1 /* ---------------------------------------------------------------------------
2 * $Id: Schedule.c,v 1.191 2004/02/27 15:58:54 simonmar Exp $
4 * (c) The GHC Team, 1998-2003
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 distrib. 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 * --------------------------------------------------------------------------*/
21 * Version with support for distributed memory parallelism aka GUM (WAY=mp):
23 The main scheduling loop in GUM iterates until a finish message is received.
24 In that case a global flag @receivedFinish@ is set and this instance of
25 the RTS shuts down. See ghc/rts/parallel/HLComms.c:processMessages()
26 for the handling of incoming messages, such as PP_FINISH.
27 Note that in the parallel case we have a system manager that coordinates
28 different PEs, each of which are running one instance of the RTS.
29 See ghc/rts/parallel/SysMan.c for the main routine of the parallel program.
30 From this routine processes executing ghc/rts/Main.c are spawned. -- HWL
32 * Version with support for simulating parallel execution aka GranSim (WAY=mg):
34 The main scheduling code in GranSim is quite different from that in std
35 (concurrent) Haskell: while concurrent Haskell just iterates over the
36 threads in the runnable queue, GranSim is event driven, i.e. it iterates
37 over the events in the global event queue. -- HWL
40 #include "PosixSource.h"
47 #include "StgStartup.h"
49 #define COMPILING_SCHEDULER
51 #include "StgMiscClosures.h"
53 #include "Interpreter.h"
54 #include "Exception.h"
61 #include "ThreadLabels.h"
63 #include "Proftimer.h"
66 #if defined(GRAN) || defined(PAR)
67 # include "GranSimRts.h"
69 # include "ParallelRts.h"
70 # include "Parallel.h"
71 # include "ParallelDebug.h"
76 #include "Capability.h"
77 #include "OSThreads.h"
80 #ifdef HAVE_SYS_TYPES_H
81 #include <sys/types.h>
96 #define USED_IN_THREADED_RTS
98 #define USED_IN_THREADED_RTS STG_UNUSED
101 #ifdef RTS_SUPPORTS_THREADS
102 #define USED_WHEN_RTS_SUPPORTS_THREADS
104 #define USED_WHEN_RTS_SUPPORTS_THREADS STG_UNUSED
107 /* Main thread queue.
108 * Locks required: sched_mutex.
110 StgMainThread *main_threads = NULL;
113 * Locks required: sched_mutex.
117 StgTSO* ActiveTSO = NULL; /* for assigning system costs; GranSim-Light only */
118 /* rtsTime TimeOfNextEvent, EndOfTimeSlice; now in GranSim.c */
121 In GranSim we have a runnable and a blocked queue for each processor.
122 In order to minimise code changes new arrays run_queue_hds/tls
123 are created. run_queue_hd is then a short cut (macro) for
124 run_queue_hds[CurrentProc] (see GranSim.h).
127 StgTSO *run_queue_hds[MAX_PROC], *run_queue_tls[MAX_PROC];
128 StgTSO *blocked_queue_hds[MAX_PROC], *blocked_queue_tls[MAX_PROC];
129 StgTSO *ccalling_threadss[MAX_PROC];
130 /* We use the same global list of threads (all_threads) in GranSim as in
131 the std RTS (i.e. we are cheating). However, we don't use this list in
132 the GranSim specific code at the moment (so we are only potentially
137 StgTSO *run_queue_hd = NULL;
138 StgTSO *run_queue_tl = NULL;
139 StgTSO *blocked_queue_hd = NULL;
140 StgTSO *blocked_queue_tl = NULL;
141 StgTSO *sleeping_queue = NULL; /* perhaps replace with a hash table? */
145 /* Linked list of all threads.
146 * Used for detecting garbage collected threads.
148 StgTSO *all_threads = NULL;
150 /* When a thread performs a safe C call (_ccall_GC, using old
151 * terminology), it gets put on the suspended_ccalling_threads
152 * list. Used by the garbage collector.
154 static StgTSO *suspended_ccalling_threads;
156 static StgTSO *threadStackOverflow(StgTSO *tso);
158 /* KH: The following two flags are shared memory locations. There is no need
159 to lock them, since they are only unset at the end of a scheduler
163 /* flag set by signal handler to precipitate a context switch */
164 nat context_switch = 0;
166 /* if this flag is set as well, give up execution */
167 rtsBool interrupted = rtsFalse;
169 /* Next thread ID to allocate.
170 * Locks required: thread_id_mutex
172 static StgThreadID next_thread_id = 1;
175 * Pointers to the state of the current thread.
176 * Rule of thumb: if CurrentTSO != NULL, then we're running a Haskell
177 * thread. If CurrentTSO == NULL, then we're at the scheduler level.
180 /* The smallest stack size that makes any sense is:
181 * RESERVED_STACK_WORDS (so we can get back from the stack overflow)
182 * + sizeofW(StgStopFrame) (the stg_stop_thread_info frame)
183 * + 1 (the closure to enter)
185 * + 1 (spare slot req'd by stg_ap_v_ret)
187 * A thread with this stack will bomb immediately with a stack
188 * overflow, which will increase its stack size.
191 #define MIN_STACK_WORDS (RESERVED_STACK_WORDS + sizeofW(StgStopFrame) + 3)
198 /* This is used in `TSO.h' and gcc 2.96 insists that this variable actually
199 * exists - earlier gccs apparently didn't.
204 static rtsBool ready_to_gc;
207 * Set to TRUE when entering a shutdown state (via shutdownHaskellAndExit()) --
208 * in an MT setting, needed to signal that a worker thread shouldn't hang around
209 * in the scheduler when it is out of work.
211 static rtsBool shutting_down_scheduler = rtsFalse;
213 void addToBlockedQueue ( StgTSO *tso );
215 static void schedule ( StgMainThread *mainThread, Capability *initialCapability );
216 void interruptStgRts ( void );
218 static void detectBlackHoles ( void );
220 #if defined(RTS_SUPPORTS_THREADS)
221 /* ToDo: carefully document the invariants that go together
222 * with these synchronisation objects.
224 Mutex sched_mutex = INIT_MUTEX_VAR;
225 Mutex term_mutex = INIT_MUTEX_VAR;
227 #endif /* RTS_SUPPORTS_THREADS */
231 rtsTime TimeOfLastYield;
232 rtsBool emitSchedule = rtsTrue;
236 static char *whatNext_strs[] = {
246 StgTSO * createSparkThread(rtsSpark spark);
247 StgTSO * activateSpark (rtsSpark spark);
250 /* ----------------------------------------------------------------------------
252 * ------------------------------------------------------------------------- */
254 #if defined(RTS_SUPPORTS_THREADS)
255 static rtsBool startingWorkerThread = rtsFalse;
257 static void taskStart(void);
261 ACQUIRE_LOCK(&sched_mutex);
263 RELEASE_LOCK(&sched_mutex);
267 startSchedulerTaskIfNecessary(void)
269 if(run_queue_hd != END_TSO_QUEUE
270 || blocked_queue_hd != END_TSO_QUEUE
271 || sleeping_queue != END_TSO_QUEUE)
273 if(!startingWorkerThread)
274 { // we don't want to start another worker thread
275 // just because the last one hasn't yet reached the
276 // "waiting for capability" state
277 startingWorkerThread = rtsTrue;
278 startTask(taskStart);
284 /* ---------------------------------------------------------------------------
285 Main scheduling loop.
287 We use round-robin scheduling, each thread returning to the
288 scheduler loop when one of these conditions is detected:
291 * timer expires (thread yields)
296 Locking notes: we acquire the scheduler lock once at the beginning
297 of the scheduler loop, and release it when
299 * running a thread, or
300 * waiting for work, or
301 * waiting for a GC to complete.
304 In a GranSim setup this loop iterates over the global event queue.
305 This revolves around the global event queue, which determines what
306 to do next. Therefore, it's more complicated than either the
307 concurrent or the parallel (GUM) setup.
310 GUM iterates over incoming messages.
311 It starts with nothing to do (thus CurrentTSO == END_TSO_QUEUE),
312 and sends out a fish whenever it has nothing to do; in-between
313 doing the actual reductions (shared code below) it processes the
314 incoming messages and deals with delayed operations
315 (see PendingFetches).
316 This is not the ugliest code you could imagine, but it's bloody close.
318 ------------------------------------------------------------------------ */
320 schedule( StgMainThread *mainThread USED_WHEN_RTS_SUPPORTS_THREADS,
321 Capability *initialCapability )
324 Capability *cap = initialCapability;
325 StgThreadReturnCode ret;
333 rtsBool receivedFinish = rtsFalse;
335 nat tp_size, sp_size; // stats only
338 rtsBool was_interrupted = rtsFalse;
339 StgTSOWhatNext prev_what_next;
341 // Pre-condition: sched_mutex is held.
343 #if defined(RTS_SUPPORTS_THREADS)
345 // in the threaded case, the capability is either passed in via the
346 // initialCapability parameter, or initialized inside the scheduler
350 sched_belch("### NEW SCHEDULER LOOP (main thr: %p, cap: %p)",
351 mainThread, initialCapability);
354 // simply initialise it in the non-threaded case
355 grabCapability(&cap);
359 /* set up first event to get things going */
360 /* ToDo: assign costs for system setup and init MainTSO ! */
361 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
363 CurrentTSO, (StgClosure*)NULL, (rtsSpark*)NULL);
366 fprintf(stderr, "GRAN: Init CurrentTSO (in schedule) = %p\n", CurrentTSO);
367 G_TSO(CurrentTSO, 5));
369 if (RtsFlags.GranFlags.Light) {
370 /* Save current time; GranSim Light only */
371 CurrentTSO->gran.clock = CurrentTime[CurrentProc];
374 event = get_next_event();
376 while (event!=(rtsEvent*)NULL) {
377 /* Choose the processor with the next event */
378 CurrentProc = event->proc;
379 CurrentTSO = event->tso;
383 while (!receivedFinish) { /* set by processMessages */
384 /* when receiving PP_FINISH message */
386 #else // everything except GRAN and PAR
392 IF_DEBUG(scheduler, printAllThreads());
394 #if defined(RTS_SUPPORTS_THREADS)
395 // Yield the capability to higher-priority tasks if necessary.
398 yieldCapability(&cap);
401 // If we do not currently hold a capability, we wait for one
404 waitForCapability(&sched_mutex, &cap,
405 mainThread ? &mainThread->bound_thread_cond : NULL);
408 // We now have a capability...
412 // If we're interrupted (the user pressed ^C, or some other
413 // termination condition occurred), kill all the currently running
417 IF_DEBUG(scheduler, sched_belch("interrupted"));
418 interrupted = rtsFalse;
419 was_interrupted = rtsTrue;
420 #if defined(RTS_SUPPORTS_THREADS)
421 // In the threaded RTS, deadlock detection doesn't work,
422 // so just exit right away.
423 prog_belch("interrupted");
424 releaseCapability(cap);
425 RELEASE_LOCK(&sched_mutex);
426 shutdownHaskellAndExit(EXIT_SUCCESS);
433 // Go through the list of main threads and wake up any
434 // clients whose computations have finished. ToDo: this
435 // should be done more efficiently without a linear scan
436 // of the main threads list, somehow...
438 #if defined(RTS_SUPPORTS_THREADS)
440 StgMainThread *m, **prev;
441 prev = &main_threads;
442 for (m = main_threads; m != NULL; prev = &m->link, m = m->link) {
443 if (m->tso->what_next == ThreadComplete
444 || m->tso->what_next == ThreadKilled)
448 if (m->tso->what_next == ThreadComplete)
452 // NOTE: return val is tso->sp[1] (see StgStartup.hc)
453 *(m->ret) = (StgClosure *)m->tso->sp[1];
465 m->stat = Interrupted;
475 removeThreadLabel((StgWord)m->tso->id);
477 releaseCapability(cap);
482 // The current OS thread can not handle the fact that
483 // the Haskell thread "m" has ended. "m" is bound;
484 // the scheduler loop in it's bound OS thread has to
485 // return, so let's pass our capability directly to
487 passCapability(&m->bound_thread_cond);
494 #else /* not threaded */
497 /* in GUM do this only on the Main PE */
500 /* If our main thread has finished or been killed, return.
503 StgMainThread *m = main_threads;
504 if (m->tso->what_next == ThreadComplete
505 || m->tso->what_next == ThreadKilled) {
507 removeThreadLabel((StgWord)m->tso->id);
509 main_threads = main_threads->link;
510 if (m->tso->what_next == ThreadComplete) {
511 // We finished successfully, fill in the return value
512 // NOTE: return val is tso->sp[1] (see StgStartup.hc)
513 if (m->ret) { *(m->ret) = (StgClosure *)m->tso->sp[1]; };
517 if (m->ret) { *(m->ret) = NULL; };
518 if (was_interrupted) {
519 m->stat = Interrupted;
530 #if defined(RTS_USER_SIGNALS)
531 // check for signals each time around the scheduler
532 if (signals_pending()) {
533 RELEASE_LOCK(&sched_mutex); /* ToDo: kill */
534 startSignalHandlers();
535 ACQUIRE_LOCK(&sched_mutex);
539 /* Check whether any waiting threads need to be woken up. If the
540 * run queue is empty, and there are no other tasks running, we
541 * can wait indefinitely for something to happen.
