1 /* ---------------------------------------------------------------------------
2 * $Id: Schedule.c,v 1.194 2004/03/13 00:56:45 sof 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);
262 startingWorkerThread = rtsFalse;
264 RELEASE_LOCK(&sched_mutex);
268 startSchedulerTaskIfNecessary(void)
270 if(run_queue_hd != END_TSO_QUEUE
271 || blocked_queue_hd != END_TSO_QUEUE
272 || sleeping_queue != END_TSO_QUEUE)
274 if(!startingWorkerThread)
275 { // we don't want to start another worker thread
276 // just because the last one hasn't yet reached the
277 // "waiting for capability" state
278 startingWorkerThread = rtsTrue;
279 startTask(taskStart);
285 /* ---------------------------------------------------------------------------
286 Main scheduling loop.
288 We use round-robin scheduling, each thread returning to the
289 scheduler loop when one of these conditions is detected:
292 * timer expires (thread yields)
297 Locking notes: we acquire the scheduler lock once at the beginning
298 of the scheduler loop, and release it when
300 * running a thread, or
301 * waiting for work, or
302 * waiting for a GC to complete.
305 In a GranSim setup this loop iterates over the global event queue.
306 This revolves around the global event queue, which determines what
307 to do next. Therefore, it's more complicated than either the
308 concurrent or the parallel (GUM) setup.
311 GUM iterates over incoming messages.
312 It starts with nothing to do (thus CurrentTSO == END_TSO_QUEUE),
313 and sends out a fish whenever it has nothing to do; in-between
314 doing the actual reductions (shared code below) it processes the
315 incoming messages and deals with delayed operations
316 (see PendingFetches).
317 This is not the ugliest code you could imagine, but it's bloody close.
319 ------------------------------------------------------------------------ */
321 schedule( StgMainThread *mainThread USED_WHEN_RTS_SUPPORTS_THREADS,
322 Capability *initialCapability )
326 StgThreadReturnCode ret;
334 rtsBool receivedFinish = rtsFalse;
336 nat tp_size, sp_size; // stats only
339 rtsBool was_interrupted = rtsFalse;
340 StgTSOWhatNext prev_what_next;
342 // Pre-condition: sched_mutex is held.
343 // We might have a capability, passed in as initialCapability.
344 cap = initialCapability;
346 #if defined(RTS_SUPPORTS_THREADS)
348 // in the threaded case, the capability is either passed in via the
349 // initialCapability parameter, or initialized inside the scheduler
353 sched_belch("### NEW SCHEDULER LOOP (main thr: %p, cap: %p)",
354 mainThread, initialCapability);
357 // simply initialise it in the non-threaded case
358 grabCapability(&cap);
362 /* set up first event to get things going */
363 /* ToDo: assign costs for system setup and init MainTSO ! */
364 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
366 CurrentTSO, (StgClosure*)NULL, (rtsSpark*)NULL);
369 fprintf(stderr, "GRAN: Init CurrentTSO (in schedule) = %p\n", CurrentTSO);
370 G_TSO(CurrentTSO, 5));
372 if (RtsFlags.GranFlags.Light) {
373 /* Save current time; GranSim Light only */
374 CurrentTSO->gran.clock = CurrentTime[CurrentProc];
377 event = get_next_event();
379 while (event!=(rtsEvent*)NULL) {
380 /* Choose the processor with the next event */
381 CurrentProc = event->proc;
382 CurrentTSO = event->tso;
386 while (!receivedFinish) { /* set by processMessages */
387 /* when receiving PP_FINISH message */
389 #else // everything except GRAN and PAR
395 IF_DEBUG(scheduler, printAllThreads());
397 #if defined(RTS_SUPPORTS_THREADS)
398 // Yield the capability to higher-priority tasks if necessary.
401 yieldCapability(&cap);
404 // If we do not currently hold a capability, we wait for one
407 waitForCapability(&sched_mutex, &cap,
408 mainThread ? &mainThread->bound_thread_cond : NULL);
411 // We now have a capability...
415 // If we're interrupted (the user pressed ^C, or some other
416 // termination condition occurred), kill all the currently running
420 IF_DEBUG(scheduler, sched_belch("interrupted"));
421 interrupted = rtsFalse;
422 was_interrupted = rtsTrue;
423 #if defined(RTS_SUPPORTS_THREADS)
424 // In the threaded RTS, deadlock detection doesn't work,
425 // so just exit right away.
426 prog_belch("interrupted");
427 releaseCapability(cap);
428 RELEASE_LOCK(&sched_mutex);
429 shutdownHaskellAndExit(EXIT_SUCCESS);
435 #if defined(RTS_USER_SIGNALS)
436 // check for signals each time around the scheduler
437 if (signals_pending()) {
438 RELEASE_LOCK(&sched_mutex); /* ToDo: kill */
439 startSignalHandlers();
440 ACQUIRE_LOCK(&sched_mutex);
445 // Check whether any waiting threads need to be woken up. If the
446 // run queue is empty, and there are no other tasks running, we
447 // can wait indefinitely for something to happen.
449 if ( !EMPTY_QUEUE(blocked_queue_hd) || !EMPTY_QUEUE(sleeping_queue)
450 #if defined(RTS_SUPPORTS_THREADS)
455 awaitEvent( EMPTY_RUN_QUEUE() );
457 // we can be interrupted while waiting for I/O...
458 if (interrupted) continue;
461 * Detect deadlock: when we have no threads to run, there are no
462 * threads waiting on I/O or sleeping, and all the other tasks are
463 * waiting for work, we must have a deadlock of some description.
465 * We first try to find threads blocked on themselves (ie. black
466 * holes), and generate NonTermination exceptions where necessary.
468 * If no threads are black holed, we have a deadlock situation, so
469 * inform all the main threads.
471 #if !defined(PAR) && !defined(RTS_SUPPORTS_THREADS)
472 if ( EMPTY_THREAD_QUEUES() )
474 IF_DEBUG(scheduler, sched_belch("deadlocked, forcing major GC..."));
475 // Garbage collection can release some new threads due to
476 // either (a) finalizers or (b) threads resurrected because
477 // they are about to be send BlockedOnDeadMVar. Any threads
478 // thus released will be immediately runnable.
479 GarbageCollect(GetRoots,rtsTrue);
481 if ( !EMPTY_RUN_QUEUE() ) { goto not_deadlocked; }
484 sched_belch("still deadlocked, checking for black holes..."));
487 if ( !EMPTY_RUN_QUEUE() ) { goto not_deadlocked; }
489 #if defined(RTS_USER_SIGNALS)
490 /* If we have user-installed signal handlers, then wait
491 * for signals to arrive rather then bombing out with a
494 if ( anyUserHandlers() ) {
496 sched_belch("still deadlocked, waiting for signals..."));
500 // we might be interrupted...
501 if (interrupted) { continue; }
503 if (signals_pending()) {
504 RELEASE_LOCK(&sched_mutex);
505 startSignalHandlers();
506 ACQUIRE_LOCK(&sched_mutex);
508 ASSERT(!EMPTY_RUN_QUEUE());
513 /* Probably a real deadlock. Send the current main thread the
514 * Deadlock exception (or in the SMP build, send *all* main
515 * threads the deadlock exception, since none of them can make
521 switch (m->tso->why_blocked) {
522 case BlockedOnBlackHole:
523 case BlockedOnException:
525 raiseAsync(m->tso, (StgClosure *)NonTermination_closure);
528 barf("deadlock: main thread blocked in a strange way");
534 #elif defined(RTS_SUPPORTS_THREADS)
535 // ToDo: add deadlock detection in threaded RTS
537 // ToDo: add deadlock detection in GUM (similar to SMP) -- HWL
540 #if defined(RTS_SUPPORTS_THREADS)
541 if ( EMPTY_RUN_QUEUE() ) {
542 continue; // nothing to do
547 if (RtsFlags.GranFlags.Light)
548 GranSimLight_enter_system(event, &ActiveTSO); // adjust ActiveTSO etc
550 /* adjust time based on time-stamp */
551 if (event->time > CurrentTime[CurrentProc] &&
552 event->evttype != ContinueThread)
553 CurrentTime[CurrentProc] = event->time;
555 /* Deal with the idle PEs (may issue FindWork or MoveSpark events) */
556 if (!RtsFlags.GranFlags.Light)
559 IF_DEBUG(gran, fprintf(stderr, "GRAN: switch by event-type\n"));
561 /* main event dispatcher in GranSim */
562 switch (event->evttype) {
563 /* Should just be continuing execution */
565 IF_DEBUG(gran, fprintf(stderr, "GRAN: doing ContinueThread\n"));
566 /* ToDo: check assertion
567 ASSERT(run_queue_hd != (StgTSO*)NULL &&
568 run_queue_hd != END_TSO_QUEUE);
570 /* Ignore ContinueThreads for fetching threads (if synchr comm) */
571 if (!RtsFlags.GranFlags.DoAsyncFetch &&
572 procStatus[CurrentProc]==Fetching) {
573 belch("ghuH: Spurious ContinueThread while Fetching ignored; TSO %d (%p) [PE %d]",
574 CurrentTSO->id, CurrentTSO, CurrentProc);
577 /* Ignore ContinueThreads for completed threads */
578 if (CurrentTSO->what_next == ThreadComplete) {
579 belch("ghuH: found a ContinueThread event for completed thread %d (%p) [PE %d] (ignoring ContinueThread)",
580 CurrentTSO->id, CurrentTSO, CurrentProc);
583 /* Ignore ContinueThreads for threads that are being migrated */
584 if (PROCS(CurrentTSO)==Nowhere) {
585 belch("ghuH: trying to run the migrating TSO %d (%p) [PE %d] (ignoring ContinueThread)",
586 CurrentTSO->id, CurrentTSO, CurrentProc);
589 /* The thread should be at the beginning of the run queue */
590 if (CurrentTSO!=run_queue_hds[CurrentProc]) {
591 belch("ghuH: TSO %d (%p) [PE %d] is not at the start of the run_queue when doing a ContinueThread",
592 CurrentTSO->id, CurrentTSO, CurrentProc);
593 break; // run the thread anyway
596 new_event(proc, proc, CurrentTime[proc],
598 (StgTSO*)NULL, (StgClosure*)NULL, (rtsSpark*)NULL);
600 */ /* Catches superfluous CONTINUEs -- should be unnecessary */
601 break; // now actually run the thread; DaH Qu'vam yImuHbej
604 do_the_fetchnode(event);
605 goto next_thread; /* handle next event in event queue */
608 do_the_globalblock(event);
609 goto next_thread; /* handle next event in event queue */
612 do_the_fetchreply(event);
613 goto next_thread; /* handle next event in event queue */
615 case UnblockThread: /* Move from the blocked queue to the tail of */
616 do_the_unblock(event);
617 goto next_thread; /* handle next event in event queue */
619 case ResumeThread: /* Move from the blocked queue to the tail of */
620 /* the runnable queue ( i.e. Qu' SImqa'lu') */
621 event->tso->gran.blocktime +=
622 CurrentTime[CurrentProc] - event->tso->gran.blockedat;
623 do_the_startthread(event);
624 goto next_thread; /* handle next event in event queue */
627 do_the_startthread(event);
628 goto next_thread; /* handle next event in event queue */
631 do_the_movethread(event);
632 goto next_thread; /* handle next event in event queue */
635 do_the_movespark(event);
636 goto next_thread; /* handle next event in event queue */
639 do_the_findwork(event);
640 goto next_thread; /* handle next event in event queue */
643 barf("Illegal event type %u\n", event->evttype);
646 /* This point was scheduler_loop in the old RTS */
648 IF_DEBUG(gran, belch("GRAN: after main switch"));
650 TimeOfLastEvent = CurrentTime[CurrentProc];
651 TimeOfNextEvent = get_time_of_next_event();
652 IgnoreEvents=(TimeOfNextEvent==0); // HWL HACK
653 // CurrentTSO = ThreadQueueHd;
655 IF_DEBUG(gran, belch("GRAN: time of next event is: %ld",
658 if (RtsFlags.GranFlags.Light)
659 GranSimLight_leave_system(event, &ActiveTSO);
661 EndOfTimeSlice = CurrentTime[CurrentProc]+RtsFlags.GranFlags.time_slice;
664 belch("GRAN: end of time-slice is %#lx", EndOfTimeSlice));
666 /* in a GranSim setup the TSO stays on the run queue */
668 /* Take a thread from the run queue. */
669 POP_RUN_QUEUE(t); // take_off_run_queue(t);
672 fprintf(stderr, "GRAN: About to run current thread, which is\n");
675 context_switch = 0; // turned on via GranYield, checking events and time slice
678 DumpGranEvent(GR_SCHEDULE, t));
680 procStatus[CurrentProc] = Busy;
683 if (PendingFetches != END_BF_QUEUE) {
687 /* ToDo: phps merge with spark activation above */
688 /* check whether we have local work and send requests if we have none */
689 if (EMPTY_RUN_QUEUE()) { /* no runnable threads */
690 /* :-[ no local threads => look out for local sparks */
691 /* the spark pool for the current PE */
692 pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable
693 if (advisory_thread_count < RtsFlags.ParFlags.maxThreads &&
694 pool->hd < pool->tl) {
696 * ToDo: add GC code check that we really have enough heap afterwards!!
