1 /* ---------------------------------------------------------------------------
2 * $Id: Schedule.c,v 1.159 2002/12/11 15:36:50 simonmar Exp $
4 * (c) The GHC Team, 1998-2000
8 * Different GHC ways use this scheduler quite differently (see comments below)
9 * Here is the global picture:
11 * WAY Name CPP flag What's it for
12 * --------------------------------------
13 * mp GUM PAR Parallel execution on a distributed memory machine
14 * s SMP SMP Parallel execution on a shared memory machine
15 * mg GranSim GRAN Simulation of parallel execution
16 * md GUM/GdH DIST Distributed execution (based on GUM)
18 * --------------------------------------------------------------------------*/
20 //@node Main scheduling code, , ,
21 //@section Main scheduling code
24 * Version with scheduler monitor support for SMPs (WAY=s):
26 This design provides a high-level API to create and schedule threads etc.
27 as documented in the SMP design document.
29 It uses a monitor design controlled by a single mutex to exercise control
30 over accesses to shared data structures, and builds on the Posix threads
33 The majority of state is shared. In order to keep essential per-task state,
34 there is a Capability structure, which contains all the information
35 needed to run a thread: its STG registers, a pointer to its TSO, a
36 nursery etc. During STG execution, a pointer to the capability is
37 kept in a register (BaseReg).
39 In a non-SMP build, there is one global capability, namely MainRegTable.
43 * Version with support for distributed memory parallelism aka GUM (WAY=mp):
45 The main scheduling loop in GUM iterates until a finish message is received.
46 In that case a global flag @receivedFinish@ is set and this instance of
47 the RTS shuts down. See ghc/rts/parallel/HLComms.c:processMessages()
48 for the handling of incoming messages, such as PP_FINISH.
49 Note that in the parallel case we have a system manager that coordinates
50 different PEs, each of which are running one instance of the RTS.
51 See ghc/rts/parallel/SysMan.c for the main routine of the parallel program.
52 From this routine processes executing ghc/rts/Main.c are spawned. -- HWL
54 * Version with support for simulating parallel execution aka GranSim (WAY=mg):
56 The main scheduling code in GranSim is quite different from that in std
57 (concurrent) Haskell: while concurrent Haskell just iterates over the
58 threads in the runnable queue, GranSim is event driven, i.e. it iterates
59 over the events in the global event queue. -- HWL
64 //* Variables and Data structures::
65 //* Main scheduling loop::
66 //* Suspend and Resume::
68 //* Garbage Collextion Routines::
69 //* Blocking Queue Routines::
70 //* Exception Handling Routines::
71 //* Debugging Routines::
75 //@node Includes, Variables and Data structures, Main scheduling code, Main scheduling code
76 //@subsection Includes
78 #include "PosixSource.h"
85 #include "StgStartup.h"
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;
370 StgTSOWhatNext prev_what_next;
372 ACQUIRE_LOCK(&sched_mutex);
374 #if defined(RTS_SUPPORTS_THREADS)
375 waitForWorkCapability(&sched_mutex, &cap, rtsFalse);
377 /* simply initialise it in the non-threaded case */
378 grabCapability(&cap);
382 /* set up first event to get things going */
383 /* ToDo: assign costs for system setup and init MainTSO ! */
384 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
386 CurrentTSO, (StgClosure*)NULL, (rtsSpark*)NULL);
389 fprintf(stderr, "GRAN: Init CurrentTSO (in schedule) = %p\n", CurrentTSO);
390 G_TSO(CurrentTSO, 5));
392 if (RtsFlags.GranFlags.Light) {
393 /* Save current time; GranSim Light only */
394 CurrentTSO->gran.clock = CurrentTime[CurrentProc];
397 event = get_next_event();
399 while (event!=(rtsEvent*)NULL) {
400 /* Choose the processor with the next event */
401 CurrentProc = event->proc;
402 CurrentTSO = event->tso;
406 while (!receivedFinish) { /* set by processMessages */
407 /* when receiving PP_FINISH message */
414 IF_DEBUG(scheduler, printAllThreads());
416 #if defined(RTS_SUPPORTS_THREADS)
417 /* Check to see whether there are any worker threads
418 waiting to deposit external call results. If so,
419 yield our capability */
420 yieldToReturningWorker(&sched_mutex, &cap);
423 /* If we're interrupted (the user pressed ^C, or some other
424 * termination condition occurred), kill all the currently running
428 IF_DEBUG(scheduler, sched_belch("interrupted"));
430 interrupted = rtsFalse;
431 was_interrupted = rtsTrue;
434 /* Go through the list of main threads and wake up any
435 * clients whose computations have finished. ToDo: this
436 * should be done more efficiently without a linear scan
437 * of the main threads list, somehow...
439 #if defined(RTS_SUPPORTS_THREADS)
441 StgMainThread *m, **prev;
442 prev = &main_threads;
443 for (m = main_threads; m != NULL; m = m->link) {
444 switch (m->tso->what_next) {
447 // NOTE: return val is tso->sp[1] (see StgStartup.hc)
448 *(m->ret) = (StgClosure *)m->tso->sp[1];
452 broadcastCondition(&m->wakeup);
454 removeThreadLabel((StgWord)m->tso);
458 if (m->ret) *(m->ret) = NULL;
460 if (was_interrupted) {
461 m->stat = Interrupted;
465 broadcastCondition(&m->wakeup);
467 removeThreadLabel((StgWord)m->tso);
476 #else /* not threaded */
479 /* in GUM do this only on the Main PE */
482 /* If our main thread has finished or been killed, return.
485 StgMainThread *m = main_threads;
486 if (m->tso->what_next == ThreadComplete
487 || m->tso->what_next == ThreadKilled) {
489 removeThreadLabel((StgWord)m->tso);
491 main_threads = main_threads->link;
492 if (m->tso->what_next == ThreadComplete) {
493 // We finished successfully, fill in the return value
494 // NOTE: return val is tso->sp[1] (see StgStartup.hc)
495 if (m->ret) { *(m->ret) = (StgClosure *)m->tso->sp[1]; };
499 if (m->ret) { *(m->ret) = NULL; };
500 if (was_interrupted) {
501 m->stat = Interrupted;
511 /* Top up the run queue from our spark pool. We try to make the
512 * number of threads in the run queue equal to the number of
515 * Disable spark support in SMP for now, non-essential & requires
516 * a little bit of work to make it compile cleanly. -- sof 1/02.
518 #if 0 /* defined(SMP) */
520 nat n = getFreeCapabilities();
521 StgTSO *tso = run_queue_hd;
523 /* Count the run queue */
524 while (n > 0 && tso != END_TSO_QUEUE) {
531 spark = findSpark(rtsFalse);
533 break; /* no more sparks in the pool */
535 /* I'd prefer this to be done in activateSpark -- HWL */
536 /* tricky - it needs to hold the scheduler lock and
537 * not try to re-acquire it -- SDM */
538 createSparkThread(spark);
540 sched_belch("==^^ turning spark of closure %p into a thread",
541 (StgClosure *)spark));
544 /* We need to wake up the other tasks if we just created some
547 if (getFreeCapabilities() - n > 1) {
548 signalCondition( &thread_ready_cond );
553 /* check for signals each time around the scheduler */
554 #ifndef mingw32_TARGET_OS
555 if (signals_pending()) {
556 RELEASE_LOCK(&sched_mutex); /* ToDo: kill */
557 startSignalHandlers();
558 ACQUIRE_LOCK(&sched_mutex);
562 /* Check whether any waiting threads need to be woken up. If the
563 * run queue is empty, and there are no other tasks running, we
564 * can wait indefinitely for something to happen.
565 * ToDo: what if another client comes along & requests another
568 if ( !EMPTY_QUEUE(blocked_queue_hd) || !EMPTY_QUEUE(sleeping_queue) ) {
569 awaitEvent( EMPTY_RUN_QUEUE()
571 && allFreeCapabilities()
575 /* we can be interrupted while waiting for I/O... */
576 if (interrupted) continue;
579 * Detect deadlock: when we have no threads to run, there are no
580 * threads waiting on I/O or sleeping, and all the other tasks are
581 * waiting for work, we must have a deadlock of some description.
583 * We first try to find threads blocked on themselves (ie. black
584 * holes), and generate NonTermination exceptions where necessary.
586 * If no threads are black holed, we have a deadlock situation, so
587 * inform all the main threads.
590 if ( EMPTY_THREAD_QUEUES()
591 #if defined(RTS_SUPPORTS_THREADS)
592 && EMPTY_QUEUE(suspended_ccalling_threads)
595 && allFreeCapabilities()
599 IF_DEBUG(scheduler, sched_belch("deadlocked, forcing major GC..."));
600 #if defined(THREADED_RTS)
601 /* and SMP mode ..? */
602 releaseCapability(cap);
604 // Garbage collection can release some new threads due to
605 // either (a) finalizers or (b) threads resurrected because
606 // they are about to be send BlockedOnDeadMVar. Any threads
607 // thus released will be immediately runnable.
608 GarbageCollect(GetRoots,rtsTrue);
610 if ( !EMPTY_RUN_QUEUE() ) { goto not_deadlocked; }
613 sched_belch("still deadlocked, checking for black holes..."));
616 if ( !EMPTY_RUN_QUEUE() ) { goto not_deadlocked; }
618 #ifndef mingw32_TARGET_OS
619 /* If we have user-installed signal handlers, then wait
620 * for signals to arrive rather then bombing out with a
623 #if defined(RTS_SUPPORTS_THREADS)
624 if ( 0 ) { /* hmm..what to do? Simply stop waiting for
625 a signal with no runnable threads (or I/O
626 suspended ones) leads nowhere quick.
627 For now, simply shut down when we reach this
630 ToDo: define precisely under what conditions
631 the Scheduler should shut down in an MT setting.
634 if ( anyUserHandlers() ) {
637 sched_belch("still deadlocked, waiting for signals..."));
641 // we might be interrupted...
642 if (interrupted) { continue; }
644 if (signals_pending()) {
645 RELEASE_LOCK(&sched_mutex);
646 startSignalHandlers();
647 ACQUIRE_LOCK(&sched_mutex);
649 ASSERT(!EMPTY_RUN_QUEUE());
654 /* Probably a real deadlock. Send the current main thread the
655 * Deadlock exception (or in the SMP build, send *all* main
656 * threads the deadlock exception, since none of them can make
661 #if defined(RTS_SUPPORTS_THREADS)
662 for (m = main_threads; m != NULL; m = m->link) {
663 switch (m->tso->why_blocked) {
664 case BlockedOnBlackHole:
665 raiseAsync(m->tso, (StgClosure *)NonTermination_closure);
667 case BlockedOnException:
669 raiseAsync(m->tso, (StgClosure *)Deadlock_closure);
672 barf("deadlock: main thread blocked in a strange way");
677 switch (m->tso->why_blocked) {
678 case BlockedOnBlackHole:
679 raiseAsync(m->tso, (StgClosure *)NonTermination_closure);
681 case BlockedOnException:
683 raiseAsync(m->tso, (StgClosure *)Deadlock_closure);
686 barf("deadlock: main thread blocked in a strange way");
691 #if defined(RTS_SUPPORTS_THREADS)
692 /* ToDo: revisit conditions (and mechanism) for shutting
693 down a multi-threaded world */
694 IF_DEBUG(scheduler, sched_belch("all done, i think...shutting down."));
695 RELEASE_LOCK(&sched_mutex);
703 /* ToDo: add deadlock detection in GUM (similar to SMP) -- HWL */
707 /* If there's a GC pending, don't do anything until it has
711 IF_DEBUG(scheduler,sched_belch("waiting for GC"));
712 waitCondition( &gc_pending_cond, &sched_mutex );
716 #if defined(RTS_SUPPORTS_THREADS)
717 /* block until we've got a thread on the run queue and a free
721 if ( EMPTY_RUN_QUEUE() ) {
722 /* Give up our capability */
723 releaseCapability(cap);
725 /* If we're in the process of shutting down (& running the
726 * a batch of finalisers), don't wait around.
