1 /* ---------------------------------------------------------------------------
2 * $Id: Schedule.c,v 1.157 2002/10/22 11:01:19 simonmar Exp $
4 * (c) The GHC Team, 1998-2000
8 * Different GHC ways use this scheduler quite differently (see comments below)
9 * Here is the global picture:
11 * WAY Name CPP flag What's it for
12 * --------------------------------------
13 * mp GUM PAR Parallel execution on a distributed memory machine
14 * s SMP SMP Parallel execution on a shared memory machine
15 * mg GranSim GRAN Simulation of parallel execution
16 * md GUM/GdH DIST Distributed execution (based on GUM)
18 * --------------------------------------------------------------------------*/
20 //@node Main scheduling code, , ,
21 //@section Main scheduling code
24 * Version with scheduler monitor support for SMPs (WAY=s):
26 This design provides a high-level API to create and schedule threads etc.
27 as documented in the SMP design document.
29 It uses a monitor design controlled by a single mutex to exercise control
30 over accesses to shared data structures, and builds on the Posix threads
33 The majority of state is shared. In order to keep essential per-task state,
34 there is a Capability structure, which contains all the information
35 needed to run a thread: its STG registers, a pointer to its TSO, a
36 nursery etc. During STG execution, a pointer to the capability is
37 kept in a register (BaseReg).
39 In a non-SMP build, there is one global capability, namely MainRegTable.
43 * Version with support for distributed memory parallelism aka GUM (WAY=mp):
45 The main scheduling loop in GUM iterates until a finish message is received.
46 In that case a global flag @receivedFinish@ is set and this instance of
47 the RTS shuts down. See ghc/rts/parallel/HLComms.c:processMessages()
48 for the handling of incoming messages, such as PP_FINISH.
49 Note that in the parallel case we have a system manager that coordinates
50 different PEs, each of which are running one instance of the RTS.
51 See ghc/rts/parallel/SysMan.c for the main routine of the parallel program.
52 From this routine processes executing ghc/rts/Main.c are spawned. -- HWL
54 * Version with support for simulating parallel execution aka GranSim (WAY=mg):
56 The main scheduling code in GranSim is quite different from that in std
57 (concurrent) Haskell: while concurrent Haskell just iterates over the
58 threads in the runnable queue, GranSim is event driven, i.e. it iterates
59 over the events in the global event queue. -- HWL
64 //* Variables and Data structures::
65 //* Main scheduling loop::
66 //* Suspend and Resume::
68 //* Garbage Collextion Routines::
69 //* Blocking Queue Routines::
70 //* Exception Handling Routines::
71 //* Debugging Routines::
75 //@node Includes, Variables and Data structures, Main scheduling code, Main scheduling code
76 //@subsection Includes
78 #include "PosixSource.h"
85 #include "StgStartup.h"
87 #define COMPILING_SCHEDULER
89 #include "StgMiscClosures.h"
91 #include "Interpreter.h"
92 #include "Exception.h"
100 #include "ThreadLabels.h"
102 #include "Proftimer.h"
103 #include "ProfHeap.h"
105 #if defined(GRAN) || defined(PAR)
106 # include "GranSimRts.h"
107 # include "GranSim.h"
108 # include "ParallelRts.h"
109 # include "Parallel.h"
110 # include "ParallelDebug.h"
111 # include "FetchMe.h"
115 #include "Capability.h"
116 #include "OSThreads.h"
119 #ifdef HAVE_SYS_TYPES_H
120 #include <sys/types.h>
130 //@node Variables and Data structures, Prototypes, Includes, Main scheduling code
131 //@subsection Variables and Data structures
133 /* Main thread queue.
134 * Locks required: sched_mutex.
136 StgMainThread *main_threads = NULL;
139 * Locks required: sched_mutex.
143 StgTSO* ActiveTSO = NULL; /* for assigning system costs; GranSim-Light only */
144 /* rtsTime TimeOfNextEvent, EndOfTimeSlice; now in GranSim.c */
147 In GranSim we have a runnable and a blocked queue for each processor.
148 In order to minimise code changes new arrays run_queue_hds/tls
149 are created. run_queue_hd is then a short cut (macro) for
150 run_queue_hds[CurrentProc] (see GranSim.h).
153 StgTSO *run_queue_hds[MAX_PROC], *run_queue_tls[MAX_PROC];
154 StgTSO *blocked_queue_hds[MAX_PROC], *blocked_queue_tls[MAX_PROC];
155 StgTSO *ccalling_threadss[MAX_PROC];
156 /* We use the same global list of threads (all_threads) in GranSim as in
157 the std RTS (i.e. we are cheating). However, we don't use this list in
158 the GranSim specific code at the moment (so we are only potentially
163 StgTSO *run_queue_hd = NULL;
164 StgTSO *run_queue_tl = NULL;
165 StgTSO *blocked_queue_hd = NULL;
166 StgTSO *blocked_queue_tl = NULL;
167 StgTSO *sleeping_queue = NULL; /* perhaps replace with a hash table? */
171 /* Linked list of all threads.
172 * Used for detecting garbage collected threads.
174 StgTSO *all_threads = NULL;
176 /* When a thread performs a safe C call (_ccall_GC, using old
177 * terminology), it gets put on the suspended_ccalling_threads
178 * list. Used by the garbage collector.
180 static StgTSO *suspended_ccalling_threads;
182 static StgTSO *threadStackOverflow(StgTSO *tso);
184 /* KH: The following two flags are shared memory locations. There is no need
185 to lock them, since they are only unset at the end of a scheduler
189 /* flag set by signal handler to precipitate a context switch */
190 //@cindex context_switch
191 nat context_switch = 0;
193 /* if this flag is set as well, give up execution */
194 //@cindex interrupted
195 rtsBool interrupted = rtsFalse;
197 /* Next thread ID to allocate.
198 * Locks required: thread_id_mutex
200 //@cindex next_thread_id
201 static StgThreadID next_thread_id = 1;
204 * Pointers to the state of the current thread.
205 * Rule of thumb: if CurrentTSO != NULL, then we're running a Haskell
206 * thread. If CurrentTSO == NULL, then we're at the scheduler level.
209 /* The smallest stack size that makes any sense is:
210 * RESERVED_STACK_WORDS (so we can get back from the stack overflow)
211 * + sizeofW(StgStopFrame) (the stg_stop_thread_info frame)
212 * + 1 (the realworld token for an IO thread)
213 * + 1 (the closure to enter)
215 * A thread with this stack will bomb immediately with a stack
216 * overflow, which will increase its stack size.
219 #define MIN_STACK_WORDS (RESERVED_STACK_WORDS + sizeofW(StgStopFrame) + 2)
226 /* This is used in `TSO.h' and gcc 2.96 insists that this variable actually
227 * exists - earlier gccs apparently didn't.
232 static rtsBool ready_to_gc;
235 * Set to TRUE when entering a shutdown state (via shutdownHaskellAndExit()) --
236 * in an MT setting, needed to signal that a worker thread shouldn't hang around
237 * in the scheduler when it is out of work.
239 static rtsBool shutting_down_scheduler = rtsFalse;
241 void addToBlockedQueue ( StgTSO *tso );
243 static void schedule ( void );
244 void interruptStgRts ( void );
246 static void detectBlackHoles ( void );
249 static void sched_belch(char *s, ...);
252 #if defined(RTS_SUPPORTS_THREADS)
253 /* ToDo: carefully document the invariants that go together
254 * with these synchronisation objects.
256 Mutex sched_mutex = INIT_MUTEX_VAR;
257 Mutex term_mutex = INIT_MUTEX_VAR;
260 * A heavyweight solution to the problem of protecting
261 * the thread_id from concurrent update.
263 Mutex thread_id_mutex = INIT_MUTEX_VAR;
267 static Condition gc_pending_cond = INIT_COND_VAR;
271 #endif /* RTS_SUPPORTS_THREADS */
275 rtsTime TimeOfLastYield;
276 rtsBool emitSchedule = rtsTrue;
280 static char *whatNext_strs[] = {
290 StgTSO * createSparkThread(rtsSpark spark);
291 StgTSO * activateSpark (rtsSpark spark);
295 * The thread state for the main thread.
296 // ToDo: check whether not needed any more
300 #if defined(PAR) || defined(RTS_SUPPORTS_THREADS)
301 static void taskStart(void);
312 //@node Main scheduling loop, Suspend and Resume, Prototypes, Main scheduling code
313 //@subsection Main scheduling loop
315 /* ---------------------------------------------------------------------------
316 Main scheduling loop.
318 We use round-robin scheduling, each thread returning to the
319 scheduler loop when one of these conditions is detected:
322 * timer expires (thread yields)
327 Locking notes: we acquire the scheduler lock once at the beginning
328 of the scheduler loop, and release it when
330 * running a thread, or
331 * waiting for work, or
332 * waiting for a GC to complete.
335 In a GranSim setup this loop iterates over the global event queue.
336 This revolves around the global event queue, which determines what
337 to do next. Therefore, it's more complicated than either the
338 concurrent or the parallel (GUM) setup.
341 GUM iterates over incoming messages.
342 It starts with nothing to do (thus CurrentTSO == END_TSO_QUEUE),
343 and sends out a fish whenever it has nothing to do; in-between
344 doing the actual reductions (shared code below) it processes the
345 incoming messages and deals with delayed operations
346 (see PendingFetches).
347 This is not the ugliest code you could imagine, but it's bloody close.
349 ------------------------------------------------------------------------ */
356 StgThreadReturnCode ret;
364 rtsBool receivedFinish = rtsFalse;
366 nat tp_size, sp_size; // stats only
369 rtsBool was_interrupted = rtsFalse;
371 ACQUIRE_LOCK(&sched_mutex);
373 #if defined(RTS_SUPPORTS_THREADS)
374 waitForWorkCapability(&sched_mutex, &cap, rtsFalse);
376 /* simply initialise it in the non-threaded case */
377 grabCapability(&cap);
381 /* set up first event to get things going */
382 /* ToDo: assign costs for system setup and init MainTSO ! */
383 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
385 CurrentTSO, (StgClosure*)NULL, (rtsSpark*)NULL);
388 fprintf(stderr, "GRAN: Init CurrentTSO (in schedule) = %p\n", CurrentTSO);
389 G_TSO(CurrentTSO, 5));
391 if (RtsFlags.GranFlags.Light) {
392 /* Save current time; GranSim Light only */
393 CurrentTSO->gran.clock = CurrentTime[CurrentProc];
396 event = get_next_event();
398 while (event!=(rtsEvent*)NULL) {
399 /* Choose the processor with the next event */
400 CurrentProc = event->proc;
401 CurrentTSO = event->tso;
405 while (!receivedFinish) { /* set by processMessages */
406 /* when receiving PP_FINISH message */
413 IF_DEBUG(scheduler, printAllThreads());
415 #if defined(RTS_SUPPORTS_THREADS)
416 /* Check to see whether there are any worker threads
417 waiting to deposit external call results. If so,
418 yield our capability */
419 yieldToReturningWorker(&sched_mutex, &cap);
422 /* If we're interrupted (the user pressed ^C, or some other
423 * termination condition occurred), kill all the currently running
427 IF_DEBUG(scheduler, sched_belch("interrupted"));
429 interrupted = rtsFalse;
430 was_interrupted = rtsTrue;
433 /* Go through the list of main threads and wake up any
434 * clients whose computations have finished. ToDo: this
435 * should be done more efficiently without a linear scan
436 * of the main threads list, somehow...
438 #if defined(RTS_SUPPORTS_THREADS)
440 StgMainThread *m, **prev;
441 prev = &main_threads;
442 for (m = main_threads; m != NULL; m = m->link) {
443 switch (m->tso->what_next) {
446 *(m->ret) = (StgClosure *)m->tso->sp[0];
450 broadcastCondition(&m->wakeup);
452 removeThreadLabel((StgWord)m->tso);
456 if (m->ret) *(m->ret) = NULL;
458 if (was_interrupted) {
459 m->stat = Interrupted;
463 broadcastCondition(&m->wakeup);
465 removeThreadLabel((StgWord)m->tso);
474 #else /* not threaded */
477 /* in GUM do this only on the Main PE */
480 /* If our main thread has finished or been killed, return.
483 StgMainThread *m = main_threads;
484 if (m->tso->what_next == ThreadComplete
485 || m->tso->what_next == ThreadKilled) {
487 removeThreadLabel((StgWord)m->tso);
489 main_threads = main_threads->link;
490 if (m->tso->what_next == ThreadComplete) {
491 /* we finished successfully, fill in the return value */
492 if (m->ret) { *(m->ret) = (StgClosure *)m->tso->sp[0]; };
496 if (m->ret) { *(m->ret) = NULL; };
497 if (was_interrupted) {
498 m->stat = Interrupted;
508 /* Top up the run queue from our spark pool. We try to make the
509 * number of threads in the run queue equal to the number of
512 * Disable spark support in SMP for now, non-essential & requires
513 * a little bit of work to make it compile cleanly. -- sof 1/02.
515 #if 0 /* defined(SMP) */
517 nat n = getFreeCapabilities();
518 StgTSO *tso = run_queue_hd;
520 /* Count the run queue */
521 while (n > 0 && tso != END_TSO_QUEUE) {
528 spark = findSpark(rtsFalse);
530 break; /* no more sparks in the pool */
532 /* I'd prefer this to be done in activateSpark -- HWL */
533 /* tricky - it needs to hold the scheduler lock and
534 * not try to re-acquire it -- SDM */
535 createSparkThread(spark);
537 sched_belch("==^^ turning spark of closure %p into a thread",
538 (StgClosure *)spark));
541 /* We need to wake up the other tasks if we just created some
544 if (getFreeCapabilities() - n > 1) {
545 signalCondition( &thread_ready_cond );
550 /* check for signals each time around the scheduler */
551 #ifndef mingw32_TARGET_OS
552 if (signals_pending()) {
553 RELEASE_LOCK(&sched_mutex); /* ToDo: kill */
554 startSignalHandlers();
555 ACQUIRE_LOCK(&sched_mutex);
559 /* Check whether any waiting threads need to be woken up. If the
560 * run queue is empty, and there are no other tasks running, we
561 * can wait indefinitely for something to happen.
