1 /* ---------------------------------------------------------------------------
2 * $Id: Schedule.c,v 1.168 2003/04/08 15:53:51 sof Exp $
4 * (c) The GHC Team, 1998-2000
8 * Different GHC ways use this scheduler quite differently (see comments below)
9 * Here is the global picture:
11 * WAY Name CPP flag What's it for
12 * --------------------------------------
13 * mp GUM PAR Parallel execution on a distributed memory machine
14 * s SMP SMP Parallel execution on a shared memory machine
15 * mg GranSim GRAN Simulation of parallel execution
16 * md GUM/GdH DIST Distributed execution (based on GUM)
18 * --------------------------------------------------------------------------*/
20 //@node Main scheduling code, , ,
21 //@section Main scheduling code
24 * Version with scheduler monitor support for SMPs (WAY=s):
26 This design provides a high-level API to create and schedule threads etc.
27 as documented in the SMP design document.
29 It uses a monitor design controlled by a single mutex to exercise control
30 over accesses to shared data structures, and builds on the Posix threads
33 The majority of state is shared. In order to keep essential per-task state,
34 there is a Capability structure, which contains all the information
35 needed to run a thread: its STG registers, a pointer to its TSO, a
36 nursery etc. During STG execution, a pointer to the capability is
37 kept in a register (BaseReg).
39 In a non-SMP build, there is one global capability, namely MainRegTable.
43 * Version with support for distributed memory parallelism aka GUM (WAY=mp):
45 The main scheduling loop in GUM iterates until a finish message is received.
46 In that case a global flag @receivedFinish@ is set and this instance of
47 the RTS shuts down. See ghc/rts/parallel/HLComms.c:processMessages()
48 for the handling of incoming messages, such as PP_FINISH.
49 Note that in the parallel case we have a system manager that coordinates
50 different PEs, each of which are running one instance of the RTS.
51 See ghc/rts/parallel/SysMan.c for the main routine of the parallel program.
52 From this routine processes executing ghc/rts/Main.c are spawned. -- HWL
54 * Version with support for simulating parallel execution aka GranSim (WAY=mg):
56 The main scheduling code in GranSim is quite different from that in std
57 (concurrent) Haskell: while concurrent Haskell just iterates over the
58 threads in the runnable queue, GranSim is event driven, i.e. it iterates
59 over the events in the global event queue. -- HWL
64 //* Variables and Data structures::
65 //* Main scheduling loop::
66 //* Suspend and Resume::
68 //* Garbage Collextion Routines::
69 //* Blocking Queue Routines::
70 //* Exception Handling Routines::
71 //* Debugging Routines::
75 //@node Includes, Variables and Data structures, Main scheduling code, Main scheduling code
76 //@subsection Includes
78 #include "PosixSource.h"
85 #include "StgStartup.h"
87 #define COMPILING_SCHEDULER
89 #include "StgMiscClosures.h"
91 #include "Interpreter.h"
92 #include "Exception.h"
99 #include "ThreadLabels.h"
101 #include "Proftimer.h"
102 #include "ProfHeap.h"
104 #if defined(GRAN) || defined(PAR)
105 # include "GranSimRts.h"
106 # include "GranSim.h"
107 # include "ParallelRts.h"
108 # include "Parallel.h"
109 # include "ParallelDebug.h"
110 # include "FetchMe.h"
114 #include "Capability.h"
115 #include "OSThreads.h"
118 #ifdef HAVE_SYS_TYPES_H
119 #include <sys/types.h>
129 //@node Variables and Data structures, Prototypes, Includes, Main scheduling code
130 //@subsection Variables and Data structures
132 /* Main thread queue.
133 * Locks required: sched_mutex.
135 StgMainThread *main_threads = NULL;
138 // Pointer to the thread that executes main
139 // When this thread is finished, the program terminates
140 // by calling shutdownHaskellAndExit.
141 // It would be better to add a call to shutdownHaskellAndExit
142 // to the Main.main wrapper and to remove this hack.
143 StgMainThread *main_main_thread = NULL;
147 * Locks required: sched_mutex.
151 StgTSO* ActiveTSO = NULL; /* for assigning system costs; GranSim-Light only */
152 /* rtsTime TimeOfNextEvent, EndOfTimeSlice; now in GranSim.c */
155 In GranSim we have a runnable and a blocked queue for each processor.
156 In order to minimise code changes new arrays run_queue_hds/tls
157 are created. run_queue_hd is then a short cut (macro) for
158 run_queue_hds[CurrentProc] (see GranSim.h).
161 StgTSO *run_queue_hds[MAX_PROC], *run_queue_tls[MAX_PROC];
162 StgTSO *blocked_queue_hds[MAX_PROC], *blocked_queue_tls[MAX_PROC];
163 StgTSO *ccalling_threadss[MAX_PROC];
164 /* We use the same global list of threads (all_threads) in GranSim as in
165 the std RTS (i.e. we are cheating). However, we don't use this list in
166 the GranSim specific code at the moment (so we are only potentially
171 StgTSO *run_queue_hd = NULL;
172 StgTSO *run_queue_tl = NULL;
173 StgTSO *blocked_queue_hd = NULL;
174 StgTSO *blocked_queue_tl = NULL;
175 StgTSO *sleeping_queue = NULL; /* perhaps replace with a hash table? */
179 /* Linked list of all threads.
180 * Used for detecting garbage collected threads.
182 StgTSO *all_threads = NULL;
184 /* When a thread performs a safe C call (_ccall_GC, using old
185 * terminology), it gets put on the suspended_ccalling_threads
186 * list. Used by the garbage collector.
188 static StgTSO *suspended_ccalling_threads;
190 static StgTSO *threadStackOverflow(StgTSO *tso);
192 /* KH: The following two flags are shared memory locations. There is no need
193 to lock them, since they are only unset at the end of a scheduler
197 /* flag set by signal handler to precipitate a context switch */
198 //@cindex context_switch
199 nat context_switch = 0;
201 /* if this flag is set as well, give up execution */
202 //@cindex interrupted
203 rtsBool interrupted = rtsFalse;
205 /* Next thread ID to allocate.
206 * Locks required: thread_id_mutex
208 //@cindex next_thread_id
209 static StgThreadID next_thread_id = 1;
212 * Pointers to the state of the current thread.
213 * Rule of thumb: if CurrentTSO != NULL, then we're running a Haskell
214 * thread. If CurrentTSO == NULL, then we're at the scheduler level.
217 /* The smallest stack size that makes any sense is:
218 * RESERVED_STACK_WORDS (so we can get back from the stack overflow)
219 * + sizeofW(StgStopFrame) (the stg_stop_thread_info frame)
220 * + 1 (the realworld token for an IO thread)
221 * + 1 (the closure to enter)
223 * A thread with this stack will bomb immediately with a stack
224 * overflow, which will increase its stack size.
227 #define MIN_STACK_WORDS (RESERVED_STACK_WORDS + sizeofW(StgStopFrame) + 2)
234 /* This is used in `TSO.h' and gcc 2.96 insists that this variable actually
235 * exists - earlier gccs apparently didn't.
240 static rtsBool ready_to_gc;
243 * Set to TRUE when entering a shutdown state (via shutdownHaskellAndExit()) --
244 * in an MT setting, needed to signal that a worker thread shouldn't hang around
245 * in the scheduler when it is out of work.
247 static rtsBool shutting_down_scheduler = rtsFalse;
249 void addToBlockedQueue ( StgTSO *tso );
251 static void schedule ( void );
252 void interruptStgRts ( void );
254 static void detectBlackHoles ( void );
257 static void sched_belch(char *s, ...);
260 #if defined(RTS_SUPPORTS_THREADS)
261 /* ToDo: carefully document the invariants that go together
262 * with these synchronisation objects.
264 Mutex sched_mutex = INIT_MUTEX_VAR;
265 Mutex term_mutex = INIT_MUTEX_VAR;
268 * A heavyweight solution to the problem of protecting
269 * the thread_id from concurrent update.
271 Mutex thread_id_mutex = INIT_MUTEX_VAR;
275 static Condition gc_pending_cond = INIT_COND_VAR;
279 #endif /* RTS_SUPPORTS_THREADS */
283 rtsTime TimeOfLastYield;
284 rtsBool emitSchedule = rtsTrue;
288 static char *whatNext_strs[] = {
298 StgTSO * createSparkThread(rtsSpark spark);
299 StgTSO * activateSpark (rtsSpark spark);
303 * The thread state for the main thread.
304 // ToDo: check whether not needed any more
308 #if defined(PAR) || defined(RTS_SUPPORTS_THREADS)
309 static void taskStart(void);
317 #if defined(RTS_SUPPORTS_THREADS)
319 startSchedulerTask(void)
321 startTask(taskStart);
325 //@node Main scheduling loop, Suspend and Resume, Prototypes, Main scheduling code
326 //@subsection Main scheduling loop
328 /* ---------------------------------------------------------------------------
329 Main scheduling loop.
331 We use round-robin scheduling, each thread returning to the
332 scheduler loop when one of these conditions is detected:
335 * timer expires (thread yields)
340 Locking notes: we acquire the scheduler lock once at the beginning
341 of the scheduler loop, and release it when
343 * running a thread, or
344 * waiting for work, or
345 * waiting for a GC to complete.
348 In a GranSim setup this loop iterates over the global event queue.
349 This revolves around the global event queue, which determines what
350 to do next. Therefore, it's more complicated than either the
351 concurrent or the parallel (GUM) setup.
354 GUM iterates over incoming messages.
355 It starts with nothing to do (thus CurrentTSO == END_TSO_QUEUE),
356 and sends out a fish whenever it has nothing to do; in-between
357 doing the actual reductions (shared code below) it processes the
358 incoming messages and deals with delayed operations
359 (see PendingFetches).
360 This is not the ugliest code you could imagine, but it's bloody close.
362 ------------------------------------------------------------------------ */
369 StgThreadReturnCode ret;
377 rtsBool receivedFinish = rtsFalse;
379 nat tp_size, sp_size; // stats only
382 rtsBool was_interrupted = rtsFalse;
383 StgTSOWhatNext prev_what_next;
385 ACQUIRE_LOCK(&sched_mutex);
387 #if defined(RTS_SUPPORTS_THREADS)
388 waitForWorkCapability(&sched_mutex, &cap, rtsFalse);
389 IF_DEBUG(scheduler, sched_belch("worker thread (osthread %p): entering RTS", osThreadId()));
391 /* simply initialise it in the non-threaded case */
392 grabCapability(&cap);
396 /* set up first event to get things going */
397 /* ToDo: assign costs for system setup and init MainTSO ! */
398 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
400 CurrentTSO, (StgClosure*)NULL, (rtsSpark*)NULL);
403 fprintf(stderr, "GRAN: Init CurrentTSO (in schedule) = %p\n", CurrentTSO);
404 G_TSO(CurrentTSO, 5));
406 if (RtsFlags.GranFlags.Light) {
407 /* Save current time; GranSim Light only */
408 CurrentTSO->gran.clock = CurrentTime[CurrentProc];
411 event = get_next_event();
413 while (event!=(rtsEvent*)NULL) {
414 /* Choose the processor with the next event */
415 CurrentProc = event->proc;
416 CurrentTSO = event->tso;
420 while (!receivedFinish) { /* set by processMessages */
421 /* when receiving PP_FINISH message */
428 IF_DEBUG(scheduler, printAllThreads());
430 #if defined(RTS_SUPPORTS_THREADS)
431 /* Check to see whether there are any worker threads
432 waiting to deposit external call results. If so,
433 yield our capability */
434 yieldToReturningWorker(&sched_mutex, &cap);
437 /* If we're interrupted (the user pressed ^C, or some other
438 * termination condition occurred), kill all the currently running
442 IF_DEBUG(scheduler, sched_belch("interrupted"));
443 interrupted = rtsFalse;
444 was_interrupted = rtsTrue;
445 #if defined(RTS_SUPPORTS_THREADS)
446 // In the threaded RTS, deadlock detection doesn't work,
447 // so just exit right away.
448 prog_belch("interrupted");
449 releaseCapability(cap);
450 startTask(taskStart); // thread-safe-call to shutdownHaskellAndExit
451 RELEASE_LOCK(&sched_mutex);
452 shutdownHaskellAndExit(EXIT_SUCCESS);
458 /* Go through the list of main threads and wake up any
459 * clients whose computations have finished. ToDo: this
460 * should be done more efficiently without a linear scan
461 * of the main threads list, somehow...
463 #if defined(RTS_SUPPORTS_THREADS)
465 StgMainThread *m, **prev;
466 prev = &main_threads;
467 for (m = main_threads; m != NULL; prev = &m->link, m = m->link) {
468 switch (m->tso->what_next) {
471 // NOTE: return val is tso->sp[1] (see StgStartup.hc)
472 *(m->ret) = (StgClosure *)m->tso->sp[1];
476 broadcastCondition(&m->wakeup);
478 removeThreadLabel((StgWord)m->tso);
480 if(m == main_main_thread)
482 releaseCapability(cap);
483 startTask(taskStart); // thread-safe-call to shutdownHaskellAndExit
484 RELEASE_LOCK(&sched_mutex);
485 shutdownHaskellAndExit(EXIT_SUCCESS);
489 if (m->ret) *(m->ret) = NULL;
491 if (was_interrupted) {
492 m->stat = Interrupted;
496 broadcastCondition(&m->wakeup);
498 removeThreadLabel((StgWord)m->tso);
500 if(m == main_main_thread)
502 releaseCapability(cap);
503 startTask(taskStart); // thread-safe-call to shutdownHaskellAndExit
504 RELEASE_LOCK(&sched_mutex);
505 shutdownHaskellAndExit(EXIT_SUCCESS);
514 #else /* not threaded */
517 /* in GUM do this only on the Main PE */
520 /* If our main thread has finished or been killed, return.
523 StgMainThread *m = main_threads;
524 if (m->tso->what_next == ThreadComplete
525 || m->tso->what_next == ThreadKilled) {
527 removeThreadLabel((StgWord)m->tso);
529 main_threads = main_threads->link;
530 if (m->tso->what_next == ThreadComplete) {
531 // We finished successfully, fill in the return value
532 // NOTE: return val is tso->sp[1] (see StgStartup.hc)
533 if (m->ret) { *(m->ret) = (StgClosure *)m->tso->sp[1]; };
537 if (m->ret) { *(m->ret) = NULL; };
538 if (was_interrupted) {
539 m->stat = Interrupted;
549 /* Top up the run queue from our spark pool. We try to make the
550 * number of threads in the run queue equal to the number of
553 * Disable spark support in SMP for now, non-essential & requires
554 * a little bit of work to make it compile cleanly. -- sof 1/02.
