1 /* ---------------------------------------------------------------------------
2 * $Id: Schedule.c,v 1.138 2002/04/23 06:34:27 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"
100 #include "Proftimer.h"
101 #include "ProfHeap.h"
103 #if defined(GRAN) || defined(PAR)
104 # include "GranSimRts.h"
105 # include "GranSim.h"
106 # include "ParallelRts.h"
107 # include "Parallel.h"
108 # include "ParallelDebug.h"
109 # include "FetchMe.h"
113 #include "Capability.h"
114 #include "OSThreads.h"
117 #ifdef HAVE_SYS_TYPES_H
118 #include <sys/types.h>
126 //@node Variables and Data structures, Prototypes, Includes, Main scheduling code
127 //@subsection Variables and Data structures
129 /* Main thread queue.
130 * Locks required: sched_mutex.
132 StgMainThread *main_threads;
135 * Locks required: sched_mutex.
139 StgTSO* ActiveTSO = NULL; /* for assigning system costs; GranSim-Light only */
140 /* rtsTime TimeOfNextEvent, EndOfTimeSlice; now in GranSim.c */
143 In GranSim we have a runnable and a blocked queue for each processor.
144 In order to minimise code changes new arrays run_queue_hds/tls
145 are created. run_queue_hd is then a short cut (macro) for
146 run_queue_hds[CurrentProc] (see GranSim.h).
149 StgTSO *run_queue_hds[MAX_PROC], *run_queue_tls[MAX_PROC];
150 StgTSO *blocked_queue_hds[MAX_PROC], *blocked_queue_tls[MAX_PROC];
151 StgTSO *ccalling_threadss[MAX_PROC];
152 /* We use the same global list of threads (all_threads) in GranSim as in
153 the std RTS (i.e. we are cheating). However, we don't use this list in
154 the GranSim specific code at the moment (so we are only potentially
159 StgTSO *run_queue_hd, *run_queue_tl;
160 StgTSO *blocked_queue_hd, *blocked_queue_tl;
161 StgTSO *sleeping_queue; /* perhaps replace with a hash table? */
165 /* Linked list of all threads.
166 * Used for detecting garbage collected threads.
170 /* When a thread performs a safe C call (_ccall_GC, using old
171 * terminology), it gets put on the suspended_ccalling_threads
172 * list. Used by the garbage collector.
174 static StgTSO *suspended_ccalling_threads;
176 static StgTSO *threadStackOverflow(StgTSO *tso);
178 /* KH: The following two flags are shared memory locations. There is no need
179 to lock them, since they are only unset at the end of a scheduler
183 /* flag set by signal handler to precipitate a context switch */
184 //@cindex context_switch
187 /* if this flag is set as well, give up execution */
188 //@cindex interrupted
191 /* Next thread ID to allocate.
192 * Locks required: sched_mutex
194 //@cindex next_thread_id
195 StgThreadID next_thread_id = 1;
198 * Pointers to the state of the current thread.
199 * Rule of thumb: if CurrentTSO != NULL, then we're running a Haskell
200 * thread. If CurrentTSO == NULL, then we're at the scheduler level.
203 /* The smallest stack size that makes any sense is:
204 * RESERVED_STACK_WORDS (so we can get back from the stack overflow)
205 * + sizeofW(StgStopFrame) (the stg_stop_thread_info frame)
206 * + 1 (the realworld token for an IO thread)
207 * + 1 (the closure to enter)
209 * A thread with this stack will bomb immediately with a stack
210 * overflow, which will increase its stack size.
213 #define MIN_STACK_WORDS (RESERVED_STACK_WORDS + sizeofW(StgStopFrame) + 2)
220 /* This is used in `TSO.h' and gcc 2.96 insists that this variable actually
221 * exists - earlier gccs apparently didn't.
228 void addToBlockedQueue ( StgTSO *tso );
230 static void schedule ( void );
231 void interruptStgRts ( void );
233 static StgTSO * createThread_ ( nat size, rtsBool have_lock, StgInt pri );
235 static StgTSO * createThread_ ( nat size, rtsBool have_lock );
238 static void detectBlackHoles ( void );
241 static void sched_belch(char *s, ...);
244 #if defined(RTS_SUPPORTS_THREADS)
245 /* ToDo: carefully document the invariants that go together
246 * with these synchronisation objects.
248 Mutex sched_mutex = INIT_MUTEX_VAR;
249 Mutex term_mutex = INIT_MUTEX_VAR;
252 static Condition gc_pending_cond = INIT_COND_VAR;
256 #endif /* RTS_SUPPORTS_THREADS */
260 rtsTime TimeOfLastYield;
261 rtsBool emitSchedule = rtsTrue;
265 char *whatNext_strs[] = {
273 char *threadReturnCode_strs[] = {
274 "HeapOverflow", /* might also be StackOverflow */
283 StgTSO * createSparkThread(rtsSpark spark);
284 StgTSO * activateSpark (rtsSpark spark);
288 * The thread state for the main thread.
289 // ToDo: check whether not needed any more
293 #if defined(PAR) || defined(RTS_SUPPORTS_THREADS)
294 static void taskStart(void);
305 //@node Main scheduling loop, Suspend and Resume, Prototypes, Main scheduling code
306 //@subsection Main scheduling loop
308 /* ---------------------------------------------------------------------------
309 Main scheduling loop.
311 We use round-robin scheduling, each thread returning to the
312 scheduler loop when one of these conditions is detected:
315 * timer expires (thread yields)
320 Locking notes: we acquire the scheduler lock once at the beginning
321 of the scheduler loop, and release it when
323 * running a thread, or
324 * waiting for work, or
325 * waiting for a GC to complete.
328 In a GranSim setup this loop iterates over the global event queue.
329 This revolves around the global event queue, which determines what
330 to do next. Therefore, it's more complicated than either the
331 concurrent or the parallel (GUM) setup.
334 GUM iterates over incoming messages.
335 It starts with nothing to do (thus CurrentTSO == END_TSO_QUEUE),
336 and sends out a fish whenever it has nothing to do; in-between
337 doing the actual reductions (shared code below) it processes the
338 incoming messages and deals with delayed operations
339 (see PendingFetches).
340 This is not the ugliest code you could imagine, but it's bloody close.
342 ------------------------------------------------------------------------ */
349 StgThreadReturnCode ret;
357 rtsBool receivedFinish = rtsFalse;
359 nat tp_size, sp_size; // stats only
362 rtsBool was_interrupted = rtsFalse;
364 ACQUIRE_LOCK(&sched_mutex);
366 #if defined(RTS_SUPPORTS_THREADS)
367 waitForWorkCapability(&sched_mutex, &cap, rtsFalse);
369 /* simply initialise it in the non-threaded case */
370 grabCapability(&cap);
374 /* set up first event to get things going */
375 /* ToDo: assign costs for system setup and init MainTSO ! */
376 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
378 CurrentTSO, (StgClosure*)NULL, (rtsSpark*)NULL);
381 fprintf(stderr, "GRAN: Init CurrentTSO (in schedule) = %p\n", CurrentTSO);
382 G_TSO(CurrentTSO, 5));
384 if (RtsFlags.GranFlags.Light) {
385 /* Save current time; GranSim Light only */
386 CurrentTSO->gran.clock = CurrentTime[CurrentProc];
389 event = get_next_event();
391 while (event!=(rtsEvent*)NULL) {
392 /* Choose the processor with the next event */
393 CurrentProc = event->proc;
394 CurrentTSO = event->tso;
398 while (!receivedFinish) { /* set by processMessages */
399 /* when receiving PP_FINISH message */
406 IF_DEBUG(scheduler, printAllThreads());
408 #if defined(RTS_SUPPORTS_THREADS)
409 /* Check to see whether there are any worker threads
410 waiting to deposit external call results. If so,
411 yield our capability */
412 yieldToReturningWorker(&sched_mutex, &cap);
415 /* If we're interrupted (the user pressed ^C, or some other
416 * termination condition occurred), kill all the currently running
420 IF_DEBUG(scheduler, sched_belch("interrupted"));
422 interrupted = rtsFalse;
423 was_interrupted = rtsTrue;
426 /* Go through the list of main threads and wake up any
427 * clients whose computations have finished. ToDo: this
428 * should be done more efficiently without a linear scan
429 * of the main threads list, somehow...
431 #if defined(RTS_SUPPORTS_THREADS)
433 StgMainThread *m, **prev;
434 prev = &main_threads;
435 for (m = main_threads; m != NULL; m = m->link) {
436 switch (m->tso->what_next) {
439 *(m->ret) = (StgClosure *)m->tso->sp[0];
443 broadcastCondition(&m->wakeup);
449 if (m->ret) *(m->ret) = NULL;
451 if (was_interrupted) {
452 m->stat = Interrupted;
456 broadcastCondition(&m->wakeup);
467 #else /* not threaded */
470 /* in GUM do this only on the Main PE */
473 /* If our main thread has finished or been killed, return.
476 StgMainThread *m = main_threads;
477 if (m->tso->what_next == ThreadComplete
478 || m->tso->what_next == ThreadKilled) {
482 main_threads = main_threads->link;
483 if (m->tso->what_next == ThreadComplete) {
484 /* we finished successfully, fill in the return value */
485 if (m->ret) { *(m->ret) = (StgClosure *)m->tso->sp[0]; };
489 if (m->ret) { *(m->ret) = NULL; };
490 if (was_interrupted) {
491 m->stat = Interrupted;
501 /* Top up the run queue from our spark pool. We try to make the
502 * number of threads in the run queue equal to the number of
505 * Disable spark support in SMP for now, non-essential & requires
506 * a little bit of work to make it compile cleanly. -- sof 1/02.
508 #if 0 /* defined(SMP) */
510 nat n = getFreeCapabilities();
511 StgTSO *tso = run_queue_hd;
513 /* Count the run queue */
514 while (n > 0 && tso != END_TSO_QUEUE) {
521 spark = findSpark(rtsFalse);
523 break; /* no more sparks in the pool */
525 /* I'd prefer this to be done in activateSpark -- HWL */
526 /* tricky - it needs to hold the scheduler lock and
527 * not try to re-acquire it -- SDM */
528 createSparkThread(spark);
530 sched_belch("==^^ turning spark of closure %p into a thread",
531 (StgClosure *)spark));
534 /* We need to wake up the other tasks if we just created some
537 if (getFreeCapabilities() - n > 1) {
538 signalCondition( &thread_ready_cond );
543 /* check for signals each time around the scheduler */
544 #ifndef mingw32_TARGET_OS
545 if (signals_pending()) {
546 RELEASE_LOCK(&sched_mutex); /* ToDo: kill */
547 startSignalHandlers();
548 ACQUIRE_LOCK(&sched_mutex);
552 /* Check whether any waiting threads need to be woken up. If the
553 * run queue is empty, and there are no other tasks running, we
554 * can wait indefinitely for something to happen.
555 * ToDo: what if another client comes along & requests another
558 if ( !EMPTY_QUEUE(blocked_queue_hd) || !EMPTY_QUEUE(sleeping_queue) ) {
559 awaitEvent( EMPTY_RUN_QUEUE()
561 && allFreeCapabilities()
565 /* we can be interrupted while waiting for I/O... */
566 if (interrupted) continue;
569 * Detect deadlock: when we have no threads to run, there are no
570 * threads waiting on I/O or sleeping, and all the other tasks are
571 * waiting for work, we must have a deadlock of some description.
573 * We first try to find threads blocked on themselves (ie. black
574 * holes), and generate NonTermination exceptions where necessary.
576 * If no threads are black holed, we have a deadlock situation, so
577 * inform all the main threads.
580 if ( EMPTY_THREAD_QUEUES()
581 #if defined(RTS_SUPPORTS_THREADS)
582 && EMPTY_QUEUE(suspended_ccalling_threads)
585 && allFreeCapabilities()
589 IF_DEBUG(scheduler, sched_belch("deadlocked, forcing major GC..."));
590 #if defined(THREADED_RTS)
591 /* and SMP mode ..? */
592 releaseCapability(cap);
594 // Garbage collection can release some new threads due to
595 // either (a) finalizers or (b) threads resurrected because
596 // they are about to be send BlockedOnDeadMVar. Any threads
597 // thus released will be immediately runnable.
598 GarbageCollect(GetRoots,rtsTrue);
600 if ( !EMPTY_RUN_QUEUE() ) { goto not_deadlocked; }
603 sched_belch("still deadlocked, checking for black holes..."));
606 if ( !EMPTY_RUN_QUEUE() ) { goto not_deadlocked; }
608 #ifndef mingw32_TARGET_OS
609 /* If we have user-installed signal handlers, then wait
610 * for signals to arrive rather then bombing out with a
613 #if defined(RTS_SUPPORTS_THREADS)
614 if ( 0 ) { /* hmm..what to do? Simply stop waiting for
615 a signal with no runnable threads (or I/O
616 suspended ones) leads nowhere quick.
617 For now, simply shut down when we reach this
620 ToDo: define precisely under what conditions
621 the Scheduler should shut down in an MT setting.
624 if ( anyUserHandlers() ) {
627 sched_belch("still deadlocked, waiting for signals..."));
631 // we might be interrupted...
