1 /* ---------------------------------------------------------------------------
2 * $Id: Schedule.c,v 1.148 2002/07/17 09:21:50 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"
99 #include "ThreadLabels.h"
101 #include "Proftimer.h"
102 #include "ProfHeap.h"
104 #if defined(GRAN) || defined(PAR)
105 # include "GranSimRts.h"
106 # include "GranSim.h"
107 # include "ParallelRts.h"
108 # include "Parallel.h"
109 # include "ParallelDebug.h"
110 # include "FetchMe.h"
114 #include "Capability.h"
115 #include "OSThreads.h"
118 #ifdef HAVE_SYS_TYPES_H
119 #include <sys/types.h>
129 //@node Variables and Data structures, Prototypes, Includes, Main scheduling code
130 //@subsection Variables and Data structures
132 /* Main thread queue.
133 * Locks required: sched_mutex.
135 StgMainThread *main_threads;
138 * Locks required: sched_mutex.
142 StgTSO* ActiveTSO = NULL; /* for assigning system costs; GranSim-Light only */
143 /* rtsTime TimeOfNextEvent, EndOfTimeSlice; now in GranSim.c */
146 In GranSim we have a runnable and a blocked queue for each processor.
147 In order to minimise code changes new arrays run_queue_hds/tls
148 are created. run_queue_hd is then a short cut (macro) for
149 run_queue_hds[CurrentProc] (see GranSim.h).
152 StgTSO *run_queue_hds[MAX_PROC], *run_queue_tls[MAX_PROC];
153 StgTSO *blocked_queue_hds[MAX_PROC], *blocked_queue_tls[MAX_PROC];
154 StgTSO *ccalling_threadss[MAX_PROC];
155 /* We use the same global list of threads (all_threads) in GranSim as in
156 the std RTS (i.e. we are cheating). However, we don't use this list in
157 the GranSim specific code at the moment (so we are only potentially
162 StgTSO *run_queue_hd, *run_queue_tl;
163 StgTSO *blocked_queue_hd, *blocked_queue_tl;
164 StgTSO *sleeping_queue; /* perhaps replace with a hash table? */
168 /* Linked list of all threads.
169 * Used for detecting garbage collected threads.
173 /* When a thread performs a safe C call (_ccall_GC, using old
174 * terminology), it gets put on the suspended_ccalling_threads
175 * list. Used by the garbage collector.
177 static StgTSO *suspended_ccalling_threads;
179 static StgTSO *threadStackOverflow(StgTSO *tso);
181 /* KH: The following two flags are shared memory locations. There is no need
182 to lock them, since they are only unset at the end of a scheduler
186 /* flag set by signal handler to precipitate a context switch */
187 //@cindex context_switch
190 /* if this flag is set as well, give up execution */
191 //@cindex interrupted
194 /* Next thread ID to allocate.
195 * Locks required: thread_id_mutex
197 //@cindex next_thread_id
198 StgThreadID next_thread_id = 1;
201 * Pointers to the state of the current thread.
202 * Rule of thumb: if CurrentTSO != NULL, then we're running a Haskell
203 * thread. If CurrentTSO == NULL, then we're at the scheduler level.
206 /* The smallest stack size that makes any sense is:
207 * RESERVED_STACK_WORDS (so we can get back from the stack overflow)
208 * + sizeofW(StgStopFrame) (the stg_stop_thread_info frame)
209 * + 1 (the realworld token for an IO thread)
210 * + 1 (the closure to enter)
212 * A thread with this stack will bomb immediately with a stack
213 * overflow, which will increase its stack size.
216 #define MIN_STACK_WORDS (RESERVED_STACK_WORDS + sizeofW(StgStopFrame) + 2)
223 /* This is used in `TSO.h' and gcc 2.96 insists that this variable actually
224 * exists - earlier gccs apparently didn't.
232 * Set to TRUE when entering a shutdown state (via shutdownHaskellAndExit()) --
233 * in an MT setting, needed to signal that a worker thread shouldn't hang around
234 * in the scheduler when it is out of work.
236 static rtsBool shutting_down_scheduler = rtsFalse;
238 void addToBlockedQueue ( StgTSO *tso );
240 static void schedule ( void );
241 void interruptStgRts ( void );
243 static void detectBlackHoles ( void );
246 static void sched_belch(char *s, ...);
249 #if defined(RTS_SUPPORTS_THREADS)
250 /* ToDo: carefully document the invariants that go together
251 * with these synchronisation objects.
253 Mutex sched_mutex = INIT_MUTEX_VAR;
254 Mutex term_mutex = INIT_MUTEX_VAR;
257 * A heavyweight solution to the problem of protecting
258 * the thread_id from concurrent update.
260 Mutex thread_id_mutex = INIT_MUTEX_VAR;
264 static Condition gc_pending_cond = INIT_COND_VAR;
268 #endif /* RTS_SUPPORTS_THREADS */
272 rtsTime TimeOfLastYield;
273 rtsBool emitSchedule = rtsTrue;
277 char *whatNext_strs[] = {
285 char *threadReturnCode_strs[] = {
286 "HeapOverflow", /* might also be StackOverflow */
295 StgTSO * createSparkThread(rtsSpark spark);
296 StgTSO * activateSpark (rtsSpark spark);
300 * The thread state for the main thread.
301 // ToDo: check whether not needed any more
305 #if defined(PAR) || defined(RTS_SUPPORTS_THREADS)
306 static void taskStart(void);
317 //@node Main scheduling loop, Suspend and Resume, Prototypes, Main scheduling code
318 //@subsection Main scheduling loop
320 /* ---------------------------------------------------------------------------
321 Main scheduling loop.
323 We use round-robin scheduling, each thread returning to the
324 scheduler loop when one of these conditions is detected:
327 * timer expires (thread yields)
332 Locking notes: we acquire the scheduler lock once at the beginning
333 of the scheduler loop, and release it when
335 * running a thread, or
336 * waiting for work, or
337 * waiting for a GC to complete.
340 In a GranSim setup this loop iterates over the global event queue.
341 This revolves around the global event queue, which determines what
342 to do next. Therefore, it's more complicated than either the
343 concurrent or the parallel (GUM) setup.
346 GUM iterates over incoming messages.
347 It starts with nothing to do (thus CurrentTSO == END_TSO_QUEUE),
348 and sends out a fish whenever it has nothing to do; in-between
349 doing the actual reductions (shared code below) it processes the
350 incoming messages and deals with delayed operations
351 (see PendingFetches).
352 This is not the ugliest code you could imagine, but it's bloody close.
354 ------------------------------------------------------------------------ */
361 StgThreadReturnCode ret;
369 rtsBool receivedFinish = rtsFalse;
371 nat tp_size, sp_size; // stats only
374 rtsBool was_interrupted = rtsFalse;
376 ACQUIRE_LOCK(&sched_mutex);
378 #if defined(RTS_SUPPORTS_THREADS)
379 waitForWorkCapability(&sched_mutex, &cap, rtsFalse);
381 /* simply initialise it in the non-threaded case */
382 grabCapability(&cap);
386 /* set up first event to get things going */
387 /* ToDo: assign costs for system setup and init MainTSO ! */
388 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
390 CurrentTSO, (StgClosure*)NULL, (rtsSpark*)NULL);
393 fprintf(stderr, "GRAN: Init CurrentTSO (in schedule) = %p\n", CurrentTSO);
394 G_TSO(CurrentTSO, 5));
396 if (RtsFlags.GranFlags.Light) {
397 /* Save current time; GranSim Light only */
398 CurrentTSO->gran.clock = CurrentTime[CurrentProc];
401 event = get_next_event();
403 while (event!=(rtsEvent*)NULL) {
404 /* Choose the processor with the next event */
405 CurrentProc = event->proc;
406 CurrentTSO = event->tso;
410 while (!receivedFinish) { /* set by processMessages */
411 /* when receiving PP_FINISH message */
418 IF_DEBUG(scheduler, printAllThreads());
420 #if defined(RTS_SUPPORTS_THREADS)
421 /* Check to see whether there are any worker threads
422 waiting to deposit external call results. If so,
423 yield our capability */
424 yieldToReturningWorker(&sched_mutex, &cap);
427 /* If we're interrupted (the user pressed ^C, or some other
428 * termination condition occurred), kill all the currently running
432 IF_DEBUG(scheduler, sched_belch("interrupted"));
434 interrupted = rtsFalse;
435 was_interrupted = rtsTrue;
438 /* Go through the list of main threads and wake up any
439 * clients whose computations have finished. ToDo: this
440 * should be done more efficiently without a linear scan
441 * of the main threads list, somehow...
443 #if defined(RTS_SUPPORTS_THREADS)
445 StgMainThread *m, **prev;
446 prev = &main_threads;
447 for (m = main_threads; m != NULL; m = m->link) {
448 switch (m->tso->what_next) {
451 *(m->ret) = (StgClosure *)m->tso->sp[0];
455 broadcastCondition(&m->wakeup);
457 removeThreadLabel(m->tso);
461 if (m->ret) *(m->ret) = NULL;
463 if (was_interrupted) {
464 m->stat = Interrupted;
468 broadcastCondition(&m->wakeup);
470 removeThreadLabel(m->tso);
479 #else /* not threaded */
482 /* in GUM do this only on the Main PE */
485 /* If our main thread has finished or been killed, return.
488 StgMainThread *m = main_threads;
489 if (m->tso->what_next == ThreadComplete
490 || m->tso->what_next == ThreadKilled) {
492 removeThreadLabel((StgWord)m->tso);
494 main_threads = main_threads->link;
495 if (m->tso->what_next == ThreadComplete) {
496 /* we finished successfully, fill in the return value */
497 if (m->ret) { *(m->ret) = (StgClosure *)m->tso->sp[0]; };
501 if (m->ret) { *(m->ret) = NULL; };
502 if (was_interrupted) {
503 m->stat = Interrupted;
513 /* Top up the run queue from our spark pool. We try to make the
514 * number of threads in the run queue equal to the number of
517 * Disable spark support in SMP for now, non-essential & requires
518 * a little bit of work to make it compile cleanly. -- sof 1/02.
520 #if 0 /* defined(SMP) */
522 nat n = getFreeCapabilities();
523 StgTSO *tso = run_queue_hd;
525 /* Count the run queue */
526 while (n > 0 && tso != END_TSO_QUEUE) {
533 spark = findSpark(rtsFalse);
535 break; /* no more sparks in the pool */
537 /* I'd prefer this to be done in activateSpark -- HWL */
538 /* tricky - it needs to hold the scheduler lock and
539 * not try to re-acquire it -- SDM */
540 createSparkThread(spark);
542 sched_belch("==^^ turning spark of closure %p into a thread",
543 (StgClosure *)spark));
546 /* We need to wake up the other tasks if we just created some
549 if (getFreeCapabilities() - n > 1) {
550 signalCondition( &thread_ready_cond );
555 /* check for signals each time around the scheduler */
556 #ifndef mingw32_TARGET_OS
557 if (signals_pending()) {
558 RELEASE_LOCK(&sched_mutex); /* ToDo: kill */
559 startSignalHandlers();
560 ACQUIRE_LOCK(&sched_mutex);
564 /* Check whether any waiting threads need to be woken up. If the
565 * run queue is empty, and there are no other tasks running, we
566 * can wait indefinitely for something to happen.
567 * ToDo: what if another client comes along & requests another
570 if ( !EMPTY_QUEUE(blocked_queue_hd) || !EMPTY_QUEUE(sleeping_queue) ) {
571 awaitEvent( EMPTY_RUN_QUEUE()
573 && allFreeCapabilities()
577 /* we can be interrupted while waiting for I/O... */
578 if (interrupted) continue;
581 * Detect deadlock: when we have no threads to run, there are no
582 * threads waiting on I/O or sleeping, and all the other tasks are
583 * waiting for work, we must have a deadlock of some description.
585 * We first try to find threads blocked on themselves (ie. black
586 * holes), and generate NonTermination exceptions where necessary.
588 * If no threads are black holed, we have a deadlock situation, so
589 * inform all the main threads.
592 if ( EMPTY_THREAD_QUEUES()
593 #if defined(RTS_SUPPORTS_THREADS)
594 && EMPTY_QUEUE(suspended_ccalling_threads)
597 && allFreeCapabilities()
601 IF_DEBUG(scheduler, sched_belch("deadlocked, forcing major GC..."));
602 #if defined(THREADED_RTS)
603 /* and SMP mode ..? */
604 releaseCapability(cap);
606 // Garbage collection can release some new threads due to
607 // either (a) finalizers or (b) threads resurrected because
608 // they are about to be send BlockedOnDeadMVar. Any threads
609 // thus released will be immediately runnable.
610 GarbageCollect(GetRoots,rtsTrue);
612 if ( !EMPTY_RUN_QUEUE() ) { goto not_deadlocked; }
615 sched_belch("still deadlocked, checking for black holes..."));
618 if ( !EMPTY_RUN_QUEUE() ) { goto not_deadlocked; }
620 #ifndef mingw32_TARGET_OS
621 /* If we have user-installed signal handlers, then wait
622 * for signals to arrive rather then bombing out with a
625 #if defined(RTS_SUPPORTS_THREADS)
626 if ( 0 ) { /* hmm..what to do? Simply stop waiting for
627 a signal with no runnable threads (or I/O
628 suspended ones) leads nowhere quick.
629 For now, simply shut down when we reach this
632 ToDo: define precisely under what conditions
633 the Scheduler should shut down in an MT setting.
