1 /* ---------------------------------------------------------------------------
2 * $Id: Schedule.c,v 1.149 2002/07/19 00:06:05 sof 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((StgWord)m->tso);
461 if (m->ret) *(m->ret) = NULL;
463 if (was_interrupted) {
464 m->stat = Interrupted;
468 broadcastCondition(&m->wakeup);
470 removeThreadLabel((StgWord)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 rtsBool killerIsMainThread = rtsFalse;
1449 IF_DEBUG(scheduler,sched_belch("forking!"));
1452 if (pid) { /* parent */
1454 /* just return the pid */
1456 } else { /* child */
1457 /* wipe all other threads */
1458 run_queue_hd = run_queue_tl = tso;
1459 tso->link = END_TSO_QUEUE;
1461 /* When clearing out the threads, we need to ensure
1462 that a 'main thread' is left behind.
1463 careful about leaving a main thread behind.
1465 ==> if the killing thread isn't a main thread, we
1468 for (m = main_threads; m != NULL; m = m->link) {
1469 if (m->tso->id == tso->id) {
1470 killerIsMainThread=rtsTrue;
1475 /* DO NOT TOUCH THE QUEUES directly because most of the code around
1476 us is picky about finding the thread still in its queue when
1477 handling the deleteThread() */
1479 if (!killerIsMainThread) {
1480 /* Add it to main_threads */
1481 m = stgMallocBytes(sizeof(StgMainThread), "forkProcess");
1484 m->ret = NULL; /* can't really do better */
1486 #if defined(RTS_SUPPORTS_THREADS)
1487 initCondition(&m->wakeup);
1489 /* Hook it up to the main_threads list. */
1490 m->link = main_threads;
1493 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
1496 /* Don't kill current thread */
1497 if (t->id == tso->id) continue;
1498 if (!killerIsMainThread) {
1500 /* Signal the abrupt completion of a now-killed main thread. */
1501 for (m = main_threads; m != NULL; m = m->link) {
1502 if (m->tso->id == t->id) {
1504 if (m->ret) { *(m->ret) = NULL; }
1505 #if defined(RTS_SUPPORTS_THREADS)
1506 broadcastCondition(&m->wakeup);
1509 removeThreadLabel((StgWord)m->tso);
1515 /* ToDo..?: kill other entries along main_threads except the
1516 * killing (main) thread.
1522 barf("forkProcess#: primop not implemented for mingw32, sorry! (%u)\n", tso->id);
1523 /* pointlessly printing out the TSOs 'id' to avoid CC unused warning. */
1525 #endif /* mingw32 */
1528 /* ---------------------------------------------------------------------------
1529 * deleteAllThreads(): kill all the live threads.
1531 * This is used when we catch a user interrupt (^C), before performing
1532 * any necessary cleanups and running finalizers.
1534 * Locks: sched_mutex held.
1535 * ------------------------------------------------------------------------- */
1537 void deleteAllThreads ( void )
1540 IF_DEBUG(scheduler,sched_belch("deleting all threads"));
1541 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
1542 next = t->global_link;
1545 run_queue_hd = run_queue_tl = END_TSO_QUEUE;
1546 blocked_queue_hd = blocked_queue_tl = END_TSO_QUEUE;
1547 sleeping_queue = END_TSO_QUEUE;
1550 /* startThread and insertThread are now in GranSim.c -- HWL */
1553 //@node Suspend and Resume, Run queue code, Main scheduling loop, Main scheduling code
1554 //@subsection Suspend and Resume
1556 /* ---------------------------------------------------------------------------
1557 * Suspending & resuming Haskell threads.
1559 * When making a "safe" call to C (aka _ccall_GC), the task gives back
1560 * its capability before calling the C function. This allows another
1561 * task to pick up the capability and carry on running Haskell
1562 * threads. It also means that if the C call blocks, it won't lock
1565 * The Haskell thread making the C call is put to sleep for the
1566 * duration of the call, on the susepended_ccalling_threads queue. We
1567 * give out a token to the task, which it can use to resume the thread
1568 * on return from the C function.
1569 * ------------------------------------------------------------------------- */
1572 suspendThread( StgRegTable *reg,
1574 #if !defined(RTS_SUPPORTS_THREADS) && !defined(DEBUG)
1582 /* assume that *reg is a pointer to the StgRegTable part
1585 cap = (Capability *)((void *)reg - sizeof(StgFunTable));
1587 ACQUIRE_LOCK(&sched_mutex);
1590 sched_belch("thread %d did a _ccall_gc (is_concurrent: %d)", cap->r.rCurrentTSO->id,concCall));
1592 threadPaused(cap->r.rCurrentTSO);
1593 cap->r.rCurrentTSO->link = suspended_ccalling_threads;
1594 suspended_ccalling_threads = cap->r.rCurrentTSO;
1596 #if defined(RTS_SUPPORTS_THREADS)
1597 cap->r.rCurrentTSO->why_blocked = BlockedOnCCall;
1600 /* Use the thread ID as the token; it should be unique */
1601 tok = cap->r.rCurrentTSO->id;
1603 /* Hand back capability */
1604 releaseCapability(cap);
1606 #if defined(RTS_SUPPORTS_THREADS)
1607 /* Preparing to leave the RTS, so ensure there's a native thread/task
1608 waiting to take over.
1610 ToDo: optimise this and only create a new task if there's a need
1611 for one (i.e., if there's only one Concurrent Haskell thread alive,
1612 there's no need to create a new task).
1614 IF_DEBUG(scheduler, sched_belch("worker thread (%d): leaving RTS", tok));
1616 startTask(taskStart);
1620 /* Other threads _might_ be available for execution; signal this */
1622 RELEASE_LOCK(&sched_mutex);
1627 resumeThread( StgInt tok,
1629 #if !defined(RTS_SUPPORTS_THREADS)
1634 StgTSO *tso, **prev;
1637 #if defined(RTS_SUPPORTS_THREADS)
1638 /* Wait for permission to re-enter the RTS with the result. */
1640 ACQUIRE_LOCK(&sched_mutex);
1641 grabReturnCapability(&sched_mutex, &cap);
1643 grabCapability(&cap);
1646 grabCapability(&cap);
1649 /* Remove the thread off of the suspended list */
1650 prev = &suspended_ccalling_threads;
1651 for (tso = suspended_ccalling_threads;
1652 tso != END_TSO_QUEUE;
1653 prev = &tso->link, tso = tso->link) {
1654 if (tso->id == (StgThreadID)tok) {
1659 if (tso == END_TSO_QUEUE) {
1660 barf("resumeThread: thread not found");
1662 tso->link = END_TSO_QUEUE;
1663 /* Reset blocking status */
1664 tso->why_blocked = NotBlocked;
1666 cap->r.rCurrentTSO = tso;
1667 RELEASE_LOCK(&sched_mutex);
1672 /* ---------------------------------------------------------------------------
1674 * ------------------------------------------------------------------------ */
1675 static void unblockThread(StgTSO *tso);
1677 /* ---------------------------------------------------------------------------
1678 * Comparing Thread ids.
1680 * This is used from STG land in the implementation of the
1681 * instances of Eq/Ord for ThreadIds.
1682 * ------------------------------------------------------------------------ */
1684 int cmp_thread(const StgTSO *tso1, const StgTSO *tso2)
1686 StgThreadID id1 = tso1->id;
1687 StgThreadID id2 = tso2->id;
1689 if (id1 < id2) return (-1);
1690 if (id1 > id2) return 1;
1694 /* ---------------------------------------------------------------------------
1695 * Fetching the ThreadID from an StgTSO.
1697 * This is used in the implementation of Show for ThreadIds.
1698 * ------------------------------------------------------------------------ */
1699 int rts_getThreadId(const StgTSO *tso)
1705 void labelThread(StgTSO *tso, char *label)
1710 /* Caveat: Once set, you can only set the thread name to "" */
1711 len = strlen(label)+1;
1714 fprintf(stderr,"insufficient memory for labelThread!\n");
1716 strncpy(buf,label,len);
1717 /* Update will free the old memory for us */
1718 updateThreadLabel((StgWord)tso,buf);
1722 /* ---------------------------------------------------------------------------
1723 Create a new thread.
1725 The new thread starts with the given stack size. Before the
1726 scheduler can run, however, this thread needs to have a closure
1727 (and possibly some arguments) pushed on its stack. See
1728 pushClosure() in Schedule.h.
1730 createGenThread() and createIOThread() (in SchedAPI.h) are
1731 convenient packaged versions of this function.
