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 // We have run some Haskell code: there might be blackhole-blocked
711 // threads to wake up now.
712 if ( blackhole_queue != END_TSO_QUEUE ) {
713 blackholes_need_checking = rtsTrue;
716 cap->r.rInHaskell = rtsFalse;
718 // The TSO might have moved, eg. if it re-entered the RTS and a GC
719 // happened. So find the new location:
720 t = cap->r.rCurrentTSO;
722 // And save the current errno in this thread.
723 t->saved_errno = errno;
725 // ----------------------------------------------------------------------
727 /* Costs for the scheduler are assigned to CCS_SYSTEM */
728 #if defined(PROFILING)
733 ACQUIRE_LOCK(&sched_mutex);
735 #if defined(RTS_SUPPORTS_THREADS)
736 IF_DEBUG(scheduler,debugBelch("sched (task %p): ", osThreadId()););
737 #elif !defined(GRAN) && !defined(PARALLEL_HASKELL)
738 IF_DEBUG(scheduler,debugBelch("sched: "););
741 schedulePostRunThread();
743 ready_to_gc = rtsFalse;
747 ready_to_gc = scheduleHandleHeapOverflow(cap,t);
751 scheduleHandleStackOverflow(t);
755 if (scheduleHandleYield(t, prev_what_next)) {
756 // shortcut for switching between compiler/interpreter:
762 scheduleHandleThreadBlocked(t);
766 if (scheduleHandleThreadFinished(mainThread, cap, t)) return;;
770 barf("schedule: invalid thread return code %d", (int)ret);
773 if (scheduleDoHeapProfile(ready_to_gc)) { ready_to_gc = rtsFalse; }
774 if (ready_to_gc) { scheduleDoGC(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.
1910 // Someone else is already trying to GC
1911 if (waiting_for_gc) return;
1912 waiting_for_gc = rtsTrue;
1914 while (n_capabilities > 0) {
1915 IF_DEBUG(scheduler, sched_belch("ready_to_gc, grabbing all the capabilies (%d left)", n_capabilities));
1916 waitForReturnCapability(&sched_mutex, &cap);
1918 caps[n_capabilities] = cap;
1921 waiting_for_gc = rtsFalse;
1924 /* Kick any transactions which are invalid back to their
1925 * atomically frames. When next scheduled they will try to
1926 * commit, this commit will fail and they will retry.
1928 for (t = all_threads; t != END_TSO_QUEUE; t = t -> link) {
1929 if (t -> what_next != ThreadRelocated && t -> trec != NO_TREC && t -> why_blocked == NotBlocked) {
1930 if (!stmValidateNestOfTransactions (t -> trec)) {
1931 IF_DEBUG(stm, sched_belch("trec %p found wasting its time", t));
1933 // strip the stack back to the ATOMICALLY_FRAME, aborting
1934 // the (nested) transaction, and saving the stack of any
1935 // partially-evaluated thunks on the heap.
1936 raiseAsync_(t, NULL, rtsTrue);
1939 ASSERT(get_itbl((StgClosure *)t->sp)->type == ATOMICALLY_FRAME);
1945 // so this happens periodically:
1946 scheduleCheckBlackHoles();
1948 IF_DEBUG(scheduler, printAllThreads());
1950 /* everybody back, start the GC.
1951 * Could do it in this thread, or signal a condition var
1952 * to do it in another thread. Either way, we need to
1953 * broadcast on gc_pending_cond afterward.
1955 #if defined(RTS_SUPPORTS_THREADS)
1956 IF_DEBUG(scheduler,sched_belch("doing GC"));
1958 GarbageCollect(GetRoots, force_major);
1962 // release our stash of capabilities.
1964 for (i = 0; i < RtsFlags.ParFlags.nNodes-1; i++) {
1965 releaseCapability(caps[i]);
1971 /* add a ContinueThread event to continue execution of current thread */
1972 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1974 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1976 debugBelch("GRAN: eventq and runnableq after Garbage collection:\n\n");
1982 /* ---------------------------------------------------------------------------
1983 * rtsSupportsBoundThreads(): is the RTS built to support bound threads?
1984 * used by Control.Concurrent for error checking.
1985 * ------------------------------------------------------------------------- */
1988 rtsSupportsBoundThreads(void)
1990 #if defined(RTS_SUPPORTS_THREADS)
1997 /* ---------------------------------------------------------------------------
1998 * isThreadBound(tso): check whether tso is bound to an OS thread.
1999 * ------------------------------------------------------------------------- */
2002 isThreadBound(StgTSO* tso USED_IN_THREADED_RTS)
2004 #if defined(RTS_SUPPORTS_THREADS)
2005 return (tso->main != NULL);
2010 /* ---------------------------------------------------------------------------
2011 * Singleton fork(). Do not copy any running threads.
2012 * ------------------------------------------------------------------------- */
2014 #ifndef mingw32_HOST_OS
2015 #define FORKPROCESS_PRIMOP_SUPPORTED
2018 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
2020 deleteThreadImmediately(StgTSO *tso);
2023 forkProcess(HsStablePtr *entry
2024 #ifndef FORKPROCESS_PRIMOP_SUPPORTED
2029 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
2035 IF_DEBUG(scheduler,sched_belch("forking!"));
2036 rts_lock(); // This not only acquires sched_mutex, it also
2037 // makes sure that no other threads are running
2041 if (pid) { /* parent */
2043 /* just return the pid */
2047 } else { /* child */
2050 // delete all threads
2051 run_queue_hd = run_queue_tl = END_TSO_QUEUE;
2053 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
2056 // don't allow threads to catch the ThreadKilled exception
2057 deleteThreadImmediately(t);
2060 // wipe the main thread list
2061 while((m = main_threads) != NULL) {
2062 main_threads = m->link;
2063 # ifdef THREADED_RTS
2064 closeCondition(&m->bound_thread_cond);
2069 rc = rts_evalStableIO(entry, NULL); // run the action
2070 rts_checkSchedStatus("forkProcess",rc);
2074 hs_exit(); // clean up and exit
2077 #else /* !FORKPROCESS_PRIMOP_SUPPORTED */
2078 barf("forkProcess#: primop not supported, sorry!\n");
2083 /* ---------------------------------------------------------------------------
2084 * deleteAllThreads(): kill all the live threads.
2086 * This is used when we catch a user interrupt (^C), before performing
2087 * any necessary cleanups and running finalizers.
2089 * Locks: sched_mutex held.
2090 * ------------------------------------------------------------------------- */
2093 deleteAllThreads ( void )
2096 IF_DEBUG(scheduler,sched_belch("deleting all threads"));
2097 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
2098 if (t->what_next == ThreadRelocated) {
2101 next = t->global_link;
2106 // The run queue now contains a bunch of ThreadKilled threads. We
2107 // must not throw these away: the main thread(s) will be in there
2108 // somewhere, and the main scheduler loop has to deal with it.
2109 // Also, the run queue is the only thing keeping these threads from
2110 // being GC'd, and we don't want the "main thread has been GC'd" panic.
2112 ASSERT(blocked_queue_hd == END_TSO_QUEUE);
2113 ASSERT(blackhole_queue == END_TSO_QUEUE);
2114 ASSERT(sleeping_queue == END_TSO_QUEUE);
2117 /* startThread and insertThread are now in GranSim.c -- HWL */
2120 /* ---------------------------------------------------------------------------
2121 * Suspending & resuming Haskell threads.
2123 * When making a "safe" call to C (aka _ccall_GC), the task gives back
2124 * its capability before calling the C function. This allows another
2125 * task to pick up the capability and carry on running Haskell
2126 * threads. It also means that if the C call blocks, it won't lock
2129 * The Haskell thread making the C call is put to sleep for the
2130 * duration of the call, on the susepended_ccalling_threads queue. We
2131 * give out a token to the task, which it can use to resume the thread
2132 * on return from the C function.
2133 * ------------------------------------------------------------------------- */
2136 suspendThread( StgRegTable *reg )
2140 int saved_errno = errno;
2142 /* assume that *reg is a pointer to the StgRegTable part
2145 cap = (Capability *)((void *)((unsigned char*)reg - sizeof(StgFunTable)));
2147 ACQUIRE_LOCK(&sched_mutex);
2150 sched_belch("thread %d did a _ccall_gc", cap->r.rCurrentTSO->id));
2152 // XXX this might not be necessary --SDM
2153 cap->r.rCurrentTSO->what_next = ThreadRunGHC;
2155 threadPaused(cap->r.rCurrentTSO);
2156 cap->r.rCurrentTSO->link = suspended_ccalling_threads;
2157 suspended_ccalling_threads = cap->r.rCurrentTSO;
2159 if(cap->r.rCurrentTSO->blocked_exceptions == NULL) {
2160 cap->r.rCurrentTSO->why_blocked = BlockedOnCCall;
2161 cap->r.rCurrentTSO->blocked_exceptions = END_TSO_QUEUE;
2163 cap->r.rCurrentTSO->why_blocked = BlockedOnCCall_NoUnblockExc;
2166 /* Use the thread ID as the token; it should be unique */
2167 tok = cap->r.rCurrentTSO->id;
2169 /* Hand back capability */
2170 cap->r.rInHaskell = rtsFalse;
2171 releaseCapability(cap);
2173 #if defined(RTS_SUPPORTS_THREADS)
2174 /* Preparing to leave the RTS, so ensure there's a native thread/task
2175 waiting to take over.
