1 /* ---------------------------------------------------------------------------
2 * $Id: Schedule.c,v 1.134 2002/03/12 13:57:11 simonmar Exp $
4 * (c) The GHC Team, 1998-2000
8 * Different GHC ways use this scheduler quite differently (see comments below)
9 * Here is the global picture:
11 * WAY Name CPP flag What's it for
12 * --------------------------------------
13 * mp GUM PAR Parallel execution on a distributed memory machine
14 * s SMP SMP Parallel execution on a shared memory machine
15 * mg GranSim GRAN Simulation of parallel execution
16 * md GUM/GdH DIST Distributed execution (based on GUM)
18 * --------------------------------------------------------------------------*/
20 //@node Main scheduling code, , ,
21 //@section Main scheduling code
24 * Version with scheduler monitor support for SMPs (WAY=s):
26 This design provides a high-level API to create and schedule threads etc.
27 as documented in the SMP design document.
29 It uses a monitor design controlled by a single mutex to exercise control
30 over accesses to shared data structures, and builds on the Posix threads
33 The majority of state is shared. In order to keep essential per-task state,
34 there is a Capability structure, which contains all the information
35 needed to run a thread: its STG registers, a pointer to its TSO, a
36 nursery etc. During STG execution, a pointer to the capability is
37 kept in a register (BaseReg).
39 In a non-SMP build, there is one global capability, namely MainRegTable.
43 * Version with support for distributed memory parallelism aka GUM (WAY=mp):
45 The main scheduling loop in GUM iterates until a finish message is received.
46 In that case a global flag @receivedFinish@ is set and this instance of
47 the RTS shuts down. See ghc/rts/parallel/HLComms.c:processMessages()
48 for the handling of incoming messages, such as PP_FINISH.
49 Note that in the parallel case we have a system manager that coordinates
50 different PEs, each of which are running one instance of the RTS.
51 See ghc/rts/parallel/SysMan.c for the main routine of the parallel program.
52 From this routine processes executing ghc/rts/Main.c are spawned. -- HWL
54 * Version with support for simulating parallel execution aka GranSim (WAY=mg):
56 The main scheduling code in GranSim is quite different from that in std
57 (concurrent) Haskell: while concurrent Haskell just iterates over the
58 threads in the runnable queue, GranSim is event driven, i.e. it iterates
59 over the events in the global event queue. -- HWL
64 //* Variables and Data structures::
65 //* Main scheduling loop::
66 //* Suspend and Resume::
68 //* Garbage Collextion Routines::
69 //* Blocking Queue Routines::
70 //* Exception Handling Routines::
71 //* Debugging Routines::
75 //@node Includes, Variables and Data structures, Main scheduling code, Main scheduling code
76 //@subsection Includes
78 #include "PosixSource.h"
85 #include "StgStartup.h"
88 #include "StgMiscClosures.h"
90 #include "Interpreter.h"
91 #include "Exception.h"
100 #include "Proftimer.h"
101 #include "ProfHeap.h"
103 #if defined(GRAN) || defined(PAR)
104 # include "GranSimRts.h"
105 # include "GranSim.h"
106 # include "ParallelRts.h"
107 # include "Parallel.h"
108 # include "ParallelDebug.h"
109 # include "FetchMe.h"
113 #include "Capability.h"
114 #include "OSThreads.h"
119 //@node Variables and Data structures, Prototypes, Includes, Main scheduling code
120 //@subsection Variables and Data structures
122 /* Main thread queue.
123 * Locks required: sched_mutex.
125 StgMainThread *main_threads;
128 * Locks required: sched_mutex.
132 StgTSO* ActiveTSO = NULL; /* for assigning system costs; GranSim-Light only */
133 /* rtsTime TimeOfNextEvent, EndOfTimeSlice; now in GranSim.c */
136 In GranSim we have a runnable and a blocked queue for each processor.
137 In order to minimise code changes new arrays run_queue_hds/tls
138 are created. run_queue_hd is then a short cut (macro) for
139 run_queue_hds[CurrentProc] (see GranSim.h).
142 StgTSO *run_queue_hds[MAX_PROC], *run_queue_tls[MAX_PROC];
143 StgTSO *blocked_queue_hds[MAX_PROC], *blocked_queue_tls[MAX_PROC];
144 StgTSO *ccalling_threadss[MAX_PROC];
145 /* We use the same global list of threads (all_threads) in GranSim as in
146 the std RTS (i.e. we are cheating). However, we don't use this list in
147 the GranSim specific code at the moment (so we are only potentially
152 StgTSO *run_queue_hd, *run_queue_tl;
153 StgTSO *blocked_queue_hd, *blocked_queue_tl;
154 StgTSO *sleeping_queue; /* perhaps replace with a hash table? */
158 /* Linked list of all threads.
159 * Used for detecting garbage collected threads.
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 static StgTSO *threadStackOverflow(StgTSO *tso);
171 /* KH: The following two flags are shared memory locations. There is no need
172 to lock them, since they are only unset at the end of a scheduler
176 /* flag set by signal handler to precipitate a context switch */
177 //@cindex context_switch
180 /* if this flag is set as well, give up execution */
181 //@cindex interrupted
184 /* Next thread ID to allocate.
185 * Locks required: sched_mutex
187 //@cindex next_thread_id
188 StgThreadID next_thread_id = 1;
191 * Pointers to the state of the current thread.
192 * Rule of thumb: if CurrentTSO != NULL, then we're running a Haskell
193 * thread. If CurrentTSO == NULL, then we're at the scheduler level.
196 /* The smallest stack size that makes any sense is:
197 * RESERVED_STACK_WORDS (so we can get back from the stack overflow)
198 * + sizeofW(StgStopFrame) (the stg_stop_thread_info frame)
199 * + 1 (the realworld token for an IO thread)
200 * + 1 (the closure to enter)
202 * A thread with this stack will bomb immediately with a stack
203 * overflow, which will increase its stack size.
206 #define MIN_STACK_WORDS (RESERVED_STACK_WORDS + sizeofW(StgStopFrame) + 2)
213 /* This is used in `TSO.h' and gcc 2.96 insists that this variable actually
214 * exists - earlier gccs apparently didn't.
221 void addToBlockedQueue ( StgTSO *tso );
223 static void schedule ( void );
224 void interruptStgRts ( void );
226 static StgTSO * createThread_ ( nat size, rtsBool have_lock, StgInt pri );
228 static StgTSO * createThread_ ( nat size, rtsBool have_lock );
231 static void detectBlackHoles ( void );
234 static void sched_belch(char *s, ...);
237 #if defined(RTS_SUPPORTS_THREADS)
238 /* ToDo: carefully document the invariants that go together
239 * with these synchronisation objects.
241 Mutex sched_mutex = INIT_MUTEX_VAR;
242 Mutex term_mutex = INIT_MUTEX_VAR;
245 static Condition gc_pending_cond = INIT_COND_VAR;
249 #endif /* RTS_SUPPORTS_THREADS */
253 rtsTime TimeOfLastYield;
254 rtsBool emitSchedule = rtsTrue;
258 char *whatNext_strs[] = {
266 char *threadReturnCode_strs[] = {
267 "HeapOverflow", /* might also be StackOverflow */
276 StgTSO * createSparkThread(rtsSpark spark);
277 StgTSO * activateSpark (rtsSpark spark);
281 * The thread state for the main thread.
282 // ToDo: check whether not needed any more
286 #if defined(PAR) || defined(RTS_SUPPORTS_THREADS)
287 static void taskStart(void);
298 //@node Main scheduling loop, Suspend and Resume, Prototypes, Main scheduling code
299 //@subsection Main scheduling loop
301 /* ---------------------------------------------------------------------------
302 Main scheduling loop.
304 We use round-robin scheduling, each thread returning to the
305 scheduler loop when one of these conditions is detected:
308 * timer expires (thread yields)
313 Locking notes: we acquire the scheduler lock once at the beginning
314 of the scheduler loop, and release it when
316 * running a thread, or
317 * waiting for work, or
318 * waiting for a GC to complete.
321 In a GranSim setup this loop iterates over the global event queue.
322 This revolves around the global event queue, which determines what
323 to do next. Therefore, it's more complicated than either the
324 concurrent or the parallel (GUM) setup.
327 GUM iterates over incoming messages.
328 It starts with nothing to do (thus CurrentTSO == END_TSO_QUEUE),
329 and sends out a fish whenever it has nothing to do; in-between
330 doing the actual reductions (shared code below) it processes the
331 incoming messages and deals with delayed operations
332 (see PendingFetches).
333 This is not the ugliest code you could imagine, but it's bloody close.
335 ------------------------------------------------------------------------ */
342 StgThreadReturnCode ret;
350 rtsBool receivedFinish = rtsFalse;
352 nat tp_size, sp_size; // stats only
355 rtsBool was_interrupted = rtsFalse;
357 ACQUIRE_LOCK(&sched_mutex);
359 #if defined(RTS_SUPPORTS_THREADS)
360 /* Check to see whether there are any worker threads
361 waiting to deposit external call results. If so,
362 yield our capability */
363 yieldToReturningWorker(&sched_mutex, cap);
365 waitForWorkCapability(&sched_mutex, &cap, rtsFalse);
369 /* set up first event to get things going */
370 /* ToDo: assign costs for system setup and init MainTSO ! */
371 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
373 CurrentTSO, (StgClosure*)NULL, (rtsSpark*)NULL);
376 fprintf(stderr, "GRAN: Init CurrentTSO (in schedule) = %p\n", CurrentTSO);
377 G_TSO(CurrentTSO, 5));
379 if (RtsFlags.GranFlags.Light) {
380 /* Save current time; GranSim Light only */
381 CurrentTSO->gran.clock = CurrentTime[CurrentProc];
384 event = get_next_event();
386 while (event!=(rtsEvent*)NULL) {
387 /* Choose the processor with the next event */
388 CurrentProc = event->proc;
389 CurrentTSO = event->tso;
393 while (!receivedFinish) { /* set by processMessages */
394 /* when receiving PP_FINISH message */
401 IF_DEBUG(scheduler, printAllThreads());
403 /* If we're interrupted (the user pressed ^C, or some other
404 * termination condition occurred), kill all the currently running
408 IF_DEBUG(scheduler, sched_belch("interrupted"));
410 interrupted = rtsFalse;
411 was_interrupted = rtsTrue;
414 /* Go through the list of main threads and wake up any
415 * clients whose computations have finished. ToDo: this
416 * should be done more efficiently without a linear scan
417 * of the main threads list, somehow...
419 #if defined(RTS_SUPPORTS_THREADS)
421 StgMainThread *m, **prev;
422 prev = &main_threads;
423 for (m = main_threads; m != NULL; m = m->link) {
424 switch (m->tso->what_next) {
427 *(m->ret) = (StgClosure *)m->tso->sp[0];
431 broadcastCondition(&m->wakeup);
434 if (m->ret) *(m->ret) = NULL;
436 if (was_interrupted) {
437 m->stat = Interrupted;
441 broadcastCondition(&m->wakeup);
449 #else /* not threaded */
452 /* in GUM do this only on the Main PE */
455 /* If our main thread has finished or been killed, return.
458 StgMainThread *m = main_threads;
459 if (m->tso->what_next == ThreadComplete
460 || m->tso->what_next == ThreadKilled) {
461 main_threads = main_threads->link;
462 if (m->tso->what_next == ThreadComplete) {
463 /* we finished successfully, fill in the return value */
464 if (m->ret) { *(m->ret) = (StgClosure *)m->tso->sp[0]; };
468 if (m->ret) { *(m->ret) = NULL; };
469 if (was_interrupted) {
470 m->stat = Interrupted;
480 /* Top up the run queue from our spark pool. We try to make the
481 * number of threads in the run queue equal to the number of
484 * Disable spark support in SMP for now, non-essential & requires
485 * a little bit of work to make it compile cleanly. -- sof 1/02.
487 #if 0 /* defined(SMP) */
489 nat n = getFreeCapabilities();
490 StgTSO *tso = run_queue_hd;
492 /* Count the run queue */
493 while (n > 0 && tso != END_TSO_QUEUE) {
500 spark = findSpark(rtsFalse);
502 break; /* no more sparks in the pool */
504 /* I'd prefer this to be done in activateSpark -- HWL */
505 /* tricky - it needs to hold the scheduler lock and
506 * not try to re-acquire it -- SDM */
507 createSparkThread(spark);
509 sched_belch("==^^ turning spark of closure %p into a thread",
510 (StgClosure *)spark));
513 /* We need to wake up the other tasks if we just created some
516 if (getFreeCapabilities() - n > 1) {
517 signalCondition( &thread_ready_cond );
522 /* check for signals each time around the scheduler */
523 #ifndef mingw32_TARGET_OS
524 if (signals_pending()) {
525 startSignalHandlers();
529 /* Check whether any waiting threads need to be woken up. If the
530 * run queue is empty, and there are no other tasks running, we
531 * can wait indefinitely for something to happen.
