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.
1489 bd = allocGroup( blocks );
1490 cap->r.rNursery->n_blocks += blocks;
1492 // link the new group into the list
1493 bd->link = cap->r.rCurrentNursery;
1494 bd->u.back = cap->r.rCurrentNursery->u.back;
1495 if (cap->r.rCurrentNursery->u.back != NULL) {
1496 cap->r.rCurrentNursery->u.back->link = bd;
1499 ASSERT(g0s0->blocks == cap->r.rCurrentNursery &&
1500 g0s0 == cap->r.rNursery);
1502 cap->r.rNursery->blocks = bd;
1504 cap->r.rCurrentNursery->u.back = bd;
1506 // initialise it as a nursery block. We initialise the
1507 // step, gen_no, and flags field of *every* sub-block in
1508 // this large block, because this is easier than making
1509 // sure that we always find the block head of a large
1510 // block whenever we call Bdescr() (eg. evacuate() and
1511 // isAlive() in the GC would both have to do this, at
1515 for (x = bd; x < bd + blocks; x++) {
1516 x->step = cap->r.rNursery;
1522 // This assert can be a killer if the app is doing lots
1523 // of large block allocations.
1524 IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
1526 // now update the nursery to point to the new block
1527 cap->r.rCurrentNursery = bd;
1529 // we might be unlucky and have another thread get on the
1530 // run queue before us and steal the large block, but in that
1531 // case the thread will just end up requesting another large
1533 PUSH_ON_RUN_QUEUE(t);
1534 return rtsFalse; /* not actually GC'ing */
1539 debugBelch("--<< thread %ld (%s) stopped: HeapOverflow\n",
1540 (long)t->id, whatNext_strs[t->what_next]));
1542 ASSERT(!is_on_queue(t,CurrentProc));
1543 #elif defined(PARALLEL_HASKELL)
1544 /* Currently we emit a DESCHEDULE event before GC in GUM.
1545 ToDo: either add separate event to distinguish SYSTEM time from rest
1546 or just nuke this DESCHEDULE (and the following SCHEDULE) */
1547 if (0 && RtsFlags.ParFlags.ParStats.Full) {
1548 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
1549 GR_DESCHEDULE, t, (StgClosure *)NULL, 0, 0);
1550 emitSchedule = rtsTrue;
1554 PUSH_ON_RUN_QUEUE(t);
1556 /* actual GC is done at the end of the while loop in schedule() */
1559 /* -----------------------------------------------------------------------------
1560 * Handle a thread that returned to the scheduler with ThreadStackOverflow
1561 * ASSUMES: sched_mutex
1562 * -------------------------------------------------------------------------- */
1565 scheduleHandleStackOverflow( StgTSO *t)
1567 IF_DEBUG(scheduler,debugBelch("--<< thread %ld (%s) stopped, StackOverflow\n",
1568 (long)t->id, whatNext_strs[t->what_next]));
1569 /* just adjust the stack for this thread, then pop it back
1573 /* enlarge the stack */
1574 StgTSO *new_t = threadStackOverflow(t);
1576 /* This TSO has moved, so update any pointers to it from the
1577 * main thread stack. It better not be on any other queues...
1578 * (it shouldn't be).
1580 if (t->main != NULL) {
1581 t->main->tso = new_t;
1583 PUSH_ON_RUN_QUEUE(new_t);
1587 /* -----------------------------------------------------------------------------
1588 * Handle a thread that returned to the scheduler with ThreadYielding
1589 * ASSUMES: sched_mutex
1590 * -------------------------------------------------------------------------- */
1593 scheduleHandleYield( StgTSO *t, nat prev_what_next )
1595 // Reset the context switch flag. We don't do this just before
1596 // running the thread, because that would mean we would lose ticks
1597 // during GC, which can lead to unfair scheduling (a thread hogs
1598 // the CPU because the tick always arrives during GC). This way
1599 // penalises threads that do a lot of allocation, but that seems
1600 // better than the alternative.
1603 /* put the thread back on the run queue. Then, if we're ready to
1604 * GC, check whether this is the last task to stop. If so, wake
1605 * up the GC thread. getThread will block during a GC until the
1609 if (t->what_next != prev_what_next) {
1610 debugBelch("--<< thread %ld (%s) stopped to switch evaluators\n",
1611 (long)t->id, whatNext_strs[t->what_next]);
1613 debugBelch("--<< thread %ld (%s) stopped, yielding\n",
1614 (long)t->id, whatNext_strs[t->what_next]);
1619 //debugBelch("&& Doing sanity check on yielding TSO %ld.", t->id);
1621 ASSERT(t->link == END_TSO_QUEUE);
1623 // Shortcut if we're just switching evaluators: don't bother
1624 // doing stack squeezing (which can be expensive), just run the
1626 if (t->what_next != prev_what_next) {
1631 ASSERT(!is_on_queue(t,CurrentProc));
1634 //debugBelch("&& Doing sanity check on all ThreadQueues (and their TSOs).");
1635 checkThreadQsSanity(rtsTrue));
1642 /* add a ContinueThread event to actually process the thread */
1643 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1645 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1647 debugBelch("GRAN: eventq and runnableq after adding yielded thread to queue again:\n");
1654 /* -----------------------------------------------------------------------------
1655 * Handle a thread that returned to the scheduler with ThreadBlocked
1656 * ASSUMES: sched_mutex
1657 * -------------------------------------------------------------------------- */
1660 scheduleHandleThreadBlocked( StgTSO *t
1661 #if !defined(GRAN) && !defined(DEBUG)
1668 debugBelch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: \n",
1669 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)));
1670 if (t->block_info.closure!=(StgClosure*)NULL) print_bq(t->block_info.closure));
1672 // ??? needed; should emit block before
1674 DumpGranEvent(GR_DESCHEDULE, t));
1675 prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
1678 ASSERT(procStatus[CurrentProc]==Busy ||
1679 ((procStatus[CurrentProc]==Fetching) &&
1680 (t->block_info.closure!=(StgClosure*)NULL)));
1681 if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
1682 !(!RtsFlags.GranFlags.DoAsyncFetch &&
1683 procStatus[CurrentProc]==Fetching))
1684 procStatus[CurrentProc] = Idle;
1688 debugBelch("--<< thread %ld (%p; %s) stopped, blocking on node %p with BQ: \n",
1689 t->id, t, whatNext_strs[t->what_next], t->block_info.closure));
1692 if (t->block_info.closure!=(StgClosure*)NULL)
1693 print_bq(t->block_info.closure));
1695 /* Send a fetch (if BlockedOnGA) and dump event to log file */
1698 /* whatever we schedule next, we must log that schedule */
1699 emitSchedule = rtsTrue;
1703 // We don't need to do anything. The thread is blocked, and it
1704 // has tidied up its stack and placed itself on whatever queue
1705 // it needs to be on.
1708 ASSERT(t->why_blocked != NotBlocked);
1709 // This might not be true under SMP: we don't have
1710 // exclusive access to this TSO, so someone might have
1711 // woken it up by now. This actually happens: try
1712 // conc023 +RTS -N2.
1716 debugBelch("--<< thread %d (%s) stopped: ",
1717 t->id, whatNext_strs[t->what_next]);
1718 printThreadBlockage(t);
1721 /* Only for dumping event to log file
1722 ToDo: do I need this in GranSim, too?
1728 /* -----------------------------------------------------------------------------
1729 * Handle a thread that returned to the scheduler with ThreadFinished
1730 * ASSUMES: sched_mutex
1731 * -------------------------------------------------------------------------- */
1734 scheduleHandleThreadFinished( StgMainThread *mainThread
1735 USED_WHEN_RTS_SUPPORTS_THREADS,
1739 /* Need to check whether this was a main thread, and if so,
1740 * return with the return value.
1742 * We also end up here if the thread kills itself with an
1743 * uncaught exception, see Exception.cmm.
1745 IF_DEBUG(scheduler,debugBelch("--++ thread %d (%s) finished\n",
1746 t->id, whatNext_strs[t->what_next]));
1749 endThread(t, CurrentProc); // clean-up the thread
1750 #elif defined(PARALLEL_HASKELL)
1751 /* For now all are advisory -- HWL */
1752 //if(t->priority==AdvisoryPriority) ??
1753 advisory_thread_count--; // JB: Caution with this counter, buggy!
1756 if(t->dist.priority==RevalPriority)
1760 # if defined(EDENOLD)
1761 // the thread could still have an outport... (BUG)
1762 if (t->eden.outport != -1) {
1763 // delete the outport for the tso which has finished...
1764 IF_PAR_DEBUG(eden_ports,
1765 debugBelch("WARNING: Scheduler removes outport %d for TSO %d.\n",
1766 t->eden.outport, t->id));
1769 // thread still in the process (HEAVY BUG! since outport has just been closed...)
1770 if (t->eden.epid != -1) {
1771 IF_PAR_DEBUG(eden_ports,
1772 debugBelch("WARNING: Scheduler removes TSO %d from process %d .\n",
1773 t->id, t->eden.epid));
1774 removeTSOfromProcess(t);
1779 if (RtsFlags.ParFlags.ParStats.Full &&
1780 !RtsFlags.ParFlags.ParStats.Suppressed)
1781 DumpEndEvent(CURRENT_PROC, t, rtsFalse /* not mandatory */);
1783 // t->par only contains statistics: left out for now...
1785 debugBelch("**** end thread: ended sparked thread %d (%lx); sparkname: %lx\n",
1786 t->id,t,t->par.sparkname));
1788 #endif // PARALLEL_HASKELL
1791 // Check whether the thread that just completed was a main
1792 // thread, and if so return with the result.
1794 // There is an assumption here that all thread completion goes
1795 // through this point; we need to make sure that if a thread
1796 // ends up in the ThreadKilled state, that it stays on the run
1797 // queue so it can be dealt with here.
1800 #if defined(RTS_SUPPORTS_THREADS)
1803 mainThread->tso == t
1807 // We are a bound thread: this must be our thread that just
1809 ASSERT(mainThread->tso == t);
1811 if (t->what_next == ThreadComplete) {
1812 if (mainThread->ret) {
1813 // NOTE: return val is tso->sp[1] (see StgStartup.hc)
1814 *(mainThread->ret) = (StgClosure *)mainThread->tso->sp[1];
1816 mainThread->stat = Success;
1818 if (mainThread->ret) {
1819 *(mainThread->ret) = NULL;
1822 mainThread->stat = Interrupted;
1824 mainThread->stat = Killed;
1828 removeThreadLabel((StgWord)mainThread->tso->id);
1830 if (mainThread->prev == NULL) {
1831 ASSERT(mainThread == main_threads);
1832 main_threads = mainThread->link;
1834 mainThread->prev->link = mainThread->link;
1836 if (mainThread->link != NULL) {
1837 mainThread->link->prev = mainThread->prev;
1839 releaseCapability(cap);
1840 return rtsTrue; // tells schedule() to return
1843 #ifdef RTS_SUPPORTS_THREADS
1844 ASSERT(t->main == NULL);
1846 if (t->main != NULL) {
1847 // Must be a main thread that is not the topmost one. Leave
1848 // it on the run queue until the stack has unwound to the
1849 // point where we can deal with this. Leaving it on the run
1850 // queue also ensures that the garbage collector knows about
1851 // this thread and its return value (it gets dropped from the
1852 // all_threads list so there's no other way to find it).