2177 IF_DEBUG(scheduler, sched_belch("worker (token %d): leaving RTS", tok));
2180 RELEASE_LOCK(&sched_mutex);
2182 errno = saved_errno;
2187 resumeThread( StgInt tok )
2189 StgTSO *tso, **prev;
2191 int saved_errno = errno;
2193 #if defined(RTS_SUPPORTS_THREADS)
2194 /* Wait for permission to re-enter the RTS with the result. */
2195 ACQUIRE_LOCK(&sched_mutex);
2196 waitForReturnCapability(&sched_mutex, &cap);
2198 IF_DEBUG(scheduler, sched_belch("worker (token %d): re-entering RTS", tok));
2200 grabCapability(&cap);
2203 /* Remove the thread off of the suspended list */
2204 prev = &suspended_ccalling_threads;
2205 for (tso = suspended_ccalling_threads;
2206 tso != END_TSO_QUEUE;
2207 prev = &tso->link, tso = tso->link) {
2208 if (tso->id == (StgThreadID)tok) {
2213 if (tso == END_TSO_QUEUE) {
2214 barf("resumeThread: thread not found");
2216 tso->link = END_TSO_QUEUE;
2218 if(tso->why_blocked == BlockedOnCCall) {
2219 awakenBlockedQueueNoLock(tso->blocked_exceptions);
2220 tso->blocked_exceptions = NULL;
2223 /* Reset blocking status */
2224 tso->why_blocked = NotBlocked;
2226 cap->r.rCurrentTSO = tso;
2227 cap->r.rInHaskell = rtsTrue;
2228 RELEASE_LOCK(&sched_mutex);
2229 errno = saved_errno;
2233 /* ---------------------------------------------------------------------------
2234 * Comparing Thread ids.
2236 * This is used from STG land in the implementation of the
2237 * instances of Eq/Ord for ThreadIds.
2238 * ------------------------------------------------------------------------ */
2241 cmp_thread(StgPtr tso1, StgPtr tso2)
2243 StgThreadID id1 = ((StgTSO *)tso1)->id;
2244 StgThreadID id2 = ((StgTSO *)tso2)->id;
2246 if (id1 < id2) return (-1);
2247 if (id1 > id2) return 1;
2251 /* ---------------------------------------------------------------------------
2252 * Fetching the ThreadID from an StgTSO.
2254 * This is used in the implementation of Show for ThreadIds.
2255 * ------------------------------------------------------------------------ */
2257 rts_getThreadId(StgPtr tso)
2259 return ((StgTSO *)tso)->id;
2264 labelThread(StgPtr tso, char *label)
2269 /* Caveat: Once set, you can only set the thread name to "" */
2270 len = strlen(label)+1;
2271 buf = stgMallocBytes(len * sizeof(char), "Schedule.c:labelThread()");
2272 strncpy(buf,label,len);
2273 /* Update will free the old memory for us */
2274 updateThreadLabel(((StgTSO *)tso)->id,buf);
2278 /* ---------------------------------------------------------------------------
2279 Create a new thread.
2281 The new thread starts with the given stack size. Before the
2282 scheduler can run, however, this thread needs to have a closure
2283 (and possibly some arguments) pushed on its stack. See
2284 pushClosure() in Schedule.h.
2286 createGenThread() and createIOThread() (in SchedAPI.h) are
2287 convenient packaged versions of this function.
2289 currently pri (priority) is only used in a GRAN setup -- HWL
2290 ------------------------------------------------------------------------ */
2292 /* currently pri (priority) is only used in a GRAN setup -- HWL */
2294 createThread(nat size, StgInt pri)
2297 createThread(nat size)
2304 /* First check whether we should create a thread at all */
2305 #if defined(PARALLEL_HASKELL)
2306 /* check that no more than RtsFlags.ParFlags.maxThreads threads are created */
2307 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads) {
2309 debugBelch("{createThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)\n",
2310 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
2311 return END_TSO_QUEUE;
2317 ASSERT(!RtsFlags.GranFlags.Light || CurrentProc==0);
2320 // ToDo: check whether size = stack_size - TSO_STRUCT_SIZEW
2322 /* catch ridiculously small stack sizes */
2323 if (size < MIN_STACK_WORDS + TSO_STRUCT_SIZEW) {
2324 size = MIN_STACK_WORDS + TSO_STRUCT_SIZEW;
2327 stack_size = size - TSO_STRUCT_SIZEW;
2329 tso = (StgTSO *)allocate(size);
2330 TICK_ALLOC_TSO(stack_size, 0);
2332 SET_HDR(tso, &stg_TSO_info, CCS_SYSTEM);
2334 SET_GRAN_HDR(tso, ThisPE);
2337 // Always start with the compiled code evaluator
2338 tso->what_next = ThreadRunGHC;
2340 tso->id = next_thread_id++;
2341 tso->why_blocked = NotBlocked;
2342 tso->blocked_exceptions = NULL;
2344 tso->saved_errno = 0;
2347 tso->stack_size = stack_size;
2348 tso->max_stack_size = round_to_mblocks(RtsFlags.GcFlags.maxStkSize)
2350 tso->sp = (P_)&(tso->stack) + stack_size;
2352 tso->trec = NO_TREC;
2355 tso->prof.CCCS = CCS_MAIN;
2358 /* put a stop frame on the stack */
2359 tso->sp -= sizeofW(StgStopFrame);
2360 SET_HDR((StgClosure*)tso->sp,(StgInfoTable *)&stg_stop_thread_info,CCS_SYSTEM);
2361 tso->link = END_TSO_QUEUE;
2365 /* uses more flexible routine in GranSim */
2366 insertThread(tso, CurrentProc);
2368 /* In a non-GranSim setup the pushing of a TSO onto the runq is separated
2374 if (RtsFlags.GranFlags.GranSimStats.Full)
2375 DumpGranEvent(GR_START,tso);
2376 #elif defined(PARALLEL_HASKELL)
2377 if (RtsFlags.ParFlags.ParStats.Full)
2378 DumpGranEvent(GR_STARTQ,tso);
2379 /* HACk to avoid SCHEDULE
2383 /* Link the new thread on the global thread list.
2385 tso->global_link = all_threads;
2389 tso->dist.priority = MandatoryPriority; //by default that is...
2393 tso->gran.pri = pri;
2395 tso->gran.magic = TSO_MAGIC; // debugging only
2397 tso->gran.sparkname = 0;
2398 tso->gran.startedat = CURRENT_TIME;
2399 tso->gran.exported = 0;
2400 tso->gran.basicblocks = 0;
2401 tso->gran.allocs = 0;
2402 tso->gran.exectime = 0;
2403 tso->gran.fetchtime = 0;
2404 tso->gran.fetchcount = 0;
2405 tso->gran.blocktime = 0;
2406 tso->gran.blockcount = 0;
2407 tso->gran.blockedat = 0;
2408 tso->gran.globalsparks = 0;
2409 tso->gran.localsparks = 0;
2410 if (RtsFlags.GranFlags.Light)
2411 tso->gran.clock = Now; /* local clock */
2413 tso->gran.clock = 0;
2415 IF_DEBUG(gran,printTSO(tso));
2416 #elif defined(PARALLEL_HASKELL)
2418 tso->par.magic = TSO_MAGIC; // debugging only
2420 tso->par.sparkname = 0;
2421 tso->par.startedat = CURRENT_TIME;
2422 tso->par.exported = 0;
2423 tso->par.basicblocks = 0;
2424 tso->par.allocs = 0;
2425 tso->par.exectime = 0;
2426 tso->par.fetchtime = 0;
2427 tso->par.fetchcount = 0;
2428 tso->par.blocktime = 0;
2429 tso->par.blockcount = 0;
2430 tso->par.blockedat = 0;
2431 tso->par.globalsparks = 0;
2432 tso->par.localsparks = 0;
2436 globalGranStats.tot_threads_created++;
2437 globalGranStats.threads_created_on_PE[CurrentProc]++;
2438 globalGranStats.tot_sq_len += spark_queue_len(CurrentProc);
2439 globalGranStats.tot_sq_probes++;
2440 #elif defined(PARALLEL_HASKELL)
2441 // collect parallel global statistics (currently done together with GC stats)
2442 if (RtsFlags.ParFlags.ParStats.Global &&
2443 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
2444 //debugBelch("Creating thread %d @ %11.2f\n", tso->id, usertime());
2445 globalParStats.tot_threads_created++;
2451 sched_belch("==__ schedule: Created TSO %d (%p);",
2452 CurrentProc, tso, tso->id));
2453 #elif defined(PARALLEL_HASKELL)
2454 IF_PAR_DEBUG(verbose,
2455 sched_belch("==__ schedule: Created TSO %d (%p); %d threads active",
2456 (long)tso->id, tso, advisory_thread_count));
2458 IF_DEBUG(scheduler,sched_belch("created thread %ld, stack size = %lx words",
2459 (long)tso->id, (long)tso->stack_size));
2466 all parallel thread creation calls should fall through the following routine.
