1 /* ---------------------------------------------------------------------------
2 * $Id: Schedule.c,v 1.132 2002/02/18 17:27:24 sof Exp $
4 * (c) The GHC Team, 1998-2000
8 * Different GHC ways use this scheduler quite differently (see comments below)
9 * Here is the global picture:
11 * WAY Name CPP flag What's it for
12 * --------------------------------------
13 * mp GUM PAR Parallel execution on a distributed memory machine
14 * s SMP SMP Parallel execution on a shared memory machine
15 * mg GranSim GRAN Simulation of parallel execution
16 * md GUM/GdH DIST Distributed execution (based on GUM)
18 * --------------------------------------------------------------------------*/
20 //@node Main scheduling code, , ,
21 //@section Main scheduling code
24 * Version with scheduler monitor support for SMPs (WAY=s):
26 This design provides a high-level API to create and schedule threads etc.
27 as documented in the SMP design document.
29 It uses a monitor design controlled by a single mutex to exercise control
30 over accesses to shared data structures, and builds on the Posix threads
33 The majority of state is shared. In order to keep essential per-task state,
34 there is a Capability structure, which contains all the information
35 needed to run a thread: its STG registers, a pointer to its TSO, a
36 nursery etc. During STG execution, a pointer to the capability is
37 kept in a register (BaseReg).
39 In a non-SMP build, there is one global capability, namely MainRegTable.
43 * Version with support for distributed memory parallelism aka GUM (WAY=mp):
45 The main scheduling loop in GUM iterates until a finish message is received.
46 In that case a global flag @receivedFinish@ is set and this instance of
47 the RTS shuts down. See ghc/rts/parallel/HLComms.c:processMessages()
48 for the handling of incoming messages, such as PP_FINISH.
49 Note that in the parallel case we have a system manager that coordinates
50 different PEs, each of which are running one instance of the RTS.
51 See ghc/rts/parallel/SysMan.c for the main routine of the parallel program.
52 From this routine processes executing ghc/rts/Main.c are spawned. -- HWL
54 * Version with support for simulating parallel execution aka GranSim (WAY=mg):
56 The main scheduling code in GranSim is quite different from that in std
57 (concurrent) Haskell: while concurrent Haskell just iterates over the
58 threads in the runnable queue, GranSim is event driven, i.e. it iterates
59 over the events in the global event queue. -- HWL
64 //* Variables and Data structures::
65 //* Main scheduling loop::
66 //* Suspend and Resume::
68 //* Garbage Collextion Routines::
69 //* Blocking Queue Routines::
70 //* Exception Handling Routines::
71 //* Debugging Routines::
75 //@node Includes, Variables and Data structures, Main scheduling code, Main scheduling code
76 //@subsection Includes
78 #include "PosixSource.h"
85 #include "StgStartup.h"
88 #include "StgMiscClosures.h"
90 #include "Interpreter.h"
91 #include "Exception.h"
100 #include "Proftimer.h"
101 #include "ProfHeap.h"
103 #if defined(GRAN) || defined(PAR)
104 # include "GranSimRts.h"
105 # include "GranSim.h"
106 # include "ParallelRts.h"
107 # include "Parallel.h"
108 # include "ParallelDebug.h"
109 # include "FetchMe.h"
113 #include "Capability.h"
114 #include "OSThreads.h"
119 //@node Variables and Data structures, Prototypes, Includes, Main scheduling code
120 //@subsection Variables and Data structures
124 * These are the threads which clients have requested that we run.
126 * In a 'threaded' build, we might have several concurrent clients all
127 * waiting for results, and each one will wait on a condition variable
128 * until the result is available.
130 * In non-SMP, clients are strictly nested: the first client calls
131 * into the RTS, which might call out again to C with a _ccall_GC, and
132 * eventually re-enter the RTS.
134 * Main threads information is kept in a linked list:
136 //@cindex StgMainThread
137 typedef struct StgMainThread_ {
139 SchedulerStatus stat;
141 #if defined(RTS_SUPPORTS_THREADS)
144 struct StgMainThread_ *link;
147 /* Main thread queue.
148 * Locks required: sched_mutex.
150 static StgMainThread *main_threads;
153 * Locks required: sched_mutex.
157 StgTSO* ActiveTSO = NULL; /* for assigning system costs; GranSim-Light only */
158 /* rtsTime TimeOfNextEvent, EndOfTimeSlice; now in GranSim.c */
161 In GranSim we have a runnable and a blocked queue for each processor.
162 In order to minimise code changes new arrays run_queue_hds/tls
163 are created. run_queue_hd is then a short cut (macro) for
164 run_queue_hds[CurrentProc] (see GranSim.h).
167 StgTSO *run_queue_hds[MAX_PROC], *run_queue_tls[MAX_PROC];
168 StgTSO *blocked_queue_hds[MAX_PROC], *blocked_queue_tls[MAX_PROC];
169 StgTSO *ccalling_threadss[MAX_PROC];
170 /* We use the same global list of threads (all_threads) in GranSim as in
171 the std RTS (i.e. we are cheating). However, we don't use this list in
172 the GranSim specific code at the moment (so we are only potentially
177 StgTSO *run_queue_hd, *run_queue_tl;
178 StgTSO *blocked_queue_hd, *blocked_queue_tl;
179 StgTSO *sleeping_queue; /* perhaps replace with a hash table? */
183 /* Linked list of all threads.
184 * Used for detecting garbage collected threads.
188 /* When a thread performs a safe C call (_ccall_GC, using old
189 * terminology), it gets put on the suspended_ccalling_threads
190 * list. Used by the garbage collector.
192 static StgTSO *suspended_ccalling_threads;
194 static StgTSO *threadStackOverflow(StgTSO *tso);
196 /* KH: The following two flags are shared memory locations. There is no need
197 to lock them, since they are only unset at the end of a scheduler
201 /* flag set by signal handler to precipitate a context switch */
202 //@cindex context_switch
205 /* if this flag is set as well, give up execution */
206 //@cindex interrupted
209 /* Next thread ID to allocate.
210 * Locks required: sched_mutex
212 //@cindex next_thread_id
213 StgThreadID next_thread_id = 1;
216 * Pointers to the state of the current thread.
217 * Rule of thumb: if CurrentTSO != NULL, then we're running a Haskell
218 * thread. If CurrentTSO == NULL, then we're at the scheduler level.
221 /* The smallest stack size that makes any sense is:
222 * RESERVED_STACK_WORDS (so we can get back from the stack overflow)
223 * + sizeofW(StgStopFrame) (the stg_stop_thread_info frame)
224 * + 1 (the realworld token for an IO thread)
225 * + 1 (the closure to enter)
227 * A thread with this stack will bomb immediately with a stack
228 * overflow, which will increase its stack size.
231 #define MIN_STACK_WORDS (RESERVED_STACK_WORDS + sizeofW(StgStopFrame) + 2)
238 /* This is used in `TSO.h' and gcc 2.96 insists that this variable actually
239 * exists - earlier gccs apparently didn't.
246 void addToBlockedQueue ( StgTSO *tso );
248 static void schedule ( void );
249 void interruptStgRts ( void );
251 static StgTSO * createThread_ ( nat size, rtsBool have_lock, StgInt pri );
253 static StgTSO * createThread_ ( nat size, rtsBool have_lock );
256 static void detectBlackHoles ( void );
259 static void sched_belch(char *s, ...);
262 #if defined(RTS_SUPPORTS_THREADS)
263 /* ToDo: carefully document the invariants that go together
264 * with these synchronisation objects.
266 Mutex sched_mutex = INIT_MUTEX_VAR;
267 Mutex term_mutex = INIT_MUTEX_VAR;
270 static Condition gc_pending_cond = INIT_COND_VAR;
274 #endif /* RTS_SUPPORTS_THREADS */
278 rtsTime TimeOfLastYield;
279 rtsBool emitSchedule = rtsTrue;
283 char *whatNext_strs[] = {
291 char *threadReturnCode_strs[] = {
292 "HeapOverflow", /* might also be StackOverflow */
301 StgTSO * createSparkThread(rtsSpark spark);
302 StgTSO * activateSpark (rtsSpark spark);
306 * The thread state for the main thread.
307 // ToDo: check whether not needed any more
311 #if defined(PAR) || defined(RTS_SUPPORTS_THREADS)
312 static void taskStart(void);
323 //@node Main scheduling loop, Suspend and Resume, Prototypes, Main scheduling code
324 //@subsection Main scheduling loop
326 /* ---------------------------------------------------------------------------
327 Main scheduling loop.
329 We use round-robin scheduling, each thread returning to the
330 scheduler loop when one of these conditions is detected:
333 * timer expires (thread yields)
338 Locking notes: we acquire the scheduler lock once at the beginning
339 of the scheduler loop, and release it when
341 * running a thread, or
342 * waiting for work, or
343 * waiting for a GC to complete.
346 In a GranSim setup this loop iterates over the global event queue.
347 This revolves around the global event queue, which determines what
348 to do next. Therefore, it's more complicated than either the
349 concurrent or the parallel (GUM) setup.
352 GUM iterates over incoming messages.
353 It starts with nothing to do (thus CurrentTSO == END_TSO_QUEUE),
354 and sends out a fish whenever it has nothing to do; in-between
355 doing the actual reductions (shared code below) it processes the
356 incoming messages and deals with delayed operations
357 (see PendingFetches).
358 This is not the ugliest code you could imagine, but it's bloody close.
360 ------------------------------------------------------------------------ */
367 StgThreadReturnCode ret;
375 rtsBool receivedFinish = rtsFalse;
377 nat tp_size, sp_size; // stats only
380 rtsBool was_interrupted = rtsFalse;
382 ACQUIRE_LOCK(&sched_mutex);
384 #if defined(RTS_SUPPORTS_THREADS)
385 /* Check to see whether there are any worker threads
386 waiting to deposit external call results. If so,
387 yield our capability */
388 yieldToReturningWorker(&sched_mutex, cap);
390 waitForWorkCapability(&sched_mutex, &cap, rtsFalse);
394 /* set up first event to get things going */
395 /* ToDo: assign costs for system setup and init MainTSO ! */
396 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
398 CurrentTSO, (StgClosure*)NULL, (rtsSpark*)NULL);
401 fprintf(stderr, "GRAN: Init CurrentTSO (in schedule) = %p\n", CurrentTSO);
402 G_TSO(CurrentTSO, 5));
404 if (RtsFlags.GranFlags.Light) {
405 /* Save current time; GranSim Light only */
406 CurrentTSO->gran.clock = CurrentTime[CurrentProc];
409 event = get_next_event();
411 while (event!=(rtsEvent*)NULL) {
412 /* Choose the processor with the next event */
413 CurrentProc = event->proc;
414 CurrentTSO = event->tso;
418 while (!receivedFinish) { /* set by processMessages */
419 /* when receiving PP_FINISH message */
426 IF_DEBUG(scheduler, printAllThreads());
428 /* If we're interrupted (the user pressed ^C, or some other
429 * termination condition occurred), kill all the currently running
433 IF_DEBUG(scheduler, sched_belch("interrupted"));
435 interrupted = rtsFalse;
436 was_interrupted = rtsTrue;
439 /* Go through the list of main threads and wake up any
440 * clients whose computations have finished. ToDo: this
441 * should be done more efficiently without a linear scan
442 * of the main threads list, somehow...
444 #if defined(RTS_SUPPORTS_THREADS)
446 StgMainThread *m, **prev;
447 prev = &main_threads;
448 for (m = main_threads; m != NULL; m = m->link) {
449 switch (m->tso->what_next) {
452 *(m->ret) = (StgClosure *)m->tso->sp[0];
456 broadcastCondition(&m->wakeup);
459 if (m->ret) *(m->ret) = NULL;
461 if (was_interrupted) {
462 m->stat = Interrupted;
466 broadcastCondition(&m->wakeup);
474 #else /* not threaded */
477 /* in GUM do this only on the Main PE */
480 /* If our main thread has finished or been killed, return.
483 StgMainThread *m = main_threads;
484 if (m->tso->what_next == ThreadComplete
485 || m->tso->what_next == ThreadKilled) {
486 main_threads = main_threads->link;
487 if (m->tso->what_next == ThreadComplete) {
488 /* we finished successfully, fill in the return value */
489 if (m->ret) { *(m->ret) = (StgClosure *)m->tso->sp[0]; };
493 if (m->ret) { *(m->ret) = NULL; };
494 if (was_interrupted) {
495 m->stat = Interrupted;
505 /* Top up the run queue from our spark pool. We try to make the
506 * number of threads in the run queue equal to the number of
509 * Disable spark support in SMP for now, non-essential & requires
510 * a little bit of work to make it compile cleanly. -- sof 1/02.
512 #if 0 /* defined(SMP) */
514 nat n = getFreeCapabilities();
515 StgTSO *tso = run_queue_hd;
517 /* Count the run queue */
518 while (n > 0 && tso != END_TSO_QUEUE) {
525 spark = findSpark(rtsFalse);
527 break; /* no more sparks in the pool */
529 /* I'd prefer this to be done in activateSpark -- HWL */
530 /* tricky - it needs to hold the scheduler lock and
531 * not try to re-acquire it -- SDM */
532 createSparkThread(spark);
534 sched_belch("==^^ turning spark of closure %p into a thread",
535 (StgClosure *)spark));
538 /* We need to wake up the other tasks if we just created some
541 if (getFreeCapabilities() - n > 1) {
542 signalCondition( &thread_ready_cond );
547 /* check for signals each time around the scheduler */
548 #ifndef mingw32_TARGET_OS
549 if (signals_pending()) {
550 startSignalHandlers();
554 /* Check whether any waiting threads need to be woken up. If the
555 * run queue is empty, and there are no other tasks running, we
556 * can wait indefinitely for something to happen.