556 #if 0 /* defined(SMP) */
558 nat n = getFreeCapabilities();
559 StgTSO *tso = run_queue_hd;
561 /* Count the run queue */
562 while (n > 0 && tso != END_TSO_QUEUE) {
569 spark = findSpark(rtsFalse);
571 break; /* no more sparks in the pool */
573 /* I'd prefer this to be done in activateSpark -- HWL */
574 /* tricky - it needs to hold the scheduler lock and
575 * not try to re-acquire it -- SDM */
576 createSparkThread(spark);
578 sched_belch("==^^ turning spark of closure %p into a thread",
579 (StgClosure *)spark));
582 /* We need to wake up the other tasks if we just created some
585 if (getFreeCapabilities() - n > 1) {
586 signalCondition( &thread_ready_cond );
591 /* check for signals each time around the scheduler */
592 #if defined(RTS_USER_SIGNALS)
593 if (signals_pending()) {
594 RELEASE_LOCK(&sched_mutex); /* ToDo: kill */
595 startSignalHandlers();
596 ACQUIRE_LOCK(&sched_mutex);
600 /* Check whether any waiting threads need to be woken up. If the
601 * run queue is empty, and there are no other tasks running, we
602 * can wait indefinitely for something to happen.
604 if ( !EMPTY_QUEUE(blocked_queue_hd) || !EMPTY_QUEUE(sleeping_queue)
605 #if defined(RTS_SUPPORTS_THREADS) && !defined(SMP)
610 awaitEvent( EMPTY_RUN_QUEUE()
612 && allFreeCapabilities()
616 /* we can be interrupted while waiting for I/O... */
617 if (interrupted) continue;
620 * Detect deadlock: when we have no threads to run, there are no
621 * threads waiting on I/O or sleeping, and all the other tasks are
622 * waiting for work, we must have a deadlock of some description.
624 * We first try to find threads blocked on themselves (ie. black
625 * holes), and generate NonTermination exceptions where necessary.
627 * If no threads are black holed, we have a deadlock situation, so
628 * inform all the main threads.
630 #if !defined(PAR) && !defined(RTS_SUPPORTS_THREADS)
631 if ( EMPTY_THREAD_QUEUES()
632 #if defined(RTS_SUPPORTS_THREADS)
633 && EMPTY_QUEUE(suspended_ccalling_threads)
636 && allFreeCapabilities()
640 IF_DEBUG(scheduler, sched_belch("deadlocked, forcing major GC..."));
641 #if defined(THREADED_RTS)
642 /* and SMP mode ..? */
643 releaseCapability(cap);
645 // Garbage collection can release some new threads due to
646 // either (a) finalizers or (b) threads resurrected because
647 // they are about to be send BlockedOnDeadMVar. Any threads
648 // thus released will be immediately runnable.
649 GarbageCollect(GetRoots,rtsTrue);
651 if ( !EMPTY_RUN_QUEUE() ) { goto not_deadlocked; }
654 sched_belch("still deadlocked, checking for black holes..."));
657 if ( !EMPTY_RUN_QUEUE() ) { goto not_deadlocked; }
659 #if defined(RTS_USER_SIGNALS)
660 /* If we have user-installed signal handlers, then wait
661 * for signals to arrive rather then bombing out with a
664 #if defined(RTS_SUPPORTS_THREADS)
665 if ( 0 ) { /* hmm..what to do? Simply stop waiting for
666 a signal with no runnable threads (or I/O
667 suspended ones) leads nowhere quick.
668 For now, simply shut down when we reach this
671 ToDo: define precisely under what conditions
672 the Scheduler should shut down in an MT setting.
675 if ( anyUserHandlers() ) {
678 sched_belch("still deadlocked, waiting for signals..."));
682 // we might be interrupted...
683 if (interrupted) { continue; }
685 if (signals_pending()) {
686 RELEASE_LOCK(&sched_mutex);
687 startSignalHandlers();
688 ACQUIRE_LOCK(&sched_mutex);
690 ASSERT(!EMPTY_RUN_QUEUE());
695 /* Probably a real deadlock. Send the current main thread the
696 * Deadlock exception (or in the SMP build, send *all* main
697 * threads the deadlock exception, since none of them can make
702 #if defined(RTS_SUPPORTS_THREADS)
703 for (m = main_threads; m != NULL; m = m->link) {
704 switch (m->tso->why_blocked) {
705 case BlockedOnBlackHole:
706 raiseAsync(m->tso, (StgClosure *)NonTermination_closure);
708 case BlockedOnException:
710 raiseAsync(m->tso, (StgClosure *)Deadlock_closure);
713 barf("deadlock: main thread blocked in a strange way");
718 switch (m->tso->why_blocked) {
719 case BlockedOnBlackHole:
720 raiseAsync(m->tso, (StgClosure *)NonTermination_closure);
722 case BlockedOnException:
724 raiseAsync(m->tso, (StgClosure *)Deadlock_closure);
727 barf("deadlock: main thread blocked in a strange way");
732 #if defined(RTS_SUPPORTS_THREADS)
733 /* ToDo: revisit conditions (and mechanism) for shutting
734 down a multi-threaded world */
735 IF_DEBUG(scheduler, sched_belch("all done, i think...shutting down."));
736 RELEASE_LOCK(&sched_mutex);
743 #elif defined(RTS_SUPPORTS_THREADS)
744 /* ToDo: add deadlock detection in threaded RTS */
746 /* ToDo: add deadlock detection in GUM (similar to SMP) -- HWL */
750 /* If there's a GC pending, don't do anything until it has
754 IF_DEBUG(scheduler,sched_belch("waiting for GC"));
755 waitCondition( &gc_pending_cond, &sched_mutex );
759 #if defined(RTS_SUPPORTS_THREADS)
761 /* block until we've got a thread on the run queue and a free
765 if ( EMPTY_RUN_QUEUE() ) {
766 /* Give up our capability */
767 releaseCapability(cap);
769 /* If we're in the process of shutting down (& running the
770 * a batch of finalisers), don't wait around.
772 if ( shutting_down_scheduler ) {
773 RELEASE_LOCK(&sched_mutex);
776 IF_DEBUG(scheduler, sched_belch("thread %d: waiting for work", osThreadId()));
777 waitForWorkCapability(&sched_mutex, &cap, rtsTrue);
778 IF_DEBUG(scheduler, sched_belch("thread %d: work now available", osThreadId()));
781 if ( EMPTY_RUN_QUEUE() ) {
782 continue; // nothing to do
788 if (RtsFlags.GranFlags.Light)
789 GranSimLight_enter_system(event, &ActiveTSO); // adjust ActiveTSO etc
791 /* adjust time based on time-stamp */
792 if (event->time > CurrentTime[CurrentProc] &&
793 event->evttype != ContinueThread)
794 CurrentTime[CurrentProc] = event->time;
796 /* Deal with the idle PEs (may issue FindWork or MoveSpark events) */
797 if (!RtsFlags.GranFlags.Light)
800 IF_DEBUG(gran, fprintf(stderr, "GRAN: switch by event-type\n"));
802 /* main event dispatcher in GranSim */
803 switch (event->evttype) {
804 /* Should just be continuing execution */
806 IF_DEBUG(gran, fprintf(stderr, "GRAN: doing ContinueThread\n"));
807 /* ToDo: check assertion
808 ASSERT(run_queue_hd != (StgTSO*)NULL &&
809 run_queue_hd != END_TSO_QUEUE);
811 /* Ignore ContinueThreads for fetching threads (if synchr comm) */
812 if (!RtsFlags.GranFlags.DoAsyncFetch &&
813 procStatus[CurrentProc]==Fetching) {
814 belch("ghuH: Spurious ContinueThread while Fetching ignored; TSO %d (%p) [PE %d]",
815 CurrentTSO->id, CurrentTSO, CurrentProc);
818 /* Ignore ContinueThreads for completed threads */
819 if (CurrentTSO->what_next == ThreadComplete) {
820 belch("ghuH: found a ContinueThread event for completed thread %d (%p) [PE %d] (ignoring ContinueThread)",
821 CurrentTSO->id, CurrentTSO, CurrentProc);
824 /* Ignore ContinueThreads for threads that are being migrated */
825 if (PROCS(CurrentTSO)==Nowhere) {
826 belch("ghuH: trying to run the migrating TSO %d (%p) [PE %d] (ignoring ContinueThread)",
827 CurrentTSO->id, CurrentTSO, CurrentProc);
830 /* The thread should be at the beginning of the run queue */
831 if (CurrentTSO!=run_queue_hds[CurrentProc]) {
832 belch("ghuH: TSO %d (%p) [PE %d] is not at the start of the run_queue when doing a ContinueThread",
833 CurrentTSO->id, CurrentTSO, CurrentProc);
834 break; // run the thread anyway
837 new_event(proc, proc, CurrentTime[proc],
839 (StgTSO*)NULL, (StgClosure*)NULL, (rtsSpark*)NULL);
841 */ /* Catches superfluous CONTINUEs -- should be unnecessary */
842 break; // now actually run the thread; DaH Qu'vam yImuHbej
845 do_the_fetchnode(event);
846 goto next_thread; /* handle next event in event queue */
849 do_the_globalblock(event);
850 goto next_thread; /* handle next event in event queue */
853 do_the_fetchreply(event);
854 goto next_thread; /* handle next event in event queue */
856 case UnblockThread: /* Move from the blocked queue to the tail of */
857 do_the_unblock(event);
858 goto next_thread; /* handle next event in event queue */
860 case ResumeThread: /* Move from the blocked queue to the tail of */
861 /* the runnable queue ( i.e. Qu' SImqa'lu') */
862 event->tso->gran.blocktime +=
863 CurrentTime[CurrentProc] - event->tso->gran.blockedat;
864 do_the_startthread(event);
865 goto next_thread; /* handle next event in event queue */
868 do_the_startthread(event);
869 goto next_thread; /* handle next event in event queue */
872 do_the_movethread(event);
873 goto next_thread; /* handle next event in event queue */
876 do_the_movespark(event);
877 goto next_thread; /* handle next event in event queue */
880 do_the_findwork(event);
881 goto next_thread; /* handle next event in event queue */
884 barf("Illegal event type %u\n", event->evttype);
887 /* This point was scheduler_loop in the old RTS */
889 IF_DEBUG(gran, belch("GRAN: after main switch"));
891 TimeOfLastEvent = CurrentTime[CurrentProc];
892 TimeOfNextEvent = get_time_of_next_event();
893 IgnoreEvents=(TimeOfNextEvent==0); // HWL HACK
894 // CurrentTSO = ThreadQueueHd;
896 IF_DEBUG(gran, belch("GRAN: time of next event is: %ld",
899 if (RtsFlags.GranFlags.Light)
900 GranSimLight_leave_system(event, &ActiveTSO);
902 EndOfTimeSlice = CurrentTime[CurrentProc]+RtsFlags.GranFlags.time_slice;
905 belch("GRAN: end of time-slice is %#lx", EndOfTimeSlice));
907 /* in a GranSim setup the TSO stays on the run queue */
909 /* Take a thread from the run queue. */
910 t = POP_RUN_QUEUE(); // take_off_run_queue(t);
913 fprintf(stderr, "GRAN: About to run current thread, which is\n");
916 context_switch = 0; // turned on via GranYield, checking events and time slice
919 DumpGranEvent(GR_SCHEDULE, t));
921 procStatus[CurrentProc] = Busy;
924 if (PendingFetches != END_BF_QUEUE) {
928 /* ToDo: phps merge with spark activation above */
929 /* check whether we have local work and send requests if we have none */
930 if (EMPTY_RUN_QUEUE()) { /* no runnable threads */
931 /* :-[ no local threads => look out for local sparks */
932 /* the spark pool for the current PE */
933 pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable
934 if (advisory_thread_count < RtsFlags.ParFlags.maxThreads &&
935 pool->hd < pool->tl) {
937 * ToDo: add GC code check that we really have enough heap afterwards!!
939 * If we're here (no runnable threads) and we have pending
940 * sparks, we must have a space problem. Get enough space
941 * to turn one of those pending sparks into a
945 spark = findSpark(rtsFalse); /* get a spark */
946 if (spark != (rtsSpark) NULL) {
947 tso = activateSpark(spark); /* turn the spark into a thread */
948 IF_PAR_DEBUG(schedule,
949 belch("==== schedule: Created TSO %d (%p); %d threads active",
950 tso->id, tso, advisory_thread_count));
952 if (tso==END_TSO_QUEUE) { /* failed to activate spark->back to loop */
953 belch("==^^ failed to activate spark");
955 } /* otherwise fall through & pick-up new tso */
957 IF_PAR_DEBUG(verbose,
958 belch("==^^ no local sparks (spark pool contains only NFs: %d)",
959 spark_queue_len(pool)));
964 /* If we still have no work we need to send a FISH to get a spark
967 if (EMPTY_RUN_QUEUE()) {
968 /* =8-[ no local sparks => look for work on other PEs */
970 * We really have absolutely no work. Send out a fish
971 * (there may be some out there already), and wait for
972 * something to arrive. We clearly can't run any threads
973 * until a SCHEDULE or RESUME arrives, and so that's what
974 * we're hoping to see. (Of course, we still have to
975 * respond to other types of messages.)
