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 /* if this flag is set as well, give up execution */
178 rtsBool interrupted = rtsFalse;
180 /* Next thread ID to allocate.
181 * Locks required: thread_id_mutex
183 static StgThreadID next_thread_id = 1;
186 * Pointers to the state of the current thread.
187 * Rule of thumb: if CurrentTSO != NULL, then we're running a Haskell
188 * thread. If CurrentTSO == NULL, then we're at the scheduler level.
191 /* The smallest stack size that makes any sense is:
192 * RESERVED_STACK_WORDS (so we can get back from the stack overflow)
193 * + sizeofW(StgStopFrame) (the stg_stop_thread_info frame)
194 * + 1 (the closure to enter)
196 * + 1 (spare slot req'd by stg_ap_v_ret)
198 * A thread with this stack will bomb immediately with a stack
199 * overflow, which will increase its stack size.
202 #define MIN_STACK_WORDS (RESERVED_STACK_WORDS + sizeofW(StgStopFrame) + 3)
209 /* This is used in `TSO.h' and gcc 2.96 insists that this variable actually
210 * exists - earlier gccs apparently didn't.
216 * Set to TRUE when entering a shutdown state (via shutdownHaskellAndExit()) --
217 * in an MT setting, needed to signal that a worker thread shouldn't hang around
218 * in the scheduler when it is out of work.
220 static rtsBool shutting_down_scheduler = rtsFalse;
222 #if defined(RTS_SUPPORTS_THREADS)
223 /* ToDo: carefully document the invariants that go together
224 * with these synchronisation objects.
226 Mutex sched_mutex = INIT_MUTEX_VAR;
227 Mutex term_mutex = INIT_MUTEX_VAR;
229 #endif /* RTS_SUPPORTS_THREADS */
231 #if defined(PARALLEL_HASKELL)
233 rtsTime TimeOfLastYield;
234 rtsBool emitSchedule = rtsTrue;
238 static char *whatNext_strs[] = {
248 /* -----------------------------------------------------------------------------
249 * static function prototypes
250 * -------------------------------------------------------------------------- */
252 #if defined(RTS_SUPPORTS_THREADS)
253 static void taskStart(void);
256 static void schedule( StgMainThread *mainThread USED_WHEN_RTS_SUPPORTS_THREADS,
257 Capability *initialCapability );
260 // These function all encapsulate parts of the scheduler loop, and are
261 // abstracted only to make the structure and control flow of the
262 // scheduler clearer.
264 static void schedulePreLoop(void);
265 static void scheduleStartSignalHandlers(void);
266 static void scheduleCheckBlockedThreads(void);
267 static void scheduleCheckBlackHoles(void);
268 static void scheduleDetectDeadlock(void);
270 static StgTSO *scheduleProcessEvent(rtsEvent *event);
272 #if defined(PARALLEL_HASKELL)
273 static StgTSO *scheduleSendPendingMessages(void);
274 static void scheduleActivateSpark(void);
275 static rtsBool scheduleGetRemoteWork(rtsBool *receivedFinish);
277 #if defined(PAR) || defined(GRAN)
278 static void scheduleGranParReport(void);
280 static void schedulePostRunThread(void);
281 static rtsBool scheduleHandleHeapOverflow( Capability *cap, StgTSO *t );
282 static void scheduleHandleStackOverflow( StgTSO *t);
283 static rtsBool scheduleHandleYield( StgTSO *t, nat prev_what_next );
284 static void scheduleHandleThreadBlocked( StgTSO *t );
285 static rtsBool scheduleHandleThreadFinished( StgMainThread *mainThread,
286 Capability *cap, StgTSO *t );
287 static rtsBool scheduleDoHeapProfile(rtsBool ready_to_gc);
288 static void scheduleDoGC(Capability *cap);
290 static void unblockThread(StgTSO *tso);
291 static rtsBool checkBlackHoles(void);
292 static SchedulerStatus waitThread_(/*out*/StgMainThread* m,
293 Capability *initialCapability
295 static void scheduleThread_ (StgTSO* tso);
296 static void AllRoots(evac_fn evac);
298 static StgTSO *threadStackOverflow(StgTSO *tso);
300 static void raiseAsync_(StgTSO *tso, StgClosure *exception,
301 rtsBool stop_at_atomically);
303 static void printThreadBlockage(StgTSO *tso);
304 static void printThreadStatus(StgTSO *tso);
306 #if defined(PARALLEL_HASKELL)
307 StgTSO * createSparkThread(rtsSpark spark);
308 StgTSO * activateSpark (rtsSpark spark);
311 /* ----------------------------------------------------------------------------
313 * ------------------------------------------------------------------------- */
315 #if defined(RTS_SUPPORTS_THREADS)
316 static nat startingWorkerThread = 0;
321 ACQUIRE_LOCK(&sched_mutex);
322 startingWorkerThread--;
325 RELEASE_LOCK(&sched_mutex);
329 startSchedulerTaskIfNecessary(void)
331 if ( !EMPTY_RUN_QUEUE()
332 && !shutting_down_scheduler // not if we're shutting down
333 && startingWorkerThread==0)
335 // we don't want to start another worker thread
336 // just because the last one hasn't yet reached the
337 // "waiting for capability" state
338 startingWorkerThread++;
339 if (!maybeStartNewWorker(taskStart)) {
340 startingWorkerThread--;
346 /* -----------------------------------------------------------------------------
347 * Putting a thread on the run queue: different scheduling policies
348 * -------------------------------------------------------------------------- */
351 addToRunQueue( StgTSO *t )
353 #if defined(PARALLEL_HASKELL)
354 if (RtsFlags.ParFlags.doFairScheduling) {
355 // this does round-robin scheduling; good for concurrency
356 APPEND_TO_RUN_QUEUE(t);
358 // this does unfair scheduling; good for parallelism
359 PUSH_ON_RUN_QUEUE(t);
362 // this does round-robin scheduling; good for concurrency
363 APPEND_TO_RUN_QUEUE(t);
367 /* ---------------------------------------------------------------------------
368 Main scheduling loop.
370 We use round-robin scheduling, each thread returning to the
371 scheduler loop when one of these conditions is detected:
374 * timer expires (thread yields)
379 Locking notes: we acquire the scheduler lock once at the beginning
380 of the scheduler loop, and release it when
382 * running a thread, or
383 * waiting for work, or
384 * waiting for a GC to complete.
387 In a GranSim setup this loop iterates over the global event queue.
388 This revolves around the global event queue, which determines what
389 to do next. Therefore, it's more complicated than either the
390 concurrent or the parallel (GUM) setup.
393 GUM iterates over incoming messages.
394 It starts with nothing to do (thus CurrentTSO == END_TSO_QUEUE),
395 and sends out a fish whenever it has nothing to do; in-between
396 doing the actual reductions (shared code below) it processes the
397 incoming messages and deals with delayed operations
398 (see PendingFetches).
399 This is not the ugliest code you could imagine, but it's bloody close.
401 ------------------------------------------------------------------------ */
404 schedule( StgMainThread *mainThread USED_WHEN_RTS_SUPPORTS_THREADS,
405 Capability *initialCapability )
409 StgThreadReturnCode ret;
412 #elif defined(PARALLEL_HASKELL)
415 rtsBool receivedFinish = rtsFalse;
417 nat tp_size, sp_size; // stats only
423 // Pre-condition: sched_mutex is held.
424 // We might have a capability, passed in as initialCapability.
425 cap = initialCapability;
427 #if !defined(RTS_SUPPORTS_THREADS)
428 // simply initialise it in the non-threaded case
429 grabCapability(&cap);
433 sched_belch("### NEW SCHEDULER LOOP (main thr: %p, cap: %p)",
434 mainThread, initialCapability);
439 // -----------------------------------------------------------
440 // Scheduler loop starts here:
442 #if defined(PARALLEL_HASKELL)
443 #define TERMINATION_CONDITION (!receivedFinish)
445 #define TERMINATION_CONDITION ((event = get_next_event()) != (rtsEvent*)NULL)
447 #define TERMINATION_CONDITION rtsTrue
450 while (TERMINATION_CONDITION) {
453 /* Choose the processor with the next event */
454 CurrentProc = event->proc;
455 CurrentTSO = event->tso;
458 IF_DEBUG(scheduler, printAllThreads());
460 #if defined(RTS_SUPPORTS_THREADS)
461 // Yield the capability to higher-priority tasks if necessary.
464 yieldCapability(&cap);
467 // If we do not currently hold a capability, we wait for one
470 waitForCapability(&sched_mutex, &cap,
471 mainThread ? &mainThread->bound_thread_cond : NULL);
474 // We now have a capability...
477 // Check whether we have re-entered the RTS from Haskell without
478 // going via suspendThread()/resumeThread (i.e. a 'safe' foreign
480 if (cap->r.rInHaskell) {
481 errorBelch("schedule: re-entered unsafely.\n"
482 " Perhaps a 'foreign import unsafe' should be 'safe'?");
487 // Test for interruption. If interrupted==rtsTrue, then either
488 // we received a keyboard interrupt (^C), or the scheduler is
489 // trying to shut down all the tasks (shutting_down_scheduler) in
493 if (shutting_down_scheduler) {
494 IF_DEBUG(scheduler, sched_belch("shutting down"));
495 releaseCapability(cap);
497 mainThread->stat = Interrupted;
498 mainThread->ret = NULL;
502 IF_DEBUG(scheduler, sched_belch("interrupted"));
507 #if defined(not_yet) && defined(SMP)
509 // Top up the run queue from our spark pool. We try to make the
510 // number of threads in the run queue equal to the number of
511 // free capabilities.
515 if (EMPTY_RUN_QUEUE()) {
516 spark = findSpark(rtsFalse);
518 break; /* no more sparks in the pool */
520 createSparkThread(spark);
522 sched_belch("==^^ turning spark of closure %p into a thread",
523 (StgClosure *)spark));
529 scheduleStartSignalHandlers();
531 // Only check the black holes here if we've nothing else to do.
532 // During normal execution, the black hole list only gets checked
533 // at GC time, to avoid repeatedly traversing this possibly long
534 // list each time around the scheduler.
535 if (EMPTY_RUN_QUEUE()) { scheduleCheckBlackHoles(); }
537 scheduleCheckBlockedThreads();
539 scheduleDetectDeadlock();
541 // Normally, the only way we can get here with no threads to
542 // run is if a keyboard interrupt received during
543 // scheduleCheckBlockedThreads() or scheduleDetectDeadlock().
544 // Additionally, it is not fatal for the
545 // threaded RTS to reach here with no threads to run.
547 // win32: might be here due to awaitEvent() being abandoned
548 // as a result of a console event having been delivered.
549 if ( EMPTY_RUN_QUEUE() ) {
550 #if !defined(RTS_SUPPORTS_THREADS) && !defined(mingw32_HOST_OS)
553 continue; // nothing to do
556 #if defined(PARALLEL_HASKELL)
557 scheduleSendPendingMessages();
558 if (EMPTY_RUN_QUEUE() && scheduleActivateSpark())
562 ASSERT(next_fish_to_send_at==0); // i.e. no delayed fishes left!
565 /* If we still have no work we need to send a FISH to get a spark
567 if (EMPTY_RUN_QUEUE()) {
568 if (!scheduleGetRemoteWork(&receivedFinish)) continue;
569 ASSERT(rtsFalse); // should not happen at the moment
571 // from here: non-empty run queue.
572 // TODO: merge above case with this, only one call processMessages() !
573 if (PacketsWaiting()) { /* process incoming messages, if
574 any pending... only in else
575 because getRemoteWork waits for
577 receivedFinish = processMessages();
582 scheduleProcessEvent(event);
586 // Get a thread to run
588 ASSERT(run_queue_hd != END_TSO_QUEUE);
591 #if defined(GRAN) || defined(PAR)
592 scheduleGranParReport(); // some kind of debuging output
594 // Sanity check the thread we're about to run. This can be
595 // expensive if there is lots of thread switching going on...
596 IF_DEBUG(sanity,checkTSO(t));
599 #if defined(RTS_SUPPORTS_THREADS)
600 // Check whether we can run this thread in the current task.
601 // If not, we have to pass our capability to the right task.
603 StgMainThread *m = t->main;
610 sched_belch("### Running thread %d in bound thread", t->id));
611 // yes, the Haskell thread is bound to the current native thread
616 sched_belch("### thread %d bound to another OS thread", t->id));
617 // no, bound to a different Haskell thread: pass to that thread
618 PUSH_ON_RUN_QUEUE(t);
619 passCapability(&m->bound_thread_cond);
625 if(mainThread != NULL)
626 // The thread we want to run is bound.
629 sched_belch("### this OS thread cannot run thread %d", t->id));
630 // no, the current native thread is bound to a different
631 // Haskell thread, so pass it to any worker thread
632 PUSH_ON_RUN_QUEUE(t);
633 passCapabilityToWorker();
640 cap->r.rCurrentTSO = t;
642 /* context switches are now initiated by the timer signal, unless
643 * the user specified "context switch as often as possible", with
646 if ((RtsFlags.ConcFlags.ctxtSwitchTicks == 0
647 && (run_queue_hd != END_TSO_QUEUE
648 || blocked_queue_hd != END_TSO_QUEUE
649 || sleeping_queue != END_TSO_QUEUE)))
654 RELEASE_LOCK(&sched_mutex);
656 IF_DEBUG(scheduler, sched_belch("-->> running thread %ld %s ...",
657 (long)t->id, whatNext_strs[t->what_next]));
659 #if defined(PROFILING)
660 startHeapProfTimer();
663 // ----------------------------------------------------------------------
664 // Run the current thread
666 prev_what_next = t->what_next;
668 errno = t->saved_errno;
669 cap->r.rInHaskell = rtsTrue;
671 switch (prev_what_next) {
675 /* Thread already finished, return to scheduler. */
676 ret = ThreadFinished;
680 ret = StgRun((StgFunPtr) stg_returnToStackTop, &cap->r);
683 case ThreadInterpret:
684 ret = interpretBCO(cap);
688 barf("schedule: invalid what_next field");
691 // We have run some Haskell code: there might be blackhole-blocked
692 // threads to wake up now.
693 if ( blackhole_queue != END_TSO_QUEUE ) {
694 blackholes_need_checking = rtsTrue;
697 cap->r.rInHaskell = rtsFalse;
699 // The TSO might have moved, eg. if it re-entered the RTS and a GC
700 // happened. So find the new location:
701 t = cap->r.rCurrentTSO;
703 // And save the current errno in this thread.