636 if ( anyUserHandlers() ) {
639 sched_belch("still deadlocked, waiting for signals..."));
643 // we might be interrupted...
644 if (interrupted) { continue; }
646 if (signals_pending()) {
647 RELEASE_LOCK(&sched_mutex);
648 startSignalHandlers();
649 ACQUIRE_LOCK(&sched_mutex);
651 ASSERT(!EMPTY_RUN_QUEUE());
656 /* Probably a real deadlock. Send the current main thread the
657 * Deadlock exception (or in the SMP build, send *all* main
658 * threads the deadlock exception, since none of them can make
663 #if defined(RTS_SUPPORTS_THREADS)
664 for (m = main_threads; m != NULL; m = m->link) {
665 switch (m->tso->why_blocked) {
666 case BlockedOnBlackHole:
667 raiseAsync(m->tso, (StgClosure *)NonTermination_closure);
669 case BlockedOnException:
671 raiseAsync(m->tso, (StgClosure *)Deadlock_closure);
674 barf("deadlock: main thread blocked in a strange way");
679 switch (m->tso->why_blocked) {
680 case BlockedOnBlackHole:
681 raiseAsync(m->tso, (StgClosure *)NonTermination_closure);
683 case BlockedOnException:
685 raiseAsync(m->tso, (StgClosure *)Deadlock_closure);
688 barf("deadlock: main thread blocked in a strange way");
693 #if defined(RTS_SUPPORTS_THREADS)
694 /* ToDo: revisit conditions (and mechanism) for shutting
695 down a multi-threaded world */
696 IF_DEBUG(scheduler, sched_belch("all done, i think...shutting down."));
697 RELEASE_LOCK(&sched_mutex);
705 /* ToDo: add deadlock detection in GUM (similar to SMP) -- HWL */
709 /* If there's a GC pending, don't do anything until it has
713 IF_DEBUG(scheduler,sched_belch("waiting for GC"));
714 waitCondition( &gc_pending_cond, &sched_mutex );
718 #if defined(RTS_SUPPORTS_THREADS)
719 /* block until we've got a thread on the run queue and a free
723 if ( EMPTY_RUN_QUEUE() ) {
724 /* Give up our capability */
725 releaseCapability(cap);
727 /* If we're in the process of shutting down (& running the
728 * a batch of finalisers), don't wait around.
730 if ( shutting_down_scheduler ) {
731 RELEASE_LOCK(&sched_mutex);
734 IF_DEBUG(scheduler, sched_belch("thread %d: waiting for work", osThreadId()));
735 waitForWorkCapability(&sched_mutex, &cap, rtsTrue);
736 IF_DEBUG(scheduler, sched_belch("thread %d: work now available", osThreadId()));
741 if (RtsFlags.GranFlags.Light)
742 GranSimLight_enter_system(event, &ActiveTSO); // adjust ActiveTSO etc
744 /* adjust time based on time-stamp */
745 if (event->time > CurrentTime[CurrentProc] &&
746 event->evttype != ContinueThread)
747 CurrentTime[CurrentProc] = event->time;
749 /* Deal with the idle PEs (may issue FindWork or MoveSpark events) */
750 if (!RtsFlags.GranFlags.Light)
753 IF_DEBUG(gran, fprintf(stderr, "GRAN: switch by event-type\n"));
755 /* main event dispatcher in GranSim */
756 switch (event->evttype) {
757 /* Should just be continuing execution */
759 IF_DEBUG(gran, fprintf(stderr, "GRAN: doing ContinueThread\n"));
760 /* ToDo: check assertion
761 ASSERT(run_queue_hd != (StgTSO*)NULL &&
762 run_queue_hd != END_TSO_QUEUE);
764 /* Ignore ContinueThreads for fetching threads (if synchr comm) */
765 if (!RtsFlags.GranFlags.DoAsyncFetch &&
766 procStatus[CurrentProc]==Fetching) {
767 belch("ghuH: Spurious ContinueThread while Fetching ignored; TSO %d (%p) [PE %d]",
768 CurrentTSO->id, CurrentTSO, CurrentProc);
771 /* Ignore ContinueThreads for completed threads */
772 if (CurrentTSO->what_next == ThreadComplete) {
773 belch("ghuH: found a ContinueThread event for completed thread %d (%p) [PE %d] (ignoring ContinueThread)",
774 CurrentTSO->id, CurrentTSO, CurrentProc);
777 /* Ignore ContinueThreads for threads that are being migrated */
778 if (PROCS(CurrentTSO)==Nowhere) {
779 belch("ghuH: trying to run the migrating TSO %d (%p) [PE %d] (ignoring ContinueThread)",
780 CurrentTSO->id, CurrentTSO, CurrentProc);
783 /* The thread should be at the beginning of the run queue */
784 if (CurrentTSO!=run_queue_hds[CurrentProc]) {
785 belch("ghuH: TSO %d (%p) [PE %d] is not at the start of the run_queue when doing a ContinueThread",
786 CurrentTSO->id, CurrentTSO, CurrentProc);
787 break; // run the thread anyway
790 new_event(proc, proc, CurrentTime[proc],
792 (StgTSO*)NULL, (StgClosure*)NULL, (rtsSpark*)NULL);
794 */ /* Catches superfluous CONTINUEs -- should be unnecessary */
795 break; // now actually run the thread; DaH Qu'vam yImuHbej
798 do_the_fetchnode(event);
799 goto next_thread; /* handle next event in event queue */
802 do_the_globalblock(event);
803 goto next_thread; /* handle next event in event queue */
806 do_the_fetchreply(event);
807 goto next_thread; /* handle next event in event queue */
809 case UnblockThread: /* Move from the blocked queue to the tail of */
810 do_the_unblock(event);
811 goto next_thread; /* handle next event in event queue */
813 case ResumeThread: /* Move from the blocked queue to the tail of */
814 /* the runnable queue ( i.e. Qu' SImqa'lu') */
815 event->tso->gran.blocktime +=
816 CurrentTime[CurrentProc] - event->tso->gran.blockedat;
817 do_the_startthread(event);
818 goto next_thread; /* handle next event in event queue */
821 do_the_startthread(event);
822 goto next_thread; /* handle next event in event queue */
825 do_the_movethread(event);
826 goto next_thread; /* handle next event in event queue */
829 do_the_movespark(event);
830 goto next_thread; /* handle next event in event queue */
833 do_the_findwork(event);
834 goto next_thread; /* handle next event in event queue */
837 barf("Illegal event type %u\n", event->evttype);
840 /* This point was scheduler_loop in the old RTS */
842 IF_DEBUG(gran, belch("GRAN: after main switch"));
844 TimeOfLastEvent = CurrentTime[CurrentProc];
845 TimeOfNextEvent = get_time_of_next_event();
846 IgnoreEvents=(TimeOfNextEvent==0); // HWL HACK
847 // CurrentTSO = ThreadQueueHd;
849 IF_DEBUG(gran, belch("GRAN: time of next event is: %ld",
852 if (RtsFlags.GranFlags.Light)
853 GranSimLight_leave_system(event, &ActiveTSO);
855 EndOfTimeSlice = CurrentTime[CurrentProc]+RtsFlags.GranFlags.time_slice;
858 belch("GRAN: end of time-slice is %#lx", EndOfTimeSlice));
860 /* in a GranSim setup the TSO stays on the run queue */
862 /* Take a thread from the run queue. */
863 t = POP_RUN_QUEUE(); // take_off_run_queue(t);
866 fprintf(stderr, "GRAN: About to run current thread, which is\n");
869 context_switch = 0; // turned on via GranYield, checking events and time slice
872 DumpGranEvent(GR_SCHEDULE, t));
874 procStatus[CurrentProc] = Busy;
877 if (PendingFetches != END_BF_QUEUE) {
881 /* ToDo: phps merge with spark activation above */
882 /* check whether we have local work and send requests if we have none */
883 if (EMPTY_RUN_QUEUE()) { /* no runnable threads */
884 /* :-[ no local threads => look out for local sparks */
885 /* the spark pool for the current PE */
886 pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable
887 if (advisory_thread_count < RtsFlags.ParFlags.maxThreads &&
888 pool->hd < pool->tl) {
890 * ToDo: add GC code check that we really have enough heap afterwards!!
892 * If we're here (no runnable threads) and we have pending
893 * sparks, we must have a space problem. Get enough space
894 * to turn one of those pending sparks into a
898 spark = findSpark(rtsFalse); /* get a spark */
899 if (spark != (rtsSpark) NULL) {
900 tso = activateSpark(spark); /* turn the spark into a thread */
901 IF_PAR_DEBUG(schedule,
902 belch("==== schedule: Created TSO %d (%p); %d threads active",
903 tso->id, tso, advisory_thread_count));
905 if (tso==END_TSO_QUEUE) { /* failed to activate spark->back to loop */
906 belch("==^^ failed to activate spark");
908 } /* otherwise fall through & pick-up new tso */
910 IF_PAR_DEBUG(verbose,
911 belch("==^^ no local sparks (spark pool contains only NFs: %d)",
912 spark_queue_len(pool)));
917 /* If we still have no work we need to send a FISH to get a spark
920 if (EMPTY_RUN_QUEUE()) {
921 /* =8-[ no local sparks => look for work on other PEs */
923 * We really have absolutely no work. Send out a fish
924 * (there may be some out there already), and wait for
925 * something to arrive. We clearly can't run any threads
926 * until a SCHEDULE or RESUME arrives, and so that's what
927 * we're hoping to see. (Of course, we still have to
928 * respond to other types of messages.)
930 TIME now = msTime() /*CURRENT_TIME*/;
931 IF_PAR_DEBUG(verbose,
932 belch("-- now=%ld", now));
933 IF_PAR_DEBUG(verbose,
934 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
935 (last_fish_arrived_at!=0 &&
936 last_fish_arrived_at+RtsFlags.ParFlags.fishDelay > now)) {
937 belch("--$$ delaying FISH until %ld (last fish %ld, delay %ld, now %ld)",
938 last_fish_arrived_at+RtsFlags.ParFlags.fishDelay,
939 last_fish_arrived_at,
940 RtsFlags.ParFlags.fishDelay, now);
943 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
944 (last_fish_arrived_at==0 ||
945 (last_fish_arrived_at+RtsFlags.ParFlags.fishDelay <= now))) {
946 /* outstandingFishes is set in sendFish, processFish;
947 avoid flooding system with fishes via delay */
949 sendFish(pe, mytid, NEW_FISH_AGE, NEW_FISH_HISTORY,
952 // Global statistics: count no. of fishes
953 if (RtsFlags.ParFlags.ParStats.Global &&
954 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
955 globalParStats.tot_fish_mess++;
959 receivedFinish = processMessages();
962 } else if (PacketsWaiting()) { /* Look for incoming messages */
963 receivedFinish = processMessages();
966 /* Now we are sure that we have some work available */
967 ASSERT(run_queue_hd != END_TSO_QUEUE);
969 /* Take a thread from the run queue, if we have work */
970 t = POP_RUN_QUEUE(); // take_off_run_queue(END_TSO_QUEUE);
971 IF_DEBUG(sanity,checkTSO(t));
973 /* ToDo: write something to the log-file
974 if (RTSflags.ParFlags.granSimStats && !sameThread)
975 DumpGranEvent(GR_SCHEDULE, RunnableThreadsHd);
979 /* the spark pool for the current PE */
980 pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable
983 belch("--=^ %d threads, %d sparks on [%#x]",
984 run_queue_len(), spark_queue_len(pool), CURRENT_PROC));
987 if (0 && RtsFlags.ParFlags.ParStats.Full &&
988 t && LastTSO && t->id != LastTSO->id &&
989 LastTSO->why_blocked == NotBlocked &&
990 LastTSO->what_next != ThreadComplete) {
991 // if previously scheduled TSO not blocked we have to record the context switch
992 DumpVeryRawGranEvent(TimeOfLastYield, CURRENT_PROC, CURRENT_PROC,
993 GR_DESCHEDULE, LastTSO, (StgClosure *)NULL, 0, 0);
996 if (RtsFlags.ParFlags.ParStats.Full &&
997 (emitSchedule /* forced emit */ ||
998 (t && LastTSO && t->id != LastTSO->id))) {
1000 we are running a different TSO, so write a schedule event to log file
1001 NB: If we use fair scheduling we also have to write a deschedule
1002 event for LastTSO; with unfair scheduling we know that the
1003 previous tso has blocked whenever we switch to another tso, so
1004 we don't need it in GUM for now
1006 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
1007 GR_SCHEDULE, t, (StgClosure *)NULL, 0, 0);
1008 emitSchedule = rtsFalse;
1012 #else /* !GRAN && !PAR */
1014 /* grab a thread from the run queue */
1015 ASSERT(run_queue_hd != END_TSO_QUEUE);
1016 t = POP_RUN_QUEUE();
1017 // Sanity check the thread we're about to run. This can be
1018 // expensive if there is lots of thread switching going on...