562 * ToDo: what if another client comes along & requests another
565 if ( !EMPTY_QUEUE(blocked_queue_hd) || !EMPTY_QUEUE(sleeping_queue) ) {
566 awaitEvent( EMPTY_RUN_QUEUE()
568 && allFreeCapabilities()
572 /* we can be interrupted while waiting for I/O... */
573 if (interrupted) continue;
576 * Detect deadlock: when we have no threads to run, there are no
577 * threads waiting on I/O or sleeping, and all the other tasks are
578 * waiting for work, we must have a deadlock of some description.
580 * We first try to find threads blocked on themselves (ie. black
581 * holes), and generate NonTermination exceptions where necessary.
583 * If no threads are black holed, we have a deadlock situation, so
584 * inform all the main threads.
587 if ( EMPTY_THREAD_QUEUES()
588 #if defined(RTS_SUPPORTS_THREADS)
589 && EMPTY_QUEUE(suspended_ccalling_threads)
592 && allFreeCapabilities()
596 IF_DEBUG(scheduler, sched_belch("deadlocked, forcing major GC..."));
597 #if defined(THREADED_RTS)
598 /* and SMP mode ..? */
599 releaseCapability(cap);
601 // Garbage collection can release some new threads due to
602 // either (a) finalizers or (b) threads resurrected because
603 // they are about to be send BlockedOnDeadMVar. Any threads
604 // thus released will be immediately runnable.
605 GarbageCollect(GetRoots,rtsTrue);
607 if ( !EMPTY_RUN_QUEUE() ) { goto not_deadlocked; }
610 sched_belch("still deadlocked, checking for black holes..."));
613 if ( !EMPTY_RUN_QUEUE() ) { goto not_deadlocked; }
615 #ifndef mingw32_TARGET_OS
616 /* If we have user-installed signal handlers, then wait
617 * for signals to arrive rather then bombing out with a
620 #if defined(RTS_SUPPORTS_THREADS)
621 if ( 0 ) { /* hmm..what to do? Simply stop waiting for
622 a signal with no runnable threads (or I/O
623 suspended ones) leads nowhere quick.
624 For now, simply shut down when we reach this
627 ToDo: define precisely under what conditions
628 the Scheduler should shut down in an MT setting.
631 if ( anyUserHandlers() ) {
634 sched_belch("still deadlocked, waiting for signals..."));
638 // we might be interrupted...
639 if (interrupted) { continue; }
641 if (signals_pending()) {
642 RELEASE_LOCK(&sched_mutex);
643 startSignalHandlers();
644 ACQUIRE_LOCK(&sched_mutex);
646 ASSERT(!EMPTY_RUN_QUEUE());
651 /* Probably a real deadlock. Send the current main thread the
652 * Deadlock exception (or in the SMP build, send *all* main
653 * threads the deadlock exception, since none of them can make
658 #if defined(RTS_SUPPORTS_THREADS)
659 for (m = main_threads; m != NULL; m = m->link) {
660 switch (m->tso->why_blocked) {
661 case BlockedOnBlackHole:
662 raiseAsync(m->tso, (StgClosure *)NonTermination_closure);
664 case BlockedOnException:
666 raiseAsync(m->tso, (StgClosure *)Deadlock_closure);
669 barf("deadlock: main thread blocked in a strange way");
674 switch (m->tso->why_blocked) {
675 case BlockedOnBlackHole:
676 raiseAsync(m->tso, (StgClosure *)NonTermination_closure);
678 case BlockedOnException:
680 raiseAsync(m->tso, (StgClosure *)Deadlock_closure);
683 barf("deadlock: main thread blocked in a strange way");
688 #if defined(RTS_SUPPORTS_THREADS)
689 /* ToDo: revisit conditions (and mechanism) for shutting
690 down a multi-threaded world */
691 IF_DEBUG(scheduler, sched_belch("all done, i think...shutting down."));
692 RELEASE_LOCK(&sched_mutex);
700 /* ToDo: add deadlock detection in GUM (similar to SMP) -- HWL */
704 /* If there's a GC pending, don't do anything until it has
708 IF_DEBUG(scheduler,sched_belch("waiting for GC"));
709 waitCondition( &gc_pending_cond, &sched_mutex );
713 #if defined(RTS_SUPPORTS_THREADS)
714 /* block until we've got a thread on the run queue and a free
718 if ( EMPTY_RUN_QUEUE() ) {
719 /* Give up our capability */
720 releaseCapability(cap);
722 /* If we're in the process of shutting down (& running the
723 * a batch of finalisers), don't wait around.
725 if ( shutting_down_scheduler ) {
726 RELEASE_LOCK(&sched_mutex);
729 IF_DEBUG(scheduler, sched_belch("thread %d: waiting for work", osThreadId()));
730 waitForWorkCapability(&sched_mutex, &cap, rtsTrue);
731 IF_DEBUG(scheduler, sched_belch("thread %d: work now available", osThreadId()));
736 if (RtsFlags.GranFlags.Light)
737 GranSimLight_enter_system(event, &ActiveTSO); // adjust ActiveTSO etc
739 /* adjust time based on time-stamp */
740 if (event->time > CurrentTime[CurrentProc] &&
741 event->evttype != ContinueThread)
742 CurrentTime[CurrentProc] = event->time;
744 /* Deal with the idle PEs (may issue FindWork or MoveSpark events) */
745 if (!RtsFlags.GranFlags.Light)
748 IF_DEBUG(gran, fprintf(stderr, "GRAN: switch by event-type\n"));
750 /* main event dispatcher in GranSim */
751 switch (event->evttype) {
752 /* Should just be continuing execution */
754 IF_DEBUG(gran, fprintf(stderr, "GRAN: doing ContinueThread\n"));
755 /* ToDo: check assertion
756 ASSERT(run_queue_hd != (StgTSO*)NULL &&
757 run_queue_hd != END_TSO_QUEUE);
759 /* Ignore ContinueThreads for fetching threads (if synchr comm) */
760 if (!RtsFlags.GranFlags.DoAsyncFetch &&
761 procStatus[CurrentProc]==Fetching) {
762 belch("ghuH: Spurious ContinueThread while Fetching ignored; TSO %d (%p) [PE %d]",
763 CurrentTSO->id, CurrentTSO, CurrentProc);
766 /* Ignore ContinueThreads for completed threads */
767 if (CurrentTSO->what_next == ThreadComplete) {
768 belch("ghuH: found a ContinueThread event for completed thread %d (%p) [PE %d] (ignoring ContinueThread)",
769 CurrentTSO->id, CurrentTSO, CurrentProc);
772 /* Ignore ContinueThreads for threads that are being migrated */
773 if (PROCS(CurrentTSO)==Nowhere) {
774 belch("ghuH: trying to run the migrating TSO %d (%p) [PE %d] (ignoring ContinueThread)",
775 CurrentTSO->id, CurrentTSO, CurrentProc);
778 /* The thread should be at the beginning of the run queue */
779 if (CurrentTSO!=run_queue_hds[CurrentProc]) {
780 belch("ghuH: TSO %d (%p) [PE %d] is not at the start of the run_queue when doing a ContinueThread",
781 CurrentTSO->id, CurrentTSO, CurrentProc);
782 break; // run the thread anyway
785 new_event(proc, proc, CurrentTime[proc],
787 (StgTSO*)NULL, (StgClosure*)NULL, (rtsSpark*)NULL);
789 */ /* Catches superfluous CONTINUEs -- should be unnecessary */
790 break; // now actually run the thread; DaH Qu'vam yImuHbej
793 do_the_fetchnode(event);
794 goto next_thread; /* handle next event in event queue */
797 do_the_globalblock(event);
798 goto next_thread; /* handle next event in event queue */
801 do_the_fetchreply(event);
802 goto next_thread; /* handle next event in event queue */
804 case UnblockThread: /* Move from the blocked queue to the tail of */
805 do_the_unblock(event);
806 goto next_thread; /* handle next event in event queue */
808 case ResumeThread: /* Move from the blocked queue to the tail of */
809 /* the runnable queue ( i.e. Qu' SImqa'lu') */
810 event->tso->gran.blocktime +=
811 CurrentTime[CurrentProc] - event->tso->gran.blockedat;
812 do_the_startthread(event);
813 goto next_thread; /* handle next event in event queue */
816 do_the_startthread(event);
817 goto next_thread; /* handle next event in event queue */
820 do_the_movethread(event);
821 goto next_thread; /* handle next event in event queue */
824 do_the_movespark(event);
825 goto next_thread; /* handle next event in event queue */
828 do_the_findwork(event);
829 goto next_thread; /* handle next event in event queue */
832 barf("Illegal event type %u\n", event->evttype);
835 /* This point was scheduler_loop in the old RTS */
837 IF_DEBUG(gran, belch("GRAN: after main switch"));
839 TimeOfLastEvent = CurrentTime[CurrentProc];
840 TimeOfNextEvent = get_time_of_next_event();
841 IgnoreEvents=(TimeOfNextEvent==0); // HWL HACK
842 // CurrentTSO = ThreadQueueHd;
844 IF_DEBUG(gran, belch("GRAN: time of next event is: %ld",
847 if (RtsFlags.GranFlags.Light)
848 GranSimLight_leave_system(event, &ActiveTSO);
850 EndOfTimeSlice = CurrentTime[CurrentProc]+RtsFlags.GranFlags.time_slice;
853 belch("GRAN: end of time-slice is %#lx", EndOfTimeSlice));
855 /* in a GranSim setup the TSO stays on the run queue */
857 /* Take a thread from the run queue. */
858 t = POP_RUN_QUEUE(); // take_off_run_queue(t);
861 fprintf(stderr, "GRAN: About to run current thread, which is\n");
864 context_switch = 0; // turned on via GranYield, checking events and time slice
867 DumpGranEvent(GR_SCHEDULE, t));
869 procStatus[CurrentProc] = Busy;
872 if (PendingFetches != END_BF_QUEUE) {
876 /* ToDo: phps merge with spark activation above */
877 /* check whether we have local work and send requests if we have none */
878 if (EMPTY_RUN_QUEUE()) { /* no runnable threads */
879 /* :-[ no local threads => look out for local sparks */
880 /* the spark pool for the current PE */
881 pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable
882 if (advisory_thread_count < RtsFlags.ParFlags.maxThreads &&
883 pool->hd < pool->tl) {
885 * ToDo: add GC code check that we really have enough heap afterwards!!
887 * If we're here (no runnable threads) and we have pending
888 * sparks, we must have a space problem. Get enough space
889 * to turn one of those pending sparks into a
893 spark = findSpark(rtsFalse); /* get a spark */
894 if (spark != (rtsSpark) NULL) {
895 tso = activateSpark(spark); /* turn the spark into a thread */
896 IF_PAR_DEBUG(schedule,
897 belch("==== schedule: Created TSO %d (%p); %d threads active",
898 tso->id, tso, advisory_thread_count));
900 if (tso==END_TSO_QUEUE) { /* failed to activate spark->back to loop */
901 belch("==^^ failed to activate spark");
903 } /* otherwise fall through & pick-up new tso */
905 IF_PAR_DEBUG(verbose,
906 belch("==^^ no local sparks (spark pool contains only NFs: %d)",
907 spark_queue_len(pool)));
912 /* If we still have no work we need to send a FISH to get a spark
915 if (EMPTY_RUN_QUEUE()) {
916 /* =8-[ no local sparks => look for work on other PEs */
918 * We really have absolutely no work. Send out a fish
919 * (there may be some out there already), and wait for
920 * something to arrive. We clearly can't run any threads
921 * until a SCHEDULE or RESUME arrives, and so that's what
922 * we're hoping to see. (Of course, we still have to
923 * respond to other types of messages.)
925 TIME now = msTime() /*CURRENT_TIME*/;
926 IF_PAR_DEBUG(verbose,
927 belch("-- now=%ld", now));
928 IF_PAR_DEBUG(verbose,
929 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
930 (last_fish_arrived_at!=0 &&
931 last_fish_arrived_at+RtsFlags.ParFlags.fishDelay > now)) {
932 belch("--$$ delaying FISH until %ld (last fish %ld, delay %ld, now %ld)",
933 last_fish_arrived_at+RtsFlags.ParFlags.fishDelay,
934 last_fish_arrived_at,
935 RtsFlags.ParFlags.fishDelay, now);
938 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
939 (last_fish_arrived_at==0 ||
940 (last_fish_arrived_at+RtsFlags.ParFlags.fishDelay <= now))) {
941 /* outstandingFishes is set in sendFish, processFish;
942 avoid flooding system with fishes via delay */
944 sendFish(pe, mytid, NEW_FISH_AGE, NEW_FISH_HISTORY,
947 // Global statistics: count no. of fishes
948 if (RtsFlags.ParFlags.ParStats.Global &&
949 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
950 globalParStats.tot_fish_mess++;
954 receivedFinish = processMessages();
957 } else if (PacketsWaiting()) { /* Look for incoming messages */
958 receivedFinish = processMessages();
961 /* Now we are sure that we have some work available */
962 ASSERT(run_queue_hd != END_TSO_QUEUE);
964 /* Take a thread from the run queue, if we have work */
965 t = POP_RUN_QUEUE(); // take_off_run_queue(END_TSO_QUEUE);
966 IF_DEBUG(sanity,checkTSO(t));
968 /* ToDo: write something to the log-file
969 if (RTSflags.ParFlags.granSimStats && !sameThread)
970 DumpGranEvent(GR_SCHEDULE, RunnableThreadsHd);
974 /* the spark pool for the current PE */
975 pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable
978 belch("--=^ %d threads, %d sparks on [%#x]",
979 run_queue_len(), spark_queue_len(pool), CURRENT_PROC));
982 if (0 && RtsFlags.ParFlags.ParStats.Full &&
983 t && LastTSO && t->id != LastTSO->id &&
984 LastTSO->why_blocked == NotBlocked &&
985 LastTSO->what_next != ThreadComplete) {
986 // if previously scheduled TSO not blocked we have to record the context switch
987 DumpVeryRawGranEvent(TimeOfLastYield, CURRENT_PROC, CURRENT_PROC,
988 GR_DESCHEDULE, LastTSO, (StgClosure *)NULL, 0, 0);
991 if (RtsFlags.ParFlags.ParStats.Full &&
992 (emitSchedule /* forced emit */ ||
993 (t && LastTSO && t->id != LastTSO->id))) {
995 we are running a different TSO, so write a schedule event to log file
996 NB: If we use fair scheduling we also have to write a deschedule
997 event for LastTSO; with unfair scheduling we know that the
998 previous tso has blocked whenever we switch to another tso, so
999 we don't need it in GUM for now
1001 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
1002 GR_SCHEDULE, t, (StgClosure *)NULL, 0, 0);
1003 emitSchedule = rtsFalse;
1007 #else /* !GRAN && !PAR */
1009 /* grab a thread from the run queue */
1010 ASSERT(run_queue_hd != END_TSO_QUEUE);
1011 t = POP_RUN_QUEUE();
1012 // Sanity check the thread we're about to run. This can be
1013 // expensive if there is lots of thread switching going on...