1733 currently pri (priority) is only used in a GRAN setup -- HWL
1734 ------------------------------------------------------------------------ */
1735 //@cindex createThread
1737 /* currently pri (priority) is only used in a GRAN setup -- HWL */
1739 createThread(nat size, StgInt pri)
1742 createThread(nat size)
1749 /* First check whether we should create a thread at all */
1751 /* check that no more than RtsFlags.ParFlags.maxThreads threads are created */
1752 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads) {
1754 belch("{createThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
1755 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
1756 return END_TSO_QUEUE;
1762 ASSERT(!RtsFlags.GranFlags.Light || CurrentProc==0);
1765 // ToDo: check whether size = stack_size - TSO_STRUCT_SIZEW
1767 /* catch ridiculously small stack sizes */
1768 if (size < MIN_STACK_WORDS + TSO_STRUCT_SIZEW) {
1769 size = MIN_STACK_WORDS + TSO_STRUCT_SIZEW;
1772 stack_size = size - TSO_STRUCT_SIZEW;
1774 tso = (StgTSO *)allocate(size);
1775 TICK_ALLOC_TSO(stack_size, 0);
1777 SET_HDR(tso, &stg_TSO_info, CCS_SYSTEM);
1779 SET_GRAN_HDR(tso, ThisPE);
1781 tso->what_next = ThreadEnterGHC;
1783 /* tso->id needs to be unique. For now we use a heavyweight mutex to
1784 * protect the increment operation on next_thread_id.
1785 * In future, we could use an atomic increment instead.
1787 ACQUIRE_LOCK(&thread_id_mutex);
1788 tso->id = next_thread_id++;
1789 RELEASE_LOCK(&thread_id_mutex);
1791 tso->why_blocked = NotBlocked;
1792 tso->blocked_exceptions = NULL;
1794 tso->stack_size = stack_size;
1795 tso->max_stack_size = round_to_mblocks(RtsFlags.GcFlags.maxStkSize)
1797 tso->sp = (P_)&(tso->stack) + stack_size;
1800 tso->prof.CCCS = CCS_MAIN;
1803 /* put a stop frame on the stack */
1804 tso->sp -= sizeofW(StgStopFrame);
1805 SET_HDR((StgClosure*)tso->sp,(StgInfoTable *)&stg_stop_thread_info,CCS_SYSTEM);
1806 tso->su = (StgUpdateFrame*)tso->sp;
1810 tso->link = END_TSO_QUEUE;
1811 /* uses more flexible routine in GranSim */
1812 insertThread(tso, CurrentProc);
1814 /* In a non-GranSim setup the pushing of a TSO onto the runq is separated
1820 if (RtsFlags.GranFlags.GranSimStats.Full)
1821 DumpGranEvent(GR_START,tso);
1823 if (RtsFlags.ParFlags.ParStats.Full)
1824 DumpGranEvent(GR_STARTQ,tso);
1825 /* HACk to avoid SCHEDULE
1829 /* Link the new thread on the global thread list.
1831 tso->global_link = all_threads;
1835 tso->dist.priority = MandatoryPriority; //by default that is...
1839 tso->gran.pri = pri;
1841 tso->gran.magic = TSO_MAGIC; // debugging only
1843 tso->gran.sparkname = 0;
1844 tso->gran.startedat = CURRENT_TIME;
1845 tso->gran.exported = 0;
1846 tso->gran.basicblocks = 0;
1847 tso->gran.allocs = 0;
1848 tso->gran.exectime = 0;
1849 tso->gran.fetchtime = 0;
1850 tso->gran.fetchcount = 0;
1851 tso->gran.blocktime = 0;
1852 tso->gran.blockcount = 0;
1853 tso->gran.blockedat = 0;
1854 tso->gran.globalsparks = 0;
1855 tso->gran.localsparks = 0;
1856 if (RtsFlags.GranFlags.Light)
1857 tso->gran.clock = Now; /* local clock */
1859 tso->gran.clock = 0;
1861 IF_DEBUG(gran,printTSO(tso));
1864 tso->par.magic = TSO_MAGIC; // debugging only
1866 tso->par.sparkname = 0;
1867 tso->par.startedat = CURRENT_TIME;
1868 tso->par.exported = 0;
1869 tso->par.basicblocks = 0;
1870 tso->par.allocs = 0;
1871 tso->par.exectime = 0;
1872 tso->par.fetchtime = 0;
1873 tso->par.fetchcount = 0;
1874 tso->par.blocktime = 0;
1875 tso->par.blockcount = 0;
1876 tso->par.blockedat = 0;
1877 tso->par.globalsparks = 0;
1878 tso->par.localsparks = 0;
1882 globalGranStats.tot_threads_created++;
1883 globalGranStats.threads_created_on_PE[CurrentProc]++;
1884 globalGranStats.tot_sq_len += spark_queue_len(CurrentProc);
1885 globalGranStats.tot_sq_probes++;
1887 // collect parallel global statistics (currently done together with GC stats)
1888 if (RtsFlags.ParFlags.ParStats.Global &&
1889 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
1890 //fprintf(stderr, "Creating thread %d @ %11.2f\n", tso->id, usertime());
1891 globalParStats.tot_threads_created++;
1897 belch("==__ schedule: Created TSO %d (%p);",
1898 CurrentProc, tso, tso->id));
1900 IF_PAR_DEBUG(verbose,
1901 belch("==__ schedule: Created TSO %d (%p); %d threads active",
1902 tso->id, tso, advisory_thread_count));
1904 IF_DEBUG(scheduler,sched_belch("created thread %ld, stack size = %lx words",
1905 tso->id, tso->stack_size));
1912 all parallel thread creation calls should fall through the following routine.
1915 createSparkThread(rtsSpark spark)
1917 ASSERT(spark != (rtsSpark)NULL);
1918 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads)
1920 barf("{createSparkThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
1921 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
1922 return END_TSO_QUEUE;
1926 tso = createThread(RtsFlags.GcFlags.initialStkSize);
1927 if (tso==END_TSO_QUEUE)
1928 barf("createSparkThread: Cannot create TSO");
1930 tso->priority = AdvisoryPriority;
1932 pushClosure(tso,spark);
1933 PUSH_ON_RUN_QUEUE(tso);
1934 advisory_thread_count++;
1941 Turn a spark into a thread.
1942 ToDo: fix for SMP (needs to acquire SCHED_MUTEX!)
1945 //@cindex activateSpark
1947 activateSpark (rtsSpark spark)
1951 tso = createSparkThread(spark);
1952 if (RtsFlags.ParFlags.ParStats.Full) {
1953 //ASSERT(run_queue_hd == END_TSO_QUEUE); // I think ...
1954 IF_PAR_DEBUG(verbose,
1955 belch("==^^ activateSpark: turning spark of closure %p (%s) into a thread",
1956 (StgClosure *)spark, info_type((StgClosure *)spark)));
1958 // ToDo: fwd info on local/global spark to thread -- HWL
1959 // tso->gran.exported = spark->exported;
1960 // tso->gran.locked = !spark->global;
1961 // tso->gran.sparkname = spark->name;
1967 static SchedulerStatus waitThread_(/*out*/StgMainThread* m
1968 #if defined(THREADED_RTS)
1969 , rtsBool blockWaiting
1974 /* ---------------------------------------------------------------------------
1977 * scheduleThread puts a thread on the head of the runnable queue.
1978 * This will usually be done immediately after a thread is created.
1979 * The caller of scheduleThread must create the thread using e.g.
1980 * createThread and push an appropriate closure
1981 * on this thread's stack before the scheduler is invoked.
1982 * ------------------------------------------------------------------------ */
1984 static void scheduleThread_ (StgTSO* tso, rtsBool createTask);
1987 scheduleThread_(StgTSO *tso
1988 , rtsBool createTask
1989 #if !defined(THREADED_RTS)
1994 ACQUIRE_LOCK(&sched_mutex);
1996 /* Put the new thread on the head of the runnable queue. The caller
1997 * better push an appropriate closure on this thread's stack
1998 * beforehand. In the SMP case, the thread may start running as
1999 * soon as we release the scheduler lock below.
2001 PUSH_ON_RUN_QUEUE(tso);
2002 #if defined(THREADED_RTS)
2003 /* If main() is scheduling a thread, don't bother creating a
2007 startTask(taskStart);
2013 IF_DEBUG(scheduler,printTSO(tso));
2015 RELEASE_LOCK(&sched_mutex);
2018 void scheduleThread(StgTSO* tso)
2020 scheduleThread_(tso, rtsFalse);
2024 scheduleWaitThread(StgTSO* tso, /*[out]*/HaskellObj* ret)
2028 m = stgMallocBytes(sizeof(StgMainThread), "waitThread");
2032 #if defined(RTS_SUPPORTS_THREADS)
2033 initCondition(&m->wakeup);
2036 /* Put the thread on the main-threads list prior to scheduling the TSO.
2037 Failure to do so introduces a race condition in the MT case (as
2038 identified by Wolfgang Thaller), whereby the new task/OS thread
2039 created by scheduleThread_() would complete prior to the thread
2040 that spawned it managed to put 'itself' on the main-threads list.
2041 The upshot of it all being that the worker thread wouldn't get to
2042 signal the completion of the its work item for the main thread to
2043 see (==> it got stuck waiting.) -- sof 6/02.
