1 /* ---------------------------------------------------------------------------
2 * $Id: Schedule.c,v 1.199 2004/08/09 14:27:53 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);
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 if(!startTask(taskStart))
281 startingWorkerThread = rtsFalse;
288 /* ---------------------------------------------------------------------------
289 Main scheduling loop.
291 We use round-robin scheduling, each thread returning to the
292 scheduler loop when one of these conditions is detected:
295 * timer expires (thread yields)
300 Locking notes: we acquire the scheduler lock once at the beginning
301 of the scheduler loop, and release it when
303 * running a thread, or
304 * waiting for work, or
305 * waiting for a GC to complete.
308 In a GranSim setup this loop iterates over the global event queue.
309 This revolves around the global event queue, which determines what
310 to do next. Therefore, it's more complicated than either the
311 concurrent or the parallel (GUM) setup.
314 GUM iterates over incoming messages.
315 It starts with nothing to do (thus CurrentTSO == END_TSO_QUEUE),
316 and sends out a fish whenever it has nothing to do; in-between
317 doing the actual reductions (shared code below) it processes the
318 incoming messages and deals with delayed operations
319 (see PendingFetches).
320 This is not the ugliest code you could imagine, but it's bloody close.
322 ------------------------------------------------------------------------ */
324 schedule( StgMainThread *mainThread USED_WHEN_RTS_SUPPORTS_THREADS,
325 Capability *initialCapability )
329 StgThreadReturnCode ret;
337 rtsBool receivedFinish = rtsFalse;
339 nat tp_size, sp_size; // stats only
342 rtsBool was_interrupted = rtsFalse;
343 StgTSOWhatNext prev_what_next;
345 // Pre-condition: sched_mutex is held.
346 // We might have a capability, passed in as initialCapability.
347 cap = initialCapability;
349 #if defined(RTS_SUPPORTS_THREADS)
351 // in the threaded case, the capability is either passed in via the
352 // initialCapability parameter, or initialized inside the scheduler
356 sched_belch("### NEW SCHEDULER LOOP (main thr: %p, cap: %p)",
357 mainThread, initialCapability);
360 // simply initialise it in the non-threaded case
361 grabCapability(&cap);
365 /* set up first event to get things going */
366 /* ToDo: assign costs for system setup and init MainTSO ! */
367 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
369 CurrentTSO, (StgClosure*)NULL, (rtsSpark*)NULL);
372 fprintf(stderr, "GRAN: Init CurrentTSO (in schedule) = %p\n", CurrentTSO);
373 G_TSO(CurrentTSO, 5));
375 if (RtsFlags.GranFlags.Light) {
376 /* Save current time; GranSim Light only */
377 CurrentTSO->gran.clock = CurrentTime[CurrentProc];
380 event = get_next_event();
382 while (event!=(rtsEvent*)NULL) {
383 /* Choose the processor with the next event */
384 CurrentProc = event->proc;
385 CurrentTSO = event->tso;
389 while (!receivedFinish) { /* set by processMessages */
390 /* when receiving PP_FINISH message */
392 #else // everything except GRAN and PAR
398 IF_DEBUG(scheduler, printAllThreads());
400 #if defined(RTS_SUPPORTS_THREADS)
401 // Yield the capability to higher-priority tasks if necessary.
404 yieldCapability(&cap);
407 // If we do not currently hold a capability, we wait for one
410 waitForCapability(&sched_mutex, &cap,
411 mainThread ? &mainThread->bound_thread_cond : NULL);
414 // We now have a capability...
418 // If we're interrupted (the user pressed ^C, or some other
419 // termination condition occurred), kill all the currently running
423 IF_DEBUG(scheduler, sched_belch("interrupted"));
424 interrupted = rtsFalse;
425 was_interrupted = rtsTrue;
426 #if defined(RTS_SUPPORTS_THREADS)
427 // In the threaded RTS, deadlock detection doesn't work,
428 // so just exit right away.
429 prog_belch("interrupted");
430 releaseCapability(cap);
431 RELEASE_LOCK(&sched_mutex);
432 shutdownHaskellAndExit(EXIT_SUCCESS);
438 #if defined(RTS_USER_SIGNALS)
439 // check for signals each time around the scheduler
440 if (signals_pending()) {
441 RELEASE_LOCK(&sched_mutex); /* ToDo: kill */
442 startSignalHandlers();
443 ACQUIRE_LOCK(&sched_mutex);
448 // Check whether any waiting threads need to be woken up. If the
449 // run queue is empty, and there are no other tasks running, we
450 // can wait indefinitely for something to happen.
452 if ( !EMPTY_QUEUE(blocked_queue_hd) || !EMPTY_QUEUE(sleeping_queue)
453 #if defined(RTS_SUPPORTS_THREADS)
458 awaitEvent( EMPTY_RUN_QUEUE() );
460 // we can be interrupted while waiting for I/O...
461 if (interrupted) continue;
464 * Detect deadlock: when we have no threads to run, there are no
465 * threads waiting on I/O or sleeping, and all the other tasks are
466 * waiting for work, we must have a deadlock of some description.
468 * We first try to find threads blocked on themselves (ie. black
469 * holes), and generate NonTermination exceptions where necessary.
471 * If no threads are black holed, we have a deadlock situation, so
472 * inform all the main threads.
474 #if !defined(PAR) && !defined(RTS_SUPPORTS_THREADS)
475 if ( EMPTY_THREAD_QUEUES() )
477 IF_DEBUG(scheduler, sched_belch("deadlocked, forcing major GC..."));
478 // Garbage collection can release some new threads due to
479 // either (a) finalizers or (b) threads resurrected because
480 // they are about to be send BlockedOnDeadMVar. Any threads
481 // thus released will be immediately runnable.
482 GarbageCollect(GetRoots,rtsTrue);
484 if ( !EMPTY_RUN_QUEUE() ) { goto not_deadlocked; }
487 sched_belch("still deadlocked, checking for black holes..."));
490 if ( !EMPTY_RUN_QUEUE() ) { goto not_deadlocked; }
492 #if defined(RTS_USER_SIGNALS)
493 /* If we have user-installed signal handlers, then wait
494 * for signals to arrive rather then bombing out with a
497 if ( anyUserHandlers() ) {
499 sched_belch("still deadlocked, waiting for signals..."));
503 // we might be interrupted...
504 if (interrupted) { continue; }
506 if (signals_pending()) {
507 RELEASE_LOCK(&sched_mutex);
508 startSignalHandlers();
509 ACQUIRE_LOCK(&sched_mutex);
511 ASSERT(!EMPTY_RUN_QUEUE());
516 /* Probably a real deadlock. Send the current main thread the
517 * Deadlock exception (or in the SMP build, send *all* main
518 * threads the deadlock exception, since none of them can make
524 switch (m->tso->why_blocked) {
525 case BlockedOnBlackHole:
526 case BlockedOnException:
528 raiseAsync(m->tso, (StgClosure *)NonTermination_closure);
531 barf("deadlock: main thread blocked in a strange way");
537 #elif defined(RTS_SUPPORTS_THREADS)
538 // ToDo: add deadlock detection in threaded RTS
540 // ToDo: add deadlock detection in GUM (similar to SMP) -- HWL
543 #if defined(RTS_SUPPORTS_THREADS)
544 if ( EMPTY_RUN_QUEUE() ) {
545 continue; // nothing to do
550 if (RtsFlags.GranFlags.Light)
551 GranSimLight_enter_system(event, &ActiveTSO); // adjust ActiveTSO etc
553 /* adjust time based on time-stamp */
554 if (event->time > CurrentTime[CurrentProc] &&
555 event->evttype != ContinueThread)
556 CurrentTime[CurrentProc] = event->time;
558 /* Deal with the idle PEs (may issue FindWork or MoveSpark events) */
559 if (!RtsFlags.GranFlags.Light)
562 IF_DEBUG(gran, fprintf(stderr, "GRAN: switch by event-type\n"));
564 /* main event dispatcher in GranSim */
565 switch (event->evttype) {
566 /* Should just be continuing execution */
568 IF_DEBUG(gran, fprintf(stderr, "GRAN: doing ContinueThread\n"));
569 /* ToDo: check assertion
570 ASSERT(run_queue_hd != (StgTSO*)NULL &&
571 run_queue_hd != END_TSO_QUEUE);
573 /* Ignore ContinueThreads for fetching threads (if synchr comm) */
574 if (!RtsFlags.GranFlags.DoAsyncFetch &&
575 procStatus[CurrentProc]==Fetching) {
576 belch("ghuH: Spurious ContinueThread while Fetching ignored; TSO %d (%p) [PE %d]",
577 CurrentTSO->id, CurrentTSO, CurrentProc);
580 /* Ignore ContinueThreads for completed threads */
581 if (CurrentTSO->what_next == ThreadComplete) {
582 belch("ghuH: found a ContinueThread event for completed thread %d (%p) [PE %d] (ignoring ContinueThread)",
583 CurrentTSO->id, CurrentTSO, CurrentProc);
586 /* Ignore ContinueThreads for threads that are being migrated */
587 if (PROCS(CurrentTSO)==Nowhere) {
588 belch("ghuH: trying to run the migrating TSO %d (%p) [PE %d] (ignoring ContinueThread)",
589 CurrentTSO->id, CurrentTSO, CurrentProc);
592 /* The thread should be at the beginning of the run queue */
593 if (CurrentTSO!=run_queue_hds[CurrentProc]) {
594 belch("ghuH: TSO %d (%p) [PE %d] is not at the start of the run_queue when doing a ContinueThread",
595 CurrentTSO->id, CurrentTSO, CurrentProc);
596 break; // run the thread anyway
599 new_event(proc, proc, CurrentTime[proc],
601 (StgTSO*)NULL, (StgClosure*)NULL, (rtsSpark*)NULL);
603 */ /* Catches superfluous CONTINUEs -- should be unnecessary */
604 break; // now actually run the thread; DaH Qu'vam yImuHbej
607 do_the_fetchnode(event);
608 goto next_thread; /* handle next event in event queue */
611 do_the_globalblock(event);
612 goto next_thread; /* handle next event in event queue */
615 do_the_fetchreply(event);
616 goto next_thread; /* handle next event in event queue */
618 case UnblockThread: /* Move from the blocked queue to the tail of */
619 do_the_unblock(event);
620 goto next_thread; /* handle next event in event queue */
622 case ResumeThread: /* Move from the blocked queue to the tail of */
623 /* the runnable queue ( i.e. Qu' SImqa'lu') */
624 event->tso->gran.blocktime +=
625 CurrentTime[CurrentProc] - event->tso->gran.blockedat;
626 do_the_startthread(event);
627 goto next_thread; /* handle next event in event queue */
630 do_the_startthread(event);
631 goto next_thread; /* handle next event in event queue */
634 do_the_movethread(event);
635 goto next_thread; /* handle next event in event queue */
638 do_the_movespark(event);
639 goto next_thread; /* handle next event in event queue */
642 do_the_findwork(event);
643 goto next_thread; /* handle next event in event queue */
646 barf("Illegal event type %u\n", event->evttype);
649 /* This point was scheduler_loop in the old RTS */
651 IF_DEBUG(gran, belch("GRAN: after main switch"));
653 TimeOfLastEvent = CurrentTime[CurrentProc];
654 TimeOfNextEvent = get_time_of_next_event();
655 IgnoreEvents=(TimeOfNextEvent==0); // HWL HACK
656 // CurrentTSO = ThreadQueueHd;
658 IF_DEBUG(gran, belch("GRAN: time of next event is: %ld",
661 if (RtsFlags.GranFlags.Light)
662 GranSimLight_leave_system(event, &ActiveTSO);
664 EndOfTimeSlice = CurrentTime[CurrentProc]+RtsFlags.GranFlags.time_slice;
667 belch("GRAN: end of time-slice is %#lx", EndOfTimeSlice));
669 /* in a GranSim setup the TSO stays on the run queue */
671 /* Take a thread from the run queue. */
672 POP_RUN_QUEUE(t); // take_off_run_queue(t);
675 fprintf(stderr, "GRAN: About to run current thread, which is\n");
678 context_switch = 0; // turned on via GranYield, checking events and time slice
681 DumpGranEvent(GR_SCHEDULE, t));
683 procStatus[CurrentProc] = Busy;
686 if (PendingFetches != END_BF_QUEUE) {
690 /* ToDo: phps merge with spark activation above */
691 /* check whether we have local work and send requests if we have none */
692 if (EMPTY_RUN_QUEUE()) { /* no runnable threads */
693 /* :-[ no local threads => look out for local sparks */
694 /* the spark pool for the current PE */
695 pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable
696 if (advisory_thread_count < RtsFlags.ParFlags.maxThreads &&
697 pool->hd < pool->tl) {
699 * ToDo: add GC code check that we really have enough heap afterwards!!