1853 APPEND_TO_RUN_QUEUE(t);
1859 /* -----------------------------------------------------------------------------
1860 * Perform a heap census, if PROFILING
1861 * -------------------------------------------------------------------------- */
1864 scheduleDoHeapProfile( rtsBool ready_to_gc STG_UNUSED )
1866 #if defined(PROFILING)
1867 // When we have +RTS -i0 and we're heap profiling, do a census at
1868 // every GC. This lets us get repeatable runs for debugging.
1869 if (performHeapProfile ||
1870 (RtsFlags.ProfFlags.profileInterval==0 &&
1871 RtsFlags.ProfFlags.doHeapProfile && ready_to_gc)) {
1872 GarbageCollect(GetRoots, rtsTrue);
1874 performHeapProfile = rtsFalse;
1875 return rtsTrue; // true <=> we already GC'd
1881 /* -----------------------------------------------------------------------------
1882 * Perform a garbage collection if necessary
1883 * ASSUMES: sched_mutex
1884 * -------------------------------------------------------------------------- */
1887 scheduleDoGC( rtsBool force_major )
1892 static rtsBool waiting_for_gc;
1893 int n_capabilities = RtsFlags.ParFlags.nNodes - 1;
1894 // subtract one because we're already holding one.
1895 Capability *caps[n_capabilities];
1899 // In order to GC, there must be no threads running Haskell code.
1900 // Therefore, the GC thread needs to hold *all* the capabilities,
1901 // and release them after the GC has completed.
1903 // This seems to be the simplest way: previous attempts involved
1904 // making all the threads with capabilities give up their
1905 // capabilities and sleep except for the *last* one, which
1906 // actually did the GC. But it's quite hard to arrange for all
1907 // the other tasks to sleep and stay asleep.
1909 // This does mean that there will be multiple entries in the
1910 // thread->capability hash table for the current thread, but
1911 // they will be removed as normal when the capabilities are
1915 // Someone else is already trying to GC
1916 if (waiting_for_gc) return;
1917 waiting_for_gc = rtsTrue;
1919 while (n_capabilities > 0) {
1920 IF_DEBUG(scheduler, sched_belch("ready_to_gc, grabbing all the capabilies (%d left)", n_capabilities));
1921 waitForReturnCapability(&sched_mutex, &cap);
1923 caps[n_capabilities] = cap;
1926 waiting_for_gc = rtsFalse;
1929 /* Kick any transactions which are invalid back to their
1930 * atomically frames. When next scheduled they will try to
1931 * commit, this commit will fail and they will retry.
1936 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
1937 if (t->what_next == ThreadRelocated) {
1940 next = t->global_link;
1941 if (t -> trec != NO_TREC && t -> why_blocked == NotBlocked) {
1942 if (!stmValidateNestOfTransactions (t -> trec)) {
1943 IF_DEBUG(stm, sched_belch("trec %p found wasting its time", t));
1945 // strip the stack back to the ATOMICALLY_FRAME, aborting
1946 // the (nested) transaction, and saving the stack of any
1947 // partially-evaluated thunks on the heap.
1948 raiseAsync_(t, NULL, rtsTrue);
1951 ASSERT(get_itbl((StgClosure *)t->sp)->type == ATOMICALLY_FRAME);
1959 // so this happens periodically:
1960 scheduleCheckBlackHoles();
1962 IF_DEBUG(scheduler, printAllThreads());
1964 /* everybody back, start the GC.
1965 * Could do it in this thread, or signal a condition var
1966 * to do it in another thread. Either way, we need to
1967 * broadcast on gc_pending_cond afterward.
1969 #if defined(RTS_SUPPORTS_THREADS)
1970 IF_DEBUG(scheduler,sched_belch("doing GC"));
1972 GarbageCollect(GetRoots, force_major);
1976 // release our stash of capabilities.
1978 for (i = 0; i < RtsFlags.ParFlags.nNodes-1; i++) {
1979 releaseCapability(caps[i]);
1985 /* add a ContinueThread event to continue execution of current thread */
1986 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1988 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1990 debugBelch("GRAN: eventq and runnableq after Garbage collection:\n\n");
1996 /* ---------------------------------------------------------------------------
1997 * rtsSupportsBoundThreads(): is the RTS built to support bound threads?
1998 * used by Control.Concurrent for error checking.
1999 * ------------------------------------------------------------------------- */
2002 rtsSupportsBoundThreads(void)
2004 #if defined(RTS_SUPPORTS_THREADS)
2011 /* ---------------------------------------------------------------------------
2012 * isThreadBound(tso): check whether tso is bound to an OS thread.
2013 * ------------------------------------------------------------------------- */
2016 isThreadBound(StgTSO* tso USED_IN_THREADED_RTS)
2018 #if defined(RTS_SUPPORTS_THREADS)
2019 return (tso->main != NULL);
2024 /* ---------------------------------------------------------------------------
2025 * Singleton fork(). Do not copy any running threads.
2026 * ------------------------------------------------------------------------- */
2028 #ifndef mingw32_HOST_OS
2029 #define FORKPROCESS_PRIMOP_SUPPORTED
2032 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
2034 deleteThreadImmediately(StgTSO *tso);
2037 forkProcess(HsStablePtr *entry
2038 #ifndef FORKPROCESS_PRIMOP_SUPPORTED
2043 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
2049 IF_DEBUG(scheduler,sched_belch("forking!"));
2050 rts_lock(); // This not only acquires sched_mutex, it also
2051 // makes sure that no other threads are running
2055 if (pid) { /* parent */
2057 /* just return the pid */
2061 } else { /* child */
2064 // delete all threads
2065 run_queue_hd = run_queue_tl = END_TSO_QUEUE;
2067 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
2070 // don't allow threads to catch the ThreadKilled exception
2071 deleteThreadImmediately(t);
2074 // wipe the main thread list
2075 while((m = main_threads) != NULL) {
2076 main_threads = m->link;
2077 # ifdef THREADED_RTS
2078 closeCondition(&m->bound_thread_cond);
2083 rc = rts_evalStableIO(entry, NULL); // run the action
2084 rts_checkSchedStatus("forkProcess",rc);
2088 hs_exit(); // clean up and exit
2091 #else /* !FORKPROCESS_PRIMOP_SUPPORTED */
2092 barf("forkProcess#: primop not supported, sorry!\n");
2097 /* ---------------------------------------------------------------------------
2098 * deleteAllThreads(): kill all the live threads.
2100 * This is used when we catch a user interrupt (^C), before performing
2101 * any necessary cleanups and running finalizers.
2103 * Locks: sched_mutex held.
2104 * ------------------------------------------------------------------------- */
2107 deleteAllThreads ( void )
2110 IF_DEBUG(scheduler,sched_belch("deleting all threads"));
2111 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
2112 if (t->what_next == ThreadRelocated) {
2115 next = t->global_link;
2120 // The run queue now contains a bunch of ThreadKilled threads. We
2121 // must not throw these away: the main thread(s) will be in there
2122 // somewhere, and the main scheduler loop has to deal with it.
2123 // Also, the run queue is the only thing keeping these threads from
2124 // being GC'd, and we don't want the "main thread has been GC'd" panic.
2126 ASSERT(blocked_queue_hd == END_TSO_QUEUE);
2127 ASSERT(blackhole_queue == END_TSO_QUEUE);
2128 ASSERT(sleeping_queue == END_TSO_QUEUE);
2131 /* startThread and insertThread are now in GranSim.c -- HWL */
2134 /* ---------------------------------------------------------------------------
2135 * Suspending & resuming Haskell threads.
2137 * When making a "safe" call to C (aka _ccall_GC), the task gives back
2138 * its capability before calling the C function. This allows another
2139 * task to pick up the capability and carry on running Haskell
2140 * threads. It also means that if the C call blocks, it won't lock
2143 * The Haskell thread making the C call is put to sleep for the
2144 * duration of the call, on the susepended_ccalling_threads queue. We
2145 * give out a token to the task, which it can use to resume the thread
2146 * on return from the C function.
2147 * ------------------------------------------------------------------------- */
2150 suspendThread( StgRegTable *reg )
2154 int saved_errno = errno;
2156 /* assume that *reg is a pointer to the StgRegTable part
2159 cap = (Capability *)((void *)((unsigned char*)reg - sizeof(StgFunTable)));
2161 ACQUIRE_LOCK(&sched_mutex);
2164 sched_belch("thread %d did a _ccall_gc", cap->r.rCurrentTSO->id));
2166 // XXX this might not be necessary --SDM
2167 cap->r.rCurrentTSO->what_next = ThreadRunGHC;
2169 threadPaused(cap->r.rCurrentTSO);
2170 cap->r.rCurrentTSO->link = suspended_ccalling_threads;
2171 suspended_ccalling_threads = cap->r.rCurrentTSO;
2173 if(cap->r.rCurrentTSO->blocked_exceptions == NULL) {
2174 cap->r.rCurrentTSO->why_blocked = BlockedOnCCall;
2175 cap->r.rCurrentTSO->blocked_exceptions = END_TSO_QUEUE;
2177 cap->r.rCurrentTSO->why_blocked = BlockedOnCCall_NoUnblockExc;
2180 /* Use the thread ID as the token; it should be unique */
2181 tok = cap->r.rCurrentTSO->id;
2183 /* Hand back capability */
2184 cap->r.rInHaskell = rtsFalse;
2185 releaseCapability(cap);
2187 #if defined(RTS_SUPPORTS_THREADS)
2188 /* Preparing to leave the RTS, so ensure there's a native thread/task
2189 waiting to take over.