2045 ACQUIRE_LOCK(&sched_mutex);
2046 IF_DEBUG(scheduler, sched_belch("== scheduler: waiting for thread (%d)\n", tso->id));
2048 m->link = main_threads;
2051 /* Inefficient (scheduleThread_() acquires it again right away),
2052 * but obviously correct.
2054 RELEASE_LOCK(&sched_mutex);
2056 scheduleThread_(tso, rtsTrue);
2057 #if defined(THREADED_RTS)
2058 return waitThread_(m, rtsTrue);
2060 return waitThread_(m);
2064 /* ---------------------------------------------------------------------------
2067 * Initialise the scheduler. This resets all the queues - if the
2068 * queues contained any threads, they'll be garbage collected at the
2071 * ------------------------------------------------------------------------ */
2075 term_handler(int sig STG_UNUSED)
2078 ACQUIRE_LOCK(&term_mutex);
2080 RELEASE_LOCK(&term_mutex);
2091 for (i=0; i<=MAX_PROC; i++) {
2092 run_queue_hds[i] = END_TSO_QUEUE;
2093 run_queue_tls[i] = END_TSO_QUEUE;
2094 blocked_queue_hds[i] = END_TSO_QUEUE;
2095 blocked_queue_tls[i] = END_TSO_QUEUE;
2096 ccalling_threadss[i] = END_TSO_QUEUE;
2097 sleeping_queue = END_TSO_QUEUE;
2100 run_queue_hd = END_TSO_QUEUE;
2101 run_queue_tl = END_TSO_QUEUE;
2102 blocked_queue_hd = END_TSO_QUEUE;
2103 blocked_queue_tl = END_TSO_QUEUE;
2104 sleeping_queue = END_TSO_QUEUE;
2107 suspended_ccalling_threads = END_TSO_QUEUE;
2109 main_threads = NULL;
2110 all_threads = END_TSO_QUEUE;
2115 RtsFlags.ConcFlags.ctxtSwitchTicks =
2116 RtsFlags.ConcFlags.ctxtSwitchTime / TICK_MILLISECS;
2118 #if defined(RTS_SUPPORTS_THREADS)
2119 /* Initialise the mutex and condition variables used by
2121 initMutex(&sched_mutex);
2122 initMutex(&term_mutex);
2123 initMutex(&thread_id_mutex);
2125 initCondition(&thread_ready_cond);
2129 initCondition(&gc_pending_cond);
2132 #if defined(RTS_SUPPORTS_THREADS)
2133 ACQUIRE_LOCK(&sched_mutex);
2136 /* Install the SIGHUP handler */
2139 struct sigaction action,oact;
2141 action.sa_handler = term_handler;
2142 sigemptyset(&action.sa_mask);
2143 action.sa_flags = 0;
2144 if (sigaction(SIGTERM, &action, &oact) != 0) {
2145 barf("can't install TERM handler");
2150 /* A capability holds the state a native thread needs in
2151 * order to execute STG code. At least one capability is
2152 * floating around (only SMP builds have more than one).
2156 #if defined(RTS_SUPPORTS_THREADS)
2157 /* start our haskell execution tasks */
2159 startTaskManager(RtsFlags.ParFlags.nNodes, taskStart);
2161 startTaskManager(0,taskStart);
2165 #if /* defined(SMP) ||*/ defined(PAR)
2169 #if defined(RTS_SUPPORTS_THREADS)
2170 RELEASE_LOCK(&sched_mutex);
2176 exitScheduler( void )
2178 #if defined(RTS_SUPPORTS_THREADS)
2181 shutting_down_scheduler = rtsTrue;
2184 /* -----------------------------------------------------------------------------
2185 Managing the per-task allocation areas.
2187 Each capability comes with an allocation area. These are
2188 fixed-length block lists into which allocation can be done.
2190 ToDo: no support for two-space collection at the moment???
2191 -------------------------------------------------------------------------- */
2193 /* -----------------------------------------------------------------------------
2194 * waitThread is the external interface for running a new computation
2195 * and waiting for the result.
2197 * In the non-SMP case, we create a new main thread, push it on the
2198 * main-thread stack, and invoke the scheduler to run it. The
2199 * scheduler will return when the top main thread on the stack has
2200 * completed or died, and fill in the necessary fields of the
2201 * main_thread structure.
2203 * In the SMP case, we create a main thread as before, but we then
2204 * create a new condition variable and sleep on it. When our new
2205 * main thread has completed, we'll be woken up and the status/result
2206 * will be in the main_thread struct.
2207 * -------------------------------------------------------------------------- */
2210 howManyThreadsAvail ( void )
2214 for (q = run_queue_hd; q != END_TSO_QUEUE; q = q->link)
2216 for (q = blocked_queue_hd; q != END_TSO_QUEUE; q = q->link)
2218 for (q = sleeping_queue; q != END_TSO_QUEUE; q = q->link)
2224 finishAllThreads ( void )
2227 while (run_queue_hd != END_TSO_QUEUE) {
2228 waitThread ( run_queue_hd, NULL);
2230 while (blocked_queue_hd != END_TSO_QUEUE) {
2231 waitThread ( blocked_queue_hd, NULL);
2233 while (sleeping_queue != END_TSO_QUEUE) {
2234 waitThread ( blocked_queue_hd, NULL);
2237 (blocked_queue_hd != END_TSO_QUEUE ||
2238 run_queue_hd != END_TSO_QUEUE ||
2239 sleeping_queue != END_TSO_QUEUE);
2243 waitThread(StgTSO *tso, /*out*/StgClosure **ret)
2247 m = stgMallocBytes(sizeof(StgMainThread), "waitThread");
2251 #if defined(RTS_SUPPORTS_THREADS)
2252 initCondition(&m->wakeup);
2255 /* see scheduleWaitThread() comment */
2256 ACQUIRE_LOCK(&sched_mutex);
2257 IF_DEBUG(scheduler, sched_belch("== scheduler: waiting for thread (%d)\n", tso->id));
2258 m->link = main_threads;
2260 RELEASE_LOCK(&sched_mutex);
2262 IF_DEBUG(scheduler, sched_belch("== scheduler: waiting for thread (%d)\n", tso->id));
2263 #if defined(THREADED_RTS)
2264 return waitThread_(m, rtsFalse);
2266 return waitThread_(m);
2272 waitThread_(StgMainThread* m
2273 #if defined(THREADED_RTS)
2274 , rtsBool blockWaiting
2278 SchedulerStatus stat;
2280 IF_DEBUG(scheduler, sched_belch("== scheduler: new main thread (%d)\n", m->tso->id));
2282 #if defined(RTS_SUPPORTS_THREADS)
2284 # if defined(THREADED_RTS)
2285 if (!blockWaiting) {
2286 /* In the threaded case, the OS thread that called main()
2287 * gets to enter the RTS directly without going via another
2291 ASSERT(m->stat != NoStatus);
2295 ACQUIRE_LOCK(&sched_mutex);
2297 waitCondition(&m->wakeup, &sched_mutex);
2298 } while (m->stat == NoStatus);
2301 /* GranSim specific init */
2302 CurrentTSO = m->tso; // the TSO to run
2303 procStatus[MainProc] = Busy; // status of main PE
2304 CurrentProc = MainProc; // PE to run it on
2308 RELEASE_LOCK(&sched_mutex);
2310 ASSERT(m->stat != NoStatus);
2315 #if defined(RTS_SUPPORTS_THREADS)
2316 closeCondition(&m->wakeup);
2319 IF_DEBUG(scheduler, fprintf(stderr, "== scheduler: main thread (%d) finished\n",
2323 #if defined(THREADED_RTS)
2326 RELEASE_LOCK(&sched_mutex);
2331 //@node Run queue code, Garbage Collextion Routines, Suspend and Resume, Main scheduling code
2332 //@subsection Run queue code
2336 NB: In GranSim we have many run queues; run_queue_hd is actually a macro
2337 unfolding to run_queue_hds[CurrentProc], thus CurrentProc is an
2338 implicit global variable that has to be correct when calling these
2342 /* Put the new thread on the head of the runnable queue.
2343 * The caller of createThread better push an appropriate closure
2344 * on this thread's stack before the scheduler is invoked.
2346 static /* inline */ void
2347 add_to_run_queue(tso)
2350 ASSERT(tso!=run_queue_hd && tso!=run_queue_tl);
2351 tso->link = run_queue_hd;
2353 if (run_queue_tl == END_TSO_QUEUE) {
2358 /* Put the new thread at the end of the runnable queue. */
2359 static /* inline */ void
2360 push_on_run_queue(tso)
2363 ASSERT(get_itbl((StgClosure *)tso)->type == TSO);
2364 ASSERT(run_queue_hd!=NULL && run_queue_tl!=NULL);
2365 ASSERT(tso!=run_queue_hd && tso!=run_queue_tl);
2366 if (run_queue_hd == END_TSO_QUEUE) {
2369 run_queue_tl->link = tso;
2375 Should be inlined because it's used very often in schedule. The tso
2376 argument is actually only needed in GranSim, where we want to have the
2377 possibility to schedule *any* TSO on the run queue, irrespective of the
2378 actual ordering. Therefore, if tso is not the nil TSO then we traverse
2379 the run queue and dequeue the tso, adjusting the links in the queue.