701 * If we're here (no runnable threads) and we have pending
702 * sparks, we must have a space problem. Get enough space
703 * to turn one of those pending sparks into a
707 spark = findSpark(rtsFalse); /* get a spark */
708 if (spark != (rtsSpark) NULL) {
709 tso = activateSpark(spark); /* turn the spark into a thread */
710 IF_PAR_DEBUG(schedule,
711 belch("==== schedule: Created TSO %d (%p); %d threads active",
712 tso->id, tso, advisory_thread_count));
714 if (tso==END_TSO_QUEUE) { /* failed to activate spark->back to loop */
715 belch("==^^ failed to activate spark");
717 } /* otherwise fall through & pick-up new tso */
719 IF_PAR_DEBUG(verbose,
720 belch("==^^ no local sparks (spark pool contains only NFs: %d)",
721 spark_queue_len(pool)));
726 /* If we still have no work we need to send a FISH to get a spark
729 if (EMPTY_RUN_QUEUE()) {
730 /* =8-[ no local sparks => look for work on other PEs */
732 * We really have absolutely no work. Send out a fish
733 * (there may be some out there already), and wait for
734 * something to arrive. We clearly can't run any threads
735 * until a SCHEDULE or RESUME arrives, and so that's what
736 * we're hoping to see. (Of course, we still have to
737 * respond to other types of messages.)
739 TIME now = msTime() /*CURRENT_TIME*/;
740 IF_PAR_DEBUG(verbose,
741 belch("-- now=%ld", now));
742 IF_PAR_DEBUG(verbose,
743 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
744 (last_fish_arrived_at!=0 &&
745 last_fish_arrived_at+RtsFlags.ParFlags.fishDelay > now)) {
746 belch("--$$ delaying FISH until %ld (last fish %ld, delay %ld, now %ld)",
747 last_fish_arrived_at+RtsFlags.ParFlags.fishDelay,
748 last_fish_arrived_at,
749 RtsFlags.ParFlags.fishDelay, now);
752 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
753 (last_fish_arrived_at==0 ||
754 (last_fish_arrived_at+RtsFlags.ParFlags.fishDelay <= now))) {
755 /* outstandingFishes is set in sendFish, processFish;
756 avoid flooding system with fishes via delay */
758 sendFish(pe, mytid, NEW_FISH_AGE, NEW_FISH_HISTORY,
761 // Global statistics: count no. of fishes
762 if (RtsFlags.ParFlags.ParStats.Global &&
763 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
764 globalParStats.tot_fish_mess++;
768 receivedFinish = processMessages();
771 } else if (PacketsWaiting()) { /* Look for incoming messages */
772 receivedFinish = processMessages();
775 /* Now we are sure that we have some work available */
776 ASSERT(run_queue_hd != END_TSO_QUEUE);
778 /* Take a thread from the run queue, if we have work */
779 POP_RUN_QUEUE(t); // take_off_run_queue(END_TSO_QUEUE);
780 IF_DEBUG(sanity,checkTSO(t));
782 /* ToDo: write something to the log-file
783 if (RTSflags.ParFlags.granSimStats && !sameThread)
784 DumpGranEvent(GR_SCHEDULE, RunnableThreadsHd);
788 /* the spark pool for the current PE */
789 pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable
792 belch("--=^ %d threads, %d sparks on [%#x]",
793 run_queue_len(), spark_queue_len(pool), CURRENT_PROC));
796 if (0 && RtsFlags.ParFlags.ParStats.Full &&
797 t && LastTSO && t->id != LastTSO->id &&
798 LastTSO->why_blocked == NotBlocked &&
799 LastTSO->what_next != ThreadComplete) {
800 // if previously scheduled TSO not blocked we have to record the context switch
801 DumpVeryRawGranEvent(TimeOfLastYield, CURRENT_PROC, CURRENT_PROC,
802 GR_DESCHEDULE, LastTSO, (StgClosure *)NULL, 0, 0);
805 if (RtsFlags.ParFlags.ParStats.Full &&
806 (emitSchedule /* forced emit */ ||
807 (t && LastTSO && t->id != LastTSO->id))) {
809 we are running a different TSO, so write a schedule event to log file
810 NB: If we use fair scheduling we also have to write a deschedule
811 event for LastTSO; with unfair scheduling we know that the
812 previous tso has blocked whenever we switch to another tso, so
813 we don't need it in GUM for now
815 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
816 GR_SCHEDULE, t, (StgClosure *)NULL, 0, 0);
817 emitSchedule = rtsFalse;
821 #else /* !GRAN && !PAR */
823 // grab a thread from the run queue
824 ASSERT(run_queue_hd != END_TSO_QUEUE);
827 // Sanity check the thread we're about to run. This can be
828 // expensive if there is lots of thread switching going on...
829 IF_DEBUG(sanity,checkTSO(t));
834 StgMainThread *m = t->main;
841 sched_belch("### Running thread %d in bound thread", t->id));
842 // yes, the Haskell thread is bound to the current native thread
847 sched_belch("### thread %d bound to another OS thread", t->id));
848 // no, bound to a different Haskell thread: pass to that thread
849 PUSH_ON_RUN_QUEUE(t);
850 passCapability(&m->bound_thread_cond);
856 if(mainThread != NULL)
857 // The thread we want to run is bound.
860 sched_belch("### this OS thread cannot run thread %d", t->id));
861 // no, the current native thread is bound to a different
862 // Haskell thread, so pass it to any worker thread
863 PUSH_ON_RUN_QUEUE(t);
864 passCapabilityToWorker();
871 cap->r.rCurrentTSO = t;
873 /* context switches are now initiated by the timer signal, unless
874 * the user specified "context switch as often as possible", with
877 if ((RtsFlags.ConcFlags.ctxtSwitchTicks == 0
878 && (run_queue_hd != END_TSO_QUEUE
879 || blocked_queue_hd != END_TSO_QUEUE
880 || sleeping_queue != END_TSO_QUEUE)))
885 RELEASE_LOCK(&sched_mutex);
887 IF_DEBUG(scheduler, sched_belch("-->> running thread %ld %s ...",
888 t->id, whatNext_strs[t->what_next]));
891 startHeapProfTimer();
894 /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
895 /* Run the current thread
897 prev_what_next = t->what_next;
899 errno = t->saved_errno;
901 switch (prev_what_next) {
905 /* Thread already finished, return to scheduler. */
906 ret = ThreadFinished;
910 ret = StgRun((StgFunPtr) stg_returnToStackTop, &cap->r);
913 case ThreadInterpret:
914 ret = interpretBCO(cap);
918 barf("schedule: invalid what_next field");
921 // The TSO might have moved, so find the new location:
922 t = cap->r.rCurrentTSO;
924 // And save the current errno in this thread.
925 t->saved_errno = errno;
927 /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
929 /* Costs for the scheduler are assigned to CCS_SYSTEM */
935 ACQUIRE_LOCK(&sched_mutex);
937 #ifdef RTS_SUPPORTS_THREADS
938 IF_DEBUG(scheduler,fprintf(stderr,"sched (task %p): ", osThreadId()););
939 #elif !defined(GRAN) && !defined(PAR)
940 IF_DEBUG(scheduler,fprintf(stderr,"sched: "););
944 /* HACK 675: if the last thread didn't yield, make sure to print a
945 SCHEDULE event to the log file when StgRunning the next thread, even
946 if it is the same one as before */
948 TimeOfLastYield = CURRENT_TIME;
954 IF_DEBUG(gran, DumpGranEvent(GR_DESCHEDULE, t));
955 globalGranStats.tot_heapover++;
957 globalParStats.tot_heapover++;
960 // did the task ask for a large block?
961 if (cap->r.rHpAlloc > BLOCK_SIZE_W) {
962 // if so, get one and push it on the front of the nursery.
966 blocks = (nat)BLOCK_ROUND_UP(cap->r.rHpAlloc * sizeof(W_)) / BLOCK_SIZE;
968 IF_DEBUG(scheduler,belch("--<< thread %ld (%s) stopped: requesting a large block (size %d)",
969 t->id, whatNext_strs[t->what_next], blocks));
971 // don't do this if it would push us over the
972 // alloc_blocks_lim limit; we'll GC first.
973 if (alloc_blocks + blocks < alloc_blocks_lim) {
975 alloc_blocks += blocks;
976 bd = allocGroup( blocks );
978 // link the new group into the list
979 bd->link = cap->r.rCurrentNursery;
980 bd->u.back = cap->r.rCurrentNursery->u.back;
981 if (cap->r.rCurrentNursery->u.back != NULL) {
982 cap->r.rCurrentNursery->u.back->link = bd;
984 ASSERT(g0s0->blocks == cap->r.rCurrentNursery &&
985 g0s0->blocks == cap->r.rNursery);
986 cap->r.rNursery = g0s0->blocks = bd;
988 cap->r.rCurrentNursery->u.back = bd;
990 // initialise it as a nursery block. We initialise the
991 // step, gen_no, and flags field of *every* sub-block in
992 // this large block, because this is easier than making
993 // sure that we always find the block head of a large
994 // block whenever we call Bdescr() (eg. evacuate() and
995 // isAlive() in the GC would both have to do this, at
999 for (x = bd; x < bd + blocks; x++) {
1006 // don't forget to update the block count in g0s0.
1007 g0s0->n_blocks += blocks;
1008 // This assert can be a killer if the app is doing lots
1009 // of large block allocations.
1010 ASSERT(countBlocks(g0s0->blocks) == g0s0->n_blocks);
1012 // now update the nursery to point to the new block
1013 cap->r.rCurrentNursery = bd;
1015 // we might be unlucky and have another thread get on the
1016 // run queue before us and steal the large block, but in that
1017 // case the thread will just end up requesting another large
1019 PUSH_ON_RUN_QUEUE(t);
1024 /* make all the running tasks block on a condition variable,
1025 * maybe set context_switch and wait till they all pile in,
1026 * then have them wait on a GC condition variable.
1028 IF_DEBUG(scheduler,belch("--<< thread %ld (%s) stopped: HeapOverflow",
1029 t->id, whatNext_strs[t->what_next]));
1032 ASSERT(!is_on_queue(t,CurrentProc));
1034 /* Currently we emit a DESCHEDULE event before GC in GUM.
1035 ToDo: either add separate event to distinguish SYSTEM time from rest
1036 or just nuke this DESCHEDULE (and the following SCHEDULE) */
1037 if (0 && RtsFlags.ParFlags.ParStats.Full) {
1038 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
1039 GR_DESCHEDULE, t, (StgClosure *)NULL, 0, 0);
1040 emitSchedule = rtsTrue;
1044 ready_to_gc = rtsTrue;
1045 context_switch = 1; /* stop other threads ASAP */
1046 PUSH_ON_RUN_QUEUE(t);
1047 /* actual GC is done at the end of the while loop */
1053 DumpGranEvent(GR_DESCHEDULE, t));
1054 globalGranStats.tot_stackover++;
1057 // DumpGranEvent(GR_DESCHEDULE, t);
1058 globalParStats.tot_stackover++;
1060 IF_DEBUG(scheduler,belch("--<< thread %ld (%s) stopped, StackOverflow",
1061 t->id, whatNext_strs[t->what_next]));
1062 /* just adjust the stack for this thread, then pop it back
1067 /* enlarge the stack */
1068 StgTSO *new_t = threadStackOverflow(t);
1070 /* This TSO has moved, so update any pointers to it from the
1071 * main thread stack. It better not be on any other queues...