2191 IF_DEBUG(scheduler, sched_belch("worker (token %d): leaving RTS", tok));
2194 RELEASE_LOCK(&sched_mutex);
2196 errno = saved_errno;
2201 resumeThread( StgInt tok )
2203 StgTSO *tso, **prev;
2205 int saved_errno = errno;
2207 #if defined(RTS_SUPPORTS_THREADS)
2208 /* Wait for permission to re-enter the RTS with the result. */
2209 ACQUIRE_LOCK(&sched_mutex);
2210 waitForReturnCapability(&sched_mutex, &cap);
2212 IF_DEBUG(scheduler, sched_belch("worker (token %d): re-entering RTS", tok));
2214 grabCapability(&cap);
2217 /* Remove the thread off of the suspended list */
2218 prev = &suspended_ccalling_threads;
2219 for (tso = suspended_ccalling_threads;
2220 tso != END_TSO_QUEUE;
2221 prev = &tso->link, tso = tso->link) {
2222 if (tso->id == (StgThreadID)tok) {
2227 if (tso == END_TSO_QUEUE) {
2228 barf("resumeThread: thread not found");
2230 tso->link = END_TSO_QUEUE;
2232 if(tso->why_blocked == BlockedOnCCall) {
2233 awakenBlockedQueueNoLock(tso->blocked_exceptions);
2234 tso->blocked_exceptions = NULL;
2237 /* Reset blocking status */
2238 tso->why_blocked = NotBlocked;
2240 cap->r.rCurrentTSO = tso;
2241 cap->r.rInHaskell = rtsTrue;
2242 RELEASE_LOCK(&sched_mutex);
2243 errno = saved_errno;
2247 /* ---------------------------------------------------------------------------
2248 * Comparing Thread ids.
2250 * This is used from STG land in the implementation of the
2251 * instances of Eq/Ord for ThreadIds.
2252 * ------------------------------------------------------------------------ */
2255 cmp_thread(StgPtr tso1, StgPtr tso2)
2257 StgThreadID id1 = ((StgTSO *)tso1)->id;
2258 StgThreadID id2 = ((StgTSO *)tso2)->id;
2260 if (id1 < id2) return (-1);
2261 if (id1 > id2) return 1;
2265 /* ---------------------------------------------------------------------------
2266 * Fetching the ThreadID from an StgTSO.
2268 * This is used in the implementation of Show for ThreadIds.
2269 * ------------------------------------------------------------------------ */
2271 rts_getThreadId(StgPtr tso)
2273 return ((StgTSO *)tso)->id;
2278 labelThread(StgPtr tso, char *label)
2283 /* Caveat: Once set, you can only set the thread name to "" */
2284 len = strlen(label)+1;
2285 buf = stgMallocBytes(len * sizeof(char), "Schedule.c:labelThread()");
2286 strncpy(buf,label,len);
2287 /* Update will free the old memory for us */
2288 updateThreadLabel(((StgTSO *)tso)->id,buf);
2292 /* ---------------------------------------------------------------------------
2293 Create a new thread.
2295 The new thread starts with the given stack size. Before the
2296 scheduler can run, however, this thread needs to have a closure
2297 (and possibly some arguments) pushed on its stack. See
2298 pushClosure() in Schedule.h.
2300 createGenThread() and createIOThread() (in SchedAPI.h) are
2301 convenient packaged versions of this function.
2303 currently pri (priority) is only used in a GRAN setup -- HWL
2304 ------------------------------------------------------------------------ */
2306 /* currently pri (priority) is only used in a GRAN setup -- HWL */
2308 createThread(nat size, StgInt pri)
2311 createThread(nat size)
2317 /* First check whether we should create a thread at all */
2318 #if defined(PARALLEL_HASKELL)
2319 /* check that no more than RtsFlags.ParFlags.maxThreads threads are created */
2320 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads) {
2322 debugBelch("{createThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)\n",
2323 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
2324 return END_TSO_QUEUE;
2330 ASSERT(!RtsFlags.GranFlags.Light || CurrentProc==0);
2333 // ToDo: check whether size = stack_size - TSO_STRUCT_SIZEW
2335 /* catch ridiculously small stack sizes */
2336 if (size < MIN_STACK_WORDS + TSO_STRUCT_SIZEW) {
2337 size = MIN_STACK_WORDS + TSO_STRUCT_SIZEW;
2340 stack_size = size - TSO_STRUCT_SIZEW;
2342 tso = (StgTSO *)allocate(size);
2343 TICK_ALLOC_TSO(stack_size, 0);
2345 SET_HDR(tso, &stg_TSO_info, CCS_SYSTEM);
2347 SET_GRAN_HDR(tso, ThisPE);
2350 // Always start with the compiled code evaluator
2351 tso->what_next = ThreadRunGHC;
2353 tso->id = next_thread_id++;
2354 tso->why_blocked = NotBlocked;
2355 tso->blocked_exceptions = NULL;
2357 tso->saved_errno = 0;
2360 tso->stack_size = stack_size;
2361 tso->max_stack_size = round_to_mblocks(RtsFlags.GcFlags.maxStkSize)
2363 tso->sp = (P_)&(tso->stack) + stack_size;
2365 tso->trec = NO_TREC;
2368 tso->prof.CCCS = CCS_MAIN;
2371 /* put a stop frame on the stack */
2372 tso->sp -= sizeofW(StgStopFrame);
2373 SET_HDR((StgClosure*)tso->sp,(StgInfoTable *)&stg_stop_thread_info,CCS_SYSTEM);
2374 tso->link = END_TSO_QUEUE;
2378 /* uses more flexible routine in GranSim */
2379 insertThread(tso, CurrentProc);
2381 /* In a non-GranSim setup the pushing of a TSO onto the runq is separated
2387 if (RtsFlags.GranFlags.GranSimStats.Full)
2388 DumpGranEvent(GR_START,tso);
2389 #elif defined(PARALLEL_HASKELL)
2390 if (RtsFlags.ParFlags.ParStats.Full)
2391 DumpGranEvent(GR_STARTQ,tso);
2392 /* HACk to avoid SCHEDULE
2396 /* Link the new thread on the global thread list.
2398 tso->global_link = all_threads;
2402 tso->dist.priority = MandatoryPriority; //by default that is...
2406 tso->gran.pri = pri;
2408 tso->gran.magic = TSO_MAGIC; // debugging only
2410 tso->gran.sparkname = 0;
2411 tso->gran.startedat = CURRENT_TIME;
2412 tso->gran.exported = 0;
2413 tso->gran.basicblocks = 0;
2414 tso->gran.allocs = 0;
2415 tso->gran.exectime = 0;
2416 tso->gran.fetchtime = 0;
2417 tso->gran.fetchcount = 0;
2418 tso->gran.blocktime = 0;
2419 tso->gran.blockcount = 0;
2420 tso->gran.blockedat = 0;
2421 tso->gran.globalsparks = 0;
2422 tso->gran.localsparks = 0;
2423 if (RtsFlags.GranFlags.Light)
2424 tso->gran.clock = Now; /* local clock */
2426 tso->gran.clock = 0;
2428 IF_DEBUG(gran,printTSO(tso));
2429 #elif defined(PARALLEL_HASKELL)
2431 tso->par.magic = TSO_MAGIC; // debugging only
2433 tso->par.sparkname = 0;
2434 tso->par.startedat = CURRENT_TIME;
2435 tso->par.exported = 0;
2436 tso->par.basicblocks = 0;
2437 tso->par.allocs = 0;
2438 tso->par.exectime = 0;
2439 tso->par.fetchtime = 0;
2440 tso->par.fetchcount = 0;
2441 tso->par.blocktime = 0;
2442 tso->par.blockcount = 0;
2443 tso->par.blockedat = 0;
2444 tso->par.globalsparks = 0;
2445 tso->par.localsparks = 0;
2449 globalGranStats.tot_threads_created++;
2450 globalGranStats.threads_created_on_PE[CurrentProc]++;
2451 globalGranStats.tot_sq_len += spark_queue_len(CurrentProc);
2452 globalGranStats.tot_sq_probes++;
2453 #elif defined(PARALLEL_HASKELL)
2454 // collect parallel global statistics (currently done together with GC stats)
2455 if (RtsFlags.ParFlags.ParStats.Global &&
2456 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
2457 //debugBelch("Creating thread %d @ %11.2f\n", tso->id, usertime());
2458 globalParStats.tot_threads_created++;
2464 sched_belch("==__ schedule: Created TSO %d (%p);",
2465 CurrentProc, tso, tso->id));
2466 #elif defined(PARALLEL_HASKELL)
2467 IF_PAR_DEBUG(verbose,
2468 sched_belch("==__ schedule: Created TSO %d (%p); %d threads active",
2469 (long)tso->id, tso, advisory_thread_count));
2471 IF_DEBUG(scheduler,sched_belch("created thread %ld, stack size = %lx words",
2472 (long)tso->id, (long)tso->stack_size));
2479 all parallel thread creation calls should fall through the following routine.
2482 createThreadFromSpark(rtsSpark spark)
2484 ASSERT(spark != (rtsSpark)NULL);
2485 // JB: TAKE CARE OF THIS COUNTER! BUGGY
2486 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads)
2488 barf("{createSparkThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
2489 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
2490 return END_TSO_QUEUE;
2494 tso = createThread(RtsFlags.GcFlags.initialStkSize);
2495 if (tso==END_TSO_QUEUE)
2496 barf("createSparkThread: Cannot create TSO");
2498 tso->priority = AdvisoryPriority;
2500 pushClosure(tso,spark);
2502 advisory_thread_count++; // JB: TAKE CARE OF THIS COUNTER! BUGGY
2509 Turn a spark into a thread.
2510 ToDo: fix for SMP (needs to acquire SCHED_MUTEX!)
2514 activateSpark (rtsSpark spark)
2518 tso = createSparkThread(spark);
2519 if (RtsFlags.ParFlags.ParStats.Full) {
2520 //ASSERT(run_queue_hd == END_TSO_QUEUE); // I think ...
2521 IF_PAR_DEBUG(verbose,
2522 debugBelch("==^^ activateSpark: turning spark of closure %p (%s) into a thread\n",
2523 (StgClosure *)spark, info_type((StgClosure *)spark)));
2525 // ToDo: fwd info on local/global spark to thread -- HWL
2526 // tso->gran.exported = spark->exported;
2527 // tso->gran.locked = !spark->global;
2528 // tso->gran.sparkname = spark->name;
2534 /* ---------------------------------------------------------------------------
2537 * scheduleThread puts a thread on the head of the runnable queue.
2538 * This will usually be done immediately after a thread is created.
2539 * The caller of scheduleThread must create the thread using e.g.
2540 * createThread and push an appropriate closure
2541 * on this thread's stack before the scheduler is invoked.
2542 * ------------------------------------------------------------------------ */
2545 scheduleThreadLocked(StgTSO *tso)
2547 // The thread goes at the *end* of the run-queue, to avoid possible
2548 // starvation of any threads already on the queue.
