = mainThread->link;
1838 if (mainThread->link != NULL) {
1839 mainThread->link->prev = mainThread->prev;
1841 releaseCapability(cap);
1842 return rtsTrue; // tells schedule() to return
1845 #ifdef RTS_SUPPORTS_THREADS
1846 ASSERT(t->main == NULL);
1848 if (t->main != NULL) {
1849 // Must be a main thread that is not the topmost one. Leave
1850 // it on the run queue until the stack has unwound to the
1851 // point where we can deal with this. Leaving it on the run
1852 // queue also ensures that the garbage collector knows about
1853 // this thread and its return value (it gets dropped from the
1854 // all_threads list so there's no other way to find it).
1855 APPEND_TO_RUN_QUEUE(t);
1861 /* -----------------------------------------------------------------------------
1862 * Perform a heap census, if PROFILING
1863 * -------------------------------------------------------------------------- */
1866 scheduleDoHeapProfile( rtsBool ready_to_gc STG_UNUSED )
1868 #if defined(PROFILING)
1869 // When we have +RTS -i0 and we're heap profiling, do a census at
1870 // every GC. This lets us get repeatable runs for debugging.
1871 if (performHeapProfile ||
1872 (RtsFlags.ProfFlags.profileInterval==0 &&
1873 RtsFlags.ProfFlags.doHeapProfile && ready_to_gc)) {
1874 GarbageCollect(GetRoots, rtsTrue);
1876 performHeapProfile = rtsFalse;
1877 return rtsTrue; // true <=> we already GC'd
1883 /* -----------------------------------------------------------------------------
1884 * Perform a garbage collection if necessary
1885 * ASSUMES: sched_mutex
1886 * -------------------------------------------------------------------------- */
1889 scheduleDoGC( rtsBool force_major )
1894 static rtsBool waiting_for_gc;
1895 int n_capabilities = RtsFlags.ParFlags.nNodes - 1;
1896 // subtract one because we're already holding one.
1897 Capability *caps[n_capabilities];
1901 // In order to GC, there must be no threads running Haskell code.
1902 // Therefore, the GC thread needs to hold *all* the capabilities,
1903 // and release them after the GC has completed.
1905 // This seems to be the simplest way: previous attempts involved
1906 // making all the threads with capabilities give up their
1907 // capabilities and sleep except for the *last* one, which
1908 // actually did the GC. But it's quite hard to arrange for all
1909 // the other tasks to sleep and stay asleep.
1911 // This does mean that there will be multiple entries in the
1912 // thread->capability hash table for the current thread, but
1913 // they will be removed as normal when the capabilities are
1917 // Someone else is already trying to GC
1918 if (waiting_for_gc) return;
1919 waiting_for_gc = rtsTrue;
1921 while (n_capabilities > 0) {
1922 IF_DEBUG(scheduler, sched_belch("ready_to_gc, grabbing all the capabilies (%d left)", n_capabilities));
1923 waitForReturnCapability(&sched_mutex, &cap);
1925 caps[n_capabilities] = cap;
1928 waiting_for_gc = rtsFalse;
1931 /* Kick any transactions which are invalid back to their
1932 * atomically frames. When next scheduled they will try to
1933 * commit, this commit will fail and they will retry.
1938 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
1939 if (t->what_next == ThreadRelocated) {
1942 next = t->global_link;
1943 if (t -> trec != NO_TREC && t -> why_blocked == NotBlocked) {
1944 if (!stmValidateNestOfTransactions (t -> trec)) {
1945 IF_DEBUG(stm, sched_belch("trec %p found wasting its time", t));
1947 // strip the stack back to the ATOMICALLY_FRAME, aborting
1948 // the (nested) transaction, and saving the stack of any
1949 // partially-evaluated thunks on the heap.
1950 raiseAsync_(t, NULL, rtsTrue);
1953 ASSERT(get_itbl((StgClosure *)t->sp)->type == ATOMICALLY_FRAME);
1961 // so this happens periodically:
1962 scheduleCheckBlackHoles();
1964 IF_DEBUG(scheduler, printAllThreads());
1966 /* everybody back, start the GC.
1967 * Could do it in this thread, or signal a condition var
1968 * to do it in another thread. Either way, we need to
1969 * broadcast on gc_pending_cond afterward.
1971 #if defined(RTS_SUPPORTS_THREADS)
1972 IF_DEBUG(scheduler,sched_belch("doing GC"));
1974 GarbageCollect(GetRoots, force_major);
1978 // release our stash of capabilities.
1980 for (i = 0; i < RtsFlags.ParFlags.nNodes-1; i++) {
1981 releaseCapability(caps[i]);
1987 /* add a ContinueThread event to continue execution of current thread */
1988 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1990 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1992 debugBelch("GRAN: eventq and runnableq after Garbage collection:\n\n");
1998 /* ---------------------------------------------------------------------------
1999 * rtsSupportsBoundThreads(): is the RTS built to support bound threads?
2000 * used by Control.Concurrent for error checking.
2001 * ------------------------------------------------------------------------- */
2004 rtsSupportsBoundThreads(void)
2006 #if defined(RTS_SUPPORTS_THREADS)
2013 /* ---------------------------------------------------------------------------
2014 * isThreadBound(tso): check whether tso is bound to an OS thread.
2015 * ------------------------------------------------------------------------- */
2018 isThreadBound(StgTSO* tso USED_IN_THREADED_RTS)
2020 #if defined(RTS_SUPPORTS_THREADS)
2021 return (tso->main != NULL);
2026 /* ---------------------------------------------------------------------------
2027 * Singleton fork(). Do not copy any running threads.
2028 * ------------------------------------------------------------------------- */
2030 #ifndef mingw32_HOST_OS
2031 #define FORKPROCESS_PRIMOP_SUPPORTED
2034 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
2036 deleteThreadImmediately(StgTSO *tso);
2039 forkProcess(HsStablePtr *entry
2040 #ifndef FORKPROCESS_PRIMOP_SUPPORTED
2045 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
2051 IF_DEBUG(scheduler,sched_belch("forking!"));
2052 rts_lock(); // This not only acquires sched_mutex, it also
2053 // makes sure that no other threads are running
2057 if (pid) { /* parent */
2059 /* just return the pid */
2063 } else { /* child */
2066 // delete all threads
2067 run_queue_hd = run_queue_tl = END_TSO_QUEUE;
2069 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
2072 // don't allow threads to catch the ThreadKilled exception
2073 deleteThreadImmediately(t);
2076 // wipe the main thread list
2077 while((m = main_threads) != NULL) {
2078 main_threads = m->link;
2079 # ifdef THREADED_RTS
2080 closeCondition(&m->bound_thread_cond);
2085 rc = rts_evalStableIO(entry, NULL); // run the action
2086 rts_checkSchedStatus("forkProcess",rc);
2090 hs_exit(); // clean up and exit
2093 #else /* !FORKPROCESS_PRIMOP_SUPPORTED */
2094 barf("forkProcess#: primop not supported, sorry!\n");
2099 /* ---------------------------------------------------------------------------
2100 * deleteAllThreads(): kill all the live threads.
2102 * This is used when we catch a user interrupt (^C), before performing
2103 * any necessary cleanups and running finalizers.
2105 * Locks: sched_mutex held.
2106 * ------------------------------------------------------------------------- */
2109 deleteAllThreads ( void )
2112 IF_DEBUG(scheduler,sched_belch("deleting all threads"));
2113 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
2114 if (t->what_next == ThreadRelocated) {
2117 next = t->global_link;
2122 // The run queue now contains a bunch of ThreadKilled threads. We
2123 // must not throw these away: the main thread(s) will be in there
2124 // somewhere, and the main scheduler loop has to deal with it.
2125 // Also, the run queue is the only thing keeping these threads from
2126 // being GC'd, and we don't want the "main thread has been GC'd" panic.
2128 ASSERT(blocked_queue_hd == END_TSO_QUEUE);
2129 ASSERT(blackhole_queue == END_TSO_QUEUE);
2130 ASSERT(sleeping_queue == END_TSO_QUEUE);
2133 /* startThread and insertThread are now in GranSim.c -- HWL */
2136 /* ---------------------------------------------------------------------------
2137 * Suspending & resuming Haskell threads.
2139 * When making a "safe" call to C (aka _ccall_GC), the task gives back
2140 * its capability before calling the C function. This allows another
2141 * task to pick up the capability and carry on running Haskell
2142 * threads. It also means that if the C call blocks, it won't lock
2145 * The Haskell thread making the C call is put to sleep for the
2146 * duration of the call, on the susepended_ccalling_threads queue. We
2147 * give out a token to the task, which it can use to resume the thread
2148 * on return from the C function.
2149 * ------------------------------------------------------------------------- */
2152 suspendThread( StgRegTable *reg )
2156 int saved_errno = errno;
2158 /* assume that *reg is a pointer to the StgRegTable part
2161 cap = (Capability *)((void *)((unsigned char*)reg - sizeof(StgFunTable)));
2163 ACQUIRE_LOCK(&sched_mutex);
2166 sched_belch("thread %d did a _ccall_gc", cap->r.rCurrentTSO->id));
2168 // XXX this might not be necessary --SDM
2169 cap->r.rCurrentTSO->what_next = ThreadRunGHC;
2171 threadPaused(cap->r.rCurrentTSO);
2172 cap->r.rCurrentTSO->link = suspended_ccalling_threads;
2173 suspended_ccalling_threads = cap->r.rCurrentTSO;
2175 if(cap->r.rCurrentTSO->blocked_exceptions == NULL) {
2176 cap->r.rCurrentTSO->why_blocked = BlockedOnCCall;
2177 cap->r.rCurrentTSO->blocked_exceptions = END_TSO_QUEUE;
2179 cap->r.rCurrentTSO->why_blocked = BlockedOnCCall_NoUnblockExc;
2182 /* Use the thread ID as the token; it should be unique */
2183 tok = cap->r.rCurrentTSO->id;
2185 /* Hand back capability */
2186 cap->r.rInHaskell = rtsFalse;
2187 releaseCapability(cap);
2189 #if defined(RTS_SUPPORTS_THREADS)
2190 /* Preparing to leave the RTS, so ensure there's a native thread/task
2191 waiting to take over.