557 * ToDo: what if another client comes along & requests another
560 if ( !EMPTY_QUEUE(blocked_queue_hd) || !EMPTY_QUEUE(sleeping_queue) ) {
561 awaitEvent( EMPTY_RUN_QUEUE()
563 && allFreeCapabilities()
567 /* we can be interrupted while waiting for I/O... */
568 if (interrupted) continue;
571 * Detect deadlock: when we have no threads to run, there are no
572 * threads waiting on I/O or sleeping, and all the other tasks are
573 * waiting for work, we must have a deadlock of some description.
575 * We first try to find threads blocked on themselves (ie. black
576 * holes), and generate NonTermination exceptions where necessary.
578 * If no threads are black holed, we have a deadlock situation, so
579 * inform all the main threads.
582 if ( EMPTY_RUN_QUEUE()
583 && EMPTY_QUEUE(blocked_queue_hd)
584 && EMPTY_QUEUE(sleeping_queue)
585 #if defined(RTS_SUPPORTS_THREADS)
586 && EMPTY_QUEUE(suspended_ccalling_threads)
589 && allFreeCapabilities()
593 IF_DEBUG(scheduler, sched_belch("deadlocked, forcing major GC..."));
594 #if defined(THREADED_RTS)
595 /* and SMP mode ..? */
596 releaseCapability(cap);
598 GarbageCollect(GetRoots,rtsTrue);
599 if ( EMPTY_QUEUE(blocked_queue_hd)
601 && EMPTY_QUEUE(sleeping_queue) ) {
603 IF_DEBUG(scheduler, sched_belch("still deadlocked, checking for black holes..."));
606 /* No black holes, so probably a real deadlock. Send the
607 * current main thread the Deadlock exception (or in the SMP
608 * build, send *all* main threads the deadlock exception,
609 * since none of them can make progress).
611 if ( EMPTY_RUN_QUEUE() ) {
613 #if defined(RTS_SUPPORTS_THREADS)
614 for (m = main_threads; m != NULL; m = m->link) {
615 switch (m->tso->why_blocked) {
616 case BlockedOnBlackHole:
617 raiseAsync(m->tso, (StgClosure *)NonTermination_closure);
619 case BlockedOnException:
621 raiseAsync(m->tso, (StgClosure *)Deadlock_closure);
624 barf("deadlock: main thread blocked in a strange way");
629 switch (m->tso->why_blocked) {
630 case BlockedOnBlackHole:
631 raiseAsync(m->tso, (StgClosure *)NonTermination_closure);
633 case BlockedOnException:
635 raiseAsync(m->tso, (StgClosure *)Deadlock_closure);
638 barf("deadlock: main thread blocked in a strange way");
642 #if defined(RTS_SUPPORTS_THREADS)
643 /* ToDo: revisit conditions (and mechanism) for shutting
644 down a multi-threaded world */
645 if ( EMPTY_RUN_QUEUE() ) {
646 IF_DEBUG(scheduler, sched_belch("all done, i think...shutting down."));
647 shutdownHaskellAndExit(0);
650 ASSERT( !EMPTY_RUN_QUEUE() );
654 /* ToDo: add deadlock detection in GUM (similar to SMP) -- HWL */
658 /* If there's a GC pending, don't do anything until it has
662 IF_DEBUG(scheduler,sched_belch("waiting for GC"));
663 waitCondition( &gc_pending_cond, &sched_mutex );
667 #if defined(RTS_SUPPORTS_THREADS)
668 /* block until we've got a thread on the run queue and a free
672 if ( EMPTY_RUN_QUEUE() ) {
673 /* Give up our capability */
674 releaseCapability(cap);
675 IF_DEBUG(scheduler, sched_belch("thread %d: waiting for work", osThreadId()));
676 waitForWorkCapability(&sched_mutex, &cap, rtsTrue);
677 IF_DEBUG(scheduler, sched_belch("thread %d: work now available", osThreadId()));
679 while ( EMPTY_RUN_QUEUE() ) {
680 waitForWorkCapability(&sched_mutex, &cap);
681 IF_DEBUG(scheduler, sched_belch("thread %d: work now available", osThreadId()));
688 if (RtsFlags.GranFlags.Light)
689 GranSimLight_enter_system(event, &ActiveTSO); // adjust ActiveTSO etc
691 /* adjust time based on time-stamp */
692 if (event->time > CurrentTime[CurrentProc] &&
693 event->evttype != ContinueThread)
694 CurrentTime[CurrentProc] = event->time;
696 /* Deal with the idle PEs (may issue FindWork or MoveSpark events) */
697 if (!RtsFlags.GranFlags.Light)
700 IF_DEBUG(gran, fprintf(stderr, "GRAN: switch by event-type\n"));
702 /* main event dispatcher in GranSim */
703 switch (event->evttype) {
704 /* Should just be continuing execution */
706 IF_DEBUG(gran, fprintf(stderr, "GRAN: doing ContinueThread\n"));
707 /* ToDo: check assertion
708 ASSERT(run_queue_hd != (StgTSO*)NULL &&
709 run_queue_hd != END_TSO_QUEUE);
711 /* Ignore ContinueThreads for fetching threads (if synchr comm) */
712 if (!RtsFlags.GranFlags.DoAsyncFetch &&
713 procStatus[CurrentProc]==Fetching) {
714 belch("ghuH: Spurious ContinueThread while Fetching ignored; TSO %d (%p) [PE %d]",
715 CurrentTSO->id, CurrentTSO, CurrentProc);
718 /* Ignore ContinueThreads for completed threads */
719 if (CurrentTSO->what_next == ThreadComplete) {
720 belch("ghuH: found a ContinueThread event for completed thread %d (%p) [PE %d] (ignoring ContinueThread)",
721 CurrentTSO->id, CurrentTSO, CurrentProc);
724 /* Ignore ContinueThreads for threads that are being migrated */
725 if (PROCS(CurrentTSO)==Nowhere) {
726 belch("ghuH: trying to run the migrating TSO %d (%p) [PE %d] (ignoring ContinueThread)",
727 CurrentTSO->id, CurrentTSO, CurrentProc);
730 /* The thread should be at the beginning of the run queue */
731 if (CurrentTSO!=run_queue_hds[CurrentProc]) {
732 belch("ghuH: TSO %d (%p) [PE %d] is not at the start of the run_queue when doing a ContinueThread",
733 CurrentTSO->id, CurrentTSO, CurrentProc);
734 break; // run the thread anyway
737 new_event(proc, proc, CurrentTime[proc],
739 (StgTSO*)NULL, (StgClosure*)NULL, (rtsSpark*)NULL);
741 */ /* Catches superfluous CONTINUEs -- should be unnecessary */
742 break; // now actually run the thread; DaH Qu'vam yImuHbej
745 do_the_fetchnode(event);
746 goto next_thread; /* handle next event in event queue */
749 do_the_globalblock(event);
750 goto next_thread; /* handle next event in event queue */
753 do_the_fetchreply(event);
754 goto next_thread; /* handle next event in event queue */
756 case UnblockThread: /* Move from the blocked queue to the tail of */
757 do_the_unblock(event);
758 goto next_thread; /* handle next event in event queue */
760 case ResumeThread: /* Move from the blocked queue to the tail of */
761 /* the runnable queue ( i.e. Qu' SImqa'lu') */
762 event->tso->gran.blocktime +=
763 CurrentTime[CurrentProc] - event->tso->gran.blockedat;
764 do_the_startthread(event);
765 goto next_thread; /* handle next event in event queue */
768 do_the_startthread(event);
769 goto next_thread; /* handle next event in event queue */
772 do_the_movethread(event);
773 goto next_thread; /* handle next event in event queue */
776 do_the_movespark(event);
777 goto next_thread; /* handle next event in event queue */
780 do_the_findwork(event);
781 goto next_thread; /* handle next event in event queue */
784 barf("Illegal event type %u\n", event->evttype);
787 /* This point was scheduler_loop in the old RTS */
789 IF_DEBUG(gran, belch("GRAN: after main switch"));
791 TimeOfLastEvent = CurrentTime[CurrentProc];
792 TimeOfNextEvent = get_time_of_next_event();
793 IgnoreEvents=(TimeOfNextEvent==0); // HWL HACK
794 // CurrentTSO = ThreadQueueHd;
796 IF_DEBUG(gran, belch("GRAN: time of next event is: %ld",
799 if (RtsFlags.GranFlags.Light)
800 GranSimLight_leave_system(event, &ActiveTSO);
802 EndOfTimeSlice = CurrentTime[CurrentProc]+RtsFlags.GranFlags.time_slice;
805 belch("GRAN: end of time-slice is %#lx", EndOfTimeSlice));
807 /* in a GranSim setup the TSO stays on the run queue */
809 /* Take a thread from the run queue. */
810 t = POP_RUN_QUEUE(); // take_off_run_queue(t);
813 fprintf(stderr, "GRAN: About to run current thread, which is\n");
816 context_switch = 0; // turned on via GranYield, checking events and time slice
819 DumpGranEvent(GR_SCHEDULE, t));
821 procStatus[CurrentProc] = Busy;
824 if (PendingFetches != END_BF_QUEUE) {
828 /* ToDo: phps merge with spark activation above */
829 /* check whether we have local work and send requests if we have none */
830 if (EMPTY_RUN_QUEUE()) { /* no runnable threads */
831 /* :-[ no local threads => look out for local sparks */
832 /* the spark pool for the current PE */
833 pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable
834 if (advisory_thread_count < RtsFlags.ParFlags.maxThreads &&
835 pool->hd < pool->tl) {
837 * ToDo: add GC code check that we really have enough heap afterwards!!
839 * If we're here (no runnable threads) and we have pending
840 * sparks, we must have a space problem. Get enough space
841 * to turn one of those pending sparks into a
845 spark = findSpark(rtsFalse); /* get a spark */
846 if (spark != (rtsSpark) NULL) {
847 tso = activateSpark(spark); /* turn the spark into a thread */
848 IF_PAR_DEBUG(schedule,
849 belch("==== schedule: Created TSO %d (%p); %d threads active",
850 tso->id, tso, advisory_thread_count));
852 if (tso==END_TSO_QUEUE) { /* failed to activate spark->back to loop */
853 belch("==^^ failed to activate spark");
855 } /* otherwise fall through & pick-up new tso */
857 IF_PAR_DEBUG(verbose,
858 belch("==^^ no local sparks (spark pool contains only NFs: %d)",
859 spark_queue_len(pool)));
864 /* If we still have no work we need to send a FISH to get a spark
867 if (EMPTY_RUN_QUEUE()) {
868 /* =8-[ no local sparks => look for work on other PEs */
870 * We really have absolutely no work. Send out a fish
871 * (there may be some out there already), and wait for
872 * something to arrive. We clearly can't run any threads
873 * until a SCHEDULE or RESUME arrives, and so that's what
874 * we're hoping to see. (Of course, we still have to
875 * respond to other types of messages.)
877 TIME now = msTime() /*CURRENT_TIME*/;
878 IF_PAR_DEBUG(verbose,
879 belch("-- now=%ld", now));
880 IF_PAR_DEBUG(verbose,
881 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
882 (last_fish_arrived_at!=0 &&
883 last_fish_arrived_at+RtsFlags.ParFlags.fishDelay > now)) {
884 belch("--$$ delaying FISH until %ld (last fish %ld, delay %ld, now %ld)",
885 last_fish_arrived_at+RtsFlags.ParFlags.fishDelay,
886 last_fish_arrived_at,
887 RtsFlags.ParFlags.fishDelay, now);
890 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
891 (last_fish_arrived_at==0 ||
892 (last_fish_arrived_at+RtsFlags.ParFlags.fishDelay <= now))) {
893 /* outstandingFishes is set in sendFish, processFish;
894 avoid flooding system with fishes via delay */
896 sendFish(pe, mytid, NEW_FISH_AGE, NEW_FISH_HISTORY,
899 // Global statistics: count no. of fishes
900 if (RtsFlags.ParFlags.ParStats.Global &&
901 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
902 globalParStats.tot_fish_mess++;
906 receivedFinish = processMessages();
909 } else if (PacketsWaiting()) { /* Look for incoming messages */
910 receivedFinish = processMessages();
913 /* Now we are sure that we have some work available */
914 ASSERT(run_queue_hd != END_TSO_QUEUE);
916 /* Take a thread from the run queue, if we have work */
917 t = POP_RUN_QUEUE(); // take_off_run_queue(END_TSO_QUEUE);
918 IF_DEBUG(sanity,checkTSO(t));
920 /* ToDo: write something to the log-file
921 if (RTSflags.ParFlags.granSimStats && !sameThread)
922 DumpGranEvent(GR_SCHEDULE, RunnableThreadsHd);
926 /* the spark pool for the current PE */
927 pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable
930 belch("--=^ %d threads, %d sparks on [%#x]",
931 run_queue_len(), spark_queue_len(pool), CURRENT_PROC));
934 if (0 && RtsFlags.ParFlags.ParStats.Full &&
935 t && LastTSO && t->id != LastTSO->id &&
936 LastTSO->why_blocked == NotBlocked &&
937 LastTSO->what_next != ThreadComplete) {
938 // if previously scheduled TSO not blocked we have to record the context switch
939 DumpVeryRawGranEvent(TimeOfLastYield, CURRENT_PROC, CURRENT_PROC,
940 GR_DESCHEDULE, LastTSO, (StgClosure *)NULL, 0, 0);
943 if (RtsFlags.ParFlags.ParStats.Full &&
944 (emitSchedule /* forced emit */ ||
945 (t && LastTSO && t->id != LastTSO->id))) {
947 we are running a different TSO, so write a schedule event to log file
948 NB: If we use fair scheduling we also have to write a deschedule
949 event for LastTSO; with unfair scheduling we know that the
950 previous tso has blocked whenever we switch to another tso, so
951 we don't need it in GUM for now
953 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
954 GR_SCHEDULE, t, (StgClosure *)NULL, 0, 0);
955 emitSchedule = rtsFalse;
959 #else /* !GRAN && !PAR */
961 /* grab a thread from the run queue */
962 ASSERT(run_queue_hd != END_TSO_QUEUE);
964 // Sanity check the thread we're about to run. This can be
965 // expensive if there is lots of thread switching going on...