1014 IF_DEBUG(sanity,checkTSO(t));
1017 cap->r.rCurrentTSO = t;
1019 /* context switches are now initiated by the timer signal, unless
1020 * the user specified "context switch as often as possible", with
1025 RtsFlags.ProfFlags.profileInterval == 0 ||
1027 (RtsFlags.ConcFlags.ctxtSwitchTicks == 0
1028 && (run_queue_hd != END_TSO_QUEUE
1029 || blocked_queue_hd != END_TSO_QUEUE
1030 || sleeping_queue != END_TSO_QUEUE)))
1035 RELEASE_LOCK(&sched_mutex);
1037 IF_DEBUG(scheduler, sched_belch("-->> Running TSO %ld (%p) %s ...",
1038 t->id, t, whatNext_strs[t->what_next]));
1041 startHeapProfTimer();
1044 /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
1045 /* Run the current thread
1047 switch (cap->r.rCurrentTSO->what_next) {
1049 case ThreadComplete:
1050 /* Thread already finished, return to scheduler. */
1051 ret = ThreadFinished;
1053 case ThreadEnterGHC:
1054 ret = StgRun((StgFunPtr) stg_enterStackTop, &cap->r);
1057 ret = StgRun((StgFunPtr) stg_returnToStackTop, &cap->r);
1059 case ThreadEnterInterp:
1060 ret = interpretBCO(cap);
1063 barf("schedule: invalid what_next field");
1065 /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
1067 /* Costs for the scheduler are assigned to CCS_SYSTEM */
1069 stopHeapProfTimer();
1073 ACQUIRE_LOCK(&sched_mutex);
1076 IF_DEBUG(scheduler,fprintf(stderr,"scheduler (task %ld): ", osThreadId()););
1077 #elif !defined(GRAN) && !defined(PAR)
1078 IF_DEBUG(scheduler,fprintf(stderr,"scheduler: "););
1080 t = cap->r.rCurrentTSO;
1083 /* HACK 675: if the last thread didn't yield, make sure to print a
1084 SCHEDULE event to the log file when StgRunning the next thread, even
1085 if it is the same one as before */
1087 TimeOfLastYield = CURRENT_TIME;
1093 IF_DEBUG(gran, DumpGranEvent(GR_DESCHEDULE, t));
1094 globalGranStats.tot_heapover++;
1096 globalParStats.tot_heapover++;
1099 // did the task ask for a large block?
1100 if (cap->r.rHpAlloc > BLOCK_SIZE_W) {
1101 // if so, get one and push it on the front of the nursery.
1105 blocks = (nat)BLOCK_ROUND_UP(cap->r.rHpAlloc * sizeof(W_)) / BLOCK_SIZE;
1107 IF_DEBUG(scheduler,belch("--<< thread %ld (%p; %s) stopped: requesting a large block (size %d)",
1109 whatNext_strs[t->what_next], blocks));
1111 // don't do this if it would push us over the
1112 // alloc_blocks_lim limit; we'll GC first.
1113 if (alloc_blocks + blocks < alloc_blocks_lim) {
1115 alloc_blocks += blocks;
1116 bd = allocGroup( blocks );
1118 // link the new group into the list
1119 bd->link = cap->r.rCurrentNursery;
1120 bd->u.back = cap->r.rCurrentNursery->u.back;
1121 if (cap->r.rCurrentNursery->u.back != NULL) {
1122 cap->r.rCurrentNursery->u.back->link = bd;
1124 ASSERT(g0s0->blocks == cap->r.rCurrentNursery &&
1125 g0s0->blocks == cap->r.rNursery);
1126 cap->r.rNursery = g0s0->blocks = bd;
1128 cap->r.rCurrentNursery->u.back = bd;
1130 // initialise it as a nursery block. We initialise the
1131 // step, gen_no, and flags field of *every* sub-block in
1132 // this large block, because this is easier than making
1133 // sure that we always find the block head of a large
1134 // block whenever we call Bdescr() (eg. evacuate() and
1135 // isAlive() in the GC would both have to do this, at
1139 for (x = bd; x < bd + blocks; x++) {
1147 // don't forget to update the block count in g0s0.
1148 g0s0->n_blocks += blocks;
1149 ASSERT(countBlocks(g0s0->blocks) == g0s0->n_blocks);
1151 // now update the nursery to point to the new block
1152 cap->r.rCurrentNursery = bd;
1154 // we might be unlucky and have another thread get on the
1155 // run queue before us and steal the large block, but in that
1156 // case the thread will just end up requesting another large
1158 PUSH_ON_RUN_QUEUE(t);
1163 /* make all the running tasks block on a condition variable,
1164 * maybe set context_switch and wait till they all pile in,
1165 * then have them wait on a GC condition variable.
1167 IF_DEBUG(scheduler,belch("--<< thread %ld (%p; %s) stopped: HeapOverflow",
1168 t->id, t, whatNext_strs[t->what_next]));
1171 ASSERT(!is_on_queue(t,CurrentProc));
1173 /* Currently we emit a DESCHEDULE event before GC in GUM.
1174 ToDo: either add separate event to distinguish SYSTEM time from rest
1175 or just nuke this DESCHEDULE (and the following SCHEDULE) */
1176 if (0 && RtsFlags.ParFlags.ParStats.Full) {
1177 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
1178 GR_DESCHEDULE, t, (StgClosure *)NULL, 0, 0);
1179 emitSchedule = rtsTrue;
1183 ready_to_gc = rtsTrue;
1184 context_switch = 1; /* stop other threads ASAP */
1185 PUSH_ON_RUN_QUEUE(t);
1186 /* actual GC is done at the end of the while loop */
1192 DumpGranEvent(GR_DESCHEDULE, t));
1193 globalGranStats.tot_stackover++;
1196 // DumpGranEvent(GR_DESCHEDULE, t);
1197 globalParStats.tot_stackover++;
1199 IF_DEBUG(scheduler,belch("--<< thread %ld (%p; %s) stopped, StackOverflow",
1200 t->id, t, whatNext_strs[t->what_next]));
1201 /* just adjust the stack for this thread, then pop it back
1207 /* enlarge the stack */
1208 StgTSO *new_t = threadStackOverflow(t);
1210 /* This TSO has moved, so update any pointers to it from the
1211 * main thread stack. It better not be on any other queues...
1212 * (it shouldn't be).
1214 for (m = main_threads; m != NULL; m = m->link) {
1219 threadPaused(new_t);
1220 PUSH_ON_RUN_QUEUE(new_t);
1224 case ThreadYielding:
1227 DumpGranEvent(GR_DESCHEDULE, t));
1228 globalGranStats.tot_yields++;
1231 // DumpGranEvent(GR_DESCHEDULE, t);
1232 globalParStats.tot_yields++;
1234 /* put the thread back on the run queue. Then, if we're ready to
1235 * GC, check whether this is the last task to stop. If so, wake
1236 * up the GC thread. getThread will block during a GC until the
1240 if (t->what_next == ThreadEnterInterp) {
1241 /* ToDo: or maybe a timer expired when we were in Hugs?
1242 * or maybe someone hit ctrl-C
1244 belch("--<< thread %ld (%p; %s) stopped to switch to Hugs",
1245 t->id, t, whatNext_strs[t->what_next]);
1247 belch("--<< thread %ld (%p; %s) stopped, yielding",
1248 t->id, t, whatNext_strs[t->what_next]);
1255 //belch("&& Doing sanity check on yielding TSO %ld.", t->id);
1257 ASSERT(t->link == END_TSO_QUEUE);
1259 ASSERT(!is_on_queue(t,CurrentProc));
1262 //belch("&& Doing sanity check on all ThreadQueues (and their TSOs).");
1263 checkThreadQsSanity(rtsTrue));
1266 if (RtsFlags.ParFlags.doFairScheduling) {
1267 /* this does round-robin scheduling; good for concurrency */
1268 APPEND_TO_RUN_QUEUE(t);
1270 /* this does unfair scheduling; good for parallelism */
1271 PUSH_ON_RUN_QUEUE(t);
1274 /* this does round-robin scheduling; good for concurrency */
1275 APPEND_TO_RUN_QUEUE(t);
1278 /* add a ContinueThread event to actually process the thread */
1279 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1281 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1283 belch("GRAN: eventq and runnableq after adding yielded thread to queue again:");
1292 belch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: ",
1293 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)));
1294 if (t->block_info.closure!=(StgClosure*)NULL) print_bq(t->block_info.closure));
1296 // ??? needed; should emit block before
1298 DumpGranEvent(GR_DESCHEDULE, t));
1299 prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
1302 ASSERT(procStatus[CurrentProc]==Busy ||
1303 ((procStatus[CurrentProc]==Fetching) &&
1304 (t->block_info.closure!=(StgClosure*)NULL)));
1305 if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
1306 !(!RtsFlags.GranFlags.DoAsyncFetch &&
1307 procStatus[CurrentProc]==Fetching))
1308 procStatus[CurrentProc] = Idle;
1312 belch("--<< thread %ld (%p; %s) stopped, blocking on node %p with BQ: ",
1313 t->id, t, whatNext_strs[t->what_next], t->block_info.closure));
1316 if (t->block_info.closure!=(StgClosure*)NULL)
1317 print_bq(t->block_info.closure));
1319 /* Send a fetch (if BlockedOnGA) and dump event to log file */
1322 /* whatever we schedule next, we must log that schedule */
1323 emitSchedule = rtsTrue;
1326 /* don't need to do anything. Either the thread is blocked on
1327 * I/O, in which case we'll have called addToBlockedQueue
1328 * previously, or it's blocked on an MVar or Blackhole, in which
1329 * case it'll be on the relevant queue already.
1332 fprintf(stderr, "--<< thread %d (%p) stopped: ", t->id, t);
1333 printThreadBlockage(t);
1334 fprintf(stderr, "\n"));
1336 /* Only for dumping event to log file
1337 ToDo: do I need this in GranSim, too?
1344 case ThreadFinished:
1345 /* Need to check whether this was a main thread, and if so, signal
1346 * the task that started it with the return value. If we have no
1347 * more main threads, we probably need to stop all the tasks until
1350 /* We also end up here if the thread kills itself with an
1351 * uncaught exception, see Exception.hc.
1353 IF_DEBUG(scheduler,belch("--++ thread %d (%p) finished", t->id, t));
1355 endThread(t, CurrentProc); // clean-up the thread
1357 /* For now all are advisory -- HWL */
1358 //if(t->priority==AdvisoryPriority) ??
1359 advisory_thread_count--;
1362 if(t->dist.priority==RevalPriority)
1366 if (RtsFlags.ParFlags.ParStats.Full &&
1367 !RtsFlags.ParFlags.ParStats.Suppressed)
1368 DumpEndEvent(CURRENT_PROC, t, rtsFalse /* not mandatory */);
1373 barf("schedule: invalid thread return code %d", (int)ret);
1377 if (RtsFlags.ProfFlags.profileInterval==0 || performHeapProfile) {
1378 GarbageCollect(GetRoots, rtsTrue);
1380 performHeapProfile = rtsFalse;
1381 ready_to_gc = rtsFalse; // we already GC'd
1387 && allFreeCapabilities()
1390 /* everybody back, start the GC.
1391 * Could do it in this thread, or signal a condition var
1392 * to do it in another thread. Either way, we need to
1393 * broadcast on gc_pending_cond afterward.
1395 #if defined(RTS_SUPPORTS_THREADS)
1396 IF_DEBUG(scheduler,sched_belch("doing GC"));
1398 GarbageCollect(GetRoots,rtsFalse);
1399 ready_to_gc = rtsFalse;
1401 broadcastCondition(&gc_pending_cond);
1404 /* add a ContinueThread event to continue execution of current thread */
1405 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1407 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1409 fprintf(stderr, "GRAN: eventq and runnableq after Garbage collection:\n");
1417 IF_GRAN_DEBUG(unused,
1418 print_eventq(EventHd));
1420 event = get_next_event();
1423 /* ToDo: wait for next message to arrive rather than busy wait */
1426 } /* end of while(1) */
1428 IF_PAR_DEBUG(verbose,
1429 belch("== Leaving schedule() after having received Finish"));
1432 /* ---------------------------------------------------------------------------
1433 * Singleton fork(). Do not copy any running threads.
1434 * ------------------------------------------------------------------------- */
1436 StgInt forkProcess(StgTSO* tso) {
1438 #ifndef mingw32_TARGET_OS
1444 IF_DEBUG(scheduler,sched_belch("forking!"));
1447 if (pid) { /* parent */
1449 /* just return the pid */
1451 } else { /* child */
1452 /* wipe all other threads */
1453 run_queue_hd = run_queue_tl = tso;
1454 tso->link = END_TSO_QUEUE;
1456 /* When clearing out the threads, we need to ensure
1457 that a 'main thread' is left behind; if there isn't,
1458 the Scheduler will shutdown next time it is entered.
1460 ==> we don't kill a thread that's on the main_threads
1461 list (nor the current thread.)