728 if ( shutting_down_scheduler ) {
729 RELEASE_LOCK(&sched_mutex);
732 IF_DEBUG(scheduler, sched_belch("thread %d: waiting for work", osThreadId()));
733 waitForWorkCapability(&sched_mutex, &cap, rtsTrue);
734 IF_DEBUG(scheduler, sched_belch("thread %d: work now available", osThreadId()));
739 if (RtsFlags.GranFlags.Light)
740 GranSimLight_enter_system(event, &ActiveTSO); // adjust ActiveTSO etc
742 /* adjust time based on time-stamp */
743 if (event->time > CurrentTime[CurrentProc] &&
744 event->evttype != ContinueThread)
745 CurrentTime[CurrentProc] = event->time;
747 /* Deal with the idle PEs (may issue FindWork or MoveSpark events) */
748 if (!RtsFlags.GranFlags.Light)
751 IF_DEBUG(gran, fprintf(stderr, "GRAN: switch by event-type\n"));
753 /* main event dispatcher in GranSim */
754 switch (event->evttype) {
755 /* Should just be continuing execution */
757 IF_DEBUG(gran, fprintf(stderr, "GRAN: doing ContinueThread\n"));
758 /* ToDo: check assertion
759 ASSERT(run_queue_hd != (StgTSO*)NULL &&
760 run_queue_hd != END_TSO_QUEUE);
762 /* Ignore ContinueThreads for fetching threads (if synchr comm) */
763 if (!RtsFlags.GranFlags.DoAsyncFetch &&
764 procStatus[CurrentProc]==Fetching) {
765 belch("ghuH: Spurious ContinueThread while Fetching ignored; TSO %d (%p) [PE %d]",
766 CurrentTSO->id, CurrentTSO, CurrentProc);
769 /* Ignore ContinueThreads for completed threads */
770 if (CurrentTSO->what_next == ThreadComplete) {
771 belch("ghuH: found a ContinueThread event for completed thread %d (%p) [PE %d] (ignoring ContinueThread)",
772 CurrentTSO->id, CurrentTSO, CurrentProc);
775 /* Ignore ContinueThreads for threads that are being migrated */
776 if (PROCS(CurrentTSO)==Nowhere) {
777 belch("ghuH: trying to run the migrating TSO %d (%p) [PE %d] (ignoring ContinueThread)",
778 CurrentTSO->id, CurrentTSO, CurrentProc);
781 /* The thread should be at the beginning of the run queue */
782 if (CurrentTSO!=run_queue_hds[CurrentProc]) {
783 belch("ghuH: TSO %d (%p) [PE %d] is not at the start of the run_queue when doing a ContinueThread",
784 CurrentTSO->id, CurrentTSO, CurrentProc);
785 break; // run the thread anyway
788 new_event(proc, proc, CurrentTime[proc],
790 (StgTSO*)NULL, (StgClosure*)NULL, (rtsSpark*)NULL);
792 */ /* Catches superfluous CONTINUEs -- should be unnecessary */
793 break; // now actually run the thread; DaH Qu'vam yImuHbej
796 do_the_fetchnode(event);
797 goto next_thread; /* handle next event in event queue */
800 do_the_globalblock(event);
801 goto next_thread; /* handle next event in event queue */
804 do_the_fetchreply(event);
805 goto next_thread; /* handle next event in event queue */
807 case UnblockThread: /* Move from the blocked queue to the tail of */
808 do_the_unblock(event);
809 goto next_thread; /* handle next event in event queue */
811 case ResumeThread: /* Move from the blocked queue to the tail of */
812 /* the runnable queue ( i.e. Qu' SImqa'lu') */
813 event->tso->gran.blocktime +=
814 CurrentTime[CurrentProc] - event->tso->gran.blockedat;
815 do_the_startthread(event);
816 goto next_thread; /* handle next event in event queue */
819 do_the_startthread(event);
820 goto next_thread; /* handle next event in event queue */
823 do_the_movethread(event);
824 goto next_thread; /* handle next event in event queue */
827 do_the_movespark(event);
828 goto next_thread; /* handle next event in event queue */
831 do_the_findwork(event);
832 goto next_thread; /* handle next event in event queue */
835 barf("Illegal event type %u\n", event->evttype);
838 /* This point was scheduler_loop in the old RTS */
840 IF_DEBUG(gran, belch("GRAN: after main switch"));
842 TimeOfLastEvent = CurrentTime[CurrentProc];
843 TimeOfNextEvent = get_time_of_next_event();
844 IgnoreEvents=(TimeOfNextEvent==0); // HWL HACK
845 // CurrentTSO = ThreadQueueHd;
847 IF_DEBUG(gran, belch("GRAN: time of next event is: %ld",
850 if (RtsFlags.GranFlags.Light)
851 GranSimLight_leave_system(event, &ActiveTSO);
853 EndOfTimeSlice = CurrentTime[CurrentProc]+RtsFlags.GranFlags.time_slice;
856 belch("GRAN: end of time-slice is %#lx", EndOfTimeSlice));
858 /* in a GranSim setup the TSO stays on the run queue */
860 /* Take a thread from the run queue. */
861 t = POP_RUN_QUEUE(); // take_off_run_queue(t);
864 fprintf(stderr, "GRAN: About to run current thread, which is\n");
867 context_switch = 0; // turned on via GranYield, checking events and time slice
870 DumpGranEvent(GR_SCHEDULE, t));
872 procStatus[CurrentProc] = Busy;
875 if (PendingFetches != END_BF_QUEUE) {
879 /* ToDo: phps merge with spark activation above */
880 /* check whether we have local work and send requests if we have none */
881 if (EMPTY_RUN_QUEUE()) { /* no runnable threads */
882 /* :-[ no local threads => look out for local sparks */
883 /* the spark pool for the current PE */
884 pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable
885 if (advisory_thread_count < RtsFlags.ParFlags.maxThreads &&
886 pool->hd < pool->tl) {
888 * ToDo: add GC code check that we really have enough heap afterwards!!
890 * If we're here (no runnable threads) and we have pending
891 * sparks, we must have a space problem. Get enough space
892 * to turn one of those pending sparks into a
896 spark = findSpark(rtsFalse); /* get a spark */
897 if (spark != (rtsSpark) NULL) {
898 tso = activateSpark(spark); /* turn the spark into a thread */
899 IF_PAR_DEBUG(schedule,
900 belch("==== schedule: Created TSO %d (%p); %d threads active",
901 tso->id, tso, advisory_thread_count));
903 if (tso==END_TSO_QUEUE) { /* failed to activate spark->back to loop */
904 belch("==^^ failed to activate spark");
906 } /* otherwise fall through & pick-up new tso */
908 IF_PAR_DEBUG(verbose,
909 belch("==^^ no local sparks (spark pool contains only NFs: %d)",
910 spark_queue_len(pool)));
915 /* If we still have no work we need to send a FISH to get a spark
918 if (EMPTY_RUN_QUEUE()) {
919 /* =8-[ no local sparks => look for work on other PEs */
921 * We really have absolutely no work. Send out a fish
922 * (there may be some out there already), and wait for
923 * something to arrive. We clearly can't run any threads
924 * until a SCHEDULE or RESUME arrives, and so that's what
925 * we're hoping to see. (Of course, we still have to
926 * respond to other types of messages.)
928 TIME now = msTime() /*CURRENT_TIME*/;
929 IF_PAR_DEBUG(verbose,
930 belch("-- now=%ld", now));
931 IF_PAR_DEBUG(verbose,
932 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
933 (last_fish_arrived_at!=0 &&
934 last_fish_arrived_at+RtsFlags.ParFlags.fishDelay > now)) {
935 belch("--$$ delaying FISH until %ld (last fish %ld, delay %ld, now %ld)",
936 last_fish_arrived_at+RtsFlags.ParFlags.fishDelay,
937 last_fish_arrived_at,
938 RtsFlags.ParFlags.fishDelay, now);
941 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
942 (last_fish_arrived_at==0 ||
943 (last_fish_arrived_at+RtsFlags.ParFlags.fishDelay <= now))) {
944 /* outstandingFishes is set in sendFish, processFish;
945 avoid flooding system with fishes via delay */
947 sendFish(pe, mytid, NEW_FISH_AGE, NEW_FISH_HISTORY,
950 // Global statistics: count no. of fishes
951 if (RtsFlags.ParFlags.ParStats.Global &&
952 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
953 globalParStats.tot_fish_mess++;
957 receivedFinish = processMessages();
960 } else if (PacketsWaiting()) { /* Look for incoming messages */
961 receivedFinish = processMessages();
964 /* Now we are sure that we have some work available */
965 ASSERT(run_queue_hd != END_TSO_QUEUE);
967 /* Take a thread from the run queue, if we have work */
968 t = POP_RUN_QUEUE(); // take_off_run_queue(END_TSO_QUEUE);
969 IF_DEBUG(sanity,checkTSO(t));
971 /* ToDo: write something to the log-file
972 if (RTSflags.ParFlags.granSimStats && !sameThread)
973 DumpGranEvent(GR_SCHEDULE, RunnableThreadsHd);
977 /* the spark pool for the current PE */
978 pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable
981 belch("--=^ %d threads, %d sparks on [%#x]",
982 run_queue_len(), spark_queue_len(pool), CURRENT_PROC));
985 if (0 && RtsFlags.ParFlags.ParStats.Full &&
986 t && LastTSO && t->id != LastTSO->id &&
987 LastTSO->why_blocked == NotBlocked &&
988 LastTSO->what_next != ThreadComplete) {
989 // if previously scheduled TSO not blocked we have to record the context switch
990 DumpVeryRawGranEvent(TimeOfLastYield, CURRENT_PROC, CURRENT_PROC,
991 GR_DESCHEDULE, LastTSO, (StgClosure *)NULL, 0, 0);
994 if (RtsFlags.ParFlags.ParStats.Full &&
995 (emitSchedule /* forced emit */ ||
996 (t && LastTSO && t->id != LastTSO->id))) {
998 we are running a different TSO, so write a schedule event to log file
999 NB: If we use fair scheduling we also have to write a deschedule
1000 event for LastTSO; with unfair scheduling we know that the
1001 previous tso has blocked whenever we switch to another tso, so
1002 we don't need it in GUM for now
1004 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
1005 GR_SCHEDULE, t, (StgClosure *)NULL, 0, 0);
1006 emitSchedule = rtsFalse;
1010 #else /* !GRAN && !PAR */
1012 /* grab a thread from the run queue */
1013 ASSERT(run_queue_hd != END_TSO_QUEUE);
1014 t = POP_RUN_QUEUE();
1015 // Sanity check the thread we're about to run. This can be
1016 // expensive if there is lots of thread switching going on...
1017 IF_DEBUG(sanity,checkTSO(t));
1020 cap->r.rCurrentTSO = t;
1022 /* context switches are now initiated by the timer signal, unless
1023 * the user specified "context switch as often as possible", with
1026 if ((RtsFlags.ConcFlags.ctxtSwitchTicks == 0
1027 && (run_queue_hd != END_TSO_QUEUE
1028 || blocked_queue_hd != END_TSO_QUEUE
1029 || sleeping_queue != END_TSO_QUEUE)))
1034 RELEASE_LOCK(&sched_mutex);
1036 IF_DEBUG(scheduler, sched_belch("-->> running thread %ld %s ...",
1037 t->id, whatNext_strs[t->what_next]));
1040 startHeapProfTimer();
1043 /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
1044 /* Run the current thread
1046 prev_what_next = t->what_next;
1047 switch (prev_what_next) {
1049 case ThreadComplete:
1050 /* Thread already finished, return to scheduler. */
1051 ret = ThreadFinished;
1054 ret = StgRun((StgFunPtr) stg_returnToStackTop, &cap->r);
1056 case ThreadInterpret:
1057 ret = interpretBCO(cap);
1060 barf("schedule: invalid what_next field");
1062 /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
1064 /* Costs for the scheduler are assigned to CCS_SYSTEM */
1066 stopHeapProfTimer();
1070 ACQUIRE_LOCK(&sched_mutex);
1073 IF_DEBUG(scheduler,fprintf(stderr,"scheduler (task %ld): ", osThreadId()););
1074 #elif !defined(GRAN) && !defined(PAR)
1075 IF_DEBUG(scheduler,fprintf(stderr,"scheduler: "););
1077 t = cap->r.rCurrentTSO;
1080 /* HACK 675: if the last thread didn't yield, make sure to print a
1081 SCHEDULE event to the log file when StgRunning the next thread, even
1082 if it is the same one as before */
1084 TimeOfLastYield = CURRENT_TIME;
1090 IF_DEBUG(gran, DumpGranEvent(GR_DESCHEDULE, t));
1091 globalGranStats.tot_heapover++;
1093 globalParStats.tot_heapover++;
1096 // did the task ask for a large block?
1097 if (cap->r.rHpAlloc > BLOCK_SIZE_W) {
1098 // if so, get one and push it on the front of the nursery.
1102 blocks = (nat)BLOCK_ROUND_UP(cap->r.rHpAlloc * sizeof(W_)) / BLOCK_SIZE;
1104 IF_DEBUG(scheduler,belch("--<< thread %ld (%s) stopped: requesting a large block (size %d)",
1105 t->id, whatNext_strs[t->what_next], blocks));
1107 // don't do this if it would push us over the
1108 // alloc_blocks_lim limit; we'll GC first.
1109 if (alloc_blocks + blocks < alloc_blocks_lim) {
1111 alloc_blocks += blocks;
1112 bd = allocGroup( blocks );
1114 // link the new group into the list
1115 bd->link = cap->r.rCurrentNursery;
1116 bd->u.back = cap->r.rCurrentNursery->u.back;
1117 if (cap->r.rCurrentNursery->u.back != NULL) {
1118 cap->r.rCurrentNursery->u.back->link = bd;
1120 ASSERT(g0s0->blocks == cap->r.rCurrentNursery &&
1121 g0s0->blocks == cap->r.rNursery);
1122 cap->r.rNursery = g0s0->blocks = bd;
1124 cap->r.rCurrentNursery->u.back = bd;
1126 // initialise it as a nursery block. We initialise the
1127 // step, gen_no, and flags field of *every* sub-block in
1128 // this large block, because this is easier than making
1129 // sure that we always find the block head of a large
1130 // block whenever we call Bdescr() (eg. evacuate() and
1131 // isAlive() in the GC would both have to do this, at
1135 for (x = bd; x < bd + blocks; x++) {
1142 // don't forget to update the block count in g0s0.