1072 * (it shouldn't be).
1074 if (t->main != NULL) {
1075 t->main->tso = new_t;
1077 PUSH_ON_RUN_QUEUE(new_t);
1081 case ThreadYielding:
1082 // Reset the context switch flag. We don't do this just before
1083 // running the thread, because that would mean we would lose ticks
1084 // during GC, which can lead to unfair scheduling (a thread hogs
1085 // the CPU because the tick always arrives during GC). This way
1086 // penalises threads that do a lot of allocation, but that seems
1087 // better than the alternative.
1092 DumpGranEvent(GR_DESCHEDULE, t));
1093 globalGranStats.tot_yields++;
1096 // DumpGranEvent(GR_DESCHEDULE, t);
1097 globalParStats.tot_yields++;
1099 /* put the thread back on the run queue. Then, if we're ready to
1100 * GC, check whether this is the last task to stop. If so, wake
1101 * up the GC thread. getThread will block during a GC until the
1105 if (t->what_next != prev_what_next) {
1106 belch("--<< thread %ld (%s) stopped to switch evaluators",
1107 t->id, whatNext_strs[t->what_next]);
1109 belch("--<< thread %ld (%s) stopped, yielding",
1110 t->id, whatNext_strs[t->what_next]);
1115 //belch("&& Doing sanity check on yielding TSO %ld.", t->id);
1117 ASSERT(t->link == END_TSO_QUEUE);
1119 // Shortcut if we're just switching evaluators: don't bother
1120 // doing stack squeezing (which can be expensive), just run the
1122 if (t->what_next != prev_what_next) {
1129 ASSERT(!is_on_queue(t,CurrentProc));
1132 //belch("&& Doing sanity check on all ThreadQueues (and their TSOs).");
1133 checkThreadQsSanity(rtsTrue));
1137 if (RtsFlags.ParFlags.doFairScheduling) {
1138 /* this does round-robin scheduling; good for concurrency */
1139 APPEND_TO_RUN_QUEUE(t);
1141 /* this does unfair scheduling; good for parallelism */
1142 PUSH_ON_RUN_QUEUE(t);
1145 // this does round-robin scheduling; good for concurrency
1146 APPEND_TO_RUN_QUEUE(t);
1150 /* add a ContinueThread event to actually process the thread */
1151 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1153 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1155 belch("GRAN: eventq and runnableq after adding yielded thread to queue again:");
1164 belch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: ",
1165 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)));
1166 if (t->block_info.closure!=(StgClosure*)NULL) print_bq(t->block_info.closure));
1168 // ??? needed; should emit block before
1170 DumpGranEvent(GR_DESCHEDULE, t));
1171 prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
1174 ASSERT(procStatus[CurrentProc]==Busy ||
1175 ((procStatus[CurrentProc]==Fetching) &&
1176 (t->block_info.closure!=(StgClosure*)NULL)));
1177 if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
1178 !(!RtsFlags.GranFlags.DoAsyncFetch &&
1179 procStatus[CurrentProc]==Fetching))
1180 procStatus[CurrentProc] = Idle;
1184 belch("--<< thread %ld (%p; %s) stopped, blocking on node %p with BQ: ",
1185 t->id, t, whatNext_strs[t->what_next], t->block_info.closure));
1188 if (t->block_info.closure!=(StgClosure*)NULL)
1189 print_bq(t->block_info.closure));
1191 /* Send a fetch (if BlockedOnGA) and dump event to log file */
1194 /* whatever we schedule next, we must log that schedule */
1195 emitSchedule = rtsTrue;
1198 /* don't need to do anything. Either the thread is blocked on
1199 * I/O, in which case we'll have called addToBlockedQueue
1200 * previously, or it's blocked on an MVar or Blackhole, in which
1201 * case it'll be on the relevant queue already.
1204 fprintf(stderr, "--<< thread %d (%s) stopped: ",
1205 t->id, whatNext_strs[t->what_next]);
1206 printThreadBlockage(t);
1207 fprintf(stderr, "\n"));
1210 /* Only for dumping event to log file
1211 ToDo: do I need this in GranSim, too?
1218 case ThreadFinished:
1219 /* Need to check whether this was a main thread, and if so, signal
1220 * the task that started it with the return value. If we have no
1221 * more main threads, we probably need to stop all the tasks until
1224 /* We also end up here if the thread kills itself with an
1225 * uncaught exception, see Exception.hc.
1227 IF_DEBUG(scheduler,belch("--++ thread %d (%s) finished",
1228 t->id, whatNext_strs[t->what_next]));
1230 endThread(t, CurrentProc); // clean-up the thread
1232 /* For now all are advisory -- HWL */
1233 //if(t->priority==AdvisoryPriority) ??
1234 advisory_thread_count--;
1237 if(t->dist.priority==RevalPriority)
1241 if (RtsFlags.ParFlags.ParStats.Full &&
1242 !RtsFlags.ParFlags.ParStats.Suppressed)
1243 DumpEndEvent(CURRENT_PROC, t, rtsFalse /* not mandatory */);
1247 // Check whether the thread that just completed was a main
1248 // thread, and if so return with the result.
1250 // There is an assumption here that all thread completion goes
1251 // through this point; we need to make sure that if a thread
1252 // ends up in the ThreadKilled state, that it stays on the run
1253 // queue so it can be dealt with here.
1256 #if defined(RTS_SUPPORTS_THREADS)
1259 mainThread->tso == t
1263 // We are a bound thread: this must be our thread that just
1265 ASSERT(mainThread->tso == t);
1267 if (t->what_next == ThreadComplete) {
1268 if (mainThread->ret) {
1269 // NOTE: return val is tso->sp[1] (see StgStartup.hc)
1270 *(mainThread->ret) = (StgClosure *)mainThread->tso->sp[1];
1272 mainThread->stat = Success;
1274 if (mainThread->ret) {
1275 *(mainThread->ret) = NULL;
1277 if (was_interrupted) {
1278 mainThread->stat = Interrupted;
1280 mainThread->stat = Killed;
1284 removeThreadLabel((StgWord)mainThread->tso->id);
1286 if (mainThread->prev == NULL) {
1287 main_threads = mainThread->link;
1289 mainThread->prev->link = mainThread->link;
1291 if (mainThread->link != NULL) {
1292 mainThread->link->prev = NULL;
1294 releaseCapability(cap);
1298 #ifdef RTS_SUPPORTS_THREADS
1299 ASSERT(t->main == NULL);
1301 if (t->main != NULL) {
1302 // Must be a main thread that is not the topmost one. Leave
1303 // it on the run queue until the stack has unwound to the
1304 // point where we can deal with this. Leaving it on the run
1305 // queue also ensures that the garbage collector knows about
1306 // this thread and its return value (it gets dropped from the
1307 // all_threads list so there's no other way to find it).
1308 APPEND_TO_RUN_QUEUE(t);
1314 barf("schedule: invalid thread return code %d", (int)ret);
1318 // When we have +RTS -i0 and we're heap profiling, do a census at
1319 // every GC. This lets us get repeatable runs for debugging.
1320 if (performHeapProfile ||
1321 (RtsFlags.ProfFlags.profileInterval==0 &&
1322 RtsFlags.ProfFlags.doHeapProfile && ready_to_gc)) {
1323 GarbageCollect(GetRoots, rtsTrue);
1325 performHeapProfile = rtsFalse;
1326 ready_to_gc = rtsFalse; // we already GC'd
1331 /* everybody back, start the GC.
1332 * Could do it in this thread, or signal a condition var
1333 * to do it in another thread. Either way, we need to
1334 * broadcast on gc_pending_cond afterward.
1336 #if defined(RTS_SUPPORTS_THREADS)
1337 IF_DEBUG(scheduler,sched_belch("doing GC"));
1339 GarbageCollect(GetRoots,rtsFalse);
1340 ready_to_gc = rtsFalse;
1342 /* add a ContinueThread event to continue execution of current thread */
1343 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1345 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1347 fprintf(stderr, "GRAN: eventq and runnableq after Garbage collection:\n");
1355 IF_GRAN_DEBUG(unused,
1356 print_eventq(EventHd));
1358 event = get_next_event();
1361 /* ToDo: wait for next message to arrive rather than busy wait */
1364 } /* end of while(1) */
1366 IF_PAR_DEBUG(verbose,
1367 belch("== Leaving schedule() after having received Finish"));
1370 /* ---------------------------------------------------------------------------
1371 * rtsSupportsBoundThreads(): is the RTS built to support bound threads?
1372 * used by Control.Concurrent for error checking.
1373 * ------------------------------------------------------------------------- */
1376 rtsSupportsBoundThreads(void)
1385 /* ---------------------------------------------------------------------------
1386 * isThreadBound(tso): check whether tso is bound to an OS thread.
1387 * ------------------------------------------------------------------------- */
1390 isThreadBound(StgTSO* tso USED_IN_THREADED_RTS)
1393 return (tso->main != NULL);
1398 /* ---------------------------------------------------------------------------
1399 * Singleton fork(). Do not copy any running threads.
1400 * ------------------------------------------------------------------------- */
1402 #ifndef mingw32_TARGET_OS
1403 #define FORKPROCESS_PRIMOP_SUPPORTED
1406 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
1408 deleteThreadImmediately(StgTSO *tso);
1411 forkProcess(HsStablePtr *entry
1412 #ifndef FORKPROCESS_PRIMOP_SUPPORTED
1417 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
1423 IF_DEBUG(scheduler,sched_belch("forking!"));
1424 rts_lock(); // This not only acquires sched_mutex, it also
1425 // makes sure that no other threads are running
1429 if (pid) { /* parent */
1431 /* just return the pid */
1435 } else { /* child */
1438 // delete all threads
1439 run_queue_hd = run_queue_tl = END_TSO_QUEUE;
1441 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
1444 // don't allow threads to catch the ThreadKilled exception
1445 deleteThreadImmediately(t);
1448 // wipe the main thread list
1449 while((m = main_threads) != NULL) {
1450 main_threads = m->link;
1451 # ifdef THREADED_RTS
1452 closeCondition(&m->bound_thread_cond);
1457 # ifdef RTS_SUPPORTS_THREADS
1458 resetTaskManagerAfterFork(); // tell startTask() and friends that
1459 startingWorkerThread = rtsFalse; // we have no worker threads any more
1460 resetWorkerWakeupPipeAfterFork();
1463 rc = rts_evalStableIO(entry, NULL); // run the action
1464 rts_checkSchedStatus("forkProcess",rc);
1468 hs_exit(); // clean up and exit
1471 #else /* !FORKPROCESS_PRIMOP_SUPPORTED */
1472 barf("forkProcess#: primop not supported, sorry!\n");
1477 /* ---------------------------------------------------------------------------
1478 * deleteAllThreads(): kill all the live threads.
1480 * This is used when we catch a user interrupt (^C), before performing
1481 * any necessary cleanups and running finalizers.
1483 * Locks: sched_mutex held.
1484 * ------------------------------------------------------------------------- */
1487 deleteAllThreads ( void )
1490 IF_DEBUG(scheduler,sched_belch("deleting all threads"));
1491 for (t = all_threads; t != END_TSO_QUEUE; t = next) {
1492 next = t->global_link;
1496 // The run queue now contains a bunch of ThreadKilled threads. We
1497 // must not throw these away: the main thread(s) will be in there
1498 // somewhere, and the main scheduler loop has to deal with it.
1499 // Also, the run queue is the only thing keeping these threads from
1500 // being GC'd, and we don't want the "main thread has been GC'd" panic.
1502 ASSERT(blocked_queue_hd == END_TSO_QUEUE);
1503 ASSERT(sleeping_queue == END_TSO_QUEUE);
1506 /* startThread and insertThread are now in GranSim.c -- HWL */
1509 /* ---------------------------------------------------------------------------
1510 * Suspending & resuming Haskell threads.