1463 [ Attempts at implementing the more ambitious scheme of
1464 killing the main_threads also, and then adding the
1465 current thread onto the main_threads list if it wasn't
1466 there already, failed -- waitThread() (for one) wasn't
1467 up to it. If it proves to be desirable to also kill
1468 the main threads, then this scheme will have to be
1469 revisited (and fully debugged!)
1474 /* DO NOT TOUCH THE QUEUES directly because most of the code around
1475 us is picky about finding the thread still in its queue when
1476 handling the deleteThread() */
1478 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
1481 /* Don't kill the current thread.. */
1482 if (t->id == tso->id) continue;
1484 /* ..or a main thread */
1485 for (m = main_threads; m != NULL; m = m->link) {
1486 if (m->tso->id == t->id) {
1498 barf("forkProcess#: primop not implemented for mingw32, sorry! (%u)\n", tso->id);
1499 /* pointlessly printing out the TSOs 'id' to avoid CC unused warning. */
1501 #endif /* mingw32 */
1504 /* ---------------------------------------------------------------------------
1505 * deleteAllThreads(): kill all the live threads.
1507 * This is used when we catch a user interrupt (^C), before performing
1508 * any necessary cleanups and running finalizers.
1510 * Locks: sched_mutex held.
1511 * ------------------------------------------------------------------------- */
1513 void deleteAllThreads ( void )
1516 IF_DEBUG(scheduler,sched_belch("deleting all threads"));
1517 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
1518 next = t->global_link;
1521 run_queue_hd = run_queue_tl = END_TSO_QUEUE;
1522 blocked_queue_hd = blocked_queue_tl = END_TSO_QUEUE;
1523 sleeping_queue = END_TSO_QUEUE;
1526 /* startThread and insertThread are now in GranSim.c -- HWL */
1529 //@node Suspend and Resume, Run queue code, Main scheduling loop, Main scheduling code
1530 //@subsection Suspend and Resume
1532 /* ---------------------------------------------------------------------------
1533 * Suspending & resuming Haskell threads.
1535 * When making a "safe" call to C (aka _ccall_GC), the task gives back
1536 * its capability before calling the C function. This allows another
1537 * task to pick up the capability and carry on running Haskell
1538 * threads. It also means that if the C call blocks, it won't lock
1541 * The Haskell thread making the C call is put to sleep for the
1542 * duration of the call, on the susepended_ccalling_threads queue. We
1543 * give out a token to the task, which it can use to resume the thread
1544 * on return from the C function.
1545 * ------------------------------------------------------------------------- */
1548 suspendThread( StgRegTable *reg,
1550 #if !defined(RTS_SUPPORTS_THREADS) && !defined(DEBUG)
1558 /* assume that *reg is a pointer to the StgRegTable part
1561 cap = (Capability *)((void *)reg - sizeof(StgFunTable));
1563 ACQUIRE_LOCK(&sched_mutex);
1566 sched_belch("thread %d did a _ccall_gc (is_concurrent: %d)", cap->r.rCurrentTSO->id,concCall));
1568 threadPaused(cap->r.rCurrentTSO);
1569 cap->r.rCurrentTSO->link = suspended_ccalling_threads;
1570 suspended_ccalling_threads = cap->r.rCurrentTSO;
1572 #if defined(RTS_SUPPORTS_THREADS)
1573 cap->r.rCurrentTSO->why_blocked = BlockedOnCCall;
1576 /* Use the thread ID as the token; it should be unique */
1577 tok = cap->r.rCurrentTSO->id;
1579 /* Hand back capability */
1580 releaseCapability(cap);
1582 #if defined(RTS_SUPPORTS_THREADS)
1583 /* Preparing to leave the RTS, so ensure there's a native thread/task
1584 waiting to take over.
1586 ToDo: optimise this and only create a new task if there's a need
1587 for one (i.e., if there's only one Concurrent Haskell thread alive,
1588 there's no need to create a new task).
1590 IF_DEBUG(scheduler, sched_belch("worker thread (%d): leaving RTS", tok));
1592 startTask(taskStart);
1596 /* Other threads _might_ be available for execution; signal this */
1598 RELEASE_LOCK(&sched_mutex);
1603 resumeThread( StgInt tok,
1605 #if !defined(RTS_SUPPORTS_THREADS)
1610 StgTSO *tso, **prev;
1613 #if defined(RTS_SUPPORTS_THREADS)
1614 /* Wait for permission to re-enter the RTS with the result. */
1616 ACQUIRE_LOCK(&sched_mutex);
1617 grabReturnCapability(&sched_mutex, &cap);
1619 grabCapability(&cap);
1622 grabCapability(&cap);
1625 /* Remove the thread off of the suspended list */
1626 prev = &suspended_ccalling_threads;
1627 for (tso = suspended_ccalling_threads;
1628 tso != END_TSO_QUEUE;
1629 prev = &tso->link, tso = tso->link) {
1630 if (tso->id == (StgThreadID)tok) {
1635 if (tso == END_TSO_QUEUE) {
1636 barf("resumeThread: thread not found");
1638 tso->link = END_TSO_QUEUE;
1639 /* Reset blocking status */
1640 tso->why_blocked = NotBlocked;
1642 cap->r.rCurrentTSO = tso;
1643 RELEASE_LOCK(&sched_mutex);
1648 /* ---------------------------------------------------------------------------
1650 * ------------------------------------------------------------------------ */
1651 static void unblockThread(StgTSO *tso);
1653 /* ---------------------------------------------------------------------------
1654 * Comparing Thread ids.
1656 * This is used from STG land in the implementation of the
1657 * instances of Eq/Ord for ThreadIds.
1658 * ------------------------------------------------------------------------ */
1661 cmp_thread(StgPtr tso1, StgPtr tso2)
1663 StgThreadID id1 = ((StgTSO *)tso1)->id;
1664 StgThreadID id2 = ((StgTSO *)tso2)->id;
1666 if (id1 < id2) return (-1);
1667 if (id1 > id2) return 1;
1671 /* ---------------------------------------------------------------------------
1672 * Fetching the ThreadID from an StgTSO.
1674 * This is used in the implementation of Show for ThreadIds.
1675 * ------------------------------------------------------------------------ */
1677 rts_getThreadId(StgPtr tso)
1679 return ((StgTSO *)tso)->id;
1684 labelThread(StgPtr tso, char *label)
1689 /* Caveat: Once set, you can only set the thread name to "" */
1690 len = strlen(label)+1;
1693 fprintf(stderr,"insufficient memory for labelThread!\n");
1695 strncpy(buf,label,len);
1696 /* Update will free the old memory for us */
1697 updateThreadLabel((StgWord)tso,buf);
1701 /* ---------------------------------------------------------------------------
1702 Create a new thread.
1704 The new thread starts with the given stack size. Before the
1705 scheduler can run, however, this thread needs to have a closure
1706 (and possibly some arguments) pushed on its stack. See
1707 pushClosure() in Schedule.h.
1709 createGenThread() and createIOThread() (in SchedAPI.h) are
1710 convenient packaged versions of this function.
1712 currently pri (priority) is only used in a GRAN setup -- HWL
1713 ------------------------------------------------------------------------ */
1714 //@cindex createThread
1716 /* currently pri (priority) is only used in a GRAN setup -- HWL */
1718 createThread(nat size, StgInt pri)
1721 createThread(nat size)
1728 /* First check whether we should create a thread at all */
1730 /* check that no more than RtsFlags.ParFlags.maxThreads threads are created */
1731 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads) {
1733 belch("{createThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
1734 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
1735 return END_TSO_QUEUE;
1741 ASSERT(!RtsFlags.GranFlags.Light || CurrentProc==0);
1744 // ToDo: check whether size = stack_size - TSO_STRUCT_SIZEW
1746 /* catch ridiculously small stack sizes */
1747 if (size < MIN_STACK_WORDS + TSO_STRUCT_SIZEW) {
1748 size = MIN_STACK_WORDS + TSO_STRUCT_SIZEW;
1751 stack_size = size - TSO_STRUCT_SIZEW;
1753 tso = (StgTSO *)allocate(size);
1754 TICK_ALLOC_TSO(stack_size, 0);
1756 SET_HDR(tso, &stg_TSO_info, CCS_SYSTEM);
1758 SET_GRAN_HDR(tso, ThisPE);
1760 tso->what_next = ThreadEnterGHC;
1762 /* tso->id needs to be unique. For now we use a heavyweight mutex to
1763 * protect the increment operation on next_thread_id.
1764 * In future, we could use an atomic increment instead.
1766 ACQUIRE_LOCK(&thread_id_mutex);
1767 tso->id = next_thread_id++;
1768 RELEASE_LOCK(&thread_id_mutex);
1770 tso->why_blocked = NotBlocked;
1771 tso->blocked_exceptions = NULL;
1773 tso->stack_size = stack_size;
1774 tso->max_stack_size = round_to_mblocks(RtsFlags.GcFlags.maxStkSize)
1776 tso->sp = (P_)&(tso->stack) + stack_size;
1779 tso->prof.CCCS = CCS_MAIN;
1782 /* put a stop frame on the stack */
1783 tso->sp -= sizeofW(StgStopFrame);
1784 SET_HDR((StgClosure*)tso->sp,(StgInfoTable *)&stg_stop_thread_info,CCS_SYSTEM);
1785 tso->su = (StgUpdateFrame*)tso->sp;
1789 tso->link = END_TSO_QUEUE;
1790 /* uses more flexible routine in GranSim */
1791 insertThread(tso, CurrentProc);
1793 /* In a non-GranSim setup the pushing of a TSO onto the runq is separated
1799 if (RtsFlags.GranFlags.GranSimStats.Full)
1800 DumpGranEvent(GR_START,tso);
1802 if (RtsFlags.ParFlags.ParStats.Full)
1803 DumpGranEvent(GR_STARTQ,tso);
1804 /* HACk to avoid SCHEDULE
1808 /* Link the new thread on the global thread list.
1810 tso->global_link = all_threads;
1814 tso->dist.priority = MandatoryPriority; //by default that is...
1818 tso->gran.pri = pri;
1820 tso->gran.magic = TSO_MAGIC; // debugging only
1822 tso->gran.sparkname = 0;
1823 tso->gran.startedat = CURRENT_TIME;
1824 tso->gran.exported = 0;
1825 tso->gran.basicblocks = 0;
1826 tso->gran.allocs = 0;
1827 tso->gran.exectime = 0;
1828 tso->gran.fetchtime = 0;
1829 tso->gran.fetchcount = 0;
1830 tso->gran.blocktime = 0;
1831 tso->gran.blockcount = 0;
1832 tso->gran.blockedat = 0;
1833 tso->gran.globalsparks = 0;
1834 tso->gran.localsparks = 0;
1835 if (RtsFlags.GranFlags.Light)
1836 tso->gran.clock = Now; /* local clock */
1838 tso->gran.clock = 0;
1840 IF_DEBUG(gran,printTSO(tso));
1843 tso->par.magic = TSO_MAGIC; // debugging only
1845 tso->par.sparkname = 0;
1846 tso->par.startedat = CURRENT_TIME;
1847 tso->par.exported = 0;
1848 tso->par.basicblocks = 0;
1849 tso->par.allocs = 0;
1850 tso->par.exectime = 0;
1851 tso->par.fetchtime = 0;
1852 tso->par.fetchcount = 0;
1853 tso->par.blocktime = 0;
1854 tso->par.blockcount = 0;
1855 tso->par.blockedat = 0;
1856 tso->par.globalsparks = 0;
1857 tso->par.localsparks = 0;
1861 globalGranStats.tot_threads_created++;
1862 globalGranStats.threads_created_on_PE[CurrentProc]++;
1863 globalGranStats.tot_sq_len += spark_queue_len(CurrentProc);
1864 globalGranStats.tot_sq_probes++;
1866 // collect parallel global statistics (currently done together with GC stats)
1867 if (RtsFlags.ParFlags.ParStats.Global &&
1868 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
1869 //fprintf(stderr, "Creating thread %d @ %11.2f\n", tso->id, usertime());
1870 globalParStats.tot_threads_created++;
1876 belch("==__ schedule: Created TSO %d (%p);",
1877 CurrentProc, tso, tso->id));
1879 IF_PAR_DEBUG(verbose,
1880 belch("==__ schedule: Created TSO %d (%p); %d threads active",
1881 tso->id, tso, advisory_thread_count));
1883 IF_DEBUG(scheduler,sched_belch("created thread %ld, stack size = %lx words",
1884 tso->id, tso->stack_size));
1891 all parallel thread creation calls should fall through the following routine.
1894 createSparkThread(rtsSpark spark)
1896 ASSERT(spark != (rtsSpark)NULL);
1897 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads)
1899 barf("{createSparkThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
1900 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
1901 return END_TSO_QUEUE;
1905 tso = createThread(RtsFlags.GcFlags.initialStkSize);
1906 if (tso==END_TSO_QUEUE)
1907 barf("createSparkThread: Cannot create TSO");
1909 tso->priority = AdvisoryPriority;
1911 pushClosure(tso,spark);
1912 PUSH_ON_RUN_QUEUE(tso);
1913 advisory_thread_count++;
1920 Turn a spark into a thread.
1921 ToDo: fix for SMP (needs to acquire SCHED_MUTEX!)
1924 //@cindex activateSpark
1926 activateSpark (rtsSpark spark)
1930 tso = createSparkThread(spark);
1931 if (RtsFlags.ParFlags.ParStats.Full) {
1932 //ASSERT(run_queue_hd == END_TSO_QUEUE); // I think ...
1933 IF_PAR_DEBUG(verbose,
1934 belch("==^^ activateSpark: turning spark of closure %p (%s) into a thread",
1935 (StgClosure *)spark, info_type((StgClosure *)spark)));
1937 // ToDo: fwd info on local/global spark to thread -- HWL
1938 // tso->gran.exported = spark->exported;
1939 // tso->gran.locked = !spark->global;
1940 // tso->gran.sparkname = spark->name;
1946 static SchedulerStatus waitThread_(/*out*/StgMainThread* m
1947 #if defined(THREADED_RTS)
1948 , rtsBool blockWaiting
1953 /* ---------------------------------------------------------------------------
1956 * scheduleThread puts a thread on the head of the runnable queue.