704 t->saved_errno = errno;
706 // ----------------------------------------------------------------------
708 /* Costs for the scheduler are assigned to CCS_SYSTEM */
709 #if defined(PROFILING)
714 ACQUIRE_LOCK(&sched_mutex);
716 #if defined(RTS_SUPPORTS_THREADS)
717 IF_DEBUG(scheduler,debugBelch("sched (task %p): ", osThreadId()););
718 #elif !defined(GRAN) && !defined(PARALLEL_HASKELL)
719 IF_DEBUG(scheduler,debugBelch("sched: "););
722 schedulePostRunThread();
724 ready_to_gc = rtsFalse;
728 ready_to_gc = scheduleHandleHeapOverflow(cap,t);
732 scheduleHandleStackOverflow(t);
736 if (scheduleHandleYield(t, prev_what_next)) {
737 // shortcut for switching between compiler/interpreter:
743 scheduleHandleThreadBlocked(t);
748 if (scheduleHandleThreadFinished(mainThread, cap, t)) return;;
752 barf("schedule: invalid thread return code %d", (int)ret);
755 if (scheduleDoHeapProfile(ready_to_gc)) { ready_to_gc = rtsFalse; }
756 if (ready_to_gc) { scheduleDoGC(cap); }
757 } /* end of while() */
759 IF_PAR_DEBUG(verbose,
760 debugBelch("== Leaving schedule() after having received Finish\n"));
763 /* ----------------------------------------------------------------------------
764 * Setting up the scheduler loop
765 * ASSUMES: sched_mutex
766 * ------------------------------------------------------------------------- */
769 schedulePreLoop(void)
772 /* set up first event to get things going */
773 /* ToDo: assign costs for system setup and init MainTSO ! */
774 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
776 CurrentTSO, (StgClosure*)NULL, (rtsSpark*)NULL);
779 debugBelch("GRAN: Init CurrentTSO (in schedule) = %p\n",
781 G_TSO(CurrentTSO, 5));
783 if (RtsFlags.GranFlags.Light) {
784 /* Save current time; GranSim Light only */
785 CurrentTSO->gran.clock = CurrentTime[CurrentProc];
790 /* ----------------------------------------------------------------------------
791 * Start any pending signal handlers
792 * ASSUMES: sched_mutex
793 * ------------------------------------------------------------------------- */
796 scheduleStartSignalHandlers(void)
798 #if defined(RTS_USER_SIGNALS) && !defined(RTS_SUPPORTS_THREADS)
799 if (signals_pending()) {
800 RELEASE_LOCK(&sched_mutex); /* ToDo: kill */
801 startSignalHandlers();
802 ACQUIRE_LOCK(&sched_mutex);
807 /* ----------------------------------------------------------------------------
808 * Check for blocked threads that can be woken up.
809 * ASSUMES: sched_mutex
810 * ------------------------------------------------------------------------- */
813 scheduleCheckBlockedThreads(void)
816 // Check whether any waiting threads need to be woken up. If the
817 // run queue is empty, and there are no other tasks running, we
818 // can wait indefinitely for something to happen.
820 if ( !EMPTY_QUEUE(blocked_queue_hd) || !EMPTY_QUEUE(sleeping_queue) )
822 #if defined(RTS_SUPPORTS_THREADS)
823 // We shouldn't be here...
824 barf("schedule: awaitEvent() in threaded RTS");
826 awaitEvent( EMPTY_RUN_QUEUE() && !blackholes_need_checking );
831 /* ----------------------------------------------------------------------------
832 * Check for threads blocked on BLACKHOLEs that can be woken up
833 * ASSUMES: sched_mutex
834 * ------------------------------------------------------------------------- */
836 scheduleCheckBlackHoles( void )
838 if ( blackholes_need_checking )
841 blackholes_need_checking = rtsFalse;
845 /* ----------------------------------------------------------------------------
846 * Detect deadlock conditions and attempt to resolve them.
847 * ASSUMES: sched_mutex
848 * ------------------------------------------------------------------------- */
851 scheduleDetectDeadlock(void)
854 * Detect deadlock: when we have no threads to run, there are no
855 * threads blocked, waiting for I/O, or sleeping, and all the
856 * other tasks are waiting for work, we must have a deadlock of
859 if ( EMPTY_THREAD_QUEUES() )
861 #if !defined(PARALLEL_HASKELL) && !defined(RTS_SUPPORTS_THREADS)
862 IF_DEBUG(scheduler, sched_belch("deadlocked, forcing major GC..."));
864 // Garbage collection can release some new threads due to
865 // either (a) finalizers or (b) threads resurrected because
866 // they are unreachable and will therefore be sent an
867 // exception. Any threads thus released will be immediately
869 GarbageCollect(GetRoots,rtsTrue);
870 if ( !EMPTY_RUN_QUEUE() ) return;
872 #if defined(RTS_USER_SIGNALS)
873 /* If we have user-installed signal handlers, then wait
874 * for signals to arrive rather then bombing out with a
877 if ( anyUserHandlers() ) {
879 sched_belch("still deadlocked, waiting for signals..."));
883 if (signals_pending()) {
884 RELEASE_LOCK(&sched_mutex);
885 startSignalHandlers();
886 ACQUIRE_LOCK(&sched_mutex);
889 // either we have threads to run, or we were interrupted:
890 ASSERT(!EMPTY_RUN_QUEUE() || interrupted);
894 /* Probably a real deadlock. Send the current main thread the
895 * Deadlock exception (or in the SMP build, send *all* main
896 * threads the deadlock exception, since none of them can make
902 switch (m->tso->why_blocked) {
904 case BlockedOnBlackHole:
905 case BlockedOnException:
907 raiseAsync(m->tso, (StgClosure *)NonTermination_closure);
910 barf("deadlock: main thread blocked in a strange way");
914 #elif defined(RTS_SUPPORTS_THREADS)
915 // ToDo: add deadlock detection in threaded RTS
916 #elif defined(PARALLEL_HASKELL)
917 // ToDo: add deadlock detection in GUM (similar to SMP) -- HWL
922 /* ----------------------------------------------------------------------------
923 * Process an event (GRAN only)
924 * ------------------------------------------------------------------------- */
928 scheduleProcessEvent(rtsEvent *event)
932 if (RtsFlags.GranFlags.Light)
933 GranSimLight_enter_system(event, &ActiveTSO); // adjust ActiveTSO etc
935 /* adjust time based on time-stamp */
936 if (event->time > CurrentTime[CurrentProc] &&
937 event->evttype != ContinueThread)
938 CurrentTime[CurrentProc] = event->time;
940 /* Deal with the idle PEs (may issue FindWork or MoveSpark events) */
941 if (!RtsFlags.GranFlags.Light)
944 IF_DEBUG(gran, debugBelch("GRAN: switch by event-type\n"));
946 /* main event dispatcher in GranSim */
947 switch (event->evttype) {
948 /* Should just be continuing execution */
950 IF_DEBUG(gran, debugBelch("GRAN: doing ContinueThread\n"));
951 /* ToDo: check assertion
952 ASSERT(run_queue_hd != (StgTSO*)NULL &&
953 run_queue_hd != END_TSO_QUEUE);
955 /* Ignore ContinueThreads for fetching threads (if synchr comm) */
956 if (!RtsFlags.GranFlags.DoAsyncFetch &&
957 procStatus[CurrentProc]==Fetching) {
958 debugBelch("ghuH: Spurious ContinueThread while Fetching ignored; TSO %d (%p) [PE %d]\n",
959 CurrentTSO->id, CurrentTSO, CurrentProc);
962 /* Ignore ContinueThreads for completed threads */
963 if (CurrentTSO->what_next == ThreadComplete) {
964 debugBelch("ghuH: found a ContinueThread event for completed thread %d (%p) [PE %d] (ignoring ContinueThread)\n",
965 CurrentTSO->id, CurrentTSO, CurrentProc);
968 /* Ignore ContinueThreads for threads that are being migrated */
969 if (PROCS(CurrentTSO)==Nowhere) {
970 debugBelch("ghuH: trying to run the migrating TSO %d (%p) [PE %d] (ignoring ContinueThread)\n",
971 CurrentTSO->id, CurrentTSO, CurrentProc);
974 /* The thread should be at the beginning of the run queue */
975 if (CurrentTSO!=run_queue_hds[CurrentProc]) {
976 debugBelch("ghuH: TSO %d (%p) [PE %d] is not at the start of the run_queue when doing a ContinueThread\n",
977 CurrentTSO->id, CurrentTSO, CurrentProc);
978 break; // run the thread anyway
981 new_event(proc, proc, CurrentTime[proc],
983 (StgTSO*)NULL, (StgClosure*)NULL, (rtsSpark*)NULL);
985 */ /* Catches superfluous CONTINUEs -- should be unnecessary */
986 break; // now actually run the thread; DaH Qu'vam yImuHbej
989 do_the_fetchnode(event);
990 goto next_thread; /* handle next event in event queue */
993 do_the_globalblock(event);
994 goto next_thread; /* handle next event in event queue */
997 do_the_fetchreply(event);
998 goto next_thread; /* handle next event in event queue */
1000 case UnblockThread: /* Move from the blocked queue to the tail of */
1001 do_the_unblock(event);
1002 goto next_thread; /* handle next event in event queue */
1004 case ResumeThread: /* Move from the blocked queue to the tail of */
1005 /* the runnable queue ( i.e. Qu' SImqa'lu') */
1006 event->tso->gran.blocktime +=
1007 CurrentTime[CurrentProc] - event->tso->gran.blockedat;
1008 do_the_startthread(event);
1009 goto next_thread; /* handle next event in event queue */
1012 do_the_startthread(event);
1013 goto next_thread; /* handle next event in event queue */
1016 do_the_movethread(event);
1017 goto next_thread; /* handle next event in event queue */
1020 do_the_movespark(event);
1021 goto next_thread; /* handle next event in event queue */
1024 do_the_findwork(event);
1025 goto next_thread; /* handle next event in event queue */
1028 barf("Illegal event type %u\n", event->evttype);
1031 /* This point was scheduler_loop in the old RTS */
1033 IF_DEBUG(gran, debugBelch("GRAN: after main switch\n"));
1035 TimeOfLastEvent = CurrentTime[CurrentProc];
1036 TimeOfNextEvent = get_time_of_next_event();
1037 IgnoreEvents=(TimeOfNextEvent==0); // HWL HACK
1038 // CurrentTSO = ThreadQueueHd;
1040 IF_DEBUG(gran, debugBelch("GRAN: time of next event is: %ld\n",
1043 if (RtsFlags.GranFlags.Light)
1044 GranSimLight_leave_system(event, &ActiveTSO);
1046 EndOfTimeSlice = CurrentTime[CurrentProc]+RtsFlags.GranFlags.time_slice;
1049 debugBelch("GRAN: end of time-slice is %#lx\n", EndOfTimeSlice));
1051 /* in a GranSim setup the TSO stays on the run queue */
1053 /* Take a thread from the run queue. */
1054 POP_RUN_QUEUE(t); // take_off_run_queue(t);
1057 debugBelch("GRAN: About to run current thread, which is\n");
1060 context_switch = 0; // turned on via GranYield, checking events and time slice
1063 DumpGranEvent(GR_SCHEDULE, t));
1065 procStatus[CurrentProc] = Busy;
1069 /* ----------------------------------------------------------------------------
1070 * Send pending messages (PARALLEL_HASKELL only)
1071 * ------------------------------------------------------------------------- */
1073 #if defined(PARALLEL_HASKELL)
1075 scheduleSendPendingMessages(void)
1081 # if defined(PAR) // global Mem.Mgmt., omit for now
1082 if (PendingFetches != END_BF_QUEUE) {
1087 if (RtsFlags.ParFlags.BufferTime) {
1088 // if we use message buffering, we must send away all message
1089 // packets which have become too old...
1095 /* ----------------------------------------------------------------------------
1096 * Activate spark threads (PARALLEL_HASKELL only)
1097 * ------------------------------------------------------------------------- */
1099 #if defined(PARALLEL_HASKELL)
1101 scheduleActivateSpark(void)
1104 ASSERT(EMPTY_RUN_QUEUE());
1105 /* We get here if the run queue is empty and want some work.
1106 We try to turn a spark into a thread, and add it to the run queue,
1107 from where it will be picked up in the next iteration of the scheduler
1111 /* :-[ no local threads => look out for local sparks */
1112 /* the spark pool for the current PE */
1113 pool = &(cap.r.rSparks); // JB: cap = (old) MainCap
1114 if (advisory_thread_count < RtsFlags.ParFlags.maxThreads &&
1115 pool->hd < pool->tl) {
1117 * ToDo: add GC code check that we really have enough heap afterwards!!
1119 * If we're here (no runnable threads) and we have pending
1120 * sparks, we must have a space problem. Get enough space
1121 * to turn one of those pending sparks into a
1125 spark = findSpark(rtsFalse); /* get a spark */
1126 if (spark != (rtsSpark) NULL) {
1127 tso = createThreadFromSpark(spark); /* turn the spark into a thread */
1128 IF_PAR_DEBUG(fish, // schedule,
1129 debugBelch("==== schedule: Created TSO %d (%p); %d threads active\n",
1130 tso->id, tso, advisory_thread_count));
1132 if (tso==END_TSO_QUEUE) { /* failed to activate spark->back to loop */
1133 IF_PAR_DEBUG(fish, // schedule,
1134 debugBelch("==^^ failed to create thread from spark @ %lx\n",
1136 return rtsFalse; /* failed to generate a thread */
1137 } /* otherwise fall through & pick-up new tso */
1139 IF_PAR_DEBUG(fish, // schedule,
1140 debugBelch("==^^ no local sparks (spark pool contains only NFs: %d)\n",
1141 spark_queue_len(pool)));
1142 return rtsFalse; /* failed to generate a thread */
1144 return rtsTrue; /* success in generating a thread */
1145 } else { /* no more threads permitted or pool empty */
1146 return rtsFalse; /* failed to generateThread */
1149 tso = NULL; // avoid compiler warning only
1150 return rtsFalse; /* dummy in non-PAR setup */
1153 #endif // PARALLEL_HASKELL
1155 /* ----------------------------------------------------------------------------
1156 * Get work from a remote node (PARALLEL_HASKELL only)
1157 * ------------------------------------------------------------------------- */
1159 #if defined(PARALLEL_HASKELL)
1161 scheduleGetRemoteWork(rtsBool *receivedFinish)
1163 ASSERT(EMPTY_RUN_QUEUE());
1165 if (RtsFlags.ParFlags.BufferTime) {
1166 IF_PAR_DEBUG(verbose,
1167 debugBelch("...send all pending data,"));
1170 for (i=1; i<=nPEs; i++)
1171 sendImmediately(i); // send all messages away immediately
1175 //++EDEN++ idle() , i.e. send all buffers, wait for work
1176 // suppress fishing in EDEN... just look for incoming messages
1177 // (blocking receive)
1178 IF_PAR_DEBUG(verbose,
1179 debugBelch("...wait for incoming messages...\n"));
1180 *receivedFinish = processMessages(); // blocking receive...
1182 // and reenter scheduling loop after having received something
1183 // (return rtsFalse below)
1185 # else /* activate SPARKS machinery */
1186 /* We get here, if we have no work, tried to activate a local spark, but still
1187 have no work. We try to get a remote spark, by sending a FISH message.
1188 Thread migration should be added here, and triggered when a sequence of
1189 fishes returns without work. */
1190 delay = (RtsFlags.ParFlags.fishDelay!=0ll ? RtsFlags.ParFlags.fishDelay : 0ll);
1192 /* =8-[ no local sparks => look for work on other PEs */
1194 * We really have absolutely no work. Send out a fish
1195 * (there may be some out there already), and wait for
1196 * something to arrive. We clearly can't run any threads
1197 * until a SCHEDULE or RESUME arrives, and so that's what
1198 * we're hoping to see. (Of course, we still have to
1199 * respond to other types of messages.)
1201 rtsTime now = msTime() /*CURRENT_TIME*/;
1202 IF_PAR_DEBUG(verbose,
1203 debugBelch("-- now=%ld\n", now));
1204 IF_PAR_DEBUG(fish, // verbose,
1205 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
1206 (last_fish_arrived_at!=0 &&
1207 last_fish_arrived_at+delay > now)) {
1208 debugBelch("--$$ <%llu> delaying FISH until %llu (last fish %llu, delay %llu)\n",
1209 now, last_fish_arrived_at+delay,
1210 last_fish_arrived_at,
1214 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
1215 advisory_thread_count < RtsFlags.ParFlags.maxThreads) { // send a FISH, but when?