2193 IF_DEBUG(scheduler, sched_belch("worker (token %d): leaving RTS", tok));
2196 RELEASE_LOCK(&sched_mutex);
2198 errno = saved_errno;
2203 resumeThread( StgInt tok )
2205 StgTSO *tso, **prev;
2207 int saved_errno = errno;
2209 #if defined(RTS_SUPPORTS_THREADS)
2210 /* Wait for permission to re-enter the RTS with the result. */
2211 ACQUIRE_LOCK(&sched_mutex);
2212 waitForReturnCapability(&sched_mutex, &cap);
2214 IF_DEBUG(scheduler, sched_belch("worker (token %d): re-entering RTS", tok));
2216 grabCapability(&cap);
2219 /* Remove the thread off of the suspended list */
2220 prev = &suspended_ccalling_threads;
2221 for (tso = suspended_ccalling_threads;
2222 tso != END_TSO_QUEUE;
2223 prev = &tso->link, tso = tso->link) {
2224 if (tso->id == (StgThreadID)tok) {
2229 if (tso == END_TSO_QUEUE) {
2230 barf("resumeThread: thread not found");
2232 tso->link = END_TSO_QUEUE;
2234 if(tso->why_blocked == BlockedOnCCall) {
2235 awakenBlockedQueueNoLock(tso->blocked_exceptions);
2236 tso->blocked_exceptions = NULL;
2239 /* Reset blocking status */
2240 tso->why_blocked = NotBlocked;
2242 cap->r.rCurrentTSO = tso;
2243 cap->r.rInHaskell = rtsTrue;
2244 RELEASE_LOCK(&sched_mutex);
2245 errno = saved_errno;
2249 /* ---------------------------------------------------------------------------
2250 * Comparing Thread ids.
2252 * This is used from STG land in the implementation of the
2253 * instances of Eq/Ord for ThreadIds.
2254 * ------------------------------------------------------------------------ */
2257 cmp_thread(StgPtr tso1, StgPtr tso2)
2259 StgThreadID id1 = ((StgTSO *)tso1)->id;
2260 StgThreadID id2 = ((StgTSO *)tso2)->id;
2262 if (id1 < id2) return (-1);
2263 if (id1 > id2) return 1;
2267 /* ---------------------------------------------------------------------------
2268 * Fetching the ThreadID from an StgTSO.
2270 * This is used in the implementation of Show for ThreadIds.
2271 * ------------------------------------------------------------------------ */
2273 rts_getThreadId(StgPtr tso)
2275 return ((StgTSO *)tso)->id;
2280 labelThread(StgPtr tso, char *label)
2285 /* Caveat: Once set, you can only set the thread name to "" */
2286 len = strlen(label)+1;
2287 buf = stgMallocBytes(len * sizeof(char), "Schedule.c:labelThread()");
2288 strncpy(buf,label,len);
2289 /* Update will free the old memory for us */
2290 updateThreadLabel(((StgTSO *)tso)->id,buf);
2294 /* ---------------------------------------------------------------------------
2295 Create a new thread.
2297 The new thread starts with the given stack size. Before the
2298 scheduler can run, however, this thread needs to have a closure
2299 (and possibly some arguments) pushed on its stack. See
2300 pushClosure() in Schedule.h.
2302 createGenThread() and createIOThread() (in SchedAPI.h) are
2303 convenient packaged versions of this function.
2305 currently pri (priority) is only used in a GRAN setup -- HWL
2306 ------------------------------------------------------------------------ */
2308 /* currently pri (priority) is only used in a GRAN setup -- HWL */
2310 createThread(nat size, StgInt pri)
2313 createThread(nat size)
2319 /* First check whether we should create a thread at all */
2320 #if defined(PARALLEL_HASKELL)
2321 /* check that no more than RtsFlags.ParFlags.maxThreads threads are created */
2322 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads) {
2324 debugBelch("{createThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)\n",
2325 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
2326 return END_TSO_QUEUE;
2332 ASSERT(!RtsFlags.GranFlags.Light || CurrentProc==0);
2335 // ToDo: check whether size = stack_size - TSO_STRUCT_SIZEW
2337 /* catch ridiculously small stack sizes */
2338 if (size < MIN_STACK_WORDS + TSO_STRUCT_SIZEW) {
2339 size = MIN_STACK_WORDS + TSO_STRUCT_SIZEW;
2342 stack_size = size - TSO_STRUCT_SIZEW;
2344 tso = (StgTSO *)allocate(size);
2345 TICK_ALLOC_TSO(stack_size, 0);
2347 SET_HDR(tso, &stg_TSO_info, CCS_SYSTEM);
2349 SET_GRAN_HDR(tso, ThisPE);
2352 // Always start with the compiled code evaluator
2353 tso->what_next = ThreadRunGHC;
2355 tso->id = next_thread_id++;
2356 tso->why_blocked = NotBlocked;
2357 tso->blocked_exceptions = NULL;
2359 tso->saved_errno = 0;
2362 tso->stack_size = stack_size;
2363 tso->max_stack_size = round_to_mblocks(RtsFlags.GcFlags.maxStkSize)
2365 tso->sp = (P_)&(tso->stack) + stack_size;
2367 tso->trec = NO_TREC;
2370 tso->prof.CCCS = CCS_MAIN;
2373 /* put a stop frame on the stack */
2374 tso->sp -= sizeofW(StgStopFrame);
2375 SET_HDR((StgClosure*)tso->sp,(StgInfoTable *)&stg_stop_thread_info,CCS_SYSTEM);
2376 tso->link = END_TSO_QUEUE;
2380 /* uses more flexible routine in GranSim */
2381 insertThread(tso, CurrentProc);
2383 /* In a non-GranSim setup the pushing of a TSO onto the runq is separated
2389 if (RtsFlags.GranFlags.GranSimStats.Full)
2390 DumpGranEvent(GR_START,tso);
2391 #elif defined(PARALLEL_HASKELL)
2392 if (RtsFlags.ParFlags.ParStats.Full)
2393 DumpGranEvent(GR_STARTQ,tso);
2394 /* HACk to avoid SCHEDULE
2398 /* Link the new thread on the global thread list.
2400 tso->global_link = all_threads;
2404 tso->dist.priority = MandatoryPriority; //by default that is...
2408 tso->gran.pri = pri;
2410 tso->gran.magic = TSO_MAGIC; // debugging only
2412 tso->gran.sparkname = 0;
2413 tso->gran.startedat = CURRENT_TIME;
2414 tso->gran.exported = 0;
2415 tso->gran.basicblocks = 0;
2416 tso->gran.allocs = 0;
2417 tso->gran.exectime = 0;
2418 tso->gran.fetchtime = 0;
2419 tso->gran.fetchcount = 0;
2420 tso->gran.blocktime = 0;
2421 tso->gran.blockcount = 0;
2422 tso->gran.blockedat = 0;
2423 tso->gran.globalsparks = 0;
2424 tso->gran.localsparks = 0;
2425 if (RtsFlags.GranFlags.Light)
2426 tso->gran.clock = Now; /* local clock */
2428 tso->gran.clock = 0;
2430 IF_DEBUG(gran,printTSO(tso));
2431 #elif defined(PARALLEL_HASKELL)
2433 tso->par.magic = TSO_MAGIC; // debugging only
2435 tso->par.sparkname = 0;
2436 tso->par.startedat = CURRENT_TIME;
2437 tso->par.exported = 0;
2438 tso->par.basicblocks = 0;
2439 tso->par.allocs = 0;
2440 tso->par.exectime = 0;
2441 tso->par.fetchtime = 0;
2442 tso->par.fetchcount = 0;
2443 tso->par.blocktime = 0;
2444 tso->par.blockcount = 0;
2445 tso->par.blockedat = 0;
2446 tso->par.globalsparks = 0;
2447 tso->par.localsparks = 0;
2451 globalGranStats.tot_threads_created++;
2452 globalGranStats.threads_created_on_PE[CurrentProc]++;
2453 globalGranStats.tot_sq_len += spark_queue_len(CurrentProc);
2454 globalGranStats.tot_sq_probes++;
2455 #elif defined(PARALLEL_HASKELL)
2456 // collect parallel global statistics (currently done together with GC stats)
2457 if (RtsFlags.ParFlags.ParStats.Global &&
2458 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
2459 //debugBelch("Creating thread %d @ %11.2f\n", tso->id, usertime());
2460 globalParStats.tot_threads_created++;
2466 sched_belch("==__ schedule: Created TSO %d (%p);",
2467 CurrentProc, tso, tso->id));
2468 #elif defined(PARALLEL_HASKELL)
2469 IF_PAR_DEBUG(verbose,
2470 sched_belch("==__ schedule: Created TSO %d (%p); %d threads active",
2471 (long)tso->id, tso, advisory_thread_count));
2473 IF_DEBUG(scheduler,sched_belch("created thread %ld, stack size = %lx words",
2474 (long)tso->id, (long)tso->stack_size));
2481 all parallel thread creation calls should fall through the following routine.