966 IF_DEBUG(sanity,checkTSO(t));
969 grabCapability(&cap);
970 cap->r.rCurrentTSO = t;
972 /* context switches are now initiated by the timer signal, unless
973 * the user specified "context switch as often as possible", with
978 RtsFlags.ProfFlags.profileInterval == 0 ||
980 (RtsFlags.ConcFlags.ctxtSwitchTicks == 0
981 && (run_queue_hd != END_TSO_QUEUE
982 || blocked_queue_hd != END_TSO_QUEUE
983 || sleeping_queue != END_TSO_QUEUE)))
988 RELEASE_LOCK(&sched_mutex);
990 IF_DEBUG(scheduler, sched_belch("-->> Running TSO %ld (%p) %s ...",
991 t->id, t, whatNext_strs[t->what_next]));
994 startHeapProfTimer();
997 /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
998 /* Run the current thread
1000 switch (cap->r.rCurrentTSO->what_next) {
1002 case ThreadComplete:
1003 /* Thread already finished, return to scheduler. */
1004 ret = ThreadFinished;
1006 case ThreadEnterGHC:
1007 ret = StgRun((StgFunPtr) stg_enterStackTop, &cap->r);
1010 ret = StgRun((StgFunPtr) stg_returnToStackTop, &cap->r);
1012 case ThreadEnterInterp:
1013 ret = interpretBCO(cap);
1016 barf("schedule: invalid what_next field");
1018 /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
1020 /* Costs for the scheduler are assigned to CCS_SYSTEM */
1022 stopHeapProfTimer();
1026 ACQUIRE_LOCK(&sched_mutex);
1029 IF_DEBUG(scheduler,fprintf(stderr,"scheduler (task %ld): ", osThreadId()););
1030 #elif !defined(GRAN) && !defined(PAR)
1031 IF_DEBUG(scheduler,fprintf(stderr,"scheduler: "););
1033 t = cap->r.rCurrentTSO;
1036 /* HACK 675: if the last thread didn't yield, make sure to print a
1037 SCHEDULE event to the log file when StgRunning the next thread, even
1038 if it is the same one as before */
1040 TimeOfLastYield = CURRENT_TIME;
1046 IF_DEBUG(gran, DumpGranEvent(GR_DESCHEDULE, t));
1047 globalGranStats.tot_heapover++;
1049 globalParStats.tot_heapover++;
1052 // did the task ask for a large block?
1053 if (cap->r.rHpAlloc > BLOCK_SIZE_W) {
1054 // if so, get one and push it on the front of the nursery.
1058 blocks = (nat)BLOCK_ROUND_UP(cap->r.rHpAlloc * sizeof(W_)) / BLOCK_SIZE;
1060 IF_DEBUG(scheduler,belch("--<< thread %ld (%p; %s) stopped: requesting a large block (size %d)",
1062 whatNext_strs[t->what_next], blocks));
1064 // don't do this if it would push us over the
1065 // alloc_blocks_lim limit; we'll GC first.
1066 if (alloc_blocks + blocks < alloc_blocks_lim) {
1068 alloc_blocks += blocks;
1069 bd = allocGroup( blocks );
1071 // link the new group into the list
1072 bd->link = cap->r.rCurrentNursery;
1073 bd->u.back = cap->r.rCurrentNursery->u.back;
1074 if (cap->r.rCurrentNursery->u.back != NULL) {
1075 cap->r.rCurrentNursery->u.back->link = bd;
1077 ASSERT(g0s0->blocks == cap->r.rCurrentNursery &&
1078 g0s0->blocks == cap->r.rNursery);
1079 cap->r.rNursery = g0s0->blocks = bd;
1081 cap->r.rCurrentNursery->u.back = bd;
1083 // initialise it as a nursery block
1087 bd->free = bd->start;
1089 // don't forget to update the block count in g0s0.
1090 g0s0->n_blocks += blocks;
1091 ASSERT(countBlocks(g0s0->blocks) == g0s0->n_blocks);
1093 // now update the nursery to point to the new block
1094 cap->r.rCurrentNursery = bd;
1096 // we might be unlucky and have another thread get on the
1097 // run queue before us and steal the large block, but in that
1098 // case the thread will just end up requesting another large
1100 PUSH_ON_RUN_QUEUE(t);
1105 /* make all the running tasks block on a condition variable,
1106 * maybe set context_switch and wait till they all pile in,
1107 * then have them wait on a GC condition variable.
1109 IF_DEBUG(scheduler,belch("--<< thread %ld (%p; %s) stopped: HeapOverflow",
1110 t->id, t, whatNext_strs[t->what_next]));
1113 ASSERT(!is_on_queue(t,CurrentProc));
1115 /* Currently we emit a DESCHEDULE event before GC in GUM.
1116 ToDo: either add separate event to distinguish SYSTEM time from rest
1117 or just nuke this DESCHEDULE (and the following SCHEDULE) */
1118 if (0 && RtsFlags.ParFlags.ParStats.Full) {
1119 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
1120 GR_DESCHEDULE, t, (StgClosure *)NULL, 0, 0);
1121 emitSchedule = rtsTrue;
1125 ready_to_gc = rtsTrue;
1126 context_switch = 1; /* stop other threads ASAP */
1127 PUSH_ON_RUN_QUEUE(t);
1128 /* actual GC is done at the end of the while loop */
1134 DumpGranEvent(GR_DESCHEDULE, t));
1135 globalGranStats.tot_stackover++;
1138 // DumpGranEvent(GR_DESCHEDULE, t);
1139 globalParStats.tot_stackover++;
1141 IF_DEBUG(scheduler,belch("--<< thread %ld (%p; %s) stopped, StackOverflow",
1142 t->id, t, whatNext_strs[t->what_next]));
1143 /* just adjust the stack for this thread, then pop it back
1149 /* enlarge the stack */
1150 StgTSO *new_t = threadStackOverflow(t);
1152 /* This TSO has moved, so update any pointers to it from the
1153 * main thread stack. It better not be on any other queues...
1154 * (it shouldn't be).
1156 for (m = main_threads; m != NULL; m = m->link) {
1161 threadPaused(new_t);
1162 PUSH_ON_RUN_QUEUE(new_t);
1166 case ThreadYielding:
1169 DumpGranEvent(GR_DESCHEDULE, t));
1170 globalGranStats.tot_yields++;
1173 // DumpGranEvent(GR_DESCHEDULE, t);
1174 globalParStats.tot_yields++;
1176 /* put the thread back on the run queue. Then, if we're ready to
1177 * GC, check whether this is the last task to stop. If so, wake
1178 * up the GC thread. getThread will block during a GC until the
1182 if (t->what_next == ThreadEnterInterp) {
1183 /* ToDo: or maybe a timer expired when we were in Hugs?
1184 * or maybe someone hit ctrl-C
1186 belch("--<< thread %ld (%p; %s) stopped to switch to Hugs",
1187 t->id, t, whatNext_strs[t->what_next]);
1189 belch("--<< thread %ld (%p; %s) stopped, yielding",
1190 t->id, t, whatNext_strs[t->what_next]);
1197 //belch("&& Doing sanity check on yielding TSO %ld.", t->id);
1199 ASSERT(t->link == END_TSO_QUEUE);
1201 ASSERT(!is_on_queue(t,CurrentProc));
1204 //belch("&& Doing sanity check on all ThreadQueues (and their TSOs).");
1205 checkThreadQsSanity(rtsTrue));
1208 if (RtsFlags.ParFlags.doFairScheduling) {
1209 /* this does round-robin scheduling; good for concurrency */
1210 APPEND_TO_RUN_QUEUE(t);
1212 /* this does unfair scheduling; good for parallelism */
1213 PUSH_ON_RUN_QUEUE(t);
1216 /* this does round-robin scheduling; good for concurrency */
1217 APPEND_TO_RUN_QUEUE(t);
1220 /* add a ContinueThread event to actually process the thread */
1221 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1223 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1225 belch("GRAN: eventq and runnableq after adding yielded thread to queue again:");
1234 belch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: ",
1235 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)));
1236 if (t->block_info.closure!=(StgClosure*)NULL) print_bq(t->block_info.closure));
1238 // ??? needed; should emit block before
1240 DumpGranEvent(GR_DESCHEDULE, t));
1241 prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
1244 ASSERT(procStatus[CurrentProc]==Busy ||
1245 ((procStatus[CurrentProc]==Fetching) &&
1246 (t->block_info.closure!=(StgClosure*)NULL)));
1247 if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
1248 !(!RtsFlags.GranFlags.DoAsyncFetch &&
1249 procStatus[CurrentProc]==Fetching))
1250 procStatus[CurrentProc] = Idle;
1254 belch("--<< thread %ld (%p; %s) stopped, blocking on node %p with BQ: ",
1255 t->id, t, whatNext_strs[t->what_next], t->block_info.closure));
1258 if (t->block_info.closure!=(StgClosure*)NULL)
1259 print_bq(t->block_info.closure));
1261 /* Send a fetch (if BlockedOnGA) and dump event to log file */
1264 /* whatever we schedule next, we must log that schedule */
1265 emitSchedule = rtsTrue;
1268 /* don't need to do anything. Either the thread is blocked on
1269 * I/O, in which case we'll have called addToBlockedQueue
1270 * previously, or it's blocked on an MVar or Blackhole, in which
1271 * case it'll be on the relevant queue already.
1274 fprintf(stderr, "--<< thread %d (%p) stopped: ", t->id, t);
1275 printThreadBlockage(t);
1276 fprintf(stderr, "\n"));
1278 /* Only for dumping event to log file
1279 ToDo: do I need this in GranSim, too?
1286 case ThreadFinished:
1287 /* Need to check whether this was a main thread, and if so, signal
1288 * the task that started it with the return value. If we have no
1289 * more main threads, we probably need to stop all the tasks until
1292 /* We also end up here if the thread kills itself with an
1293 * uncaught exception, see Exception.hc.
1295 IF_DEBUG(scheduler,belch("--++ thread %d (%p) finished", t->id, t));
1297 endThread(t, CurrentProc); // clean-up the thread
1299 /* For now all are advisory -- HWL */
1300 //if(t->priority==AdvisoryPriority) ??
1301 advisory_thread_count--;
1304 if(t->dist.priority==RevalPriority)
1308 if (RtsFlags.ParFlags.ParStats.Full &&
1309 !RtsFlags.ParFlags.ParStats.Suppressed)
1310 DumpEndEvent(CURRENT_PROC, t, rtsFalse /* not mandatory */);
1315 barf("schedule: invalid thread return code %d", (int)ret);
1318 #if defined(RTS_SUPPORTS_THREADS)
1319 /* I don't understand what this re-grab is doing -- sof */
1320 grabCapability(&cap);
1324 if (RtsFlags.ProfFlags.profileInterval==0 || performHeapProfile) {
1325 GarbageCollect(GetRoots, rtsTrue);
1327 performHeapProfile = rtsFalse;
1328 ready_to_gc = rtsFalse; // we already GC'd
1334 && allFreeCapabilities()
1337 /* everybody back, start the GC.
1338 * Could do it in this thread, or signal a condition var
1339 * to do it in another thread. Either way, we need to
1340 * broadcast on gc_pending_cond afterward.
1342 #if defined(RTS_SUPPORTS_THREADS)
1343 IF_DEBUG(scheduler,sched_belch("doing GC"));
1345 GarbageCollect(GetRoots,rtsFalse);
1346 ready_to_gc = rtsFalse;
1348 broadcastCondition(&gc_pending_cond);
1351 /* add a ContinueThread event to continue execution of current thread */
1352 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1354 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1356 fprintf(stderr, "GRAN: eventq and runnableq after Garbage collection:\n");
1364 IF_GRAN_DEBUG(unused,
1365 print_eventq(EventHd));
1367 event = get_next_event();
1370 /* ToDo: wait for next message to arrive rather than busy wait */
1373 } /* end of while(1) */
1375 IF_PAR_DEBUG(verbose,
1376 belch("== Leaving schedule() after having received Finish"));
1379 /* ---------------------------------------------------------------------------
1380 * deleteAllThreads(): kill all the live threads.