2549 APPEND_TO_RUN_QUEUE(tso);
2554 scheduleThread(StgTSO* tso)
2556 ACQUIRE_LOCK(&sched_mutex);
2557 scheduleThreadLocked(tso);
2558 RELEASE_LOCK(&sched_mutex);
2561 #if defined(RTS_SUPPORTS_THREADS)
2562 static Condition bound_cond_cache;
2563 static int bound_cond_cache_full = 0;
2568 scheduleWaitThread(StgTSO* tso, /*[out]*/HaskellObj* ret,
2569 Capability *initialCapability)
2571 // Precondition: sched_mutex must be held
2574 m = stgMallocBytes(sizeof(StgMainThread), "waitThread");
2579 m->link = main_threads;
2581 if (main_threads != NULL) {
2582 main_threads->prev = m;
2586 #if defined(RTS_SUPPORTS_THREADS)
2587 // Allocating a new condition for each thread is expensive, so we
2588 // cache one. This is a pretty feeble hack, but it helps speed up
2589 // consecutive call-ins quite a bit.
2590 if (bound_cond_cache_full) {
2591 m->bound_thread_cond = bound_cond_cache;
2592 bound_cond_cache_full = 0;
2594 initCondition(&m->bound_thread_cond);
2598 /* Put the thread on the main-threads list prior to scheduling the TSO.
2599 Failure to do so introduces a race condition in the MT case (as
2600 identified by Wolfgang Thaller), whereby the new task/OS thread
2601 created by scheduleThread_() would complete prior to the thread
2602 that spawned it managed to put 'itself' on the main-threads list.
2603 The upshot of it all being that the worker thread wouldn't get to
2604 signal the completion of the its work item for the main thread to
2605 see (==> it got stuck waiting.) -- sof 6/02.
2607 IF_DEBUG(scheduler, sched_belch("waiting for thread (%d)", tso->id));
2609 APPEND_TO_RUN_QUEUE(tso);
2610 // NB. Don't call threadRunnable() here, because the thread is
2611 // bound and only runnable by *this* OS thread, so waking up other
2612 // workers will just slow things down.
2614 return waitThread_(m, initialCapability);
2617 /* ---------------------------------------------------------------------------
2620 * Initialise the scheduler. This resets all the queues - if the
2621 * queues contained any threads, they'll be garbage collected at the
2624 * ------------------------------------------------------------------------ */
2632 for (i=0; i<=MAX_PROC; i++) {
2633 run_queue_hds[i] = END_TSO_QUEUE;
2634 run_queue_tls[i] = END_TSO_QUEUE;
2635 blocked_queue_hds[i] = END_TSO_QUEUE;
2636 blocked_queue_tls[i] = END_TSO_QUEUE;
2637 ccalling_threadss[i] = END_TSO_QUEUE;
2638 blackhole_queue[i] = END_TSO_QUEUE;
2639 sleeping_queue = END_TSO_QUEUE;
2642 run_queue_hd = END_TSO_QUEUE;
2643 run_queue_tl = END_TSO_QUEUE;
2644 blocked_queue_hd = END_TSO_QUEUE;
2645 blocked_queue_tl = END_TSO_QUEUE;
2646 blackhole_queue = END_TSO_QUEUE;
2647 sleeping_queue = END_TSO_QUEUE;
2650 suspended_ccalling_threads = END_TSO_QUEUE;
2652 main_threads = NULL;
2653 all_threads = END_TSO_QUEUE;
2658 RtsFlags.ConcFlags.ctxtSwitchTicks =
2659 RtsFlags.ConcFlags.ctxtSwitchTime / TICK_MILLISECS;
2661 #if defined(RTS_SUPPORTS_THREADS)
2662 /* Initialise the mutex and condition variables used by
2664 initMutex(&sched_mutex);
2665 initMutex(&term_mutex);
2668 ACQUIRE_LOCK(&sched_mutex);
2670 /* A capability holds the state a native thread needs in
2671 * order to execute STG code. At least one capability is
2672 * floating around (only SMP builds have more than one).
2676 #if defined(RTS_SUPPORTS_THREADS)
2681 /* eagerly start some extra workers */
2682 startingWorkerThread = RtsFlags.ParFlags.nNodes;
2683 startTasks(RtsFlags.ParFlags.nNodes, taskStart);
2686 #if /* defined(SMP) ||*/ defined(PARALLEL_HASKELL)
2690 RELEASE_LOCK(&sched_mutex);
2694 exitScheduler( void )
2696 interrupted = rtsTrue;
2697 shutting_down_scheduler = rtsTrue;
2698 #if defined(RTS_SUPPORTS_THREADS)
2699 if (threadIsTask(osThreadId())) { taskStop(); }
2704 /* ----------------------------------------------------------------------------
2705 Managing the per-task allocation areas.
2707 Each capability comes with an allocation area. These are
2708 fixed-length block lists into which allocation can be done.
2710 ToDo: no support for two-space collection at the moment???
2711 ------------------------------------------------------------------------- */
2713 static SchedulerStatus
2714 waitThread_(StgMainThread* m, Capability *initialCapability)
2716 SchedulerStatus stat;
2718 // Precondition: sched_mutex must be held.
2719 IF_DEBUG(scheduler, sched_belch("new main thread (%d)", m->tso->id));
2722 /* GranSim specific init */
2723 CurrentTSO = m->tso; // the TSO to run
2724 procStatus[MainProc] = Busy; // status of main PE
2725 CurrentProc = MainProc; // PE to run it on
2726 schedule(m,initialCapability);
2728 schedule(m,initialCapability);
2729 ASSERT(m->stat != NoStatus);
2734 #if defined(RTS_SUPPORTS_THREADS)
2735 // Free the condition variable, returning it to the cache if possible.
2736 if (!bound_cond_cache_full) {
2737 bound_cond_cache = m->bound_thread_cond;
2738 bound_cond_cache_full = 1;
2740 closeCondition(&m->bound_thread_cond);
2744 IF_DEBUG(scheduler, sched_belch("main thread (%d) finished", m->tso->id));
2747 // Postcondition: sched_mutex still held
2751 /* ---------------------------------------------------------------------------
2752 Where are the roots that we know about?
2754 - all the threads on the runnable queue
2755 - all the threads on the blocked queue
2756 - all the threads on the sleeping queue
2757 - all the thread currently executing a _ccall_GC
2758 - all the "main threads"
2760 ------------------------------------------------------------------------ */
2762 /* This has to be protected either by the scheduler monitor, or by the
2763 garbage collection monitor (probably the latter).
2768 GetRoots( evac_fn evac )
2773 for (i=0; i<=RtsFlags.GranFlags.proc; i++) {
2774 if ((run_queue_hds[i] != END_TSO_QUEUE) && ((run_queue_hds[i] != NULL)))
2775 evac((StgClosure **)&run_queue_hds[i]);
2776 if ((run_queue_tls[i] != END_TSO_QUEUE) && ((run_queue_tls[i] != NULL)))
2777 evac((StgClosure **)&run_queue_tls[i]);
2779 if ((blocked_queue_hds[i] != END_TSO_QUEUE) && ((blocked_queue_hds[i] != NULL)))
2780 evac((StgClosure **)&blocked_queue_hds[i]);
2781 if ((blocked_queue_tls[i] != END_TSO_QUEUE) && ((blocked_queue_tls[i] != NULL)))
2782 evac((StgClosure **)&blocked_queue_tls[i]);
2783 if ((ccalling_threadss[i] != END_TSO_QUEUE) && ((ccalling_threadss[i] != NULL)))
2784 evac((StgClosure **)&ccalling_threads[i]);
2791 if (run_queue_hd != END_TSO_QUEUE) {
2792 ASSERT(run_queue_tl != END_TSO_QUEUE);
2793 evac((StgClosure **)&run_queue_hd);
2794 evac((StgClosure **)&run_queue_tl);
2797 if (blocked_queue_hd != END_TSO_QUEUE) {
2798 ASSERT(blocked_queue_tl != END_TSO_QUEUE);
2799 evac((StgClosure **)&blocked_queue_hd);
2800 evac((StgClosure **)&blocked_queue_tl);
2803 if (sleeping_queue != END_TSO_QUEUE) {
2804 evac((StgClosure **)&sleeping_queue);
2808 if (blackhole_queue != END_TSO_QUEUE) {
2809 evac((StgClosure **)&blackhole_queue);
2812 if (suspended_ccalling_threads != END_TSO_QUEUE) {
2813 evac((StgClosure **)&suspended_ccalling_threads);
2816 #if defined(PARALLEL_HASKELL) || defined(GRAN)
2817 markSparkQueue(evac);
2820 #if defined(RTS_USER_SIGNALS)
2821 // mark the signal handlers (signals should be already blocked)
2822 markSignalHandlers(evac);
2826 /* -----------------------------------------------------------------------------
2829 This is the interface to the garbage collector from Haskell land.
2830 We provide this so that external C code can allocate and garbage
2831 collect when called from Haskell via _ccall_GC.
2833 It might be useful to provide an interface whereby the programmer
2834 can specify more roots (ToDo).
2836 This needs to be protected by the GC condition variable above. KH.
2837 -------------------------------------------------------------------------- */
2839 static void (*extra_roots)(evac_fn);
2844 /* Obligated to hold this lock upon entry */
2845 ACQUIRE_LOCK(&sched_mutex);
2846 GarbageCollect(GetRoots,rtsFalse);
2847 RELEASE_LOCK(&sched_mutex);
2851 performMajorGC(void)
2853 ACQUIRE_LOCK(&sched_mutex);
2854 GarbageCollect(GetRoots,rtsTrue);
2855 RELEASE_LOCK(&sched_mutex);
2859 AllRoots(evac_fn evac)
2861 GetRoots(evac); // the scheduler's roots
2862 extra_roots(evac); // the user's roots
2866 performGCWithRoots(void (*get_roots)(evac_fn))
2868 ACQUIRE_LOCK(&sched_mutex);
2869 extra_roots = get_roots;
2870 GarbageCollect(AllRoots,rtsFalse);
2871 RELEASE_LOCK(&sched_mutex);
2874 /* -----------------------------------------------------------------------------
2877 If the thread has reached its maximum stack size, then raise the
2878 StackOverflow exception in the offending thread. Otherwise
2879 relocate the TSO into a larger chunk of memory and adjust its stack
2881 -------------------------------------------------------------------------- */
2884 threadStackOverflow(StgTSO *tso)
2886 nat new_stack_size, stack_words;
2891 IF_DEBUG(sanity,checkTSO(tso));
2892 if (tso->stack_size >= tso->max_stack_size) {
2895 debugBelch("@@ threadStackOverflow of TSO %ld (%p): stack too large (now %ld; max is %ld)\n",
2896 (long)tso->id, tso, (long)tso->stack_size, (long)tso->max_stack_size);
2897 /* If we're debugging, just print out the top of the stack */
2898 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2901 /* Send this thread the StackOverflow exception */
2902 raiseAsync(tso, (StgClosure *)stackOverflow_closure);
2906 /* Try to double the current stack size. If that takes us over the
2907 * maximum stack size for this thread, then use the maximum instead.
2908 * Finally round up so the TSO ends up as a whole number of blocks.