698 * If we're here (no runnable threads) and we have pending
699 * sparks, we must have a space problem. Get enough space
700 * to turn one of those pending sparks into a
704 spark = findSpark(rtsFalse); /* get a spark */
705 if (spark != (rtsSpark) NULL) {
706 tso = activateSpark(spark); /* turn the spark into a thread */
707 IF_PAR_DEBUG(schedule,
708 belch("==== schedule: Created TSO %d (%p); %d threads active",
709 tso->id, tso, advisory_thread_count));
711 if (tso==END_TSO_QUEUE) { /* failed to activate spark->back to loop */
712 belch("==^^ failed to activate spark");
714 } /* otherwise fall through & pick-up new tso */
716 IF_PAR_DEBUG(verbose,
717 belch("==^^ no local sparks (spark pool contains only NFs: %d)",
718 spark_queue_len(pool)));
723 /* If we still have no work we need to send a FISH to get a spark
726 if (EMPTY_RUN_QUEUE()) {
727 /* =8-[ no local sparks => look for work on other PEs */
729 * We really have absolutely no work. Send out a fish
730 * (there may be some out there already), and wait for
731 * something to arrive. We clearly can't run any threads
732 * until a SCHEDULE or RESUME arrives, and so that's what
733 * we're hoping to see. (Of course, we still have to
734 * respond to other types of messages.)
736 TIME now = msTime() /*CURRENT_TIME*/;
737 IF_PAR_DEBUG(verbose,
738 belch("-- now=%ld", now));
739 IF_PAR_DEBUG(verbose,
740 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
741 (last_fish_arrived_at!=0 &&
742 last_fish_arrived_at+RtsFlags.ParFlags.fishDelay > now)) {
743 belch("--$$ delaying FISH until %ld (last fish %ld, delay %ld, now %ld)",
744 last_fish_arrived_at+RtsFlags.ParFlags.fishDelay,
745 last_fish_arrived_at,
746 RtsFlags.ParFlags.fishDelay, now);
749 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
750 (last_fish_arrived_at==0 ||
751 (last_fish_arrived_at+RtsFlags.ParFlags.fishDelay <= now))) {
752 /* outstandingFishes is set in sendFish, processFish;
753 avoid flooding system with fishes via delay */
755 sendFish(pe, mytid, NEW_FISH_AGE, NEW_FISH_HISTORY,
758 // Global statistics: count no. of fishes
759 if (RtsFlags.ParFlags.ParStats.Global &&
760 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
761 globalParStats.tot_fish_mess++;
765 receivedFinish = processMessages();
768 } else if (PacketsWaiting()) { /* Look for incoming messages */
769 receivedFinish = processMessages();
772 /* Now we are sure that we have some work available */
773 ASSERT(run_queue_hd != END_TSO_QUEUE);
775 /* Take a thread from the run queue, if we have work */
776 POP_RUN_QUEUE(t); // take_off_run_queue(END_TSO_QUEUE);
777 IF_DEBUG(sanity,checkTSO(t));
779 /* ToDo: write something to the log-file
780 if (RTSflags.ParFlags.granSimStats && !sameThread)
781 DumpGranEvent(GR_SCHEDULE, RunnableThreadsHd);
785 /* the spark pool for the current PE */
786 pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable
789 belch("--=^ %d threads, %d sparks on [%#x]",
790 run_queue_len(), spark_queue_len(pool), CURRENT_PROC));
793 if (0 && RtsFlags.ParFlags.ParStats.Full &&
794 t && LastTSO && t->id != LastTSO->id &&
795 LastTSO->why_blocked == NotBlocked &&
796 LastTSO->what_next != ThreadComplete) {
797 // if previously scheduled TSO not blocked we have to record the context switch
798 DumpVeryRawGranEvent(TimeOfLastYield, CURRENT_PROC, CURRENT_PROC,
799 GR_DESCHEDULE, LastTSO, (StgClosure *)NULL, 0, 0);
802 if (RtsFlags.ParFlags.ParStats.Full &&
803 (emitSchedule /* forced emit */ ||
804 (t && LastTSO && t->id != LastTSO->id))) {
806 we are running a different TSO, so write a schedule event to log file
807 NB: If we use fair scheduling we also have to write a deschedule
808 event for LastTSO; with unfair scheduling we know that the
809 previous tso has blocked whenever we switch to another tso, so
810 we don't need it in GUM for now
812 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
813 GR_SCHEDULE, t, (StgClosure *)NULL, 0, 0);
814 emitSchedule = rtsFalse;
818 #else /* !GRAN && !PAR */
820 // grab a thread from the run queue
821 ASSERT(run_queue_hd != END_TSO_QUEUE);
824 // Sanity check the thread we're about to run. This can be
825 // expensive if there is lots of thread switching going on...
826 IF_DEBUG(sanity,checkTSO(t));
831 StgMainThread *m = t->main;
838 sched_belch("### Running thread %d in bound thread", t->id));
839 // yes, the Haskell thread is bound to the current native thread
844 sched_belch("### thread %d bound to another OS thread", t->id));
845 // no, bound to a different Haskell thread: pass to that thread
846 PUSH_ON_RUN_QUEUE(t);
847 passCapability(&m->bound_thread_cond);
853 if(mainThread != NULL)
854 // The thread we want to run is bound.
857 sched_belch("### this OS thread cannot run thread %d", t->id));
858 // no, the current native thread is bound to a different
859 // Haskell thread, so pass it to any worker thread
860 PUSH_ON_RUN_QUEUE(t);
861 passCapabilityToWorker();
868 cap->r.rCurrentTSO = t;
870 /* context switches are now initiated by the timer signal, unless
871 * the user specified "context switch as often as possible", with
874 if ((RtsFlags.ConcFlags.ctxtSwitchTicks == 0
875 && (run_queue_hd != END_TSO_QUEUE
876 || blocked_queue_hd != END_TSO_QUEUE
877 || sleeping_queue != END_TSO_QUEUE)))
884 RELEASE_LOCK(&sched_mutex);
886 IF_DEBUG(scheduler, sched_belch("-->> running thread %ld %s ...",
887 t->id, whatNext_strs[t->what_next]));
890 startHeapProfTimer();
893 /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
894 /* Run the current thread
896 prev_what_next = t->what_next;
897 switch (prev_what_next) {
900 /* Thread already finished, return to scheduler. */
901 ret = ThreadFinished;
904 errno = t->saved_errno;
905 ret = StgRun((StgFunPtr) stg_returnToStackTop, &cap->r);
906 t->saved_errno = errno;
908 case ThreadInterpret:
909 ret = interpretBCO(cap);
912 barf("schedule: invalid what_next field");
914 /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
916 /* Costs for the scheduler are assigned to CCS_SYSTEM */
922 ACQUIRE_LOCK(&sched_mutex);
924 #ifdef RTS_SUPPORTS_THREADS
925 IF_DEBUG(scheduler,fprintf(stderr,"sched (task %p): ", osThreadId()););
926 #elif !defined(GRAN) && !defined(PAR)
927 IF_DEBUG(scheduler,fprintf(stderr,"sched: "););
929 t = cap->r.rCurrentTSO;
932 /* HACK 675: if the last thread didn't yield, make sure to print a
933 SCHEDULE event to the log file when StgRunning the next thread, even
934 if it is the same one as before */
936 TimeOfLastYield = CURRENT_TIME;
942 IF_DEBUG(gran, DumpGranEvent(GR_DESCHEDULE, t));
943 globalGranStats.tot_heapover++;
945 globalParStats.tot_heapover++;
948 // did the task ask for a large block?
949 if (cap->r.rHpAlloc > BLOCK_SIZE_W) {
950 // if so, get one and push it on the front of the nursery.
954 blocks = (nat)BLOCK_ROUND_UP(cap->r.rHpAlloc * sizeof(W_)) / BLOCK_SIZE;
956 IF_DEBUG(scheduler,belch("--<< thread %ld (%s) stopped: requesting a large block (size %d)",
957 t->id, whatNext_strs[t->what_next], blocks));
959 // don't do this if it would push us over the
960 // alloc_blocks_lim limit; we'll GC first.
961 if (alloc_blocks + blocks < alloc_blocks_lim) {
963 alloc_blocks += blocks;
964 bd = allocGroup( blocks );
966 // link the new group into the list
967 bd->link = cap->r.rCurrentNursery;
968 bd->u.back = cap->r.rCurrentNursery->u.back;
969 if (cap->r.rCurrentNursery->u.back != NULL) {
970 cap->r.rCurrentNursery->u.back->link = bd;
972 ASSERT(g0s0->blocks == cap->r.rCurrentNursery &&
973 g0s0->blocks == cap->r.rNursery);
974 cap->r.rNursery = g0s0->blocks = bd;
976 cap->r.rCurrentNursery->u.back = bd;
978 // initialise it as a nursery block. We initialise the
979 // step, gen_no, and flags field of *every* sub-block in
980 // this large block, because this is easier than making
981 // sure that we always find the block head of a large
982 // block whenever we call Bdescr() (eg. evacuate() and
983 // isAlive() in the GC would both have to do this, at
987 for (x = bd; x < bd + blocks; x++) {
994 // don't forget to update the block count in g0s0.
995 g0s0->n_blocks += blocks;
996 // This assert can be a killer if the app is doing lots
997 // of large block allocations.
998 ASSERT(countBlocks(g0s0->blocks) == g0s0->n_blocks);
1000 // now update the nursery to point to the new block
1001 cap->r.rCurrentNursery = bd;
1003 // we might be unlucky and have another thread get on the
1004 // run queue before us and steal the large block, but in that
1005 // case the thread will just end up requesting another large
1007 PUSH_ON_RUN_QUEUE(t);
1012 /* make all the running tasks block on a condition variable,
1013 * maybe set context_switch and wait till they all pile in,
1014 * then have them wait on a GC condition variable.
1016 IF_DEBUG(scheduler,belch("--<< thread %ld (%s) stopped: HeapOverflow",
1017 t->id, whatNext_strs[t->what_next]));
1020 ASSERT(!is_on_queue(t,CurrentProc));
1022 /* Currently we emit a DESCHEDULE event before GC in GUM.