1957 * This will usually be done immediately after a thread is created.
1958 * The caller of scheduleThread must create the thread using e.g.
1959 * createThread and push an appropriate closure
1960 * on this thread's stack before the scheduler is invoked.
1961 * ------------------------------------------------------------------------ */
1963 static void scheduleThread_ (StgTSO* tso, rtsBool createTask);
1966 scheduleThread_(StgTSO *tso
1967 , rtsBool createTask
1968 #if !defined(THREADED_RTS)
1973 ACQUIRE_LOCK(&sched_mutex);
1975 /* Put the new thread on the head of the runnable queue. The caller
1976 * better push an appropriate closure on this thread's stack
1977 * beforehand. In the SMP case, the thread may start running as
1978 * soon as we release the scheduler lock below.
1980 PUSH_ON_RUN_QUEUE(tso);
1981 #if defined(THREADED_RTS)
1982 /* If main() is scheduling a thread, don't bother creating a
1986 startTask(taskStart);
1992 IF_DEBUG(scheduler,printTSO(tso));
1994 RELEASE_LOCK(&sched_mutex);
1997 void scheduleThread(StgTSO* tso)
1999 scheduleThread_(tso, rtsFalse);
2003 scheduleWaitThread(StgTSO* tso, /*[out]*/HaskellObj* ret)
2007 m = stgMallocBytes(sizeof(StgMainThread), "waitThread");
2011 #if defined(RTS_SUPPORTS_THREADS)
2012 initCondition(&m->wakeup);
2015 /* Put the thread on the main-threads list prior to scheduling the TSO.
2016 Failure to do so introduces a race condition in the MT case (as
2017 identified by Wolfgang Thaller), whereby the new task/OS thread
2018 created by scheduleThread_() would complete prior to the thread
2019 that spawned it managed to put 'itself' on the main-threads list.
2020 The upshot of it all being that the worker thread wouldn't get to
2021 signal the completion of the its work item for the main thread to
2022 see (==> it got stuck waiting.) -- sof 6/02.
2024 ACQUIRE_LOCK(&sched_mutex);
2025 IF_DEBUG(scheduler, sched_belch("== scheduler: waiting for thread (%d)\n", tso->id));
2027 m->link = main_threads;
2030 /* Inefficient (scheduleThread_() acquires it again right away),
2031 * but obviously correct.
2033 RELEASE_LOCK(&sched_mutex);
2035 scheduleThread_(tso, rtsTrue);
2036 #if defined(THREADED_RTS)
2037 return waitThread_(m, rtsTrue);
2039 return waitThread_(m);
2043 /* ---------------------------------------------------------------------------
2046 * Initialise the scheduler. This resets all the queues - if the
2047 * queues contained any threads, they'll be garbage collected at the
2050 * ------------------------------------------------------------------------ */
2054 term_handler(int sig STG_UNUSED)
2057 ACQUIRE_LOCK(&term_mutex);
2059 RELEASE_LOCK(&term_mutex);
2070 for (i=0; i<=MAX_PROC; i++) {
2071 run_queue_hds[i] = END_TSO_QUEUE;
2072 run_queue_tls[i] = END_TSO_QUEUE;
2073 blocked_queue_hds[i] = END_TSO_QUEUE;
2074 blocked_queue_tls[i] = END_TSO_QUEUE;
2075 ccalling_threadss[i] = END_TSO_QUEUE;
2076 sleeping_queue = END_TSO_QUEUE;
2079 run_queue_hd = END_TSO_QUEUE;
2080 run_queue_tl = END_TSO_QUEUE;
2081 blocked_queue_hd = END_TSO_QUEUE;
2082 blocked_queue_tl = END_TSO_QUEUE;
2083 sleeping_queue = END_TSO_QUEUE;
2086 suspended_ccalling_threads = END_TSO_QUEUE;
2088 main_threads = NULL;
2089 all_threads = END_TSO_QUEUE;
2094 RtsFlags.ConcFlags.ctxtSwitchTicks =
2095 RtsFlags.ConcFlags.ctxtSwitchTime / TICK_MILLISECS;
2097 #if defined(RTS_SUPPORTS_THREADS)
2098 /* Initialise the mutex and condition variables used by
2100 initMutex(&sched_mutex);
2101 initMutex(&term_mutex);
2102 initMutex(&thread_id_mutex);
2104 initCondition(&thread_ready_cond);
2108 initCondition(&gc_pending_cond);
2111 #if defined(RTS_SUPPORTS_THREADS)
2112 ACQUIRE_LOCK(&sched_mutex);
2115 /* Install the SIGHUP handler */
2118 struct sigaction action,oact;
2120 action.sa_handler = term_handler;
2121 sigemptyset(&action.sa_mask);
2122 action.sa_flags = 0;
2123 if (sigaction(SIGTERM, &action, &oact) != 0) {
2124 barf("can't install TERM handler");
2129 /* A capability holds the state a native thread needs in
2130 * order to execute STG code. At least one capability is
2131 * floating around (only SMP builds have more than one).
2135 #if defined(RTS_SUPPORTS_THREADS)
2136 /* start our haskell execution tasks */
2138 startTaskManager(RtsFlags.ParFlags.nNodes, taskStart);
2140 startTaskManager(0,taskStart);
2144 #if /* defined(SMP) ||*/ defined(PAR)
2148 #if defined(RTS_SUPPORTS_THREADS)
2149 RELEASE_LOCK(&sched_mutex);
2155 exitScheduler( void )
2157 #if defined(RTS_SUPPORTS_THREADS)
2160 shutting_down_scheduler = rtsTrue;
2163 /* -----------------------------------------------------------------------------
2164 Managing the per-task allocation areas.
2166 Each capability comes with an allocation area. These are
2167 fixed-length block lists into which allocation can be done.
2169 ToDo: no support for two-space collection at the moment???
2170 -------------------------------------------------------------------------- */
2172 /* -----------------------------------------------------------------------------
2173 * waitThread is the external interface for running a new computation
2174 * and waiting for the result.
2176 * In the non-SMP case, we create a new main thread, push it on the
2177 * main-thread stack, and invoke the scheduler to run it. The
2178 * scheduler will return when the top main thread on the stack has
2179 * completed or died, and fill in the necessary fields of the
2180 * main_thread structure.
2182 * In the SMP case, we create a main thread as before, but we then
2183 * create a new condition variable and sleep on it. When our new
2184 * main thread has completed, we'll be woken up and the status/result
2185 * will be in the main_thread struct.
2186 * -------------------------------------------------------------------------- */
2189 howManyThreadsAvail ( void )
2193 for (q = run_queue_hd; q != END_TSO_QUEUE; q = q->link)
2195 for (q = blocked_queue_hd; q != END_TSO_QUEUE; q = q->link)
2197 for (q = sleeping_queue; q != END_TSO_QUEUE; q = q->link)
2203 finishAllThreads ( void )
2206 while (run_queue_hd != END_TSO_QUEUE) {
2207 waitThread ( run_queue_hd, NULL);
2209 while (blocked_queue_hd != END_TSO_QUEUE) {
2210 waitThread ( blocked_queue_hd, NULL);
2212 while (sleeping_queue != END_TSO_QUEUE) {
2213 waitThread ( blocked_queue_hd, NULL);
2216 (blocked_queue_hd != END_TSO_QUEUE ||
2217 run_queue_hd != END_TSO_QUEUE ||
2218 sleeping_queue != END_TSO_QUEUE);
2222 waitThread(StgTSO *tso, /*out*/StgClosure **ret)
2226 m = stgMallocBytes(sizeof(StgMainThread), "waitThread");
2230 #if defined(RTS_SUPPORTS_THREADS)
2231 initCondition(&m->wakeup);
2234 /* see scheduleWaitThread() comment */
2235 ACQUIRE_LOCK(&sched_mutex);
2236 IF_DEBUG(scheduler, sched_belch("== scheduler: waiting for thread (%d)\n", tso->id));
2237 m->link = main_threads;
2239 RELEASE_LOCK(&sched_mutex);
2241 IF_DEBUG(scheduler, sched_belch("== scheduler: waiting for thread (%d)\n", tso->id));
2242 #if defined(THREADED_RTS)
2243 return waitThread_(m, rtsFalse);
2245 return waitThread_(m);
2251 waitThread_(StgMainThread* m
2252 #if defined(THREADED_RTS)
2253 , rtsBool blockWaiting
2257 SchedulerStatus stat;
2259 IF_DEBUG(scheduler, sched_belch("== scheduler: new main thread (%d)\n", m->tso->id));
2261 #if defined(RTS_SUPPORTS_THREADS)
2263 # if defined(THREADED_RTS)
2264 if (!blockWaiting) {
2265 /* In the threaded case, the OS thread that called main()
2266 * gets to enter the RTS directly without going via another
2270 ASSERT(m->stat != NoStatus);
2274 ACQUIRE_LOCK(&sched_mutex);
2276 waitCondition(&m->wakeup, &sched_mutex);
2277 } while (m->stat == NoStatus);
2280 /* GranSim specific init */
2281 CurrentTSO = m->tso; // the TSO to run
2282 procStatus[MainProc] = Busy; // status of main PE
2283 CurrentProc = MainProc; // PE to run it on
2287 RELEASE_LOCK(&sched_mutex);
2289 ASSERT(m->stat != NoStatus);
2294 #if defined(RTS_SUPPORTS_THREADS)
2295 closeCondition(&m->wakeup);
2298 IF_DEBUG(scheduler, fprintf(stderr, "== scheduler: main thread (%d) finished\n",
2302 #if defined(THREADED_RTS)
2305 RELEASE_LOCK(&sched_mutex);
2310 //@node Run queue code, Garbage Collextion Routines, Suspend and Resume, Main scheduling code
2311 //@subsection Run queue code
2315 NB: In GranSim we have many run queues; run_queue_hd is actually a macro
2316 unfolding to run_queue_hds[CurrentProc], thus CurrentProc is an
2317 implicit global variable that has to be correct when calling these
2321 /* Put the new thread on the head of the runnable queue.
2322 * The caller of createThread better push an appropriate closure
2323 * on this thread's stack before the scheduler is invoked.
2325 static /* inline */ void
2326 add_to_run_queue(tso)
2329 ASSERT(tso!=run_queue_hd && tso!=run_queue_tl);
2330 tso->link = run_queue_hd;
2332 if (run_queue_tl == END_TSO_QUEUE) {
2337 /* Put the new thread at the end of the runnable queue. */
2338 static /* inline */ void
2339 push_on_run_queue(tso)
2342 ASSERT(get_itbl((StgClosure *)tso)->type == TSO);
2343 ASSERT(run_queue_hd!=NULL && run_queue_tl!=NULL);
2344 ASSERT(tso!=run_queue_hd && tso!=run_queue_tl);
2345 if (run_queue_hd == END_TSO_QUEUE) {
2348 run_queue_tl->link = tso;
2354 Should be inlined because it's used very often in schedule. The tso
2355 argument is actually only needed in GranSim, where we want to have the
2356 possibility to schedule *any* TSO on the run queue, irrespective of the
2357 actual ordering. Therefore, if tso is not the nil TSO then we traverse
2358 the run queue and dequeue the tso, adjusting the links in the queue.
2360 //@cindex take_off_run_queue
2361 static /* inline */ StgTSO*
2362 take_off_run_queue(StgTSO *tso) {
2366 qetlaHbogh Qu' ngaSbogh ghomDaQ {tso} yIteq!
2368 if tso is specified, unlink that tso from the run_queue (doesn't have
2369 to be at the beginning of the queue); GranSim only
2371 if (tso!=END_TSO_QUEUE) {
2372 /* find tso in queue */
2373 for (t=run_queue_hd, prev=END_TSO_QUEUE;
2374 t!=END_TSO_QUEUE && t!=tso;
2378 /* now actually dequeue the tso */
2379 if (prev!=END_TSO_QUEUE) {
2380 ASSERT(run_queue_hd!=t);
2381 prev->link = t->link;
2383 /* t is at beginning of thread queue */
2384 ASSERT(run_queue_hd==t);
2385 run_queue_hd = t->link;
2387 /* t is at end of thread queue */
2388 if (t->link==END_TSO_QUEUE) {
2389 ASSERT(t==run_queue_tl);
2390 run_queue_tl = prev;
2392 ASSERT(run_queue_tl!=t);
2394 t->link = END_TSO_QUEUE;
2396 /* take tso from the beginning of the queue; std concurrent code */
2398 if (t != END_TSO_QUEUE) {
2399 run_queue_hd = t->link;
2400 t->link = END_TSO_QUEUE;
2401 if (run_queue_hd == END_TSO_QUEUE) {
2402 run_queue_tl = END_TSO_QUEUE;
2411 //@node Garbage Collextion Routines, Blocking Queue Routines, Run queue code, Main scheduling code
2412 //@subsection Garbage Collextion Routines
2414 /* ---------------------------------------------------------------------------
2415 Where are the roots that we know about?
2417 - all the threads on the runnable queue
2418 - all the threads on the blocked queue
2419 - all the threads on the sleeping queue
2420 - all the thread currently executing a _ccall_GC
2421 - all the "main threads"
2423 ------------------------------------------------------------------------ */
2425 /* This has to be protected either by the scheduler monitor, or by the
2426 garbage collection monitor (probably the latter).