2910 new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
2911 new_tso_size = (lnat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) +
2912 TSO_STRUCT_SIZE)/sizeof(W_);
2913 new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */
2914 new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
2916 IF_DEBUG(scheduler, debugBelch("== sched: increasing stack size from %d words to %d.\n", tso->stack_size, new_stack_size));
2918 dest = (StgTSO *)allocate(new_tso_size);
2919 TICK_ALLOC_TSO(new_stack_size,0);
2921 /* copy the TSO block and the old stack into the new area */
2922 memcpy(dest,tso,TSO_STRUCT_SIZE);
2923 stack_words = tso->stack + tso->stack_size - tso->sp;
2924 new_sp = (P_)dest + new_tso_size - stack_words;
2925 memcpy(new_sp, tso->sp, stack_words * sizeof(W_));
2927 /* relocate the stack pointers... */
2929 dest->stack_size = new_stack_size;
2931 /* Mark the old TSO as relocated. We have to check for relocated
2932 * TSOs in the garbage collector and any primops that deal with TSOs.
2934 * It's important to set the sp value to just beyond the end
2935 * of the stack, so we don't attempt to scavenge any part of the
2938 tso->what_next = ThreadRelocated;
2940 tso->sp = (P_)&(tso->stack[tso->stack_size]);
2941 tso->why_blocked = NotBlocked;
2943 IF_PAR_DEBUG(verbose,
2944 debugBelch("@@ threadStackOverflow of TSO %d (now at %p): stack size increased to %ld\n",
2945 tso->id, tso, tso->stack_size);
2946 /* If we're debugging, just print out the top of the stack */
2947 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2950 IF_DEBUG(sanity,checkTSO(tso));
2952 IF_DEBUG(scheduler,printTSO(dest));
2958 /* ---------------------------------------------------------------------------
2959 Wake up a queue that was blocked on some resource.
2960 ------------------------------------------------------------------------ */
2964 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2967 #elif defined(PARALLEL_HASKELL)
2969 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2971 /* write RESUME events to log file and
2972 update blocked and fetch time (depending on type of the orig closure) */
2973 if (RtsFlags.ParFlags.ParStats.Full) {
2974 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
2975 GR_RESUMEQ, ((StgTSO *)bqe), ((StgTSO *)bqe)->block_info.closure,
2976 0, 0 /* spark_queue_len(ADVISORY_POOL) */);
2977 if (EMPTY_RUN_QUEUE())
2978 emitSchedule = rtsTrue;
2980 switch (get_itbl(node)->type) {
2982 ((StgTSO *)bqe)->par.fetchtime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2987 ((StgTSO *)bqe)->par.blocktime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2994 barf("{unblockOneLocked}Daq Qagh: unexpected closure in blocking queue");
3001 StgBlockingQueueElement *
3002 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
3005 PEs node_loc, tso_loc;
3007 node_loc = where_is(node); // should be lifted out of loop
3008 tso = (StgTSO *)bqe; // wastes an assignment to get the type right
3009 tso_loc = where_is((StgClosure *)tso);
3010 if (IS_LOCAL_TO(PROCS(node),tso_loc)) { // TSO is local
3011 /* !fake_fetch => TSO is on CurrentProc is same as IS_LOCAL_TO */
3012 ASSERT(CurrentProc!=node_loc || tso_loc==CurrentProc);
3013 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.lunblocktime;
3014 // insertThread(tso, node_loc);
3015 new_event(tso_loc, tso_loc, CurrentTime[CurrentProc],
3017 tso, node, (rtsSpark*)NULL);
3018 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
3021 } else { // TSO is remote (actually should be FMBQ)
3022 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.mpacktime +
3023 RtsFlags.GranFlags.Costs.gunblocktime +
3024 RtsFlags.GranFlags.Costs.latency;
3025 new_event(tso_loc, CurrentProc, CurrentTime[CurrentProc],
3027 tso, node, (rtsSpark*)NULL);
3028 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
3031 /* the thread-queue-overhead is accounted for in either Resume or UnblockThread */
3033 debugBelch(" %s TSO %d (%p) [PE %d] (block_info.closure=%p) (next=%p) ,",
3034 (node_loc==tso_loc ? "Local" : "Global"),
3035 tso->id, tso, CurrentProc, tso->block_info.closure, tso->link));
3036 tso->block_info.closure = NULL;
3037 IF_DEBUG(scheduler,debugBelch("-- Waking up thread %ld (%p)\n",
3040 #elif defined(PARALLEL_HASKELL)
3041 StgBlockingQueueElement *
3042 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
3044 StgBlockingQueueElement *next;
3046 switch (get_itbl(bqe)->type) {
3048 ASSERT(((StgTSO *)bqe)->why_blocked != NotBlocked);
3049 /* if it's a TSO just push it onto the run_queue */
3051 ((StgTSO *)bqe)->link = END_TSO_QUEUE; // debugging?
3052 APPEND_TO_RUN_QUEUE((StgTSO *)bqe);
3054 unblockCount(bqe, node);
3055 /* reset blocking status after dumping event */
3056 ((StgTSO *)bqe)->why_blocked = NotBlocked;
3060 /* if it's a BLOCKED_FETCH put it on the PendingFetches list */
3062 bqe->link = (StgBlockingQueueElement *)PendingFetches;
3063 PendingFetches = (StgBlockedFetch *)bqe;
3067 /* can ignore this case in a non-debugging setup;
3068 see comments on RBHSave closures above */
3070 /* check that the closure is an RBHSave closure */
3071 ASSERT(get_itbl((StgClosure *)bqe) == &stg_RBH_Save_0_info ||
3072 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_1_info ||
3073 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_2_info);
3077 barf("{unblockOneLocked}Daq Qagh: Unexpected IP (%#lx; %s) in blocking queue at %#lx\n",
3078 get_itbl((StgClosure *)bqe), info_type((StgClosure *)bqe),
3082 IF_PAR_DEBUG(bq, debugBelch(", %p (%s)\n", bqe, info_type((StgClosure*)bqe)));
3086 #else /* !GRAN && !PARALLEL_HASKELL */
3088 unblockOneLocked(StgTSO *tso)
3092 ASSERT(get_itbl(tso)->type == TSO);
3093 ASSERT(tso->why_blocked != NotBlocked);
3094 tso->why_blocked = NotBlocked;
3096 tso->link = END_TSO_QUEUE;
3097 APPEND_TO_RUN_QUEUE(tso);
3099 IF_DEBUG(scheduler,sched_belch("waking up thread %ld", (long)tso->id));
3104 #if defined(GRAN) || defined(PARALLEL_HASKELL)
3105 INLINE_ME StgBlockingQueueElement *
3106 unblockOne(StgBlockingQueueElement *bqe, StgClosure *node)
3108 ACQUIRE_LOCK(&sched_mutex);
3109 bqe = unblockOneLocked(bqe, node);
3110 RELEASE_LOCK(&sched_mutex);
3115 unblockOne(StgTSO *tso)
3117 ACQUIRE_LOCK(&sched_mutex);
3118 tso = unblockOneLocked(tso);
3119 RELEASE_LOCK(&sched_mutex);
3126 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
3128 StgBlockingQueueElement *bqe;
3133 debugBelch("##-_ AwBQ for node %p on PE %d @ %ld by TSO %d (%p): \n", \
3134 node, CurrentProc, CurrentTime[CurrentProc],
3135 CurrentTSO->id, CurrentTSO));
3137 node_loc = where_is(node);
3139 ASSERT(q == END_BQ_QUEUE ||
3140 get_itbl(q)->type == TSO || // q is either a TSO or an RBHSave
3141 get_itbl(q)->type == CONSTR); // closure (type constructor)
3142 ASSERT(is_unique(node));
3144 /* FAKE FETCH: magically copy the node to the tso's proc;
3145 no Fetch necessary because in reality the node should not have been
3146 moved to the other PE in the first place
3148 if (CurrentProc!=node_loc) {
3150 debugBelch("## node %p is on PE %d but CurrentProc is %d (TSO %d); assuming fake fetch and adjusting bitmask (old: %#x)\n",
3151 node, node_loc, CurrentProc, CurrentTSO->id,
3152 // CurrentTSO, where_is(CurrentTSO),
3153 node->header.gran.procs));
3154 node->header.gran.procs = (node->header.gran.procs) | PE_NUMBER(CurrentProc);
3156 debugBelch("## new bitmask of node %p is %#x\n",
3157 node, node->header.gran.procs));
3158 if (RtsFlags.GranFlags.GranSimStats.Global) {
3159 globalGranStats.tot_fake_fetches++;
3164 // ToDo: check: ASSERT(CurrentProc==node_loc);
3165 while (get_itbl(bqe)->type==TSO) { // q != END_TSO_QUEUE) {
3168 bqe points to the current element in the queue
3169 next points to the next element in the queue
3171 //tso = (StgTSO *)bqe; // wastes an assignment to get the type right
3172 //tso_loc = where_is(tso);
3174 bqe = unblockOneLocked(bqe, node);
3177 /* if this is the BQ of an RBH, we have to put back the info ripped out of
3178 the closure to make room for the anchor of the BQ */
3179 if (bqe!=END_BQ_QUEUE) {
3180 ASSERT(get_itbl(node)->type == RBH && get_itbl(bqe)->type == CONSTR);
3182 ASSERT((info_ptr==&RBH_Save_0_info) ||
3183 (info_ptr==&RBH_Save_1_info) ||
3184 (info_ptr==&RBH_Save_2_info));
3186 /* cf. convertToRBH in RBH.c for writing the RBHSave closure */
3187 ((StgRBH *)node)->blocking_queue = (StgBlockingQueueElement *)((StgRBHSave *)bqe)->payload[0];
3188 ((StgRBH *)node)->mut_link = (StgMutClosure *)((StgRBHSave *)bqe)->payload[1];
3191 debugBelch("## Filled in RBH_Save for %p (%s) at end of AwBQ\n",
3192 node, info_type(node)));
3195 /* statistics gathering */
3196 if (RtsFlags.GranFlags.GranSimStats.Global) {
3197 // globalGranStats.tot_bq_processing_time += bq_processing_time;
3198 globalGranStats.tot_bq_len += len; // total length of all bqs awakened
3199 // globalGranStats.tot_bq_len_local += len_local; // same for local TSOs only
3200 globalGranStats.tot_awbq++; // total no. of bqs awakened
3203 debugBelch("## BQ Stats of %p: [%d entries] %s\n",
3204 node, len, (bqe!=END_BQ_QUEUE) ? "RBH" : ""));
3206 #elif defined(PARALLEL_HASKELL)
3208 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
3210 StgBlockingQueueElement *bqe;
3212 ACQUIRE_LOCK(&sched_mutex);
3214 IF_PAR_DEBUG(verbose,
3215 debugBelch("##-_ AwBQ for node %p on [%x]: \n",
3219 if(get_itbl(q)->type == CONSTR || q==END_BQ_QUEUE) {
3220 IF_PAR_DEBUG(verbose, debugBelch("## ... nothing to unblock so lets just return. RFP (BUG?)\n"));
3225 ASSERT(q == END_BQ_QUEUE ||
3226 get_itbl(q)->type == TSO ||
3227 get_itbl(q)->type == BLOCKED_FETCH ||
3228 get_itbl(q)->type == CONSTR);
3231 while (get_itbl(bqe)->type==TSO ||
3232 get_itbl(bqe)->type==BLOCKED_FETCH) {
3233 bqe = unblockOneLocked(bqe, node);
3235 RELEASE_LOCK(&sched_mutex);
3238 #else /* !GRAN && !PARALLEL_HASKELL */
3241 awakenBlockedQueueNoLock(StgTSO *tso)
3243 if (tso == NULL) return; // hack; see bug #1235728, and comments in
3245 while (tso != END_TSO_QUEUE) {
3246 tso = unblockOneLocked(tso);
3251 awakenBlockedQueue(StgTSO *tso)
3253 if (tso == NULL) return; // hack; see bug #1235728, and comments in
3255 ACQUIRE_LOCK(&sched_mutex);
3256 while (tso != END_TSO_QUEUE) {
3257 tso = unblockOneLocked(tso);
3259 RELEASE_LOCK(&sched_mutex);
3263 /* ---------------------------------------------------------------------------
3265 - usually called inside a signal handler so it mustn't do anything fancy.