1023 ToDo: either add separate event to distinguish SYSTEM time from rest
1024 or just nuke this DESCHEDULE (and the following SCHEDULE) */
1025 if (0 && RtsFlags.ParFlags.ParStats.Full) {
1026 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
1027 GR_DESCHEDULE, t, (StgClosure *)NULL, 0, 0);
1028 emitSchedule = rtsTrue;
1032 ready_to_gc = rtsTrue;
1033 context_switch = 1; /* stop other threads ASAP */
1034 PUSH_ON_RUN_QUEUE(t);
1035 /* actual GC is done at the end of the while loop */
1041 DumpGranEvent(GR_DESCHEDULE, t));
1042 globalGranStats.tot_stackover++;
1045 // DumpGranEvent(GR_DESCHEDULE, t);
1046 globalParStats.tot_stackover++;
1048 IF_DEBUG(scheduler,belch("--<< thread %ld (%s) stopped, StackOverflow",
1049 t->id, whatNext_strs[t->what_next]));
1050 /* just adjust the stack for this thread, then pop it back
1055 /* enlarge the stack */
1056 StgTSO *new_t = threadStackOverflow(t);
1058 /* This TSO has moved, so update any pointers to it from the
1059 * main thread stack. It better not be on any other queues...
1060 * (it shouldn't be).
1062 if (t->main != NULL) {
1063 t->main->tso = new_t;
1065 threadPaused(new_t);
1066 PUSH_ON_RUN_QUEUE(new_t);
1070 case ThreadYielding:
1073 DumpGranEvent(GR_DESCHEDULE, t));
1074 globalGranStats.tot_yields++;
1077 // DumpGranEvent(GR_DESCHEDULE, t);
1078 globalParStats.tot_yields++;
1080 /* put the thread back on the run queue. Then, if we're ready to
1081 * GC, check whether this is the last task to stop. If so, wake
1082 * up the GC thread. getThread will block during a GC until the
1086 if (t->what_next != prev_what_next) {
1087 belch("--<< thread %ld (%s) stopped to switch evaluators",
1088 t->id, whatNext_strs[t->what_next]);
1090 belch("--<< thread %ld (%s) stopped, yielding",
1091 t->id, whatNext_strs[t->what_next]);
1096 //belch("&& Doing sanity check on yielding TSO %ld.", t->id);
1098 ASSERT(t->link == END_TSO_QUEUE);
1100 // Shortcut if we're just switching evaluators: don't bother
1101 // doing stack squeezing (which can be expensive), just run the
1103 if (t->what_next != prev_what_next) {
1110 ASSERT(!is_on_queue(t,CurrentProc));
1113 //belch("&& Doing sanity check on all ThreadQueues (and their TSOs).");
1114 checkThreadQsSanity(rtsTrue));
1118 if (RtsFlags.ParFlags.doFairScheduling) {
1119 /* this does round-robin scheduling; good for concurrency */
1120 APPEND_TO_RUN_QUEUE(t);
1122 /* this does unfair scheduling; good for parallelism */
1123 PUSH_ON_RUN_QUEUE(t);
1126 // this does round-robin scheduling; good for concurrency
1127 APPEND_TO_RUN_QUEUE(t);
1131 /* add a ContinueThread event to actually process the thread */
1132 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1134 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1136 belch("GRAN: eventq and runnableq after adding yielded thread to queue again:");
1145 belch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: ",
1146 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)));
1147 if (t->block_info.closure!=(StgClosure*)NULL) print_bq(t->block_info.closure));
1149 // ??? needed; should emit block before
1151 DumpGranEvent(GR_DESCHEDULE, t));
1152 prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
1155 ASSERT(procStatus[CurrentProc]==Busy ||
1156 ((procStatus[CurrentProc]==Fetching) &&
1157 (t->block_info.closure!=(StgClosure*)NULL)));
1158 if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
1159 !(!RtsFlags.GranFlags.DoAsyncFetch &&
1160 procStatus[CurrentProc]==Fetching))
1161 procStatus[CurrentProc] = Idle;
1165 belch("--<< thread %ld (%p; %s) stopped, blocking on node %p with BQ: ",
1166 t->id, t, whatNext_strs[t->what_next], t->block_info.closure));
1169 if (t->block_info.closure!=(StgClosure*)NULL)
1170 print_bq(t->block_info.closure));
1172 /* Send a fetch (if BlockedOnGA) and dump event to log file */
1175 /* whatever we schedule next, we must log that schedule */
1176 emitSchedule = rtsTrue;
1179 /* don't need to do anything. Either the thread is blocked on
1180 * I/O, in which case we'll have called addToBlockedQueue
1181 * previously, or it's blocked on an MVar or Blackhole, in which
1182 * case it'll be on the relevant queue already.
1185 fprintf(stderr, "--<< thread %d (%s) stopped: ",
1186 t->id, whatNext_strs[t->what_next]);
1187 printThreadBlockage(t);
1188 fprintf(stderr, "\n"));
1191 /* Only for dumping event to log file
1192 ToDo: do I need this in GranSim, too?
1199 case ThreadFinished:
1200 /* Need to check whether this was a main thread, and if so, signal
1201 * the task that started it with the return value. If we have no
1202 * more main threads, we probably need to stop all the tasks until
1205 /* We also end up here if the thread kills itself with an
1206 * uncaught exception, see Exception.hc.
1208 IF_DEBUG(scheduler,belch("--++ thread %d (%s) finished",
1209 t->id, whatNext_strs[t->what_next]));
1211 endThread(t, CurrentProc); // clean-up the thread
1213 /* For now all are advisory -- HWL */
1214 //if(t->priority==AdvisoryPriority) ??
1215 advisory_thread_count--;
1218 if(t->dist.priority==RevalPriority)
1222 if (RtsFlags.ParFlags.ParStats.Full &&
1223 !RtsFlags.ParFlags.ParStats.Suppressed)
1224 DumpEndEvent(CURRENT_PROC, t, rtsFalse /* not mandatory */);
1228 // Check whether the thread that just completed was a main
1229 // thread, and if so return with the result.
1231 // There is an assumption here that all thread completion goes
1232 // through this point; we need to make sure that if a thread
1233 // ends up in the ThreadKilled state, that it stays on the run
1234 // queue so it can be dealt with here.
1237 #if defined(RTS_SUPPORTS_THREADS)
1240 mainThread->tso == t
1244 // We are a bound thread: this must be our thread that just
1246 ASSERT(mainThread->tso == t);
1248 if (t->what_next == ThreadComplete) {
1249 if (mainThread->ret) {
1250 // NOTE: return val is tso->sp[1] (see StgStartup.hc)
1251 *(mainThread->ret) = (StgClosure *)mainThread->tso->sp[1];
1253 mainThread->stat = Success;
1255 if (mainThread->ret) {
1256 *(mainThread->ret) = NULL;
1258 if (was_interrupted) {
1259 mainThread->stat = Interrupted;
1261 mainThread->stat = Killed;
1265 removeThreadLabel((StgWord)mainThread->tso->id);
1267 if (mainThread->prev == NULL) {
1268 main_threads = mainThread->link;
1270 mainThread->prev->link = mainThread->link;
1272 if (mainThread->link != NULL) {
1273 mainThread->link->prev = NULL;
1275 releaseCapability(cap);
1279 #ifdef RTS_SUPPORTS_THREADS
1280 ASSERT(t->main == NULL);
1282 if (t->main != NULL) {
1283 // Must be a main thread that is not the topmost one. Leave
1284 // it on the run queue until the stack has unwound to the
1285 // point where we can deal with this. Leaving it on the run
1286 // queue also ensures that the garbage collector knows about
1287 // this thread and its return value (it gets dropped from the
1288 // all_threads list so there's no other way to find it).
1289 APPEND_TO_RUN_QUEUE(t);
1295 barf("schedule: invalid thread return code %d", (int)ret);
1299 // When we have +RTS -i0 and we're heap profiling, do a census at
1300 // every GC. This lets us get repeatable runs for debugging.
1301 if (performHeapProfile ||
1302 (RtsFlags.ProfFlags.profileInterval==0 &&
1303 RtsFlags.ProfFlags.doHeapProfile && ready_to_gc)) {
1304 GarbageCollect(GetRoots, rtsTrue);
1306 performHeapProfile = rtsFalse;
1307 ready_to_gc = rtsFalse; // we already GC'd
1312 /* everybody back, start the GC.
1313 * Could do it in this thread, or signal a condition var
1314 * to do it in another thread. Either way, we need to
1315 * broadcast on gc_pending_cond afterward.
1317 #if defined(RTS_SUPPORTS_THREADS)
1318 IF_DEBUG(scheduler,sched_belch("doing GC"));
1320 GarbageCollect(GetRoots,rtsFalse);
1321 ready_to_gc = rtsFalse;
1323 /* add a ContinueThread event to continue execution of current thread */
1324 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1326 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1328 fprintf(stderr, "GRAN: eventq and runnableq after Garbage collection:\n");
1336 IF_GRAN_DEBUG(unused,
1337 print_eventq(EventHd));
1339 event = get_next_event();
1342 /* ToDo: wait for next message to arrive rather than busy wait */
1345 } /* end of while(1) */
1347 IF_PAR_DEBUG(verbose,
1348 belch("== Leaving schedule() after having received Finish"));
1351 /* ---------------------------------------------------------------------------
1352 * rtsSupportsBoundThreads(): is the RTS built to support bound threads?
1353 * used by Control.Concurrent for error checking.
1354 * ------------------------------------------------------------------------- */
1357 rtsSupportsBoundThreads(void)
1366 /* ---------------------------------------------------------------------------
1367 * isThreadBound(tso): check whether tso is bound to an OS thread.
1368 * ------------------------------------------------------------------------- */
1371 isThreadBound(StgTSO* tso USED_IN_THREADED_RTS)
1374 return (tso->main != NULL);
1379 /* ---------------------------------------------------------------------------
1380 * Singleton fork(). Do not copy any running threads.
1381 * ------------------------------------------------------------------------- */
1383 #ifndef mingw32_TARGET_OS
1384 #define FORKPROCESS_PRIMOP_SUPPORTED
1387 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
1389 deleteThreadImmediately(StgTSO *tso);
1392 forkProcess(HsStablePtr *entry
1393 #ifndef FORKPROCESS_PRIMOP_SUPPORTED
1398 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
1404 IF_DEBUG(scheduler,sched_belch("forking!"));
1405 rts_lock(); // This not only acquires sched_mutex, it also
1406 // makes sure that no other threads are running
1410 if (pid) { /* parent */
1412 /* just return the pid */
1416 } else { /* child */
1419 // delete all threads
1420 run_queue_hd = run_queue_tl = END_TSO_QUEUE;
1422 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
1425 // don't allow threads to catch the ThreadKilled exception
1426 deleteThreadImmediately(t);
1429 // wipe the main thread list
1430 while((m = main_threads) != NULL) {
1431 main_threads = m->link;
1432 # ifdef THREADED_RTS
1433 closeCondition(&m->bound_thread_cond);
1438 # ifdef RTS_SUPPORTS_THREADS
1439 resetTaskManagerAfterFork(); // tell startTask() and friends that
1440 startingWorkerThread = rtsFalse; // we have no worker threads any more
1441 resetWorkerWakeupPipeAfterFork();
1444 rc = rts_evalStableIO(entry, NULL); // run the action
1445 rts_checkSchedStatus("forkProcess",rc);
1449 hs_exit(); // clean up and exit
1452 #else /* !FORKPROCESS_PRIMOP_SUPPORTED */
1453 barf("forkProcess#: primop not supported, sorry!\n");
1458 /* ---------------------------------------------------------------------------
1459 * deleteAllThreads(): kill all the live threads.
1461 * This is used when we catch a user interrupt (^C), before performing
1462 * any necessary cleanups and running finalizers.
1464 * Locks: sched_mutex held.
1465 * ------------------------------------------------------------------------- */
1468 deleteAllThreads ( void )
1471 IF_DEBUG(scheduler,sched_belch("deleting all threads"));
1472 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
1473 next = t->global_link;
1476 run_queue_hd = run_queue_tl = END_TSO_QUEUE;
1477 blocked_queue_hd = blocked_queue_tl = END_TSO_QUEUE;
1478 sleeping_queue = END_TSO_QUEUE;
1481 /* startThread and insertThread are now in GranSim.c -- HWL */
1484 /* ---------------------------------------------------------------------------
1485 * Suspending & resuming Haskell threads.