532 * ToDo: what if another client comes along & requests another
535 if ( !EMPTY_QUEUE(blocked_queue_hd) || !EMPTY_QUEUE(sleeping_queue) ) {
536 awaitEvent( EMPTY_RUN_QUEUE()
538 && allFreeCapabilities()
542 /* we can be interrupted while waiting for I/O... */
543 if (interrupted) continue;
546 * Detect deadlock: when we have no threads to run, there are no
547 * threads waiting on I/O or sleeping, and all the other tasks are
548 * waiting for work, we must have a deadlock of some description.
550 * We first try to find threads blocked on themselves (ie. black
551 * holes), and generate NonTermination exceptions where necessary.
553 * If no threads are black holed, we have a deadlock situation, so
554 * inform all the main threads.
557 if ( EMPTY_THREAD_QUEUES()
558 #if defined(RTS_SUPPORTS_THREADS)
559 && EMPTY_QUEUE(suspended_ccalling_threads)
562 && allFreeCapabilities()
566 IF_DEBUG(scheduler, sched_belch("deadlocked, forcing major GC..."));
567 #if defined(THREADED_RTS)
568 /* and SMP mode ..? */
569 releaseCapability(cap);
571 // Garbage collection can release some new threads due to
572 // either (a) finalizers or (b) threads resurrected because
573 // they are about to be send BlockedOnDeadMVar. Any threads
574 // thus released will be immediately runnable.
575 GarbageCollect(GetRoots,rtsTrue);
577 if ( !EMPTY_RUN_QUEUE() ) { goto not_deadlocked; }
580 sched_belch("still deadlocked, checking for black holes..."));
583 if ( !EMPTY_RUN_QUEUE() ) { goto not_deadlocked; }
585 #ifndef mingw32_TARGET_OS
586 /* If we have user-installed signal handlers, then wait
587 * for signals to arrive rather then bombing out with a
590 if ( anyUserHandlers() ) {
592 sched_belch("still deadlocked, waiting for signals..."));
596 // we might be interrupted...
597 if (interrupted) { continue; }
599 if (signals_pending()) {
600 startSignalHandlers();
602 ASSERT(!EMPTY_RUN_QUEUE());
607 /* Probably a real deadlock. Send the current main thread the
608 * Deadlock exception (or in the SMP build, send *all* main
609 * threads the deadlock exception, since none of them can make
614 #if defined(RTS_SUPPORTS_THREADS)
615 for (m = main_threads; m != NULL; m = m->link) {
616 switch (m->tso->why_blocked) {
617 case BlockedOnBlackHole:
618 raiseAsync(m->tso, (StgClosure *)NonTermination_closure);
620 case BlockedOnException:
622 raiseAsync(m->tso, (StgClosure *)Deadlock_closure);
625 barf("deadlock: main thread blocked in a strange way");
630 switch (m->tso->why_blocked) {
631 case BlockedOnBlackHole:
632 raiseAsync(m->tso, (StgClosure *)NonTermination_closure);
634 case BlockedOnException:
636 raiseAsync(m->tso, (StgClosure *)Deadlock_closure);
639 barf("deadlock: main thread blocked in a strange way");
644 #if defined(RTS_SUPPORTS_THREADS)
645 /* ToDo: revisit conditions (and mechanism) for shutting
646 down a multi-threaded world */
647 IF_DEBUG(scheduler, sched_belch("all done, i think...shutting down."));
648 shutdownHaskellAndExit(0);
654 /* ToDo: add deadlock detection in GUM (similar to SMP) -- HWL */
658 /* If there's a GC pending, don't do anything until it has
662 IF_DEBUG(scheduler,sched_belch("waiting for GC"));
663 waitCondition( &gc_pending_cond, &sched_mutex );
667 #if defined(RTS_SUPPORTS_THREADS)
668 /* block until we've got a thread on the run queue and a free
672 if ( EMPTY_RUN_QUEUE() ) {
673 /* Give up our capability */
674 releaseCapability(cap);
675 IF_DEBUG(scheduler, sched_belch("thread %d: waiting for work", osThreadId()));
676 waitForWorkCapability(&sched_mutex, &cap, rtsTrue);
677 IF_DEBUG(scheduler, sched_belch("thread %d: work now available", osThreadId()));
679 while ( EMPTY_RUN_QUEUE() ) {
680 waitForWorkCapability(&sched_mutex, &cap);
681 IF_DEBUG(scheduler, sched_belch("thread %d: work now available", osThreadId()));
688 if (RtsFlags.GranFlags.Light)
689 GranSimLight_enter_system(event, &ActiveTSO); // adjust ActiveTSO etc
691 /* adjust time based on time-stamp */
692 if (event->time > CurrentTime[CurrentProc] &&
693 event->evttype != ContinueThread)
694 CurrentTime[CurrentProc] = event->time;
696 /* Deal with the idle PEs (may issue FindWork or MoveSpark events) */
697 if (!RtsFlags.GranFlags.Light)
700 IF_DEBUG(gran, fprintf(stderr, "GRAN: switch by event-type\n"));
702 /* main event dispatcher in GranSim */
703 switch (event->evttype) {
704 /* Should just be continuing execution */
706 IF_DEBUG(gran, fprintf(stderr, "GRAN: doing ContinueThread\n"));
707 /* ToDo: check assertion
708 ASSERT(run_queue_hd != (StgTSO*)NULL &&
709 run_queue_hd != END_TSO_QUEUE);
711 /* Ignore ContinueThreads for fetching threads (if synchr comm) */
712 if (!RtsFlags.GranFlags.DoAsyncFetch &&
713 procStatus[CurrentProc]==Fetching) {
714 belch("ghuH: Spurious ContinueThread while Fetching ignored; TSO %d (%p) [PE %d]",
715 CurrentTSO->id, CurrentTSO, CurrentProc);
718 /* Ignore ContinueThreads for completed threads */
719 if (CurrentTSO->what_next == ThreadComplete) {
720 belch("ghuH: found a ContinueThread event for completed thread %d (%p) [PE %d] (ignoring ContinueThread)",
721 CurrentTSO->id, CurrentTSO, CurrentProc);
724 /* Ignore ContinueThreads for threads that are being migrated */
725 if (PROCS(CurrentTSO)==Nowhere) {
726 belch("ghuH: trying to run the migrating TSO %d (%p) [PE %d] (ignoring ContinueThread)",
727 CurrentTSO->id, CurrentTSO, CurrentProc);
730 /* The thread should be at the beginning of the run queue */
731 if (CurrentTSO!=run_queue_hds[CurrentProc]) {
732 belch("ghuH: TSO %d (%p) [PE %d] is not at the start of the run_queue when doing a ContinueThread",
733 CurrentTSO->id, CurrentTSO, CurrentProc);
734 break; // run the thread anyway
737 new_event(proc, proc, CurrentTime[proc],
739 (StgTSO*)NULL, (StgClosure*)NULL, (rtsSpark*)NULL);
741 */ /* Catches superfluous CONTINUEs -- should be unnecessary */
742 break; // now actually run the thread; DaH Qu'vam yImuHbej
745 do_the_fetchnode(event);
746 goto next_thread; /* handle next event in event queue */
749 do_the_globalblock(event);
750 goto next_thread; /* handle next event in event queue */
753 do_the_fetchreply(event);
754 goto next_thread; /* handle next event in event queue */
756 case UnblockThread: /* Move from the blocked queue to the tail of */
757 do_the_unblock(event);
758 goto next_thread; /* handle next event in event queue */
760 case ResumeThread: /* Move from the blocked queue to the tail of */
761 /* the runnable queue ( i.e. Qu' SImqa'lu') */
762 event->tso->gran.blocktime +=
763 CurrentTime[CurrentProc] - event->tso->gran.blockedat;
764 do_the_startthread(event);
765 goto next_thread; /* handle next event in event queue */
768 do_the_startthread(event);
769 goto next_thread; /* handle next event in event queue */
772 do_the_movethread(event);
773 goto next_thread; /* handle next event in event queue */
776 do_the_movespark(event);
777 goto next_thread; /* handle next event in event queue */
780 do_the_findwork(event);
781 goto next_thread; /* handle next event in event queue */
784 barf("Illegal event type %u\n", event->evttype);
787 /* This point was scheduler_loop in the old RTS */
789 IF_DEBUG(gran, belch("GRAN: after main switch"));
791 TimeOfLastEvent = CurrentTime[CurrentProc];
792 TimeOfNextEvent = get_time_of_next_event();
793 IgnoreEvents=(TimeOfNextEvent==0); // HWL HACK
794 // CurrentTSO = ThreadQueueHd;
796 IF_DEBUG(gran, belch("GRAN: time of next event is: %ld",
799 if (RtsFlags.GranFlags.Light)
800 GranSimLight_leave_system(event, &ActiveTSO);
802 EndOfTimeSlice = CurrentTime[CurrentProc]+RtsFlags.GranFlags.time_slice;
805 belch("GRAN: end of time-slice is %#lx", EndOfTimeSlice));
807 /* in a GranSim setup the TSO stays on the run queue */
809 /* Take a thread from the run queue. */
810 t = POP_RUN_QUEUE(); // take_off_run_queue(t);
813 fprintf(stderr, "GRAN: About to run current thread, which is\n");
816 context_switch = 0; // turned on via GranYield, checking events and time slice
819 DumpGranEvent(GR_SCHEDULE, t));
821 procStatus[CurrentProc] = Busy;
824 if (PendingFetches != END_BF_QUEUE) {
828 /* ToDo: phps merge with spark activation above */
829 /* check whether we have local work and send requests if we have none */
830 if (EMPTY_RUN_QUEUE()) { /* no runnable threads */
831 /* :-[ no local threads => look out for local sparks */
832 /* the spark pool for the current PE */
833 pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable
834 if (advisory_thread_count < RtsFlags.ParFlags.maxThreads &&
835 pool->hd < pool->tl) {
837 * ToDo: add GC code check that we really have enough heap afterwards!!
839 * If we're here (no runnable threads) and we have pending
840 * sparks, we must have a space problem. Get enough space
841 * to turn one of those pending sparks into a
845 spark = findSpark(rtsFalse); /* get a spark */
846 if (spark != (rtsSpark) NULL) {
847 tso = activateSpark(spark); /* turn the spark into a thread */
848 IF_PAR_DEBUG(schedule,
849 belch("==== schedule: Created TSO %d (%p); %d threads active",
850 tso->id, tso, advisory_thread_count));
852 if (tso==END_TSO_QUEUE) { /* failed to activate spark->back to loop */
853 belch("==^^ failed to activate spark");
855 } /* otherwise fall through & pick-up new tso */
857 IF_PAR_DEBUG(verbose,
858 belch("==^^ no local sparks (spark pool contains only NFs: %d)",
859 spark_queue_len(pool)));
864 /* If we still have no work we need to send a FISH to get a spark
867 if (EMPTY_RUN_QUEUE()) {
868 /* =8-[ no local sparks => look for work on other PEs */
870 * We really have absolutely no work. Send out a fish
871 * (there may be some out there already), and wait for
872 * something to arrive. We clearly can't run any threads
873 * until a SCHEDULE or RESUME arrives, and so that's what
874 * we're hoping to see. (Of course, we still have to
875 * respond to other types of messages.)
877 TIME now = msTime() /*CURRENT_TIME*/;
878 IF_PAR_DEBUG(verbose,
879 belch("-- now=%ld", now));
880 IF_PAR_DEBUG(verbose,
881 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
882 (last_fish_arrived_at!=0 &&
883 last_fish_arrived_at+RtsFlags.ParFlags.fishDelay > now)) {
884 belch("--$$ delaying FISH until %ld (last fish %ld, delay %ld, now %ld)",
885 last_fish_arrived_at+RtsFlags.ParFlags.fishDelay,
886 last_fish_arrived_at,
887 RtsFlags.ParFlags.fishDelay, now);
890 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
891 (last_fish_arrived_at==0 ||
892 (last_fish_arrived_at+RtsFlags.ParFlags.fishDelay <= now))) {
893 /* outstandingFishes is set in sendFish, processFish;
894 avoid flooding system with fishes via delay */
896 sendFish(pe, mytid, NEW_FISH_AGE, NEW_FISH_HISTORY,
899 // Global statistics: count no. of fishes
900 if (RtsFlags.ParFlags.ParStats.Global &&
901 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
902 globalParStats.tot_fish_mess++;
906 receivedFinish = processMessages();
909 } else if (PacketsWaiting()) { /* Look for incoming messages */
910 receivedFinish = processMessages();
913 /* Now we are sure that we have some work available */
914 ASSERT(run_queue_hd != END_TSO_QUEUE);
916 /* Take a thread from the run queue, if we have work */
917 t = POP_RUN_QUEUE(); // take_off_run_queue(END_TSO_QUEUE);
918 IF_DEBUG(sanity,checkTSO(t));
920 /* ToDo: write something to the log-file
921 if (RTSflags.ParFlags.granSimStats && !sameThread)
922 DumpGranEvent(GR_SCHEDULE, RunnableThreadsHd);
926 /* the spark pool for the current PE */
927 pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable
930 belch("--=^ %d threads, %d sparks on [%#x]",
931 run_queue_len(), spark_queue_len(pool), CURRENT_PROC));
934 if (0 && RtsFlags.ParFlags.ParStats.Full &&
935 t && LastTSO && t->id != LastTSO->id &&
936 LastTSO->why_blocked == NotBlocked &&
937 LastTSO->what_next != ThreadComplete) {
938 // if previously scheduled TSO not blocked we have to record the context switch
939 DumpVeryRawGranEvent(TimeOfLastYield, CURRENT_PROC, CURRENT_PROC,
940 GR_DESCHEDULE, LastTSO, (StgClosure *)NULL, 0, 0);
943 if (RtsFlags.ParFlags.ParStats.Full &&
944 (emitSchedule /* forced emit */ ||
945 (t && LastTSO && t->id != LastTSO->id))) {
947 we are running a different TSO, so write a schedule event to log file
948 NB: If we use fair scheduling we also have to write a deschedule
949 event for LastTSO; with unfair scheduling we know that the
950 previous tso has blocked whenever we switch to another tso, so
951 we don't need it in GUM for now
953 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
954 GR_SCHEDULE, t, (StgClosure *)NULL, 0, 0);
955 emitSchedule = rtsFalse;
959 #else /* !GRAN && !PAR */
961 /* grab a thread from the run queue */
962 ASSERT(run_queue_hd != END_TSO_QUEUE);
964 // Sanity check the thread we're about to run. This can be
965 // expensive if there is lots of thread switching going on...