1143 g0s0->n_blocks += blocks;
1144 // This assert can be a killer if the app is doing lots
1145 // of large block allocations.
1146 ASSERT(countBlocks(g0s0->blocks) == g0s0->n_blocks);
1148 // now update the nursery to point to the new block
1149 cap->r.rCurrentNursery = bd;
1151 // we might be unlucky and have another thread get on the
1152 // run queue before us and steal the large block, but in that
1153 // case the thread will just end up requesting another large
1155 PUSH_ON_RUN_QUEUE(t);
1160 /* make all the running tasks block on a condition variable,
1161 * maybe set context_switch and wait till they all pile in,
1162 * then have them wait on a GC condition variable.
1164 IF_DEBUG(scheduler,belch("--<< thread %ld (%s) stopped: HeapOverflow",
1165 t->id, whatNext_strs[t->what_next]));
1168 ASSERT(!is_on_queue(t,CurrentProc));
1170 /* Currently we emit a DESCHEDULE event before GC in GUM.
1171 ToDo: either add separate event to distinguish SYSTEM time from rest
1172 or just nuke this DESCHEDULE (and the following SCHEDULE) */
1173 if (0 && RtsFlags.ParFlags.ParStats.Full) {
1174 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
1175 GR_DESCHEDULE, t, (StgClosure *)NULL, 0, 0);
1176 emitSchedule = rtsTrue;
1180 ready_to_gc = rtsTrue;
1181 context_switch = 1; /* stop other threads ASAP */
1182 PUSH_ON_RUN_QUEUE(t);
1183 /* actual GC is done at the end of the while loop */
1189 DumpGranEvent(GR_DESCHEDULE, t));
1190 globalGranStats.tot_stackover++;
1193 // DumpGranEvent(GR_DESCHEDULE, t);
1194 globalParStats.tot_stackover++;
1196 IF_DEBUG(scheduler,belch("--<< thread %ld (%s) stopped, StackOverflow",
1197 t->id, whatNext_strs[t->what_next]));
1198 /* just adjust the stack for this thread, then pop it back
1204 /* enlarge the stack */
1205 StgTSO *new_t = threadStackOverflow(t);
1207 /* This TSO has moved, so update any pointers to it from the
1208 * main thread stack. It better not be on any other queues...
1209 * (it shouldn't be).
1211 for (m = main_threads; m != NULL; m = m->link) {
1216 threadPaused(new_t);
1217 PUSH_ON_RUN_QUEUE(new_t);
1221 case ThreadYielding:
1224 DumpGranEvent(GR_DESCHEDULE, t));
1225 globalGranStats.tot_yields++;
1228 // DumpGranEvent(GR_DESCHEDULE, t);
1229 globalParStats.tot_yields++;
1231 /* put the thread back on the run queue. Then, if we're ready to
1232 * GC, check whether this is the last task to stop. If so, wake
1233 * up the GC thread. getThread will block during a GC until the
1237 if (t->what_next != prev_what_next) {
1238 /* ToDo: or maybe a timer expired when we were in Hugs?
1239 * or maybe someone hit ctrl-C
1241 belch("--<< thread %ld (%s) stopped to switch evaluators",
1242 t->id, whatNext_strs[t->what_next]);
1244 belch("--<< thread %ld (%s) stopped, yielding",
1245 t->id, whatNext_strs[t->what_next]);
1252 //belch("&& Doing sanity check on yielding TSO %ld.", t->id);
1254 ASSERT(t->link == END_TSO_QUEUE);
1256 ASSERT(!is_on_queue(t,CurrentProc));
1259 //belch("&& Doing sanity check on all ThreadQueues (and their TSOs).");
1260 checkThreadQsSanity(rtsTrue));
1263 if (RtsFlags.ParFlags.doFairScheduling) {
1264 /* this does round-robin scheduling; good for concurrency */
1265 APPEND_TO_RUN_QUEUE(t);
1267 /* this does unfair scheduling; good for parallelism */
1268 PUSH_ON_RUN_QUEUE(t);
1271 if (t->what_next != prev_what_next) {
1272 // switching evaluators; don't context-switch
1273 PUSH_ON_RUN_QUEUE(t);
1275 // this does round-robin scheduling; good for concurrency
1276 APPEND_TO_RUN_QUEUE(t);
1280 /* add a ContinueThread event to actually process the thread */
1281 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1283 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1285 belch("GRAN: eventq and runnableq after adding yielded thread to queue again:");
1294 belch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: ",
1295 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)));
1296 if (t->block_info.closure!=(StgClosure*)NULL) print_bq(t->block_info.closure));
1298 // ??? needed; should emit block before
1300 DumpGranEvent(GR_DESCHEDULE, t));
1301 prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
1304 ASSERT(procStatus[CurrentProc]==Busy ||
1305 ((procStatus[CurrentProc]==Fetching) &&
1306 (t->block_info.closure!=(StgClosure*)NULL)));
1307 if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
1308 !(!RtsFlags.GranFlags.DoAsyncFetch &&
1309 procStatus[CurrentProc]==Fetching))
1310 procStatus[CurrentProc] = Idle;
1314 belch("--<< thread %ld (%p; %s) stopped, blocking on node %p with BQ: ",
1315 t->id, t, whatNext_strs[t->what_next], t->block_info.closure));
1318 if (t->block_info.closure!=(StgClosure*)NULL)
1319 print_bq(t->block_info.closure));
1321 /* Send a fetch (if BlockedOnGA) and dump event to log file */
1324 /* whatever we schedule next, we must log that schedule */
1325 emitSchedule = rtsTrue;
1328 /* don't need to do anything. Either the thread is blocked on
1329 * I/O, in which case we'll have called addToBlockedQueue
1330 * previously, or it's blocked on an MVar or Blackhole, in which
1331 * case it'll be on the relevant queue already.
1334 fprintf(stderr, "--<< thread %d (%s) stopped: ",
1335 t->id, whatNext_strs[t->what_next]);
1336 printThreadBlockage(t);
1337 fprintf(stderr, "\n"));
1339 /* Only for dumping event to log file
1340 ToDo: do I need this in GranSim, too?
1347 case ThreadFinished:
1348 /* Need to check whether this was a main thread, and if so, signal
1349 * the task that started it with the return value. If we have no
1350 * more main threads, we probably need to stop all the tasks until
1353 /* We also end up here if the thread kills itself with an
1354 * uncaught exception, see Exception.hc.
1356 IF_DEBUG(scheduler,belch("--++ thread %d (%s) finished",
1357 t->id, whatNext_strs[t->what_next]));
1359 endThread(t, CurrentProc); // clean-up the thread
1361 /* For now all are advisory -- HWL */
1362 //if(t->priority==AdvisoryPriority) ??
1363 advisory_thread_count--;
1366 if(t->dist.priority==RevalPriority)
1370 if (RtsFlags.ParFlags.ParStats.Full &&
1371 !RtsFlags.ParFlags.ParStats.Suppressed)
1372 DumpEndEvent(CURRENT_PROC, t, rtsFalse /* not mandatory */);
1377 barf("schedule: invalid thread return code %d", (int)ret);
1381 // When we have +RTS -i0 and we're heap profiling, do a census at
1382 // every GC. This lets us get repeatable runs for debugging.
1383 if (performHeapProfile ||
1384 (RtsFlags.ProfFlags.profileInterval==0 &&
1385 RtsFlags.ProfFlags.doHeapProfile && ready_to_gc)) {
1386 GarbageCollect(GetRoots, rtsTrue);
1388 performHeapProfile = rtsFalse;
1389 ready_to_gc = rtsFalse; // we already GC'd
1395 && allFreeCapabilities()
1398 /* everybody back, start the GC.
1399 * Could do it in this thread, or signal a condition var
1400 * to do it in another thread. Either way, we need to
1401 * broadcast on gc_pending_cond afterward.
1403 #if defined(RTS_SUPPORTS_THREADS)
1404 IF_DEBUG(scheduler,sched_belch("doing GC"));
1406 GarbageCollect(GetRoots,rtsFalse);
1407 ready_to_gc = rtsFalse;
1409 broadcastCondition(&gc_pending_cond);
1412 /* add a ContinueThread event to continue execution of current thread */
1413 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1415 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1417 fprintf(stderr, "GRAN: eventq and runnableq after Garbage collection:\n");
1425 IF_GRAN_DEBUG(unused,
1426 print_eventq(EventHd));
1428 event = get_next_event();
1431 /* ToDo: wait for next message to arrive rather than busy wait */
1434 } /* end of while(1) */
1436 IF_PAR_DEBUG(verbose,
1437 belch("== Leaving schedule() after having received Finish"));
1440 /* ---------------------------------------------------------------------------
1441 * Singleton fork(). Do not copy any running threads.
1442 * ------------------------------------------------------------------------- */
1444 StgInt forkProcess(StgTSO* tso) {
1446 #ifndef mingw32_TARGET_OS
1452 IF_DEBUG(scheduler,sched_belch("forking!"));
1455 if (pid) { /* parent */
1457 /* just return the pid */
1459 } else { /* child */
1460 /* wipe all other threads */
1461 run_queue_hd = run_queue_tl = tso;
1462 tso->link = END_TSO_QUEUE;
1464 /* When clearing out the threads, we need to ensure
1465 that a 'main thread' is left behind; if there isn't,
1466 the Scheduler will shutdown next time it is entered.
1468 ==> we don't kill a thread that's on the main_threads
1469 list (nor the current thread.)
1471 [ Attempts at implementing the more ambitious scheme of
1472 killing the main_threads also, and then adding the
1473 current thread onto the main_threads list if it wasn't
1474 there already, failed -- waitThread() (for one) wasn't
1475 up to it. If it proves to be desirable to also kill
1476 the main threads, then this scheme will have to be
1477 revisited (and fully debugged!)
1482 /* DO NOT TOUCH THE QUEUES directly because most of the code around
1483 us is picky about finding the thread still in its queue when
1484 handling the deleteThread() */
1486 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
1489 /* Don't kill the current thread.. */
1490 if (t->id == tso->id) continue;
1492 /* ..or a main thread */
1493 for (m = main_threads; m != NULL; m = m->link) {
1494 if (m->tso->id == t->id) {
1506 barf("forkProcess#: primop not implemented for mingw32, sorry! (%u)\n", tso->id);
1507 /* pointlessly printing out the TSOs 'id' to avoid CC unused warning. */
1509 #endif /* mingw32 */
1512 /* ---------------------------------------------------------------------------
1513 * deleteAllThreads(): kill all the live threads.
1515 * This is used when we catch a user interrupt (^C), before performing
1516 * any necessary cleanups and running finalizers.
1518 * Locks: sched_mutex held.
1519 * ------------------------------------------------------------------------- */
1521 void deleteAllThreads ( void )
1524 IF_DEBUG(scheduler,sched_belch("deleting all threads"));
1525 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
1526 next = t->global_link;
1529 run_queue_hd = run_queue_tl = END_TSO_QUEUE;
1530 blocked_queue_hd = blocked_queue_tl = END_TSO_QUEUE;
1531 sleeping_queue = END_TSO_QUEUE;
1534 /* startThread and insertThread are now in GranSim.c -- HWL */
1537 //@node Suspend and Resume, Run queue code, Main scheduling loop, Main scheduling code
1538 //@subsection Suspend and Resume
1540 /* ---------------------------------------------------------------------------
1541 * Suspending & resuming Haskell threads.
1543 * When making a "safe" call to C (aka _ccall_GC), the task gives back
1544 * its capability before calling the C function. This allows another
1545 * task to pick up the capability and carry on running Haskell
1546 * threads. It also means that if the C call blocks, it won't lock
1549 * The Haskell thread making the C call is put to sleep for the
1550 * duration of the call, on the susepended_ccalling_threads queue. We
1551 * give out a token to the task, which it can use to resume the thread
1552 * on return from the C function.
1553 * ------------------------------------------------------------------------- */
1556 suspendThread( StgRegTable *reg,
1558 #if !defined(RTS_SUPPORTS_THREADS) && !defined(DEBUG)
1566 /* assume that *reg is a pointer to the StgRegTable part
1569 cap = (Capability *)((void *)reg - sizeof(StgFunTable));
1571 ACQUIRE_LOCK(&sched_mutex);
1574 sched_belch("thread %d did a _ccall_gc (is_concurrent: %d)", cap->r.rCurrentTSO->id,concCall));
1576 // XXX this might not be necessary --SDM
1577 cap->r.rCurrentTSO->what_next = ThreadRunGHC;
1579 threadPaused(cap->r.rCurrentTSO);
1580 cap->r.rCurrentTSO->link = suspended_ccalling_threads;
1581 suspended_ccalling_threads = cap->r.rCurrentTSO;
1583 #if defined(RTS_SUPPORTS_THREADS)
1584 cap->r.rCurrentTSO->why_blocked = BlockedOnCCall;
1587 /* Use the thread ID as the token; it should be unique */
1588 tok = cap->r.rCurrentTSO->id;
1590 /* Hand back capability */
1591 releaseCapability(cap);
1593 #if defined(RTS_SUPPORTS_THREADS)
1594 /* Preparing to leave the RTS, so ensure there's a native thread/task
1595 waiting to take over.