543 if ( !EMPTY_QUEUE(blocked_queue_hd) || !EMPTY_QUEUE(sleeping_queue)
544 #if defined(RTS_SUPPORTS_THREADS)
549 awaitEvent( EMPTY_RUN_QUEUE() );
551 /* we can be interrupted while waiting for I/O... */
552 if (interrupted) continue;
555 * Detect deadlock: when we have no threads to run, there are no
556 * threads waiting on I/O or sleeping, and all the other tasks are
557 * waiting for work, we must have a deadlock of some description.
559 * We first try to find threads blocked on themselves (ie. black
560 * holes), and generate NonTermination exceptions where necessary.
562 * If no threads are black holed, we have a deadlock situation, so
563 * inform all the main threads.
565 #if !defined(PAR) && !defined(RTS_SUPPORTS_THREADS)
566 if ( EMPTY_THREAD_QUEUES() )
568 IF_DEBUG(scheduler, sched_belch("deadlocked, forcing major GC..."));
569 // Garbage collection can release some new threads due to
570 // either (a) finalizers or (b) threads resurrected because
571 // they are about to be send BlockedOnDeadMVar. Any threads
572 // thus released will be immediately runnable.
573 GarbageCollect(GetRoots,rtsTrue);
575 if ( !EMPTY_RUN_QUEUE() ) { goto not_deadlocked; }
578 sched_belch("still deadlocked, checking for black holes..."));
581 if ( !EMPTY_RUN_QUEUE() ) { goto not_deadlocked; }
583 #if defined(RTS_USER_SIGNALS)
584 /* If we have user-installed signal handlers, then wait
585 * for signals to arrive rather then bombing out with a
588 if ( anyUserHandlers() ) {
590 sched_belch("still deadlocked, waiting for signals..."));
594 // we might be interrupted...
595 if (interrupted) { continue; }
597 if (signals_pending()) {
598 RELEASE_LOCK(&sched_mutex);
599 startSignalHandlers();
600 ACQUIRE_LOCK(&sched_mutex);
602 ASSERT(!EMPTY_RUN_QUEUE());
607 /* Probably a real deadlock. Send the current main thread the
608 * Deadlock exception (or in the SMP build, send *all* main
609 * threads the deadlock exception, since none of them can make
615 switch (m->tso->why_blocked) {
616 case BlockedOnBlackHole:
617 raiseAsync(m->tso, (StgClosure *)NonTermination_closure);
619 case BlockedOnException:
621 raiseAsync(m->tso, (StgClosure *)Deadlock_closure);
624 barf("deadlock: main thread blocked in a strange way");
630 #elif defined(RTS_SUPPORTS_THREADS)
631 // ToDo: add deadlock detection in threaded RTS
633 // ToDo: add deadlock detection in GUM (similar to SMP) -- HWL
636 #if defined(RTS_SUPPORTS_THREADS)
637 if ( EMPTY_RUN_QUEUE() ) {
638 continue; // nothing to do
643 if (RtsFlags.GranFlags.Light)
644 GranSimLight_enter_system(event, &ActiveTSO); // adjust ActiveTSO etc
646 /* adjust time based on time-stamp */
647 if (event->time > CurrentTime[CurrentProc] &&
648 event->evttype != ContinueThread)
649 CurrentTime[CurrentProc] = event->time;
651 /* Deal with the idle PEs (may issue FindWork or MoveSpark events) */
652 if (!RtsFlags.GranFlags.Light)
655 IF_DEBUG(gran, fprintf(stderr, "GRAN: switch by event-type\n"));
657 /* main event dispatcher in GranSim */
658 switch (event->evttype) {
659 /* Should just be continuing execution */
661 IF_DEBUG(gran, fprintf(stderr, "GRAN: doing ContinueThread\n"));
662 /* ToDo: check assertion
663 ASSERT(run_queue_hd != (StgTSO*)NULL &&
664 run_queue_hd != END_TSO_QUEUE);
666 /* Ignore ContinueThreads for fetching threads (if synchr comm) */
667 if (!RtsFlags.GranFlags.DoAsyncFetch &&
668 procStatus[CurrentProc]==Fetching) {
669 belch("ghuH: Spurious ContinueThread while Fetching ignored; TSO %d (%p) [PE %d]",
670 CurrentTSO->id, CurrentTSO, CurrentProc);
673 /* Ignore ContinueThreads for completed threads */
674 if (CurrentTSO->what_next == ThreadComplete) {
675 belch("ghuH: found a ContinueThread event for completed thread %d (%p) [PE %d] (ignoring ContinueThread)",
676 CurrentTSO->id, CurrentTSO, CurrentProc);
679 /* Ignore ContinueThreads for threads that are being migrated */
680 if (PROCS(CurrentTSO)==Nowhere) {
681 belch("ghuH: trying to run the migrating TSO %d (%p) [PE %d] (ignoring ContinueThread)",
682 CurrentTSO->id, CurrentTSO, CurrentProc);
685 /* The thread should be at the beginning of the run queue */
686 if (CurrentTSO!=run_queue_hds[CurrentProc]) {
687 belch("ghuH: TSO %d (%p) [PE %d] is not at the start of the run_queue when doing a ContinueThread",
688 CurrentTSO->id, CurrentTSO, CurrentProc);
689 break; // run the thread anyway
692 new_event(proc, proc, CurrentTime[proc],
694 (StgTSO*)NULL, (StgClosure*)NULL, (rtsSpark*)NULL);
696 */ /* Catches superfluous CONTINUEs -- should be unnecessary */
697 break; // now actually run the thread; DaH Qu'vam yImuHbej
700 do_the_fetchnode(event);
701 goto next_thread; /* handle next event in event queue */
704 do_the_globalblock(event);
705 goto next_thread; /* handle next event in event queue */
708 do_the_fetchreply(event);
709 goto next_thread; /* handle next event in event queue */
711 case UnblockThread: /* Move from the blocked queue to the tail of */
712 do_the_unblock(event);
713 goto next_thread; /* handle next event in event queue */
715 case ResumeThread: /* Move from the blocked queue to the tail of */
716 /* the runnable queue ( i.e. Qu' SImqa'lu') */
717 event->tso->gran.blocktime +=
718 CurrentTime[CurrentProc] - event->tso->gran.blockedat;
719 do_the_startthread(event);
720 goto next_thread; /* handle next event in event queue */
723 do_the_startthread(event);
724 goto next_thread; /* handle next event in event queue */
727 do_the_movethread(event);
728 goto next_thread; /* handle next event in event queue */
731 do_the_movespark(event);
732 goto next_thread; /* handle next event in event queue */
735 do_the_findwork(event);
736 goto next_thread; /* handle next event in event queue */
739 barf("Illegal event type %u\n", event->evttype);
742 /* This point was scheduler_loop in the old RTS */
744 IF_DEBUG(gran, belch("GRAN: after main switch"));
746 TimeOfLastEvent = CurrentTime[CurrentProc];
747 TimeOfNextEvent = get_time_of_next_event();
748 IgnoreEvents=(TimeOfNextEvent==0); // HWL HACK
749 // CurrentTSO = ThreadQueueHd;
751 IF_DEBUG(gran, belch("GRAN: time of next event is: %ld",
754 if (RtsFlags.GranFlags.Light)
755 GranSimLight_leave_system(event, &ActiveTSO);
757 EndOfTimeSlice = CurrentTime[CurrentProc]+RtsFlags.GranFlags.time_slice;
760 belch("GRAN: end of time-slice is %#lx", EndOfTimeSlice));
762 /* in a GranSim setup the TSO stays on the run queue */
764 /* Take a thread from the run queue. */
765 POP_RUN_QUEUE(t); // take_off_run_queue(t);
768 fprintf(stderr, "GRAN: About to run current thread, which is\n");
771 context_switch = 0; // turned on via GranYield, checking events and time slice
774 DumpGranEvent(GR_SCHEDULE, t));
776 procStatus[CurrentProc] = Busy;
779 if (PendingFetches != END_BF_QUEUE) {
783 /* ToDo: phps merge with spark activation above */
784 /* check whether we have local work and send requests if we have none */
785 if (EMPTY_RUN_QUEUE()) { /* no runnable threads */
786 /* :-[ no local threads => look out for local sparks */
787 /* the spark pool for the current PE */
788 pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable
789 if (advisory_thread_count < RtsFlags.ParFlags.maxThreads &&
790 pool->hd < pool->tl) {
792 * ToDo: add GC code check that we really have enough heap afterwards!!
794 * If we're here (no runnable threads) and we have pending
795 * sparks, we must have a space problem. Get enough space
796 * to turn one of those pending sparks into a
800 spark = findSpark(rtsFalse); /* get a spark */
801 if (spark != (rtsSpark) NULL) {
802 tso = activateSpark(spark); /* turn the spark into a thread */
803 IF_PAR_DEBUG(schedule,
804 belch("==== schedule: Created TSO %d (%p); %d threads active",
805 tso->id, tso, advisory_thread_count));
807 if (tso==END_TSO_QUEUE) { /* failed to activate spark->back to loop */
808 belch("==^^ failed to activate spark");
810 } /* otherwise fall through & pick-up new tso */
812 IF_PAR_DEBUG(verbose,
813 belch("==^^ no local sparks (spark pool contains only NFs: %d)",
814 spark_queue_len(pool)));
819 /* If we still have no work we need to send a FISH to get a spark
822 if (EMPTY_RUN_QUEUE()) {
823 /* =8-[ no local sparks => look for work on other PEs */
825 * We really have absolutely no work. Send out a fish
826 * (there may be some out there already), and wait for
827 * something to arrive. We clearly can't run any threads
828 * until a SCHEDULE or RESUME arrives, and so that's what
829 * we're hoping to see. (Of course, we still have to
830 * respond to other types of messages.)
832 TIME now = msTime() /*CURRENT_TIME*/;
833 IF_PAR_DEBUG(verbose,
834 belch("-- now=%ld", now));
835 IF_PAR_DEBUG(verbose,
836 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
837 (last_fish_arrived_at!=0 &&
838 last_fish_arrived_at+RtsFlags.ParFlags.fishDelay > now)) {
839 belch("--$$ delaying FISH until %ld (last fish %ld, delay %ld, now %ld)",
840 last_fish_arrived_at+RtsFlags.ParFlags.fishDelay,
841 last_fish_arrived_at,
842 RtsFlags.ParFlags.fishDelay, now);
845 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
846 (last_fish_arrived_at==0 ||
847 (last_fish_arrived_at+RtsFlags.ParFlags.fishDelay <= now))) {
848 /* outstandingFishes is set in sendFish, processFish;
849 avoid flooding system with fishes via delay */
851 sendFish(pe, mytid, NEW_FISH_AGE, NEW_FISH_HISTORY,
854 // Global statistics: count no. of fishes
855 if (RtsFlags.ParFlags.ParStats.Global &&
856 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
857 globalParStats.tot_fish_mess++;
861 receivedFinish = processMessages();
864 } else if (PacketsWaiting()) { /* Look for incoming messages */
865 receivedFinish = processMessages();
868 /* Now we are sure that we have some work available */
869 ASSERT(run_queue_hd != END_TSO_QUEUE);
871 /* Take a thread from the run queue, if we have work */
872 POP_RUN_QUEUE(t); // take_off_run_queue(END_TSO_QUEUE);
873 IF_DEBUG(sanity,checkTSO(t));
875 /* ToDo: write something to the log-file
876 if (RTSflags.ParFlags.granSimStats && !sameThread)
877 DumpGranEvent(GR_SCHEDULE, RunnableThreadsHd);
881 /* the spark pool for the current PE */
882 pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable
885 belch("--=^ %d threads, %d sparks on [%#x]",
886 run_queue_len(), spark_queue_len(pool), CURRENT_PROC));
889 if (0 && RtsFlags.ParFlags.ParStats.Full &&
890 t && LastTSO && t->id != LastTSO->id &&
891 LastTSO->why_blocked == NotBlocked &&
892 LastTSO->what_next != ThreadComplete) {
893 // if previously scheduled TSO not blocked we have to record the context switch
894 DumpVeryRawGranEvent(TimeOfLastYield, CURRENT_PROC, CURRENT_PROC,
895 GR_DESCHEDULE, LastTSO, (StgClosure *)NULL, 0, 0);
898 if (RtsFlags.ParFlags.ParStats.Full &&
899 (emitSchedule /* forced emit */ ||
900 (t && LastTSO && t->id != LastTSO->id))) {
902 we are running a different TSO, so write a schedule event to log file
903 NB: If we use fair scheduling we also have to write a deschedule
904 event for LastTSO; with unfair scheduling we know that the
905 previous tso has blocked whenever we switch to another tso, so
906 we don't need it in GUM for now
908 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
909 GR_SCHEDULE, t, (StgClosure *)NULL, 0, 0);
910 emitSchedule = rtsFalse;
914 #else /* !GRAN && !PAR */
916 // grab a thread from the run queue
917 ASSERT(run_queue_hd != END_TSO_QUEUE);
920 // Sanity check the thread we're about to run. This can be
921 // expensive if there is lots of thread switching going on...