1216 if (last_fish_arrived_at==0 ||
1217 (last_fish_arrived_at+delay <= now)) { // send FISH now!
1218 /* outstandingFishes is set in sendFish, processFish;
1219 avoid flooding system with fishes via delay */
1220 next_fish_to_send_at = 0;
1222 /* ToDo: this should be done in the main scheduling loop to avoid the
1223 busy wait here; not so bad if fish delay is very small */
1224 int iq = 0; // DEBUGGING -- HWL
1225 next_fish_to_send_at = last_fish_arrived_at+delay; // remember when to send
1226 /* send a fish when ready, but process messages that arrive in the meantime */
1228 if (PacketsWaiting()) {
1230 *receivedFinish = processMessages();
1233 } while (!*receivedFinish || now<next_fish_to_send_at);
1234 // JB: This means the fish could become obsolete, if we receive
1235 // work. Better check for work again?
1236 // last line: while (!receivedFinish || !haveWork || now<...)
1237 // next line: if (receivedFinish || haveWork )
1239 if (*receivedFinish) // no need to send a FISH if we are finishing anyway
1240 return rtsFalse; // NB: this will leave scheduler loop
1241 // immediately after return!
1243 IF_PAR_DEBUG(fish, // verbose,
1244 debugBelch("--$$ <%llu> sent delayed fish (%d processMessages); active/total threads=%d/%d\n",now,iq,run_queue_len(),advisory_thread_count));
1248 // JB: IMHO, this should all be hidden inside sendFish(...)
1250 sendFish(pe, thisPE, NEW_FISH_AGE, NEW_FISH_HISTORY,
1253 // Global statistics: count no. of fishes
1254 if (RtsFlags.ParFlags.ParStats.Global &&
1255 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
1256 globalParStats.tot_fish_mess++;
1260 /* delayed fishes must have been sent by now! */
1261 next_fish_to_send_at = 0;
1264 *receivedFinish = processMessages();
1265 # endif /* SPARKS */
1268 /* NB: this function always returns rtsFalse, meaning the scheduler
1269 loop continues with the next iteration;
1271 return code means success in finding work; we enter this function
1272 if there is no local work, thus have to send a fish which takes
1273 time until it arrives with work; in the meantime we should process
1274 messages in the main loop;
1277 #endif // PARALLEL_HASKELL
1279 /* ----------------------------------------------------------------------------
1280 * PAR/GRAN: Report stats & debugging info(?)
1281 * ------------------------------------------------------------------------- */
1283 #if defined(PAR) || defined(GRAN)
1285 scheduleGranParReport(void)
1287 ASSERT(run_queue_hd != END_TSO_QUEUE);
1289 /* Take a thread from the run queue, if we have work */
1290 POP_RUN_QUEUE(t); // take_off_run_queue(END_TSO_QUEUE);
1292 /* If this TSO has got its outport closed in the meantime,
1293 * it mustn't be run. Instead, we have to clean it up as if it was finished.
1294 * It has to be marked as TH_DEAD for this purpose.
1295 * If it is TH_TERM instead, it is supposed to have finished in the normal way.
1297 JB: TODO: investigate wether state change field could be nuked
1298 entirely and replaced by the normal tso state (whatnext
1299 field). All we want to do is to kill tsos from outside.
1302 /* ToDo: write something to the log-file
1303 if (RTSflags.ParFlags.granSimStats && !sameThread)
1304 DumpGranEvent(GR_SCHEDULE, RunnableThreadsHd);
1308 /* the spark pool for the current PE */
1309 pool = &(cap.r.rSparks); // cap = (old) MainCap
1312 debugBelch("--=^ %d threads, %d sparks on [%#x]\n",
1313 run_queue_len(), spark_queue_len(pool), CURRENT_PROC));
1316 debugBelch("--=^ %d threads, %d sparks on [%#x]\n",
1317 run_queue_len(), spark_queue_len(pool), CURRENT_PROC));
1319 if (RtsFlags.ParFlags.ParStats.Full &&
1320 (t->par.sparkname != (StgInt)0) && // only log spark generated threads
1321 (emitSchedule || // forced emit
1322 (t && LastTSO && t->id != LastTSO->id))) {
1324 we are running a different TSO, so write a schedule event to log file
1325 NB: If we use fair scheduling we also have to write a deschedule
1326 event for LastTSO; with unfair scheduling we know that the
1327 previous tso has blocked whenever we switch to another tso, so
1328 we don't need it in GUM for now
1330 IF_PAR_DEBUG(fish, // schedule,
1331 debugBelch("____ scheduling spark generated thread %d (%lx) (%lx) via a forced emit\n",t->id,t,t->par.sparkname));
1333 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
1334 GR_SCHEDULE, t, (StgClosure *)NULL, 0, 0);
1335 emitSchedule = rtsFalse;
1340 /* ----------------------------------------------------------------------------
1341 * After running a thread...
1342 * ASSUMES: sched_mutex
1343 * ------------------------------------------------------------------------- */
1346 schedulePostRunThread(void)
1349 /* HACK 675: if the last thread didn't yield, make sure to print a
1350 SCHEDULE event to the log file when StgRunning the next thread, even
1351 if it is the same one as before */
1353 TimeOfLastYield = CURRENT_TIME;
1356 /* some statistics gathering in the parallel case */
1358 #if defined(GRAN) || defined(PAR) || defined(EDEN)
1362 IF_DEBUG(gran, DumpGranEvent(GR_DESCHEDULE, t));
1363 globalGranStats.tot_heapover++;
1365 globalParStats.tot_heapover++;
1372 DumpGranEvent(GR_DESCHEDULE, t));
1373 globalGranStats.tot_stackover++;
1376 // DumpGranEvent(GR_DESCHEDULE, t);
1377 globalParStats.tot_stackover++;
1381 case ThreadYielding:
1384 DumpGranEvent(GR_DESCHEDULE, t));
1385 globalGranStats.tot_yields++;
1388 // DumpGranEvent(GR_DESCHEDULE, t);
1389 globalParStats.tot_yields++;
1396 debugBelch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: ",
1397 t->id, t, whatNext_strs[t->what_next], t->block_info.closure,
1398 (t->block_info.closure==(StgClosure*)NULL ? 99 : where_is(t->block_info.closure)));
1399 if (t->block_info.closure!=(StgClosure*)NULL)
1400 print_bq(t->block_info.closure);
1403 // ??? needed; should emit block before
1405 DumpGranEvent(GR_DESCHEDULE, t));
1406 prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
1409 ASSERT(procStatus[CurrentProc]==Busy ||
1410 ((procStatus[CurrentProc]==Fetching) &&
1411 (t->block_info.closure!=(StgClosure*)NULL)));
1412 if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
1413 !(!RtsFlags.GranFlags.DoAsyncFetch &&
1414 procStatus[CurrentProc]==Fetching))
1415 procStatus[CurrentProc] = Idle;
1418 //++PAR++ blockThread() writes the event (change?)
1422 case ThreadFinished:
1426 barf("parGlobalStats: unknown return code");
1432 /* -----------------------------------------------------------------------------
1433 * Handle a thread that returned to the scheduler with ThreadHeepOverflow
1434 * ASSUMES: sched_mutex
1435 * -------------------------------------------------------------------------- */
1438 scheduleHandleHeapOverflow( Capability *cap, StgTSO *t )
1440 // did the task ask for a large block?
1441 if (cap->r.rHpAlloc > BLOCK_SIZE) {
1442 // if so, get one and push it on the front of the nursery.
1446 blocks = (lnat)BLOCK_ROUND_UP(cap->r.rHpAlloc) / BLOCK_SIZE;
1449 debugBelch("--<< thread %ld (%s) stopped: requesting a large block (size %ld)\n",
1450 (long)t->id, whatNext_strs[t->what_next], blocks));
1452 // don't do this if it would push us over the
1453 // alloc_blocks_lim limit; we'll GC first.
1454 if (alloc_blocks + blocks < alloc_blocks_lim) {
1456 alloc_blocks += blocks;
1457 bd = allocGroup( blocks );
1459 // link the new group into the list
1460 bd->link = cap->r.rCurrentNursery;
1461 bd->u.back = cap->r.rCurrentNursery->u.back;
1462 if (cap->r.rCurrentNursery->u.back != NULL) {
1463 cap->r.rCurrentNursery->u.back->link = bd;
1466 ASSERT(g0s0->blocks == cap->r.rCurrentNursery &&
1467 g0s0 == cap->r.rNursery);
1470 cap->r.rNursery->blocks = bd;
1472 cap->r.rCurrentNursery->u.back = bd;
1474 // initialise it as a nursery block. We initialise the
1475 // step, gen_no, and flags field of *every* sub-block in
1476 // this large block, because this is easier than making
1477 // sure that we always find the block head of a large
1478 // block whenever we call Bdescr() (eg. evacuate() and
1479 // isAlive() in the GC would both have to do this, at
1483 for (x = bd; x < bd + blocks; x++) {
1491 // don't forget to update the block count in g0s0.
1492 g0s0->n_blocks += blocks;
1494 // This assert can be a killer if the app is doing lots
1495 // of large block allocations.
1496 ASSERT(countBlocks(g0s0->blocks) == g0s0->n_blocks);
1499 // now update the nursery to point to the new block
1500 cap->r.rCurrentNursery = bd;
1502 // we might be unlucky and have another thread get on the
1503 // run queue before us and steal the large block, but in that
1504 // case the thread will just end up requesting another large
1506 PUSH_ON_RUN_QUEUE(t);
1507 return rtsFalse; /* not actually GC'ing */
1511 /* make all the running tasks block on a condition variable,
1512 * maybe set context_switch and wait till they all pile in,
1513 * then have them wait on a GC condition variable.
1516 debugBelch("--<< thread %ld (%s) stopped: HeapOverflow\n",
1517 (long)t->id, whatNext_strs[t->what_next]));
1520 ASSERT(!is_on_queue(t,CurrentProc));
1521 #elif defined(PARALLEL_HASKELL)
1522 /* Currently we emit a DESCHEDULE event before GC in GUM.
1523 ToDo: either add separate event to distinguish SYSTEM time from rest
1524 or just nuke this DESCHEDULE (and the following SCHEDULE) */
1525 if (0 && RtsFlags.ParFlags.ParStats.Full) {
1526 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
1527 GR_DESCHEDULE, t, (StgClosure *)NULL, 0, 0);
1528 emitSchedule = rtsTrue;
1532 PUSH_ON_RUN_QUEUE(t);
1534 /* actual GC is done at the end of the while loop in schedule() */
1537 /* -----------------------------------------------------------------------------
1538 * Handle a thread that returned to the scheduler with ThreadStackOverflow
1539 * ASSUMES: sched_mutex
1540 * -------------------------------------------------------------------------- */
1543 scheduleHandleStackOverflow( StgTSO *t)
1545 IF_DEBUG(scheduler,debugBelch("--<< thread %ld (%s) stopped, StackOverflow\n",
1546 (long)t->id, whatNext_strs[t->what_next]));
1547 /* just adjust the stack for this thread, then pop it back
1552 /* enlarge the stack */
1553 StgTSO *new_t = threadStackOverflow(t);
1555 /* This TSO has moved, so update any pointers to it from the
1556 * main thread stack. It better not be on any other queues...
1557 * (it shouldn't be).
1559 if (t->main != NULL) {
1560 t->main->tso = new_t;
1562 PUSH_ON_RUN_QUEUE(new_t);
1566 /* -----------------------------------------------------------------------------
1567 * Handle a thread that returned to the scheduler with ThreadYielding
1568 * ASSUMES: sched_mutex
1569 * -------------------------------------------------------------------------- */
1572 scheduleHandleYield( StgTSO *t, nat prev_what_next )
1574 // Reset the context switch flag. We don't do this just before
1575 // running the thread, because that would mean we would lose ticks
1576 // during GC, which can lead to unfair scheduling (a thread hogs
1577 // the CPU because the tick always arrives during GC). This way
1578 // penalises threads that do a lot of allocation, but that seems
1579 // better than the alternative.
1582 /* put the thread back on the run queue. Then, if we're ready to
1583 * GC, check whether this is the last task to stop. If so, wake
1584 * up the GC thread. getThread will block during a GC until the
1588 if (t->what_next != prev_what_next) {
1589 debugBelch("--<< thread %ld (%s) stopped to switch evaluators\n",
1590 (long)t->id, whatNext_strs[t->what_next]);
1592 debugBelch("--<< thread %ld (%s) stopped, yielding\n",
1593 (long)t->id, whatNext_strs[t->what_next]);
1598 //debugBelch("&& Doing sanity check on yielding TSO %ld.", t->id);
1600 ASSERT(t->link == END_TSO_QUEUE);
1602 // Shortcut if we're just switching evaluators: don't bother
1603 // doing stack squeezing (which can be expensive), just run the
1605 if (t->what_next != prev_what_next) {
1612 ASSERT(!is_on_queue(t,CurrentProc));
1615 //debugBelch("&& Doing sanity check on all ThreadQueues (and their TSOs).");
1616 checkThreadQsSanity(rtsTrue));
1623 /* add a ContinueThread event to actually process the thread */
1624 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1626 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1628 debugBelch("GRAN: eventq and runnableq after adding yielded thread to queue again:\n");
1635 /* -----------------------------------------------------------------------------
1636 * Handle a thread that returned to the scheduler with ThreadBlocked
1637 * ASSUMES: sched_mutex
1638 * -------------------------------------------------------------------------- */
1641 scheduleHandleThreadBlocked( StgTSO *t
1642 #if !defined(GRAN) && !defined(DEBUG)
1649 debugBelch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: \n",
1650 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)));
1651 if (t->block_info.closure!=(StgClosure*)NULL) print_bq(t->block_info.closure));
1653 // ??? needed; should emit block before
1655 DumpGranEvent(GR_DESCHEDULE, t));
1656 prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
1659 ASSERT(procStatus[CurrentProc]==Busy ||
1660 ((procStatus[CurrentProc]==Fetching) &&
1661 (t->block_info.closure!=(StgClosure*)NULL)));
1662 if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
1663 !(!RtsFlags.GranFlags.DoAsyncFetch &&
1664 procStatus[CurrentProc]==Fetching))
1665 procStatus[CurrentProc] = Idle;
1669 debugBelch("--<< thread %ld (%p; %s) stopped, blocking on node %p with BQ: \n",
1670 t->id, t, whatNext_strs[t->what_next], t->block_info.closure));
1673 if (t->block_info.closure!=(StgClosure*)NULL)
1674 print_bq(t->block_info.closure));
1676 /* Send a fetch (if BlockedOnGA) and dump event to log file */
1679 /* whatever we schedule next, we must log that schedule */
1680 emitSchedule = rtsTrue;
1683 /* don't need to do anything. Either the thread is blocked on
1684 * I/O, in which case we'll have called addToBlockedQueue
1685 * previously, or it's blocked on an MVar or Blackhole, in which
1686 * case it'll be on the relevant queue already.