3266 ------------------------------------------------------------------------ */
3269 interruptStgRts(void)
3274 /* ToDo: if invoked from a signal handler, this threadRunnable
3275 * only works if there's another thread (not this one) waiting to
3280 /* -----------------------------------------------------------------------------
3283 This is for use when we raise an exception in another thread, which
3285 This has nothing to do with the UnblockThread event in GranSim. -- HWL
3286 -------------------------------------------------------------------------- */
3288 #if defined(GRAN) || defined(PARALLEL_HASKELL)
3290 NB: only the type of the blocking queue is different in GranSim and GUM
3291 the operations on the queue-elements are the same
3292 long live polymorphism!
3294 Locks: sched_mutex is held upon entry and exit.
3298 unblockThread(StgTSO *tso)
3300 StgBlockingQueueElement *t, **last;
3302 switch (tso->why_blocked) {
3305 return; /* not blocked */
3308 // Be careful: nothing to do here! We tell the scheduler that the thread
3309 // is runnable and we leave it to the stack-walking code to abort the
3310 // transaction while unwinding the stack. We should perhaps have a debugging
3311 // test to make sure that this really happens and that the 'zombie' transaction
3312 // does not get committed.
3316 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
3318 StgBlockingQueueElement *last_tso = END_BQ_QUEUE;
3319 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
3321 last = (StgBlockingQueueElement **)&mvar->head;
3322 for (t = (StgBlockingQueueElement *)mvar->head;
3324 last = &t->link, last_tso = t, t = t->link) {
3325 if (t == (StgBlockingQueueElement *)tso) {
3326 *last = (StgBlockingQueueElement *)tso->link;
3327 if (mvar->tail == tso) {
3328 mvar->tail = (StgTSO *)last_tso;
3333 barf("unblockThread (MVAR): TSO not found");
3336 case BlockedOnBlackHole:
3337 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
3339 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
3341 last = &bq->blocking_queue;
3342 for (t = bq->blocking_queue;
3344 last = &t->link, t = t->link) {
3345 if (t == (StgBlockingQueueElement *)tso) {
3346 *last = (StgBlockingQueueElement *)tso->link;
3350 barf("unblockThread (BLACKHOLE): TSO not found");
3353 case BlockedOnException:
3355 StgTSO *target = tso->block_info.tso;
3357 ASSERT(get_itbl(target)->type == TSO);
3359 if (target->what_next == ThreadRelocated) {
3360 target = target->link;
3361 ASSERT(get_itbl(target)->type == TSO);
3364 ASSERT(target->blocked_exceptions != NULL);
3366 last = (StgBlockingQueueElement **)&target->blocked_exceptions;
3367 for (t = (StgBlockingQueueElement *)target->blocked_exceptions;
3369 last = &t->link, t = t->link) {
3370 ASSERT(get_itbl(t)->type == TSO);
3371 if (t == (StgBlockingQueueElement *)tso) {
3372 *last = (StgBlockingQueueElement *)tso->link;
3376 barf("unblockThread (Exception): TSO not found");
3380 case BlockedOnWrite:
3381 #if defined(mingw32_HOST_OS)
3382 case BlockedOnDoProc:
3385 /* take TSO off blocked_queue */
3386 StgBlockingQueueElement *prev = NULL;
3387 for (t = (StgBlockingQueueElement *)blocked_queue_hd; t != END_BQ_QUEUE;
3388 prev = t, t = t->link) {
3389 if (t == (StgBlockingQueueElement *)tso) {
3391 blocked_queue_hd = (StgTSO *)t->link;
3392 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
3393 blocked_queue_tl = END_TSO_QUEUE;
3396 prev->link = t->link;
3397 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
3398 blocked_queue_tl = (StgTSO *)prev;
3401 #if defined(mingw32_HOST_OS)
3402 /* (Cooperatively) signal that the worker thread should abort
3405 abandonWorkRequest(tso->block_info.async_result->reqID);
3410 barf("unblockThread (I/O): TSO not found");
3413 case BlockedOnDelay:
3415 /* take TSO off sleeping_queue */
3416 StgBlockingQueueElement *prev = NULL;
3417 for (t = (StgBlockingQueueElement *)sleeping_queue; t != END_BQ_QUEUE;
3418 prev = t, t = t->link) {
3419 if (t == (StgBlockingQueueElement *)tso) {
3421 sleeping_queue = (StgTSO *)t->link;
3423 prev->link = t->link;
3428 barf("unblockThread (delay): TSO not found");
3432 barf("unblockThread");
3436 tso->link = END_TSO_QUEUE;
3437 tso->why_blocked = NotBlocked;
3438 tso->block_info.closure = NULL;
3439 PUSH_ON_RUN_QUEUE(tso);
3443 unblockThread(StgTSO *tso)
3447 /* To avoid locking unnecessarily. */
3448 if (tso->why_blocked == NotBlocked) {
3452 switch (tso->why_blocked) {
3455 // Be careful: nothing to do here! We tell the scheduler that the thread
3456 // is runnable and we leave it to the stack-walking code to abort the
3457 // transaction while unwinding the stack. We should perhaps have a debugging
3458 // test to make sure that this really happens and that the 'zombie' transaction
3459 // does not get committed.
3463 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
3465 StgTSO *last_tso = END_TSO_QUEUE;
3466 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
3469 for (t = mvar->head; t != END_TSO_QUEUE;
3470 last = &t->link, last_tso = t, t = t->link) {
3473 if (mvar->tail == tso) {
3474 mvar->tail = last_tso;
3479 barf("unblockThread (MVAR): TSO not found");
3482 case BlockedOnBlackHole:
3484 last = &blackhole_queue;
3485 for (t = blackhole_queue; t != END_TSO_QUEUE;
3486 last = &t->link, t = t->link) {
3492 barf("unblockThread (BLACKHOLE): TSO not found");
3495 case BlockedOnException:
3497 StgTSO *target = tso->block_info.tso;
3499 ASSERT(get_itbl(target)->type == TSO);
3501 while (target->what_next == ThreadRelocated) {
3502 target = target->link;
3503 ASSERT(get_itbl(target)->type == TSO);
3506 ASSERT(target->blocked_exceptions != NULL);
3508 last = &target->blocked_exceptions;
3509 for (t = target->blocked_exceptions; t != END_TSO_QUEUE;
3510 last = &t->link, t = t->link) {
3511 ASSERT(get_itbl(t)->type == TSO);
3517 barf("unblockThread (Exception): TSO not found");
3521 case BlockedOnWrite:
3522 #if defined(mingw32_HOST_OS)
3523 case BlockedOnDoProc:
3526 StgTSO *prev = NULL;
3527 for (t = blocked_queue_hd; t != END_TSO_QUEUE;
3528 prev = t, t = t->link) {
3531 blocked_queue_hd = t->link;
3532 if (blocked_queue_tl == t) {
3533 blocked_queue_tl = END_TSO_QUEUE;
3536 prev->link = t->link;
3537 if (blocked_queue_tl == t) {
3538 blocked_queue_tl = prev;
3541 #if defined(mingw32_HOST_OS)
3542 /* (Cooperatively) signal that the worker thread should abort
3545 abandonWorkRequest(tso->block_info.async_result->reqID);
3550 barf("unblockThread (I/O): TSO not found");
3553 case BlockedOnDelay:
3555 StgTSO *prev = NULL;
3556 for (t = sleeping_queue; t != END_TSO_QUEUE;
3557 prev = t, t = t->link) {
3560 sleeping_queue = t->link;
3562 prev->link = t->link;
3567 barf("unblockThread (delay): TSO not found");
3571 barf("unblockThread");
3575 tso->link = END_TSO_QUEUE;
3576 tso->why_blocked = NotBlocked;
3577 tso->block_info.closure = NULL;
3578 APPEND_TO_RUN_QUEUE(tso);
3582 /* -----------------------------------------------------------------------------
3585 * Check the blackhole_queue for threads that can be woken up. We do
3586 * this periodically: before every GC, and whenever the run queue is
3589 * An elegant solution might be to just wake up all the blocked
3590 * threads with awakenBlockedQueue occasionally: they'll go back to
3591 * sleep again if the object is still a BLACKHOLE. Unfortunately this
3592 * doesn't give us a way to tell whether we've actually managed to
3593 * wake up any threads, so we would be busy-waiting.
3595 * -------------------------------------------------------------------------- */
3598 checkBlackHoles( void )
3601 rtsBool any_woke_up = rtsFalse;
3604 IF_DEBUG(scheduler, sched_belch("checking threads blocked on black holes"));
3606 // ASSUMES: sched_mutex
3607 prev = &blackhole_queue;
3608 t = blackhole_queue;
3609 while (t != END_TSO_QUEUE) {
3610 ASSERT(t->why_blocked == BlockedOnBlackHole);
3611 type = get_itbl(t->block_info.closure)->type;
3612 if (type != BLACKHOLE && type != CAF_BLACKHOLE) {
3613 t = unblockOneLocked(t);
3615 any_woke_up = rtsTrue;
3625 /* -----------------------------------------------------------------------------
3628 * The following function implements the magic for raising an
3629 * asynchronous exception in an existing thread.
3631 * We first remove the thread from any queue on which it might be
3632 * blocked. The possible blockages are MVARs and BLACKHOLE_BQs.