977 TIME now = msTime() /*CURRENT_TIME*/;
978 IF_PAR_DEBUG(verbose,
979 belch("-- now=%ld", now));
980 IF_PAR_DEBUG(verbose,
981 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
982 (last_fish_arrived_at!=0 &&
983 last_fish_arrived_at+RtsFlags.ParFlags.fishDelay > now)) {
984 belch("--$$ delaying FISH until %ld (last fish %ld, delay %ld, now %ld)",
985 last_fish_arrived_at+RtsFlags.ParFlags.fishDelay,
986 last_fish_arrived_at,
987 RtsFlags.ParFlags.fishDelay, now);
990 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
991 (last_fish_arrived_at==0 ||
992 (last_fish_arrived_at+RtsFlags.ParFlags.fishDelay <= now))) {
993 /* outstandingFishes is set in sendFish, processFish;
994 avoid flooding system with fishes via delay */
996 sendFish(pe, mytid, NEW_FISH_AGE, NEW_FISH_HISTORY,
999 // Global statistics: count no. of fishes
1000 if (RtsFlags.ParFlags.ParStats.Global &&
1001 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
1002 globalParStats.tot_fish_mess++;
1006 receivedFinish = processMessages();
1009 } else if (PacketsWaiting()) { /* Look for incoming messages */
1010 receivedFinish = processMessages();
1013 /* Now we are sure that we have some work available */
1014 ASSERT(run_queue_hd != END_TSO_QUEUE);
1016 /* Take a thread from the run queue, if we have work */
1017 t = POP_RUN_QUEUE(); // take_off_run_queue(END_TSO_QUEUE);
1018 IF_DEBUG(sanity,checkTSO(t));
1020 /* ToDo: write something to the log-file
1021 if (RTSflags.ParFlags.granSimStats && !sameThread)
1022 DumpGranEvent(GR_SCHEDULE, RunnableThreadsHd);
1026 /* the spark pool for the current PE */
1027 pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable
1030 belch("--=^ %d threads, %d sparks on [%#x]",
1031 run_queue_len(), spark_queue_len(pool), CURRENT_PROC));
1034 if (0 && RtsFlags.ParFlags.ParStats.Full &&
1035 t && LastTSO && t->id != LastTSO->id &&
1036 LastTSO->why_blocked == NotBlocked &&
1037 LastTSO->what_next != ThreadComplete) {
1038 // if previously scheduled TSO not blocked we have to record the context switch
1039 DumpVeryRawGranEvent(TimeOfLastYield, CURRENT_PROC, CURRENT_PROC,
1040 GR_DESCHEDULE, LastTSO, (StgClosure *)NULL, 0, 0);
1043 if (RtsFlags.ParFlags.ParStats.Full &&
1044 (emitSchedule /* forced emit */ ||
1045 (t && LastTSO && t->id != LastTSO->id))) {
1047 we are running a different TSO, so write a schedule event to log file
1048 NB: If we use fair scheduling we also have to write a deschedule
1049 event for LastTSO; with unfair scheduling we know that the
1050 previous tso has blocked whenever we switch to another tso, so
1051 we don't need it in GUM for now
1053 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
1054 GR_SCHEDULE, t, (StgClosure *)NULL, 0, 0);
1055 emitSchedule = rtsFalse;
1059 #else /* !GRAN && !PAR */
1061 /* grab a thread from the run queue */
1062 ASSERT(run_queue_hd != END_TSO_QUEUE);
1063 t = POP_RUN_QUEUE();
1064 // Sanity check the thread we're about to run. This can be
1065 // expensive if there is lots of thread switching going on...
1066 IF_DEBUG(sanity,checkTSO(t));
1069 cap->r.rCurrentTSO = t;
1071 /* context switches are now initiated by the timer signal, unless
1072 * the user specified "context switch as often as possible", with
1075 if ((RtsFlags.ConcFlags.ctxtSwitchTicks == 0
1076 && (run_queue_hd != END_TSO_QUEUE
1077 || blocked_queue_hd != END_TSO_QUEUE
1078 || sleeping_queue != END_TSO_QUEUE)))
1085 RELEASE_LOCK(&sched_mutex);
1087 IF_DEBUG(scheduler, sched_belch("-->> running thread %ld %s ...",
1088 t->id, whatNext_strs[t->what_next]));
1091 startHeapProfTimer();
1094 /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
1095 /* Run the current thread
1097 prev_what_next = t->what_next;
1098 switch (prev_what_next) {
1100 case ThreadComplete:
1101 /* Thread already finished, return to scheduler. */
1102 ret = ThreadFinished;
1105 ret = StgRun((StgFunPtr) stg_returnToStackTop, &cap->r);
1107 case ThreadInterpret:
1108 ret = interpretBCO(cap);
1111 barf("schedule: invalid what_next field");
1113 /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
1115 /* Costs for the scheduler are assigned to CCS_SYSTEM */
1117 stopHeapProfTimer();
1121 ACQUIRE_LOCK(&sched_mutex);
1123 #ifdef RTS_SUPPORTS_THREADS
1124 IF_DEBUG(scheduler,fprintf(stderr,"scheduler (task %ld): ", osThreadId()););
1125 #elif !defined(GRAN) && !defined(PAR)
1126 IF_DEBUG(scheduler,fprintf(stderr,"scheduler: "););
1128 t = cap->r.rCurrentTSO;
1131 /* HACK 675: if the last thread didn't yield, make sure to print a
1132 SCHEDULE event to the log file when StgRunning the next thread, even
1133 if it is the same one as before */
1135 TimeOfLastYield = CURRENT_TIME;
1141 IF_DEBUG(gran, DumpGranEvent(GR_DESCHEDULE, t));
1142 globalGranStats.tot_heapover++;
1144 globalParStats.tot_heapover++;
1147 // did the task ask for a large block?
1148 if (cap->r.rHpAlloc > BLOCK_SIZE_W) {
1149 // if so, get one and push it on the front of the nursery.
1153 blocks = (nat)BLOCK_ROUND_UP(cap->r.rHpAlloc * sizeof(W_)) / BLOCK_SIZE;
1155 IF_DEBUG(scheduler,belch("--<< thread %ld (%s) stopped: requesting a large block (size %d)",
1156 t->id, whatNext_strs[t->what_next], blocks));
1158 // don't do this if it would push us over the
1159 // alloc_blocks_lim limit; we'll GC first.
1160 if (alloc_blocks + blocks < alloc_blocks_lim) {
1162 alloc_blocks += blocks;
1163 bd = allocGroup( blocks );
1165 // link the new group into the list
1166 bd->link = cap->r.rCurrentNursery;
1167 bd->u.back = cap->r.rCurrentNursery->u.back;
1168 if (cap->r.rCurrentNursery->u.back != NULL) {
1169 cap->r.rCurrentNursery->u.back->link = bd;
1171 ASSERT(g0s0->blocks == cap->r.rCurrentNursery &&
1172 g0s0->blocks == cap->r.rNursery);
1173 cap->r.rNursery = g0s0->blocks = bd;
1175 cap->r.rCurrentNursery->u.back = bd;
1177 // initialise it as a nursery block. We initialise the
1178 // step, gen_no, and flags field of *every* sub-block in
1179 // this large block, because this is easier than making
1180 // sure that we always find the block head of a large
1181 // block whenever we call Bdescr() (eg. evacuate() and
1182 // isAlive() in the GC would both have to do this, at
1186 for (x = bd; x < bd + blocks; x++) {
1193 // don't forget to update the block count in g0s0.
1194 g0s0->n_blocks += blocks;
1195 // This assert can be a killer if the app is doing lots
1196 // of large block allocations.
1197 ASSERT(countBlocks(g0s0->blocks) == g0s0->n_blocks);
1199 // now update the nursery to point to the new block
1200 cap->r.rCurrentNursery = bd;
1202 // we might be unlucky and have another thread get on the
1203 // run queue before us and steal the large block, but in that
1204 // case the thread will just end up requesting another large
1206 PUSH_ON_RUN_QUEUE(t);
1211 /* make all the running tasks block on a condition variable,
1212 * maybe set context_switch and wait till they all pile in,
1213 * then have them wait on a GC condition variable.
1215 IF_DEBUG(scheduler,belch("--<< thread %ld (%s) stopped: HeapOverflow",
1216 t->id, whatNext_strs[t->what_next]));
1219 ASSERT(!is_on_queue(t,CurrentProc));
1221 /* Currently we emit a DESCHEDULE event before GC in GUM.
1222 ToDo: either add separate event to distinguish SYSTEM time from rest
1223 or just nuke this DESCHEDULE (and the following SCHEDULE) */
1224 if (0 && RtsFlags.ParFlags.ParStats.Full) {
1225 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
1226 GR_DESCHEDULE, t, (StgClosure *)NULL, 0, 0);
1227 emitSchedule = rtsTrue;
1231 ready_to_gc = rtsTrue;
1232 context_switch = 1; /* stop other threads ASAP */
1233 PUSH_ON_RUN_QUEUE(t);
1234 /* actual GC is done at the end of the while loop */
1240 DumpGranEvent(GR_DESCHEDULE, t));
1241 globalGranStats.tot_stackover++;
1244 // DumpGranEvent(GR_DESCHEDULE, t);
1245 globalParStats.tot_stackover++;
1247 IF_DEBUG(scheduler,belch("--<< thread %ld (%s) stopped, StackOverflow",
1248 t->id, whatNext_strs[t->what_next]));
1249 /* just adjust the stack for this thread, then pop it back
1255 /* enlarge the stack */
1256 StgTSO *new_t = threadStackOverflow(t);
1258 /* This TSO has moved, so update any pointers to it from the
1259 * main thread stack. It better not be on any other queues...
1260 * (it shouldn't be).
1262 for (m = main_threads; m != NULL; m = m->link) {
1267 threadPaused(new_t);
1268 PUSH_ON_RUN_QUEUE(new_t);
1272 case ThreadYielding:
1275 DumpGranEvent(GR_DESCHEDULE, t));
1276 globalGranStats.tot_yields++;
1279 // DumpGranEvent(GR_DESCHEDULE, t);
1280 globalParStats.tot_yields++;
1282 /* put the thread back on the run queue. Then, if we're ready to
1283 * GC, check whether this is the last task to stop. If so, wake
1284 * up the GC thread. getThread will block during a GC until the
1288 if (t->what_next != prev_what_next) {
1289 belch("--<< thread %ld (%s) stopped to switch evaluators",
1290 t->id, whatNext_strs[t->what_next]);
1292 belch("--<< thread %ld (%s) stopped, yielding",
1293 t->id, whatNext_strs[t->what_next]);
1298 //belch("&& Doing sanity check on yielding TSO %ld.", t->id);
1300 ASSERT(t->link == END_TSO_QUEUE);
1302 // Shortcut if we're just switching evaluators: don't bother
1303 // doing stack squeezing (which can be expensive), just run the
1305 if (t->what_next != prev_what_next) {
1312 ASSERT(!is_on_queue(t,CurrentProc));
1315 //belch("&& Doing sanity check on all ThreadQueues (and their TSOs).");
1316 checkThreadQsSanity(rtsTrue));
1320 if (RtsFlags.ParFlags.doFairScheduling) {
1321 /* this does round-robin scheduling; good for concurrency */
1322 APPEND_TO_RUN_QUEUE(t);
1324 /* this does unfair scheduling; good for parallelism */
1325 PUSH_ON_RUN_QUEUE(t);
1328 // this does round-robin scheduling; good for concurrency
1329 APPEND_TO_RUN_QUEUE(t);
1333 /* add a ContinueThread event to actually process the thread */
1334 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1336 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1338 belch("GRAN: eventq and runnableq after adding yielded thread to queue again:");
1347 belch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: ",
1348 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)));
1349 if (t->block_info.closure!=(StgClosure*)NULL) print_bq(t->block_info.closure));
1351 // ??? needed; should emit block before
1353 DumpGranEvent(GR_DESCHEDULE, t));
1354 prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
1357 ASSERT(procStatus[CurrentProc]==Busy ||
1358 ((procStatus[CurrentProc]==Fetching) &&
1359 (t->block_info.closure!=(StgClosure*)NULL)));
1360 if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
1361 !(!RtsFlags.GranFlags.DoAsyncFetch &&
1362 procStatus[CurrentProc]==Fetching))
1363 procStatus[CurrentProc] = Idle;
1367 belch("--<< thread %ld (%p; %s) stopped, blocking on node %p with BQ: ",
1368 t->id, t, whatNext_strs[t->what_next], t->block_info.closure));
1371 if (t->block_info.closure!=(StgClosure*)NULL)
1372 print_bq(t->block_info.closure));
1374 /* Send a fetch (if BlockedOnGA) and dump event to log file */
1377 /* whatever we schedule next, we must log that schedule */
1378 emitSchedule = rtsTrue;
1381 /* don't need to do anything. Either the thread is blocked on
1382 * I/O, in which case we'll have called addToBlockedQueue
1383 * previously, or it's blocked on an MVar or Blackhole, in which
1384 * case it'll be on the relevant queue already.
1387 fprintf(stderr, "--<< thread %d (%s) stopped: ",
1388 t->id, whatNext_strs[t->what_next]);
1389 printThreadBlockage(t);
1390 fprintf(stderr, "\n"));
1392 /* Only for dumping event to log file
1393 ToDo: do I need this in GranSim, too?
1400 case ThreadFinished:
1401 /* Need to check whether this was a main thread, and if so, signal
1402 * the task that started it with the return value. If we have no
1403 * more main threads, we probably need to stop all the tasks until
1406 /* We also end up here if the thread kills itself with an
1407 * uncaught exception, see Exception.hc.
1409 IF_DEBUG(scheduler,belch("--++ thread %d (%s) finished",
1410 t->id, whatNext_strs[t->what_next]));
1412 endThread(t, CurrentProc); // clean-up the thread
1414 /* For now all are advisory -- HWL */
1415 //if(t->priority==AdvisoryPriority) ??
1416 advisory_thread_count--;
1419 if(t->dist.priority==RevalPriority)
1423 if (RtsFlags.ParFlags.ParStats.Full &&
1424 !RtsFlags.ParFlags.ParStats.Suppressed)
1425 DumpEndEvent(CURRENT_PROC, t, rtsFalse /* not mandatory */);
1430 barf("schedule: invalid thread return code %d", (int)ret);
1434 // When we have +RTS -i0 and we're heap profiling, do a census at
1435 // every GC. This lets us get repeatable runs for debugging.
1436 if (performHeapProfile ||
1437 (RtsFlags.ProfFlags.profileInterval==0 &&
1438 RtsFlags.ProfFlags.doHeapProfile && ready_to_gc)) {
1439 GarbageCollect(GetRoots, rtsTrue);
1441 performHeapProfile = rtsFalse;
1442 ready_to_gc = rtsFalse; // we already GC'd
1448 && allFreeCapabilities()
1451 /* everybody back, start the GC.
1452 * Could do it in this thread, or signal a condition var
1453 * to do it in another thread. Either way, we need to
1454 * broadcast on gc_pending_cond afterward.