1382 * This is used when we catch a user interrupt (^C), before performing
1383 * any necessary cleanups and running finalizers.
1384 * ------------------------------------------------------------------------- */
1386 void deleteAllThreads ( void )
1389 IF_DEBUG(scheduler,sched_belch("deleting all threads"));
1390 for (t = run_queue_hd; t != END_TSO_QUEUE; t = next) {
1394 for (t = blocked_queue_hd; t != END_TSO_QUEUE; t = next) {
1398 for (t = sleeping_queue; t != END_TSO_QUEUE; t = next) {
1402 run_queue_hd = run_queue_tl = END_TSO_QUEUE;
1403 blocked_queue_hd = blocked_queue_tl = END_TSO_QUEUE;
1404 sleeping_queue = END_TSO_QUEUE;
1407 /* startThread and insertThread are now in GranSim.c -- HWL */
1410 //@node Suspend and Resume, Run queue code, Main scheduling loop, Main scheduling code
1411 //@subsection Suspend and Resume
1413 /* ---------------------------------------------------------------------------
1414 * Suspending & resuming Haskell threads.
1416 * When making a "safe" call to C (aka _ccall_GC), the task gives back
1417 * its capability before calling the C function. This allows another
1418 * task to pick up the capability and carry on running Haskell
1419 * threads. It also means that if the C call blocks, it won't lock
1422 * The Haskell thread making the C call is put to sleep for the
1423 * duration of the call, on the susepended_ccalling_threads queue. We
1424 * give out a token to the task, which it can use to resume the thread
1425 * on return from the C function.
1426 * ------------------------------------------------------------------------- */
1429 suspendThread( StgRegTable *reg,
1431 #if !defined(RTS_SUPPORTS_THREADS)
1439 /* assume that *reg is a pointer to the StgRegTable part
1442 cap = (Capability *)((void *)reg - sizeof(StgFunTable));
1444 ACQUIRE_LOCK(&sched_mutex);
1447 sched_belch("thread %d did a _ccall_gc", cap->r.rCurrentTSO->id));
1449 threadPaused(cap->r.rCurrentTSO);
1450 cap->r.rCurrentTSO->link = suspended_ccalling_threads;
1451 suspended_ccalling_threads = cap->r.rCurrentTSO;
1453 #if defined(RTS_SUPPORTS_THREADS)
1454 cap->r.rCurrentTSO->why_blocked = BlockedOnCCall;
1457 /* Use the thread ID as the token; it should be unique */
1458 tok = cap->r.rCurrentTSO->id;
1460 /* Hand back capability */
1461 releaseCapability(cap);
1463 #if defined(RTS_SUPPORTS_THREADS)
1464 /* Preparing to leave the RTS, so ensure there's a native thread/task
1465 waiting to take over.
1467 ToDo: optimise this and only create a new task if there's a need
1468 for one (i.e., if there's only one Concurrent Haskell thread alive,
1469 there's no need to create a new task).
1471 IF_DEBUG(scheduler, sched_belch("worker thread (%d): leaving RTS", tok));
1473 startTask(taskStart);
1477 /* Other threads _might_ be available for execution; signal this */
1479 RELEASE_LOCK(&sched_mutex);
1484 resumeThread( StgInt tok,
1486 #if !defined(RTS_SUPPORTS_THREADS)
1491 StgTSO *tso, **prev;
1494 #if defined(RTS_SUPPORTS_THREADS)
1495 /* Wait for permission to re-enter the RTS with the result. */
1497 grabReturnCapability(&sched_mutex, &cap);
1499 grabCapability(&cap);
1502 grabCapability(&cap);
1505 /* Remove the thread off of the suspended list */
1506 prev = &suspended_ccalling_threads;
1507 for (tso = suspended_ccalling_threads;
1508 tso != END_TSO_QUEUE;
1509 prev = &tso->link, tso = tso->link) {
1510 if (tso->id == (StgThreadID)tok) {
1515 if (tso == END_TSO_QUEUE) {
1516 barf("resumeThread: thread not found");
1518 tso->link = END_TSO_QUEUE;
1519 /* Reset blocking status */
1520 tso->why_blocked = NotBlocked;
1522 RELEASE_LOCK(&sched_mutex);
1524 cap->r.rCurrentTSO = tso;
1529 /* ---------------------------------------------------------------------------
1531 * ------------------------------------------------------------------------ */
1532 static void unblockThread(StgTSO *tso);
1534 /* ---------------------------------------------------------------------------
1535 * Comparing Thread ids.
1537 * This is used from STG land in the implementation of the
1538 * instances of Eq/Ord for ThreadIds.
1539 * ------------------------------------------------------------------------ */
1541 int cmp_thread(const StgTSO *tso1, const StgTSO *tso2)
1543 StgThreadID id1 = tso1->id;
1544 StgThreadID id2 = tso2->id;
1546 if (id1 < id2) return (-1);
1547 if (id1 > id2) return 1;
1551 /* ---------------------------------------------------------------------------
1552 * Fetching the ThreadID from an StgTSO.
1554 * This is used in the implementation of Show for ThreadIds.
1555 * ------------------------------------------------------------------------ */
1556 int rts_getThreadId(const StgTSO *tso)
1561 /* ---------------------------------------------------------------------------
1562 Create a new thread.
1564 The new thread starts with the given stack size. Before the
1565 scheduler can run, however, this thread needs to have a closure
1566 (and possibly some arguments) pushed on its stack. See
1567 pushClosure() in Schedule.h.
1569 createGenThread() and createIOThread() (in SchedAPI.h) are
1570 convenient packaged versions of this function.
1572 currently pri (priority) is only used in a GRAN setup -- HWL
1573 ------------------------------------------------------------------------ */
1574 //@cindex createThread
1576 /* currently pri (priority) is only used in a GRAN setup -- HWL */
1578 createThread(nat stack_size, StgInt pri)
1580 return createThread_(stack_size, rtsFalse, pri);
1584 createThread_(nat size, rtsBool have_lock, StgInt pri)
1588 createThread(nat stack_size)
1590 return createThread_(stack_size, rtsFalse);
1594 createThread_(nat size, rtsBool have_lock)
1601 /* First check whether we should create a thread at all */
1603 /* check that no more than RtsFlags.ParFlags.maxThreads threads are created */
1604 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads) {
1606 belch("{createThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
1607 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
1608 return END_TSO_QUEUE;
1614 ASSERT(!RtsFlags.GranFlags.Light || CurrentProc==0);
1617 // ToDo: check whether size = stack_size - TSO_STRUCT_SIZEW
1619 /* catch ridiculously small stack sizes */
1620 if (size < MIN_STACK_WORDS + TSO_STRUCT_SIZEW) {
1621 size = MIN_STACK_WORDS + TSO_STRUCT_SIZEW;
1624 stack_size = size - TSO_STRUCT_SIZEW;
1626 tso = (StgTSO *)allocate(size);
1627 TICK_ALLOC_TSO(stack_size, 0);
1629 SET_HDR(tso, &stg_TSO_info, CCS_SYSTEM);
1631 SET_GRAN_HDR(tso, ThisPE);
1633 tso->what_next = ThreadEnterGHC;
1635 /* tso->id needs to be unique. For now we use a heavyweight mutex to
1636 * protect the increment operation on next_thread_id.
1637 * In future, we could use an atomic increment instead.
1639 if (!have_lock) { ACQUIRE_LOCK(&sched_mutex); }
1640 tso->id = next_thread_id++;
1641 if (!have_lock) { RELEASE_LOCK(&sched_mutex); }
1643 tso->why_blocked = NotBlocked;
1644 tso->blocked_exceptions = NULL;
1646 tso->stack_size = stack_size;
1647 tso->max_stack_size = round_to_mblocks(RtsFlags.GcFlags.maxStkSize)
1649 tso->sp = (P_)&(tso->stack) + stack_size;
1652 tso->prof.CCCS = CCS_MAIN;
1655 /* put a stop frame on the stack */
1656 tso->sp -= sizeofW(StgStopFrame);
1657 SET_HDR((StgClosure*)tso->sp,(StgInfoTable *)&stg_stop_thread_info,CCS_SYSTEM);
1658 tso->su = (StgUpdateFrame*)tso->sp;
1662 tso->link = END_TSO_QUEUE;
1663 /* uses more flexible routine in GranSim */
1664 insertThread(tso, CurrentProc);
1666 /* In a non-GranSim setup the pushing of a TSO onto the runq is separated
1672 if (RtsFlags.GranFlags.GranSimStats.Full)
1673 DumpGranEvent(GR_START,tso);
1675 if (RtsFlags.ParFlags.ParStats.Full)
1676 DumpGranEvent(GR_STARTQ,tso);
1677 /* HACk to avoid SCHEDULE
1681 /* Link the new thread on the global thread list.
1683 tso->global_link = all_threads;
1687 tso->dist.priority = MandatoryPriority; //by default that is...
1691 tso->gran.pri = pri;
1693 tso->gran.magic = TSO_MAGIC; // debugging only
1695 tso->gran.sparkname = 0;
1696 tso->gran.startedat = CURRENT_TIME;
1697 tso->gran.exported = 0;
1698 tso->gran.basicblocks = 0;
1699 tso->gran.allocs = 0;
1700 tso->gran.exectime = 0;
1701 tso->gran.fetchtime = 0;
1702 tso->gran.fetchcount = 0;
1703 tso->gran.blocktime = 0;
1704 tso->gran.blockcount = 0;
1705 tso->gran.blockedat = 0;
1706 tso->gran.globalsparks = 0;
1707 tso->gran.localsparks = 0;
1708 if (RtsFlags.GranFlags.Light)
1709 tso->gran.clock = Now; /* local clock */
1711 tso->gran.clock = 0;
1713 IF_DEBUG(gran,printTSO(tso));
1716 tso->par.magic = TSO_MAGIC; // debugging only
1718 tso->par.sparkname = 0;
1719 tso->par.startedat = CURRENT_TIME;
1720 tso->par.exported = 0;
1721 tso->par.basicblocks = 0;
1722 tso->par.allocs = 0;
1723 tso->par.exectime = 0;
1724 tso->par.fetchtime = 0;
1725 tso->par.fetchcount = 0;
1726 tso->par.blocktime = 0;
1727 tso->par.blockcount = 0;
1728 tso->par.blockedat = 0;
1729 tso->par.globalsparks = 0;
1730 tso->par.localsparks = 0;
1734 globalGranStats.tot_threads_created++;
1735 globalGranStats.threads_created_on_PE[CurrentProc]++;
1736 globalGranStats.tot_sq_len += spark_queue_len(CurrentProc);
1737 globalGranStats.tot_sq_probes++;
1739 // collect parallel global statistics (currently done together with GC stats)
1740 if (RtsFlags.ParFlags.ParStats.Global &&
1741 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
1742 //fprintf(stderr, "Creating thread %d @ %11.2f\n", tso->id, usertime());
1743 globalParStats.tot_threads_created++;
1749 belch("==__ schedule: Created TSO %d (%p);",
1750 CurrentProc, tso, tso->id));
1752 IF_PAR_DEBUG(verbose,
1753 belch("==__ schedule: Created TSO %d (%p); %d threads active",
1754 tso->id, tso, advisory_thread_count));
1756 IF_DEBUG(scheduler,sched_belch("created thread %ld, stack size = %lx words",
1757 tso->id, tso->stack_size));
1764 all parallel thread creation calls should fall through the following routine.
1767 createSparkThread(rtsSpark spark)
1769 ASSERT(spark != (rtsSpark)NULL);
1770 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads)
1772 barf("{createSparkThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
1773 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
1774 return END_TSO_QUEUE;
1778 tso = createThread_(RtsFlags.GcFlags.initialStkSize, rtsTrue);
1779 if (tso==END_TSO_QUEUE)
1780 barf("createSparkThread: Cannot create TSO");
1782 tso->priority = AdvisoryPriority;
1784 pushClosure(tso,spark);
1785 PUSH_ON_RUN_QUEUE(tso);
1786 advisory_thread_count++;
1793 Turn a spark into a thread.
1794 ToDo: fix for SMP (needs to acquire SCHED_MUTEX!)
1797 //@cindex activateSpark
1799 activateSpark (rtsSpark spark)
1803 tso = createSparkThread(spark);
1804 if (RtsFlags.ParFlags.ParStats.Full) {
1805 //ASSERT(run_queue_hd == END_TSO_QUEUE); // I think ...
1806 IF_PAR_DEBUG(verbose,
1807 belch("==^^ activateSpark: turning spark of closure %p (%s) into a thread",
1808 (StgClosure *)spark, info_type((StgClosure *)spark)));
1810 // ToDo: fwd info on local/global spark to thread -- HWL
1811 // tso->gran.exported = spark->exported;
1812 // tso->gran.locked = !spark->global;
1813 // tso->gran.sparkname = spark->name;
1819 /* ---------------------------------------------------------------------------
1822 * scheduleThread puts a thread on the head of the runnable queue.
1823 * This will usually be done immediately after a thread is created.
1824 * The caller of scheduleThread must create the thread using e.g.