1688 ASSERT(t->why_blocked != NotBlocked);
1690 debugBelch("--<< thread %d (%s) stopped: ",
1691 t->id, whatNext_strs[t->what_next]);
1692 printThreadBlockage(t);
1695 /* Only for dumping event to log file
1696 ToDo: do I need this in GranSim, too?
1702 /* -----------------------------------------------------------------------------
1703 * Handle a thread that returned to the scheduler with ThreadFinished
1704 * ASSUMES: sched_mutex
1705 * -------------------------------------------------------------------------- */
1708 scheduleHandleThreadFinished( StgMainThread *mainThread
1709 USED_WHEN_RTS_SUPPORTS_THREADS,
1713 /* Need to check whether this was a main thread, and if so,
1714 * return with the return value.
1716 * We also end up here if the thread kills itself with an
1717 * uncaught exception, see Exception.cmm.
1719 IF_DEBUG(scheduler,debugBelch("--++ thread %d (%s) finished\n",
1720 t->id, whatNext_strs[t->what_next]));
1723 endThread(t, CurrentProc); // clean-up the thread
1724 #elif defined(PARALLEL_HASKELL)
1725 /* For now all are advisory -- HWL */
1726 //if(t->priority==AdvisoryPriority) ??
1727 advisory_thread_count--; // JB: Caution with this counter, buggy!
1730 if(t->dist.priority==RevalPriority)
1734 # if defined(EDENOLD)
1735 // the thread could still have an outport... (BUG)
1736 if (t->eden.outport != -1) {
1737 // delete the outport for the tso which has finished...
1738 IF_PAR_DEBUG(eden_ports,
1739 debugBelch("WARNING: Scheduler removes outport %d for TSO %d.\n",
1740 t->eden.outport, t->id));
1743 // thread still in the process (HEAVY BUG! since outport has just been closed...)
1744 if (t->eden.epid != -1) {
1745 IF_PAR_DEBUG(eden_ports,
1746 debugBelch("WARNING: Scheduler removes TSO %d from process %d .\n",
1747 t->id, t->eden.epid));
1748 removeTSOfromProcess(t);
1753 if (RtsFlags.ParFlags.ParStats.Full &&
1754 !RtsFlags.ParFlags.ParStats.Suppressed)
1755 DumpEndEvent(CURRENT_PROC, t, rtsFalse /* not mandatory */);
1757 // t->par only contains statistics: left out for now...
1759 debugBelch("**** end thread: ended sparked thread %d (%lx); sparkname: %lx\n",
1760 t->id,t,t->par.sparkname));
1762 #endif // PARALLEL_HASKELL
1765 // Check whether the thread that just completed was a main
1766 // thread, and if so return with the result.
1768 // There is an assumption here that all thread completion goes
1769 // through this point; we need to make sure that if a thread
1770 // ends up in the ThreadKilled state, that it stays on the run
1771 // queue so it can be dealt with here.
1774 #if defined(RTS_SUPPORTS_THREADS)
1777 mainThread->tso == t
1781 // We are a bound thread: this must be our thread that just
1783 ASSERT(mainThread->tso == t);
1785 if (t->what_next == ThreadComplete) {
1786 if (mainThread->ret) {
1787 // NOTE: return val is tso->sp[1] (see StgStartup.hc)
1788 *(mainThread->ret) = (StgClosure *)mainThread->tso->sp[1];
1790 mainThread->stat = Success;
1792 if (mainThread->ret) {
1793 *(mainThread->ret) = NULL;
1796 mainThread->stat = Interrupted;
1798 mainThread->stat = Killed;
1802 removeThreadLabel((StgWord)mainThread->tso->id);
1804 if (mainThread->prev == NULL) {
1805 main_threads = mainThread->link;
1807 mainThread->prev->link = mainThread->link;
1809 if (mainThread->link != NULL) {
1810 mainThread->link->prev = mainThread->prev;
1812 releaseCapability(cap);
1813 return rtsTrue; // tells schedule() to return
1816 #ifdef RTS_SUPPORTS_THREADS
1817 ASSERT(t->main == NULL);
1819 if (t->main != NULL) {
1820 // Must be a main thread that is not the topmost one. Leave
1821 // it on the run queue until the stack has unwound to the
1822 // point where we can deal with this. Leaving it on the run
1823 // queue also ensures that the garbage collector knows about
1824 // this thread and its return value (it gets dropped from the
1825 // all_threads list so there's no other way to find it).
1826 APPEND_TO_RUN_QUEUE(t);
1832 /* -----------------------------------------------------------------------------
1833 * Perform a heap census, if PROFILING
1834 * -------------------------------------------------------------------------- */
1837 scheduleDoHeapProfile( rtsBool ready_to_gc STG_UNUSED )
1839 #if defined(PROFILING)
1840 // When we have +RTS -i0 and we're heap profiling, do a census at
1841 // every GC. This lets us get repeatable runs for debugging.
1842 if (performHeapProfile ||
1843 (RtsFlags.ProfFlags.profileInterval==0 &&
1844 RtsFlags.ProfFlags.doHeapProfile && ready_to_gc)) {
1845 GarbageCollect(GetRoots, rtsTrue);
1847 performHeapProfile = rtsFalse;
1848 return rtsTrue; // true <=> we already GC'd
1854 /* -----------------------------------------------------------------------------
1855 * Perform a garbage collection if necessary
1856 * ASSUMES: sched_mutex
1857 * -------------------------------------------------------------------------- */
1860 scheduleDoGC( Capability *cap STG_UNUSED )
1864 static rtsBool waiting_for_gc;
1865 int n_capabilities = RtsFlags.ParFlags.nNodes - 1;
1866 // subtract one because we're already holding one.
1867 Capability *caps[n_capabilities];
1871 // In order to GC, there must be no threads running Haskell code.
1872 // Therefore, the GC thread needs to hold *all* the capabilities,
1873 // and release them after the GC has completed.
1875 // This seems to be the simplest way: previous attempts involved
1876 // making all the threads with capabilities give up their
1877 // capabilities and sleep except for the *last* one, which
1878 // actually did the GC. But it's quite hard to arrange for all
1879 // the other tasks to sleep and stay asleep.
1882 // Someone else is already trying to GC
1883 if (waiting_for_gc) return;
1884 waiting_for_gc = rtsTrue;
1886 caps[n_capabilities] = cap;
1887 while (n_capabilities > 0) {
1888 IF_DEBUG(scheduler, sched_belch("ready_to_gc, grabbing all the capabilies (%d left)", n_capabilities));
1889 waitForReturnCapability(&sched_mutex, &cap);
1891 caps[n_capabilities] = cap;
1894 waiting_for_gc = rtsFalse;
1897 /* Kick any transactions which are invalid back to their
1898 * atomically frames. When next scheduled they will try to
1899 * commit, this commit will fail and they will retry.
1901 for (t = all_threads; t != END_TSO_QUEUE; t = t -> link) {
1902 if (t -> what_next != ThreadRelocated && t -> trec != NO_TREC && t -> why_blocked == NotBlocked) {
1903 if (!stmValidateTransaction (t -> trec)) {
1904 IF_DEBUG(stm, sched_belch("trec %p found wasting its time", t));
1906 // strip the stack back to the ATOMICALLY_FRAME, aborting
1907 // the (nested) transaction, and saving the stack of any
1908 // partially-evaluated thunks on the heap.
1909 raiseAsync_(t, NULL, rtsTrue);
1912 ASSERT(get_itbl((StgClosure *)t->sp)->type == ATOMICALLY_FRAME);
1918 // so this happens periodically:
1919 scheduleCheckBlackHoles();
1921 /* everybody back, start the GC.
1922 * Could do it in this thread, or signal a condition var
1923 * to do it in another thread. Either way, we need to
1924 * broadcast on gc_pending_cond afterward.
1926 #if defined(RTS_SUPPORTS_THREADS)
1927 IF_DEBUG(scheduler,sched_belch("doing GC"));
1929 GarbageCollect(GetRoots,rtsFalse);
1933 // release our stash of capabilities.
1935 for (i = 0; i < RtsFlags.ParFlags.nNodes-1; i++) {
1936 releaseCapability(caps[i]);
1942 /* add a ContinueThread event to continue execution of current thread */
1943 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1945 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1947 debugBelch("GRAN: eventq and runnableq after Garbage collection:\n\n");
1953 /* ---------------------------------------------------------------------------
1954 * rtsSupportsBoundThreads(): is the RTS built to support bound threads?
1955 * used by Control.Concurrent for error checking.
1956 * ------------------------------------------------------------------------- */
1959 rtsSupportsBoundThreads(void)
1968 /* ---------------------------------------------------------------------------
1969 * isThreadBound(tso): check whether tso is bound to an OS thread.
1970 * ------------------------------------------------------------------------- */
1973 isThreadBound(StgTSO* tso USED_IN_THREADED_RTS)
1976 return (tso->main != NULL);
1981 /* ---------------------------------------------------------------------------
1982 * Singleton fork(). Do not copy any running threads.
1983 * ------------------------------------------------------------------------- */
1985 #ifndef mingw32_HOST_OS
1986 #define FORKPROCESS_PRIMOP_SUPPORTED
1989 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
1991 deleteThreadImmediately(StgTSO *tso);
1994 forkProcess(HsStablePtr *entry
1995 #ifndef FORKPROCESS_PRIMOP_SUPPORTED
2000 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
2006 IF_DEBUG(scheduler,sched_belch("forking!"));
2007 rts_lock(); // This not only acquires sched_mutex, it also
2008 // makes sure that no other threads are running
2012 if (pid) { /* parent */
2014 /* just return the pid */
2018 } else { /* child */
2021 // delete all threads
2022 run_queue_hd = run_queue_tl = END_TSO_QUEUE;
2024 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
2027 // don't allow threads to catch the ThreadKilled exception
2028 deleteThreadImmediately(t);
2031 // wipe the main thread list
2032 while((m = main_threads) != NULL) {
2033 main_threads = m->link;
2034 # ifdef THREADED_RTS
2035 closeCondition(&m->bound_thread_cond);
2040 rc = rts_evalStableIO(entry, NULL); // run the action
2041 rts_checkSchedStatus("forkProcess",rc);
2045 hs_exit(); // clean up and exit
2048 #else /* !FORKPROCESS_PRIMOP_SUPPORTED */
2049 barf("forkProcess#: primop not supported, sorry!\n");
2054 /* ---------------------------------------------------------------------------
2055 * deleteAllThreads(): kill all the live threads.
2057 * This is used when we catch a user interrupt (^C), before performing
2058 * any necessary cleanups and running finalizers.
2060 * Locks: sched_mutex held.
2061 * ------------------------------------------------------------------------- */
2064 deleteAllThreads ( void )
2067 IF_DEBUG(scheduler,sched_belch("deleting all threads"));
2068 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
2069 next = t->global_link;
2073 // The run queue now contains a bunch of ThreadKilled threads. We
2074 // must not throw these away: the main thread(s) will be in there
2075 // somewhere, and the main scheduler loop has to deal with it.
2076 // Also, the run queue is the only thing keeping these threads from
2077 // being GC'd, and we don't want the "main thread has been GC'd" panic.
2079 ASSERT(blocked_queue_hd == END_TSO_QUEUE);
2080 ASSERT(blackhole_queue == END_TSO_QUEUE);
2081 ASSERT(sleeping_queue == END_TSO_QUEUE);
2084 /* startThread and insertThread are now in GranSim.c -- HWL */
2087 /* ---------------------------------------------------------------------------
2088 * Suspending & resuming Haskell threads.
2090 * When making a "safe" call to C (aka _ccall_GC), the task gives back
2091 * its capability before calling the C function. This allows another
2092 * task to pick up the capability and carry on running Haskell
2093 * threads. It also means that if the C call blocks, it won't lock
2096 * The Haskell thread making the C call is put to sleep for the
2097 * duration of the call, on the susepended_ccalling_threads queue. We
2098 * give out a token to the task, which it can use to resume the thread
2099 * on return from the C function.
2100 * ------------------------------------------------------------------------- */
2103 suspendThread( StgRegTable *reg )
2107 int saved_errno = errno;
2109 /* assume that *reg is a pointer to the StgRegTable part
2112 cap = (Capability *)((void *)((unsigned char*)reg - sizeof(StgFunTable)));
2114 ACQUIRE_LOCK(&sched_mutex);
2117 sched_belch("thread %d did a _ccall_gc", cap->r.rCurrentTSO->id));
2119 // XXX this might not be necessary --SDM
2120 cap->r.rCurrentTSO->what_next = ThreadRunGHC;
2122 threadPaused(cap->r.rCurrentTSO);
2123 cap->r.rCurrentTSO->link = suspended_ccalling_threads;
2124 suspended_ccalling_threads = cap->r.rCurrentTSO;
2126 if(cap->r.rCurrentTSO->blocked_exceptions == NULL) {
2127 cap->r.rCurrentTSO->why_blocked = BlockedOnCCall;
2128 cap->r.rCurrentTSO->blocked_exceptions = END_TSO_QUEUE;
2130 cap->r.rCurrentTSO->why_blocked = BlockedOnCCall_NoUnblockExc;
2133 /* Use the thread ID as the token; it should be unique */
2134 tok = cap->r.rCurrentTSO->id;
2136 /* Hand back capability */
2137 cap->r.rInHaskell = rtsFalse;
2138 releaseCapability(cap);
2140 #if defined(RTS_SUPPORTS_THREADS)
2141 /* Preparing to leave the RTS, so ensure there's a native thread/task
2142 waiting to take over.
2144 IF_DEBUG(scheduler, sched_belch("worker (token %d): leaving RTS", tok));
2147 RELEASE_LOCK(&sched_mutex);
2149 errno = saved_errno;
2154 resumeThread( StgInt tok )
2156 StgTSO *tso, **prev;
2158 int saved_errno = errno;
2160 #if defined(RTS_SUPPORTS_THREADS)
2161 /* Wait for permission to re-enter the RTS with the result. */
2162 ACQUIRE_LOCK(&sched_mutex);
2163 waitForReturnCapability(&sched_mutex, &cap);
2165 IF_DEBUG(scheduler, sched_belch("worker (token %d): re-entering RTS", tok));
2167 grabCapability(&cap);
2170 /* Remove the thread off of the suspended list */
2171 prev = &suspended_ccalling_threads;
2172 for (tso = suspended_ccalling_threads;
2173 tso != END_TSO_QUEUE;
2174 prev = &tso->link, tso = tso->link) {
2175 if (tso->id == (StgThreadID)tok) {
2180 if (tso == END_TSO_QUEUE) {
2181 barf("resumeThread: thread not found");
2183 tso->link = END_TSO_QUEUE;
2185 if(tso->why_blocked == BlockedOnCCall) {
2186 awakenBlockedQueueNoLock(tso->blocked_exceptions);
2187 tso->blocked_exceptions = NULL;
2190 /* Reset blocking status */
2191 tso->why_blocked = NotBlocked;
2193 cap->r.rCurrentTSO = tso;
2194 cap->r.rInHaskell = rtsTrue;
2195 RELEASE_LOCK(&sched_mutex);
2196 errno = saved_errno;
2200 /* ---------------------------------------------------------------------------
2201 * Comparing Thread ids.
2203 * This is used from STG land in the implementation of the
2204 * instances of Eq/Ord for ThreadIds.
2205 * ------------------------------------------------------------------------ */
2208 cmp_thread(StgPtr tso1, StgPtr tso2)
2210 StgThreadID id1 = ((StgTSO *)tso1)->id;
2211 StgThreadID id2 = ((StgTSO *)tso2)->id;
2213 if (id1 < id2) return (-1);
2214 if (id1 > id2) return 1;
2218 /* ---------------------------------------------------------------------------
2219 * Fetching the ThreadID from an StgTSO.