966 IF_DEBUG(sanity,checkTSO(t));
969 grabCapability(&cap);
970 cap->r.rCurrentTSO = t;
972 /* context switches are now initiated by the timer signal, unless
973 * the user specified "context switch as often as possible", with
978 RtsFlags.ProfFlags.profileInterval == 0 ||
980 (RtsFlags.ConcFlags.ctxtSwitchTicks == 0
981 && (run_queue_hd != END_TSO_QUEUE
982 || blocked_queue_hd != END_TSO_QUEUE
983 || sleeping_queue != END_TSO_QUEUE)))
988 RELEASE_LOCK(&sched_mutex);
990 IF_DEBUG(scheduler, sched_belch("-->> Running TSO %ld (%p) %s ...",
991 t->id, t, whatNext_strs[t->what_next]));
994 startHeapProfTimer();
997 /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
998 /* Run the current thread
1000 switch (cap->r.rCurrentTSO->what_next) {
1002 case ThreadComplete:
1003 /* Thread already finished, return to scheduler. */
1004 ret = ThreadFinished;
1006 case ThreadEnterGHC:
1007 ret = StgRun((StgFunPtr) stg_enterStackTop, &cap->r);
1010 ret = StgRun((StgFunPtr) stg_returnToStackTop, &cap->r);
1012 case ThreadEnterInterp:
1013 ret = interpretBCO(cap);
1016 barf("schedule: invalid what_next field");
1018 /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
1020 /* Costs for the scheduler are assigned to CCS_SYSTEM */
1022 stopHeapProfTimer();
1026 ACQUIRE_LOCK(&sched_mutex);
1029 IF_DEBUG(scheduler,fprintf(stderr,"scheduler (task %ld): ", osThreadId()););
1030 #elif !defined(GRAN) && !defined(PAR)
1031 IF_DEBUG(scheduler,fprintf(stderr,"scheduler: "););
1033 t = cap->r.rCurrentTSO;
1036 /* HACK 675: if the last thread didn't yield, make sure to print a
1037 SCHEDULE event to the log file when StgRunning the next thread, even
1038 if it is the same one as before */
1040 TimeOfLastYield = CURRENT_TIME;
1046 IF_DEBUG(gran, DumpGranEvent(GR_DESCHEDULE, t));
1047 globalGranStats.tot_heapover++;
1049 globalParStats.tot_heapover++;
1052 // did the task ask for a large block?
1053 if (cap->r.rHpAlloc > BLOCK_SIZE_W) {
1054 // if so, get one and push it on the front of the nursery.
1058 blocks = (nat)BLOCK_ROUND_UP(cap->r.rHpAlloc * sizeof(W_)) / BLOCK_SIZE;
1060 IF_DEBUG(scheduler,belch("--<< thread %ld (%p; %s) stopped: requesting a large block (size %d)",
1062 whatNext_strs[t->what_next], blocks));
1064 // don't do this if it would push us over the
1065 // alloc_blocks_lim limit; we'll GC first.
1066 if (alloc_blocks + blocks < alloc_blocks_lim) {
1068 alloc_blocks += blocks;
1069 bd = allocGroup( blocks );
1071 // link the new group into the list
1072 bd->link = cap->r.rCurrentNursery;
1073 bd->u.back = cap->r.rCurrentNursery->u.back;
1074 if (cap->r.rCurrentNursery->u.back != NULL) {
1075 cap->r.rCurrentNursery->u.back->link = bd;
1077 ASSERT(g0s0->blocks == cap->r.rCurrentNursery &&
1078 g0s0->blocks == cap->r.rNursery);
1079 cap->r.rNursery = g0s0->blocks = bd;
1081 cap->r.rCurrentNursery->u.back = bd;
1083 // initialise it as a nursery block
1087 bd->free = bd->start;
1089 // don't forget to update the block count in g0s0.
1090 g0s0->n_blocks += blocks;
1091 ASSERT(countBlocks(g0s0->blocks) == g0s0->n_blocks);
1093 // now update the nursery to point to the new block
1094 cap->r.rCurrentNursery = bd;
1096 // we might be unlucky and have another thread get on the
1097 // run queue before us and steal the large block, but in that
1098 // case the thread will just end up requesting another large
1100 PUSH_ON_RUN_QUEUE(t);
1105 /* make all the running tasks block on a condition variable,
1106 * maybe set context_switch and wait till they all pile in,
1107 * then have them wait on a GC condition variable.
1109 IF_DEBUG(scheduler,belch("--<< thread %ld (%p; %s) stopped: HeapOverflow",
1110 t->id, t, whatNext_strs[t->what_next]));
1113 ASSERT(!is_on_queue(t,CurrentProc));
1115 /* Currently we emit a DESCHEDULE event before GC in GUM.
1116 ToDo: either add separate event to distinguish SYSTEM time from rest
1117 or just nuke this DESCHEDULE (and the following SCHEDULE) */
1118 if (0 && RtsFlags.ParFlags.ParStats.Full) {
1119 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
1120 GR_DESCHEDULE, t, (StgClosure *)NULL, 0, 0);
1121 emitSchedule = rtsTrue;
1125 ready_to_gc = rtsTrue;
1126 context_switch = 1; /* stop other threads ASAP */
1127 PUSH_ON_RUN_QUEUE(t);
1128 /* actual GC is done at the end of the while loop */
1134 DumpGranEvent(GR_DESCHEDULE, t));
1135 globalGranStats.tot_stackover++;
1138 // DumpGranEvent(GR_DESCHEDULE, t);
1139 globalParStats.tot_stackover++;
1141 IF_DEBUG(scheduler,belch("--<< thread %ld (%p; %s) stopped, StackOverflow",
1142 t->id, t, whatNext_strs[t->what_next]));
1143 /* just adjust the stack for this thread, then pop it back
1149 /* enlarge the stack */
1150 StgTSO *new_t = threadStackOverflow(t);
1152 /* This TSO has moved, so update any pointers to it from the
1153 * main thread stack. It better not be on any other queues...
1154 * (it shouldn't be).
1156 for (m = main_threads; m != NULL; m = m->link) {
1161 threadPaused(new_t);
1162 PUSH_ON_RUN_QUEUE(new_t);
1166 case ThreadYielding:
1169 DumpGranEvent(GR_DESCHEDULE, t));
1170 globalGranStats.tot_yields++;
1173 // DumpGranEvent(GR_DESCHEDULE, t);
1174 globalParStats.tot_yields++;
1176 /* put the thread back on the run queue. Then, if we're ready to
1177 * GC, check whether this is the last task to stop. If so, wake
1178 * up the GC thread. getThread will block during a GC until the
1182 if (t->what_next == ThreadEnterInterp) {
1183 /* ToDo: or maybe a timer expired when we were in Hugs?
1184 * or maybe someone hit ctrl-C
1186 belch("--<< thread %ld (%p; %s) stopped to switch to Hugs",
1187 t->id, t, whatNext_strs[t->what_next]);
1189 belch("--<< thread %ld (%p; %s) stopped, yielding",
1190 t->id, t, whatNext_strs[t->what_next]);
1197 //belch("&& Doing sanity check on yielding TSO %ld.", t->id);
1199 ASSERT(t->link == END_TSO_QUEUE);
1201 ASSERT(!is_on_queue(t,CurrentProc));
1204 //belch("&& Doing sanity check on all ThreadQueues (and their TSOs).");
1205 checkThreadQsSanity(rtsTrue));
1208 if (RtsFlags.ParFlags.doFairScheduling) {
1209 /* this does round-robin scheduling; good for concurrency */
1210 APPEND_TO_RUN_QUEUE(t);
1212 /* this does unfair scheduling; good for parallelism */
1213 PUSH_ON_RUN_QUEUE(t);
1216 /* this does round-robin scheduling; good for concurrency */
1217 APPEND_TO_RUN_QUEUE(t);
1220 /* add a ContinueThread event to actually process the thread */
1221 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1223 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1225 belch("GRAN: eventq and runnableq after adding yielded thread to queue again:");
1234 belch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: ",
1235 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)));
1236 if (t->block_info.closure!=(StgClosure*)NULL) print_bq(t->block_info.closure));
1238 // ??? needed; should emit block before
1240 DumpGranEvent(GR_DESCHEDULE, t));
1241 prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
1244 ASSERT(procStatus[CurrentProc]==Busy ||
1245 ((procStatus[CurrentProc]==Fetching) &&
1246 (t->block_info.closure!=(StgClosure*)NULL)));
1247 if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
1248 !(!RtsFlags.GranFlags.DoAsyncFetch &&
1249 procStatus[CurrentProc]==Fetching))
1250 procStatus[CurrentProc] = Idle;
1254 belch("--<< thread %ld (%p; %s) stopped, blocking on node %p with BQ: ",
1255 t->id, t, whatNext_strs[t->what_next], t->block_info.closure));
1258 if (t->block_info.closure!=(StgClosure*)NULL)
1259 print_bq(t->block_info.closure));
1261 /* Send a fetch (if BlockedOnGA) and dump event to log file */
1264 /* whatever we schedule next, we must log that schedule */
1265 emitSchedule = rtsTrue;
1268 /* don't need to do anything. Either the thread is blocked on
1269 * I/O, in which case we'll have called addToBlockedQueue
1270 * previously, or it's blocked on an MVar or Blackhole, in which
1271 * case it'll be on the relevant queue already.
1274 fprintf(stderr, "--<< thread %d (%p) stopped: ", t->id, t);
1275 printThreadBlockage(t);
1276 fprintf(stderr, "\n"));
1278 /* Only for dumping event to log file
1279 ToDo: do I need this in GranSim, too?
1286 case ThreadFinished:
1287 /* Need to check whether this was a main thread, and if so, signal
1288 * the task that started it with the return value. If we have no
1289 * more main threads, we probably need to stop all the tasks until
1292 /* We also end up here if the thread kills itself with an
1293 * uncaught exception, see Exception.hc.
1295 IF_DEBUG(scheduler,belch("--++ thread %d (%p) finished", t->id, t));
1297 endThread(t, CurrentProc); // clean-up the thread
1299 /* For now all are advisory -- HWL */
1300 //if(t->priority==AdvisoryPriority) ??
1301 advisory_thread_count--;
1304 if(t->dist.priority==RevalPriority)
1308 if (RtsFlags.ParFlags.ParStats.Full &&
1309 !RtsFlags.ParFlags.ParStats.Suppressed)
1310 DumpEndEvent(CURRENT_PROC, t, rtsFalse /* not mandatory */);
1315 barf("schedule: invalid thread return code %d", (int)ret);
1318 #if defined(RTS_SUPPORTS_THREADS)
1319 /* I don't understand what this re-grab is doing -- sof */
1320 grabCapability(&cap);
1324 if (RtsFlags.ProfFlags.profileInterval==0 || performHeapProfile) {
1325 GarbageCollect(GetRoots, rtsTrue);
1327 performHeapProfile = rtsFalse;
1328 ready_to_gc = rtsFalse; // we already GC'd
1334 && allFreeCapabilities()
1337 /* everybody back, start the GC.
1338 * Could do it in this thread, or signal a condition var
1339 * to do it in another thread. Either way, we need to
1340 * broadcast on gc_pending_cond afterward.
1342 #if defined(RTS_SUPPORTS_THREADS)
1343 IF_DEBUG(scheduler,sched_belch("doing GC"));
1345 GarbageCollect(GetRoots,rtsFalse);
1346 ready_to_gc = rtsFalse;
1348 broadcastCondition(&gc_pending_cond);
1351 /* add a ContinueThread event to continue execution of current thread */
1352 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1354 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1356 fprintf(stderr, "GRAN: eventq and runnableq after Garbage collection:\n");
1364 IF_GRAN_DEBUG(unused,
1365 print_eventq(EventHd));
1367 event = get_next_event();
1370 /* ToDo: wait for next message to arrive rather than busy wait */
1373 } /* end of while(1) */
1375 IF_PAR_DEBUG(verbose,
1376 belch("== Leaving schedule() after having received Finish"));
1379 /* ---------------------------------------------------------------------------
1380 * deleteAllThreads(): kill all the live threads.