1825 * createThread and push an appropriate closure
1826 * on this thread's stack before the scheduler is invoked.
1827 * ------------------------------------------------------------------------ */
1829 static void scheduleThread_ (StgTSO* tso, rtsBool createTask);
1832 scheduleThread_(StgTSO *tso
1833 , rtsBool createTask
1834 #if !defined(THREADED_RTS)
1839 ACQUIRE_LOCK(&sched_mutex);
1841 /* Put the new thread on the head of the runnable queue. The caller
1842 * better push an appropriate closure on this thread's stack
1843 * beforehand. In the SMP case, the thread may start running as
1844 * soon as we release the scheduler lock below.
1846 PUSH_ON_RUN_QUEUE(tso);
1847 #if defined(THREADED_RTS)
1848 /* If main() is scheduling a thread, don't bother creating a
1852 startTask(taskStart);
1858 IF_DEBUG(scheduler,printTSO(tso));
1860 RELEASE_LOCK(&sched_mutex);
1863 void scheduleThread(StgTSO* tso)
1865 return scheduleThread_(tso, rtsFalse);
1868 void scheduleExtThread(StgTSO* tso)
1870 return scheduleThread_(tso, rtsTrue);
1873 /* ---------------------------------------------------------------------------
1876 * Initialise the scheduler. This resets all the queues - if the
1877 * queues contained any threads, they'll be garbage collected at the
1880 * ------------------------------------------------------------------------ */
1884 term_handler(int sig STG_UNUSED)
1887 ACQUIRE_LOCK(&term_mutex);
1889 RELEASE_LOCK(&term_mutex);
1900 for (i=0; i<=MAX_PROC; i++) {
1901 run_queue_hds[i] = END_TSO_QUEUE;
1902 run_queue_tls[i] = END_TSO_QUEUE;
1903 blocked_queue_hds[i] = END_TSO_QUEUE;
1904 blocked_queue_tls[i] = END_TSO_QUEUE;
1905 ccalling_threadss[i] = END_TSO_QUEUE;
1906 sleeping_queue = END_TSO_QUEUE;
1909 run_queue_hd = END_TSO_QUEUE;
1910 run_queue_tl = END_TSO_QUEUE;
1911 blocked_queue_hd = END_TSO_QUEUE;
1912 blocked_queue_tl = END_TSO_QUEUE;
1913 sleeping_queue = END_TSO_QUEUE;
1916 suspended_ccalling_threads = END_TSO_QUEUE;
1918 main_threads = NULL;
1919 all_threads = END_TSO_QUEUE;
1924 RtsFlags.ConcFlags.ctxtSwitchTicks =
1925 RtsFlags.ConcFlags.ctxtSwitchTime / TICK_MILLISECS;
1927 #if defined(RTS_SUPPORTS_THREADS)
1928 /* Initialise the mutex and condition variables used by
1930 initMutex(&sched_mutex);
1931 initMutex(&term_mutex);
1933 initCondition(&thread_ready_cond);
1937 initCondition(&gc_pending_cond);
1940 #if defined(RTS_SUPPORTS_THREADS)
1941 ACQUIRE_LOCK(&sched_mutex);
1944 /* Install the SIGHUP handler */
1947 struct sigaction action,oact;
1949 action.sa_handler = term_handler;
1950 sigemptyset(&action.sa_mask);
1951 action.sa_flags = 0;
1952 if (sigaction(SIGTERM, &action, &oact) != 0) {
1953 barf("can't install TERM handler");
1958 /* A capability holds the state a native thread needs in
1959 * order to execute STG code. At least one capability is
1960 * floating around (only SMP builds have more than one).
1964 #if defined(RTS_SUPPORTS_THREADS)
1965 /* start our haskell execution tasks */
1967 startTaskManager(RtsFlags.ParFlags.nNodes, taskStart);
1969 startTaskManager(0,taskStart);
1973 #if /* defined(SMP) ||*/ defined(PAR)
1977 #if defined(RTS_SUPPORTS_THREADS)
1978 RELEASE_LOCK(&sched_mutex);
1984 exitScheduler( void )
1986 #if defined(RTS_SUPPORTS_THREADS)
1991 /* -----------------------------------------------------------------------------
1992 Managing the per-task allocation areas.
1994 Each capability comes with an allocation area. These are
1995 fixed-length block lists into which allocation can be done.
1997 ToDo: no support for two-space collection at the moment???
1998 -------------------------------------------------------------------------- */
2000 /* -----------------------------------------------------------------------------
2001 * waitThread is the external interface for running a new computation
2002 * and waiting for the result.
2004 * In the non-SMP case, we create a new main thread, push it on the
2005 * main-thread stack, and invoke the scheduler to run it. The
2006 * scheduler will return when the top main thread on the stack has
2007 * completed or died, and fill in the necessary fields of the
2008 * main_thread structure.
2010 * In the SMP case, we create a main thread as before, but we then
2011 * create a new condition variable and sleep on it. When our new
2012 * main thread has completed, we'll be woken up and the status/result
2013 * will be in the main_thread struct.
2014 * -------------------------------------------------------------------------- */
2017 howManyThreadsAvail ( void )
2021 for (q = run_queue_hd; q != END_TSO_QUEUE; q = q->link)
2023 for (q = blocked_queue_hd; q != END_TSO_QUEUE; q = q->link)
2025 for (q = sleeping_queue; q != END_TSO_QUEUE; q = q->link)
2031 finishAllThreads ( void )
2034 while (run_queue_hd != END_TSO_QUEUE) {
2035 waitThread ( run_queue_hd, NULL);
2037 while (blocked_queue_hd != END_TSO_QUEUE) {
2038 waitThread ( blocked_queue_hd, NULL);
2040 while (sleeping_queue != END_TSO_QUEUE) {
2041 waitThread ( blocked_queue_hd, NULL);
2044 (blocked_queue_hd != END_TSO_QUEUE ||
2045 run_queue_hd != END_TSO_QUEUE ||
2046 sleeping_queue != END_TSO_QUEUE);
2050 waitThread(StgTSO *tso, /*out*/StgClosure **ret)
2052 #if defined(THREADED_RTS)
2053 return waitThread_(tso,ret, rtsFalse);
2055 return waitThread_(tso,ret);
2060 waitThread_(StgTSO *tso,
2061 /*out*/StgClosure **ret
2062 #if defined(THREADED_RTS)
2063 , rtsBool blockWaiting
2068 SchedulerStatus stat;
2070 ACQUIRE_LOCK(&sched_mutex);
2072 m = stgMallocBytes(sizeof(StgMainThread), "waitThread");
2077 #if defined(RTS_SUPPORTS_THREADS)
2078 initCondition(&m->wakeup);
2081 m->link = main_threads;
2084 IF_DEBUG(scheduler, sched_belch("== scheduler: new main thread (%d)\n", m->tso->id));
2086 #if defined(RTS_SUPPORTS_THREADS)
2088 # if defined(THREADED_RTS)
2089 if (!blockWaiting) {
2090 /* In the threaded case, the OS thread that called main()
2091 * gets to enter the RTS directly without going via another
2094 RELEASE_LOCK(&sched_mutex);
2096 ASSERT(m->stat != NoStatus);
2100 IF_DEBUG(scheduler, sched_belch("sfoo"));
2102 waitCondition(&m->wakeup, &sched_mutex);
2103 } while (m->stat == NoStatus);
2106 /* GranSim specific init */
2107 CurrentTSO = m->tso; // the TSO to run
2108 procStatus[MainProc] = Busy; // status of main PE
2109 CurrentProc = MainProc; // PE to run it on
2113 RELEASE_LOCK(&sched_mutex);
2115 ASSERT(m->stat != NoStatus);
2120 #if defined(RTS_SUPPORTS_THREADS)
2121 closeCondition(&m->wakeup);
2124 IF_DEBUG(scheduler, fprintf(stderr, "== scheduler: main thread (%d) finished\n",
2128 #if defined(THREADED_RTS)
2131 RELEASE_LOCK(&sched_mutex);
2136 //@node Run queue code, Garbage Collextion Routines, Suspend and Resume, Main scheduling code
2137 //@subsection Run queue code
2141 NB: In GranSim we have many run queues; run_queue_hd is actually a macro
2142 unfolding to run_queue_hds[CurrentProc], thus CurrentProc is an
2143 implicit global variable that has to be correct when calling these
2147 /* Put the new thread on the head of the runnable queue.
2148 * The caller of createThread better push an appropriate closure
2149 * on this thread's stack before the scheduler is invoked.
2151 static /* inline */ void
2152 add_to_run_queue(tso)
2155 ASSERT(tso!=run_queue_hd && tso!=run_queue_tl);
2156 tso->link = run_queue_hd;
2158 if (run_queue_tl == END_TSO_QUEUE) {
2163 /* Put the new thread at the end of the runnable queue. */
2164 static /* inline */ void
2165 push_on_run_queue(tso)
2168 ASSERT(get_itbl((StgClosure *)tso)->type == TSO);
2169 ASSERT(run_queue_hd!=NULL && run_queue_tl!=NULL);
2170 ASSERT(tso!=run_queue_hd && tso!=run_queue_tl);
2171 if (run_queue_hd == END_TSO_QUEUE) {
2174 run_queue_tl->link = tso;
2180 Should be inlined because it's used very often in schedule. The tso
2181 argument is actually only needed in GranSim, where we want to have the
2182 possibility to schedule *any* TSO on the run queue, irrespective of the
2183 actual ordering. Therefore, if tso is not the nil TSO then we traverse
2184 the run queue and dequeue the tso, adjusting the links in the queue.
2186 //@cindex take_off_run_queue
2187 static /* inline */ StgTSO*
2188 take_off_run_queue(StgTSO *tso) {
2192 qetlaHbogh Qu' ngaSbogh ghomDaQ {tso} yIteq!
2194 if tso is specified, unlink that tso from the run_queue (doesn't have
2195 to be at the beginning of the queue); GranSim only
2197 if (tso!=END_TSO_QUEUE) {
2198 /* find tso in queue */
2199 for (t=run_queue_hd, prev=END_TSO_QUEUE;
2200 t!=END_TSO_QUEUE && t!=tso;
2204 /* now actually dequeue the tso */
2205 if (prev!=END_TSO_QUEUE) {
2206 ASSERT(run_queue_hd!=t);
2207 prev->link = t->link;
2209 /* t is at beginning of thread queue */
2210 ASSERT(run_queue_hd==t);
2211 run_queue_hd = t->link;
2213 /* t is at end of thread queue */
2214 if (t->link==END_TSO_QUEUE) {
2215 ASSERT(t==run_queue_tl);
2216 run_queue_tl = prev;
2218 ASSERT(run_queue_tl!=t);
2220 t->link = END_TSO_QUEUE;
2222 /* take tso from the beginning of the queue; std concurrent code */
2224 if (t != END_TSO_QUEUE) {
2225 run_queue_hd = t->link;
2226 t->link = END_TSO_QUEUE;
2227 if (run_queue_hd == END_TSO_QUEUE) {
2228 run_queue_tl = END_TSO_QUEUE;
2237 //@node Garbage Collextion Routines, Blocking Queue Routines, Run queue code, Main scheduling code
2238 //@subsection Garbage Collextion Routines
2240 /* ---------------------------------------------------------------------------
2241 Where are the roots that we know about?
2243 - all the threads on the runnable queue
2244 - all the threads on the blocked queue
2245 - all the threads on the sleeping queue
2246 - all the thread currently executing a _ccall_GC
2247 - all the "main threads"
2249 ------------------------------------------------------------------------ */
2251 /* This has to be protected either by the scheduler monitor, or by the
2252 garbage collection monitor (probably the latter).
2257 GetRoots(evac_fn evac)
2264 for (i=0; i<=RtsFlags.GranFlags.proc; i++) {
2265 if ((run_queue_hds[i] != END_TSO_QUEUE) && ((run_queue_hds[i] != NULL)))
2266 evac((StgClosure **)&run_queue_hds[i]);
2267 if ((run_queue_tls[i] != END_TSO_QUEUE) && ((run_queue_tls[i] != NULL)))
2268 evac((StgClosure **)&run_queue_tls[i]);
2270 if ((blocked_queue_hds[i] != END_TSO_QUEUE) && ((blocked_queue_hds[i] != NULL)))
2271 evac((StgClosure **)&blocked_queue_hds[i]);
2272 if ((blocked_queue_tls[i] != END_TSO_QUEUE) && ((blocked_queue_tls[i] != NULL)))
2273 evac((StgClosure **)&blocked_queue_tls[i]);
2274 if ((ccalling_threadss[i] != END_TSO_QUEUE) && ((ccalling_threadss[i] != NULL)))
2275 evac((StgClosure **)&ccalling_threads[i]);
2282 if (run_queue_hd != END_TSO_QUEUE) {
2283 ASSERT(run_queue_tl != END_TSO_QUEUE);
2284 evac((StgClosure **)&run_queue_hd);
2285 evac((StgClosure **)&run_queue_tl);
2288 if (blocked_queue_hd != END_TSO_QUEUE) {
2289 ASSERT(blocked_queue_tl != END_TSO_QUEUE);
2290 evac((StgClosure **)&blocked_queue_hd);
2291 evac((StgClosure **)&blocked_queue_tl);
2294 if (sleeping_queue != END_TSO_QUEUE) {
2295 evac((StgClosure **)&sleeping_queue);
2299 for (m = main_threads; m != NULL; m = m->link) {
2300 evac((StgClosure **)&m->tso);
2302 if (suspended_ccalling_threads != END_TSO_QUEUE) {
2303 evac((StgClosure **)&suspended_ccalling_threads);
2306 #if defined(PAR) || defined(GRAN)
2307 markSparkQueue(evac);
2311 /* -----------------------------------------------------------------------------
2314 This is the interface to the garbage collector from Haskell land.
2315 We provide this so that external C code can allocate and garbage
2316 collect when called from Haskell via _ccall_GC.