1487 * When making a "safe" call to C (aka _ccall_GC), the task gives back
1488 * its capability before calling the C function. This allows another
1489 * task to pick up the capability and carry on running Haskell
1490 * threads. It also means that if the C call blocks, it won't lock
1493 * The Haskell thread making the C call is put to sleep for the
1494 * duration of the call, on the susepended_ccalling_threads queue. We
1495 * give out a token to the task, which it can use to resume the thread
1496 * on return from the C function.
1497 * ------------------------------------------------------------------------- */
1500 suspendThread( StgRegTable *reg,
1509 int saved_errno = errno;
1511 /* assume that *reg is a pointer to the StgRegTable part
1514 cap = (Capability *)((void *)((unsigned char*)reg - sizeof(StgFunTable)));
1516 ACQUIRE_LOCK(&sched_mutex);
1519 sched_belch("thread %d did a _ccall_gc (is_concurrent: %d)", cap->r.rCurrentTSO->id,concCall));
1521 // XXX this might not be necessary --SDM
1522 cap->r.rCurrentTSO->what_next = ThreadRunGHC;
1524 threadPaused(cap->r.rCurrentTSO);
1525 cap->r.rCurrentTSO->link = suspended_ccalling_threads;
1526 suspended_ccalling_threads = cap->r.rCurrentTSO;
1528 if(cap->r.rCurrentTSO->blocked_exceptions == NULL) {
1529 cap->r.rCurrentTSO->why_blocked = BlockedOnCCall;
1530 cap->r.rCurrentTSO->blocked_exceptions = END_TSO_QUEUE;
1532 cap->r.rCurrentTSO->why_blocked = BlockedOnCCall_NoUnblockExc;
1535 /* Use the thread ID as the token; it should be unique */
1536 tok = cap->r.rCurrentTSO->id;
1538 /* Hand back capability */
1539 releaseCapability(cap);
1541 #if defined(RTS_SUPPORTS_THREADS)
1542 /* Preparing to leave the RTS, so ensure there's a native thread/task
1543 waiting to take over.
1545 IF_DEBUG(scheduler, sched_belch("worker (token %d): leaving RTS", tok));
1548 /* Other threads _might_ be available for execution; signal this */
1550 RELEASE_LOCK(&sched_mutex);
1552 errno = saved_errno;
1557 resumeThread( StgInt tok,
1558 rtsBool concCall STG_UNUSED )
1560 StgTSO *tso, **prev;
1562 int saved_errno = errno;
1564 #if defined(RTS_SUPPORTS_THREADS)
1565 /* Wait for permission to re-enter the RTS with the result. */
1566 ACQUIRE_LOCK(&sched_mutex);
1567 waitForReturnCapability(&sched_mutex, &cap);
1569 IF_DEBUG(scheduler, sched_belch("worker (token %d): re-entering RTS", tok));
1571 grabCapability(&cap);
1574 /* Remove the thread off of the suspended list */
1575 prev = &suspended_ccalling_threads;
1576 for (tso = suspended_ccalling_threads;
1577 tso != END_TSO_QUEUE;
1578 prev = &tso->link, tso = tso->link) {
1579 if (tso->id == (StgThreadID)tok) {
1584 if (tso == END_TSO_QUEUE) {
1585 barf("resumeThread: thread not found");
1587 tso->link = END_TSO_QUEUE;
1589 if(tso->why_blocked == BlockedOnCCall) {
1590 awakenBlockedQueueNoLock(tso->blocked_exceptions);
1591 tso->blocked_exceptions = NULL;
1594 /* Reset blocking status */
1595 tso->why_blocked = NotBlocked;
1597 cap->r.rCurrentTSO = tso;
1598 RELEASE_LOCK(&sched_mutex);
1599 errno = saved_errno;
1604 /* ---------------------------------------------------------------------------
1606 * ------------------------------------------------------------------------ */
1607 static void unblockThread(StgTSO *tso);
1609 /* ---------------------------------------------------------------------------
1610 * Comparing Thread ids.
1612 * This is used from STG land in the implementation of the
1613 * instances of Eq/Ord for ThreadIds.
1614 * ------------------------------------------------------------------------ */
1617 cmp_thread(StgPtr tso1, StgPtr tso2)
1619 StgThreadID id1 = ((StgTSO *)tso1)->id;
1620 StgThreadID id2 = ((StgTSO *)tso2)->id;
1622 if (id1 < id2) return (-1);
1623 if (id1 > id2) return 1;
1627 /* ---------------------------------------------------------------------------
1628 * Fetching the ThreadID from an StgTSO.
1630 * This is used in the implementation of Show for ThreadIds.
1631 * ------------------------------------------------------------------------ */
1633 rts_getThreadId(StgPtr tso)
1635 return ((StgTSO *)tso)->id;
1640 labelThread(StgPtr tso, char *label)
1645 /* Caveat: Once set, you can only set the thread name to "" */
1646 len = strlen(label)+1;
1647 buf = stgMallocBytes(len * sizeof(char), "Schedule.c:labelThread()");
1648 strncpy(buf,label,len);
1649 /* Update will free the old memory for us */
1650 updateThreadLabel(((StgTSO *)tso)->id,buf);
1654 /* ---------------------------------------------------------------------------
1655 Create a new thread.
1657 The new thread starts with the given stack size. Before the
1658 scheduler can run, however, this thread needs to have a closure
1659 (and possibly some arguments) pushed on its stack. See
1660 pushClosure() in Schedule.h.
1662 createGenThread() and createIOThread() (in SchedAPI.h) are
1663 convenient packaged versions of this function.
1665 currently pri (priority) is only used in a GRAN setup -- HWL
1666 ------------------------------------------------------------------------ */
1668 /* currently pri (priority) is only used in a GRAN setup -- HWL */
1670 createThread(nat size, StgInt pri)
1673 createThread(nat size)
1680 /* First check whether we should create a thread at all */
1682 /* check that no more than RtsFlags.ParFlags.maxThreads threads are created */
1683 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads) {
1685 belch("{createThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
1686 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
1687 return END_TSO_QUEUE;
1693 ASSERT(!RtsFlags.GranFlags.Light || CurrentProc==0);
1696 // ToDo: check whether size = stack_size - TSO_STRUCT_SIZEW
1698 /* catch ridiculously small stack sizes */
1699 if (size < MIN_STACK_WORDS + TSO_STRUCT_SIZEW) {
1700 size = MIN_STACK_WORDS + TSO_STRUCT_SIZEW;
1703 stack_size = size - TSO_STRUCT_SIZEW;
1705 tso = (StgTSO *)allocate(size);
1706 TICK_ALLOC_TSO(stack_size, 0);
1708 SET_HDR(tso, &stg_TSO_info, CCS_SYSTEM);
1710 SET_GRAN_HDR(tso, ThisPE);
1713 // Always start with the compiled code evaluator
1714 tso->what_next = ThreadRunGHC;
1716 tso->id = next_thread_id++;
1717 tso->why_blocked = NotBlocked;
1718 tso->blocked_exceptions = NULL;
1720 tso->saved_errno = 0;
1723 tso->stack_size = stack_size;
1724 tso->max_stack_size = round_to_mblocks(RtsFlags.GcFlags.maxStkSize)
1726 tso->sp = (P_)&(tso->stack) + stack_size;
1729 tso->prof.CCCS = CCS_MAIN;
1732 /* put a stop frame on the stack */
1733 tso->sp -= sizeofW(StgStopFrame);
1734 SET_HDR((StgClosure*)tso->sp,(StgInfoTable *)&stg_stop_thread_info,CCS_SYSTEM);
1737 tso->link = END_TSO_QUEUE;
1738 /* uses more flexible routine in GranSim */
1739 insertThread(tso, CurrentProc);
1741 /* In a non-GranSim setup the pushing of a TSO onto the runq is separated
1747 if (RtsFlags.GranFlags.GranSimStats.Full)
1748 DumpGranEvent(GR_START,tso);
1750 if (RtsFlags.ParFlags.ParStats.Full)
1751 DumpGranEvent(GR_STARTQ,tso);
1752 /* HACk to avoid SCHEDULE
1756 /* Link the new thread on the global thread list.
1758 tso->global_link = all_threads;
1762 tso->dist.priority = MandatoryPriority; //by default that is...
1766 tso->gran.pri = pri;
1768 tso->gran.magic = TSO_MAGIC; // debugging only
1770 tso->gran.sparkname = 0;
1771 tso->gran.startedat = CURRENT_TIME;
1772 tso->gran.exported = 0;
1773 tso->gran.basicblocks = 0;
1774 tso->gran.allocs = 0;
1775 tso->gran.exectime = 0;
1776 tso->gran.fetchtime = 0;
1777 tso->gran.fetchcount = 0;
1778 tso->gran.blocktime = 0;
1779 tso->gran.blockcount = 0;
1780 tso->gran.blockedat = 0;
1781 tso->gran.globalsparks = 0;
1782 tso->gran.localsparks = 0;
1783 if (RtsFlags.GranFlags.Light)
1784 tso->gran.clock = Now; /* local clock */
1786 tso->gran.clock = 0;
1788 IF_DEBUG(gran,printTSO(tso));
1791 tso->par.magic = TSO_MAGIC; // debugging only
1793 tso->par.sparkname = 0;
1794 tso->par.startedat = CURRENT_TIME;
1795 tso->par.exported = 0;
1796 tso->par.basicblocks = 0;
1797 tso->par.allocs = 0;
1798 tso->par.exectime = 0;
1799 tso->par.fetchtime = 0;
1800 tso->par.fetchcount = 0;
1801 tso->par.blocktime = 0;
1802 tso->par.blockcount = 0;
1803 tso->par.blockedat = 0;
1804 tso->par.globalsparks = 0;
1805 tso->par.localsparks = 0;
1809 globalGranStats.tot_threads_created++;
1810 globalGranStats.threads_created_on_PE[CurrentProc]++;
1811 globalGranStats.tot_sq_len += spark_queue_len(CurrentProc);
1812 globalGranStats.tot_sq_probes++;
1814 // collect parallel global statistics (currently done together with GC stats)
1815 if (RtsFlags.ParFlags.ParStats.Global &&
1816 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
1817 //fprintf(stderr, "Creating thread %d @ %11.2f\n", tso->id, usertime());
1818 globalParStats.tot_threads_created++;
1824 belch("==__ schedule: Created TSO %d (%p);",
1825 CurrentProc, tso, tso->id));
1827 IF_PAR_DEBUG(verbose,
1828 belch("==__ schedule: Created TSO %d (%p); %d threads active",
1829 tso->id, tso, advisory_thread_count));
1831 IF_DEBUG(scheduler,sched_belch("created thread %ld, stack size = %lx words",
1832 tso->id, tso->stack_size));
1839 all parallel thread creation calls should fall through the following routine.
1842 createSparkThread(rtsSpark spark)
1844 ASSERT(spark != (rtsSpark)NULL);
1845 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads)
1847 barf("{createSparkThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
1848 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
1849 return END_TSO_QUEUE;
1853 tso = createThread(RtsFlags.GcFlags.initialStkSize);
1854 if (tso==END_TSO_QUEUE)
1855 barf("createSparkThread: Cannot create TSO");
1857 tso->priority = AdvisoryPriority;
1859 pushClosure(tso,spark);
1860 PUSH_ON_RUN_QUEUE(tso);
1861 advisory_thread_count++;
1868 Turn a spark into a thread.
1869 ToDo: fix for SMP (needs to acquire SCHED_MUTEX!)
1873 activateSpark (rtsSpark spark)
1877 tso = createSparkThread(spark);
1878 if (RtsFlags.ParFlags.ParStats.Full) {
1879 //ASSERT(run_queue_hd == END_TSO_QUEUE); // I think ...