1382 * This is used when we catch a user interrupt (^C), before performing
1383 * any necessary cleanups and running finalizers.
1384 * ------------------------------------------------------------------------- */
1386 void deleteAllThreads ( void )
1389 IF_DEBUG(scheduler,sched_belch("deleting all threads"));
1390 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
1391 next = t->global_link;
1394 run_queue_hd = run_queue_tl = END_TSO_QUEUE;
1395 blocked_queue_hd = blocked_queue_tl = END_TSO_QUEUE;
1396 sleeping_queue = END_TSO_QUEUE;
1399 /* startThread and insertThread are now in GranSim.c -- HWL */
1402 //@node Suspend and Resume, Run queue code, Main scheduling loop, Main scheduling code
1403 //@subsection Suspend and Resume
1405 /* ---------------------------------------------------------------------------
1406 * Suspending & resuming Haskell threads.
1408 * When making a "safe" call to C (aka _ccall_GC), the task gives back
1409 * its capability before calling the C function. This allows another
1410 * task to pick up the capability and carry on running Haskell
1411 * threads. It also means that if the C call blocks, it won't lock
1414 * The Haskell thread making the C call is put to sleep for the
1415 * duration of the call, on the susepended_ccalling_threads queue. We
1416 * give out a token to the task, which it can use to resume the thread
1417 * on return from the C function.
1418 * ------------------------------------------------------------------------- */
1421 suspendThread( StgRegTable *reg,
1423 #if !defined(RTS_SUPPORTS_THREADS)
1431 /* assume that *reg is a pointer to the StgRegTable part
1434 cap = (Capability *)((void *)reg - sizeof(StgFunTable));
1436 ACQUIRE_LOCK(&sched_mutex);
1439 sched_belch("thread %d did a _ccall_gc", cap->r.rCurrentTSO->id));
1441 threadPaused(cap->r.rCurrentTSO);
1442 cap->r.rCurrentTSO->link = suspended_ccalling_threads;
1443 suspended_ccalling_threads = cap->r.rCurrentTSO;
1445 #if defined(RTS_SUPPORTS_THREADS)
1446 cap->r.rCurrentTSO->why_blocked = BlockedOnCCall;
1449 /* Use the thread ID as the token; it should be unique */
1450 tok = cap->r.rCurrentTSO->id;
1452 /* Hand back capability */
1453 releaseCapability(cap);
1455 #if defined(RTS_SUPPORTS_THREADS)
1456 /* Preparing to leave the RTS, so ensure there's a native thread/task
1457 waiting to take over.
1459 ToDo: optimise this and only create a new task if there's a need
1460 for one (i.e., if there's only one Concurrent Haskell thread alive,
1461 there's no need to create a new task).
1463 IF_DEBUG(scheduler, sched_belch("worker thread (%d): leaving RTS", tok));
1465 startTask(taskStart);
1469 /* Other threads _might_ be available for execution; signal this */
1471 RELEASE_LOCK(&sched_mutex);
1476 resumeThread( StgInt tok,
1478 #if !defined(RTS_SUPPORTS_THREADS)
1483 StgTSO *tso, **prev;
1486 #if defined(RTS_SUPPORTS_THREADS)
1487 /* Wait for permission to re-enter the RTS with the result. */
1489 grabReturnCapability(&sched_mutex, &cap);
1491 grabCapability(&cap);
1494 grabCapability(&cap);
1497 /* Remove the thread off of the suspended list */
1498 prev = &suspended_ccalling_threads;
1499 for (tso = suspended_ccalling_threads;
1500 tso != END_TSO_QUEUE;
1501 prev = &tso->link, tso = tso->link) {
1502 if (tso->id == (StgThreadID)tok) {
1507 if (tso == END_TSO_QUEUE) {
1508 barf("resumeThread: thread not found");
1510 tso->link = END_TSO_QUEUE;
1511 /* Reset blocking status */
1512 tso->why_blocked = NotBlocked;
1514 RELEASE_LOCK(&sched_mutex);
1516 cap->r.rCurrentTSO = tso;
1521 /* ---------------------------------------------------------------------------
1523 * ------------------------------------------------------------------------ */
1524 static void unblockThread(StgTSO *tso);
1526 /* ---------------------------------------------------------------------------
1527 * Comparing Thread ids.
1529 * This is used from STG land in the implementation of the
1530 * instances of Eq/Ord for ThreadIds.
1531 * ------------------------------------------------------------------------ */
1533 int cmp_thread(const StgTSO *tso1, const StgTSO *tso2)
1535 StgThreadID id1 = tso1->id;
1536 StgThreadID id2 = tso2->id;
1538 if (id1 < id2) return (-1);
1539 if (id1 > id2) return 1;
1543 /* ---------------------------------------------------------------------------
1544 * Fetching the ThreadID from an StgTSO.
1546 * This is used in the implementation of Show for ThreadIds.
1547 * ------------------------------------------------------------------------ */
1548 int rts_getThreadId(const StgTSO *tso)
1553 /* ---------------------------------------------------------------------------
1554 Create a new thread.
1556 The new thread starts with the given stack size. Before the
1557 scheduler can run, however, this thread needs to have a closure
1558 (and possibly some arguments) pushed on its stack. See
1559 pushClosure() in Schedule.h.
1561 createGenThread() and createIOThread() (in SchedAPI.h) are
1562 convenient packaged versions of this function.
1564 currently pri (priority) is only used in a GRAN setup -- HWL
1565 ------------------------------------------------------------------------ */
1566 //@cindex createThread
1568 /* currently pri (priority) is only used in a GRAN setup -- HWL */
1570 createThread(nat stack_size, StgInt pri)
1572 return createThread_(stack_size, rtsFalse, pri);
1576 createThread_(nat size, rtsBool have_lock, StgInt pri)
1580 createThread(nat stack_size)
1582 return createThread_(stack_size, rtsFalse);
1586 createThread_(nat size, rtsBool have_lock)
1593 /* First check whether we should create a thread at all */
1595 /* check that no more than RtsFlags.ParFlags.maxThreads threads are created */
1596 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads) {
1598 belch("{createThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
1599 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
1600 return END_TSO_QUEUE;
1606 ASSERT(!RtsFlags.GranFlags.Light || CurrentProc==0);
1609 // ToDo: check whether size = stack_size - TSO_STRUCT_SIZEW
1611 /* catch ridiculously small stack sizes */
1612 if (size < MIN_STACK_WORDS + TSO_STRUCT_SIZEW) {
1613 size = MIN_STACK_WORDS + TSO_STRUCT_SIZEW;
1616 stack_size = size - TSO_STRUCT_SIZEW;
1618 tso = (StgTSO *)allocate(size);
1619 TICK_ALLOC_TSO(stack_size, 0);
1621 SET_HDR(tso, &stg_TSO_info, CCS_SYSTEM);
1623 SET_GRAN_HDR(tso, ThisPE);
1625 tso->what_next = ThreadEnterGHC;
1627 /* tso->id needs to be unique. For now we use a heavyweight mutex to
1628 * protect the increment operation on next_thread_id.
1629 * In future, we could use an atomic increment instead.
1631 if (!have_lock) { ACQUIRE_LOCK(&sched_mutex); }
1632 tso->id = next_thread_id++;
1633 if (!have_lock) { RELEASE_LOCK(&sched_mutex); }
1635 tso->why_blocked = NotBlocked;
1636 tso->blocked_exceptions = NULL;
1638 tso->stack_size = stack_size;
1639 tso->max_stack_size = round_to_mblocks(RtsFlags.GcFlags.maxStkSize)
1641 tso->sp = (P_)&(tso->stack) + stack_size;
1644 tso->prof.CCCS = CCS_MAIN;
1647 /* put a stop frame on the stack */
1648 tso->sp -= sizeofW(StgStopFrame);
1649 SET_HDR((StgClosure*)tso->sp,(StgInfoTable *)&stg_stop_thread_info,CCS_SYSTEM);
1650 tso->su = (StgUpdateFrame*)tso->sp;
1654 tso->link = END_TSO_QUEUE;
1655 /* uses more flexible routine in GranSim */
1656 insertThread(tso, CurrentProc);
1658 /* In a non-GranSim setup the pushing of a TSO onto the runq is separated
1664 if (RtsFlags.GranFlags.GranSimStats.Full)
1665 DumpGranEvent(GR_START,tso);
1667 if (RtsFlags.ParFlags.ParStats.Full)
1668 DumpGranEvent(GR_STARTQ,tso);
1669 /* HACk to avoid SCHEDULE
1673 /* Link the new thread on the global thread list.
1675 tso->global_link = all_threads;
1679 tso->dist.priority = MandatoryPriority; //by default that is...
1683 tso->gran.pri = pri;
1685 tso->gran.magic = TSO_MAGIC; // debugging only
1687 tso->gran.sparkname = 0;
1688 tso->gran.startedat = CURRENT_TIME;
1689 tso->gran.exported = 0;
1690 tso->gran.basicblocks = 0;
1691 tso->gran.allocs = 0;
1692 tso->gran.exectime = 0;
1693 tso->gran.fetchtime = 0;
1694 tso->gran.fetchcount = 0;
1695 tso->gran.blocktime = 0;
1696 tso->gran.blockcount = 0;
1697 tso->gran.blockedat = 0;
1698 tso->gran.globalsparks = 0;
1699 tso->gran.localsparks = 0;
1700 if (RtsFlags.GranFlags.Light)
1701 tso->gran.clock = Now; /* local clock */
1703 tso->gran.clock = 0;
1705 IF_DEBUG(gran,printTSO(tso));
1708 tso->par.magic = TSO_MAGIC; // debugging only
1710 tso->par.sparkname = 0;
1711 tso->par.startedat = CURRENT_TIME;
1712 tso->par.exported = 0;
1713 tso->par.basicblocks = 0;
1714 tso->par.allocs = 0;
1715 tso->par.exectime = 0;
1716 tso->par.fetchtime = 0;
1717 tso->par.fetchcount = 0;
1718 tso->par.blocktime = 0;
1719 tso->par.blockcount = 0;
1720 tso->par.blockedat = 0;
1721 tso->par.globalsparks = 0;
1722 tso->par.localsparks = 0;
1726 globalGranStats.tot_threads_created++;
1727 globalGranStats.threads_created_on_PE[CurrentProc]++;
1728 globalGranStats.tot_sq_len += spark_queue_len(CurrentProc);
1729 globalGranStats.tot_sq_probes++;
1731 // collect parallel global statistics (currently done together with GC stats)
1732 if (RtsFlags.ParFlags.ParStats.Global &&
1733 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
1734 //fprintf(stderr, "Creating thread %d @ %11.2f\n", tso->id, usertime());
1735 globalParStats.tot_threads_created++;
1741 belch("==__ schedule: Created TSO %d (%p);",
1742 CurrentProc, tso, tso->id));
1744 IF_PAR_DEBUG(verbose,
1745 belch("==__ schedule: Created TSO %d (%p); %d threads active",
1746 tso->id, tso, advisory_thread_count));
1748 IF_DEBUG(scheduler,sched_belch("created thread %ld, stack size = %lx words",
1749 tso->id, tso->stack_size));
1756 all parallel thread creation calls should fall through the following routine.
1759 createSparkThread(rtsSpark spark)
1761 ASSERT(spark != (rtsSpark)NULL);
1762 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads)
1764 barf("{createSparkThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
1765 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
1766 return END_TSO_QUEUE;
1770 tso = createThread_(RtsFlags.GcFlags.initialStkSize, rtsTrue);
1771 if (tso==END_TSO_QUEUE)
1772 barf("createSparkThread: Cannot create TSO");
1774 tso->priority = AdvisoryPriority;
1776 pushClosure(tso,spark);
1777 PUSH_ON_RUN_QUEUE(tso);
1778 advisory_thread_count++;
1785 Turn a spark into a thread.
1786 ToDo: fix for SMP (needs to acquire SCHED_MUTEX!)
1789 //@cindex activateSpark
1791 activateSpark (rtsSpark spark)
1795 tso = createSparkThread(spark);
1796 if (RtsFlags.ParFlags.ParStats.Full) {
1797 //ASSERT(run_queue_hd == END_TSO_QUEUE); // I think ...
1798 IF_PAR_DEBUG(verbose,
1799 belch("==^^ activateSpark: turning spark of closure %p (%s) into a thread",
1800 (StgClosure *)spark, info_type((StgClosure *)spark)));
1802 // ToDo: fwd info on local/global spark to thread -- HWL
1803 // tso->gran.exported = spark->exported;
1804 // tso->gran.locked = !spark->global;
1805 // tso->gran.sparkname = spark->name;
1811 /* ---------------------------------------------------------------------------
1814 * scheduleThread puts a thread on the head of the runnable queue.
1815 * This will usually be done immediately after a thread is created.
1816 * The caller of scheduleThread must create the thread using e.g.
1817 * createThread and push an appropriate closure
1818 * on this thread's stack before the scheduler is invoked.