1456 #if defined(RTS_SUPPORTS_THREADS)
1457 IF_DEBUG(scheduler,sched_belch("doing GC"));
1459 GarbageCollect(GetRoots,rtsFalse);
1460 ready_to_gc = rtsFalse;
1462 broadcastCondition(&gc_pending_cond);
1465 /* add a ContinueThread event to continue execution of current thread */
1466 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1468 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1470 fprintf(stderr, "GRAN: eventq and runnableq after Garbage collection:\n");
1478 IF_GRAN_DEBUG(unused,
1479 print_eventq(EventHd));
1481 event = get_next_event();
1484 /* ToDo: wait for next message to arrive rather than busy wait */
1487 } /* end of while(1) */
1489 IF_PAR_DEBUG(verbose,
1490 belch("== Leaving schedule() after having received Finish"));
1493 /* ---------------------------------------------------------------------------
1494 * Singleton fork(). Do not copy any running threads.
1495 * ------------------------------------------------------------------------- */
1497 StgInt forkProcess(StgTSO* tso) {
1499 #ifndef mingw32_TARGET_OS
1505 IF_DEBUG(scheduler,sched_belch("forking!"));
1508 if (pid) { /* parent */
1510 /* just return the pid */
1512 } else { /* child */
1513 /* wipe all other threads */
1514 run_queue_hd = run_queue_tl = tso;
1515 tso->link = END_TSO_QUEUE;
1517 /* When clearing out the threads, we need to ensure
1518 that a 'main thread' is left behind; if there isn't,
1519 the Scheduler will shutdown next time it is entered.
1521 ==> we don't kill a thread that's on the main_threads
1522 list (nor the current thread.)
1524 [ Attempts at implementing the more ambitious scheme of
1525 killing the main_threads also, and then adding the
1526 current thread onto the main_threads list if it wasn't
1527 there already, failed -- waitThread() (for one) wasn't
1528 up to it. If it proves to be desirable to also kill
1529 the main threads, then this scheme will have to be
1530 revisited (and fully debugged!)
1535 /* DO NOT TOUCH THE QUEUES directly because most of the code around
1536 us is picky about finding the thread still in its queue when
1537 handling the deleteThread() */
1539 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
1542 /* Don't kill the current thread.. */
1543 if (t->id == tso->id) continue;
1545 /* ..or a main thread */
1546 for (m = main_threads; m != NULL; m = m->link) {
1547 if (m->tso->id == t->id) {
1559 barf("forkProcess#: primop not implemented for mingw32, sorry! (%u)\n", tso->id);
1560 /* pointlessly printing out the TSOs 'id' to avoid CC unused warning. */
1562 #endif /* mingw32 */
1565 /* ---------------------------------------------------------------------------
1566 * deleteAllThreads(): kill all the live threads.
1568 * This is used when we catch a user interrupt (^C), before performing
1569 * any necessary cleanups and running finalizers.
1571 * Locks: sched_mutex held.
1572 * ------------------------------------------------------------------------- */
1574 void deleteAllThreads ( void )
1577 IF_DEBUG(scheduler,sched_belch("deleting all threads"));
1578 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
1579 next = t->global_link;
1582 run_queue_hd = run_queue_tl = END_TSO_QUEUE;
1583 blocked_queue_hd = blocked_queue_tl = END_TSO_QUEUE;
1584 sleeping_queue = END_TSO_QUEUE;
1587 /* startThread and insertThread are now in GranSim.c -- HWL */
1590 //@node Suspend and Resume, Run queue code, Main scheduling loop, Main scheduling code
1591 //@subsection Suspend and Resume
1593 /* ---------------------------------------------------------------------------
1594 * Suspending & resuming Haskell threads.
1596 * When making a "safe" call to C (aka _ccall_GC), the task gives back
1597 * its capability before calling the C function. This allows another
1598 * task to pick up the capability and carry on running Haskell
1599 * threads. It also means that if the C call blocks, it won't lock
1602 * The Haskell thread making the C call is put to sleep for the
1603 * duration of the call, on the susepended_ccalling_threads queue. We
1604 * give out a token to the task, which it can use to resume the thread
1605 * on return from the C function.
1606 * ------------------------------------------------------------------------- */
1609 suspendThread( StgRegTable *reg,
1611 #if !defined(RTS_SUPPORTS_THREADS) && !defined(DEBUG)
1619 /* assume that *reg is a pointer to the StgRegTable part
1622 cap = (Capability *)((void *)reg - sizeof(StgFunTable));
1624 ACQUIRE_LOCK(&sched_mutex);
1627 sched_belch("thread %d did a _ccall_gc (is_concurrent: %d)", cap->r.rCurrentTSO->id,concCall));
1629 // XXX this might not be necessary --SDM
1630 cap->r.rCurrentTSO->what_next = ThreadRunGHC;
1632 threadPaused(cap->r.rCurrentTSO);
1633 cap->r.rCurrentTSO->link = suspended_ccalling_threads;
1634 suspended_ccalling_threads = cap->r.rCurrentTSO;
1636 #if defined(RTS_SUPPORTS_THREADS)
1637 if(cap->r.rCurrentTSO->blocked_exceptions == NULL)
1639 cap->r.rCurrentTSO->why_blocked = BlockedOnCCall;
1640 cap->r.rCurrentTSO->blocked_exceptions = END_TSO_QUEUE;
1644 cap->r.rCurrentTSO->why_blocked = BlockedOnCCall_NoUnblockExc;
1648 /* Use the thread ID as the token; it should be unique */
1649 tok = cap->r.rCurrentTSO->id;
1651 /* Hand back capability */
1652 releaseCapability(cap);
1654 #if defined(RTS_SUPPORTS_THREADS)
1655 /* Preparing to leave the RTS, so ensure there's a native thread/task
1656 waiting to take over.
1658 IF_DEBUG(scheduler, sched_belch("worker thread (%d, osthread %p): leaving RTS", tok, osThreadId()));
1659 //if (concCall) { // implementing "safe" as opposed to "threadsafe" is more difficult
1660 startTask(taskStart);
1664 /* Other threads _might_ be available for execution; signal this */
1666 RELEASE_LOCK(&sched_mutex);
1671 resumeThread( StgInt tok,
1672 rtsBool concCall STG_UNUSED )
1674 StgTSO *tso, **prev;
1677 #if defined(RTS_SUPPORTS_THREADS)
1678 /* Wait for permission to re-enter the RTS with the result. */
1679 ACQUIRE_LOCK(&sched_mutex);
1680 grabReturnCapability(&sched_mutex, &cap);
1682 IF_DEBUG(scheduler, sched_belch("worker thread (%d, osthread %p): re-entering RTS", tok, osThreadId()));
1684 grabCapability(&cap);
1687 /* Remove the thread off of the suspended list */
1688 prev = &suspended_ccalling_threads;
1689 for (tso = suspended_ccalling_threads;
1690 tso != END_TSO_QUEUE;
1691 prev = &tso->link, tso = tso->link) {
1692 if (tso->id == (StgThreadID)tok) {
1697 if (tso == END_TSO_QUEUE) {
1698 barf("resumeThread: thread not found");
1700 tso->link = END_TSO_QUEUE;
1702 #if defined(RTS_SUPPORTS_THREADS)
1703 if(tso->why_blocked == BlockedOnCCall)
1705 awakenBlockedQueueNoLock(tso->blocked_exceptions);
1706 tso->blocked_exceptions = NULL;
1710 /* Reset blocking status */
1711 tso->why_blocked = NotBlocked;
1713 cap->r.rCurrentTSO = tso;
1714 #if defined(RTS_SUPPORTS_THREADS)
1715 RELEASE_LOCK(&sched_mutex);
1721 /* ---------------------------------------------------------------------------
1723 * ------------------------------------------------------------------------ */
1724 static void unblockThread(StgTSO *tso);
1726 /* ---------------------------------------------------------------------------
1727 * Comparing Thread ids.
1729 * This is used from STG land in the implementation of the
1730 * instances of Eq/Ord for ThreadIds.
1731 * ------------------------------------------------------------------------ */
1734 cmp_thread(StgPtr tso1, StgPtr tso2)
1736 StgThreadID id1 = ((StgTSO *)tso1)->id;
1737 StgThreadID id2 = ((StgTSO *)tso2)->id;
1739 if (id1 < id2) return (-1);
1740 if (id1 > id2) return 1;
1744 /* ---------------------------------------------------------------------------
1745 * Fetching the ThreadID from an StgTSO.
1747 * This is used in the implementation of Show for ThreadIds.
1748 * ------------------------------------------------------------------------ */
1750 rts_getThreadId(StgPtr tso)
1752 return ((StgTSO *)tso)->id;
1757 labelThread(StgPtr tso, char *label)
1762 /* Caveat: Once set, you can only set the thread name to "" */
1763 len = strlen(label)+1;
1764 buf = stgMallocBytes(len * sizeof(char), "Schedule.c:labelThread()");
1765 strncpy(buf,label,len);
1766 /* Update will free the old memory for us */
1767 updateThreadLabel((StgWord)tso,buf);
1771 /* ---------------------------------------------------------------------------
1772 Create a new thread.
1774 The new thread starts with the given stack size. Before the
1775 scheduler can run, however, this thread needs to have a closure
1776 (and possibly some arguments) pushed on its stack. See
1777 pushClosure() in Schedule.h.
1779 createGenThread() and createIOThread() (in SchedAPI.h) are
1780 convenient packaged versions of this function.
1782 currently pri (priority) is only used in a GRAN setup -- HWL
1783 ------------------------------------------------------------------------ */
1784 //@cindex createThread
1786 /* currently pri (priority) is only used in a GRAN setup -- HWL */
1788 createThread(nat size, StgInt pri)
1791 createThread(nat size)
1798 /* First check whether we should create a thread at all */
1800 /* check that no more than RtsFlags.ParFlags.maxThreads threads are created */
1801 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads) {
1803 belch("{createThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
1804 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
1805 return END_TSO_QUEUE;
1811 ASSERT(!RtsFlags.GranFlags.Light || CurrentProc==0);
1814 // ToDo: check whether size = stack_size - TSO_STRUCT_SIZEW
1816 /* catch ridiculously small stack sizes */
1817 if (size < MIN_STACK_WORDS + TSO_STRUCT_SIZEW) {
1818 size = MIN_STACK_WORDS + TSO_STRUCT_SIZEW;
1821 stack_size = size - TSO_STRUCT_SIZEW;
1823 tso = (StgTSO *)allocate(size);
1824 TICK_ALLOC_TSO(stack_size, 0);
1826 SET_HDR(tso, &stg_TSO_info, CCS_SYSTEM);
1828 SET_GRAN_HDR(tso, ThisPE);
1831 // Always start with the compiled code evaluator
1832 tso->what_next = ThreadRunGHC;
1834 /* tso->id needs to be unique. For now we use a heavyweight mutex to
1835 * protect the increment operation on next_thread_id.
1836 * In future, we could use an atomic increment instead.
1838 ACQUIRE_LOCK(&thread_id_mutex);
1839 tso->id = next_thread_id++;
1840 RELEASE_LOCK(&thread_id_mutex);
1842 tso->why_blocked = NotBlocked;
1843 tso->blocked_exceptions = NULL;
1845 tso->stack_size = stack_size;
1846 tso->max_stack_size = round_to_mblocks(RtsFlags.GcFlags.maxStkSize)
1848 tso->sp = (P_)&(tso->stack) + stack_size;
1851 tso->prof.CCCS = CCS_MAIN;
1854 /* put a stop frame on the stack */
1855 tso->sp -= sizeofW(StgStopFrame);
1856 SET_HDR((StgClosure*)tso->sp,(StgInfoTable *)&stg_stop_thread_info,CCS_SYSTEM);
1859 tso->link = END_TSO_QUEUE;
1860 /* uses more flexible routine in GranSim */
1861 insertThread(tso, CurrentProc);
1863 /* In a non-GranSim setup the pushing of a TSO onto the runq is separated
1869 if (RtsFlags.GranFlags.GranSimStats.Full)
1870 DumpGranEvent(GR_START,tso);
1872 if (RtsFlags.ParFlags.ParStats.Full)
1873 DumpGranEvent(GR_STARTQ,tso);
1874 /* HACk to avoid SCHEDULE
1878 /* Link the new thread on the global thread list.
1880 tso->global_link = all_threads;
1884 tso->dist.priority = MandatoryPriority; //by default that is...
1888 tso->gran.pri = pri;
1890 tso->gran.magic = TSO_MAGIC; // debugging only
1892 tso->gran.sparkname = 0;
1893 tso->gran.startedat = CURRENT_TIME;
1894 tso->gran.exported = 0;
1895 tso->gran.basicblocks = 0;
1896 tso->gran.allocs = 0;
1897 tso->gran.exectime = 0;
1898 tso->gran.fetchtime = 0;
1899 tso->gran.fetchcount = 0;
1900 tso->gran.blocktime = 0;
1901 tso->gran.blockcount = 0;
1902 tso->gran.blockedat = 0;
1903 tso->gran.globalsparks = 0;
1904 tso->gran.localsparks = 0;
1905 if (RtsFlags.GranFlags.Light)
1906 tso->gran.clock = Now; /* local clock */
1908 tso->gran.clock = 0;
1910 IF_DEBUG(gran,printTSO(tso));
1913 tso->par.magic = TSO_MAGIC; // debugging only
1915 tso->par.sparkname = 0;
1916 tso->par.startedat = CURRENT_TIME;
1917 tso->par.exported = 0;
1918 tso->par.basicblocks = 0;
1919 tso->par.allocs = 0;
1920 tso->par.exectime = 0;
1921 tso->par.fetchtime = 0;
1922 tso->par.fetchcount = 0;
1923 tso->par.blocktime = 0;
1924 tso->par.blockcount = 0;
1925 tso->par.blockedat = 0;
1926 tso->par.globalsparks = 0;
1927 tso->par.localsparks = 0;
1931 globalGranStats.tot_threads_created++;
1932 globalGranStats.threads_created_on_PE[CurrentProc]++;
1933 globalGranStats.tot_sq_len += spark_queue_len(CurrentProc);
1934 globalGranStats.tot_sq_probes++;
1936 // collect parallel global statistics (currently done together with GC stats)
1937 if (RtsFlags.ParFlags.ParStats.Global &&
1938 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
1939 //fprintf(stderr, "Creating thread %d @ %11.2f\n", tso->id, usertime());
1940 globalParStats.tot_threads_created++;
1946 belch("==__ schedule: Created TSO %d (%p);",
1947 CurrentProc, tso, tso->id));
1949 IF_PAR_DEBUG(verbose,
1950 belch("==__ schedule: Created TSO %d (%p); %d threads active",
1951 tso->id, tso, advisory_thread_count));
1953 IF_DEBUG(scheduler,sched_belch("created thread %ld, stack size = %lx words",
1954 tso->id, tso->stack_size));
1961 all parallel thread creation calls should fall through the following routine.