1880 IF_PAR_DEBUG(verbose,
1881 belch("==^^ activateSpark: turning spark of closure %p (%s) into a thread",
1882 (StgClosure *)spark, info_type((StgClosure *)spark)));
1884 // ToDo: fwd info on local/global spark to thread -- HWL
1885 // tso->gran.exported = spark->exported;
1886 // tso->gran.locked = !spark->global;
1887 // tso->gran.sparkname = spark->name;
1893 static SchedulerStatus waitThread_(/*out*/StgMainThread* m,
1894 Capability *initialCapability
1898 /* ---------------------------------------------------------------------------
1901 * scheduleThread puts a thread on the head of the runnable queue.
1902 * This will usually be done immediately after a thread is created.
1903 * The caller of scheduleThread must create the thread using e.g.
1904 * createThread and push an appropriate closure
1905 * on this thread's stack before the scheduler is invoked.
1906 * ------------------------------------------------------------------------ */
1908 static void scheduleThread_ (StgTSO* tso);
1911 scheduleThread_(StgTSO *tso)
1913 // Precondition: sched_mutex must be held.
1914 PUSH_ON_RUN_QUEUE(tso);
1919 scheduleThread(StgTSO* tso)
1921 ACQUIRE_LOCK(&sched_mutex);
1922 scheduleThread_(tso);
1923 RELEASE_LOCK(&sched_mutex);
1926 #if defined(RTS_SUPPORTS_THREADS)
1927 static Condition bound_cond_cache;
1928 static int bound_cond_cache_full = 0;
1933 scheduleWaitThread(StgTSO* tso, /*[out]*/HaskellObj* ret,
1934 Capability *initialCapability)
1936 // Precondition: sched_mutex must be held
1939 m = stgMallocBytes(sizeof(StgMainThread), "waitThread");
1944 m->link = main_threads;
1946 if (main_threads != NULL) {
1947 main_threads->prev = m;
1951 #if defined(RTS_SUPPORTS_THREADS)
1952 // Allocating a new condition for each thread is expensive, so we
1953 // cache one. This is a pretty feeble hack, but it helps speed up
1954 // consecutive call-ins quite a bit.
1955 if (bound_cond_cache_full) {
1956 m->bound_thread_cond = bound_cond_cache;
1957 bound_cond_cache_full = 0;
1959 initCondition(&m->bound_thread_cond);
1963 /* Put the thread on the main-threads list prior to scheduling the TSO.
1964 Failure to do so introduces a race condition in the MT case (as
1965 identified by Wolfgang Thaller), whereby the new task/OS thread
1966 created by scheduleThread_() would complete prior to the thread
1967 that spawned it managed to put 'itself' on the main-threads list.
1968 The upshot of it all being that the worker thread wouldn't get to
1969 signal the completion of the its work item for the main thread to
1970 see (==> it got stuck waiting.) -- sof 6/02.
1972 IF_DEBUG(scheduler, sched_belch("waiting for thread (%d)", tso->id));
1974 PUSH_ON_RUN_QUEUE(tso);
1975 // NB. Don't call THREAD_RUNNABLE() here, because the thread is
1976 // bound and only runnable by *this* OS thread, so waking up other
1977 // workers will just slow things down.
1979 return waitThread_(m, initialCapability);
1982 /* ---------------------------------------------------------------------------
1985 * Initialise the scheduler. This resets all the queues - if the
1986 * queues contained any threads, they'll be garbage collected at the
1989 * ------------------------------------------------------------------------ */
1997 for (i=0; i<=MAX_PROC; i++) {
1998 run_queue_hds[i] = END_TSO_QUEUE;
1999 run_queue_tls[i] = END_TSO_QUEUE;
2000 blocked_queue_hds[i] = END_TSO_QUEUE;
2001 blocked_queue_tls[i] = END_TSO_QUEUE;
2002 ccalling_threadss[i] = END_TSO_QUEUE;
2003 sleeping_queue = END_TSO_QUEUE;
2006 run_queue_hd = END_TSO_QUEUE;
2007 run_queue_tl = END_TSO_QUEUE;
2008 blocked_queue_hd = END_TSO_QUEUE;
2009 blocked_queue_tl = END_TSO_QUEUE;
2010 sleeping_queue = END_TSO_QUEUE;
2013 suspended_ccalling_threads = END_TSO_QUEUE;
2015 main_threads = NULL;
2016 all_threads = END_TSO_QUEUE;
2021 RtsFlags.ConcFlags.ctxtSwitchTicks =
2022 RtsFlags.ConcFlags.ctxtSwitchTime / TICK_MILLISECS;
2024 #if defined(RTS_SUPPORTS_THREADS)
2025 /* Initialise the mutex and condition variables used by
2027 initMutex(&sched_mutex);
2028 initMutex(&term_mutex);
2031 ACQUIRE_LOCK(&sched_mutex);
2033 /* A capability holds the state a native thread needs in
2034 * order to execute STG code. At least one capability is
2035 * floating around (only SMP builds have more than one).
2039 #if defined(RTS_SUPPORTS_THREADS)
2040 /* start our haskell execution tasks */
2041 startTaskManager(0,taskStart);
2044 #if /* defined(SMP) ||*/ defined(PAR)
2048 RELEASE_LOCK(&sched_mutex);
2052 exitScheduler( void )
2054 #if defined(RTS_SUPPORTS_THREADS)
2057 shutting_down_scheduler = rtsTrue;
2060 /* ----------------------------------------------------------------------------
2061 Managing the per-task allocation areas.
2063 Each capability comes with an allocation area. These are
2064 fixed-length block lists into which allocation can be done.
2066 ToDo: no support for two-space collection at the moment???
2067 ------------------------------------------------------------------------- */
2071 waitThread_(StgMainThread* m, Capability *initialCapability)
2073 SchedulerStatus stat;
2075 // Precondition: sched_mutex must be held.
2076 IF_DEBUG(scheduler, sched_belch("new main thread (%d)", m->tso->id));
2079 /* GranSim specific init */
2080 CurrentTSO = m->tso; // the TSO to run
2081 procStatus[MainProc] = Busy; // status of main PE
2082 CurrentProc = MainProc; // PE to run it on
2083 schedule(m,initialCapability);
2085 schedule(m,initialCapability);
2086 ASSERT(m->stat != NoStatus);
2091 #if defined(RTS_SUPPORTS_THREADS)
2092 // Free the condition variable, returning it to the cache if possible.
2093 if (!bound_cond_cache_full) {
2094 bound_cond_cache = m->bound_thread_cond;
2095 bound_cond_cache_full = 1;
2097 closeCondition(&m->bound_thread_cond);
2101 IF_DEBUG(scheduler, sched_belch("main thread (%d) finished", m->tso->id));
2104 // Postcondition: sched_mutex still held
2108 /* ---------------------------------------------------------------------------
2109 Where are the roots that we know about?
2111 - all the threads on the runnable queue
2112 - all the threads on the blocked queue
2113 - all the threads on the sleeping queue
2114 - all the thread currently executing a _ccall_GC
2115 - all the "main threads"
2117 ------------------------------------------------------------------------ */
2119 /* This has to be protected either by the scheduler monitor, or by the
2120 garbage collection monitor (probably the latter).
2125 GetRoots( evac_fn evac )
2130 for (i=0; i<=RtsFlags.GranFlags.proc; i++) {
2131 if ((run_queue_hds[i] != END_TSO_QUEUE) && ((run_queue_hds[i] != NULL)))
2132 evac((StgClosure **)&run_queue_hds[i]);
2133 if ((run_queue_tls[i] != END_TSO_QUEUE) && ((run_queue_tls[i] != NULL)))
2134 evac((StgClosure **)&run_queue_tls[i]);
2136 if ((blocked_queue_hds[i] != END_TSO_QUEUE) && ((blocked_queue_hds[i] != NULL)))
2137 evac((StgClosure **)&blocked_queue_hds[i]);
2138 if ((blocked_queue_tls[i] != END_TSO_QUEUE) && ((blocked_queue_tls[i] != NULL)))
2139 evac((StgClosure **)&blocked_queue_tls[i]);
2140 if ((ccalling_threadss[i] != END_TSO_QUEUE) && ((ccalling_threadss[i] != NULL)))
2141 evac((StgClosure **)&ccalling_threads[i]);
2148 if (run_queue_hd != END_TSO_QUEUE) {
2149 ASSERT(run_queue_tl != END_TSO_QUEUE);
2150 evac((StgClosure **)&run_queue_hd);
2151 evac((StgClosure **)&run_queue_tl);
2154 if (blocked_queue_hd != END_TSO_QUEUE) {
2155 ASSERT(blocked_queue_tl != END_TSO_QUEUE);
2156 evac((StgClosure **)&blocked_queue_hd);
2157 evac((StgClosure **)&blocked_queue_tl);
2160 if (sleeping_queue != END_TSO_QUEUE) {
2161 evac((StgClosure **)&sleeping_queue);
2165 if (suspended_ccalling_threads != END_TSO_QUEUE) {
2166 evac((StgClosure **)&suspended_ccalling_threads);
2169 #if defined(PAR) || defined(GRAN)
2170 markSparkQueue(evac);
2173 #if defined(RTS_USER_SIGNALS)
2174 // mark the signal handlers (signals should be already blocked)
2175 markSignalHandlers(evac);
2179 /* -----------------------------------------------------------------------------
2182 This is the interface to the garbage collector from Haskell land.
2183 We provide this so that external C code can allocate and garbage
2184 collect when called from Haskell via _ccall_GC.
2186 It might be useful to provide an interface whereby the programmer
2187 can specify more roots (ToDo).
2189 This needs to be protected by the GC condition variable above. KH.
2190 -------------------------------------------------------------------------- */
2192 static void (*extra_roots)(evac_fn);
2197 /* Obligated to hold this lock upon entry */
2198 ACQUIRE_LOCK(&sched_mutex);
2199 GarbageCollect(GetRoots,rtsFalse);
2200 RELEASE_LOCK(&sched_mutex);
2204 performMajorGC(void)
2206 ACQUIRE_LOCK(&sched_mutex);
2207 GarbageCollect(GetRoots,rtsTrue);
2208 RELEASE_LOCK(&sched_mutex);
2212 AllRoots(evac_fn evac)
2214 GetRoots(evac); // the scheduler's roots
2215 extra_roots(evac); // the user's roots
2219 performGCWithRoots(void (*get_roots)(evac_fn))
2221 ACQUIRE_LOCK(&sched_mutex);
2222 extra_roots = get_roots;
2223 GarbageCollect(AllRoots,rtsFalse);
2224 RELEASE_LOCK(&sched_mutex);
2227 /* -----------------------------------------------------------------------------
2230 If the thread has reached its maximum stack size, then raise the
2231 StackOverflow exception in the offending thread. Otherwise
2232 relocate the TSO into a larger chunk of memory and adjust its stack
2234 -------------------------------------------------------------------------- */
2237 threadStackOverflow(StgTSO *tso)
2239 nat new_stack_size, new_tso_size, stack_words;
2243 IF_DEBUG(sanity,checkTSO(tso));
2244 if (tso->stack_size >= tso->max_stack_size) {
2247 belch("@@ threadStackOverflow of TSO %d (%p): stack too large (now %ld; max is %ld)",
2248 tso->id, tso, tso->stack_size, tso->max_stack_size);
2249 /* If we're debugging, just print out the top of the stack */
2250 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2253 /* Send this thread the StackOverflow exception */
2254 raiseAsync(tso, (StgClosure *)stackOverflow_closure);
2258 /* Try to double the current stack size. If that takes us over the
2259 * maximum stack size for this thread, then use the maximum instead.
2260 * Finally round up so the TSO ends up as a whole number of blocks.