1819 * ------------------------------------------------------------------------ */
1821 static void scheduleThread_ (StgTSO* tso, rtsBool createTask);
1824 scheduleThread_(StgTSO *tso
1825 , rtsBool createTask
1826 #if !defined(THREADED_RTS)
1831 ACQUIRE_LOCK(&sched_mutex);
1833 /* Put the new thread on the head of the runnable queue. The caller
1834 * better push an appropriate closure on this thread's stack
1835 * beforehand. In the SMP case, the thread may start running as
1836 * soon as we release the scheduler lock below.
1838 PUSH_ON_RUN_QUEUE(tso);
1839 #if defined(THREADED_RTS)
1840 /* If main() is scheduling a thread, don't bother creating a
1844 startTask(taskStart);
1850 IF_DEBUG(scheduler,printTSO(tso));
1852 RELEASE_LOCK(&sched_mutex);
1855 void scheduleThread(StgTSO* tso)
1857 return scheduleThread_(tso, rtsFalse);
1860 void scheduleExtThread(StgTSO* tso)
1862 return scheduleThread_(tso, rtsTrue);
1865 /* ---------------------------------------------------------------------------
1868 * Initialise the scheduler. This resets all the queues - if the
1869 * queues contained any threads, they'll be garbage collected at the
1872 * ------------------------------------------------------------------------ */
1876 term_handler(int sig STG_UNUSED)
1879 ACQUIRE_LOCK(&term_mutex);
1881 RELEASE_LOCK(&term_mutex);
1892 for (i=0; i<=MAX_PROC; i++) {
1893 run_queue_hds[i] = END_TSO_QUEUE;
1894 run_queue_tls[i] = END_TSO_QUEUE;
1895 blocked_queue_hds[i] = END_TSO_QUEUE;
1896 blocked_queue_tls[i] = END_TSO_QUEUE;
1897 ccalling_threadss[i] = END_TSO_QUEUE;
1898 sleeping_queue = END_TSO_QUEUE;
1901 run_queue_hd = END_TSO_QUEUE;
1902 run_queue_tl = END_TSO_QUEUE;
1903 blocked_queue_hd = END_TSO_QUEUE;
1904 blocked_queue_tl = END_TSO_QUEUE;
1905 sleeping_queue = END_TSO_QUEUE;
1908 suspended_ccalling_threads = END_TSO_QUEUE;
1910 main_threads = NULL;
1911 all_threads = END_TSO_QUEUE;
1916 RtsFlags.ConcFlags.ctxtSwitchTicks =
1917 RtsFlags.ConcFlags.ctxtSwitchTime / TICK_MILLISECS;
1919 #if defined(RTS_SUPPORTS_THREADS)
1920 /* Initialise the mutex and condition variables used by
1922 initMutex(&sched_mutex);
1923 initMutex(&term_mutex);
1925 initCondition(&thread_ready_cond);
1929 initCondition(&gc_pending_cond);
1932 #if defined(RTS_SUPPORTS_THREADS)
1933 ACQUIRE_LOCK(&sched_mutex);
1936 /* Install the SIGHUP handler */
1939 struct sigaction action,oact;
1941 action.sa_handler = term_handler;
1942 sigemptyset(&action.sa_mask);
1943 action.sa_flags = 0;
1944 if (sigaction(SIGTERM, &action, &oact) != 0) {
1945 barf("can't install TERM handler");
1950 /* A capability holds the state a native thread needs in
1951 * order to execute STG code. At least one capability is
1952 * floating around (only SMP builds have more than one).
1956 #if defined(RTS_SUPPORTS_THREADS)
1957 /* start our haskell execution tasks */
1959 startTaskManager(RtsFlags.ParFlags.nNodes, taskStart);
1961 startTaskManager(0,taskStart);
1965 #if /* defined(SMP) ||*/ defined(PAR)
1969 #if defined(RTS_SUPPORTS_THREADS)
1970 RELEASE_LOCK(&sched_mutex);
1976 exitScheduler( void )
1978 #if defined(RTS_SUPPORTS_THREADS)
1983 /* -----------------------------------------------------------------------------
1984 Managing the per-task allocation areas.
1986 Each capability comes with an allocation area. These are
1987 fixed-length block lists into which allocation can be done.
1989 ToDo: no support for two-space collection at the moment???
1990 -------------------------------------------------------------------------- */
1992 /* -----------------------------------------------------------------------------
1993 * waitThread is the external interface for running a new computation
1994 * and waiting for the result.
1996 * In the non-SMP case, we create a new main thread, push it on the
1997 * main-thread stack, and invoke the scheduler to run it. The
1998 * scheduler will return when the top main thread on the stack has
1999 * completed or died, and fill in the necessary fields of the
2000 * main_thread structure.
2002 * In the SMP case, we create a main thread as before, but we then
2003 * create a new condition variable and sleep on it. When our new
2004 * main thread has completed, we'll be woken up and the status/result
2005 * will be in the main_thread struct.
2006 * -------------------------------------------------------------------------- */
2009 howManyThreadsAvail ( void )
2013 for (q = run_queue_hd; q != END_TSO_QUEUE; q = q->link)
2015 for (q = blocked_queue_hd; q != END_TSO_QUEUE; q = q->link)
2017 for (q = sleeping_queue; q != END_TSO_QUEUE; q = q->link)
2023 finishAllThreads ( void )
2026 while (run_queue_hd != END_TSO_QUEUE) {
2027 waitThread ( run_queue_hd, NULL);
2029 while (blocked_queue_hd != END_TSO_QUEUE) {
2030 waitThread ( blocked_queue_hd, NULL);
2032 while (sleeping_queue != END_TSO_QUEUE) {
2033 waitThread ( blocked_queue_hd, NULL);
2036 (blocked_queue_hd != END_TSO_QUEUE ||
2037 run_queue_hd != END_TSO_QUEUE ||
2038 sleeping_queue != END_TSO_QUEUE);
2042 waitThread(StgTSO *tso, /*out*/StgClosure **ret)
2044 #if defined(THREADED_RTS)
2045 return waitThread_(tso,ret, rtsFalse);
2047 return waitThread_(tso,ret);
2052 waitThread_(StgTSO *tso,
2053 /*out*/StgClosure **ret
2054 #if defined(THREADED_RTS)
2055 , rtsBool blockWaiting
2060 SchedulerStatus stat;
2062 ACQUIRE_LOCK(&sched_mutex);
2064 m = stgMallocBytes(sizeof(StgMainThread), "waitThread");
2069 #if defined(RTS_SUPPORTS_THREADS)
2070 initCondition(&m->wakeup);
2073 m->link = main_threads;
2076 IF_DEBUG(scheduler, sched_belch("== scheduler: new main thread (%d)\n", m->tso->id));
2078 #if defined(RTS_SUPPORTS_THREADS)
2080 # if defined(THREADED_RTS)
2081 if (!blockWaiting) {
2082 /* In the threaded case, the OS thread that called main()
2083 * gets to enter the RTS directly without going via another
2086 RELEASE_LOCK(&sched_mutex);
2088 ASSERT(m->stat != NoStatus);
2092 IF_DEBUG(scheduler, sched_belch("sfoo"));
2094 waitCondition(&m->wakeup, &sched_mutex);
2095 } while (m->stat == NoStatus);
2098 /* GranSim specific init */
2099 CurrentTSO = m->tso; // the TSO to run
2100 procStatus[MainProc] = Busy; // status of main PE
2101 CurrentProc = MainProc; // PE to run it on
2105 RELEASE_LOCK(&sched_mutex);
2107 ASSERT(m->stat != NoStatus);
2112 #if defined(RTS_SUPPORTS_THREADS)
2113 closeCondition(&m->wakeup);
2116 IF_DEBUG(scheduler, fprintf(stderr, "== scheduler: main thread (%d) finished\n",
2120 #if defined(THREADED_RTS)
2123 RELEASE_LOCK(&sched_mutex);
2128 //@node Run queue code, Garbage Collextion Routines, Suspend and Resume, Main scheduling code
2129 //@subsection Run queue code
2133 NB: In GranSim we have many run queues; run_queue_hd is actually a macro
2134 unfolding to run_queue_hds[CurrentProc], thus CurrentProc is an
2135 implicit global variable that has to be correct when calling these
2139 /* Put the new thread on the head of the runnable queue.
2140 * The caller of createThread better push an appropriate closure
2141 * on this thread's stack before the scheduler is invoked.
2143 static /* inline */ void
2144 add_to_run_queue(tso)
2147 ASSERT(tso!=run_queue_hd && tso!=run_queue_tl);
2148 tso->link = run_queue_hd;
2150 if (run_queue_tl == END_TSO_QUEUE) {
2155 /* Put the new thread at the end of the runnable queue. */
2156 static /* inline */ void
2157 push_on_run_queue(tso)
2160 ASSERT(get_itbl((StgClosure *)tso)->type == TSO);
2161 ASSERT(run_queue_hd!=NULL && run_queue_tl!=NULL);
2162 ASSERT(tso!=run_queue_hd && tso!=run_queue_tl);
2163 if (run_queue_hd == END_TSO_QUEUE) {
2166 run_queue_tl->link = tso;
2172 Should be inlined because it's used very often in schedule. The tso
2173 argument is actually only needed in GranSim, where we want to have the
2174 possibility to schedule *any* TSO on the run queue, irrespective of the
2175 actual ordering. Therefore, if tso is not the nil TSO then we traverse
2176 the run queue and dequeue the tso, adjusting the links in the queue.
2178 //@cindex take_off_run_queue
2179 static /* inline */ StgTSO*
2180 take_off_run_queue(StgTSO *tso) {
2184 qetlaHbogh Qu' ngaSbogh ghomDaQ {tso} yIteq!
2186 if tso is specified, unlink that tso from the run_queue (doesn't have
2187 to be at the beginning of the queue); GranSim only
2189 if (tso!=END_TSO_QUEUE) {
2190 /* find tso in queue */
2191 for (t=run_queue_hd, prev=END_TSO_QUEUE;
2192 t!=END_TSO_QUEUE && t!=tso;
2196 /* now actually dequeue the tso */
2197 if (prev!=END_TSO_QUEUE) {
2198 ASSERT(run_queue_hd!=t);
2199 prev->link = t->link;
2201 /* t is at beginning of thread queue */
2202 ASSERT(run_queue_hd==t);
2203 run_queue_hd = t->link;
2205 /* t is at end of thread queue */
2206 if (t->link==END_TSO_QUEUE) {
2207 ASSERT(t==run_queue_tl);
2208 run_queue_tl = prev;
2210 ASSERT(run_queue_tl!=t);
2212 t->link = END_TSO_QUEUE;
2214 /* take tso from the beginning of the queue; std concurrent code */
2216 if (t != END_TSO_QUEUE) {
2217 run_queue_hd = t->link;
2218 t->link = END_TSO_QUEUE;
2219 if (run_queue_hd == END_TSO_QUEUE) {
2220 run_queue_tl = END_TSO_QUEUE;
2229 //@node Garbage Collextion Routines, Blocking Queue Routines, Run queue code, Main scheduling code
2230 //@subsection Garbage Collextion Routines
2232 /* ---------------------------------------------------------------------------
2233 Where are the roots that we know about?
2235 - all the threads on the runnable queue
2236 - all the threads on the blocked queue
2237 - all the threads on the sleeping queue
2238 - all the thread currently executing a _ccall_GC
2239 - all the "main threads"
2241 ------------------------------------------------------------------------ */
2243 /* This has to be protected either by the scheduler monitor, or by the
2244 garbage collection monitor (probably the latter).
2249 GetRoots(evac_fn evac)
2256 for (i=0; i<=RtsFlags.GranFlags.proc; i++) {
2257 if ((run_queue_hds[i] != END_TSO_QUEUE) && ((run_queue_hds[i] != NULL)))
2258 evac((StgClosure **)&run_queue_hds[i]);
2259 if ((run_queue_tls[i] != END_TSO_QUEUE) && ((run_queue_tls[i] != NULL)))
2260 evac((StgClosure **)&run_queue_tls[i]);
2262 if ((blocked_queue_hds[i] != END_TSO_QUEUE) && ((blocked_queue_hds[i] != NULL)))
2263 evac((StgClosure **)&blocked_queue_hds[i]);
2264 if ((blocked_queue_tls[i] != END_TSO_QUEUE) && ((blocked_queue_tls[i] != NULL)))
2265 evac((StgClosure **)&blocked_queue_tls[i]);
2266 if ((ccalling_threadss[i] != END_TSO_QUEUE) && ((ccalling_threadss[i] != NULL)))
2267 evac((StgClosure **)&ccalling_threads[i]);
2274 if (run_queue_hd != END_TSO_QUEUE) {
2275 ASSERT(run_queue_tl != END_TSO_QUEUE);
2276 evac((StgClosure **)&run_queue_hd);
2277 evac((StgClosure **)&run_queue_tl);
2280 if (blocked_queue_hd != END_TSO_QUEUE) {
2281 ASSERT(blocked_queue_tl != END_TSO_QUEUE);
2282 evac((StgClosure **)&blocked_queue_hd);
2283 evac((StgClosure **)&blocked_queue_tl);
2286 if (sleeping_queue != END_TSO_QUEUE) {
2287 evac((StgClosure **)&sleeping_queue);
2291 if (suspended_ccalling_threads != END_TSO_QUEUE) {
2292 evac((StgClosure **)&suspended_ccalling_threads);
2295 #if defined(PAR) || defined(GRAN)
2296 markSparkQueue(evac);
2300 /* -----------------------------------------------------------------------------
2303 This is the interface to the garbage collector from Haskell land.
2304 We provide this so that external C code can allocate and garbage
2305 collect when called from Haskell via _ccall_GC.