1964 createSparkThread(rtsSpark spark)
1966 ASSERT(spark != (rtsSpark)NULL);
1967 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads)
1969 barf("{createSparkThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
1970 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
1971 return END_TSO_QUEUE;
1975 tso = createThread(RtsFlags.GcFlags.initialStkSize);
1976 if (tso==END_TSO_QUEUE)
1977 barf("createSparkThread: Cannot create TSO");
1979 tso->priority = AdvisoryPriority;
1981 pushClosure(tso,spark);
1982 PUSH_ON_RUN_QUEUE(tso);
1983 advisory_thread_count++;
1990 Turn a spark into a thread.
1991 ToDo: fix for SMP (needs to acquire SCHED_MUTEX!)
1994 //@cindex activateSpark
1996 activateSpark (rtsSpark spark)
2000 tso = createSparkThread(spark);
2001 if (RtsFlags.ParFlags.ParStats.Full) {
2002 //ASSERT(run_queue_hd == END_TSO_QUEUE); // I think ...
2003 IF_PAR_DEBUG(verbose,
2004 belch("==^^ activateSpark: turning spark of closure %p (%s) into a thread",
2005 (StgClosure *)spark, info_type((StgClosure *)spark)));
2007 // ToDo: fwd info on local/global spark to thread -- HWL
2008 // tso->gran.exported = spark->exported;
2009 // tso->gran.locked = !spark->global;
2010 // tso->gran.sparkname = spark->name;
2016 static SchedulerStatus waitThread_(/*out*/StgMainThread* m
2017 #if defined(THREADED_RTS)
2018 , rtsBool blockWaiting
2023 /* ---------------------------------------------------------------------------
2026 * scheduleThread puts a thread on the head of the runnable queue.
2027 * This will usually be done immediately after a thread is created.
2028 * The caller of scheduleThread must create the thread using e.g.
2029 * createThread and push an appropriate closure
2030 * on this thread's stack before the scheduler is invoked.
2031 * ------------------------------------------------------------------------ */
2033 static void scheduleThread_ (StgTSO* tso);
2036 scheduleThread_(StgTSO *tso)
2038 // Precondition: sched_mutex must be held.
2040 /* Put the new thread on the head of the runnable queue. The caller
2041 * better push an appropriate closure on this thread's stack
2042 * beforehand. In the SMP case, the thread may start running as
2043 * soon as we release the scheduler lock below.
2045 PUSH_ON_RUN_QUEUE(tso);
2049 IF_DEBUG(scheduler,printTSO(tso));
2053 void scheduleThread(StgTSO* tso)
2055 ACQUIRE_LOCK(&sched_mutex);
2056 scheduleThread_(tso);
2057 RELEASE_LOCK(&sched_mutex);
2061 scheduleWaitThread(StgTSO* tso, /*[out]*/HaskellObj* ret)
2062 { // Precondition: sched_mutex must be held
2065 m = stgMallocBytes(sizeof(StgMainThread), "waitThread");
2069 #if defined(RTS_SUPPORTS_THREADS)
2070 initCondition(&m->wakeup);
2073 /* Put the thread on the main-threads list prior to scheduling the TSO.
2074 Failure to do so introduces a race condition in the MT case (as
2075 identified by Wolfgang Thaller), whereby the new task/OS thread
2076 created by scheduleThread_() would complete prior to the thread
2077 that spawned it managed to put 'itself' on the main-threads list.
2078 The upshot of it all being that the worker thread wouldn't get to
2079 signal the completion of the its work item for the main thread to
2080 see (==> it got stuck waiting.) -- sof 6/02.
2082 IF_DEBUG(scheduler, sched_belch("waiting for thread (%d)\n", tso->id));
2084 m->link = main_threads;
2087 scheduleThread_(tso);
2088 #if defined(THREADED_RTS)
2089 return waitThread_(m, rtsTrue);
2091 return waitThread_(m);
2095 /* ---------------------------------------------------------------------------
2098 * Initialise the scheduler. This resets all the queues - if the
2099 * queues contained any threads, they'll be garbage collected at the
2102 * ------------------------------------------------------------------------ */
2106 term_handler(int sig STG_UNUSED)
2109 ACQUIRE_LOCK(&term_mutex);
2111 RELEASE_LOCK(&term_mutex);
2122 for (i=0; i<=MAX_PROC; i++) {
2123 run_queue_hds[i] = END_TSO_QUEUE;
2124 run_queue_tls[i] = END_TSO_QUEUE;
2125 blocked_queue_hds[i] = END_TSO_QUEUE;
2126 blocked_queue_tls[i] = END_TSO_QUEUE;
2127 ccalling_threadss[i] = END_TSO_QUEUE;
2128 sleeping_queue = END_TSO_QUEUE;
2131 run_queue_hd = END_TSO_QUEUE;
2132 run_queue_tl = END_TSO_QUEUE;
2133 blocked_queue_hd = END_TSO_QUEUE;
2134 blocked_queue_tl = END_TSO_QUEUE;
2135 sleeping_queue = END_TSO_QUEUE;
2138 suspended_ccalling_threads = END_TSO_QUEUE;
2140 main_threads = NULL;
2141 all_threads = END_TSO_QUEUE;
2146 RtsFlags.ConcFlags.ctxtSwitchTicks =
2147 RtsFlags.ConcFlags.ctxtSwitchTime / TICK_MILLISECS;
2149 #if defined(RTS_SUPPORTS_THREADS)
2150 /* Initialise the mutex and condition variables used by
2152 initMutex(&sched_mutex);
2153 initMutex(&term_mutex);
2154 initMutex(&thread_id_mutex);
2156 initCondition(&thread_ready_cond);
2160 initCondition(&gc_pending_cond);
2163 #if defined(RTS_SUPPORTS_THREADS)
2164 ACQUIRE_LOCK(&sched_mutex);
2167 /* Install the SIGHUP handler */
2170 struct sigaction action,oact;
2172 action.sa_handler = term_handler;
2173 sigemptyset(&action.sa_mask);
2174 action.sa_flags = 0;
2175 if (sigaction(SIGTERM, &action, &oact) != 0) {
2176 barf("can't install TERM handler");
2181 /* A capability holds the state a native thread needs in
2182 * order to execute STG code. At least one capability is
2183 * floating around (only SMP builds have more than one).
2187 #if defined(RTS_SUPPORTS_THREADS)
2188 /* start our haskell execution tasks */
2190 startTaskManager(RtsFlags.ParFlags.nNodes, taskStart);
2192 startTaskManager(0,taskStart);
2196 #if /* defined(SMP) ||*/ defined(PAR)
2200 #if defined(RTS_SUPPORTS_THREADS)
2201 RELEASE_LOCK(&sched_mutex);
2207 exitScheduler( void )
2209 #if defined(RTS_SUPPORTS_THREADS)
2212 shutting_down_scheduler = rtsTrue;
2215 /* -----------------------------------------------------------------------------
2216 Managing the per-task allocation areas.
2218 Each capability comes with an allocation area. These are
2219 fixed-length block lists into which allocation can be done.
2221 ToDo: no support for two-space collection at the moment???
2222 -------------------------------------------------------------------------- */
2224 /* -----------------------------------------------------------------------------
2225 * waitThread is the external interface for running a new computation
2226 * and waiting for the result.
2228 * In the non-SMP case, we create a new main thread, push it on the
2229 * main-thread stack, and invoke the scheduler to run it. The
2230 * scheduler will return when the top main thread on the stack has
2231 * completed or died, and fill in the necessary fields of the
2232 * main_thread structure.
2234 * In the SMP case, we create a main thread as before, but we then
2235 * create a new condition variable and sleep on it. When our new
2236 * main thread has completed, we'll be woken up and the status/result
2237 * will be in the main_thread struct.
2238 * -------------------------------------------------------------------------- */
2241 howManyThreadsAvail ( void )
2245 for (q = run_queue_hd; q != END_TSO_QUEUE; q = q->link)
2247 for (q = blocked_queue_hd; q != END_TSO_QUEUE; q = q->link)
2249 for (q = sleeping_queue; q != END_TSO_QUEUE; q = q->link)
2255 finishAllThreads ( void )
2258 while (run_queue_hd != END_TSO_QUEUE) {
2259 waitThread ( run_queue_hd, NULL);
2261 while (blocked_queue_hd != END_TSO_QUEUE) {
2262 waitThread ( blocked_queue_hd, NULL);
2264 while (sleeping_queue != END_TSO_QUEUE) {
2265 waitThread ( blocked_queue_hd, NULL);
2268 (blocked_queue_hd != END_TSO_QUEUE ||
2269 run_queue_hd != END_TSO_QUEUE ||
2270 sleeping_queue != END_TSO_QUEUE);
2274 waitThread(StgTSO *tso, /*out*/StgClosure **ret)
2277 SchedulerStatus stat;
2279 m = stgMallocBytes(sizeof(StgMainThread), "waitThread");
2283 #if defined(RTS_SUPPORTS_THREADS)
2284 initCondition(&m->wakeup);
2287 /* see scheduleWaitThread() comment */
2288 ACQUIRE_LOCK(&sched_mutex);
2289 m->link = main_threads;
2292 IF_DEBUG(scheduler, sched_belch("waiting for thread %d", tso->id));
2293 #if defined(THREADED_RTS)
2294 stat = waitThread_(m, rtsFalse);
2296 stat = waitThread_(m);
2298 RELEASE_LOCK(&sched_mutex);
2304 waitThread_(StgMainThread* m
2305 #if defined(THREADED_RTS)
2306 , rtsBool blockWaiting
2310 SchedulerStatus stat;
2312 // Precondition: sched_mutex must be held.
2313 IF_DEBUG(scheduler, sched_belch("== scheduler: new main thread (%d)\n", m->tso->id));
2315 #if defined(RTS_SUPPORTS_THREADS)
2317 # if defined(THREADED_RTS)
2318 if (!blockWaiting) {
2319 /* In the threaded case, the OS thread that called main()
2320 * gets to enter the RTS directly without going via another
2323 main_main_thread = m;
2324 RELEASE_LOCK(&sched_mutex);
2326 ACQUIRE_LOCK(&sched_mutex);
2327 main_main_thread = NULL;
2328 ASSERT(m->stat != NoStatus);
2333 waitCondition(&m->wakeup, &sched_mutex);
2334 } while (m->stat == NoStatus);
2337 /* GranSim specific init */
2338 CurrentTSO = m->tso; // the TSO to run
2339 procStatus[MainProc] = Busy; // status of main PE
2340 CurrentProc = MainProc; // PE to run it on
2342 RELEASE_LOCK(&sched_mutex);
2345 RELEASE_LOCK(&sched_mutex);
2347 ASSERT(m->stat != NoStatus);
2352 #if defined(RTS_SUPPORTS_THREADS)
2353 closeCondition(&m->wakeup);
2356 IF_DEBUG(scheduler, fprintf(stderr, "== scheduler: main thread (%d) finished\n",
2360 // Postcondition: sched_mutex still held
2364 //@node Run queue code, Garbage Collextion Routines, Suspend and Resume, Main scheduling code
2365 //@subsection Run queue code
2369 NB: In GranSim we have many run queues; run_queue_hd is actually a macro
2370 unfolding to run_queue_hds[CurrentProc], thus CurrentProc is an
2371 implicit global variable that has to be correct when calling these
2375 /* Put the new thread on the head of the runnable queue.
2376 * The caller of createThread better push an appropriate closure
2377 * on this thread's stack before the scheduler is invoked.
2379 static /* inline */ void
2380 add_to_run_queue(tso)
2383 ASSERT(tso!=run_queue_hd && tso!=run_queue_tl);
2384 tso->link = run_queue_hd;
2386 if (run_queue_tl == END_TSO_QUEUE) {
2391 /* Put the new thread at the end of the runnable queue. */
2392 static /* inline */ void
2393 push_on_run_queue(tso)
2396 ASSERT(get_itbl((StgClosure *)tso)->type == TSO);
2397 ASSERT(run_queue_hd!=NULL && run_queue_tl!=NULL);
2398 ASSERT(tso!=run_queue_hd && tso!=run_queue_tl);
2399 if (run_queue_hd == END_TSO_QUEUE) {
2402 run_queue_tl->link = tso;
2408 Should be inlined because it's used very often in schedule. The tso
2409 argument is actually only needed in GranSim, where we want to have the
2410 possibility to schedule *any* TSO on the run queue, irrespective of the
2411 actual ordering. Therefore, if tso is not the nil TSO then we traverse
2412 the run queue and dequeue the tso, adjusting the links in the queue.
2414 //@cindex take_off_run_queue
2415 static /* inline */ StgTSO*
2416 take_off_run_queue(StgTSO *tso) {
2420 qetlaHbogh Qu' ngaSbogh ghomDaQ {tso} yIteq!
2422 if tso is specified, unlink that tso from the run_queue (doesn't have
2423 to be at the beginning of the queue); GranSim only
2425 if (tso!=END_TSO_QUEUE) {
2426 /* find tso in queue */
2427 for (t=run_queue_hd, prev=END_TSO_QUEUE;
2428 t!=END_TSO_QUEUE && t!=tso;
2432 /* now actually dequeue the tso */
2433 if (prev!=END_TSO_QUEUE) {
2434 ASSERT(run_queue_hd!=t);
2435 prev->link = t->link;
2437 /* t is at beginning of thread queue */
2438 ASSERT(run_queue_hd==t);
2439 run_queue_hd = t->link;
2441 /* t is at end of thread queue */
2442 if (t->link==END_TSO_QUEUE) {
2443 ASSERT(t==run_queue_tl);
2444 run_queue_tl = prev;
2446 ASSERT(run_queue_tl!=t);
2448 t->link = END_TSO_QUEUE;
2450 /* take tso from the beginning of the queue; std concurrent code */
2452 if (t != END_TSO_QUEUE) {
2453 run_queue_hd = t->link;
2454 t->link = END_TSO_QUEUE;
2455 if (run_queue_hd == END_TSO_QUEUE) {
2456 run_queue_tl = END_TSO_QUEUE;
2465 //@node Garbage Collextion Routines, Blocking Queue Routines, Run queue code, Main scheduling code
2466 //@subsection Garbage Collextion Routines
2468 /* ---------------------------------------------------------------------------
2469 Where are the roots that we know about?
2471 - all the threads on the runnable queue
2472 - all the threads on the blocked queue
2473 - all the threads on the sleeping queue
2474 - all the thread currently executing a _ccall_GC
2475 - all the "main threads"
2477 ------------------------------------------------------------------------ */
2479 /* This has to be protected either by the scheduler monitor, or by the
2480 garbage collection monitor (probably the latter).