922 IF_DEBUG(sanity,checkTSO(t));
928 for(m = main_threads; m; m = m->link)
939 sched_belch("### Running thread %d in bound thread", t->id));
940 // yes, the Haskell thread is bound to the current native thread
945 sched_belch("### thread %d bound to another OS thread", t->id));
946 // no, bound to a different Haskell thread: pass to that thread
947 PUSH_ON_RUN_QUEUE(t);
948 passCapability(&m->bound_thread_cond);
954 if(mainThread != NULL)
955 // The thread we want to run is bound.
958 sched_belch("### this OS thread cannot run thread %d", t->id));
959 // no, the current native thread is bound to a different
960 // Haskell thread, so pass it to any worker thread
961 PUSH_ON_RUN_QUEUE(t);
962 passCapabilityToWorker();
969 cap->r.rCurrentTSO = t;
971 /* context switches are now initiated by the timer signal, unless
972 * the user specified "context switch as often as possible", with
975 if ((RtsFlags.ConcFlags.ctxtSwitchTicks == 0
976 && (run_queue_hd != END_TSO_QUEUE
977 || blocked_queue_hd != END_TSO_QUEUE
978 || sleeping_queue != END_TSO_QUEUE)))
985 RELEASE_LOCK(&sched_mutex);
987 IF_DEBUG(scheduler, sched_belch("-->> running thread %ld %s ...",
988 t->id, whatNext_strs[t->what_next]));
991 startHeapProfTimer();
994 /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
995 /* Run the current thread
997 prev_what_next = t->what_next;
998 switch (prev_what_next) {
1000 case ThreadComplete:
1001 /* Thread already finished, return to scheduler. */
1002 ret = ThreadFinished;
1005 errno = t->saved_errno;
1006 ret = StgRun((StgFunPtr) stg_returnToStackTop, &cap->r);
1007 t->saved_errno = errno;
1009 case ThreadInterpret:
1010 ret = interpretBCO(cap);
1013 barf("schedule: invalid what_next field");
1015 /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
1017 /* Costs for the scheduler are assigned to CCS_SYSTEM */
1019 stopHeapProfTimer();
1023 ACQUIRE_LOCK(&sched_mutex);
1025 #ifdef RTS_SUPPORTS_THREADS
1026 IF_DEBUG(scheduler,fprintf(stderr,"sched (task %p): ", osThreadId()););
1027 #elif !defined(GRAN) && !defined(PAR)
1028 IF_DEBUG(scheduler,fprintf(stderr,"sched: "););
1030 t = cap->r.rCurrentTSO;
1033 /* HACK 675: if the last thread didn't yield, make sure to print a
1034 SCHEDULE event to the log file when StgRunning the next thread, even
1035 if it is the same one as before */
1037 TimeOfLastYield = CURRENT_TIME;
1043 IF_DEBUG(gran, DumpGranEvent(GR_DESCHEDULE, t));
1044 globalGranStats.tot_heapover++;
1046 globalParStats.tot_heapover++;
1049 // did the task ask for a large block?
1050 if (cap->r.rHpAlloc > BLOCK_SIZE_W) {
1051 // if so, get one and push it on the front of the nursery.
1055 blocks = (nat)BLOCK_ROUND_UP(cap->r.rHpAlloc * sizeof(W_)) / BLOCK_SIZE;
1057 IF_DEBUG(scheduler,belch("--<< thread %ld (%s) stopped: requesting a large block (size %d)",
1058 t->id, whatNext_strs[t->what_next], blocks));
1060 // don't do this if it would push us over the
1061 // alloc_blocks_lim limit; we'll GC first.
1062 if (alloc_blocks + blocks < alloc_blocks_lim) {
1064 alloc_blocks += blocks;
1065 bd = allocGroup( blocks );
1067 // link the new group into the list
1068 bd->link = cap->r.rCurrentNursery;
1069 bd->u.back = cap->r.rCurrentNursery->u.back;
1070 if (cap->r.rCurrentNursery->u.back != NULL) {
1071 cap->r.rCurrentNursery->u.back->link = bd;
1073 ASSERT(g0s0->blocks == cap->r.rCurrentNursery &&
1074 g0s0->blocks == cap->r.rNursery);
1075 cap->r.rNursery = g0s0->blocks = bd;
1077 cap->r.rCurrentNursery->u.back = bd;
1079 // initialise it as a nursery block. We initialise the
1080 // step, gen_no, and flags field of *every* sub-block in
1081 // this large block, because this is easier than making
1082 // sure that we always find the block head of a large
1083 // block whenever we call Bdescr() (eg. evacuate() and
1084 // isAlive() in the GC would both have to do this, at
1088 for (x = bd; x < bd + blocks; x++) {
1095 // don't forget to update the block count in g0s0.
1096 g0s0->n_blocks += blocks;
1097 // This assert can be a killer if the app is doing lots
1098 // of large block allocations.
1099 ASSERT(countBlocks(g0s0->blocks) == g0s0->n_blocks);
1101 // now update the nursery to point to the new block
1102 cap->r.rCurrentNursery = bd;
1104 // we might be unlucky and have another thread get on the
1105 // run queue before us and steal the large block, but in that
1106 // case the thread will just end up requesting another large
1108 PUSH_ON_RUN_QUEUE(t);
1113 /* make all the running tasks block on a condition variable,
1114 * maybe set context_switch and wait till they all pile in,
1115 * then have them wait on a GC condition variable.
1117 IF_DEBUG(scheduler,belch("--<< thread %ld (%s) stopped: HeapOverflow",
1118 t->id, whatNext_strs[t->what_next]));
1121 ASSERT(!is_on_queue(t,CurrentProc));
1123 /* Currently we emit a DESCHEDULE event before GC in GUM.
1124 ToDo: either add separate event to distinguish SYSTEM time from rest
1125 or just nuke this DESCHEDULE (and the following SCHEDULE) */
1126 if (0 && RtsFlags.ParFlags.ParStats.Full) {
1127 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
1128 GR_DESCHEDULE, t, (StgClosure *)NULL, 0, 0);
1129 emitSchedule = rtsTrue;
1133 ready_to_gc = rtsTrue;
1134 context_switch = 1; /* stop other threads ASAP */
1135 PUSH_ON_RUN_QUEUE(t);
1136 /* actual GC is done at the end of the while loop */
1142 DumpGranEvent(GR_DESCHEDULE, t));
1143 globalGranStats.tot_stackover++;
1146 // DumpGranEvent(GR_DESCHEDULE, t);
1147 globalParStats.tot_stackover++;
1149 IF_DEBUG(scheduler,belch("--<< thread %ld (%s) stopped, StackOverflow",
1150 t->id, whatNext_strs[t->what_next]));
1151 /* just adjust the stack for this thread, then pop it back
1157 /* enlarge the stack */
1158 StgTSO *new_t = threadStackOverflow(t);
1160 /* This TSO has moved, so update any pointers to it from the
1161 * main thread stack. It better not be on any other queues...
1162 * (it shouldn't be).
1164 for (m = main_threads; m != NULL; m = m->link) {
1169 threadPaused(new_t);
1170 PUSH_ON_RUN_QUEUE(new_t);
1174 case ThreadYielding:
1177 DumpGranEvent(GR_DESCHEDULE, t));
1178 globalGranStats.tot_yields++;
1181 // DumpGranEvent(GR_DESCHEDULE, t);
1182 globalParStats.tot_yields++;
1184 /* put the thread back on the run queue. Then, if we're ready to
1185 * GC, check whether this is the last task to stop. If so, wake
1186 * up the GC thread. getThread will block during a GC until the
1190 if (t->what_next != prev_what_next) {
1191 belch("--<< thread %ld (%s) stopped to switch evaluators",
1192 t->id, whatNext_strs[t->what_next]);
1194 belch("--<< thread %ld (%s) stopped, yielding",
1195 t->id, whatNext_strs[t->what_next]);
1200 //belch("&& Doing sanity check on yielding TSO %ld.", t->id);
1202 ASSERT(t->link == END_TSO_QUEUE);
1204 // Shortcut if we're just switching evaluators: don't bother
1205 // doing stack squeezing (which can be expensive), just run the
1207 if (t->what_next != prev_what_next) {
1214 ASSERT(!is_on_queue(t,CurrentProc));
1217 //belch("&& Doing sanity check on all ThreadQueues (and their TSOs).");
1218 checkThreadQsSanity(rtsTrue));
1222 if (RtsFlags.ParFlags.doFairScheduling) {
1223 /* this does round-robin scheduling; good for concurrency */
1224 APPEND_TO_RUN_QUEUE(t);
1226 /* this does unfair scheduling; good for parallelism */
1227 PUSH_ON_RUN_QUEUE(t);
1230 // this does round-robin scheduling; good for concurrency
1231 APPEND_TO_RUN_QUEUE(t);
1235 /* add a ContinueThread event to actually process the thread */
1236 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1238 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1240 belch("GRAN: eventq and runnableq after adding yielded thread to queue again:");
1249 belch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: ",
1250 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)));
1251 if (t->block_info.closure!=(StgClosure*)NULL) print_bq(t->block_info.closure));
1253 // ??? needed; should emit block before
1255 DumpGranEvent(GR_DESCHEDULE, t));
1256 prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
1259 ASSERT(procStatus[CurrentProc]==Busy ||
1260 ((procStatus[CurrentProc]==Fetching) &&
1261 (t->block_info.closure!=(StgClosure*)NULL)));
1262 if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
1263 !(!RtsFlags.GranFlags.DoAsyncFetch &&
1264 procStatus[CurrentProc]==Fetching))
1265 procStatus[CurrentProc] = Idle;
1269 belch("--<< thread %ld (%p; %s) stopped, blocking on node %p with BQ: ",
1270 t->id, t, whatNext_strs[t->what_next], t->block_info.closure));
1273 if (t->block_info.closure!=(StgClosure*)NULL)
1274 print_bq(t->block_info.closure));
1276 /* Send a fetch (if BlockedOnGA) and dump event to log file */
1279 /* whatever we schedule next, we must log that schedule */
1280 emitSchedule = rtsTrue;
1283 /* don't need to do anything. Either the thread is blocked on
1284 * I/O, in which case we'll have called addToBlockedQueue
1285 * previously, or it's blocked on an MVar or Blackhole, in which
1286 * case it'll be on the relevant queue already.
1289 fprintf(stderr, "--<< thread %d (%s) stopped: ",
1290 t->id, whatNext_strs[t->what_next]);
1291 printThreadBlockage(t);
1292 fprintf(stderr, "\n"));
1295 /* Only for dumping event to log file
1296 ToDo: do I need this in GranSim, too?
1303 case ThreadFinished:
1304 /* Need to check whether this was a main thread, and if so, signal
1305 * the task that started it with the return value. If we have no
1306 * more main threads, we probably need to stop all the tasks until
1309 /* We also end up here if the thread kills itself with an
1310 * uncaught exception, see Exception.hc.
1312 IF_DEBUG(scheduler,belch("--++ thread %d (%s) finished",
1313 t->id, whatNext_strs[t->what_next]));
1315 endThread(t, CurrentProc); // clean-up the thread
1317 /* For now all are advisory -- HWL */
1318 //if(t->priority==AdvisoryPriority) ??
1319 advisory_thread_count--;
1322 if(t->dist.priority==RevalPriority)
1326 if (RtsFlags.ParFlags.ParStats.Full &&
1327 !RtsFlags.ParFlags.ParStats.Suppressed)
1328 DumpEndEvent(CURRENT_PROC, t, rtsFalse /* not mandatory */);
1333 barf("schedule: invalid thread return code %d", (int)ret);
1337 // When we have +RTS -i0 and we're heap profiling, do a census at
1338 // every GC. This lets us get repeatable runs for debugging.
1339 if (performHeapProfile ||
1340 (RtsFlags.ProfFlags.profileInterval==0 &&
1341 RtsFlags.ProfFlags.doHeapProfile && ready_to_gc)) {
1342 GarbageCollect(GetRoots, rtsTrue);
1344 performHeapProfile = rtsFalse;
1345 ready_to_gc = rtsFalse; // we already GC'd
1350 /* everybody back, start the GC.