1019 IF_DEBUG(sanity,checkTSO(t));
1022 cap->r.rCurrentTSO = t;
1024 /* context switches are now initiated by the timer signal, unless
1025 * the user specified "context switch as often as possible", with
1030 RtsFlags.ProfFlags.profileInterval == 0 ||
1032 (RtsFlags.ConcFlags.ctxtSwitchTicks == 0
1033 && (run_queue_hd != END_TSO_QUEUE
1034 || blocked_queue_hd != END_TSO_QUEUE
1035 || sleeping_queue != END_TSO_QUEUE)))
1040 RELEASE_LOCK(&sched_mutex);
1042 IF_DEBUG(scheduler, sched_belch("-->> Running TSO %ld (%p) %s ...",
1043 t->id, t, whatNext_strs[t->what_next]));
1046 startHeapProfTimer();
1049 /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
1050 /* Run the current thread
1052 switch (cap->r.rCurrentTSO->what_next) {
1054 case ThreadComplete:
1055 /* Thread already finished, return to scheduler. */
1056 ret = ThreadFinished;
1058 case ThreadEnterGHC:
1059 ret = StgRun((StgFunPtr) stg_enterStackTop, &cap->r);
1062 ret = StgRun((StgFunPtr) stg_returnToStackTop, &cap->r);
1064 case ThreadEnterInterp:
1065 ret = interpretBCO(cap);
1068 barf("schedule: invalid what_next field");
1070 /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
1072 /* Costs for the scheduler are assigned to CCS_SYSTEM */
1074 stopHeapProfTimer();
1078 ACQUIRE_LOCK(&sched_mutex);
1081 IF_DEBUG(scheduler,fprintf(stderr,"scheduler (task %ld): ", osThreadId()););
1082 #elif !defined(GRAN) && !defined(PAR)
1083 IF_DEBUG(scheduler,fprintf(stderr,"scheduler: "););
1085 t = cap->r.rCurrentTSO;
1088 /* HACK 675: if the last thread didn't yield, make sure to print a
1089 SCHEDULE event to the log file when StgRunning the next thread, even
1090 if it is the same one as before */
1092 TimeOfLastYield = CURRENT_TIME;
1098 IF_DEBUG(gran, DumpGranEvent(GR_DESCHEDULE, t));
1099 globalGranStats.tot_heapover++;
1101 globalParStats.tot_heapover++;
1104 // did the task ask for a large block?
1105 if (cap->r.rHpAlloc > BLOCK_SIZE_W) {
1106 // if so, get one and push it on the front of the nursery.
1110 blocks = (nat)BLOCK_ROUND_UP(cap->r.rHpAlloc * sizeof(W_)) / BLOCK_SIZE;
1112 IF_DEBUG(scheduler,belch("--<< thread %ld (%p; %s) stopped: requesting a large block (size %d)",
1114 whatNext_strs[t->what_next], blocks));
1116 // don't do this if it would push us over the
1117 // alloc_blocks_lim limit; we'll GC first.
1118 if (alloc_blocks + blocks < alloc_blocks_lim) {
1120 alloc_blocks += blocks;
1121 bd = allocGroup( blocks );
1123 // link the new group into the list
1124 bd->link = cap->r.rCurrentNursery;
1125 bd->u.back = cap->r.rCurrentNursery->u.back;
1126 if (cap->r.rCurrentNursery->u.back != NULL) {
1127 cap->r.rCurrentNursery->u.back->link = bd;
1129 ASSERT(g0s0->blocks == cap->r.rCurrentNursery &&
1130 g0s0->blocks == cap->r.rNursery);
1131 cap->r.rNursery = g0s0->blocks = bd;
1133 cap->r.rCurrentNursery->u.back = bd;
1135 // initialise it as a nursery block. We initialise the
1136 // step, gen_no, and flags field of *every* sub-block in
1137 // this large block, because this is easier than making
1138 // sure that we always find the block head of a large
1139 // block whenever we call Bdescr() (eg. evacuate() and
1140 // isAlive() in the GC would both have to do this, at
1144 for (x = bd; x < bd + blocks; x++) {
1152 // don't forget to update the block count in g0s0.
1153 g0s0->n_blocks += blocks;
1154 ASSERT(countBlocks(g0s0->blocks) == g0s0->n_blocks);
1156 // now update the nursery to point to the new block
1157 cap->r.rCurrentNursery = bd;
1159 // we might be unlucky and have another thread get on the
1160 // run queue before us and steal the large block, but in that
1161 // case the thread will just end up requesting another large
1163 PUSH_ON_RUN_QUEUE(t);
1168 /* make all the running tasks block on a condition variable,
1169 * maybe set context_switch and wait till they all pile in,
1170 * then have them wait on a GC condition variable.
1172 IF_DEBUG(scheduler,belch("--<< thread %ld (%p; %s) stopped: HeapOverflow",
1173 t->id, t, whatNext_strs[t->what_next]));
1176 ASSERT(!is_on_queue(t,CurrentProc));
1178 /* Currently we emit a DESCHEDULE event before GC in GUM.
1179 ToDo: either add separate event to distinguish SYSTEM time from rest
1180 or just nuke this DESCHEDULE (and the following SCHEDULE) */
1181 if (0 && RtsFlags.ParFlags.ParStats.Full) {
1182 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
1183 GR_DESCHEDULE, t, (StgClosure *)NULL, 0, 0);
1184 emitSchedule = rtsTrue;
1188 ready_to_gc = rtsTrue;
1189 context_switch = 1; /* stop other threads ASAP */
1190 PUSH_ON_RUN_QUEUE(t);
1191 /* actual GC is done at the end of the while loop */
1197 DumpGranEvent(GR_DESCHEDULE, t));
1198 globalGranStats.tot_stackover++;
1201 // DumpGranEvent(GR_DESCHEDULE, t);
1202 globalParStats.tot_stackover++;
1204 IF_DEBUG(scheduler,belch("--<< thread %ld (%p; %s) stopped, StackOverflow",
1205 t->id, t, whatNext_strs[t->what_next]));
1206 /* just adjust the stack for this thread, then pop it back
1212 /* enlarge the stack */
1213 StgTSO *new_t = threadStackOverflow(t);
1215 /* This TSO has moved, so update any pointers to it from the
1216 * main thread stack. It better not be on any other queues...
1217 * (it shouldn't be).
1219 for (m = main_threads; m != NULL; m = m->link) {
1224 threadPaused(new_t);
1225 PUSH_ON_RUN_QUEUE(new_t);
1229 case ThreadYielding:
1232 DumpGranEvent(GR_DESCHEDULE, t));
1233 globalGranStats.tot_yields++;
1236 // DumpGranEvent(GR_DESCHEDULE, t);
1237 globalParStats.tot_yields++;
1239 /* put the thread back on the run queue. Then, if we're ready to
1240 * GC, check whether this is the last task to stop. If so, wake
1241 * up the GC thread. getThread will block during a GC until the
1245 if (t->what_next == ThreadEnterInterp) {
1246 /* ToDo: or maybe a timer expired when we were in Hugs?
1247 * or maybe someone hit ctrl-C
1249 belch("--<< thread %ld (%p; %s) stopped to switch to Hugs",
1250 t->id, t, whatNext_strs[t->what_next]);
1252 belch("--<< thread %ld (%p; %s) stopped, yielding",
1253 t->id, t, whatNext_strs[t->what_next]);
1260 //belch("&& Doing sanity check on yielding TSO %ld.", t->id);
1262 ASSERT(t->link == END_TSO_QUEUE);
1264 ASSERT(!is_on_queue(t,CurrentProc));
1267 //belch("&& Doing sanity check on all ThreadQueues (and their TSOs).");
1268 checkThreadQsSanity(rtsTrue));
1271 if (RtsFlags.ParFlags.doFairScheduling) {
1272 /* this does round-robin scheduling; good for concurrency */
1273 APPEND_TO_RUN_QUEUE(t);
1275 /* this does unfair scheduling; good for parallelism */
1276 PUSH_ON_RUN_QUEUE(t);
1279 /* this does round-robin scheduling; good for concurrency */
1280 APPEND_TO_RUN_QUEUE(t);
1283 /* add a ContinueThread event to actually process the thread */
1284 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1286 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1288 belch("GRAN: eventq and runnableq after adding yielded thread to queue again:");
1297 belch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: ",
1298 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)));
1299 if (t->block_info.closure!=(StgClosure*)NULL) print_bq(t->block_info.closure));
1301 // ??? needed; should emit block before
1303 DumpGranEvent(GR_DESCHEDULE, t));
1304 prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
1307 ASSERT(procStatus[CurrentProc]==Busy ||
1308 ((procStatus[CurrentProc]==Fetching) &&
1309 (t->block_info.closure!=(StgClosure*)NULL)));
1310 if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
1311 !(!RtsFlags.GranFlags.DoAsyncFetch &&
1312 procStatus[CurrentProc]==Fetching))
1313 procStatus[CurrentProc] = Idle;
1317 belch("--<< thread %ld (%p; %s) stopped, blocking on node %p with BQ: ",
1318 t->id, t, whatNext_strs[t->what_next], t->block_info.closure));
1321 if (t->block_info.closure!=(StgClosure*)NULL)
1322 print_bq(t->block_info.closure));
1324 /* Send a fetch (if BlockedOnGA) and dump event to log file */
1327 /* whatever we schedule next, we must log that schedule */
1328 emitSchedule = rtsTrue;
1331 /* don't need to do anything. Either the thread is blocked on
1332 * I/O, in which case we'll have called addToBlockedQueue
1333 * previously, or it's blocked on an MVar or Blackhole, in which
1334 * case it'll be on the relevant queue already.
1337 fprintf(stderr, "--<< thread %d (%p) stopped: ", t->id, t);
1338 printThreadBlockage(t);
1339 fprintf(stderr, "\n"));
1341 /* Only for dumping event to log file
1342 ToDo: do I need this in GranSim, too?
1349 case ThreadFinished:
1350 /* Need to check whether this was a main thread, and if so, signal
1351 * the task that started it with the return value. If we have no
1352 * more main threads, we probably need to stop all the tasks until
1355 /* We also end up here if the thread kills itself with an
1356 * uncaught exception, see Exception.hc.
1358 IF_DEBUG(scheduler,belch("--++ thread %d (%p) finished", t->id, t));
1360 endThread(t, CurrentProc); // clean-up the thread
1362 /* For now all are advisory -- HWL */
1363 //if(t->priority==AdvisoryPriority) ??
1364 advisory_thread_count--;
1367 if(t->dist.priority==RevalPriority)
1371 if (RtsFlags.ParFlags.ParStats.Full &&
1372 !RtsFlags.ParFlags.ParStats.Suppressed)
1373 DumpEndEvent(CURRENT_PROC, t, rtsFalse /* not mandatory */);
1378 barf("schedule: invalid thread return code %d", (int)ret);
1382 if (RtsFlags.ProfFlags.profileInterval==0 || performHeapProfile) {
1383 GarbageCollect(GetRoots, rtsTrue);
1385 performHeapProfile = rtsFalse;
1386 ready_to_gc = rtsFalse; // we already GC'd
1392 && allFreeCapabilities()
1395 /* everybody back, start the GC.
1396 * Could do it in this thread, or signal a condition var
1397 * to do it in another thread. Either way, we need to
1398 * broadcast on gc_pending_cond afterward.
1400 #if defined(RTS_SUPPORTS_THREADS)
1401 IF_DEBUG(scheduler,sched_belch("doing GC"));
1403 GarbageCollect(GetRoots,rtsFalse);
1404 ready_to_gc = rtsFalse;
1406 broadcastCondition(&gc_pending_cond);
1409 /* add a ContinueThread event to continue execution of current thread */
1410 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1412 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1414 fprintf(stderr, "GRAN: eventq and runnableq after Garbage collection:\n");
1422 IF_GRAN_DEBUG(unused,
1423 print_eventq(EventHd));
1425 event = get_next_event();
1428 /* ToDo: wait for next message to arrive rather than busy wait */
1431 } /* end of while(1) */
1433 IF_PAR_DEBUG(verbose,
1434 belch("== Leaving schedule() after having received Finish"));
1437 /* ---------------------------------------------------------------------------
1438 * Singleton fork(). Do not copy any running threads.
1439 * ------------------------------------------------------------------------- */
1441 StgInt forkProcess(StgTSO* tso) {
1443 #ifndef mingw32_TARGET_OS
1447 IF_DEBUG(scheduler,sched_belch("forking!"));
1450 if (pid) { /* parent */
1452 /* just return the pid */
1454 } else { /* child */
1455 /* wipe all other threads */
1457 tso->link = END_TSO_QUEUE;
1459 /* DO NOT TOUCH THE QUEUES directly because most of the code around
1460 us is picky about finding the threat still in its queue when
1461 handling the deleteThread() */
1463 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
1465 if (t->id != tso->id) {
1472 barf("forkProcess#: primop not implemented for mingw32, sorry! (%u)\n", tso->id);
1473 /* pointlessly printing out the TSOs 'id' to avoid CC unused warning. */
1475 #endif /* mingw32 */
1478 /* ---------------------------------------------------------------------------
1479 * deleteAllThreads(): kill all the live threads.
1481 * This is used when we catch a user interrupt (^C), before performing
1482 * any necessary cleanups and running finalizers.
1484 * Locks: sched_mutex held.
1485 * ------------------------------------------------------------------------- */
1487 void deleteAllThreads ( void )
1490 IF_DEBUG(scheduler,sched_belch("deleting all threads"));
1491 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
1492 next = t->global_link;
1495 run_queue_hd = run_queue_tl = END_TSO_QUEUE;
1496 blocked_queue_hd = blocked_queue_tl = END_TSO_QUEUE;
1497 sleeping_queue = END_TSO_QUEUE;
1500 /* startThread and insertThread are now in GranSim.c -- HWL */
1503 //@node Suspend and Resume, Run queue code, Main scheduling loop, Main scheduling code
1504 //@subsection Suspend and Resume
1506 /* ---------------------------------------------------------------------------
1507 * Suspending & resuming Haskell threads.