2485 GetRoots(evac_fn evac)
2490 for (i=0; i<=RtsFlags.GranFlags.proc; i++) {
2491 if ((run_queue_hds[i] != END_TSO_QUEUE) && ((run_queue_hds[i] != NULL)))
2492 evac((StgClosure **)&run_queue_hds[i]);
2493 if ((run_queue_tls[i] != END_TSO_QUEUE) && ((run_queue_tls[i] != NULL)))
2494 evac((StgClosure **)&run_queue_tls[i]);
2496 if ((blocked_queue_hds[i] != END_TSO_QUEUE) && ((blocked_queue_hds[i] != NULL)))
2497 evac((StgClosure **)&blocked_queue_hds[i]);
2498 if ((blocked_queue_tls[i] != END_TSO_QUEUE) && ((blocked_queue_tls[i] != NULL)))
2499 evac((StgClosure **)&blocked_queue_tls[i]);
2500 if ((ccalling_threadss[i] != END_TSO_QUEUE) && ((ccalling_threadss[i] != NULL)))
2501 evac((StgClosure **)&ccalling_threads[i]);
2508 if (run_queue_hd != END_TSO_QUEUE) {
2509 ASSERT(run_queue_tl != END_TSO_QUEUE);
2510 evac((StgClosure **)&run_queue_hd);
2511 evac((StgClosure **)&run_queue_tl);
2514 if (blocked_queue_hd != END_TSO_QUEUE) {
2515 ASSERT(blocked_queue_tl != END_TSO_QUEUE);
2516 evac((StgClosure **)&blocked_queue_hd);
2517 evac((StgClosure **)&blocked_queue_tl);
2520 if (sleeping_queue != END_TSO_QUEUE) {
2521 evac((StgClosure **)&sleeping_queue);
2525 if (suspended_ccalling_threads != END_TSO_QUEUE) {
2526 evac((StgClosure **)&suspended_ccalling_threads);
2529 #if defined(PAR) || defined(GRAN)
2530 markSparkQueue(evac);
2533 #if defined(RTS_USER_SIGNALS)
2534 // mark the signal handlers (signals should be already blocked)
2535 markSignalHandlers(evac);
2538 // main threads which have completed need to be retained until they
2539 // are dealt with in the main scheduler loop. They won't be
2540 // retained any other way: the GC will drop them from the
2541 // all_threads list, so we have to be careful to treat them as roots
2545 for (m = main_threads; m != NULL; m = m->link) {
2546 switch (m->tso->what_next) {
2547 case ThreadComplete:
2549 evac((StgClosure **)&m->tso);
2558 /* -----------------------------------------------------------------------------
2561 This is the interface to the garbage collector from Haskell land.
2562 We provide this so that external C code can allocate and garbage
2563 collect when called from Haskell via _ccall_GC.
2565 It might be useful to provide an interface whereby the programmer
2566 can specify more roots (ToDo).
2568 This needs to be protected by the GC condition variable above. KH.
2569 -------------------------------------------------------------------------- */
2571 static void (*extra_roots)(evac_fn);
2576 /* Obligated to hold this lock upon entry */
2577 ACQUIRE_LOCK(&sched_mutex);
2578 GarbageCollect(GetRoots,rtsFalse);
2579 RELEASE_LOCK(&sched_mutex);
2583 performMajorGC(void)
2585 ACQUIRE_LOCK(&sched_mutex);
2586 GarbageCollect(GetRoots,rtsTrue);
2587 RELEASE_LOCK(&sched_mutex);
2591 AllRoots(evac_fn evac)
2593 GetRoots(evac); // the scheduler's roots
2594 extra_roots(evac); // the user's roots
2598 performGCWithRoots(void (*get_roots)(evac_fn))
2600 ACQUIRE_LOCK(&sched_mutex);
2601 extra_roots = get_roots;
2602 GarbageCollect(AllRoots,rtsFalse);
2603 RELEASE_LOCK(&sched_mutex);
2606 /* -----------------------------------------------------------------------------
2609 If the thread has reached its maximum stack size, then raise the
2610 StackOverflow exception in the offending thread. Otherwise
2611 relocate the TSO into a larger chunk of memory and adjust its stack
2613 -------------------------------------------------------------------------- */
2616 threadStackOverflow(StgTSO *tso)
2618 nat new_stack_size, new_tso_size, stack_words;
2622 IF_DEBUG(sanity,checkTSO(tso));
2623 if (tso->stack_size >= tso->max_stack_size) {
2626 belch("@@ threadStackOverflow of TSO %d (%p): stack too large (now %ld; max is %ld",
2627 tso->id, tso, tso->stack_size, tso->max_stack_size);
2628 /* If we're debugging, just print out the top of the stack */
2629 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2632 /* Send this thread the StackOverflow exception */
2633 raiseAsync(tso, (StgClosure *)stackOverflow_closure);
2637 /* Try to double the current stack size. If that takes us over the
2638 * maximum stack size for this thread, then use the maximum instead.
2639 * Finally round up so the TSO ends up as a whole number of blocks.
2641 new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
2642 new_tso_size = (nat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) +
2643 TSO_STRUCT_SIZE)/sizeof(W_);
2644 new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */
2645 new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
2647 IF_DEBUG(scheduler, fprintf(stderr,"== scheduler: increasing stack size from %d words to %d.\n", tso->stack_size, new_stack_size));
2649 dest = (StgTSO *)allocate(new_tso_size);
2650 TICK_ALLOC_TSO(new_stack_size,0);
2652 /* copy the TSO block and the old stack into the new area */
2653 memcpy(dest,tso,TSO_STRUCT_SIZE);
2654 stack_words = tso->stack + tso->stack_size - tso->sp;
2655 new_sp = (P_)dest + new_tso_size - stack_words;
2656 memcpy(new_sp, tso->sp, stack_words * sizeof(W_));
2658 /* relocate the stack pointers... */
2660 dest->stack_size = new_stack_size;
2662 /* Mark the old TSO as relocated. We have to check for relocated
2663 * TSOs in the garbage collector and any primops that deal with TSOs.
2665 * It's important to set the sp value to just beyond the end
2666 * of the stack, so we don't attempt to scavenge any part of the
2669 tso->what_next = ThreadRelocated;
2671 tso->sp = (P_)&(tso->stack[tso->stack_size]);
2672 tso->why_blocked = NotBlocked;
2673 dest->mut_link = NULL;
2675 IF_PAR_DEBUG(verbose,
2676 belch("@@ threadStackOverflow of TSO %d (now at %p): stack size increased to %ld",
2677 tso->id, tso, tso->stack_size);
2678 /* If we're debugging, just print out the top of the stack */
2679 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2682 IF_DEBUG(sanity,checkTSO(tso));
2684 IF_DEBUG(scheduler,printTSO(dest));
2690 //@node Blocking Queue Routines, Exception Handling Routines, Garbage Collextion Routines, Main scheduling code
2691 //@subsection Blocking Queue Routines
2693 /* ---------------------------------------------------------------------------
2694 Wake up a queue that was blocked on some resource.
2695 ------------------------------------------------------------------------ */
2699 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2704 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2706 /* write RESUME events to log file and
2707 update blocked and fetch time (depending on type of the orig closure) */
2708 if (RtsFlags.ParFlags.ParStats.Full) {
2709 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
2710 GR_RESUMEQ, ((StgTSO *)bqe), ((StgTSO *)bqe)->block_info.closure,
2711 0, 0 /* spark_queue_len(ADVISORY_POOL) */);
2712 if (EMPTY_RUN_QUEUE())
2713 emitSchedule = rtsTrue;
2715 switch (get_itbl(node)->type) {
2717 ((StgTSO *)bqe)->par.fetchtime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2722 ((StgTSO *)bqe)->par.blocktime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2729 barf("{unblockOneLocked}Daq Qagh: unexpected closure in blocking queue");
2736 static StgBlockingQueueElement *
2737 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2740 PEs node_loc, tso_loc;
2742 node_loc = where_is(node); // should be lifted out of loop
2743 tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2744 tso_loc = where_is((StgClosure *)tso);
2745 if (IS_LOCAL_TO(PROCS(node),tso_loc)) { // TSO is local
2746 /* !fake_fetch => TSO is on CurrentProc is same as IS_LOCAL_TO */
2747 ASSERT(CurrentProc!=node_loc || tso_loc==CurrentProc);
2748 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.lunblocktime;
2749 // insertThread(tso, node_loc);
2750 new_event(tso_loc, tso_loc, CurrentTime[CurrentProc],
2752 tso, node, (rtsSpark*)NULL);
2753 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2756 } else { // TSO is remote (actually should be FMBQ)
2757 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.mpacktime +
2758 RtsFlags.GranFlags.Costs.gunblocktime +
2759 RtsFlags.GranFlags.Costs.latency;
2760 new_event(tso_loc, CurrentProc, CurrentTime[CurrentProc],
2762 tso, node, (rtsSpark*)NULL);
2763 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2766 /* the thread-queue-overhead is accounted for in either Resume or UnblockThread */
2768 fprintf(stderr," %s TSO %d (%p) [PE %d] (block_info.closure=%p) (next=%p) ,",
2769 (node_loc==tso_loc ? "Local" : "Global"),
2770 tso->id, tso, CurrentProc, tso->block_info.closure, tso->link));
2771 tso->block_info.closure = NULL;
2772 IF_DEBUG(scheduler,belch("-- Waking up thread %ld (%p)",
2776 static StgBlockingQueueElement *
2777 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2779 StgBlockingQueueElement *next;
2781 switch (get_itbl(bqe)->type) {
2783 ASSERT(((StgTSO *)bqe)->why_blocked != NotBlocked);
2784 /* if it's a TSO just push it onto the run_queue */
2786 // ((StgTSO *)bqe)->link = END_TSO_QUEUE; // debugging?
2787 PUSH_ON_RUN_QUEUE((StgTSO *)bqe);
2789 unblockCount(bqe, node);
2790 /* reset blocking status after dumping event */
2791 ((StgTSO *)bqe)->why_blocked = NotBlocked;
2795 /* if it's a BLOCKED_FETCH put it on the PendingFetches list */
2797 bqe->link = (StgBlockingQueueElement *)PendingFetches;
2798 PendingFetches = (StgBlockedFetch *)bqe;
2802 /* can ignore this case in a non-debugging setup;
2803 see comments on RBHSave closures above */
2805 /* check that the closure is an RBHSave closure */
2806 ASSERT(get_itbl((StgClosure *)bqe) == &stg_RBH_Save_0_info ||
2807 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_1_info ||
2808 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_2_info);
2812 barf("{unblockOneLocked}Daq Qagh: Unexpected IP (%#lx; %s) in blocking queue at %#lx\n",
2813 get_itbl((StgClosure *)bqe), info_type((StgClosure *)bqe),
2817 IF_PAR_DEBUG(bq, fprintf(stderr, ", %p (%s)", bqe, info_type((StgClosure*)bqe)));
2821 #else /* !GRAN && !PAR */
2823 unblockOneLocked(StgTSO *tso)
2827 ASSERT(get_itbl(tso)->type == TSO);
2828 ASSERT(tso->why_blocked != NotBlocked);
2829 tso->why_blocked = NotBlocked;
2831 PUSH_ON_RUN_QUEUE(tso);
2833 IF_DEBUG(scheduler,sched_belch("waking up thread %ld", tso->id));
2838 #if defined(GRAN) || defined(PAR)
2839 inline StgBlockingQueueElement *
2840 unblockOne(StgBlockingQueueElement *bqe, StgClosure *node)
2842 ACQUIRE_LOCK(&sched_mutex);
2843 bqe = unblockOneLocked(bqe, node);
2844 RELEASE_LOCK(&sched_mutex);
2849 unblockOne(StgTSO *tso)
2851 ACQUIRE_LOCK(&sched_mutex);
2852 tso = unblockOneLocked(tso);
2853 RELEASE_LOCK(&sched_mutex);
2860 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2862 StgBlockingQueueElement *bqe;
2867 belch("##-_ AwBQ for node %p on PE %d @ %ld by TSO %d (%p): ", \
2868 node, CurrentProc, CurrentTime[CurrentProc],
2869 CurrentTSO->id, CurrentTSO));
2871 node_loc = where_is(node);
2873 ASSERT(q == END_BQ_QUEUE ||
2874 get_itbl(q)->type == TSO || // q is either a TSO or an RBHSave
2875 get_itbl(q)->type == CONSTR); // closure (type constructor)
2876 ASSERT(is_unique(node));
2878 /* FAKE FETCH: magically copy the node to the tso's proc;
2879 no Fetch necessary because in reality the node should not have been
2880 moved to the other PE in the first place
2882 if (CurrentProc!=node_loc) {
2884 belch("## node %p is on PE %d but CurrentProc is %d (TSO %d); assuming fake fetch and adjusting bitmask (old: %#x)",
2885 node, node_loc, CurrentProc, CurrentTSO->id,
2886 // CurrentTSO, where_is(CurrentTSO),
2887 node->header.gran.procs));
2888 node->header.gran.procs = (node->header.gran.procs) | PE_NUMBER(CurrentProc);
2890 belch("## new bitmask of node %p is %#x",
2891 node, node->header.gran.procs));
2892 if (RtsFlags.GranFlags.GranSimStats.Global) {
2893 globalGranStats.tot_fake_fetches++;
2898 // ToDo: check: ASSERT(CurrentProc==node_loc);
2899 while (get_itbl(bqe)->type==TSO) { // q != END_TSO_QUEUE) {
2902 bqe points to the current element in the queue
2903 next points to the next element in the queue
2905 //tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2906 //tso_loc = where_is(tso);
2908 bqe = unblockOneLocked(bqe, node);
2911 /* if this is the BQ of an RBH, we have to put back the info ripped out of
2912 the closure to make room for the anchor of the BQ */
2913 if (bqe!=END_BQ_QUEUE) {
2914 ASSERT(get_itbl(node)->type == RBH && get_itbl(bqe)->type == CONSTR);
2916 ASSERT((info_ptr==&RBH_Save_0_info) ||
2917 (info_ptr==&RBH_Save_1_info) ||
2918 (info_ptr==&RBH_Save_2_info));
2920 /* cf. convertToRBH in RBH.c for writing the RBHSave closure */
2921 ((StgRBH *)node)->blocking_queue = (StgBlockingQueueElement *)((StgRBHSave *)bqe)->payload[0];
2922 ((StgRBH *)node)->mut_link = (StgMutClosure *)((StgRBHSave *)bqe)->payload[1];
2925 belch("## Filled in RBH_Save for %p (%s) at end of AwBQ",
2926 node, info_type(node)));
2929 /* statistics gathering */
2930 if (RtsFlags.GranFlags.GranSimStats.Global) {
2931 // globalGranStats.tot_bq_processing_time += bq_processing_time;
2932 globalGranStats.tot_bq_len += len; // total length of all bqs awakened
2933 // globalGranStats.tot_bq_len_local += len_local; // same for local TSOs only
2934 globalGranStats.tot_awbq++; // total no. of bqs awakened
2937 fprintf(stderr,"## BQ Stats of %p: [%d entries] %s\n",
2938 node, len, (bqe!=END_BQ_QUEUE) ? "RBH" : ""));
2942 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2944 StgBlockingQueueElement *bqe;
2946 ACQUIRE_LOCK(&sched_mutex);
2948 IF_PAR_DEBUG(verbose,
2949 belch("##-_ AwBQ for node %p on [%x]: ",
2953 if(get_itbl(q)->type == CONSTR || q==END_BQ_QUEUE) {
2954 IF_PAR_DEBUG(verbose, belch("## ... nothing to unblock so lets just return. RFP (BUG?)"));
2959 ASSERT(q == END_BQ_QUEUE ||
2960 get_itbl(q)->type == TSO ||
2961 get_itbl(q)->type == BLOCKED_FETCH ||
2962 get_itbl(q)->type == CONSTR);
2965 while (get_itbl(bqe)->type==TSO ||
2966 get_itbl(bqe)->type==BLOCKED_FETCH) {
2967 bqe = unblockOneLocked(bqe, node);
2969 RELEASE_LOCK(&sched_mutex);
2972 #else /* !GRAN && !PAR */
2974 #ifdef RTS_SUPPORTS_THREADS
2976 awakenBlockedQueueNoLock(StgTSO *tso)
2978 while (tso != END_TSO_QUEUE) {
2979 tso = unblockOneLocked(tso);
2985 awakenBlockedQueue(StgTSO *tso)
2987 ACQUIRE_LOCK(&sched_mutex);
2988 while (tso != END_TSO_QUEUE) {
2989 tso = unblockOneLocked(tso);
2991 RELEASE_LOCK(&sched_mutex);
2995 //@node Exception Handling Routines, Debugging Routines, Blocking Queue Routines, Main scheduling code
2996 //@subsection Exception Handling Routines
2998 /* ---------------------------------------------------------------------------
3000 - usually called inside a signal handler so it mustn't do anything fancy.
