1 /* ---------------------------------------------------------------------------
2 * $Id: Schedule.c,v 1.90 2001/02/11 17:51:08 simonmar Exp $
4 * (c) The GHC Team, 1998-2000
8 * The main scheduling code in GranSim is quite different from that in std
9 * (concurrent) Haskell: while concurrent Haskell just iterates over the
10 * threads in the runnable queue, GranSim is event driven, i.e. it iterates
11 * over the events in the global event queue. -- HWL
12 * --------------------------------------------------------------------------*/
14 //@node Main scheduling code, , ,
15 //@section Main scheduling code
17 /* Version with scheduler monitor support for SMPs.
19 This design provides a high-level API to create and schedule threads etc.
20 as documented in the SMP design document.
22 It uses a monitor design controlled by a single mutex to exercise control
23 over accesses to shared data structures, and builds on the Posix threads
26 The majority of state is shared. In order to keep essential per-task state,
27 there is a Capability structure, which contains all the information
28 needed to run a thread: its STG registers, a pointer to its TSO, a
29 nursery etc. During STG execution, a pointer to the capability is
30 kept in a register (BaseReg).
32 In a non-SMP build, there is one global capability, namely MainRegTable.
39 //* Variables and Data structures::
40 //* Main scheduling loop::
41 //* Suspend and Resume::
43 //* Garbage Collextion Routines::
44 //* Blocking Queue Routines::
45 //* Exception Handling Routines::
46 //* Debugging Routines::
50 //@node Includes, Variables and Data structures, Main scheduling code, Main scheduling code
51 //@subsection Includes
59 #include "StgStartup.h"
63 #include "StgMiscClosures.h"
65 #include "Interpreter.h"
66 #include "Exception.h"
74 #if defined(GRAN) || defined(PAR)
75 # include "GranSimRts.h"
77 # include "ParallelRts.h"
78 # include "Parallel.h"
79 # include "ParallelDebug.h"
87 //@node Variables and Data structures, Prototypes, Includes, Main scheduling code
88 //@subsection Variables and Data structures
92 * These are the threads which clients have requested that we run.
94 * In an SMP build, we might have several concurrent clients all
95 * waiting for results, and each one will wait on a condition variable
96 * until the result is available.
98 * In non-SMP, clients are strictly nested: the first client calls
99 * into the RTS, which might call out again to C with a _ccall_GC, and
100 * eventually re-enter the RTS.
102 * Main threads information is kept in a linked list:
104 //@cindex StgMainThread
105 typedef struct StgMainThread_ {
107 SchedulerStatus stat;
110 pthread_cond_t wakeup;
112 struct StgMainThread_ *link;
115 /* Main thread queue.
116 * Locks required: sched_mutex.
118 static StgMainThread *main_threads;
121 * Locks required: sched_mutex.
125 StgTSO* ActiveTSO = NULL; /* for assigning system costs; GranSim-Light only */
126 /* rtsTime TimeOfNextEvent, EndOfTimeSlice; now in GranSim.c */
129 In GranSim we have a runable and a blocked queue for each processor.
130 In order to minimise code changes new arrays run_queue_hds/tls
131 are created. run_queue_hd is then a short cut (macro) for
132 run_queue_hds[CurrentProc] (see GranSim.h).
135 StgTSO *run_queue_hds[MAX_PROC], *run_queue_tls[MAX_PROC];
136 StgTSO *blocked_queue_hds[MAX_PROC], *blocked_queue_tls[MAX_PROC];
137 StgTSO *ccalling_threadss[MAX_PROC];
138 /* We use the same global list of threads (all_threads) in GranSim as in
139 the std RTS (i.e. we are cheating). However, we don't use this list in
140 the GranSim specific code at the moment (so we are only potentially
145 StgTSO *run_queue_hd, *run_queue_tl;
146 StgTSO *blocked_queue_hd, *blocked_queue_tl;
147 StgTSO *sleeping_queue; /* perhaps replace with a hash table? */
151 /* Linked list of all threads.
152 * Used for detecting garbage collected threads.
156 /* Threads suspended in _ccall_GC.
158 static StgTSO *suspended_ccalling_threads;
160 static void GetRoots(void);
161 static StgTSO *threadStackOverflow(StgTSO *tso);
163 /* KH: The following two flags are shared memory locations. There is no need
164 to lock them, since they are only unset at the end of a scheduler
168 /* flag set by signal handler to precipitate a context switch */
169 //@cindex context_switch
172 /* if this flag is set as well, give up execution */
173 //@cindex interrupted
176 /* Next thread ID to allocate.
177 * Locks required: sched_mutex
179 //@cindex next_thread_id
180 StgThreadID next_thread_id = 1;
183 * Pointers to the state of the current thread.
184 * Rule of thumb: if CurrentTSO != NULL, then we're running a Haskell
185 * thread. If CurrentTSO == NULL, then we're at the scheduler level.
188 /* The smallest stack size that makes any sense is:
189 * RESERVED_STACK_WORDS (so we can get back from the stack overflow)
190 * + sizeofW(StgStopFrame) (the stg_stop_thread_info frame)
191 * + 1 (the realworld token for an IO thread)
192 * + 1 (the closure to enter)
194 * A thread with this stack will bomb immediately with a stack
195 * overflow, which will increase its stack size.
198 #define MIN_STACK_WORDS (RESERVED_STACK_WORDS + sizeofW(StgStopFrame) + 2)
200 /* Free capability list.
201 * Locks required: sched_mutex.
204 //@cindex free_capabilities
205 //@cindex n_free_capabilities
206 Capability *free_capabilities; /* Available capabilities for running threads */
207 nat n_free_capabilities; /* total number of available capabilities */
209 //@cindex MainRegTable
210 Capability MainRegTable; /* for non-SMP, we have one global capability */
217 /* This is used in `TSO.h' and gcc 2.96 insists that this variable actually
218 * exists - earlier gccs apparently didn't.
225 /* All our current task ids, saved in case we need to kill them later.
232 void addToBlockedQueue ( StgTSO *tso );
234 static void schedule ( void );
235 void interruptStgRts ( void );
237 static StgTSO * createThread_ ( nat size, rtsBool have_lock, StgInt pri );
239 static StgTSO * createThread_ ( nat size, rtsBool have_lock );
242 static void detectBlackHoles ( void );
245 static void sched_belch(char *s, ...);
249 //@cindex sched_mutex
251 //@cindex thread_ready_cond
252 //@cindex gc_pending_cond
253 pthread_mutex_t sched_mutex = PTHREAD_MUTEX_INITIALIZER;
254 pthread_mutex_t term_mutex = PTHREAD_MUTEX_INITIALIZER;
255 pthread_cond_t thread_ready_cond = PTHREAD_COND_INITIALIZER;
256 pthread_cond_t gc_pending_cond = PTHREAD_COND_INITIALIZER;
263 rtsTime TimeOfLastYield;
267 char *whatNext_strs[] = {
275 char *threadReturnCode_strs[] = {
276 "HeapOverflow", /* might also be StackOverflow */
285 * The thread state for the main thread.
286 // ToDo: check whether not needed any more
290 //@node Main scheduling loop, Suspend and Resume, Prototypes, Main scheduling code
291 //@subsection Main scheduling loop
293 /* ---------------------------------------------------------------------------
294 Main scheduling loop.
296 We use round-robin scheduling, each thread returning to the
297 scheduler loop when one of these conditions is detected:
300 * timer expires (thread yields)
305 Locking notes: we acquire the scheduler lock once at the beginning
306 of the scheduler loop, and release it when
308 * running a thread, or
309 * waiting for work, or
310 * waiting for a GC to complete.
313 In a GranSim setup this loop iterates over the global event queue.
314 This revolves around the global event queue, which determines what
315 to do next. Therefore, it's more complicated than either the
316 concurrent or the parallel (GUM) setup.
319 GUM iterates over incoming messages.
320 It starts with nothing to do (thus CurrentTSO == END_TSO_QUEUE),
321 and sends out a fish whenever it has nothing to do; in-between
322 doing the actual reductions (shared code below) it processes the
323 incoming messages and deals with delayed operations
324 (see PendingFetches).
325 This is not the ugliest code you could imagine, but it's bloody close.
327 ------------------------------------------------------------------------ */
334 StgThreadReturnCode ret;
343 rtsBool was_interrupted = rtsFalse;
345 ACQUIRE_LOCK(&sched_mutex);
349 /* set up first event to get things going */
350 /* ToDo: assign costs for system setup and init MainTSO ! */
351 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
353 CurrentTSO, (StgClosure*)NULL, (rtsSpark*)NULL);
356 fprintf(stderr, "GRAN: Init CurrentTSO (in schedule) = %p\n", CurrentTSO);
357 G_TSO(CurrentTSO, 5));
359 if (RtsFlags.GranFlags.Light) {
360 /* Save current time; GranSim Light only */
361 CurrentTSO->gran.clock = CurrentTime[CurrentProc];
364 event = get_next_event();
366 while (event!=(rtsEvent*)NULL) {
367 /* Choose the processor with the next event */
368 CurrentProc = event->proc;
369 CurrentTSO = event->tso;
373 while (!GlobalStopPending) { /* GlobalStopPending set in par_exit */
381 IF_DEBUG(scheduler, printAllThreads());
383 /* If we're interrupted (the user pressed ^C, or some other
384 * termination condition occurred), kill all the currently running
388 IF_DEBUG(scheduler, sched_belch("interrupted"));
390 interrupted = rtsFalse;
391 was_interrupted = rtsTrue;
394 /* Go through the list of main threads and wake up any
395 * clients whose computations have finished. ToDo: this
396 * should be done more efficiently without a linear scan
397 * of the main threads list, somehow...
401 StgMainThread *m, **prev;
402 prev = &main_threads;
403 for (m = main_threads; m != NULL; m = m->link) {
404 switch (m->tso->what_next) {
407 *(m->ret) = (StgClosure *)m->tso->sp[0];
411 pthread_cond_broadcast(&m->wakeup);
415 if (was_interrupted) {
416 m->stat = Interrupted;
420 pthread_cond_broadcast(&m->wakeup);
430 /* in GUM do this only on the Main PE */
433 /* If our main thread has finished or been killed, return.
436 StgMainThread *m = main_threads;
437 if (m->tso->what_next == ThreadComplete
438 || m->tso->what_next == ThreadKilled) {
439 main_threads = main_threads->link;
440 if (m->tso->what_next == ThreadComplete) {
441 /* we finished successfully, fill in the return value */
442 if (m->ret) { *(m->ret) = (StgClosure *)m->tso->sp[0]; };
446 if (was_interrupted) {
447 m->stat = Interrupted;
457 /* Top up the run queue from our spark pool. We try to make the
458 * number of threads in the run queue equal to the number of
463 nat n = n_free_capabilities;
464 StgTSO *tso = run_queue_hd;
466 /* Count the run queue */
467 while (n > 0 && tso != END_TSO_QUEUE) {
476 break; /* no more sparks in the pool */
478 /* I'd prefer this to be done in activateSpark -- HWL */
479 /* tricky - it needs to hold the scheduler lock and
480 * not try to re-acquire it -- SDM */
482 tso = createThread_(RtsFlags.GcFlags.initialStkSize, rtsTrue);
483 pushClosure(tso,spark);
484 PUSH_ON_RUN_QUEUE(tso);
486 advisory_thread_count++;
490 sched_belch("turning spark of closure %p into a thread",
491 (StgClosure *)spark));
494 /* We need to wake up the other tasks if we just created some
497 if (n_free_capabilities - n > 1) {
498 pthread_cond_signal(&thread_ready_cond);
503 /* Check whether any waiting threads need to be woken up. If the
504 * run queue is empty, and there are no other tasks running, we
505 * can wait indefinitely for something to happen.