1351 * Could do it in this thread, or signal a condition var
1352 * to do it in another thread. Either way, we need to
1353 * broadcast on gc_pending_cond afterward.
1355 #if defined(RTS_SUPPORTS_THREADS)
1356 IF_DEBUG(scheduler,sched_belch("doing GC"));
1358 GarbageCollect(GetRoots,rtsFalse);
1359 ready_to_gc = rtsFalse;
1361 /* add a ContinueThread event to continue execution of current thread */
1362 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1364 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1366 fprintf(stderr, "GRAN: eventq and runnableq after Garbage collection:\n");
1374 IF_GRAN_DEBUG(unused,
1375 print_eventq(EventHd));
1377 event = get_next_event();
1380 /* ToDo: wait for next message to arrive rather than busy wait */
1383 } /* end of while(1) */
1385 IF_PAR_DEBUG(verbose,
1386 belch("== Leaving schedule() after having received Finish"));
1389 /* ---------------------------------------------------------------------------
1390 * rtsSupportsBoundThreads(): is the RTS built to support bound threads?
1391 * used by Control.Concurrent for error checking.
1392 * ------------------------------------------------------------------------- */
1395 rtsSupportsBoundThreads(void)
1404 /* ---------------------------------------------------------------------------
1405 * isThreadBound(tso): check whether tso is bound to an OS thread.
1406 * ------------------------------------------------------------------------- */
1409 isThreadBound(StgTSO* tso USED_IN_THREADED_RTS)
1413 for(m = main_threads; m; m = m->link)
1422 /* ---------------------------------------------------------------------------
1423 * Singleton fork(). Do not copy any running threads.
1424 * ------------------------------------------------------------------------- */
1427 deleteThreadImmediately(StgTSO *tso);
1430 forkProcess(HsStablePtr *entry)
1432 #ifndef mingw32_TARGET_OS
1438 IF_DEBUG(scheduler,sched_belch("forking!"));
1439 rts_lock(); // This not only acquires sched_mutex, it also
1440 // makes sure that no other threads are running
1444 if (pid) { /* parent */
1446 /* just return the pid */
1450 } else { /* child */
1453 // delete all threads
1454 run_queue_hd = run_queue_tl = END_TSO_QUEUE;
1456 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
1459 // don't allow threads to catch the ThreadKilled exception
1460 deleteThreadImmediately(t);
1463 // wipe the main thread list
1464 while((m = main_threads) != NULL) {
1465 main_threads = m->link;
1467 closeCondition(&m->bound_thread_cond);
1472 #ifdef RTS_SUPPORTS_THREADS
1473 resetTaskManagerAfterFork(); // tell startTask() and friends that
1474 startingWorkerThread = rtsFalse; // we have no worker threads any more
1475 resetWorkerWakeupPipeAfterFork();
1478 rc = rts_evalStableIO(entry, NULL); // run the action
1479 rts_checkSchedStatus("forkProcess",rc);
1483 hs_exit(); // clean up and exit
1487 barf("forkProcess#: primop not implemented for mingw32, sorry!\n");
1489 #endif /* mingw32 */
1492 /* ---------------------------------------------------------------------------
1493 * deleteAllThreads(): kill all the live threads.
1495 * This is used when we catch a user interrupt (^C), before performing
1496 * any necessary cleanups and running finalizers.
1498 * Locks: sched_mutex held.
1499 * ------------------------------------------------------------------------- */
1502 deleteAllThreads ( void )
1505 IF_DEBUG(scheduler,sched_belch("deleting all threads"));
1506 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
1507 next = t->global_link;
1510 run_queue_hd = run_queue_tl = END_TSO_QUEUE;
1511 blocked_queue_hd = blocked_queue_tl = END_TSO_QUEUE;
1512 sleeping_queue = END_TSO_QUEUE;
1515 /* startThread and insertThread are now in GranSim.c -- HWL */
1518 /* ---------------------------------------------------------------------------
1519 * Suspending & resuming Haskell threads.
1521 * When making a "safe" call to C (aka _ccall_GC), the task gives back
1522 * its capability before calling the C function. This allows another
1523 * task to pick up the capability and carry on running Haskell
1524 * threads. It also means that if the C call blocks, it won't lock
1527 * The Haskell thread making the C call is put to sleep for the
1528 * duration of the call, on the susepended_ccalling_threads queue. We
1529 * give out a token to the task, which it can use to resume the thread
1530 * on return from the C function.
1531 * ------------------------------------------------------------------------- */
1534 suspendThread( StgRegTable *reg,
1543 int saved_errno = errno;
1545 /* assume that *reg is a pointer to the StgRegTable part
1548 cap = (Capability *)((void *)((unsigned char*)reg - sizeof(StgFunTable)));
1550 ACQUIRE_LOCK(&sched_mutex);
1553 sched_belch("thread %d did a _ccall_gc (is_concurrent: %d)", cap->r.rCurrentTSO->id,concCall));
1555 // XXX this might not be necessary --SDM
1556 cap->r.rCurrentTSO->what_next = ThreadRunGHC;
1558 threadPaused(cap->r.rCurrentTSO);
1559 cap->r.rCurrentTSO->link = suspended_ccalling_threads;
1560 suspended_ccalling_threads = cap->r.rCurrentTSO;
1562 #if defined(RTS_SUPPORTS_THREADS)
1563 if(cap->r.rCurrentTSO->blocked_exceptions == NULL)
1565 cap->r.rCurrentTSO->why_blocked = BlockedOnCCall;
1566 cap->r.rCurrentTSO->blocked_exceptions = END_TSO_QUEUE;
1570 cap->r.rCurrentTSO->why_blocked = BlockedOnCCall_NoUnblockExc;
1574 /* Use the thread ID as the token; it should be unique */
1575 tok = cap->r.rCurrentTSO->id;
1577 /* Hand back capability */
1578 releaseCapability(cap);
1580 #if defined(RTS_SUPPORTS_THREADS)
1581 /* Preparing to leave the RTS, so ensure there's a native thread/task
1582 waiting to take over.
1584 IF_DEBUG(scheduler, sched_belch("worker (token %d): leaving RTS", tok));
1587 /* Other threads _might_ be available for execution; signal this */
1589 RELEASE_LOCK(&sched_mutex);
1591 errno = saved_errno;
1596 resumeThread( StgInt tok,
1597 rtsBool concCall STG_UNUSED )
1599 StgTSO *tso, **prev;
1601 int saved_errno = errno;
1603 #if defined(RTS_SUPPORTS_THREADS)
1604 /* Wait for permission to re-enter the RTS with the result. */
1605 ACQUIRE_LOCK(&sched_mutex);
1606 waitForReturnCapability(&sched_mutex, &cap);
1608 IF_DEBUG(scheduler, sched_belch("worker (token %d): re-entering RTS", tok));
1610 grabCapability(&cap);
1613 /* Remove the thread off of the suspended list */
1614 prev = &suspended_ccalling_threads;
1615 for (tso = suspended_ccalling_threads;
1616 tso != END_TSO_QUEUE;
1617 prev = &tso->link, tso = tso->link) {
1618 if (tso->id == (StgThreadID)tok) {
1623 if (tso == END_TSO_QUEUE) {
1624 barf("resumeThread: thread not found");
1626 tso->link = END_TSO_QUEUE;
1628 #if defined(RTS_SUPPORTS_THREADS)
1629 if(tso->why_blocked == BlockedOnCCall)
1631 awakenBlockedQueueNoLock(tso->blocked_exceptions);
1632 tso->blocked_exceptions = NULL;
1636 /* Reset blocking status */
1637 tso->why_blocked = NotBlocked;
1639 cap->r.rCurrentTSO = tso;
1640 RELEASE_LOCK(&sched_mutex);
1641 errno = saved_errno;
1646 /* ---------------------------------------------------------------------------
1648 * ------------------------------------------------------------------------ */
1649 static void unblockThread(StgTSO *tso);
1651 /* ---------------------------------------------------------------------------
1652 * Comparing Thread ids.
1654 * This is used from STG land in the implementation of the
1655 * instances of Eq/Ord for ThreadIds.
1656 * ------------------------------------------------------------------------ */
1659 cmp_thread(StgPtr tso1, StgPtr tso2)
1661 StgThreadID id1 = ((StgTSO *)tso1)->id;
1662 StgThreadID id2 = ((StgTSO *)tso2)->id;
1664 if (id1 < id2) return (-1);
1665 if (id1 > id2) return 1;
1669 /* ---------------------------------------------------------------------------
1670 * Fetching the ThreadID from an StgTSO.
1672 * This is used in the implementation of Show for ThreadIds.
1673 * ------------------------------------------------------------------------ */
1675 rts_getThreadId(StgPtr tso)
1677 return ((StgTSO *)tso)->id;
1682 labelThread(StgPtr tso, char *label)
1687 /* Caveat: Once set, you can only set the thread name to "" */
1688 len = strlen(label)+1;
1689 buf = stgMallocBytes(len * sizeof(char), "Schedule.c:labelThread()");
1690 strncpy(buf,label,len);
1691 /* Update will free the old memory for us */
1692 updateThreadLabel(((StgTSO *)tso)->id,buf);
1696 /* ---------------------------------------------------------------------------
1697 Create a new thread.
1699 The new thread starts with the given stack size. Before the
1700 scheduler can run, however, this thread needs to have a closure
1701 (and possibly some arguments) pushed on its stack. See
1702 pushClosure() in Schedule.h.
1704 createGenThread() and createIOThread() (in SchedAPI.h) are
1705 convenient packaged versions of this function.
1707 currently pri (priority) is only used in a GRAN setup -- HWL
1708 ------------------------------------------------------------------------ */
1710 /* currently pri (priority) is only used in a GRAN setup -- HWL */
1712 createThread(nat size, StgInt pri)
1715 createThread(nat size)
1722 /* First check whether we should create a thread at all */
1724 /* check that no more than RtsFlags.ParFlags.maxThreads threads are created */
1725 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads) {
1727 belch("{createThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
1728 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
1729 return END_TSO_QUEUE;
1735 ASSERT(!RtsFlags.GranFlags.Light || CurrentProc==0);
1738 // ToDo: check whether size = stack_size - TSO_STRUCT_SIZEW
1740 /* catch ridiculously small stack sizes */
1741 if (size < MIN_STACK_WORDS + TSO_STRUCT_SIZEW) {
1742 size = MIN_STACK_WORDS + TSO_STRUCT_SIZEW;
1745 stack_size = size - TSO_STRUCT_SIZEW;
1747 tso = (StgTSO *)allocate(size);
1748 TICK_ALLOC_TSO(stack_size, 0);
1750 SET_HDR(tso, &stg_TSO_info, CCS_SYSTEM);
1752 SET_GRAN_HDR(tso, ThisPE);
1755 // Always start with the compiled code evaluator
1756 tso->what_next = ThreadRunGHC;
1758 tso->id = next_thread_id++;
1759 tso->why_blocked = NotBlocked;
1760 tso->blocked_exceptions = NULL;
1762 tso->saved_errno = 0;
1764 tso->stack_size = stack_size;
1765 tso->max_stack_size = round_to_mblocks(RtsFlags.GcFlags.maxStkSize)
1767 tso->sp = (P_)&(tso->stack) + stack_size;
1770 tso->prof.CCCS = CCS_MAIN;
1773 /* put a stop frame on the stack */
1774 tso->sp -= sizeofW(StgStopFrame);
1775 SET_HDR((StgClosure*)tso->sp,(StgInfoTable *)&stg_stop_thread_info,CCS_SYSTEM);
1778 tso->link = END_TSO_QUEUE;
1779 /* uses more flexible routine in GranSim */
1780 insertThread(tso, CurrentProc);
1782 /* In a non-GranSim setup the pushing of a TSO onto the runq is separated
1788 if (RtsFlags.GranFlags.GranSimStats.Full)
1789 DumpGranEvent(GR_START,tso);
1791 if (RtsFlags.ParFlags.ParStats.Full)
1792 DumpGranEvent(GR_STARTQ,tso);
1793 /* HACk to avoid SCHEDULE
1797 /* Link the new thread on the global thread list.
1799 tso->global_link = all_threads;
1803 tso->dist.priority = MandatoryPriority; //by default that is...