3001 ------------------------------------------------------------------------ */
3004 interruptStgRts(void)
3010 /* -----------------------------------------------------------------------------
3013 This is for use when we raise an exception in another thread, which
3015 This has nothing to do with the UnblockThread event in GranSim. -- HWL
3016 -------------------------------------------------------------------------- */
3018 #if defined(GRAN) || defined(PAR)
3020 NB: only the type of the blocking queue is different in GranSim and GUM
3021 the operations on the queue-elements are the same
3022 long live polymorphism!
3024 Locks: sched_mutex is held upon entry and exit.
3028 unblockThread(StgTSO *tso)
3030 StgBlockingQueueElement *t, **last;
3032 switch (tso->why_blocked) {
3035 return; /* not blocked */
3038 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
3040 StgBlockingQueueElement *last_tso = END_BQ_QUEUE;
3041 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
3043 last = (StgBlockingQueueElement **)&mvar->head;
3044 for (t = (StgBlockingQueueElement *)mvar->head;
3046 last = &t->link, last_tso = t, t = t->link) {
3047 if (t == (StgBlockingQueueElement *)tso) {
3048 *last = (StgBlockingQueueElement *)tso->link;
3049 if (mvar->tail == tso) {
3050 mvar->tail = (StgTSO *)last_tso;
3055 barf("unblockThread (MVAR): TSO not found");
3058 case BlockedOnBlackHole:
3059 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
3061 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
3063 last = &bq->blocking_queue;
3064 for (t = bq->blocking_queue;
3066 last = &t->link, t = t->link) {
3067 if (t == (StgBlockingQueueElement *)tso) {
3068 *last = (StgBlockingQueueElement *)tso->link;
3072 barf("unblockThread (BLACKHOLE): TSO not found");
3075 case BlockedOnException:
3077 StgTSO *target = tso->block_info.tso;
3079 ASSERT(get_itbl(target)->type == TSO);
3081 if (target->what_next == ThreadRelocated) {
3082 target = target->link;
3083 ASSERT(get_itbl(target)->type == TSO);
3086 ASSERT(target->blocked_exceptions != NULL);
3088 last = (StgBlockingQueueElement **)&target->blocked_exceptions;
3089 for (t = (StgBlockingQueueElement *)target->blocked_exceptions;
3091 last = &t->link, t = t->link) {
3092 ASSERT(get_itbl(t)->type == TSO);
3093 if (t == (StgBlockingQueueElement *)tso) {
3094 *last = (StgBlockingQueueElement *)tso->link;
3098 barf("unblockThread (Exception): TSO not found");
3102 case BlockedOnWrite:
3104 /* take TSO off blocked_queue */
3105 StgBlockingQueueElement *prev = NULL;
3106 for (t = (StgBlockingQueueElement *)blocked_queue_hd; t != END_BQ_QUEUE;
3107 prev = t, t = t->link) {
3108 if (t == (StgBlockingQueueElement *)tso) {
3110 blocked_queue_hd = (StgTSO *)t->link;
3111 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
3112 blocked_queue_tl = END_TSO_QUEUE;
3115 prev->link = t->link;
3116 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
3117 blocked_queue_tl = (StgTSO *)prev;
3123 barf("unblockThread (I/O): TSO not found");
3126 case BlockedOnDelay:
3128 /* take TSO off sleeping_queue */
3129 StgBlockingQueueElement *prev = NULL;
3130 for (t = (StgBlockingQueueElement *)sleeping_queue; t != END_BQ_QUEUE;
3131 prev = t, t = t->link) {
3132 if (t == (StgBlockingQueueElement *)tso) {
3134 sleeping_queue = (StgTSO *)t->link;
3136 prev->link = t->link;
3141 barf("unblockThread (I/O): TSO not found");
3145 barf("unblockThread");
3149 tso->link = END_TSO_QUEUE;
3150 tso->why_blocked = NotBlocked;
3151 tso->block_info.closure = NULL;
3152 PUSH_ON_RUN_QUEUE(tso);
3156 unblockThread(StgTSO *tso)
3160 /* To avoid locking unnecessarily. */
3161 if (tso->why_blocked == NotBlocked) {
3165 switch (tso->why_blocked) {
3168 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
3170 StgTSO *last_tso = END_TSO_QUEUE;
3171 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
3174 for (t = mvar->head; t != END_TSO_QUEUE;
3175 last = &t->link, last_tso = t, t = t->link) {
3178 if (mvar->tail == tso) {
3179 mvar->tail = last_tso;
3184 barf("unblockThread (MVAR): TSO not found");
3187 case BlockedOnBlackHole:
3188 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
3190 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
3192 last = &bq->blocking_queue;
3193 for (t = bq->blocking_queue; t != END_TSO_QUEUE;
3194 last = &t->link, t = t->link) {
3200 barf("unblockThread (BLACKHOLE): TSO not found");
3203 case BlockedOnException:
3205 StgTSO *target = tso->block_info.tso;
3207 ASSERT(get_itbl(target)->type == TSO);
3209 while (target->what_next == ThreadRelocated) {
3210 target = target->link;
3211 ASSERT(get_itbl(target)->type == TSO);
3214 ASSERT(target->blocked_exceptions != NULL);
3216 last = &target->blocked_exceptions;
3217 for (t = target->blocked_exceptions; t != END_TSO_QUEUE;
3218 last = &t->link, t = t->link) {
3219 ASSERT(get_itbl(t)->type == TSO);
3225 barf("unblockThread (Exception): TSO not found");
3229 case BlockedOnWrite:
3231 StgTSO *prev = NULL;
3232 for (t = blocked_queue_hd; t != END_TSO_QUEUE;
3233 prev = t, t = t->link) {
3236 blocked_queue_hd = t->link;
3237 if (blocked_queue_tl == t) {
3238 blocked_queue_tl = END_TSO_QUEUE;
3241 prev->link = t->link;
3242 if (blocked_queue_tl == t) {
3243 blocked_queue_tl = prev;
3249 barf("unblockThread (I/O): TSO not found");
3252 case BlockedOnDelay:
3254 StgTSO *prev = NULL;
3255 for (t = sleeping_queue; t != END_TSO_QUEUE;
3256 prev = t, t = t->link) {
3259 sleeping_queue = t->link;
3261 prev->link = t->link;
3266 barf("unblockThread (I/O): TSO not found");
3270 barf("unblockThread");
3274 tso->link = END_TSO_QUEUE;
3275 tso->why_blocked = NotBlocked;
3276 tso->block_info.closure = NULL;
3277 PUSH_ON_RUN_QUEUE(tso);
3281 /* -----------------------------------------------------------------------------
3284 * The following function implements the magic for raising an
3285 * asynchronous exception in an existing thread.
3287 * We first remove the thread from any queue on which it might be
3288 * blocked. The possible blockages are MVARs and BLACKHOLE_BQs.
3290 * We strip the stack down to the innermost CATCH_FRAME, building
3291 * thunks in the heap for all the active computations, so they can
3292 * be restarted if necessary. When we reach a CATCH_FRAME, we build
3293 * an application of the handler to the exception, and push it on
3294 * the top of the stack.
3296 * How exactly do we save all the active computations? We create an
3297 * AP_STACK for every UpdateFrame on the stack. Entering one of these
3298 * AP_STACKs pushes everything from the corresponding update frame
3299 * upwards onto the stack. (Actually, it pushes everything up to the
3300 * next update frame plus a pointer to the next AP_STACK object.
3301 * Entering the next AP_STACK object pushes more onto the stack until we
3302 * reach the last AP_STACK object - at which point the stack should look
3303 * exactly as it did when we killed the TSO and we can continue
3304 * execution by entering the closure on top of the stack.
3306 * We can also kill a thread entirely - this happens if either (a) the
3307 * exception passed to raiseAsync is NULL, or (b) there's no
3308 * CATCH_FRAME on the stack. In either case, we strip the entire
3309 * stack and replace the thread with a zombie.
3311 * Locks: sched_mutex held upon entry nor exit.
3313 * -------------------------------------------------------------------------- */
3316 deleteThread(StgTSO *tso)
3318 raiseAsync(tso,NULL);
3322 raiseAsyncWithLock(StgTSO *tso, StgClosure *exception)
3324 /* When raising async exs from contexts where sched_mutex isn't held;
3325 use raiseAsyncWithLock(). */
3326 ACQUIRE_LOCK(&sched_mutex);
3327 raiseAsync(tso,exception);
3328 RELEASE_LOCK(&sched_mutex);
3332 raiseAsync(StgTSO *tso, StgClosure *exception)
3334 StgRetInfoTable *info;
3337 // Thread already dead?
3338 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
3343 sched_belch("raising exception in thread %ld.", tso->id));
3345 // Remove it from any blocking queues
3350 // The stack freezing code assumes there's a closure pointer on
3351 // the top of the stack, so we have to arrange that this is the case...
3353 if (sp[0] == (W_)&stg_enter_info) {
3357 sp[0] = (W_)&stg_dummy_ret_closure;
3363 // 1. Let the top of the stack be the "current closure"
3365 // 2. Walk up the stack until we find either an UPDATE_FRAME or a
3368 // 3. If it's an UPDATE_FRAME, then make an AP_STACK containing the
3369 // current closure applied to the chunk of stack up to (but not
3370 // including) the update frame. This closure becomes the "current
3371 // closure". Go back to step 2.
3373 // 4. If it's a CATCH_FRAME, then leave the exception handler on
3374 // top of the stack applied to the exception.
3376 // 5. If it's a STOP_FRAME, then kill the thread.
3381 info = get_ret_itbl((StgClosure *)frame);
3383 while (info->i.type != UPDATE_FRAME
3384 && (info->i.type != CATCH_FRAME || exception == NULL)
3385 && info->i.type != STOP_FRAME) {
3386 frame += stack_frame_sizeW((StgClosure *)frame);
3387 info = get_ret_itbl((StgClosure *)frame);
3390 switch (info->i.type) {
3393 // If we find a CATCH_FRAME, and we've got an exception to raise,
3394 // then build the THUNK raise(exception), and leave it on
3395 // top of the CATCH_FRAME ready to enter.
3399 StgCatchFrame *cf = (StgCatchFrame *)frame;
3403 // we've got an exception to raise, so let's pass it to the
3404 // handler in this frame.
3406 raise = (StgClosure *)allocate(sizeofW(StgClosure)+1);
3407 TICK_ALLOC_SE_THK(1,0);
3408 SET_HDR(raise,&stg_raise_info,cf->header.prof.ccs);
3409 raise->payload[0] = exception;
3411 // throw away the stack from Sp up to the CATCH_FRAME.
3415 /* Ensure that async excpetions are blocked now, so we don't get
3416 * a surprise exception before we get around to executing the
3419 if (tso->blocked_exceptions == NULL) {
3420 tso->blocked_exceptions = END_TSO_QUEUE;
3423 /* Put the newly-built THUNK on top of the stack, ready to execute
3424 * when the thread restarts.