506 * ToDo: what if another client comes along & requests another
509 if (blocked_queue_hd != END_TSO_QUEUE || sleeping_queue != END_TSO_QUEUE) {
511 (run_queue_hd == END_TSO_QUEUE)
513 && (n_free_capabilities == RtsFlags.ParFlags.nNodes)
517 /* we can be interrupted while waiting for I/O... */
518 if (interrupted) continue;
520 /* check for signals each time around the scheduler */
521 #ifndef mingw32_TARGET_OS
522 if (signals_pending()) {
523 start_signal_handlers();
528 * Detect deadlock: when we have no threads to run, there are no
529 * threads waiting on I/O or sleeping, and all the other tasks are
530 * waiting for work, we must have a deadlock of some description.
532 * We first try to find threads blocked on themselves (ie. black
533 * holes), and generate NonTermination exceptions where necessary.
535 * If no threads are black holed, we have a deadlock situation, so
536 * inform all the main threads.
539 if (blocked_queue_hd == END_TSO_QUEUE
540 && run_queue_hd == END_TSO_QUEUE
541 && sleeping_queue == END_TSO_QUEUE
542 && (n_free_capabilities == RtsFlags.ParFlags.nNodes))
544 IF_DEBUG(scheduler, sched_belch("deadlocked, checking for black holes..."));
546 if (run_queue_hd == END_TSO_QUEUE) {
548 for (m = main_threads; m != NULL; m = m->link) {
551 pthread_cond_broadcast(&m->wakeup);
557 if (blocked_queue_hd == END_TSO_QUEUE
558 && run_queue_hd == END_TSO_QUEUE
559 && sleeping_queue == END_TSO_QUEUE)
561 IF_DEBUG(scheduler, sched_belch("deadlocked, checking for black holes..."));
563 if (run_queue_hd == END_TSO_QUEUE) {
564 StgMainThread *m = main_threads;
567 main_threads = m->link;
574 /* If there's a GC pending, don't do anything until it has
578 IF_DEBUG(scheduler,sched_belch("waiting for GC"));
579 pthread_cond_wait(&gc_pending_cond, &sched_mutex);
582 /* block until we've got a thread on the run queue and a free
585 while (run_queue_hd == END_TSO_QUEUE || free_capabilities == NULL) {
586 IF_DEBUG(scheduler, sched_belch("waiting for work"));
587 pthread_cond_wait(&thread_ready_cond, &sched_mutex);
588 IF_DEBUG(scheduler, sched_belch("work now available"));
594 if (RtsFlags.GranFlags.Light)
595 GranSimLight_enter_system(event, &ActiveTSO); // adjust ActiveTSO etc
597 /* adjust time based on time-stamp */
598 if (event->time > CurrentTime[CurrentProc] &&
599 event->evttype != ContinueThread)
600 CurrentTime[CurrentProc] = event->time;
602 /* Deal with the idle PEs (may issue FindWork or MoveSpark events) */
603 if (!RtsFlags.GranFlags.Light)
606 IF_DEBUG(gran, fprintf(stderr, "GRAN: switch by event-type\n"))
608 /* main event dispatcher in GranSim */
609 switch (event->evttype) {
610 /* Should just be continuing execution */
612 IF_DEBUG(gran, fprintf(stderr, "GRAN: doing ContinueThread\n"));
613 /* ToDo: check assertion
614 ASSERT(run_queue_hd != (StgTSO*)NULL &&
615 run_queue_hd != END_TSO_QUEUE);
617 /* Ignore ContinueThreads for fetching threads (if synchr comm) */
618 if (!RtsFlags.GranFlags.DoAsyncFetch &&
619 procStatus[CurrentProc]==Fetching) {
620 belch("ghuH: Spurious ContinueThread while Fetching ignored; TSO %d (%p) [PE %d]",
621 CurrentTSO->id, CurrentTSO, CurrentProc);
624 /* Ignore ContinueThreads for completed threads */
625 if (CurrentTSO->what_next == ThreadComplete) {
626 belch("ghuH: found a ContinueThread event for completed thread %d (%p) [PE %d] (ignoring ContinueThread)",
627 CurrentTSO->id, CurrentTSO, CurrentProc);
630 /* Ignore ContinueThreads for threads that are being migrated */
631 if (PROCS(CurrentTSO)==Nowhere) {
632 belch("ghuH: trying to run the migrating TSO %d (%p) [PE %d] (ignoring ContinueThread)",
633 CurrentTSO->id, CurrentTSO, CurrentProc);
636 /* The thread should be at the beginning of the run queue */
637 if (CurrentTSO!=run_queue_hds[CurrentProc]) {
638 belch("ghuH: TSO %d (%p) [PE %d] is not at the start of the run_queue when doing a ContinueThread",
639 CurrentTSO->id, CurrentTSO, CurrentProc);
640 break; // run the thread anyway
643 new_event(proc, proc, CurrentTime[proc],
645 (StgTSO*)NULL, (StgClosure*)NULL, (rtsSpark*)NULL);
647 */ /* Catches superfluous CONTINUEs -- should be unnecessary */
648 break; // now actually run the thread; DaH Qu'vam yImuHbej
651 do_the_fetchnode(event);
652 goto next_thread; /* handle next event in event queue */
655 do_the_globalblock(event);
656 goto next_thread; /* handle next event in event queue */
659 do_the_fetchreply(event);
660 goto next_thread; /* handle next event in event queue */
662 case UnblockThread: /* Move from the blocked queue to the tail of */
663 do_the_unblock(event);
664 goto next_thread; /* handle next event in event queue */
666 case ResumeThread: /* Move from the blocked queue to the tail of */
667 /* the runnable queue ( i.e. Qu' SImqa'lu') */
668 event->tso->gran.blocktime +=
669 CurrentTime[CurrentProc] - event->tso->gran.blockedat;
670 do_the_startthread(event);
671 goto next_thread; /* handle next event in event queue */
674 do_the_startthread(event);
675 goto next_thread; /* handle next event in event queue */
678 do_the_movethread(event);
679 goto next_thread; /* handle next event in event queue */
682 do_the_movespark(event);
683 goto next_thread; /* handle next event in event queue */
686 do_the_findwork(event);
687 goto next_thread; /* handle next event in event queue */
690 barf("Illegal event type %u\n", event->evttype);
693 /* This point was scheduler_loop in the old RTS */
695 IF_DEBUG(gran, belch("GRAN: after main switch"));
697 TimeOfLastEvent = CurrentTime[CurrentProc];
698 TimeOfNextEvent = get_time_of_next_event();
699 IgnoreEvents=(TimeOfNextEvent==0); // HWL HACK
700 // CurrentTSO = ThreadQueueHd;
702 IF_DEBUG(gran, belch("GRAN: time of next event is: %ld",
705 if (RtsFlags.GranFlags.Light)
706 GranSimLight_leave_system(event, &ActiveTSO);
708 EndOfTimeSlice = CurrentTime[CurrentProc]+RtsFlags.GranFlags.time_slice;
711 belch("GRAN: end of time-slice is %#lx", EndOfTimeSlice));
713 /* in a GranSim setup the TSO stays on the run queue */
715 /* Take a thread from the run queue. */
716 t = POP_RUN_QUEUE(); // take_off_run_queue(t);
719 fprintf(stderr, "GRAN: About to run current thread, which is\n");
722 context_switch = 0; // turned on via GranYield, checking events and time slice
725 DumpGranEvent(GR_SCHEDULE, t));
727 procStatus[CurrentProc] = Busy;
731 if (PendingFetches != END_BF_QUEUE) {
735 /* ToDo: phps merge with spark activation above */
736 /* check whether we have local work and send requests if we have none */
737 if (run_queue_hd == END_TSO_QUEUE) { /* no runnable threads */
738 /* :-[ no local threads => look out for local sparks */
739 /* the spark pool for the current PE */
740 pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable
741 if (advisory_thread_count < RtsFlags.ParFlags.maxThreads &&
742 pool->hd < pool->tl) {
744 * ToDo: add GC code check that we really have enough heap afterwards!!
746 * If we're here (no runnable threads) and we have pending
747 * sparks, we must have a space problem. Get enough space
748 * to turn one of those pending sparks into a
752 spark = findSpark(); /* get a spark */
753 if (spark != (rtsSpark) NULL) {
754 tso = activateSpark(spark); /* turn the spark into a thread */
755 IF_PAR_DEBUG(schedule,
756 belch("==== schedule: Created TSO %d (%p); %d threads active",
757 tso->id, tso, advisory_thread_count));
759 if (tso==END_TSO_QUEUE) { /* failed to activate spark->back to loop */
760 belch("==^^ failed to activate spark");
762 } /* otherwise fall through & pick-up new tso */
764 IF_PAR_DEBUG(verbose,
765 belch("==^^ no local sparks (spark pool contains only NFs: %d)",
766 spark_queue_len(pool)));
770 /* =8-[ no local sparks => look for work on other PEs */
773 * We really have absolutely no work. Send out a fish
774 * (there may be some out there already), and wait for
775 * something to arrive. We clearly can't run any threads
776 * until a SCHEDULE or RESUME arrives, and so that's what
777 * we're hoping to see. (Of course, we still have to
778 * respond to other types of messages.)
781 outstandingFishes < RtsFlags.ParFlags.maxFishes ) { // &&
782 // (last_fish_arrived_at+FISH_DELAY < CURRENT_TIME)) {
783 /* fishing set in sendFish, processFish;
784 avoid flooding system with fishes via delay */
786 sendFish(pe, mytid, NEW_FISH_AGE, NEW_FISH_HISTORY,
794 } else if (PacketsWaiting()) { /* Look for incoming messages */
798 /* Now we are sure that we have some work available */
799 ASSERT(run_queue_hd != END_TSO_QUEUE);
800 /* Take a thread from the run queue, if we have work */
801 t = POP_RUN_QUEUE(); // take_off_run_queue(END_TSO_QUEUE);
803 /* ToDo: write something to the log-file
804 if (RTSflags.ParFlags.granSimStats && !sameThread)
805 DumpGranEvent(GR_SCHEDULE, RunnableThreadsHd);
809 /* the spark pool for the current PE */
810 pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable
812 IF_DEBUG(scheduler, belch("--^^ %d sparks on [%#x] (hd=%x; tl=%x; base=%x, lim=%x)",
813 spark_queue_len(pool),
815 pool->hd, pool->tl, pool->base, pool->lim));
817 IF_DEBUG(scheduler, belch("--== %d threads on [%#x] (hd=%x; tl=%x)",
818 run_queue_len(), CURRENT_PROC,
819 run_queue_hd, run_queue_tl));
824 we are running a different TSO, so write a schedule event to log file
825 NB: If we use fair scheduling we also have to write a deschedule
826 event for LastTSO; with unfair scheduling we know that the
827 previous tso has blocked whenever we switch to another tso, so
828 we don't need it in GUM for now
830 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
831 GR_SCHEDULE, t, (StgClosure *)NULL, 0, 0);
835 #else /* !GRAN && !PAR */
837 /* grab a thread from the run queue
839 ASSERT(run_queue_hd != END_TSO_QUEUE);
841 IF_DEBUG(sanity,checkTSO(t));
848 cap = free_capabilities;
849 free_capabilities = cap->link;
850 n_free_capabilities--;
855 cap->rCurrentTSO = t;
857 /* context switches are now initiated by the timer signal, unless
858 * the user specified "context switch as often as possible", with
861 if (RtsFlags.ConcFlags.ctxtSwitchTicks == 0
862 && (run_queue_hd != END_TSO_QUEUE
863 || blocked_queue_hd != END_TSO_QUEUE
864 || sleeping_queue != END_TSO_QUEUE))
869 RELEASE_LOCK(&sched_mutex);
871 IF_DEBUG(scheduler, sched_belch("-->> Running TSO %ld (%p) %s ...",
872 t->id, t, whatNext_strs[t->what_next]));
874 /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
875 /* Run the current thread
877 switch (cap->rCurrentTSO->what_next) {
880 /* Thread already finished, return to scheduler. */
881 ret = ThreadFinished;
884 ret = StgRun((StgFunPtr) stg_enterStackTop, cap);
887 ret = StgRun((StgFunPtr) stg_returnToStackTop, cap);
889 case ThreadEnterInterp:
890 ret = interpretBCO(cap);
893 barf("schedule: invalid what_next field");
895 /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
897 /* Costs for the scheduler are assigned to CCS_SYSTEM */
902 ACQUIRE_LOCK(&sched_mutex);
905 IF_DEBUG(scheduler,fprintf(stderr,"scheduler (task %ld): ", pthread_self()););
906 #elif !defined(GRAN) && !defined(PAR)
907 IF_DEBUG(scheduler,fprintf(stderr,"scheduler: "););
909 t = cap->rCurrentTSO;
912 /* HACK 675: if the last thread didn't yield, make sure to print a
913 SCHEDULE event to the log file when StgRunning the next thread, even
914 if it is the same one as before */
915 LastTSO = t; //(ret == ThreadBlocked) ? END_TSO_QUEUE : t;
916 TimeOfLastYield = CURRENT_TIME;
921 /* make all the running tasks block on a condition variable,
922 * maybe set context_switch and wait till they all pile in,
923 * then have them wait on a GC condition variable.