2431 GetRoots(evac_fn evac)
2436 for (i=0; i<=RtsFlags.GranFlags.proc; i++) {
2437 if ((run_queue_hds[i] != END_TSO_QUEUE) && ((run_queue_hds[i] != NULL)))
2438 evac((StgClosure **)&run_queue_hds[i]);
2439 if ((run_queue_tls[i] != END_TSO_QUEUE) && ((run_queue_tls[i] != NULL)))
2440 evac((StgClosure **)&run_queue_tls[i]);
2442 if ((blocked_queue_hds[i] != END_TSO_QUEUE) && ((blocked_queue_hds[i] != NULL)))
2443 evac((StgClosure **)&blocked_queue_hds[i]);
2444 if ((blocked_queue_tls[i] != END_TSO_QUEUE) && ((blocked_queue_tls[i] != NULL)))
2445 evac((StgClosure **)&blocked_queue_tls[i]);
2446 if ((ccalling_threadss[i] != END_TSO_QUEUE) && ((ccalling_threadss[i] != NULL)))
2447 evac((StgClosure **)&ccalling_threads[i]);
2454 if (run_queue_hd != END_TSO_QUEUE) {
2455 ASSERT(run_queue_tl != END_TSO_QUEUE);
2456 evac((StgClosure **)&run_queue_hd);
2457 evac((StgClosure **)&run_queue_tl);
2460 if (blocked_queue_hd != END_TSO_QUEUE) {
2461 ASSERT(blocked_queue_tl != END_TSO_QUEUE);
2462 evac((StgClosure **)&blocked_queue_hd);
2463 evac((StgClosure **)&blocked_queue_tl);
2466 if (sleeping_queue != END_TSO_QUEUE) {
2467 evac((StgClosure **)&sleeping_queue);
2471 if (suspended_ccalling_threads != END_TSO_QUEUE) {
2472 evac((StgClosure **)&suspended_ccalling_threads);
2475 #if defined(PAR) || defined(GRAN)
2476 markSparkQueue(evac);
2479 #ifndef mingw32_TARGET_OS
2480 // mark the signal handlers (signals should be already blocked)
2481 markSignalHandlers(evac);
2484 // main threads which have completed need to be retained until they
2485 // are dealt with in the main scheduler loop. They won't be
2486 // retained any other way: the GC will drop them from the
2487 // all_threads list, so we have to be careful to treat them as roots
2491 for (m = main_threads; m != NULL; m = m->link) {
2492 switch (m->tso->what_next) {
2493 case ThreadComplete:
2495 evac((StgClosure **)&m->tso);
2504 /* -----------------------------------------------------------------------------
2507 This is the interface to the garbage collector from Haskell land.
2508 We provide this so that external C code can allocate and garbage
2509 collect when called from Haskell via _ccall_GC.
2511 It might be useful to provide an interface whereby the programmer
2512 can specify more roots (ToDo).
2514 This needs to be protected by the GC condition variable above. KH.
2515 -------------------------------------------------------------------------- */
2517 static void (*extra_roots)(evac_fn);
2522 /* Obligated to hold this lock upon entry */
2523 ACQUIRE_LOCK(&sched_mutex);
2524 GarbageCollect(GetRoots,rtsFalse);
2525 RELEASE_LOCK(&sched_mutex);
2529 performMajorGC(void)
2531 ACQUIRE_LOCK(&sched_mutex);
2532 GarbageCollect(GetRoots,rtsTrue);
2533 RELEASE_LOCK(&sched_mutex);
2537 AllRoots(evac_fn evac)
2539 GetRoots(evac); // the scheduler's roots
2540 extra_roots(evac); // the user's roots
2544 performGCWithRoots(void (*get_roots)(evac_fn))
2546 ACQUIRE_LOCK(&sched_mutex);
2547 extra_roots = get_roots;
2548 GarbageCollect(AllRoots,rtsFalse);
2549 RELEASE_LOCK(&sched_mutex);
2552 /* -----------------------------------------------------------------------------
2555 If the thread has reached its maximum stack size, then raise the
2556 StackOverflow exception in the offending thread. Otherwise
2557 relocate the TSO into a larger chunk of memory and adjust its stack
2559 -------------------------------------------------------------------------- */
2562 threadStackOverflow(StgTSO *tso)
2564 nat new_stack_size, new_tso_size, diff, stack_words;
2568 IF_DEBUG(sanity,checkTSO(tso));
2569 if (tso->stack_size >= tso->max_stack_size) {
2572 belch("@@ threadStackOverflow of TSO %d (%p): stack too large (now %ld; max is %ld",
2573 tso->id, tso, tso->stack_size, tso->max_stack_size);
2574 /* If we're debugging, just print out the top of the stack */
2575 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2578 /* Send this thread the StackOverflow exception */
2579 raiseAsync(tso, (StgClosure *)stackOverflow_closure);
2583 /* Try to double the current stack size. If that takes us over the
2584 * maximum stack size for this thread, then use the maximum instead.
2585 * Finally round up so the TSO ends up as a whole number of blocks.
2587 new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
2588 new_tso_size = (nat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) +
2589 TSO_STRUCT_SIZE)/sizeof(W_);
2590 new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */
2591 new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
2593 IF_DEBUG(scheduler, fprintf(stderr,"== scheduler: increasing stack size from %d words to %d.\n", tso->stack_size, new_stack_size));
2595 dest = (StgTSO *)allocate(new_tso_size);
2596 TICK_ALLOC_TSO(new_stack_size,0);
2598 /* copy the TSO block and the old stack into the new area */
2599 memcpy(dest,tso,TSO_STRUCT_SIZE);
2600 stack_words = tso->stack + tso->stack_size - tso->sp;
2601 new_sp = (P_)dest + new_tso_size - stack_words;
2602 memcpy(new_sp, tso->sp, stack_words * sizeof(W_));
2604 /* relocate the stack pointers... */
2605 diff = (P_)new_sp - (P_)tso->sp; /* In *words* */
2606 dest->su = (StgUpdateFrame *) ((P_)dest->su + diff);
2608 dest->stack_size = new_stack_size;
2610 /* and relocate the update frame list */
2611 relocate_stack(dest, diff);
2613 /* Mark the old TSO as relocated. We have to check for relocated
2614 * TSOs in the garbage collector and any primops that deal with TSOs.
2616 * It's important to set the sp and su values to just beyond the end
2617 * of the stack, so we don't attempt to scavenge any part of the
2620 tso->what_next = ThreadRelocated;
2622 tso->sp = (P_)&(tso->stack[tso->stack_size]);
2623 tso->su = (StgUpdateFrame *)tso->sp;
2624 tso->why_blocked = NotBlocked;
2625 dest->mut_link = NULL;
2627 IF_PAR_DEBUG(verbose,
2628 belch("@@ threadStackOverflow of TSO %d (now at %p): stack size increased to %ld",
2629 tso->id, tso, tso->stack_size);
2630 /* If we're debugging, just print out the top of the stack */
2631 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2634 IF_DEBUG(sanity,checkTSO(tso));
2636 IF_DEBUG(scheduler,printTSO(dest));
2642 //@node Blocking Queue Routines, Exception Handling Routines, Garbage Collextion Routines, Main scheduling code
2643 //@subsection Blocking Queue Routines
2645 /* ---------------------------------------------------------------------------
2646 Wake up a queue that was blocked on some resource.
2647 ------------------------------------------------------------------------ */
2651 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2656 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2658 /* write RESUME events to log file and
2659 update blocked and fetch time (depending on type of the orig closure) */
2660 if (RtsFlags.ParFlags.ParStats.Full) {
2661 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
2662 GR_RESUMEQ, ((StgTSO *)bqe), ((StgTSO *)bqe)->block_info.closure,
2663 0, 0 /* spark_queue_len(ADVISORY_POOL) */);
2664 if (EMPTY_RUN_QUEUE())
2665 emitSchedule = rtsTrue;
2667 switch (get_itbl(node)->type) {
2669 ((StgTSO *)bqe)->par.fetchtime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2674 ((StgTSO *)bqe)->par.blocktime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2681 barf("{unblockOneLocked}Daq Qagh: unexpected closure in blocking queue");
2688 static StgBlockingQueueElement *
2689 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2692 PEs node_loc, tso_loc;
2694 node_loc = where_is(node); // should be lifted out of loop
2695 tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2696 tso_loc = where_is((StgClosure *)tso);
2697 if (IS_LOCAL_TO(PROCS(node),tso_loc)) { // TSO is local
2698 /* !fake_fetch => TSO is on CurrentProc is same as IS_LOCAL_TO */
2699 ASSERT(CurrentProc!=node_loc || tso_loc==CurrentProc);
2700 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.lunblocktime;
2701 // insertThread(tso, node_loc);
2702 new_event(tso_loc, tso_loc, CurrentTime[CurrentProc],
2704 tso, node, (rtsSpark*)NULL);
2705 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2708 } else { // TSO is remote (actually should be FMBQ)
2709 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.mpacktime +
2710 RtsFlags.GranFlags.Costs.gunblocktime +
2711 RtsFlags.GranFlags.Costs.latency;
2712 new_event(tso_loc, CurrentProc, CurrentTime[CurrentProc],
2714 tso, node, (rtsSpark*)NULL);
2715 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2718 /* the thread-queue-overhead is accounted for in either Resume or UnblockThread */
2720 fprintf(stderr," %s TSO %d (%p) [PE %d] (block_info.closure=%p) (next=%p) ,",
2721 (node_loc==tso_loc ? "Local" : "Global"),
2722 tso->id, tso, CurrentProc, tso->block_info.closure, tso->link));
2723 tso->block_info.closure = NULL;
2724 IF_DEBUG(scheduler,belch("-- Waking up thread %ld (%p)",
2728 static StgBlockingQueueElement *
2729 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2731 StgBlockingQueueElement *next;
2733 switch (get_itbl(bqe)->type) {
2735 ASSERT(((StgTSO *)bqe)->why_blocked != NotBlocked);
2736 /* if it's a TSO just push it onto the run_queue */
2738 // ((StgTSO *)bqe)->link = END_TSO_QUEUE; // debugging?
2739 PUSH_ON_RUN_QUEUE((StgTSO *)bqe);
2741 unblockCount(bqe, node);
2742 /* reset blocking status after dumping event */
2743 ((StgTSO *)bqe)->why_blocked = NotBlocked;
2747 /* if it's a BLOCKED_FETCH put it on the PendingFetches list */
2749 bqe->link = (StgBlockingQueueElement *)PendingFetches;
2750 PendingFetches = (StgBlockedFetch *)bqe;
2754 /* can ignore this case in a non-debugging setup;
2755 see comments on RBHSave closures above */
2757 /* check that the closure is an RBHSave closure */
2758 ASSERT(get_itbl((StgClosure *)bqe) == &stg_RBH_Save_0_info ||
2759 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_1_info ||
2760 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_2_info);
2764 barf("{unblockOneLocked}Daq Qagh: Unexpected IP (%#lx; %s) in blocking queue at %#lx\n",
2765 get_itbl((StgClosure *)bqe), info_type((StgClosure *)bqe),
2769 IF_PAR_DEBUG(bq, fprintf(stderr, ", %p (%s)", bqe, info_type((StgClosure*)bqe)));
2773 #else /* !GRAN && !PAR */
2775 unblockOneLocked(StgTSO *tso)
2779 ASSERT(get_itbl(tso)->type == TSO);
2780 ASSERT(tso->why_blocked != NotBlocked);
2781 tso->why_blocked = NotBlocked;
2783 PUSH_ON_RUN_QUEUE(tso);
2785 IF_DEBUG(scheduler,sched_belch("waking up thread %ld", tso->id));
2790 #if defined(GRAN) || defined(PAR)
2791 inline StgBlockingQueueElement *
2792 unblockOne(StgBlockingQueueElement *bqe, StgClosure *node)
2794 ACQUIRE_LOCK(&sched_mutex);
2795 bqe = unblockOneLocked(bqe, node);
2796 RELEASE_LOCK(&sched_mutex);
2801 unblockOne(StgTSO *tso)
2803 ACQUIRE_LOCK(&sched_mutex);
2804 tso = unblockOneLocked(tso);
2805 RELEASE_LOCK(&sched_mutex);
2812 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2814 StgBlockingQueueElement *bqe;
2819 belch("##-_ AwBQ for node %p on PE %d @ %ld by TSO %d (%p): ", \
2820 node, CurrentProc, CurrentTime[CurrentProc],
2821 CurrentTSO->id, CurrentTSO));
2823 node_loc = where_is(node);
2825 ASSERT(q == END_BQ_QUEUE ||
2826 get_itbl(q)->type == TSO || // q is either a TSO or an RBHSave
2827 get_itbl(q)->type == CONSTR); // closure (type constructor)
2828 ASSERT(is_unique(node));
2830 /* FAKE FETCH: magically copy the node to the tso's proc;
2831 no Fetch necessary because in reality the node should not have been
2832 moved to the other PE in the first place
2834 if (CurrentProc!=node_loc) {
2836 belch("## node %p is on PE %d but CurrentProc is %d (TSO %d); assuming fake fetch and adjusting bitmask (old: %#x)",
2837 node, node_loc, CurrentProc, CurrentTSO->id,
2838 // CurrentTSO, where_is(CurrentTSO),
2839 node->header.gran.procs));
2840 node->header.gran.procs = (node->header.gran.procs) | PE_NUMBER(CurrentProc);
2842 belch("## new bitmask of node %p is %#x",
2843 node, node->header.gran.procs));
2844 if (RtsFlags.GranFlags.GranSimStats.Global) {
2845 globalGranStats.tot_fake_fetches++;
2850 // ToDo: check: ASSERT(CurrentProc==node_loc);
2851 while (get_itbl(bqe)->type==TSO) { // q != END_TSO_QUEUE) {
2854 bqe points to the current element in the queue
2855 next points to the next element in the queue
2857 //tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2858 //tso_loc = where_is(tso);
2860 bqe = unblockOneLocked(bqe, node);
2863 /* if this is the BQ of an RBH, we have to put back the info ripped out of
2864 the closure to make room for the anchor of the BQ */
2865 if (bqe!=END_BQ_QUEUE) {
2866 ASSERT(get_itbl(node)->type == RBH && get_itbl(bqe)->type == CONSTR);
2868 ASSERT((info_ptr==&RBH_Save_0_info) ||
2869 (info_ptr==&RBH_Save_1_info) ||
2870 (info_ptr==&RBH_Save_2_info));
2872 /* cf. convertToRBH in RBH.c for writing the RBHSave closure */
2873 ((StgRBH *)node)->blocking_queue = (StgBlockingQueueElement *)((StgRBHSave *)bqe)->payload[0];
2874 ((StgRBH *)node)->mut_link = (StgMutClosure *)((StgRBHSave *)bqe)->payload[1];
2877 belch("## Filled in RBH_Save for %p (%s) at end of AwBQ",
2878 node, info_type(node)));
2881 /* statistics gathering */
2882 if (RtsFlags.GranFlags.GranSimStats.Global) {
2883 // globalGranStats.tot_bq_processing_time += bq_processing_time;
2884 globalGranStats.tot_bq_len += len; // total length of all bqs awakened
2885 // globalGranStats.tot_bq_len_local += len_local; // same for local TSOs only
2886 globalGranStats.tot_awbq++; // total no. of bqs awakened
2889 fprintf(stderr,"## BQ Stats of %p: [%d entries] %s\n",
2890 node, len, (bqe!=END_BQ_QUEUE) ? "RBH" : ""));
2894 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2896 StgBlockingQueueElement *bqe;
2898 ACQUIRE_LOCK(&sched_mutex);
2900 IF_PAR_DEBUG(verbose,
2901 belch("##-_ AwBQ for node %p on [%x]: ",
2905 if(get_itbl(q)->type == CONSTR || q==END_BQ_QUEUE) {
2906 IF_PAR_DEBUG(verbose, belch("## ... nothing to unblock so lets just return. RFP (BUG?)"));
2911 ASSERT(q == END_BQ_QUEUE ||
2912 get_itbl(q)->type == TSO ||
2913 get_itbl(q)->type == BLOCKED_FETCH ||
2914 get_itbl(q)->type == CONSTR);
2917 while (get_itbl(bqe)->type==TSO ||
2918 get_itbl(bqe)->type==BLOCKED_FETCH) {
2919 bqe = unblockOneLocked(bqe, node);
2921 RELEASE_LOCK(&sched_mutex);
2924 #else /* !GRAN && !PAR */
2926 awakenBlockedQueue(StgTSO *tso)
2928 ACQUIRE_LOCK(&sched_mutex);
2929 while (tso != END_TSO_QUEUE) {
2930 tso = unblockOneLocked(tso);
2932 RELEASE_LOCK(&sched_mutex);
2936 //@node Exception Handling Routines, Debugging Routines, Blocking Queue Routines, Main scheduling code
2937 //@subsection Exception Handling Routines
2939 /* ---------------------------------------------------------------------------
2941 - usually called inside a signal handler so it mustn't do anything fancy.