1509 * When making a "safe" call to C (aka _ccall_GC), the task gives back
1510 * its capability before calling the C function. This allows another
1511 * task to pick up the capability and carry on running Haskell
1512 * threads. It also means that if the C call blocks, it won't lock
1515 * The Haskell thread making the C call is put to sleep for the
1516 * duration of the call, on the susepended_ccalling_threads queue. We
1517 * give out a token to the task, which it can use to resume the thread
1518 * on return from the C function.
1519 * ------------------------------------------------------------------------- */
1522 suspendThread( StgRegTable *reg,
1524 #if !defined(RTS_SUPPORTS_THREADS) && !defined(DEBUG)
1532 /* assume that *reg is a pointer to the StgRegTable part
1535 cap = (Capability *)((void *)reg - sizeof(StgFunTable));
1537 ACQUIRE_LOCK(&sched_mutex);
1540 sched_belch("thread %d did a _ccall_gc (is_concurrent: %d)", cap->r.rCurrentTSO->id,concCall));
1542 threadPaused(cap->r.rCurrentTSO);
1543 cap->r.rCurrentTSO->link = suspended_ccalling_threads;
1544 suspended_ccalling_threads = cap->r.rCurrentTSO;
1546 #if defined(RTS_SUPPORTS_THREADS)
1547 cap->r.rCurrentTSO->why_blocked = BlockedOnCCall;
1550 /* Use the thread ID as the token; it should be unique */
1551 tok = cap->r.rCurrentTSO->id;
1553 /* Hand back capability */
1554 releaseCapability(cap);
1556 #if defined(RTS_SUPPORTS_THREADS)
1557 /* Preparing to leave the RTS, so ensure there's a native thread/task
1558 waiting to take over.
1560 ToDo: optimise this and only create a new task if there's a need
1561 for one (i.e., if there's only one Concurrent Haskell thread alive,
1562 there's no need to create a new task).
1564 IF_DEBUG(scheduler, sched_belch("worker thread (%d): leaving RTS", tok));
1566 startTask(taskStart);
1570 /* Other threads _might_ be available for execution; signal this */
1572 RELEASE_LOCK(&sched_mutex);
1577 resumeThread( StgInt tok,
1579 #if !defined(RTS_SUPPORTS_THREADS)
1584 StgTSO *tso, **prev;
1587 #if defined(RTS_SUPPORTS_THREADS)
1588 /* Wait for permission to re-enter the RTS with the result. */
1590 ACQUIRE_LOCK(&sched_mutex);
1591 grabReturnCapability(&sched_mutex, &cap);
1593 grabCapability(&cap);
1596 grabCapability(&cap);
1599 /* Remove the thread off of the suspended list */
1600 prev = &suspended_ccalling_threads;
1601 for (tso = suspended_ccalling_threads;
1602 tso != END_TSO_QUEUE;
1603 prev = &tso->link, tso = tso->link) {
1604 if (tso->id == (StgThreadID)tok) {
1609 if (tso == END_TSO_QUEUE) {
1610 barf("resumeThread: thread not found");
1612 tso->link = END_TSO_QUEUE;
1613 /* Reset blocking status */
1614 tso->why_blocked = NotBlocked;
1616 cap->r.rCurrentTSO = tso;
1617 RELEASE_LOCK(&sched_mutex);
1622 /* ---------------------------------------------------------------------------
1624 * ------------------------------------------------------------------------ */
1625 static void unblockThread(StgTSO *tso);
1627 /* ---------------------------------------------------------------------------
1628 * Comparing Thread ids.
1630 * This is used from STG land in the implementation of the
1631 * instances of Eq/Ord for ThreadIds.
1632 * ------------------------------------------------------------------------ */
1634 int cmp_thread(const StgTSO *tso1, const StgTSO *tso2)
1636 StgThreadID id1 = tso1->id;
1637 StgThreadID id2 = tso2->id;
1639 if (id1 < id2) return (-1);
1640 if (id1 > id2) return 1;
1644 /* ---------------------------------------------------------------------------
1645 * Fetching the ThreadID from an StgTSO.
1647 * This is used in the implementation of Show for ThreadIds.
1648 * ------------------------------------------------------------------------ */
1649 int rts_getThreadId(const StgTSO *tso)
1655 void labelThread(StgTSO *tso, char *label)
1660 /* Caveat: Once set, you can only set the thread name to "" */
1661 len = strlen(label)+1;
1664 fprintf(stderr,"insufficient memory for labelThread!\n");
1666 strncpy(buf,label,len);
1667 /* Update will free the old memory for us */
1668 updateThreadLabel((StgWord)tso,buf);
1672 /* ---------------------------------------------------------------------------
1673 Create a new thread.
1675 The new thread starts with the given stack size. Before the
1676 scheduler can run, however, this thread needs to have a closure
1677 (and possibly some arguments) pushed on its stack. See
1678 pushClosure() in Schedule.h.
1680 createGenThread() and createIOThread() (in SchedAPI.h) are
1681 convenient packaged versions of this function.
1683 currently pri (priority) is only used in a GRAN setup -- HWL
1684 ------------------------------------------------------------------------ */
1685 //@cindex createThread
1687 /* currently pri (priority) is only used in a GRAN setup -- HWL */
1689 createThread(nat size, StgInt pri)
1692 createThread(nat size)
1699 /* First check whether we should create a thread at all */
1701 /* check that no more than RtsFlags.ParFlags.maxThreads threads are created */
1702 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads) {
1704 belch("{createThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
1705 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
1706 return END_TSO_QUEUE;
1712 ASSERT(!RtsFlags.GranFlags.Light || CurrentProc==0);
1715 // ToDo: check whether size = stack_size - TSO_STRUCT_SIZEW
1717 /* catch ridiculously small stack sizes */
1718 if (size < MIN_STACK_WORDS + TSO_STRUCT_SIZEW) {
1719 size = MIN_STACK_WORDS + TSO_STRUCT_SIZEW;
1722 stack_size = size - TSO_STRUCT_SIZEW;
1724 tso = (StgTSO *)allocate(size);
1725 TICK_ALLOC_TSO(stack_size, 0);
1727 SET_HDR(tso, &stg_TSO_info, CCS_SYSTEM);
1729 SET_GRAN_HDR(tso, ThisPE);
1731 tso->what_next = ThreadEnterGHC;
1733 /* tso->id needs to be unique. For now we use a heavyweight mutex to
1734 * protect the increment operation on next_thread_id.
1735 * In future, we could use an atomic increment instead.
1737 ACQUIRE_LOCK(&thread_id_mutex);
1738 tso->id = next_thread_id++;
1739 RELEASE_LOCK(&thread_id_mutex);
1741 tso->why_blocked = NotBlocked;
1742 tso->blocked_exceptions = NULL;
1744 tso->stack_size = stack_size;
1745 tso->max_stack_size = round_to_mblocks(RtsFlags.GcFlags.maxStkSize)
1747 tso->sp = (P_)&(tso->stack) + stack_size;
1750 tso->prof.CCCS = CCS_MAIN;
1753 /* put a stop frame on the stack */
1754 tso->sp -= sizeofW(StgStopFrame);
1755 SET_HDR((StgClosure*)tso->sp,(StgInfoTable *)&stg_stop_thread_info,CCS_SYSTEM);
1756 tso->su = (StgUpdateFrame*)tso->sp;
1760 tso->link = END_TSO_QUEUE;
1761 /* uses more flexible routine in GranSim */
1762 insertThread(tso, CurrentProc);
1764 /* In a non-GranSim setup the pushing of a TSO onto the runq is separated
1770 if (RtsFlags.GranFlags.GranSimStats.Full)
1771 DumpGranEvent(GR_START,tso);
1773 if (RtsFlags.ParFlags.ParStats.Full)
1774 DumpGranEvent(GR_STARTQ,tso);
1775 /* HACk to avoid SCHEDULE
1779 /* Link the new thread on the global thread list.
1781 tso->global_link = all_threads;
1785 tso->dist.priority = MandatoryPriority; //by default that is...
1789 tso->gran.pri = pri;
1791 tso->gran.magic = TSO_MAGIC; // debugging only
1793 tso->gran.sparkname = 0;
1794 tso->gran.startedat = CURRENT_TIME;
1795 tso->gran.exported = 0;
1796 tso->gran.basicblocks = 0;
1797 tso->gran.allocs = 0;
1798 tso->gran.exectime = 0;
1799 tso->gran.fetchtime = 0;
1800 tso->gran.fetchcount = 0;
1801 tso->gran.blocktime = 0;
1802 tso->gran.blockcount = 0;
1803 tso->gran.blockedat = 0;
1804 tso->gran.globalsparks = 0;
1805 tso->gran.localsparks = 0;
1806 if (RtsFlags.GranFlags.Light)
1807 tso->gran.clock = Now; /* local clock */
1809 tso->gran.clock = 0;
1811 IF_DEBUG(gran,printTSO(tso));
1814 tso->par.magic = TSO_MAGIC; // debugging only
1816 tso->par.sparkname = 0;
1817 tso->par.startedat = CURRENT_TIME;
1818 tso->par.exported = 0;
1819 tso->par.basicblocks = 0;
1820 tso->par.allocs = 0;
1821 tso->par.exectime = 0;
1822 tso->par.fetchtime = 0;
1823 tso->par.fetchcount = 0;
1824 tso->par.blocktime = 0;
1825 tso->par.blockcount = 0;
1826 tso->par.blockedat = 0;
1827 tso->par.globalsparks = 0;
1828 tso->par.localsparks = 0;
1832 globalGranStats.tot_threads_created++;
1833 globalGranStats.threads_created_on_PE[CurrentProc]++;
1834 globalGranStats.tot_sq_len += spark_queue_len(CurrentProc);
1835 globalGranStats.tot_sq_probes++;
1837 // collect parallel global statistics (currently done together with GC stats)
1838 if (RtsFlags.ParFlags.ParStats.Global &&
1839 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
1840 //fprintf(stderr, "Creating thread %d @ %11.2f\n", tso->id, usertime());
1841 globalParStats.tot_threads_created++;
1847 belch("==__ schedule: Created TSO %d (%p);",
1848 CurrentProc, tso, tso->id));
1850 IF_PAR_DEBUG(verbose,
1851 belch("==__ schedule: Created TSO %d (%p); %d threads active",
1852 tso->id, tso, advisory_thread_count));
1854 IF_DEBUG(scheduler,sched_belch("created thread %ld, stack size = %lx words",
1855 tso->id, tso->stack_size));
1862 all parallel thread creation calls should fall through the following routine.
1865 createSparkThread(rtsSpark spark)
1867 ASSERT(spark != (rtsSpark)NULL);
1868 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads)
1870 barf("{createSparkThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
1871 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
1872 return END_TSO_QUEUE;
1876 tso = createThread(RtsFlags.GcFlags.initialStkSize);
1877 if (tso==END_TSO_QUEUE)
1878 barf("createSparkThread: Cannot create TSO");
1880 tso->priority = AdvisoryPriority;
1882 pushClosure(tso,spark);
1883 PUSH_ON_RUN_QUEUE(tso);
1884 advisory_thread_count++;
1891 Turn a spark into a thread.
1892 ToDo: fix for SMP (needs to acquire SCHED_MUTEX!)
1895 //@cindex activateSpark
1897 activateSpark (rtsSpark spark)
1901 tso = createSparkThread(spark);
1902 if (RtsFlags.ParFlags.ParStats.Full) {
1903 //ASSERT(run_queue_hd == END_TSO_QUEUE); // I think ...
1904 IF_PAR_DEBUG(verbose,
1905 belch("==^^ activateSpark: turning spark of closure %p (%s) into a thread",
1906 (StgClosure *)spark, info_type((StgClosure *)spark)));
1908 // ToDo: fwd info on local/global spark to thread -- HWL
1909 // tso->gran.exported = spark->exported;
1910 // tso->gran.locked = !spark->global;
1911 // tso->gran.sparkname = spark->name;
1917 static SchedulerStatus waitThread_(/*out*/StgMainThread* m
1918 #if defined(THREADED_RTS)
1919 , rtsBool blockWaiting
1924 /* ---------------------------------------------------------------------------
1927 * scheduleThread puts a thread on the head of the runnable queue.
1928 * This will usually be done immediately after a thread is created.
1929 * The caller of scheduleThread must create the thread using e.g.
1930 * createThread and push an appropriate closure
1931 * on this thread's stack before the scheduler is invoked.
1932 * ------------------------------------------------------------------------ */
1934 static void scheduleThread_ (StgTSO* tso, rtsBool createTask);
1937 scheduleThread_(StgTSO *tso
1938 , rtsBool createTask
1939 #if !defined(THREADED_RTS)
1944 ACQUIRE_LOCK(&sched_mutex);
1946 /* Put the new thread on the head of the runnable queue. The caller
1947 * better push an appropriate closure on this thread's stack
1948 * beforehand. In the SMP case, the thread may start running as
1949 * soon as we release the scheduler lock below.
1951 PUSH_ON_RUN_QUEUE(tso);
1952 #if defined(THREADED_RTS)
1953 /* If main() is scheduling a thread, don't bother creating a
1957 startTask(taskStart);
1963 IF_DEBUG(scheduler,printTSO(tso));
1965 RELEASE_LOCK(&sched_mutex);
1968 void scheduleThread(StgTSO* tso)
1970 scheduleThread_(tso, rtsFalse);
1974 scheduleWaitThread(StgTSO* tso, /*[out]*/HaskellObj* ret)
1978 m = stgMallocBytes(sizeof(StgMainThread), "waitThread");
1982 #if defined(RTS_SUPPORTS_THREADS)
1983 initCondition(&m->wakeup);
1986 /* Put the thread on the main-threads list prior to scheduling the TSO.