2484 createThreadFromSpark(rtsSpark spark)
2486 ASSERT(spark != (rtsSpark)NULL);
2487 // JB: TAKE CARE OF THIS COUNTER! BUGGY
2488 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads)
2490 barf("{createSparkThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
2491 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
2492 return END_TSO_QUEUE;
2496 tso = createThread(RtsFlags.GcFlags.initialStkSize);
2497 if (tso==END_TSO_QUEUE)
2498 barf("createSparkThread: Cannot create TSO");
2500 tso->priority = AdvisoryPriority;
2502 pushClosure(tso,spark);
2504 advisory_thread_count++; // JB: TAKE CARE OF THIS COUNTER! BUGGY
2511 Turn a spark into a thread.
2512 ToDo: fix for SMP (needs to acquire SCHED_MUTEX!)
2516 activateSpark (rtsSpark spark)
2520 tso = createSparkThread(spark);
2521 if (RtsFlags.ParFlags.ParStats.Full) {
2522 //ASSERT(run_queue_hd == END_TSO_QUEUE); // I think ...
2523 IF_PAR_DEBUG(verbose,
2524 debugBelch("==^^ activateSpark: turning spark of closure %p (%s) into a thread\n",
2525 (StgClosure *)spark, info_type((StgClosure *)spark)));
2527 // ToDo: fwd info on local/global spark to thread -- HWL
2528 // tso->gran.exported = spark->exported;
2529 // tso->gran.locked = !spark->global;
2530 // tso->gran.sparkname = spark->name;
2536 /* ---------------------------------------------------------------------------
2539 * scheduleThread puts a thread on the head of the runnable queue.
2540 * This will usually be done immediately after a thread is created.
2541 * The caller of scheduleThread must create the thread using e.g.
2542 * createThread and push an appropriate closure
2543 * on this thread's stack before the scheduler is invoked.
2544 * ------------------------------------------------------------------------ */
2547 scheduleThreadLocked(StgTSO *tso)
2549 // The thread goes at the *end* of the run-queue, to avoid possible
2550 // starvation of any threads already on the queue.
2551 APPEND_TO_RUN_QUEUE(tso);
2556 scheduleThread(StgTSO* tso)
2558 ACQUIRE_LOCK(&sched_mutex);
2559 scheduleThreadLocked(tso);
2560 RELEASE_LOCK(&sched_mutex);
2563 #if defined(RTS_SUPPORTS_THREADS)
2564 static Condition bound_cond_cache;
2565 static int bound_cond_cache_full = 0;
2570 scheduleWaitThread(StgTSO* tso, /*[out]*/HaskellObj* ret,
2571 Capability *initialCapability)
2573 // Precondition: sched_mutex must be held
2576 m = stgMallocBytes(sizeof(StgMainThread), "waitThread");
2581 m->link = main_threads;
2583 if (main_threads != NULL) {
2584 main_threads->prev = m;
2588 #if defined(RTS_SUPPORTS_THREADS)
2589 // Allocating a new condition for each thread is expensive, so we
2590 // cache one. This is a pretty feeble hack, but it helps speed up
2591 // consecutive call-ins quite a bit.
2592 if (bound_cond_cache_full) {
2593 m->bound_thread_cond = bound_cond_cache;
2594 bound_cond_cache_full = 0;
2596 initCondition(&m->bound_thread_cond);
2600 /* Put the thread on the main-threads list prior to scheduling the TSO.
2601 Failure to do so introduces a race condition in the MT case (as
2602 identified by Wolfgang Thaller), whereby the new task/OS thread
2603 created by scheduleThread_() would complete prior to the thread
2604 that spawned it managed to put 'itself' on the main-threads list.
2605 The upshot of it all being that the worker thread wouldn't get to
2606 signal the completion of the its work item for the main thread to
2607 see (==> it got stuck waiting.) -- sof 6/02.
2609 IF_DEBUG(scheduler, sched_belch("waiting for thread (%d)", tso->id));
2611 APPEND_TO_RUN_QUEUE(tso);
2612 // NB. Don't call threadRunnable() here, because the thread is
2613 // bound and only runnable by *this* OS thread, so waking up other
2614 // workers will just slow things down.
2616 return waitThread_(m, initialCapability);
2619 /* ---------------------------------------------------------------------------
2622 * Initialise the scheduler. This resets all the queues - if the
2623 * queues contained any threads, they'll be garbage collected at the
2626 * ------------------------------------------------------------------------ */
2634 for (i=0; i<=MAX_PROC; i++) {
2635 run_queue_hds[i] = END_TSO_QUEUE;
2636 run_queue_tls[i] = END_TSO_QUEUE;
2637 blocked_queue_hds[i] = END_TSO_QUEUE;
2638 blocked_queue_tls[i] = END_TSO_QUEUE;
2639 ccalling_threadss[i] = END_TSO_QUEUE;
2640 blackhole_queue[i] = END_TSO_QUEUE;
2641 sleeping_queue = END_TSO_QUEUE;
2644 run_queue_hd = END_TSO_QUEUE;
2645 run_queue_tl = END_TSO_QUEUE;
2646 blocked_queue_hd = END_TSO_QUEUE;
2647 blocked_queue_tl = END_TSO_QUEUE;
2648 blackhole_queue = END_TSO_QUEUE;
2649 sleeping_queue = END_TSO_QUEUE;
2652 suspended_ccalling_threads = END_TSO_QUEUE;
2654 main_threads = NULL;
2655 all_threads = END_TSO_QUEUE;
2660 RtsFlags.ConcFlags.ctxtSwitchTicks =
2661 RtsFlags.ConcFlags.ctxtSwitchTime / TICK_MILLISECS;
2663 #if defined(RTS_SUPPORTS_THREADS)
2664 /* Initialise the mutex and condition variables used by
2666 initMutex(&sched_mutex);
2667 initMutex(&term_mutex);
2670 ACQUIRE_LOCK(&sched_mutex);
2672 /* A capability holds the state a native thread needs in
2673 * order to execute STG code. At least one capability is
2674 * floating around (only SMP builds have more than one).
2678 #if defined(RTS_SUPPORTS_THREADS)
2683 /* eagerly start some extra workers */
2684 startingWorkerThread = RtsFlags.ParFlags.nNodes;
2685 startTasks(RtsFlags.ParFlags.nNodes, taskStart);
2688 #if /* defined(SMP) ||*/ defined(PARALLEL_HASKELL)
2692 RELEASE_LOCK(&sched_mutex);
2696 exitScheduler( void )
2698 interrupted = rtsTrue;
2699 shutting_down_scheduler = rtsTrue;
2701 #if defined(RTS_SUPPORTS_THREADS)
2702 if (threadIsTask(osThreadId())) { taskStop(); }
2705 // What can we do here? There are a bunch of worker threads, it
2706 // might be nice to let them exit cleanly. There may be some main
2707 // threads in the run queue; we should let them return to their
2708 // callers with an Interrupted state. We can't in general wait
2709 // for all the running Tasks to stop, because some might be off in
2710 // a C call that is blocked.
2712 // Letting the run queue drain is the safest thing. That lets any
2713 // main threads return that can return, and cleans up all the
2714 // runnable threads. Then we grab all the Capabilities to stop
2715 // anything unexpected happening while we shut down.
2717 // ToDo: this doesn't let us get the time stats from the worker
2718 // tasks, because they haven't called taskStop().
2720 ACQUIRE_LOCK(&sched_mutex);
2723 for (i = 1000; i > 0; i--) {
2724 if (EMPTY_RUN_QUEUE()) {
2725 IF_DEBUG(scheduler, sched_belch("run queue is empty"));
2728 IF_DEBUG(scheduler, sched_belch("yielding"));
2729 RELEASE_LOCK(&sched_mutex);
2732 ACQUIRE_LOCK(&sched_mutex);
2739 int n_capabilities = RtsFlags.ParFlags.nNodes;
2740 Capability *caps[n_capabilities];
2743 while (n_capabilities > 0) {
2744 IF_DEBUG(scheduler, sched_belch("exitScheduler: grabbing all the capabilies (%d left)", n_capabilities));
2745 waitForReturnCapability(&sched_mutex, &cap);
2747 caps[n_capabilities] = cap;
2753 waitForReturnCapability(&sched_mutex, &cap);
2759 /* ----------------------------------------------------------------------------
2760 Managing the per-task allocation areas.
2762 Each capability comes with an allocation area. These are
2763 fixed-length block lists into which allocation can be done.
2765 ToDo: no support for two-space collection at the moment???
2766 ------------------------------------------------------------------------- */
2768 static SchedulerStatus
2769 waitThread_(StgMainThread* m, Capability *initialCapability)
2771 SchedulerStatus stat;
2773 // Precondition: sched_mutex must be held.
2774 IF_DEBUG(scheduler, sched_belch("new main thread (%d)", m->tso->id));
2777 /* GranSim specific init */
2778 CurrentTSO = m->tso; // the TSO to run
2779 procStatus[MainProc] = Busy; // status of main PE
2780 CurrentProc = MainProc; // PE to run it on
2781 schedule(m,initialCapability);
2783 schedule(m,initialCapability);
2784 ASSERT(m->stat != NoStatus);
2789 #if defined(RTS_SUPPORTS_THREADS)
2790 // Free the condition variable, returning it to the cache if possible.
2791 if (!bound_cond_cache_full) {
2792 bound_cond_cache = m->bound_thread_cond;
2793 bound_cond_cache_full = 1;
2795 closeCondition(&m->bound_thread_cond);
2799 IF_DEBUG(scheduler, sched_belch("main thread (%d) finished", m->tso->id));
2802 // Postcondition: sched_mutex still held
2806 /* ---------------------------------------------------------------------------
2807 Where are the roots that we know about?
2809 - all the threads on the runnable queue
2810 - all the threads on the blocked queue
2811 - all the threads on the sleeping queue
2812 - all the thread currently executing a _ccall_GC
2813 - all the "main threads"
2815 ------------------------------------------------------------------------ */
2817 /* This has to be protected either by the scheduler monitor, or by the
2818 garbage collection monitor (probably the latter).