2381 //@cindex take_off_run_queue
2382 static /* inline */ StgTSO*
2383 take_off_run_queue(StgTSO *tso) {
2387 qetlaHbogh Qu' ngaSbogh ghomDaQ {tso} yIteq!
2389 if tso is specified, unlink that tso from the run_queue (doesn't have
2390 to be at the beginning of the queue); GranSim only
2392 if (tso!=END_TSO_QUEUE) {
2393 /* find tso in queue */
2394 for (t=run_queue_hd, prev=END_TSO_QUEUE;
2395 t!=END_TSO_QUEUE && t!=tso;
2399 /* now actually dequeue the tso */
2400 if (prev!=END_TSO_QUEUE) {
2401 ASSERT(run_queue_hd!=t);
2402 prev->link = t->link;
2404 /* t is at beginning of thread queue */
2405 ASSERT(run_queue_hd==t);
2406 run_queue_hd = t->link;
2408 /* t is at end of thread queue */
2409 if (t->link==END_TSO_QUEUE) {
2410 ASSERT(t==run_queue_tl);
2411 run_queue_tl = prev;
2413 ASSERT(run_queue_tl!=t);
2415 t->link = END_TSO_QUEUE;
2417 /* take tso from the beginning of the queue; std concurrent code */
2419 if (t != END_TSO_QUEUE) {
2420 run_queue_hd = t->link;
2421 t->link = END_TSO_QUEUE;
2422 if (run_queue_hd == END_TSO_QUEUE) {
2423 run_queue_tl = END_TSO_QUEUE;
2432 //@node Garbage Collextion Routines, Blocking Queue Routines, Run queue code, Main scheduling code
2433 //@subsection Garbage Collextion Routines
2435 /* ---------------------------------------------------------------------------
2436 Where are the roots that we know about?
2438 - all the threads on the runnable queue
2439 - all the threads on the blocked queue
2440 - all the threads on the sleeping queue
2441 - all the thread currently executing a _ccall_GC
2442 - all the "main threads"
2444 ------------------------------------------------------------------------ */
2446 /* This has to be protected either by the scheduler monitor, or by the
2447 garbage collection monitor (probably the latter).
2452 GetRoots(evac_fn evac)
2457 for (i=0; i<=RtsFlags.GranFlags.proc; i++) {
2458 if ((run_queue_hds[i] != END_TSO_QUEUE) && ((run_queue_hds[i] != NULL)))
2459 evac((StgClosure **)&run_queue_hds[i]);
2460 if ((run_queue_tls[i] != END_TSO_QUEUE) && ((run_queue_tls[i] != NULL)))
2461 evac((StgClosure **)&run_queue_tls[i]);
2463 if ((blocked_queue_hds[i] != END_TSO_QUEUE) && ((blocked_queue_hds[i] != NULL)))
2464 evac((StgClosure **)&blocked_queue_hds[i]);
2465 if ((blocked_queue_tls[i] != END_TSO_QUEUE) && ((blocked_queue_tls[i] != NULL)))
2466 evac((StgClosure **)&blocked_queue_tls[i]);
2467 if ((ccalling_threadss[i] != END_TSO_QUEUE) && ((ccalling_threadss[i] != NULL)))
2468 evac((StgClosure **)&ccalling_threads[i]);
2475 if (run_queue_hd != END_TSO_QUEUE) {
2476 ASSERT(run_queue_tl != END_TSO_QUEUE);
2477 evac((StgClosure **)&run_queue_hd);
2478 evac((StgClosure **)&run_queue_tl);
2481 if (blocked_queue_hd != END_TSO_QUEUE) {
2482 ASSERT(blocked_queue_tl != END_TSO_QUEUE);
2483 evac((StgClosure **)&blocked_queue_hd);
2484 evac((StgClosure **)&blocked_queue_tl);
2487 if (sleeping_queue != END_TSO_QUEUE) {
2488 evac((StgClosure **)&sleeping_queue);
2492 if (suspended_ccalling_threads != END_TSO_QUEUE) {
2493 evac((StgClosure **)&suspended_ccalling_threads);
2496 #if defined(PAR) || defined(GRAN)
2497 markSparkQueue(evac);
2501 /* -----------------------------------------------------------------------------
2504 This is the interface to the garbage collector from Haskell land.
2505 We provide this so that external C code can allocate and garbage
2506 collect when called from Haskell via _ccall_GC.
2508 It might be useful to provide an interface whereby the programmer
2509 can specify more roots (ToDo).
2511 This needs to be protected by the GC condition variable above. KH.
2512 -------------------------------------------------------------------------- */
2514 void (*extra_roots)(evac_fn);
2519 /* Obligated to hold this lock upon entry */
2520 ACQUIRE_LOCK(&sched_mutex);
2521 GarbageCollect(GetRoots,rtsFalse);
2522 RELEASE_LOCK(&sched_mutex);
2526 performMajorGC(void)
2528 ACQUIRE_LOCK(&sched_mutex);
2529 GarbageCollect(GetRoots,rtsTrue);
2530 RELEASE_LOCK(&sched_mutex);
2534 AllRoots(evac_fn evac)
2536 GetRoots(evac); // the scheduler's roots
2537 extra_roots(evac); // the user's roots
2541 performGCWithRoots(void (*get_roots)(evac_fn))
2543 ACQUIRE_LOCK(&sched_mutex);
2544 extra_roots = get_roots;
2545 GarbageCollect(AllRoots,rtsFalse);
2546 RELEASE_LOCK(&sched_mutex);
2549 /* -----------------------------------------------------------------------------
2552 If the thread has reached its maximum stack size, then raise the
2553 StackOverflow exception in the offending thread. Otherwise
2554 relocate the TSO into a larger chunk of memory and adjust its stack
2556 -------------------------------------------------------------------------- */
2559 threadStackOverflow(StgTSO *tso)
2561 nat new_stack_size, new_tso_size, diff, stack_words;
2565 IF_DEBUG(sanity,checkTSO(tso));
2566 if (tso->stack_size >= tso->max_stack_size) {
2569 belch("@@ threadStackOverflow of TSO %d (%p): stack too large (now %ld; max is %ld",
2570 tso->id, tso, tso->stack_size, tso->max_stack_size);
2571 /* If we're debugging, just print out the top of the stack */
2572 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2575 /* Send this thread the StackOverflow exception */
2576 raiseAsync(tso, (StgClosure *)stackOverflow_closure);
2580 /* Try to double the current stack size. If that takes us over the
2581 * maximum stack size for this thread, then use the maximum instead.
2582 * Finally round up so the TSO ends up as a whole number of blocks.
2584 new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
2585 new_tso_size = (nat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) +
2586 TSO_STRUCT_SIZE)/sizeof(W_);
2587 new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */
2588 new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
2590 IF_DEBUG(scheduler, fprintf(stderr,"== scheduler: increasing stack size from %d words to %d.\n", tso->stack_size, new_stack_size));
2592 dest = (StgTSO *)allocate(new_tso_size);
2593 TICK_ALLOC_TSO(new_stack_size,0);
2595 /* copy the TSO block and the old stack into the new area */
2596 memcpy(dest,tso,TSO_STRUCT_SIZE);
2597 stack_words = tso->stack + tso->stack_size - tso->sp;
2598 new_sp = (P_)dest + new_tso_size - stack_words;
2599 memcpy(new_sp, tso->sp, stack_words * sizeof(W_));
2601 /* relocate the stack pointers... */
2602 diff = (P_)new_sp - (P_)tso->sp; /* In *words* */
2603 dest->su = (StgUpdateFrame *) ((P_)dest->su + diff);
2605 dest->stack_size = new_stack_size;
2607 /* and relocate the update frame list */
2608 relocate_stack(dest, diff);
2610 /* Mark the old TSO as relocated. We have to check for relocated
2611 * TSOs in the garbage collector and any primops that deal with TSOs.