1987 Failure to do so introduces a race condition in the MT case (as
1988 identified by Wolfgang Thaller), whereby the new task/OS thread
1989 created by scheduleThread_() would complete prior to the thread
1990 that spawned it managed to put 'itself' on the main-threads list.
1991 The upshot of it all being that the worker thread wouldn't get to
1992 signal the completion of the its work item for the main thread to
1993 see (==> it got stuck waiting.) -- sof 6/02.
1995 ACQUIRE_LOCK(&sched_mutex);
1996 IF_DEBUG(scheduler, sched_belch("== scheduler: waiting for thread (%d)\n", tso->id));
1998 m->link = main_threads;
2001 /* Inefficient (scheduleThread_() acquires it again right away),
2002 * but obviously correct.
2004 RELEASE_LOCK(&sched_mutex);
2006 scheduleThread_(tso, rtsTrue);
2007 #if defined(THREADED_RTS)
2008 return waitThread_(m, rtsTrue);
2010 return waitThread_(m);
2014 /* ---------------------------------------------------------------------------
2017 * Initialise the scheduler. This resets all the queues - if the
2018 * queues contained any threads, they'll be garbage collected at the
2021 * ------------------------------------------------------------------------ */
2025 term_handler(int sig STG_UNUSED)
2028 ACQUIRE_LOCK(&term_mutex);
2030 RELEASE_LOCK(&term_mutex);
2041 for (i=0; i<=MAX_PROC; i++) {
2042 run_queue_hds[i] = END_TSO_QUEUE;
2043 run_queue_tls[i] = END_TSO_QUEUE;
2044 blocked_queue_hds[i] = END_TSO_QUEUE;
2045 blocked_queue_tls[i] = END_TSO_QUEUE;
2046 ccalling_threadss[i] = END_TSO_QUEUE;
2047 sleeping_queue = END_TSO_QUEUE;
2050 run_queue_hd = END_TSO_QUEUE;
2051 run_queue_tl = END_TSO_QUEUE;
2052 blocked_queue_hd = END_TSO_QUEUE;
2053 blocked_queue_tl = END_TSO_QUEUE;
2054 sleeping_queue = END_TSO_QUEUE;
2057 suspended_ccalling_threads = END_TSO_QUEUE;
2059 main_threads = NULL;
2060 all_threads = END_TSO_QUEUE;
2065 RtsFlags.ConcFlags.ctxtSwitchTicks =
2066 RtsFlags.ConcFlags.ctxtSwitchTime / TICK_MILLISECS;
2068 #if defined(RTS_SUPPORTS_THREADS)
2069 /* Initialise the mutex and condition variables used by
2071 initMutex(&sched_mutex);
2072 initMutex(&term_mutex);
2073 initMutex(&thread_id_mutex);
2075 initCondition(&thread_ready_cond);
2079 initCondition(&gc_pending_cond);
2082 #if defined(RTS_SUPPORTS_THREADS)
2083 ACQUIRE_LOCK(&sched_mutex);
2086 /* Install the SIGHUP handler */
2089 struct sigaction action,oact;
2091 action.sa_handler = term_handler;
2092 sigemptyset(&action.sa_mask);
2093 action.sa_flags = 0;
2094 if (sigaction(SIGTERM, &action, &oact) != 0) {
2095 barf("can't install TERM handler");
2100 /* A capability holds the state a native thread needs in
2101 * order to execute STG code. At least one capability is
2102 * floating around (only SMP builds have more than one).
2106 #if defined(RTS_SUPPORTS_THREADS)
2107 /* start our haskell execution tasks */
2109 startTaskManager(RtsFlags.ParFlags.nNodes, taskStart);
2111 startTaskManager(0,taskStart);
2115 #if /* defined(SMP) ||*/ defined(PAR)
2119 #if defined(RTS_SUPPORTS_THREADS)
2120 RELEASE_LOCK(&sched_mutex);
2126 exitScheduler( void )
2128 #if defined(RTS_SUPPORTS_THREADS)
2131 shutting_down_scheduler = rtsTrue;
2134 /* -----------------------------------------------------------------------------
2135 Managing the per-task allocation areas.
2137 Each capability comes with an allocation area. These are
2138 fixed-length block lists into which allocation can be done.
2140 ToDo: no support for two-space collection at the moment???
2141 -------------------------------------------------------------------------- */
2143 /* -----------------------------------------------------------------------------
2144 * waitThread is the external interface for running a new computation
2145 * and waiting for the result.
2147 * In the non-SMP case, we create a new main thread, push it on the
2148 * main-thread stack, and invoke the scheduler to run it. The
2149 * scheduler will return when the top main thread on the stack has
2150 * completed or died, and fill in the necessary fields of the
2151 * main_thread structure.
2153 * In the SMP case, we create a main thread as before, but we then
2154 * create a new condition variable and sleep on it. When our new
2155 * main thread has completed, we'll be woken up and the status/result
2156 * will be in the main_thread struct.
2157 * -------------------------------------------------------------------------- */
2160 howManyThreadsAvail ( void )
2164 for (q = run_queue_hd; q != END_TSO_QUEUE; q = q->link)
2166 for (q = blocked_queue_hd; q != END_TSO_QUEUE; q = q->link)
2168 for (q = sleeping_queue; q != END_TSO_QUEUE; q = q->link)
2174 finishAllThreads ( void )
2177 while (run_queue_hd != END_TSO_QUEUE) {
2178 waitThread ( run_queue_hd, NULL);
2180 while (blocked_queue_hd != END_TSO_QUEUE) {
2181 waitThread ( blocked_queue_hd, NULL);
2183 while (sleeping_queue != END_TSO_QUEUE) {
2184 waitThread ( blocked_queue_hd, NULL);
2187 (blocked_queue_hd != END_TSO_QUEUE ||
2188 run_queue_hd != END_TSO_QUEUE ||
2189 sleeping_queue != END_TSO_QUEUE);
2193 waitThread(StgTSO *tso, /*out*/StgClosure **ret)
2197 m = stgMallocBytes(sizeof(StgMainThread), "waitThread");
2201 #if defined(RTS_SUPPORTS_THREADS)
2202 initCondition(&m->wakeup);
2205 /* see scheduleWaitThread() comment */
2206 ACQUIRE_LOCK(&sched_mutex);
2207 IF_DEBUG(scheduler, sched_belch("== scheduler: waiting for thread (%d)\n", tso->id));
2208 m->link = main_threads;
2210 RELEASE_LOCK(&sched_mutex);
2212 IF_DEBUG(scheduler, sched_belch("== scheduler: waiting for thread (%d)\n", tso->id));
2213 #if defined(THREADED_RTS)
2214 return waitThread_(m, rtsFalse);
2216 return waitThread_(m);
2222 waitThread_(StgMainThread* m
2223 #if defined(THREADED_RTS)
2224 , rtsBool blockWaiting
2228 SchedulerStatus stat;
2230 IF_DEBUG(scheduler, sched_belch("== scheduler: new main thread (%d)\n", m->tso->id));
2232 #if defined(RTS_SUPPORTS_THREADS)
2234 # if defined(THREADED_RTS)
2235 if (!blockWaiting) {
2236 /* In the threaded case, the OS thread that called main()
2237 * gets to enter the RTS directly without going via another
2241 ASSERT(m->stat != NoStatus);
2245 ACQUIRE_LOCK(&sched_mutex);
2247 waitCondition(&m->wakeup, &sched_mutex);
2248 } while (m->stat == NoStatus);
2251 /* GranSim specific init */
2252 CurrentTSO = m->tso; // the TSO to run
2253 procStatus[MainProc] = Busy; // status of main PE
2254 CurrentProc = MainProc; // PE to run it on
2258 RELEASE_LOCK(&sched_mutex);
2260 ASSERT(m->stat != NoStatus);
2265 #if defined(RTS_SUPPORTS_THREADS)
2266 closeCondition(&m->wakeup);
2269 IF_DEBUG(scheduler, fprintf(stderr, "== scheduler: main thread (%d) finished\n",
2273 #if defined(THREADED_RTS)
2276 RELEASE_LOCK(&sched_mutex);
2281 //@node Run queue code, Garbage Collextion Routines, Suspend and Resume, Main scheduling code
2282 //@subsection Run queue code
2286 NB: In GranSim we have many run queues; run_queue_hd is actually a macro
2287 unfolding to run_queue_hds[CurrentProc], thus CurrentProc is an
2288 implicit global variable that has to be correct when calling these
2292 /* Put the new thread on the head of the runnable queue.
2293 * The caller of createThread better push an appropriate closure
2294 * on this thread's stack before the scheduler is invoked.
2296 static /* inline */ void
2297 add_to_run_queue(tso)
2300 ASSERT(tso!=run_queue_hd && tso!=run_queue_tl);
2301 tso->link = run_queue_hd;
2303 if (run_queue_tl == END_TSO_QUEUE) {
2308 /* Put the new thread at the end of the runnable queue. */
2309 static /* inline */ void
2310 push_on_run_queue(tso)
2313 ASSERT(get_itbl((StgClosure *)tso)->type == TSO);
2314 ASSERT(run_queue_hd!=NULL && run_queue_tl!=NULL);
2315 ASSERT(tso!=run_queue_hd && tso!=run_queue_tl);
2316 if (run_queue_hd == END_TSO_QUEUE) {
2319 run_queue_tl->link = tso;
2325 Should be inlined because it's used very often in schedule. The tso
2326 argument is actually only needed in GranSim, where we want to have the
2327 possibility to schedule *any* TSO on the run queue, irrespective of the
2328 actual ordering. Therefore, if tso is not the nil TSO then we traverse
2329 the run queue and dequeue the tso, adjusting the links in the queue.
2331 //@cindex take_off_run_queue
2332 static /* inline */ StgTSO*
2333 take_off_run_queue(StgTSO *tso) {
2337 qetlaHbogh Qu' ngaSbogh ghomDaQ {tso} yIteq!
2339 if tso is specified, unlink that tso from the run_queue (doesn't have
2340 to be at the beginning of the queue); GranSim only
2342 if (tso!=END_TSO_QUEUE) {
2343 /* find tso in queue */
2344 for (t=run_queue_hd, prev=END_TSO_QUEUE;
2345 t!=END_TSO_QUEUE && t!=tso;
2349 /* now actually dequeue the tso */
2350 if (prev!=END_TSO_QUEUE) {
2351 ASSERT(run_queue_hd!=t);
2352 prev->link = t->link;
2354 /* t is at beginning of thread queue */
2355 ASSERT(run_queue_hd==t);
2356 run_queue_hd = t->link;
2358 /* t is at end of thread queue */
2359 if (t->link==END_TSO_QUEUE) {
2360 ASSERT(t==run_queue_tl);
2361 run_queue_tl = prev;
2363 ASSERT(run_queue_tl!=t);
2365 t->link = END_TSO_QUEUE;
2367 /* take tso from the beginning of the queue; std concurrent code */
2369 if (t != END_TSO_QUEUE) {
2370 run_queue_hd = t->link;
2371 t->link = END_TSO_QUEUE;
2372 if (run_queue_hd == END_TSO_QUEUE) {
2373 run_queue_tl = END_TSO_QUEUE;
2382 //@node Garbage Collextion Routines, Blocking Queue Routines, Run queue code, Main scheduling code
2383 //@subsection Garbage Collextion Routines
2385 /* ---------------------------------------------------------------------------
2386 Where are the roots that we know about?
2388 - all the threads on the runnable queue
2389 - all the threads on the blocked queue
2390 - all the threads on the sleeping queue
2391 - all the thread currently executing a _ccall_GC
2392 - all the "main threads"
2394 ------------------------------------------------------------------------ */
2396 /* This has to be protected either by the scheduler monitor, or by the
2397 garbage collection monitor (probably the latter).
2402 GetRoots(evac_fn evac)
2407 for (i=0; i<=RtsFlags.GranFlags.proc; i++) {
2408 if ((run_queue_hds[i] != END_TSO_QUEUE) && ((run_queue_hds[i] != NULL)))
2409 evac((StgClosure **)&run_queue_hds[i]);
2410 if ((run_queue_tls[i] != END_TSO_QUEUE) && ((run_queue_tls[i] != NULL)))
2411 evac((StgClosure **)&run_queue_tls[i]);
2413 if ((blocked_queue_hds[i] != END_TSO_QUEUE) && ((blocked_queue_hds[i] != NULL)))
2414 evac((StgClosure **)&blocked_queue_hds[i]);
2415 if ((blocked_queue_tls[i] != END_TSO_QUEUE) && ((blocked_queue_tls[i] != NULL)))
2416 evac((StgClosure **)&blocked_queue_tls[i]);
2417 if ((ccalling_threadss[i] != END_TSO_QUEUE) && ((ccalling_threadss[i] != NULL)))
2418 evac((StgClosure **)&ccalling_threads[i]);
2425 if (run_queue_hd != END_TSO_QUEUE) {
2426 ASSERT(run_queue_tl != END_TSO_QUEUE);
2427 evac((StgClosure **)&run_queue_hd);
2428 evac((StgClosure **)&run_queue_tl);
2431 if (blocked_queue_hd != END_TSO_QUEUE) {
2432 ASSERT(blocked_queue_tl != END_TSO_QUEUE);
2433 evac((StgClosure **)&blocked_queue_hd);
2434 evac((StgClosure **)&blocked_queue_tl);
2437 if (sleeping_queue != END_TSO_QUEUE) {
2438 evac((StgClosure **)&sleeping_queue);
2442 if (suspended_ccalling_threads != END_TSO_QUEUE) {
2443 evac((StgClosure **)&suspended_ccalling_threads);
2446 #if defined(PAR) || defined(GRAN)
2447 markSparkQueue(evac);
2451 /* -----------------------------------------------------------------------------
2454 This is the interface to the garbage collector from Haskell land.
2455 We provide this so that external C code can allocate and garbage
2456 collect when called from Haskell via _ccall_GC.