1597 ToDo: optimise this and only create a new task if there's a need
1598 for one (i.e., if there's only one Concurrent Haskell thread alive,
1599 there's no need to create a new task).
1601 IF_DEBUG(scheduler, sched_belch("worker thread (%d): leaving RTS", tok));
1603 startTask(taskStart);
1607 /* Other threads _might_ be available for execution; signal this */
1609 RELEASE_LOCK(&sched_mutex);
1614 resumeThread( StgInt tok,
1616 #if !defined(RTS_SUPPORTS_THREADS)
1621 StgTSO *tso, **prev;
1624 #if defined(RTS_SUPPORTS_THREADS)
1625 /* Wait for permission to re-enter the RTS with the result. */
1627 ACQUIRE_LOCK(&sched_mutex);
1628 grabReturnCapability(&sched_mutex, &cap);
1630 grabCapability(&cap);
1633 grabCapability(&cap);
1636 /* Remove the thread off of the suspended list */
1637 prev = &suspended_ccalling_threads;
1638 for (tso = suspended_ccalling_threads;
1639 tso != END_TSO_QUEUE;
1640 prev = &tso->link, tso = tso->link) {
1641 if (tso->id == (StgThreadID)tok) {
1646 if (tso == END_TSO_QUEUE) {
1647 barf("resumeThread: thread not found");
1649 tso->link = END_TSO_QUEUE;
1650 /* Reset blocking status */
1651 tso->why_blocked = NotBlocked;
1653 cap->r.rCurrentTSO = tso;
1654 RELEASE_LOCK(&sched_mutex);
1659 /* ---------------------------------------------------------------------------
1661 * ------------------------------------------------------------------------ */
1662 static void unblockThread(StgTSO *tso);
1664 /* ---------------------------------------------------------------------------
1665 * Comparing Thread ids.
1667 * This is used from STG land in the implementation of the
1668 * instances of Eq/Ord for ThreadIds.
1669 * ------------------------------------------------------------------------ */
1672 cmp_thread(StgPtr tso1, StgPtr tso2)
1674 StgThreadID id1 = ((StgTSO *)tso1)->id;
1675 StgThreadID id2 = ((StgTSO *)tso2)->id;
1677 if (id1 < id2) return (-1);
1678 if (id1 > id2) return 1;
1682 /* ---------------------------------------------------------------------------
1683 * Fetching the ThreadID from an StgTSO.
1685 * This is used in the implementation of Show for ThreadIds.
1686 * ------------------------------------------------------------------------ */
1688 rts_getThreadId(StgPtr tso)
1690 return ((StgTSO *)tso)->id;
1695 labelThread(StgPtr tso, char *label)
1700 /* Caveat: Once set, you can only set the thread name to "" */
1701 len = strlen(label)+1;
1704 fprintf(stderr,"insufficient memory for labelThread!\n");
1706 strncpy(buf,label,len);
1707 /* Update will free the old memory for us */
1708 updateThreadLabel((StgWord)tso,buf);
1712 /* ---------------------------------------------------------------------------
1713 Create a new thread.
1715 The new thread starts with the given stack size. Before the
1716 scheduler can run, however, this thread needs to have a closure
1717 (and possibly some arguments) pushed on its stack. See
1718 pushClosure() in Schedule.h.
1720 createGenThread() and createIOThread() (in SchedAPI.h) are
1721 convenient packaged versions of this function.
1723 currently pri (priority) is only used in a GRAN setup -- HWL
1724 ------------------------------------------------------------------------ */
1725 //@cindex createThread
1727 /* currently pri (priority) is only used in a GRAN setup -- HWL */
1729 createThread(nat size, StgInt pri)
1732 createThread(nat size)
1739 /* First check whether we should create a thread at all */
1741 /* check that no more than RtsFlags.ParFlags.maxThreads threads are created */
1742 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads) {
1744 belch("{createThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
1745 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
1746 return END_TSO_QUEUE;
1752 ASSERT(!RtsFlags.GranFlags.Light || CurrentProc==0);
1755 // ToDo: check whether size = stack_size - TSO_STRUCT_SIZEW
1757 /* catch ridiculously small stack sizes */
1758 if (size < MIN_STACK_WORDS + TSO_STRUCT_SIZEW) {
1759 size = MIN_STACK_WORDS + TSO_STRUCT_SIZEW;
1762 stack_size = size - TSO_STRUCT_SIZEW;
1764 tso = (StgTSO *)allocate(size);
1765 TICK_ALLOC_TSO(stack_size, 0);
1767 SET_HDR(tso, &stg_TSO_info, CCS_SYSTEM);
1769 SET_GRAN_HDR(tso, ThisPE);
1772 // Always start with the compiled code evaluator
1773 tso->what_next = ThreadRunGHC;
1775 /* tso->id needs to be unique. For now we use a heavyweight mutex to
1776 * protect the increment operation on next_thread_id.
1777 * In future, we could use an atomic increment instead.
1779 ACQUIRE_LOCK(&thread_id_mutex);
1780 tso->id = next_thread_id++;
1781 RELEASE_LOCK(&thread_id_mutex);
1783 tso->why_blocked = NotBlocked;
1784 tso->blocked_exceptions = NULL;
1786 tso->stack_size = stack_size;
1787 tso->max_stack_size = round_to_mblocks(RtsFlags.GcFlags.maxStkSize)
1789 tso->sp = (P_)&(tso->stack) + stack_size;
1792 tso->prof.CCCS = CCS_MAIN;
1795 /* put a stop frame on the stack */
1796 tso->sp -= sizeofW(StgStopFrame);
1797 SET_HDR((StgClosure*)tso->sp,(StgInfoTable *)&stg_stop_thread_info,CCS_SYSTEM);
1800 tso->link = END_TSO_QUEUE;
1801 /* uses more flexible routine in GranSim */
1802 insertThread(tso, CurrentProc);
1804 /* In a non-GranSim setup the pushing of a TSO onto the runq is separated
1810 if (RtsFlags.GranFlags.GranSimStats.Full)
1811 DumpGranEvent(GR_START,tso);
1813 if (RtsFlags.ParFlags.ParStats.Full)
1814 DumpGranEvent(GR_STARTQ,tso);
1815 /* HACk to avoid SCHEDULE
1819 /* Link the new thread on the global thread list.
1821 tso->global_link = all_threads;
1825 tso->dist.priority = MandatoryPriority; //by default that is...
1829 tso->gran.pri = pri;
1831 tso->gran.magic = TSO_MAGIC; // debugging only
1833 tso->gran.sparkname = 0;
1834 tso->gran.startedat = CURRENT_TIME;
1835 tso->gran.exported = 0;
1836 tso->gran.basicblocks = 0;
1837 tso->gran.allocs = 0;
1838 tso->gran.exectime = 0;
1839 tso->gran.fetchtime = 0;
1840 tso->gran.fetchcount = 0;
1841 tso->gran.blocktime = 0;
1842 tso->gran.blockcount = 0;
1843 tso->gran.blockedat = 0;
1844 tso->gran.globalsparks = 0;
1845 tso->gran.localsparks = 0;
1846 if (RtsFlags.GranFlags.Light)
1847 tso->gran.clock = Now; /* local clock */
1849 tso->gran.clock = 0;
1851 IF_DEBUG(gran,printTSO(tso));
1854 tso->par.magic = TSO_MAGIC; // debugging only
1856 tso->par.sparkname = 0;
1857 tso->par.startedat = CURRENT_TIME;
1858 tso->par.exported = 0;
1859 tso->par.basicblocks = 0;
1860 tso->par.allocs = 0;
1861 tso->par.exectime = 0;
1862 tso->par.fetchtime = 0;
1863 tso->par.fetchcount = 0;
1864 tso->par.blocktime = 0;
1865 tso->par.blockcount = 0;
1866 tso->par.blockedat = 0;
1867 tso->par.globalsparks = 0;
1868 tso->par.localsparks = 0;
1872 globalGranStats.tot_threads_created++;
1873 globalGranStats.threads_created_on_PE[CurrentProc]++;
1874 globalGranStats.tot_sq_len += spark_queue_len(CurrentProc);
1875 globalGranStats.tot_sq_probes++;
1877 // collect parallel global statistics (currently done together with GC stats)
1878 if (RtsFlags.ParFlags.ParStats.Global &&
1879 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
1880 //fprintf(stderr, "Creating thread %d @ %11.2f\n", tso->id, usertime());
1881 globalParStats.tot_threads_created++;
1887 belch("==__ schedule: Created TSO %d (%p);",
1888 CurrentProc, tso, tso->id));
1890 IF_PAR_DEBUG(verbose,
1891 belch("==__ schedule: Created TSO %d (%p); %d threads active",
1892 tso->id, tso, advisory_thread_count));
1894 IF_DEBUG(scheduler,sched_belch("created thread %ld, stack size = %lx words",
1895 tso->id, tso->stack_size));
1902 all parallel thread creation calls should fall through the following routine.
1905 createSparkThread(rtsSpark spark)
1907 ASSERT(spark != (rtsSpark)NULL);
1908 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads)
1910 barf("{createSparkThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
1911 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
1912 return END_TSO_QUEUE;
1916 tso = createThread(RtsFlags.GcFlags.initialStkSize);
1917 if (tso==END_TSO_QUEUE)
1918 barf("createSparkThread: Cannot create TSO");
1920 tso->priority = AdvisoryPriority;
1922 pushClosure(tso,spark);
1923 PUSH_ON_RUN_QUEUE(tso);
1924 advisory_thread_count++;
1931 Turn a spark into a thread.
1932 ToDo: fix for SMP (needs to acquire SCHED_MUTEX!)
1935 //@cindex activateSpark
1937 activateSpark (rtsSpark spark)
1941 tso = createSparkThread(spark);
1942 if (RtsFlags.ParFlags.ParStats.Full) {
1943 //ASSERT(run_queue_hd == END_TSO_QUEUE); // I think ...
1944 IF_PAR_DEBUG(verbose,
1945 belch("==^^ activateSpark: turning spark of closure %p (%s) into a thread",
1946 (StgClosure *)spark, info_type((StgClosure *)spark)));
1948 // ToDo: fwd info on local/global spark to thread -- HWL
1949 // tso->gran.exported = spark->exported;
1950 // tso->gran.locked = !spark->global;
1951 // tso->gran.sparkname = spark->name;
1957 static SchedulerStatus waitThread_(/*out*/StgMainThread* m
1958 #if defined(THREADED_RTS)
1959 , rtsBool blockWaiting
1964 /* ---------------------------------------------------------------------------
1967 * scheduleThread puts a thread on the head of the runnable queue.
1968 * This will usually be done immediately after a thread is created.
1969 * The caller of scheduleThread must create the thread using e.g.
1970 * createThread and push an appropriate closure
1971 * on this thread's stack before the scheduler is invoked.
1972 * ------------------------------------------------------------------------ */
1974 static void scheduleThread_ (StgTSO* tso, rtsBool createTask);
1977 scheduleThread_(StgTSO *tso
1978 , rtsBool createTask
1979 #if !defined(THREADED_RTS)
1984 // Precondition: sched_mutex must be held.
1986 /* Put the new thread on the head of the runnable queue. The caller
1987 * better push an appropriate closure on this thread's stack
1988 * beforehand. In the SMP case, the thread may start running as
1989 * soon as we release the scheduler lock below.
1991 PUSH_ON_RUN_QUEUE(tso);
1992 #if defined(THREADED_RTS)
1993 /* If main() is scheduling a thread, don't bother creating a
1997 startTask(taskStart);
2003 IF_DEBUG(scheduler,printTSO(tso));
2007 void scheduleThread(StgTSO* tso)
2009 ACQUIRE_LOCK(&sched_mutex);
2010 scheduleThread_(tso, rtsFalse);
2011 RELEASE_LOCK(&sched_mutex);
2015 scheduleWaitThread(StgTSO* tso, /*[out]*/HaskellObj* ret)
2019 m = stgMallocBytes(sizeof(StgMainThread), "waitThread");
2023 #if defined(RTS_SUPPORTS_THREADS)
2024 initCondition(&m->wakeup);
2027 /* Put the thread on the main-threads list prior to scheduling the TSO.
2028 Failure to do so introduces a race condition in the MT case (as
2029 identified by Wolfgang Thaller), whereby the new task/OS thread
2030 created by scheduleThread_() would complete prior to the thread
2031 that spawned it managed to put 'itself' on the main-threads list.
2032 The upshot of it all being that the worker thread wouldn't get to
2033 signal the completion of the its work item for the main thread to
2034 see (==> it got stuck waiting.) -- sof 6/02.