3634 * We strip the stack down to the innermost CATCH_FRAME, building
3635 * thunks in the heap for all the active computations, so they can
3636 * be restarted if necessary. When we reach a CATCH_FRAME, we build
3637 * an application of the handler to the exception, and push it on
3638 * the top of the stack.
3640 * How exactly do we save all the active computations? We create an
3641 * AP_STACK for every UpdateFrame on the stack. Entering one of these
3642 * AP_STACKs pushes everything from the corresponding update frame
3643 * upwards onto the stack. (Actually, it pushes everything up to the
3644 * next update frame plus a pointer to the next AP_STACK object.
3645 * Entering the next AP_STACK object pushes more onto the stack until we
3646 * reach the last AP_STACK object - at which point the stack should look
3647 * exactly as it did when we killed the TSO and we can continue
3648 * execution by entering the closure on top of the stack.
3650 * We can also kill a thread entirely - this happens if either (a) the
3651 * exception passed to raiseAsync is NULL, or (b) there's no
3652 * CATCH_FRAME on the stack. In either case, we strip the entire
3653 * stack and replace the thread with a zombie.
3655 * Locks: sched_mutex held upon entry nor exit.
3657 * -------------------------------------------------------------------------- */
3660 deleteThread(StgTSO *tso)
3662 if (tso->why_blocked != BlockedOnCCall &&
3663 tso->why_blocked != BlockedOnCCall_NoUnblockExc) {
3664 raiseAsync(tso,NULL);
3668 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
3670 deleteThreadImmediately(StgTSO *tso)
3671 { // for forkProcess only:
3672 // delete thread without giving it a chance to catch the KillThread exception
3674 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
3678 if (tso->why_blocked != BlockedOnCCall &&
3679 tso->why_blocked != BlockedOnCCall_NoUnblockExc) {
3683 tso->what_next = ThreadKilled;
3688 raiseAsyncWithLock(StgTSO *tso, StgClosure *exception)
3690 /* When raising async exs from contexts where sched_mutex isn't held;
3691 use raiseAsyncWithLock(). */
3692 ACQUIRE_LOCK(&sched_mutex);
3693 raiseAsync(tso,exception);
3694 RELEASE_LOCK(&sched_mutex);
3698 raiseAsync(StgTSO *tso, StgClosure *exception)
3700 raiseAsync_(tso, exception, rtsFalse);
3704 raiseAsync_(StgTSO *tso, StgClosure *exception, rtsBool stop_at_atomically)
3706 StgRetInfoTable *info;
3709 // Thread already dead?
3710 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
3715 sched_belch("raising exception in thread %ld.", (long)tso->id));
3717 // Remove it from any blocking queues
3722 // The stack freezing code assumes there's a closure pointer on
3723 // the top of the stack, so we have to arrange that this is the case...
3725 if (sp[0] == (W_)&stg_enter_info) {
3729 sp[0] = (W_)&stg_dummy_ret_closure;
3735 // 1. Let the top of the stack be the "current closure"
3737 // 2. Walk up the stack until we find either an UPDATE_FRAME or a
3740 // 3. If it's an UPDATE_FRAME, then make an AP_STACK containing the
3741 // current closure applied to the chunk of stack up to (but not
3742 // including) the update frame. This closure becomes the "current
3743 // closure". Go back to step 2.
3745 // 4. If it's a CATCH_FRAME, then leave the exception handler on
3746 // top of the stack applied to the exception.
3748 // 5. If it's a STOP_FRAME, then kill the thread.
3750 // NB: if we pass an ATOMICALLY_FRAME then abort the associated
3757 info = get_ret_itbl((StgClosure *)frame);
3759 while (info->i.type != UPDATE_FRAME
3760 && (info->i.type != CATCH_FRAME || exception == NULL)
3761 && info->i.type != STOP_FRAME
3762 && (info->i.type != ATOMICALLY_FRAME || stop_at_atomically == rtsFalse))
3764 if (info->i.type == CATCH_RETRY_FRAME || info->i.type == ATOMICALLY_FRAME) {
3765 // IF we find an ATOMICALLY_FRAME then we abort the
3766 // current transaction and propagate the exception. In
3767 // this case (unlike ordinary exceptions) we do not care
3768 // whether the transaction is valid or not because its
3769 // possible validity cannot have caused the exception
3770 // and will not be visible after the abort.
3772 debugBelch("Found atomically block delivering async exception\n"));
3773 stmAbortTransaction(tso -> trec);
3774 tso -> trec = stmGetEnclosingTRec(tso -> trec);
3776 frame += stack_frame_sizeW((StgClosure *)frame);
3777 info = get_ret_itbl((StgClosure *)frame);
3780 switch (info->i.type) {
3782 case ATOMICALLY_FRAME:
3783 ASSERT(stop_at_atomically);
3784 ASSERT(stmGetEnclosingTRec(tso->trec) == NO_TREC);
3785 stmCondemnTransaction(tso -> trec);
3789 // R1 is not a register: the return convention for IO in
3790 // this case puts the return value on the stack, so we
3791 // need to set up the stack to return to the atomically
3792 // frame properly...
3793 tso->sp = frame - 2;
3794 tso->sp[1] = (StgWord) &stg_NO_FINALIZER_closure; // why not?
3795 tso->sp[0] = (StgWord) &stg_ut_1_0_unreg_info;
3797 tso->what_next = ThreadRunGHC;
3801 // If we find a CATCH_FRAME, and we've got an exception to raise,
3802 // then build the THUNK raise(exception), and leave it on
3803 // top of the CATCH_FRAME ready to enter.
3807 StgCatchFrame *cf = (StgCatchFrame *)frame;
3811 // we've got an exception to raise, so let's pass it to the
3812 // handler in this frame.
3814 raise = (StgThunk *)allocate(sizeofW(StgThunk)+MIN_UPD_SIZE);
3815 TICK_ALLOC_SE_THK(1,0);
3816 SET_HDR(raise,&stg_raise_info,cf->header.prof.ccs);
3817 raise->payload[0] = exception;
3819 // throw away the stack from Sp up to the CATCH_FRAME.
3823 /* Ensure that async excpetions are blocked now, so we don't get
3824 * a surprise exception before we get around to executing the
3827 if (tso->blocked_exceptions == NULL) {
3828 tso->blocked_exceptions = END_TSO_QUEUE;
3831 /* Put the newly-built THUNK on top of the stack, ready to execute
3832 * when the thread restarts.
3835 sp[-1] = (W_)&stg_enter_info;
3837 tso->what_next = ThreadRunGHC;
3838 IF_DEBUG(sanity, checkTSO(tso));
3847 // First build an AP_STACK consisting of the stack chunk above the
3848 // current update frame, with the top word on the stack as the
3851 words = frame - sp - 1;
3852 ap = (StgAP_STACK *)allocate(AP_STACK_sizeW(words));
3855 ap->fun = (StgClosure *)sp[0];
3857 for(i=0; i < (nat)words; ++i) {
3858 ap->payload[i] = (StgClosure *)*sp++;
3861 SET_HDR(ap,&stg_AP_STACK_info,
3862 ((StgClosure *)frame)->header.prof.ccs /* ToDo */);
3863 TICK_ALLOC_UP_THK(words+1,0);
3866 debugBelch("sched: Updating ");
3867 printPtr((P_)((StgUpdateFrame *)frame)->updatee);
3868 debugBelch(" with ");
3869 printObj((StgClosure *)ap);
3872 // Replace the updatee with an indirection - happily
3873 // this will also wake up any threads currently
3874 // waiting on the result.
3876 // Warning: if we're in a loop, more than one update frame on
3877 // the stack may point to the same object. Be careful not to
3878 // overwrite an IND_OLDGEN in this case, because we'll screw
3879 // up the mutable lists. To be on the safe side, don't
3880 // overwrite any kind of indirection at all. See also
3881 // threadSqueezeStack in GC.c, where we have to make a similar
3884 if (!closure_IND(((StgUpdateFrame *)frame)->updatee)) {
3885 // revert the black hole
3886 UPD_IND_NOLOCK(((StgUpdateFrame *)frame)->updatee,
3889 sp += sizeofW(StgUpdateFrame) - 1;
3890 sp[0] = (W_)ap; // push onto stack
3895 // We've stripped the entire stack, the thread is now dead.
3896 sp += sizeofW(StgStopFrame);
3897 tso->what_next = ThreadKilled;
3908 /* -----------------------------------------------------------------------------
3909 raiseExceptionHelper
3911 This function is called by the raise# primitve, just so that we can
3912 move some of the tricky bits of raising an exception from C-- into
3913 C. Who knows, it might be a useful re-useable thing here too.
3914 -------------------------------------------------------------------------- */
3917 raiseExceptionHelper (StgTSO *tso, StgClosure *exception)
3919 StgThunk *raise_closure = NULL;
3921 StgRetInfoTable *info;
3923 // This closure represents the expression 'raise# E' where E
3924 // is the exception raise. It is used to overwrite all the
3925 // thunks which are currently under evaluataion.
3929 // LDV profiling: stg_raise_info has THUNK as its closure
3930 // type. Since a THUNK takes at least MIN_UPD_SIZE words in its
3931 // payload, MIN_UPD_SIZE is more approprate than 1. It seems that
3932 // 1 does not cause any problem unless profiling is performed.
3933 // However, when LDV profiling goes on, we need to linearly scan
3934 // small object pool, where raise_closure is stored, so we should
3935 // use MIN_UPD_SIZE.
3937 // raise_closure = (StgClosure *)RET_STGCALL1(P_,allocate,
3938 // sizeofW(StgClosure)+1);
3942 // Walk up the stack, looking for the catch frame. On the way,
3943 // we update any closures pointed to from update frames with the
3944 // raise closure that we just built.
3948 info = get_ret_itbl((StgClosure *)p);
3949 next = p + stack_frame_sizeW((StgClosure *)p);
3950 switch (info->i.type) {
3953 // Only create raise_closure if we need to.
3954 if (raise_closure == NULL) {
3956 (StgThunk *)allocate(sizeofW(StgThunk)+MIN_UPD_SIZE);
3957 SET_HDR(raise_closure, &stg_raise_info, CCCS);
3958 raise_closure->payload[0] = exception;
3960 UPD_IND(((StgUpdateFrame *)p)->updatee,(StgClosure *)raise_closure);
3964 case ATOMICALLY_FRAME:
3965 IF_DEBUG(stm, debugBelch("Found ATOMICALLY_FRAME at %p\n", p));
3967 return ATOMICALLY_FRAME;
3973 case CATCH_STM_FRAME:
3974 IF_DEBUG(stm, debugBelch("Found CATCH_STM_FRAME at %p\n", p));
3976 return CATCH_STM_FRAME;
3982 case CATCH_RETRY_FRAME:
3991 /* -----------------------------------------------------------------------------
3992 findRetryFrameHelper
3994 This function is called by the retry# primitive. It traverses the stack
3995 leaving tso->sp referring to the frame which should handle the retry.