2458 It might be useful to provide an interface whereby the programmer
2459 can specify more roots (ToDo).
2461 This needs to be protected by the GC condition variable above. KH.
2462 -------------------------------------------------------------------------- */
2464 void (*extra_roots)(evac_fn);
2469 /* Obligated to hold this lock upon entry */
2470 ACQUIRE_LOCK(&sched_mutex);
2471 GarbageCollect(GetRoots,rtsFalse);
2472 RELEASE_LOCK(&sched_mutex);
2476 performMajorGC(void)
2478 ACQUIRE_LOCK(&sched_mutex);
2479 GarbageCollect(GetRoots,rtsTrue);
2480 RELEASE_LOCK(&sched_mutex);
2484 AllRoots(evac_fn evac)
2486 GetRoots(evac); // the scheduler's roots
2487 extra_roots(evac); // the user's roots
2491 performGCWithRoots(void (*get_roots)(evac_fn))
2493 ACQUIRE_LOCK(&sched_mutex);
2494 extra_roots = get_roots;
2495 GarbageCollect(AllRoots,rtsFalse);
2496 RELEASE_LOCK(&sched_mutex);
2499 /* -----------------------------------------------------------------------------
2502 If the thread has reached its maximum stack size, then raise the
2503 StackOverflow exception in the offending thread. Otherwise
2504 relocate the TSO into a larger chunk of memory and adjust its stack
2506 -------------------------------------------------------------------------- */
2509 threadStackOverflow(StgTSO *tso)
2511 nat new_stack_size, new_tso_size, diff, stack_words;
2515 IF_DEBUG(sanity,checkTSO(tso));
2516 if (tso->stack_size >= tso->max_stack_size) {
2519 belch("@@ threadStackOverflow of TSO %d (%p): stack too large (now %ld; max is %ld",
2520 tso->id, tso, tso->stack_size, tso->max_stack_size);
2521 /* If we're debugging, just print out the top of the stack */
2522 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2525 /* Send this thread the StackOverflow exception */
2526 raiseAsync(tso, (StgClosure *)stackOverflow_closure);
2530 /* Try to double the current stack size. If that takes us over the
2531 * maximum stack size for this thread, then use the maximum instead.
2532 * Finally round up so the TSO ends up as a whole number of blocks.
2534 new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
2535 new_tso_size = (nat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) +
2536 TSO_STRUCT_SIZE)/sizeof(W_);
2537 new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */
2538 new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
2540 IF_DEBUG(scheduler, fprintf(stderr,"== scheduler: increasing stack size from %d words to %d.\n", tso->stack_size, new_stack_size));
2542 dest = (StgTSO *)allocate(new_tso_size);
2543 TICK_ALLOC_TSO(new_stack_size,0);
2545 /* copy the TSO block and the old stack into the new area */
2546 memcpy(dest,tso,TSO_STRUCT_SIZE);
2547 stack_words = tso->stack + tso->stack_size - tso->sp;
2548 new_sp = (P_)dest + new_tso_size - stack_words;
2549 memcpy(new_sp, tso->sp, stack_words * sizeof(W_));
2551 /* relocate the stack pointers... */
2552 diff = (P_)new_sp - (P_)tso->sp; /* In *words* */
2553 dest->su = (StgUpdateFrame *) ((P_)dest->su + diff);
2555 dest->stack_size = new_stack_size;
2557 /* and relocate the update frame list */
2558 relocate_stack(dest, diff);
2560 /* Mark the old TSO as relocated. We have to check for relocated
2561 * TSOs in the garbage collector and any primops that deal with TSOs.
2563 * It's important to set the sp and su values to just beyond the end
2564 * of the stack, so we don't attempt to scavenge any part of the
2567 tso->what_next = ThreadRelocated;
2569 tso->sp = (P_)&(tso->stack[tso->stack_size]);
2570 tso->su = (StgUpdateFrame *)tso->sp;
2571 tso->why_blocked = NotBlocked;
2572 dest->mut_link = NULL;
2574 IF_PAR_DEBUG(verbose,
2575 belch("@@ threadStackOverflow of TSO %d (now at %p): stack size increased to %ld",
2576 tso->id, tso, tso->stack_size);
2577 /* If we're debugging, just print out the top of the stack */
2578 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2581 IF_DEBUG(sanity,checkTSO(tso));
2583 IF_DEBUG(scheduler,printTSO(dest));
2589 //@node Blocking Queue Routines, Exception Handling Routines, Garbage Collextion Routines, Main scheduling code
2590 //@subsection Blocking Queue Routines
2592 /* ---------------------------------------------------------------------------
2593 Wake up a queue that was blocked on some resource.
2594 ------------------------------------------------------------------------ */
2598 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2603 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2605 /* write RESUME events to log file and
2606 update blocked and fetch time (depending on type of the orig closure) */
2607 if (RtsFlags.ParFlags.ParStats.Full) {
2608 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
2609 GR_RESUMEQ, ((StgTSO *)bqe), ((StgTSO *)bqe)->block_info.closure,
2610 0, 0 /* spark_queue_len(ADVISORY_POOL) */);
2611 if (EMPTY_RUN_QUEUE())
2612 emitSchedule = rtsTrue;
2614 switch (get_itbl(node)->type) {
2616 ((StgTSO *)bqe)->par.fetchtime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2621 ((StgTSO *)bqe)->par.blocktime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2628 barf("{unblockOneLocked}Daq Qagh: unexpected closure in blocking queue");
2635 static StgBlockingQueueElement *
2636 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2639 PEs node_loc, tso_loc;
2641 node_loc = where_is(node); // should be lifted out of loop
2642 tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2643 tso_loc = where_is((StgClosure *)tso);
2644 if (IS_LOCAL_TO(PROCS(node),tso_loc)) { // TSO is local
2645 /* !fake_fetch => TSO is on CurrentProc is same as IS_LOCAL_TO */
2646 ASSERT(CurrentProc!=node_loc || tso_loc==CurrentProc);
2647 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.lunblocktime;
2648 // insertThread(tso, node_loc);
2649 new_event(tso_loc, tso_loc, CurrentTime[CurrentProc],
2651 tso, node, (rtsSpark*)NULL);
2652 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2655 } else { // TSO is remote (actually should be FMBQ)
2656 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.mpacktime +
2657 RtsFlags.GranFlags.Costs.gunblocktime +
2658 RtsFlags.GranFlags.Costs.latency;
2659 new_event(tso_loc, CurrentProc, CurrentTime[CurrentProc],
2661 tso, node, (rtsSpark*)NULL);
2662 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2665 /* the thread-queue-overhead is accounted for in either Resume or UnblockThread */
2667 fprintf(stderr," %s TSO %d (%p) [PE %d] (block_info.closure=%p) (next=%p) ,",
2668 (node_loc==tso_loc ? "Local" : "Global"),
2669 tso->id, tso, CurrentProc, tso->block_info.closure, tso->link));
2670 tso->block_info.closure = NULL;
2671 IF_DEBUG(scheduler,belch("-- Waking up thread %ld (%p)",
2675 static StgBlockingQueueElement *
2676 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2678 StgBlockingQueueElement *next;
2680 switch (get_itbl(bqe)->type) {
2682 ASSERT(((StgTSO *)bqe)->why_blocked != NotBlocked);
2683 /* if it's a TSO just push it onto the run_queue */
2685 // ((StgTSO *)bqe)->link = END_TSO_QUEUE; // debugging?
2686 PUSH_ON_RUN_QUEUE((StgTSO *)bqe);
2688 unblockCount(bqe, node);
2689 /* reset blocking status after dumping event */
2690 ((StgTSO *)bqe)->why_blocked = NotBlocked;
2694 /* if it's a BLOCKED_FETCH put it on the PendingFetches list */
2696 bqe->link = (StgBlockingQueueElement *)PendingFetches;
2697 PendingFetches = (StgBlockedFetch *)bqe;
2701 /* can ignore this case in a non-debugging setup;
2702 see comments on RBHSave closures above */
2704 /* check that the closure is an RBHSave closure */
2705 ASSERT(get_itbl((StgClosure *)bqe) == &stg_RBH_Save_0_info ||
2706 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_1_info ||
2707 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_2_info);
2711 barf("{unblockOneLocked}Daq Qagh: Unexpected IP (%#lx; %s) in blocking queue at %#lx\n",
2712 get_itbl((StgClosure *)bqe), info_type((StgClosure *)bqe),
2716 IF_PAR_DEBUG(bq, fprintf(stderr, ", %p (%s)", bqe, info_type((StgClosure*)bqe)));
2720 #else /* !GRAN && !PAR */
2722 unblockOneLocked(StgTSO *tso)
2726 ASSERT(get_itbl(tso)->type == TSO);
2727 ASSERT(tso->why_blocked != NotBlocked);
2728 tso->why_blocked = NotBlocked;
2730 PUSH_ON_RUN_QUEUE(tso);
2732 IF_DEBUG(scheduler,sched_belch("waking up thread %ld", tso->id));
2737 #if defined(GRAN) || defined(PAR)
2738 inline StgBlockingQueueElement *
2739 unblockOne(StgBlockingQueueElement *bqe, StgClosure *node)
2741 ACQUIRE_LOCK(&sched_mutex);
2742 bqe = unblockOneLocked(bqe, node);
2743 RELEASE_LOCK(&sched_mutex);
2748 unblockOne(StgTSO *tso)
2750 ACQUIRE_LOCK(&sched_mutex);
2751 tso = unblockOneLocked(tso);
2752 RELEASE_LOCK(&sched_mutex);
2759 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2761 StgBlockingQueueElement *bqe;
2766 belch("##-_ AwBQ for node %p on PE %d @ %ld by TSO %d (%p): ", \
2767 node, CurrentProc, CurrentTime[CurrentProc],
2768 CurrentTSO->id, CurrentTSO));
2770 node_loc = where_is(node);
2772 ASSERT(q == END_BQ_QUEUE ||
2773 get_itbl(q)->type == TSO || // q is either a TSO or an RBHSave
2774 get_itbl(q)->type == CONSTR); // closure (type constructor)
2775 ASSERT(is_unique(node));
2777 /* FAKE FETCH: magically copy the node to the tso's proc;
2778 no Fetch necessary because in reality the node should not have been
2779 moved to the other PE in the first place
2781 if (CurrentProc!=node_loc) {
2783 belch("## node %p is on PE %d but CurrentProc is %d (TSO %d); assuming fake fetch and adjusting bitmask (old: %#x)",
2784 node, node_loc, CurrentProc, CurrentTSO->id,
2785 // CurrentTSO, where_is(CurrentTSO),
2786 node->header.gran.procs));
2787 node->header.gran.procs = (node->header.gran.procs) | PE_NUMBER(CurrentProc);
2789 belch("## new bitmask of node %p is %#x",
2790 node, node->header.gran.procs));
2791 if (RtsFlags.GranFlags.GranSimStats.Global) {
2792 globalGranStats.tot_fake_fetches++;
2797 // ToDo: check: ASSERT(CurrentProc==node_loc);
2798 while (get_itbl(bqe)->type==TSO) { // q != END_TSO_QUEUE) {
2801 bqe points to the current element in the queue
2802 next points to the next element in the queue
2804 //tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2805 //tso_loc = where_is(tso);
2807 bqe = unblockOneLocked(bqe, node);
2810 /* if this is the BQ of an RBH, we have to put back the info ripped out of
2811 the closure to make room for the anchor of the BQ */
2812 if (bqe!=END_BQ_QUEUE) {
2813 ASSERT(get_itbl(node)->type == RBH && get_itbl(bqe)->type == CONSTR);
2815 ASSERT((info_ptr==&RBH_Save_0_info) ||
2816 (info_ptr==&RBH_Save_1_info) ||
2817 (info_ptr==&RBH_Save_2_info));
2819 /* cf. convertToRBH in RBH.c for writing the RBHSave closure */
2820 ((StgRBH *)node)->blocking_queue = (StgBlockingQueueElement *)((StgRBHSave *)bqe)->payload[0];
2821 ((StgRBH *)node)->mut_link = (StgMutClosure *)((StgRBHSave *)bqe)->payload[1];
2824 belch("## Filled in RBH_Save for %p (%s) at end of AwBQ",
2825 node, info_type(node)));
2828 /* statistics gathering */
2829 if (RtsFlags.GranFlags.GranSimStats.Global) {
2830 // globalGranStats.tot_bq_processing_time += bq_processing_time;
2831 globalGranStats.tot_bq_len += len; // total length of all bqs awakened
2832 // globalGranStats.tot_bq_len_local += len_local; // same for local TSOs only
2833 globalGranStats.tot_awbq++; // total no. of bqs awakened
2836 fprintf(stderr,"## BQ Stats of %p: [%d entries] %s\n",
2837 node, len, (bqe!=END_BQ_QUEUE) ? "RBH" : ""));
2841 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2843 StgBlockingQueueElement *bqe;
2845 ACQUIRE_LOCK(&sched_mutex);
2847 IF_PAR_DEBUG(verbose,
2848 belch("##-_ AwBQ for node %p on [%x]: ",
2852 if(get_itbl(q)->type == CONSTR || q==END_BQ_QUEUE) {
2853 IF_PAR_DEBUG(verbose, belch("## ... nothing to unblock so lets just return. RFP (BUG?)"));
2858 ASSERT(q == END_BQ_QUEUE ||
2859 get_itbl(q)->type == TSO ||
2860 get_itbl(q)->type == BLOCKED_FETCH ||
2861 get_itbl(q)->type == CONSTR);
2864 while (get_itbl(bqe)->type==TSO ||
2865 get_itbl(bqe)->type==BLOCKED_FETCH) {
2866 bqe = unblockOneLocked(bqe, node);
2868 RELEASE_LOCK(&sched_mutex);
2871 #else /* !GRAN && !PAR */
2873 awakenBlockedQueue(StgTSO *tso)
2875 ACQUIRE_LOCK(&sched_mutex);
2876 while (tso != END_TSO_QUEUE) {
2877 tso = unblockOneLocked(tso);
2879 RELEASE_LOCK(&sched_mutex);
2883 //@node Exception Handling Routines, Debugging Routines, Blocking Queue Routines, Main scheduling code
2884 //@subsection Exception Handling Routines
2886 /* ---------------------------------------------------------------------------
2888 - usually called inside a signal handler so it mustn't do anything fancy.