2823 GetRoots( evac_fn evac )
2828 for (i=0; i<=RtsFlags.GranFlags.proc; i++) {
2829 if ((run_queue_hds[i] != END_TSO_QUEUE) && ((run_queue_hds[i] != NULL)))
2830 evac((StgClosure **)&run_queue_hds[i]);
2831 if ((run_queue_tls[i] != END_TSO_QUEUE) && ((run_queue_tls[i] != NULL)))
2832 evac((StgClosure **)&run_queue_tls[i]);
2834 if ((blocked_queue_hds[i] != END_TSO_QUEUE) && ((blocked_queue_hds[i] != NULL)))
2835 evac((StgClosure **)&blocked_queue_hds[i]);
2836 if ((blocked_queue_tls[i] != END_TSO_QUEUE) && ((blocked_queue_tls[i] != NULL)))
2837 evac((StgClosure **)&blocked_queue_tls[i]);
2838 if ((ccalling_threadss[i] != END_TSO_QUEUE) && ((ccalling_threadss[i] != NULL)))
2839 evac((StgClosure **)&ccalling_threads[i]);
2846 if (run_queue_hd != END_TSO_QUEUE) {
2847 ASSERT(run_queue_tl != END_TSO_QUEUE);
2848 evac((StgClosure **)&run_queue_hd);
2849 evac((StgClosure **)&run_queue_tl);
2852 if (blocked_queue_hd != END_TSO_QUEUE) {
2853 ASSERT(blocked_queue_tl != END_TSO_QUEUE);
2854 evac((StgClosure **)&blocked_queue_hd);
2855 evac((StgClosure **)&blocked_queue_tl);
2858 if (sleeping_queue != END_TSO_QUEUE) {
2859 evac((StgClosure **)&sleeping_queue);
2863 if (blackhole_queue != END_TSO_QUEUE) {
2864 evac((StgClosure **)&blackhole_queue);
2867 if (suspended_ccalling_threads != END_TSO_QUEUE) {
2868 evac((StgClosure **)&suspended_ccalling_threads);
2871 #if defined(PARALLEL_HASKELL) || defined(GRAN)
2872 markSparkQueue(evac);
2875 #if defined(RTS_USER_SIGNALS)
2876 // mark the signal handlers (signals should be already blocked)
2877 markSignalHandlers(evac);
2881 /* -----------------------------------------------------------------------------
2884 This is the interface to the garbage collector from Haskell land.
2885 We provide this so that external C code can allocate and garbage
2886 collect when called from Haskell via _ccall_GC.
2888 It might be useful to provide an interface whereby the programmer
2889 can specify more roots (ToDo).
2891 This needs to be protected by the GC condition variable above. KH.
2892 -------------------------------------------------------------------------- */
2894 static void (*extra_roots)(evac_fn);
2899 /* Obligated to hold this lock upon entry */
2900 ACQUIRE_LOCK(&sched_mutex);
2901 GarbageCollect(GetRoots,rtsFalse);
2902 RELEASE_LOCK(&sched_mutex);
2906 performMajorGC(void)
2908 ACQUIRE_LOCK(&sched_mutex);
2909 GarbageCollect(GetRoots,rtsTrue);
2910 RELEASE_LOCK(&sched_mutex);
2914 AllRoots(evac_fn evac)
2916 GetRoots(evac); // the scheduler's roots
2917 extra_roots(evac); // the user's roots
2921 performGCWithRoots(void (*get_roots)(evac_fn))
2923 ACQUIRE_LOCK(&sched_mutex);
2924 extra_roots = get_roots;
2925 GarbageCollect(AllRoots,rtsFalse);
2926 RELEASE_LOCK(&sched_mutex);
2929 /* -----------------------------------------------------------------------------
2932 If the thread has reached its maximum stack size, then raise the
2933 StackOverflow exception in the offending thread. Otherwise
2934 relocate the TSO into a larger chunk of memory and adjust its stack
2936 -------------------------------------------------------------------------- */
2939 threadStackOverflow(StgTSO *tso)
2941 nat new_stack_size, stack_words;
2946 IF_DEBUG(sanity,checkTSO(tso));
2947 if (tso->stack_size >= tso->max_stack_size) {
2950 debugBelch("@@ threadStackOverflow of TSO %ld (%p): stack too large (now %ld; max is %ld)\n",
2951 (long)tso->id, tso, (long)tso->stack_size, (long)tso->max_stack_size);
2952 /* If we're debugging, just print out the top of the stack */
2953 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2956 /* Send this thread the StackOverflow exception */
2957 raiseAsync(tso, (StgClosure *)stackOverflow_closure);
2961 /* Try to double the current stack size. If that takes us over the
2962 * maximum stack size for this thread, then use the maximum instead.
2963 * Finally round up so the TSO ends up as a whole number of blocks.
2965 new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
2966 new_tso_size = (lnat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) +
2967 TSO_STRUCT_SIZE)/sizeof(W_);
2968 new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */
2969 new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
2971 IF_DEBUG(scheduler, debugBelch("== sched: increasing stack size from %d words to %d.\n", tso->stack_size, new_stack_size));
2973 dest = (StgTSO *)allocate(new_tso_size);
2974 TICK_ALLOC_TSO(new_stack_size,0);
2976 /* copy the TSO block and the old stack into the new area */
2977 memcpy(dest,tso,TSO_STRUCT_SIZE);
2978 stack_words = tso->stack + tso->stack_size - tso->sp;
2979 new_sp = (P_)dest + new_tso_size - stack_words;
2980 memcpy(new_sp, tso->sp, stack_words * sizeof(W_));
2982 /* relocate the stack pointers... */
2984 dest->stack_size = new_stack_size;
2986 /* Mark the old TSO as relocated. We have to check for relocated
2987 * TSOs in the garbage collector and any primops that deal with TSOs.
2989 * It's important to set the sp value to just beyond the end
2990 * of the stack, so we don't attempt to scavenge any part of the
2993 tso->what_next = ThreadRelocated;
2995 tso->sp = (P_)&(tso->stack[tso->stack_size]);
2996 tso->why_blocked = NotBlocked;
2998 IF_PAR_DEBUG(verbose,
2999 debugBelch("@@ threadStackOverflow of TSO %d (now at %p): stack size increased to %ld\n",
3000 tso->id, tso, tso->stack_size);
3001 /* If we're debugging, just print out the top of the stack */
3002 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
3005 IF_DEBUG(sanity,checkTSO(tso));
3007 IF_DEBUG(scheduler,printTSO(dest));
3013 /* ---------------------------------------------------------------------------
3014 Wake up a queue that was blocked on some resource.
3015 ------------------------------------------------------------------------ */
3019 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
3022 #elif defined(PARALLEL_HASKELL)
3024 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
3026 /* write RESUME events to log file and
3027 update blocked and fetch time (depending on type of the orig closure) */
3028 if (RtsFlags.ParFlags.ParStats.Full) {
3029 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
3030 GR_RESUMEQ, ((StgTSO *)bqe), ((StgTSO *)bqe)->block_info.closure,
3031 0, 0 /* spark_queue_len(ADVISORY_POOL) */);
3032 if (EMPTY_RUN_QUEUE())
3033 emitSchedule = rtsTrue;
3035 switch (get_itbl(node)->type) {
3037 ((StgTSO *)bqe)->par.fetchtime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
3042 ((StgTSO *)bqe)->par.blocktime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
3049 barf("{unblockOneLocked}Daq Qagh: unexpected closure in blocking queue");
3056 StgBlockingQueueElement *
3057 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
3060 PEs node_loc, tso_loc;
3062 node_loc = where_is(node); // should be lifted out of loop
3063 tso = (StgTSO *)bqe; // wastes an assignment to get the type right
3064 tso_loc = where_is((StgClosure *)tso);
3065 if (IS_LOCAL_TO(PROCS(node),tso_loc)) { // TSO is local
3066 /* !fake_fetch => TSO is on CurrentProc is same as IS_LOCAL_TO */
3067 ASSERT(CurrentProc!=node_loc || tso_loc==CurrentProc);
3068 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.lunblocktime;
3069 // insertThread(tso, node_loc);
3070 new_event(tso_loc, tso_loc, CurrentTime[CurrentProc],
3072 tso, node, (rtsSpark*)NULL);
3073 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
3076 } else { // TSO is remote (actually should be FMBQ)
3077 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.mpacktime +
3078 RtsFlags.GranFlags.Costs.gunblocktime +
3079 RtsFlags.GranFlags.Costs.latency;
3080 new_event(tso_loc, CurrentProc, CurrentTime[CurrentProc],
3082 tso, node, (rtsSpark*)NULL);
3083 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
3086 /* the thread-queue-overhead is accounted for in either Resume or UnblockThread */
3088 debugBelch(" %s TSO %d (%p) [PE %d] (block_info.closure=%p) (next=%p) ,",
3089 (node_loc==tso_loc ? "Local" : "Global"),
3090 tso->id, tso, CurrentProc, tso->block_info.closure, tso->link));
3091 tso->block_info.closure = NULL;
3092 IF_DEBUG(scheduler,debugBelch("-- Waking up thread %ld (%p)\n",
3095 #elif defined(PARALLEL_HASKELL)
3096 StgBlockingQueueElement *
3097 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
3099 StgBlockingQueueElement *next;
3101 switch (get_itbl(bqe)->type) {
3103 ASSERT(((StgTSO *)bqe)->why_blocked != NotBlocked);
3104 /* if it's a TSO just push it onto the run_queue */
3106 ((StgTSO *)bqe)->link = END_TSO_QUEUE; // debugging?