925 IF_DEBUG(scheduler,belch("--<< thread %ld (%p; %s) stopped: HeapOverflow",
926 t->id, t, whatNext_strs[t->what_next]));
929 ASSERT(!is_on_queue(t,CurrentProc));
932 ready_to_gc = rtsTrue;
933 context_switch = 1; /* stop other threads ASAP */
934 PUSH_ON_RUN_QUEUE(t);
935 /* actual GC is done at the end of the while loop */
939 IF_DEBUG(scheduler,belch("--<< thread %ld (%p; %s) stopped, StackOverflow",
940 t->id, t, whatNext_strs[t->what_next]));
941 /* just adjust the stack for this thread, then pop it back
947 /* enlarge the stack */
948 StgTSO *new_t = threadStackOverflow(t);
950 /* This TSO has moved, so update any pointers to it from the
951 * main thread stack. It better not be on any other queues...
954 for (m = main_threads; m != NULL; m = m->link) {
960 PUSH_ON_RUN_QUEUE(new_t);
967 DumpGranEvent(GR_DESCHEDULE, t));
968 globalGranStats.tot_yields++;
971 DumpGranEvent(GR_DESCHEDULE, t));
973 /* put the thread back on the run queue. Then, if we're ready to
974 * GC, check whether this is the last task to stop. If so, wake
975 * up the GC thread. getThread will block during a GC until the
979 if (t->what_next == ThreadEnterInterp) {
980 /* ToDo: or maybe a timer expired when we were in Hugs?
981 * or maybe someone hit ctrl-C
983 belch("--<< thread %ld (%p; %s) stopped to switch to Hugs",
984 t->id, t, whatNext_strs[t->what_next]);
986 belch("--<< thread %ld (%p; %s) stopped, yielding",
987 t->id, t, whatNext_strs[t->what_next]);
994 //belch("&& Doing sanity check on yielding TSO %ld.", t->id);
996 ASSERT(t->link == END_TSO_QUEUE);
998 ASSERT(!is_on_queue(t,CurrentProc));
1001 //belch("&& Doing sanity check on all ThreadQueues (and their TSOs).");
1002 checkThreadQsSanity(rtsTrue));
1004 APPEND_TO_RUN_QUEUE(t);
1006 /* add a ContinueThread event to actually process the thread */
1007 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1009 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1011 belch("GRAN: eventq and runnableq after adding yielded thread to queue again:");
1020 belch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: ",
1021 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)));
1022 if (t->block_info.closure!=(StgClosure*)NULL) print_bq(t->block_info.closure));
1024 // ??? needed; should emit block before
1026 DumpGranEvent(GR_DESCHEDULE, t));
1027 prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
1030 ASSERT(procStatus[CurrentProc]==Busy ||
1031 ((procStatus[CurrentProc]==Fetching) &&
1032 (t->block_info.closure!=(StgClosure*)NULL)));
1033 if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
1034 !(!RtsFlags.GranFlags.DoAsyncFetch &&
1035 procStatus[CurrentProc]==Fetching))
1036 procStatus[CurrentProc] = Idle;
1040 DumpGranEvent(GR_DESCHEDULE, t));
1042 /* Send a fetch (if BlockedOnGA) and dump event to log file */
1046 belch("--<< thread %ld (%p; %s) stopped, blocking on node %p with BQ: ",
1047 t->id, t, whatNext_strs[t->what_next], t->block_info.closure);
1048 if (t->block_info.closure!=(StgClosure*)NULL) print_bq(t->block_info.closure));
1051 /* don't need to do anything. Either the thread is blocked on
1052 * I/O, in which case we'll have called addToBlockedQueue
1053 * previously, or it's blocked on an MVar or Blackhole, in which
1054 * case it'll be on the relevant queue already.
1057 fprintf(stderr, "--<< thread %d (%p) stopped: ", t->id, t);
1058 printThreadBlockage(t);
1059 fprintf(stderr, "\n"));
1061 /* Only for dumping event to log file
1062 ToDo: do I need this in GranSim, too?
1069 case ThreadFinished:
1070 /* Need to check whether this was a main thread, and if so, signal
1071 * the task that started it with the return value. If we have no
1072 * more main threads, we probably need to stop all the tasks until
1075 /* We also end up here if the thread kills itself with an
1076 * uncaught exception, see Exception.hc.
1078 IF_DEBUG(scheduler,belch("--++ thread %d (%p) finished", t->id, t));
1080 endThread(t, CurrentProc); // clean-up the thread
1082 advisory_thread_count--;
1083 if (RtsFlags.ParFlags.ParStats.Full)
1084 DumpEndEvent(CURRENT_PROC, t, rtsFalse /* not mandatory */);
1089 barf("schedule: invalid thread return code %d", (int)ret);
1093 cap->link = free_capabilities;
1094 free_capabilities = cap;
1095 n_free_capabilities++;
1099 if (ready_to_gc && n_free_capabilities == RtsFlags.ParFlags.nNodes)
1104 /* everybody back, start the GC.
1105 * Could do it in this thread, or signal a condition var
1106 * to do it in another thread. Either way, we need to
1107 * broadcast on gc_pending_cond afterward.
1110 IF_DEBUG(scheduler,sched_belch("doing GC"));
1112 GarbageCollect(GetRoots,rtsFalse);
1113 ready_to_gc = rtsFalse;
1115 pthread_cond_broadcast(&gc_pending_cond);
1118 /* add a ContinueThread event to continue execution of current thread */
1119 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1121 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1123 fprintf(stderr, "GRAN: eventq and runnableq after Garbage collection:\n");
1130 IF_GRAN_DEBUG(unused,
1131 print_eventq(EventHd));
1133 event = get_next_event();
1137 /* ToDo: wait for next message to arrive rather than busy wait */
1142 t = take_off_run_queue(END_TSO_QUEUE);
1145 } /* end of while(1) */
1148 /* ---------------------------------------------------------------------------
1149 * deleteAllThreads(): kill all the live threads.
1151 * This is used when we catch a user interrupt (^C), before performing
1152 * any necessary cleanups and running finalizers.
1153 * ------------------------------------------------------------------------- */
1155 void deleteAllThreads ( void )
1158 IF_DEBUG(scheduler,sched_belch("deleting all threads"));
1159 for (t = run_queue_hd; t != END_TSO_QUEUE; t = t->link) {
1162 for (t = blocked_queue_hd; t != END_TSO_QUEUE; t = t->link) {
1165 for (t = sleeping_queue; t != END_TSO_QUEUE; t = t->link) {
1168 run_queue_hd = run_queue_tl = END_TSO_QUEUE;
1169 blocked_queue_hd = blocked_queue_tl = END_TSO_QUEUE;
1170 sleeping_queue = END_TSO_QUEUE;
1173 /* startThread and insertThread are now in GranSim.c -- HWL */
1175 //@node Suspend and Resume, Run queue code, Main scheduling loop, Main scheduling code
1176 //@subsection Suspend and Resume
1178 /* ---------------------------------------------------------------------------
1179 * Suspending & resuming Haskell threads.
1181 * When making a "safe" call to C (aka _ccall_GC), the task gives back
1182 * its capability before calling the C function. This allows another
1183 * task to pick up the capability and carry on running Haskell
1184 * threads. It also means that if the C call blocks, it won't lock
1187 * The Haskell thread making the C call is put to sleep for the
1188 * duration of the call, on the susepended_ccalling_threads queue. We
1189 * give out a token to the task, which it can use to resume the thread
1190 * on return from the C function.
1191 * ------------------------------------------------------------------------- */
1194 suspendThread( Capability *cap )
1198 ACQUIRE_LOCK(&sched_mutex);
1201 sched_belch("thread %d did a _ccall_gc", cap->rCurrentTSO->id));
1203 threadPaused(cap->rCurrentTSO);
1204 cap->rCurrentTSO->link = suspended_ccalling_threads;
1205 suspended_ccalling_threads = cap->rCurrentTSO;
1207 /* Use the thread ID as the token; it should be unique */
1208 tok = cap->rCurrentTSO->id;
1211 cap->link = free_capabilities;
1212 free_capabilities = cap;
1213 n_free_capabilities++;
1216 RELEASE_LOCK(&sched_mutex);
1221 resumeThread( StgInt tok )
1223 StgTSO *tso, **prev;
1226 ACQUIRE_LOCK(&sched_mutex);
1228 prev = &suspended_ccalling_threads;
1229 for (tso = suspended_ccalling_threads;
1230 tso != END_TSO_QUEUE;
1231 prev = &tso->link, tso = tso->link) {
1232 if (tso->id == (StgThreadID)tok) {
1237 if (tso == END_TSO_QUEUE) {
1238 barf("resumeThread: thread not found");
1240 tso->link = END_TSO_QUEUE;
1243 while (free_capabilities == NULL) {
1244 IF_DEBUG(scheduler, sched_belch("waiting to resume"));
1245 pthread_cond_wait(&thread_ready_cond, &sched_mutex);
1246 IF_DEBUG(scheduler, sched_belch("resuming thread %d", tso->id));
1248 cap = free_capabilities;
1249 free_capabilities = cap->link;
1250 n_free_capabilities--;
1252 cap = &MainRegTable;
1255 cap->rCurrentTSO = tso;
1257 RELEASE_LOCK(&sched_mutex);
1262 /* ---------------------------------------------------------------------------
1264 * ------------------------------------------------------------------------ */
1265 static void unblockThread(StgTSO *tso);
1267 /* ---------------------------------------------------------------------------
1268 * Comparing Thread ids.
1270 * This is used from STG land in the implementation of the
1271 * instances of Eq/Ord for ThreadIds.
1272 * ------------------------------------------------------------------------ */
1274 int cmp_thread(const StgTSO *tso1, const StgTSO *tso2)
1276 StgThreadID id1 = tso1->id;
1277 StgThreadID id2 = tso2->id;
1279 if (id1 < id2) return (-1);
1280 if (id1 > id2) return 1;
1284 /* ---------------------------------------------------------------------------
1285 Create a new thread.
1287 The new thread starts with the given stack size. Before the
1288 scheduler can run, however, this thread needs to have a closure
1289 (and possibly some arguments) pushed on its stack. See
1290 pushClosure() in Schedule.h.