2889 ------------------------------------------------------------------------ */
2892 interruptStgRts(void)
2898 /* -----------------------------------------------------------------------------
2901 This is for use when we raise an exception in another thread, which
2903 This has nothing to do with the UnblockThread event in GranSim. -- HWL
2904 -------------------------------------------------------------------------- */
2906 #if defined(GRAN) || defined(PAR)
2908 NB: only the type of the blocking queue is different in GranSim and GUM
2909 the operations on the queue-elements are the same
2910 long live polymorphism!
2912 Locks: sched_mutex is held upon entry and exit.
2916 unblockThread(StgTSO *tso)
2918 StgBlockingQueueElement *t, **last;
2920 switch (tso->why_blocked) {
2923 return; /* not blocked */
2926 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
2928 StgBlockingQueueElement *last_tso = END_BQ_QUEUE;
2929 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
2931 last = (StgBlockingQueueElement **)&mvar->head;
2932 for (t = (StgBlockingQueueElement *)mvar->head;
2934 last = &t->link, last_tso = t, t = t->link) {
2935 if (t == (StgBlockingQueueElement *)tso) {
2936 *last = (StgBlockingQueueElement *)tso->link;
2937 if (mvar->tail == tso) {
2938 mvar->tail = (StgTSO *)last_tso;
2943 barf("unblockThread (MVAR): TSO not found");
2946 case BlockedOnBlackHole:
2947 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
2949 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
2951 last = &bq->blocking_queue;
2952 for (t = bq->blocking_queue;
2954 last = &t->link, t = t->link) {
2955 if (t == (StgBlockingQueueElement *)tso) {
2956 *last = (StgBlockingQueueElement *)tso->link;
2960 barf("unblockThread (BLACKHOLE): TSO not found");
2963 case BlockedOnException:
2965 StgTSO *target = tso->block_info.tso;
2967 ASSERT(get_itbl(target)->type == TSO);
2969 if (target->what_next == ThreadRelocated) {
2970 target = target->link;
2971 ASSERT(get_itbl(target)->type == TSO);
2974 ASSERT(target->blocked_exceptions != NULL);
2976 last = (StgBlockingQueueElement **)&target->blocked_exceptions;
2977 for (t = (StgBlockingQueueElement *)target->blocked_exceptions;
2979 last = &t->link, t = t->link) {
2980 ASSERT(get_itbl(t)->type == TSO);
2981 if (t == (StgBlockingQueueElement *)tso) {
2982 *last = (StgBlockingQueueElement *)tso->link;
2986 barf("unblockThread (Exception): TSO not found");
2990 case BlockedOnWrite:
2992 /* take TSO off blocked_queue */
2993 StgBlockingQueueElement *prev = NULL;
2994 for (t = (StgBlockingQueueElement *)blocked_queue_hd; t != END_BQ_QUEUE;
2995 prev = t, t = t->link) {
2996 if (t == (StgBlockingQueueElement *)tso) {
2998 blocked_queue_hd = (StgTSO *)t->link;
2999 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
3000 blocked_queue_tl = END_TSO_QUEUE;
3003 prev->link = t->link;
3004 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
3005 blocked_queue_tl = (StgTSO *)prev;
3011 barf("unblockThread (I/O): TSO not found");
3014 case BlockedOnDelay:
3016 /* take TSO off sleeping_queue */
3017 StgBlockingQueueElement *prev = NULL;
3018 for (t = (StgBlockingQueueElement *)sleeping_queue; t != END_BQ_QUEUE;
3019 prev = t, t = t->link) {
3020 if (t == (StgBlockingQueueElement *)tso) {
3022 sleeping_queue = (StgTSO *)t->link;
3024 prev->link = t->link;
3029 barf("unblockThread (I/O): TSO not found");
3033 barf("unblockThread");
3037 tso->link = END_TSO_QUEUE;
3038 tso->why_blocked = NotBlocked;
3039 tso->block_info.closure = NULL;
3040 PUSH_ON_RUN_QUEUE(tso);
3044 unblockThread(StgTSO *tso)
3048 /* To avoid locking unnecessarily. */
3049 if (tso->why_blocked == NotBlocked) {
3053 switch (tso->why_blocked) {
3056 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
3058 StgTSO *last_tso = END_TSO_QUEUE;
3059 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
3062 for (t = mvar->head; t != END_TSO_QUEUE;
3063 last = &t->link, last_tso = t, t = t->link) {
3066 if (mvar->tail == tso) {
3067 mvar->tail = last_tso;
3072 barf("unblockThread (MVAR): TSO not found");
3075 case BlockedOnBlackHole:
3076 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
3078 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
3080 last = &bq->blocking_queue;
3081 for (t = bq->blocking_queue; t != END_TSO_QUEUE;
3082 last = &t->link, t = t->link) {
3088 barf("unblockThread (BLACKHOLE): TSO not found");
3091 case BlockedOnException:
3093 StgTSO *target = tso->block_info.tso;
3095 ASSERT(get_itbl(target)->type == TSO);
3097 while (target->what_next == ThreadRelocated) {
3098 target = target->link;
3099 ASSERT(get_itbl(target)->type == TSO);
3102 ASSERT(target->blocked_exceptions != NULL);
3104 last = &target->blocked_exceptions;
3105 for (t = target->blocked_exceptions; t != END_TSO_QUEUE;
3106 last = &t->link, t = t->link) {
3107 ASSERT(get_itbl(t)->type == TSO);
3113 barf("unblockThread (Exception): TSO not found");
3117 case BlockedOnWrite:
3119 StgTSO *prev = NULL;
3120 for (t = blocked_queue_hd; t != END_TSO_QUEUE;
3121 prev = t, t = t->link) {
3124 blocked_queue_hd = t->link;
3125 if (blocked_queue_tl == t) {
3126 blocked_queue_tl = END_TSO_QUEUE;
3129 prev->link = t->link;
3130 if (blocked_queue_tl == t) {
3131 blocked_queue_tl = prev;
3137 barf("unblockThread (I/O): TSO not found");
3140 case BlockedOnDelay:
3142 StgTSO *prev = NULL;
3143 for (t = sleeping_queue; t != END_TSO_QUEUE;
3144 prev = t, t = t->link) {
3147 sleeping_queue = t->link;
3149 prev->link = t->link;
3154 barf("unblockThread (I/O): TSO not found");
3158 barf("unblockThread");
3162 tso->link = END_TSO_QUEUE;
3163 tso->why_blocked = NotBlocked;
3164 tso->block_info.closure = NULL;
3165 PUSH_ON_RUN_QUEUE(tso);
3169 /* -----------------------------------------------------------------------------
3172 * The following function implements the magic for raising an
3173 * asynchronous exception in an existing thread.
3175 * We first remove the thread from any queue on which it might be
3176 * blocked. The possible blockages are MVARs and BLACKHOLE_BQs.
3178 * We strip the stack down to the innermost CATCH_FRAME, building
3179 * thunks in the heap for all the active computations, so they can
3180 * be restarted if necessary. When we reach a CATCH_FRAME, we build
3181 * an application of the handler to the exception, and push it on
3182 * the top of the stack.
3184 * How exactly do we save all the active computations? We create an
3185 * AP_UPD for every UpdateFrame on the stack. Entering one of these
3186 * AP_UPDs pushes everything from the corresponding update frame
3187 * upwards onto the stack. (Actually, it pushes everything up to the
3188 * next update frame plus a pointer to the next AP_UPD object.
3189 * Entering the next AP_UPD object pushes more onto the stack until we
3190 * reach the last AP_UPD object - at which point the stack should look
3191 * exactly as it did when we killed the TSO and we can continue
3192 * execution by entering the closure on top of the stack.
3194 * We can also kill a thread entirely - this happens if either (a) the
3195 * exception passed to raiseAsync is NULL, or (b) there's no
3196 * CATCH_FRAME on the stack. In either case, we strip the entire
3197 * stack and replace the thread with a zombie.
3199 * Locks: sched_mutex held upon entry nor exit.
3201 * -------------------------------------------------------------------------- */
3204 deleteThread(StgTSO *tso)
3206 raiseAsync(tso,NULL);
3210 raiseAsyncWithLock(StgTSO *tso, StgClosure *exception)
3212 /* When raising async exs from contexts where sched_mutex isn't held;
3213 use raiseAsyncWithLock(). */
3214 ACQUIRE_LOCK(&sched_mutex);
3215 raiseAsync(tso,exception);
3216 RELEASE_LOCK(&sched_mutex);
3220 raiseAsync(StgTSO *tso, StgClosure *exception)
3222 StgUpdateFrame* su = tso->su;
3223 StgPtr sp = tso->sp;
3225 /* Thread already dead? */
3226 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
3230 IF_DEBUG(scheduler, sched_belch("raising exception in thread %ld.", tso->id));
3232 /* Remove it from any blocking queues */
3235 IF_DEBUG(scheduler, sched_belch("raising exception in thread %ld.", tso->id));
3236 /* The stack freezing code assumes there's a closure pointer on
3237 * the top of the stack. This isn't always the case with compiled
3238 * code, so we have to push a dummy closure on the top which just
3239 * returns to the next return address on the stack.
3241 if ( LOOKS_LIKE_GHC_INFO((void*)*sp) ) {
3242 *(--sp) = (W_)&stg_dummy_ret_closure;
3246 nat words = ((P_)su - (P_)sp) - 1;
3250 /* If we find a CATCH_FRAME, and we've got an exception to raise,
3251 * then build the THUNK raise(exception), and leave it on
3252 * top of the CATCH_FRAME ready to enter.
3254 if (get_itbl(su)->type == CATCH_FRAME && exception != NULL) {
3256 StgCatchFrame *cf = (StgCatchFrame *)su;
3260 /* we've got an exception to raise, so let's pass it to the
3261 * handler in this frame.
3263 raise = (StgClosure *)allocate(sizeofW(StgClosure)+1);
3264 TICK_ALLOC_SE_THK(1,0);
3265 SET_HDR(raise,&stg_raise_info,cf->header.prof.ccs);
3266 raise->payload[0] = exception;
3268 /* throw away the stack from Sp up to the CATCH_FRAME.
3272 /* Ensure that async excpetions are blocked now, so we don't get
3273 * a surprise exception before we get around to executing the
3276 if (tso->blocked_exceptions == NULL) {
3277 tso->blocked_exceptions = END_TSO_QUEUE;
3280 /* Put the newly-built THUNK on top of the stack, ready to execute
3281 * when the thread restarts.
3286 tso->what_next = ThreadEnterGHC;
3287 IF_DEBUG(sanity, checkTSO(tso));
3291 /* First build an AP_UPD consisting of the stack chunk above the
3292 * current update frame, with the top word on the stack as the
3295 ap = (StgAP_UPD *)allocate(AP_sizeW(words));
3300 ap->fun = (StgClosure *)sp[0];
3302 for(i=0; i < (nat)words; ++i) {
3303 ap->payload[i] = (StgClosure *)*sp++;
3306 switch (get_itbl(su)->type) {
3310 SET_HDR(ap,&stg_AP_UPD_info,su->header.prof.ccs /* ToDo */);
3311 TICK_ALLOC_UP_THK(words+1,0);
3314 fprintf(stderr, "scheduler: Updating ");
3315 printPtr((P_)su->updatee);
3316 fprintf(stderr, " with ");
3317 printObj((StgClosure *)ap);
3320 /* Replace the updatee with an indirection - happily
3321 * this will also wake up any threads currently
3322 * waiting on the result.
3324 * Warning: if we're in a loop, more than one update frame on
3325 * the stack may point to the same object. Be careful not to
3326 * overwrite an IND_OLDGEN in this case, because we'll screw
3327 * up the mutable lists. To be on the safe side, don't
3328 * overwrite any kind of indirection at all. See also
3329 * threadSqueezeStack in GC.c, where we have to make a similar
3332 if (!closure_IND(su->updatee)) {
3333 UPD_IND_NOLOCK(su->updatee,ap); /* revert the black hole */
3336 sp += sizeofW(StgUpdateFrame) -1;
3337 sp[0] = (W_)ap; /* push onto stack */
3343 StgCatchFrame *cf = (StgCatchFrame *)su;
3346 /* We want a PAP, not an AP_UPD. Fortunately, the
3347 * layout's the same.