1807 tso->gran.pri = pri;
1809 tso->gran.magic = TSO_MAGIC; // debugging only
1811 tso->gran.sparkname = 0;
1812 tso->gran.startedat = CURRENT_TIME;
1813 tso->gran.exported = 0;
1814 tso->gran.basicblocks = 0;
1815 tso->gran.allocs = 0;
1816 tso->gran.exectime = 0;
1817 tso->gran.fetchtime = 0;
1818 tso->gran.fetchcount = 0;
1819 tso->gran.blocktime = 0;
1820 tso->gran.blockcount = 0;
1821 tso->gran.blockedat = 0;
1822 tso->gran.globalsparks = 0;
1823 tso->gran.localsparks = 0;
1824 if (RtsFlags.GranFlags.Light)
1825 tso->gran.clock = Now; /* local clock */
1827 tso->gran.clock = 0;
1829 IF_DEBUG(gran,printTSO(tso));
1832 tso->par.magic = TSO_MAGIC; // debugging only
1834 tso->par.sparkname = 0;
1835 tso->par.startedat = CURRENT_TIME;
1836 tso->par.exported = 0;
1837 tso->par.basicblocks = 0;
1838 tso->par.allocs = 0;
1839 tso->par.exectime = 0;
1840 tso->par.fetchtime = 0;
1841 tso->par.fetchcount = 0;
1842 tso->par.blocktime = 0;
1843 tso->par.blockcount = 0;
1844 tso->par.blockedat = 0;
1845 tso->par.globalsparks = 0;
1846 tso->par.localsparks = 0;
1850 globalGranStats.tot_threads_created++;
1851 globalGranStats.threads_created_on_PE[CurrentProc]++;
1852 globalGranStats.tot_sq_len += spark_queue_len(CurrentProc);
1853 globalGranStats.tot_sq_probes++;
1855 // collect parallel global statistics (currently done together with GC stats)
1856 if (RtsFlags.ParFlags.ParStats.Global &&
1857 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
1858 //fprintf(stderr, "Creating thread %d @ %11.2f\n", tso->id, usertime());
1859 globalParStats.tot_threads_created++;
1865 belch("==__ schedule: Created TSO %d (%p);",
1866 CurrentProc, tso, tso->id));
1868 IF_PAR_DEBUG(verbose,
1869 belch("==__ schedule: Created TSO %d (%p); %d threads active",
1870 tso->id, tso, advisory_thread_count));
1872 IF_DEBUG(scheduler,sched_belch("created thread %ld, stack size = %lx words",
1873 tso->id, tso->stack_size));
1880 all parallel thread creation calls should fall through the following routine.
1883 createSparkThread(rtsSpark spark)
1885 ASSERT(spark != (rtsSpark)NULL);
1886 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads)
1888 barf("{createSparkThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
1889 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
1890 return END_TSO_QUEUE;
1894 tso = createThread(RtsFlags.GcFlags.initialStkSize);
1895 if (tso==END_TSO_QUEUE)
1896 barf("createSparkThread: Cannot create TSO");
1898 tso->priority = AdvisoryPriority;
1900 pushClosure(tso,spark);
1901 PUSH_ON_RUN_QUEUE(tso);
1902 advisory_thread_count++;
1909 Turn a spark into a thread.
1910 ToDo: fix for SMP (needs to acquire SCHED_MUTEX!)
1914 activateSpark (rtsSpark spark)
1918 tso = createSparkThread(spark);
1919 if (RtsFlags.ParFlags.ParStats.Full) {
1920 //ASSERT(run_queue_hd == END_TSO_QUEUE); // I think ...
1921 IF_PAR_DEBUG(verbose,
1922 belch("==^^ activateSpark: turning spark of closure %p (%s) into a thread",
1923 (StgClosure *)spark, info_type((StgClosure *)spark)));
1925 // ToDo: fwd info on local/global spark to thread -- HWL
1926 // tso->gran.exported = spark->exported;
1927 // tso->gran.locked = !spark->global;
1928 // tso->gran.sparkname = spark->name;
1934 static SchedulerStatus waitThread_(/*out*/StgMainThread* m,
1935 Capability *initialCapability
1939 /* ---------------------------------------------------------------------------
1942 * scheduleThread puts a thread on the head of the runnable queue.
1943 * This will usually be done immediately after a thread is created.
1944 * The caller of scheduleThread must create the thread using e.g.
1945 * createThread and push an appropriate closure
1946 * on this thread's stack before the scheduler is invoked.
1947 * ------------------------------------------------------------------------ */
1949 static void scheduleThread_ (StgTSO* tso);
1952 scheduleThread_(StgTSO *tso)
1954 // Precondition: sched_mutex must be held.
1955 PUSH_ON_RUN_QUEUE(tso);
1960 scheduleThread(StgTSO* tso)
1962 ACQUIRE_LOCK(&sched_mutex);
1963 scheduleThread_(tso);
1964 RELEASE_LOCK(&sched_mutex);
1967 #if defined(RTS_SUPPORTS_THREADS)
1968 static Condition bound_cond_cache;
1969 static int bound_cond_cache_full = 0;
1974 scheduleWaitThread(StgTSO* tso, /*[out]*/HaskellObj* ret,
1975 Capability *initialCapability)
1977 // Precondition: sched_mutex must be held
1980 m = stgMallocBytes(sizeof(StgMainThread), "waitThread");
1984 #if defined(RTS_SUPPORTS_THREADS)
1985 // Allocating a new condition for each thread is expensive, so we
1986 // cache one. This is a pretty feeble hack, but it helps speed up
1987 // consecutive call-ins quite a bit.
1988 if (bound_cond_cache_full) {
1989 m->bound_thread_cond = bound_cond_cache;
1990 bound_cond_cache_full = 0;
1992 initCondition(&m->bound_thread_cond);
1996 /* Put the thread on the main-threads list prior to scheduling the TSO.
1997 Failure to do so introduces a race condition in the MT case (as
1998 identified by Wolfgang Thaller), whereby the new task/OS thread
1999 created by scheduleThread_() would complete prior to the thread
2000 that spawned it managed to put 'itself' on the main-threads list.
2001 The upshot of it all being that the worker thread wouldn't get to
2002 signal the completion of the its work item for the main thread to
2003 see (==> it got stuck waiting.) -- sof 6/02.
2005 IF_DEBUG(scheduler, sched_belch("waiting for thread (%d)", tso->id));
2007 m->link = main_threads;
2010 PUSH_ON_RUN_QUEUE(tso);
2011 // NB. Don't call THREAD_RUNNABLE() here, because the thread is
2012 // bound and only runnable by *this* OS thread, so waking up other
2013 // workers will just slow things down.
2015 return waitThread_(m, initialCapability);
2018 /* ---------------------------------------------------------------------------
2021 * Initialise the scheduler. This resets all the queues - if the
2022 * queues contained any threads, they'll be garbage collected at the
2025 * ------------------------------------------------------------------------ */
2033 for (i=0; i<=MAX_PROC; i++) {
2034 run_queue_hds[i] = END_TSO_QUEUE;
2035 run_queue_tls[i] = END_TSO_QUEUE;
2036 blocked_queue_hds[i] = END_TSO_QUEUE;
2037 blocked_queue_tls[i] = END_TSO_QUEUE;
2038 ccalling_threadss[i] = END_TSO_QUEUE;
2039 sleeping_queue = END_TSO_QUEUE;
2042 run_queue_hd = END_TSO_QUEUE;
2043 run_queue_tl = END_TSO_QUEUE;
2044 blocked_queue_hd = END_TSO_QUEUE;
2045 blocked_queue_tl = END_TSO_QUEUE;
2046 sleeping_queue = END_TSO_QUEUE;
2049 suspended_ccalling_threads = END_TSO_QUEUE;
2051 main_threads = NULL;
2052 all_threads = END_TSO_QUEUE;
2057 RtsFlags.ConcFlags.ctxtSwitchTicks =
2058 RtsFlags.ConcFlags.ctxtSwitchTime / TICK_MILLISECS;
2060 #if defined(RTS_SUPPORTS_THREADS)
2061 /* Initialise the mutex and condition variables used by
2063 initMutex(&sched_mutex);
2064 initMutex(&term_mutex);
2067 ACQUIRE_LOCK(&sched_mutex);
2069 /* A capability holds the state a native thread needs in
2070 * order to execute STG code. At least one capability is
2071 * floating around (only SMP builds have more than one).
2075 #if defined(RTS_SUPPORTS_THREADS)
2076 /* start our haskell execution tasks */
2077 startTaskManager(0,taskStart);
2080 #if /* defined(SMP) ||*/ defined(PAR)
2084 RELEASE_LOCK(&sched_mutex);
2088 exitScheduler( void )
2090 #if defined(RTS_SUPPORTS_THREADS)
2093 shutting_down_scheduler = rtsTrue;
2096 /* ----------------------------------------------------------------------------
2097 Managing the per-task allocation areas.
2099 Each capability comes with an allocation area. These are
2100 fixed-length block lists into which allocation can be done.
2102 ToDo: no support for two-space collection at the moment???
2103 ------------------------------------------------------------------------- */
2107 waitThread_(StgMainThread* m, Capability *initialCapability)
2109 SchedulerStatus stat;
2111 // Precondition: sched_mutex must be held.
2112 IF_DEBUG(scheduler, sched_belch("new main thread (%d)", m->tso->id));
2115 /* GranSim specific init */
2116 CurrentTSO = m->tso; // the TSO to run
2117 procStatus[MainProc] = Busy; // status of main PE
2118 CurrentProc = MainProc; // PE to run it on
2119 schedule(m,initialCapability);
2121 schedule(m,initialCapability);
2122 ASSERT(m->stat != NoStatus);
2127 #if defined(RTS_SUPPORTS_THREADS)
2128 // Free the condition variable, returning it to the cache if possible.
2129 if (!bound_cond_cache_full) {
2130 bound_cond_cache = m->bound_thread_cond;
2131 bound_cond_cache_full = 1;
2133 closeCondition(&m->bound_thread_cond);
2137 IF_DEBUG(scheduler, sched_belch("main thread (%d) finished", m->tso->id));
2140 // Postcondition: sched_mutex still held
2144 /* ---------------------------------------------------------------------------
2145 Where are the roots that we know about?
2147 - all the threads on the runnable queue
2148 - all the threads on the blocked queue
2149 - all the threads on the sleeping queue
2150 - all the thread currently executing a _ccall_GC
2151 - all the "main threads"
2153 ------------------------------------------------------------------------ */
2155 /* This has to be protected either by the scheduler monitor, or by the
2156 garbage collection monitor (probably the latter).
2161 GetRoots( evac_fn evac )
2166 for (i=0; i<=RtsFlags.GranFlags.proc; i++) {
2167 if ((run_queue_hds[i] != END_TSO_QUEUE) && ((run_queue_hds[i] != NULL)))
2168 evac((StgClosure **)&run_queue_hds[i]);
2169 if ((run_queue_tls[i] != END_TSO_QUEUE) && ((run_queue_tls[i] != NULL)))
2170 evac((StgClosure **)&run_queue_tls[i]);
2172 if ((blocked_queue_hds[i] != END_TSO_QUEUE) && ((blocked_queue_hds[i] != NULL)))
2173 evac((StgClosure **)&blocked_queue_hds[i]);
2174 if ((blocked_queue_tls[i] != END_TSO_QUEUE) && ((blocked_queue_tls[i] != NULL)))
2175 evac((StgClosure **)&blocked_queue_tls[i]);
2176 if ((ccalling_threadss[i] != END_TSO_QUEUE) && ((ccalling_threadss[i] != NULL)))
2177 evac((StgClosure **)&ccalling_threads[i]);
2184 if (run_queue_hd != END_TSO_QUEUE) {
2185 ASSERT(run_queue_tl != END_TSO_QUEUE);
2186 evac((StgClosure **)&run_queue_hd);
2187 evac((StgClosure **)&run_queue_tl);
2190 if (blocked_queue_hd != END_TSO_QUEUE) {
2191 ASSERT(blocked_queue_tl != END_TSO_QUEUE);
2192 evac((StgClosure **)&blocked_queue_hd);
2193 evac((StgClosure **)&blocked_queue_tl);
2196 if (sleeping_queue != END_TSO_QUEUE) {
2197 evac((StgClosure **)&sleeping_queue);
2201 if (suspended_ccalling_threads != END_TSO_QUEUE) {
2202 evac((StgClosure **)&suspended_ccalling_threads);
2205 #if defined(PAR) || defined(GRAN)
2206 markSparkQueue(evac);
2209 #if defined(RTS_USER_SIGNALS)
2210 // mark the signal handlers (signals should be already blocked)
2211 markSignalHandlers(evac);
2214 // main threads which have completed need to be retained until they
2215 // are dealt with in the main scheduler loop. They won't be
2216 // retained any other way: the GC will drop them from the
2217 // all_threads list, so we have to be careful to treat them as roots
2221 for (m = main_threads; m != NULL; m = m->link) {
2222 switch (m->tso->what_next) {
2223 case ThreadComplete:
2225 evac((StgClosure **)&m->tso);
2234 /* -----------------------------------------------------------------------------
2237 This is the interface to the garbage collector from Haskell land.
2238 We provide this so that external C code can allocate and garbage
2239 collect when called from Haskell via _ccall_GC.
2241 It might be useful to provide an interface whereby the programmer
2242 can specify more roots (ToDo).