1292 createGenThread() and createIOThread() (in SchedAPI.h) are
1293 convenient packaged versions of this function.
1295 currently pri (priority) is only used in a GRAN setup -- HWL
1296 ------------------------------------------------------------------------ */
1297 //@cindex createThread
1299 /* currently pri (priority) is only used in a GRAN setup -- HWL */
1301 createThread(nat stack_size, StgInt pri)
1303 return createThread_(stack_size, rtsFalse, pri);
1307 createThread_(nat size, rtsBool have_lock, StgInt pri)
1311 createThread(nat stack_size)
1313 return createThread_(stack_size, rtsFalse);
1317 createThread_(nat size, rtsBool have_lock)
1324 /* First check whether we should create a thread at all */
1326 /* check that no more than RtsFlags.ParFlags.maxThreads threads are created */
1327 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads) {
1329 belch("{createThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
1330 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
1331 return END_TSO_QUEUE;
1337 ASSERT(!RtsFlags.GranFlags.Light || CurrentProc==0);
1340 // ToDo: check whether size = stack_size - TSO_STRUCT_SIZEW
1342 /* catch ridiculously small stack sizes */
1343 if (size < MIN_STACK_WORDS + TSO_STRUCT_SIZEW) {
1344 size = MIN_STACK_WORDS + TSO_STRUCT_SIZEW;
1347 stack_size = size - TSO_STRUCT_SIZEW;
1349 tso = (StgTSO *)allocate(size);
1350 TICK_ALLOC_TSO(size-TSO_STRUCT_SIZEW, 0);
1352 SET_HDR(tso, &stg_TSO_info, CCS_SYSTEM);
1354 SET_GRAN_HDR(tso, ThisPE);
1356 tso->what_next = ThreadEnterGHC;
1358 /* tso->id needs to be unique. For now we use a heavyweight mutex to
1359 * protect the increment operation on next_thread_id.
1360 * In future, we could use an atomic increment instead.
1362 if (!have_lock) { ACQUIRE_LOCK(&sched_mutex); }
1363 tso->id = next_thread_id++;
1364 if (!have_lock) { RELEASE_LOCK(&sched_mutex); }
1366 tso->why_blocked = NotBlocked;
1367 tso->blocked_exceptions = NULL;
1369 tso->stack_size = stack_size;
1370 tso->max_stack_size = round_to_mblocks(RtsFlags.GcFlags.maxStkSize)
1372 tso->sp = (P_)&(tso->stack) + stack_size;
1375 tso->prof.CCCS = CCS_MAIN;
1378 /* put a stop frame on the stack */
1379 tso->sp -= sizeofW(StgStopFrame);
1380 SET_HDR((StgClosure*)tso->sp,(StgInfoTable *)&stg_stop_thread_info,CCS_SYSTEM);
1381 tso->su = (StgUpdateFrame*)tso->sp;
1385 tso->link = END_TSO_QUEUE;
1386 /* uses more flexible routine in GranSim */
1387 insertThread(tso, CurrentProc);
1389 /* In a non-GranSim setup the pushing of a TSO onto the runq is separated
1394 #if defined(GRAN) || defined(PAR)
1395 DumpGranEvent(GR_START,tso);
1398 /* Link the new thread on the global thread list.
1400 tso->global_link = all_threads;
1404 tso->gran.pri = pri;
1406 tso->gran.magic = TSO_MAGIC; // debugging only
1408 tso->gran.sparkname = 0;
1409 tso->gran.startedat = CURRENT_TIME;
1410 tso->gran.exported = 0;
1411 tso->gran.basicblocks = 0;
1412 tso->gran.allocs = 0;
1413 tso->gran.exectime = 0;
1414 tso->gran.fetchtime = 0;
1415 tso->gran.fetchcount = 0;
1416 tso->gran.blocktime = 0;
1417 tso->gran.blockcount = 0;
1418 tso->gran.blockedat = 0;
1419 tso->gran.globalsparks = 0;
1420 tso->gran.localsparks = 0;
1421 if (RtsFlags.GranFlags.Light)
1422 tso->gran.clock = Now; /* local clock */
1424 tso->gran.clock = 0;
1426 IF_DEBUG(gran,printTSO(tso));
1429 tso->par.magic = TSO_MAGIC; // debugging only
1431 tso->par.sparkname = 0;
1432 tso->par.startedat = CURRENT_TIME;
1433 tso->par.exported = 0;
1434 tso->par.basicblocks = 0;
1435 tso->par.allocs = 0;
1436 tso->par.exectime = 0;
1437 tso->par.fetchtime = 0;
1438 tso->par.fetchcount = 0;
1439 tso->par.blocktime = 0;
1440 tso->par.blockcount = 0;
1441 tso->par.blockedat = 0;
1442 tso->par.globalsparks = 0;
1443 tso->par.localsparks = 0;
1447 globalGranStats.tot_threads_created++;
1448 globalGranStats.threads_created_on_PE[CurrentProc]++;
1449 globalGranStats.tot_sq_len += spark_queue_len(CurrentProc);
1450 globalGranStats.tot_sq_probes++;
1455 belch("==__ schedule: Created TSO %d (%p);",
1456 CurrentProc, tso, tso->id));
1458 IF_PAR_DEBUG(verbose,
1459 belch("==__ schedule: Created TSO %d (%p); %d threads active",
1460 tso->id, tso, advisory_thread_count));
1462 IF_DEBUG(scheduler,sched_belch("created thread %ld, stack size = %lx words",
1463 tso->id, tso->stack_size));
1469 Turn a spark into a thread.
1470 ToDo: fix for SMP (needs to acquire SCHED_MUTEX!)
1473 //@cindex activateSpark
1475 activateSpark (rtsSpark spark)
1479 ASSERT(spark != (rtsSpark)NULL);
1480 tso = createThread_(RtsFlags.GcFlags.initialStkSize, rtsTrue);
1481 if (tso!=END_TSO_QUEUE) {
1482 pushClosure(tso,spark);
1483 PUSH_ON_RUN_QUEUE(tso);
1484 advisory_thread_count++;
1486 if (RtsFlags.ParFlags.ParStats.Full) {
1487 //ASSERT(run_queue_hd == END_TSO_QUEUE); // I think ...
1488 IF_PAR_DEBUG(verbose,
1489 belch("==^^ activateSpark: turning spark of closure %p (%s) into a thread",
1490 (StgClosure *)spark, info_type((StgClosure *)spark)));
1493 barf("activateSpark: Cannot create TSO");
1495 // ToDo: fwd info on local/global spark to thread -- HWL
1496 // tso->gran.exported = spark->exported;
1497 // tso->gran.locked = !spark->global;
1498 // tso->gran.sparkname = spark->name;
1504 /* ---------------------------------------------------------------------------
1507 * scheduleThread puts a thread on the head of the runnable queue.
1508 * This will usually be done immediately after a thread is created.
1509 * The caller of scheduleThread must create the thread using e.g.
1510 * createThread and push an appropriate closure
1511 * on this thread's stack before the scheduler is invoked.
1512 * ------------------------------------------------------------------------ */
1515 scheduleThread(StgTSO *tso)
1517 if (tso==END_TSO_QUEUE){
1522 ACQUIRE_LOCK(&sched_mutex);
1524 /* Put the new thread on the head of the runnable queue. The caller
1525 * better push an appropriate closure on this thread's stack
1526 * beforehand. In the SMP case, the thread may start running as
1527 * soon as we release the scheduler lock below.
1529 PUSH_ON_RUN_QUEUE(tso);
1533 IF_DEBUG(scheduler,printTSO(tso));
1535 RELEASE_LOCK(&sched_mutex);
1538 /* ---------------------------------------------------------------------------
1541 * Start up Posix threads to run each of the scheduler tasks.
1542 * I believe the task ids are not needed in the system as defined.
1544 * ------------------------------------------------------------------------ */
1546 #if defined(PAR) || defined(SMP)
1548 taskStart( void *arg STG_UNUSED )
1550 rts_evalNothing(NULL);
1554 /* ---------------------------------------------------------------------------
1557 * Initialise the scheduler. This resets all the queues - if the
1558 * queues contained any threads, they'll be garbage collected at the
1561 * This now calls startTasks(), so should only be called once! KH @ 25/10/99
1562 * ------------------------------------------------------------------------ */
1566 term_handler(int sig STG_UNUSED)
1569 ACQUIRE_LOCK(&term_mutex);
1571 RELEASE_LOCK(&term_mutex);
1576 //@cindex initScheduler
1583 for (i=0; i<=MAX_PROC; i++) {
1584 run_queue_hds[i] = END_TSO_QUEUE;
1585 run_queue_tls[i] = END_TSO_QUEUE;
1586 blocked_queue_hds[i] = END_TSO_QUEUE;
1587 blocked_queue_tls[i] = END_TSO_QUEUE;
1588 ccalling_threadss[i] = END_TSO_QUEUE;
1589 sleeping_queue = END_TSO_QUEUE;
1592 run_queue_hd = END_TSO_QUEUE;
1593 run_queue_tl = END_TSO_QUEUE;
1594 blocked_queue_hd = END_TSO_QUEUE;
1595 blocked_queue_tl = END_TSO_QUEUE;
1596 sleeping_queue = END_TSO_QUEUE;
1599 suspended_ccalling_threads = END_TSO_QUEUE;
1601 main_threads = NULL;
1602 all_threads = END_TSO_QUEUE;
1607 RtsFlags.ConcFlags.ctxtSwitchTicks =
1608 RtsFlags.ConcFlags.ctxtSwitchTime / TICK_MILLISECS;
1610 /* Install the SIGHUP handler */
1613 struct sigaction action,oact;
1615 action.sa_handler = term_handler;
1616 sigemptyset(&action.sa_mask);
1617 action.sa_flags = 0;
1618 if (sigaction(SIGTERM, &action, &oact) != 0) {
1619 barf("can't install TERM handler");
1625 /* Allocate N Capabilities */
1628 Capability *cap, *prev;
1631 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
1632 cap = stgMallocBytes(sizeof(Capability), "initScheduler:capabilities");
1636 free_capabilities = cap;
1637 n_free_capabilities = RtsFlags.ParFlags.nNodes;
1639 IF_DEBUG(scheduler,fprintf(stderr,"scheduler: Allocated %d capabilities\n",
1640 n_free_capabilities););
1643 #if defined(SMP) || defined(PAR)
1656 /* make some space for saving all the thread ids */
1657 task_ids = stgMallocBytes(RtsFlags.ParFlags.nNodes * sizeof(task_info),
1658 "initScheduler:task_ids");
1660 /* and create all the threads */
1661 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
1662 r = pthread_create(&tid,NULL,taskStart,NULL);
1664 barf("startTasks: Can't create new Posix thread");
1666 task_ids[i].id = tid;
1667 task_ids[i].mut_time = 0.0;
1668 task_ids[i].mut_etime = 0.0;
1669 task_ids[i].gc_time = 0.0;
1670 task_ids[i].gc_etime = 0.0;
1671 task_ids[i].elapsedtimestart = elapsedtime();
1672 IF_DEBUG(scheduler,fprintf(stderr,"scheduler: Started task: %ld\n",tid););
1678 exitScheduler( void )
1683 /* Don't want to use pthread_cancel, since we'd have to install
1684 * these silly exception handlers (pthread_cleanup_{push,pop}) around
1688 /* Cancel all our tasks */
1689 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
1690 pthread_cancel(task_ids[i].id);
1693 /* Wait for all the tasks to terminate */
1694 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
1695 IF_DEBUG(scheduler,fprintf(stderr,"scheduler: waiting for task %ld\n",
1697 pthread_join(task_ids[i].id, NULL);
1701 /* Send 'em all a SIGHUP. That should shut 'em up.