2469 createThreadFromSpark(rtsSpark spark)
2471 ASSERT(spark != (rtsSpark)NULL);
2472 // JB: TAKE CARE OF THIS COUNTER! BUGGY
2473 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads)
2475 barf("{createSparkThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
2476 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
2477 return END_TSO_QUEUE;
2481 tso = createThread(RtsFlags.GcFlags.initialStkSize);
2482 if (tso==END_TSO_QUEUE)
2483 barf("createSparkThread: Cannot create TSO");
2485 tso->priority = AdvisoryPriority;
2487 pushClosure(tso,spark);
2489 advisory_thread_count++; // JB: TAKE CARE OF THIS COUNTER! BUGGY
2496 Turn a spark into a thread.
2497 ToDo: fix for SMP (needs to acquire SCHED_MUTEX!)
2501 activateSpark (rtsSpark spark)
2505 tso = createSparkThread(spark);
2506 if (RtsFlags.ParFlags.ParStats.Full) {
2507 //ASSERT(run_queue_hd == END_TSO_QUEUE); // I think ...
2508 IF_PAR_DEBUG(verbose,
2509 debugBelch("==^^ activateSpark: turning spark of closure %p (%s) into a thread\n",
2510 (StgClosure *)spark, info_type((StgClosure *)spark)));
2512 // ToDo: fwd info on local/global spark to thread -- HWL
2513 // tso->gran.exported = spark->exported;
2514 // tso->gran.locked = !spark->global;
2515 // tso->gran.sparkname = spark->name;
2521 /* ---------------------------------------------------------------------------
2524 * scheduleThread puts a thread on the head of the runnable queue.
2525 * This will usually be done immediately after a thread is created.
2526 * The caller of scheduleThread must create the thread using e.g.
2527 * createThread and push an appropriate closure
2528 * on this thread's stack before the scheduler is invoked.
2529 * ------------------------------------------------------------------------ */
2532 scheduleThread_(StgTSO *tso)
2534 // The thread goes at the *end* of the run-queue, to avoid possible
2535 // starvation of any threads already on the queue.
2536 APPEND_TO_RUN_QUEUE(tso);
2541 scheduleThread(StgTSO* tso)
2543 ACQUIRE_LOCK(&sched_mutex);
2544 scheduleThread_(tso);
2545 RELEASE_LOCK(&sched_mutex);
2548 #if defined(RTS_SUPPORTS_THREADS)
2549 static Condition bound_cond_cache;
2550 static int bound_cond_cache_full = 0;
2555 scheduleWaitThread(StgTSO* tso, /*[out]*/HaskellObj* ret,
2556 Capability *initialCapability)
2558 // Precondition: sched_mutex must be held
2561 m = stgMallocBytes(sizeof(StgMainThread), "waitThread");
2566 m->link = main_threads;
2568 if (main_threads != NULL) {
2569 main_threads->prev = m;
2573 #if defined(RTS_SUPPORTS_THREADS)
2574 // Allocating a new condition for each thread is expensive, so we
2575 // cache one. This is a pretty feeble hack, but it helps speed up
2576 // consecutive call-ins quite a bit.
2577 if (bound_cond_cache_full) {
2578 m->bound_thread_cond = bound_cond_cache;
2579 bound_cond_cache_full = 0;
2581 initCondition(&m->bound_thread_cond);
2585 /* Put the thread on the main-threads list prior to scheduling the TSO.
2586 Failure to do so introduces a race condition in the MT case (as
2587 identified by Wolfgang Thaller), whereby the new task/OS thread
2588 created by scheduleThread_() would complete prior to the thread
2589 that spawned it managed to put 'itself' on the main-threads list.
2590 The upshot of it all being that the worker thread wouldn't get to
2591 signal the completion of the its work item for the main thread to
2592 see (==> it got stuck waiting.) -- sof 6/02.
2594 IF_DEBUG(scheduler, sched_belch("waiting for thread (%d)", tso->id));
2596 APPEND_TO_RUN_QUEUE(tso);
2597 // NB. Don't call threadRunnable() here, because the thread is
2598 // bound and only runnable by *this* OS thread, so waking up other
2599 // workers will just slow things down.
2601 return waitThread_(m, initialCapability);
2604 /* ---------------------------------------------------------------------------
2607 * Initialise the scheduler. This resets all the queues - if the
2608 * queues contained any threads, they'll be garbage collected at the
2611 * ------------------------------------------------------------------------ */
2619 for (i=0; i<=MAX_PROC; i++) {
2620 run_queue_hds[i] = END_TSO_QUEUE;
2621 run_queue_tls[i] = END_TSO_QUEUE;
2622 blocked_queue_hds[i] = END_TSO_QUEUE;
2623 blocked_queue_tls[i] = END_TSO_QUEUE;
2624 ccalling_threadss[i] = END_TSO_QUEUE;
2625 blackhole_queue[i] = END_TSO_QUEUE;
2626 sleeping_queue = END_TSO_QUEUE;
2629 run_queue_hd = END_TSO_QUEUE;
2630 run_queue_tl = END_TSO_QUEUE;
2631 blocked_queue_hd = END_TSO_QUEUE;
2632 blocked_queue_tl = END_TSO_QUEUE;
2633 blackhole_queue = END_TSO_QUEUE;
2634 sleeping_queue = END_TSO_QUEUE;
2637 suspended_ccalling_threads = END_TSO_QUEUE;
2639 main_threads = NULL;
2640 all_threads = END_TSO_QUEUE;
2645 RtsFlags.ConcFlags.ctxtSwitchTicks =
2646 RtsFlags.ConcFlags.ctxtSwitchTime / TICK_MILLISECS;
2648 #if defined(RTS_SUPPORTS_THREADS)
2649 /* Initialise the mutex and condition variables used by
2651 initMutex(&sched_mutex);
2652 initMutex(&term_mutex);
2655 ACQUIRE_LOCK(&sched_mutex);
2657 /* A capability holds the state a native thread needs in
2658 * order to execute STG code. At least one capability is
2659 * floating around (only SMP builds have more than one).
2663 #if defined(RTS_SUPPORTS_THREADS)
2668 /* eagerly start some extra workers */
2669 startingWorkerThread = RtsFlags.ParFlags.nNodes;
2670 startTasks(RtsFlags.ParFlags.nNodes, taskStart);
2673 #if /* defined(SMP) ||*/ defined(PARALLEL_HASKELL)
2677 RELEASE_LOCK(&sched_mutex);
2681 exitScheduler( void )
2683 interrupted = rtsTrue;
2684 shutting_down_scheduler = rtsTrue;
2685 #if defined(RTS_SUPPORTS_THREADS)
2686 if (threadIsTask(osThreadId())) { taskStop(); }
2691 /* ----------------------------------------------------------------------------
2692 Managing the per-task allocation areas.
2694 Each capability comes with an allocation area. These are
2695 fixed-length block lists into which allocation can be done.
2697 ToDo: no support for two-space collection at the moment???
2698 ------------------------------------------------------------------------- */
2700 static SchedulerStatus
2701 waitThread_(StgMainThread* m, Capability *initialCapability)
2703 SchedulerStatus stat;
2705 // Precondition: sched_mutex must be held.
2706 IF_DEBUG(scheduler, sched_belch("new main thread (%d)", m->tso->id));
2709 /* GranSim specific init */
2710 CurrentTSO = m->tso; // the TSO to run
2711 procStatus[MainProc] = Busy; // status of main PE
2712 CurrentProc = MainProc; // PE to run it on
2713 schedule(m,initialCapability);
2715 schedule(m,initialCapability);
2716 ASSERT(m->stat != NoStatus);
2721 #if defined(RTS_SUPPORTS_THREADS)
2722 // Free the condition variable, returning it to the cache if possible.
2723 if (!bound_cond_cache_full) {
2724 bound_cond_cache = m->bound_thread_cond;
2725 bound_cond_cache_full = 1;
2727 closeCondition(&m->bound_thread_cond);
2731 IF_DEBUG(scheduler, sched_belch("main thread (%d) finished", m->tso->id));
2734 // Postcondition: sched_mutex still held
2738 /* ---------------------------------------------------------------------------
2739 Where are the roots that we know about?
2741 - all the threads on the runnable queue
2742 - all the threads on the blocked queue
2743 - all the threads on the sleeping queue
2744 - all the thread currently executing a _ccall_GC
2745 - all the "main threads"
2747 ------------------------------------------------------------------------ */
2749 /* This has to be protected either by the scheduler monitor, or by the
2750 garbage collection monitor (probably the latter).