632 if (interrupted) { continue; }
634 if (signals_pending()) {
635 RELEASE_LOCK(&sched_mutex);
636 startSignalHandlers();
637 ACQUIRE_LOCK(&sched_mutex);
639 ASSERT(!EMPTY_RUN_QUEUE());
644 /* Probably a real deadlock. Send the current main thread the
645 * Deadlock exception (or in the SMP build, send *all* main
646 * threads the deadlock exception, since none of them can make
651 #if defined(RTS_SUPPORTS_THREADS)
652 for (m = main_threads; m != NULL; m = m->link) {
653 switch (m->tso->why_blocked) {
654 case BlockedOnBlackHole:
655 raiseAsync(m->tso, (StgClosure *)NonTermination_closure);
657 case BlockedOnException:
659 raiseAsync(m->tso, (StgClosure *)Deadlock_closure);
662 barf("deadlock: main thread blocked in a strange way");
667 switch (m->tso->why_blocked) {
668 case BlockedOnBlackHole:
669 raiseAsync(m->tso, (StgClosure *)NonTermination_closure);
671 case BlockedOnException:
673 raiseAsync(m->tso, (StgClosure *)Deadlock_closure);
676 barf("deadlock: main thread blocked in a strange way");
681 #if defined(RTS_SUPPORTS_THREADS)
682 /* ToDo: revisit conditions (and mechanism) for shutting
683 down a multi-threaded world */
684 IF_DEBUG(scheduler, sched_belch("all done, i think...shutting down."));
685 shutdownHaskellAndExit(0);
691 /* ToDo: add deadlock detection in GUM (similar to SMP) -- HWL */
695 /* If there's a GC pending, don't do anything until it has
699 IF_DEBUG(scheduler,sched_belch("waiting for GC"));
700 waitCondition( &gc_pending_cond, &sched_mutex );
704 #if defined(RTS_SUPPORTS_THREADS)
705 /* block until we've got a thread on the run queue and a free
709 if ( EMPTY_RUN_QUEUE() ) {
710 /* Give up our capability */
711 releaseCapability(cap);
712 IF_DEBUG(scheduler, sched_belch("thread %d: waiting for work", osThreadId()));
713 waitForWorkCapability(&sched_mutex, &cap, rtsTrue);
714 IF_DEBUG(scheduler, sched_belch("thread %d: work now available", osThreadId()));
716 while ( EMPTY_RUN_QUEUE() ) {
717 waitForWorkCapability(&sched_mutex, &cap);
718 IF_DEBUG(scheduler, sched_belch("thread %d: work now available", osThreadId()));
725 if (RtsFlags.GranFlags.Light)
726 GranSimLight_enter_system(event, &ActiveTSO); // adjust ActiveTSO etc
728 /* adjust time based on time-stamp */
729 if (event->time > CurrentTime[CurrentProc] &&
730 event->evttype != ContinueThread)
731 CurrentTime[CurrentProc] = event->time;
733 /* Deal with the idle PEs (may issue FindWork or MoveSpark events) */
734 if (!RtsFlags.GranFlags.Light)
737 IF_DEBUG(gran, fprintf(stderr, "GRAN: switch by event-type\n"));
739 /* main event dispatcher in GranSim */
740 switch (event->evttype) {
741 /* Should just be continuing execution */
743 IF_DEBUG(gran, fprintf(stderr, "GRAN: doing ContinueThread\n"));
744 /* ToDo: check assertion
745 ASSERT(run_queue_hd != (StgTSO*)NULL &&
746 run_queue_hd != END_TSO_QUEUE);
748 /* Ignore ContinueThreads for fetching threads (if synchr comm) */
749 if (!RtsFlags.GranFlags.DoAsyncFetch &&
750 procStatus[CurrentProc]==Fetching) {
751 belch("ghuH: Spurious ContinueThread while Fetching ignored; TSO %d (%p) [PE %d]",
752 CurrentTSO->id, CurrentTSO, CurrentProc);
755 /* Ignore ContinueThreads for completed threads */
756 if (CurrentTSO->what_next == ThreadComplete) {
757 belch("ghuH: found a ContinueThread event for completed thread %d (%p) [PE %d] (ignoring ContinueThread)",
758 CurrentTSO->id, CurrentTSO, CurrentProc);
761 /* Ignore ContinueThreads for threads that are being migrated */
762 if (PROCS(CurrentTSO)==Nowhere) {
763 belch("ghuH: trying to run the migrating TSO %d (%p) [PE %d] (ignoring ContinueThread)",
764 CurrentTSO->id, CurrentTSO, CurrentProc);
767 /* The thread should be at the beginning of the run queue */
768 if (CurrentTSO!=run_queue_hds[CurrentProc]) {
769 belch("ghuH: TSO %d (%p) [PE %d] is not at the start of the run_queue when doing a ContinueThread",
770 CurrentTSO->id, CurrentTSO, CurrentProc);
771 break; // run the thread anyway
774 new_event(proc, proc, CurrentTime[proc],
776 (StgTSO*)NULL, (StgClosure*)NULL, (rtsSpark*)NULL);
778 */ /* Catches superfluous CONTINUEs -- should be unnecessary */
779 break; // now actually run the thread; DaH Qu'vam yImuHbej
782 do_the_fetchnode(event);
783 goto next_thread; /* handle next event in event queue */
786 do_the_globalblock(event);
787 goto next_thread; /* handle next event in event queue */
790 do_the_fetchreply(event);
791 goto next_thread; /* handle next event in event queue */
793 case UnblockThread: /* Move from the blocked queue to the tail of */
794 do_the_unblock(event);
795 goto next_thread; /* handle next event in event queue */
797 case ResumeThread: /* Move from the blocked queue to the tail of */
798 /* the runnable queue ( i.e. Qu' SImqa'lu') */
799 event->tso->gran.blocktime +=
800 CurrentTime[CurrentProc] - event->tso->gran.blockedat;
801 do_the_startthread(event);
802 goto next_thread; /* handle next event in event queue */
805 do_the_startthread(event);
806 goto next_thread; /* handle next event in event queue */
809 do_the_movethread(event);
810 goto next_thread; /* handle next event in event queue */
813 do_the_movespark(event);
814 goto next_thread; /* handle next event in event queue */
817 do_the_findwork(event);
818 goto next_thread; /* handle next event in event queue */
821 barf("Illegal event type %u\n", event->evttype);
824 /* This point was scheduler_loop in the old RTS */
826 IF_DEBUG(gran, belch("GRAN: after main switch"));
828 TimeOfLastEvent = CurrentTime[CurrentProc];
829 TimeOfNextEvent = get_time_of_next_event();
830 IgnoreEvents=(TimeOfNextEvent==0); // HWL HACK
831 // CurrentTSO = ThreadQueueHd;
833 IF_DEBUG(gran, belch("GRAN: time of next event is: %ld",
836 if (RtsFlags.GranFlags.Light)
837 GranSimLight_leave_system(event, &ActiveTSO);
839 EndOfTimeSlice = CurrentTime[CurrentProc]+RtsFlags.GranFlags.time_slice;
842 belch("GRAN: end of time-slice is %#lx", EndOfTimeSlice));
844 /* in a GranSim setup the TSO stays on the run queue */
846 /* Take a thread from the run queue. */
847 t = POP_RUN_QUEUE(); // take_off_run_queue(t);
850 fprintf(stderr, "GRAN: About to run current thread, which is\n");
853 context_switch = 0; // turned on via GranYield, checking events and time slice
856 DumpGranEvent(GR_SCHEDULE, t));
858 procStatus[CurrentProc] = Busy;
861 if (PendingFetches != END_BF_QUEUE) {
865 /* ToDo: phps merge with spark activation above */
866 /* check whether we have local work and send requests if we have none */
867 if (EMPTY_RUN_QUEUE()) { /* no runnable threads */
868 /* :-[ no local threads => look out for local sparks */
869 /* the spark pool for the current PE */
870 pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable
871 if (advisory_thread_count < RtsFlags.ParFlags.maxThreads &&
872 pool->hd < pool->tl) {
874 * ToDo: add GC code check that we really have enough heap afterwards!!
876 * If we're here (no runnable threads) and we have pending
877 * sparks, we must have a space problem. Get enough space
878 * to turn one of those pending sparks into a
882 spark = findSpark(rtsFalse); /* get a spark */
883 if (spark != (rtsSpark) NULL) {
884 tso = activateSpark(spark); /* turn the spark into a thread */
885 IF_PAR_DEBUG(schedule,
886 belch("==== schedule: Created TSO %d (%p); %d threads active",
887 tso->id, tso, advisory_thread_count));
889 if (tso==END_TSO_QUEUE) { /* failed to activate spark->back to loop */
890 belch("==^^ failed to activate spark");
892 } /* otherwise fall through & pick-up new tso */
894 IF_PAR_DEBUG(verbose,
895 belch("==^^ no local sparks (spark pool contains only NFs: %d)",
896 spark_queue_len(pool)));
901 /* If we still have no work we need to send a FISH to get a spark
904 if (EMPTY_RUN_QUEUE()) {
905 /* =8-[ no local sparks => look for work on other PEs */
907 * We really have absolutely no work. Send out a fish
908 * (there may be some out there already), and wait for
909 * something to arrive. We clearly can't run any threads
910 * until a SCHEDULE or RESUME arrives, and so that's what
911 * we're hoping to see. (Of course, we still have to
912 * respond to other types of messages.)
914 TIME now = msTime() /*CURRENT_TIME*/;
915 IF_PAR_DEBUG(verbose,
916 belch("-- now=%ld", now));
917 IF_PAR_DEBUG(verbose,
918 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
919 (last_fish_arrived_at!=0 &&
920 last_fish_arrived_at+RtsFlags.ParFlags.fishDelay > now)) {
921 belch("--$$ delaying FISH until %ld (last fish %ld, delay %ld, now %ld)",
922 last_fish_arrived_at+RtsFlags.ParFlags.fishDelay,
923 last_fish_arrived_at,
924 RtsFlags.ParFlags.fishDelay, now);
927 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
928 (last_fish_arrived_at==0 ||
929 (last_fish_arrived_at+RtsFlags.ParFlags.fishDelay <= now))) {
930 /* outstandingFishes is set in sendFish, processFish;
931 avoid flooding system with fishes via delay */
933 sendFish(pe, mytid, NEW_FISH_AGE, NEW_FISH_HISTORY,
936 // Global statistics: count no. of fishes
937 if (RtsFlags.ParFlags.ParStats.Global &&
938 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
939 globalParStats.tot_fish_mess++;
943 receivedFinish = processMessages();
946 } else if (PacketsWaiting()) { /* Look for incoming messages */
947 receivedFinish = processMessages();
950 /* Now we are sure that we have some work available */
951 ASSERT(run_queue_hd != END_TSO_QUEUE);
953 /* Take a thread from the run queue, if we have work */
954 t = POP_RUN_QUEUE(); // take_off_run_queue(END_TSO_QUEUE);
955 IF_DEBUG(sanity,checkTSO(t));
957 /* ToDo: write something to the log-file
958 if (RTSflags.ParFlags.granSimStats && !sameThread)
959 DumpGranEvent(GR_SCHEDULE, RunnableThreadsHd);
963 /* the spark pool for the current PE */
964 pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable
967 belch("--=^ %d threads, %d sparks on [%#x]",
968 run_queue_len(), spark_queue_len(pool), CURRENT_PROC));
971 if (0 && RtsFlags.ParFlags.ParStats.Full &&
972 t && LastTSO && t->id != LastTSO->id &&
973 LastTSO->why_blocked == NotBlocked &&
974 LastTSO->what_next != ThreadComplete) {
975 // if previously scheduled TSO not blocked we have to record the context switch
976 DumpVeryRawGranEvent(TimeOfLastYield, CURRENT_PROC, CURRENT_PROC,
977 GR_DESCHEDULE, LastTSO, (StgClosure *)NULL, 0, 0);
980 if (RtsFlags.ParFlags.ParStats.Full &&
981 (emitSchedule /* forced emit */ ||
982 (t && LastTSO && t->id != LastTSO->id))) {
984 we are running a different TSO, so write a schedule event to log file
985 NB: If we use fair scheduling we also have to write a deschedule
986 event for LastTSO; with unfair scheduling we know that the
987 previous tso has blocked whenever we switch to another tso, so
988 we don't need it in GUM for now
990 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
991 GR_SCHEDULE, t, (StgClosure *)NULL, 0, 0);
992 emitSchedule = rtsFalse;
996 #else /* !GRAN && !PAR */
998 /* grab a thread from the run queue */
999 ASSERT(run_queue_hd != END_TSO_QUEUE);
1000 t = POP_RUN_QUEUE();
1001 // Sanity check the thread we're about to run. This can be
1002 // expensive if there is lots of thread switching going on...