2942 ------------------------------------------------------------------------ */
2945 interruptStgRts(void)
2951 /* -----------------------------------------------------------------------------
2954 This is for use when we raise an exception in another thread, which
2956 This has nothing to do with the UnblockThread event in GranSim. -- HWL
2957 -------------------------------------------------------------------------- */
2959 #if defined(GRAN) || defined(PAR)
2961 NB: only the type of the blocking queue is different in GranSim and GUM
2962 the operations on the queue-elements are the same
2963 long live polymorphism!
2965 Locks: sched_mutex is held upon entry and exit.
2969 unblockThread(StgTSO *tso)
2971 StgBlockingQueueElement *t, **last;
2973 switch (tso->why_blocked) {
2976 return; /* not blocked */
2979 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
2981 StgBlockingQueueElement *last_tso = END_BQ_QUEUE;
2982 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
2984 last = (StgBlockingQueueElement **)&mvar->head;
2985 for (t = (StgBlockingQueueElement *)mvar->head;
2987 last = &t->link, last_tso = t, t = t->link) {
2988 if (t == (StgBlockingQueueElement *)tso) {
2989 *last = (StgBlockingQueueElement *)tso->link;
2990 if (mvar->tail == tso) {
2991 mvar->tail = (StgTSO *)last_tso;
2996 barf("unblockThread (MVAR): TSO not found");
2999 case BlockedOnBlackHole:
3000 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
3002 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
3004 last = &bq->blocking_queue;
3005 for (t = bq->blocking_queue;
3007 last = &t->link, t = t->link) {
3008 if (t == (StgBlockingQueueElement *)tso) {
3009 *last = (StgBlockingQueueElement *)tso->link;
3013 barf("unblockThread (BLACKHOLE): TSO not found");
3016 case BlockedOnException:
3018 StgTSO *target = tso->block_info.tso;
3020 ASSERT(get_itbl(target)->type == TSO);
3022 if (target->what_next == ThreadRelocated) {
3023 target = target->link;
3024 ASSERT(get_itbl(target)->type == TSO);
3027 ASSERT(target->blocked_exceptions != NULL);
3029 last = (StgBlockingQueueElement **)&target->blocked_exceptions;
3030 for (t = (StgBlockingQueueElement *)target->blocked_exceptions;
3032 last = &t->link, t = t->link) {
3033 ASSERT(get_itbl(t)->type == TSO);
3034 if (t == (StgBlockingQueueElement *)tso) {
3035 *last = (StgBlockingQueueElement *)tso->link;
3039 barf("unblockThread (Exception): TSO not found");
3043 case BlockedOnWrite:
3045 /* take TSO off blocked_queue */
3046 StgBlockingQueueElement *prev = NULL;
3047 for (t = (StgBlockingQueueElement *)blocked_queue_hd; t != END_BQ_QUEUE;
3048 prev = t, t = t->link) {
3049 if (t == (StgBlockingQueueElement *)tso) {
3051 blocked_queue_hd = (StgTSO *)t->link;
3052 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
3053 blocked_queue_tl = END_TSO_QUEUE;
3056 prev->link = t->link;
3057 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
3058 blocked_queue_tl = (StgTSO *)prev;
3064 barf("unblockThread (I/O): TSO not found");
3067 case BlockedOnDelay:
3069 /* take TSO off sleeping_queue */
3070 StgBlockingQueueElement *prev = NULL;
3071 for (t = (StgBlockingQueueElement *)sleeping_queue; t != END_BQ_QUEUE;
3072 prev = t, t = t->link) {
3073 if (t == (StgBlockingQueueElement *)tso) {
3075 sleeping_queue = (StgTSO *)t->link;
3077 prev->link = t->link;
3082 barf("unblockThread (I/O): TSO not found");
3086 barf("unblockThread");
3090 tso->link = END_TSO_QUEUE;
3091 tso->why_blocked = NotBlocked;
3092 tso->block_info.closure = NULL;
3093 PUSH_ON_RUN_QUEUE(tso);
3097 unblockThread(StgTSO *tso)
3101 /* To avoid locking unnecessarily. */
3102 if (tso->why_blocked == NotBlocked) {
3106 switch (tso->why_blocked) {
3109 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
3111 StgTSO *last_tso = END_TSO_QUEUE;
3112 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
3115 for (t = mvar->head; t != END_TSO_QUEUE;
3116 last = &t->link, last_tso = t, t = t->link) {
3119 if (mvar->tail == tso) {
3120 mvar->tail = last_tso;
3125 barf("unblockThread (MVAR): TSO not found");
3128 case BlockedOnBlackHole:
3129 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
3131 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
3133 last = &bq->blocking_queue;
3134 for (t = bq->blocking_queue; t != END_TSO_QUEUE;
3135 last = &t->link, t = t->link) {
3141 barf("unblockThread (BLACKHOLE): TSO not found");
3144 case BlockedOnException:
3146 StgTSO *target = tso->block_info.tso;
3148 ASSERT(get_itbl(target)->type == TSO);
3150 while (target->what_next == ThreadRelocated) {
3151 target = target->link;
3152 ASSERT(get_itbl(target)->type == TSO);
3155 ASSERT(target->blocked_exceptions != NULL);
3157 last = &target->blocked_exceptions;
3158 for (t = target->blocked_exceptions; t != END_TSO_QUEUE;
3159 last = &t->link, t = t->link) {
3160 ASSERT(get_itbl(t)->type == TSO);
3166 barf("unblockThread (Exception): TSO not found");
3170 case BlockedOnWrite:
3172 StgTSO *prev = NULL;
3173 for (t = blocked_queue_hd; t != END_TSO_QUEUE;
3174 prev = t, t = t->link) {
3177 blocked_queue_hd = t->link;
3178 if (blocked_queue_tl == t) {
3179 blocked_queue_tl = END_TSO_QUEUE;
3182 prev->link = t->link;
3183 if (blocked_queue_tl == t) {
3184 blocked_queue_tl = prev;
3190 barf("unblockThread (I/O): TSO not found");
3193 case BlockedOnDelay:
3195 StgTSO *prev = NULL;
3196 for (t = sleeping_queue; t != END_TSO_QUEUE;
3197 prev = t, t = t->link) {
3200 sleeping_queue = t->link;
3202 prev->link = t->link;
3207 barf("unblockThread (I/O): TSO not found");
3211 barf("unblockThread");
3215 tso->link = END_TSO_QUEUE;
3216 tso->why_blocked = NotBlocked;
3217 tso->block_info.closure = NULL;
3218 PUSH_ON_RUN_QUEUE(tso);
3222 /* -----------------------------------------------------------------------------
3225 * The following function implements the magic for raising an
3226 * asynchronous exception in an existing thread.
3228 * We first remove the thread from any queue on which it might be
3229 * blocked. The possible blockages are MVARs and BLACKHOLE_BQs.
3231 * We strip the stack down to the innermost CATCH_FRAME, building
3232 * thunks in the heap for all the active computations, so they can
3233 * be restarted if necessary. When we reach a CATCH_FRAME, we build
3234 * an application of the handler to the exception, and push it on
3235 * the top of the stack.
3237 * How exactly do we save all the active computations? We create an
3238 * AP_UPD for every UpdateFrame on the stack. Entering one of these
3239 * AP_UPDs pushes everything from the corresponding update frame
3240 * upwards onto the stack. (Actually, it pushes everything up to the
3241 * next update frame plus a pointer to the next AP_UPD object.
3242 * Entering the next AP_UPD object pushes more onto the stack until we
3243 * reach the last AP_UPD object - at which point the stack should look
3244 * exactly as it did when we killed the TSO and we can continue
3245 * execution by entering the closure on top of the stack.
3247 * We can also kill a thread entirely - this happens if either (a) the
3248 * exception passed to raiseAsync is NULL, or (b) there's no
3249 * CATCH_FRAME on the stack. In either case, we strip the entire
3250 * stack and replace the thread with a zombie.
3252 * Locks: sched_mutex held upon entry nor exit.
3254 * -------------------------------------------------------------------------- */
3257 deleteThread(StgTSO *tso)
3259 raiseAsync(tso,NULL);
3263 raiseAsyncWithLock(StgTSO *tso, StgClosure *exception)
3265 /* When raising async exs from contexts where sched_mutex isn't held;
3266 use raiseAsyncWithLock(). */
3267 ACQUIRE_LOCK(&sched_mutex);
3268 raiseAsync(tso,exception);
3269 RELEASE_LOCK(&sched_mutex);
3273 raiseAsync(StgTSO *tso, StgClosure *exception)
3275 StgUpdateFrame* su = tso->su;
3276 StgPtr sp = tso->sp;
3278 /* Thread already dead? */
3279 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
3283 IF_DEBUG(scheduler, sched_belch("raising exception in thread %ld.", tso->id));
3285 /* Remove it from any blocking queues */
3288 /* The stack freezing code assumes there's a closure pointer on
3289 * the top of the stack. This isn't always the case with compiled
3290 * code, so we have to push a dummy closure on the top which just
3291 * returns to the next return address on the stack.
3293 if ( LOOKS_LIKE_GHC_INFO((void*)*sp) ) {
3294 *(--sp) = (W_)&stg_dummy_ret_closure;
3298 nat words = ((P_)su - (P_)sp) - 1;
3302 ASSERT((P_)su > (P_)sp);
3304 /* If we find a CATCH_FRAME, and we've got an exception to raise,
3305 * then build the THUNK raise(exception), and leave it on
3306 * top of the CATCH_FRAME ready to enter.
3308 if (get_itbl(su)->type == CATCH_FRAME && exception != NULL) {
3310 StgCatchFrame *cf = (StgCatchFrame *)su;
3314 /* we've got an exception to raise, so let's pass it to the
3315 * handler in this frame.
3317 raise = (StgClosure *)allocate(sizeofW(StgClosure)+1);
3318 TICK_ALLOC_SE_THK(1,0);
3319 SET_HDR(raise,&stg_raise_info,cf->header.prof.ccs);
3320 raise->payload[0] = exception;
3322 /* throw away the stack from Sp up to the CATCH_FRAME.
3326 /* Ensure that async excpetions are blocked now, so we don't get
3327 * a surprise exception before we get around to executing the
3330 if (tso->blocked_exceptions == NULL) {
3331 tso->blocked_exceptions = END_TSO_QUEUE;
3334 /* Put the newly-built THUNK on top of the stack, ready to execute
3335 * when the thread restarts.
3340 tso->what_next = ThreadEnterGHC;
3341 IF_DEBUG(sanity, checkTSO(tso));
3345 /* First build an AP_UPD consisting of the stack chunk above the
3346 * current update frame, with the top word on the stack as the
3349 ap = (StgAP_UPD *)allocate(AP_sizeW(words));
3352 ap->fun = (StgClosure *)sp[0];
3354 for(i=0; i < (nat)words; ++i) {
3355 ap->payload[i] = (StgClosure *)*sp++;
3358 switch (get_itbl(su)->type) {
3362 SET_HDR(ap,&stg_AP_UPD_info,su->header.prof.ccs /* ToDo */);
3363 TICK_ALLOC_UP_THK(words+1,0);
3366 fprintf(stderr, "scheduler: Updating ");
3367 printPtr((P_)su->updatee);
3368 fprintf(stderr, " with ");
3369 printObj((StgClosure *)ap);
3372 /* Replace the updatee with an indirection - happily
3373 * this will also wake up any threads currently
3374 * waiting on the result.