2755 GetRoots( evac_fn evac )
2760 for (i=0; i<=RtsFlags.GranFlags.proc; i++) {
2761 if ((run_queue_hds[i] != END_TSO_QUEUE) && ((run_queue_hds[i] != NULL)))
2762 evac((StgClosure **)&run_queue_hds[i]);
2763 if ((run_queue_tls[i] != END_TSO_QUEUE) && ((run_queue_tls[i] != NULL)))
2764 evac((StgClosure **)&run_queue_tls[i]);
2766 if ((blocked_queue_hds[i] != END_TSO_QUEUE) && ((blocked_queue_hds[i] != NULL)))
2767 evac((StgClosure **)&blocked_queue_hds[i]);
2768 if ((blocked_queue_tls[i] != END_TSO_QUEUE) && ((blocked_queue_tls[i] != NULL)))
2769 evac((StgClosure **)&blocked_queue_tls[i]);
2770 if ((ccalling_threadss[i] != END_TSO_QUEUE) && ((ccalling_threadss[i] != NULL)))
2771 evac((StgClosure **)&ccalling_threads[i]);
2778 if (run_queue_hd != END_TSO_QUEUE) {
2779 ASSERT(run_queue_tl != END_TSO_QUEUE);
2780 evac((StgClosure **)&run_queue_hd);
2781 evac((StgClosure **)&run_queue_tl);
2784 if (blocked_queue_hd != END_TSO_QUEUE) {
2785 ASSERT(blocked_queue_tl != END_TSO_QUEUE);
2786 evac((StgClosure **)&blocked_queue_hd);
2787 evac((StgClosure **)&blocked_queue_tl);
2790 if (sleeping_queue != END_TSO_QUEUE) {
2791 evac((StgClosure **)&sleeping_queue);
2795 if (blackhole_queue != END_TSO_QUEUE) {
2796 evac((StgClosure **)&blackhole_queue);
2799 if (suspended_ccalling_threads != END_TSO_QUEUE) {
2800 evac((StgClosure **)&suspended_ccalling_threads);
2803 #if defined(PARALLEL_HASKELL) || defined(GRAN)
2804 markSparkQueue(evac);
2807 #if defined(RTS_USER_SIGNALS)
2808 // mark the signal handlers (signals should be already blocked)
2809 markSignalHandlers(evac);
2813 /* -----------------------------------------------------------------------------
2816 This is the interface to the garbage collector from Haskell land.
2817 We provide this so that external C code can allocate and garbage
2818 collect when called from Haskell via _ccall_GC.
2820 It might be useful to provide an interface whereby the programmer
2821 can specify more roots (ToDo).
2823 This needs to be protected by the GC condition variable above. KH.
2824 -------------------------------------------------------------------------- */
2826 static void (*extra_roots)(evac_fn);
2831 /* Obligated to hold this lock upon entry */
2832 ACQUIRE_LOCK(&sched_mutex);
2833 GarbageCollect(GetRoots,rtsFalse);
2834 RELEASE_LOCK(&sched_mutex);
2838 performMajorGC(void)
2840 ACQUIRE_LOCK(&sched_mutex);
2841 GarbageCollect(GetRoots,rtsTrue);
2842 RELEASE_LOCK(&sched_mutex);
2846 AllRoots(evac_fn evac)
2848 GetRoots(evac); // the scheduler's roots
2849 extra_roots(evac); // the user's roots
2853 performGCWithRoots(void (*get_roots)(evac_fn))
2855 ACQUIRE_LOCK(&sched_mutex);
2856 extra_roots = get_roots;
2857 GarbageCollect(AllRoots,rtsFalse);
2858 RELEASE_LOCK(&sched_mutex);
2861 /* -----------------------------------------------------------------------------
2864 If the thread has reached its maximum stack size, then raise the
2865 StackOverflow exception in the offending thread. Otherwise
2866 relocate the TSO into a larger chunk of memory and adjust its stack
2868 -------------------------------------------------------------------------- */
2871 threadStackOverflow(StgTSO *tso)
2873 nat new_stack_size, stack_words;
2878 IF_DEBUG(sanity,checkTSO(tso));
2879 if (tso->stack_size >= tso->max_stack_size) {
2882 debugBelch("@@ threadStackOverflow of TSO %ld (%p): stack too large (now %ld; max is %ld)\n",
2883 (long)tso->id, tso, (long)tso->stack_size, (long)tso->max_stack_size);
2884 /* If we're debugging, just print out the top of the stack */
2885 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2888 /* Send this thread the StackOverflow exception */
2889 raiseAsync(tso, (StgClosure *)stackOverflow_closure);
2893 /* Try to double the current stack size. If that takes us over the
2894 * maximum stack size for this thread, then use the maximum instead.
2895 * Finally round up so the TSO ends up as a whole number of blocks.
2897 new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
2898 new_tso_size = (lnat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) +
2899 TSO_STRUCT_SIZE)/sizeof(W_);
2900 new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */
2901 new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
2903 IF_DEBUG(scheduler, debugBelch("== sched: increasing stack size from %d words to %d.\n", tso->stack_size, new_stack_size));
2905 dest = (StgTSO *)allocate(new_tso_size);
2906 TICK_ALLOC_TSO(new_stack_size,0);
2908 /* copy the TSO block and the old stack into the new area */
2909 memcpy(dest,tso,TSO_STRUCT_SIZE);
2910 stack_words = tso->stack + tso->stack_size - tso->sp;
2911 new_sp = (P_)dest + new_tso_size - stack_words;
2912 memcpy(new_sp, tso->sp, stack_words * sizeof(W_));
2914 /* relocate the stack pointers... */
2916 dest->stack_size = new_stack_size;
2918 /* Mark the old TSO as relocated. We have to check for relocated
2919 * TSOs in the garbage collector and any primops that deal with TSOs.
2921 * It's important to set the sp value to just beyond the end
2922 * of the stack, so we don't attempt to scavenge any part of the
2925 tso->what_next = ThreadRelocated;
2927 tso->sp = (P_)&(tso->stack[tso->stack_size]);
2928 tso->why_blocked = NotBlocked;
2930 IF_PAR_DEBUG(verbose,
2931 debugBelch("@@ threadStackOverflow of TSO %d (now at %p): stack size increased to %ld\n",
2932 tso->id, tso, tso->stack_size);
2933 /* If we're debugging, just print out the top of the stack */
2934 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2937 IF_DEBUG(sanity,checkTSO(tso));
2939 IF_DEBUG(scheduler,printTSO(dest));
2945 /* ---------------------------------------------------------------------------
2946 Wake up a queue that was blocked on some resource.
2947 ------------------------------------------------------------------------ */
2951 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2954 #elif defined(PARALLEL_HASKELL)
2956 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2958 /* write RESUME events to log file and
2959 update blocked and fetch time (depending on type of the orig closure) */
2960 if (RtsFlags.ParFlags.ParStats.Full) {
2961 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
2962 GR_RESUMEQ, ((StgTSO *)bqe), ((StgTSO *)bqe)->block_info.closure,
2963 0, 0 /* spark_queue_len(ADVISORY_POOL) */);
2964 if (EMPTY_RUN_QUEUE())
2965 emitSchedule = rtsTrue;
2967 switch (get_itbl(node)->type) {
2969 ((StgTSO *)bqe)->par.fetchtime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2974 ((StgTSO *)bqe)->par.blocktime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2981 barf("{unblockOneLocked}Daq Qagh: unexpected closure in blocking queue");
2988 StgBlockingQueueElement *
2989 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2992 PEs node_loc, tso_loc;
2994 node_loc = where_is(node); // should be lifted out of loop
2995 tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2996 tso_loc = where_is((StgClosure *)tso);
2997 if (IS_LOCAL_TO(PROCS(node),tso_loc)) { // TSO is local
2998 /* !fake_fetch => TSO is on CurrentProc is same as IS_LOCAL_TO */
2999 ASSERT(CurrentProc!=node_loc || tso_loc==CurrentProc);
3000 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.lunblocktime;
3001 // insertThread(tso, node_loc);
3002 new_event(tso_loc, tso_loc, CurrentTime[CurrentProc],
3004 tso, node, (rtsSpark*)NULL);
3005 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
3008 } else { // TSO is remote (actually should be FMBQ)
3009 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.mpacktime +
3010 RtsFlags.GranFlags.Costs.gunblocktime +
3011 RtsFlags.GranFlags.Costs.latency;
3012 new_event(tso_loc, CurrentProc, CurrentTime[CurrentProc],
3014 tso, node, (rtsSpark*)NULL);
3015 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
3018 /* the thread-queue-overhead is accounted for in either Resume or UnblockThread */
3020 debugBelch(" %s TSO %d (%p) [PE %d] (block_info.closure=%p) (next=%p) ,",
3021 (node_loc==tso_loc ? "Local" : "Global"),
3022 tso->id, tso, CurrentProc, tso->block_info.closure, tso->link));
3023 tso->block_info.closure = NULL;
3024 IF_DEBUG(scheduler,debugBelch("-- Waking up thread %ld (%p)\n",
3027 #elif defined(PARALLEL_HASKELL)
3028 StgBlockingQueueElement *
3029 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
3031 StgBlockingQueueElement *next;
3033 switch (get_itbl(bqe)->type) {
3035 ASSERT(((StgTSO *)bqe)->why_blocked != NotBlocked);
3036 /* if it's a TSO just push it onto the run_queue */
3038 ((StgTSO *)bqe)->link = END_TSO_QUEUE; // debugging?