3427 sp[-1] = (W_)&stg_enter_info;
3429 tso->what_next = ThreadRunGHC;
3430 IF_DEBUG(sanity, checkTSO(tso));
3439 // First build an AP_STACK consisting of the stack chunk above the
3440 // current update frame, with the top word on the stack as the
3443 words = frame - sp - 1;
3444 ap = (StgAP_STACK *)allocate(PAP_sizeW(words));
3447 ap->fun = (StgClosure *)sp[0];
3449 for(i=0; i < (nat)words; ++i) {
3450 ap->payload[i] = (StgClosure *)*sp++;
3453 SET_HDR(ap,&stg_AP_STACK_info,
3454 ((StgClosure *)frame)->header.prof.ccs /* ToDo */);
3455 TICK_ALLOC_UP_THK(words+1,0);
3458 fprintf(stderr, "scheduler: Updating ");
3459 printPtr((P_)((StgUpdateFrame *)frame)->updatee);
3460 fprintf(stderr, " with ");
3461 printObj((StgClosure *)ap);
3464 // Replace the updatee with an indirection - happily
3465 // this will also wake up any threads currently
3466 // waiting on the result.
3468 // Warning: if we're in a loop, more than one update frame on
3469 // the stack may point to the same object. Be careful not to
3470 // overwrite an IND_OLDGEN in this case, because we'll screw
3471 // up the mutable lists. To be on the safe side, don't
3472 // overwrite any kind of indirection at all. See also
3473 // threadSqueezeStack in GC.c, where we have to make a similar
3476 if (!closure_IND(((StgUpdateFrame *)frame)->updatee)) {
3477 // revert the black hole
3478 UPD_IND_NOLOCK(((StgUpdateFrame *)frame)->updatee,ap);
3480 sp += sizeofW(StgUpdateFrame) - 1;
3481 sp[0] = (W_)ap; // push onto stack
3486 // We've stripped the entire stack, the thread is now dead.
3487 sp += sizeofW(StgStopFrame);
3488 tso->what_next = ThreadKilled;
3499 /* -----------------------------------------------------------------------------
3500 resurrectThreads is called after garbage collection on the list of
3501 threads found to be garbage. Each of these threads will be woken
3502 up and sent a signal: BlockedOnDeadMVar if the thread was blocked
3503 on an MVar, or NonTermination if the thread was blocked on a Black
3506 Locks: sched_mutex isn't held upon entry nor exit.
3507 -------------------------------------------------------------------------- */
3510 resurrectThreads( StgTSO *threads )
3514 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
3515 next = tso->global_link;
3516 tso->global_link = all_threads;
3518 IF_DEBUG(scheduler, sched_belch("resurrecting thread %d", tso->id));
3520 switch (tso->why_blocked) {
3522 case BlockedOnException:
3523 /* Called by GC - sched_mutex lock is currently held. */
3524 raiseAsync(tso,(StgClosure *)BlockedOnDeadMVar_closure);
3526 case BlockedOnBlackHole:
3527 raiseAsync(tso,(StgClosure *)NonTermination_closure);
3530 /* This might happen if the thread was blocked on a black hole
3531 * belonging to a thread that we've just woken up (raiseAsync
3532 * can wake up threads, remember...).
3536 barf("resurrectThreads: thread blocked in a strange way");
3541 /* -----------------------------------------------------------------------------
3542 * Blackhole detection: if we reach a deadlock, test whether any
3543 * threads are blocked on themselves. Any threads which are found to
3544 * be self-blocked get sent a NonTermination exception.
3546 * This is only done in a deadlock situation in order to avoid
3547 * performance overhead in the normal case.
3549 * Locks: sched_mutex is held upon entry and exit.
3550 * -------------------------------------------------------------------------- */
3553 detectBlackHoles( void )
3555 StgTSO *tso = all_threads;
3557 StgClosure *blocked_on;
3558 StgRetInfoTable *info;
3560 for (tso = all_threads; tso != END_TSO_QUEUE; tso = tso->global_link) {
3562 while (tso->what_next == ThreadRelocated) {
3564 ASSERT(get_itbl(tso)->type == TSO);
3567 if (tso->why_blocked != BlockedOnBlackHole) {
3570 blocked_on = tso->block_info.closure;
3572 frame = (StgClosure *)tso->sp;
3575 info = get_ret_itbl(frame);
3576 switch (info->i.type) {
3578 if (((StgUpdateFrame *)frame)->updatee == blocked_on) {
3579 /* We are blocking on one of our own computations, so
3580 * send this thread the NonTermination exception.
3583 sched_belch("thread %d is blocked on itself", tso->id));
3584 raiseAsync(tso, (StgClosure *)NonTermination_closure);
3588 frame = (StgClosure *) ((StgUpdateFrame *)frame + 1);
3594 // normal stack frames; do nothing except advance the pointer
3596 (StgPtr)frame += stack_frame_sizeW(frame);
3603 //@node Debugging Routines, Index, Exception Handling Routines, Main scheduling code
3604 //@subsection Debugging Routines
3606 /* -----------------------------------------------------------------------------
3607 * Debugging: why is a thread blocked
3608 * [Also provides useful information when debugging threaded programs
3609 * at the Haskell source code level, so enable outside of DEBUG. --sof 7/02]
3610 -------------------------------------------------------------------------- */
3614 printThreadBlockage(StgTSO *tso)
3616 switch (tso->why_blocked) {
3618 fprintf(stderr,"is blocked on read from fd %d", tso->block_info.fd);
3620 case BlockedOnWrite:
3621 fprintf(stderr,"is blocked on write to fd %d", tso->block_info.fd);
3623 case BlockedOnDelay:
3624 fprintf(stderr,"is blocked until %d", tso->block_info.target);
3627 fprintf(stderr,"is blocked on an MVar");
3629 case BlockedOnException:
3630 fprintf(stderr,"is blocked on delivering an exception to thread %d",
3631 tso->block_info.tso->id);
3633 case BlockedOnBlackHole:
3634 fprintf(stderr,"is blocked on a black hole");
3637 fprintf(stderr,"is not blocked");
3641 fprintf(stderr,"is blocked on global address; local FM_BQ is %p (%s)",
3642 tso->block_info.closure, info_type(tso->block_info.closure));
3644 case BlockedOnGA_NoSend:
3645 fprintf(stderr,"is blocked on global address (no send); local FM_BQ is %p (%s)",
3646 tso->block_info.closure, info_type(tso->block_info.closure));
3649 #if defined(RTS_SUPPORTS_THREADS)
3650 case BlockedOnCCall:
3651 fprintf(stderr,"is blocked on an external call");
3653 case BlockedOnCCall_NoUnblockExc:
3654 fprintf(stderr,"is blocked on an external call (exceptions were already blocked)");
3658 barf("printThreadBlockage: strange tso->why_blocked: %d for TSO %d (%d)",
3659 tso->why_blocked, tso->id, tso);
3665 printThreadStatus(StgTSO *tso)
3667 switch (tso->what_next) {
3669 fprintf(stderr,"has been killed");
3671 case ThreadComplete:
3672 fprintf(stderr,"has completed");
3675 printThreadBlockage(tso);
3680 printAllThreads(void)
3686 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
3687 ullong_format_string(TIME_ON_PROC(CurrentProc),
3688 time_string, rtsFalse/*no commas!*/);
3690 fprintf(stderr, "all threads at [%s]:\n", time_string);
3692 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
3693 ullong_format_string(CURRENT_TIME,
3694 time_string, rtsFalse/*no commas!*/);
3696 fprintf(stderr,"all threads at [%s]:\n", time_string);
3698 fprintf(stderr,"all threads:\n");
3701 for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
3702 fprintf(stderr, "\tthread %d @ %p ", t->id, (void *)t);
3703 label = lookupThreadLabel((StgWord)t);
3704 if (label) fprintf(stderr,"[\"%s\"] ",(char *)label);
3705 printThreadStatus(t);
3706 fprintf(stderr,"\n");
3713 Print a whole blocking queue attached to node (debugging only).
3718 print_bq (StgClosure *node)
3720 StgBlockingQueueElement *bqe;
3724 fprintf(stderr,"## BQ of closure %p (%s): ",
3725 node, info_type(node));
3727 /* should cover all closures that may have a blocking queue */
3728 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
3729 get_itbl(node)->type == FETCH_ME_BQ ||
3730 get_itbl(node)->type == RBH ||
3731 get_itbl(node)->type == MVAR);
3733 ASSERT(node!=(StgClosure*)NULL); // sanity check
3735 print_bqe(((StgBlockingQueue*)node)->blocking_queue);
3739 Print a whole blocking queue starting with the element bqe.
3742 print_bqe (StgBlockingQueueElement *bqe)
3747 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
3749 for (end = (bqe==END_BQ_QUEUE);
3750 !end; // iterate until bqe points to a CONSTR
3751 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE),
3752 bqe = end ? END_BQ_QUEUE : bqe->link) {
3753 ASSERT(bqe != END_BQ_QUEUE); // sanity check
3754 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
3755 /* types of closures that may appear in a blocking queue */
3756 ASSERT(get_itbl(bqe)->type == TSO ||
3757 get_itbl(bqe)->type == BLOCKED_FETCH ||
3758 get_itbl(bqe)->type == CONSTR);
3759 /* only BQs of an RBH end with an RBH_Save closure */
3760 //ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
3762 switch (get_itbl(bqe)->type) {
3764 fprintf(stderr," TSO %u (%x),",
3765 ((StgTSO *)bqe)->id, ((StgTSO *)bqe));
3768 fprintf(stderr," BF (node=%p, ga=((%x, %d, %x)),",
3769 ((StgBlockedFetch *)bqe)->node,
3770 ((StgBlockedFetch *)bqe)->ga.payload.gc.gtid,
3771 ((StgBlockedFetch *)bqe)->ga.payload.gc.slot,
3772 ((StgBlockedFetch *)bqe)->ga.weight);
3775 fprintf(stderr," %s (IP %p),",
3776 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
3777 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
3778 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
3779 "RBH_Save_?"), get_itbl(bqe));
3782 barf("Unexpected closure type %s in blocking queue", // of %p (%s)",
3783 info_type((StgClosure *)bqe)); // , node, info_type(node));
3787 fputc('\n', stderr);
3789 # elif defined(GRAN)
3791 print_bq (StgClosure *node)
3793 StgBlockingQueueElement *bqe;
3794 PEs node_loc, tso_loc;
3797 /* should cover all closures that may have a blocking queue */
3798 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
3799 get_itbl(node)->type == FETCH_ME_BQ ||
3800 get_itbl(node)->type == RBH);
3802 ASSERT(node!=(StgClosure*)NULL); // sanity check
3803 node_loc = where_is(node);
3805 fprintf(stderr,"## BQ of closure %p (%s) on [PE %d]: ",
3806 node, info_type(node), node_loc);
3809 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
3811 for (bqe = ((StgBlockingQueue*)node)->blocking_queue, end = (bqe==END_BQ_QUEUE);
3812 !end; // iterate until bqe points to a CONSTR
3813 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) {
3814 ASSERT(bqe != END_BQ_QUEUE); // sanity check
3815 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
3816 /* types of closures that may appear in a blocking queue */
3817 ASSERT(get_itbl(bqe)->type == TSO ||
3818 get_itbl(bqe)->type == CONSTR);
3819 /* only BQs of an RBH end with an RBH_Save closure */
3820 ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
3822 tso_loc = where_is((StgClosure *)bqe);
3823 switch (get_itbl(bqe)->type) {
3825 fprintf(stderr," TSO %d (%p) on [PE %d],",
3826 ((StgTSO *)bqe)->id, (StgTSO *)bqe, tso_loc);
3829 fprintf(stderr," %s (IP %p),",
3830 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
3831 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
3832 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
3833 "RBH_Save_?"), get_itbl(bqe));
3836 barf("Unexpected closure type %s in blocking queue of %p (%s)",
3837 info_type((StgClosure *)bqe), node, info_type(node));
3841 fputc('\n', stderr);
3845 Nice and easy: only TSOs on the blocking queue
3848 print_bq (StgClosure *node)
3852 ASSERT(node!=(StgClosure*)NULL); // sanity check
3853 for (tso = ((StgBlockingQueue*)node)->blocking_queue;
3854 tso != END_TSO_QUEUE;
3856 ASSERT(tso!=NULL && tso!=END_TSO_QUEUE); // sanity check
3857 ASSERT(get_itbl(tso)->type == TSO); // guess what, sanity check
3858 fprintf(stderr," TSO %d (%p),", tso->id, tso);
3860 fputc('\n', stderr);
3871 for (i=0, tso=run_queue_hd;
3872 tso != END_TSO_QUEUE;
3881 sched_belch(char *s, ...)
3886 fprintf(stderr, "scheduler (task %ld): ", osThreadId());
3888 fprintf(stderr, "== ");
3890 fprintf(stderr, "scheduler: ");
3892 vfprintf(stderr, s, ap);
3893 fprintf(stderr, "\n");
3900 //@node Index, , Debugging Routines, Main scheduling code
3904 //* StgMainThread:: @cindex\s-+StgMainThread
3905 //* awaken_blocked_queue:: @cindex\s-+awaken_blocked_queue
3906 //* blocked_queue_hd:: @cindex\s-+blocked_queue_hd
3907 //* blocked_queue_tl:: @cindex\s-+blocked_queue_tl
3908 //* context_switch:: @cindex\s-+context_switch
3909 //* createThread:: @cindex\s-+createThread
3910 //* gc_pending_cond:: @cindex\s-+gc_pending_cond
3911 //* initScheduler:: @cindex\s-+initScheduler
3912 //* interrupted:: @cindex\s-+interrupted
3913 //* next_thread_id:: @cindex\s-+next_thread_id
3914 //* print_bq:: @cindex\s-+print_bq
3915 //* run_queue_hd:: @cindex\s-+run_queue_hd
3916 //* run_queue_tl:: @cindex\s-+run_queue_tl
3917 //* sched_mutex:: @cindex\s-+sched_mutex
3918 //* schedule:: @cindex\s-+schedule
3919 //* take_off_run_queue:: @cindex\s-+take_off_run_queue
3920 //* term_mutex:: @cindex\s-+term_mutex