1512 * When making a "safe" call to C (aka _ccall_GC), the task gives back
1513 * its capability before calling the C function. This allows another
1514 * task to pick up the capability and carry on running Haskell
1515 * threads. It also means that if the C call blocks, it won't lock
1518 * The Haskell thread making the C call is put to sleep for the
1519 * duration of the call, on the susepended_ccalling_threads queue. We
1520 * give out a token to the task, which it can use to resume the thread
1521 * on return from the C function.
1522 * ------------------------------------------------------------------------- */
1525 suspendThread( StgRegTable *reg,
1534 int saved_errno = errno;
1536 /* assume that *reg is a pointer to the StgRegTable part
1539 cap = (Capability *)((void *)((unsigned char*)reg - sizeof(StgFunTable)));
1541 ACQUIRE_LOCK(&sched_mutex);
1544 sched_belch("thread %d did a _ccall_gc (is_concurrent: %d)", cap->r.rCurrentTSO->id,concCall));
1546 // XXX this might not be necessary --SDM
1547 cap->r.rCurrentTSO->what_next = ThreadRunGHC;
1549 threadPaused(cap->r.rCurrentTSO);
1550 cap->r.rCurrentTSO->link = suspended_ccalling_threads;
1551 suspended_ccalling_threads = cap->r.rCurrentTSO;
1553 if(cap->r.rCurrentTSO->blocked_exceptions == NULL) {
1554 cap->r.rCurrentTSO->why_blocked = BlockedOnCCall;
1555 cap->r.rCurrentTSO->blocked_exceptions = END_TSO_QUEUE;
1557 cap->r.rCurrentTSO->why_blocked = BlockedOnCCall_NoUnblockExc;
1560 /* Use the thread ID as the token; it should be unique */
1561 tok = cap->r.rCurrentTSO->id;
1563 /* Hand back capability */
1564 releaseCapability(cap);
1566 #if defined(RTS_SUPPORTS_THREADS)
1567 /* Preparing to leave the RTS, so ensure there's a native thread/task
1568 waiting to take over.
1570 IF_DEBUG(scheduler, sched_belch("worker (token %d): leaving RTS", tok));
1573 /* Other threads _might_ be available for execution; signal this */
1575 RELEASE_LOCK(&sched_mutex);
1577 errno = saved_errno;
1582 resumeThread( StgInt tok,
1583 rtsBool concCall STG_UNUSED )
1585 StgTSO *tso, **prev;
1587 int saved_errno = errno;
1589 #if defined(RTS_SUPPORTS_THREADS)
1590 /* Wait for permission to re-enter the RTS with the result. */
1591 ACQUIRE_LOCK(&sched_mutex);
1592 waitForReturnCapability(&sched_mutex, &cap);
1594 IF_DEBUG(scheduler, sched_belch("worker (token %d): re-entering RTS", tok));
1596 grabCapability(&cap);
1599 /* Remove the thread off of the suspended list */
1600 prev = &suspended_ccalling_threads;
1601 for (tso = suspended_ccalling_threads;
1602 tso != END_TSO_QUEUE;
1603 prev = &tso->link, tso = tso->link) {
1604 if (tso->id == (StgThreadID)tok) {
1609 if (tso == END_TSO_QUEUE) {
1610 barf("resumeThread: thread not found");
1612 tso->link = END_TSO_QUEUE;
1614 if(tso->why_blocked == BlockedOnCCall) {
1615 awakenBlockedQueueNoLock(tso->blocked_exceptions);
1616 tso->blocked_exceptions = NULL;
1619 /* Reset blocking status */
1620 tso->why_blocked = NotBlocked;
1622 cap->r.rCurrentTSO = tso;
1623 RELEASE_LOCK(&sched_mutex);
1624 errno = saved_errno;
1629 /* ---------------------------------------------------------------------------
1631 * ------------------------------------------------------------------------ */
1632 static void unblockThread(StgTSO *tso);
1634 /* ---------------------------------------------------------------------------
1635 * Comparing Thread ids.
1637 * This is used from STG land in the implementation of the
1638 * instances of Eq/Ord for ThreadIds.
1639 * ------------------------------------------------------------------------ */
1642 cmp_thread(StgPtr tso1, StgPtr tso2)
1644 StgThreadID id1 = ((StgTSO *)tso1)->id;
1645 StgThreadID id2 = ((StgTSO *)tso2)->id;
1647 if (id1 < id2) return (-1);
1648 if (id1 > id2) return 1;
1652 /* ---------------------------------------------------------------------------
1653 * Fetching the ThreadID from an StgTSO.
1655 * This is used in the implementation of Show for ThreadIds.
1656 * ------------------------------------------------------------------------ */
1658 rts_getThreadId(StgPtr tso)
1660 return ((StgTSO *)tso)->id;
1665 labelThread(StgPtr tso, char *label)
1670 /* Caveat: Once set, you can only set the thread name to "" */
1671 len = strlen(label)+1;
1672 buf = stgMallocBytes(len * sizeof(char), "Schedule.c:labelThread()");
1673 strncpy(buf,label,len);
1674 /* Update will free the old memory for us */
1675 updateThreadLabel(((StgTSO *)tso)->id,buf);
1679 /* ---------------------------------------------------------------------------
1680 Create a new thread.
1682 The new thread starts with the given stack size. Before the
1683 scheduler can run, however, this thread needs to have a closure
1684 (and possibly some arguments) pushed on its stack. See
1685 pushClosure() in Schedule.h.
1687 createGenThread() and createIOThread() (in SchedAPI.h) are
1688 convenient packaged versions of this function.
1690 currently pri (priority) is only used in a GRAN setup -- HWL
1691 ------------------------------------------------------------------------ */
1693 /* currently pri (priority) is only used in a GRAN setup -- HWL */
1695 createThread(nat size, StgInt pri)
1698 createThread(nat size)
1705 /* First check whether we should create a thread at all */
1707 /* check that no more than RtsFlags.ParFlags.maxThreads threads are created */
1708 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads) {
1710 belch("{createThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
1711 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
1712 return END_TSO_QUEUE;
1718 ASSERT(!RtsFlags.GranFlags.Light || CurrentProc==0);
1721 // ToDo: check whether size = stack_size - TSO_STRUCT_SIZEW
1723 /* catch ridiculously small stack sizes */
1724 if (size < MIN_STACK_WORDS + TSO_STRUCT_SIZEW) {
1725 size = MIN_STACK_WORDS + TSO_STRUCT_SIZEW;
1728 stack_size = size - TSO_STRUCT_SIZEW;
1730 tso = (StgTSO *)allocate(size);
1731 TICK_ALLOC_TSO(stack_size, 0);
1733 SET_HDR(tso, &stg_TSO_info, CCS_SYSTEM);
1735 SET_GRAN_HDR(tso, ThisPE);
1738 // Always start with the compiled code evaluator
1739 tso->what_next = ThreadRunGHC;
1741 tso->id = next_thread_id++;
1742 tso->why_blocked = NotBlocked;
1743 tso->blocked_exceptions = NULL;
1745 tso->saved_errno = 0;
1748 tso->stack_size = stack_size;
1749 tso->max_stack_size = round_to_mblocks(RtsFlags.GcFlags.maxStkSize)
1751 tso->sp = (P_)&(tso->stack) + stack_size;
1754 tso->prof.CCCS = CCS_MAIN;
1757 /* put a stop frame on the stack */
1758 tso->sp -= sizeofW(StgStopFrame);
1759 SET_HDR((StgClosure*)tso->sp,(StgInfoTable *)&stg_stop_thread_info,CCS_SYSTEM);
1760 tso->link = END_TSO_QUEUE;
1764 /* uses more flexible routine in GranSim */
1765 insertThread(tso, CurrentProc);
1767 /* In a non-GranSim setup the pushing of a TSO onto the runq is separated
1773 if (RtsFlags.GranFlags.GranSimStats.Full)
1774 DumpGranEvent(GR_START,tso);
1776 if (RtsFlags.ParFlags.ParStats.Full)
1777 DumpGranEvent(GR_STARTQ,tso);
1778 /* HACk to avoid SCHEDULE
1782 /* Link the new thread on the global thread list.
1784 tso->global_link = all_threads;
1788 tso->dist.priority = MandatoryPriority; //by default that is...
1792 tso->gran.pri = pri;
1794 tso->gran.magic = TSO_MAGIC; // debugging only
1796 tso->gran.sparkname = 0;
1797 tso->gran.startedat = CURRENT_TIME;
1798 tso->gran.exported = 0;
1799 tso->gran.basicblocks = 0;
1800 tso->gran.allocs = 0;
1801 tso->gran.exectime = 0;
1802 tso->gran.fetchtime = 0;
1803 tso->gran.fetchcount = 0;
1804 tso->gran.blocktime = 0;
1805 tso->gran.blockcount = 0;
1806 tso->gran.blockedat = 0;
1807 tso->gran.globalsparks = 0;
1808 tso->gran.localsparks = 0;
1809 if (RtsFlags.GranFlags.Light)
1810 tso->gran.clock = Now; /* local clock */
1812 tso->gran.clock = 0;
1814 IF_DEBUG(gran,printTSO(tso));
1817 tso->par.magic = TSO_MAGIC; // debugging only
1819 tso->par.sparkname = 0;
1820 tso->par.startedat = CURRENT_TIME;
1821 tso->par.exported = 0;
1822 tso->par.basicblocks = 0;
1823 tso->par.allocs = 0;
1824 tso->par.exectime = 0;
1825 tso->par.fetchtime = 0;
1826 tso->par.fetchcount = 0;
1827 tso->par.blocktime = 0;
1828 tso->par.blockcount = 0;
1829 tso->par.blockedat = 0;
1830 tso->par.globalsparks = 0;
1831 tso->par.localsparks = 0;
1835 globalGranStats.tot_threads_created++;
1836 globalGranStats.threads_created_on_PE[CurrentProc]++;
1837 globalGranStats.tot_sq_len += spark_queue_len(CurrentProc);
1838 globalGranStats.tot_sq_probes++;
1840 // collect parallel global statistics (currently done together with GC stats)
1841 if (RtsFlags.ParFlags.ParStats.Global &&
1842 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
1843 //fprintf(stderr, "Creating thread %d @ %11.2f\n", tso->id, usertime());
1844 globalParStats.tot_threads_created++;
1850 belch("==__ schedule: Created TSO %d (%p);",
1851 CurrentProc, tso, tso->id));
1853 IF_PAR_DEBUG(verbose,
1854 belch("==__ schedule: Created TSO %d (%p); %d threads active",
1855 tso->id, tso, advisory_thread_count));
1857 IF_DEBUG(scheduler,sched_belch("created thread %ld, stack size = %lx words",
1858 tso->id, tso->stack_size));
1865 all parallel thread creation calls should fall through the following routine.
1868 createSparkThread(rtsSpark spark)
1870 ASSERT(spark != (rtsSpark)NULL);
1871 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads)
1873 barf("{createSparkThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
1874 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
1875 return END_TSO_QUEUE;
1879 tso = createThread(RtsFlags.GcFlags.initialStkSize);
1880 if (tso==END_TSO_QUEUE)
1881 barf("createSparkThread: Cannot create TSO");
1883 tso->priority = AdvisoryPriority;
1885 pushClosure(tso,spark);
1886 PUSH_ON_RUN_QUEUE(tso);
1887 advisory_thread_count++;
1894 Turn a spark into a thread.
1895 ToDo: fix for SMP (needs to acquire SCHED_MUTEX!)
1899 activateSpark (rtsSpark spark)
1903 tso = createSparkThread(spark);
1904 if (RtsFlags.ParFlags.ParStats.Full) {
1905 //ASSERT(run_queue_hd == END_TSO_QUEUE); // I think ...