2307 It might be useful to provide an interface whereby the programmer
2308 can specify more roots (ToDo).
2310 This needs to be protected by the GC condition variable above. KH.
2311 -------------------------------------------------------------------------- */
2313 void (*extra_roots)(evac_fn);
2318 /* Obligated to hold this lock upon entry */
2319 ACQUIRE_LOCK(&sched_mutex);
2320 GarbageCollect(GetRoots,rtsFalse);
2321 RELEASE_LOCK(&sched_mutex);
2325 performMajorGC(void)
2327 ACQUIRE_LOCK(&sched_mutex);
2328 GarbageCollect(GetRoots,rtsTrue);
2329 RELEASE_LOCK(&sched_mutex);
2333 AllRoots(evac_fn evac)
2335 GetRoots(evac); // the scheduler's roots
2336 extra_roots(evac); // the user's roots
2340 performGCWithRoots(void (*get_roots)(evac_fn))
2342 ACQUIRE_LOCK(&sched_mutex);
2343 extra_roots = get_roots;
2344 GarbageCollect(AllRoots,rtsFalse);
2345 RELEASE_LOCK(&sched_mutex);
2348 /* -----------------------------------------------------------------------------
2351 If the thread has reached its maximum stack size, then raise the
2352 StackOverflow exception in the offending thread. Otherwise
2353 relocate the TSO into a larger chunk of memory and adjust its stack
2355 -------------------------------------------------------------------------- */
2358 threadStackOverflow(StgTSO *tso)
2360 nat new_stack_size, new_tso_size, diff, stack_words;
2364 IF_DEBUG(sanity,checkTSO(tso));
2365 if (tso->stack_size >= tso->max_stack_size) {
2368 belch("@@ threadStackOverflow of TSO %d (%p): stack too large (now %ld; max is %ld",
2369 tso->id, tso, tso->stack_size, tso->max_stack_size);
2370 /* If we're debugging, just print out the top of the stack */
2371 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2374 /* Send this thread the StackOverflow exception */
2375 raiseAsync(tso, (StgClosure *)stackOverflow_closure);
2379 /* Try to double the current stack size. If that takes us over the
2380 * maximum stack size for this thread, then use the maximum instead.
2381 * Finally round up so the TSO ends up as a whole number of blocks.
2383 new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
2384 new_tso_size = (nat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) +
2385 TSO_STRUCT_SIZE)/sizeof(W_);
2386 new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */
2387 new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
2389 IF_DEBUG(scheduler, fprintf(stderr,"== scheduler: increasing stack size from %d words to %d.\n", tso->stack_size, new_stack_size));
2391 dest = (StgTSO *)allocate(new_tso_size);
2392 TICK_ALLOC_TSO(new_stack_size,0);
2394 /* copy the TSO block and the old stack into the new area */
2395 memcpy(dest,tso,TSO_STRUCT_SIZE);
2396 stack_words = tso->stack + tso->stack_size - tso->sp;
2397 new_sp = (P_)dest + new_tso_size - stack_words;
2398 memcpy(new_sp, tso->sp, stack_words * sizeof(W_));
2400 /* relocate the stack pointers... */
2401 diff = (P_)new_sp - (P_)tso->sp; /* In *words* */
2402 dest->su = (StgUpdateFrame *) ((P_)dest->su + diff);
2404 dest->stack_size = new_stack_size;
2406 /* and relocate the update frame list */
2407 relocate_stack(dest, diff);
2409 /* Mark the old TSO as relocated. We have to check for relocated
2410 * TSOs in the garbage collector and any primops that deal with TSOs.
2412 * It's important to set the sp and su values to just beyond the end
2413 * of the stack, so we don't attempt to scavenge any part of the
2416 tso->what_next = ThreadRelocated;
2418 tso->sp = (P_)&(tso->stack[tso->stack_size]);
2419 tso->su = (StgUpdateFrame *)tso->sp;
2420 tso->why_blocked = NotBlocked;
2421 dest->mut_link = NULL;
2423 IF_PAR_DEBUG(verbose,
2424 belch("@@ threadStackOverflow of TSO %d (now at %p): stack size increased to %ld",
2425 tso->id, tso, tso->stack_size);
2426 /* If we're debugging, just print out the top of the stack */
2427 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2430 IF_DEBUG(sanity,checkTSO(tso));
2432 IF_DEBUG(scheduler,printTSO(dest));
2438 //@node Blocking Queue Routines, Exception Handling Routines, Garbage Collextion Routines, Main scheduling code
2439 //@subsection Blocking Queue Routines
2441 /* ---------------------------------------------------------------------------
2442 Wake up a queue that was blocked on some resource.
2443 ------------------------------------------------------------------------ */
2447 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2452 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2454 /* write RESUME events to log file and
2455 update blocked and fetch time (depending on type of the orig closure) */
2456 if (RtsFlags.ParFlags.ParStats.Full) {
2457 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
2458 GR_RESUMEQ, ((StgTSO *)bqe), ((StgTSO *)bqe)->block_info.closure,
2459 0, 0 /* spark_queue_len(ADVISORY_POOL) */);
2460 if (EMPTY_RUN_QUEUE())
2461 emitSchedule = rtsTrue;
2463 switch (get_itbl(node)->type) {
2465 ((StgTSO *)bqe)->par.fetchtime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2470 ((StgTSO *)bqe)->par.blocktime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2477 barf("{unblockOneLocked}Daq Qagh: unexpected closure in blocking queue");
2484 static StgBlockingQueueElement *
2485 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2488 PEs node_loc, tso_loc;
2490 node_loc = where_is(node); // should be lifted out of loop
2491 tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2492 tso_loc = where_is((StgClosure *)tso);
2493 if (IS_LOCAL_TO(PROCS(node),tso_loc)) { // TSO is local
2494 /* !fake_fetch => TSO is on CurrentProc is same as IS_LOCAL_TO */
2495 ASSERT(CurrentProc!=node_loc || tso_loc==CurrentProc);
2496 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.lunblocktime;
2497 // insertThread(tso, node_loc);
2498 new_event(tso_loc, tso_loc, CurrentTime[CurrentProc],
2500 tso, node, (rtsSpark*)NULL);
2501 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2504 } else { // TSO is remote (actually should be FMBQ)
2505 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.mpacktime +
2506 RtsFlags.GranFlags.Costs.gunblocktime +
2507 RtsFlags.GranFlags.Costs.latency;
2508 new_event(tso_loc, CurrentProc, CurrentTime[CurrentProc],
2510 tso, node, (rtsSpark*)NULL);
2511 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2514 /* the thread-queue-overhead is accounted for in either Resume or UnblockThread */
2516 fprintf(stderr," %s TSO %d (%p) [PE %d] (block_info.closure=%p) (next=%p) ,",
2517 (node_loc==tso_loc ? "Local" : "Global"),
2518 tso->id, tso, CurrentProc, tso->block_info.closure, tso->link));
2519 tso->block_info.closure = NULL;
2520 IF_DEBUG(scheduler,belch("-- Waking up thread %ld (%p)",
2524 static StgBlockingQueueElement *
2525 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2527 StgBlockingQueueElement *next;
2529 switch (get_itbl(bqe)->type) {
2531 ASSERT(((StgTSO *)bqe)->why_blocked != NotBlocked);
2532 /* if it's a TSO just push it onto the run_queue */
2534 // ((StgTSO *)bqe)->link = END_TSO_QUEUE; // debugging?
2535 PUSH_ON_RUN_QUEUE((StgTSO *)bqe);
2537 unblockCount(bqe, node);
2538 /* reset blocking status after dumping event */
2539 ((StgTSO *)bqe)->why_blocked = NotBlocked;
2543 /* if it's a BLOCKED_FETCH put it on the PendingFetches list */
2545 bqe->link = (StgBlockingQueueElement *)PendingFetches;
2546 PendingFetches = (StgBlockedFetch *)bqe;
2550 /* can ignore this case in a non-debugging setup;
2551 see comments on RBHSave closures above */
2553 /* check that the closure is an RBHSave closure */
2554 ASSERT(get_itbl((StgClosure *)bqe) == &stg_RBH_Save_0_info ||
2555 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_1_info ||
2556 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_2_info);
2560 barf("{unblockOneLocked}Daq Qagh: Unexpected IP (%#lx; %s) in blocking queue at %#lx\n",
2561 get_itbl((StgClosure *)bqe), info_type((StgClosure *)bqe),
2565 IF_PAR_DEBUG(bq, fprintf(stderr, ", %p (%s)", bqe, info_type((StgClosure*)bqe)));
2569 #else /* !GRAN && !PAR */
2571 unblockOneLocked(StgTSO *tso)
2575 ASSERT(get_itbl(tso)->type == TSO);
2576 ASSERT(tso->why_blocked != NotBlocked);
2577 tso->why_blocked = NotBlocked;
2579 PUSH_ON_RUN_QUEUE(tso);
2581 IF_DEBUG(scheduler,sched_belch("waking up thread %ld", tso->id));
2586 #if defined(GRAN) || defined(PAR)
2587 inline StgBlockingQueueElement *
2588 unblockOne(StgBlockingQueueElement *bqe, StgClosure *node)
2590 ACQUIRE_LOCK(&sched_mutex);
2591 bqe = unblockOneLocked(bqe, node);
2592 RELEASE_LOCK(&sched_mutex);
2597 unblockOne(StgTSO *tso)
2599 ACQUIRE_LOCK(&sched_mutex);
2600 tso = unblockOneLocked(tso);
2601 RELEASE_LOCK(&sched_mutex);
2608 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2610 StgBlockingQueueElement *bqe;
2615 belch("##-_ AwBQ for node %p on PE %d @ %ld by TSO %d (%p): ", \
2616 node, CurrentProc, CurrentTime[CurrentProc],
2617 CurrentTSO->id, CurrentTSO));
2619 node_loc = where_is(node);
2621 ASSERT(q == END_BQ_QUEUE ||
2622 get_itbl(q)->type == TSO || // q is either a TSO or an RBHSave
2623 get_itbl(q)->type == CONSTR); // closure (type constructor)
2624 ASSERT(is_unique(node));
2626 /* FAKE FETCH: magically copy the node to the tso's proc;
2627 no Fetch necessary because in reality the node should not have been
2628 moved to the other PE in the first place
2630 if (CurrentProc!=node_loc) {
2632 belch("## node %p is on PE %d but CurrentProc is %d (TSO %d); assuming fake fetch and adjusting bitmask (old: %#x)",
2633 node, node_loc, CurrentProc, CurrentTSO->id,
2634 // CurrentTSO, where_is(CurrentTSO),
2635 node->header.gran.procs));
2636 node->header.gran.procs = (node->header.gran.procs) | PE_NUMBER(CurrentProc);
2638 belch("## new bitmask of node %p is %#x",
2639 node, node->header.gran.procs));
2640 if (RtsFlags.GranFlags.GranSimStats.Global) {
2641 globalGranStats.tot_fake_fetches++;
2646 // ToDo: check: ASSERT(CurrentProc==node_loc);
2647 while (get_itbl(bqe)->type==TSO) { // q != END_TSO_QUEUE) {
2650 bqe points to the current element in the queue
2651 next points to the next element in the queue
2653 //tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2654 //tso_loc = where_is(tso);
2656 bqe = unblockOneLocked(bqe, node);
2659 /* if this is the BQ of an RBH, we have to put back the info ripped out of
2660 the closure to make room for the anchor of the BQ */
2661 if (bqe!=END_BQ_QUEUE) {
2662 ASSERT(get_itbl(node)->type == RBH && get_itbl(bqe)->type == CONSTR);
2664 ASSERT((info_ptr==&RBH_Save_0_info) ||
2665 (info_ptr==&RBH_Save_1_info) ||
2666 (info_ptr==&RBH_Save_2_info));
2668 /* cf. convertToRBH in RBH.c for writing the RBHSave closure */
2669 ((StgRBH *)node)->blocking_queue = (StgBlockingQueueElement *)((StgRBHSave *)bqe)->payload[0];
2670 ((StgRBH *)node)->mut_link = (StgMutClosure *)((StgRBHSave *)bqe)->payload[1];
2673 belch("## Filled in RBH_Save for %p (%s) at end of AwBQ",
2674 node, info_type(node)));
2677 /* statistics gathering */
2678 if (RtsFlags.GranFlags.GranSimStats.Global) {
2679 // globalGranStats.tot_bq_processing_time += bq_processing_time;
2680 globalGranStats.tot_bq_len += len; // total length of all bqs awakened
2681 // globalGranStats.tot_bq_len_local += len_local; // same for local TSOs only
2682 globalGranStats.tot_awbq++; // total no. of bqs awakened
2685 fprintf(stderr,"## BQ Stats of %p: [%d entries] %s\n",
2686 node, len, (bqe!=END_BQ_QUEUE) ? "RBH" : ""));
2690 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2692 StgBlockingQueueElement *bqe;
2694 ACQUIRE_LOCK(&sched_mutex);
2696 IF_PAR_DEBUG(verbose,
2697 belch("##-_ AwBQ for node %p on [%x]: ",
2701 if(get_itbl(q)->type == CONSTR || q==END_BQ_QUEUE) {
2702 IF_PAR_DEBUG(verbose, belch("## ... nothing to unblock so lets just return. RFP (BUG?)"));
2707 ASSERT(q == END_BQ_QUEUE ||
2708 get_itbl(q)->type == TSO ||
2709 get_itbl(q)->type == BLOCKED_FETCH ||
2710 get_itbl(q)->type == CONSTR);
2713 while (get_itbl(bqe)->type==TSO ||
2714 get_itbl(bqe)->type==BLOCKED_FETCH) {
2715 bqe = unblockOneLocked(bqe, node);
2717 RELEASE_LOCK(&sched_mutex);
2720 #else /* !GRAN && !PAR */
2722 awakenBlockedQueue(StgTSO *tso)
2724 ACQUIRE_LOCK(&sched_mutex);
2725 while (tso != END_TSO_QUEUE) {
2726 tso = unblockOneLocked(tso);
2728 RELEASE_LOCK(&sched_mutex);
2732 //@node Exception Handling Routines, Debugging Routines, Blocking Queue Routines, Main scheduling code
2733 //@subsection Exception Handling Routines
2735 /* ---------------------------------------------------------------------------
2737 - usually called inside a signal handler so it mustn't do anything fancy.