2613 * It's important to set the sp and su values to just beyond the end
2614 * of the stack, so we don't attempt to scavenge any part of the
2617 tso->what_next = ThreadRelocated;
2619 tso->sp = (P_)&(tso->stack[tso->stack_size]);
2620 tso->su = (StgUpdateFrame *)tso->sp;
2621 tso->why_blocked = NotBlocked;
2622 dest->mut_link = NULL;
2624 IF_PAR_DEBUG(verbose,
2625 belch("@@ threadStackOverflow of TSO %d (now at %p): stack size increased to %ld",
2626 tso->id, tso, tso->stack_size);
2627 /* If we're debugging, just print out the top of the stack */
2628 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2631 IF_DEBUG(sanity,checkTSO(tso));
2633 IF_DEBUG(scheduler,printTSO(dest));
2639 //@node Blocking Queue Routines, Exception Handling Routines, Garbage Collextion Routines, Main scheduling code
2640 //@subsection Blocking Queue Routines
2642 /* ---------------------------------------------------------------------------
2643 Wake up a queue that was blocked on some resource.
2644 ------------------------------------------------------------------------ */
2648 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2653 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2655 /* write RESUME events to log file and
2656 update blocked and fetch time (depending on type of the orig closure) */
2657 if (RtsFlags.ParFlags.ParStats.Full) {
2658 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
2659 GR_RESUMEQ, ((StgTSO *)bqe), ((StgTSO *)bqe)->block_info.closure,
2660 0, 0 /* spark_queue_len(ADVISORY_POOL) */);
2661 if (EMPTY_RUN_QUEUE())
2662 emitSchedule = rtsTrue;
2664 switch (get_itbl(node)->type) {
2666 ((StgTSO *)bqe)->par.fetchtime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2671 ((StgTSO *)bqe)->par.blocktime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2678 barf("{unblockOneLocked}Daq Qagh: unexpected closure in blocking queue");
2685 static StgBlockingQueueElement *
2686 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2689 PEs node_loc, tso_loc;
2691 node_loc = where_is(node); // should be lifted out of loop
2692 tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2693 tso_loc = where_is((StgClosure *)tso);
2694 if (IS_LOCAL_TO(PROCS(node),tso_loc)) { // TSO is local
2695 /* !fake_fetch => TSO is on CurrentProc is same as IS_LOCAL_TO */
2696 ASSERT(CurrentProc!=node_loc || tso_loc==CurrentProc);
2697 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.lunblocktime;
2698 // insertThread(tso, node_loc);
2699 new_event(tso_loc, tso_loc, CurrentTime[CurrentProc],
2701 tso, node, (rtsSpark*)NULL);
2702 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2705 } else { // TSO is remote (actually should be FMBQ)
2706 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.mpacktime +
2707 RtsFlags.GranFlags.Costs.gunblocktime +
2708 RtsFlags.GranFlags.Costs.latency;
2709 new_event(tso_loc, CurrentProc, CurrentTime[CurrentProc],
2711 tso, node, (rtsSpark*)NULL);
2712 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2715 /* the thread-queue-overhead is accounted for in either Resume or UnblockThread */
2717 fprintf(stderr," %s TSO %d (%p) [PE %d] (block_info.closure=%p) (next=%p) ,",
2718 (node_loc==tso_loc ? "Local" : "Global"),
2719 tso->id, tso, CurrentProc, tso->block_info.closure, tso->link));
2720 tso->block_info.closure = NULL;
2721 IF_DEBUG(scheduler,belch("-- Waking up thread %ld (%p)",
2725 static StgBlockingQueueElement *
2726 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2728 StgBlockingQueueElement *next;
2730 switch (get_itbl(bqe)->type) {
2732 ASSERT(((StgTSO *)bqe)->why_blocked != NotBlocked);
2733 /* if it's a TSO just push it onto the run_queue */
2735 // ((StgTSO *)bqe)->link = END_TSO_QUEUE; // debugging?
2736 PUSH_ON_RUN_QUEUE((StgTSO *)bqe);
2738 unblockCount(bqe, node);
2739 /* reset blocking status after dumping event */
2740 ((StgTSO *)bqe)->why_blocked = NotBlocked;
2744 /* if it's a BLOCKED_FETCH put it on the PendingFetches list */
2746 bqe->link = (StgBlockingQueueElement *)PendingFetches;
2747 PendingFetches = (StgBlockedFetch *)bqe;
2751 /* can ignore this case in a non-debugging setup;
2752 see comments on RBHSave closures above */
2754 /* check that the closure is an RBHSave closure */
2755 ASSERT(get_itbl((StgClosure *)bqe) == &stg_RBH_Save_0_info ||
2756 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_1_info ||
2757 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_2_info);
2761 barf("{unblockOneLocked}Daq Qagh: Unexpected IP (%#lx; %s) in blocking queue at %#lx\n",
2762 get_itbl((StgClosure *)bqe), info_type((StgClosure *)bqe),
2766 IF_PAR_DEBUG(bq, fprintf(stderr, ", %p (%s)", bqe, info_type((StgClosure*)bqe)));
2770 #else /* !GRAN && !PAR */
2772 unblockOneLocked(StgTSO *tso)
2776 ASSERT(get_itbl(tso)->type == TSO);
2777 ASSERT(tso->why_blocked != NotBlocked);
2778 tso->why_blocked = NotBlocked;
2780 PUSH_ON_RUN_QUEUE(tso);
2782 IF_DEBUG(scheduler,sched_belch("waking up thread %ld", tso->id));
2787 #if defined(GRAN) || defined(PAR)
2788 inline StgBlockingQueueElement *
2789 unblockOne(StgBlockingQueueElement *bqe, StgClosure *node)
2791 ACQUIRE_LOCK(&sched_mutex);
2792 bqe = unblockOneLocked(bqe, node);
2793 RELEASE_LOCK(&sched_mutex);
2798 unblockOne(StgTSO *tso)
2800 ACQUIRE_LOCK(&sched_mutex);
2801 tso = unblockOneLocked(tso);
2802 RELEASE_LOCK(&sched_mutex);
2809 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2811 StgBlockingQueueElement *bqe;
2816 belch("##-_ AwBQ for node %p on PE %d @ %ld by TSO %d (%p): ", \
2817 node, CurrentProc, CurrentTime[CurrentProc],
2818 CurrentTSO->id, CurrentTSO));
2820 node_loc = where_is(node);
2822 ASSERT(q == END_BQ_QUEUE ||
2823 get_itbl(q)->type == TSO || // q is either a TSO or an RBHSave
2824 get_itbl(q)->type == CONSTR); // closure (type constructor)
2825 ASSERT(is_unique(node));
2827 /* FAKE FETCH: magically copy the node to the tso's proc;
2828 no Fetch necessary because in reality the node should not have been
2829 moved to the other PE in the first place
2831 if (CurrentProc!=node_loc) {
2833 belch("## node %p is on PE %d but CurrentProc is %d (TSO %d); assuming fake fetch and adjusting bitmask (old: %#x)",
2834 node, node_loc, CurrentProc, CurrentTSO->id,
2835 // CurrentTSO, where_is(CurrentTSO),
2836 node->header.gran.procs));
2837 node->header.gran.procs = (node->header.gran.procs) | PE_NUMBER(CurrentProc);
2839 belch("## new bitmask of node %p is %#x",
2840 node, node->header.gran.procs));
2841 if (RtsFlags.GranFlags.GranSimStats.Global) {
2842 globalGranStats.tot_fake_fetches++;
2847 // ToDo: check: ASSERT(CurrentProc==node_loc);
2848 while (get_itbl(bqe)->type==TSO) { // q != END_TSO_QUEUE) {
2851 bqe points to the current element in the queue
2852 next points to the next element in the queue
2854 //tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2855 //tso_loc = where_is(tso);
2857 bqe = unblockOneLocked(bqe, node);
2860 /* if this is the BQ of an RBH, we have to put back the info ripped out of
2861 the closure to make room for the anchor of the BQ */
2862 if (bqe!=END_BQ_QUEUE) {
2863 ASSERT(get_itbl(node)->type == RBH && get_itbl(bqe)->type == CONSTR);
2865 ASSERT((info_ptr==&RBH_Save_0_info) ||
2866 (info_ptr==&RBH_Save_1_info) ||
2867 (info_ptr==&RBH_Save_2_info));
2869 /* cf. convertToRBH in RBH.c for writing the RBHSave closure */
2870 ((StgRBH *)node)->blocking_queue = (StgBlockingQueueElement *)((StgRBHSave *)bqe)->payload[0];
2871 ((StgRBH *)node)->mut_link = (StgMutClosure *)((StgRBHSave *)bqe)->payload[1];
2874 belch("## Filled in RBH_Save for %p (%s) at end of AwBQ",
2875 node, info_type(node)));
2878 /* statistics gathering */
2879 if (RtsFlags.GranFlags.GranSimStats.Global) {
2880 // globalGranStats.tot_bq_processing_time += bq_processing_time;
2881 globalGranStats.tot_bq_len += len; // total length of all bqs awakened
2882 // globalGranStats.tot_bq_len_local += len_local; // same for local TSOs only
2883 globalGranStats.tot_awbq++; // total no. of bqs awakened
2886 fprintf(stderr,"## BQ Stats of %p: [%d entries] %s\n",
2887 node, len, (bqe!=END_BQ_QUEUE) ? "RBH" : ""));
2891 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2893 StgBlockingQueueElement *bqe;
2895 ACQUIRE_LOCK(&sched_mutex);
2897 IF_PAR_DEBUG(verbose,
2898 belch("##-_ AwBQ for node %p on [%x]: ",
2902 if(get_itbl(q)->type == CONSTR || q==END_BQ_QUEUE) {
2903 IF_PAR_DEBUG(verbose, belch("## ... nothing to unblock so lets just return. RFP (BUG?)"));
2908 ASSERT(q == END_BQ_QUEUE ||
2909 get_itbl(q)->type == TSO ||
2910 get_itbl(q)->type == BLOCKED_FETCH ||
2911 get_itbl(q)->type == CONSTR);
2914 while (get_itbl(bqe)->type==TSO ||
2915 get_itbl(bqe)->type==BLOCKED_FETCH) {
2916 bqe = unblockOneLocked(bqe, node);
2918 RELEASE_LOCK(&sched_mutex);
2921 #else /* !GRAN && !PAR */
2923 awakenBlockedQueue(StgTSO *tso)
2925 ACQUIRE_LOCK(&sched_mutex);
2926 while (tso != END_TSO_QUEUE) {
2927 tso = unblockOneLocked(tso);
2929 RELEASE_LOCK(&sched_mutex);
2933 //@node Exception Handling Routines, Debugging Routines, Blocking Queue Routines, Main scheduling code
2934 //@subsection Exception Handling Routines
2936 /* ---------------------------------------------------------------------------
2938 - usually called inside a signal handler so it mustn't do anything fancy.