2318 It might be useful to provide an interface whereby the programmer
2319 can specify more roots (ToDo).
2321 This needs to be protected by the GC condition variable above. KH.
2322 -------------------------------------------------------------------------- */
2324 void (*extra_roots)(evac_fn);
2329 /* Obligated to hold this lock upon entry */
2330 ACQUIRE_LOCK(&sched_mutex);
2331 GarbageCollect(GetRoots,rtsFalse);
2332 RELEASE_LOCK(&sched_mutex);
2336 performMajorGC(void)
2338 ACQUIRE_LOCK(&sched_mutex);
2339 GarbageCollect(GetRoots,rtsTrue);
2340 RELEASE_LOCK(&sched_mutex);
2344 AllRoots(evac_fn evac)
2346 GetRoots(evac); // the scheduler's roots
2347 extra_roots(evac); // the user's roots
2351 performGCWithRoots(void (*get_roots)(evac_fn))
2353 ACQUIRE_LOCK(&sched_mutex);
2354 extra_roots = get_roots;
2355 GarbageCollect(AllRoots,rtsFalse);
2356 RELEASE_LOCK(&sched_mutex);
2359 /* -----------------------------------------------------------------------------
2362 If the thread has reached its maximum stack size, then raise the
2363 StackOverflow exception in the offending thread. Otherwise
2364 relocate the TSO into a larger chunk of memory and adjust its stack
2366 -------------------------------------------------------------------------- */
2369 threadStackOverflow(StgTSO *tso)
2371 nat new_stack_size, new_tso_size, diff, stack_words;
2375 IF_DEBUG(sanity,checkTSO(tso));
2376 if (tso->stack_size >= tso->max_stack_size) {
2379 belch("@@ threadStackOverflow of TSO %d (%p): stack too large (now %ld; max is %ld",
2380 tso->id, tso, tso->stack_size, tso->max_stack_size);
2381 /* If we're debugging, just print out the top of the stack */
2382 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2385 /* Send this thread the StackOverflow exception */
2386 raiseAsync(tso, (StgClosure *)stackOverflow_closure);
2390 /* Try to double the current stack size. If that takes us over the
2391 * maximum stack size for this thread, then use the maximum instead.
2392 * Finally round up so the TSO ends up as a whole number of blocks.
2394 new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
2395 new_tso_size = (nat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) +
2396 TSO_STRUCT_SIZE)/sizeof(W_);
2397 new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */
2398 new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
2400 IF_DEBUG(scheduler, fprintf(stderr,"== scheduler: increasing stack size from %d words to %d.\n", tso->stack_size, new_stack_size));
2402 dest = (StgTSO *)allocate(new_tso_size);
2403 TICK_ALLOC_TSO(new_stack_size,0);
2405 /* copy the TSO block and the old stack into the new area */
2406 memcpy(dest,tso,TSO_STRUCT_SIZE);
2407 stack_words = tso->stack + tso->stack_size - tso->sp;
2408 new_sp = (P_)dest + new_tso_size - stack_words;
2409 memcpy(new_sp, tso->sp, stack_words * sizeof(W_));
2411 /* relocate the stack pointers... */
2412 diff = (P_)new_sp - (P_)tso->sp; /* In *words* */
2413 dest->su = (StgUpdateFrame *) ((P_)dest->su + diff);
2415 dest->stack_size = new_stack_size;
2417 /* and relocate the update frame list */
2418 relocate_stack(dest, diff);
2420 /* Mark the old TSO as relocated. We have to check for relocated
2421 * TSOs in the garbage collector and any primops that deal with TSOs.
2423 * It's important to set the sp and su values to just beyond the end
2424 * of the stack, so we don't attempt to scavenge any part of the
2427 tso->what_next = ThreadRelocated;
2429 tso->sp = (P_)&(tso->stack[tso->stack_size]);
2430 tso->su = (StgUpdateFrame *)tso->sp;
2431 tso->why_blocked = NotBlocked;
2432 dest->mut_link = NULL;
2434 IF_PAR_DEBUG(verbose,
2435 belch("@@ threadStackOverflow of TSO %d (now at %p): stack size increased to %ld",
2436 tso->id, tso, tso->stack_size);
2437 /* If we're debugging, just print out the top of the stack */
2438 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2441 IF_DEBUG(sanity,checkTSO(tso));
2443 IF_DEBUG(scheduler,printTSO(dest));
2449 //@node Blocking Queue Routines, Exception Handling Routines, Garbage Collextion Routines, Main scheduling code
2450 //@subsection Blocking Queue Routines
2452 /* ---------------------------------------------------------------------------
2453 Wake up a queue that was blocked on some resource.
2454 ------------------------------------------------------------------------ */
2458 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2463 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2465 /* write RESUME events to log file and
2466 update blocked and fetch time (depending on type of the orig closure) */
2467 if (RtsFlags.ParFlags.ParStats.Full) {
2468 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
2469 GR_RESUMEQ, ((StgTSO *)bqe), ((StgTSO *)bqe)->block_info.closure,
2470 0, 0 /* spark_queue_len(ADVISORY_POOL) */);
2471 if (EMPTY_RUN_QUEUE())
2472 emitSchedule = rtsTrue;
2474 switch (get_itbl(node)->type) {
2476 ((StgTSO *)bqe)->par.fetchtime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2481 ((StgTSO *)bqe)->par.blocktime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2488 barf("{unblockOneLocked}Daq Qagh: unexpected closure in blocking queue");
2495 static StgBlockingQueueElement *
2496 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2499 PEs node_loc, tso_loc;
2501 node_loc = where_is(node); // should be lifted out of loop
2502 tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2503 tso_loc = where_is((StgClosure *)tso);
2504 if (IS_LOCAL_TO(PROCS(node),tso_loc)) { // TSO is local
2505 /* !fake_fetch => TSO is on CurrentProc is same as IS_LOCAL_TO */
2506 ASSERT(CurrentProc!=node_loc || tso_loc==CurrentProc);
2507 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.lunblocktime;
2508 // insertThread(tso, node_loc);
2509 new_event(tso_loc, tso_loc, CurrentTime[CurrentProc],
2511 tso, node, (rtsSpark*)NULL);
2512 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2515 } else { // TSO is remote (actually should be FMBQ)
2516 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.mpacktime +
2517 RtsFlags.GranFlags.Costs.gunblocktime +
2518 RtsFlags.GranFlags.Costs.latency;
2519 new_event(tso_loc, CurrentProc, CurrentTime[CurrentProc],
2521 tso, node, (rtsSpark*)NULL);
2522 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2525 /* the thread-queue-overhead is accounted for in either Resume or UnblockThread */
2527 fprintf(stderr," %s TSO %d (%p) [PE %d] (block_info.closure=%p) (next=%p) ,",
2528 (node_loc==tso_loc ? "Local" : "Global"),
2529 tso->id, tso, CurrentProc, tso->block_info.closure, tso->link));
2530 tso->block_info.closure = NULL;
2531 IF_DEBUG(scheduler,belch("-- Waking up thread %ld (%p)",
2535 static StgBlockingQueueElement *
2536 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2538 StgBlockingQueueElement *next;
2540 switch (get_itbl(bqe)->type) {
2542 ASSERT(((StgTSO *)bqe)->why_blocked != NotBlocked);
2543 /* if it's a TSO just push it onto the run_queue */
2545 // ((StgTSO *)bqe)->link = END_TSO_QUEUE; // debugging?
2546 PUSH_ON_RUN_QUEUE((StgTSO *)bqe);
2548 unblockCount(bqe, node);
2549 /* reset blocking status after dumping event */
2550 ((StgTSO *)bqe)->why_blocked = NotBlocked;
2554 /* if it's a BLOCKED_FETCH put it on the PendingFetches list */
2556 bqe->link = (StgBlockingQueueElement *)PendingFetches;
2557 PendingFetches = (StgBlockedFetch *)bqe;
2561 /* can ignore this case in a non-debugging setup;
2562 see comments on RBHSave closures above */
2564 /* check that the closure is an RBHSave closure */
2565 ASSERT(get_itbl((StgClosure *)bqe) == &stg_RBH_Save_0_info ||
2566 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_1_info ||
2567 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_2_info);
2571 barf("{unblockOneLocked}Daq Qagh: Unexpected IP (%#lx; %s) in blocking queue at %#lx\n",
2572 get_itbl((StgClosure *)bqe), info_type((StgClosure *)bqe),
2576 IF_PAR_DEBUG(bq, fprintf(stderr, ", %p (%s)", bqe, info_type((StgClosure*)bqe)));
2580 #else /* !GRAN && !PAR */
2582 unblockOneLocked(StgTSO *tso)
2586 ASSERT(get_itbl(tso)->type == TSO);
2587 ASSERT(tso->why_blocked != NotBlocked);
2588 tso->why_blocked = NotBlocked;
2590 PUSH_ON_RUN_QUEUE(tso);
2592 IF_DEBUG(scheduler,sched_belch("waking up thread %ld", tso->id));
2597 #if defined(GRAN) || defined(PAR)
2598 inline StgBlockingQueueElement *
2599 unblockOne(StgBlockingQueueElement *bqe, StgClosure *node)
2601 ACQUIRE_LOCK(&sched_mutex);
2602 bqe = unblockOneLocked(bqe, node);
2603 RELEASE_LOCK(&sched_mutex);
2608 unblockOne(StgTSO *tso)
2610 ACQUIRE_LOCK(&sched_mutex);
2611 tso = unblockOneLocked(tso);
2612 RELEASE_LOCK(&sched_mutex);
2619 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2621 StgBlockingQueueElement *bqe;
2626 belch("##-_ AwBQ for node %p on PE %d @ %ld by TSO %d (%p): ", \
2627 node, CurrentProc, CurrentTime[CurrentProc],
2628 CurrentTSO->id, CurrentTSO));
2630 node_loc = where_is(node);
2632 ASSERT(q == END_BQ_QUEUE ||
2633 get_itbl(q)->type == TSO || // q is either a TSO or an RBHSave
2634 get_itbl(q)->type == CONSTR); // closure (type constructor)
2635 ASSERT(is_unique(node));
2637 /* FAKE FETCH: magically copy the node to the tso's proc;
2638 no Fetch necessary because in reality the node should not have been
2639 moved to the other PE in the first place
2641 if (CurrentProc!=node_loc) {
2643 belch("## node %p is on PE %d but CurrentProc is %d (TSO %d); assuming fake fetch and adjusting bitmask (old: %#x)",
2644 node, node_loc, CurrentProc, CurrentTSO->id,
2645 // CurrentTSO, where_is(CurrentTSO),
2646 node->header.gran.procs));
2647 node->header.gran.procs = (node->header.gran.procs) | PE_NUMBER(CurrentProc);
2649 belch("## new bitmask of node %p is %#x",
2650 node, node->header.gran.procs));
2651 if (RtsFlags.GranFlags.GranSimStats.Global) {
2652 globalGranStats.tot_fake_fetches++;
2657 // ToDo: check: ASSERT(CurrentProc==node_loc);
2658 while (get_itbl(bqe)->type==TSO) { // q != END_TSO_QUEUE) {
2661 bqe points to the current element in the queue
2662 next points to the next element in the queue
2664 //tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2665 //tso_loc = where_is(tso);
2667 bqe = unblockOneLocked(bqe, node);
2670 /* if this is the BQ of an RBH, we have to put back the info ripped out of
2671 the closure to make room for the anchor of the BQ */
2672 if (bqe!=END_BQ_QUEUE) {
2673 ASSERT(get_itbl(node)->type == RBH && get_itbl(bqe)->type == CONSTR);
2675 ASSERT((info_ptr==&RBH_Save_0_info) ||
2676 (info_ptr==&RBH_Save_1_info) ||
2677 (info_ptr==&RBH_Save_2_info));
2679 /* cf. convertToRBH in RBH.c for writing the RBHSave closure */
2680 ((StgRBH *)node)->blocking_queue = (StgBlockingQueueElement *)((StgRBHSave *)bqe)->payload[0];
2681 ((StgRBH *)node)->mut_link = (StgMutClosure *)((StgRBHSave *)bqe)->payload[1];
2684 belch("## Filled in RBH_Save for %p (%s) at end of AwBQ",
2685 node, info_type(node)));
2688 /* statistics gathering */
2689 if (RtsFlags.GranFlags.GranSimStats.Global) {
2690 // globalGranStats.tot_bq_processing_time += bq_processing_time;
2691 globalGranStats.tot_bq_len += len; // total length of all bqs awakened
2692 // globalGranStats.tot_bq_len_local += len_local; // same for local TSOs only
2693 globalGranStats.tot_awbq++; // total no. of bqs awakened
2696 fprintf(stderr,"## BQ Stats of %p: [%d entries] %s\n",
2697 node, len, (bqe!=END_BQ_QUEUE) ? "RBH" : ""));
2701 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2703 StgBlockingQueueElement *bqe;
2705 ACQUIRE_LOCK(&sched_mutex);
2707 IF_PAR_DEBUG(verbose,
2708 belch("##-_ AwBQ for node %p on [%x]: ",
2712 if(get_itbl(q)->type == CONSTR || q==END_BQ_QUEUE) {
2713 IF_PAR_DEBUG(verbose, belch("## ... nothing to unblock so lets just return. RFP (BUG?)"));
2718 ASSERT(q == END_BQ_QUEUE ||
2719 get_itbl(q)->type == TSO ||
2720 get_itbl(q)->type == BLOCKED_FETCH ||
2721 get_itbl(q)->type == CONSTR);
2724 while (get_itbl(bqe)->type==TSO ||
2725 get_itbl(bqe)->type==BLOCKED_FETCH) {
2726 bqe = unblockOneLocked(bqe, node);
2728 RELEASE_LOCK(&sched_mutex);
2731 #else /* !GRAN && !PAR */
2733 awakenBlockedQueue(StgTSO *tso)
2735 ACQUIRE_LOCK(&sched_mutex);
2736 while (tso != END_TSO_QUEUE) {
2737 tso = unblockOneLocked(tso);
2739 RELEASE_LOCK(&sched_mutex);
2743 //@node Exception Handling Routines, Debugging Routines, Blocking Queue Routines, Main scheduling code
2744 //@subsection Exception Handling Routines
2746 /* ---------------------------------------------------------------------------
2748 - usually called inside a signal handler so it mustn't do anything fancy.