3997 This should either be a CATCH_RETRY_FRAME (if the retry# is within an orElse#)
3998 or should be a ATOMICALLY_FRAME (if the retry# reaches the top level).
4000 We skip CATCH_STM_FRAMEs because retries are not considered to be exceptions,
4001 despite the similar implementation.
4003 We should not expect to see CATCH_FRAME or STOP_FRAME because those should
4004 not be created within memory transactions.
4005 -------------------------------------------------------------------------- */
4008 findRetryFrameHelper (StgTSO *tso)
4011 StgRetInfoTable *info;
4015 info = get_ret_itbl((StgClosure *)p);
4016 next = p + stack_frame_sizeW((StgClosure *)p);
4017 switch (info->i.type) {
4019 case ATOMICALLY_FRAME:
4020 IF_DEBUG(stm, debugBelch("Found ATOMICALLY_FRAME at %p during retrry\n", p));
4022 return ATOMICALLY_FRAME;
4024 case CATCH_RETRY_FRAME:
4025 IF_DEBUG(stm, debugBelch("Found CATCH_RETRY_FRAME at %p during retrry\n", p));
4027 return CATCH_RETRY_FRAME;
4029 case CATCH_STM_FRAME:
4031 ASSERT(info->i.type != CATCH_FRAME);
4032 ASSERT(info->i.type != STOP_FRAME);
4039 /* -----------------------------------------------------------------------------
4040 resurrectThreads is called after garbage collection on the list of
4041 threads found to be garbage. Each of these threads will be woken
4042 up and sent a signal: BlockedOnDeadMVar if the thread was blocked
4043 on an MVar, or NonTermination if the thread was blocked on a Black
4046 Locks: sched_mutex isn't held upon entry nor exit.
4047 -------------------------------------------------------------------------- */
4050 resurrectThreads( StgTSO *threads )
4054 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
4055 next = tso->global_link;
4056 tso->global_link = all_threads;
4058 IF_DEBUG(scheduler, sched_belch("resurrecting thread %d", tso->id));
4060 switch (tso->why_blocked) {
4062 case BlockedOnException:
4063 /* Called by GC - sched_mutex lock is currently held. */
4064 raiseAsync(tso,(StgClosure *)BlockedOnDeadMVar_closure);
4066 case BlockedOnBlackHole:
4067 raiseAsync(tso,(StgClosure *)NonTermination_closure);
4070 raiseAsync(tso,(StgClosure *)BlockedIndefinitely_closure);
4073 /* This might happen if the thread was blocked on a black hole
4074 * belonging to a thread that we've just woken up (raiseAsync
4075 * can wake up threads, remember...).
4079 barf("resurrectThreads: thread blocked in a strange way");
4084 /* ----------------------------------------------------------------------------
4085 * Debugging: why is a thread blocked
4086 * [Also provides useful information when debugging threaded programs
4087 * at the Haskell source code level, so enable outside of DEBUG. --sof 7/02]
4088 ------------------------------------------------------------------------- */
4091 printThreadBlockage(StgTSO *tso)
4093 switch (tso->why_blocked) {
4095 debugBelch("is blocked on read from fd %d", (int)(tso->block_info.fd));
4097 case BlockedOnWrite:
4098 debugBelch("is blocked on write to fd %d", (int)(tso->block_info.fd));
4100 #if defined(mingw32_HOST_OS)
4101 case BlockedOnDoProc:
4102 debugBelch("is blocked on proc (request: %ld)", tso->block_info.async_result->reqID);
4105 case BlockedOnDelay:
4106 debugBelch("is blocked until %ld", (long)(tso->block_info.target));
4109 debugBelch("is blocked on an MVar @ %p", tso->block_info.closure);
4111 case BlockedOnException:
4112 debugBelch("is blocked on delivering an exception to thread %d",
4113 tso->block_info.tso->id);
4115 case BlockedOnBlackHole:
4116 debugBelch("is blocked on a black hole");
4119 debugBelch("is not blocked");
4121 #if defined(PARALLEL_HASKELL)
4123 debugBelch("is blocked on global address; local FM_BQ is %p (%s)",
4124 tso->block_info.closure, info_type(tso->block_info.closure));
4126 case BlockedOnGA_NoSend:
4127 debugBelch("is blocked on global address (no send); local FM_BQ is %p (%s)",
4128 tso->block_info.closure, info_type(tso->block_info.closure));
4131 case BlockedOnCCall:
4132 debugBelch("is blocked on an external call");
4134 case BlockedOnCCall_NoUnblockExc:
4135 debugBelch("is blocked on an external call (exceptions were already blocked)");
4138 debugBelch("is blocked on an STM operation");
4141 barf("printThreadBlockage: strange tso->why_blocked: %d for TSO %d (%d)",
4142 tso->why_blocked, tso->id, tso);
4147 printThreadStatus(StgTSO *tso)
4149 switch (tso->what_next) {
4151 debugBelch("has been killed");
4153 case ThreadComplete:
4154 debugBelch("has completed");
4157 printThreadBlockage(tso);
4162 printAllThreads(void)
4167 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
4168 ullong_format_string(TIME_ON_PROC(CurrentProc),
4169 time_string, rtsFalse/*no commas!*/);
4171 debugBelch("all threads at [%s]:\n", time_string);
4172 # elif defined(PARALLEL_HASKELL)
4173 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
4174 ullong_format_string(CURRENT_TIME,
4175 time_string, rtsFalse/*no commas!*/);
4177 debugBelch("all threads at [%s]:\n", time_string);
4179 debugBelch("all threads:\n");
4182 for (t = all_threads; t != END_TSO_QUEUE; ) {
4183 debugBelch("\tthread %4d @ %p ", t->id, (void *)t);
4186 void *label = lookupThreadLabel(t->id);
4187 if (label) debugBelch("[\"%s\"] ",(char *)label);
4190 if (t->what_next == ThreadRelocated) {
4191 debugBelch("has been relocated...\n");
4194 printThreadStatus(t);
4205 printThreadQueue(StgTSO *t)
4208 for (; t != END_TSO_QUEUE; t = t->link) {
4209 debugBelch("\tthread %d @ %p ", t->id, (void *)t);
4210 if (t->what_next == ThreadRelocated) {
4211 debugBelch("has been relocated...\n");
4213 printThreadStatus(t);
4218 debugBelch("%d threads on queue\n", i);
4222 Print a whole blocking queue attached to node (debugging only).
4224 # if defined(PARALLEL_HASKELL)
4226 print_bq (StgClosure *node)
4228 StgBlockingQueueElement *bqe;
4232 debugBelch("## BQ of closure %p (%s): ",
4233 node, info_type(node));
4235 /* should cover all closures that may have a blocking queue */
4236 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
4237 get_itbl(node)->type == FETCH_ME_BQ ||
4238 get_itbl(node)->type == RBH ||
4239 get_itbl(node)->type == MVAR);
4241 ASSERT(node!=(StgClosure*)NULL); // sanity check
4243 print_bqe(((StgBlockingQueue*)node)->blocking_queue);
4247 Print a whole blocking queue starting with the element bqe.
4250 print_bqe (StgBlockingQueueElement *bqe)
4255 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
4257 for (end = (bqe==END_BQ_QUEUE);
4258 !end; // iterate until bqe points to a CONSTR
4259 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE),
4260 bqe = end ? END_BQ_QUEUE : bqe->link) {
4261 ASSERT(bqe != END_BQ_QUEUE); // sanity check
4262 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
4263 /* types of closures that may appear in a blocking queue */
4264 ASSERT(get_itbl(bqe)->type == TSO ||
4265 get_itbl(bqe)->type == BLOCKED_FETCH ||
4266 get_itbl(bqe)->type == CONSTR);
4267 /* only BQs of an RBH end with an RBH_Save closure */
4268 //ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
4270 switch (get_itbl(bqe)->type) {
4272 debugBelch(" TSO %u (%x),",
4273 ((StgTSO *)bqe)->id, ((StgTSO *)bqe));
4276 debugBelch(" BF (node=%p, ga=((%x, %d, %x)),",
4277 ((StgBlockedFetch *)bqe)->node,
4278 ((StgBlockedFetch *)bqe)->ga.payload.gc.gtid,
4279 ((StgBlockedFetch *)bqe)->ga.payload.gc.slot,
4280 ((StgBlockedFetch *)bqe)->ga.weight);
4283 debugBelch(" %s (IP %p),",
4284 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
4285 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
4286 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
4287 "RBH_Save_?"), get_itbl(bqe));
4290 barf("Unexpected closure type %s in blocking queue", // of %p (%s)",
4291 info_type((StgClosure *)bqe)); // , node, info_type(node));
4297 # elif defined(GRAN)
4299 print_bq (StgClosure *node)
4301 StgBlockingQueueElement *bqe;
4302 PEs node_loc, tso_loc;
4305 /* should cover all closures that may have a blocking queue */
4306 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
4307 get_itbl(node)->type == FETCH_ME_BQ ||
4308 get_itbl(node)->type == RBH);
4310 ASSERT(node!=(StgClosure*)NULL); // sanity check
4311 node_loc = where_is(node);
4313 debugBelch("## BQ of closure %p (%s) on [PE %d]: ",
4314 node, info_type(node), node_loc);
4317 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
4319 for (bqe = ((StgBlockingQueue*)node)->blocking_queue, end = (bqe==END_BQ_QUEUE);
4320 !end; // iterate until bqe points to a CONSTR
4321 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) {
4322 ASSERT(bqe != END_BQ_QUEUE); // sanity check
4323 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
4324 /* types of closures that may appear in a blocking queue */
4325 ASSERT(get_itbl(bqe)->type == TSO ||
4326 get_itbl(bqe)->type == CONSTR);
4327 /* only BQs of an RBH end with an RBH_Save closure */
4328 ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
4330 tso_loc = where_is((StgClosure *)bqe);
4331 switch (get_itbl(bqe)->type) {
4333 debugBelch(" TSO %d (%p) on [PE %d],",
4334 ((StgTSO *)bqe)->id, (StgTSO *)bqe, tso_loc);
4337 debugBelch(" %s (IP %p),",
4338 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
4339 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
4340 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
4341 "RBH_Save_?"), get_itbl(bqe));
4344 barf("Unexpected closure type %s in blocking queue of %p (%s)",
4345 info_type((StgClosure *)bqe), node, info_type(node));
4353 #if defined(PARALLEL_HASKELL)
4360 for (i=0, tso=run_queue_hd;
4361 tso != END_TSO_QUEUE;
4370 sched_belch(char *s, ...)
4374 #ifdef RTS_SUPPORTS_THREADS
4375 debugBelch("sched (task %p): ", osThreadId());
4376 #elif defined(PARALLEL_HASKELL)
4379 debugBelch("sched: ");