3039 APPEND_TO_RUN_QUEUE((StgTSO *)bqe);
3041 unblockCount(bqe, node);
3042 /* reset blocking status after dumping event */
3043 ((StgTSO *)bqe)->why_blocked = NotBlocked;
3047 /* if it's a BLOCKED_FETCH put it on the PendingFetches list */
3049 bqe->link = (StgBlockingQueueElement *)PendingFetches;
3050 PendingFetches = (StgBlockedFetch *)bqe;
3054 /* can ignore this case in a non-debugging setup;
3055 see comments on RBHSave closures above */
3057 /* check that the closure is an RBHSave closure */
3058 ASSERT(get_itbl((StgClosure *)bqe) == &stg_RBH_Save_0_info ||
3059 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_1_info ||
3060 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_2_info);
3064 barf("{unblockOneLocked}Daq Qagh: Unexpected IP (%#lx; %s) in blocking queue at %#lx\n",
3065 get_itbl((StgClosure *)bqe), info_type((StgClosure *)bqe),
3069 IF_PAR_DEBUG(bq, debugBelch(", %p (%s)\n", bqe, info_type((StgClosure*)bqe)));
3073 #else /* !GRAN && !PARALLEL_HASKELL */
3075 unblockOneLocked(StgTSO *tso)
3079 ASSERT(get_itbl(tso)->type == TSO);
3080 ASSERT(tso->why_blocked != NotBlocked);
3081 tso->why_blocked = NotBlocked;
3083 tso->link = END_TSO_QUEUE;
3084 APPEND_TO_RUN_QUEUE(tso);
3086 IF_DEBUG(scheduler,sched_belch("waking up thread %ld", (long)tso->id));
3091 #if defined(GRAN) || defined(PARALLEL_HASKELL)
3092 INLINE_ME StgBlockingQueueElement *
3093 unblockOne(StgBlockingQueueElement *bqe, StgClosure *node)
3095 ACQUIRE_LOCK(&sched_mutex);
3096 bqe = unblockOneLocked(bqe, node);
3097 RELEASE_LOCK(&sched_mutex);
3102 unblockOne(StgTSO *tso)
3104 ACQUIRE_LOCK(&sched_mutex);
3105 tso = unblockOneLocked(tso);
3106 RELEASE_LOCK(&sched_mutex);
3113 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
3115 StgBlockingQueueElement *bqe;
3120 debugBelch("##-_ AwBQ for node %p on PE %d @ %ld by TSO %d (%p): \n", \
3121 node, CurrentProc, CurrentTime[CurrentProc],
3122 CurrentTSO->id, CurrentTSO));
3124 node_loc = where_is(node);
3126 ASSERT(q == END_BQ_QUEUE ||
3127 get_itbl(q)->type == TSO || // q is either a TSO or an RBHSave
3128 get_itbl(q)->type == CONSTR); // closure (type constructor)
3129 ASSERT(is_unique(node));
3131 /* FAKE FETCH: magically copy the node to the tso's proc;
3132 no Fetch necessary because in reality the node should not have been
3133 moved to the other PE in the first place
3135 if (CurrentProc!=node_loc) {
3137 debugBelch("## node %p is on PE %d but CurrentProc is %d (TSO %d); assuming fake fetch and adjusting bitmask (old: %#x)\n",
3138 node, node_loc, CurrentProc, CurrentTSO->id,
3139 // CurrentTSO, where_is(CurrentTSO),
3140 node->header.gran.procs));
3141 node->header.gran.procs = (node->header.gran.procs) | PE_NUMBER(CurrentProc);
3143 debugBelch("## new bitmask of node %p is %#x\n",
3144 node, node->header.gran.procs));
3145 if (RtsFlags.GranFlags.GranSimStats.Global) {
3146 globalGranStats.tot_fake_fetches++;
3151 // ToDo: check: ASSERT(CurrentProc==node_loc);
3152 while (get_itbl(bqe)->type==TSO) { // q != END_TSO_QUEUE) {
3155 bqe points to the current element in the queue
3156 next points to the next element in the queue
3158 //tso = (StgTSO *)bqe; // wastes an assignment to get the type right
3159 //tso_loc = where_is(tso);
3161 bqe = unblockOneLocked(bqe, node);
3164 /* if this is the BQ of an RBH, we have to put back the info ripped out of
3165 the closure to make room for the anchor of the BQ */
3166 if (bqe!=END_BQ_QUEUE) {
3167 ASSERT(get_itbl(node)->type == RBH && get_itbl(bqe)->type == CONSTR);
3169 ASSERT((info_ptr==&RBH_Save_0_info) ||
3170 (info_ptr==&RBH_Save_1_info) ||
3171 (info_ptr==&RBH_Save_2_info));
3173 /* cf. convertToRBH in RBH.c for writing the RBHSave closure */
3174 ((StgRBH *)node)->blocking_queue = (StgBlockingQueueElement *)((StgRBHSave *)bqe)->payload[0];
3175 ((StgRBH *)node)->mut_link = (StgMutClosure *)((StgRBHSave *)bqe)->payload[1];
3178 debugBelch("## Filled in RBH_Save for %p (%s) at end of AwBQ\n",
3179 node, info_type(node)));
3182 /* statistics gathering */
3183 if (RtsFlags.GranFlags.GranSimStats.Global) {
3184 // globalGranStats.tot_bq_processing_time += bq_processing_time;
3185 globalGranStats.tot_bq_len += len; // total length of all bqs awakened
3186 // globalGranStats.tot_bq_len_local += len_local; // same for local TSOs only
3187 globalGranStats.tot_awbq++; // total no. of bqs awakened
3190 debugBelch("## BQ Stats of %p: [%d entries] %s\n",
3191 node, len, (bqe!=END_BQ_QUEUE) ? "RBH" : ""));
3193 #elif defined(PARALLEL_HASKELL)
3195 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
3197 StgBlockingQueueElement *bqe;
3199 ACQUIRE_LOCK(&sched_mutex);
3201 IF_PAR_DEBUG(verbose,
3202 debugBelch("##-_ AwBQ for node %p on [%x]: \n",
3206 if(get_itbl(q)->type == CONSTR || q==END_BQ_QUEUE) {
3207 IF_PAR_DEBUG(verbose, debugBelch("## ... nothing to unblock so lets just return. RFP (BUG?)\n"));
3212 ASSERT(q == END_BQ_QUEUE ||
3213 get_itbl(q)->type == TSO ||
3214 get_itbl(q)->type == BLOCKED_FETCH ||
3215 get_itbl(q)->type == CONSTR);
3218 while (get_itbl(bqe)->type==TSO ||
3219 get_itbl(bqe)->type==BLOCKED_FETCH) {
3220 bqe = unblockOneLocked(bqe, node);
3222 RELEASE_LOCK(&sched_mutex);
3225 #else /* !GRAN && !PARALLEL_HASKELL */
3228 awakenBlockedQueueNoLock(StgTSO *tso)
3230 while (tso != END_TSO_QUEUE) {
3231 tso = unblockOneLocked(tso);
3236 awakenBlockedQueue(StgTSO *tso)
3238 ACQUIRE_LOCK(&sched_mutex);
3239 while (tso != END_TSO_QUEUE) {
3240 tso = unblockOneLocked(tso);
3242 RELEASE_LOCK(&sched_mutex);
3246 /* ---------------------------------------------------------------------------
3248 - usually called inside a signal handler so it mustn't do anything fancy.
3249 ------------------------------------------------------------------------ */
3252 interruptStgRts(void)
3257 /* ToDo: if invoked from a signal handler, this threadRunnable
3258 * only works if there's another thread (not this one) waiting to
3263 /* -----------------------------------------------------------------------------
3266 This is for use when we raise an exception in another thread, which
3268 This has nothing to do with the UnblockThread event in GranSim. -- HWL
3269 -------------------------------------------------------------------------- */
3271 #if defined(GRAN) || defined(PARALLEL_HASKELL)
3273 NB: only the type of the blocking queue is different in GranSim and GUM
3274 the operations on the queue-elements are the same
3275 long live polymorphism!
3277 Locks: sched_mutex is held upon entry and exit.
3281 unblockThread(StgTSO *tso)
3283 StgBlockingQueueElement *t, **last;
3285 switch (tso->why_blocked) {
3288 return; /* not blocked */
3291 // Be careful: nothing to do here! We tell the scheduler that the thread
3292 // is runnable and we leave it to the stack-walking code to abort the
3293 // transaction while unwinding the stack. We should perhaps have a debugging
3294 // test to make sure that this really happens and that the 'zombie' transaction
3295 // does not get committed.