1003 IF_DEBUG(sanity,checkTSO(t));
1006 cap->r.rCurrentTSO = t;
1008 /* context switches are now initiated by the timer signal, unless
1009 * the user specified "context switch as often as possible", with
1014 RtsFlags.ProfFlags.profileInterval == 0 ||
1016 (RtsFlags.ConcFlags.ctxtSwitchTicks == 0
1017 && (run_queue_hd != END_TSO_QUEUE
1018 || blocked_queue_hd != END_TSO_QUEUE
1019 || sleeping_queue != END_TSO_QUEUE)))
1024 RELEASE_LOCK(&sched_mutex);
1026 IF_DEBUG(scheduler, sched_belch("-->> Running TSO %ld (%p) %s ...",
1027 t->id, t, whatNext_strs[t->what_next]));
1030 startHeapProfTimer();
1033 /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
1034 /* Run the current thread
1036 switch (cap->r.rCurrentTSO->what_next) {
1038 case ThreadComplete:
1039 /* Thread already finished, return to scheduler. */
1040 ret = ThreadFinished;
1042 case ThreadEnterGHC:
1043 ret = StgRun((StgFunPtr) stg_enterStackTop, &cap->r);
1046 ret = StgRun((StgFunPtr) stg_returnToStackTop, &cap->r);
1048 case ThreadEnterInterp:
1049 ret = interpretBCO(cap);
1052 barf("schedule: invalid what_next field");
1054 /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
1056 /* Costs for the scheduler are assigned to CCS_SYSTEM */
1058 stopHeapProfTimer();
1062 ACQUIRE_LOCK(&sched_mutex);
1065 IF_DEBUG(scheduler,fprintf(stderr,"scheduler (task %ld): ", osThreadId()););
1066 #elif !defined(GRAN) && !defined(PAR)
1067 IF_DEBUG(scheduler,fprintf(stderr,"scheduler: "););
1069 t = cap->r.rCurrentTSO;
1072 /* HACK 675: if the last thread didn't yield, make sure to print a
1073 SCHEDULE event to the log file when StgRunning the next thread, even
1074 if it is the same one as before */
1076 TimeOfLastYield = CURRENT_TIME;
1082 IF_DEBUG(gran, DumpGranEvent(GR_DESCHEDULE, t));
1083 globalGranStats.tot_heapover++;
1085 globalParStats.tot_heapover++;
1088 // did the task ask for a large block?
1089 if (cap->r.rHpAlloc > BLOCK_SIZE_W) {
1090 // if so, get one and push it on the front of the nursery.
1094 blocks = (nat)BLOCK_ROUND_UP(cap->r.rHpAlloc * sizeof(W_)) / BLOCK_SIZE;
1096 IF_DEBUG(scheduler,belch("--<< thread %ld (%p; %s) stopped: requesting a large block (size %d)",
1098 whatNext_strs[t->what_next], blocks));
1100 // don't do this if it would push us over the
1101 // alloc_blocks_lim limit; we'll GC first.
1102 if (alloc_blocks + blocks < alloc_blocks_lim) {
1104 alloc_blocks += blocks;
1105 bd = allocGroup( blocks );
1107 // link the new group into the list
1108 bd->link = cap->r.rCurrentNursery;
1109 bd->u.back = cap->r.rCurrentNursery->u.back;
1110 if (cap->r.rCurrentNursery->u.back != NULL) {
1111 cap->r.rCurrentNursery->u.back->link = bd;
1113 ASSERT(g0s0->blocks == cap->r.rCurrentNursery &&
1114 g0s0->blocks == cap->r.rNursery);
1115 cap->r.rNursery = g0s0->blocks = bd;
1117 cap->r.rCurrentNursery->u.back = bd;
1119 // initialise it as a nursery block
1123 bd->free = bd->start;
1125 // don't forget to update the block count in g0s0.
1126 g0s0->n_blocks += blocks;
1127 ASSERT(countBlocks(g0s0->blocks) == g0s0->n_blocks);
1129 // now update the nursery to point to the new block
1130 cap->r.rCurrentNursery = bd;
1132 // we might be unlucky and have another thread get on the
1133 // run queue before us and steal the large block, but in that
1134 // case the thread will just end up requesting another large
1136 PUSH_ON_RUN_QUEUE(t);
1141 /* make all the running tasks block on a condition variable,
1142 * maybe set context_switch and wait till they all pile in,
1143 * then have them wait on a GC condition variable.
1145 IF_DEBUG(scheduler,belch("--<< thread %ld (%p; %s) stopped: HeapOverflow",
1146 t->id, t, whatNext_strs[t->what_next]));
1149 ASSERT(!is_on_queue(t,CurrentProc));
1151 /* Currently we emit a DESCHEDULE event before GC in GUM.
1152 ToDo: either add separate event to distinguish SYSTEM time from rest
1153 or just nuke this DESCHEDULE (and the following SCHEDULE) */
1154 if (0 && RtsFlags.ParFlags.ParStats.Full) {
1155 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
1156 GR_DESCHEDULE, t, (StgClosure *)NULL, 0, 0);
1157 emitSchedule = rtsTrue;
1161 ready_to_gc = rtsTrue;
1162 context_switch = 1; /* stop other threads ASAP */
1163 PUSH_ON_RUN_QUEUE(t);
1164 /* actual GC is done at the end of the while loop */
1170 DumpGranEvent(GR_DESCHEDULE, t));
1171 globalGranStats.tot_stackover++;
1174 // DumpGranEvent(GR_DESCHEDULE, t);
1175 globalParStats.tot_stackover++;
1177 IF_DEBUG(scheduler,belch("--<< thread %ld (%p; %s) stopped, StackOverflow",
1178 t->id, t, whatNext_strs[t->what_next]));
1179 /* just adjust the stack for this thread, then pop it back
1185 /* enlarge the stack */
1186 StgTSO *new_t = threadStackOverflow(t);
1188 /* This TSO has moved, so update any pointers to it from the
1189 * main thread stack. It better not be on any other queues...
1190 * (it shouldn't be).
1192 for (m = main_threads; m != NULL; m = m->link) {
1197 threadPaused(new_t);
1198 PUSH_ON_RUN_QUEUE(new_t);
1202 case ThreadYielding:
1205 DumpGranEvent(GR_DESCHEDULE, t));
1206 globalGranStats.tot_yields++;
1209 // DumpGranEvent(GR_DESCHEDULE, t);
1210 globalParStats.tot_yields++;
1212 /* put the thread back on the run queue. Then, if we're ready to
1213 * GC, check whether this is the last task to stop. If so, wake
1214 * up the GC thread. getThread will block during a GC until the
1218 if (t->what_next == ThreadEnterInterp) {
1219 /* ToDo: or maybe a timer expired when we were in Hugs?
1220 * or maybe someone hit ctrl-C
1222 belch("--<< thread %ld (%p; %s) stopped to switch to Hugs",
1223 t->id, t, whatNext_strs[t->what_next]);
1225 belch("--<< thread %ld (%p; %s) stopped, yielding",
1226 t->id, t, whatNext_strs[t->what_next]);
1233 //belch("&& Doing sanity check on yielding TSO %ld.", t->id);
1235 ASSERT(t->link == END_TSO_QUEUE);
1237 ASSERT(!is_on_queue(t,CurrentProc));
1240 //belch("&& Doing sanity check on all ThreadQueues (and their TSOs).");
1241 checkThreadQsSanity(rtsTrue));
1244 if (RtsFlags.ParFlags.doFairScheduling) {
1245 /* this does round-robin scheduling; good for concurrency */
1246 APPEND_TO_RUN_QUEUE(t);
1248 /* this does unfair scheduling; good for parallelism */
1249 PUSH_ON_RUN_QUEUE(t);
1252 /* this does round-robin scheduling; good for concurrency */
1253 APPEND_TO_RUN_QUEUE(t);
1256 /* add a ContinueThread event to actually process the thread */
1257 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1259 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1261 belch("GRAN: eventq and runnableq after adding yielded thread to queue again:");
1270 belch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: ",
1271 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)));
1272 if (t->block_info.closure!=(StgClosure*)NULL) print_bq(t->block_info.closure));
1274 // ??? needed; should emit block before
1276 DumpGranEvent(GR_DESCHEDULE, t));
1277 prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
1280 ASSERT(procStatus[CurrentProc]==Busy ||
1281 ((procStatus[CurrentProc]==Fetching) &&
1282 (t->block_info.closure!=(StgClosure*)NULL)));
1283 if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
1284 !(!RtsFlags.GranFlags.DoAsyncFetch &&
1285 procStatus[CurrentProc]==Fetching))
1286 procStatus[CurrentProc] = Idle;
1290 belch("--<< thread %ld (%p; %s) stopped, blocking on node %p with BQ: ",
1291 t->id, t, whatNext_strs[t->what_next], t->block_info.closure));
1294 if (t->block_info.closure!=(StgClosure*)NULL)
1295 print_bq(t->block_info.closure));
1297 /* Send a fetch (if BlockedOnGA) and dump event to log file */
1300 /* whatever we schedule next, we must log that schedule */
1301 emitSchedule = rtsTrue;
1304 /* don't need to do anything. Either the thread is blocked on
1305 * I/O, in which case we'll have called addToBlockedQueue
1306 * previously, or it's blocked on an MVar or Blackhole, in which
1307 * case it'll be on the relevant queue already.
1310 fprintf(stderr, "--<< thread %d (%p) stopped: ", t->id, t);
1311 printThreadBlockage(t);
1312 fprintf(stderr, "\n"));
1314 /* Only for dumping event to log file
1315 ToDo: do I need this in GranSim, too?
1322 case ThreadFinished:
1323 /* Need to check whether this was a main thread, and if so, signal
1324 * the task that started it with the return value. If we have no
1325 * more main threads, we probably need to stop all the tasks until
1328 /* We also end up here if the thread kills itself with an
1329 * uncaught exception, see Exception.hc.
1331 IF_DEBUG(scheduler,belch("--++ thread %d (%p) finished", t->id, t));
1333 endThread(t, CurrentProc); // clean-up the thread
1335 /* For now all are advisory -- HWL */
1336 //if(t->priority==AdvisoryPriority) ??
1337 advisory_thread_count--;
1340 if(t->dist.priority==RevalPriority)
1344 if (RtsFlags.ParFlags.ParStats.Full &&
1345 !RtsFlags.ParFlags.ParStats.Suppressed)
1346 DumpEndEvent(CURRENT_PROC, t, rtsFalse /* not mandatory */);
1351 barf("schedule: invalid thread return code %d", (int)ret);
1355 if (RtsFlags.ProfFlags.profileInterval==0 || performHeapProfile) {
1356 GarbageCollect(GetRoots, rtsTrue);
1358 performHeapProfile = rtsFalse;
1359 ready_to_gc = rtsFalse; // we already GC'd
1365 && allFreeCapabilities()
1368 /* everybody back, start the GC.
1369 * Could do it in this thread, or signal a condition var
1370 * to do it in another thread. Either way, we need to
1371 * broadcast on gc_pending_cond afterward.
1373 #if defined(RTS_SUPPORTS_THREADS)
1374 IF_DEBUG(scheduler,sched_belch("doing GC"));
1376 GarbageCollect(GetRoots,rtsFalse);
1377 ready_to_gc = rtsFalse;
1379 broadcastCondition(&gc_pending_cond);
1382 /* add a ContinueThread event to continue execution of current thread */
1383 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1385 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1387 fprintf(stderr, "GRAN: eventq and runnableq after Garbage collection:\n");
1395 IF_GRAN_DEBUG(unused,
1396 print_eventq(EventHd));
1398 event = get_next_event();
1401 /* ToDo: wait for next message to arrive rather than busy wait */
1404 } /* end of while(1) */
1406 IF_PAR_DEBUG(verbose,
1407 belch("== Leaving schedule() after having received Finish"));
1410 /* ---------------------------------------------------------------------------
1411 * Singleton fork(). Do not copy any running threads.
1412 * ------------------------------------------------------------------------- */
1414 StgInt forkProcess(StgTSO* tso) {
1416 #ifndef mingw32_TARGET_OS
1420 IF_DEBUG(scheduler,sched_belch("forking!"));
1423 if (pid) { /* parent */
1425 /* just return the pid */
1427 } else { /* child */
1428 /* wipe all other threads */
1430 tso->link = END_TSO_QUEUE;
1432 /* DO NOT TOUCH THE QUEUES directly because most of the code around
1433 us is picky about finding the threat still in its queue when
1434 handling the deleteThread() */
1436 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
1438 if (t->id != tso->id) {
1445 barf("forkProcess#: primop not implemented for mingw32, sorry!");
1447 #endif /* mingw32 */
1450 /* ---------------------------------------------------------------------------
1451 * deleteAllThreads(): kill all the live threads.
1453 * This is used when we catch a user interrupt (^C), before performing
1454 * any necessary cleanups and running finalizers.
1456 * Locks: sched_mutex held.
1457 * ------------------------------------------------------------------------- */
1459 void deleteAllThreads ( void )
1462 IF_DEBUG(scheduler,sched_belch("deleting all threads"));
1463 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
1464 next = t->global_link;
1467 run_queue_hd = run_queue_tl = END_TSO_QUEUE;
1468 blocked_queue_hd = blocked_queue_tl = END_TSO_QUEUE;
1469 sleeping_queue = END_TSO_QUEUE;
1472 /* startThread and insertThread are now in GranSim.c -- HWL */
1475 //@node Suspend and Resume, Run queue code, Main scheduling loop, Main scheduling code
1476 //@subsection Suspend and Resume
1478 /* ---------------------------------------------------------------------------
1479 * Suspending & resuming Haskell threads.