2749 ------------------------------------------------------------------------ */
2752 interruptStgRts(void)
2758 /* -----------------------------------------------------------------------------
2761 This is for use when we raise an exception in another thread, which
2763 This has nothing to do with the UnblockThread event in GranSim. -- HWL
2764 -------------------------------------------------------------------------- */
2766 #if defined(GRAN) || defined(PAR)
2768 NB: only the type of the blocking queue is different in GranSim and GUM
2769 the operations on the queue-elements are the same
2770 long live polymorphism!
2773 unblockThread(StgTSO *tso)
2775 StgBlockingQueueElement *t, **last;
2777 ACQUIRE_LOCK(&sched_mutex);
2778 switch (tso->why_blocked) {
2781 return; /* not blocked */
2784 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
2786 StgBlockingQueueElement *last_tso = END_BQ_QUEUE;
2787 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
2789 last = (StgBlockingQueueElement **)&mvar->head;
2790 for (t = (StgBlockingQueueElement *)mvar->head;
2792 last = &t->link, last_tso = t, t = t->link) {
2793 if (t == (StgBlockingQueueElement *)tso) {
2794 *last = (StgBlockingQueueElement *)tso->link;
2795 if (mvar->tail == tso) {
2796 mvar->tail = (StgTSO *)last_tso;
2801 barf("unblockThread (MVAR): TSO not found");
2804 case BlockedOnBlackHole:
2805 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
2807 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
2809 last = &bq->blocking_queue;
2810 for (t = bq->blocking_queue;
2812 last = &t->link, t = t->link) {
2813 if (t == (StgBlockingQueueElement *)tso) {
2814 *last = (StgBlockingQueueElement *)tso->link;
2818 barf("unblockThread (BLACKHOLE): TSO not found");
2821 case BlockedOnException:
2823 StgTSO *target = tso->block_info.tso;
2825 ASSERT(get_itbl(target)->type == TSO);
2827 if (target->what_next == ThreadRelocated) {
2828 target = target->link;
2829 ASSERT(get_itbl(target)->type == TSO);
2832 ASSERT(target->blocked_exceptions != NULL);
2834 last = (StgBlockingQueueElement **)&target->blocked_exceptions;
2835 for (t = (StgBlockingQueueElement *)target->blocked_exceptions;
2837 last = &t->link, t = t->link) {
2838 ASSERT(get_itbl(t)->type == TSO);
2839 if (t == (StgBlockingQueueElement *)tso) {
2840 *last = (StgBlockingQueueElement *)tso->link;
2844 barf("unblockThread (Exception): TSO not found");
2848 case BlockedOnWrite:
2850 /* take TSO off blocked_queue */
2851 StgBlockingQueueElement *prev = NULL;
2852 for (t = (StgBlockingQueueElement *)blocked_queue_hd; t != END_BQ_QUEUE;
2853 prev = t, t = t->link) {
2854 if (t == (StgBlockingQueueElement *)tso) {
2856 blocked_queue_hd = (StgTSO *)t->link;
2857 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
2858 blocked_queue_tl = END_TSO_QUEUE;
2861 prev->link = t->link;
2862 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
2863 blocked_queue_tl = (StgTSO *)prev;
2869 barf("unblockThread (I/O): TSO not found");
2872 case BlockedOnDelay:
2874 /* take TSO off sleeping_queue */
2875 StgBlockingQueueElement *prev = NULL;
2876 for (t = (StgBlockingQueueElement *)sleeping_queue; t != END_BQ_QUEUE;
2877 prev = t, t = t->link) {
2878 if (t == (StgBlockingQueueElement *)tso) {
2880 sleeping_queue = (StgTSO *)t->link;
2882 prev->link = t->link;
2887 barf("unblockThread (I/O): TSO not found");
2891 barf("unblockThread");
2895 tso->link = END_TSO_QUEUE;
2896 tso->why_blocked = NotBlocked;
2897 tso->block_info.closure = NULL;
2898 PUSH_ON_RUN_QUEUE(tso);
2899 RELEASE_LOCK(&sched_mutex);
2903 unblockThread(StgTSO *tso)
2907 ACQUIRE_LOCK(&sched_mutex);
2908 switch (tso->why_blocked) {
2911 return; /* not blocked */
2914 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
2916 StgTSO *last_tso = END_TSO_QUEUE;
2917 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
2920 for (t = mvar->head; t != END_TSO_QUEUE;
2921 last = &t->link, last_tso = t, t = t->link) {
2924 if (mvar->tail == tso) {
2925 mvar->tail = last_tso;
2930 barf("unblockThread (MVAR): TSO not found");
2933 case BlockedOnBlackHole:
2934 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
2936 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
2938 last = &bq->blocking_queue;
2939 for (t = bq->blocking_queue; t != END_TSO_QUEUE;
2940 last = &t->link, t = t->link) {
2946 barf("unblockThread (BLACKHOLE): TSO not found");
2949 case BlockedOnException:
2951 StgTSO *target = tso->block_info.tso;
2953 ASSERT(get_itbl(target)->type == TSO);
2955 while (target->what_next == ThreadRelocated) {
2956 target = target->link;
2957 ASSERT(get_itbl(target)->type == TSO);
2960 ASSERT(target->blocked_exceptions != NULL);
2962 last = &target->blocked_exceptions;
2963 for (t = target->blocked_exceptions; t != END_TSO_QUEUE;
2964 last = &t->link, t = t->link) {
2965 ASSERT(get_itbl(t)->type == TSO);
2971 barf("unblockThread (Exception): TSO not found");
2975 case BlockedOnWrite:
2977 StgTSO *prev = NULL;
2978 for (t = blocked_queue_hd; t != END_TSO_QUEUE;
2979 prev = t, t = t->link) {
2982 blocked_queue_hd = t->link;
2983 if (blocked_queue_tl == t) {
2984 blocked_queue_tl = END_TSO_QUEUE;
2987 prev->link = t->link;
2988 if (blocked_queue_tl == t) {
2989 blocked_queue_tl = prev;
2995 barf("unblockThread (I/O): TSO not found");
2998 case BlockedOnDelay:
3000 StgTSO *prev = NULL;
3001 for (t = sleeping_queue; t != END_TSO_QUEUE;
3002 prev = t, t = t->link) {
3005 sleeping_queue = t->link;
3007 prev->link = t->link;
3012 barf("unblockThread (I/O): TSO not found");
3016 barf("unblockThread");
3020 tso->link = END_TSO_QUEUE;
3021 tso->why_blocked = NotBlocked;
3022 tso->block_info.closure = NULL;
3023 PUSH_ON_RUN_QUEUE(tso);
3024 RELEASE_LOCK(&sched_mutex);
3028 /* -----------------------------------------------------------------------------
3031 * The following function implements the magic for raising an
3032 * asynchronous exception in an existing thread.
3034 * We first remove the thread from any queue on which it might be
3035 * blocked. The possible blockages are MVARs and BLACKHOLE_BQs.
3037 * We strip the stack down to the innermost CATCH_FRAME, building
3038 * thunks in the heap for all the active computations, so they can
3039 * be restarted if necessary. When we reach a CATCH_FRAME, we build
3040 * an application of the handler to the exception, and push it on
3041 * the top of the stack.
3043 * How exactly do we save all the active computations? We create an
3044 * AP_UPD for every UpdateFrame on the stack. Entering one of these
3045 * AP_UPDs pushes everything from the corresponding update frame
3046 * upwards onto the stack. (Actually, it pushes everything up to the
3047 * next update frame plus a pointer to the next AP_UPD object.
3048 * Entering the next AP_UPD object pushes more onto the stack until we
3049 * reach the last AP_UPD object - at which point the stack should look
3050 * exactly as it did when we killed the TSO and we can continue
3051 * execution by entering the closure on top of the stack.
3053 * We can also kill a thread entirely - this happens if either (a) the
3054 * exception passed to raiseAsync is NULL, or (b) there's no
3055 * CATCH_FRAME on the stack. In either case, we strip the entire
3056 * stack and replace the thread with a zombie.
3058 * -------------------------------------------------------------------------- */
3061 deleteThread(StgTSO *tso)
3063 raiseAsync(tso,NULL);
3067 raiseAsync(StgTSO *tso, StgClosure *exception)
3069 StgUpdateFrame* su = tso->su;
3070 StgPtr sp = tso->sp;
3072 /* Thread already dead? */
3073 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
3077 IF_DEBUG(scheduler, sched_belch("raising exception in thread %ld.", tso->id));
3079 /* Remove it from any blocking queues */
3082 /* The stack freezing code assumes there's a closure pointer on
3083 * the top of the stack. This isn't always the case with compiled
3084 * code, so we have to push a dummy closure on the top which just
3085 * returns to the next return address on the stack.
3087 if ( LOOKS_LIKE_GHC_INFO((void*)*sp) ) {
3088 *(--sp) = (W_)&stg_dummy_ret_closure;
3092 nat words = ((P_)su - (P_)sp) - 1;
3096 /* If we find a CATCH_FRAME, and we've got an exception to raise,
3097 * then build the THUNK raise(exception), and leave it on
3098 * top of the CATCH_FRAME ready to enter.
3100 if (get_itbl(su)->type == CATCH_FRAME && exception != NULL) {
3101 StgCatchFrame *cf = (StgCatchFrame *)su;
3104 /* we've got an exception to raise, so let's pass it to the
3105 * handler in this frame.
3107 raise = (StgClosure *)allocate(sizeofW(StgClosure)+1);
3108 TICK_ALLOC_SE_THK(1,0);
3109 SET_HDR(raise,&stg_raise_info,cf->header.prof.ccs);
3110 raise->payload[0] = exception;
3112 /* throw away the stack from Sp up to the CATCH_FRAME.
3116 /* Ensure that async excpetions are blocked now, so we don't get
3117 * a surprise exception before we get around to executing the
3120 if (tso->blocked_exceptions == NULL) {
3121 tso->blocked_exceptions = END_TSO_QUEUE;
3124 /* Put the newly-built THUNK on top of the stack, ready to execute
3125 * when the thread restarts.
3130 tso->what_next = ThreadEnterGHC;
3131 IF_DEBUG(sanity, checkTSO(tso));
3135 /* First build an AP_UPD consisting of the stack chunk above the
3136 * current update frame, with the top word on the stack as the
3139 ap = (StgAP_UPD *)allocate(AP_sizeW(words));
3144 ap->fun = (StgClosure *)sp[0];
3146 for(i=0; i < (nat)words; ++i) {
3147 ap->payload[i] = (StgClosure *)*sp++;
3150 switch (get_itbl(su)->type) {
3154 SET_HDR(ap,&stg_AP_UPD_info,su->header.prof.ccs /* ToDo */);
3155 TICK_ALLOC_UP_THK(words+1,0);
3158 fprintf(stderr, "scheduler: Updating ");
3159 printPtr((P_)su->updatee);
3160 fprintf(stderr, " with ");
3161 printObj((StgClosure *)ap);
3164 /* Replace the updatee with an indirection - happily
3165 * this will also wake up any threads currently
3166 * waiting on the result.
3168 * Warning: if we're in a loop, more than one update frame on
3169 * the stack may point to the same object. Be careful not to
3170 * overwrite an IND_OLDGEN in this case, because we'll screw
3171 * up the mutable lists. To be on the safe side, don't
3172 * overwrite any kind of indirection at all. See also
3173 * threadSqueezeStack in GC.c, where we have to make a similar
3176 if (!closure_IND(su->updatee)) {
3177 UPD_IND_NOLOCK(su->updatee,ap); /* revert the black hole */
3180 sp += sizeofW(StgUpdateFrame) -1;
3181 sp[0] = (W_)ap; /* push onto stack */
3187 StgCatchFrame *cf = (StgCatchFrame *)su;
3190 /* We want a PAP, not an AP_UPD. Fortunately, the
3191 * layout's the same.