3299 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
3301 StgBlockingQueueElement *last_tso = END_BQ_QUEUE;
3302 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
3304 last = (StgBlockingQueueElement **)&mvar->head;
3305 for (t = (StgBlockingQueueElement *)mvar->head;
3307 last = &t->link, last_tso = t, t = t->link) {
3308 if (t == (StgBlockingQueueElement *)tso) {
3309 *last = (StgBlockingQueueElement *)tso->link;
3310 if (mvar->tail == tso) {
3311 mvar->tail = (StgTSO *)last_tso;
3316 barf("unblockThread (MVAR): TSO not found");
3319 case BlockedOnBlackHole:
3320 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
3322 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
3324 last = &bq->blocking_queue;
3325 for (t = bq->blocking_queue;
3327 last = &t->link, t = t->link) {
3328 if (t == (StgBlockingQueueElement *)tso) {
3329 *last = (StgBlockingQueueElement *)tso->link;
3333 barf("unblockThread (BLACKHOLE): TSO not found");
3336 case BlockedOnException:
3338 StgTSO *target = tso->block_info.tso;
3340 ASSERT(get_itbl(target)->type == TSO);
3342 if (target->what_next == ThreadRelocated) {
3343 target = target->link;
3344 ASSERT(get_itbl(target)->type == TSO);
3347 ASSERT(target->blocked_exceptions != NULL);
3349 last = (StgBlockingQueueElement **)&target->blocked_exceptions;
3350 for (t = (StgBlockingQueueElement *)target->blocked_exceptions;
3352 last = &t->link, t = t->link) {
3353 ASSERT(get_itbl(t)->type == TSO);
3354 if (t == (StgBlockingQueueElement *)tso) {
3355 *last = (StgBlockingQueueElement *)tso->link;
3359 barf("unblockThread (Exception): TSO not found");
3363 case BlockedOnWrite:
3364 #if defined(mingw32_HOST_OS)
3365 case BlockedOnDoProc:
3368 /* take TSO off blocked_queue */
3369 StgBlockingQueueElement *prev = NULL;
3370 for (t = (StgBlockingQueueElement *)blocked_queue_hd; t != END_BQ_QUEUE;
3371 prev = t, t = t->link) {
3372 if (t == (StgBlockingQueueElement *)tso) {
3374 blocked_queue_hd = (StgTSO *)t->link;
3375 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
3376 blocked_queue_tl = END_TSO_QUEUE;
3379 prev->link = t->link;
3380 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
3381 blocked_queue_tl = (StgTSO *)prev;
3384 #if defined(mingw32_HOST_OS)
3385 /* (Cooperatively) signal that the worker thread should abort
3388 abandonWorkRequest(tso->block_info.async_result->reqID);
3393 barf("unblockThread (I/O): TSO not found");
3396 case BlockedOnDelay:
3398 /* take TSO off sleeping_queue */
3399 StgBlockingQueueElement *prev = NULL;
3400 for (t = (StgBlockingQueueElement *)sleeping_queue; t != END_BQ_QUEUE;
3401 prev = t, t = t->link) {
3402 if (t == (StgBlockingQueueElement *)tso) {
3404 sleeping_queue = (StgTSO *)t->link;
3406 prev->link = t->link;
3411 barf("unblockThread (delay): TSO not found");
3415 barf("unblockThread");
3419 tso->link = END_TSO_QUEUE;
3420 tso->why_blocked = NotBlocked;
3421 tso->block_info.closure = NULL;
3422 PUSH_ON_RUN_QUEUE(tso);
3426 unblockThread(StgTSO *tso)
3430 /* To avoid locking unnecessarily. */
3431 if (tso->why_blocked == NotBlocked) {
3435 switch (tso->why_blocked) {
3438 // Be careful: nothing to do here! We tell the scheduler that the thread
3439 // is runnable and we leave it to the stack-walking code to abort the
3440 // transaction while unwinding the stack. We should perhaps have a debugging
3441 // test to make sure that this really happens and that the 'zombie' transaction
3442 // does not get committed.
3446 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
3448 StgTSO *last_tso = END_TSO_QUEUE;
3449 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
3452 for (t = mvar->head; t != END_TSO_QUEUE;
3453 last = &t->link, last_tso = t, t = t->link) {
3456 if (mvar->tail == tso) {
3457 mvar->tail = last_tso;
3462 barf("unblockThread (MVAR): TSO not found");
3465 case BlockedOnBlackHole:
3467 last = &blackhole_queue;
3468 for (t = blackhole_queue; t != END_TSO_QUEUE;
3469 last = &t->link, t = t->link) {
3475 barf("unblockThread (BLACKHOLE): TSO not found");
3478 case BlockedOnException:
3480 StgTSO *target = tso->block_info.tso;
3482 ASSERT(get_itbl(target)->type == TSO);
3484 while (target->what_next == ThreadRelocated) {
3485 target = target->link;
3486 ASSERT(get_itbl(target)->type == TSO);
3489 ASSERT(target->blocked_exceptions != NULL);
3491 last = &target->blocked_exceptions;
3492 for (t = target->blocked_exceptions; t != END_TSO_QUEUE;
3493 last = &t->link, t = t->link) {
3494 ASSERT(get_itbl(t)->type == TSO);
3500 barf("unblockThread (Exception): TSO not found");
3504 case BlockedOnWrite:
3505 #if defined(mingw32_HOST_OS)
3506 case BlockedOnDoProc:
3509 StgTSO *prev = NULL;
3510 for (t = blocked_queue_hd; t != END_TSO_QUEUE;
3511 prev = t, t = t->link) {
3514 blocked_queue_hd = t->link;
3515 if (blocked_queue_tl == t) {
3516 blocked_queue_tl = END_TSO_QUEUE;
3519 prev->link = t->link;
3520 if (blocked_queue_tl == t) {
3521 blocked_queue_tl = prev;
3524 #if defined(mingw32_HOST_OS)
3525 /* (Cooperatively) signal that the worker thread should abort
3528 abandonWorkRequest(tso->block_info.async_result->reqID);
3533 barf("unblockThread (I/O): TSO not found");
3536 case BlockedOnDelay:
3538 StgTSO *prev = NULL;
3539 for (t = sleeping_queue; t != END_TSO_QUEUE;
3540 prev = t, t = t->link) {
3543 sleeping_queue = t->link;
3545 prev->link = t->link;
3550 barf("unblockThread (delay): TSO not found");
3554 barf("unblockThread");
3558 tso->link = END_TSO_QUEUE;
3559 tso->why_blocked = NotBlocked;
3560 tso->block_info.closure = NULL;
3561 APPEND_TO_RUN_QUEUE(tso);
3565 /* -----------------------------------------------------------------------------
3568 * Check the blackhole_queue for threads that can be woken up. We do
3569 * this periodically: before every GC, and whenever the run queue is
3572 * An elegant solution might be to just wake up all the blocked
3573 * threads with awakenBlockedQueue occasionally: they'll go back to
3574 * sleep again if the object is still a BLACKHOLE. Unfortunately this
3575 * doesn't give us a way to tell whether we've actually managed to
3576 * wake up any threads, so we would be busy-waiting.
3578 * -------------------------------------------------------------------------- */
3581 checkBlackHoles( void )
3584 rtsBool any_woke_up = rtsFalse;
3587 IF_DEBUG(scheduler, sched_belch("checking threads blocked on black holes"));
3589 // ASSUMES: sched_mutex
3590 prev = &blackhole_queue;
3591 t = blackhole_queue;
3592 while (t != END_TSO_QUEUE) {
3593 ASSERT(t->why_blocked == BlockedOnBlackHole);
3594 type = get_itbl(t->block_info.closure)->type;
3595 if (type != BLACKHOLE && type != CAF_BLACKHOLE) {
3596 t = unblockOneLocked(t);
3598 any_woke_up = rtsTrue;
3608 /* -----------------------------------------------------------------------------
3611 * The following function implements the magic for raising an
3612 * asynchronous exception in an existing thread.
3614 * We first remove the thread from any queue on which it might be
3615 * blocked. The possible blockages are MVARs and BLACKHOLE_BQs.
3617 * We strip the stack down to the innermost CATCH_FRAME, building
3618 * thunks in the heap for all the active computations, so they can
3619 * be restarted if necessary. When we reach a CATCH_FRAME, we build
3620 * an application of the handler to the exception, and push it on
3621 * the top of the stack.
3623 * How exactly do we save all the active computations? We create an
3624 * AP_STACK for every UpdateFrame on the stack. Entering one of these
3625 * AP_STACKs pushes everything from the corresponding update frame
3626 * upwards onto the stack. (Actually, it pushes everything up to the
3627 * next update frame plus a pointer to the next AP_STACK object.
3628 * Entering the next AP_STACK object pushes more onto the stack until we
3629 * reach the last AP_STACK object - at which point the stack should look
3630 * exactly as it did when we killed the TSO and we can continue
3631 * execution by entering the closure on top of the stack.
3633 * We can also kill a thread entirely - this happens if either (a) the
3634 * exception passed to raiseAsync is NULL, or (b) there's no
3635 * CATCH_FRAME on the stack. In either case, we strip the entire
3636 * stack and replace the thread with a zombie.
3638 * Locks: sched_mutex held upon entry nor exit.
3640 * -------------------------------------------------------------------------- */
3643 deleteThread(StgTSO *tso)
3645 if (tso->why_blocked != BlockedOnCCall &&
3646 tso->why_blocked != BlockedOnCCall_NoUnblockExc) {
3647 raiseAsync(tso,NULL);
3651 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
3653 deleteThreadImmediately(StgTSO *tso)
3654 { // for forkProcess only:
3655 // delete thread without giving it a chance to catch the KillThread exception
3657 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
3661 if (tso->why_blocked != BlockedOnCCall &&
3662 tso->why_blocked != BlockedOnCCall_NoUnblockExc) {
3666 tso->what_next = ThreadKilled;
3671 raiseAsyncWithLock(StgTSO *tso, StgClosure *exception)
3673 /* When raising async exs from contexts where sched_mutex isn't held;
3674 use raiseAsyncWithLock(). */
3675 ACQUIRE_LOCK(&sched_mutex);
3676 raiseAsync(tso,exception);
3677 RELEASE_LOCK(&sched_mutex);
3681 raiseAsync(StgTSO *tso, StgClosure *exception)
3683 raiseAsync_(tso, exception, rtsFalse);
3687 raiseAsync_(StgTSO *tso, StgClosure *exception, rtsBool stop_at_atomically)
3689 StgRetInfoTable *info;
3692 // Thread already dead?
3693 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
3698 sched_belch("raising exception in thread %ld.", (long)tso->id));
3700 // Remove it from any blocking queues
3705 // The stack freezing code assumes there's a closure pointer on
3706 // the top of the stack, so we have to arrange that this is the case...
3708 if (sp[0] == (W_)&stg_enter_info) {
3712 sp[0] = (W_)&stg_dummy_ret_closure;
3718 // 1. Let the top of the stack be the "current closure"
3720 // 2. Walk up the stack until we find either an UPDATE_FRAME or a
3723 // 3. If it's an UPDATE_FRAME, then make an AP_STACK containing the
3724 // current closure applied to the chunk of stack up to (but not
3725 // including) the update frame. This closure becomes the "current
3726 // closure". Go back to step 2.