2262 new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
2263 new_tso_size = (nat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) +
2264 TSO_STRUCT_SIZE)/sizeof(W_);
2265 new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */
2266 new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
2268 IF_DEBUG(scheduler, fprintf(stderr,"== sched: increasing stack size from %d words to %d.\n", tso->stack_size, new_stack_size));
2270 dest = (StgTSO *)allocate(new_tso_size);
2271 TICK_ALLOC_TSO(new_stack_size,0);
2273 /* copy the TSO block and the old stack into the new area */
2274 memcpy(dest,tso,TSO_STRUCT_SIZE);
2275 stack_words = tso->stack + tso->stack_size - tso->sp;
2276 new_sp = (P_)dest + new_tso_size - stack_words;
2277 memcpy(new_sp, tso->sp, stack_words * sizeof(W_));
2279 /* relocate the stack pointers... */
2281 dest->stack_size = new_stack_size;
2283 /* Mark the old TSO as relocated. We have to check for relocated
2284 * TSOs in the garbage collector and any primops that deal with TSOs.
2286 * It's important to set the sp value to just beyond the end
2287 * of the stack, so we don't attempt to scavenge any part of the
2290 tso->what_next = ThreadRelocated;
2292 tso->sp = (P_)&(tso->stack[tso->stack_size]);
2293 tso->why_blocked = NotBlocked;
2294 dest->mut_link = NULL;
2296 IF_PAR_DEBUG(verbose,
2297 belch("@@ threadStackOverflow of TSO %d (now at %p): stack size increased to %ld",
2298 tso->id, tso, tso->stack_size);
2299 /* If we're debugging, just print out the top of the stack */
2300 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2303 IF_DEBUG(sanity,checkTSO(tso));
2305 IF_DEBUG(scheduler,printTSO(dest));
2311 /* ---------------------------------------------------------------------------
2312 Wake up a queue that was blocked on some resource.
2313 ------------------------------------------------------------------------ */
2317 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2322 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2324 /* write RESUME events to log file and
2325 update blocked and fetch time (depending on type of the orig closure) */
2326 if (RtsFlags.ParFlags.ParStats.Full) {
2327 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
2328 GR_RESUMEQ, ((StgTSO *)bqe), ((StgTSO *)bqe)->block_info.closure,
2329 0, 0 /* spark_queue_len(ADVISORY_POOL) */);
2330 if (EMPTY_RUN_QUEUE())
2331 emitSchedule = rtsTrue;
2333 switch (get_itbl(node)->type) {
2335 ((StgTSO *)bqe)->par.fetchtime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2340 ((StgTSO *)bqe)->par.blocktime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2347 barf("{unblockOneLocked}Daq Qagh: unexpected closure in blocking queue");
2354 static StgBlockingQueueElement *
2355 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2358 PEs node_loc, tso_loc;
2360 node_loc = where_is(node); // should be lifted out of loop
2361 tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2362 tso_loc = where_is((StgClosure *)tso);
2363 if (IS_LOCAL_TO(PROCS(node),tso_loc)) { // TSO is local
2364 /* !fake_fetch => TSO is on CurrentProc is same as IS_LOCAL_TO */
2365 ASSERT(CurrentProc!=node_loc || tso_loc==CurrentProc);
2366 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.lunblocktime;
2367 // insertThread(tso, node_loc);
2368 new_event(tso_loc, tso_loc, CurrentTime[CurrentProc],
2370 tso, node, (rtsSpark*)NULL);
2371 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2374 } else { // TSO is remote (actually should be FMBQ)
2375 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.mpacktime +
2376 RtsFlags.GranFlags.Costs.gunblocktime +
2377 RtsFlags.GranFlags.Costs.latency;
2378 new_event(tso_loc, CurrentProc, CurrentTime[CurrentProc],
2380 tso, node, (rtsSpark*)NULL);
2381 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2384 /* the thread-queue-overhead is accounted for in either Resume or UnblockThread */
2386 fprintf(stderr," %s TSO %d (%p) [PE %d] (block_info.closure=%p) (next=%p) ,",
2387 (node_loc==tso_loc ? "Local" : "Global"),
2388 tso->id, tso, CurrentProc, tso->block_info.closure, tso->link));
2389 tso->block_info.closure = NULL;
2390 IF_DEBUG(scheduler,belch("-- Waking up thread %ld (%p)",
2394 static StgBlockingQueueElement *
2395 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2397 StgBlockingQueueElement *next;
2399 switch (get_itbl(bqe)->type) {
2401 ASSERT(((StgTSO *)bqe)->why_blocked != NotBlocked);
2402 /* if it's a TSO just push it onto the run_queue */
2404 // ((StgTSO *)bqe)->link = END_TSO_QUEUE; // debugging?
2405 PUSH_ON_RUN_QUEUE((StgTSO *)bqe);
2407 unblockCount(bqe, node);
2408 /* reset blocking status after dumping event */
2409 ((StgTSO *)bqe)->why_blocked = NotBlocked;
2413 /* if it's a BLOCKED_FETCH put it on the PendingFetches list */
2415 bqe->link = (StgBlockingQueueElement *)PendingFetches;
2416 PendingFetches = (StgBlockedFetch *)bqe;
2420 /* can ignore this case in a non-debugging setup;
2421 see comments on RBHSave closures above */
2423 /* check that the closure is an RBHSave closure */
2424 ASSERT(get_itbl((StgClosure *)bqe) == &stg_RBH_Save_0_info ||
2425 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_1_info ||
2426 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_2_info);
2430 barf("{unblockOneLocked}Daq Qagh: Unexpected IP (%#lx; %s) in blocking queue at %#lx\n",
2431 get_itbl((StgClosure *)bqe), info_type((StgClosure *)bqe),
2435 IF_PAR_DEBUG(bq, fprintf(stderr, ", %p (%s)", bqe, info_type((StgClosure*)bqe)));
2439 #else /* !GRAN && !PAR */
2441 unblockOneLocked(StgTSO *tso)
2445 ASSERT(get_itbl(tso)->type == TSO);
2446 ASSERT(tso->why_blocked != NotBlocked);
2447 tso->why_blocked = NotBlocked;
2449 PUSH_ON_RUN_QUEUE(tso);
2451 IF_DEBUG(scheduler,sched_belch("waking up thread %ld", tso->id));
2456 #if defined(GRAN) || defined(PAR)
2457 INLINE_ME StgBlockingQueueElement *
2458 unblockOne(StgBlockingQueueElement *bqe, StgClosure *node)
2460 ACQUIRE_LOCK(&sched_mutex);
2461 bqe = unblockOneLocked(bqe, node);
2462 RELEASE_LOCK(&sched_mutex);
2467 unblockOne(StgTSO *tso)
2469 ACQUIRE_LOCK(&sched_mutex);
2470 tso = unblockOneLocked(tso);
2471 RELEASE_LOCK(&sched_mutex);
2478 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2480 StgBlockingQueueElement *bqe;
2485 belch("##-_ AwBQ for node %p on PE %d @ %ld by TSO %d (%p): ", \
2486 node, CurrentProc, CurrentTime[CurrentProc],
2487 CurrentTSO->id, CurrentTSO));
2489 node_loc = where_is(node);
2491 ASSERT(q == END_BQ_QUEUE ||
2492 get_itbl(q)->type == TSO || // q is either a TSO or an RBHSave
2493 get_itbl(q)->type == CONSTR); // closure (type constructor)
2494 ASSERT(is_unique(node));
2496 /* FAKE FETCH: magically copy the node to the tso's proc;
2497 no Fetch necessary because in reality the node should not have been
2498 moved to the other PE in the first place
2500 if (CurrentProc!=node_loc) {
2502 belch("## node %p is on PE %d but CurrentProc is %d (TSO %d); assuming fake fetch and adjusting bitmask (old: %#x)",
2503 node, node_loc, CurrentProc, CurrentTSO->id,
2504 // CurrentTSO, where_is(CurrentTSO),
2505 node->header.gran.procs));
2506 node->header.gran.procs = (node->header.gran.procs) | PE_NUMBER(CurrentProc);
2508 belch("## new bitmask of node %p is %#x",
2509 node, node->header.gran.procs));
2510 if (RtsFlags.GranFlags.GranSimStats.Global) {
2511 globalGranStats.tot_fake_fetches++;
2516 // ToDo: check: ASSERT(CurrentProc==node_loc);
2517 while (get_itbl(bqe)->type==TSO) { // q != END_TSO_QUEUE) {
2520 bqe points to the current element in the queue
2521 next points to the next element in the queue
2523 //tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2524 //tso_loc = where_is(tso);
2526 bqe = unblockOneLocked(bqe, node);
2529 /* if this is the BQ of an RBH, we have to put back the info ripped out of
2530 the closure to make room for the anchor of the BQ */
2531 if (bqe!=END_BQ_QUEUE) {
2532 ASSERT(get_itbl(node)->type == RBH && get_itbl(bqe)->type == CONSTR);
2534 ASSERT((info_ptr==&RBH_Save_0_info) ||
2535 (info_ptr==&RBH_Save_1_info) ||
2536 (info_ptr==&RBH_Save_2_info));
2538 /* cf. convertToRBH in RBH.c for writing the RBHSave closure */
2539 ((StgRBH *)node)->blocking_queue = (StgBlockingQueueElement *)((StgRBHSave *)bqe)->payload[0];
2540 ((StgRBH *)node)->mut_link = (StgMutClosure *)((StgRBHSave *)bqe)->payload[1];
2543 belch("## Filled in RBH_Save for %p (%s) at end of AwBQ",
2544 node, info_type(node)));
2547 /* statistics gathering */
2548 if (RtsFlags.GranFlags.GranSimStats.Global) {
2549 // globalGranStats.tot_bq_processing_time += bq_processing_time;
2550 globalGranStats.tot_bq_len += len; // total length of all bqs awakened
2551 // globalGranStats.tot_bq_len_local += len_local; // same for local TSOs only
2552 globalGranStats.tot_awbq++; // total no. of bqs awakened
2555 fprintf(stderr,"## BQ Stats of %p: [%d entries] %s\n",
2556 node, len, (bqe!=END_BQ_QUEUE) ? "RBH" : ""));
2560 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2562 StgBlockingQueueElement *bqe;
2564 ACQUIRE_LOCK(&sched_mutex);
2566 IF_PAR_DEBUG(verbose,
2567 belch("##-_ AwBQ for node %p on [%x]: ",
2571 if(get_itbl(q)->type == CONSTR || q==END_BQ_QUEUE) {
2572 IF_PAR_DEBUG(verbose, belch("## ... nothing to unblock so lets just return. RFP (BUG?)"));
2577 ASSERT(q == END_BQ_QUEUE ||
2578 get_itbl(q)->type == TSO ||
2579 get_itbl(q)->type == BLOCKED_FETCH ||
2580 get_itbl(q)->type == CONSTR);
2583 while (get_itbl(bqe)->type==TSO ||
2584 get_itbl(bqe)->type==BLOCKED_FETCH) {
2585 bqe = unblockOneLocked(bqe, node);
2587 RELEASE_LOCK(&sched_mutex);
2590 #else /* !GRAN && !PAR */
2593 awakenBlockedQueueNoLock(StgTSO *tso)
2595 while (tso != END_TSO_QUEUE) {
2596 tso = unblockOneLocked(tso);
2601 awakenBlockedQueue(StgTSO *tso)
2603 ACQUIRE_LOCK(&sched_mutex);
2604 while (tso != END_TSO_QUEUE) {
2605 tso = unblockOneLocked(tso);
2607 RELEASE_LOCK(&sched_mutex);
2611 /* ---------------------------------------------------------------------------
2613 - usually called inside a signal handler so it mustn't do anything fancy.
2614 ------------------------------------------------------------------------ */
2617 interruptStgRts(void)
2621 #ifdef RTS_SUPPORTS_THREADS
2622 wakeBlockedWorkerThread();
2626 /* -----------------------------------------------------------------------------
2629 This is for use when we raise an exception in another thread, which
2631 This has nothing to do with the UnblockThread event in GranSim. -- HWL
2632 -------------------------------------------------------------------------- */
2634 #if defined(GRAN) || defined(PAR)
2636 NB: only the type of the blocking queue is different in GranSim and GUM
2637 the operations on the queue-elements are the same
2638 long live polymorphism!