2036 ACQUIRE_LOCK(&sched_mutex);
2037 IF_DEBUG(scheduler, sched_belch("waiting for thread (%d)\n", tso->id));
2039 m->link = main_threads;
2042 scheduleThread_(tso, rtsTrue);
2043 #if defined(THREADED_RTS)
2044 return waitThread_(m, rtsTrue); // waitThread_ releases sched_mutex
2046 return waitThread_(m);
2050 /* ---------------------------------------------------------------------------
2053 * Initialise the scheduler. This resets all the queues - if the
2054 * queues contained any threads, they'll be garbage collected at the
2057 * ------------------------------------------------------------------------ */
2061 term_handler(int sig STG_UNUSED)
2064 ACQUIRE_LOCK(&term_mutex);
2066 RELEASE_LOCK(&term_mutex);
2077 for (i=0; i<=MAX_PROC; i++) {
2078 run_queue_hds[i] = END_TSO_QUEUE;
2079 run_queue_tls[i] = END_TSO_QUEUE;
2080 blocked_queue_hds[i] = END_TSO_QUEUE;
2081 blocked_queue_tls[i] = END_TSO_QUEUE;
2082 ccalling_threadss[i] = END_TSO_QUEUE;
2083 sleeping_queue = END_TSO_QUEUE;
2086 run_queue_hd = END_TSO_QUEUE;
2087 run_queue_tl = END_TSO_QUEUE;
2088 blocked_queue_hd = END_TSO_QUEUE;
2089 blocked_queue_tl = END_TSO_QUEUE;
2090 sleeping_queue = END_TSO_QUEUE;
2093 suspended_ccalling_threads = END_TSO_QUEUE;
2095 main_threads = NULL;
2096 all_threads = END_TSO_QUEUE;
2101 RtsFlags.ConcFlags.ctxtSwitchTicks =
2102 RtsFlags.ConcFlags.ctxtSwitchTime / TICK_MILLISECS;
2104 #if defined(RTS_SUPPORTS_THREADS)
2105 /* Initialise the mutex and condition variables used by
2107 initMutex(&sched_mutex);
2108 initMutex(&term_mutex);
2109 initMutex(&thread_id_mutex);
2111 initCondition(&thread_ready_cond);
2115 initCondition(&gc_pending_cond);
2118 #if defined(RTS_SUPPORTS_THREADS)
2119 ACQUIRE_LOCK(&sched_mutex);
2122 /* Install the SIGHUP handler */
2125 struct sigaction action,oact;
2127 action.sa_handler = term_handler;
2128 sigemptyset(&action.sa_mask);
2129 action.sa_flags = 0;
2130 if (sigaction(SIGTERM, &action, &oact) != 0) {
2131 barf("can't install TERM handler");
2136 /* A capability holds the state a native thread needs in
2137 * order to execute STG code. At least one capability is
2138 * floating around (only SMP builds have more than one).
2142 #if defined(RTS_SUPPORTS_THREADS)
2143 /* start our haskell execution tasks */
2145 startTaskManager(RtsFlags.ParFlags.nNodes, taskStart);
2147 startTaskManager(0,taskStart);
2151 #if /* defined(SMP) ||*/ defined(PAR)
2155 #if defined(RTS_SUPPORTS_THREADS)
2156 RELEASE_LOCK(&sched_mutex);
2162 exitScheduler( void )
2164 #if defined(RTS_SUPPORTS_THREADS)
2167 shutting_down_scheduler = rtsTrue;
2170 /* -----------------------------------------------------------------------------
2171 Managing the per-task allocation areas.
2173 Each capability comes with an allocation area. These are
2174 fixed-length block lists into which allocation can be done.
2176 ToDo: no support for two-space collection at the moment???
2177 -------------------------------------------------------------------------- */
2179 /* -----------------------------------------------------------------------------
2180 * waitThread is the external interface for running a new computation
2181 * and waiting for the result.
2183 * In the non-SMP case, we create a new main thread, push it on the
2184 * main-thread stack, and invoke the scheduler to run it. The
2185 * scheduler will return when the top main thread on the stack has
2186 * completed or died, and fill in the necessary fields of the
2187 * main_thread structure.
2189 * In the SMP case, we create a main thread as before, but we then
2190 * create a new condition variable and sleep on it. When our new
2191 * main thread has completed, we'll be woken up and the status/result
2192 * will be in the main_thread struct.
2193 * -------------------------------------------------------------------------- */
2196 howManyThreadsAvail ( void )
2200 for (q = run_queue_hd; q != END_TSO_QUEUE; q = q->link)
2202 for (q = blocked_queue_hd; q != END_TSO_QUEUE; q = q->link)
2204 for (q = sleeping_queue; q != END_TSO_QUEUE; q = q->link)
2210 finishAllThreads ( void )
2213 while (run_queue_hd != END_TSO_QUEUE) {
2214 waitThread ( run_queue_hd, NULL);
2216 while (blocked_queue_hd != END_TSO_QUEUE) {
2217 waitThread ( blocked_queue_hd, NULL);
2219 while (sleeping_queue != END_TSO_QUEUE) {
2220 waitThread ( blocked_queue_hd, NULL);
2223 (blocked_queue_hd != END_TSO_QUEUE ||
2224 run_queue_hd != END_TSO_QUEUE ||
2225 sleeping_queue != END_TSO_QUEUE);
2229 waitThread(StgTSO *tso, /*out*/StgClosure **ret)
2233 m = stgMallocBytes(sizeof(StgMainThread), "waitThread");
2237 #if defined(RTS_SUPPORTS_THREADS)
2238 initCondition(&m->wakeup);
2241 /* see scheduleWaitThread() comment */
2242 ACQUIRE_LOCK(&sched_mutex);
2243 m->link = main_threads;
2246 IF_DEBUG(scheduler, sched_belch("waiting for thread %d", tso->id));
2247 #if defined(THREADED_RTS)
2248 return waitThread_(m, rtsFalse); // waitThread_ releases sched_mutex
2250 return waitThread_(m);
2256 waitThread_(StgMainThread* m
2257 #if defined(THREADED_RTS)
2258 , rtsBool blockWaiting
2262 SchedulerStatus stat;
2264 // Precondition: sched_mutex must be held.
2265 IF_DEBUG(scheduler, sched_belch("== scheduler: new main thread (%d)\n", m->tso->id));
2267 #if defined(RTS_SUPPORTS_THREADS)
2269 # if defined(THREADED_RTS)
2270 if (!blockWaiting) {
2271 /* In the threaded case, the OS thread that called main()
2272 * gets to enter the RTS directly without going via another
2275 RELEASE_LOCK(&sched_mutex);
2277 ASSERT(m->stat != NoStatus);
2282 waitCondition(&m->wakeup, &sched_mutex);
2283 } while (m->stat == NoStatus);
2286 /* GranSim specific init */
2287 CurrentTSO = m->tso; // the TSO to run
2288 procStatus[MainProc] = Busy; // status of main PE
2289 CurrentProc = MainProc; // PE to run it on
2291 RELEASE_LOCK(&sched_mutex);
2294 RELEASE_LOCK(&sched_mutex);
2296 ASSERT(m->stat != NoStatus);
2301 #if defined(RTS_SUPPORTS_THREADS)
2302 closeCondition(&m->wakeup);
2305 IF_DEBUG(scheduler, fprintf(stderr, "== scheduler: main thread (%d) finished\n",
2309 #if defined(THREADED_RTS)
2312 RELEASE_LOCK(&sched_mutex);
2314 // Postcondition: sched_mutex must not be held
2318 //@node Run queue code, Garbage Collextion Routines, Suspend and Resume, Main scheduling code
2319 //@subsection Run queue code
2323 NB: In GranSim we have many run queues; run_queue_hd is actually a macro
2324 unfolding to run_queue_hds[CurrentProc], thus CurrentProc is an
2325 implicit global variable that has to be correct when calling these
2329 /* Put the new thread on the head of the runnable queue.
2330 * The caller of createThread better push an appropriate closure
2331 * on this thread's stack before the scheduler is invoked.
2333 static /* inline */ void
2334 add_to_run_queue(tso)
2337 ASSERT(tso!=run_queue_hd && tso!=run_queue_tl);
2338 tso->link = run_queue_hd;
2340 if (run_queue_tl == END_TSO_QUEUE) {
2345 /* Put the new thread at the end of the runnable queue. */
2346 static /* inline */ void
2347 push_on_run_queue(tso)
2350 ASSERT(get_itbl((StgClosure *)tso)->type == TSO);
2351 ASSERT(run_queue_hd!=NULL && run_queue_tl!=NULL);
2352 ASSERT(tso!=run_queue_hd && tso!=run_queue_tl);
2353 if (run_queue_hd == END_TSO_QUEUE) {
2356 run_queue_tl->link = tso;
2362 Should be inlined because it's used very often in schedule. The tso
2363 argument is actually only needed in GranSim, where we want to have the
2364 possibility to schedule *any* TSO on the run queue, irrespective of the
2365 actual ordering. Therefore, if tso is not the nil TSO then we traverse
2366 the run queue and dequeue the tso, adjusting the links in the queue.
2368 //@cindex take_off_run_queue
2369 static /* inline */ StgTSO*
2370 take_off_run_queue(StgTSO *tso) {
2374 qetlaHbogh Qu' ngaSbogh ghomDaQ {tso} yIteq!
2376 if tso is specified, unlink that tso from the run_queue (doesn't have
2377 to be at the beginning of the queue); GranSim only
2379 if (tso!=END_TSO_QUEUE) {
2380 /* find tso in queue */
2381 for (t=run_queue_hd, prev=END_TSO_QUEUE;
2382 t!=END_TSO_QUEUE && t!=tso;
2386 /* now actually dequeue the tso */
2387 if (prev!=END_TSO_QUEUE) {
2388 ASSERT(run_queue_hd!=t);
2389 prev->link = t->link;
2391 /* t is at beginning of thread queue */
2392 ASSERT(run_queue_hd==t);
2393 run_queue_hd = t->link;
2395 /* t is at end of thread queue */
2396 if (t->link==END_TSO_QUEUE) {
2397 ASSERT(t==run_queue_tl);
2398 run_queue_tl = prev;
2400 ASSERT(run_queue_tl!=t);
2402 t->link = END_TSO_QUEUE;
2404 /* take tso from the beginning of the queue; std concurrent code */
2406 if (t != END_TSO_QUEUE) {
2407 run_queue_hd = t->link;
2408 t->link = END_TSO_QUEUE;
2409 if (run_queue_hd == END_TSO_QUEUE) {
2410 run_queue_tl = END_TSO_QUEUE;
2419 //@node Garbage Collextion Routines, Blocking Queue Routines, Run queue code, Main scheduling code
2420 //@subsection Garbage Collextion Routines
2422 /* ---------------------------------------------------------------------------
2423 Where are the roots that we know about?
2425 - all the threads on the runnable queue
2426 - all the threads on the blocked queue
2427 - all the threads on the sleeping queue
2428 - all the thread currently executing a _ccall_GC
2429 - all the "main threads"
2431 ------------------------------------------------------------------------ */
2433 /* This has to be protected either by the scheduler monitor, or by the
2434 garbage collection monitor (probably the latter).
2439 GetRoots(evac_fn evac)
2444 for (i=0; i<=RtsFlags.GranFlags.proc; i++) {
2445 if ((run_queue_hds[i] != END_TSO_QUEUE) && ((run_queue_hds[i] != NULL)))
2446 evac((StgClosure **)&run_queue_hds[i]);
2447 if ((run_queue_tls[i] != END_TSO_QUEUE) && ((run_queue_tls[i] != NULL)))
2448 evac((StgClosure **)&run_queue_tls[i]);
2450 if ((blocked_queue_hds[i] != END_TSO_QUEUE) && ((blocked_queue_hds[i] != NULL)))
2451 evac((StgClosure **)&blocked_queue_hds[i]);
2452 if ((blocked_queue_tls[i] != END_TSO_QUEUE) && ((blocked_queue_tls[i] != NULL)))
2453 evac((StgClosure **)&blocked_queue_tls[i]);
2454 if ((ccalling_threadss[i] != END_TSO_QUEUE) && ((ccalling_threadss[i] != NULL)))
2455 evac((StgClosure **)&ccalling_threads[i]);
2462 if (run_queue_hd != END_TSO_QUEUE) {
2463 ASSERT(run_queue_tl != END_TSO_QUEUE);
2464 evac((StgClosure **)&run_queue_hd);
2465 evac((StgClosure **)&run_queue_tl);
2468 if (blocked_queue_hd != END_TSO_QUEUE) {
2469 ASSERT(blocked_queue_tl != END_TSO_QUEUE);
2470 evac((StgClosure **)&blocked_queue_hd);
2471 evac((StgClosure **)&blocked_queue_tl);
2474 if (sleeping_queue != END_TSO_QUEUE) {
2475 evac((StgClosure **)&sleeping_queue);
2479 if (suspended_ccalling_threads != END_TSO_QUEUE) {
2480 evac((StgClosure **)&suspended_ccalling_threads);
2483 #if defined(PAR) || defined(GRAN)
2484 markSparkQueue(evac);
2487 #ifndef mingw32_TARGET_OS
2488 // mark the signal handlers (signals should be already blocked)
2489 markSignalHandlers(evac);
2492 // main threads which have completed need to be retained until they
2493 // are dealt with in the main scheduler loop. They won't be
2494 // retained any other way: the GC will drop them from the
2495 // all_threads list, so we have to be careful to treat them as roots
2499 for (m = main_threads; m != NULL; m = m->link) {
2500 switch (m->tso->what_next) {
2501 case ThreadComplete:
2503 evac((StgClosure **)&m->tso);
2512 /* -----------------------------------------------------------------------------
2515 This is the interface to the garbage collector from Haskell land.
2516 We provide this so that external C code can allocate and garbage
2517 collect when called from Haskell via _ccall_GC.