3107 APPEND_TO_RUN_QUEUE((StgTSO *)bqe);
3109 unblockCount(bqe, node);
3110 /* reset blocking status after dumping event */
3111 ((StgTSO *)bqe)->why_blocked = NotBlocked;
3115 /* if it's a BLOCKED_FETCH put it on the PendingFetches list */
3117 bqe->link = (StgBlockingQueueElement *)PendingFetches;
3118 PendingFetches = (StgBlockedFetch *)bqe;
3122 /* can ignore this case in a non-debugging setup;
3123 see comments on RBHSave closures above */
3125 /* check that the closure is an RBHSave closure */
3126 ASSERT(get_itbl((StgClosure *)bqe) == &stg_RBH_Save_0_info ||
3127 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_1_info ||
3128 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_2_info);
3132 barf("{unblockOneLocked}Daq Qagh: Unexpected IP (%#lx; %s) in blocking queue at %#lx\n",
3133 get_itbl((StgClosure *)bqe), info_type((StgClosure *)bqe),
3137 IF_PAR_DEBUG(bq, debugBelch(", %p (%s)\n", bqe, info_type((StgClosure*)bqe)));
3141 #else /* !GRAN && !PARALLEL_HASKELL */
3143 unblockOneLocked(StgTSO *tso)
3147 ASSERT(get_itbl(tso)->type == TSO);
3148 ASSERT(tso->why_blocked != NotBlocked);
3149 tso->why_blocked = NotBlocked;
3151 tso->link = END_TSO_QUEUE;
3152 APPEND_TO_RUN_QUEUE(tso);
3154 IF_DEBUG(scheduler,sched_belch("waking up thread %ld", (long)tso->id));
3159 #if defined(GRAN) || defined(PARALLEL_HASKELL)
3160 INLINE_ME StgBlockingQueueElement *
3161 unblockOne(StgBlockingQueueElement *bqe, StgClosure *node)
3163 ACQUIRE_LOCK(&sched_mutex);
3164 bqe = unblockOneLocked(bqe, node);
3165 RELEASE_LOCK(&sched_mutex);
3170 unblockOne(StgTSO *tso)
3172 ACQUIRE_LOCK(&sched_mutex);
3173 tso = unblockOneLocked(tso);
3174 RELEASE_LOCK(&sched_mutex);
3181 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
3183 StgBlockingQueueElement *bqe;
3188 debugBelch("##-_ AwBQ for node %p on PE %d @ %ld by TSO %d (%p): \n", \
3189 node, CurrentProc, CurrentTime[CurrentProc],
3190 CurrentTSO->id, CurrentTSO));
3192 node_loc = where_is(node);
3194 ASSERT(q == END_BQ_QUEUE ||
3195 get_itbl(q)->type == TSO || // q is either a TSO or an RBHSave
3196 get_itbl(q)->type == CONSTR); // closure (type constructor)
3197 ASSERT(is_unique(node));
3199 /* FAKE FETCH: magically copy the node to the tso's proc;
3200 no Fetch necessary because in reality the node should not have been
3201 moved to the other PE in the first place
3203 if (CurrentProc!=node_loc) {
3205 debugBelch("## node %p is on PE %d but CurrentProc is %d (TSO %d); assuming fake fetch and adjusting bitmask (old: %#x)\n",
3206 node, node_loc, CurrentProc, CurrentTSO->id,
3207 // CurrentTSO, where_is(CurrentTSO),
3208 node->header.gran.procs));
3209 node->header.gran.procs = (node->header.gran.procs) | PE_NUMBER(CurrentProc);
3211 debugBelch("## new bitmask of node %p is %#x\n",
3212 node, node->header.gran.procs));
3213 if (RtsFlags.GranFlags.GranSimStats.Global) {
3214 globalGranStats.tot_fake_fetches++;
3219 // ToDo: check: ASSERT(CurrentProc==node_loc);
3220 while (get_itbl(bqe)->type==TSO) { // q != END_TSO_QUEUE) {
3223 bqe points to the current element in the queue
3224 next points to the next element in the queue
3226 //tso = (StgTSO *)bqe; // wastes an assignment to get the type right
3227 //tso_loc = where_is(tso);
3229 bqe = unblockOneLocked(bqe, node);
3232 /* if this is the BQ of an RBH, we have to put back the info ripped out of
3233 the closure to make room for the anchor of the BQ */
3234 if (bqe!=END_BQ_QUEUE) {
3235 ASSERT(get_itbl(node)->type == RBH && get_itbl(bqe)->type == CONSTR);
3237 ASSERT((info_ptr==&RBH_Save_0_info) ||
3238 (info_ptr==&RBH_Save_1_info) ||
3239 (info_ptr==&RBH_Save_2_info));
3241 /* cf. convertToRBH in RBH.c for writing the RBHSave closure */
3242 ((StgRBH *)node)->blocking_queue = (StgBlockingQueueElement *)((StgRBHSave *)bqe)->payload[0];
3243 ((StgRBH *)node)->mut_link = (StgMutClosure *)((StgRBHSave *)bqe)->payload[1];
3246 debugBelch("## Filled in RBH_Save for %p (%s) at end of AwBQ\n",
3247 node, info_type(node)));
3250 /* statistics gathering */
3251 if (RtsFlags.GranFlags.GranSimStats.Global) {
3252 // globalGranStats.tot_bq_processing_time += bq_processing_time;
3253 globalGranStats.tot_bq_len += len; // total length of all bqs awakened
3254 // globalGranStats.tot_bq_len_local += len_local; // same for local TSOs only
3255 globalGranStats.tot_awbq++; // total no. of bqs awakened
3258 debugBelch("## BQ Stats of %p: [%d entries] %s\n",
3259 node, len, (bqe!=END_BQ_QUEUE) ? "RBH" : ""));
3261 #elif defined(PARALLEL_HASKELL)
3263 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
3265 StgBlockingQueueElement *bqe;
3267 ACQUIRE_LOCK(&sched_mutex);
3269 IF_PAR_DEBUG(verbose,
3270 debugBelch("##-_ AwBQ for node %p on [%x]: \n",
3274 if(get_itbl(q)->type == CONSTR || q==END_BQ_QUEUE) {
3275 IF_PAR_DEBUG(verbose, debugBelch("## ... nothing to unblock so lets just return. RFP (BUG?)\n"));
3280 ASSERT(q == END_BQ_QUEUE ||
3281 get_itbl(q)->type == TSO ||
3282 get_itbl(q)->type == BLOCKED_FETCH ||
3283 get_itbl(q)->type == CONSTR);
3286 while (get_itbl(bqe)->type==TSO ||
3287 get_itbl(bqe)->type==BLOCKED_FETCH) {
3288 bqe = unblockOneLocked(bqe, node);
3290 RELEASE_LOCK(&sched_mutex);
3293 #else /* !GRAN && !PARALLEL_HASKELL */
3296 awakenBlockedQueueNoLock(StgTSO *tso)
3298 if (tso == NULL) return; // hack; see bug #1235728, and comments in
3300 while (tso != END_TSO_QUEUE) {
3301 tso = unblockOneLocked(tso);
3306 awakenBlockedQueue(StgTSO *tso)
3308 if (tso == NULL) return; // hack; see bug #1235728, and comments in
3310 ACQUIRE_LOCK(&sched_mutex);
3311 while (tso != END_TSO_QUEUE) {
3312 tso = unblockOneLocked(tso);
3314 RELEASE_LOCK(&sched_mutex);
3318 /* ---------------------------------------------------------------------------
3320 - usually called inside a signal handler so it mustn't do anything fancy.
3321 ------------------------------------------------------------------------ */
3324 interruptStgRts(void)
3329 /* ToDo: if invoked from a signal handler, this threadRunnable
3330 * only works if there's another thread (not this one) waiting to
3335 /* -----------------------------------------------------------------------------
3338 This is for use when we raise an exception in another thread, which
3340 This has nothing to do with the UnblockThread event in GranSim. -- HWL
3341 -------------------------------------------------------------------------- */
3343 #if defined(GRAN) || defined(PARALLEL_HASKELL)
3345 NB: only the type of the blocking queue is different in GranSim and GUM
3346 the operations on the queue-elements are the same
3347 long live polymorphism!
3349 Locks: sched_mutex is held upon entry and exit.
3353 unblockThread(StgTSO *tso)
3355 StgBlockingQueueElement *t, **last;
3357 switch (tso->why_blocked) {
3360 return; /* not blocked */
3363 // Be careful: nothing to do here! We tell the scheduler that the thread
3364 // is runnable and we leave it to the stack-walking code to abort the
3365 // transaction while unwinding the stack. We should perhaps have a debugging
3366 // test to make sure that this really happens and that the 'zombie' transaction
3367 // does not get committed.