1703 await_death = RtsFlags.ParFlags.nNodes;
1704 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
1705 pthread_kill(task_ids[i].id,SIGTERM);
1707 while (await_death > 0) {
1713 /* -----------------------------------------------------------------------------
1714 Managing the per-task allocation areas.
1716 Each capability comes with an allocation area. These are
1717 fixed-length block lists into which allocation can be done.
1719 ToDo: no support for two-space collection at the moment???
1720 -------------------------------------------------------------------------- */
1722 /* -----------------------------------------------------------------------------
1723 * waitThread is the external interface for running a new computation
1724 * and waiting for the result.
1726 * In the non-SMP case, we create a new main thread, push it on the
1727 * main-thread stack, and invoke the scheduler to run it. The
1728 * scheduler will return when the top main thread on the stack has
1729 * completed or died, and fill in the necessary fields of the
1730 * main_thread structure.
1732 * In the SMP case, we create a main thread as before, but we then
1733 * create a new condition variable and sleep on it. When our new
1734 * main thread has completed, we'll be woken up and the status/result
1735 * will be in the main_thread struct.
1736 * -------------------------------------------------------------------------- */
1739 howManyThreadsAvail ( void )
1743 for (q = run_queue_hd; q != END_TSO_QUEUE; q = q->link)
1745 for (q = blocked_queue_hd; q != END_TSO_QUEUE; q = q->link)
1747 for (q = sleeping_queue; q != END_TSO_QUEUE; q = q->link)
1753 finishAllThreads ( void )
1756 while (run_queue_hd != END_TSO_QUEUE) {
1757 waitThread ( run_queue_hd, NULL );
1759 while (blocked_queue_hd != END_TSO_QUEUE) {
1760 waitThread ( blocked_queue_hd, NULL );
1762 while (sleeping_queue != END_TSO_QUEUE) {
1763 waitThread ( blocked_queue_hd, NULL );
1766 (blocked_queue_hd != END_TSO_QUEUE ||
1767 run_queue_hd != END_TSO_QUEUE ||
1768 sleeping_queue != END_TSO_QUEUE);
1772 waitThread(StgTSO *tso, /*out*/StgClosure **ret)
1775 SchedulerStatus stat;
1777 ACQUIRE_LOCK(&sched_mutex);
1779 m = stgMallocBytes(sizeof(StgMainThread), "waitThread");
1785 pthread_cond_init(&m->wakeup, NULL);
1788 m->link = main_threads;
1791 IF_DEBUG(scheduler, fprintf(stderr, "scheduler: new main thread (%d)\n",
1796 pthread_cond_wait(&m->wakeup, &sched_mutex);
1797 } while (m->stat == NoStatus);
1799 /* GranSim specific init */
1800 CurrentTSO = m->tso; // the TSO to run
1801 procStatus[MainProc] = Busy; // status of main PE
1802 CurrentProc = MainProc; // PE to run it on
1807 ASSERT(m->stat != NoStatus);
1813 pthread_cond_destroy(&m->wakeup);
1816 IF_DEBUG(scheduler, fprintf(stderr, "scheduler: main thread (%d) finished\n",
1820 RELEASE_LOCK(&sched_mutex);
1825 //@node Run queue code, Garbage Collextion Routines, Suspend and Resume, Main scheduling code
1826 //@subsection Run queue code
1830 NB: In GranSim we have many run queues; run_queue_hd is actually a macro
1831 unfolding to run_queue_hds[CurrentProc], thus CurrentProc is an
1832 implicit global variable that has to be correct when calling these
1836 /* Put the new thread on the head of the runnable queue.
1837 * The caller of createThread better push an appropriate closure
1838 * on this thread's stack before the scheduler is invoked.
1840 static /* inline */ void
1841 add_to_run_queue(tso)
1844 ASSERT(tso!=run_queue_hd && tso!=run_queue_tl);
1845 tso->link = run_queue_hd;
1847 if (run_queue_tl == END_TSO_QUEUE) {
1852 /* Put the new thread at the end of the runnable queue. */
1853 static /* inline */ void
1854 push_on_run_queue(tso)
1857 ASSERT(get_itbl((StgClosure *)tso)->type == TSO);
1858 ASSERT(run_queue_hd!=NULL && run_queue_tl!=NULL);
1859 ASSERT(tso!=run_queue_hd && tso!=run_queue_tl);
1860 if (run_queue_hd == END_TSO_QUEUE) {
1863 run_queue_tl->link = tso;
1869 Should be inlined because it's used very often in schedule. The tso
1870 argument is actually only needed in GranSim, where we want to have the
1871 possibility to schedule *any* TSO on the run queue, irrespective of the
1872 actual ordering. Therefore, if tso is not the nil TSO then we traverse
1873 the run queue and dequeue the tso, adjusting the links in the queue.
1875 //@cindex take_off_run_queue
1876 static /* inline */ StgTSO*
1877 take_off_run_queue(StgTSO *tso) {
1881 qetlaHbogh Qu' ngaSbogh ghomDaQ {tso} yIteq!
1883 if tso is specified, unlink that tso from the run_queue (doesn't have
1884 to be at the beginning of the queue); GranSim only
1886 if (tso!=END_TSO_QUEUE) {
1887 /* find tso in queue */
1888 for (t=run_queue_hd, prev=END_TSO_QUEUE;
1889 t!=END_TSO_QUEUE && t!=tso;
1893 /* now actually dequeue the tso */
1894 if (prev!=END_TSO_QUEUE) {
1895 ASSERT(run_queue_hd!=t);
1896 prev->link = t->link;
1898 /* t is at beginning of thread queue */
1899 ASSERT(run_queue_hd==t);
1900 run_queue_hd = t->link;
1902 /* t is at end of thread queue */
1903 if (t->link==END_TSO_QUEUE) {
1904 ASSERT(t==run_queue_tl);
1905 run_queue_tl = prev;
1907 ASSERT(run_queue_tl!=t);
1909 t->link = END_TSO_QUEUE;
1911 /* take tso from the beginning of the queue; std concurrent code */
1913 if (t != END_TSO_QUEUE) {
1914 run_queue_hd = t->link;
1915 t->link = END_TSO_QUEUE;
1916 if (run_queue_hd == END_TSO_QUEUE) {
1917 run_queue_tl = END_TSO_QUEUE;
1926 //@node Garbage Collextion Routines, Blocking Queue Routines, Run queue code, Main scheduling code
1927 //@subsection Garbage Collextion Routines
1929 /* ---------------------------------------------------------------------------
1930 Where are the roots that we know about?
1932 - all the threads on the runnable queue
1933 - all the threads on the blocked queue
1934 - all the threads on the sleeping queue
1935 - all the thread currently executing a _ccall_GC
1936 - all the "main threads"
1938 ------------------------------------------------------------------------ */
1940 /* This has to be protected either by the scheduler monitor, or by the
1941 garbage collection monitor (probably the latter).
1945 static void GetRoots(void)
1952 for (i=0; i<=RtsFlags.GranFlags.proc; i++) {
1953 if ((run_queue_hds[i] != END_TSO_QUEUE) && ((run_queue_hds[i] != NULL)))
1954 run_queue_hds[i] = (StgTSO *)MarkRoot((StgClosure *)run_queue_hds[i]);
1955 if ((run_queue_tls[i] != END_TSO_QUEUE) && ((run_queue_tls[i] != NULL)))
1956 run_queue_tls[i] = (StgTSO *)MarkRoot((StgClosure *)run_queue_tls[i]);
1958 if ((blocked_queue_hds[i] != END_TSO_QUEUE) && ((blocked_queue_hds[i] != NULL)))
1959 blocked_queue_hds[i] = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_hds[i]);
1960 if ((blocked_queue_tls[i] != END_TSO_QUEUE) && ((blocked_queue_tls[i] != NULL)))
1961 blocked_queue_tls[i] = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_tls[i]);
1962 if ((ccalling_threadss[i] != END_TSO_QUEUE) && ((ccalling_threadss[i] != NULL)))
1963 ccalling_threadss[i] = (StgTSO *)MarkRoot((StgClosure *)ccalling_threadss[i]);
1970 if (run_queue_hd != END_TSO_QUEUE) {
1971 ASSERT(run_queue_tl != END_TSO_QUEUE);
1972 run_queue_hd = (StgTSO *)MarkRoot((StgClosure *)run_queue_hd);
1973 run_queue_tl = (StgTSO *)MarkRoot((StgClosure *)run_queue_tl);
1976 if (blocked_queue_hd != END_TSO_QUEUE) {
1977 ASSERT(blocked_queue_tl != END_TSO_QUEUE);
1978 blocked_queue_hd = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_hd);
1979 blocked_queue_tl = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_tl);
1982 if (sleeping_queue != END_TSO_QUEUE) {
1983 sleeping_queue = (StgTSO *)MarkRoot((StgClosure *)sleeping_queue);
1987 for (m = main_threads; m != NULL; m = m->link) {
1988 m->tso = (StgTSO *)MarkRoot((StgClosure *)m->tso);
1990 if (suspended_ccalling_threads != END_TSO_QUEUE)
1991 suspended_ccalling_threads =
1992 (StgTSO *)MarkRoot((StgClosure *)suspended_ccalling_threads);
1994 #if defined(SMP) || defined(PAR) || defined(GRAN)
1999 /* -----------------------------------------------------------------------------
2002 This is the interface to the garbage collector from Haskell land.
2003 We provide this so that external C code can allocate and garbage
2004 collect when called from Haskell via _ccall_GC.
2006 It might be useful to provide an interface whereby the programmer
2007 can specify more roots (ToDo).
2009 This needs to be protected by the GC condition variable above. KH.
2010 -------------------------------------------------------------------------- */
2012 void (*extra_roots)(void);
2017 GarbageCollect(GetRoots,rtsFalse);
2021 performMajorGC(void)
2023 GarbageCollect(GetRoots,rtsTrue);
2029 GetRoots(); /* the scheduler's roots */
2030 extra_roots(); /* the user's roots */
2034 performGCWithRoots(void (*get_roots)(void))
2036 extra_roots = get_roots;
2038 GarbageCollect(AllRoots,rtsFalse);
2041 /* -----------------------------------------------------------------------------
2044 If the thread has reached its maximum stack size, then raise the
2045 StackOverflow exception in the offending thread. Otherwise
2046 relocate the TSO into a larger chunk of memory and adjust its stack
2048 -------------------------------------------------------------------------- */
2051 threadStackOverflow(StgTSO *tso)
2053 nat new_stack_size, new_tso_size, diff, stack_words;
2057 IF_DEBUG(sanity,checkTSO(tso));
2058 if (tso->stack_size >= tso->max_stack_size) {
2061 belch("@@ threadStackOverflow of TSO %d (%p): stack too large (now %ld; max is %ld",
2062 tso->id, tso, tso->stack_size, tso->max_stack_size);
2063 /* If we're debugging, just print out the top of the stack */
2064 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2067 /* Send this thread the StackOverflow exception */
2068 raiseAsync(tso, (StgClosure *)stackOverflow_closure);
2072 /* Try to double the current stack size. If that takes us over the
2073 * maximum stack size for this thread, then use the maximum instead.
2074 * Finally round up so the TSO ends up as a whole number of blocks.