2738 ------------------------------------------------------------------------ */
2741 interruptStgRts(void)
2747 /* -----------------------------------------------------------------------------
2750 This is for use when we raise an exception in another thread, which
2752 This has nothing to do with the UnblockThread event in GranSim. -- HWL
2753 -------------------------------------------------------------------------- */
2755 #if defined(GRAN) || defined(PAR)
2757 NB: only the type of the blocking queue is different in GranSim and GUM
2758 the operations on the queue-elements are the same
2759 long live polymorphism!
2762 unblockThread(StgTSO *tso)
2764 StgBlockingQueueElement *t, **last;
2766 ACQUIRE_LOCK(&sched_mutex);
2767 switch (tso->why_blocked) {
2770 return; /* not blocked */
2773 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
2775 StgBlockingQueueElement *last_tso = END_BQ_QUEUE;
2776 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
2778 last = (StgBlockingQueueElement **)&mvar->head;
2779 for (t = (StgBlockingQueueElement *)mvar->head;
2781 last = &t->link, last_tso = t, t = t->link) {
2782 if (t == (StgBlockingQueueElement *)tso) {
2783 *last = (StgBlockingQueueElement *)tso->link;
2784 if (mvar->tail == tso) {
2785 mvar->tail = (StgTSO *)last_tso;
2790 barf("unblockThread (MVAR): TSO not found");
2793 case BlockedOnBlackHole:
2794 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
2796 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
2798 last = &bq->blocking_queue;
2799 for (t = bq->blocking_queue;
2801 last = &t->link, t = t->link) {
2802 if (t == (StgBlockingQueueElement *)tso) {
2803 *last = (StgBlockingQueueElement *)tso->link;
2807 barf("unblockThread (BLACKHOLE): TSO not found");
2810 case BlockedOnException:
2812 StgTSO *target = tso->block_info.tso;
2814 ASSERT(get_itbl(target)->type == TSO);
2816 if (target->what_next == ThreadRelocated) {
2817 target = target->link;
2818 ASSERT(get_itbl(target)->type == TSO);
2821 ASSERT(target->blocked_exceptions != NULL);
2823 last = (StgBlockingQueueElement **)&target->blocked_exceptions;
2824 for (t = (StgBlockingQueueElement *)target->blocked_exceptions;
2826 last = &t->link, t = t->link) {
2827 ASSERT(get_itbl(t)->type == TSO);
2828 if (t == (StgBlockingQueueElement *)tso) {
2829 *last = (StgBlockingQueueElement *)tso->link;
2833 barf("unblockThread (Exception): TSO not found");
2837 case BlockedOnWrite:
2839 /* take TSO off blocked_queue */
2840 StgBlockingQueueElement *prev = NULL;
2841 for (t = (StgBlockingQueueElement *)blocked_queue_hd; t != END_BQ_QUEUE;
2842 prev = t, t = t->link) {
2843 if (t == (StgBlockingQueueElement *)tso) {
2845 blocked_queue_hd = (StgTSO *)t->link;
2846 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
2847 blocked_queue_tl = END_TSO_QUEUE;
2850 prev->link = t->link;
2851 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
2852 blocked_queue_tl = (StgTSO *)prev;
2858 barf("unblockThread (I/O): TSO not found");
2861 case BlockedOnDelay:
2863 /* take TSO off sleeping_queue */
2864 StgBlockingQueueElement *prev = NULL;
2865 for (t = (StgBlockingQueueElement *)sleeping_queue; t != END_BQ_QUEUE;
2866 prev = t, t = t->link) {
2867 if (t == (StgBlockingQueueElement *)tso) {
2869 sleeping_queue = (StgTSO *)t->link;
2871 prev->link = t->link;
2876 barf("unblockThread (I/O): TSO not found");
2880 barf("unblockThread");
2884 tso->link = END_TSO_QUEUE;
2885 tso->why_blocked = NotBlocked;
2886 tso->block_info.closure = NULL;
2887 PUSH_ON_RUN_QUEUE(tso);
2888 RELEASE_LOCK(&sched_mutex);
2892 unblockThread(StgTSO *tso)
2896 ACQUIRE_LOCK(&sched_mutex);
2897 switch (tso->why_blocked) {
2900 return; /* not blocked */
2903 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
2905 StgTSO *last_tso = END_TSO_QUEUE;
2906 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
2909 for (t = mvar->head; t != END_TSO_QUEUE;
2910 last = &t->link, last_tso = t, t = t->link) {
2913 if (mvar->tail == tso) {
2914 mvar->tail = last_tso;
2919 barf("unblockThread (MVAR): TSO not found");
2922 case BlockedOnBlackHole:
2923 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
2925 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
2927 last = &bq->blocking_queue;
2928 for (t = bq->blocking_queue; t != END_TSO_QUEUE;
2929 last = &t->link, t = t->link) {
2935 barf("unblockThread (BLACKHOLE): TSO not found");
2938 case BlockedOnException:
2940 StgTSO *target = tso->block_info.tso;
2942 ASSERT(get_itbl(target)->type == TSO);
2944 while (target->what_next == ThreadRelocated) {
2945 target = target->link;
2946 ASSERT(get_itbl(target)->type == TSO);
2949 ASSERT(target->blocked_exceptions != NULL);
2951 last = &target->blocked_exceptions;
2952 for (t = target->blocked_exceptions; t != END_TSO_QUEUE;
2953 last = &t->link, t = t->link) {
2954 ASSERT(get_itbl(t)->type == TSO);
2960 barf("unblockThread (Exception): TSO not found");
2964 case BlockedOnWrite:
2966 StgTSO *prev = NULL;
2967 for (t = blocked_queue_hd; t != END_TSO_QUEUE;
2968 prev = t, t = t->link) {
2971 blocked_queue_hd = t->link;
2972 if (blocked_queue_tl == t) {
2973 blocked_queue_tl = END_TSO_QUEUE;
2976 prev->link = t->link;
2977 if (blocked_queue_tl == t) {
2978 blocked_queue_tl = prev;
2984 barf("unblockThread (I/O): TSO not found");
2987 case BlockedOnDelay:
2989 StgTSO *prev = NULL;
2990 for (t = sleeping_queue; t != END_TSO_QUEUE;
2991 prev = t, t = t->link) {
2994 sleeping_queue = t->link;
2996 prev->link = t->link;
3001 barf("unblockThread (I/O): TSO not found");
3005 barf("unblockThread");
3009 tso->link = END_TSO_QUEUE;
3010 tso->why_blocked = NotBlocked;
3011 tso->block_info.closure = NULL;
3012 PUSH_ON_RUN_QUEUE(tso);
3013 RELEASE_LOCK(&sched_mutex);
3017 /* -----------------------------------------------------------------------------
3020 * The following function implements the magic for raising an
3021 * asynchronous exception in an existing thread.
3023 * We first remove the thread from any queue on which it might be
3024 * blocked. The possible blockages are MVARs and BLACKHOLE_BQs.
3026 * We strip the stack down to the innermost CATCH_FRAME, building
3027 * thunks in the heap for all the active computations, so they can
3028 * be restarted if necessary. When we reach a CATCH_FRAME, we build
3029 * an application of the handler to the exception, and push it on
3030 * the top of the stack.
3032 * How exactly do we save all the active computations? We create an
3033 * AP_UPD for every UpdateFrame on the stack. Entering one of these
3034 * AP_UPDs pushes everything from the corresponding update frame
3035 * upwards onto the stack. (Actually, it pushes everything up to the
3036 * next update frame plus a pointer to the next AP_UPD object.
3037 * Entering the next AP_UPD object pushes more onto the stack until we
3038 * reach the last AP_UPD object - at which point the stack should look
3039 * exactly as it did when we killed the TSO and we can continue
3040 * execution by entering the closure on top of the stack.
3042 * We can also kill a thread entirely - this happens if either (a) the
3043 * exception passed to raiseAsync is NULL, or (b) there's no
3044 * CATCH_FRAME on the stack. In either case, we strip the entire
3045 * stack and replace the thread with a zombie.
3047 * -------------------------------------------------------------------------- */
3050 deleteThread(StgTSO *tso)
3052 raiseAsync(tso,NULL);
3056 raiseAsync(StgTSO *tso, StgClosure *exception)
3058 StgUpdateFrame* su = tso->su;
3059 StgPtr sp = tso->sp;
3061 /* Thread already dead? */
3062 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
3066 IF_DEBUG(scheduler, sched_belch("raising exception in thread %ld.", tso->id));
3068 /* Remove it from any blocking queues */
3071 /* The stack freezing code assumes there's a closure pointer on
3072 * the top of the stack. This isn't always the case with compiled
3073 * code, so we have to push a dummy closure on the top which just
3074 * returns to the next return address on the stack.
3076 if ( LOOKS_LIKE_GHC_INFO((void*)*sp) ) {
3077 *(--sp) = (W_)&stg_dummy_ret_closure;
3081 nat words = ((P_)su - (P_)sp) - 1;
3085 /* If we find a CATCH_FRAME, and we've got an exception to raise,
3086 * then build the THUNK raise(exception), and leave it on
3087 * top of the CATCH_FRAME ready to enter.
3089 if (get_itbl(su)->type == CATCH_FRAME && exception != NULL) {
3090 StgCatchFrame *cf = (StgCatchFrame *)su;
3093 /* we've got an exception to raise, so let's pass it to the
3094 * handler in this frame.
3096 raise = (StgClosure *)allocate(sizeofW(StgClosure)+1);
3097 TICK_ALLOC_SE_THK(1,0);
3098 SET_HDR(raise,&stg_raise_info,cf->header.prof.ccs);
3099 raise->payload[0] = exception;
3101 /* throw away the stack from Sp up to the CATCH_FRAME.
3105 /* Ensure that async excpetions are blocked now, so we don't get
3106 * a surprise exception before we get around to executing the
3109 if (tso->blocked_exceptions == NULL) {
3110 tso->blocked_exceptions = END_TSO_QUEUE;
3113 /* Put the newly-built THUNK on top of the stack, ready to execute
3114 * when the thread restarts.
3119 tso->what_next = ThreadEnterGHC;
3120 IF_DEBUG(sanity, checkTSO(tso));
3124 /* First build an AP_UPD consisting of the stack chunk above the
3125 * current update frame, with the top word on the stack as the
3128 ap = (StgAP_UPD *)allocate(AP_sizeW(words));
3133 ap->fun = (StgClosure *)sp[0];
3135 for(i=0; i < (nat)words; ++i) {
3136 ap->payload[i] = (StgClosure *)*sp++;
3139 switch (get_itbl(su)->type) {
3143 SET_HDR(ap,&stg_AP_UPD_info,su->header.prof.ccs /* ToDo */);
3144 TICK_ALLOC_UP_THK(words+1,0);
3147 fprintf(stderr, "scheduler: Updating ");
3148 printPtr((P_)su->updatee);
3149 fprintf(stderr, " with ");
3150 printObj((StgClosure *)ap);
3153 /* Replace the updatee with an indirection - happily
3154 * this will also wake up any threads currently
3155 * waiting on the result.
3157 * Warning: if we're in a loop, more than one update frame on
3158 * the stack may point to the same object. Be careful not to
3159 * overwrite an IND_OLDGEN in this case, because we'll screw
3160 * up the mutable lists. To be on the safe side, don't
3161 * overwrite any kind of indirection at all. See also
3162 * threadSqueezeStack in GC.c, where we have to make a similar
3165 if (!closure_IND(su->updatee)) {
3166 UPD_IND_NOLOCK(su->updatee,ap); /* revert the black hole */
3169 sp += sizeofW(StgUpdateFrame) -1;
3170 sp[0] = (W_)ap; /* push onto stack */
3176 StgCatchFrame *cf = (StgCatchFrame *)su;
3179 /* We want a PAP, not an AP_UPD. Fortunately, the
3180 * layout's the same.