3371 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
3373 StgBlockingQueueElement *last_tso = END_BQ_QUEUE;
3374 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
3376 last = (StgBlockingQueueElement **)&mvar->head;
3377 for (t = (StgBlockingQueueElement *)mvar->head;
3379 last = &t->link, last_tso = t, t = t->link) {
3380 if (t == (StgBlockingQueueElement *)tso) {
3381 *last = (StgBlockingQueueElement *)tso->link;
3382 if (mvar->tail == tso) {
3383 mvar->tail = (StgTSO *)last_tso;
3388 barf("unblockThread (MVAR): TSO not found");
3391 case BlockedOnBlackHole:
3392 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
3394 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
3396 last = &bq->blocking_queue;
3397 for (t = bq->blocking_queue;
3399 last = &t->link, t = t->link) {
3400 if (t == (StgBlockingQueueElement *)tso) {
3401 *last = (StgBlockingQueueElement *)tso->link;
3405 barf("unblockThread (BLACKHOLE): TSO not found");
3408 case BlockedOnException:
3410 StgTSO *target = tso->block_info.tso;
3412 ASSERT(get_itbl(target)->type == TSO);
3414 if (target->what_next == ThreadRelocated) {
3415 target = target->link;
3416 ASSERT(get_itbl(target)->type == TSO);
3419 ASSERT(target->blocked_exceptions != NULL);
3421 last = (StgBlockingQueueElement **)&target->blocked_exceptions;
3422 for (t = (StgBlockingQueueElement *)target->blocked_exceptions;
3424 last = &t->link, t = t->link) {
3425 ASSERT(get_itbl(t)->type == TSO);
3426 if (t == (StgBlockingQueueElement *)tso) {
3427 *last = (StgBlockingQueueElement *)tso->link;
3431 barf("unblockThread (Exception): TSO not found");
3435 case BlockedOnWrite:
3436 #if defined(mingw32_HOST_OS)
3437 case BlockedOnDoProc:
3440 /* take TSO off blocked_queue */
3441 StgBlockingQueueElement *prev = NULL;
3442 for (t = (StgBlockingQueueElement *)blocked_queue_hd; t != END_BQ_QUEUE;
3443 prev = t, t = t->link) {
3444 if (t == (StgBlockingQueueElement *)tso) {
3446 blocked_queue_hd = (StgTSO *)t->link;
3447 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
3448 blocked_queue_tl = END_TSO_QUEUE;
3451 prev->link = t->link;
3452 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
3453 blocked_queue_tl = (StgTSO *)prev;
3456 #if defined(mingw32_HOST_OS)
3457 /* (Cooperatively) signal that the worker thread should abort
3460 abandonWorkRequest(tso->block_info.async_result->reqID);
3465 barf("unblockThread (I/O): TSO not found");
3468 case BlockedOnDelay:
3470 /* take TSO off sleeping_queue */
3471 StgBlockingQueueElement *prev = NULL;
3472 for (t = (StgBlockingQueueElement *)sleeping_queue; t != END_BQ_QUEUE;
3473 prev = t, t = t->link) {
3474 if (t == (StgBlockingQueueElement *)tso) {
3476 sleeping_queue = (StgTSO *)t->link;
3478 prev->link = t->link;
3483 barf("unblockThread (delay): TSO not found");
3487 barf("unblockThread");
3491 tso->link = END_TSO_QUEUE;
3492 tso->why_blocked = NotBlocked;
3493 tso->block_info.closure = NULL;
3494 PUSH_ON_RUN_QUEUE(tso);
3498 unblockThread(StgTSO *tso)
3502 /* To avoid locking unnecessarily. */
3503 if (tso->why_blocked == NotBlocked) {
3507 switch (tso->why_blocked) {
3510 // Be careful: nothing to do here! We tell the scheduler that the thread
3511 // is runnable and we leave it to the stack-walking code to abort the
3512 // transaction while unwinding the stack. We should perhaps have a debugging
3513 // test to make sure that this really happens and that the 'zombie' transaction
3514 // does not get committed.
3518 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
3520 StgTSO *last_tso = END_TSO_QUEUE;
3521 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
3524 for (t = mvar->head; t != END_TSO_QUEUE;
3525 last = &t->link, last_tso = t, t = t->link) {
3528 if (mvar->tail == tso) {
3529 mvar->tail = last_tso;
3534 barf("unblockThread (MVAR): TSO not found");
3537 case BlockedOnBlackHole:
3539 last = &blackhole_queue;
3540 for (t = blackhole_queue; t != END_TSO_QUEUE;
3541 last = &t->link, t = t->link) {
3547 barf("unblockThread (BLACKHOLE): TSO not found");
3550 case BlockedOnException:
3552 StgTSO *target = tso->block_info.tso;
3554 ASSERT(get_itbl(target)->type == TSO);
3556 while (target->what_next == ThreadRelocated) {
3557 target = target->link;
3558 ASSERT(get_itbl(target)->type == TSO);
3561 ASSERT(target->blocked_exceptions != NULL);
3563 last = &target->blocked_exceptions;
3564 for (t = target->blocked_exceptions; t != END_TSO_QUEUE;
3565 last = &t->link, t = t->link) {
3566 ASSERT(get_itbl(t)->type == TSO);
3572 barf("unblockThread (Exception): TSO not found");
3576 case BlockedOnWrite:
3577 #if defined(mingw32_HOST_OS)
3578 case BlockedOnDoProc:
3581 StgTSO *prev = NULL;
3582 for (t = blocked_queue_hd; t != END_TSO_QUEUE;
3583 prev = t, t = t->link) {
3586 blocked_queue_hd = t->link;
3587 if (blocked_queue_tl == t) {
3588 blocked_queue_tl = END_TSO_QUEUE;
3591 prev->link = t->link;
3592 if (blocked_queue_tl == t) {
3593 blocked_queue_tl = prev;
3596 #if defined(mingw32_HOST_OS)
3597 /* (Cooperatively) signal that the worker thread should abort
3600 abandonWorkRequest(tso->block_info.async_result->reqID);
3605 barf("unblockThread (I/O): TSO not found");
3608 case BlockedOnDelay:
3610 StgTSO *prev = NULL;
3611 for (t = sleeping_queue; t != END_TSO_QUEUE;
3612 prev = t, t = t->link) {
3615 sleeping_queue = t->link;
3617 prev->link = t->link;
3622 barf("unblockThread (delay): TSO not found");
3626 barf("unblockThread");
3630 tso->link = END_TSO_QUEUE;
3631 tso->why_blocked = NotBlocked;
3632 tso->block_info.closure = NULL;
3633 APPEND_TO_RUN_QUEUE(tso);
3637 /* -----------------------------------------------------------------------------
3640 * Check the blackhole_queue for threads that can be woken up. We do
3641 * this periodically: before every GC, and whenever the run queue is
3644 * An elegant solution might be to just wake up all the blocked
3645 * threads with awakenBlockedQueue occasionally: they'll go back to
3646 * sleep again if the object is still a BLACKHOLE. Unfortunately this
3647 * doesn't give us a way to tell whether we've actually managed to
3648 * wake up any threads, so we would be busy-waiting.
3650 * -------------------------------------------------------------------------- */
3653 checkBlackHoles( void )
3656 rtsBool any_woke_up = rtsFalse;
3659 IF_DEBUG(scheduler, sched_belch("checking threads blocked on black holes"));
3661 // ASSUMES: sched_mutex
3662 prev = &blackhole_queue;
3663 t = blackhole_queue;
3664 while (t != END_TSO_QUEUE) {
3665 ASSERT(t->why_blocked == BlockedOnBlackHole);
3666 type = get_itbl(t->block_info.closure)->type;
3667 if (type != BLACKHOLE && type != CAF_BLACKHOLE) {
3668 IF_DEBUG(sanity,checkTSO(t));
3669 t = unblockOneLocked(t);
3671 any_woke_up = rtsTrue;
3681 /* -----------------------------------------------------------------------------
3684 * The following function implements the magic for raising an
3685 * asynchronous exception in an existing thread.
3687 * We first remove the thread from any queue on which it might be
3688 * blocked. The possible blockages are MVARs and BLACKHOLE_BQs.
3690 * We strip the stack down to the innermost CATCH_FRAME, building
3691 * thunks in the heap for all the active computations, so they can
3692 * be restarted if necessary. When we reach a CATCH_FRAME, we build
3693 * an application of the handler to the exception, and push it on
3694 * the top of the stack.
3696 * How exactly do we save all the active computations? We create an
3697 * AP_STACK for every UpdateFrame on the stack. Entering one of these
3698 * AP_STACKs pushes everything from the corresponding update frame
3699 * upwards onto the stack. (Actually, it pushes everything up to the
3700 * next update frame plus a pointer to the next AP_STACK object.
3701 * Entering the next AP_STACK object pushes more onto the stack until we
3702 * reach the last AP_STACK object - at which point the stack should look
3703 * exactly as it did when we killed the TSO and we can continue
3704 * execution by entering the closure on top of the stack.
3706 * We can also kill a thread entirely - this happens if either (a) the
3707 * exception passed to raiseAsync is NULL, or (b) there's no
3708 * CATCH_FRAME on the stack. In either case, we strip the entire
3709 * stack and replace the thread with a zombie.
3711 * Locks: sched_mutex held upon entry nor exit.
3713 * -------------------------------------------------------------------------- */
3716 deleteThread(StgTSO *tso)
3718 if (tso->why_blocked != BlockedOnCCall &&
3719 tso->why_blocked != BlockedOnCCall_NoUnblockExc) {
3720 raiseAsync(tso,NULL);
3724 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
3726 deleteThreadImmediately(StgTSO *tso)
3727 { // for forkProcess only:
3728 // delete thread without giving it a chance to catch the KillThread exception
3730 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
3734 if (tso->why_blocked != BlockedOnCCall &&
3735 tso->why_blocked != BlockedOnCCall_NoUnblockExc) {
3739 tso->what_next = ThreadKilled;
3744 raiseAsyncWithLock(StgTSO *tso, StgClosure *exception)
3746 /* When raising async exs from contexts where sched_mutex isn't held;
3747 use raiseAsyncWithLock(). */
3748 ACQUIRE_LOCK(&sched_mutex);
3749 raiseAsync(tso,exception);
3750 RELEASE_LOCK(&sched_mutex);
3754 raiseAsync(StgTSO *tso, StgClosure *exception)
3756 raiseAsync_(tso, exception, rtsFalse);
3760 raiseAsync_(StgTSO *tso, StgClosure *exception, rtsBool stop_at_atomically)
3762 StgRetInfoTable *info;
3765 // Thread already dead?