2244 This needs to be protected by the GC condition variable above. KH.
2245 -------------------------------------------------------------------------- */
2247 static void (*extra_roots)(evac_fn);
2252 /* Obligated to hold this lock upon entry */
2253 ACQUIRE_LOCK(&sched_mutex);
2254 GarbageCollect(GetRoots,rtsFalse);
2255 RELEASE_LOCK(&sched_mutex);
2259 performMajorGC(void)
2261 ACQUIRE_LOCK(&sched_mutex);
2262 GarbageCollect(GetRoots,rtsTrue);
2263 RELEASE_LOCK(&sched_mutex);
2267 AllRoots(evac_fn evac)
2269 GetRoots(evac); // the scheduler's roots
2270 extra_roots(evac); // the user's roots
2274 performGCWithRoots(void (*get_roots)(evac_fn))
2276 ACQUIRE_LOCK(&sched_mutex);
2277 extra_roots = get_roots;
2278 GarbageCollect(AllRoots,rtsFalse);
2279 RELEASE_LOCK(&sched_mutex);
2282 /* -----------------------------------------------------------------------------
2285 If the thread has reached its maximum stack size, then raise the
2286 StackOverflow exception in the offending thread. Otherwise
2287 relocate the TSO into a larger chunk of memory and adjust its stack
2289 -------------------------------------------------------------------------- */
2292 threadStackOverflow(StgTSO *tso)
2294 nat new_stack_size, new_tso_size, stack_words;
2298 IF_DEBUG(sanity,checkTSO(tso));
2299 if (tso->stack_size >= tso->max_stack_size) {
2302 belch("@@ threadStackOverflow of TSO %d (%p): stack too large (now %ld; max is %ld)",
2303 tso->id, tso, tso->stack_size, tso->max_stack_size);
2304 /* If we're debugging, just print out the top of the stack */
2305 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2308 /* Send this thread the StackOverflow exception */
2309 raiseAsync(tso, (StgClosure *)stackOverflow_closure);
2313 /* Try to double the current stack size. If that takes us over the
2314 * maximum stack size for this thread, then use the maximum instead.
2315 * Finally round up so the TSO ends up as a whole number of blocks.
2317 new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
2318 new_tso_size = (nat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) +
2319 TSO_STRUCT_SIZE)/sizeof(W_);
2320 new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */
2321 new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
2323 IF_DEBUG(scheduler, fprintf(stderr,"== sched: increasing stack size from %d words to %d.\n", tso->stack_size, new_stack_size));
2325 dest = (StgTSO *)allocate(new_tso_size);
2326 TICK_ALLOC_TSO(new_stack_size,0);
2328 /* copy the TSO block and the old stack into the new area */
2329 memcpy(dest,tso,TSO_STRUCT_SIZE);
2330 stack_words = tso->stack + tso->stack_size - tso->sp;
2331 new_sp = (P_)dest + new_tso_size - stack_words;
2332 memcpy(new_sp, tso->sp, stack_words * sizeof(W_));
2334 /* relocate the stack pointers... */
2336 dest->stack_size = new_stack_size;
2338 /* Mark the old TSO as relocated. We have to check for relocated
2339 * TSOs in the garbage collector and any primops that deal with TSOs.
2341 * It's important to set the sp value to just beyond the end
2342 * of the stack, so we don't attempt to scavenge any part of the
2345 tso->what_next = ThreadRelocated;
2347 tso->sp = (P_)&(tso->stack[tso->stack_size]);
2348 tso->why_blocked = NotBlocked;
2349 dest->mut_link = NULL;
2351 IF_PAR_DEBUG(verbose,
2352 belch("@@ threadStackOverflow of TSO %d (now at %p): stack size increased to %ld",
2353 tso->id, tso, tso->stack_size);
2354 /* If we're debugging, just print out the top of the stack */
2355 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2358 IF_DEBUG(sanity,checkTSO(tso));
2360 IF_DEBUG(scheduler,printTSO(dest));
2366 /* ---------------------------------------------------------------------------
2367 Wake up a queue that was blocked on some resource.
2368 ------------------------------------------------------------------------ */
2372 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2377 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2379 /* write RESUME events to log file and
2380 update blocked and fetch time (depending on type of the orig closure) */
2381 if (RtsFlags.ParFlags.ParStats.Full) {
2382 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
2383 GR_RESUMEQ, ((StgTSO *)bqe), ((StgTSO *)bqe)->block_info.closure,
2384 0, 0 /* spark_queue_len(ADVISORY_POOL) */);
2385 if (EMPTY_RUN_QUEUE())
2386 emitSchedule = rtsTrue;
2388 switch (get_itbl(node)->type) {
2390 ((StgTSO *)bqe)->par.fetchtime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2395 ((StgTSO *)bqe)->par.blocktime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2402 barf("{unblockOneLocked}Daq Qagh: unexpected closure in blocking queue");
2409 static StgBlockingQueueElement *
2410 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2413 PEs node_loc, tso_loc;
2415 node_loc = where_is(node); // should be lifted out of loop
2416 tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2417 tso_loc = where_is((StgClosure *)tso);
2418 if (IS_LOCAL_TO(PROCS(node),tso_loc)) { // TSO is local
2419 /* !fake_fetch => TSO is on CurrentProc is same as IS_LOCAL_TO */
2420 ASSERT(CurrentProc!=node_loc || tso_loc==CurrentProc);
2421 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.lunblocktime;
2422 // insertThread(tso, node_loc);
2423 new_event(tso_loc, tso_loc, CurrentTime[CurrentProc],
2425 tso, node, (rtsSpark*)NULL);
2426 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2429 } else { // TSO is remote (actually should be FMBQ)
2430 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.mpacktime +
2431 RtsFlags.GranFlags.Costs.gunblocktime +
2432 RtsFlags.GranFlags.Costs.latency;
2433 new_event(tso_loc, CurrentProc, CurrentTime[CurrentProc],
2435 tso, node, (rtsSpark*)NULL);
2436 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2439 /* the thread-queue-overhead is accounted for in either Resume or UnblockThread */
2441 fprintf(stderr," %s TSO %d (%p) [PE %d] (block_info.closure=%p) (next=%p) ,",
2442 (node_loc==tso_loc ? "Local" : "Global"),
2443 tso->id, tso, CurrentProc, tso->block_info.closure, tso->link));
2444 tso->block_info.closure = NULL;
2445 IF_DEBUG(scheduler,belch("-- Waking up thread %ld (%p)",
2449 static StgBlockingQueueElement *
2450 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2452 StgBlockingQueueElement *next;
2454 switch (get_itbl(bqe)->type) {
2456 ASSERT(((StgTSO *)bqe)->why_blocked != NotBlocked);
2457 /* if it's a TSO just push it onto the run_queue */
2459 // ((StgTSO *)bqe)->link = END_TSO_QUEUE; // debugging?
2460 PUSH_ON_RUN_QUEUE((StgTSO *)bqe);
2462 unblockCount(bqe, node);
2463 /* reset blocking status after dumping event */
2464 ((StgTSO *)bqe)->why_blocked = NotBlocked;
2468 /* if it's a BLOCKED_FETCH put it on the PendingFetches list */
2470 bqe->link = (StgBlockingQueueElement *)PendingFetches;
2471 PendingFetches = (StgBlockedFetch *)bqe;
2475 /* can ignore this case in a non-debugging setup;
2476 see comments on RBHSave closures above */
2478 /* check that the closure is an RBHSave closure */
2479 ASSERT(get_itbl((StgClosure *)bqe) == &stg_RBH_Save_0_info ||
2480 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_1_info ||
2481 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_2_info);
2485 barf("{unblockOneLocked}Daq Qagh: Unexpected IP (%#lx; %s) in blocking queue at %#lx\n",
2486 get_itbl((StgClosure *)bqe), info_type((StgClosure *)bqe),
2490 IF_PAR_DEBUG(bq, fprintf(stderr, ", %p (%s)", bqe, info_type((StgClosure*)bqe)));
2494 #else /* !GRAN && !PAR */
2496 unblockOneLocked(StgTSO *tso)
2500 ASSERT(get_itbl(tso)->type == TSO);
2501 ASSERT(tso->why_blocked != NotBlocked);
2502 tso->why_blocked = NotBlocked;
2504 PUSH_ON_RUN_QUEUE(tso);
2506 IF_DEBUG(scheduler,sched_belch("waking up thread %ld", tso->id));
2511 #if defined(GRAN) || defined(PAR)
2512 INLINE_ME StgBlockingQueueElement *
2513 unblockOne(StgBlockingQueueElement *bqe, StgClosure *node)
2515 ACQUIRE_LOCK(&sched_mutex);
2516 bqe = unblockOneLocked(bqe, node);
2517 RELEASE_LOCK(&sched_mutex);
2522 unblockOne(StgTSO *tso)
2524 ACQUIRE_LOCK(&sched_mutex);
2525 tso = unblockOneLocked(tso);
2526 RELEASE_LOCK(&sched_mutex);
2533 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2535 StgBlockingQueueElement *bqe;
2540 belch("##-_ AwBQ for node %p on PE %d @ %ld by TSO %d (%p): ", \
2541 node, CurrentProc, CurrentTime[CurrentProc],
2542 CurrentTSO->id, CurrentTSO));
2544 node_loc = where_is(node);
2546 ASSERT(q == END_BQ_QUEUE ||
2547 get_itbl(q)->type == TSO || // q is either a TSO or an RBHSave
2548 get_itbl(q)->type == CONSTR); // closure (type constructor)
2549 ASSERT(is_unique(node));
2551 /* FAKE FETCH: magically copy the node to the tso's proc;
2552 no Fetch necessary because in reality the node should not have been
2553 moved to the other PE in the first place
2555 if (CurrentProc!=node_loc) {
2557 belch("## node %p is on PE %d but CurrentProc is %d (TSO %d); assuming fake fetch and adjusting bitmask (old: %#x)",
2558 node, node_loc, CurrentProc, CurrentTSO->id,
2559 // CurrentTSO, where_is(CurrentTSO),
2560 node->header.gran.procs));
2561 node->header.gran.procs = (node->header.gran.procs) | PE_NUMBER(CurrentProc);
2563 belch("## new bitmask of node %p is %#x",
2564 node, node->header.gran.procs));
2565 if (RtsFlags.GranFlags.GranSimStats.Global) {
2566 globalGranStats.tot_fake_fetches++;
2571 // ToDo: check: ASSERT(CurrentProc==node_loc);
2572 while (get_itbl(bqe)->type==TSO) { // q != END_TSO_QUEUE) {
2575 bqe points to the current element in the queue
2576 next points to the next element in the queue
2578 //tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2579 //tso_loc = where_is(tso);
2581 bqe = unblockOneLocked(bqe, node);
2584 /* if this is the BQ of an RBH, we have to put back the info ripped out of
2585 the closure to make room for the anchor of the BQ */
2586 if (bqe!=END_BQ_QUEUE) {
2587 ASSERT(get_itbl(node)->type == RBH && get_itbl(bqe)->type == CONSTR);
2589 ASSERT((info_ptr==&RBH_Save_0_info) ||
2590 (info_ptr==&RBH_Save_1_info) ||
2591 (info_ptr==&RBH_Save_2_info));
2593 /* cf. convertToRBH in RBH.c for writing the RBHSave closure */
2594 ((StgRBH *)node)->blocking_queue = (StgBlockingQueueElement *)((StgRBHSave *)bqe)->payload[0];
2595 ((StgRBH *)node)->mut_link = (StgMutClosure *)((StgRBHSave *)bqe)->payload[1];
2598 belch("## Filled in RBH_Save for %p (%s) at end of AwBQ",
2599 node, info_type(node)));
2602 /* statistics gathering */
2603 if (RtsFlags.GranFlags.GranSimStats.Global) {
2604 // globalGranStats.tot_bq_processing_time += bq_processing_time;
2605 globalGranStats.tot_bq_len += len; // total length of all bqs awakened
2606 // globalGranStats.tot_bq_len_local += len_local; // same for local TSOs only
2607 globalGranStats.tot_awbq++; // total no. of bqs awakened
2610 fprintf(stderr,"## BQ Stats of %p: [%d entries] %s\n",
2611 node, len, (bqe!=END_BQ_QUEUE) ? "RBH" : ""));
2615 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2617 StgBlockingQueueElement *bqe;
2619 ACQUIRE_LOCK(&sched_mutex);
2621 IF_PAR_DEBUG(verbose,
2622 belch("##-_ AwBQ for node %p on [%x]: ",
2626 if(get_itbl(q)->type == CONSTR || q==END_BQ_QUEUE) {
2627 IF_PAR_DEBUG(verbose, belch("## ... nothing to unblock so lets just return. RFP (BUG?)"));
2632 ASSERT(q == END_BQ_QUEUE ||
2633 get_itbl(q)->type == TSO ||
2634 get_itbl(q)->type == BLOCKED_FETCH ||
2635 get_itbl(q)->type == CONSTR);
2638 while (get_itbl(bqe)->type==TSO ||
2639 get_itbl(bqe)->type==BLOCKED_FETCH) {
2640 bqe = unblockOneLocked(bqe, node);
2642 RELEASE_LOCK(&sched_mutex);
2645 #else /* !GRAN && !PAR */
2647 #ifdef RTS_SUPPORTS_THREADS
2649 awakenBlockedQueueNoLock(StgTSO *tso)
2651 while (tso != END_TSO_QUEUE) {
2652 tso = unblockOneLocked(tso);
2658 awakenBlockedQueue(StgTSO *tso)
2660 ACQUIRE_LOCK(&sched_mutex);
2661 while (tso != END_TSO_QUEUE) {
2662 tso = unblockOneLocked(tso);
2664 RELEASE_LOCK(&sched_mutex);
2668 /* ---------------------------------------------------------------------------
2670 - usually called inside a signal handler so it mustn't do anything fancy.