3376 * Warning: if we're in a loop, more than one update frame on
3377 * the stack may point to the same object. Be careful not to
3378 * overwrite an IND_OLDGEN in this case, because we'll screw
3379 * up the mutable lists. To be on the safe side, don't
3380 * overwrite any kind of indirection at all. See also
3381 * threadSqueezeStack in GC.c, where we have to make a similar
3384 if (!closure_IND(su->updatee)) {
3385 UPD_IND_NOLOCK(su->updatee,ap); /* revert the black hole */
3388 sp += sizeofW(StgUpdateFrame) -1;
3389 sp[0] = (W_)ap; /* push onto stack */
3395 StgCatchFrame *cf = (StgCatchFrame *)su;
3398 /* We want a PAP, not an AP_UPD. Fortunately, the
3399 * layout's the same.
3401 SET_HDR(ap,&stg_PAP_info,su->header.prof.ccs /* ToDo */);
3402 TICK_ALLOC_UPD_PAP(words+1,0);
3404 /* now build o = FUN(catch,ap,handler) */
3405 o = (StgClosure *)allocate(sizeofW(StgClosure)+2);
3406 TICK_ALLOC_FUN(2,0);
3407 SET_HDR(o,&stg_catch_info,su->header.prof.ccs /* ToDo */);
3408 o->payload[0] = (StgClosure *)ap;
3409 o->payload[1] = cf->handler;
3412 fprintf(stderr, "scheduler: Built ");
3413 printObj((StgClosure *)o);
3416 /* pop the old handler and put o on the stack */
3418 sp += sizeofW(StgCatchFrame) - 1;
3425 StgSeqFrame *sf = (StgSeqFrame *)su;
3428 SET_HDR(ap,&stg_PAP_info,su->header.prof.ccs /* ToDo */);
3429 TICK_ALLOC_UPD_PAP(words+1,0);
3431 /* now build o = FUN(seq,ap) */
3432 o = (StgClosure *)allocate(sizeofW(StgClosure)+1);
3433 TICK_ALLOC_SE_THK(1,0);
3434 SET_HDR(o,&stg_seq_info,su->header.prof.ccs /* ToDo */);
3435 o->payload[0] = (StgClosure *)ap;
3438 fprintf(stderr, "scheduler: Built ");
3439 printObj((StgClosure *)o);
3442 /* pop the old handler and put o on the stack */
3444 sp += sizeofW(StgSeqFrame) - 1;
3450 /* We've stripped the entire stack, the thread is now dead. */
3451 sp += sizeofW(StgStopFrame) - 1;
3452 sp[0] = (W_)exception; /* save the exception */
3453 tso->what_next = ThreadKilled;
3454 tso->su = (StgUpdateFrame *)(sp+1);
3465 /* -----------------------------------------------------------------------------
3466 resurrectThreads is called after garbage collection on the list of
3467 threads found to be garbage. Each of these threads will be woken
3468 up and sent a signal: BlockedOnDeadMVar if the thread was blocked
3469 on an MVar, or NonTermination if the thread was blocked on a Black
3472 Locks: sched_mutex isn't held upon entry nor exit.
3473 -------------------------------------------------------------------------- */
3476 resurrectThreads( StgTSO *threads )
3480 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
3481 next = tso->global_link;
3482 tso->global_link = all_threads;
3484 IF_DEBUG(scheduler, sched_belch("resurrecting thread %d", tso->id));
3486 switch (tso->why_blocked) {
3488 case BlockedOnException:
3489 /* Called by GC - sched_mutex lock is currently held. */
3490 raiseAsync(tso,(StgClosure *)BlockedOnDeadMVar_closure);
3492 case BlockedOnBlackHole:
3493 raiseAsync(tso,(StgClosure *)NonTermination_closure);
3496 /* This might happen if the thread was blocked on a black hole
3497 * belonging to a thread that we've just woken up (raiseAsync
3498 * can wake up threads, remember...).
3502 barf("resurrectThreads: thread blocked in a strange way");
3507 /* -----------------------------------------------------------------------------
3508 * Blackhole detection: if we reach a deadlock, test whether any
3509 * threads are blocked on themselves. Any threads which are found to
3510 * be self-blocked get sent a NonTermination exception.
3512 * This is only done in a deadlock situation in order to avoid
3513 * performance overhead in the normal case.
3515 * Locks: sched_mutex is held upon entry and exit.
3516 * -------------------------------------------------------------------------- */
3519 detectBlackHoles( void )
3521 StgTSO *t = all_threads;
3522 StgUpdateFrame *frame;
3523 StgClosure *blocked_on;
3525 for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
3527 while (t->what_next == ThreadRelocated) {
3529 ASSERT(get_itbl(t)->type == TSO);
3532 if (t->why_blocked != BlockedOnBlackHole) {
3536 blocked_on = t->block_info.closure;
3538 for (frame = t->su; ; frame = frame->link) {
3539 switch (get_itbl(frame)->type) {
3542 if (frame->updatee == blocked_on) {
3543 /* We are blocking on one of our own computations, so
3544 * send this thread the NonTermination exception.
3547 sched_belch("thread %d is blocked on itself", t->id));
3548 raiseAsync(t, (StgClosure *)NonTermination_closure);
3569 //@node Debugging Routines, Index, Exception Handling Routines, Main scheduling code
3570 //@subsection Debugging Routines
3572 /* -----------------------------------------------------------------------------
3573 * Debugging: why is a thread blocked
3574 * [Also provides useful information when debugging threaded programs
3575 * at the Haskell source code level, so enable outside of DEBUG. --sof 7/02]
3576 -------------------------------------------------------------------------- */
3580 printThreadBlockage(StgTSO *tso)
3582 switch (tso->why_blocked) {
3584 fprintf(stderr,"is blocked on read from fd %d", tso->block_info.fd);
3586 case BlockedOnWrite:
3587 fprintf(stderr,"is blocked on write to fd %d", tso->block_info.fd);
3589 case BlockedOnDelay:
3590 fprintf(stderr,"is blocked until %d", tso->block_info.target);
3593 fprintf(stderr,"is blocked on an MVar");
3595 case BlockedOnException:
3596 fprintf(stderr,"is blocked on delivering an exception to thread %d",
3597 tso->block_info.tso->id);
3599 case BlockedOnBlackHole:
3600 fprintf(stderr,"is blocked on a black hole");
3603 fprintf(stderr,"is not blocked");
3607 fprintf(stderr,"is blocked on global address; local FM_BQ is %p (%s)",
3608 tso->block_info.closure, info_type(tso->block_info.closure));
3610 case BlockedOnGA_NoSend:
3611 fprintf(stderr,"is blocked on global address (no send); local FM_BQ is %p (%s)",
3612 tso->block_info.closure, info_type(tso->block_info.closure));
3615 #if defined(RTS_SUPPORTS_THREADS)
3616 case BlockedOnCCall:
3617 fprintf(stderr,"is blocked on an external call");
3621 barf("printThreadBlockage: strange tso->why_blocked: %d for TSO %d (%d)",
3622 tso->why_blocked, tso->id, tso);
3628 printThreadStatus(StgTSO *tso)
3630 switch (tso->what_next) {
3632 fprintf(stderr,"has been killed");
3634 case ThreadComplete:
3635 fprintf(stderr,"has completed");
3638 printThreadBlockage(tso);
3643 printAllThreads(void)
3649 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
3650 ullong_format_string(TIME_ON_PROC(CurrentProc),
3651 time_string, rtsFalse/*no commas!*/);
3653 fprintf(stderr, "all threads at [%s]:\n", time_string);
3655 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
3656 ullong_format_string(CURRENT_TIME,
3657 time_string, rtsFalse/*no commas!*/);
3659 fprintf(stderr,"all threads at [%s]:\n", time_string);
3661 fprintf(stderr,"all threads:\n");
3664 for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
3665 fprintf(stderr, "\tthread %d @ %p ", t->id, (void *)t);
3666 label = lookupThreadLabel((StgWord)t);
3667 if (label) fprintf(stderr,"[\"%s\"] ",(char *)label);
3668 printThreadStatus(t);
3669 fprintf(stderr,"\n");
3676 Print a whole blocking queue attached to node (debugging only).
3681 print_bq (StgClosure *node)
3683 StgBlockingQueueElement *bqe;
3687 fprintf(stderr,"## BQ of closure %p (%s): ",
3688 node, info_type(node));
3690 /* should cover all closures that may have a blocking queue */
3691 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
3692 get_itbl(node)->type == FETCH_ME_BQ ||
3693 get_itbl(node)->type == RBH ||
3694 get_itbl(node)->type == MVAR);
3696 ASSERT(node!=(StgClosure*)NULL); // sanity check
3698 print_bqe(((StgBlockingQueue*)node)->blocking_queue);
3702 Print a whole blocking queue starting with the element bqe.
3705 print_bqe (StgBlockingQueueElement *bqe)
3710 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
3712 for (end = (bqe==END_BQ_QUEUE);
3713 !end; // iterate until bqe points to a CONSTR
3714 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE),
3715 bqe = end ? END_BQ_QUEUE : bqe->link) {
3716 ASSERT(bqe != END_BQ_QUEUE); // sanity check
3717 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
3718 /* types of closures that may appear in a blocking queue */
3719 ASSERT(get_itbl(bqe)->type == TSO ||
3720 get_itbl(bqe)->type == BLOCKED_FETCH ||
3721 get_itbl(bqe)->type == CONSTR);
3722 /* only BQs of an RBH end with an RBH_Save closure */
3723 //ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
3725 switch (get_itbl(bqe)->type) {
3727 fprintf(stderr," TSO %u (%x),",
3728 ((StgTSO *)bqe)->id, ((StgTSO *)bqe));
3731 fprintf(stderr," BF (node=%p, ga=((%x, %d, %x)),",
3732 ((StgBlockedFetch *)bqe)->node,
3733 ((StgBlockedFetch *)bqe)->ga.payload.gc.gtid,
3734 ((StgBlockedFetch *)bqe)->ga.payload.gc.slot,
3735 ((StgBlockedFetch *)bqe)->ga.weight);
3738 fprintf(stderr," %s (IP %p),",
3739 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
3740 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
3741 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
3742 "RBH_Save_?"), get_itbl(bqe));
3745 barf("Unexpected closure type %s in blocking queue", // of %p (%s)",
3746 info_type((StgClosure *)bqe)); // , node, info_type(node));
3750 fputc('\n', stderr);
3752 # elif defined(GRAN)
3754 print_bq (StgClosure *node)
3756 StgBlockingQueueElement *bqe;
3757 PEs node_loc, tso_loc;
3760 /* should cover all closures that may have a blocking queue */
3761 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
3762 get_itbl(node)->type == FETCH_ME_BQ ||
3763 get_itbl(node)->type == RBH);
3765 ASSERT(node!=(StgClosure*)NULL); // sanity check
3766 node_loc = where_is(node);
3768 fprintf(stderr,"## BQ of closure %p (%s) on [PE %d]: ",
3769 node, info_type(node), node_loc);
3772 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
3774 for (bqe = ((StgBlockingQueue*)node)->blocking_queue, end = (bqe==END_BQ_QUEUE);
3775 !end; // iterate until bqe points to a CONSTR
3776 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) {
3777 ASSERT(bqe != END_BQ_QUEUE); // sanity check
3778 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
3779 /* types of closures that may appear in a blocking queue */
3780 ASSERT(get_itbl(bqe)->type == TSO ||
3781 get_itbl(bqe)->type == CONSTR);
3782 /* only BQs of an RBH end with an RBH_Save closure */
3783 ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
3785 tso_loc = where_is((StgClosure *)bqe);
3786 switch (get_itbl(bqe)->type) {
3788 fprintf(stderr," TSO %d (%p) on [PE %d],",
3789 ((StgTSO *)bqe)->id, (StgTSO *)bqe, tso_loc);
3792 fprintf(stderr," %s (IP %p),",
3793 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
3794 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
3795 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
3796 "RBH_Save_?"), get_itbl(bqe));
3799 barf("Unexpected closure type %s in blocking queue of %p (%s)",
3800 info_type((StgClosure *)bqe), node, info_type(node));
3804 fputc('\n', stderr);
3808 Nice and easy: only TSOs on the blocking queue
3811 print_bq (StgClosure *node)
3815 ASSERT(node!=(StgClosure*)NULL); // sanity check
3816 for (tso = ((StgBlockingQueue*)node)->blocking_queue;
3817 tso != END_TSO_QUEUE;
3819 ASSERT(tso!=NULL && tso!=END_TSO_QUEUE); // sanity check
3820 ASSERT(get_itbl(tso)->type == TSO); // guess what, sanity check
3821 fprintf(stderr," TSO %d (%p),", tso->id, tso);
3823 fputc('\n', stderr);
3834 for (i=0, tso=run_queue_hd;
3835 tso != END_TSO_QUEUE;
3844 sched_belch(char *s, ...)
3849 fprintf(stderr, "scheduler (task %ld): ", osThreadId());
3851 fprintf(stderr, "== ");
3853 fprintf(stderr, "scheduler: ");
3855 vfprintf(stderr, s, ap);
3856 fprintf(stderr, "\n");
3863 //@node Index, , Debugging Routines, Main scheduling code
3867 //* StgMainThread:: @cindex\s-+StgMainThread
3868 //* awaken_blocked_queue:: @cindex\s-+awaken_blocked_queue
3869 //* blocked_queue_hd:: @cindex\s-+blocked_queue_hd
3870 //* blocked_queue_tl:: @cindex\s-+blocked_queue_tl
3871 //* context_switch:: @cindex\s-+context_switch
3872 //* createThread:: @cindex\s-+createThread
3873 //* gc_pending_cond:: @cindex\s-+gc_pending_cond
3874 //* initScheduler:: @cindex\s-+initScheduler
3875 //* interrupted:: @cindex\s-+interrupted
3876 //* next_thread_id:: @cindex\s-+next_thread_id
3877 //* print_bq:: @cindex\s-+print_bq
3878 //* run_queue_hd:: @cindex\s-+run_queue_hd
3879 //* run_queue_tl:: @cindex\s-+run_queue_tl
3880 //* sched_mutex:: @cindex\s-+sched_mutex
3881 //* schedule:: @cindex\s-+schedule
3882 //* take_off_run_queue:: @cindex\s-+take_off_run_queue
3883 //* term_mutex:: @cindex\s-+term_mutex