2076 new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
2077 new_tso_size = (nat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) +
2078 TSO_STRUCT_SIZE)/sizeof(W_);
2079 new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */
2080 new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
2082 IF_DEBUG(scheduler, fprintf(stderr,"scheduler: increasing stack size from %d words to %d.\n", tso->stack_size, new_stack_size));
2084 dest = (StgTSO *)allocate(new_tso_size);
2085 TICK_ALLOC_TSO(new_tso_size-sizeofW(StgTSO),0);
2087 /* copy the TSO block and the old stack into the new area */
2088 memcpy(dest,tso,TSO_STRUCT_SIZE);
2089 stack_words = tso->stack + tso->stack_size - tso->sp;
2090 new_sp = (P_)dest + new_tso_size - stack_words;
2091 memcpy(new_sp, tso->sp, stack_words * sizeof(W_));
2093 /* relocate the stack pointers... */
2094 diff = (P_)new_sp - (P_)tso->sp; /* In *words* */
2095 dest->su = (StgUpdateFrame *) ((P_)dest->su + diff);
2097 dest->stack_size = new_stack_size;
2099 /* and relocate the update frame list */
2100 relocate_TSO(tso, dest);
2102 /* Mark the old TSO as relocated. We have to check for relocated
2103 * TSOs in the garbage collector and any primops that deal with TSOs.
2105 * It's important to set the sp and su values to just beyond the end
2106 * of the stack, so we don't attempt to scavenge any part of the
2109 tso->what_next = ThreadRelocated;
2111 tso->sp = (P_)&(tso->stack[tso->stack_size]);
2112 tso->su = (StgUpdateFrame *)tso->sp;
2113 tso->why_blocked = NotBlocked;
2114 dest->mut_link = NULL;
2116 IF_PAR_DEBUG(verbose,
2117 belch("@@ threadStackOverflow of TSO %d (now at %p): stack size increased to %ld",
2118 tso->id, tso, tso->stack_size);
2119 /* If we're debugging, just print out the top of the stack */
2120 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2123 IF_DEBUG(sanity,checkTSO(tso));
2125 IF_DEBUG(scheduler,printTSO(dest));
2131 //@node Blocking Queue Routines, Exception Handling Routines, Garbage Collextion Routines, Main scheduling code
2132 //@subsection Blocking Queue Routines
2134 /* ---------------------------------------------------------------------------
2135 Wake up a queue that was blocked on some resource.
2136 ------------------------------------------------------------------------ */
2140 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2145 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2147 /* write RESUME events to log file and
2148 update blocked and fetch time (depending on type of the orig closure) */
2149 if (RtsFlags.ParFlags.ParStats.Full) {
2150 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
2151 GR_RESUME, ((StgTSO *)bqe), ((StgTSO *)bqe)->block_info.closure,
2152 0, 0 /* spark_queue_len(ADVISORY_POOL) */);
2154 switch (get_itbl(node)->type) {
2156 ((StgTSO *)bqe)->par.fetchtime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2161 ((StgTSO *)bqe)->par.blocktime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2164 barf("{unblockOneLocked}Daq Qagh: unexpected closure in blocking queue");
2171 static StgBlockingQueueElement *
2172 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2175 PEs node_loc, tso_loc;
2177 node_loc = where_is(node); // should be lifted out of loop
2178 tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2179 tso_loc = where_is((StgClosure *)tso);
2180 if (IS_LOCAL_TO(PROCS(node),tso_loc)) { // TSO is local
2181 /* !fake_fetch => TSO is on CurrentProc is same as IS_LOCAL_TO */
2182 ASSERT(CurrentProc!=node_loc || tso_loc==CurrentProc);
2183 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.lunblocktime;
2184 // insertThread(tso, node_loc);
2185 new_event(tso_loc, tso_loc, CurrentTime[CurrentProc],
2187 tso, node, (rtsSpark*)NULL);
2188 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2191 } else { // TSO is remote (actually should be FMBQ)
2192 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.mpacktime +
2193 RtsFlags.GranFlags.Costs.gunblocktime +
2194 RtsFlags.GranFlags.Costs.latency;
2195 new_event(tso_loc, CurrentProc, CurrentTime[CurrentProc],
2197 tso, node, (rtsSpark*)NULL);
2198 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2201 /* the thread-queue-overhead is accounted for in either Resume or UnblockThread */
2203 fprintf(stderr," %s TSO %d (%p) [PE %d] (block_info.closure=%p) (next=%p) ,",
2204 (node_loc==tso_loc ? "Local" : "Global"),
2205 tso->id, tso, CurrentProc, tso->block_info.closure, tso->link));
2206 tso->block_info.closure = NULL;
2207 IF_DEBUG(scheduler,belch("-- Waking up thread %ld (%p)",
2211 static StgBlockingQueueElement *
2212 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2214 StgBlockingQueueElement *next;
2216 switch (get_itbl(bqe)->type) {
2218 ASSERT(((StgTSO *)bqe)->why_blocked != NotBlocked);
2219 /* if it's a TSO just push it onto the run_queue */
2221 // ((StgTSO *)bqe)->link = END_TSO_QUEUE; // debugging?
2222 PUSH_ON_RUN_QUEUE((StgTSO *)bqe);
2224 unblockCount(bqe, node);
2225 /* reset blocking status after dumping event */
2226 ((StgTSO *)bqe)->why_blocked = NotBlocked;
2230 /* if it's a BLOCKED_FETCH put it on the PendingFetches list */
2232 bqe->link = PendingFetches;
2233 PendingFetches = bqe;
2237 /* can ignore this case in a non-debugging setup;
2238 see comments on RBHSave closures above */
2240 /* check that the closure is an RBHSave closure */
2241 ASSERT(get_itbl((StgClosure *)bqe) == &stg_RBH_Save_0_info ||
2242 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_1_info ||
2243 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_2_info);
2247 barf("{unblockOneLocked}Daq Qagh: Unexpected IP (%#lx; %s) in blocking queue at %#lx\n",
2248 get_itbl((StgClosure *)bqe), info_type((StgClosure *)bqe),
2252 // IF_DEBUG(scheduler,sched_belch("waking up thread %ld", tso->id));
2256 #else /* !GRAN && !PAR */
2258 unblockOneLocked(StgTSO *tso)
2262 ASSERT(get_itbl(tso)->type == TSO);
2263 ASSERT(tso->why_blocked != NotBlocked);
2264 tso->why_blocked = NotBlocked;
2266 PUSH_ON_RUN_QUEUE(tso);
2268 IF_DEBUG(scheduler,sched_belch("waking up thread %ld", tso->id));
2273 #if defined(GRAN) || defined(PAR)
2274 inline StgBlockingQueueElement *
2275 unblockOne(StgBlockingQueueElement *bqe, StgClosure *node)
2277 ACQUIRE_LOCK(&sched_mutex);
2278 bqe = unblockOneLocked(bqe, node);
2279 RELEASE_LOCK(&sched_mutex);
2284 unblockOne(StgTSO *tso)
2286 ACQUIRE_LOCK(&sched_mutex);
2287 tso = unblockOneLocked(tso);
2288 RELEASE_LOCK(&sched_mutex);
2295 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2297 StgBlockingQueueElement *bqe;
2302 belch("## AwBQ for node %p on PE %d @ %ld by TSO %d (%p): ", \
2303 node, CurrentProc, CurrentTime[CurrentProc],
2304 CurrentTSO->id, CurrentTSO));
2306 node_loc = where_is(node);
2308 ASSERT(get_itbl(q)->type == TSO || // q is either a TSO or an RBHSave
2309 get_itbl(q)->type == CONSTR); // closure (type constructor)
2310 ASSERT(is_unique(node));
2312 /* FAKE FETCH: magically copy the node to the tso's proc;
2313 no Fetch necessary because in reality the node should not have been
2314 moved to the other PE in the first place
2316 if (CurrentProc!=node_loc) {
2318 belch("## node %p is on PE %d but CurrentProc is %d (TSO %d); assuming fake fetch and adjusting bitmask (old: %#x)",
2319 node, node_loc, CurrentProc, CurrentTSO->id,
2320 // CurrentTSO, where_is(CurrentTSO),
2321 node->header.gran.procs));
2322 node->header.gran.procs = (node->header.gran.procs) | PE_NUMBER(CurrentProc);
2324 belch("## new bitmask of node %p is %#x",
2325 node, node->header.gran.procs));
2326 if (RtsFlags.GranFlags.GranSimStats.Global) {
2327 globalGranStats.tot_fake_fetches++;
2332 // ToDo: check: ASSERT(CurrentProc==node_loc);
2333 while (get_itbl(bqe)->type==TSO) { // q != END_TSO_QUEUE) {
2336 bqe points to the current element in the queue
2337 next points to the next element in the queue
2339 //tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2340 //tso_loc = where_is(tso);
2342 bqe = unblockOneLocked(bqe, node);
2345 /* if this is the BQ of an RBH, we have to put back the info ripped out of
2346 the closure to make room for the anchor of the BQ */
2347 if (bqe!=END_BQ_QUEUE) {
2348 ASSERT(get_itbl(node)->type == RBH && get_itbl(bqe)->type == CONSTR);
2350 ASSERT((info_ptr==&RBH_Save_0_info) ||
2351 (info_ptr==&RBH_Save_1_info) ||
2352 (info_ptr==&RBH_Save_2_info));
2354 /* cf. convertToRBH in RBH.c for writing the RBHSave closure */
2355 ((StgRBH *)node)->blocking_queue = (StgBlockingQueueElement *)((StgRBHSave *)bqe)->payload[0];
2356 ((StgRBH *)node)->mut_link = (StgMutClosure *)((StgRBHSave *)bqe)->payload[1];
2359 belch("## Filled in RBH_Save for %p (%s) at end of AwBQ",
2360 node, info_type(node)));
2363 /* statistics gathering */
2364 if (RtsFlags.GranFlags.GranSimStats.Global) {
2365 // globalGranStats.tot_bq_processing_time += bq_processing_time;
2366 globalGranStats.tot_bq_len += len; // total length of all bqs awakened
2367 // globalGranStats.tot_bq_len_local += len_local; // same for local TSOs only
2368 globalGranStats.tot_awbq++; // total no. of bqs awakened
2371 fprintf(stderr,"## BQ Stats of %p: [%d entries] %s\n",
2372 node, len, (bqe!=END_BQ_QUEUE) ? "RBH" : ""));
2376 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2378 StgBlockingQueueElement *bqe, *next;
2380 ACQUIRE_LOCK(&sched_mutex);
2382 IF_PAR_DEBUG(verbose,
2383 belch("## AwBQ for node %p on [%x]: ",
2386 ASSERT(get_itbl(q)->type == TSO ||
2387 get_itbl(q)->type == BLOCKED_FETCH ||
2388 get_itbl(q)->type == CONSTR);
2391 while (get_itbl(bqe)->type==TSO ||
2392 get_itbl(bqe)->type==BLOCKED_FETCH) {
2393 bqe = unblockOneLocked(bqe, node);
2395 RELEASE_LOCK(&sched_mutex);
2398 #else /* !GRAN && !PAR */
2400 awakenBlockedQueue(StgTSO *tso)
2402 ACQUIRE_LOCK(&sched_mutex);
2403 while (tso != END_TSO_QUEUE) {
2404 tso = unblockOneLocked(tso);
2406 RELEASE_LOCK(&sched_mutex);
2410 //@node Exception Handling Routines, Debugging Routines, Blocking Queue Routines, Main scheduling code
2411 //@subsection Exception Handling Routines
2413 /* ---------------------------------------------------------------------------
2415 - usually called inside a signal handler so it mustn't do anything fancy.
2416 ------------------------------------------------------------------------ */
2419 interruptStgRts(void)
2425 /* -----------------------------------------------------------------------------
2428 This is for use when we raise an exception in another thread, which
2430 This has nothing to do with the UnblockThread event in GranSim. -- HWL
2431 -------------------------------------------------------------------------- */
2433 #if defined(GRAN) || defined(PAR)
2435 NB: only the type of the blocking queue is different in GranSim and GUM
2436 the operations on the queue-elements are the same
2437 long live polymorphism!