1906 IF_PAR_DEBUG(verbose,
1907 belch("==^^ activateSpark: turning spark of closure %p (%s) into a thread",
1908 (StgClosure *)spark, info_type((StgClosure *)spark)));
1910 // ToDo: fwd info on local/global spark to thread -- HWL
1911 // tso->gran.exported = spark->exported;
1912 // tso->gran.locked = !spark->global;
1913 // tso->gran.sparkname = spark->name;
1919 static SchedulerStatus waitThread_(/*out*/StgMainThread* m,
1920 Capability *initialCapability
1924 /* ---------------------------------------------------------------------------
1927 * scheduleThread puts a thread on the head of the runnable queue.
1928 * This will usually be done immediately after a thread is created.
1929 * The caller of scheduleThread must create the thread using e.g.
1930 * createThread and push an appropriate closure
1931 * on this thread's stack before the scheduler is invoked.
1932 * ------------------------------------------------------------------------ */
1934 static void scheduleThread_ (StgTSO* tso);
1937 scheduleThread_(StgTSO *tso)
1939 // Precondition: sched_mutex must be held.
1940 // The thread goes at the *end* of the run-queue, to avoid possible
1941 // starvation of any threads already on the queue.
1942 APPEND_TO_RUN_QUEUE(tso);
1947 scheduleThread(StgTSO* tso)
1949 ACQUIRE_LOCK(&sched_mutex);
1950 scheduleThread_(tso);
1951 RELEASE_LOCK(&sched_mutex);
1954 #if defined(RTS_SUPPORTS_THREADS)
1955 static Condition bound_cond_cache;
1956 static int bound_cond_cache_full = 0;
1961 scheduleWaitThread(StgTSO* tso, /*[out]*/HaskellObj* ret,
1962 Capability *initialCapability)
1964 // Precondition: sched_mutex must be held
1967 m = stgMallocBytes(sizeof(StgMainThread), "waitThread");
1972 m->link = main_threads;
1974 if (main_threads != NULL) {
1975 main_threads->prev = m;
1979 #if defined(RTS_SUPPORTS_THREADS)
1980 // Allocating a new condition for each thread is expensive, so we
1981 // cache one. This is a pretty feeble hack, but it helps speed up
1982 // consecutive call-ins quite a bit.
1983 if (bound_cond_cache_full) {
1984 m->bound_thread_cond = bound_cond_cache;
1985 bound_cond_cache_full = 0;
1987 initCondition(&m->bound_thread_cond);
1991 /* Put the thread on the main-threads list prior to scheduling the TSO.
1992 Failure to do so introduces a race condition in the MT case (as
1993 identified by Wolfgang Thaller), whereby the new task/OS thread
1994 created by scheduleThread_() would complete prior to the thread
1995 that spawned it managed to put 'itself' on the main-threads list.
1996 The upshot of it all being that the worker thread wouldn't get to
1997 signal the completion of the its work item for the main thread to
1998 see (==> it got stuck waiting.) -- sof 6/02.
2000 IF_DEBUG(scheduler, sched_belch("waiting for thread (%d)", tso->id));
2002 APPEND_TO_RUN_QUEUE(tso);
2003 // NB. Don't call THREAD_RUNNABLE() here, because the thread is
2004 // bound and only runnable by *this* OS thread, so waking up other
2005 // workers will just slow things down.
2007 return waitThread_(m, initialCapability);
2010 /* ---------------------------------------------------------------------------
2013 * Initialise the scheduler. This resets all the queues - if the
2014 * queues contained any threads, they'll be garbage collected at the
2017 * ------------------------------------------------------------------------ */
2025 for (i=0; i<=MAX_PROC; i++) {
2026 run_queue_hds[i] = END_TSO_QUEUE;
2027 run_queue_tls[i] = END_TSO_QUEUE;
2028 blocked_queue_hds[i] = END_TSO_QUEUE;
2029 blocked_queue_tls[i] = END_TSO_QUEUE;
2030 ccalling_threadss[i] = END_TSO_QUEUE;
2031 sleeping_queue = END_TSO_QUEUE;
2034 run_queue_hd = END_TSO_QUEUE;
2035 run_queue_tl = END_TSO_QUEUE;
2036 blocked_queue_hd = END_TSO_QUEUE;
2037 blocked_queue_tl = END_TSO_QUEUE;
2038 sleeping_queue = END_TSO_QUEUE;
2041 suspended_ccalling_threads = END_TSO_QUEUE;
2043 main_threads = NULL;
2044 all_threads = END_TSO_QUEUE;
2049 RtsFlags.ConcFlags.ctxtSwitchTicks =
2050 RtsFlags.ConcFlags.ctxtSwitchTime / TICK_MILLISECS;
2052 #if defined(RTS_SUPPORTS_THREADS)
2053 /* Initialise the mutex and condition variables used by
2055 initMutex(&sched_mutex);
2056 initMutex(&term_mutex);
2059 ACQUIRE_LOCK(&sched_mutex);
2061 /* A capability holds the state a native thread needs in
2062 * order to execute STG code. At least one capability is
2063 * floating around (only SMP builds have more than one).
2067 #if defined(RTS_SUPPORTS_THREADS)
2068 /* start our haskell execution tasks */
2069 startTaskManager(0,taskStart);
2072 #if /* defined(SMP) ||*/ defined(PAR)
2076 RELEASE_LOCK(&sched_mutex);
2080 exitScheduler( void )
2082 #if defined(RTS_SUPPORTS_THREADS)
2085 shutting_down_scheduler = rtsTrue;
2088 /* ----------------------------------------------------------------------------
2089 Managing the per-task allocation areas.
2091 Each capability comes with an allocation area. These are
2092 fixed-length block lists into which allocation can be done.
2094 ToDo: no support for two-space collection at the moment???
2095 ------------------------------------------------------------------------- */
2099 waitThread_(StgMainThread* m, Capability *initialCapability)
2101 SchedulerStatus stat;
2103 // Precondition: sched_mutex must be held.
2104 IF_DEBUG(scheduler, sched_belch("new main thread (%d)", m->tso->id));
2107 /* GranSim specific init */
2108 CurrentTSO = m->tso; // the TSO to run
2109 procStatus[MainProc] = Busy; // status of main PE
2110 CurrentProc = MainProc; // PE to run it on
2111 schedule(m,initialCapability);
2113 schedule(m,initialCapability);
2114 ASSERT(m->stat != NoStatus);
2119 #if defined(RTS_SUPPORTS_THREADS)
2120 // Free the condition variable, returning it to the cache if possible.
2121 if (!bound_cond_cache_full) {
2122 bound_cond_cache = m->bound_thread_cond;
2123 bound_cond_cache_full = 1;
2125 closeCondition(&m->bound_thread_cond);
2129 IF_DEBUG(scheduler, sched_belch("main thread (%d) finished", m->tso->id));
2132 // Postcondition: sched_mutex still held
2136 /* ---------------------------------------------------------------------------
2137 Where are the roots that we know about?
2139 - all the threads on the runnable queue
2140 - all the threads on the blocked queue
2141 - all the threads on the sleeping queue
2142 - all the thread currently executing a _ccall_GC
2143 - all the "main threads"
2145 ------------------------------------------------------------------------ */
2147 /* This has to be protected either by the scheduler monitor, or by the
2148 garbage collection monitor (probably the latter).
2153 GetRoots( evac_fn evac )
2158 for (i=0; i<=RtsFlags.GranFlags.proc; i++) {
2159 if ((run_queue_hds[i] != END_TSO_QUEUE) && ((run_queue_hds[i] != NULL)))
2160 evac((StgClosure **)&run_queue_hds[i]);
2161 if ((run_queue_tls[i] != END_TSO_QUEUE) && ((run_queue_tls[i] != NULL)))
2162 evac((StgClosure **)&run_queue_tls[i]);
2164 if ((blocked_queue_hds[i] != END_TSO_QUEUE) && ((blocked_queue_hds[i] != NULL)))
2165 evac((StgClosure **)&blocked_queue_hds[i]);
2166 if ((blocked_queue_tls[i] != END_TSO_QUEUE) && ((blocked_queue_tls[i] != NULL)))
2167 evac((StgClosure **)&blocked_queue_tls[i]);
2168 if ((ccalling_threadss[i] != END_TSO_QUEUE) && ((ccalling_threadss[i] != NULL)))
2169 evac((StgClosure **)&ccalling_threads[i]);
2176 if (run_queue_hd != END_TSO_QUEUE) {
2177 ASSERT(run_queue_tl != END_TSO_QUEUE);
2178 evac((StgClosure **)&run_queue_hd);
2179 evac((StgClosure **)&run_queue_tl);
2182 if (blocked_queue_hd != END_TSO_QUEUE) {
2183 ASSERT(blocked_queue_tl != END_TSO_QUEUE);
2184 evac((StgClosure **)&blocked_queue_hd);
2185 evac((StgClosure **)&blocked_queue_tl);
2188 if (sleeping_queue != END_TSO_QUEUE) {
2189 evac((StgClosure **)&sleeping_queue);
2193 if (suspended_ccalling_threads != END_TSO_QUEUE) {
2194 evac((StgClosure **)&suspended_ccalling_threads);
2197 #if defined(PAR) || defined(GRAN)
2198 markSparkQueue(evac);
2201 #if defined(RTS_USER_SIGNALS)
2202 // mark the signal handlers (signals should be already blocked)
2203 markSignalHandlers(evac);
2207 /* -----------------------------------------------------------------------------
2210 This is the interface to the garbage collector from Haskell land.
2211 We provide this so that external C code can allocate and garbage
2212 collect when called from Haskell via _ccall_GC.
2214 It might be useful to provide an interface whereby the programmer
2215 can specify more roots (ToDo).
2217 This needs to be protected by the GC condition variable above. KH.
2218 -------------------------------------------------------------------------- */
2220 static void (*extra_roots)(evac_fn);
2225 /* Obligated to hold this lock upon entry */
2226 ACQUIRE_LOCK(&sched_mutex);
2227 GarbageCollect(GetRoots,rtsFalse);
2228 RELEASE_LOCK(&sched_mutex);
2232 performMajorGC(void)
2234 ACQUIRE_LOCK(&sched_mutex);
2235 GarbageCollect(GetRoots,rtsTrue);
2236 RELEASE_LOCK(&sched_mutex);
2240 AllRoots(evac_fn evac)
2242 GetRoots(evac); // the scheduler's roots
2243 extra_roots(evac); // the user's roots
2247 performGCWithRoots(void (*get_roots)(evac_fn))
2249 ACQUIRE_LOCK(&sched_mutex);
2250 extra_roots = get_roots;
2251 GarbageCollect(AllRoots,rtsFalse);
2252 RELEASE_LOCK(&sched_mutex);
2255 /* -----------------------------------------------------------------------------
2258 If the thread has reached its maximum stack size, then raise the
2259 StackOverflow exception in the offending thread. Otherwise
2260 relocate the TSO into a larger chunk of memory and adjust its stack
2262 -------------------------------------------------------------------------- */
2265 threadStackOverflow(StgTSO *tso)
2267 nat new_stack_size, new_tso_size, stack_words;
2271 IF_DEBUG(sanity,checkTSO(tso));
2272 if (tso->stack_size >= tso->max_stack_size) {
2275 belch("@@ threadStackOverflow of TSO %d (%p): stack too large (now %ld; max is %ld)",
2276 tso->id, tso, tso->stack_size, tso->max_stack_size);
2277 /* If we're debugging, just print out the top of the stack */
2278 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2281 /* Send this thread the StackOverflow exception */
2282 raiseAsync(tso, (StgClosure *)stackOverflow_closure);
2286 /* Try to double the current stack size. If that takes us over the
2287 * maximum stack size for this thread, then use the maximum instead.
2288 * Finally round up so the TSO ends up as a whole number of blocks.