3182 SET_HDR(ap,&stg_PAP_info,su->header.prof.ccs /* ToDo */);
3183 TICK_ALLOC_UPD_PAP(words+1,0);
3185 /* now build o = FUN(catch,ap,handler) */
3186 o = (StgClosure *)allocate(sizeofW(StgClosure)+2);
3187 TICK_ALLOC_FUN(2,0);
3188 SET_HDR(o,&stg_catch_info,su->header.prof.ccs /* ToDo */);
3189 o->payload[0] = (StgClosure *)ap;
3190 o->payload[1] = cf->handler;
3193 fprintf(stderr, "scheduler: Built ");
3194 printObj((StgClosure *)o);
3197 /* pop the old handler and put o on the stack */
3199 sp += sizeofW(StgCatchFrame) - 1;
3206 StgSeqFrame *sf = (StgSeqFrame *)su;
3209 SET_HDR(ap,&stg_PAP_info,su->header.prof.ccs /* ToDo */);
3210 TICK_ALLOC_UPD_PAP(words+1,0);
3212 /* now build o = FUN(seq,ap) */
3213 o = (StgClosure *)allocate(sizeofW(StgClosure)+1);
3214 TICK_ALLOC_SE_THK(1,0);
3215 SET_HDR(o,&stg_seq_info,su->header.prof.ccs /* ToDo */);
3216 o->payload[0] = (StgClosure *)ap;
3219 fprintf(stderr, "scheduler: Built ");
3220 printObj((StgClosure *)o);
3223 /* pop the old handler and put o on the stack */
3225 sp += sizeofW(StgSeqFrame) - 1;
3231 /* We've stripped the entire stack, the thread is now dead. */
3232 sp += sizeofW(StgStopFrame) - 1;
3233 sp[0] = (W_)exception; /* save the exception */
3234 tso->what_next = ThreadKilled;
3235 tso->su = (StgUpdateFrame *)(sp+1);
3246 /* -----------------------------------------------------------------------------
3247 resurrectThreads is called after garbage collection on the list of
3248 threads found to be garbage. Each of these threads will be woken
3249 up and sent a signal: BlockedOnDeadMVar if the thread was blocked
3250 on an MVar, or NonTermination if the thread was blocked on a Black
3252 -------------------------------------------------------------------------- */
3255 resurrectThreads( StgTSO *threads )
3259 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
3260 next = tso->global_link;
3261 tso->global_link = all_threads;
3263 IF_DEBUG(scheduler, sched_belch("resurrecting thread %d", tso->id));
3265 switch (tso->why_blocked) {
3267 case BlockedOnException:
3268 raiseAsync(tso,(StgClosure *)BlockedOnDeadMVar_closure);
3270 case BlockedOnBlackHole:
3271 raiseAsync(tso,(StgClosure *)NonTermination_closure);
3274 /* This might happen if the thread was blocked on a black hole
3275 * belonging to a thread that we've just woken up (raiseAsync
3276 * can wake up threads, remember...).
3280 barf("resurrectThreads: thread blocked in a strange way");
3285 /* -----------------------------------------------------------------------------
3286 * Blackhole detection: if we reach a deadlock, test whether any
3287 * threads are blocked on themselves. Any threads which are found to
3288 * be self-blocked get sent a NonTermination exception.
3290 * This is only done in a deadlock situation in order to avoid
3291 * performance overhead in the normal case.
3292 * -------------------------------------------------------------------------- */
3295 detectBlackHoles( void )
3297 StgTSO *t = all_threads;
3298 StgUpdateFrame *frame;
3299 StgClosure *blocked_on;
3301 for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
3303 while (t->what_next == ThreadRelocated) {
3305 ASSERT(get_itbl(t)->type == TSO);
3308 if (t->why_blocked != BlockedOnBlackHole) {
3312 blocked_on = t->block_info.closure;
3314 for (frame = t->su; ; frame = frame->link) {
3315 switch (get_itbl(frame)->type) {
3318 if (frame->updatee == blocked_on) {
3319 /* We are blocking on one of our own computations, so
3320 * send this thread the NonTermination exception.
3323 sched_belch("thread %d is blocked on itself", t->id));
3324 raiseAsync(t, (StgClosure *)NonTermination_closure);
3345 //@node Debugging Routines, Index, Exception Handling Routines, Main scheduling code
3346 //@subsection Debugging Routines
3348 /* -----------------------------------------------------------------------------
3349 Debugging: why is a thread blocked
3350 -------------------------------------------------------------------------- */
3355 printThreadBlockage(StgTSO *tso)
3357 switch (tso->why_blocked) {
3359 fprintf(stderr,"is blocked on read from fd %d", tso->block_info.fd);
3361 case BlockedOnWrite:
3362 fprintf(stderr,"is blocked on write to fd %d", tso->block_info.fd);
3364 case BlockedOnDelay:
3365 fprintf(stderr,"is blocked until %d", tso->block_info.target);
3368 fprintf(stderr,"is blocked on an MVar");
3370 case BlockedOnException:
3371 fprintf(stderr,"is blocked on delivering an exception to thread %d",
3372 tso->block_info.tso->id);
3374 case BlockedOnBlackHole:
3375 fprintf(stderr,"is blocked on a black hole");
3378 fprintf(stderr,"is not blocked");
3382 fprintf(stderr,"is blocked on global address; local FM_BQ is %p (%s)",
3383 tso->block_info.closure, info_type(tso->block_info.closure));
3385 case BlockedOnGA_NoSend:
3386 fprintf(stderr,"is blocked on global address (no send); local FM_BQ is %p (%s)",
3387 tso->block_info.closure, info_type(tso->block_info.closure));
3390 #if defined(RTS_SUPPORTS_THREADS)
3391 case BlockedOnCCall:
3392 fprintf(stderr,"is blocked on an external call");
3396 barf("printThreadBlockage: strange tso->why_blocked: %d for TSO %d (%d)",
3397 tso->why_blocked, tso->id, tso);
3402 printThreadStatus(StgTSO *tso)
3404 switch (tso->what_next) {
3406 fprintf(stderr,"has been killed");
3408 case ThreadComplete:
3409 fprintf(stderr,"has completed");
3412 printThreadBlockage(tso);
3417 printAllThreads(void)
3422 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
3423 ullong_format_string(TIME_ON_PROC(CurrentProc),
3424 time_string, rtsFalse/*no commas!*/);
3426 sched_belch("all threads at [%s]:", time_string);
3428 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
3429 ullong_format_string(CURRENT_TIME,
3430 time_string, rtsFalse/*no commas!*/);
3432 sched_belch("all threads at [%s]:", time_string);
3434 sched_belch("all threads:");
3437 for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
3438 fprintf(stderr, "\tthread %d ", t->id);
3439 printThreadStatus(t);
3440 fprintf(stderr,"\n");
3445 Print a whole blocking queue attached to node (debugging only).
3450 print_bq (StgClosure *node)
3452 StgBlockingQueueElement *bqe;
3456 fprintf(stderr,"## BQ of closure %p (%s): ",
3457 node, info_type(node));
3459 /* should cover all closures that may have a blocking queue */
3460 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
3461 get_itbl(node)->type == FETCH_ME_BQ ||
3462 get_itbl(node)->type == RBH ||
3463 get_itbl(node)->type == MVAR);
3465 ASSERT(node!=(StgClosure*)NULL); // sanity check
3467 print_bqe(((StgBlockingQueue*)node)->blocking_queue);
3471 Print a whole blocking queue starting with the element bqe.
3474 print_bqe (StgBlockingQueueElement *bqe)
3479 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
3481 for (end = (bqe==END_BQ_QUEUE);
3482 !end; // iterate until bqe points to a CONSTR
3483 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE),
3484 bqe = end ? END_BQ_QUEUE : bqe->link) {
3485 ASSERT(bqe != END_BQ_QUEUE); // sanity check
3486 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
3487 /* types of closures that may appear in a blocking queue */
3488 ASSERT(get_itbl(bqe)->type == TSO ||
3489 get_itbl(bqe)->type == BLOCKED_FETCH ||
3490 get_itbl(bqe)->type == CONSTR);
3491 /* only BQs of an RBH end with an RBH_Save closure */
3492 //ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
3494 switch (get_itbl(bqe)->type) {
3496 fprintf(stderr," TSO %u (%x),",
3497 ((StgTSO *)bqe)->id, ((StgTSO *)bqe));
3500 fprintf(stderr," BF (node=%p, ga=((%x, %d, %x)),",
3501 ((StgBlockedFetch *)bqe)->node,
3502 ((StgBlockedFetch *)bqe)->ga.payload.gc.gtid,
3503 ((StgBlockedFetch *)bqe)->ga.payload.gc.slot,
3504 ((StgBlockedFetch *)bqe)->ga.weight);
3507 fprintf(stderr," %s (IP %p),",
3508 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
3509 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
3510 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
3511 "RBH_Save_?"), get_itbl(bqe));
3514 barf("Unexpected closure type %s in blocking queue", // of %p (%s)",
3515 info_type((StgClosure *)bqe)); // , node, info_type(node));
3519 fputc('\n', stderr);
3521 # elif defined(GRAN)
3523 print_bq (StgClosure *node)
3525 StgBlockingQueueElement *bqe;
3526 PEs node_loc, tso_loc;
3529 /* should cover all closures that may have a blocking queue */
3530 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
3531 get_itbl(node)->type == FETCH_ME_BQ ||
3532 get_itbl(node)->type == RBH);
3534 ASSERT(node!=(StgClosure*)NULL); // sanity check
3535 node_loc = where_is(node);
3537 fprintf(stderr,"## BQ of closure %p (%s) on [PE %d]: ",
3538 node, info_type(node), node_loc);
3541 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
3543 for (bqe = ((StgBlockingQueue*)node)->blocking_queue, end = (bqe==END_BQ_QUEUE);
3544 !end; // iterate until bqe points to a CONSTR
3545 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) {
3546 ASSERT(bqe != END_BQ_QUEUE); // sanity check
3547 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
3548 /* types of closures that may appear in a blocking queue */
3549 ASSERT(get_itbl(bqe)->type == TSO ||
3550 get_itbl(bqe)->type == CONSTR);
3551 /* only BQs of an RBH end with an RBH_Save closure */
3552 ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
3554 tso_loc = where_is((StgClosure *)bqe);
3555 switch (get_itbl(bqe)->type) {
3557 fprintf(stderr," TSO %d (%p) on [PE %d],",
3558 ((StgTSO *)bqe)->id, (StgTSO *)bqe, tso_loc);
3561 fprintf(stderr," %s (IP %p),",
3562 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
3563 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
3564 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
3565 "RBH_Save_?"), get_itbl(bqe));
3568 barf("Unexpected closure type %s in blocking queue of %p (%s)",
3569 info_type((StgClosure *)bqe), node, info_type(node));
3573 fputc('\n', stderr);
3577 Nice and easy: only TSOs on the blocking queue
3580 print_bq (StgClosure *node)
3584 ASSERT(node!=(StgClosure*)NULL); // sanity check
3585 for (tso = ((StgBlockingQueue*)node)->blocking_queue;
3586 tso != END_TSO_QUEUE;
3588 ASSERT(tso!=NULL && tso!=END_TSO_QUEUE); // sanity check
3589 ASSERT(get_itbl(tso)->type == TSO); // guess what, sanity check
3590 fprintf(stderr," TSO %d (%p),", tso->id, tso);
3592 fputc('\n', stderr);
3603 for (i=0, tso=run_queue_hd;
3604 tso != END_TSO_QUEUE;
3613 sched_belch(char *s, ...)
3618 fprintf(stderr, "scheduler (task %ld): ", osThreadId());
3620 fprintf(stderr, "== ");
3622 fprintf(stderr, "scheduler: ");
3624 vfprintf(stderr, s, ap);
3625 fprintf(stderr, "\n");
3631 //@node Index, , Debugging Routines, Main scheduling code
3635 //* StgMainThread:: @cindex\s-+StgMainThread
3636 //* awaken_blocked_queue:: @cindex\s-+awaken_blocked_queue
3637 //* blocked_queue_hd:: @cindex\s-+blocked_queue_hd
3638 //* blocked_queue_tl:: @cindex\s-+blocked_queue_tl
3639 //* context_switch:: @cindex\s-+context_switch
3640 //* createThread:: @cindex\s-+createThread
3641 //* gc_pending_cond:: @cindex\s-+gc_pending_cond
3642 //* initScheduler:: @cindex\s-+initScheduler
3643 //* interrupted:: @cindex\s-+interrupted
3644 //* next_thread_id:: @cindex\s-+next_thread_id
3645 //* print_bq:: @cindex\s-+print_bq
3646 //* run_queue_hd:: @cindex\s-+run_queue_hd
3647 //* run_queue_tl:: @cindex\s-+run_queue_tl
3648 //* sched_mutex:: @cindex\s-+sched_mutex
3649 //* schedule:: @cindex\s-+schedule
3650 //* take_off_run_queue:: @cindex\s-+take_off_run_queue
3651 //* term_mutex:: @cindex\s-+term_mutex