3728 // 4. If it's a CATCH_FRAME, then leave the exception handler on
3729 // top of the stack applied to the exception.
3731 // 5. If it's a STOP_FRAME, then kill the thread.
3733 // NB: if we pass an ATOMICALLY_FRAME then abort the associated
3740 info = get_ret_itbl((StgClosure *)frame);
3742 while (info->i.type != UPDATE_FRAME
3743 && (info->i.type != CATCH_FRAME || exception == NULL)
3744 && info->i.type != STOP_FRAME
3745 && (info->i.type != ATOMICALLY_FRAME || stop_at_atomically == rtsFalse))
3747 if (info->i.type == CATCH_RETRY_FRAME || info->i.type == ATOMICALLY_FRAME) {
3748 // IF we find an ATOMICALLY_FRAME then we abort the
3749 // current transaction and propagate the exception. In
3750 // this case (unlike ordinary exceptions) we do not care
3751 // whether the transaction is valid or not because its
3752 // possible validity cannot have caused the exception
3753 // and will not be visible after the abort.
3755 debugBelch("Found atomically block delivering async exception\n"));
3756 stmAbortTransaction(tso -> trec);
3757 tso -> trec = stmGetEnclosingTRec(tso -> trec);
3759 frame += stack_frame_sizeW((StgClosure *)frame);
3760 info = get_ret_itbl((StgClosure *)frame);
3763 switch (info->i.type) {
3765 case ATOMICALLY_FRAME:
3766 ASSERT(stop_at_atomically);
3767 ASSERT(stmGetEnclosingTRec(tso->trec) == NO_TREC);
3768 stmCondemnTransaction(tso -> trec);
3772 // R1 is not a register: the return convention for IO in
3773 // this case puts the return value on the stack, so we
3774 // need to set up the stack to return to the atomically
3775 // frame properly...
3776 tso->sp = frame - 2;
3777 tso->sp[1] = (StgWord) &stg_NO_FINALIZER_closure; // why not?
3778 tso->sp[0] = (StgWord) &stg_ut_1_0_unreg_info;
3780 tso->what_next = ThreadRunGHC;
3784 // If we find a CATCH_FRAME, and we've got an exception to raise,
3785 // then build the THUNK raise(exception), and leave it on
3786 // top of the CATCH_FRAME ready to enter.
3790 StgCatchFrame *cf = (StgCatchFrame *)frame;
3794 // we've got an exception to raise, so let's pass it to the
3795 // handler in this frame.
3797 raise = (StgThunk *)allocate(sizeofW(StgThunk)+1);
3798 TICK_ALLOC_SE_THK(1,0);
3799 SET_HDR(raise,&stg_raise_info,cf->header.prof.ccs);
3800 raise->payload[0] = exception;
3802 // throw away the stack from Sp up to the CATCH_FRAME.
3806 /* Ensure that async excpetions are blocked now, so we don't get
3807 * a surprise exception before we get around to executing the
3810 if (tso->blocked_exceptions == NULL) {
3811 tso->blocked_exceptions = END_TSO_QUEUE;
3814 /* Put the newly-built THUNK on top of the stack, ready to execute
3815 * when the thread restarts.
3818 sp[-1] = (W_)&stg_enter_info;
3820 tso->what_next = ThreadRunGHC;
3821 IF_DEBUG(sanity, checkTSO(tso));
3830 // First build an AP_STACK consisting of the stack chunk above the
3831 // current update frame, with the top word on the stack as the
3834 words = frame - sp - 1;
3835 ap = (StgAP_STACK *)allocate(AP_STACK_sizeW(words));
3838 ap->fun = (StgClosure *)sp[0];
3840 for(i=0; i < (nat)words; ++i) {
3841 ap->payload[i] = (StgClosure *)*sp++;
3844 SET_HDR(ap,&stg_AP_STACK_info,
3845 ((StgClosure *)frame)->header.prof.ccs /* ToDo */);
3846 TICK_ALLOC_UP_THK(words+1,0);
3849 debugBelch("sched: Updating ");
3850 printPtr((P_)((StgUpdateFrame *)frame)->updatee);
3851 debugBelch(" with ");
3852 printObj((StgClosure *)ap);
3855 // Replace the updatee with an indirection - happily
3856 // this will also wake up any threads currently
3857 // waiting on the result.
3859 // Warning: if we're in a loop, more than one update frame on
3860 // the stack may point to the same object. Be careful not to
3861 // overwrite an IND_OLDGEN in this case, because we'll screw
3862 // up the mutable lists. To be on the safe side, don't
3863 // overwrite any kind of indirection at all. See also
3864 // threadSqueezeStack in GC.c, where we have to make a similar
3867 if (!closure_IND(((StgUpdateFrame *)frame)->updatee)) {
3868 // revert the black hole
3869 UPD_IND_NOLOCK(((StgUpdateFrame *)frame)->updatee,
3872 sp += sizeofW(StgUpdateFrame) - 1;
3873 sp[0] = (W_)ap; // push onto stack
3878 // We've stripped the entire stack, the thread is now dead.
3879 sp += sizeofW(StgStopFrame);
3880 tso->what_next = ThreadKilled;
3891 /* -----------------------------------------------------------------------------
3892 raiseExceptionHelper
3894 This function is called by the raise# primitve, just so that we can
3895 move some of the tricky bits of raising an exception from C-- into
3896 C. Who knows, it might be a useful re-useable thing here too.
3897 -------------------------------------------------------------------------- */
3900 raiseExceptionHelper (StgTSO *tso, StgClosure *exception)
3902 StgThunk *raise_closure = NULL;
3904 StgRetInfoTable *info;
3906 // This closure represents the expression 'raise# E' where E
3907 // is the exception raise. It is used to overwrite all the
3908 // thunks which are currently under evaluataion.
3912 // LDV profiling: stg_raise_info has THUNK as its closure
3913 // type. Since a THUNK takes at least MIN_UPD_SIZE words in its
3914 // payload, MIN_UPD_SIZE is more approprate than 1. It seems that
3915 // 1 does not cause any problem unless profiling is performed.
3916 // However, when LDV profiling goes on, we need to linearly scan
3917 // small object pool, where raise_closure is stored, so we should
3918 // use MIN_UPD_SIZE.
3920 // raise_closure = (StgClosure *)RET_STGCALL1(P_,allocate,
3921 // sizeofW(StgClosure)+1);
3925 // Walk up the stack, looking for the catch frame. On the way,
3926 // we update any closures pointed to from update frames with the
3927 // raise closure that we just built.
3931 info = get_ret_itbl((StgClosure *)p);
3932 next = p + stack_frame_sizeW((StgClosure *)p);
3933 switch (info->i.type) {
3936 // Only create raise_closure if we need to.
3937 if (raise_closure == NULL) {
3939 (StgThunk *)allocate(sizeofW(StgThunk)+MIN_UPD_SIZE);
3940 SET_HDR(raise_closure, &stg_raise_info, CCCS);
3941 raise_closure->payload[0] = exception;
3943 UPD_IND(((StgUpdateFrame *)p)->updatee,(StgClosure *)raise_closure);
3947 case ATOMICALLY_FRAME:
3948 IF_DEBUG(stm, debugBelch("Found ATOMICALLY_FRAME at %p\n", p));
3950 return ATOMICALLY_FRAME;
3956 case CATCH_STM_FRAME:
3957 IF_DEBUG(stm, debugBelch("Found CATCH_STM_FRAME at %p\n", p));
3959 return CATCH_STM_FRAME;
3965 case CATCH_RETRY_FRAME:
3974 /* -----------------------------------------------------------------------------
3975 findRetryFrameHelper
3977 This function is called by the retry# primitive. It traverses the stack
3978 leaving tso->sp referring to the frame which should handle the retry.
3980 This should either be a CATCH_RETRY_FRAME (if the retry# is within an orElse#)
3981 or should be a ATOMICALLY_FRAME (if the retry# reaches the top level).
3983 We skip CATCH_STM_FRAMEs because retries are not considered to be exceptions,
3984 despite the similar implementation.
3986 We should not expect to see CATCH_FRAME or STOP_FRAME because those should
3987 not be created within memory transactions.