1481 * When making a "safe" call to C (aka _ccall_GC), the task gives back
1482 * its capability before calling the C function. This allows another
1483 * task to pick up the capability and carry on running Haskell
1484 * threads. It also means that if the C call blocks, it won't lock
1487 * The Haskell thread making the C call is put to sleep for the
1488 * duration of the call, on the susepended_ccalling_threads queue. We
1489 * give out a token to the task, which it can use to resume the thread
1490 * on return from the C function.
1491 * ------------------------------------------------------------------------- */
1494 suspendThread( StgRegTable *reg,
1496 #if !defined(RTS_SUPPORTS_THREADS) && !defined(DEBUG)
1504 /* assume that *reg is a pointer to the StgRegTable part
1507 cap = (Capability *)((void *)reg - sizeof(StgFunTable));
1509 ACQUIRE_LOCK(&sched_mutex);
1512 sched_belch("thread %d did a _ccall_gc (is_concurrent: %d)", cap->r.rCurrentTSO->id,concCall));
1514 threadPaused(cap->r.rCurrentTSO);
1515 cap->r.rCurrentTSO->link = suspended_ccalling_threads;
1516 suspended_ccalling_threads = cap->r.rCurrentTSO;
1518 #if defined(RTS_SUPPORTS_THREADS)
1519 cap->r.rCurrentTSO->why_blocked = BlockedOnCCall;
1522 /* Use the thread ID as the token; it should be unique */
1523 tok = cap->r.rCurrentTSO->id;
1525 /* Hand back capability */
1526 releaseCapability(cap);
1528 #if defined(RTS_SUPPORTS_THREADS)
1529 /* Preparing to leave the RTS, so ensure there's a native thread/task
1530 waiting to take over.
1532 ToDo: optimise this and only create a new task if there's a need
1533 for one (i.e., if there's only one Concurrent Haskell thread alive,
1534 there's no need to create a new task).
1536 IF_DEBUG(scheduler, sched_belch("worker thread (%d): leaving RTS", tok));
1538 startTask(taskStart);
1542 /* Other threads _might_ be available for execution; signal this */
1544 RELEASE_LOCK(&sched_mutex);
1549 resumeThread( StgInt tok,
1551 #if !defined(RTS_SUPPORTS_THREADS)
1556 StgTSO *tso, **prev;
1559 #if defined(RTS_SUPPORTS_THREADS)
1560 /* Wait for permission to re-enter the RTS with the result. */
1562 ACQUIRE_LOCK(&sched_mutex);
1563 grabReturnCapability(&sched_mutex, &cap);
1565 grabCapability(&cap);
1568 grabCapability(&cap);
1571 /* Remove the thread off of the suspended list */
1572 prev = &suspended_ccalling_threads;
1573 for (tso = suspended_ccalling_threads;
1574 tso != END_TSO_QUEUE;
1575 prev = &tso->link, tso = tso->link) {
1576 if (tso->id == (StgThreadID)tok) {
1581 if (tso == END_TSO_QUEUE) {
1582 barf("resumeThread: thread not found");
1584 tso->link = END_TSO_QUEUE;
1585 /* Reset blocking status */
1586 tso->why_blocked = NotBlocked;
1588 cap->r.rCurrentTSO = tso;
1589 RELEASE_LOCK(&sched_mutex);
1594 /* ---------------------------------------------------------------------------
1596 * ------------------------------------------------------------------------ */
1597 static void unblockThread(StgTSO *tso);
1599 /* ---------------------------------------------------------------------------
1600 * Comparing Thread ids.
1602 * This is used from STG land in the implementation of the
1603 * instances of Eq/Ord for ThreadIds.
1604 * ------------------------------------------------------------------------ */
1606 int cmp_thread(const StgTSO *tso1, const StgTSO *tso2)
1608 StgThreadID id1 = tso1->id;
1609 StgThreadID id2 = tso2->id;
1611 if (id1 < id2) return (-1);
1612 if (id1 > id2) return 1;
1616 /* ---------------------------------------------------------------------------
1617 * Fetching the ThreadID from an StgTSO.
1619 * This is used in the implementation of Show for ThreadIds.
1620 * ------------------------------------------------------------------------ */
1621 int rts_getThreadId(const StgTSO *tso)
1627 void labelThread(StgTSO *tso, char *label)
1632 /* Caveat: Once set, you can only set the thread name to "" */
1633 len = strlen(label)+1;
1634 buf = realloc(tso->label,len);
1636 fprintf(stderr,"insufficient memory for labelThread!\n");
1639 strncpy(buf,label,len);
1644 /* ---------------------------------------------------------------------------
1645 Create a new thread.
1647 The new thread starts with the given stack size. Before the
1648 scheduler can run, however, this thread needs to have a closure
1649 (and possibly some arguments) pushed on its stack. See
1650 pushClosure() in Schedule.h.
1652 createGenThread() and createIOThread() (in SchedAPI.h) are
1653 convenient packaged versions of this function.
1655 currently pri (priority) is only used in a GRAN setup -- HWL
1656 ------------------------------------------------------------------------ */
1657 //@cindex createThread
1659 /* currently pri (priority) is only used in a GRAN setup -- HWL */
1661 createThread(nat stack_size, StgInt pri)
1663 return createThread_(stack_size, rtsFalse, pri);
1667 createThread_(nat size, rtsBool have_lock, StgInt pri)
1671 createThread(nat stack_size)
1673 return createThread_(stack_size, rtsFalse);
1677 createThread_(nat size, rtsBool have_lock)
1684 /* First check whether we should create a thread at all */
1686 /* check that no more than RtsFlags.ParFlags.maxThreads threads are created */
1687 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads) {
1689 belch("{createThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
1690 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
1691 return END_TSO_QUEUE;
1697 ASSERT(!RtsFlags.GranFlags.Light || CurrentProc==0);
1700 // ToDo: check whether size = stack_size - TSO_STRUCT_SIZEW
1702 /* catch ridiculously small stack sizes */
1703 if (size < MIN_STACK_WORDS + TSO_STRUCT_SIZEW) {
1704 size = MIN_STACK_WORDS + TSO_STRUCT_SIZEW;
1707 stack_size = size - TSO_STRUCT_SIZEW;
1709 tso = (StgTSO *)allocate(size);
1710 TICK_ALLOC_TSO(stack_size, 0);
1712 SET_HDR(tso, &stg_TSO_info, CCS_SYSTEM);
1714 SET_GRAN_HDR(tso, ThisPE);
1716 tso->what_next = ThreadEnterGHC;
1722 /* tso->id needs to be unique. For now we use a heavyweight mutex to
1723 * protect the increment operation on next_thread_id.
1724 * In future, we could use an atomic increment instead.
1727 if (!have_lock) { ACQUIRE_LOCK(&sched_mutex); }
1729 tso->id = next_thread_id++;
1731 if (!have_lock) { RELEASE_LOCK(&sched_mutex); }
1734 tso->why_blocked = NotBlocked;
1735 tso->blocked_exceptions = NULL;
1737 tso->stack_size = stack_size;
1738 tso->max_stack_size = round_to_mblocks(RtsFlags.GcFlags.maxStkSize)
1740 tso->sp = (P_)&(tso->stack) + stack_size;
1743 tso->prof.CCCS = CCS_MAIN;
1746 /* put a stop frame on the stack */
1747 tso->sp -= sizeofW(StgStopFrame);
1748 SET_HDR((StgClosure*)tso->sp,(StgInfoTable *)&stg_stop_thread_info,CCS_SYSTEM);
1749 tso->su = (StgUpdateFrame*)tso->sp;
1753 tso->link = END_TSO_QUEUE;
1754 /* uses more flexible routine in GranSim */
1755 insertThread(tso, CurrentProc);
1757 /* In a non-GranSim setup the pushing of a TSO onto the runq is separated
1763 if (RtsFlags.GranFlags.GranSimStats.Full)
1764 DumpGranEvent(GR_START,tso);
1766 if (RtsFlags.ParFlags.ParStats.Full)
1767 DumpGranEvent(GR_STARTQ,tso);
1768 /* HACk to avoid SCHEDULE
1772 /* Link the new thread on the global thread list.
1774 tso->global_link = all_threads;
1778 tso->dist.priority = MandatoryPriority; //by default that is...
1782 tso->gran.pri = pri;
1784 tso->gran.magic = TSO_MAGIC; // debugging only
1786 tso->gran.sparkname = 0;
1787 tso->gran.startedat = CURRENT_TIME;
1788 tso->gran.exported = 0;
1789 tso->gran.basicblocks = 0;
1790 tso->gran.allocs = 0;
1791 tso->gran.exectime = 0;
1792 tso->gran.fetchtime = 0;
1793 tso->gran.fetchcount = 0;
1794 tso->gran.blocktime = 0;
1795 tso->gran.blockcount = 0;
1796 tso->gran.blockedat = 0;
1797 tso->gran.globalsparks = 0;
1798 tso->gran.localsparks = 0;
1799 if (RtsFlags.GranFlags.Light)
1800 tso->gran.clock = Now; /* local clock */
1802 tso->gran.clock = 0;
1804 IF_DEBUG(gran,printTSO(tso));
1807 tso->par.magic = TSO_MAGIC; // debugging only
1809 tso->par.sparkname = 0;
1810 tso->par.startedat = CURRENT_TIME;
1811 tso->par.exported = 0;
1812 tso->par.basicblocks = 0;
1813 tso->par.allocs = 0;
1814 tso->par.exectime = 0;
1815 tso->par.fetchtime = 0;
1816 tso->par.fetchcount = 0;
1817 tso->par.blocktime = 0;
1818 tso->par.blockcount = 0;
1819 tso->par.blockedat = 0;
1820 tso->par.globalsparks = 0;
1821 tso->par.localsparks = 0;
1825 globalGranStats.tot_threads_created++;
1826 globalGranStats.threads_created_on_PE[CurrentProc]++;
1827 globalGranStats.tot_sq_len += spark_queue_len(CurrentProc);
1828 globalGranStats.tot_sq_probes++;
1830 // collect parallel global statistics (currently done together with GC stats)
1831 if (RtsFlags.ParFlags.ParStats.Global &&
1832 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
1833 //fprintf(stderr, "Creating thread %d @ %11.2f\n", tso->id, usertime());
1834 globalParStats.tot_threads_created++;
1840 belch("==__ schedule: Created TSO %d (%p);",
1841 CurrentProc, tso, tso->id));
1843 IF_PAR_DEBUG(verbose,
1844 belch("==__ schedule: Created TSO %d (%p); %d threads active",
1845 tso->id, tso, advisory_thread_count));
1847 IF_DEBUG(scheduler,sched_belch("created thread %ld, stack size = %lx words",
1848 tso->id, tso->stack_size));
1855 all parallel thread creation calls should fall through the following routine.
1858 createSparkThread(rtsSpark spark)
1860 ASSERT(spark != (rtsSpark)NULL);
1861 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads)
1863 barf("{createSparkThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
1864 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
1865 return END_TSO_QUEUE;
1869 tso = createThread_(RtsFlags.GcFlags.initialStkSize, rtsTrue);
1870 if (tso==END_TSO_QUEUE)
1871 barf("createSparkThread: Cannot create TSO");
1873 tso->priority = AdvisoryPriority;
1875 pushClosure(tso,spark);
1876 PUSH_ON_RUN_QUEUE(tso);
1877 advisory_thread_count++;
1884 Turn a spark into a thread.
1885 ToDo: fix for SMP (needs to acquire SCHED_MUTEX!)
1888 //@cindex activateSpark
1890 activateSpark (rtsSpark spark)
1894 tso = createSparkThread(spark);
1895 if (RtsFlags.ParFlags.ParStats.Full) {
1896 //ASSERT(run_queue_hd == END_TSO_QUEUE); // I think ...
1897 IF_PAR_DEBUG(verbose,
1898 belch("==^^ activateSpark: turning spark of closure %p (%s) into a thread",
1899 (StgClosure *)spark, info_type((StgClosure *)spark)));
1901 // ToDo: fwd info on local/global spark to thread -- HWL
1902 // tso->gran.exported = spark->exported;
1903 // tso->gran.locked = !spark->global;
1904 // tso->gran.sparkname = spark->name;
1910 /* ---------------------------------------------------------------------------
1913 * scheduleThread puts a thread on the head of the runnable queue.
1914 * This will usually be done immediately after a thread is created.
1915 * The caller of scheduleThread must create the thread using e.g.
1916 * createThread and push an appropriate closure
1917 * on this thread's stack before the scheduler is invoked.
1918 * ------------------------------------------------------------------------ */
1920 static void scheduleThread_ (StgTSO* tso, rtsBool createTask);
1923 scheduleThread_(StgTSO *tso
1924 , rtsBool createTask
1925 #if !defined(THREADED_RTS)
1930 ACQUIRE_LOCK(&sched_mutex);
1932 /* Put the new thread on the head of the runnable queue. The caller
1933 * better push an appropriate closure on this thread's stack
1934 * beforehand. In the SMP case, the thread may start running as
1935 * soon as we release the scheduler lock below.