2519 It might be useful to provide an interface whereby the programmer
2520 can specify more roots (ToDo).
2522 This needs to be protected by the GC condition variable above. KH.
2523 -------------------------------------------------------------------------- */
2525 static void (*extra_roots)(evac_fn);
2530 /* Obligated to hold this lock upon entry */
2531 ACQUIRE_LOCK(&sched_mutex);
2532 GarbageCollect(GetRoots,rtsFalse);
2533 RELEASE_LOCK(&sched_mutex);
2537 performMajorGC(void)
2539 ACQUIRE_LOCK(&sched_mutex);
2540 GarbageCollect(GetRoots,rtsTrue);
2541 RELEASE_LOCK(&sched_mutex);
2545 AllRoots(evac_fn evac)
2547 GetRoots(evac); // the scheduler's roots
2548 extra_roots(evac); // the user's roots
2552 performGCWithRoots(void (*get_roots)(evac_fn))
2554 ACQUIRE_LOCK(&sched_mutex);
2555 extra_roots = get_roots;
2556 GarbageCollect(AllRoots,rtsFalse);
2557 RELEASE_LOCK(&sched_mutex);
2560 /* -----------------------------------------------------------------------------
2563 If the thread has reached its maximum stack size, then raise the
2564 StackOverflow exception in the offending thread. Otherwise
2565 relocate the TSO into a larger chunk of memory and adjust its stack
2567 -------------------------------------------------------------------------- */
2570 threadStackOverflow(StgTSO *tso)
2572 nat new_stack_size, new_tso_size, diff, stack_words;
2576 IF_DEBUG(sanity,checkTSO(tso));
2577 if (tso->stack_size >= tso->max_stack_size) {
2580 belch("@@ threadStackOverflow of TSO %d (%p): stack too large (now %ld; max is %ld",
2581 tso->id, tso, tso->stack_size, tso->max_stack_size);
2582 /* If we're debugging, just print out the top of the stack */
2583 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2586 /* Send this thread the StackOverflow exception */
2587 raiseAsync(tso, (StgClosure *)stackOverflow_closure);
2591 /* Try to double the current stack size. If that takes us over the
2592 * maximum stack size for this thread, then use the maximum instead.
2593 * Finally round up so the TSO ends up as a whole number of blocks.
2595 new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
2596 new_tso_size = (nat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) +
2597 TSO_STRUCT_SIZE)/sizeof(W_);
2598 new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */
2599 new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
2601 IF_DEBUG(scheduler, fprintf(stderr,"== scheduler: increasing stack size from %d words to %d.\n", tso->stack_size, new_stack_size));
2603 dest = (StgTSO *)allocate(new_tso_size);
2604 TICK_ALLOC_TSO(new_stack_size,0);
2606 /* copy the TSO block and the old stack into the new area */
2607 memcpy(dest,tso,TSO_STRUCT_SIZE);
2608 stack_words = tso->stack + tso->stack_size - tso->sp;
2609 new_sp = (P_)dest + new_tso_size - stack_words;
2610 memcpy(new_sp, tso->sp, stack_words * sizeof(W_));
2612 /* relocate the stack pointers... */
2613 diff = (P_)new_sp - (P_)tso->sp; /* In *words* */
2615 dest->stack_size = new_stack_size;
2617 /* Mark the old TSO as relocated. We have to check for relocated
2618 * TSOs in the garbage collector and any primops that deal with TSOs.
2620 * It's important to set the sp value to just beyond the end
2621 * of the stack, so we don't attempt to scavenge any part of the
2624 tso->what_next = ThreadRelocated;
2626 tso->sp = (P_)&(tso->stack[tso->stack_size]);
2627 tso->why_blocked = NotBlocked;
2628 dest->mut_link = NULL;
2630 IF_PAR_DEBUG(verbose,
2631 belch("@@ threadStackOverflow of TSO %d (now at %p): stack size increased to %ld",
2632 tso->id, tso, tso->stack_size);
2633 /* If we're debugging, just print out the top of the stack */
2634 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2637 IF_DEBUG(sanity,checkTSO(tso));
2639 IF_DEBUG(scheduler,printTSO(dest));
2645 //@node Blocking Queue Routines, Exception Handling Routines, Garbage Collextion Routines, Main scheduling code
2646 //@subsection Blocking Queue Routines
2648 /* ---------------------------------------------------------------------------
2649 Wake up a queue that was blocked on some resource.
2650 ------------------------------------------------------------------------ */
2654 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2659 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2661 /* write RESUME events to log file and
2662 update blocked and fetch time (depending on type of the orig closure) */
2663 if (RtsFlags.ParFlags.ParStats.Full) {
2664 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
2665 GR_RESUMEQ, ((StgTSO *)bqe), ((StgTSO *)bqe)->block_info.closure,
2666 0, 0 /* spark_queue_len(ADVISORY_POOL) */);
2667 if (EMPTY_RUN_QUEUE())
2668 emitSchedule = rtsTrue;
2670 switch (get_itbl(node)->type) {
2672 ((StgTSO *)bqe)->par.fetchtime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2677 ((StgTSO *)bqe)->par.blocktime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2684 barf("{unblockOneLocked}Daq Qagh: unexpected closure in blocking queue");
2691 static StgBlockingQueueElement *
2692 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2695 PEs node_loc, tso_loc;
2697 node_loc = where_is(node); // should be lifted out of loop
2698 tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2699 tso_loc = where_is((StgClosure *)tso);
2700 if (IS_LOCAL_TO(PROCS(node),tso_loc)) { // TSO is local
2701 /* !fake_fetch => TSO is on CurrentProc is same as IS_LOCAL_TO */
2702 ASSERT(CurrentProc!=node_loc || tso_loc==CurrentProc);
2703 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.lunblocktime;
2704 // insertThread(tso, node_loc);
2705 new_event(tso_loc, tso_loc, CurrentTime[CurrentProc],
2707 tso, node, (rtsSpark*)NULL);
2708 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2711 } else { // TSO is remote (actually should be FMBQ)
2712 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.mpacktime +
2713 RtsFlags.GranFlags.Costs.gunblocktime +
2714 RtsFlags.GranFlags.Costs.latency;
2715 new_event(tso_loc, CurrentProc, CurrentTime[CurrentProc],
2717 tso, node, (rtsSpark*)NULL);
2718 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2721 /* the thread-queue-overhead is accounted for in either Resume or UnblockThread */
2723 fprintf(stderr," %s TSO %d (%p) [PE %d] (block_info.closure=%p) (next=%p) ,",
2724 (node_loc==tso_loc ? "Local" : "Global"),
2725 tso->id, tso, CurrentProc, tso->block_info.closure, tso->link));
2726 tso->block_info.closure = NULL;
2727 IF_DEBUG(scheduler,belch("-- Waking up thread %ld (%p)",
2731 static StgBlockingQueueElement *
2732 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2734 StgBlockingQueueElement *next;
2736 switch (get_itbl(bqe)->type) {
2738 ASSERT(((StgTSO *)bqe)->why_blocked != NotBlocked);
2739 /* if it's a TSO just push it onto the run_queue */
2741 // ((StgTSO *)bqe)->link = END_TSO_QUEUE; // debugging?
2742 PUSH_ON_RUN_QUEUE((StgTSO *)bqe);
2744 unblockCount(bqe, node);
2745 /* reset blocking status after dumping event */
2746 ((StgTSO *)bqe)->why_blocked = NotBlocked;
2750 /* if it's a BLOCKED_FETCH put it on the PendingFetches list */
2752 bqe->link = (StgBlockingQueueElement *)PendingFetches;
2753 PendingFetches = (StgBlockedFetch *)bqe;
2757 /* can ignore this case in a non-debugging setup;
2758 see comments on RBHSave closures above */
2760 /* check that the closure is an RBHSave closure */
2761 ASSERT(get_itbl((StgClosure *)bqe) == &stg_RBH_Save_0_info ||
2762 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_1_info ||
2763 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_2_info);
2767 barf("{unblockOneLocked}Daq Qagh: Unexpected IP (%#lx; %s) in blocking queue at %#lx\n",
2768 get_itbl((StgClosure *)bqe), info_type((StgClosure *)bqe),
2772 IF_PAR_DEBUG(bq, fprintf(stderr, ", %p (%s)", bqe, info_type((StgClosure*)bqe)));
2776 #else /* !GRAN && !PAR */
2778 unblockOneLocked(StgTSO *tso)
2782 ASSERT(get_itbl(tso)->type == TSO);
2783 ASSERT(tso->why_blocked != NotBlocked);
2784 tso->why_blocked = NotBlocked;
2786 PUSH_ON_RUN_QUEUE(tso);
2788 IF_DEBUG(scheduler,sched_belch("waking up thread %ld", tso->id));
2793 #if defined(GRAN) || defined(PAR)
2794 inline StgBlockingQueueElement *
2795 unblockOne(StgBlockingQueueElement *bqe, StgClosure *node)
2797 ACQUIRE_LOCK(&sched_mutex);
2798 bqe = unblockOneLocked(bqe, node);
2799 RELEASE_LOCK(&sched_mutex);
2804 unblockOne(StgTSO *tso)
2806 ACQUIRE_LOCK(&sched_mutex);
2807 tso = unblockOneLocked(tso);
2808 RELEASE_LOCK(&sched_mutex);
2815 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2817 StgBlockingQueueElement *bqe;
2822 belch("##-_ AwBQ for node %p on PE %d @ %ld by TSO %d (%p): ", \
2823 node, CurrentProc, CurrentTime[CurrentProc],
2824 CurrentTSO->id, CurrentTSO));
2826 node_loc = where_is(node);
2828 ASSERT(q == END_BQ_QUEUE ||
2829 get_itbl(q)->type == TSO || // q is either a TSO or an RBHSave
2830 get_itbl(q)->type == CONSTR); // closure (type constructor)
2831 ASSERT(is_unique(node));
2833 /* FAKE FETCH: magically copy the node to the tso's proc;
2834 no Fetch necessary because in reality the node should not have been
2835 moved to the other PE in the first place
2837 if (CurrentProc!=node_loc) {
2839 belch("## node %p is on PE %d but CurrentProc is %d (TSO %d); assuming fake fetch and adjusting bitmask (old: %#x)",
2840 node, node_loc, CurrentProc, CurrentTSO->id,
2841 // CurrentTSO, where_is(CurrentTSO),
2842 node->header.gran.procs));
2843 node->header.gran.procs = (node->header.gran.procs) | PE_NUMBER(CurrentProc);
2845 belch("## new bitmask of node %p is %#x",
2846 node, node->header.gran.procs));
2847 if (RtsFlags.GranFlags.GranSimStats.Global) {
2848 globalGranStats.tot_fake_fetches++;
2853 // ToDo: check: ASSERT(CurrentProc==node_loc);
2854 while (get_itbl(bqe)->type==TSO) { // q != END_TSO_QUEUE) {
2857 bqe points to the current element in the queue
2858 next points to the next element in the queue
2860 //tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2861 //tso_loc = where_is(tso);
2863 bqe = unblockOneLocked(bqe, node);
2866 /* if this is the BQ of an RBH, we have to put back the info ripped out of
2867 the closure to make room for the anchor of the BQ */
2868 if (bqe!=END_BQ_QUEUE) {
2869 ASSERT(get_itbl(node)->type == RBH && get_itbl(bqe)->type == CONSTR);
2871 ASSERT((info_ptr==&RBH_Save_0_info) ||
2872 (info_ptr==&RBH_Save_1_info) ||
2873 (info_ptr==&RBH_Save_2_info));
2875 /* cf. convertToRBH in RBH.c for writing the RBHSave closure */
2876 ((StgRBH *)node)->blocking_queue = (StgBlockingQueueElement *)((StgRBHSave *)bqe)->payload[0];
2877 ((StgRBH *)node)->mut_link = (StgMutClosure *)((StgRBHSave *)bqe)->payload[1];
2880 belch("## Filled in RBH_Save for %p (%s) at end of AwBQ",
2881 node, info_type(node)));
2884 /* statistics gathering */
2885 if (RtsFlags.GranFlags.GranSimStats.Global) {
2886 // globalGranStats.tot_bq_processing_time += bq_processing_time;
2887 globalGranStats.tot_bq_len += len; // total length of all bqs awakened
2888 // globalGranStats.tot_bq_len_local += len_local; // same for local TSOs only
2889 globalGranStats.tot_awbq++; // total no. of bqs awakened
2892 fprintf(stderr,"## BQ Stats of %p: [%d entries] %s\n",
2893 node, len, (bqe!=END_BQ_QUEUE) ? "RBH" : ""));
2897 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2899 StgBlockingQueueElement *bqe;
2901 ACQUIRE_LOCK(&sched_mutex);
2903 IF_PAR_DEBUG(verbose,
2904 belch("##-_ AwBQ for node %p on [%x]: ",
2908 if(get_itbl(q)->type == CONSTR || q==END_BQ_QUEUE) {
2909 IF_PAR_DEBUG(verbose, belch("## ... nothing to unblock so lets just return. RFP (BUG?)"));
2914 ASSERT(q == END_BQ_QUEUE ||
2915 get_itbl(q)->type == TSO ||
2916 get_itbl(q)->type == BLOCKED_FETCH ||
2917 get_itbl(q)->type == CONSTR);
2920 while (get_itbl(bqe)->type==TSO ||
2921 get_itbl(bqe)->type==BLOCKED_FETCH) {
2922 bqe = unblockOneLocked(bqe, node);
2924 RELEASE_LOCK(&sched_mutex);
2927 #else /* !GRAN && !PAR */
2929 awakenBlockedQueue(StgTSO *tso)
2931 ACQUIRE_LOCK(&sched_mutex);
2932 while (tso != END_TSO_QUEUE) {
2933 tso = unblockOneLocked(tso);
2935 RELEASE_LOCK(&sched_mutex);
2939 //@node Exception Handling Routines, Debugging Routines, Blocking Queue Routines, Main scheduling code
2940 //@subsection Exception Handling Routines
2942 /* ---------------------------------------------------------------------------
2944 - usually called inside a signal handler so it mustn't do anything fancy.