2290 new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
2291 new_tso_size = (nat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) +
2292 TSO_STRUCT_SIZE)/sizeof(W_);
2293 new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */
2294 new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
2296 IF_DEBUG(scheduler, fprintf(stderr,"== sched: increasing stack size from %d words to %d.\n", tso->stack_size, new_stack_size));
2298 dest = (StgTSO *)allocate(new_tso_size);
2299 TICK_ALLOC_TSO(new_stack_size,0);
2301 /* copy the TSO block and the old stack into the new area */
2302 memcpy(dest,tso,TSO_STRUCT_SIZE);
2303 stack_words = tso->stack + tso->stack_size - tso->sp;
2304 new_sp = (P_)dest + new_tso_size - stack_words;
2305 memcpy(new_sp, tso->sp, stack_words * sizeof(W_));
2307 /* relocate the stack pointers... */
2309 dest->stack_size = new_stack_size;
2311 /* Mark the old TSO as relocated. We have to check for relocated
2312 * TSOs in the garbage collector and any primops that deal with TSOs.
2314 * It's important to set the sp value to just beyond the end
2315 * of the stack, so we don't attempt to scavenge any part of the
2318 tso->what_next = ThreadRelocated;
2320 tso->sp = (P_)&(tso->stack[tso->stack_size]);
2321 tso->why_blocked = NotBlocked;
2322 dest->mut_link = NULL;
2324 IF_PAR_DEBUG(verbose,
2325 belch("@@ threadStackOverflow of TSO %d (now at %p): stack size increased to %ld",
2326 tso->id, tso, tso->stack_size);
2327 /* If we're debugging, just print out the top of the stack */
2328 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2331 IF_DEBUG(sanity,checkTSO(tso));
2333 IF_DEBUG(scheduler,printTSO(dest));
2339 /* ---------------------------------------------------------------------------
2340 Wake up a queue that was blocked on some resource.
2341 ------------------------------------------------------------------------ */
2345 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2350 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2352 /* write RESUME events to log file and
2353 update blocked and fetch time (depending on type of the orig closure) */
2354 if (RtsFlags.ParFlags.ParStats.Full) {
2355 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
2356 GR_RESUMEQ, ((StgTSO *)bqe), ((StgTSO *)bqe)->block_info.closure,
2357 0, 0 /* spark_queue_len(ADVISORY_POOL) */);
2358 if (EMPTY_RUN_QUEUE())
2359 emitSchedule = rtsTrue;
2361 switch (get_itbl(node)->type) {
2363 ((StgTSO *)bqe)->par.fetchtime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2368 ((StgTSO *)bqe)->par.blocktime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2375 barf("{unblockOneLocked}Daq Qagh: unexpected closure in blocking queue");
2382 static StgBlockingQueueElement *
2383 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2386 PEs node_loc, tso_loc;
2388 node_loc = where_is(node); // should be lifted out of loop
2389 tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2390 tso_loc = where_is((StgClosure *)tso);
2391 if (IS_LOCAL_TO(PROCS(node),tso_loc)) { // TSO is local
2392 /* !fake_fetch => TSO is on CurrentProc is same as IS_LOCAL_TO */
2393 ASSERT(CurrentProc!=node_loc || tso_loc==CurrentProc);
2394 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.lunblocktime;
2395 // insertThread(tso, node_loc);
2396 new_event(tso_loc, tso_loc, CurrentTime[CurrentProc],
2398 tso, node, (rtsSpark*)NULL);
2399 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2402 } else { // TSO is remote (actually should be FMBQ)
2403 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.mpacktime +
2404 RtsFlags.GranFlags.Costs.gunblocktime +
2405 RtsFlags.GranFlags.Costs.latency;
2406 new_event(tso_loc, CurrentProc, CurrentTime[CurrentProc],
2408 tso, node, (rtsSpark*)NULL);
2409 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2412 /* the thread-queue-overhead is accounted for in either Resume or UnblockThread */
2414 fprintf(stderr," %s TSO %d (%p) [PE %d] (block_info.closure=%p) (next=%p) ,",
2415 (node_loc==tso_loc ? "Local" : "Global"),
2416 tso->id, tso, CurrentProc, tso->block_info.closure, tso->link));
2417 tso->block_info.closure = NULL;
2418 IF_DEBUG(scheduler,belch("-- Waking up thread %ld (%p)",
2422 static StgBlockingQueueElement *
2423 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2425 StgBlockingQueueElement *next;
2427 switch (get_itbl(bqe)->type) {
2429 ASSERT(((StgTSO *)bqe)->why_blocked != NotBlocked);
2430 /* if it's a TSO just push it onto the run_queue */
2432 // ((StgTSO *)bqe)->link = END_TSO_QUEUE; // debugging?
2433 APPEND_TO_RUN_QUEUE((StgTSO *)bqe);
2435 unblockCount(bqe, node);
2436 /* reset blocking status after dumping event */
2437 ((StgTSO *)bqe)->why_blocked = NotBlocked;
2441 /* if it's a BLOCKED_FETCH put it on the PendingFetches list */
2443 bqe->link = (StgBlockingQueueElement *)PendingFetches;
2444 PendingFetches = (StgBlockedFetch *)bqe;
2448 /* can ignore this case in a non-debugging setup;
2449 see comments on RBHSave closures above */
2451 /* check that the closure is an RBHSave closure */
2452 ASSERT(get_itbl((StgClosure *)bqe) == &stg_RBH_Save_0_info ||
2453 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_1_info ||
2454 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_2_info);
2458 barf("{unblockOneLocked}Daq Qagh: Unexpected IP (%#lx; %s) in blocking queue at %#lx\n",
2459 get_itbl((StgClosure *)bqe), info_type((StgClosure *)bqe),
2463 IF_PAR_DEBUG(bq, fprintf(stderr, ", %p (%s)", bqe, info_type((StgClosure*)bqe)));
2467 #else /* !GRAN && !PAR */
2469 unblockOneLocked(StgTSO *tso)
2473 ASSERT(get_itbl(tso)->type == TSO);
2474 ASSERT(tso->why_blocked != NotBlocked);
2475 tso->why_blocked = NotBlocked;
2477 APPEND_TO_RUN_QUEUE(tso);
2479 IF_DEBUG(scheduler,sched_belch("waking up thread %ld", tso->id));
2484 #if defined(GRAN) || defined(PAR)
2485 INLINE_ME StgBlockingQueueElement *
2486 unblockOne(StgBlockingQueueElement *bqe, StgClosure *node)
2488 ACQUIRE_LOCK(&sched_mutex);
2489 bqe = unblockOneLocked(bqe, node);
2490 RELEASE_LOCK(&sched_mutex);
2495 unblockOne(StgTSO *tso)
2497 ACQUIRE_LOCK(&sched_mutex);
2498 tso = unblockOneLocked(tso);
2499 RELEASE_LOCK(&sched_mutex);
2506 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2508 StgBlockingQueueElement *bqe;
2513 belch("##-_ AwBQ for node %p on PE %d @ %ld by TSO %d (%p): ", \
2514 node, CurrentProc, CurrentTime[CurrentProc],
2515 CurrentTSO->id, CurrentTSO));
2517 node_loc = where_is(node);
2519 ASSERT(q == END_BQ_QUEUE ||
2520 get_itbl(q)->type == TSO || // q is either a TSO or an RBHSave
2521 get_itbl(q)->type == CONSTR); // closure (type constructor)
2522 ASSERT(is_unique(node));
2524 /* FAKE FETCH: magically copy the node to the tso's proc;
2525 no Fetch necessary because in reality the node should not have been
2526 moved to the other PE in the first place
2528 if (CurrentProc!=node_loc) {
2530 belch("## node %p is on PE %d but CurrentProc is %d (TSO %d); assuming fake fetch and adjusting bitmask (old: %#x)",
2531 node, node_loc, CurrentProc, CurrentTSO->id,
2532 // CurrentTSO, where_is(CurrentTSO),
2533 node->header.gran.procs));
2534 node->header.gran.procs = (node->header.gran.procs) | PE_NUMBER(CurrentProc);
2536 belch("## new bitmask of node %p is %#x",
2537 node, node->header.gran.procs));
2538 if (RtsFlags.GranFlags.GranSimStats.Global) {
2539 globalGranStats.tot_fake_fetches++;
2544 // ToDo: check: ASSERT(CurrentProc==node_loc);
2545 while (get_itbl(bqe)->type==TSO) { // q != END_TSO_QUEUE) {
2548 bqe points to the current element in the queue
2549 next points to the next element in the queue
2551 //tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2552 //tso_loc = where_is(tso);
2554 bqe = unblockOneLocked(bqe, node);
2557 /* if this is the BQ of an RBH, we have to put back the info ripped out of
2558 the closure to make room for the anchor of the BQ */
2559 if (bqe!=END_BQ_QUEUE) {
2560 ASSERT(get_itbl(node)->type == RBH && get_itbl(bqe)->type == CONSTR);
2562 ASSERT((info_ptr==&RBH_Save_0_info) ||
2563 (info_ptr==&RBH_Save_1_info) ||
2564 (info_ptr==&RBH_Save_2_info));
2566 /* cf. convertToRBH in RBH.c for writing the RBHSave closure */
2567 ((StgRBH *)node)->blocking_queue = (StgBlockingQueueElement *)((StgRBHSave *)bqe)->payload[0];
2568 ((StgRBH *)node)->mut_link = (StgMutClosure *)((StgRBHSave *)bqe)->payload[1];
2571 belch("## Filled in RBH_Save for %p (%s) at end of AwBQ",
2572 node, info_type(node)));
2575 /* statistics gathering */
2576 if (RtsFlags.GranFlags.GranSimStats.Global) {
2577 // globalGranStats.tot_bq_processing_time += bq_processing_time;
2578 globalGranStats.tot_bq_len += len; // total length of all bqs awakened
2579 // globalGranStats.tot_bq_len_local += len_local; // same for local TSOs only
2580 globalGranStats.tot_awbq++; // total no. of bqs awakened
2583 fprintf(stderr,"## BQ Stats of %p: [%d entries] %s\n",
2584 node, len, (bqe!=END_BQ_QUEUE) ? "RBH" : ""));
2588 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2590 StgBlockingQueueElement *bqe;
2592 ACQUIRE_LOCK(&sched_mutex);
2594 IF_PAR_DEBUG(verbose,
2595 belch("##-_ AwBQ for node %p on [%x]: ",
2599 if(get_itbl(q)->type == CONSTR || q==END_BQ_QUEUE) {
2600 IF_PAR_DEBUG(verbose, belch("## ... nothing to unblock so lets just return. RFP (BUG?)"));
2605 ASSERT(q == END_BQ_QUEUE ||
2606 get_itbl(q)->type == TSO ||
2607 get_itbl(q)->type == BLOCKED_FETCH ||
2608 get_itbl(q)->type == CONSTR);
2611 while (get_itbl(bqe)->type==TSO ||
2612 get_itbl(bqe)->type==BLOCKED_FETCH) {
2613 bqe = unblockOneLocked(bqe, node);
2615 RELEASE_LOCK(&sched_mutex);
2618 #else /* !GRAN && !PAR */
2621 awakenBlockedQueueNoLock(StgTSO *tso)
2623 while (tso != END_TSO_QUEUE) {
2624 tso = unblockOneLocked(tso);
2629 awakenBlockedQueue(StgTSO *tso)
2631 ACQUIRE_LOCK(&sched_mutex);
2632 while (tso != END_TSO_QUEUE) {
2633 tso = unblockOneLocked(tso);
2635 RELEASE_LOCK(&sched_mutex);
2639 /* ---------------------------------------------------------------------------
2641 - usually called inside a signal handler so it mustn't do anything fancy.
2642 ------------------------------------------------------------------------ */
2645 interruptStgRts(void)
2649 #ifdef RTS_SUPPORTS_THREADS
2650 wakeBlockedWorkerThread();
2654 /* -----------------------------------------------------------------------------
2657 This is for use when we raise an exception in another thread, which
2659 This has nothing to do with the UnblockThread event in GranSim. -- HWL
2660 -------------------------------------------------------------------------- */
2662 #if defined(GRAN) || defined(PAR)
2664 NB: only the type of the blocking queue is different in GranSim and GUM
2665 the operations on the queue-elements are the same
2666 long live polymorphism!