2939 ------------------------------------------------------------------------ */
2942 interruptStgRts(void)
2948 /* -----------------------------------------------------------------------------
2951 This is for use when we raise an exception in another thread, which
2953 This has nothing to do with the UnblockThread event in GranSim. -- HWL
2954 -------------------------------------------------------------------------- */
2956 #if defined(GRAN) || defined(PAR)
2958 NB: only the type of the blocking queue is different in GranSim and GUM
2959 the operations on the queue-elements are the same
2960 long live polymorphism!
2962 Locks: sched_mutex is held upon entry and exit.
2966 unblockThread(StgTSO *tso)
2968 StgBlockingQueueElement *t, **last;
2970 switch (tso->why_blocked) {
2973 return; /* not blocked */
2976 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
2978 StgBlockingQueueElement *last_tso = END_BQ_QUEUE;
2979 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
2981 last = (StgBlockingQueueElement **)&mvar->head;
2982 for (t = (StgBlockingQueueElement *)mvar->head;
2984 last = &t->link, last_tso = t, t = t->link) {
2985 if (t == (StgBlockingQueueElement *)tso) {
2986 *last = (StgBlockingQueueElement *)tso->link;
2987 if (mvar->tail == tso) {
2988 mvar->tail = (StgTSO *)last_tso;
2993 barf("unblockThread (MVAR): TSO not found");
2996 case BlockedOnBlackHole:
2997 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
2999 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
3001 last = &bq->blocking_queue;
3002 for (t = bq->blocking_queue;
3004 last = &t->link, t = t->link) {
3005 if (t == (StgBlockingQueueElement *)tso) {
3006 *last = (StgBlockingQueueElement *)tso->link;
3010 barf("unblockThread (BLACKHOLE): TSO not found");
3013 case BlockedOnException:
3015 StgTSO *target = tso->block_info.tso;
3017 ASSERT(get_itbl(target)->type == TSO);
3019 if (target->what_next == ThreadRelocated) {
3020 target = target->link;
3021 ASSERT(get_itbl(target)->type == TSO);
3024 ASSERT(target->blocked_exceptions != NULL);
3026 last = (StgBlockingQueueElement **)&target->blocked_exceptions;
3027 for (t = (StgBlockingQueueElement *)target->blocked_exceptions;
3029 last = &t->link, t = t->link) {
3030 ASSERT(get_itbl(t)->type == TSO);
3031 if (t == (StgBlockingQueueElement *)tso) {
3032 *last = (StgBlockingQueueElement *)tso->link;
3036 barf("unblockThread (Exception): TSO not found");
3040 case BlockedOnWrite:
3042 /* take TSO off blocked_queue */
3043 StgBlockingQueueElement *prev = NULL;
3044 for (t = (StgBlockingQueueElement *)blocked_queue_hd; t != END_BQ_QUEUE;
3045 prev = t, t = t->link) {
3046 if (t == (StgBlockingQueueElement *)tso) {
3048 blocked_queue_hd = (StgTSO *)t->link;
3049 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
3050 blocked_queue_tl = END_TSO_QUEUE;
3053 prev->link = t->link;
3054 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
3055 blocked_queue_tl = (StgTSO *)prev;
3061 barf("unblockThread (I/O): TSO not found");
3064 case BlockedOnDelay:
3066 /* take TSO off sleeping_queue */
3067 StgBlockingQueueElement *prev = NULL;
3068 for (t = (StgBlockingQueueElement *)sleeping_queue; t != END_BQ_QUEUE;
3069 prev = t, t = t->link) {
3070 if (t == (StgBlockingQueueElement *)tso) {
3072 sleeping_queue = (StgTSO *)t->link;
3074 prev->link = t->link;
3079 barf("unblockThread (I/O): TSO not found");
3083 barf("unblockThread");
3087 tso->link = END_TSO_QUEUE;
3088 tso->why_blocked = NotBlocked;
3089 tso->block_info.closure = NULL;
3090 PUSH_ON_RUN_QUEUE(tso);
3094 unblockThread(StgTSO *tso)
3098 /* To avoid locking unnecessarily. */
3099 if (tso->why_blocked == NotBlocked) {
3103 switch (tso->why_blocked) {
3106 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
3108 StgTSO *last_tso = END_TSO_QUEUE;
3109 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
3112 for (t = mvar->head; t != END_TSO_QUEUE;
3113 last = &t->link, last_tso = t, t = t->link) {
3116 if (mvar->tail == tso) {
3117 mvar->tail = last_tso;
3122 barf("unblockThread (MVAR): TSO not found");
3125 case BlockedOnBlackHole:
3126 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
3128 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
3130 last = &bq->blocking_queue;
3131 for (t = bq->blocking_queue; t != END_TSO_QUEUE;
3132 last = &t->link, t = t->link) {
3138 barf("unblockThread (BLACKHOLE): TSO not found");
3141 case BlockedOnException:
3143 StgTSO *target = tso->block_info.tso;
3145 ASSERT(get_itbl(target)->type == TSO);
3147 while (target->what_next == ThreadRelocated) {
3148 target = target->link;
3149 ASSERT(get_itbl(target)->type == TSO);
3152 ASSERT(target->blocked_exceptions != NULL);
3154 last = &target->blocked_exceptions;
3155 for (t = target->blocked_exceptions; t != END_TSO_QUEUE;
3156 last = &t->link, t = t->link) {
3157 ASSERT(get_itbl(t)->type == TSO);
3163 barf("unblockThread (Exception): TSO not found");
3167 case BlockedOnWrite:
3169 StgTSO *prev = NULL;
3170 for (t = blocked_queue_hd; t != END_TSO_QUEUE;
3171 prev = t, t = t->link) {
3174 blocked_queue_hd = t->link;
3175 if (blocked_queue_tl == t) {
3176 blocked_queue_tl = END_TSO_QUEUE;
3179 prev->link = t->link;
3180 if (blocked_queue_tl == t) {
3181 blocked_queue_tl = prev;
3187 barf("unblockThread (I/O): TSO not found");
3190 case BlockedOnDelay:
3192 StgTSO *prev = NULL;
3193 for (t = sleeping_queue; t != END_TSO_QUEUE;
3194 prev = t, t = t->link) {
3197 sleeping_queue = t->link;
3199 prev->link = t->link;
3204 barf("unblockThread (I/O): TSO not found");
3208 barf("unblockThread");
3212 tso->link = END_TSO_QUEUE;
3213 tso->why_blocked = NotBlocked;
3214 tso->block_info.closure = NULL;
3215 PUSH_ON_RUN_QUEUE(tso);
3219 /* -----------------------------------------------------------------------------
3222 * The following function implements the magic for raising an
3223 * asynchronous exception in an existing thread.
3225 * We first remove the thread from any queue on which it might be
3226 * blocked. The possible blockages are MVARs and BLACKHOLE_BQs.
3228 * We strip the stack down to the innermost CATCH_FRAME, building
3229 * thunks in the heap for all the active computations, so they can
3230 * be restarted if necessary. When we reach a CATCH_FRAME, we build
3231 * an application of the handler to the exception, and push it on
3232 * the top of the stack.
3234 * How exactly do we save all the active computations? We create an
3235 * AP_UPD for every UpdateFrame on the stack. Entering one of these
3236 * AP_UPDs pushes everything from the corresponding update frame
3237 * upwards onto the stack. (Actually, it pushes everything up to the
3238 * next update frame plus a pointer to the next AP_UPD object.
3239 * Entering the next AP_UPD object pushes more onto the stack until we
3240 * reach the last AP_UPD object - at which point the stack should look
3241 * exactly as it did when we killed the TSO and we can continue
3242 * execution by entering the closure on top of the stack.
3244 * We can also kill a thread entirely - this happens if either (a) the
3245 * exception passed to raiseAsync is NULL, or (b) there's no
3246 * CATCH_FRAME on the stack. In either case, we strip the entire
3247 * stack and replace the thread with a zombie.