2640 Locks: sched_mutex is held upon entry and exit.
2644 unblockThread(StgTSO *tso)
2646 StgBlockingQueueElement *t, **last;
2648 switch (tso->why_blocked) {
2651 return; /* not blocked */
2654 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
2656 StgBlockingQueueElement *last_tso = END_BQ_QUEUE;
2657 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
2659 last = (StgBlockingQueueElement **)&mvar->head;
2660 for (t = (StgBlockingQueueElement *)mvar->head;
2662 last = &t->link, last_tso = t, t = t->link) {
2663 if (t == (StgBlockingQueueElement *)tso) {
2664 *last = (StgBlockingQueueElement *)tso->link;
2665 if (mvar->tail == tso) {
2666 mvar->tail = (StgTSO *)last_tso;
2671 barf("unblockThread (MVAR): TSO not found");
2674 case BlockedOnBlackHole:
2675 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
2677 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
2679 last = &bq->blocking_queue;
2680 for (t = bq->blocking_queue;
2682 last = &t->link, t = t->link) {
2683 if (t == (StgBlockingQueueElement *)tso) {
2684 *last = (StgBlockingQueueElement *)tso->link;
2688 barf("unblockThread (BLACKHOLE): TSO not found");
2691 case BlockedOnException:
2693 StgTSO *target = tso->block_info.tso;
2695 ASSERT(get_itbl(target)->type == TSO);
2697 if (target->what_next == ThreadRelocated) {
2698 target = target->link;
2699 ASSERT(get_itbl(target)->type == TSO);
2702 ASSERT(target->blocked_exceptions != NULL);
2704 last = (StgBlockingQueueElement **)&target->blocked_exceptions;
2705 for (t = (StgBlockingQueueElement *)target->blocked_exceptions;
2707 last = &t->link, t = t->link) {
2708 ASSERT(get_itbl(t)->type == TSO);
2709 if (t == (StgBlockingQueueElement *)tso) {
2710 *last = (StgBlockingQueueElement *)tso->link;
2714 barf("unblockThread (Exception): TSO not found");
2718 case BlockedOnWrite:
2719 #if defined(mingw32_TARGET_OS)
2720 case BlockedOnDoProc:
2723 /* take TSO off blocked_queue */
2724 StgBlockingQueueElement *prev = NULL;
2725 for (t = (StgBlockingQueueElement *)blocked_queue_hd; t != END_BQ_QUEUE;
2726 prev = t, t = t->link) {
2727 if (t == (StgBlockingQueueElement *)tso) {
2729 blocked_queue_hd = (StgTSO *)t->link;
2730 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
2731 blocked_queue_tl = END_TSO_QUEUE;
2734 prev->link = t->link;
2735 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
2736 blocked_queue_tl = (StgTSO *)prev;
2742 barf("unblockThread (I/O): TSO not found");
2745 case BlockedOnDelay:
2747 /* take TSO off sleeping_queue */
2748 StgBlockingQueueElement *prev = NULL;
2749 for (t = (StgBlockingQueueElement *)sleeping_queue; t != END_BQ_QUEUE;
2750 prev = t, t = t->link) {
2751 if (t == (StgBlockingQueueElement *)tso) {
2753 sleeping_queue = (StgTSO *)t->link;
2755 prev->link = t->link;
2760 barf("unblockThread (delay): TSO not found");
2764 barf("unblockThread");
2768 tso->link = END_TSO_QUEUE;
2769 tso->why_blocked = NotBlocked;
2770 tso->block_info.closure = NULL;
2771 PUSH_ON_RUN_QUEUE(tso);
2775 unblockThread(StgTSO *tso)
2779 /* To avoid locking unnecessarily. */
2780 if (tso->why_blocked == NotBlocked) {
2784 switch (tso->why_blocked) {
2787 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
2789 StgTSO *last_tso = END_TSO_QUEUE;
2790 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
2793 for (t = mvar->head; t != END_TSO_QUEUE;
2794 last = &t->link, last_tso = t, t = t->link) {
2797 if (mvar->tail == tso) {
2798 mvar->tail = last_tso;
2803 barf("unblockThread (MVAR): TSO not found");
2806 case BlockedOnBlackHole:
2807 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
2809 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
2811 last = &bq->blocking_queue;
2812 for (t = bq->blocking_queue; t != END_TSO_QUEUE;
2813 last = &t->link, t = t->link) {
2819 barf("unblockThread (BLACKHOLE): TSO not found");
2822 case BlockedOnException:
2824 StgTSO *target = tso->block_info.tso;
2826 ASSERT(get_itbl(target)->type == TSO);
2828 while (target->what_next == ThreadRelocated) {
2829 target = target->link;
2830 ASSERT(get_itbl(target)->type == TSO);
2833 ASSERT(target->blocked_exceptions != NULL);
2835 last = &target->blocked_exceptions;
2836 for (t = target->blocked_exceptions; t != END_TSO_QUEUE;
2837 last = &t->link, t = t->link) {
2838 ASSERT(get_itbl(t)->type == TSO);
2844 barf("unblockThread (Exception): TSO not found");
2848 case BlockedOnWrite:
2849 #if defined(mingw32_TARGET_OS)
2850 case BlockedOnDoProc:
2853 StgTSO *prev = NULL;
2854 for (t = blocked_queue_hd; t != END_TSO_QUEUE;
2855 prev = t, t = t->link) {
2858 blocked_queue_hd = t->link;
2859 if (blocked_queue_tl == t) {
2860 blocked_queue_tl = END_TSO_QUEUE;
2863 prev->link = t->link;
2864 if (blocked_queue_tl == t) {
2865 blocked_queue_tl = prev;
2871 barf("unblockThread (I/O): TSO not found");
2874 case BlockedOnDelay:
2876 StgTSO *prev = NULL;
2877 for (t = sleeping_queue; t != END_TSO_QUEUE;
2878 prev = t, t = t->link) {
2881 sleeping_queue = t->link;
2883 prev->link = t->link;
2888 barf("unblockThread (delay): TSO not found");
2892 barf("unblockThread");
2896 tso->link = END_TSO_QUEUE;
2897 tso->why_blocked = NotBlocked;
2898 tso->block_info.closure = NULL;
2899 PUSH_ON_RUN_QUEUE(tso);
2903 /* -----------------------------------------------------------------------------
2906 * The following function implements the magic for raising an
2907 * asynchronous exception in an existing thread.
2909 * We first remove the thread from any queue on which it might be
2910 * blocked. The possible blockages are MVARs and BLACKHOLE_BQs.
2912 * We strip the stack down to the innermost CATCH_FRAME, building
2913 * thunks in the heap for all the active computations, so they can
2914 * be restarted if necessary. When we reach a CATCH_FRAME, we build
2915 * an application of the handler to the exception, and push it on
2916 * the top of the stack.
2918 * How exactly do we save all the active computations? We create an
2919 * AP_STACK for every UpdateFrame on the stack. Entering one of these
2920 * AP_STACKs pushes everything from the corresponding update frame
2921 * upwards onto the stack. (Actually, it pushes everything up to the
2922 * next update frame plus a pointer to the next AP_STACK object.
2923 * Entering the next AP_STACK object pushes more onto the stack until we
2924 * reach the last AP_STACK object - at which point the stack should look
2925 * exactly as it did when we killed the TSO and we can continue
2926 * execution by entering the closure on top of the stack.
2928 * We can also kill a thread entirely - this happens if either (a) the
2929 * exception passed to raiseAsync is NULL, or (b) there's no
2930 * CATCH_FRAME on the stack. In either case, we strip the entire
2931 * stack and replace the thread with a zombie.
2933 * Locks: sched_mutex held upon entry nor exit.
2935 * -------------------------------------------------------------------------- */
2938 deleteThread(StgTSO *tso)
2940 raiseAsync(tso,NULL);
2943 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
2945 deleteThreadImmediately(StgTSO *tso)
2946 { // for forkProcess only:
2947 // delete thread without giving it a chance to catch the KillThread exception
2949 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
2953 if (tso->why_blocked != BlockedOnCCall &&
2954 tso->why_blocked != BlockedOnCCall_NoUnblockExc) {
2958 tso->what_next = ThreadKilled;
2963 raiseAsyncWithLock(StgTSO *tso, StgClosure *exception)
2965 /* When raising async exs from contexts where sched_mutex isn't held;
2966 use raiseAsyncWithLock(). */
2967 ACQUIRE_LOCK(&sched_mutex);
2968 raiseAsync(tso,exception);
2969 RELEASE_LOCK(&sched_mutex);
2973 raiseAsync(StgTSO *tso, StgClosure *exception)
2975 StgRetInfoTable *info;
2978 // Thread already dead?
2979 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
2984 sched_belch("raising exception in thread %ld.", tso->id));
2986 // Remove it from any blocking queues
2991 // The stack freezing code assumes there's a closure pointer on
2992 // the top of the stack, so we have to arrange that this is the case...
2994 if (sp[0] == (W_)&stg_enter_info) {
2998 sp[0] = (W_)&stg_dummy_ret_closure;
3004 // 1. Let the top of the stack be the "current closure"
3006 // 2. Walk up the stack until we find either an UPDATE_FRAME or a
3009 // 3. If it's an UPDATE_FRAME, then make an AP_STACK containing the
3010 // current closure applied to the chunk of stack up to (but not
3011 // including) the update frame. This closure becomes the "current
3012 // closure". Go back to step 2.
3014 // 4. If it's a CATCH_FRAME, then leave the exception handler on
3015 // top of the stack applied to the exception.
3017 // 5. If it's a STOP_FRAME, then kill the thread.
3022 info = get_ret_itbl((StgClosure *)frame);
3024 while (info->i.type != UPDATE_FRAME
3025 && (info->i.type != CATCH_FRAME || exception == NULL)
3026 && info->i.type != STOP_FRAME) {
3027 frame += stack_frame_sizeW((StgClosure *)frame);
3028 info = get_ret_itbl((StgClosure *)frame);
3031 switch (info->i.type) {
3034 // If we find a CATCH_FRAME, and we've got an exception to raise,
3035 // then build the THUNK raise(exception), and leave it on
3036 // top of the CATCH_FRAME ready to enter.
3040 StgCatchFrame *cf = (StgCatchFrame *)frame;
3044 // we've got an exception to raise, so let's pass it to the
3045 // handler in this frame.
3047 raise = (StgClosure *)allocate(sizeofW(StgClosure)+1);
3048 TICK_ALLOC_SE_THK(1,0);
3049 SET_HDR(raise,&stg_raise_info,cf->header.prof.ccs);
3050 raise->payload[0] = exception;
3052 // throw away the stack from Sp up to the CATCH_FRAME.
3056 /* Ensure that async excpetions are blocked now, so we don't get
3057 * a surprise exception before we get around to executing the
3060 if (tso->blocked_exceptions == NULL) {
3061 tso->blocked_exceptions = END_TSO_QUEUE;
3064 /* Put the newly-built THUNK on top of the stack, ready to execute
3065 * when the thread restarts.
3068 sp[-1] = (W_)&stg_enter_info;
3070 tso->what_next = ThreadRunGHC;
3071 IF_DEBUG(sanity, checkTSO(tso));
3080 // First build an AP_STACK consisting of the stack chunk above the
3081 // current update frame, with the top word on the stack as the
3084 words = frame - sp - 1;
3085 ap = (StgAP_STACK *)allocate(PAP_sizeW(words));
3088 ap->fun = (StgClosure *)sp[0];
3090 for(i=0; i < (nat)words; ++i) {
3091 ap->payload[i] = (StgClosure *)*sp++;
3094 SET_HDR(ap,&stg_AP_STACK_info,
3095 ((StgClosure *)frame)->header.prof.ccs /* ToDo */);
3096 TICK_ALLOC_UP_THK(words+1,0);
3099 fprintf(stderr, "sched: Updating ");
3100 printPtr((P_)((StgUpdateFrame *)frame)->updatee);
3101 fprintf(stderr, " with ");
3102 printObj((StgClosure *)ap);
3105 // Replace the updatee with an indirection - happily
3106 // this will also wake up any threads currently
3107 // waiting on the result.