3193 SET_HDR(ap,&stg_PAP_info,su->header.prof.ccs /* ToDo */);
3194 TICK_ALLOC_UPD_PAP(words+1,0);
3196 /* now build o = FUN(catch,ap,handler) */
3197 o = (StgClosure *)allocate(sizeofW(StgClosure)+2);
3198 TICK_ALLOC_FUN(2,0);
3199 SET_HDR(o,&stg_catch_info,su->header.prof.ccs /* ToDo */);
3200 o->payload[0] = (StgClosure *)ap;
3201 o->payload[1] = cf->handler;
3204 fprintf(stderr, "scheduler: Built ");
3205 printObj((StgClosure *)o);
3208 /* pop the old handler and put o on the stack */
3210 sp += sizeofW(StgCatchFrame) - 1;
3217 StgSeqFrame *sf = (StgSeqFrame *)su;
3220 SET_HDR(ap,&stg_PAP_info,su->header.prof.ccs /* ToDo */);
3221 TICK_ALLOC_UPD_PAP(words+1,0);
3223 /* now build o = FUN(seq,ap) */
3224 o = (StgClosure *)allocate(sizeofW(StgClosure)+1);
3225 TICK_ALLOC_SE_THK(1,0);
3226 SET_HDR(o,&stg_seq_info,su->header.prof.ccs /* ToDo */);
3227 o->payload[0] = (StgClosure *)ap;
3230 fprintf(stderr, "scheduler: Built ");
3231 printObj((StgClosure *)o);
3234 /* pop the old handler and put o on the stack */
3236 sp += sizeofW(StgSeqFrame) - 1;
3242 /* We've stripped the entire stack, the thread is now dead. */
3243 sp += sizeofW(StgStopFrame) - 1;
3244 sp[0] = (W_)exception; /* save the exception */
3245 tso->what_next = ThreadKilled;
3246 tso->su = (StgUpdateFrame *)(sp+1);
3257 /* -----------------------------------------------------------------------------
3258 resurrectThreads is called after garbage collection on the list of
3259 threads found to be garbage. Each of these threads will be woken
3260 up and sent a signal: BlockedOnDeadMVar if the thread was blocked
3261 on an MVar, or NonTermination if the thread was blocked on a Black
3263 -------------------------------------------------------------------------- */
3266 resurrectThreads( StgTSO *threads )
3270 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
3271 next = tso->global_link;
3272 tso->global_link = all_threads;
3274 IF_DEBUG(scheduler, sched_belch("resurrecting thread %d", tso->id));
3276 switch (tso->why_blocked) {
3278 case BlockedOnException:
3279 raiseAsync(tso,(StgClosure *)BlockedOnDeadMVar_closure);
3281 case BlockedOnBlackHole:
3282 raiseAsync(tso,(StgClosure *)NonTermination_closure);
3285 /* This might happen if the thread was blocked on a black hole
3286 * belonging to a thread that we've just woken up (raiseAsync
3287 * can wake up threads, remember...).
3291 barf("resurrectThreads: thread blocked in a strange way");
3296 /* -----------------------------------------------------------------------------
3297 * Blackhole detection: if we reach a deadlock, test whether any
3298 * threads are blocked on themselves. Any threads which are found to
3299 * be self-blocked get sent a NonTermination exception.
3301 * This is only done in a deadlock situation in order to avoid
3302 * performance overhead in the normal case.
3303 * -------------------------------------------------------------------------- */
3306 detectBlackHoles( void )
3308 StgTSO *t = all_threads;
3309 StgUpdateFrame *frame;
3310 StgClosure *blocked_on;
3312 for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
3314 while (t->what_next == ThreadRelocated) {
3316 ASSERT(get_itbl(t)->type == TSO);
3319 if (t->why_blocked != BlockedOnBlackHole) {
3323 blocked_on = t->block_info.closure;
3325 for (frame = t->su; ; frame = frame->link) {
3326 switch (get_itbl(frame)->type) {
3329 if (frame->updatee == blocked_on) {
3330 /* We are blocking on one of our own computations, so
3331 * send this thread the NonTermination exception.
3334 sched_belch("thread %d is blocked on itself", t->id));
3335 raiseAsync(t, (StgClosure *)NonTermination_closure);
3356 //@node Debugging Routines, Index, Exception Handling Routines, Main scheduling code
3357 //@subsection Debugging Routines
3359 /* -----------------------------------------------------------------------------
3360 Debugging: why is a thread blocked
3361 -------------------------------------------------------------------------- */
3366 printThreadBlockage(StgTSO *tso)
3368 switch (tso->why_blocked) {
3370 fprintf(stderr,"is blocked on read from fd %d", tso->block_info.fd);
3372 case BlockedOnWrite:
3373 fprintf(stderr,"is blocked on write to fd %d", tso->block_info.fd);
3375 case BlockedOnDelay:
3376 fprintf(stderr,"is blocked until %d", tso->block_info.target);
3379 fprintf(stderr,"is blocked on an MVar");
3381 case BlockedOnException:
3382 fprintf(stderr,"is blocked on delivering an exception to thread %d",
3383 tso->block_info.tso->id);
3385 case BlockedOnBlackHole:
3386 fprintf(stderr,"is blocked on a black hole");
3389 fprintf(stderr,"is not blocked");
3393 fprintf(stderr,"is blocked on global address; local FM_BQ is %p (%s)",
3394 tso->block_info.closure, info_type(tso->block_info.closure));
3396 case BlockedOnGA_NoSend:
3397 fprintf(stderr,"is blocked on global address (no send); local FM_BQ is %p (%s)",
3398 tso->block_info.closure, info_type(tso->block_info.closure));
3401 #if defined(RTS_SUPPORTS_THREADS)
3402 case BlockedOnCCall:
3403 fprintf(stderr,"is blocked on an external call");
3407 barf("printThreadBlockage: strange tso->why_blocked: %d for TSO %d (%d)",
3408 tso->why_blocked, tso->id, tso);
3413 printThreadStatus(StgTSO *tso)
3415 switch (tso->what_next) {
3417 fprintf(stderr,"has been killed");
3419 case ThreadComplete:
3420 fprintf(stderr,"has completed");
3423 printThreadBlockage(tso);
3428 printAllThreads(void)
3433 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
3434 ullong_format_string(TIME_ON_PROC(CurrentProc),
3435 time_string, rtsFalse/*no commas!*/);
3437 sched_belch("all threads at [%s]:", time_string);
3439 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
3440 ullong_format_string(CURRENT_TIME,
3441 time_string, rtsFalse/*no commas!*/);
3443 sched_belch("all threads at [%s]:", time_string);
3445 sched_belch("all threads:");
3448 for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
3449 fprintf(stderr, "\tthread %d ", t->id);
3450 printThreadStatus(t);
3451 fprintf(stderr,"\n");
3456 Print a whole blocking queue attached to node (debugging only).
3461 print_bq (StgClosure *node)
3463 StgBlockingQueueElement *bqe;
3467 fprintf(stderr,"## BQ of closure %p (%s): ",
3468 node, info_type(node));
3470 /* should cover all closures that may have a blocking queue */
3471 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
3472 get_itbl(node)->type == FETCH_ME_BQ ||
3473 get_itbl(node)->type == RBH ||
3474 get_itbl(node)->type == MVAR);
3476 ASSERT(node!=(StgClosure*)NULL); // sanity check
3478 print_bqe(((StgBlockingQueue*)node)->blocking_queue);
3482 Print a whole blocking queue starting with the element bqe.
3485 print_bqe (StgBlockingQueueElement *bqe)
3490 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
3492 for (end = (bqe==END_BQ_QUEUE);
3493 !end; // iterate until bqe points to a CONSTR
3494 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE),
3495 bqe = end ? END_BQ_QUEUE : bqe->link) {
3496 ASSERT(bqe != END_BQ_QUEUE); // sanity check
3497 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
3498 /* types of closures that may appear in a blocking queue */
3499 ASSERT(get_itbl(bqe)->type == TSO ||
3500 get_itbl(bqe)->type == BLOCKED_FETCH ||
3501 get_itbl(bqe)->type == CONSTR);
3502 /* only BQs of an RBH end with an RBH_Save closure */
3503 //ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
3505 switch (get_itbl(bqe)->type) {
3507 fprintf(stderr," TSO %u (%x),",
3508 ((StgTSO *)bqe)->id, ((StgTSO *)bqe));
3511 fprintf(stderr," BF (node=%p, ga=((%x, %d, %x)),",
3512 ((StgBlockedFetch *)bqe)->node,
3513 ((StgBlockedFetch *)bqe)->ga.payload.gc.gtid,
3514 ((StgBlockedFetch *)bqe)->ga.payload.gc.slot,
3515 ((StgBlockedFetch *)bqe)->ga.weight);
3518 fprintf(stderr," %s (IP %p),",
3519 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
3520 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
3521 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
3522 "RBH_Save_?"), get_itbl(bqe));
3525 barf("Unexpected closure type %s in blocking queue", // of %p (%s)",
3526 info_type((StgClosure *)bqe)); // , node, info_type(node));
3530 fputc('\n', stderr);
3532 # elif defined(GRAN)
3534 print_bq (StgClosure *node)
3536 StgBlockingQueueElement *bqe;
3537 PEs node_loc, tso_loc;
3540 /* should cover all closures that may have a blocking queue */
3541 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
3542 get_itbl(node)->type == FETCH_ME_BQ ||
3543 get_itbl(node)->type == RBH);
3545 ASSERT(node!=(StgClosure*)NULL); // sanity check
3546 node_loc = where_is(node);
3548 fprintf(stderr,"## BQ of closure %p (%s) on [PE %d]: ",
3549 node, info_type(node), node_loc);
3552 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
3554 for (bqe = ((StgBlockingQueue*)node)->blocking_queue, end = (bqe==END_BQ_QUEUE);
3555 !end; // iterate until bqe points to a CONSTR
3556 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) {
3557 ASSERT(bqe != END_BQ_QUEUE); // sanity check
3558 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
3559 /* types of closures that may appear in a blocking queue */
3560 ASSERT(get_itbl(bqe)->type == TSO ||
3561 get_itbl(bqe)->type == CONSTR);
3562 /* only BQs of an RBH end with an RBH_Save closure */
3563 ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
3565 tso_loc = where_is((StgClosure *)bqe);
3566 switch (get_itbl(bqe)->type) {
3568 fprintf(stderr," TSO %d (%p) on [PE %d],",
3569 ((StgTSO *)bqe)->id, (StgTSO *)bqe, tso_loc);
3572 fprintf(stderr," %s (IP %p),",
3573 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
3574 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
3575 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
3576 "RBH_Save_?"), get_itbl(bqe));
3579 barf("Unexpected closure type %s in blocking queue of %p (%s)",
3580 info_type((StgClosure *)bqe), node, info_type(node));
3584 fputc('\n', stderr);
3588 Nice and easy: only TSOs on the blocking queue
3591 print_bq (StgClosure *node)
3595 ASSERT(node!=(StgClosure*)NULL); // sanity check
3596 for (tso = ((StgBlockingQueue*)node)->blocking_queue;
3597 tso != END_TSO_QUEUE;
3599 ASSERT(tso!=NULL && tso!=END_TSO_QUEUE); // sanity check
3600 ASSERT(get_itbl(tso)->type == TSO); // guess what, sanity check
3601 fprintf(stderr," TSO %d (%p),", tso->id, tso);
3603 fputc('\n', stderr);
3614 for (i=0, tso=run_queue_hd;
3615 tso != END_TSO_QUEUE;
3624 sched_belch(char *s, ...)
3629 fprintf(stderr, "scheduler (task %ld): ", osThreadId());
3631 fprintf(stderr, "== ");
3633 fprintf(stderr, "scheduler: ");
3635 vfprintf(stderr, s, ap);
3636 fprintf(stderr, "\n");
3642 //@node Index, , Debugging Routines, Main scheduling code
3646 //* StgMainThread:: @cindex\s-+StgMainThread
3647 //* awaken_blocked_queue:: @cindex\s-+awaken_blocked_queue
3648 //* blocked_queue_hd:: @cindex\s-+blocked_queue_hd
3649 //* blocked_queue_tl:: @cindex\s-+blocked_queue_tl
3650 //* context_switch:: @cindex\s-+context_switch
3651 //* createThread:: @cindex\s-+createThread
3652 //* gc_pending_cond:: @cindex\s-+gc_pending_cond
3653 //* initScheduler:: @cindex\s-+initScheduler
3654 //* interrupted:: @cindex\s-+interrupted
3655 //* next_thread_id:: @cindex\s-+next_thread_id
3656 //* print_bq:: @cindex\s-+print_bq
3657 //* run_queue_hd:: @cindex\s-+run_queue_hd
3658 //* run_queue_tl:: @cindex\s-+run_queue_tl
3659 //* sched_mutex:: @cindex\s-+sched_mutex
3660 //* schedule:: @cindex\s-+schedule
3661 //* take_off_run_queue:: @cindex\s-+take_off_run_queue
3662 //* term_mutex:: @cindex\s-+term_mutex