1937 PUSH_ON_RUN_QUEUE(tso);
1938 #if defined(THREADED_RTS)
1939 /* If main() is scheduling a thread, don't bother creating a
1943 startTask(taskStart);
1949 IF_DEBUG(scheduler,printTSO(tso));
1951 RELEASE_LOCK(&sched_mutex);
1954 void scheduleThread(StgTSO* tso)
1956 return scheduleThread_(tso, rtsFalse);
1959 void scheduleExtThread(StgTSO* tso)
1961 return scheduleThread_(tso, rtsTrue);
1964 /* ---------------------------------------------------------------------------
1967 * Initialise the scheduler. This resets all the queues - if the
1968 * queues contained any threads, they'll be garbage collected at the
1971 * ------------------------------------------------------------------------ */
1975 term_handler(int sig STG_UNUSED)
1978 ACQUIRE_LOCK(&term_mutex);
1980 RELEASE_LOCK(&term_mutex);
1991 for (i=0; i<=MAX_PROC; i++) {
1992 run_queue_hds[i] = END_TSO_QUEUE;
1993 run_queue_tls[i] = END_TSO_QUEUE;
1994 blocked_queue_hds[i] = END_TSO_QUEUE;
1995 blocked_queue_tls[i] = END_TSO_QUEUE;
1996 ccalling_threadss[i] = END_TSO_QUEUE;
1997 sleeping_queue = END_TSO_QUEUE;
2000 run_queue_hd = END_TSO_QUEUE;
2001 run_queue_tl = END_TSO_QUEUE;
2002 blocked_queue_hd = END_TSO_QUEUE;
2003 blocked_queue_tl = END_TSO_QUEUE;
2004 sleeping_queue = END_TSO_QUEUE;
2007 suspended_ccalling_threads = END_TSO_QUEUE;
2009 main_threads = NULL;
2010 all_threads = END_TSO_QUEUE;
2015 RtsFlags.ConcFlags.ctxtSwitchTicks =
2016 RtsFlags.ConcFlags.ctxtSwitchTime / TICK_MILLISECS;
2018 #if defined(RTS_SUPPORTS_THREADS)
2019 /* Initialise the mutex and condition variables used by
2021 initMutex(&sched_mutex);
2022 initMutex(&term_mutex);
2024 initCondition(&thread_ready_cond);
2028 initCondition(&gc_pending_cond);
2031 #if defined(RTS_SUPPORTS_THREADS)
2032 ACQUIRE_LOCK(&sched_mutex);
2035 /* Install the SIGHUP handler */
2038 struct sigaction action,oact;
2040 action.sa_handler = term_handler;
2041 sigemptyset(&action.sa_mask);
2042 action.sa_flags = 0;
2043 if (sigaction(SIGTERM, &action, &oact) != 0) {
2044 barf("can't install TERM handler");
2049 /* A capability holds the state a native thread needs in
2050 * order to execute STG code. At least one capability is
2051 * floating around (only SMP builds have more than one).
2055 #if defined(RTS_SUPPORTS_THREADS)
2056 /* start our haskell execution tasks */
2058 startTaskManager(RtsFlags.ParFlags.nNodes, taskStart);
2060 startTaskManager(0,taskStart);
2064 #if /* defined(SMP) ||*/ defined(PAR)
2068 #if defined(RTS_SUPPORTS_THREADS)
2069 RELEASE_LOCK(&sched_mutex);
2075 exitScheduler( void )
2077 #if defined(RTS_SUPPORTS_THREADS)
2082 /* -----------------------------------------------------------------------------
2083 Managing the per-task allocation areas.
2085 Each capability comes with an allocation area. These are
2086 fixed-length block lists into which allocation can be done.
2088 ToDo: no support for two-space collection at the moment???
2089 -------------------------------------------------------------------------- */
2091 /* -----------------------------------------------------------------------------
2092 * waitThread is the external interface for running a new computation
2093 * and waiting for the result.
2095 * In the non-SMP case, we create a new main thread, push it on the
2096 * main-thread stack, and invoke the scheduler to run it. The
2097 * scheduler will return when the top main thread on the stack has
2098 * completed or died, and fill in the necessary fields of the
2099 * main_thread structure.
2101 * In the SMP case, we create a main thread as before, but we then
2102 * create a new condition variable and sleep on it. When our new
2103 * main thread has completed, we'll be woken up and the status/result
2104 * will be in the main_thread struct.
2105 * -------------------------------------------------------------------------- */
2108 howManyThreadsAvail ( void )
2112 for (q = run_queue_hd; q != END_TSO_QUEUE; q = q->link)
2114 for (q = blocked_queue_hd; q != END_TSO_QUEUE; q = q->link)
2116 for (q = sleeping_queue; q != END_TSO_QUEUE; q = q->link)
2122 finishAllThreads ( void )
2125 while (run_queue_hd != END_TSO_QUEUE) {
2126 waitThread ( run_queue_hd, NULL);
2128 while (blocked_queue_hd != END_TSO_QUEUE) {
2129 waitThread ( blocked_queue_hd, NULL);
2131 while (sleeping_queue != END_TSO_QUEUE) {
2132 waitThread ( blocked_queue_hd, NULL);
2135 (blocked_queue_hd != END_TSO_QUEUE ||
2136 run_queue_hd != END_TSO_QUEUE ||
2137 sleeping_queue != END_TSO_QUEUE);
2141 waitThread(StgTSO *tso, /*out*/StgClosure **ret)
2143 IF_DEBUG(scheduler, sched_belch("== scheduler: waiting for thread (%d)\n", tso->id));
2144 #if defined(THREADED_RTS)
2145 return waitThread_(tso,ret, rtsFalse);
2147 return waitThread_(tso,ret);
2152 waitThread_(StgTSO *tso,
2153 /*out*/StgClosure **ret
2154 #if defined(THREADED_RTS)
2155 , rtsBool blockWaiting
2160 SchedulerStatus stat;
2162 ACQUIRE_LOCK(&sched_mutex);
2163 IF_DEBUG(scheduler, sched_belch("== scheduler: waiting for thread (%d)\n", tso->id));
2165 m = stgMallocBytes(sizeof(StgMainThread), "waitThread");
2170 #if defined(RTS_SUPPORTS_THREADS)
2171 initCondition(&m->wakeup);
2174 m->link = main_threads;
2177 IF_DEBUG(scheduler, sched_belch("== scheduler: new main thread (%d)\n", m->tso->id));
2179 #if defined(RTS_SUPPORTS_THREADS)
2181 # if defined(THREADED_RTS)
2182 if (!blockWaiting) {
2183 /* In the threaded case, the OS thread that called main()
2184 * gets to enter the RTS directly without going via another
2187 RELEASE_LOCK(&sched_mutex);
2189 ASSERT(m->stat != NoStatus);
2193 IF_DEBUG(scheduler, sched_belch("sfoo"));
2195 waitCondition(&m->wakeup, &sched_mutex);
2196 } while (m->stat == NoStatus);
2199 /* GranSim specific init */
2200 CurrentTSO = m->tso; // the TSO to run
2201 procStatus[MainProc] = Busy; // status of main PE
2202 CurrentProc = MainProc; // PE to run it on
2206 RELEASE_LOCK(&sched_mutex);
2208 ASSERT(m->stat != NoStatus);
2213 #if defined(RTS_SUPPORTS_THREADS)
2214 closeCondition(&m->wakeup);
2217 IF_DEBUG(scheduler, fprintf(stderr, "== scheduler: main thread (%d) finished\n",
2221 #if defined(THREADED_RTS)
2224 RELEASE_LOCK(&sched_mutex);
2229 //@node Run queue code, Garbage Collextion Routines, Suspend and Resume, Main scheduling code
2230 //@subsection Run queue code
2234 NB: In GranSim we have many run queues; run_queue_hd is actually a macro
2235 unfolding to run_queue_hds[CurrentProc], thus CurrentProc is an
2236 implicit global variable that has to be correct when calling these
2240 /* Put the new thread on the head of the runnable queue.
2241 * The caller of createThread better push an appropriate closure
2242 * on this thread's stack before the scheduler is invoked.
2244 static /* inline */ void
2245 add_to_run_queue(tso)
2248 ASSERT(tso!=run_queue_hd && tso!=run_queue_tl);
2249 tso->link = run_queue_hd;
2251 if (run_queue_tl == END_TSO_QUEUE) {
2256 /* Put the new thread at the end of the runnable queue. */
2257 static /* inline */ void
2258 push_on_run_queue(tso)
2261 ASSERT(get_itbl((StgClosure *)tso)->type == TSO);
2262 ASSERT(run_queue_hd!=NULL && run_queue_tl!=NULL);
2263 ASSERT(tso!=run_queue_hd && tso!=run_queue_tl);
2264 if (run_queue_hd == END_TSO_QUEUE) {
2267 run_queue_tl->link = tso;
2273 Should be inlined because it's used very often in schedule. The tso
2274 argument is actually only needed in GranSim, where we want to have the
2275 possibility to schedule *any* TSO on the run queue, irrespective of the
2276 actual ordering. Therefore, if tso is not the nil TSO then we traverse
2277 the run queue and dequeue the tso, adjusting the links in the queue.
2279 //@cindex take_off_run_queue
2280 static /* inline */ StgTSO*
2281 take_off_run_queue(StgTSO *tso) {
2285 qetlaHbogh Qu' ngaSbogh ghomDaQ {tso} yIteq!
2287 if tso is specified, unlink that tso from the run_queue (doesn't have
2288 to be at the beginning of the queue); GranSim only
2290 if (tso!=END_TSO_QUEUE) {
2291 /* find tso in queue */
2292 for (t=run_queue_hd, prev=END_TSO_QUEUE;
2293 t!=END_TSO_QUEUE && t!=tso;
2297 /* now actually dequeue the tso */
2298 if (prev!=END_TSO_QUEUE) {
2299 ASSERT(run_queue_hd!=t);
2300 prev->link = t->link;
2302 /* t is at beginning of thread queue */
2303 ASSERT(run_queue_hd==t);
2304 run_queue_hd = t->link;
2306 /* t is at end of thread queue */
2307 if (t->link==END_TSO_QUEUE) {
2308 ASSERT(t==run_queue_tl);
2309 run_queue_tl = prev;
2311 ASSERT(run_queue_tl!=t);
2313 t->link = END_TSO_QUEUE;
2315 /* take tso from the beginning of the queue; std concurrent code */
2317 if (t != END_TSO_QUEUE) {
2318 run_queue_hd = t->link;
2319 t->link = END_TSO_QUEUE;
2320 if (run_queue_hd == END_TSO_QUEUE) {
2321 run_queue_tl = END_TSO_QUEUE;
2330 //@node Garbage Collextion Routines, Blocking Queue Routines, Run queue code, Main scheduling code
2331 //@subsection Garbage Collextion Routines
2333 /* ---------------------------------------------------------------------------
2334 Where are the roots that we know about?
2336 - all the threads on the runnable queue
2337 - all the threads on the blocked queue
2338 - all the threads on the sleeping queue
2339 - all the thread currently executing a _ccall_GC
2340 - all the "main threads"
2342 ------------------------------------------------------------------------ */
2344 /* This has to be protected either by the scheduler monitor, or by the
2345 garbage collection monitor (probably the latter).
2350 GetRoots(evac_fn evac)
2355 for (i=0; i<=RtsFlags.GranFlags.proc; i++) {
2356 if ((run_queue_hds[i] != END_TSO_QUEUE) && ((run_queue_hds[i] != NULL)))
2357 evac((StgClosure **)&run_queue_hds[i]);
2358 if ((run_queue_tls[i] != END_TSO_QUEUE) && ((run_queue_tls[i] != NULL)))
2359 evac((StgClosure **)&run_queue_tls[i]);
2361 if ((blocked_queue_hds[i] != END_TSO_QUEUE) && ((blocked_queue_hds[i] != NULL)))
2362 evac((StgClosure **)&blocked_queue_hds[i]);
2363 if ((blocked_queue_tls[i] != END_TSO_QUEUE) && ((blocked_queue_tls[i] != NULL)))
2364 evac((StgClosure **)&blocked_queue_tls[i]);
2365 if ((ccalling_threadss[i] != END_TSO_QUEUE) && ((ccalling_threadss[i] != NULL)))
2366 evac((StgClosure **)&ccalling_threads[i]);
2373 if (run_queue_hd != END_TSO_QUEUE) {
2374 ASSERT(run_queue_tl != END_TSO_QUEUE);
2375 evac((StgClosure **)&run_queue_hd);
2376 evac((StgClosure **)&run_queue_tl);
2379 if (blocked_queue_hd != END_TSO_QUEUE) {
2380 ASSERT(blocked_queue_tl != END_TSO_QUEUE);
2381 evac((StgClosure **)&blocked_queue_hd);
2382 evac((StgClosure **)&blocked_queue_tl);
2385 if (sleeping_queue != END_TSO_QUEUE) {
2386 evac((StgClosure **)&sleeping_queue);
2390 if (suspended_ccalling_threads != END_TSO_QUEUE) {
2391 evac((StgClosure **)&suspended_ccalling_threads);
2394 #if defined(PAR) || defined(GRAN)
2395 markSparkQueue(evac);
2399 /* -----------------------------------------------------------------------------
2402 This is the interface to the garbage collector from Haskell land.
2403 We provide this so that external C code can allocate and garbage
2404 collect when called from Haskell via _ccall_GC.
2406 It might be useful to provide an interface whereby the programmer
2407 can specify more roots (ToDo).