2671 ------------------------------------------------------------------------ */
2674 interruptStgRts(void)
2678 #ifdef RTS_SUPPORTS_THREADS
2679 wakeBlockedWorkerThread();
2683 /* -----------------------------------------------------------------------------
2686 This is for use when we raise an exception in another thread, which
2688 This has nothing to do with the UnblockThread event in GranSim. -- HWL
2689 -------------------------------------------------------------------------- */
2691 #if defined(GRAN) || defined(PAR)
2693 NB: only the type of the blocking queue is different in GranSim and GUM
2694 the operations on the queue-elements are the same
2695 long live polymorphism!
2697 Locks: sched_mutex is held upon entry and exit.
2701 unblockThread(StgTSO *tso)
2703 StgBlockingQueueElement *t, **last;
2705 switch (tso->why_blocked) {
2708 return; /* not blocked */
2711 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
2713 StgBlockingQueueElement *last_tso = END_BQ_QUEUE;
2714 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
2716 last = (StgBlockingQueueElement **)&mvar->head;
2717 for (t = (StgBlockingQueueElement *)mvar->head;
2719 last = &t->link, last_tso = t, t = t->link) {
2720 if (t == (StgBlockingQueueElement *)tso) {
2721 *last = (StgBlockingQueueElement *)tso->link;
2722 if (mvar->tail == tso) {
2723 mvar->tail = (StgTSO *)last_tso;
2728 barf("unblockThread (MVAR): TSO not found");
2731 case BlockedOnBlackHole:
2732 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
2734 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
2736 last = &bq->blocking_queue;
2737 for (t = bq->blocking_queue;
2739 last = &t->link, t = t->link) {
2740 if (t == (StgBlockingQueueElement *)tso) {
2741 *last = (StgBlockingQueueElement *)tso->link;
2745 barf("unblockThread (BLACKHOLE): TSO not found");
2748 case BlockedOnException:
2750 StgTSO *target = tso->block_info.tso;
2752 ASSERT(get_itbl(target)->type == TSO);
2754 if (target->what_next == ThreadRelocated) {
2755 target = target->link;
2756 ASSERT(get_itbl(target)->type == TSO);
2759 ASSERT(target->blocked_exceptions != NULL);
2761 last = (StgBlockingQueueElement **)&target->blocked_exceptions;
2762 for (t = (StgBlockingQueueElement *)target->blocked_exceptions;
2764 last = &t->link, t = t->link) {
2765 ASSERT(get_itbl(t)->type == TSO);
2766 if (t == (StgBlockingQueueElement *)tso) {
2767 *last = (StgBlockingQueueElement *)tso->link;
2771 barf("unblockThread (Exception): TSO not found");
2775 case BlockedOnWrite:
2776 #if defined(mingw32_TARGET_OS)
2777 case BlockedOnDoProc:
2780 /* take TSO off blocked_queue */
2781 StgBlockingQueueElement *prev = NULL;
2782 for (t = (StgBlockingQueueElement *)blocked_queue_hd; t != END_BQ_QUEUE;
2783 prev = t, t = t->link) {
2784 if (t == (StgBlockingQueueElement *)tso) {
2786 blocked_queue_hd = (StgTSO *)t->link;
2787 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
2788 blocked_queue_tl = END_TSO_QUEUE;
2791 prev->link = t->link;
2792 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
2793 blocked_queue_tl = (StgTSO *)prev;
2799 barf("unblockThread (I/O): TSO not found");
2802 case BlockedOnDelay:
2804 /* take TSO off sleeping_queue */
2805 StgBlockingQueueElement *prev = NULL;
2806 for (t = (StgBlockingQueueElement *)sleeping_queue; t != END_BQ_QUEUE;
2807 prev = t, t = t->link) {
2808 if (t == (StgBlockingQueueElement *)tso) {
2810 sleeping_queue = (StgTSO *)t->link;
2812 prev->link = t->link;
2817 barf("unblockThread (delay): TSO not found");
2821 barf("unblockThread");
2825 tso->link = END_TSO_QUEUE;
2826 tso->why_blocked = NotBlocked;
2827 tso->block_info.closure = NULL;
2828 PUSH_ON_RUN_QUEUE(tso);
2832 unblockThread(StgTSO *tso)
2836 /* To avoid locking unnecessarily. */
2837 if (tso->why_blocked == NotBlocked) {
2841 switch (tso->why_blocked) {
2844 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
2846 StgTSO *last_tso = END_TSO_QUEUE;
2847 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
2850 for (t = mvar->head; t != END_TSO_QUEUE;
2851 last = &t->link, last_tso = t, t = t->link) {
2854 if (mvar->tail == tso) {
2855 mvar->tail = last_tso;
2860 barf("unblockThread (MVAR): TSO not found");
2863 case BlockedOnBlackHole:
2864 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
2866 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
2868 last = &bq->blocking_queue;
2869 for (t = bq->blocking_queue; t != END_TSO_QUEUE;
2870 last = &t->link, t = t->link) {
2876 barf("unblockThread (BLACKHOLE): TSO not found");
2879 case BlockedOnException:
2881 StgTSO *target = tso->block_info.tso;
2883 ASSERT(get_itbl(target)->type == TSO);
2885 while (target->what_next == ThreadRelocated) {
2886 target = target->link;
2887 ASSERT(get_itbl(target)->type == TSO);
2890 ASSERT(target->blocked_exceptions != NULL);
2892 last = &target->blocked_exceptions;
2893 for (t = target->blocked_exceptions; t != END_TSO_QUEUE;
2894 last = &t->link, t = t->link) {
2895 ASSERT(get_itbl(t)->type == TSO);
2901 barf("unblockThread (Exception): TSO not found");
2905 case BlockedOnWrite:
2906 #if defined(mingw32_TARGET_OS)
2907 case BlockedOnDoProc:
2910 StgTSO *prev = NULL;
2911 for (t = blocked_queue_hd; t != END_TSO_QUEUE;
2912 prev = t, t = t->link) {
2915 blocked_queue_hd = t->link;
2916 if (blocked_queue_tl == t) {
2917 blocked_queue_tl = END_TSO_QUEUE;
2920 prev->link = t->link;
2921 if (blocked_queue_tl == t) {
2922 blocked_queue_tl = prev;
2928 barf("unblockThread (I/O): TSO not found");
2931 case BlockedOnDelay:
2933 StgTSO *prev = NULL;
2934 for (t = sleeping_queue; t != END_TSO_QUEUE;
2935 prev = t, t = t->link) {
2938 sleeping_queue = t->link;
2940 prev->link = t->link;
2945 barf("unblockThread (delay): TSO not found");
2949 barf("unblockThread");
2953 tso->link = END_TSO_QUEUE;
2954 tso->why_blocked = NotBlocked;
2955 tso->block_info.closure = NULL;
2956 PUSH_ON_RUN_QUEUE(tso);
2960 /* -----------------------------------------------------------------------------
2963 * The following function implements the magic for raising an
2964 * asynchronous exception in an existing thread.
2966 * We first remove the thread from any queue on which it might be
2967 * blocked. The possible blockages are MVARs and BLACKHOLE_BQs.
2969 * We strip the stack down to the innermost CATCH_FRAME, building
2970 * thunks in the heap for all the active computations, so they can
2971 * be restarted if necessary. When we reach a CATCH_FRAME, we build
2972 * an application of the handler to the exception, and push it on
2973 * the top of the stack.
2975 * How exactly do we save all the active computations? We create an
2976 * AP_STACK for every UpdateFrame on the stack. Entering one of these
2977 * AP_STACKs pushes everything from the corresponding update frame
2978 * upwards onto the stack. (Actually, it pushes everything up to the
2979 * next update frame plus a pointer to the next AP_STACK object.
2980 * Entering the next AP_STACK object pushes more onto the stack until we
2981 * reach the last AP_STACK object - at which point the stack should look
2982 * exactly as it did when we killed the TSO and we can continue
2983 * execution by entering the closure on top of the stack.
2985 * We can also kill a thread entirely - this happens if either (a) the
2986 * exception passed to raiseAsync is NULL, or (b) there's no
2987 * CATCH_FRAME on the stack. In either case, we strip the entire
2988 * stack and replace the thread with a zombie.
2990 * Locks: sched_mutex held upon entry nor exit.
2992 * -------------------------------------------------------------------------- */
2995 deleteThread(StgTSO *tso)
2997 raiseAsync(tso,NULL);
3001 deleteThreadImmediately(StgTSO *tso)
3002 { // for forkProcess only:
3003 // delete thread without giving it a chance to catch the KillThread exception
3005 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
3008 #if defined(RTS_SUPPORTS_THREADS)
3009 if (tso->why_blocked != BlockedOnCCall
3010 && tso->why_blocked != BlockedOnCCall_NoUnblockExc)
3013 tso->what_next = ThreadKilled;
3017 raiseAsyncWithLock(StgTSO *tso, StgClosure *exception)
3019 /* When raising async exs from contexts where sched_mutex isn't held;
3020 use raiseAsyncWithLock(). */
3021 ACQUIRE_LOCK(&sched_mutex);
3022 raiseAsync(tso,exception);
3023 RELEASE_LOCK(&sched_mutex);
3027 raiseAsync(StgTSO *tso, StgClosure *exception)
3029 StgRetInfoTable *info;
3032 // Thread already dead?
3033 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
3038 sched_belch("raising exception in thread %ld.", tso->id));
3040 // Remove it from any blocking queues
3045 // The stack freezing code assumes there's a closure pointer on
3046 // the top of the stack, so we have to arrange that this is the case...
3048 if (sp[0] == (W_)&stg_enter_info) {
3052 sp[0] = (W_)&stg_dummy_ret_closure;
3058 // 1. Let the top of the stack be the "current closure"
3060 // 2. Walk up the stack until we find either an UPDATE_FRAME or a
3063 // 3. If it's an UPDATE_FRAME, then make an AP_STACK containing the
3064 // current closure applied to the chunk of stack up to (but not
3065 // including) the update frame. This closure becomes the "current
3066 // closure". Go back to step 2.
3068 // 4. If it's a CATCH_FRAME, then leave the exception handler on
3069 // top of the stack applied to the exception.
3071 // 5. If it's a STOP_FRAME, then kill the thread.
3076 info = get_ret_itbl((StgClosure *)frame);
3078 while (info->i.type != UPDATE_FRAME
3079 && (info->i.type != CATCH_FRAME || exception == NULL)
3080 && info->i.type != STOP_FRAME) {
3081 frame += stack_frame_sizeW((StgClosure *)frame);
3082 info = get_ret_itbl((StgClosure *)frame);
3085 switch (info->i.type) {
3088 // If we find a CATCH_FRAME, and we've got an exception to raise,
3089 // then build the THUNK raise(exception), and leave it on
3090 // top of the CATCH_FRAME ready to enter.
3094 StgCatchFrame *cf = (StgCatchFrame *)frame;
3098 // we've got an exception to raise, so let's pass it to the
3099 // handler in this frame.
3101 raise = (StgClosure *)allocate(sizeofW(StgClosure)+1);
3102 TICK_ALLOC_SE_THK(1,0);
3103 SET_HDR(raise,&stg_raise_info,cf->header.prof.ccs);
3104 raise->payload[0] = exception;
3106 // throw away the stack from Sp up to the CATCH_FRAME.
3110 /* Ensure that async excpetions are blocked now, so we don't get
3111 * a surprise exception before we get around to executing the
3114 if (tso->blocked_exceptions == NULL) {
3115 tso->blocked_exceptions = END_TSO_QUEUE;
3118 /* Put the newly-built THUNK on top of the stack, ready to execute
3119 * when the thread restarts.