2945 ------------------------------------------------------------------------ */
2948 interruptStgRts(void)
2954 /* -----------------------------------------------------------------------------
2957 This is for use when we raise an exception in another thread, which
2959 This has nothing to do with the UnblockThread event in GranSim. -- HWL
2960 -------------------------------------------------------------------------- */
2962 #if defined(GRAN) || defined(PAR)
2964 NB: only the type of the blocking queue is different in GranSim and GUM
2965 the operations on the queue-elements are the same
2966 long live polymorphism!
2968 Locks: sched_mutex is held upon entry and exit.
2972 unblockThread(StgTSO *tso)
2974 StgBlockingQueueElement *t, **last;
2976 switch (tso->why_blocked) {
2979 return; /* not blocked */
2982 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
2984 StgBlockingQueueElement *last_tso = END_BQ_QUEUE;
2985 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
2987 last = (StgBlockingQueueElement **)&mvar->head;
2988 for (t = (StgBlockingQueueElement *)mvar->head;
2990 last = &t->link, last_tso = t, t = t->link) {
2991 if (t == (StgBlockingQueueElement *)tso) {
2992 *last = (StgBlockingQueueElement *)tso->link;
2993 if (mvar->tail == tso) {
2994 mvar->tail = (StgTSO *)last_tso;
2999 barf("unblockThread (MVAR): TSO not found");
3002 case BlockedOnBlackHole:
3003 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
3005 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
3007 last = &bq->blocking_queue;
3008 for (t = bq->blocking_queue;
3010 last = &t->link, t = t->link) {
3011 if (t == (StgBlockingQueueElement *)tso) {
3012 *last = (StgBlockingQueueElement *)tso->link;
3016 barf("unblockThread (BLACKHOLE): TSO not found");
3019 case BlockedOnException:
3021 StgTSO *target = tso->block_info.tso;
3023 ASSERT(get_itbl(target)->type == TSO);
3025 if (target->what_next == ThreadRelocated) {
3026 target = target->link;
3027 ASSERT(get_itbl(target)->type == TSO);
3030 ASSERT(target->blocked_exceptions != NULL);
3032 last = (StgBlockingQueueElement **)&target->blocked_exceptions;
3033 for (t = (StgBlockingQueueElement *)target->blocked_exceptions;
3035 last = &t->link, t = t->link) {
3036 ASSERT(get_itbl(t)->type == TSO);
3037 if (t == (StgBlockingQueueElement *)tso) {
3038 *last = (StgBlockingQueueElement *)tso->link;
3042 barf("unblockThread (Exception): TSO not found");
3046 case BlockedOnWrite:
3048 /* take TSO off blocked_queue */
3049 StgBlockingQueueElement *prev = NULL;
3050 for (t = (StgBlockingQueueElement *)blocked_queue_hd; t != END_BQ_QUEUE;
3051 prev = t, t = t->link) {
3052 if (t == (StgBlockingQueueElement *)tso) {
3054 blocked_queue_hd = (StgTSO *)t->link;
3055 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
3056 blocked_queue_tl = END_TSO_QUEUE;
3059 prev->link = t->link;
3060 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
3061 blocked_queue_tl = (StgTSO *)prev;
3067 barf("unblockThread (I/O): TSO not found");
3070 case BlockedOnDelay:
3072 /* take TSO off sleeping_queue */
3073 StgBlockingQueueElement *prev = NULL;
3074 for (t = (StgBlockingQueueElement *)sleeping_queue; t != END_BQ_QUEUE;
3075 prev = t, t = t->link) {
3076 if (t == (StgBlockingQueueElement *)tso) {
3078 sleeping_queue = (StgTSO *)t->link;
3080 prev->link = t->link;
3085 barf("unblockThread (I/O): TSO not found");
3089 barf("unblockThread");
3093 tso->link = END_TSO_QUEUE;
3094 tso->why_blocked = NotBlocked;
3095 tso->block_info.closure = NULL;
3096 PUSH_ON_RUN_QUEUE(tso);
3100 unblockThread(StgTSO *tso)
3104 /* To avoid locking unnecessarily. */
3105 if (tso->why_blocked == NotBlocked) {
3109 switch (tso->why_blocked) {
3112 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
3114 StgTSO *last_tso = END_TSO_QUEUE;
3115 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
3118 for (t = mvar->head; t != END_TSO_QUEUE;
3119 last = &t->link, last_tso = t, t = t->link) {
3122 if (mvar->tail == tso) {
3123 mvar->tail = last_tso;
3128 barf("unblockThread (MVAR): TSO not found");
3131 case BlockedOnBlackHole:
3132 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
3134 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
3136 last = &bq->blocking_queue;
3137 for (t = bq->blocking_queue; t != END_TSO_QUEUE;
3138 last = &t->link, t = t->link) {
3144 barf("unblockThread (BLACKHOLE): TSO not found");
3147 case BlockedOnException:
3149 StgTSO *target = tso->block_info.tso;
3151 ASSERT(get_itbl(target)->type == TSO);
3153 while (target->what_next == ThreadRelocated) {
3154 target = target->link;
3155 ASSERT(get_itbl(target)->type == TSO);
3158 ASSERT(target->blocked_exceptions != NULL);
3160 last = &target->blocked_exceptions;
3161 for (t = target->blocked_exceptions; t != END_TSO_QUEUE;
3162 last = &t->link, t = t->link) {
3163 ASSERT(get_itbl(t)->type == TSO);
3169 barf("unblockThread (Exception): TSO not found");
3173 case BlockedOnWrite:
3175 StgTSO *prev = NULL;
3176 for (t = blocked_queue_hd; t != END_TSO_QUEUE;
3177 prev = t, t = t->link) {
3180 blocked_queue_hd = t->link;
3181 if (blocked_queue_tl == t) {
3182 blocked_queue_tl = END_TSO_QUEUE;
3185 prev->link = t->link;
3186 if (blocked_queue_tl == t) {
3187 blocked_queue_tl = prev;
3193 barf("unblockThread (I/O): TSO not found");
3196 case BlockedOnDelay:
3198 StgTSO *prev = NULL;
3199 for (t = sleeping_queue; t != END_TSO_QUEUE;
3200 prev = t, t = t->link) {
3203 sleeping_queue = t->link;
3205 prev->link = t->link;
3210 barf("unblockThread (I/O): TSO not found");
3214 barf("unblockThread");
3218 tso->link = END_TSO_QUEUE;
3219 tso->why_blocked = NotBlocked;
3220 tso->block_info.closure = NULL;
3221 PUSH_ON_RUN_QUEUE(tso);
3225 /* -----------------------------------------------------------------------------
3228 * The following function implements the magic for raising an
3229 * asynchronous exception in an existing thread.
3231 * We first remove the thread from any queue on which it might be
3232 * blocked. The possible blockages are MVARs and BLACKHOLE_BQs.
3234 * We strip the stack down to the innermost CATCH_FRAME, building
3235 * thunks in the heap for all the active computations, so they can
3236 * be restarted if necessary. When we reach a CATCH_FRAME, we build
3237 * an application of the handler to the exception, and push it on
3238 * the top of the stack.
3240 * How exactly do we save all the active computations? We create an
3241 * AP_STACK for every UpdateFrame on the stack. Entering one of these
3242 * AP_STACKs pushes everything from the corresponding update frame
3243 * upwards onto the stack. (Actually, it pushes everything up to the
3244 * next update frame plus a pointer to the next AP_STACK object.
3245 * Entering the next AP_STACK object pushes more onto the stack until we
3246 * reach the last AP_STACK object - at which point the stack should look
3247 * exactly as it did when we killed the TSO and we can continue
3248 * execution by entering the closure on top of the stack.
3250 * We can also kill a thread entirely - this happens if either (a) the
3251 * exception passed to raiseAsync is NULL, or (b) there's no
3252 * CATCH_FRAME on the stack. In either case, we strip the entire
3253 * stack and replace the thread with a zombie.
3255 * Locks: sched_mutex held upon entry nor exit.
3257 * -------------------------------------------------------------------------- */
3260 deleteThread(StgTSO *tso)
3262 raiseAsync(tso,NULL);
3266 raiseAsyncWithLock(StgTSO *tso, StgClosure *exception)
3268 /* When raising async exs from contexts where sched_mutex isn't held;
3269 use raiseAsyncWithLock(). */
3270 ACQUIRE_LOCK(&sched_mutex);
3271 raiseAsync(tso,exception);
3272 RELEASE_LOCK(&sched_mutex);
3276 raiseAsync(StgTSO *tso, StgClosure *exception)
3278 StgRetInfoTable *info;
3281 // Thread already dead?
3282 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
3287 sched_belch("raising exception in thread %ld.", tso->id));
3289 // Remove it from any blocking queues
3294 // The stack freezing code assumes there's a closure pointer on
3295 // the top of the stack, so we have to arrange that this is the case...
3297 if (sp[0] == (W_)&stg_enter_info) {
3301 sp[0] = (W_)&stg_dummy_ret_closure;
3307 // 1. Let the top of the stack be the "current closure"
3309 // 2. Walk up the stack until we find either an UPDATE_FRAME or a
3312 // 3. If it's an UPDATE_FRAME, then make an AP_STACK containing the
3313 // current closure applied to the chunk of stack up to (but not
3314 // including) the update frame. This closure becomes the "current
3315 // closure". Go back to step 2.
3317 // 4. If it's a CATCH_FRAME, then leave the exception handler on
3318 // top of the stack applied to the exception.
3320 // 5. If it's a STOP_FRAME, then kill the thread.
3325 info = get_ret_itbl((StgClosure *)frame);
3327 while (info->i.type != UPDATE_FRAME
3328 && (info->i.type != CATCH_FRAME || exception == NULL)
3329 && info->i.type != STOP_FRAME) {
3330 frame += stack_frame_sizeW((StgClosure *)frame);
3331 info = get_ret_itbl((StgClosure *)frame);
3334 switch (info->i.type) {
3337 // If we find a CATCH_FRAME, and we've got an exception to raise,
3338 // then build the THUNK raise(exception), and leave it on
3339 // top of the CATCH_FRAME ready to enter.
3343 StgCatchFrame *cf = (StgCatchFrame *)frame;
3347 // we've got an exception to raise, so let's pass it to the
3348 // handler in this frame.
3350 raise = (StgClosure *)allocate(sizeofW(StgClosure)+1);
3351 TICK_ALLOC_SE_THK(1,0);
3352 SET_HDR(raise,&stg_raise_info,cf->header.prof.ccs);
3353 raise->payload[0] = exception;
3355 // throw away the stack from Sp up to the CATCH_FRAME.
3359 /* Ensure that async excpetions are blocked now, so we don't get
3360 * a surprise exception before we get around to executing the
3363 if (tso->blocked_exceptions == NULL) {
3364 tso->blocked_exceptions = END_TSO_QUEUE;
3367 /* Put the newly-built THUNK on top of the stack, ready to execute
3368 * when the thread restarts.
3371 sp[-1] = (W_)&stg_enter_info;
3373 tso->what_next = ThreadRunGHC;
3374 IF_DEBUG(sanity, checkTSO(tso));
3383 // First build an AP_STACK consisting of the stack chunk above the
3384 // current update frame, with the top word on the stack as the
3387 words = frame - sp - 1;
3388 ap = (StgAP_STACK *)allocate(PAP_sizeW(words));
3391 ap->fun = (StgClosure *)sp[0];
3393 for(i=0; i < (nat)words; ++i) {
3394 ap->payload[i] = (StgClosure *)*sp++;
3397 SET_HDR(ap,&stg_AP_STACK_info,
3398 ((StgClosure *)frame)->header.prof.ccs /* ToDo */);
3399 TICK_ALLOC_UP_THK(words+1,0);
3402 fprintf(stderr, "scheduler: Updating ");
3403 printPtr((P_)((StgUpdateFrame *)frame)->updatee);
3404 fprintf(stderr, " with ");
3405 printObj((StgClosure *)ap);
3408 // Replace the updatee with an indirection - happily
3409 // this will also wake up any threads currently
3410 // waiting on the result.