2409 This needs to be protected by the GC condition variable above. KH.
2410 -------------------------------------------------------------------------- */
2412 void (*extra_roots)(evac_fn);
2417 /* Obligated to hold this lock upon entry */
2418 ACQUIRE_LOCK(&sched_mutex);
2419 GarbageCollect(GetRoots,rtsFalse);
2420 RELEASE_LOCK(&sched_mutex);
2424 performMajorGC(void)
2426 ACQUIRE_LOCK(&sched_mutex);
2427 GarbageCollect(GetRoots,rtsTrue);
2428 RELEASE_LOCK(&sched_mutex);
2432 AllRoots(evac_fn evac)
2434 GetRoots(evac); // the scheduler's roots
2435 extra_roots(evac); // the user's roots
2439 performGCWithRoots(void (*get_roots)(evac_fn))
2441 ACQUIRE_LOCK(&sched_mutex);
2442 extra_roots = get_roots;
2443 GarbageCollect(AllRoots,rtsFalse);
2444 RELEASE_LOCK(&sched_mutex);
2447 /* -----------------------------------------------------------------------------
2450 If the thread has reached its maximum stack size, then raise the
2451 StackOverflow exception in the offending thread. Otherwise
2452 relocate the TSO into a larger chunk of memory and adjust its stack
2454 -------------------------------------------------------------------------- */
2457 threadStackOverflow(StgTSO *tso)
2459 nat new_stack_size, new_tso_size, diff, stack_words;
2463 IF_DEBUG(sanity,checkTSO(tso));
2464 if (tso->stack_size >= tso->max_stack_size) {
2467 belch("@@ threadStackOverflow of TSO %d (%p): stack too large (now %ld; max is %ld",
2468 tso->id, tso, tso->stack_size, tso->max_stack_size);
2469 /* If we're debugging, just print out the top of the stack */
2470 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2473 /* Send this thread the StackOverflow exception */
2474 raiseAsync(tso, (StgClosure *)stackOverflow_closure);
2478 /* Try to double the current stack size. If that takes us over the
2479 * maximum stack size for this thread, then use the maximum instead.
2480 * Finally round up so the TSO ends up as a whole number of blocks.
2482 new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
2483 new_tso_size = (nat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) +
2484 TSO_STRUCT_SIZE)/sizeof(W_);
2485 new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */
2486 new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
2488 IF_DEBUG(scheduler, fprintf(stderr,"== scheduler: increasing stack size from %d words to %d.\n", tso->stack_size, new_stack_size));
2490 dest = (StgTSO *)allocate(new_tso_size);
2491 TICK_ALLOC_TSO(new_stack_size,0);
2493 /* copy the TSO block and the old stack into the new area */
2494 memcpy(dest,tso,TSO_STRUCT_SIZE);
2495 stack_words = tso->stack + tso->stack_size - tso->sp;
2496 new_sp = (P_)dest + new_tso_size - stack_words;
2497 memcpy(new_sp, tso->sp, stack_words * sizeof(W_));
2499 /* relocate the stack pointers... */
2500 diff = (P_)new_sp - (P_)tso->sp; /* In *words* */
2501 dest->su = (StgUpdateFrame *) ((P_)dest->su + diff);
2503 dest->stack_size = new_stack_size;
2505 /* and relocate the update frame list */
2506 relocate_stack(dest, diff);
2508 /* Mark the old TSO as relocated. We have to check for relocated
2509 * TSOs in the garbage collector and any primops that deal with TSOs.
2511 * It's important to set the sp and su values to just beyond the end
2512 * of the stack, so we don't attempt to scavenge any part of the
2515 tso->what_next = ThreadRelocated;
2517 tso->sp = (P_)&(tso->stack[tso->stack_size]);
2518 tso->su = (StgUpdateFrame *)tso->sp;
2519 tso->why_blocked = NotBlocked;
2520 dest->mut_link = NULL;
2522 IF_PAR_DEBUG(verbose,
2523 belch("@@ threadStackOverflow of TSO %d (now at %p): stack size increased to %ld",
2524 tso->id, tso, tso->stack_size);
2525 /* If we're debugging, just print out the top of the stack */
2526 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2529 IF_DEBUG(sanity,checkTSO(tso));
2531 IF_DEBUG(scheduler,printTSO(dest));
2537 //@node Blocking Queue Routines, Exception Handling Routines, Garbage Collextion Routines, Main scheduling code
2538 //@subsection Blocking Queue Routines
2540 /* ---------------------------------------------------------------------------
2541 Wake up a queue that was blocked on some resource.
2542 ------------------------------------------------------------------------ */
2546 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2551 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2553 /* write RESUME events to log file and
2554 update blocked and fetch time (depending on type of the orig closure) */
2555 if (RtsFlags.ParFlags.ParStats.Full) {
2556 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
2557 GR_RESUMEQ, ((StgTSO *)bqe), ((StgTSO *)bqe)->block_info.closure,
2558 0, 0 /* spark_queue_len(ADVISORY_POOL) */);
2559 if (EMPTY_RUN_QUEUE())
2560 emitSchedule = rtsTrue;
2562 switch (get_itbl(node)->type) {
2564 ((StgTSO *)bqe)->par.fetchtime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2569 ((StgTSO *)bqe)->par.blocktime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2576 barf("{unblockOneLocked}Daq Qagh: unexpected closure in blocking queue");
2583 static StgBlockingQueueElement *
2584 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2587 PEs node_loc, tso_loc;
2589 node_loc = where_is(node); // should be lifted out of loop
2590 tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2591 tso_loc = where_is((StgClosure *)tso);
2592 if (IS_LOCAL_TO(PROCS(node),tso_loc)) { // TSO is local
2593 /* !fake_fetch => TSO is on CurrentProc is same as IS_LOCAL_TO */
2594 ASSERT(CurrentProc!=node_loc || tso_loc==CurrentProc);
2595 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.lunblocktime;
2596 // insertThread(tso, node_loc);
2597 new_event(tso_loc, tso_loc, CurrentTime[CurrentProc],
2599 tso, node, (rtsSpark*)NULL);
2600 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2603 } else { // TSO is remote (actually should be FMBQ)
2604 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.mpacktime +
2605 RtsFlags.GranFlags.Costs.gunblocktime +
2606 RtsFlags.GranFlags.Costs.latency;
2607 new_event(tso_loc, CurrentProc, CurrentTime[CurrentProc],
2609 tso, node, (rtsSpark*)NULL);
2610 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2613 /* the thread-queue-overhead is accounted for in either Resume or UnblockThread */
2615 fprintf(stderr," %s TSO %d (%p) [PE %d] (block_info.closure=%p) (next=%p) ,",
2616 (node_loc==tso_loc ? "Local" : "Global"),
2617 tso->id, tso, CurrentProc, tso->block_info.closure, tso->link));
2618 tso->block_info.closure = NULL;
2619 IF_DEBUG(scheduler,belch("-- Waking up thread %ld (%p)",
2623 static StgBlockingQueueElement *
2624 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2626 StgBlockingQueueElement *next;
2628 switch (get_itbl(bqe)->type) {
2630 ASSERT(((StgTSO *)bqe)->why_blocked != NotBlocked);
2631 /* if it's a TSO just push it onto the run_queue */
2633 // ((StgTSO *)bqe)->link = END_TSO_QUEUE; // debugging?
2634 PUSH_ON_RUN_QUEUE((StgTSO *)bqe);
2636 unblockCount(bqe, node);
2637 /* reset blocking status after dumping event */
2638 ((StgTSO *)bqe)->why_blocked = NotBlocked;
2642 /* if it's a BLOCKED_FETCH put it on the PendingFetches list */
2644 bqe->link = (StgBlockingQueueElement *)PendingFetches;
2645 PendingFetches = (StgBlockedFetch *)bqe;
2649 /* can ignore this case in a non-debugging setup;
2650 see comments on RBHSave closures above */
2652 /* check that the closure is an RBHSave closure */
2653 ASSERT(get_itbl((StgClosure *)bqe) == &stg_RBH_Save_0_info ||
2654 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_1_info ||
2655 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_2_info);
2659 barf("{unblockOneLocked}Daq Qagh: Unexpected IP (%#lx; %s) in blocking queue at %#lx\n",
2660 get_itbl((StgClosure *)bqe), info_type((StgClosure *)bqe),
2664 IF_PAR_DEBUG(bq, fprintf(stderr, ", %p (%s)", bqe, info_type((StgClosure*)bqe)));
2668 #else /* !GRAN && !PAR */
2670 unblockOneLocked(StgTSO *tso)
2674 ASSERT(get_itbl(tso)->type == TSO);
2675 ASSERT(tso->why_blocked != NotBlocked);
2676 tso->why_blocked = NotBlocked;
2678 PUSH_ON_RUN_QUEUE(tso);
2680 IF_DEBUG(scheduler,sched_belch("waking up thread %ld", tso->id));
2685 #if defined(GRAN) || defined(PAR)
2686 inline StgBlockingQueueElement *
2687 unblockOne(StgBlockingQueueElement *bqe, StgClosure *node)
2689 ACQUIRE_LOCK(&sched_mutex);
2690 bqe = unblockOneLocked(bqe, node);
2691 RELEASE_LOCK(&sched_mutex);
2696 unblockOne(StgTSO *tso)
2698 ACQUIRE_LOCK(&sched_mutex);
2699 tso = unblockOneLocked(tso);
2700 RELEASE_LOCK(&sched_mutex);
2707 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2709 StgBlockingQueueElement *bqe;
2714 belch("##-_ AwBQ for node %p on PE %d @ %ld by TSO %d (%p): ", \
2715 node, CurrentProc, CurrentTime[CurrentProc],
2716 CurrentTSO->id, CurrentTSO));
2718 node_loc = where_is(node);
2720 ASSERT(q == END_BQ_QUEUE ||
2721 get_itbl(q)->type == TSO || // q is either a TSO or an RBHSave
2722 get_itbl(q)->type == CONSTR); // closure (type constructor)
2723 ASSERT(is_unique(node));
2725 /* FAKE FETCH: magically copy the node to the tso's proc;
2726 no Fetch necessary because in reality the node should not have been
2727 moved to the other PE in the first place
2729 if (CurrentProc!=node_loc) {
2731 belch("## node %p is on PE %d but CurrentProc is %d (TSO %d); assuming fake fetch and adjusting bitmask (old: %#x)",
2732 node, node_loc, CurrentProc, CurrentTSO->id,
2733 // CurrentTSO, where_is(CurrentTSO),
2734 node->header.gran.procs));
2735 node->header.gran.procs = (node->header.gran.procs) | PE_NUMBER(CurrentProc);
2737 belch("## new bitmask of node %p is %#x",
2738 node, node->header.gran.procs));
2739 if (RtsFlags.GranFlags.GranSimStats.Global) {
2740 globalGranStats.tot_fake_fetches++;
2745 // ToDo: check: ASSERT(CurrentProc==node_loc);
2746 while (get_itbl(bqe)->type==TSO) { // q != END_TSO_QUEUE) {
2749 bqe points to the current element in the queue
2750 next points to the next element in the queue
2752 //tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2753 //tso_loc = where_is(tso);
2755 bqe = unblockOneLocked(bqe, node);
2758 /* if this is the BQ of an RBH, we have to put back the info ripped out of
2759 the closure to make room for the anchor of the BQ */
2760 if (bqe!=END_BQ_QUEUE) {
2761 ASSERT(get_itbl(node)->type == RBH && get_itbl(bqe)->type == CONSTR);
2763 ASSERT((info_ptr==&RBH_Save_0_info) ||
2764 (info_ptr==&RBH_Save_1_info) ||
2765 (info_ptr==&RBH_Save_2_info));
2767 /* cf. convertToRBH in RBH.c for writing the RBHSave closure */
2768 ((StgRBH *)node)->blocking_queue = (StgBlockingQueueElement *)((StgRBHSave *)bqe)->payload[0];
2769 ((StgRBH *)node)->mut_link = (StgMutClosure *)((StgRBHSave *)bqe)->payload[1];
2772 belch("## Filled in RBH_Save for %p (%s) at end of AwBQ",
2773 node, info_type(node)));
2776 /* statistics gathering */
2777 if (RtsFlags.GranFlags.GranSimStats.Global) {
2778 // globalGranStats.tot_bq_processing_time += bq_processing_time;
2779 globalGranStats.tot_bq_len += len; // total length of all bqs awakened
2780 // globalGranStats.tot_bq_len_local += len_local; // same for local TSOs only
2781 globalGranStats.tot_awbq++; // total no. of bqs awakened
2784 fprintf(stderr,"## BQ Stats of %p: [%d entries] %s\n",
2785 node, len, (bqe!=END_BQ_QUEUE) ? "RBH" : ""));
2789 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2791 StgBlockingQueueElement *bqe;
2793 ACQUIRE_LOCK(&sched_mutex);
2795 IF_PAR_DEBUG(verbose,
2796 belch("##-_ AwBQ for node %p on [%x]: ",
2800 if(get_itbl(q)->type == CONSTR || q==END_BQ_QUEUE) {
2801 IF_PAR_DEBUG(verbose, belch("## ... nothing to unblock so lets just return. RFP (BUG?)"));
2806 ASSERT(q == END_BQ_QUEUE ||
2807 get_itbl(q)->type == TSO ||
2808 get_itbl(q)->type == BLOCKED_FETCH ||
2809 get_itbl(q)->type == CONSTR);
2812 while (get_itbl(bqe)->type==TSO ||
2813 get_itbl(bqe)->type==BLOCKED_FETCH) {
2814 bqe = unblockOneLocked(bqe, node);
2816 RELEASE_LOCK(&sched_mutex);
2819 #else /* !GRAN && !PAR */
2821 awakenBlockedQueue(StgTSO *tso)
2823 ACQUIRE_LOCK(&sched_mutex);
2824 while (tso != END_TSO_QUEUE) {
2825 tso = unblockOneLocked(tso);
2827 RELEASE_LOCK(&sched_mutex);
2831 //@node Exception Handling Routines, Debugging Routines, Blocking Queue Routines, Main scheduling code
2832 //@subsection Exception Handling Routines
2834 /* ---------------------------------------------------------------------------
2836 - usually called inside a signal handler so it mustn't do anything fancy.