3249 * Locks: sched_mutex held upon entry nor exit.
3251 * -------------------------------------------------------------------------- */
3254 deleteThread(StgTSO *tso)
3256 raiseAsync(tso,NULL);
3260 raiseAsyncWithLock(StgTSO *tso, StgClosure *exception)
3262 /* When raising async exs from contexts where sched_mutex isn't held;
3263 use raiseAsyncWithLock(). */
3264 ACQUIRE_LOCK(&sched_mutex);
3265 raiseAsync(tso,exception);
3266 RELEASE_LOCK(&sched_mutex);
3270 raiseAsync(StgTSO *tso, StgClosure *exception)
3272 StgUpdateFrame* su = tso->su;
3273 StgPtr sp = tso->sp;
3275 /* Thread already dead? */
3276 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
3280 IF_DEBUG(scheduler, sched_belch("raising exception in thread %ld.", tso->id));
3282 /* Remove it from any blocking queues */
3285 IF_DEBUG(scheduler, sched_belch("raising exception in thread %ld.", tso->id));
3286 /* The stack freezing code assumes there's a closure pointer on
3287 * the top of the stack. This isn't always the case with compiled
3288 * code, so we have to push a dummy closure on the top which just
3289 * returns to the next return address on the stack.
3291 if ( LOOKS_LIKE_GHC_INFO((void*)*sp) ) {
3292 *(--sp) = (W_)&stg_dummy_ret_closure;
3296 nat words = ((P_)su - (P_)sp) - 1;
3300 /* If we find a CATCH_FRAME, and we've got an exception to raise,
3301 * then build the THUNK raise(exception), and leave it on
3302 * top of the CATCH_FRAME ready to enter.
3304 if (get_itbl(su)->type == CATCH_FRAME && exception != NULL) {
3306 StgCatchFrame *cf = (StgCatchFrame *)su;
3310 /* we've got an exception to raise, so let's pass it to the
3311 * handler in this frame.
3313 raise = (StgClosure *)allocate(sizeofW(StgClosure)+1);
3314 TICK_ALLOC_SE_THK(1,0);
3315 SET_HDR(raise,&stg_raise_info,cf->header.prof.ccs);
3316 raise->payload[0] = exception;
3318 /* throw away the stack from Sp up to the CATCH_FRAME.
3322 /* Ensure that async excpetions are blocked now, so we don't get
3323 * a surprise exception before we get around to executing the
3326 if (tso->blocked_exceptions == NULL) {
3327 tso->blocked_exceptions = END_TSO_QUEUE;
3330 /* Put the newly-built THUNK on top of the stack, ready to execute
3331 * when the thread restarts.
3336 tso->what_next = ThreadEnterGHC;
3337 IF_DEBUG(sanity, checkTSO(tso));
3341 /* First build an AP_UPD consisting of the stack chunk above the
3342 * current update frame, with the top word on the stack as the
3345 ap = (StgAP_UPD *)allocate(AP_sizeW(words));
3350 ap->fun = (StgClosure *)sp[0];
3352 for(i=0; i < (nat)words; ++i) {
3353 ap->payload[i] = (StgClosure *)*sp++;
3356 switch (get_itbl(su)->type) {
3360 SET_HDR(ap,&stg_AP_UPD_info,su->header.prof.ccs /* ToDo */);
3361 TICK_ALLOC_UP_THK(words+1,0);
3364 fprintf(stderr, "scheduler: Updating ");
3365 printPtr((P_)su->updatee);
3366 fprintf(stderr, " with ");
3367 printObj((StgClosure *)ap);
3370 /* Replace the updatee with an indirection - happily
3371 * this will also wake up any threads currently
3372 * waiting on the result.
3374 * Warning: if we're in a loop, more than one update frame on
3375 * the stack may point to the same object. Be careful not to
3376 * overwrite an IND_OLDGEN in this case, because we'll screw
3377 * up the mutable lists. To be on the safe side, don't
3378 * overwrite any kind of indirection at all. See also
3379 * threadSqueezeStack in GC.c, where we have to make a similar
3382 if (!closure_IND(su->updatee)) {
3383 UPD_IND_NOLOCK(su->updatee,ap); /* revert the black hole */
3386 sp += sizeofW(StgUpdateFrame) -1;
3387 sp[0] = (W_)ap; /* push onto stack */
3393 StgCatchFrame *cf = (StgCatchFrame *)su;
3396 /* We want a PAP, not an AP_UPD. Fortunately, the
3397 * layout's the same.
3399 SET_HDR(ap,&stg_PAP_info,su->header.prof.ccs /* ToDo */);
3400 TICK_ALLOC_UPD_PAP(words+1,0);
3402 /* now build o = FUN(catch,ap,handler) */
3403 o = (StgClosure *)allocate(sizeofW(StgClosure)+2);
3404 TICK_ALLOC_FUN(2,0);
3405 SET_HDR(o,&stg_catch_info,su->header.prof.ccs /* ToDo */);
3406 o->payload[0] = (StgClosure *)ap;
3407 o->payload[1] = cf->handler;
3410 fprintf(stderr, "scheduler: Built ");
3411 printObj((StgClosure *)o);
3414 /* pop the old handler and put o on the stack */
3416 sp += sizeofW(StgCatchFrame) - 1;
3423 StgSeqFrame *sf = (StgSeqFrame *)su;
3426 SET_HDR(ap,&stg_PAP_info,su->header.prof.ccs /* ToDo */);
3427 TICK_ALLOC_UPD_PAP(words+1,0);
3429 /* now build o = FUN(seq,ap) */
3430 o = (StgClosure *)allocate(sizeofW(StgClosure)+1);
3431 TICK_ALLOC_SE_THK(1,0);
3432 SET_HDR(o,&stg_seq_info,su->header.prof.ccs /* ToDo */);
3433 o->payload[0] = (StgClosure *)ap;
3436 fprintf(stderr, "scheduler: Built ");
3437 printObj((StgClosure *)o);
3440 /* pop the old handler and put o on the stack */
3442 sp += sizeofW(StgSeqFrame) - 1;
3448 /* We've stripped the entire stack, the thread is now dead. */
3449 sp += sizeofW(StgStopFrame) - 1;
3450 sp[0] = (W_)exception; /* save the exception */
3451 tso->what_next = ThreadKilled;
3452 tso->su = (StgUpdateFrame *)(sp+1);
3463 /* -----------------------------------------------------------------------------
3464 resurrectThreads is called after garbage collection on the list of
3465 threads found to be garbage. Each of these threads will be woken
3466 up and sent a signal: BlockedOnDeadMVar if the thread was blocked
3467 on an MVar, or NonTermination if the thread was blocked on a Black
3470 Locks: sched_mutex isn't held upon entry nor exit.
3471 -------------------------------------------------------------------------- */
3474 resurrectThreads( StgTSO *threads )
3478 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
3479 next = tso->global_link;
3480 tso->global_link = all_threads;
3482 IF_DEBUG(scheduler, sched_belch("resurrecting thread %d", tso->id));
3484 switch (tso->why_blocked) {
3486 case BlockedOnException:
3487 /* Called by GC - sched_mutex lock is currently held. */
3488 raiseAsync(tso,(StgClosure *)BlockedOnDeadMVar_closure);
3490 case BlockedOnBlackHole:
3491 raiseAsync(tso,(StgClosure *)NonTermination_closure);
3494 /* This might happen if the thread was blocked on a black hole
3495 * belonging to a thread that we've just woken up (raiseAsync
3496 * can wake up threads, remember...).
3500 barf("resurrectThreads: thread blocked in a strange way");
3505 /* -----------------------------------------------------------------------------
3506 * Blackhole detection: if we reach a deadlock, test whether any
3507 * threads are blocked on themselves. Any threads which are found to
3508 * be self-blocked get sent a NonTermination exception.
3510 * This is only done in a deadlock situation in order to avoid
3511 * performance overhead in the normal case.
3513 * Locks: sched_mutex is held upon entry and exit.
3514 * -------------------------------------------------------------------------- */
3517 detectBlackHoles( void )
3519 StgTSO *t = all_threads;
3520 StgUpdateFrame *frame;
3521 StgClosure *blocked_on;
3523 for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
3525 while (t->what_next == ThreadRelocated) {
3527 ASSERT(get_itbl(t)->type == TSO);
3530 if (t->why_blocked != BlockedOnBlackHole) {
3534 blocked_on = t->block_info.closure;
3536 for (frame = t->su; ; frame = frame->link) {
3537 switch (get_itbl(frame)->type) {
3540 if (frame->updatee == blocked_on) {
3541 /* We are blocking on one of our own computations, so
3542 * send this thread the NonTermination exception.