3109 // Warning: if we're in a loop, more than one update frame on
3110 // the stack may point to the same object. Be careful not to
3111 // overwrite an IND_OLDGEN in this case, because we'll screw
3112 // up the mutable lists. To be on the safe side, don't
3113 // overwrite any kind of indirection at all. See also
3114 // threadSqueezeStack in GC.c, where we have to make a similar
3117 if (!closure_IND(((StgUpdateFrame *)frame)->updatee)) {
3118 // revert the black hole
3119 UPD_IND_NOLOCK(((StgUpdateFrame *)frame)->updatee,ap);
3121 sp += sizeofW(StgUpdateFrame) - 1;
3122 sp[0] = (W_)ap; // push onto stack
3127 // We've stripped the entire stack, the thread is now dead.
3128 sp += sizeofW(StgStopFrame);
3129 tso->what_next = ThreadKilled;
3140 /* -----------------------------------------------------------------------------
3141 resurrectThreads is called after garbage collection on the list of
3142 threads found to be garbage. Each of these threads will be woken
3143 up and sent a signal: BlockedOnDeadMVar if the thread was blocked
3144 on an MVar, or NonTermination if the thread was blocked on a Black
3147 Locks: sched_mutex isn't held upon entry nor exit.
3148 -------------------------------------------------------------------------- */
3151 resurrectThreads( StgTSO *threads )
3155 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
3156 next = tso->global_link;
3157 tso->global_link = all_threads;
3159 IF_DEBUG(scheduler, sched_belch("resurrecting thread %d", tso->id));
3161 switch (tso->why_blocked) {
3163 case BlockedOnException:
3164 /* Called by GC - sched_mutex lock is currently held. */
3165 raiseAsync(tso,(StgClosure *)BlockedOnDeadMVar_closure);
3167 case BlockedOnBlackHole:
3168 raiseAsync(tso,(StgClosure *)NonTermination_closure);
3171 /* This might happen if the thread was blocked on a black hole
3172 * belonging to a thread that we've just woken up (raiseAsync
3173 * can wake up threads, remember...).
3177 barf("resurrectThreads: thread blocked in a strange way");
3182 /* -----------------------------------------------------------------------------
3183 * Blackhole detection: if we reach a deadlock, test whether any
3184 * threads are blocked on themselves. Any threads which are found to
3185 * be self-blocked get sent a NonTermination exception.
3187 * This is only done in a deadlock situation in order to avoid
3188 * performance overhead in the normal case.
3190 * Locks: sched_mutex is held upon entry and exit.
3191 * -------------------------------------------------------------------------- */
3194 detectBlackHoles( void )
3196 StgTSO *tso = all_threads;
3198 StgClosure *blocked_on;
3199 StgRetInfoTable *info;
3201 for (tso = all_threads; tso != END_TSO_QUEUE; tso = tso->global_link) {
3203 while (tso->what_next == ThreadRelocated) {
3205 ASSERT(get_itbl(tso)->type == TSO);
3208 if (tso->why_blocked != BlockedOnBlackHole) {
3211 blocked_on = tso->block_info.closure;
3213 frame = (StgClosure *)tso->sp;
3216 info = get_ret_itbl(frame);
3217 switch (info->i.type) {
3219 if (((StgUpdateFrame *)frame)->updatee == blocked_on) {
3220 /* We are blocking on one of our own computations, so
3221 * send this thread the NonTermination exception.
3224 sched_belch("thread %d is blocked on itself", tso->id));
3225 raiseAsync(tso, (StgClosure *)NonTermination_closure);
3229 frame = (StgClosure *) ((StgUpdateFrame *)frame + 1);
3235 // normal stack frames; do nothing except advance the pointer
3237 (StgPtr)frame += stack_frame_sizeW(frame);
3244 /* ----------------------------------------------------------------------------
3245 * Debugging: why is a thread blocked
3246 * [Also provides useful information when debugging threaded programs
3247 * at the Haskell source code level, so enable outside of DEBUG. --sof 7/02]
3248 ------------------------------------------------------------------------- */
3252 printThreadBlockage(StgTSO *tso)
3254 switch (tso->why_blocked) {
3256 fprintf(stderr,"is blocked on read from fd %d", tso->block_info.fd);
3258 case BlockedOnWrite:
3259 fprintf(stderr,"is blocked on write to fd %d", tso->block_info.fd);
3261 #if defined(mingw32_TARGET_OS)
3262 case BlockedOnDoProc:
3263 fprintf(stderr,"is blocked on proc (request: %d)", tso->block_info.async_result->reqID);
3266 case BlockedOnDelay:
3267 fprintf(stderr,"is blocked until %d", tso->block_info.target);
3270 fprintf(stderr,"is blocked on an MVar");
3272 case BlockedOnException:
3273 fprintf(stderr,"is blocked on delivering an exception to thread %d",
3274 tso->block_info.tso->id);
3276 case BlockedOnBlackHole:
3277 fprintf(stderr,"is blocked on a black hole");
3280 fprintf(stderr,"is not blocked");
3284 fprintf(stderr,"is blocked on global address; local FM_BQ is %p (%s)",
3285 tso->block_info.closure, info_type(tso->block_info.closure));
3287 case BlockedOnGA_NoSend:
3288 fprintf(stderr,"is blocked on global address (no send); local FM_BQ is %p (%s)",
3289 tso->block_info.closure, info_type(tso->block_info.closure));
3292 case BlockedOnCCall:
3293 fprintf(stderr,"is blocked on an external call");
3295 case BlockedOnCCall_NoUnblockExc:
3296 fprintf(stderr,"is blocked on an external call (exceptions were already blocked)");
3299 barf("printThreadBlockage: strange tso->why_blocked: %d for TSO %d (%d)",
3300 tso->why_blocked, tso->id, tso);
3306 printThreadStatus(StgTSO *tso)
3308 switch (tso->what_next) {
3310 fprintf(stderr,"has been killed");
3312 case ThreadComplete:
3313 fprintf(stderr,"has completed");
3316 printThreadBlockage(tso);
3321 printAllThreads(void)
3327 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
3328 ullong_format_string(TIME_ON_PROC(CurrentProc),
3329 time_string, rtsFalse/*no commas!*/);
3331 fprintf(stderr, "all threads at [%s]:\n", time_string);
3333 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
3334 ullong_format_string(CURRENT_TIME,
3335 time_string, rtsFalse/*no commas!*/);
3337 fprintf(stderr,"all threads at [%s]:\n", time_string);
3339 fprintf(stderr,"all threads:\n");
3342 for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
3343 fprintf(stderr, "\tthread %d @ %p ", t->id, (void *)t);
3344 label = lookupThreadLabel(t->id);
3345 if (label) fprintf(stderr,"[\"%s\"] ",(char *)label);
3346 printThreadStatus(t);
3347 fprintf(stderr,"\n");
3354 Print a whole blocking queue attached to node (debugging only).
3358 print_bq (StgClosure *node)
3360 StgBlockingQueueElement *bqe;
3364 fprintf(stderr,"## BQ of closure %p (%s): ",
3365 node, info_type(node));
3367 /* should cover all closures that may have a blocking queue */
3368 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
3369 get_itbl(node)->type == FETCH_ME_BQ ||
3370 get_itbl(node)->type == RBH ||
3371 get_itbl(node)->type == MVAR);
3373 ASSERT(node!=(StgClosure*)NULL); // sanity check
3375 print_bqe(((StgBlockingQueue*)node)->blocking_queue);
3379 Print a whole blocking queue starting with the element bqe.
3382 print_bqe (StgBlockingQueueElement *bqe)
3387 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
3389 for (end = (bqe==END_BQ_QUEUE);
3390 !end; // iterate until bqe points to a CONSTR
3391 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE),
3392 bqe = end ? END_BQ_QUEUE : bqe->link) {
3393 ASSERT(bqe != END_BQ_QUEUE); // sanity check
3394 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
3395 /* types of closures that may appear in a blocking queue */
3396 ASSERT(get_itbl(bqe)->type == TSO ||
3397 get_itbl(bqe)->type == BLOCKED_FETCH ||
3398 get_itbl(bqe)->type == CONSTR);
3399 /* only BQs of an RBH end with an RBH_Save closure */
3400 //ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
3402 switch (get_itbl(bqe)->type) {
3404 fprintf(stderr," TSO %u (%x),",
3405 ((StgTSO *)bqe)->id, ((StgTSO *)bqe));
3408 fprintf(stderr," BF (node=%p, ga=((%x, %d, %x)),",
3409 ((StgBlockedFetch *)bqe)->node,
3410 ((StgBlockedFetch *)bqe)->ga.payload.gc.gtid,
3411 ((StgBlockedFetch *)bqe)->ga.payload.gc.slot,
3412 ((StgBlockedFetch *)bqe)->ga.weight);
3415 fprintf(stderr," %s (IP %p),",
3416 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
3417 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
3418 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
3419 "RBH_Save_?"), get_itbl(bqe));
3422 barf("Unexpected closure type %s in blocking queue", // of %p (%s)",
3423 info_type((StgClosure *)bqe)); // , node, info_type(node));
3427 fputc('\n', stderr);
3429 # elif defined(GRAN)
3431 print_bq (StgClosure *node)
3433 StgBlockingQueueElement *bqe;
3434 PEs node_loc, tso_loc;
3437 /* should cover all closures that may have a blocking queue */
3438 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
3439 get_itbl(node)->type == FETCH_ME_BQ ||
3440 get_itbl(node)->type == RBH);
3442 ASSERT(node!=(StgClosure*)NULL); // sanity check
3443 node_loc = where_is(node);
3445 fprintf(stderr,"## BQ of closure %p (%s) on [PE %d]: ",
3446 node, info_type(node), node_loc);
3449 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
3451 for (bqe = ((StgBlockingQueue*)node)->blocking_queue, end = (bqe==END_BQ_QUEUE);
3452 !end; // iterate until bqe points to a CONSTR
3453 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) {
3454 ASSERT(bqe != END_BQ_QUEUE); // sanity check
3455 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
3456 /* types of closures that may appear in a blocking queue */
3457 ASSERT(get_itbl(bqe)->type == TSO ||
3458 get_itbl(bqe)->type == CONSTR);
3459 /* only BQs of an RBH end with an RBH_Save closure */
3460 ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
3462 tso_loc = where_is((StgClosure *)bqe);
3463 switch (get_itbl(bqe)->type) {
3465 fprintf(stderr," TSO %d (%p) on [PE %d],",
3466 ((StgTSO *)bqe)->id, (StgTSO *)bqe, tso_loc);
3469 fprintf(stderr," %s (IP %p),",
3470 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
3471 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
3472 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
3473 "RBH_Save_?"), get_itbl(bqe));
3476 barf("Unexpected closure type %s in blocking queue of %p (%s)",
3477 info_type((StgClosure *)bqe), node, info_type(node));
3481 fputc('\n', stderr);
3485 Nice and easy: only TSOs on the blocking queue
3488 print_bq (StgClosure *node)
3492 ASSERT(node!=(StgClosure*)NULL); // sanity check
3493 for (tso = ((StgBlockingQueue*)node)->blocking_queue;
3494 tso != END_TSO_QUEUE;
3496 ASSERT(tso!=NULL && tso!=END_TSO_QUEUE); // sanity check
3497 ASSERT(get_itbl(tso)->type == TSO); // guess what, sanity check
3498 fprintf(stderr," TSO %d (%p),", tso->id, tso);
3500 fputc('\n', stderr);
3511 for (i=0, tso=run_queue_hd;
3512 tso != END_TSO_QUEUE;
3521 sched_belch(char *s, ...)
3525 #ifdef RTS_SUPPORTS_THREADS
3526 fprintf(stderr, "sched (task %p): ", osThreadId());
3528 fprintf(stderr, "== ");
3530 fprintf(stderr, "sched: ");
3532 vfprintf(stderr, s, ap);
3533 fprintf(stderr, "\n");