3766 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
3771 sched_belch("raising exception in thread %ld.", (long)tso->id));
3773 // Remove it from any blocking queues
3778 // The stack freezing code assumes there's a closure pointer on
3779 // the top of the stack, so we have to arrange that this is the case...
3781 if (sp[0] == (W_)&stg_enter_info) {
3785 sp[0] = (W_)&stg_dummy_ret_closure;
3791 // 1. Let the top of the stack be the "current closure"
3793 // 2. Walk up the stack until we find either an UPDATE_FRAME or a
3796 // 3. If it's an UPDATE_FRAME, then make an AP_STACK containing the
3797 // current closure applied to the chunk of stack up to (but not
3798 // including) the update frame. This closure becomes the "current
3799 // closure". Go back to step 2.
3801 // 4. If it's a CATCH_FRAME, then leave the exception handler on
3802 // top of the stack applied to the exception.
3804 // 5. If it's a STOP_FRAME, then kill the thread.
3806 // NB: if we pass an ATOMICALLY_FRAME then abort the associated
3813 info = get_ret_itbl((StgClosure *)frame);
3815 while (info->i.type != UPDATE_FRAME
3816 && (info->i.type != CATCH_FRAME || exception == NULL)
3817 && info->i.type != STOP_FRAME
3818 && (info->i.type != ATOMICALLY_FRAME || stop_at_atomically == rtsFalse))
3820 if (info->i.type == CATCH_RETRY_FRAME || info->i.type == ATOMICALLY_FRAME) {
3821 // IF we find an ATOMICALLY_FRAME then we abort the
3822 // current transaction and propagate the exception. In
3823 // this case (unlike ordinary exceptions) we do not care
3824 // whether the transaction is valid or not because its
3825 // possible validity cannot have caused the exception
3826 // and will not be visible after the abort.
3828 debugBelch("Found atomically block delivering async exception\n"));
3829 stmAbortTransaction(tso -> trec);
3830 tso -> trec = stmGetEnclosingTRec(tso -> trec);
3832 frame += stack_frame_sizeW((StgClosure *)frame);
3833 info = get_ret_itbl((StgClosure *)frame);
3836 switch (info->i.type) {
3838 case ATOMICALLY_FRAME:
3839 ASSERT(stop_at_atomically);
3840 ASSERT(stmGetEnclosingTRec(tso->trec) == NO_TREC);
3841 stmCondemnTransaction(tso -> trec);
3845 // R1 is not a register: the return convention for IO in
3846 // this case puts the return value on the stack, so we
3847 // need to set up the stack to return to the atomically
3848 // frame properly...
3849 tso->sp = frame - 2;
3850 tso->sp[1] = (StgWord) &stg_NO_FINALIZER_closure; // why not?
3851 tso->sp[0] = (StgWord) &stg_ut_1_0_unreg_info;
3853 tso->what_next = ThreadRunGHC;
3857 // If we find a CATCH_FRAME, and we've got an exception to raise,
3858 // then build the THUNK raise(exception), and leave it on
3859 // top of the CATCH_FRAME ready to enter.
3863 StgCatchFrame *cf = (StgCatchFrame *)frame;
3867 // we've got an exception to raise, so let's pass it to the
3868 // handler in this frame.
3870 raise = (StgThunk *)allocate(sizeofW(StgThunk)+MIN_UPD_SIZE);
3871 TICK_ALLOC_SE_THK(1,0);
3872 SET_HDR(raise,&stg_raise_info,cf->header.prof.ccs);
3873 raise->payload[0] = exception;
3875 // throw away the stack from Sp up to the CATCH_FRAME.
3879 /* Ensure that async excpetions are blocked now, so we don't get
3880 * a surprise exception before we get around to executing the
3883 if (tso->blocked_exceptions == NULL) {
3884 tso->blocked_exceptions = END_TSO_QUEUE;
3887 /* Put the newly-built THUNK on top of the stack, ready to execute
3888 * when the thread restarts.
3891 sp[-1] = (W_)&stg_enter_info;
3893 tso->what_next = ThreadRunGHC;
3894 IF_DEBUG(sanity, checkTSO(tso));
3903 // First build an AP_STACK consisting of the stack chunk above the
3904 // current update frame, with the top word on the stack as the
3907 words = frame - sp - 1;
3908 ap = (StgAP_STACK *)allocate(AP_STACK_sizeW(words));
3911 ap->fun = (StgClosure *)sp[0];
3913 for(i=0; i < (nat)words; ++i) {
3914 ap->payload[i] = (StgClosure *)*sp++;
3917 SET_HDR(ap,&stg_AP_STACK_info,
3918 ((StgClosure *)frame)->header.prof.ccs /* ToDo */);
3919 TICK_ALLOC_UP_THK(words+1,0);
3922 debugBelch("sched: Updating ");
3923 printPtr((P_)((StgUpdateFrame *)frame)->updatee);
3924 debugBelch(" with ");
3925 printObj((StgClosure *)ap);
3928 // Replace the updatee with an indirection - happily
3929 // this will also wake up any threads currently
3930 // waiting on the result.
3932 // Warning: if we're in a loop, more than one update frame on
3933 // the stack may point to the same object. Be careful not to
3934 // overwrite an IND_OLDGEN in this case, because we'll screw
3935 // up the mutable lists. To be on the safe side, don't
3936 // overwrite any kind of indirection at all. See also
3937 // threadSqueezeStack in GC.c, where we have to make a similar
3940 if (!closure_IND(((StgUpdateFrame *)frame)->updatee)) {
3941 // revert the black hole
3942 UPD_IND_NOLOCK(((StgUpdateFrame *)frame)->updatee,
3945 sp += sizeofW(StgUpdateFrame) - 1;
3946 sp[0] = (W_)ap; // push onto stack
3951 // We've stripped the entire stack, the thread is now dead.
3952 sp += sizeofW(StgStopFrame);
3953 tso->what_next = ThreadKilled;
3964 /* -----------------------------------------------------------------------------
3965 raiseExceptionHelper
3967 This function is called by the raise# primitve, just so that we can
3968 move some of the tricky bits of raising an exception from C-- into
3969 C. Who knows, it might be a useful re-useable thing here too.
3970 -------------------------------------------------------------------------- */
3973 raiseExceptionHelper (StgTSO *tso, StgClosure *exception)
3975 StgThunk *raise_closure = NULL;
3977 StgRetInfoTable *info;
3979 // This closure represents the expression 'raise# E' where E
3980 // is the exception raise. It is used to overwrite all the
3981 // thunks which are currently under evaluataion.
3985 // LDV profiling: stg_raise_info has THUNK as its closure
3986 // type. Since a THUNK takes at least MIN_UPD_SIZE words in its
3987 // payload, MIN_UPD_SIZE is more approprate than 1. It seems that
3988 // 1 does not cause any problem unless profiling is performed.
3989 // However, when LDV profiling goes on, we need to linearly scan
3990 // small object pool, where raise_closure is stored, so we should
3991 // use MIN_UPD_SIZE.
3993 // raise_closure = (StgClosure *)RET_STGCALL1(P_,allocate,
3994 // sizeofW(StgClosure)+1);
3998 // Walk up the stack, looking for the catch frame. On the way,
3999 // we update any closures pointed to from update frames with the
4000 // raise closure that we just built.
4004 info = get_ret_itbl((StgClosure *)p);
4005 next = p + stack_frame_sizeW((StgClosure *)p);
4006 switch (info->i.type) {
4009 // Only create raise_closure if we need to.
4010 if (raise_closure == NULL) {
4012 (StgThunk *)allocate(sizeofW(StgThunk)+MIN_UPD_SIZE);
4013 SET_HDR(raise_closure, &stg_raise_info, CCCS);
4014 raise_closure->payload[0] = exception;
4016 UPD_IND(((StgUpdateFrame *)p)->updatee,(StgClosure *)raise_closure);
4020 case ATOMICALLY_FRAME:
4021 IF_DEBUG(stm, debugBelch("Found ATOMICALLY_FRAME at %p\n", p));
4023 return ATOMICALLY_FRAME;
4029 case CATCH_STM_FRAME:
4030 IF_DEBUG(stm, debugBelch("Found CATCH_STM_FRAME at %p\n", p));
4032 return CATCH_STM_FRAME;
4038 case CATCH_RETRY_FRAME:
4047 /* -----------------------------------------------------------------------------
4048 findRetryFrameHelper
4050 This function is called by the retry# primitive. It traverses the stack
4051 leaving tso->sp referring to the frame which should handle the retry.
4053 This should either be a CATCH_RETRY_FRAME (if the retry# is within an orElse#)
4054 or should be a ATOMICALLY_FRAME (if the retry# reaches the top level).
4056 We skip CATCH_STM_FRAMEs because retries are not considered to be exceptions,
4057 despite the similar implementation.
4059 We should not expect to see CATCH_FRAME or STOP_FRAME because those should
4060 not be created within memory transactions.