3545 sched_belch("thread %d is blocked on itself", t->id));
3546 raiseAsync(t, (StgClosure *)NonTermination_closure);
3567 //@node Debugging Routines, Index, Exception Handling Routines, Main scheduling code
3568 //@subsection Debugging Routines
3570 /* -----------------------------------------------------------------------------
3571 Debugging: why is a thread blocked
3572 -------------------------------------------------------------------------- */
3577 printThreadBlockage(StgTSO *tso)
3579 switch (tso->why_blocked) {
3581 fprintf(stderr,"is blocked on read from fd %d", tso->block_info.fd);
3583 case BlockedOnWrite:
3584 fprintf(stderr,"is blocked on write to fd %d", tso->block_info.fd);
3586 case BlockedOnDelay:
3587 fprintf(stderr,"is blocked until %d", tso->block_info.target);
3590 fprintf(stderr,"is blocked on an MVar");
3592 case BlockedOnException:
3593 fprintf(stderr,"is blocked on delivering an exception to thread %d",
3594 tso->block_info.tso->id);
3596 case BlockedOnBlackHole:
3597 fprintf(stderr,"is blocked on a black hole");
3600 fprintf(stderr,"is not blocked");
3604 fprintf(stderr,"is blocked on global address; local FM_BQ is %p (%s)",
3605 tso->block_info.closure, info_type(tso->block_info.closure));
3607 case BlockedOnGA_NoSend:
3608 fprintf(stderr,"is blocked on global address (no send); local FM_BQ is %p (%s)",
3609 tso->block_info.closure, info_type(tso->block_info.closure));
3612 #if defined(RTS_SUPPORTS_THREADS)
3613 case BlockedOnCCall:
3614 fprintf(stderr,"is blocked on an external call");
3618 barf("printThreadBlockage: strange tso->why_blocked: %d for TSO %d (%d)",
3619 tso->why_blocked, tso->id, tso);
3624 printThreadStatus(StgTSO *tso)
3626 switch (tso->what_next) {
3628 fprintf(stderr,"has been killed");
3630 case ThreadComplete:
3631 fprintf(stderr,"has completed");
3634 printThreadBlockage(tso);
3639 printAllThreads(void)
3645 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
3646 ullong_format_string(TIME_ON_PROC(CurrentProc),
3647 time_string, rtsFalse/*no commas!*/);
3649 sched_belch("all threads at [%s]:", time_string);
3651 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
3652 ullong_format_string(CURRENT_TIME,
3653 time_string, rtsFalse/*no commas!*/);
3655 sched_belch("all threads at [%s]:", time_string);
3657 sched_belch("all threads:");
3660 for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
3661 fprintf(stderr, "\tthread %d @ %p ", t->id, (void *)t);
3662 label = lookupThreadLabel((StgWord)t);
3663 if (label) fprintf(stderr,"[\"%s\"] ",(char *)label);
3664 printThreadStatus(t);
3665 fprintf(stderr,"\n");
3670 Print a whole blocking queue attached to node (debugging only).
3675 print_bq (StgClosure *node)
3677 StgBlockingQueueElement *bqe;
3681 fprintf(stderr,"## BQ of closure %p (%s): ",
3682 node, info_type(node));
3684 /* should cover all closures that may have a blocking queue */
3685 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
3686 get_itbl(node)->type == FETCH_ME_BQ ||
3687 get_itbl(node)->type == RBH ||
3688 get_itbl(node)->type == MVAR);
3690 ASSERT(node!=(StgClosure*)NULL); // sanity check
3692 print_bqe(((StgBlockingQueue*)node)->blocking_queue);
3696 Print a whole blocking queue starting with the element bqe.
3699 print_bqe (StgBlockingQueueElement *bqe)
3704 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
3706 for (end = (bqe==END_BQ_QUEUE);
3707 !end; // iterate until bqe points to a CONSTR
3708 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE),
3709 bqe = end ? END_BQ_QUEUE : bqe->link) {
3710 ASSERT(bqe != END_BQ_QUEUE); // sanity check
3711 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
3712 /* types of closures that may appear in a blocking queue */
3713 ASSERT(get_itbl(bqe)->type == TSO ||
3714 get_itbl(bqe)->type == BLOCKED_FETCH ||
3715 get_itbl(bqe)->type == CONSTR);
3716 /* only BQs of an RBH end with an RBH_Save closure */
3717 //ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
3719 switch (get_itbl(bqe)->type) {
3721 fprintf(stderr," TSO %u (%x),",
3722 ((StgTSO *)bqe)->id, ((StgTSO *)bqe));
3725 fprintf(stderr," BF (node=%p, ga=((%x, %d, %x)),",
3726 ((StgBlockedFetch *)bqe)->node,
3727 ((StgBlockedFetch *)bqe)->ga.payload.gc.gtid,
3728 ((StgBlockedFetch *)bqe)->ga.payload.gc.slot,
3729 ((StgBlockedFetch *)bqe)->ga.weight);
3732 fprintf(stderr," %s (IP %p),",
3733 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
3734 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
3735 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
3736 "RBH_Save_?"), get_itbl(bqe));
3739 barf("Unexpected closure type %s in blocking queue", // of %p (%s)",
3740 info_type((StgClosure *)bqe)); // , node, info_type(node));
3744 fputc('\n', stderr);
3746 # elif defined(GRAN)
3748 print_bq (StgClosure *node)
3750 StgBlockingQueueElement *bqe;
3751 PEs node_loc, tso_loc;
3754 /* should cover all closures that may have a blocking queue */
3755 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
3756 get_itbl(node)->type == FETCH_ME_BQ ||
3757 get_itbl(node)->type == RBH);
3759 ASSERT(node!=(StgClosure*)NULL); // sanity check
3760 node_loc = where_is(node);
3762 fprintf(stderr,"## BQ of closure %p (%s) on [PE %d]: ",
3763 node, info_type(node), node_loc);
3766 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
3768 for (bqe = ((StgBlockingQueue*)node)->blocking_queue, end = (bqe==END_BQ_QUEUE);
3769 !end; // iterate until bqe points to a CONSTR
3770 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) {
3771 ASSERT(bqe != END_BQ_QUEUE); // sanity check
3772 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
3773 /* types of closures that may appear in a blocking queue */
3774 ASSERT(get_itbl(bqe)->type == TSO ||
3775 get_itbl(bqe)->type == CONSTR);
3776 /* only BQs of an RBH end with an RBH_Save closure */
3777 ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
3779 tso_loc = where_is((StgClosure *)bqe);
3780 switch (get_itbl(bqe)->type) {
3782 fprintf(stderr," TSO %d (%p) on [PE %d],",
3783 ((StgTSO *)bqe)->id, (StgTSO *)bqe, tso_loc);
3786 fprintf(stderr," %s (IP %p),",
3787 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
3788 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
3789 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
3790 "RBH_Save_?"), get_itbl(bqe));
3793 barf("Unexpected closure type %s in blocking queue of %p (%s)",
3794 info_type((StgClosure *)bqe), node, info_type(node));
3798 fputc('\n', stderr);
3802 Nice and easy: only TSOs on the blocking queue
3805 print_bq (StgClosure *node)
3809 ASSERT(node!=(StgClosure*)NULL); // sanity check
3810 for (tso = ((StgBlockingQueue*)node)->blocking_queue;
3811 tso != END_TSO_QUEUE;
3813 ASSERT(tso!=NULL && tso!=END_TSO_QUEUE); // sanity check
3814 ASSERT(get_itbl(tso)->type == TSO); // guess what, sanity check
3815 fprintf(stderr," TSO %d (%p),", tso->id, tso);
3817 fputc('\n', stderr);
3828 for (i=0, tso=run_queue_hd;
3829 tso != END_TSO_QUEUE;
3838 sched_belch(char *s, ...)
3843 fprintf(stderr, "scheduler (task %ld): ", osThreadId());
3845 fprintf(stderr, "== ");
3847 fprintf(stderr, "scheduler: ");
3849 vfprintf(stderr, s, ap);
3850 fprintf(stderr, "\n");
3857 //@node Index, , Debugging Routines, Main scheduling code
3861 //* StgMainThread:: @cindex\s-+StgMainThread
3862 //* awaken_blocked_queue:: @cindex\s-+awaken_blocked_queue
3863 //* blocked_queue_hd:: @cindex\s-+blocked_queue_hd
3864 //* blocked_queue_tl:: @cindex\s-+blocked_queue_tl
3865 //* context_switch:: @cindex\s-+context_switch
3866 //* createThread:: @cindex\s-+createThread
3867 //* gc_pending_cond:: @cindex\s-+gc_pending_cond
3868 //* initScheduler:: @cindex\s-+initScheduler
3869 //* interrupted:: @cindex\s-+interrupted
3870 //* next_thread_id:: @cindex\s-+next_thread_id
3871 //* print_bq:: @cindex\s-+print_bq
3872 //* run_queue_hd:: @cindex\s-+run_queue_hd
3873 //* run_queue_tl:: @cindex\s-+run_queue_tl
3874 //* sched_mutex:: @cindex\s-+sched_mutex
3875 //* schedule:: @cindex\s-+schedule
3876 //* take_off_run_queue:: @cindex\s-+take_off_run_queue
3877 //* term_mutex:: @cindex\s-+term_mutex