2440 unblockThread(StgTSO *tso)
2442 StgBlockingQueueElement *t, **last;
2444 ACQUIRE_LOCK(&sched_mutex);
2445 switch (tso->why_blocked) {
2448 return; /* not blocked */
2451 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
2453 StgBlockingQueueElement *last_tso = END_BQ_QUEUE;
2454 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
2456 last = (StgBlockingQueueElement **)&mvar->head;
2457 for (t = (StgBlockingQueueElement *)mvar->head;
2459 last = &t->link, last_tso = t, t = t->link) {
2460 if (t == (StgBlockingQueueElement *)tso) {
2461 *last = (StgBlockingQueueElement *)tso->link;
2462 if (mvar->tail == tso) {
2463 mvar->tail = (StgTSO *)last_tso;
2468 barf("unblockThread (MVAR): TSO not found");
2471 case BlockedOnBlackHole:
2472 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
2474 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
2476 last = &bq->blocking_queue;
2477 for (t = bq->blocking_queue;
2479 last = &t->link, t = t->link) {
2480 if (t == (StgBlockingQueueElement *)tso) {
2481 *last = (StgBlockingQueueElement *)tso->link;
2485 barf("unblockThread (BLACKHOLE): TSO not found");
2488 case BlockedOnException:
2490 StgTSO *target = tso->block_info.tso;
2492 ASSERT(get_itbl(target)->type == TSO);
2493 ASSERT(target->blocked_exceptions != NULL);
2495 last = (StgBlockingQueueElement **)&target->blocked_exceptions;
2496 for (t = (StgBlockingQueueElement *)target->blocked_exceptions;
2498 last = &t->link, t = t->link) {
2499 ASSERT(get_itbl(t)->type == TSO);
2500 if (t == (StgBlockingQueueElement *)tso) {
2501 *last = (StgBlockingQueueElement *)tso->link;
2505 barf("unblockThread (Exception): TSO not found");
2509 case BlockedOnWrite:
2511 StgBlockingQueueElement *prev = NULL;
2512 for (t = (StgBlockingQueueElement *)blocked_queue_hd; t != END_BQ_QUEUE;
2513 prev = t, t = t->link) {
2514 if (t == (StgBlockingQueueElement *)tso) {
2516 blocked_queue_hd = (StgTSO *)t->link;
2517 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
2518 blocked_queue_tl = END_TSO_QUEUE;
2521 prev->link = t->link;
2522 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
2523 blocked_queue_tl = (StgTSO *)prev;
2529 barf("unblockThread (I/O): TSO not found");
2532 case BlockedOnDelay:
2534 StgBlockingQueueElement *prev = NULL;
2535 for (t = (StgBlockingQueueElement *)sleeping_queue; t != END_BQ_QUEUE;
2536 prev = t, t = t->link) {
2537 if (t == (StgBlockingQueueElement *)tso) {
2539 sleeping_queue = (StgTSO *)t->link;
2541 prev->link = t->link;
2546 barf("unblockThread (I/O): TSO not found");
2550 barf("unblockThread");
2554 tso->link = END_TSO_QUEUE;
2555 tso->why_blocked = NotBlocked;
2556 tso->block_info.closure = NULL;
2557 PUSH_ON_RUN_QUEUE(tso);
2558 RELEASE_LOCK(&sched_mutex);
2562 unblockThread(StgTSO *tso)
2566 ACQUIRE_LOCK(&sched_mutex);
2567 switch (tso->why_blocked) {
2570 return; /* not blocked */
2573 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
2575 StgTSO *last_tso = END_TSO_QUEUE;
2576 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
2579 for (t = mvar->head; t != END_TSO_QUEUE;
2580 last = &t->link, last_tso = t, t = t->link) {
2583 if (mvar->tail == tso) {
2584 mvar->tail = last_tso;
2589 barf("unblockThread (MVAR): TSO not found");
2592 case BlockedOnBlackHole:
2593 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
2595 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
2597 last = &bq->blocking_queue;
2598 for (t = bq->blocking_queue; t != END_TSO_QUEUE;
2599 last = &t->link, t = t->link) {
2605 barf("unblockThread (BLACKHOLE): TSO not found");
2608 case BlockedOnException:
2610 StgTSO *target = tso->block_info.tso;
2612 ASSERT(get_itbl(target)->type == TSO);
2613 ASSERT(target->blocked_exceptions != NULL);
2615 last = &target->blocked_exceptions;
2616 for (t = target->blocked_exceptions; t != END_TSO_QUEUE;
2617 last = &t->link, t = t->link) {
2618 ASSERT(get_itbl(t)->type == TSO);
2624 barf("unblockThread (Exception): TSO not found");
2628 case BlockedOnWrite:
2630 StgTSO *prev = NULL;
2631 for (t = blocked_queue_hd; t != END_TSO_QUEUE;
2632 prev = t, t = t->link) {
2635 blocked_queue_hd = t->link;
2636 if (blocked_queue_tl == t) {
2637 blocked_queue_tl = END_TSO_QUEUE;
2640 prev->link = t->link;
2641 if (blocked_queue_tl == t) {
2642 blocked_queue_tl = prev;
2648 barf("unblockThread (I/O): TSO not found");
2651 case BlockedOnDelay:
2653 StgTSO *prev = NULL;
2654 for (t = sleeping_queue; t != END_TSO_QUEUE;
2655 prev = t, t = t->link) {
2658 sleeping_queue = t->link;
2660 prev->link = t->link;
2665 barf("unblockThread (I/O): TSO not found");
2669 barf("unblockThread");
2673 tso->link = END_TSO_QUEUE;
2674 tso->why_blocked = NotBlocked;
2675 tso->block_info.closure = NULL;
2676 PUSH_ON_RUN_QUEUE(tso);
2677 RELEASE_LOCK(&sched_mutex);
2681 /* -----------------------------------------------------------------------------
2684 * The following function implements the magic for raising an
2685 * asynchronous exception in an existing thread.
2687 * We first remove the thread from any queue on which it might be
2688 * blocked. The possible blockages are MVARs and BLACKHOLE_BQs.
2690 * We strip the stack down to the innermost CATCH_FRAME, building
2691 * thunks in the heap for all the active computations, so they can
2692 * be restarted if necessary. When we reach a CATCH_FRAME, we build
2693 * an application of the handler to the exception, and push it on
2694 * the top of the stack.
2696 * How exactly do we save all the active computations? We create an
2697 * AP_UPD for every UpdateFrame on the stack. Entering one of these
2698 * AP_UPDs pushes everything from the corresponding update frame
2699 * upwards onto the stack. (Actually, it pushes everything up to the
2700 * next update frame plus a pointer to the next AP_UPD object.
2701 * Entering the next AP_UPD object pushes more onto the stack until we
2702 * reach the last AP_UPD object - at which point the stack should look
2703 * exactly as it did when we killed the TSO and we can continue
2704 * execution by entering the closure on top of the stack.
2706 * We can also kill a thread entirely - this happens if either (a) the
2707 * exception passed to raiseAsync is NULL, or (b) there's no
2708 * CATCH_FRAME on the stack. In either case, we strip the entire
2709 * stack and replace the thread with a zombie.
2711 * -------------------------------------------------------------------------- */
2714 deleteThread(StgTSO *tso)
2716 raiseAsync(tso,NULL);
2720 raiseAsync(StgTSO *tso, StgClosure *exception)
2722 StgUpdateFrame* su = tso->su;
2723 StgPtr sp = tso->sp;
2725 /* Thread already dead? */
2726 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
2730 IF_DEBUG(scheduler, sched_belch("raising exception in thread %ld.", tso->id));
2732 /* Remove it from any blocking queues */
2735 /* The stack freezing code assumes there's a closure pointer on
2736 * the top of the stack. This isn't always the case with compiled
2737 * code, so we have to push a dummy closure on the top which just
2738 * returns to the next return address on the stack.
2740 if ( LOOKS_LIKE_GHC_INFO((void*)*sp) ) {
2741 *(--sp) = (W_)&stg_dummy_ret_closure;
2745 int words = ((P_)su - (P_)sp) - 1;
2749 /* If we find a CATCH_FRAME, and we've got an exception to raise,
2750 * then build PAP(handler,exception,realworld#), and leave it on
2751 * top of the stack ready to enter.
2753 if (get_itbl(su)->type == CATCH_FRAME && exception != NULL) {
2754 StgCatchFrame *cf = (StgCatchFrame *)su;
2755 /* we've got an exception to raise, so let's pass it to the
2756 * handler in this frame.
2758 ap = (StgAP_UPD *)allocate(sizeofW(StgPAP) + 2);
2759 TICK_ALLOC_UPD_PAP(3,0);
2760 SET_HDR(ap,&stg_PAP_info,cf->header.prof.ccs);
2763 ap->fun = cf->handler; /* :: Exception -> IO a */
2764 ap->payload[0] = exception;
2765 ap->payload[1] = ARG_TAG(0); /* realworld token */
2767 /* throw away the stack from Sp up to and including the
2770 sp = (P_)su + sizeofW(StgCatchFrame) - 1;
2773 /* Restore the blocked/unblocked state for asynchronous exceptions
2774 * at the CATCH_FRAME.
2776 * If exceptions were unblocked at the catch, arrange that they
2777 * are unblocked again after executing the handler by pushing an
2778 * unblockAsyncExceptions_ret stack frame.
2780 if (!cf->exceptions_blocked) {
2781 *(sp--) = (W_)&stg_unblockAsyncExceptionszh_ret_info;
2784 /* Ensure that async exceptions are blocked when running the handler.
2786 if (tso->blocked_exceptions == NULL) {
2787 tso->blocked_exceptions = END_TSO_QUEUE;
2790 /* Put the newly-built PAP on top of the stack, ready to execute
2791 * when the thread restarts.
2795 tso->what_next = ThreadEnterGHC;
2796 IF_DEBUG(sanity, checkTSO(tso));
2800 /* First build an AP_UPD consisting of the stack chunk above the
2801 * current update frame, with the top word on the stack as the
2804 ap = (StgAP_UPD *)allocate(AP_sizeW(words));
2809 ap->fun = (StgClosure *)sp[0];
2811 for(i=0; i < (nat)words; ++i) {
2812 ap->payload[i] = (StgClosure *)*sp++;
2815 switch (get_itbl(su)->type) {
2819 SET_HDR(ap,&stg_AP_UPD_info,su->header.prof.ccs /* ToDo */);
2820 TICK_ALLOC_UP_THK(words+1,0);
2823 fprintf(stderr, "scheduler: Updating ");
2824 printPtr((P_)su->updatee);
2825 fprintf(stderr, " with ");
2826 printObj((StgClosure *)ap);
2829 /* Replace the updatee with an indirection - happily
2830 * this will also wake up any threads currently
2831 * waiting on the result.
2833 UPD_IND_NOLOCK(su->updatee,ap); /* revert the black hole */
2835 sp += sizeofW(StgUpdateFrame) -1;
2836 sp[0] = (W_)ap; /* push onto stack */
2842 StgCatchFrame *cf = (StgCatchFrame *)su;
2845 /* We want a PAP, not an AP_UPD. Fortunately, the
2846 * layout's the same.