2221 * This is used in the implementation of Show for ThreadIds.
2222 * ------------------------------------------------------------------------ */
2224 rts_getThreadId(StgPtr tso)
2226 return ((StgTSO *)tso)->id;
2231 labelThread(StgPtr tso, char *label)
2236 /* Caveat: Once set, you can only set the thread name to "" */
2237 len = strlen(label)+1;
2238 buf = stgMallocBytes(len * sizeof(char), "Schedule.c:labelThread()");
2239 strncpy(buf,label,len);
2240 /* Update will free the old memory for us */
2241 updateThreadLabel(((StgTSO *)tso)->id,buf);
2245 /* ---------------------------------------------------------------------------
2246 Create a new thread.
2248 The new thread starts with the given stack size. Before the
2249 scheduler can run, however, this thread needs to have a closure
2250 (and possibly some arguments) pushed on its stack. See
2251 pushClosure() in Schedule.h.
2253 createGenThread() and createIOThread() (in SchedAPI.h) are
2254 convenient packaged versions of this function.
2256 currently pri (priority) is only used in a GRAN setup -- HWL
2257 ------------------------------------------------------------------------ */
2259 /* currently pri (priority) is only used in a GRAN setup -- HWL */
2261 createThread(nat size, StgInt pri)
2264 createThread(nat size)
2271 /* First check whether we should create a thread at all */
2272 #if defined(PARALLEL_HASKELL)
2273 /* check that no more than RtsFlags.ParFlags.maxThreads threads are created */
2274 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads) {
2276 debugBelch("{createThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)\n",
2277 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
2278 return END_TSO_QUEUE;
2284 ASSERT(!RtsFlags.GranFlags.Light || CurrentProc==0);
2287 // ToDo: check whether size = stack_size - TSO_STRUCT_SIZEW
2289 /* catch ridiculously small stack sizes */
2290 if (size < MIN_STACK_WORDS + TSO_STRUCT_SIZEW) {
2291 size = MIN_STACK_WORDS + TSO_STRUCT_SIZEW;
2294 stack_size = size - TSO_STRUCT_SIZEW;
2296 tso = (StgTSO *)allocate(size);
2297 TICK_ALLOC_TSO(stack_size, 0);
2299 SET_HDR(tso, &stg_TSO_info, CCS_SYSTEM);
2301 SET_GRAN_HDR(tso, ThisPE);
2304 // Always start with the compiled code evaluator
2305 tso->what_next = ThreadRunGHC;
2307 tso->id = next_thread_id++;
2308 tso->why_blocked = NotBlocked;
2309 tso->blocked_exceptions = NULL;
2311 tso->saved_errno = 0;
2314 tso->stack_size = stack_size;
2315 tso->max_stack_size = round_to_mblocks(RtsFlags.GcFlags.maxStkSize)
2317 tso->sp = (P_)&(tso->stack) + stack_size;
2319 tso->trec = NO_TREC;
2322 tso->prof.CCCS = CCS_MAIN;
2325 /* put a stop frame on the stack */
2326 tso->sp -= sizeofW(StgStopFrame);
2327 SET_HDR((StgClosure*)tso->sp,(StgInfoTable *)&stg_stop_thread_info,CCS_SYSTEM);
2328 tso->link = END_TSO_QUEUE;
2332 /* uses more flexible routine in GranSim */
2333 insertThread(tso, CurrentProc);
2335 /* In a non-GranSim setup the pushing of a TSO onto the runq is separated
2341 if (RtsFlags.GranFlags.GranSimStats.Full)
2342 DumpGranEvent(GR_START,tso);
2343 #elif defined(PARALLEL_HASKELL)
2344 if (RtsFlags.ParFlags.ParStats.Full)
2345 DumpGranEvent(GR_STARTQ,tso);
2346 /* HACk to avoid SCHEDULE
2350 /* Link the new thread on the global thread list.
2352 tso->global_link = all_threads;
2356 tso->dist.priority = MandatoryPriority; //by default that is...
2360 tso->gran.pri = pri;
2362 tso->gran.magic = TSO_MAGIC; // debugging only
2364 tso->gran.sparkname = 0;
2365 tso->gran.startedat = CURRENT_TIME;
2366 tso->gran.exported = 0;
2367 tso->gran.basicblocks = 0;
2368 tso->gran.allocs = 0;
2369 tso->gran.exectime = 0;
2370 tso->gran.fetchtime = 0;
2371 tso->gran.fetchcount = 0;
2372 tso->gran.blocktime = 0;
2373 tso->gran.blockcount = 0;
2374 tso->gran.blockedat = 0;
2375 tso->gran.globalsparks = 0;
2376 tso->gran.localsparks = 0;
2377 if (RtsFlags.GranFlags.Light)
2378 tso->gran.clock = Now; /* local clock */
2380 tso->gran.clock = 0;
2382 IF_DEBUG(gran,printTSO(tso));
2383 #elif defined(PARALLEL_HASKELL)
2385 tso->par.magic = TSO_MAGIC; // debugging only
2387 tso->par.sparkname = 0;
2388 tso->par.startedat = CURRENT_TIME;
2389 tso->par.exported = 0;
2390 tso->par.basicblocks = 0;
2391 tso->par.allocs = 0;
2392 tso->par.exectime = 0;
2393 tso->par.fetchtime = 0;
2394 tso->par.fetchcount = 0;
2395 tso->par.blocktime = 0;
2396 tso->par.blockcount = 0;
2397 tso->par.blockedat = 0;
2398 tso->par.globalsparks = 0;
2399 tso->par.localsparks = 0;
2403 globalGranStats.tot_threads_created++;
2404 globalGranStats.threads_created_on_PE[CurrentProc]++;
2405 globalGranStats.tot_sq_len += spark_queue_len(CurrentProc);
2406 globalGranStats.tot_sq_probes++;
2407 #elif defined(PARALLEL_HASKELL)
2408 // collect parallel global statistics (currently done together with GC stats)
2409 if (RtsFlags.ParFlags.ParStats.Global &&
2410 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
2411 //debugBelch("Creating thread %d @ %11.2f\n", tso->id, usertime());
2412 globalParStats.tot_threads_created++;
2418 sched_belch("==__ schedule: Created TSO %d (%p);",
2419 CurrentProc, tso, tso->id));
2420 #elif defined(PARALLEL_HASKELL)
2421 IF_PAR_DEBUG(verbose,
2422 sched_belch("==__ schedule: Created TSO %d (%p); %d threads active",
2423 (long)tso->id, tso, advisory_thread_count));
2425 IF_DEBUG(scheduler,sched_belch("created thread %ld, stack size = %lx words",
2426 (long)tso->id, (long)tso->stack_size));
2433 all parallel thread creation calls should fall through the following routine.
2436 createThreadFromSpark(rtsSpark spark)
2438 ASSERT(spark != (rtsSpark)NULL);
2439 // JB: TAKE CARE OF THIS COUNTER! BUGGY
2440 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads)
2442 barf("{createSparkThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
2443 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
2444 return END_TSO_QUEUE;
2448 tso = createThread(RtsFlags.GcFlags.initialStkSize);
2449 if (tso==END_TSO_QUEUE)
2450 barf("createSparkThread: Cannot create TSO");
2452 tso->priority = AdvisoryPriority;
2454 pushClosure(tso,spark);
2456 advisory_thread_count++; // JB: TAKE CARE OF THIS COUNTER! BUGGY
2463 Turn a spark into a thread.
2464 ToDo: fix for SMP (needs to acquire SCHED_MUTEX!)
2468 activateSpark (rtsSpark spark)
2472 tso = createSparkThread(spark);
2473 if (RtsFlags.ParFlags.ParStats.Full) {
2474 //ASSERT(run_queue_hd == END_TSO_QUEUE); // I think ...
2475 IF_PAR_DEBUG(verbose,
2476 debugBelch("==^^ activateSpark: turning spark of closure %p (%s) into a thread\n",
2477 (StgClosure *)spark, info_type((StgClosure *)spark)));
2479 // ToDo: fwd info on local/global spark to thread -- HWL
2480 // tso->gran.exported = spark->exported;
2481 // tso->gran.locked = !spark->global;
2482 // tso->gran.sparkname = spark->name;
2488 /* ---------------------------------------------------------------------------
2491 * scheduleThread puts a thread on the head of the runnable queue.
2492 * This will usually be done immediately after a thread is created.
2493 * The caller of scheduleThread must create the thread using e.g.
2494 * createThread and push an appropriate closure
2495 * on this thread's stack before the scheduler is invoked.
2496 * ------------------------------------------------------------------------ */
2499 scheduleThread_(StgTSO *tso)
2501 // The thread goes at the *end* of the run-queue, to avoid possible
2502 // starvation of any threads already on the queue.
2503 APPEND_TO_RUN_QUEUE(tso);
2508 scheduleThread(StgTSO* tso)
2510 ACQUIRE_LOCK(&sched_mutex);
2511 scheduleThread_(tso);
2512 RELEASE_LOCK(&sched_mutex);
2515 #if defined(RTS_SUPPORTS_THREADS)
2516 static Condition bound_cond_cache;
2517 static int bound_cond_cache_full = 0;
2522 scheduleWaitThread(StgTSO* tso, /*[out]*/HaskellObj* ret,
2523 Capability *initialCapability)
2525 // Precondition: sched_mutex must be held
2528 m = stgMallocBytes(sizeof(StgMainThread), "waitThread");
2533 m->link = main_threads;
2535 if (main_threads != NULL) {
2536 main_threads->prev = m;
2540 #if defined(RTS_SUPPORTS_THREADS)
2541 // Allocating a new condition for each thread is expensive, so we
2542 // cache one. This is a pretty feeble hack, but it helps speed up
2543 // consecutive call-ins quite a bit.
2544 if (bound_cond_cache_full) {
2545 m->bound_thread_cond = bound_cond_cache;
2546 bound_cond_cache_full = 0;
2548 initCondition(&m->bound_thread_cond);
2552 /* Put the thread on the main-threads list prior to scheduling the TSO.
2553 Failure to do so introduces a race condition in the MT case (as
2554 identified by Wolfgang Thaller), whereby the new task/OS thread
2555 created by scheduleThread_() would complete prior to the thread
2556 that spawned it managed to put 'itself' on the main-threads list.
2557 The upshot of it all being that the worker thread wouldn't get to
2558 signal the completion of the its work item for the main thread to
2559 see (==> it got stuck waiting.) -- sof 6/02.
2561 IF_DEBUG(scheduler, sched_belch("waiting for thread (%d)", tso->id));
2563 APPEND_TO_RUN_QUEUE(tso);
2564 // NB. Don't call threadRunnable() here, because the thread is
2565 // bound and only runnable by *this* OS thread, so waking up other
2566 // workers will just slow things down.
2568 return waitThread_(m, initialCapability);
2571 /* ---------------------------------------------------------------------------
2574 * Initialise the scheduler. This resets all the queues - if the
2575 * queues contained any threads, they'll be garbage collected at the
2578 * ------------------------------------------------------------------------ */
2586 for (i=0; i<=MAX_PROC; i++) {
2587 run_queue_hds[i] = END_TSO_QUEUE;
2588 run_queue_tls[i] = END_TSO_QUEUE;
2589 blocked_queue_hds[i] = END_TSO_QUEUE;
2590 blocked_queue_tls[i] = END_TSO_QUEUE;
2591 ccalling_threadss[i] = END_TSO_QUEUE;
2592 blackhole_queue[i] = END_TSO_QUEUE;
2593 sleeping_queue = END_TSO_QUEUE;
2596 run_queue_hd = END_TSO_QUEUE;
2597 run_queue_tl = END_TSO_QUEUE;
2598 blocked_queue_hd = END_TSO_QUEUE;
2599 blocked_queue_tl = END_TSO_QUEUE;
2600 blackhole_queue = END_TSO_QUEUE;
2601 sleeping_queue = END_TSO_QUEUE;
2604 suspended_ccalling_threads = END_TSO_QUEUE;
2606 main_threads = NULL;
2607 all_threads = END_TSO_QUEUE;
2612 RtsFlags.ConcFlags.ctxtSwitchTicks =
2613 RtsFlags.ConcFlags.ctxtSwitchTime / TICK_MILLISECS;
2615 #if defined(RTS_SUPPORTS_THREADS)
2616 /* Initialise the mutex and condition variables used by
2618 initMutex(&sched_mutex);
2619 initMutex(&term_mutex);
2622 ACQUIRE_LOCK(&sched_mutex);
2624 /* A capability holds the state a native thread needs in
2625 * order to execute STG code. At least one capability is
2626 * floating around (only SMP builds have more than one).
2630 #if defined(RTS_SUPPORTS_THREADS)
2635 /* eagerly start some extra workers */
2636 startingWorkerThread = RtsFlags.ParFlags.nNodes;
2637 startTasks(RtsFlags.ParFlags.nNodes, taskStart);
2640 #if /* defined(SMP) ||*/ defined(PARALLEL_HASKELL)
2644 RELEASE_LOCK(&sched_mutex);
2648 exitScheduler( void )
2650 interrupted = rtsTrue;
2651 shutting_down_scheduler = rtsTrue;
2652 #if defined(RTS_SUPPORTS_THREADS)
2653 if (threadIsTask(osThreadId())) { taskStop(); }
2658 /* ----------------------------------------------------------------------------
2659 Managing the per-task allocation areas.
2661 Each capability comes with an allocation area. These are
2662 fixed-length block lists into which allocation can be done.
2664 ToDo: no support for two-space collection at the moment???
2665 ------------------------------------------------------------------------- */
2667 static SchedulerStatus
2668 waitThread_(StgMainThread* m, Capability *initialCapability)
2670 SchedulerStatus stat;
2672 // Precondition: sched_mutex must be held.
2673 IF_DEBUG(scheduler, sched_belch("new main thread (%d)", m->tso->id));
2676 /* GranSim specific init */
2677 CurrentTSO = m->tso; // the TSO to run
2678 procStatus[MainProc] = Busy; // status of main PE
2679 CurrentProc = MainProc; // PE to run it on
2680 schedule(m,initialCapability);
2682 schedule(m,initialCapability);
2683 ASSERT(m->stat != NoStatus);
2688 #if defined(RTS_SUPPORTS_THREADS)
2689 // Free the condition variable, returning it to the cache if possible.
2690 if (!bound_cond_cache_full) {
2691 bound_cond_cache = m->bound_thread_cond;
2692 bound_cond_cache_full = 1;
2694 closeCondition(&m->bound_thread_cond);
2698 IF_DEBUG(scheduler, sched_belch("main thread (%d) finished", m->tso->id));
2701 // Postcondition: sched_mutex still held
2705 /* ---------------------------------------------------------------------------
2706 Where are the roots that we know about?
2708 - all the threads on the runnable queue
2709 - all the threads on the blocked queue
2710 - all the threads on the sleeping queue
2711 - all the thread currently executing a _ccall_GC
2712 - all the "main threads"
2714 ------------------------------------------------------------------------ */
2716 /* This has to be protected either by the scheduler monitor, or by the
2717 garbage collection monitor (probably the latter).