4061 -------------------------------------------------------------------------- */
4064 findRetryFrameHelper (StgTSO *tso)
4067 StgRetInfoTable *info;
4071 info = get_ret_itbl((StgClosure *)p);
4072 next = p + stack_frame_sizeW((StgClosure *)p);
4073 switch (info->i.type) {
4075 case ATOMICALLY_FRAME:
4076 IF_DEBUG(stm, debugBelch("Found ATOMICALLY_FRAME at %p during retrry\n", p));
4078 return ATOMICALLY_FRAME;
4080 case CATCH_RETRY_FRAME:
4081 IF_DEBUG(stm, debugBelch("Found CATCH_RETRY_FRAME at %p during retrry\n", p));
4083 return CATCH_RETRY_FRAME;
4085 case CATCH_STM_FRAME:
4087 ASSERT(info->i.type != CATCH_FRAME);
4088 ASSERT(info->i.type != STOP_FRAME);
4095 /* -----------------------------------------------------------------------------
4096 resurrectThreads is called after garbage collection on the list of
4097 threads found to be garbage. Each of these threads will be woken
4098 up and sent a signal: BlockedOnDeadMVar if the thread was blocked
4099 on an MVar, or NonTermination if the thread was blocked on a Black
4102 Locks: sched_mutex isn't held upon entry nor exit.
4103 -------------------------------------------------------------------------- */
4106 resurrectThreads( StgTSO *threads )
4110 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
4111 next = tso->global_link;
4112 tso->global_link = all_threads;
4114 IF_DEBUG(scheduler, sched_belch("resurrecting thread %d", tso->id));
4116 switch (tso->why_blocked) {
4118 case BlockedOnException:
4119 /* Called by GC - sched_mutex lock is currently held. */
4120 raiseAsync(tso,(StgClosure *)BlockedOnDeadMVar_closure);
4122 case BlockedOnBlackHole:
4123 raiseAsync(tso,(StgClosure *)NonTermination_closure);
4126 raiseAsync(tso,(StgClosure *)BlockedIndefinitely_closure);
4129 /* This might happen if the thread was blocked on a black hole
4130 * belonging to a thread that we've just woken up (raiseAsync
4131 * can wake up threads, remember...).
4135 barf("resurrectThreads: thread blocked in a strange way");
4140 /* ----------------------------------------------------------------------------
4141 * Debugging: why is a thread blocked
4142 * [Also provides useful information when debugging threaded programs
4143 * at the Haskell source code level, so enable outside of DEBUG. --sof 7/02]
4144 ------------------------------------------------------------------------- */
4147 printThreadBlockage(StgTSO *tso)
4149 switch (tso->why_blocked) {
4151 debugBelch("is blocked on read from fd %d", (int)(tso->block_info.fd));
4153 case BlockedOnWrite:
4154 debugBelch("is blocked on write to fd %d", (int)(tso->block_info.fd));
4156 #if defined(mingw32_HOST_OS)
4157 case BlockedOnDoProc:
4158 debugBelch("is blocked on proc (request: %ld)", tso->block_info.async_result->reqID);
4161 case BlockedOnDelay:
4162 debugBelch("is blocked until %ld", (long)(tso->block_info.target));
4165 debugBelch("is blocked on an MVar @ %p", tso->block_info.closure);
4167 case BlockedOnException:
4168 debugBelch("is blocked on delivering an exception to thread %d",
4169 tso->block_info.tso->id);
4171 case BlockedOnBlackHole:
4172 debugBelch("is blocked on a black hole");
4175 debugBelch("is not blocked");
4177 #if defined(PARALLEL_HASKELL)
4179 debugBelch("is blocked on global address; local FM_BQ is %p (%s)",
4180 tso->block_info.closure, info_type(tso->block_info.closure));
4182 case BlockedOnGA_NoSend:
4183 debugBelch("is blocked on global address (no send); local FM_BQ is %p (%s)",
4184 tso->block_info.closure, info_type(tso->block_info.closure));
4187 case BlockedOnCCall:
4188 debugBelch("is blocked on an external call");
4190 case BlockedOnCCall_NoUnblockExc:
4191 debugBelch("is blocked on an external call (exceptions were already blocked)");
4194 debugBelch("is blocked on an STM operation");
4197 barf("printThreadBlockage: strange tso->why_blocked: %d for TSO %d (%d)",
4198 tso->why_blocked, tso->id, tso);
4203 printThreadStatus(StgTSO *tso)
4205 switch (tso->what_next) {
4207 debugBelch("has been killed");
4209 case ThreadComplete:
4210 debugBelch("has completed");
4213 printThreadBlockage(tso);
4218 printAllThreads(void)
4223 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
4224 ullong_format_string(TIME_ON_PROC(CurrentProc),
4225 time_string, rtsFalse/*no commas!*/);
4227 debugBelch("all threads at [%s]:\n", time_string);
4228 # elif defined(PARALLEL_HASKELL)
4229 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
4230 ullong_format_string(CURRENT_TIME,
4231 time_string, rtsFalse/*no commas!*/);
4233 debugBelch("all threads at [%s]:\n", time_string);
4235 debugBelch("all threads:\n");
4238 for (t = all_threads; t != END_TSO_QUEUE; ) {
4239 debugBelch("\tthread %4d @ %p ", t->id, (void *)t);
4242 void *label = lookupThreadLabel(t->id);
4243 if (label) debugBelch("[\"%s\"] ",(char *)label);
4246 if (t->what_next == ThreadRelocated) {
4247 debugBelch("has been relocated...\n");
4250 printThreadStatus(t);
4261 printThreadQueue(StgTSO *t)
4264 for (; t != END_TSO_QUEUE; t = t->link) {
4265 debugBelch("\tthread %d @ %p ", t->id, (void *)t);
4266 if (t->what_next == ThreadRelocated) {
4267 debugBelch("has been relocated...\n");
4269 printThreadStatus(t);
4274 debugBelch("%d threads on queue\n", i);
4278 Print a whole blocking queue attached to node (debugging only).
4280 # if defined(PARALLEL_HASKELL)
4282 print_bq (StgClosure *node)
4284 StgBlockingQueueElement *bqe;
4288 debugBelch("## BQ of closure %p (%s): ",
4289 node, info_type(node));
4291 /* should cover all closures that may have a blocking queue */
4292 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
4293 get_itbl(node)->type == FETCH_ME_BQ ||
4294 get_itbl(node)->type == RBH ||
4295 get_itbl(node)->type == MVAR);
4297 ASSERT(node!=(StgClosure*)NULL); // sanity check
4299 print_bqe(((StgBlockingQueue*)node)->blocking_queue);
4303 Print a whole blocking queue starting with the element bqe.
4306 print_bqe (StgBlockingQueueElement *bqe)
4311 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
4313 for (end = (bqe==END_BQ_QUEUE);
4314 !end; // iterate until bqe points to a CONSTR
4315 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE),
4316 bqe = end ? END_BQ_QUEUE : bqe->link) {
4317 ASSERT(bqe != END_BQ_QUEUE); // sanity check
4318 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
4319 /* types of closures that may appear in a blocking queue */
4320 ASSERT(get_itbl(bqe)->type == TSO ||
4321 get_itbl(bqe)->type == BLOCKED_FETCH ||
4322 get_itbl(bqe)->type == CONSTR);
4323 /* only BQs of an RBH end with an RBH_Save closure */
4324 //ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
4326 switch (get_itbl(bqe)->type) {
4328 debugBelch(" TSO %u (%x),",
4329 ((StgTSO *)bqe)->id, ((StgTSO *)bqe));
4332 debugBelch(" BF (node=%p, ga=((%x, %d, %x)),",
4333 ((StgBlockedFetch *)bqe)->node,
4334 ((StgBlockedFetch *)bqe)->ga.payload.gc.gtid,
4335 ((StgBlockedFetch *)bqe)->ga.payload.gc.slot,
4336 ((StgBlockedFetch *)bqe)->ga.weight);
4339 debugBelch(" %s (IP %p),",
4340 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
4341 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
4342 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
4343 "RBH_Save_?"), get_itbl(bqe));
4346 barf("Unexpected closure type %s in blocking queue", // of %p (%s)",
4347 info_type((StgClosure *)bqe)); // , node, info_type(node));
4353 # elif defined(GRAN)
4355 print_bq (StgClosure *node)
4357 StgBlockingQueueElement *bqe;
4358 PEs node_loc, tso_loc;
4361 /* should cover all closures that may have a blocking queue */
4362 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
4363 get_itbl(node)->type == FETCH_ME_BQ ||
4364 get_itbl(node)->type == RBH);
4366 ASSERT(node!=(StgClosure*)NULL); // sanity check
4367 node_loc = where_is(node);
4369 debugBelch("## BQ of closure %p (%s) on [PE %d]: ",
4370 node, info_type(node), node_loc);
4373 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
4375 for (bqe = ((StgBlockingQueue*)node)->blocking_queue, end = (bqe==END_BQ_QUEUE);
4376 !end; // iterate until bqe points to a CONSTR
4377 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) {
4378 ASSERT(bqe != END_BQ_QUEUE); // sanity check
4379 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
4380 /* types of closures that may appear in a blocking queue */
4381 ASSERT(get_itbl(bqe)->type == TSO ||
4382 get_itbl(bqe)->type == CONSTR);
4383 /* only BQs of an RBH end with an RBH_Save closure */
4384 ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
4386 tso_loc = where_is((StgClosure *)bqe);
4387 switch (get_itbl(bqe)->type) {
4389 debugBelch(" TSO %d (%p) on [PE %d],",
4390 ((StgTSO *)bqe)->id, (StgTSO *)bqe, tso_loc);
4393 debugBelch(" %s (IP %p),",
4394 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
4395 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
4396 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
4397 "RBH_Save_?"), get_itbl(bqe));
4400 barf("Unexpected closure type %s in blocking queue of %p (%s)",
4401 info_type((StgClosure *)bqe), node, info_type(node));
4409 #if defined(PARALLEL_HASKELL)
4416 for (i=0, tso=run_queue_hd;
4417 tso != END_TSO_QUEUE;
4426 sched_belch(char *s, ...)
4430 #ifdef RTS_SUPPORTS_THREADS
4431 debugBelch("sched (task %p): ", osThreadId());
4432 #elif defined(PARALLEL_HASKELL)
4435 debugBelch("sched: ");