3988 -------------------------------------------------------------------------- */
3991 findRetryFrameHelper (StgTSO *tso)
3994 StgRetInfoTable *info;
3998 info = get_ret_itbl((StgClosure *)p);
3999 next = p + stack_frame_sizeW((StgClosure *)p);
4000 switch (info->i.type) {
4002 case ATOMICALLY_FRAME:
4003 IF_DEBUG(stm, debugBelch("Found ATOMICALLY_FRAME at %p during retrry\n", p));
4005 return ATOMICALLY_FRAME;
4007 case CATCH_RETRY_FRAME:
4008 IF_DEBUG(stm, debugBelch("Found CATCH_RETRY_FRAME at %p during retrry\n", p));
4010 return CATCH_RETRY_FRAME;
4012 case CATCH_STM_FRAME:
4014 ASSERT(info->i.type != CATCH_FRAME);
4015 ASSERT(info->i.type != STOP_FRAME);
4022 /* -----------------------------------------------------------------------------
4023 resurrectThreads is called after garbage collection on the list of
4024 threads found to be garbage. Each of these threads will be woken
4025 up and sent a signal: BlockedOnDeadMVar if the thread was blocked
4026 on an MVar, or NonTermination if the thread was blocked on a Black
4029 Locks: sched_mutex isn't held upon entry nor exit.
4030 -------------------------------------------------------------------------- */
4033 resurrectThreads( StgTSO *threads )
4037 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
4038 next = tso->global_link;
4039 tso->global_link = all_threads;
4041 IF_DEBUG(scheduler, sched_belch("resurrecting thread %d", tso->id));
4043 switch (tso->why_blocked) {
4045 case BlockedOnException:
4046 /* Called by GC - sched_mutex lock is currently held. */
4047 raiseAsync(tso,(StgClosure *)BlockedOnDeadMVar_closure);
4049 case BlockedOnBlackHole:
4050 raiseAsync(tso,(StgClosure *)NonTermination_closure);
4053 raiseAsync(tso,(StgClosure *)BlockedIndefinitely_closure);
4056 /* This might happen if the thread was blocked on a black hole
4057 * belonging to a thread that we've just woken up (raiseAsync
4058 * can wake up threads, remember...).
4062 barf("resurrectThreads: thread blocked in a strange way");
4067 /* ----------------------------------------------------------------------------
4068 * Debugging: why is a thread blocked
4069 * [Also provides useful information when debugging threaded programs
4070 * at the Haskell source code level, so enable outside of DEBUG. --sof 7/02]
4071 ------------------------------------------------------------------------- */
4074 printThreadBlockage(StgTSO *tso)
4076 switch (tso->why_blocked) {
4078 debugBelch("is blocked on read from fd %d", (int)(tso->block_info.fd));
4080 case BlockedOnWrite:
4081 debugBelch("is blocked on write to fd %d", (int)(tso->block_info.fd));
4083 #if defined(mingw32_HOST_OS)
4084 case BlockedOnDoProc:
4085 debugBelch("is blocked on proc (request: %ld)", tso->block_info.async_result->reqID);
4088 case BlockedOnDelay:
4089 debugBelch("is blocked until %ld", (long)(tso->block_info.target));
4092 debugBelch("is blocked on an MVar @ %p", tso->block_info.closure);
4094 case BlockedOnException:
4095 debugBelch("is blocked on delivering an exception to thread %d",
4096 tso->block_info.tso->id);
4098 case BlockedOnBlackHole:
4099 debugBelch("is blocked on a black hole");
4102 debugBelch("is not blocked");
4104 #if defined(PARALLEL_HASKELL)
4106 debugBelch("is blocked on global address; local FM_BQ is %p (%s)",
4107 tso->block_info.closure, info_type(tso->block_info.closure));
4109 case BlockedOnGA_NoSend:
4110 debugBelch("is blocked on global address (no send); local FM_BQ is %p (%s)",
4111 tso->block_info.closure, info_type(tso->block_info.closure));
4114 case BlockedOnCCall:
4115 debugBelch("is blocked on an external call");
4117 case BlockedOnCCall_NoUnblockExc:
4118 debugBelch("is blocked on an external call (exceptions were already blocked)");
4121 debugBelch("is blocked on an STM operation");
4124 barf("printThreadBlockage: strange tso->why_blocked: %d for TSO %d (%d)",
4125 tso->why_blocked, tso->id, tso);
4130 printThreadStatus(StgTSO *tso)
4132 switch (tso->what_next) {
4134 debugBelch("has been killed");
4136 case ThreadComplete:
4137 debugBelch("has completed");
4140 printThreadBlockage(tso);
4145 printAllThreads(void)
4150 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
4151 ullong_format_string(TIME_ON_PROC(CurrentProc),
4152 time_string, rtsFalse/*no commas!*/);
4154 debugBelch("all threads at [%s]:\n", time_string);
4155 # elif defined(PARALLEL_HASKELL)
4156 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
4157 ullong_format_string(CURRENT_TIME,
4158 time_string, rtsFalse/*no commas!*/);
4160 debugBelch("all threads at [%s]:\n", time_string);
4162 debugBelch("all threads:\n");
4165 for (t = all_threads; t != END_TSO_QUEUE; ) {
4166 debugBelch("\tthread %4d @ %p ", t->id, (void *)t);
4169 void *label = lookupThreadLabel(t->id);
4170 if (label) debugBelch("[\"%s\"] ",(char *)label);
4173 if (t->what_next == ThreadRelocated) {
4174 debugBelch("has been relocated...\n");
4177 printThreadStatus(t);
4188 printThreadQueue(StgTSO *t)
4191 for (; t != END_TSO_QUEUE; t = t->link) {
4192 debugBelch("\tthread %d @ %p ", t->id, (void *)t);
4193 if (t->what_next == ThreadRelocated) {
4194 debugBelch("has been relocated...\n");
4196 printThreadStatus(t);
4201 debugBelch("%d threads on queue\n", i);
4205 Print a whole blocking queue attached to node (debugging only).
4207 # if defined(PARALLEL_HASKELL)
4209 print_bq (StgClosure *node)
4211 StgBlockingQueueElement *bqe;
4215 debugBelch("## BQ of closure %p (%s): ",
4216 node, info_type(node));
4218 /* should cover all closures that may have a blocking queue */
4219 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
4220 get_itbl(node)->type == FETCH_ME_BQ ||
4221 get_itbl(node)->type == RBH ||
4222 get_itbl(node)->type == MVAR);
4224 ASSERT(node!=(StgClosure*)NULL); // sanity check
4226 print_bqe(((StgBlockingQueue*)node)->blocking_queue);
4230 Print a whole blocking queue starting with the element bqe.
4233 print_bqe (StgBlockingQueueElement *bqe)
4238 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
4240 for (end = (bqe==END_BQ_QUEUE);
4241 !end; // iterate until bqe points to a CONSTR
4242 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE),
4243 bqe = end ? END_BQ_QUEUE : bqe->link) {
4244 ASSERT(bqe != END_BQ_QUEUE); // sanity check
4245 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
4246 /* types of closures that may appear in a blocking queue */
4247 ASSERT(get_itbl(bqe)->type == TSO ||
4248 get_itbl(bqe)->type == BLOCKED_FETCH ||
4249 get_itbl(bqe)->type == CONSTR);
4250 /* only BQs of an RBH end with an RBH_Save closure */
4251 //ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
4253 switch (get_itbl(bqe)->type) {
4255 debugBelch(" TSO %u (%x),",
4256 ((StgTSO *)bqe)->id, ((StgTSO *)bqe));
4259 debugBelch(" BF (node=%p, ga=((%x, %d, %x)),",
4260 ((StgBlockedFetch *)bqe)->node,
4261 ((StgBlockedFetch *)bqe)->ga.payload.gc.gtid,
4262 ((StgBlockedFetch *)bqe)->ga.payload.gc.slot,
4263 ((StgBlockedFetch *)bqe)->ga.weight);
4266 debugBelch(" %s (IP %p),",
4267 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
4268 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
4269 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
4270 "RBH_Save_?"), get_itbl(bqe));
4273 barf("Unexpected closure type %s in blocking queue", // of %p (%s)",
4274 info_type((StgClosure *)bqe)); // , node, info_type(node));
4280 # elif defined(GRAN)
4282 print_bq (StgClosure *node)
4284 StgBlockingQueueElement *bqe;
4285 PEs node_loc, tso_loc;
4288 /* should cover all closures that may have a blocking queue */
4289 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
4290 get_itbl(node)->type == FETCH_ME_BQ ||
4291 get_itbl(node)->type == RBH);
4293 ASSERT(node!=(StgClosure*)NULL); // sanity check
4294 node_loc = where_is(node);
4296 debugBelch("## BQ of closure %p (%s) on [PE %d]: ",
4297 node, info_type(node), node_loc);
4300 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
4302 for (bqe = ((StgBlockingQueue*)node)->blocking_queue, end = (bqe==END_BQ_QUEUE);
4303 !end; // iterate until bqe points to a CONSTR
4304 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) {
4305 ASSERT(bqe != END_BQ_QUEUE); // sanity check
4306 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
4307 /* types of closures that may appear in a blocking queue */
4308 ASSERT(get_itbl(bqe)->type == TSO ||
4309 get_itbl(bqe)->type == CONSTR);
4310 /* only BQs of an RBH end with an RBH_Save closure */
4311 ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
4313 tso_loc = where_is((StgClosure *)bqe);
4314 switch (get_itbl(bqe)->type) {
4316 debugBelch(" TSO %d (%p) on [PE %d],",
4317 ((StgTSO *)bqe)->id, (StgTSO *)bqe, tso_loc);
4320 debugBelch(" %s (IP %p),",
4321 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
4322 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
4323 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
4324 "RBH_Save_?"), get_itbl(bqe));
4327 barf("Unexpected closure type %s in blocking queue of %p (%s)",
4328 info_type((StgClosure *)bqe), node, info_type(node));
4336 #if defined(PARALLEL_HASKELL)
4343 for (i=0, tso=run_queue_hd;
4344 tso != END_TSO_QUEUE;
4353 sched_belch(char *s, ...)
4357 #ifdef RTS_SUPPORTS_THREADS
4358 debugBelch("sched (task %p): ", osThreadId());
4359 #elif defined(PARALLEL_HASKELL)
4362 debugBelch("sched: ");