2722 GetRoots( evac_fn evac )
2727 for (i=0; i<=RtsFlags.GranFlags.proc; i++) {
2728 if ((run_queue_hds[i] != END_TSO_QUEUE) && ((run_queue_hds[i] != NULL)))
2729 evac((StgClosure **)&run_queue_hds[i]);
2730 if ((run_queue_tls[i] != END_TSO_QUEUE) && ((run_queue_tls[i] != NULL)))
2731 evac((StgClosure **)&run_queue_tls[i]);
2733 if ((blocked_queue_hds[i] != END_TSO_QUEUE) && ((blocked_queue_hds[i] != NULL)))
2734 evac((StgClosure **)&blocked_queue_hds[i]);
2735 if ((blocked_queue_tls[i] != END_TSO_QUEUE) && ((blocked_queue_tls[i] != NULL)))
2736 evac((StgClosure **)&blocked_queue_tls[i]);
2737 if ((ccalling_threadss[i] != END_TSO_QUEUE) && ((ccalling_threadss[i] != NULL)))
2738 evac((StgClosure **)&ccalling_threads[i]);
2745 if (run_queue_hd != END_TSO_QUEUE) {
2746 ASSERT(run_queue_tl != END_TSO_QUEUE);
2747 evac((StgClosure **)&run_queue_hd);
2748 evac((StgClosure **)&run_queue_tl);
2751 if (blocked_queue_hd != END_TSO_QUEUE) {
2752 ASSERT(blocked_queue_tl != END_TSO_QUEUE);
2753 evac((StgClosure **)&blocked_queue_hd);
2754 evac((StgClosure **)&blocked_queue_tl);
2757 if (sleeping_queue != END_TSO_QUEUE) {
2758 evac((StgClosure **)&sleeping_queue);
2762 if (blackhole_queue != END_TSO_QUEUE) {
2763 evac((StgClosure **)&blackhole_queue);
2766 if (suspended_ccalling_threads != END_TSO_QUEUE) {
2767 evac((StgClosure **)&suspended_ccalling_threads);
2770 #if defined(PARALLEL_HASKELL) || defined(GRAN)
2771 markSparkQueue(evac);
2774 #if defined(RTS_USER_SIGNALS)
2775 // mark the signal handlers (signals should be already blocked)
2776 markSignalHandlers(evac);
2780 /* -----------------------------------------------------------------------------
2783 This is the interface to the garbage collector from Haskell land.
2784 We provide this so that external C code can allocate and garbage
2785 collect when called from Haskell via _ccall_GC.
2787 It might be useful to provide an interface whereby the programmer
2788 can specify more roots (ToDo).
2790 This needs to be protected by the GC condition variable above. KH.
2791 -------------------------------------------------------------------------- */
2793 static void (*extra_roots)(evac_fn);
2798 /* Obligated to hold this lock upon entry */
2799 ACQUIRE_LOCK(&sched_mutex);
2800 GarbageCollect(GetRoots,rtsFalse);
2801 RELEASE_LOCK(&sched_mutex);
2805 performMajorGC(void)
2807 ACQUIRE_LOCK(&sched_mutex);
2808 GarbageCollect(GetRoots,rtsTrue);
2809 RELEASE_LOCK(&sched_mutex);
2813 AllRoots(evac_fn evac)
2815 GetRoots(evac); // the scheduler's roots
2816 extra_roots(evac); // the user's roots
2820 performGCWithRoots(void (*get_roots)(evac_fn))
2822 ACQUIRE_LOCK(&sched_mutex);
2823 extra_roots = get_roots;
2824 GarbageCollect(AllRoots,rtsFalse);
2825 RELEASE_LOCK(&sched_mutex);
2828 /* -----------------------------------------------------------------------------
2831 If the thread has reached its maximum stack size, then raise the
2832 StackOverflow exception in the offending thread. Otherwise
2833 relocate the TSO into a larger chunk of memory and adjust its stack
2835 -------------------------------------------------------------------------- */
2838 threadStackOverflow(StgTSO *tso)
2840 nat new_stack_size, stack_words;
2845 IF_DEBUG(sanity,checkTSO(tso));
2846 if (tso->stack_size >= tso->max_stack_size) {
2849 debugBelch("@@ threadStackOverflow of TSO %ld (%p): stack too large (now %ld; max is %ld)\n",
2850 (long)tso->id, tso, (long)tso->stack_size, (long)tso->max_stack_size);
2851 /* If we're debugging, just print out the top of the stack */
2852 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2855 /* Send this thread the StackOverflow exception */
2856 raiseAsync(tso, (StgClosure *)stackOverflow_closure);
2860 /* Try to double the current stack size. If that takes us over the
2861 * maximum stack size for this thread, then use the maximum instead.
2862 * Finally round up so the TSO ends up as a whole number of blocks.
2864 new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
2865 new_tso_size = (lnat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) +
2866 TSO_STRUCT_SIZE)/sizeof(W_);
2867 new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */
2868 new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
2870 IF_DEBUG(scheduler, debugBelch("== sched: increasing stack size from %d words to %d.\n", tso->stack_size, new_stack_size));
2872 dest = (StgTSO *)allocate(new_tso_size);
2873 TICK_ALLOC_TSO(new_stack_size,0);
2875 /* copy the TSO block and the old stack into the new area */
2876 memcpy(dest,tso,TSO_STRUCT_SIZE);
2877 stack_words = tso->stack + tso->stack_size - tso->sp;
2878 new_sp = (P_)dest + new_tso_size - stack_words;
2879 memcpy(new_sp, tso->sp, stack_words * sizeof(W_));
2881 /* relocate the stack pointers... */
2883 dest->stack_size = new_stack_size;
2885 /* Mark the old TSO as relocated. We have to check for relocated
2886 * TSOs in the garbage collector and any primops that deal with TSOs.
2888 * It's important to set the sp value to just beyond the end
2889 * of the stack, so we don't attempt to scavenge any part of the
2892 tso->what_next = ThreadRelocated;
2894 tso->sp = (P_)&(tso->stack[tso->stack_size]);
2895 tso->why_blocked = NotBlocked;
2897 IF_PAR_DEBUG(verbose,
2898 debugBelch("@@ threadStackOverflow of TSO %d (now at %p): stack size increased to %ld\n",
2899 tso->id, tso, tso->stack_size);
2900 /* If we're debugging, just print out the top of the stack */
2901 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2904 IF_DEBUG(sanity,checkTSO(tso));
2906 IF_DEBUG(scheduler,printTSO(dest));
2912 /* ---------------------------------------------------------------------------
2913 Wake up a queue that was blocked on some resource.
2914 ------------------------------------------------------------------------ */
2918 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2921 #elif defined(PARALLEL_HASKELL)
2923 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2925 /* write RESUME events to log file and
2926 update blocked and fetch time (depending on type of the orig closure) */
2927 if (RtsFlags.ParFlags.ParStats.Full) {
2928 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
2929 GR_RESUMEQ, ((StgTSO *)bqe), ((StgTSO *)bqe)->block_info.closure,
2930 0, 0 /* spark_queue_len(ADVISORY_POOL) */);
2931 if (EMPTY_RUN_QUEUE())
2932 emitSchedule = rtsTrue;
2934 switch (get_itbl(node)->type) {
2936 ((StgTSO *)bqe)->par.fetchtime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2941 ((StgTSO *)bqe)->par.blocktime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2948 barf("{unblockOneLocked}Daq Qagh: unexpected closure in blocking queue");
2955 static StgBlockingQueueElement *
2956 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2959 PEs node_loc, tso_loc;
2961 node_loc = where_is(node); // should be lifted out of loop
2962 tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2963 tso_loc = where_is((StgClosure *)tso);
2964 if (IS_LOCAL_TO(PROCS(node),tso_loc)) { // TSO is local
2965 /* !fake_fetch => TSO is on CurrentProc is same as IS_LOCAL_TO */
2966 ASSERT(CurrentProc!=node_loc || tso_loc==CurrentProc);
2967 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.lunblocktime;
2968 // insertThread(tso, node_loc);
2969 new_event(tso_loc, tso_loc, CurrentTime[CurrentProc],
2971 tso, node, (rtsSpark*)NULL);
2972 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2975 } else { // TSO is remote (actually should be FMBQ)
2976 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.mpacktime +
2977 RtsFlags.GranFlags.Costs.gunblocktime +
2978 RtsFlags.GranFlags.Costs.latency;
2979 new_event(tso_loc, CurrentProc, CurrentTime[CurrentProc],
2981 tso, node, (rtsSpark*)NULL);
2982 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2985 /* the thread-queue-overhead is accounted for in either Resume or UnblockThread */
2987 debugBelch(" %s TSO %d (%p) [PE %d] (block_info.closure=%p) (next=%p) ,",
2988 (node_loc==tso_loc ? "Local" : "Global"),
2989 tso->id, tso, CurrentProc, tso->block_info.closure, tso->link));
2990 tso->block_info.closure = NULL;
2991 IF_DEBUG(scheduler,debugBelch("-- Waking up thread %ld (%p)\n",
2994 #elif defined(PARALLEL_HASKELL)
2995 static StgBlockingQueueElement *
2996 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2998 StgBlockingQueueElement *next;
3000 switch (get_itbl(bqe)->type) {
3002 ASSERT(((StgTSO *)bqe)->why_blocked != NotBlocked);
3003 /* if it's a TSO just push it onto the run_queue */
3005 ((StgTSO *)bqe)->link = END_TSO_QUEUE; // debugging?
3006 APPEND_TO_RUN_QUEUE((StgTSO *)bqe);
3008 unblockCount(bqe, node);
3009 /* reset blocking status after dumping event */
3010 ((StgTSO *)bqe)->why_blocked = NotBlocked;
3014 /* if it's a BLOCKED_FETCH put it on the PendingFetches list */
3016 bqe->link = (StgBlockingQueueElement *)PendingFetches;
3017 PendingFetches = (StgBlockedFetch *)bqe;
3021 /* can ignore this case in a non-debugging setup;
3022 see comments on RBHSave closures above */
3024 /* check that the closure is an RBHSave closure */
3025 ASSERT(get_itbl((StgClosure *)bqe) == &stg_RBH_Save_0_info ||
3026 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_1_info ||
3027 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_2_info);
3031 barf("{unblockOneLocked}Daq Qagh: Unexpected IP (%#lx; %s) in blocking queue at %#lx\n",
3032 get_itbl((StgClosure *)bqe), info_type((StgClosure *)bqe),
3036 IF_PAR_DEBUG(bq, debugBelch(", %p (%s)\n", bqe, info_type((StgClosure*)bqe)));
3040 #else /* !GRAN && !PARALLEL_HASKELL */
3042 unblockOneLocked(StgTSO *tso)
3046 ASSERT(get_itbl(tso)->type == TSO);
3047 ASSERT(tso->why_blocked != NotBlocked);
3048 tso->why_blocked = NotBlocked;
3050 tso->link = END_TSO_QUEUE;
3051 APPEND_TO_RUN_QUEUE(tso);
3053 IF_DEBUG(scheduler,sched_belch("waking up thread %ld", (long)tso->id));
3058 #if defined(GRAN) || defined(PARALLEL_HASKELL)
3059 INLINE_ME StgBlockingQueueElement *
3060 unblockOne(StgBlockingQueueElement *bqe, StgClosure *node)
3062 ACQUIRE_LOCK(&sched_mutex);
3063 bqe = unblockOneLocked(bqe, node);
3064 RELEASE_LOCK(&sched_mutex);
3069 unblockOne(StgTSO *tso)
3071 ACQUIRE_LOCK(&sched_mutex);
3072 tso = unblockOneLocked(tso);
3073 RELEASE_LOCK(&sched_mutex);
3080 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
3082 StgBlockingQueueElement *bqe;
3087 debugBelch("##-_ AwBQ for node %p on PE %d @ %ld by TSO %d (%p): \n", \
3088 node, CurrentProc, CurrentTime[CurrentProc],
3089 CurrentTSO->id, CurrentTSO));
3091 node_loc = where_is(node);
3093 ASSERT(q == END_BQ_QUEUE ||
3094 get_itbl(q)->type == TSO || // q is either a TSO or an RBHSave
3095 get_itbl(q)->type == CONSTR); // closure (type constructor)
3096 ASSERT(is_unique(node));
3098 /* FAKE FETCH: magically copy the node to the tso's proc;
3099 no Fetch necessary because in reality the node should not have been
3100 moved to the other PE in the first place
3102 if (CurrentProc!=node_loc) {
3104 debugBelch("## node %p is on PE %d but CurrentProc is %d (TSO %d); assuming fake fetch and adjusting bitmask (old: %#x)\n",
3105 node, node_loc, CurrentProc, CurrentTSO->id,
3106 // CurrentTSO, where_is(CurrentTSO),
3107 node->header.gran.procs));
3108 node->header.gran.procs = (node->header.gran.procs) | PE_NUMBER(CurrentProc);
3110 debugBelch("## new bitmask of node %p is %#x\n",
3111 node, node->header.gran.procs));
3112 if (RtsFlags.GranFlags.GranSimStats.Global) {
3113 globalGranStats.tot_fake_fetches++;
3118 // ToDo: check: ASSERT(CurrentProc==node_loc);
3119 while (get_itbl(bqe)->type==TSO) { // q != END_TSO_QUEUE) {
3122 bqe points to the current element in the queue
3123 next points to the next element in the queue
3125 //tso = (StgTSO *)bqe; // wastes an assignment to get the type right
3126 //tso_loc = where_is(tso);
3128 bqe = unblockOneLocked(bqe, node);
3131 /* if this is the BQ of an RBH, we have to put back the info ripped out of
3132 the closure to make room for the anchor of the BQ */
3133 if (bqe!=END_BQ_QUEUE) {
3134 ASSERT(get_itbl(node)->type == RBH && get_itbl(bqe)->type == CONSTR);
3136 ASSERT((info_ptr==&RBH_Save_0_info) ||
3137 (info_ptr==&RBH_Save_1_info) ||
3138 (info_ptr==&RBH_Save_2_info));
3140 /* cf. convertToRBH in RBH.c for writing the RBHSave closure */
3141 ((StgRBH *)node)->blocking_queue = (StgBlockingQueueElement *)((StgRBHSave *)bqe)->payload[0];
3142 ((StgRBH *)node)->mut_link = (StgMutClosure *)((StgRBHSave *)bqe)->payload[1];
3145 debugBelch("## Filled in RBH_Save for %p (%s) at end of AwBQ\n",
3146 node, info_type(node)));
3149 /* statistics gathering */
3150 if (RtsFlags.GranFlags.GranSimStats.Global) {
3151 // globalGranStats.tot_bq_processing_time += bq_processing_time;
3152 globalGranStats.tot_bq_len += len; // total length of all bqs awakened
3153 // globalGranStats.tot_bq_len_local += len_local; // same for local TSOs only
3154 globalGranStats.tot_awbq++; // total no. of bqs awakened
3157 debugBelch("## BQ Stats of %p: [%d entries] %s\n",
3158 node, len, (bqe!=END_BQ_QUEUE) ? "RBH" : ""));
3160 #elif defined(PARALLEL_HASKELL)
3162 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
3164 StgBlockingQueueElement *bqe;
3166 ACQUIRE_LOCK(&sched_mutex);
3168 IF_PAR_DEBUG(verbose,
3169 debugBelch("##-_ AwBQ for node %p on [%x]: \n",
3173 if(get_itbl(q)->type == CONSTR || q==END_BQ_QUEUE) {
3174 IF_PAR_DEBUG(verbose, debugBelch("## ... nothing to unblock so lets just return. RFP (BUG?)\n"));
3179 ASSERT(q == END_BQ_QUEUE ||
3180 get_itbl(q)->type == TSO ||
3181 get_itbl(q)->type == BLOCKED_FETCH ||
3182 get_itbl(q)->type == CONSTR);
3185 while (get_itbl(bqe)->type==TSO ||
3186 get_itbl(bqe)->type==BLOCKED_FETCH) {
3187 bqe = unblockOneLocked(bqe, node);
3189 RELEASE_LOCK(&sched_mutex);
3192 #else /* !GRAN && !PARALLEL_HASKELL */
3195 awakenBlockedQueueNoLock(StgTSO *tso)
3197 while (tso != END_TSO_QUEUE) {
3198 tso = unblockOneLocked(tso);
3203 awakenBlockedQueue(StgTSO *tso)
3205 ACQUIRE_LOCK(&sched_mutex);
3206 while (tso != END_TSO_QUEUE) {
3207 tso = unblockOneLocked(tso);
3209 RELEASE_LOCK(&sched_mutex);
3213 /* ---------------------------------------------------------------------------
3215 - usually called inside a signal handler so it mustn't do anything fancy.