2837 ------------------------------------------------------------------------ */
2840 interruptStgRts(void)
2846 /* -----------------------------------------------------------------------------
2849 This is for use when we raise an exception in another thread, which
2851 This has nothing to do with the UnblockThread event in GranSim. -- HWL
2852 -------------------------------------------------------------------------- */
2854 #if defined(GRAN) || defined(PAR)
2856 NB: only the type of the blocking queue is different in GranSim and GUM
2857 the operations on the queue-elements are the same
2858 long live polymorphism!
2860 Locks: sched_mutex is held upon entry and exit.
2864 unblockThread(StgTSO *tso)
2866 StgBlockingQueueElement *t, **last;
2868 switch (tso->why_blocked) {
2871 return; /* not blocked */
2874 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
2876 StgBlockingQueueElement *last_tso = END_BQ_QUEUE;
2877 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
2879 last = (StgBlockingQueueElement **)&mvar->head;
2880 for (t = (StgBlockingQueueElement *)mvar->head;
2882 last = &t->link, last_tso = t, t = t->link) {
2883 if (t == (StgBlockingQueueElement *)tso) {
2884 *last = (StgBlockingQueueElement *)tso->link;
2885 if (mvar->tail == tso) {
2886 mvar->tail = (StgTSO *)last_tso;
2891 barf("unblockThread (MVAR): TSO not found");
2894 case BlockedOnBlackHole:
2895 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
2897 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
2899 last = &bq->blocking_queue;
2900 for (t = bq->blocking_queue;
2902 last = &t->link, t = t->link) {
2903 if (t == (StgBlockingQueueElement *)tso) {
2904 *last = (StgBlockingQueueElement *)tso->link;
2908 barf("unblockThread (BLACKHOLE): TSO not found");
2911 case BlockedOnException:
2913 StgTSO *target = tso->block_info.tso;
2915 ASSERT(get_itbl(target)->type == TSO);
2917 if (target->what_next == ThreadRelocated) {
2918 target = target->link;
2919 ASSERT(get_itbl(target)->type == TSO);
2922 ASSERT(target->blocked_exceptions != NULL);
2924 last = (StgBlockingQueueElement **)&target->blocked_exceptions;
2925 for (t = (StgBlockingQueueElement *)target->blocked_exceptions;
2927 last = &t->link, t = t->link) {
2928 ASSERT(get_itbl(t)->type == TSO);
2929 if (t == (StgBlockingQueueElement *)tso) {
2930 *last = (StgBlockingQueueElement *)tso->link;
2934 barf("unblockThread (Exception): TSO not found");
2938 case BlockedOnWrite:
2940 /* take TSO off blocked_queue */
2941 StgBlockingQueueElement *prev = NULL;
2942 for (t = (StgBlockingQueueElement *)blocked_queue_hd; t != END_BQ_QUEUE;
2943 prev = t, t = t->link) {
2944 if (t == (StgBlockingQueueElement *)tso) {
2946 blocked_queue_hd = (StgTSO *)t->link;
2947 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
2948 blocked_queue_tl = END_TSO_QUEUE;
2951 prev->link = t->link;
2952 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
2953 blocked_queue_tl = (StgTSO *)prev;
2959 barf("unblockThread (I/O): TSO not found");
2962 case BlockedOnDelay:
2964 /* take TSO off sleeping_queue */
2965 StgBlockingQueueElement *prev = NULL;
2966 for (t = (StgBlockingQueueElement *)sleeping_queue; t != END_BQ_QUEUE;
2967 prev = t, t = t->link) {
2968 if (t == (StgBlockingQueueElement *)tso) {
2970 sleeping_queue = (StgTSO *)t->link;
2972 prev->link = t->link;
2977 barf("unblockThread (I/O): TSO not found");
2981 barf("unblockThread");
2985 tso->link = END_TSO_QUEUE;
2986 tso->why_blocked = NotBlocked;
2987 tso->block_info.closure = NULL;
2988 PUSH_ON_RUN_QUEUE(tso);
2992 unblockThread(StgTSO *tso)
2996 /* To avoid locking unnecessarily. */
2997 if (tso->why_blocked == NotBlocked) {
3001 switch (tso->why_blocked) {
3004 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
3006 StgTSO *last_tso = END_TSO_QUEUE;
3007 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
3010 for (t = mvar->head; t != END_TSO_QUEUE;
3011 last = &t->link, last_tso = t, t = t->link) {
3014 if (mvar->tail == tso) {
3015 mvar->tail = last_tso;
3020 barf("unblockThread (MVAR): TSO not found");
3023 case BlockedOnBlackHole:
3024 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
3026 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
3028 last = &bq->blocking_queue;
3029 for (t = bq->blocking_queue; t != END_TSO_QUEUE;
3030 last = &t->link, t = t->link) {
3036 barf("unblockThread (BLACKHOLE): TSO not found");
3039 case BlockedOnException:
3041 StgTSO *target = tso->block_info.tso;
3043 ASSERT(get_itbl(target)->type == TSO);
3045 while (target->what_next == ThreadRelocated) {
3046 target = target->link;
3047 ASSERT(get_itbl(target)->type == TSO);
3050 ASSERT(target->blocked_exceptions != NULL);
3052 last = &target->blocked_exceptions;
3053 for (t = target->blocked_exceptions; t != END_TSO_QUEUE;
3054 last = &t->link, t = t->link) {
3055 ASSERT(get_itbl(t)->type == TSO);
3061 barf("unblockThread (Exception): TSO not found");
3065 case BlockedOnWrite:
3067 StgTSO *prev = NULL;
3068 for (t = blocked_queue_hd; t != END_TSO_QUEUE;
3069 prev = t, t = t->link) {
3072 blocked_queue_hd = t->link;
3073 if (blocked_queue_tl == t) {
3074 blocked_queue_tl = END_TSO_QUEUE;
3077 prev->link = t->link;
3078 if (blocked_queue_tl == t) {
3079 blocked_queue_tl = prev;
3085 barf("unblockThread (I/O): TSO not found");
3088 case BlockedOnDelay:
3090 StgTSO *prev = NULL;
3091 for (t = sleeping_queue; t != END_TSO_QUEUE;
3092 prev = t, t = t->link) {
3095 sleeping_queue = t->link;
3097 prev->link = t->link;
3102 barf("unblockThread (I/O): TSO not found");
3106 barf("unblockThread");
3110 tso->link = END_TSO_QUEUE;
3111 tso->why_blocked = NotBlocked;
3112 tso->block_info.closure = NULL;
3113 PUSH_ON_RUN_QUEUE(tso);
3117 /* -----------------------------------------------------------------------------
3120 * The following function implements the magic for raising an
3121 * asynchronous exception in an existing thread.
3123 * We first remove the thread from any queue on which it might be
3124 * blocked. The possible blockages are MVARs and BLACKHOLE_BQs.
3126 * We strip the stack down to the innermost CATCH_FRAME, building
3127 * thunks in the heap for all the active computations, so they can
3128 * be restarted if necessary. When we reach a CATCH_FRAME, we build
3129 * an application of the handler to the exception, and push it on
3130 * the top of the stack.
3132 * How exactly do we save all the active computations? We create an
3133 * AP_UPD for every UpdateFrame on the stack. Entering one of these
3134 * AP_UPDs pushes everything from the corresponding update frame
3135 * upwards onto the stack. (Actually, it pushes everything up to the
3136 * next update frame plus a pointer to the next AP_UPD object.
3137 * Entering the next AP_UPD object pushes more onto the stack until we
3138 * reach the last AP_UPD object - at which point the stack should look
3139 * exactly as it did when we killed the TSO and we can continue
3140 * execution by entering the closure on top of the stack.
3142 * We can also kill a thread entirely - this happens if either (a) the
3143 * exception passed to raiseAsync is NULL, or (b) there's no
3144 * CATCH_FRAME on the stack. In either case, we strip the entire
3145 * stack and replace the thread with a zombie.
3147 * Locks: sched_mutex held upon entry nor exit.
3149 * -------------------------------------------------------------------------- */
3152 deleteThread(StgTSO *tso)
3154 raiseAsync(tso,NULL);
3158 raiseAsyncWithLock(StgTSO *tso, StgClosure *exception)
3160 /* When raising async exs from contexts where sched_mutex isn't held;
3161 use raiseAsyncWithLock(). */
3162 ACQUIRE_LOCK(&sched_mutex);
3163 raiseAsync(tso,exception);
3164 RELEASE_LOCK(&sched_mutex);
3168 raiseAsync(StgTSO *tso, StgClosure *exception)
3170 StgUpdateFrame* su = tso->su;
3171 StgPtr sp = tso->sp;
3173 /* Thread already dead? */
3174 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
3178 IF_DEBUG(scheduler, sched_belch("raising exception in thread %ld.", tso->id));
3180 /* Remove it from any blocking queues */
3183 IF_DEBUG(scheduler, sched_belch("raising exception in thread %ld.", tso->id));
3184 /* The stack freezing code assumes there's a closure pointer on
3185 * the top of the stack. This isn't always the case with compiled
3186 * code, so we have to push a dummy closure on the top which just
3187 * returns to the next return address on the stack.
3189 if ( LOOKS_LIKE_GHC_INFO((void*)*sp) ) {
3190 *(--sp) = (W_)&stg_dummy_ret_closure;
3194 nat words = ((P_)su - (P_)sp) - 1;
3198 /* If we find a CATCH_FRAME, and we've got an exception to raise,
3199 * then build the THUNK raise(exception), and leave it on
3200 * top of the CATCH_FRAME ready to enter.
3202 if (get_itbl(su)->type == CATCH_FRAME && exception != NULL) {
3204 StgCatchFrame *cf = (StgCatchFrame *)su;
3208 /* we've got an exception to raise, so let's pass it to the
3209 * handler in this frame.
3211 raise = (StgClosure *)allocate(sizeofW(StgClosure)+1);
3212 TICK_ALLOC_SE_THK(1,0);
3213 SET_HDR(raise,&stg_raise_info,cf->header.prof.ccs);
3214 raise->payload[0] = exception;
3216 /* throw away the stack from Sp up to the CATCH_FRAME.
3220 /* Ensure that async excpetions are blocked now, so we don't get
3221 * a surprise exception before we get around to executing the
3224 if (tso->blocked_exceptions == NULL) {
3225 tso->blocked_exceptions = END_TSO_QUEUE;
3228 /* Put the newly-built THUNK on top of the stack, ready to execute
3229 * when the thread restarts.
3234 tso->what_next = ThreadEnterGHC;
3235 IF_DEBUG(sanity, checkTSO(tso));
3239 /* First build an AP_UPD consisting of the stack chunk above the
3240 * current update frame, with the top word on the stack as the
3243 ap = (StgAP_UPD *)allocate(AP_sizeW(words));
3248 ap->fun = (StgClosure *)sp[0];
3250 for(i=0; i < (nat)words; ++i) {
3251 ap->payload[i] = (StgClosure *)*sp++;
3254 switch (get_itbl(su)->type) {
3258 SET_HDR(ap,&stg_AP_UPD_info,su->header.prof.ccs /* ToDo */);
3259 TICK_ALLOC_UP_THK(words+1,0);
3262 fprintf(stderr, "scheduler: Updating ");
3263 printPtr((P_)su->updatee);
3264 fprintf(stderr, " with ");
3265 printObj((StgClosure *)ap);
3268 /* Replace the updatee with an indirection - happily
3269 * this will also wake up any threads currently
3270 * waiting on the result.
3272 * Warning: if we're in a loop, more than one update frame on
3273 * the stack may point to the same object. Be careful not to
3274 * overwrite an IND_OLDGEN in this case, because we'll screw
3275 * up the mutable lists. To be on the safe side, don't
3276 * overwrite any kind of indirection at all. See also
3277 * threadSqueezeStack in GC.c, where we have to make a similar
3280 if (!closure_IND(su->updatee)) {
3281 UPD_IND_NOLOCK(su->updatee,ap); /* revert the black hole */
3284 sp += sizeofW(StgUpdateFrame) -1;
3285 sp[0] = (W_)ap; /* push onto stack */
3291 StgCatchFrame *cf = (StgCatchFrame *)su;
3294 /* We want a PAP, not an AP_UPD. Fortunately, the
3295 * layout's the same.