3349 SET_HDR(ap,&stg_PAP_info,su->header.prof.ccs /* ToDo */);
3350 TICK_ALLOC_UPD_PAP(words+1,0);
3352 /* now build o = FUN(catch,ap,handler) */
3353 o = (StgClosure *)allocate(sizeofW(StgClosure)+2);
3354 TICK_ALLOC_FUN(2,0);
3355 SET_HDR(o,&stg_catch_info,su->header.prof.ccs /* ToDo */);
3356 o->payload[0] = (StgClosure *)ap;
3357 o->payload[1] = cf->handler;
3360 fprintf(stderr, "scheduler: Built ");
3361 printObj((StgClosure *)o);
3364 /* pop the old handler and put o on the stack */
3366 sp += sizeofW(StgCatchFrame) - 1;
3373 StgSeqFrame *sf = (StgSeqFrame *)su;
3376 SET_HDR(ap,&stg_PAP_info,su->header.prof.ccs /* ToDo */);
3377 TICK_ALLOC_UPD_PAP(words+1,0);
3379 /* now build o = FUN(seq,ap) */
3380 o = (StgClosure *)allocate(sizeofW(StgClosure)+1);
3381 TICK_ALLOC_SE_THK(1,0);
3382 SET_HDR(o,&stg_seq_info,su->header.prof.ccs /* ToDo */);
3383 o->payload[0] = (StgClosure *)ap;
3386 fprintf(stderr, "scheduler: Built ");
3387 printObj((StgClosure *)o);
3390 /* pop the old handler and put o on the stack */
3392 sp += sizeofW(StgSeqFrame) - 1;
3398 /* We've stripped the entire stack, the thread is now dead. */
3399 sp += sizeofW(StgStopFrame) - 1;
3400 sp[0] = (W_)exception; /* save the exception */
3401 tso->what_next = ThreadKilled;
3402 tso->su = (StgUpdateFrame *)(sp+1);
3413 /* -----------------------------------------------------------------------------
3414 resurrectThreads is called after garbage collection on the list of
3415 threads found to be garbage. Each of these threads will be woken
3416 up and sent a signal: BlockedOnDeadMVar if the thread was blocked
3417 on an MVar, or NonTermination if the thread was blocked on a Black
3420 Locks: sched_mutex isn't held upon entry nor exit.
3421 -------------------------------------------------------------------------- */
3424 resurrectThreads( StgTSO *threads )
3428 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
3429 next = tso->global_link;
3430 tso->global_link = all_threads;
3432 IF_DEBUG(scheduler, sched_belch("resurrecting thread %d", tso->id));
3434 switch (tso->why_blocked) {
3436 case BlockedOnException:
3437 /* Called by GC - sched_mutex lock is currently held. */
3438 raiseAsync(tso,(StgClosure *)BlockedOnDeadMVar_closure);
3440 case BlockedOnBlackHole:
3441 raiseAsync(tso,(StgClosure *)NonTermination_closure);
3444 /* This might happen if the thread was blocked on a black hole
3445 * belonging to a thread that we've just woken up (raiseAsync
3446 * can wake up threads, remember...).
3450 barf("resurrectThreads: thread blocked in a strange way");
3455 /* -----------------------------------------------------------------------------
3456 * Blackhole detection: if we reach a deadlock, test whether any
3457 * threads are blocked on themselves. Any threads which are found to
3458 * be self-blocked get sent a NonTermination exception.
3460 * This is only done in a deadlock situation in order to avoid
3461 * performance overhead in the normal case.
3463 * Locks: sched_mutex is held upon entry and exit.
3464 * -------------------------------------------------------------------------- */
3467 detectBlackHoles( void )
3469 StgTSO *t = all_threads;
3470 StgUpdateFrame *frame;
3471 StgClosure *blocked_on;
3473 for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
3475 while (t->what_next == ThreadRelocated) {
3477 ASSERT(get_itbl(t)->type == TSO);
3480 if (t->why_blocked != BlockedOnBlackHole) {
3484 blocked_on = t->block_info.closure;
3486 for (frame = t->su; ; frame = frame->link) {
3487 switch (get_itbl(frame)->type) {
3490 if (frame->updatee == blocked_on) {
3491 /* We are blocking on one of our own computations, so
3492 * send this thread the NonTermination exception.
3495 sched_belch("thread %d is blocked on itself", t->id));
3496 raiseAsync(t, (StgClosure *)NonTermination_closure);
3517 //@node Debugging Routines, Index, Exception Handling Routines, Main scheduling code
3518 //@subsection Debugging Routines
3520 /* -----------------------------------------------------------------------------
3521 Debugging: why is a thread blocked
3522 -------------------------------------------------------------------------- */
3527 printThreadBlockage(StgTSO *tso)
3529 switch (tso->why_blocked) {
3531 fprintf(stderr,"is blocked on read from fd %d", tso->block_info.fd);
3533 case BlockedOnWrite:
3534 fprintf(stderr,"is blocked on write to fd %d", tso->block_info.fd);
3536 case BlockedOnDelay:
3537 fprintf(stderr,"is blocked until %d", tso->block_info.target);
3540 fprintf(stderr,"is blocked on an MVar");
3542 case BlockedOnException:
3543 fprintf(stderr,"is blocked on delivering an exception to thread %d",
3544 tso->block_info.tso->id);
3546 case BlockedOnBlackHole:
3547 fprintf(stderr,"is blocked on a black hole");
3550 fprintf(stderr,"is not blocked");
3554 fprintf(stderr,"is blocked on global address; local FM_BQ is %p (%s)",
3555 tso->block_info.closure, info_type(tso->block_info.closure));
3557 case BlockedOnGA_NoSend:
3558 fprintf(stderr,"is blocked on global address (no send); local FM_BQ is %p (%s)",
3559 tso->block_info.closure, info_type(tso->block_info.closure));
3562 #if defined(RTS_SUPPORTS_THREADS)
3563 case BlockedOnCCall:
3564 fprintf(stderr,"is blocked on an external call");
3568 barf("printThreadBlockage: strange tso->why_blocked: %d for TSO %d (%d)",
3569 tso->why_blocked, tso->id, tso);
3574 printThreadStatus(StgTSO *tso)
3576 switch (tso->what_next) {
3578 fprintf(stderr,"has been killed");
3580 case ThreadComplete:
3581 fprintf(stderr,"has completed");
3584 printThreadBlockage(tso);
3589 printAllThreads(void)
3595 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
3596 ullong_format_string(TIME_ON_PROC(CurrentProc),
3597 time_string, rtsFalse/*no commas!*/);
3599 sched_belch("all threads at [%s]:", time_string);
3601 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
3602 ullong_format_string(CURRENT_TIME,
3603 time_string, rtsFalse/*no commas!*/);
3605 sched_belch("all threads at [%s]:", time_string);
3607 sched_belch("all threads:");
3610 for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
3611 fprintf(stderr, "\tthread %d @ %p ", t->id, (void *)t);
3612 label = lookupThreadLabel((StgWord)t);
3613 if (label) fprintf(stderr,"[\"%s\"] ",(char *)label);
3614 printThreadStatus(t);
3615 fprintf(stderr,"\n");
3620 Print a whole blocking queue attached to node (debugging only).
3625 print_bq (StgClosure *node)
3627 StgBlockingQueueElement *bqe;
3631 fprintf(stderr,"## BQ of closure %p (%s): ",
3632 node, info_type(node));
3634 /* should cover all closures that may have a blocking queue */
3635 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
3636 get_itbl(node)->type == FETCH_ME_BQ ||
3637 get_itbl(node)->type == RBH ||
3638 get_itbl(node)->type == MVAR);
3640 ASSERT(node!=(StgClosure*)NULL); // sanity check
3642 print_bqe(((StgBlockingQueue*)node)->blocking_queue);
3646 Print a whole blocking queue starting with the element bqe.
3649 print_bqe (StgBlockingQueueElement *bqe)
3654 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
3656 for (end = (bqe==END_BQ_QUEUE);
3657 !end; // iterate until bqe points to a CONSTR
3658 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE),
3659 bqe = end ? END_BQ_QUEUE : bqe->link) {
3660 ASSERT(bqe != END_BQ_QUEUE); // sanity check
3661 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
3662 /* types of closures that may appear in a blocking queue */
3663 ASSERT(get_itbl(bqe)->type == TSO ||
3664 get_itbl(bqe)->type == BLOCKED_FETCH ||
3665 get_itbl(bqe)->type == CONSTR);
3666 /* only BQs of an RBH end with an RBH_Save closure */
3667 //ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
3669 switch (get_itbl(bqe)->type) {
3671 fprintf(stderr," TSO %u (%x),",
3672 ((StgTSO *)bqe)->id, ((StgTSO *)bqe));
3675 fprintf(stderr," BF (node=%p, ga=((%x, %d, %x)),",
3676 ((StgBlockedFetch *)bqe)->node,
3677 ((StgBlockedFetch *)bqe)->ga.payload.gc.gtid,
3678 ((StgBlockedFetch *)bqe)->ga.payload.gc.slot,
3679 ((StgBlockedFetch *)bqe)->ga.weight);
3682 fprintf(stderr," %s (IP %p),",
3683 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
3684 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
3685 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
3686 "RBH_Save_?"), get_itbl(bqe));
3689 barf("Unexpected closure type %s in blocking queue", // of %p (%s)",
3690 info_type((StgClosure *)bqe)); // , node, info_type(node));
3694 fputc('\n', stderr);
3696 # elif defined(GRAN)
3698 print_bq (StgClosure *node)
3700 StgBlockingQueueElement *bqe;
3701 PEs node_loc, tso_loc;
3704 /* should cover all closures that may have a blocking queue */
3705 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
3706 get_itbl(node)->type == FETCH_ME_BQ ||
3707 get_itbl(node)->type == RBH);
3709 ASSERT(node!=(StgClosure*)NULL); // sanity check
3710 node_loc = where_is(node);
3712 fprintf(stderr,"## BQ of closure %p (%s) on [PE %d]: ",
3713 node, info_type(node), node_loc);
3716 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
3718 for (bqe = ((StgBlockingQueue*)node)->blocking_queue, end = (bqe==END_BQ_QUEUE);
3719 !end; // iterate until bqe points to a CONSTR
3720 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) {
3721 ASSERT(bqe != END_BQ_QUEUE); // sanity check
3722 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
3723 /* types of closures that may appear in a blocking queue */
3724 ASSERT(get_itbl(bqe)->type == TSO ||
3725 get_itbl(bqe)->type == CONSTR);
3726 /* only BQs of an RBH end with an RBH_Save closure */
3727 ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
3729 tso_loc = where_is((StgClosure *)bqe);
3730 switch (get_itbl(bqe)->type) {
3732 fprintf(stderr," TSO %d (%p) on [PE %d],",
3733 ((StgTSO *)bqe)->id, (StgTSO *)bqe, tso_loc);
3736 fprintf(stderr," %s (IP %p),",
3737 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
3738 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
3739 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
3740 "RBH_Save_?"), get_itbl(bqe));
3743 barf("Unexpected closure type %s in blocking queue of %p (%s)",
3744 info_type((StgClosure *)bqe), node, info_type(node));
3748 fputc('\n', stderr);
3752 Nice and easy: only TSOs on the blocking queue
3755 print_bq (StgClosure *node)
3759 ASSERT(node!=(StgClosure*)NULL); // sanity check
3760 for (tso = ((StgBlockingQueue*)node)->blocking_queue;
3761 tso != END_TSO_QUEUE;
3763 ASSERT(tso!=NULL && tso!=END_TSO_QUEUE); // sanity check
3764 ASSERT(get_itbl(tso)->type == TSO); // guess what, sanity check
3765 fprintf(stderr," TSO %d (%p),", tso->id, tso);
3767 fputc('\n', stderr);
3778 for (i=0, tso=run_queue_hd;
3779 tso != END_TSO_QUEUE;
3788 sched_belch(char *s, ...)
3793 fprintf(stderr, "scheduler (task %ld): ", osThreadId());
3795 fprintf(stderr, "== ");
3797 fprintf(stderr, "scheduler: ");
3799 vfprintf(stderr, s, ap);
3800 fprintf(stderr, "\n");
3807 //@node Index, , Debugging Routines, Main scheduling code
3811 //* StgMainThread:: @cindex\s-+StgMainThread
3812 //* awaken_blocked_queue:: @cindex\s-+awaken_blocked_queue
3813 //* blocked_queue_hd:: @cindex\s-+blocked_queue_hd
3814 //* blocked_queue_tl:: @cindex\s-+blocked_queue_tl
3815 //* context_switch:: @cindex\s-+context_switch
3816 //* createThread:: @cindex\s-+createThread
3817 //* gc_pending_cond:: @cindex\s-+gc_pending_cond
3818 //* initScheduler:: @cindex\s-+initScheduler
3819 //* interrupted:: @cindex\s-+interrupted
3820 //* next_thread_id:: @cindex\s-+next_thread_id
3821 //* print_bq:: @cindex\s-+print_bq
3822 //* run_queue_hd:: @cindex\s-+run_queue_hd
3823 //* run_queue_tl:: @cindex\s-+run_queue_tl
3824 //* sched_mutex:: @cindex\s-+sched_mutex
3825 //* schedule:: @cindex\s-+schedule
3826 //* take_off_run_queue:: @cindex\s-+take_off_run_queue
3827 //* term_mutex:: @cindex\s-+term_mutex