3122 sp[-1] = (W_)&stg_enter_info;
3124 tso->what_next = ThreadRunGHC;
3125 IF_DEBUG(sanity, checkTSO(tso));
3134 // First build an AP_STACK consisting of the stack chunk above the
3135 // current update frame, with the top word on the stack as the
3138 words = frame - sp - 1;
3139 ap = (StgAP_STACK *)allocate(PAP_sizeW(words));
3142 ap->fun = (StgClosure *)sp[0];
3144 for(i=0; i < (nat)words; ++i) {
3145 ap->payload[i] = (StgClosure *)*sp++;
3148 SET_HDR(ap,&stg_AP_STACK_info,
3149 ((StgClosure *)frame)->header.prof.ccs /* ToDo */);
3150 TICK_ALLOC_UP_THK(words+1,0);
3153 fprintf(stderr, "sched: Updating ");
3154 printPtr((P_)((StgUpdateFrame *)frame)->updatee);
3155 fprintf(stderr, " with ");
3156 printObj((StgClosure *)ap);
3159 // Replace the updatee with an indirection - happily
3160 // this will also wake up any threads currently
3161 // waiting on the result.
3163 // Warning: if we're in a loop, more than one update frame on
3164 // the stack may point to the same object. Be careful not to
3165 // overwrite an IND_OLDGEN in this case, because we'll screw
3166 // up the mutable lists. To be on the safe side, don't
3167 // overwrite any kind of indirection at all. See also
3168 // threadSqueezeStack in GC.c, where we have to make a similar
3171 if (!closure_IND(((StgUpdateFrame *)frame)->updatee)) {
3172 // revert the black hole
3173 UPD_IND_NOLOCK(((StgUpdateFrame *)frame)->updatee,ap);
3175 sp += sizeofW(StgUpdateFrame) - 1;
3176 sp[0] = (W_)ap; // push onto stack
3181 // We've stripped the entire stack, the thread is now dead.
3182 sp += sizeofW(StgStopFrame);
3183 tso->what_next = ThreadKilled;
3194 /* -----------------------------------------------------------------------------
3195 resurrectThreads is called after garbage collection on the list of
3196 threads found to be garbage. Each of these threads will be woken
3197 up and sent a signal: BlockedOnDeadMVar if the thread was blocked
3198 on an MVar, or NonTermination if the thread was blocked on a Black
3201 Locks: sched_mutex isn't held upon entry nor exit.
3202 -------------------------------------------------------------------------- */
3205 resurrectThreads( StgTSO *threads )
3209 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
3210 next = tso->global_link;
3211 tso->global_link = all_threads;
3213 IF_DEBUG(scheduler, sched_belch("resurrecting thread %d", tso->id));
3215 switch (tso->why_blocked) {
3217 case BlockedOnException:
3218 /* Called by GC - sched_mutex lock is currently held. */
3219 raiseAsync(tso,(StgClosure *)BlockedOnDeadMVar_closure);
3221 case BlockedOnBlackHole:
3222 raiseAsync(tso,(StgClosure *)NonTermination_closure);
3225 /* This might happen if the thread was blocked on a black hole
3226 * belonging to a thread that we've just woken up (raiseAsync
3227 * can wake up threads, remember...).
3231 barf("resurrectThreads: thread blocked in a strange way");
3236 /* -----------------------------------------------------------------------------
3237 * Blackhole detection: if we reach a deadlock, test whether any
3238 * threads are blocked on themselves. Any threads which are found to
3239 * be self-blocked get sent a NonTermination exception.
3241 * This is only done in a deadlock situation in order to avoid
3242 * performance overhead in the normal case.
3244 * Locks: sched_mutex is held upon entry and exit.
3245 * -------------------------------------------------------------------------- */
3248 detectBlackHoles( void )
3250 StgTSO *tso = all_threads;
3252 StgClosure *blocked_on;
3253 StgRetInfoTable *info;
3255 for (tso = all_threads; tso != END_TSO_QUEUE; tso = tso->global_link) {
3257 while (tso->what_next == ThreadRelocated) {
3259 ASSERT(get_itbl(tso)->type == TSO);
3262 if (tso->why_blocked != BlockedOnBlackHole) {
3265 blocked_on = tso->block_info.closure;
3267 frame = (StgClosure *)tso->sp;
3270 info = get_ret_itbl(frame);
3271 switch (info->i.type) {
3273 if (((StgUpdateFrame *)frame)->updatee == blocked_on) {
3274 /* We are blocking on one of our own computations, so
3275 * send this thread the NonTermination exception.
3278 sched_belch("thread %d is blocked on itself", tso->id));
3279 raiseAsync(tso, (StgClosure *)NonTermination_closure);
3283 frame = (StgClosure *) ((StgUpdateFrame *)frame + 1);
3289 // normal stack frames; do nothing except advance the pointer
3291 (StgPtr)frame += stack_frame_sizeW(frame);
3298 /* ----------------------------------------------------------------------------
3299 * Debugging: why is a thread blocked
3300 * [Also provides useful information when debugging threaded programs
3301 * at the Haskell source code level, so enable outside of DEBUG. --sof 7/02]
3302 ------------------------------------------------------------------------- */
3306 printThreadBlockage(StgTSO *tso)
3308 switch (tso->why_blocked) {
3310 fprintf(stderr,"is blocked on read from fd %d", tso->block_info.fd);
3312 case BlockedOnWrite:
3313 fprintf(stderr,"is blocked on write to fd %d", tso->block_info.fd);
3315 #if defined(mingw32_TARGET_OS)
3316 case BlockedOnDoProc:
3317 fprintf(stderr,"is blocked on proc (request: %d)", tso->block_info.async_result->reqID);
3320 case BlockedOnDelay:
3321 fprintf(stderr,"is blocked until %d", tso->block_info.target);
3324 fprintf(stderr,"is blocked on an MVar");
3326 case BlockedOnException:
3327 fprintf(stderr,"is blocked on delivering an exception to thread %d",
3328 tso->block_info.tso->id);
3330 case BlockedOnBlackHole:
3331 fprintf(stderr,"is blocked on a black hole");
3334 fprintf(stderr,"is not blocked");
3338 fprintf(stderr,"is blocked on global address; local FM_BQ is %p (%s)",
3339 tso->block_info.closure, info_type(tso->block_info.closure));
3341 case BlockedOnGA_NoSend:
3342 fprintf(stderr,"is blocked on global address (no send); local FM_BQ is %p (%s)",
3343 tso->block_info.closure, info_type(tso->block_info.closure));
3346 #if defined(RTS_SUPPORTS_THREADS)
3347 case BlockedOnCCall:
3348 fprintf(stderr,"is blocked on an external call");
3350 case BlockedOnCCall_NoUnblockExc:
3351 fprintf(stderr,"is blocked on an external call (exceptions were already blocked)");
3355 barf("printThreadBlockage: strange tso->why_blocked: %d for TSO %d (%d)",
3356 tso->why_blocked, tso->id, tso);
3362 printThreadStatus(StgTSO *tso)
3364 switch (tso->what_next) {
3366 fprintf(stderr,"has been killed");
3368 case ThreadComplete:
3369 fprintf(stderr,"has completed");
3372 printThreadBlockage(tso);
3377 printAllThreads(void)
3383 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
3384 ullong_format_string(TIME_ON_PROC(CurrentProc),
3385 time_string, rtsFalse/*no commas!*/);
3387 fprintf(stderr, "all threads at [%s]:\n", time_string);
3389 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
3390 ullong_format_string(CURRENT_TIME,
3391 time_string, rtsFalse/*no commas!*/);
3393 fprintf(stderr,"all threads at [%s]:\n", time_string);
3395 fprintf(stderr,"all threads:\n");
3398 for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
3399 fprintf(stderr, "\tthread %d @ %p ", t->id, (void *)t);
3400 label = lookupThreadLabel(t->id);
3401 if (label) fprintf(stderr,"[\"%s\"] ",(char *)label);
3402 printThreadStatus(t);
3403 fprintf(stderr,"\n");
3410 Print a whole blocking queue attached to node (debugging only).
3414 print_bq (StgClosure *node)
3416 StgBlockingQueueElement *bqe;
3420 fprintf(stderr,"## BQ of closure %p (%s): ",
3421 node, info_type(node));
3423 /* should cover all closures that may have a blocking queue */
3424 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
3425 get_itbl(node)->type == FETCH_ME_BQ ||
3426 get_itbl(node)->type == RBH ||
3427 get_itbl(node)->type == MVAR);
3429 ASSERT(node!=(StgClosure*)NULL); // sanity check
3431 print_bqe(((StgBlockingQueue*)node)->blocking_queue);
3435 Print a whole blocking queue starting with the element bqe.
3438 print_bqe (StgBlockingQueueElement *bqe)
3443 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
3445 for (end = (bqe==END_BQ_QUEUE);
3446 !end; // iterate until bqe points to a CONSTR
3447 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE),
3448 bqe = end ? END_BQ_QUEUE : bqe->link) {
3449 ASSERT(bqe != END_BQ_QUEUE); // sanity check
3450 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
3451 /* types of closures that may appear in a blocking queue */
3452 ASSERT(get_itbl(bqe)->type == TSO ||
3453 get_itbl(bqe)->type == BLOCKED_FETCH ||
3454 get_itbl(bqe)->type == CONSTR);
3455 /* only BQs of an RBH end with an RBH_Save closure */
3456 //ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
3458 switch (get_itbl(bqe)->type) {
3460 fprintf(stderr," TSO %u (%x),",
3461 ((StgTSO *)bqe)->id, ((StgTSO *)bqe));
3464 fprintf(stderr," BF (node=%p, ga=((%x, %d, %x)),",
3465 ((StgBlockedFetch *)bqe)->node,
3466 ((StgBlockedFetch *)bqe)->ga.payload.gc.gtid,
3467 ((StgBlockedFetch *)bqe)->ga.payload.gc.slot,
3468 ((StgBlockedFetch *)bqe)->ga.weight);
3471 fprintf(stderr," %s (IP %p),",
3472 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
3473 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
3474 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
3475 "RBH_Save_?"), get_itbl(bqe));
3478 barf("Unexpected closure type %s in blocking queue", // of %p (%s)",
3479 info_type((StgClosure *)bqe)); // , node, info_type(node));
3483 fputc('\n', stderr);
3485 # elif defined(GRAN)
3487 print_bq (StgClosure *node)
3489 StgBlockingQueueElement *bqe;
3490 PEs node_loc, tso_loc;
3493 /* should cover all closures that may have a blocking queue */
3494 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
3495 get_itbl(node)->type == FETCH_ME_BQ ||
3496 get_itbl(node)->type == RBH);
3498 ASSERT(node!=(StgClosure*)NULL); // sanity check
3499 node_loc = where_is(node);
3501 fprintf(stderr,"## BQ of closure %p (%s) on [PE %d]: ",
3502 node, info_type(node), node_loc);
3505 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
3507 for (bqe = ((StgBlockingQueue*)node)->blocking_queue, end = (bqe==END_BQ_QUEUE);
3508 !end; // iterate until bqe points to a CONSTR
3509 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) {
3510 ASSERT(bqe != END_BQ_QUEUE); // sanity check
3511 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
3512 /* types of closures that may appear in a blocking queue */
3513 ASSERT(get_itbl(bqe)->type == TSO ||
3514 get_itbl(bqe)->type == CONSTR);
3515 /* only BQs of an RBH end with an RBH_Save closure */
3516 ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
3518 tso_loc = where_is((StgClosure *)bqe);
3519 switch (get_itbl(bqe)->type) {
3521 fprintf(stderr," TSO %d (%p) on [PE %d],",
3522 ((StgTSO *)bqe)->id, (StgTSO *)bqe, tso_loc);
3525 fprintf(stderr," %s (IP %p),",
3526 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
3527 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
3528 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
3529 "RBH_Save_?"), get_itbl(bqe));
3532 barf("Unexpected closure type %s in blocking queue of %p (%s)",
3533 info_type((StgClosure *)bqe), node, info_type(node));
3537 fputc('\n', stderr);
3541 Nice and easy: only TSOs on the blocking queue
3544 print_bq (StgClosure *node)
3548 ASSERT(node!=(StgClosure*)NULL); // sanity check
3549 for (tso = ((StgBlockingQueue*)node)->blocking_queue;
3550 tso != END_TSO_QUEUE;
3552 ASSERT(tso!=NULL && tso!=END_TSO_QUEUE); // sanity check
3553 ASSERT(get_itbl(tso)->type == TSO); // guess what, sanity check
3554 fprintf(stderr," TSO %d (%p),", tso->id, tso);
3556 fputc('\n', stderr);
3567 for (i=0, tso=run_queue_hd;
3568 tso != END_TSO_QUEUE;
3577 sched_belch(char *s, ...)
3581 #ifdef RTS_SUPPORTS_THREADS
3582 fprintf(stderr, "sched (task %p): ", osThreadId());
3584 fprintf(stderr, "== ");
3586 fprintf(stderr, "sched: ");
3588 vfprintf(stderr, s, ap);
3589 fprintf(stderr, "\n");