3216 ------------------------------------------------------------------------ */
3219 interruptStgRts(void)
3225 /* -----------------------------------------------------------------------------
3228 This is for use when we raise an exception in another thread, which
3230 This has nothing to do with the UnblockThread event in GranSim. -- HWL
3231 -------------------------------------------------------------------------- */
3233 #if defined(GRAN) || defined(PARALLEL_HASKELL)
3235 NB: only the type of the blocking queue is different in GranSim and GUM
3236 the operations on the queue-elements are the same
3237 long live polymorphism!
3239 Locks: sched_mutex is held upon entry and exit.
3243 unblockThread(StgTSO *tso)
3245 StgBlockingQueueElement *t, **last;
3247 switch (tso->why_blocked) {
3250 return; /* not blocked */
3253 // Be careful: nothing to do here! We tell the scheduler that the thread
3254 // is runnable and we leave it to the stack-walking code to abort the
3255 // transaction while unwinding the stack. We should perhaps have a debugging
3256 // test to make sure that this really happens and that the 'zombie' transaction
3257 // does not get committed.
3261 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
3263 StgBlockingQueueElement *last_tso = END_BQ_QUEUE;
3264 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
3266 last = (StgBlockingQueueElement **)&mvar->head;
3267 for (t = (StgBlockingQueueElement *)mvar->head;
3269 last = &t->link, last_tso = t, t = t->link) {
3270 if (t == (StgBlockingQueueElement *)tso) {
3271 *last = (StgBlockingQueueElement *)tso->link;
3272 if (mvar->tail == tso) {
3273 mvar->tail = (StgTSO *)last_tso;
3278 barf("unblockThread (MVAR): TSO not found");
3281 case BlockedOnBlackHole:
3282 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
3284 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
3286 last = &bq->blocking_queue;
3287 for (t = bq->blocking_queue;
3289 last = &t->link, t = t->link) {
3290 if (t == (StgBlockingQueueElement *)tso) {
3291 *last = (StgBlockingQueueElement *)tso->link;
3295 barf("unblockThread (BLACKHOLE): TSO not found");
3298 case BlockedOnException:
3300 StgTSO *target = tso->block_info.tso;
3302 ASSERT(get_itbl(target)->type == TSO);
3304 if (target->what_next == ThreadRelocated) {
3305 target = target->link;
3306 ASSERT(get_itbl(target)->type == TSO);
3309 ASSERT(target->blocked_exceptions != NULL);
3311 last = (StgBlockingQueueElement **)&target->blocked_exceptions;
3312 for (t = (StgBlockingQueueElement *)target->blocked_exceptions;
3314 last = &t->link, t = t->link) {
3315 ASSERT(get_itbl(t)->type == TSO);
3316 if (t == (StgBlockingQueueElement *)tso) {
3317 *last = (StgBlockingQueueElement *)tso->link;
3321 barf("unblockThread (Exception): TSO not found");
3325 case BlockedOnWrite:
3326 #if defined(mingw32_HOST_OS)
3327 case BlockedOnDoProc:
3330 /* take TSO off blocked_queue */
3331 StgBlockingQueueElement *prev = NULL;
3332 for (t = (StgBlockingQueueElement *)blocked_queue_hd; t != END_BQ_QUEUE;
3333 prev = t, t = t->link) {
3334 if (t == (StgBlockingQueueElement *)tso) {
3336 blocked_queue_hd = (StgTSO *)t->link;
3337 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
3338 blocked_queue_tl = END_TSO_QUEUE;
3341 prev->link = t->link;
3342 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
3343 blocked_queue_tl = (StgTSO *)prev;
3349 barf("unblockThread (I/O): TSO not found");
3352 case BlockedOnDelay:
3354 /* take TSO off sleeping_queue */
3355 StgBlockingQueueElement *prev = NULL;
3356 for (t = (StgBlockingQueueElement *)sleeping_queue; t != END_BQ_QUEUE;
3357 prev = t, t = t->link) {
3358 if (t == (StgBlockingQueueElement *)tso) {
3360 sleeping_queue = (StgTSO *)t->link;
3362 prev->link = t->link;
3367 barf("unblockThread (delay): TSO not found");
3371 barf("unblockThread");
3375 tso->link = END_TSO_QUEUE;
3376 tso->why_blocked = NotBlocked;
3377 tso->block_info.closure = NULL;
3378 PUSH_ON_RUN_QUEUE(tso);
3382 unblockThread(StgTSO *tso)
3386 /* To avoid locking unnecessarily. */
3387 if (tso->why_blocked == NotBlocked) {
3391 switch (tso->why_blocked) {
3394 // Be careful: nothing to do here! We tell the scheduler that the thread
3395 // is runnable and we leave it to the stack-walking code to abort the
3396 // transaction while unwinding the stack. We should perhaps have a debugging
3397 // test to make sure that this really happens and that the 'zombie' transaction
3398 // does not get committed.
3402 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
3404 StgTSO *last_tso = END_TSO_QUEUE;
3405 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
3408 for (t = mvar->head; t != END_TSO_QUEUE;
3409 last = &t->link, last_tso = t, t = t->link) {
3412 if (mvar->tail == tso) {
3413 mvar->tail = last_tso;
3418 barf("unblockThread (MVAR): TSO not found");
3421 case BlockedOnBlackHole:
3423 last = &blackhole_queue;
3424 for (t = blackhole_queue; t != END_TSO_QUEUE;
3425 last = &t->link, t = t->link) {
3431 barf("unblockThread (BLACKHOLE): TSO not found");
3434 case BlockedOnException:
3436 StgTSO *target = tso->block_info.tso;
3438 ASSERT(get_itbl(target)->type == TSO);
3440 while (target->what_next == ThreadRelocated) {
3441 target = target->link;
3442 ASSERT(get_itbl(target)->type == TSO);
3445 ASSERT(target->blocked_exceptions != NULL);
3447 last = &target->blocked_exceptions;
3448 for (t = target->blocked_exceptions; t != END_TSO_QUEUE;
3449 last = &t->link, t = t->link) {
3450 ASSERT(get_itbl(t)->type == TSO);
3456 barf("unblockThread (Exception): TSO not found");
3460 case BlockedOnWrite:
3461 #if defined(mingw32_HOST_OS)
3462 case BlockedOnDoProc:
3465 StgTSO *prev = NULL;
3466 for (t = blocked_queue_hd; t != END_TSO_QUEUE;
3467 prev = t, t = t->link) {
3470 blocked_queue_hd = t->link;
3471 if (blocked_queue_tl == t) {
3472 blocked_queue_tl = END_TSO_QUEUE;
3475 prev->link = t->link;
3476 if (blocked_queue_tl == t) {
3477 blocked_queue_tl = prev;
3483 barf("unblockThread (I/O): TSO not found");
3486 case BlockedOnDelay:
3488 StgTSO *prev = NULL;
3489 for (t = sleeping_queue; t != END_TSO_QUEUE;
3490 prev = t, t = t->link) {
3493 sleeping_queue = t->link;
3495 prev->link = t->link;
3500 barf("unblockThread (delay): TSO not found");
3504 barf("unblockThread");
3508 tso->link = END_TSO_QUEUE;
3509 tso->why_blocked = NotBlocked;
3510 tso->block_info.closure = NULL;
3511 APPEND_TO_RUN_QUEUE(tso);
3515 /* -----------------------------------------------------------------------------
3518 * Check the blackhole_queue for threads that can be woken up. We do
3519 * this periodically: before every GC, and whenever the run queue is
3522 * An elegant solution might be to just wake up all the blocked
3523 * threads with awakenBlockedQueue occasionally: they'll go back to
3524 * sleep again if the object is still a BLACKHOLE. Unfortunately this
3525 * doesn't give us a way to tell whether we've actually managed to
3526 * wake up any threads, so we would be busy-waiting.
3528 * -------------------------------------------------------------------------- */
3531 checkBlackHoles( void )
3534 rtsBool any_woke_up = rtsFalse;
3537 IF_DEBUG(scheduler, sched_belch("checking threads blocked on black holes"));
3539 // ASSUMES: sched_mutex
3540 prev = &blackhole_queue;
3541 t = blackhole_queue;
3542 while (t != END_TSO_QUEUE) {
3543 ASSERT(t->why_blocked == BlockedOnBlackHole);
3544 type = get_itbl(t->block_info.closure)->type;
3545 if (type != BLACKHOLE && type != CAF_BLACKHOLE) {
3546 t = unblockOneLocked(t);
3548 any_woke_up = rtsTrue;
3558 /* -----------------------------------------------------------------------------
3561 * The following function implements the magic for raising an
3562 * asynchronous exception in an existing thread.
3564 * We first remove the thread from any queue on which it might be
3565 * blocked. The possible blockages are MVARs and BLACKHOLE_BQs.
3567 * We strip the stack down to the innermost CATCH_FRAME, building
3568 * thunks in the heap for all the active computations, so they can
3569 * be restarted if necessary. When we reach a CATCH_FRAME, we build
3570 * an application of the handler to the exception, and push it on
3571 * the top of the stack.
3573 * How exactly do we save all the active computations? We create an
3574 * AP_STACK for every UpdateFrame on the stack. Entering one of these
3575 * AP_STACKs pushes everything from the corresponding update frame
3576 * upwards onto the stack. (Actually, it pushes everything up to the
3577 * next update frame plus a pointer to the next AP_STACK object.
3578 * Entering the next AP_STACK object pushes more onto the stack until we
3579 * reach the last AP_STACK object - at which point the stack should look
3580 * exactly as it did when we killed the TSO and we can continue
3581 * execution by entering the closure on top of the stack.
3583 * We can also kill a thread entirely - this happens if either (a) the
3584 * exception passed to raiseAsync is NULL, or (b) there's no
3585 * CATCH_FRAME on the stack. In either case, we strip the entire
3586 * stack and replace the thread with a zombie.
3588 * Locks: sched_mutex held upon entry nor exit.
3590 * -------------------------------------------------------------------------- */
3593 deleteThread(StgTSO *tso)
3595 if (tso->why_blocked != BlockedOnCCall &&
3596 tso->why_blocked != BlockedOnCCall_NoUnblockExc) {
3597 raiseAsync(tso,NULL);
3601 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
3603 deleteThreadImmediately(StgTSO *tso)
3604 { // for forkProcess only:
3605 // delete thread without giving it a chance to catch the KillThread exception
3607 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
3611 if (tso->why_blocked != BlockedOnCCall &&
3612 tso->why_blocked != BlockedOnCCall_NoUnblockExc) {
3616 tso->what_next = ThreadKilled;
3621 raiseAsyncWithLock(StgTSO *tso, StgClosure *exception)
3623 /* When raising async exs from contexts where sched_mutex isn't held;
3624 use raiseAsyncWithLock(). */
3625 ACQUIRE_LOCK(&sched_mutex);
3626 raiseAsync(tso,exception);
3627 RELEASE_LOCK(&sched_mutex);
3631 raiseAsync(StgTSO *tso, StgClosure *exception)
3633 raiseAsync_(tso, exception, rtsFalse);
3637 raiseAsync_(StgTSO *tso, StgClosure *exception, rtsBool stop_at_atomically)
3639 StgRetInfoTable *info;
3642 // Thread already dead?
3643 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
3648 sched_belch("raising exception in thread %ld.", (long)tso->id));
3650 // Remove it from any blocking queues
3655 // The stack freezing code assumes there's a closure pointer on
3656 // the top of the stack, so we have to arrange that this is the case...
3658 if (sp[0] == (W_)&stg_enter_info) {
3662 sp[0] = (W_)&stg_dummy_ret_closure;
3668 // 1. Let the top of the stack be the "current closure"
3670 // 2. Walk up the stack until we find either an UPDATE_FRAME or a
3673 // 3. If it's an UPDATE_FRAME, then make an AP_STACK containing the
3674 // current closure applied to the chunk of stack up to (but not
3675 // including) the update frame. This closure becomes the "current
3676 // closure". Go back to step 2.
3678 // 4. If it's a CATCH_FRAME, then leave the exception handler on
3679 // top of the stack applied to the exception.
3681 // 5. If it's a STOP_FRAME, then kill the thread.
3683 // NB: if we pass an ATOMICALLY_FRAME then abort the associated
3690 info = get_ret_itbl((StgClosure *)frame);
3692 while (info->i.type != UPDATE_FRAME
3693 && (info->i.type != CATCH_FRAME || exception == NULL)
3694 && info->i.type != STOP_FRAME
3695 && (info->i.type != ATOMICALLY_FRAME || stop_at_atomically == rtsFalse))
3697 if (info->i.type == CATCH_RETRY_FRAME || info->i.type == ATOMICALLY_FRAME) {
3698 // IF we find an ATOMICALLY_FRAME then we abort the
3699 // current transaction and propagate the exception. In
3700 // this case (unlike ordinary exceptions) we do not care
3701 // whether the transaction is valid or not because its
3702 // possible validity cannot have caused the exception
3703 // and will not be visible after the abort.
3705 debugBelch("Found atomically block delivering async exception\n"));
3706 stmAbortTransaction(tso -> trec);
3707 tso -> trec = stmGetEnclosingTRec(tso -> trec);
3709 frame += stack_frame_sizeW((StgClosure *)frame);
3710 info = get_ret_itbl((StgClosure *)frame);
3713 switch (info->i.type) {
3715 case ATOMICALLY_FRAME:
3716 ASSERT(stop_at_atomically);
3717 ASSERT(stmGetEnclosingTRec(tso->trec) == NO_TREC);
3718 stmCondemnTransaction(tso -> trec);
3722 // R1 is not a register: the return convention for IO in
3723 // this case puts the return value on the stack, so we
3724 // need to set up the stack to return to the atomically
3725 // frame properly...
3726 tso->sp = frame - 2;
3727 tso->sp[1] = (StgWord) &stg_NO_FINALIZER_closure; // why not?
3728 tso->sp[0] = (StgWord) &stg_ut_1_0_unreg_info;
3730 tso->what_next = ThreadRunGHC;
3734 // If we find a CATCH_FRAME, and we've got an exception to raise,
3735 // then build the THUNK raise(exception), and leave it on
3736 // top of the CATCH_FRAME ready to enter.