2668 Locks: sched_mutex is held upon entry and exit.
2672 unblockThread(StgTSO *tso)
2674 StgBlockingQueueElement *t, **last;
2676 switch (tso->why_blocked) {
2679 return; /* not blocked */
2682 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
2684 StgBlockingQueueElement *last_tso = END_BQ_QUEUE;
2685 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
2687 last = (StgBlockingQueueElement **)&mvar->head;
2688 for (t = (StgBlockingQueueElement *)mvar->head;
2690 last = &t->link, last_tso = t, t = t->link) {
2691 if (t == (StgBlockingQueueElement *)tso) {
2692 *last = (StgBlockingQueueElement *)tso->link;
2693 if (mvar->tail == tso) {
2694 mvar->tail = (StgTSO *)last_tso;
2699 barf("unblockThread (MVAR): TSO not found");
2702 case BlockedOnBlackHole:
2703 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
2705 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
2707 last = &bq->blocking_queue;
2708 for (t = bq->blocking_queue;
2710 last = &t->link, t = t->link) {
2711 if (t == (StgBlockingQueueElement *)tso) {
2712 *last = (StgBlockingQueueElement *)tso->link;
2716 barf("unblockThread (BLACKHOLE): TSO not found");
2719 case BlockedOnException:
2721 StgTSO *target = tso->block_info.tso;
2723 ASSERT(get_itbl(target)->type == TSO);
2725 if (target->what_next == ThreadRelocated) {
2726 target = target->link;
2727 ASSERT(get_itbl(target)->type == TSO);
2730 ASSERT(target->blocked_exceptions != NULL);
2732 last = (StgBlockingQueueElement **)&target->blocked_exceptions;
2733 for (t = (StgBlockingQueueElement *)target->blocked_exceptions;
2735 last = &t->link, t = t->link) {
2736 ASSERT(get_itbl(t)->type == TSO);
2737 if (t == (StgBlockingQueueElement *)tso) {
2738 *last = (StgBlockingQueueElement *)tso->link;
2742 barf("unblockThread (Exception): TSO not found");
2746 case BlockedOnWrite:
2747 #if defined(mingw32_TARGET_OS)
2748 case BlockedOnDoProc:
2751 /* take TSO off blocked_queue */
2752 StgBlockingQueueElement *prev = NULL;
2753 for (t = (StgBlockingQueueElement *)blocked_queue_hd; t != END_BQ_QUEUE;
2754 prev = t, t = t->link) {
2755 if (t == (StgBlockingQueueElement *)tso) {
2757 blocked_queue_hd = (StgTSO *)t->link;
2758 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
2759 blocked_queue_tl = END_TSO_QUEUE;
2762 prev->link = t->link;
2763 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
2764 blocked_queue_tl = (StgTSO *)prev;
2770 barf("unblockThread (I/O): TSO not found");
2773 case BlockedOnDelay:
2775 /* take TSO off sleeping_queue */
2776 StgBlockingQueueElement *prev = NULL;
2777 for (t = (StgBlockingQueueElement *)sleeping_queue; t != END_BQ_QUEUE;
2778 prev = t, t = t->link) {
2779 if (t == (StgBlockingQueueElement *)tso) {
2781 sleeping_queue = (StgTSO *)t->link;
2783 prev->link = t->link;
2788 barf("unblockThread (delay): TSO not found");
2792 barf("unblockThread");
2796 tso->link = END_TSO_QUEUE;
2797 tso->why_blocked = NotBlocked;
2798 tso->block_info.closure = NULL;
2799 PUSH_ON_RUN_QUEUE(tso);
2803 unblockThread(StgTSO *tso)
2807 /* To avoid locking unnecessarily. */
2808 if (tso->why_blocked == NotBlocked) {
2812 switch (tso->why_blocked) {
2815 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
2817 StgTSO *last_tso = END_TSO_QUEUE;
2818 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
2821 for (t = mvar->head; t != END_TSO_QUEUE;
2822 last = &t->link, last_tso = t, t = t->link) {
2825 if (mvar->tail == tso) {
2826 mvar->tail = last_tso;
2831 barf("unblockThread (MVAR): TSO not found");
2834 case BlockedOnBlackHole:
2835 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
2837 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
2839 last = &bq->blocking_queue;
2840 for (t = bq->blocking_queue; t != END_TSO_QUEUE;
2841 last = &t->link, t = t->link) {
2847 barf("unblockThread (BLACKHOLE): TSO not found");
2850 case BlockedOnException:
2852 StgTSO *target = tso->block_info.tso;
2854 ASSERT(get_itbl(target)->type == TSO);
2856 while (target->what_next == ThreadRelocated) {
2857 target = target->link;
2858 ASSERT(get_itbl(target)->type == TSO);
2861 ASSERT(target->blocked_exceptions != NULL);
2863 last = &target->blocked_exceptions;
2864 for (t = target->blocked_exceptions; t != END_TSO_QUEUE;
2865 last = &t->link, t = t->link) {
2866 ASSERT(get_itbl(t)->type == TSO);
2872 barf("unblockThread (Exception): TSO not found");
2876 case BlockedOnWrite:
2877 #if defined(mingw32_TARGET_OS)
2878 case BlockedOnDoProc:
2881 StgTSO *prev = NULL;
2882 for (t = blocked_queue_hd; t != END_TSO_QUEUE;
2883 prev = t, t = t->link) {
2886 blocked_queue_hd = t->link;
2887 if (blocked_queue_tl == t) {
2888 blocked_queue_tl = END_TSO_QUEUE;
2891 prev->link = t->link;
2892 if (blocked_queue_tl == t) {
2893 blocked_queue_tl = prev;
2899 barf("unblockThread (I/O): TSO not found");
2902 case BlockedOnDelay:
2904 StgTSO *prev = NULL;
2905 for (t = sleeping_queue; t != END_TSO_QUEUE;
2906 prev = t, t = t->link) {
2909 sleeping_queue = t->link;
2911 prev->link = t->link;
2916 barf("unblockThread (delay): TSO not found");
2920 barf("unblockThread");
2924 tso->link = END_TSO_QUEUE;
2925 tso->why_blocked = NotBlocked;
2926 tso->block_info.closure = NULL;
2927 APPEND_TO_RUN_QUEUE(tso);
2931 /* -----------------------------------------------------------------------------
2934 * The following function implements the magic for raising an
2935 * asynchronous exception in an existing thread.
2937 * We first remove the thread from any queue on which it might be
2938 * blocked. The possible blockages are MVARs and BLACKHOLE_BQs.
2940 * We strip the stack down to the innermost CATCH_FRAME, building
2941 * thunks in the heap for all the active computations, so they can
2942 * be restarted if necessary. When we reach a CATCH_FRAME, we build
2943 * an application of the handler to the exception, and push it on
2944 * the top of the stack.
2946 * How exactly do we save all the active computations? We create an
2947 * AP_STACK for every UpdateFrame on the stack. Entering one of these
2948 * AP_STACKs pushes everything from the corresponding update frame
2949 * upwards onto the stack. (Actually, it pushes everything up to the
2950 * next update frame plus a pointer to the next AP_STACK object.
2951 * Entering the next AP_STACK object pushes more onto the stack until we
2952 * reach the last AP_STACK object - at which point the stack should look
2953 * exactly as it did when we killed the TSO and we can continue
2954 * execution by entering the closure on top of the stack.
2956 * We can also kill a thread entirely - this happens if either (a) the
2957 * exception passed to raiseAsync is NULL, or (b) there's no
2958 * CATCH_FRAME on the stack. In either case, we strip the entire
2959 * stack and replace the thread with a zombie.
2961 * Locks: sched_mutex held upon entry nor exit.
2963 * -------------------------------------------------------------------------- */
2966 deleteThread(StgTSO *tso)
2968 raiseAsync(tso,NULL);
2971 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
2973 deleteThreadImmediately(StgTSO *tso)
2974 { // for forkProcess only:
2975 // delete thread without giving it a chance to catch the KillThread exception
2977 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
2981 if (tso->why_blocked != BlockedOnCCall &&
2982 tso->why_blocked != BlockedOnCCall_NoUnblockExc) {
2986 tso->what_next = ThreadKilled;
2991 raiseAsyncWithLock(StgTSO *tso, StgClosure *exception)
2993 /* When raising async exs from contexts where sched_mutex isn't held;
2994 use raiseAsyncWithLock(). */
2995 ACQUIRE_LOCK(&sched_mutex);
2996 raiseAsync(tso,exception);
2997 RELEASE_LOCK(&sched_mutex);
3001 raiseAsync(StgTSO *tso, StgClosure *exception)
3003 StgRetInfoTable *info;
3006 // Thread already dead?
3007 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
3012 sched_belch("raising exception in thread %ld.", tso->id));
3014 // Remove it from any blocking queues
3019 // The stack freezing code assumes there's a closure pointer on
3020 // the top of the stack, so we have to arrange that this is the case...
3022 if (sp[0] == (W_)&stg_enter_info) {
3026 sp[0] = (W_)&stg_dummy_ret_closure;
3032 // 1. Let the top of the stack be the "current closure"
3034 // 2. Walk up the stack until we find either an UPDATE_FRAME or a
3037 // 3. If it's an UPDATE_FRAME, then make an AP_STACK containing the
3038 // current closure applied to the chunk of stack up to (but not
3039 // including) the update frame. This closure becomes the "current
3040 // closure". Go back to step 2.
3042 // 4. If it's a CATCH_FRAME, then leave the exception handler on
3043 // top of the stack applied to the exception.
3045 // 5. If it's a STOP_FRAME, then kill the thread.
3050 info = get_ret_itbl((StgClosure *)frame);
3052 while (info->i.type != UPDATE_FRAME
3053 && (info->i.type != CATCH_FRAME || exception == NULL)
3054 && info->i.type != STOP_FRAME) {
3055 frame += stack_frame_sizeW((StgClosure *)frame);
3056 info = get_ret_itbl((StgClosure *)frame);
3059 switch (info->i.type) {
3062 // If we find a CATCH_FRAME, and we've got an exception to raise,
3063 // then build the THUNK raise(exception), and leave it on
3064 // top of the CATCH_FRAME ready to enter.
3068 StgCatchFrame *cf = (StgCatchFrame *)frame;
3072 // we've got an exception to raise, so let's pass it to the
3073 // handler in this frame.
3075 raise = (StgClosure *)allocate(sizeofW(StgClosure)+1);
3076 TICK_ALLOC_SE_THK(1,0);
3077 SET_HDR(raise,&stg_raise_info,cf->header.prof.ccs);
3078 raise->payload[0] = exception;
3080 // throw away the stack from Sp up to the CATCH_FRAME.
3084 /* Ensure that async excpetions are blocked now, so we don't get
3085 * a surprise exception before we get around to executing the
3088 if (tso->blocked_exceptions == NULL) {
3089 tso->blocked_exceptions = END_TSO_QUEUE;
3092 /* Put the newly-built THUNK on top of the stack, ready to execute
3093 * when the thread restarts.
3096 sp[-1] = (W_)&stg_enter_info;
3098 tso->what_next = ThreadRunGHC;
3099 IF_DEBUG(sanity, checkTSO(tso));
3108 // First build an AP_STACK consisting of the stack chunk above the
3109 // current update frame, with the top word on the stack as the
3112 words = frame - sp - 1;
3113 ap = (StgAP_STACK *)allocate(PAP_sizeW(words));
3116 ap->fun = (StgClosure *)sp[0];
3118 for(i=0; i < (nat)words; ++i) {
3119 ap->payload[i] = (StgClosure *)*sp++;
3122 SET_HDR(ap,&stg_AP_STACK_info,
3123 ((StgClosure *)frame)->header.prof.ccs /* ToDo */);
3124 TICK_ALLOC_UP_THK(words+1,0);
3127 fprintf(stderr, "sched: Updating ");
3128 printPtr((P_)((StgUpdateFrame *)frame)->updatee);
3129 fprintf(stderr, " with ");
3130 printObj((StgClosure *)ap);
3133 // Replace the updatee with an indirection - happily
3134 // this will also wake up any threads currently
3135 // waiting on the result.