3297 SET_HDR(ap,&stg_PAP_info,su->header.prof.ccs /* ToDo */);
3298 TICK_ALLOC_UPD_PAP(words+1,0);
3300 /* now build o = FUN(catch,ap,handler) */
3301 o = (StgClosure *)allocate(sizeofW(StgClosure)+2);
3302 TICK_ALLOC_FUN(2,0);
3303 SET_HDR(o,&stg_catch_info,su->header.prof.ccs /* ToDo */);
3304 o->payload[0] = (StgClosure *)ap;
3305 o->payload[1] = cf->handler;
3308 fprintf(stderr, "scheduler: Built ");
3309 printObj((StgClosure *)o);
3312 /* pop the old handler and put o on the stack */
3314 sp += sizeofW(StgCatchFrame) - 1;
3321 StgSeqFrame *sf = (StgSeqFrame *)su;
3324 SET_HDR(ap,&stg_PAP_info,su->header.prof.ccs /* ToDo */);
3325 TICK_ALLOC_UPD_PAP(words+1,0);
3327 /* now build o = FUN(seq,ap) */
3328 o = (StgClosure *)allocate(sizeofW(StgClosure)+1);
3329 TICK_ALLOC_SE_THK(1,0);
3330 SET_HDR(o,&stg_seq_info,su->header.prof.ccs /* ToDo */);
3331 o->payload[0] = (StgClosure *)ap;
3334 fprintf(stderr, "scheduler: Built ");
3335 printObj((StgClosure *)o);
3338 /* pop the old handler and put o on the stack */
3340 sp += sizeofW(StgSeqFrame) - 1;
3346 /* We've stripped the entire stack, the thread is now dead. */
3347 sp += sizeofW(StgStopFrame) - 1;
3348 sp[0] = (W_)exception; /* save the exception */
3349 tso->what_next = ThreadKilled;
3350 tso->su = (StgUpdateFrame *)(sp+1);
3361 /* -----------------------------------------------------------------------------
3362 resurrectThreads is called after garbage collection on the list of
3363 threads found to be garbage. Each of these threads will be woken
3364 up and sent a signal: BlockedOnDeadMVar if the thread was blocked
3365 on an MVar, or NonTermination if the thread was blocked on a Black
3368 Locks: sched_mutex isn't held upon entry nor exit.
3369 -------------------------------------------------------------------------- */
3372 resurrectThreads( StgTSO *threads )
3376 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
3377 next = tso->global_link;
3378 tso->global_link = all_threads;
3380 IF_DEBUG(scheduler, sched_belch("resurrecting thread %d", tso->id));
3382 switch (tso->why_blocked) {
3384 case BlockedOnException:
3385 /* Called by GC - sched_mutex lock is currently held. */
3386 raiseAsync(tso,(StgClosure *)BlockedOnDeadMVar_closure);
3388 case BlockedOnBlackHole:
3389 raiseAsync(tso,(StgClosure *)NonTermination_closure);
3392 /* This might happen if the thread was blocked on a black hole
3393 * belonging to a thread that we've just woken up (raiseAsync
3394 * can wake up threads, remember...).
3398 barf("resurrectThreads: thread blocked in a strange way");
3403 /* -----------------------------------------------------------------------------
3404 * Blackhole detection: if we reach a deadlock, test whether any
3405 * threads are blocked on themselves. Any threads which are found to
3406 * be self-blocked get sent a NonTermination exception.
3408 * This is only done in a deadlock situation in order to avoid
3409 * performance overhead in the normal case.
3411 * Locks: sched_mutex is held upon entry and exit.
3412 * -------------------------------------------------------------------------- */
3415 detectBlackHoles( void )
3417 StgTSO *t = all_threads;
3418 StgUpdateFrame *frame;
3419 StgClosure *blocked_on;
3421 for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
3423 while (t->what_next == ThreadRelocated) {
3425 ASSERT(get_itbl(t)->type == TSO);
3428 if (t->why_blocked != BlockedOnBlackHole) {
3432 blocked_on = t->block_info.closure;
3434 for (frame = t->su; ; frame = frame->link) {
3435 switch (get_itbl(frame)->type) {
3438 if (frame->updatee == blocked_on) {
3439 /* We are blocking on one of our own computations, so
3440 * send this thread the NonTermination exception.
3443 sched_belch("thread %d is blocked on itself", t->id));
3444 raiseAsync(t, (StgClosure *)NonTermination_closure);
3465 //@node Debugging Routines, Index, Exception Handling Routines, Main scheduling code
3466 //@subsection Debugging Routines
3468 /* -----------------------------------------------------------------------------
3469 Debugging: why is a thread blocked
3470 -------------------------------------------------------------------------- */
3475 printThreadBlockage(StgTSO *tso)
3477 switch (tso->why_blocked) {
3479 fprintf(stderr,"is blocked on read from fd %d", tso->block_info.fd);
3481 case BlockedOnWrite:
3482 fprintf(stderr,"is blocked on write to fd %d", tso->block_info.fd);
3484 case BlockedOnDelay:
3485 fprintf(stderr,"is blocked until %d", tso->block_info.target);
3488 fprintf(stderr,"is blocked on an MVar");
3490 case BlockedOnException:
3491 fprintf(stderr,"is blocked on delivering an exception to thread %d",
3492 tso->block_info.tso->id);
3494 case BlockedOnBlackHole:
3495 fprintf(stderr,"is blocked on a black hole");
3498 fprintf(stderr,"is not blocked");
3502 fprintf(stderr,"is blocked on global address; local FM_BQ is %p (%s)",
3503 tso->block_info.closure, info_type(tso->block_info.closure));
3505 case BlockedOnGA_NoSend:
3506 fprintf(stderr,"is blocked on global address (no send); local FM_BQ is %p (%s)",
3507 tso->block_info.closure, info_type(tso->block_info.closure));
3510 #if defined(RTS_SUPPORTS_THREADS)
3511 case BlockedOnCCall:
3512 fprintf(stderr,"is blocked on an external call");
3516 barf("printThreadBlockage: strange tso->why_blocked: %d for TSO %d (%d)",
3517 tso->why_blocked, tso->id, tso);
3522 printThreadStatus(StgTSO *tso)
3524 switch (tso->what_next) {
3526 fprintf(stderr,"has been killed");
3528 case ThreadComplete:
3529 fprintf(stderr,"has completed");
3532 printThreadBlockage(tso);
3537 printAllThreads(void)
3542 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
3543 ullong_format_string(TIME_ON_PROC(CurrentProc),
3544 time_string, rtsFalse/*no commas!*/);
3546 sched_belch("all threads at [%s]:", time_string);
3548 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
3549 ullong_format_string(CURRENT_TIME,
3550 time_string, rtsFalse/*no commas!*/);
3552 sched_belch("all threads at [%s]:", time_string);
3554 sched_belch("all threads:");
3557 for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
3558 fprintf(stderr, "\tthread %d ", t->id);
3559 if (t->label) fprintf(stderr,"[\"%s\"] ",t->label);
3560 printThreadStatus(t);
3561 fprintf(stderr,"\n");
3566 Print a whole blocking queue attached to node (debugging only).
3571 print_bq (StgClosure *node)
3573 StgBlockingQueueElement *bqe;
3577 fprintf(stderr,"## BQ of closure %p (%s): ",
3578 node, info_type(node));
3580 /* should cover all closures that may have a blocking queue */
3581 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
3582 get_itbl(node)->type == FETCH_ME_BQ ||
3583 get_itbl(node)->type == RBH ||
3584 get_itbl(node)->type == MVAR);
3586 ASSERT(node!=(StgClosure*)NULL); // sanity check
3588 print_bqe(((StgBlockingQueue*)node)->blocking_queue);
3592 Print a whole blocking queue starting with the element bqe.
3595 print_bqe (StgBlockingQueueElement *bqe)
3600 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
3602 for (end = (bqe==END_BQ_QUEUE);
3603 !end; // iterate until bqe points to a CONSTR
3604 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE),
3605 bqe = end ? END_BQ_QUEUE : bqe->link) {
3606 ASSERT(bqe != END_BQ_QUEUE); // sanity check
3607 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
3608 /* types of closures that may appear in a blocking queue */
3609 ASSERT(get_itbl(bqe)->type == TSO ||
3610 get_itbl(bqe)->type == BLOCKED_FETCH ||
3611 get_itbl(bqe)->type == CONSTR);
3612 /* only BQs of an RBH end with an RBH_Save closure */
3613 //ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
3615 switch (get_itbl(bqe)->type) {
3617 fprintf(stderr," TSO %u (%x),",
3618 ((StgTSO *)bqe)->id, ((StgTSO *)bqe));
3621 fprintf(stderr," BF (node=%p, ga=((%x, %d, %x)),",
3622 ((StgBlockedFetch *)bqe)->node,
3623 ((StgBlockedFetch *)bqe)->ga.payload.gc.gtid,
3624 ((StgBlockedFetch *)bqe)->ga.payload.gc.slot,
3625 ((StgBlockedFetch *)bqe)->ga.weight);
3628 fprintf(stderr," %s (IP %p),",
3629 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
3630 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
3631 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
3632 "RBH_Save_?"), get_itbl(bqe));
3635 barf("Unexpected closure type %s in blocking queue", // of %p (%s)",
3636 info_type((StgClosure *)bqe)); // , node, info_type(node));
3640 fputc('\n', stderr);
3642 # elif defined(GRAN)
3644 print_bq (StgClosure *node)
3646 StgBlockingQueueElement *bqe;
3647 PEs node_loc, tso_loc;
3650 /* should cover all closures that may have a blocking queue */
3651 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
3652 get_itbl(node)->type == FETCH_ME_BQ ||
3653 get_itbl(node)->type == RBH);
3655 ASSERT(node!=(StgClosure*)NULL); // sanity check
3656 node_loc = where_is(node);
3658 fprintf(stderr,"## BQ of closure %p (%s) on [PE %d]: ",
3659 node, info_type(node), node_loc);
3662 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
3664 for (bqe = ((StgBlockingQueue*)node)->blocking_queue, end = (bqe==END_BQ_QUEUE);
3665 !end; // iterate until bqe points to a CONSTR
3666 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) {
3667 ASSERT(bqe != END_BQ_QUEUE); // sanity check
3668 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
3669 /* types of closures that may appear in a blocking queue */
3670 ASSERT(get_itbl(bqe)->type == TSO ||
3671 get_itbl(bqe)->type == CONSTR);
3672 /* only BQs of an RBH end with an RBH_Save closure */
3673 ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
3675 tso_loc = where_is((StgClosure *)bqe);
3676 switch (get_itbl(bqe)->type) {
3678 fprintf(stderr," TSO %d (%p) on [PE %d],",
3679 ((StgTSO *)bqe)->id, (StgTSO *)bqe, tso_loc);
3682 fprintf(stderr," %s (IP %p),",
3683 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
3684 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
3685 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
3686 "RBH_Save_?"), get_itbl(bqe));
3689 barf("Unexpected closure type %s in blocking queue of %p (%s)",
3690 info_type((StgClosure *)bqe), node, info_type(node));
3694 fputc('\n', stderr);
3698 Nice and easy: only TSOs on the blocking queue
3701 print_bq (StgClosure *node)
3705 ASSERT(node!=(StgClosure*)NULL); // sanity check
3706 for (tso = ((StgBlockingQueue*)node)->blocking_queue;
3707 tso != END_TSO_QUEUE;
3709 ASSERT(tso!=NULL && tso!=END_TSO_QUEUE); // sanity check
3710 ASSERT(get_itbl(tso)->type == TSO); // guess what, sanity check
3711 fprintf(stderr," TSO %d (%p),", tso->id, tso);
3713 fputc('\n', stderr);
3724 for (i=0, tso=run_queue_hd;
3725 tso != END_TSO_QUEUE;
3734 sched_belch(char *s, ...)
3739 fprintf(stderr, "scheduler (task %ld): ", osThreadId());
3741 fprintf(stderr, "== ");
3743 fprintf(stderr, "scheduler: ");
3745 vfprintf(stderr, s, ap);
3746 fprintf(stderr, "\n");
3752 //@node Index, , Debugging Routines, Main scheduling code
3756 //* StgMainThread:: @cindex\s-+StgMainThread
3757 //* awaken_blocked_queue:: @cindex\s-+awaken_blocked_queue
3758 //* blocked_queue_hd:: @cindex\s-+blocked_queue_hd
3759 //* blocked_queue_tl:: @cindex\s-+blocked_queue_tl
3760 //* context_switch:: @cindex\s-+context_switch
3761 //* createThread:: @cindex\s-+createThread
3762 //* gc_pending_cond:: @cindex\s-+gc_pending_cond
3763 //* initScheduler:: @cindex\s-+initScheduler
3764 //* interrupted:: @cindex\s-+interrupted
3765 //* next_thread_id:: @cindex\s-+next_thread_id
3766 //* print_bq:: @cindex\s-+print_bq
3767 //* run_queue_hd:: @cindex\s-+run_queue_hd
3768 //* run_queue_tl:: @cindex\s-+run_queue_tl
3769 //* sched_mutex:: @cindex\s-+sched_mutex
3770 //* schedule:: @cindex\s-+schedule
3771 //* take_off_run_queue:: @cindex\s-+take_off_run_queue
3772 //* term_mutex:: @cindex\s-+term_mutex