3740 StgCatchFrame *cf = (StgCatchFrame *)frame;
3744 // we've got an exception to raise, so let's pass it to the
3745 // handler in this frame.
3747 raise = (StgClosure *)allocate(sizeofW(StgClosure)+1);
3748 TICK_ALLOC_SE_THK(1,0);
3749 SET_HDR(raise,&stg_raise_info,cf->header.prof.ccs);
3750 raise->payload[0] = exception;
3752 // throw away the stack from Sp up to the CATCH_FRAME.
3756 /* Ensure that async excpetions are blocked now, so we don't get
3757 * a surprise exception before we get around to executing the
3760 if (tso->blocked_exceptions == NULL) {
3761 tso->blocked_exceptions = END_TSO_QUEUE;
3764 /* Put the newly-built THUNK on top of the stack, ready to execute
3765 * when the thread restarts.
3768 sp[-1] = (W_)&stg_enter_info;
3770 tso->what_next = ThreadRunGHC;
3771 IF_DEBUG(sanity, checkTSO(tso));
3780 // First build an AP_STACK consisting of the stack chunk above the
3781 // current update frame, with the top word on the stack as the
3784 words = frame - sp - 1;
3785 ap = (StgAP_STACK *)allocate(PAP_sizeW(words));
3788 ap->fun = (StgClosure *)sp[0];
3790 for(i=0; i < (nat)words; ++i) {
3791 ap->payload[i] = (StgClosure *)*sp++;
3794 SET_HDR(ap,&stg_AP_STACK_info,
3795 ((StgClosure *)frame)->header.prof.ccs /* ToDo */);
3796 TICK_ALLOC_UP_THK(words+1,0);
3799 debugBelch("sched: Updating ");
3800 printPtr((P_)((StgUpdateFrame *)frame)->updatee);
3801 debugBelch(" with ");
3802 printObj((StgClosure *)ap);
3805 // Replace the updatee with an indirection - happily
3806 // this will also wake up any threads currently
3807 // waiting on the result.
3809 // Warning: if we're in a loop, more than one update frame on
3810 // the stack may point to the same object. Be careful not to
3811 // overwrite an IND_OLDGEN in this case, because we'll screw
3812 // up the mutable lists. To be on the safe side, don't
3813 // overwrite any kind of indirection at all. See also
3814 // threadSqueezeStack in GC.c, where we have to make a similar
3817 if (!closure_IND(((StgUpdateFrame *)frame)->updatee)) {
3818 // revert the black hole
3819 UPD_IND_NOLOCK(((StgUpdateFrame *)frame)->updatee,
3822 sp += sizeofW(StgUpdateFrame) - 1;
3823 sp[0] = (W_)ap; // push onto stack
3828 // We've stripped the entire stack, the thread is now dead.
3829 sp += sizeofW(StgStopFrame);
3830 tso->what_next = ThreadKilled;
3841 /* -----------------------------------------------------------------------------
3842 raiseExceptionHelper
3844 This function is called by the raise# primitve, just so that we can
3845 move some of the tricky bits of raising an exception from C-- into
3846 C. Who knows, it might be a useful re-useable thing here too.
3847 -------------------------------------------------------------------------- */
3850 raiseExceptionHelper (StgTSO *tso, StgClosure *exception)
3852 StgClosure *raise_closure = NULL;
3854 StgRetInfoTable *info;
3856 // This closure represents the expression 'raise# E' where E
3857 // is the exception raise. It is used to overwrite all the
3858 // thunks which are currently under evaluataion.
3862 // LDV profiling: stg_raise_info has THUNK as its closure
3863 // type. Since a THUNK takes at least MIN_UPD_SIZE words in its
3864 // payload, MIN_UPD_SIZE is more approprate than 1. It seems that
3865 // 1 does not cause any problem unless profiling is performed.
3866 // However, when LDV profiling goes on, we need to linearly scan
3867 // small object pool, where raise_closure is stored, so we should
3868 // use MIN_UPD_SIZE.
3870 // raise_closure = (StgClosure *)RET_STGCALL1(P_,allocate,
3871 // sizeofW(StgClosure)+1);
3875 // Walk up the stack, looking for the catch frame. On the way,
3876 // we update any closures pointed to from update frames with the
3877 // raise closure that we just built.
3881 info = get_ret_itbl((StgClosure *)p);
3882 next = p + stack_frame_sizeW((StgClosure *)p);
3883 switch (info->i.type) {
3886 // Only create raise_closure if we need to.
3887 if (raise_closure == NULL) {
3889 (StgClosure *)allocate(sizeofW(StgClosure)+MIN_UPD_SIZE);
3890 SET_HDR(raise_closure, &stg_raise_info, CCCS);
3891 raise_closure->payload[0] = exception;
3893 UPD_IND(((StgUpdateFrame *)p)->updatee,raise_closure);
3897 case ATOMICALLY_FRAME:
3898 IF_DEBUG(stm, debugBelch("Found ATOMICALLY_FRAME at %p\n", p));
3900 return ATOMICALLY_FRAME;
3906 case CATCH_STM_FRAME:
3907 IF_DEBUG(stm, debugBelch("Found CATCH_STM_FRAME at %p\n", p));
3909 return CATCH_STM_FRAME;
3915 case CATCH_RETRY_FRAME:
3924 /* -----------------------------------------------------------------------------
3925 findRetryFrameHelper
3927 This function is called by the retry# primitive. It traverses the stack
3928 leaving tso->sp referring to the frame which should handle the retry.
3930 This should either be a CATCH_RETRY_FRAME (if the retry# is within an orElse#)
3931 or should be a ATOMICALLY_FRAME (if the retry# reaches the top level).
3933 We skip CATCH_STM_FRAMEs because retries are not considered to be exceptions,
3934 despite the similar implementation.
3936 We should not expect to see CATCH_FRAME or STOP_FRAME because those should
3937 not be created within memory transactions.
3938 -------------------------------------------------------------------------- */
3941 findRetryFrameHelper (StgTSO *tso)
3944 StgRetInfoTable *info;
3948 info = get_ret_itbl((StgClosure *)p);
3949 next = p + stack_frame_sizeW((StgClosure *)p);
3950 switch (info->i.type) {
3952 case ATOMICALLY_FRAME:
3953 IF_DEBUG(stm, debugBelch("Found ATOMICALLY_FRAME at %p during retrry\n", p));
3955 return ATOMICALLY_FRAME;
3957 case CATCH_RETRY_FRAME:
3958 IF_DEBUG(stm, debugBelch("Found CATCH_RETRY_FRAME at %p during retrry\n", p));
3960 return CATCH_RETRY_FRAME;
3962 case CATCH_STM_FRAME:
3964 ASSERT(info->i.type != CATCH_FRAME);
3965 ASSERT(info->i.type != STOP_FRAME);
3972 /* -----------------------------------------------------------------------------
3973 resurrectThreads is called after garbage collection on the list of
3974 threads found to be garbage. Each of these threads will be woken
3975 up and sent a signal: BlockedOnDeadMVar if the thread was blocked
3976 on an MVar, or NonTermination if the thread was blocked on a Black
3979 Locks: sched_mutex isn't held upon entry nor exit.
3980 -------------------------------------------------------------------------- */
3983 resurrectThreads( StgTSO *threads )
3987 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
3988 next = tso->global_link;
3989 tso->global_link = all_threads;
3991 IF_DEBUG(scheduler, sched_belch("resurrecting thread %d", tso->id));
3993 switch (tso->why_blocked) {
3995 case BlockedOnException:
3996 /* Called by GC - sched_mutex lock is currently held. */
3997 raiseAsync(tso,(StgClosure *)BlockedOnDeadMVar_closure);
3999 case BlockedOnBlackHole:
4000 raiseAsync(tso,(StgClosure *)NonTermination_closure);
4003 raiseAsync(tso,(StgClosure *)BlockedIndefinitely_closure);
4006 /* This might happen if the thread was blocked on a black hole
4007 * belonging to a thread that we've just woken up (raiseAsync
4008 * can wake up threads, remember...).
4012 barf("resurrectThreads: thread blocked in a strange way");
4017 /* ----------------------------------------------------------------------------
4018 * Debugging: why is a thread blocked
4019 * [Also provides useful information when debugging threaded programs
4020 * at the Haskell source code level, so enable outside of DEBUG. --sof 7/02]
4021 ------------------------------------------------------------------------- */
4024 printThreadBlockage(StgTSO *tso)
4026 switch (tso->why_blocked) {
4028 debugBelch("is blocked on read from fd %ld", tso->block_info.fd);
4030 case BlockedOnWrite:
4031 debugBelch("is blocked on write to fd %ld", tso->block_info.fd);
4033 #if defined(mingw32_HOST_OS)
4034 case BlockedOnDoProc:
4035 debugBelch("is blocked on proc (request: %ld)", tso->block_info.async_result->reqID);
4038 case BlockedOnDelay:
4039 debugBelch("is blocked until %ld", tso->block_info.target);
4042 debugBelch("is blocked on an MVar");
4044 case BlockedOnException:
4045 debugBelch("is blocked on delivering an exception to thread %d",
4046 tso->block_info.tso->id);
4048 case BlockedOnBlackHole:
4049 debugBelch("is blocked on a black hole");
4052 debugBelch("is not blocked");
4054 #if defined(PARALLEL_HASKELL)
4056 debugBelch("is blocked on global address; local FM_BQ is %p (%s)",
4057 tso->block_info.closure, info_type(tso->block_info.closure));
4059 case BlockedOnGA_NoSend:
4060 debugBelch("is blocked on global address (no send); local FM_BQ is %p (%s)",
4061 tso->block_info.closure, info_type(tso->block_info.closure));
4064 case BlockedOnCCall:
4065 debugBelch("is blocked on an external call");
4067 case BlockedOnCCall_NoUnblockExc:
4068 debugBelch("is blocked on an external call (exceptions were already blocked)");
4071 debugBelch("is blocked on an STM operation");
4074 barf("printThreadBlockage: strange tso->why_blocked: %d for TSO %d (%d)",
4075 tso->why_blocked, tso->id, tso);
4080 printThreadStatus(StgTSO *tso)
4082 switch (tso->what_next) {
4084 debugBelch("has been killed");
4086 case ThreadComplete:
4087 debugBelch("has completed");
4090 printThreadBlockage(tso);
4095 printAllThreads(void)
4100 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
4101 ullong_format_string(TIME_ON_PROC(CurrentProc),
4102 time_string, rtsFalse/*no commas!*/);
4104 debugBelch("all threads at [%s]:\n", time_string);
4105 # elif defined(PARALLEL_HASKELL)
4106 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
4107 ullong_format_string(CURRENT_TIME,
4108 time_string, rtsFalse/*no commas!*/);
4110 debugBelch("all threads at [%s]:\n", time_string);
4112 debugBelch("all threads:\n");
4115 for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
4116 debugBelch("\tthread %d @ %p ", t->id, (void *)t);
4119 void *label = lookupThreadLabel(t->id);
4120 if (label) debugBelch("[\"%s\"] ",(char *)label);
4123 printThreadStatus(t);
4131 Print a whole blocking queue attached to node (debugging only).
4133 # if defined(PARALLEL_HASKELL)
4135 print_bq (StgClosure *node)
4137 StgBlockingQueueElement *bqe;
4141 debugBelch("## BQ of closure %p (%s): ",
4142 node, info_type(node));
4144 /* should cover all closures that may have a blocking queue */
4145 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
4146 get_itbl(node)->type == FETCH_ME_BQ ||
4147 get_itbl(node)->type == RBH ||
4148 get_itbl(node)->type == MVAR);
4150 ASSERT(node!=(StgClosure*)NULL); // sanity check
4152 print_bqe(((StgBlockingQueue*)node)->blocking_queue);
4156 Print a whole blocking queue starting with the element bqe.
4159 print_bqe (StgBlockingQueueElement *bqe)
4164 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
4166 for (end = (bqe==END_BQ_QUEUE);
4167 !end; // iterate until bqe points to a CONSTR
4168 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE),
4169 bqe = end ? END_BQ_QUEUE : bqe->link) {
4170 ASSERT(bqe != END_BQ_QUEUE); // sanity check
4171 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
4172 /* types of closures that may appear in a blocking queue */
4173 ASSERT(get_itbl(bqe)->type == TSO ||
4174 get_itbl(bqe)->type == BLOCKED_FETCH ||
4175 get_itbl(bqe)->type == CONSTR);
4176 /* only BQs of an RBH end with an RBH_Save closure */
4177 //ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
4179 switch (get_itbl(bqe)->type) {
4181 debugBelch(" TSO %u (%x),",
4182 ((StgTSO *)bqe)->id, ((StgTSO *)bqe));
4185 debugBelch(" BF (node=%p, ga=((%x, %d, %x)),",
4186 ((StgBlockedFetch *)bqe)->node,
4187 ((StgBlockedFetch *)bqe)->ga.payload.gc.gtid,
4188 ((StgBlockedFetch *)bqe)->ga.payload.gc.slot,
4189 ((StgBlockedFetch *)bqe)->ga.weight);
4192 debugBelch(" %s (IP %p),",
4193 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
4194 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
4195 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
4196 "RBH_Save_?"), get_itbl(bqe));
4199 barf("Unexpected closure type %s in blocking queue", // of %p (%s)",
4200 info_type((StgClosure *)bqe)); // , node, info_type(node));
4206 # elif defined(GRAN)
4208 print_bq (StgClosure *node)
4210 StgBlockingQueueElement *bqe;
4211 PEs node_loc, tso_loc;
4214 /* should cover all closures that may have a blocking queue */
4215 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
4216 get_itbl(node)->type == FETCH_ME_BQ ||
4217 get_itbl(node)->type == RBH);
4219 ASSERT(node!=(StgClosure*)NULL); // sanity check
4220 node_loc = where_is(node);
4222 debugBelch("## BQ of closure %p (%s) on [PE %d]: ",
4223 node, info_type(node), node_loc);
4226 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
4228 for (bqe = ((StgBlockingQueue*)node)->blocking_queue, end = (bqe==END_BQ_QUEUE);
4229 !end; // iterate until bqe points to a CONSTR
4230 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) {
4231 ASSERT(bqe != END_BQ_QUEUE); // sanity check
4232 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
4233 /* types of closures that may appear in a blocking queue */
4234 ASSERT(get_itbl(bqe)->type == TSO ||
4235 get_itbl(bqe)->type == CONSTR);
4236 /* only BQs of an RBH end with an RBH_Save closure */
4237 ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
4239 tso_loc = where_is((StgClosure *)bqe);
4240 switch (get_itbl(bqe)->type) {
4242 debugBelch(" TSO %d (%p) on [PE %d],",
4243 ((StgTSO *)bqe)->id, (StgTSO *)bqe, tso_loc);
4246 debugBelch(" %s (IP %p),",
4247 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
4248 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
4249 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
4250 "RBH_Save_?"), get_itbl(bqe));
4253 barf("Unexpected closure type %s in blocking queue of %p (%s)",
4254 info_type((StgClosure *)bqe), node, info_type(node));
4262 #if defined(PARALLEL_HASKELL)
4269 for (i=0, tso=run_queue_hd;
4270 tso != END_TSO_QUEUE;
4279 sched_belch(char *s, ...)
4283 #ifdef RTS_SUPPORTS_THREADS
4284 debugBelch("sched (task %p): ", osThreadId());
4285 #elif defined(PARALLEL_HASKELL)
4288 debugBelch("sched: ");