3137 // Warning: if we're in a loop, more than one update frame on
3138 // the stack may point to the same object. Be careful not to
3139 // overwrite an IND_OLDGEN in this case, because we'll screw
3140 // up the mutable lists. To be on the safe side, don't
3141 // overwrite any kind of indirection at all. See also
3142 // threadSqueezeStack in GC.c, where we have to make a similar
3145 if (!closure_IND(((StgUpdateFrame *)frame)->updatee)) {
3146 // revert the black hole
3147 UPD_IND_NOLOCK(((StgUpdateFrame *)frame)->updatee,ap);
3149 sp += sizeofW(StgUpdateFrame) - 1;
3150 sp[0] = (W_)ap; // push onto stack
3155 // We've stripped the entire stack, the thread is now dead.
3156 sp += sizeofW(StgStopFrame);
3157 tso->what_next = ThreadKilled;
3168 /* -----------------------------------------------------------------------------
3169 resurrectThreads is called after garbage collection on the list of
3170 threads found to be garbage. Each of these threads will be woken
3171 up and sent a signal: BlockedOnDeadMVar if the thread was blocked
3172 on an MVar, or NonTermination if the thread was blocked on a Black
3175 Locks: sched_mutex isn't held upon entry nor exit.
3176 -------------------------------------------------------------------------- */
3179 resurrectThreads( StgTSO *threads )
3183 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
3184 next = tso->global_link;
3185 tso->global_link = all_threads;
3187 IF_DEBUG(scheduler, sched_belch("resurrecting thread %d", tso->id));
3189 switch (tso->why_blocked) {
3191 case BlockedOnException:
3192 /* Called by GC - sched_mutex lock is currently held. */
3193 raiseAsync(tso,(StgClosure *)BlockedOnDeadMVar_closure);
3195 case BlockedOnBlackHole:
3196 raiseAsync(tso,(StgClosure *)NonTermination_closure);
3199 /* This might happen if the thread was blocked on a black hole
3200 * belonging to a thread that we've just woken up (raiseAsync
3201 * can wake up threads, remember...).
3205 barf("resurrectThreads: thread blocked in a strange way");
3210 /* -----------------------------------------------------------------------------
3211 * Blackhole detection: if we reach a deadlock, test whether any
3212 * threads are blocked on themselves. Any threads which are found to
3213 * be self-blocked get sent a NonTermination exception.
3215 * This is only done in a deadlock situation in order to avoid
3216 * performance overhead in the normal case.
3218 * Locks: sched_mutex is held upon entry and exit.
3219 * -------------------------------------------------------------------------- */
3222 detectBlackHoles( void )
3224 StgTSO *tso = all_threads;
3226 StgClosure *blocked_on;
3227 StgRetInfoTable *info;
3229 for (tso = all_threads; tso != END_TSO_QUEUE; tso = tso->global_link) {
3231 while (tso->what_next == ThreadRelocated) {
3233 ASSERT(get_itbl(tso)->type == TSO);
3236 if (tso->why_blocked != BlockedOnBlackHole) {
3239 blocked_on = tso->block_info.closure;
3241 frame = (StgClosure *)tso->sp;
3244 info = get_ret_itbl(frame);
3245 switch (info->i.type) {
3247 if (((StgUpdateFrame *)frame)->updatee == blocked_on) {
3248 /* We are blocking on one of our own computations, so
3249 * send this thread the NonTermination exception.
3252 sched_belch("thread %d is blocked on itself", tso->id));
3253 raiseAsync(tso, (StgClosure *)NonTermination_closure);
3257 frame = (StgClosure *) ((StgUpdateFrame *)frame + 1);
3263 // normal stack frames; do nothing except advance the pointer
3265 (StgPtr)frame += stack_frame_sizeW(frame);
3272 /* ----------------------------------------------------------------------------
3273 * Debugging: why is a thread blocked
3274 * [Also provides useful information when debugging threaded programs
3275 * at the Haskell source code level, so enable outside of DEBUG. --sof 7/02]
3276 ------------------------------------------------------------------------- */
3280 printThreadBlockage(StgTSO *tso)
3282 switch (tso->why_blocked) {
3284 fprintf(stderr,"is blocked on read from fd %d", tso->block_info.fd);
3286 case BlockedOnWrite:
3287 fprintf(stderr,"is blocked on write to fd %d", tso->block_info.fd);
3289 #if defined(mingw32_TARGET_OS)
3290 case BlockedOnDoProc:
3291 fprintf(stderr,"is blocked on proc (request: %d)", tso->block_info.async_result->reqID);
3294 case BlockedOnDelay:
3295 fprintf(stderr,"is blocked until %d", tso->block_info.target);
3298 fprintf(stderr,"is blocked on an MVar");
3300 case BlockedOnException:
3301 fprintf(stderr,"is blocked on delivering an exception to thread %d",
3302 tso->block_info.tso->id);
3304 case BlockedOnBlackHole:
3305 fprintf(stderr,"is blocked on a black hole");
3308 fprintf(stderr,"is not blocked");
3312 fprintf(stderr,"is blocked on global address; local FM_BQ is %p (%s)",
3313 tso->block_info.closure, info_type(tso->block_info.closure));
3315 case BlockedOnGA_NoSend:
3316 fprintf(stderr,"is blocked on global address (no send); local FM_BQ is %p (%s)",
3317 tso->block_info.closure, info_type(tso->block_info.closure));
3320 case BlockedOnCCall:
3321 fprintf(stderr,"is blocked on an external call");
3323 case BlockedOnCCall_NoUnblockExc:
3324 fprintf(stderr,"is blocked on an external call (exceptions were already blocked)");
3327 barf("printThreadBlockage: strange tso->why_blocked: %d for TSO %d (%d)",
3328 tso->why_blocked, tso->id, tso);
3334 printThreadStatus(StgTSO *tso)
3336 switch (tso->what_next) {
3338 fprintf(stderr,"has been killed");
3340 case ThreadComplete:
3341 fprintf(stderr,"has completed");
3344 printThreadBlockage(tso);
3349 printAllThreads(void)
3355 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
3356 ullong_format_string(TIME_ON_PROC(CurrentProc),
3357 time_string, rtsFalse/*no commas!*/);
3359 fprintf(stderr, "all threads at [%s]:\n", time_string);
3361 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
3362 ullong_format_string(CURRENT_TIME,
3363 time_string, rtsFalse/*no commas!*/);
3365 fprintf(stderr,"all threads at [%s]:\n", time_string);
3367 fprintf(stderr,"all threads:\n");
3370 for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
3371 fprintf(stderr, "\tthread %d @ %p ", t->id, (void *)t);
3372 label = lookupThreadLabel(t->id);
3373 if (label) fprintf(stderr,"[\"%s\"] ",(char *)label);
3374 printThreadStatus(t);
3375 fprintf(stderr,"\n");
3382 Print a whole blocking queue attached to node (debugging only).
3386 print_bq (StgClosure *node)
3388 StgBlockingQueueElement *bqe;
3392 fprintf(stderr,"## BQ of closure %p (%s): ",
3393 node, info_type(node));
3395 /* should cover all closures that may have a blocking queue */
3396 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
3397 get_itbl(node)->type == FETCH_ME_BQ ||
3398 get_itbl(node)->type == RBH ||
3399 get_itbl(node)->type == MVAR);
3401 ASSERT(node!=(StgClosure*)NULL); // sanity check
3403 print_bqe(((StgBlockingQueue*)node)->blocking_queue);
3407 Print a whole blocking queue starting with the element bqe.
3410 print_bqe (StgBlockingQueueElement *bqe)
3415 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
3417 for (end = (bqe==END_BQ_QUEUE);
3418 !end; // iterate until bqe points to a CONSTR
3419 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE),
3420 bqe = end ? END_BQ_QUEUE : bqe->link) {
3421 ASSERT(bqe != END_BQ_QUEUE); // sanity check
3422 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
3423 /* types of closures that may appear in a blocking queue */
3424 ASSERT(get_itbl(bqe)->type == TSO ||
3425 get_itbl(bqe)->type == BLOCKED_FETCH ||
3426 get_itbl(bqe)->type == CONSTR);
3427 /* only BQs of an RBH end with an RBH_Save closure */
3428 //ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
3430 switch (get_itbl(bqe)->type) {
3432 fprintf(stderr," TSO %u (%x),",
3433 ((StgTSO *)bqe)->id, ((StgTSO *)bqe));
3436 fprintf(stderr," BF (node=%p, ga=((%x, %d, %x)),",
3437 ((StgBlockedFetch *)bqe)->node,
3438 ((StgBlockedFetch *)bqe)->ga.payload.gc.gtid,
3439 ((StgBlockedFetch *)bqe)->ga.payload.gc.slot,
3440 ((StgBlockedFetch *)bqe)->ga.weight);
3443 fprintf(stderr," %s (IP %p),",
3444 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
3445 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
3446 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
3447 "RBH_Save_?"), get_itbl(bqe));
3450 barf("Unexpected closure type %s in blocking queue", // of %p (%s)",
3451 info_type((StgClosure *)bqe)); // , node, info_type(node));
3455 fputc('\n', stderr);
3457 # elif defined(GRAN)
3459 print_bq (StgClosure *node)
3461 StgBlockingQueueElement *bqe;
3462 PEs node_loc, tso_loc;
3465 /* should cover all closures that may have a blocking queue */
3466 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
3467 get_itbl(node)->type == FETCH_ME_BQ ||
3468 get_itbl(node)->type == RBH);
3470 ASSERT(node!=(StgClosure*)NULL); // sanity check
3471 node_loc = where_is(node);
3473 fprintf(stderr,"## BQ of closure %p (%s) on [PE %d]: ",
3474 node, info_type(node), node_loc);
3477 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
3479 for (bqe = ((StgBlockingQueue*)node)->blocking_queue, end = (bqe==END_BQ_QUEUE);
3480 !end; // iterate until bqe points to a CONSTR
3481 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) {
3482 ASSERT(bqe != END_BQ_QUEUE); // sanity check
3483 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
3484 /* types of closures that may appear in a blocking queue */
3485 ASSERT(get_itbl(bqe)->type == TSO ||
3486 get_itbl(bqe)->type == CONSTR);
3487 /* only BQs of an RBH end with an RBH_Save closure */
3488 ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
3490 tso_loc = where_is((StgClosure *)bqe);
3491 switch (get_itbl(bqe)->type) {
3493 fprintf(stderr," TSO %d (%p) on [PE %d],",
3494 ((StgTSO *)bqe)->id, (StgTSO *)bqe, tso_loc);
3497 fprintf(stderr," %s (IP %p),",
3498 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
3499 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
3500 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
3501 "RBH_Save_?"), get_itbl(bqe));
3504 barf("Unexpected closure type %s in blocking queue of %p (%s)",
3505 info_type((StgClosure *)bqe), node, info_type(node));
3509 fputc('\n', stderr);
3513 Nice and easy: only TSOs on the blocking queue
3516 print_bq (StgClosure *node)
3520 ASSERT(node!=(StgClosure*)NULL); // sanity check
3521 for (tso = ((StgBlockingQueue*)node)->blocking_queue;
3522 tso != END_TSO_QUEUE;
3524 ASSERT(tso!=NULL && tso!=END_TSO_QUEUE); // sanity check
3525 ASSERT(get_itbl(tso)->type == TSO); // guess what, sanity check
3526 fprintf(stderr," TSO %d (%p),", tso->id, tso);
3528 fputc('\n', stderr);
3539 for (i=0, tso=run_queue_hd;
3540 tso != END_TSO_QUEUE;
3549 sched_belch(char *s, ...)
3553 #ifdef RTS_SUPPORTS_THREADS
3554 fprintf(stderr, "sched (task %p): ", osThreadId());
3556 fprintf(stderr, "== ");
3558 fprintf(stderr, "sched: ");
3560 vfprintf(stderr, s, ap);
3561 fprintf(stderr, "\n");