2848 SET_HDR(ap,&stg_PAP_info,su->header.prof.ccs /* ToDo */);
2849 TICK_ALLOC_UPD_PAP(words+1,0);
2851 /* now build o = FUN(catch,ap,handler) */
2852 o = (StgClosure *)allocate(sizeofW(StgClosure)+2);
2853 TICK_ALLOC_FUN(2,0);
2854 SET_HDR(o,&stg_catch_info,su->header.prof.ccs /* ToDo */);
2855 o->payload[0] = (StgClosure *)ap;
2856 o->payload[1] = cf->handler;
2859 fprintf(stderr, "scheduler: Built ");
2860 printObj((StgClosure *)o);
2863 /* pop the old handler and put o on the stack */
2865 sp += sizeofW(StgCatchFrame) - 1;
2872 StgSeqFrame *sf = (StgSeqFrame *)su;
2875 SET_HDR(ap,&stg_PAP_info,su->header.prof.ccs /* ToDo */);
2876 TICK_ALLOC_UPD_PAP(words+1,0);
2878 /* now build o = FUN(seq,ap) */
2879 o = (StgClosure *)allocate(sizeofW(StgClosure)+1);
2880 TICK_ALLOC_SE_THK(1,0);
2881 SET_HDR(o,&stg_seq_info,su->header.prof.ccs /* ToDo */);
2882 o->payload[0] = (StgClosure *)ap;
2885 fprintf(stderr, "scheduler: Built ");
2886 printObj((StgClosure *)o);
2889 /* pop the old handler and put o on the stack */
2891 sp += sizeofW(StgSeqFrame) - 1;
2897 /* We've stripped the entire stack, the thread is now dead. */
2898 sp += sizeofW(StgStopFrame) - 1;
2899 sp[0] = (W_)exception; /* save the exception */
2900 tso->what_next = ThreadKilled;
2901 tso->su = (StgUpdateFrame *)(sp+1);
2912 /* -----------------------------------------------------------------------------
2913 resurrectThreads is called after garbage collection on the list of
2914 threads found to be garbage. Each of these threads will be woken
2915 up and sent a signal: BlockedOnDeadMVar if the thread was blocked
2916 on an MVar, or NonTermination if the thread was blocked on a Black
2918 -------------------------------------------------------------------------- */
2921 resurrectThreads( StgTSO *threads )
2925 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
2926 next = tso->global_link;
2927 tso->global_link = all_threads;
2929 IF_DEBUG(scheduler, sched_belch("resurrecting thread %d", tso->id));
2931 switch (tso->why_blocked) {
2933 case BlockedOnException:
2934 raiseAsync(tso,(StgClosure *)BlockedOnDeadMVar_closure);
2936 case BlockedOnBlackHole:
2937 raiseAsync(tso,(StgClosure *)NonTermination_closure);
2940 /* This might happen if the thread was blocked on a black hole
2941 * belonging to a thread that we've just woken up (raiseAsync
2942 * can wake up threads, remember...).
2946 barf("resurrectThreads: thread blocked in a strange way");
2951 /* -----------------------------------------------------------------------------
2952 * Blackhole detection: if we reach a deadlock, test whether any
2953 * threads are blocked on themselves. Any threads which are found to
2954 * be self-blocked get sent a NonTermination exception.
2956 * This is only done in a deadlock situation in order to avoid
2957 * performance overhead in the normal case.
2958 * -------------------------------------------------------------------------- */
2961 detectBlackHoles( void )
2963 StgTSO *t = all_threads;
2964 StgUpdateFrame *frame;
2965 StgClosure *blocked_on;
2967 for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
2969 if (t->why_blocked != BlockedOnBlackHole) {
2973 blocked_on = t->block_info.closure;
2975 for (frame = t->su; ; frame = frame->link) {
2976 switch (get_itbl(frame)->type) {
2979 if (frame->updatee == blocked_on) {
2980 /* We are blocking on one of our own computations, so
2981 * send this thread the NonTermination exception.
2984 sched_belch("thread %d is blocked on itself", t->id));
2985 raiseAsync(t, (StgClosure *)NonTermination_closure);
3006 //@node Debugging Routines, Index, Exception Handling Routines, Main scheduling code
3007 //@subsection Debugging Routines
3009 /* -----------------------------------------------------------------------------
3010 Debugging: why is a thread blocked
3011 -------------------------------------------------------------------------- */
3016 printThreadBlockage(StgTSO *tso)
3018 switch (tso->why_blocked) {
3020 fprintf(stderr,"is blocked on read from fd %d", tso->block_info.fd);
3022 case BlockedOnWrite:
3023 fprintf(stderr,"is blocked on write to fd %d", tso->block_info.fd);
3025 case BlockedOnDelay:
3026 fprintf(stderr,"is blocked until %d", tso->block_info.target);
3029 fprintf(stderr,"is blocked on an MVar");
3031 case BlockedOnException:
3032 fprintf(stderr,"is blocked on delivering an exception to thread %d",
3033 tso->block_info.tso->id);
3035 case BlockedOnBlackHole:
3036 fprintf(stderr,"is blocked on a black hole");
3039 fprintf(stderr,"is not blocked");
3043 fprintf(stderr,"is blocked on global address; local FM_BQ is %p (%s)",
3044 tso->block_info.closure, info_type(tso->block_info.closure));
3046 case BlockedOnGA_NoSend:
3047 fprintf(stderr,"is blocked on global address (no send); local FM_BQ is %p (%s)",
3048 tso->block_info.closure, info_type(tso->block_info.closure));
3052 barf("printThreadBlockage: strange tso->why_blocked: %d for TSO %d (%d)",
3053 tso->why_blocked, tso->id, tso);
3058 printThreadStatus(StgTSO *tso)
3060 switch (tso->what_next) {
3062 fprintf(stderr,"has been killed");
3064 case ThreadComplete:
3065 fprintf(stderr,"has completed");
3068 printThreadBlockage(tso);
3073 printAllThreads(void)
3077 sched_belch("all threads:");
3078 for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
3079 fprintf(stderr, "\tthread %d ", t->id);
3080 printThreadStatus(t);
3081 fprintf(stderr,"\n");
3086 Print a whole blocking queue attached to node (debugging only).
3091 print_bq (StgClosure *node)
3093 StgBlockingQueueElement *bqe;
3097 fprintf(stderr,"## BQ of closure %p (%s): ",
3098 node, info_type(node));
3100 /* should cover all closures that may have a blocking queue */
3101 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
3102 get_itbl(node)->type == FETCH_ME_BQ ||
3103 get_itbl(node)->type == RBH);
3105 ASSERT(node!=(StgClosure*)NULL); // sanity check
3107 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
3109 for (bqe = ((StgBlockingQueue*)node)->blocking_queue, end = (bqe==END_BQ_QUEUE);
3110 !end; // iterate until bqe points to a CONSTR
3111 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) {
3112 ASSERT(bqe != END_BQ_QUEUE); // sanity check
3113 ASSERT(bqe != (StgTSO*)NULL); // sanity check
3114 /* types of closures that may appear in a blocking queue */
3115 ASSERT(get_itbl(bqe)->type == TSO ||
3116 get_itbl(bqe)->type == BLOCKED_FETCH ||
3117 get_itbl(bqe)->type == CONSTR);
3118 /* only BQs of an RBH end with an RBH_Save closure */
3119 ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
3121 switch (get_itbl(bqe)->type) {
3123 fprintf(stderr," TSO %d (%x),",
3124 ((StgTSO *)bqe)->id, ((StgTSO *)bqe));
3127 fprintf(stderr," BF (node=%p, ga=((%x, %d, %x)),",
3128 ((StgBlockedFetch *)bqe)->node,
3129 ((StgBlockedFetch *)bqe)->ga.payload.gc.gtid,
3130 ((StgBlockedFetch *)bqe)->ga.payload.gc.slot,
3131 ((StgBlockedFetch *)bqe)->ga.weight);
3134 fprintf(stderr," %s (IP %p),",
3135 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
3136 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
3137 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
3138 "RBH_Save_?"), get_itbl(bqe));
3141 barf("Unexpected closure type %s in blocking queue of %p (%s)",
3142 info_type(bqe), node, info_type(node));
3146 fputc('\n', stderr);
3148 # elif defined(GRAN)
3150 print_bq (StgClosure *node)
3152 StgBlockingQueueElement *bqe;
3153 PEs node_loc, tso_loc;
3156 /* should cover all closures that may have a blocking queue */
3157 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
3158 get_itbl(node)->type == FETCH_ME_BQ ||
3159 get_itbl(node)->type == RBH);
3161 ASSERT(node!=(StgClosure*)NULL); // sanity check
3162 node_loc = where_is(node);
3164 fprintf(stderr,"## BQ of closure %p (%s) on [PE %d]: ",
3165 node, info_type(node), node_loc);
3168 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
3170 for (bqe = ((StgBlockingQueue*)node)->blocking_queue, end = (bqe==END_BQ_QUEUE);
3171 !end; // iterate until bqe points to a CONSTR
3172 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) {
3173 ASSERT(bqe != END_BQ_QUEUE); // sanity check
3174 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
3175 /* types of closures that may appear in a blocking queue */
3176 ASSERT(get_itbl(bqe)->type == TSO ||
3177 get_itbl(bqe)->type == CONSTR);
3178 /* only BQs of an RBH end with an RBH_Save closure */
3179 ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
3181 tso_loc = where_is((StgClosure *)bqe);
3182 switch (get_itbl(bqe)->type) {
3184 fprintf(stderr," TSO %d (%p) on [PE %d],",
3185 ((StgTSO *)bqe)->id, (StgTSO *)bqe, tso_loc);
3188 fprintf(stderr," %s (IP %p),",
3189 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
3190 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
3191 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
3192 "RBH_Save_?"), get_itbl(bqe));
3195 barf("Unexpected closure type %s in blocking queue of %p (%s)",
3196 info_type((StgClosure *)bqe), node, info_type(node));
3200 fputc('\n', stderr);
3204 Nice and easy: only TSOs on the blocking queue
3207 print_bq (StgClosure *node)
3211 ASSERT(node!=(StgClosure*)NULL); // sanity check
3212 for (tso = ((StgBlockingQueue*)node)->blocking_queue;
3213 tso != END_TSO_QUEUE;
3215 ASSERT(tso!=NULL && tso!=END_TSO_QUEUE); // sanity check
3216 ASSERT(get_itbl(tso)->type == TSO); // guess what, sanity check
3217 fprintf(stderr," TSO %d (%p),", tso->id, tso);
3219 fputc('\n', stderr);
3230 for (i=0, tso=run_queue_hd;
3231 tso != END_TSO_QUEUE;
3240 sched_belch(char *s, ...)
3245 fprintf(stderr, "scheduler (task %ld): ", pthread_self());
3247 fprintf(stderr, "scheduler: ");
3249 vfprintf(stderr, s, ap);
3250 fprintf(stderr, "\n");
3256 //@node Index, , Debugging Routines, Main scheduling code
3260 //* MainRegTable:: @cindex\s-+MainRegTable
3261 //* StgMainThread:: @cindex\s-+StgMainThread
3262 //* awaken_blocked_queue:: @cindex\s-+awaken_blocked_queue
3263 //* blocked_queue_hd:: @cindex\s-+blocked_queue_hd
3264 //* blocked_queue_tl:: @cindex\s-+blocked_queue_tl
3265 //* context_switch:: @cindex\s-+context_switch
3266 //* createThread:: @cindex\s-+createThread
3267 //* free_capabilities:: @cindex\s-+free_capabilities
3268 //* gc_pending_cond:: @cindex\s-+gc_pending_cond
3269 //* initScheduler:: @cindex\s-+initScheduler
3270 //* interrupted:: @cindex\s-+interrupted
3271 //* n_free_capabilities:: @cindex\s-+n_free_capabilities
3272 //* next_thread_id:: @cindex\s-+next_thread_id
3273 //* print_bq:: @cindex\s-+print_bq
3274 //* run_queue_hd:: @cindex\s-+run_queue_hd
3275 //* run_queue_tl:: @cindex\s-+run_queue_tl
3276 //* sched_mutex:: @cindex\s-+sched_mutex
3277 //* schedule:: @cindex\s-+schedule
3278 //* take_off_run_queue:: @cindex\s-+take_off_run_queue
3279 //* task_ids:: @cindex\s-+task_ids
3280 //* term_mutex:: @cindex\s-+term_mutex
3281 //* thread_ready_cond:: @cindex\s-+thread_ready_cond