3412 // Warning: if we're in a loop, more than one update frame on
3413 // the stack may point to the same object. Be careful not to
3414 // overwrite an IND_OLDGEN in this case, because we'll screw
3415 // up the mutable lists. To be on the safe side, don't
3416 // overwrite any kind of indirection at all. See also
3417 // threadSqueezeStack in GC.c, where we have to make a similar
3420 if (!closure_IND(((StgUpdateFrame *)frame)->updatee)) {
3421 // revert the black hole
3422 UPD_IND_NOLOCK(((StgUpdateFrame *)frame)->updatee,ap);
3424 sp += sizeofW(StgUpdateFrame) - 1;
3425 sp[0] = (W_)ap; // push onto stack
3430 // We've stripped the entire stack, the thread is now dead.
3431 sp += sizeofW(StgStopFrame);
3432 tso->what_next = ThreadKilled;
3443 /* -----------------------------------------------------------------------------
3444 resurrectThreads is called after garbage collection on the list of
3445 threads found to be garbage. Each of these threads will be woken
3446 up and sent a signal: BlockedOnDeadMVar if the thread was blocked
3447 on an MVar, or NonTermination if the thread was blocked on a Black
3450 Locks: sched_mutex isn't held upon entry nor exit.
3451 -------------------------------------------------------------------------- */
3454 resurrectThreads( StgTSO *threads )
3458 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
3459 next = tso->global_link;
3460 tso->global_link = all_threads;
3462 IF_DEBUG(scheduler, sched_belch("resurrecting thread %d", tso->id));
3464 switch (tso->why_blocked) {
3466 case BlockedOnException:
3467 /* Called by GC - sched_mutex lock is currently held. */
3468 raiseAsync(tso,(StgClosure *)BlockedOnDeadMVar_closure);
3470 case BlockedOnBlackHole:
3471 raiseAsync(tso,(StgClosure *)NonTermination_closure);
3474 /* This might happen if the thread was blocked on a black hole
3475 * belonging to a thread that we've just woken up (raiseAsync
3476 * can wake up threads, remember...).
3480 barf("resurrectThreads: thread blocked in a strange way");
3485 /* -----------------------------------------------------------------------------
3486 * Blackhole detection: if we reach a deadlock, test whether any
3487 * threads are blocked on themselves. Any threads which are found to
3488 * be self-blocked get sent a NonTermination exception.
3490 * This is only done in a deadlock situation in order to avoid
3491 * performance overhead in the normal case.
3493 * Locks: sched_mutex is held upon entry and exit.
3494 * -------------------------------------------------------------------------- */
3497 detectBlackHoles( void )
3499 StgTSO *tso = all_threads;
3501 StgClosure *blocked_on;
3502 StgRetInfoTable *info;
3504 for (tso = all_threads; tso != END_TSO_QUEUE; tso = tso->global_link) {
3506 while (tso->what_next == ThreadRelocated) {
3508 ASSERT(get_itbl(tso)->type == TSO);
3511 if (tso->why_blocked != BlockedOnBlackHole) {
3515 blocked_on = tso->block_info.closure;
3517 frame = (StgClosure *)tso->sp;
3520 info = get_ret_itbl(frame);
3521 switch (info->i.type) {
3524 if (((StgUpdateFrame *)frame)->updatee == blocked_on) {
3525 /* We are blocking on one of our own computations, so
3526 * send this thread the NonTermination exception.
3529 sched_belch("thread %d is blocked on itself", tso->id));
3530 raiseAsync(tso, (StgClosure *)NonTermination_closure);
3534 frame = (StgClosure *) ((StgUpdateFrame *)frame + 1);
3540 // normal stack frames; do nothing except advance the pointer
3542 (StgPtr)frame += stack_frame_sizeW(frame);
3549 //@node Debugging Routines, Index, Exception Handling Routines, Main scheduling code
3550 //@subsection Debugging Routines
3552 /* -----------------------------------------------------------------------------
3553 * Debugging: why is a thread blocked
3554 * [Also provides useful information when debugging threaded programs
3555 * at the Haskell source code level, so enable outside of DEBUG. --sof 7/02]
3556 -------------------------------------------------------------------------- */
3560 printThreadBlockage(StgTSO *tso)
3562 switch (tso->why_blocked) {
3564 fprintf(stderr,"is blocked on read from fd %d", tso->block_info.fd);
3566 case BlockedOnWrite:
3567 fprintf(stderr,"is blocked on write to fd %d", tso->block_info.fd);
3569 case BlockedOnDelay:
3570 fprintf(stderr,"is blocked until %d", tso->block_info.target);
3573 fprintf(stderr,"is blocked on an MVar");
3575 case BlockedOnException:
3576 fprintf(stderr,"is blocked on delivering an exception to thread %d",
3577 tso->block_info.tso->id);
3579 case BlockedOnBlackHole:
3580 fprintf(stderr,"is blocked on a black hole");
3583 fprintf(stderr,"is not blocked");
3587 fprintf(stderr,"is blocked on global address; local FM_BQ is %p (%s)",
3588 tso->block_info.closure, info_type(tso->block_info.closure));
3590 case BlockedOnGA_NoSend:
3591 fprintf(stderr,"is blocked on global address (no send); local FM_BQ is %p (%s)",
3592 tso->block_info.closure, info_type(tso->block_info.closure));
3595 #if defined(RTS_SUPPORTS_THREADS)
3596 case BlockedOnCCall:
3597 fprintf(stderr,"is blocked on an external call");
3601 barf("printThreadBlockage: strange tso->why_blocked: %d for TSO %d (%d)",
3602 tso->why_blocked, tso->id, tso);
3608 printThreadStatus(StgTSO *tso)
3610 switch (tso->what_next) {
3612 fprintf(stderr,"has been killed");
3614 case ThreadComplete:
3615 fprintf(stderr,"has completed");
3618 printThreadBlockage(tso);
3623 printAllThreads(void)
3629 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
3630 ullong_format_string(TIME_ON_PROC(CurrentProc),
3631 time_string, rtsFalse/*no commas!*/);
3633 fprintf(stderr, "all threads at [%s]:\n", time_string);
3635 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
3636 ullong_format_string(CURRENT_TIME,
3637 time_string, rtsFalse/*no commas!*/);
3639 fprintf(stderr,"all threads at [%s]:\n", time_string);
3641 fprintf(stderr,"all threads:\n");
3644 for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
3645 fprintf(stderr, "\tthread %d @ %p ", t->id, (void *)t);
3646 label = lookupThreadLabel((StgWord)t);
3647 if (label) fprintf(stderr,"[\"%s\"] ",(char *)label);
3648 printThreadStatus(t);
3649 fprintf(stderr,"\n");
3656 Print a whole blocking queue attached to node (debugging only).
3661 print_bq (StgClosure *node)
3663 StgBlockingQueueElement *bqe;
3667 fprintf(stderr,"## BQ of closure %p (%s): ",
3668 node, info_type(node));
3670 /* should cover all closures that may have a blocking queue */
3671 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
3672 get_itbl(node)->type == FETCH_ME_BQ ||
3673 get_itbl(node)->type == RBH ||
3674 get_itbl(node)->type == MVAR);
3676 ASSERT(node!=(StgClosure*)NULL); // sanity check
3678 print_bqe(((StgBlockingQueue*)node)->blocking_queue);
3682 Print a whole blocking queue starting with the element bqe.
3685 print_bqe (StgBlockingQueueElement *bqe)
3690 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
3692 for (end = (bqe==END_BQ_QUEUE);
3693 !end; // iterate until bqe points to a CONSTR
3694 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE),
3695 bqe = end ? END_BQ_QUEUE : bqe->link) {
3696 ASSERT(bqe != END_BQ_QUEUE); // sanity check
3697 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
3698 /* types of closures that may appear in a blocking queue */
3699 ASSERT(get_itbl(bqe)->type == TSO ||
3700 get_itbl(bqe)->type == BLOCKED_FETCH ||
3701 get_itbl(bqe)->type == CONSTR);
3702 /* only BQs of an RBH end with an RBH_Save closure */
3703 //ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
3705 switch (get_itbl(bqe)->type) {
3707 fprintf(stderr," TSO %u (%x),",
3708 ((StgTSO *)bqe)->id, ((StgTSO *)bqe));
3711 fprintf(stderr," BF (node=%p, ga=((%x, %d, %x)),",
3712 ((StgBlockedFetch *)bqe)->node,
3713 ((StgBlockedFetch *)bqe)->ga.payload.gc.gtid,
3714 ((StgBlockedFetch *)bqe)->ga.payload.gc.slot,
3715 ((StgBlockedFetch *)bqe)->ga.weight);
3718 fprintf(stderr," %s (IP %p),",
3719 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
3720 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
3721 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
3722 "RBH_Save_?"), get_itbl(bqe));
3725 barf("Unexpected closure type %s in blocking queue", // of %p (%s)",
3726 info_type((StgClosure *)bqe)); // , node, info_type(node));
3730 fputc('\n', stderr);
3732 # elif defined(GRAN)
3734 print_bq (StgClosure *node)
3736 StgBlockingQueueElement *bqe;
3737 PEs node_loc, tso_loc;
3740 /* should cover all closures that may have a blocking queue */
3741 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
3742 get_itbl(node)->type == FETCH_ME_BQ ||
3743 get_itbl(node)->type == RBH);
3745 ASSERT(node!=(StgClosure*)NULL); // sanity check
3746 node_loc = where_is(node);
3748 fprintf(stderr,"## BQ of closure %p (%s) on [PE %d]: ",
3749 node, info_type(node), node_loc);
3752 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
3754 for (bqe = ((StgBlockingQueue*)node)->blocking_queue, end = (bqe==END_BQ_QUEUE);
3755 !end; // iterate until bqe points to a CONSTR
3756 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) {
3757 ASSERT(bqe != END_BQ_QUEUE); // sanity check
3758 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
3759 /* types of closures that may appear in a blocking queue */
3760 ASSERT(get_itbl(bqe)->type == TSO ||
3761 get_itbl(bqe)->type == CONSTR);
3762 /* only BQs of an RBH end with an RBH_Save closure */
3763 ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
3765 tso_loc = where_is((StgClosure *)bqe);
3766 switch (get_itbl(bqe)->type) {
3768 fprintf(stderr," TSO %d (%p) on [PE %d],",
3769 ((StgTSO *)bqe)->id, (StgTSO *)bqe, tso_loc);
3772 fprintf(stderr," %s (IP %p),",
3773 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
3774 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
3775 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
3776 "RBH_Save_?"), get_itbl(bqe));
3779 barf("Unexpected closure type %s in blocking queue of %p (%s)",
3780 info_type((StgClosure *)bqe), node, info_type(node));
3784 fputc('\n', stderr);
3788 Nice and easy: only TSOs on the blocking queue
3791 print_bq (StgClosure *node)
3795 ASSERT(node!=(StgClosure*)NULL); // sanity check
3796 for (tso = ((StgBlockingQueue*)node)->blocking_queue;
3797 tso != END_TSO_QUEUE;
3799 ASSERT(tso!=NULL && tso!=END_TSO_QUEUE); // sanity check
3800 ASSERT(get_itbl(tso)->type == TSO); // guess what, sanity check
3801 fprintf(stderr," TSO %d (%p),", tso->id, tso);
3803 fputc('\n', stderr);
3814 for (i=0, tso=run_queue_hd;
3815 tso != END_TSO_QUEUE;
3824 sched_belch(char *s, ...)
3829 fprintf(stderr, "scheduler (task %ld): ", osThreadId());
3831 fprintf(stderr, "== ");
3833 fprintf(stderr, "scheduler: ");
3835 vfprintf(stderr, s, ap);
3836 fprintf(stderr, "\n");
3843 //@node Index, , Debugging Routines, Main scheduling code
3847 //* StgMainThread:: @cindex\s-+StgMainThread
3848 //* awaken_blocked_queue:: @cindex\s-+awaken_blocked_queue
3849 //* blocked_queue_hd:: @cindex\s-+blocked_queue_hd
3850 //* blocked_queue_tl:: @cindex\s-+blocked_queue_tl
3851 //* context_switch:: @cindex\s-+context_switch
3852 //* createThread:: @cindex\s-+createThread
3853 //* gc_pending_cond:: @cindex\s-+gc_pending_cond
3854 //* initScheduler:: @cindex\s-+initScheduler
3855 //* interrupted:: @cindex\s-+interrupted
3856 //* next_thread_id:: @cindex\s-+next_thread_id
3857 //* print_bq:: @cindex\s-+print_bq
3858 //* run_queue_hd:: @cindex\s-+run_queue_hd
3859 //* run_queue_tl:: @cindex\s-+run_queue_tl
3860 //* sched_mutex:: @cindex\s-+sched_mutex
3861 //* schedule:: @cindex\s-+schedule
3862 //* take_off_run_queue:: @cindex\s-+take_off_run_queue
3863 //* term_mutex:: @cindex\s-+term_mutex