1 /* ---------------------------------------------------------------------------
2 * $Id: Schedule.c,v 1.89 2001/02/09 13:09:16 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:
892 IF_DEBUG(scheduler,sched_belch("entering interpreter"));
893 ret = interpretBCO(cap);
897 barf("Panic: entered a BCO but no bytecode interpreter in this build");
900 barf("schedule: invalid what_next field");
902 /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
904 /* Costs for the scheduler are assigned to CCS_SYSTEM */
909 ACQUIRE_LOCK(&sched_mutex);
912 IF_DEBUG(scheduler,fprintf(stderr,"scheduler (task %ld): ", pthread_self()););
913 #elif !defined(GRAN) && !defined(PAR)
914 IF_DEBUG(scheduler,fprintf(stderr,"scheduler: "););
916 t = cap->rCurrentTSO;
919 /* HACK 675: if the last thread didn't yield, make sure to print a
920 SCHEDULE event to the log file when StgRunning the next thread, even
921 if it is the same one as before */
922 LastTSO = t; //(ret == ThreadBlocked) ? END_TSO_QUEUE : t;
923 TimeOfLastYield = CURRENT_TIME;
928 /* make all the running tasks block on a condition variable,
929 * maybe set context_switch and wait till they all pile in,
930 * then have them wait on a GC condition variable.
932 IF_DEBUG(scheduler,belch("--<< thread %ld (%p; %s) stopped: HeapOverflow",
933 t->id, t, whatNext_strs[t->what_next]));
936 ASSERT(!is_on_queue(t,CurrentProc));
939 ready_to_gc = rtsTrue;
940 context_switch = 1; /* stop other threads ASAP */
941 PUSH_ON_RUN_QUEUE(t);
942 /* actual GC is done at the end of the while loop */
946 IF_DEBUG(scheduler,belch("--<< thread %ld (%p; %s) stopped, StackOverflow",
947 t->id, t, whatNext_strs[t->what_next]));
948 /* just adjust the stack for this thread, then pop it back
954 /* enlarge the stack */
955 StgTSO *new_t = threadStackOverflow(t);
957 /* This TSO has moved, so update any pointers to it from the
958 * main thread stack. It better not be on any other queues...
961 for (m = main_threads; m != NULL; m = m->link) {
967 PUSH_ON_RUN_QUEUE(new_t);
974 DumpGranEvent(GR_DESCHEDULE, t));
975 globalGranStats.tot_yields++;
978 DumpGranEvent(GR_DESCHEDULE, t));
980 /* put the thread back on the run queue. Then, if we're ready to
981 * GC, check whether this is the last task to stop. If so, wake
982 * up the GC thread. getThread will block during a GC until the
986 if (t->what_next == ThreadEnterInterp) {
987 /* ToDo: or maybe a timer expired when we were in Hugs?
988 * or maybe someone hit ctrl-C
990 belch("--<< thread %ld (%p; %s) stopped to switch to Hugs",
991 t->id, t, whatNext_strs[t->what_next]);
993 belch("--<< thread %ld (%p; %s) stopped, yielding",
994 t->id, t, whatNext_strs[t->what_next]);
1001 //belch("&& Doing sanity check on yielding TSO %ld.", t->id);
1003 ASSERT(t->link == END_TSO_QUEUE);
1005 ASSERT(!is_on_queue(t,CurrentProc));
1008 //belch("&& Doing sanity check on all ThreadQueues (and their TSOs).");
1009 checkThreadQsSanity(rtsTrue));
1011 APPEND_TO_RUN_QUEUE(t);
1013 /* add a ContinueThread event to actually process the thread */
1014 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1016 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1018 belch("GRAN: eventq and runnableq after adding yielded thread to queue again:");
1027 belch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: ",
1028 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)));
1029 if (t->block_info.closure!=(StgClosure*)NULL) print_bq(t->block_info.closure));
1031 // ??? needed; should emit block before
1033 DumpGranEvent(GR_DESCHEDULE, t));
1034 prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
1037 ASSERT(procStatus[CurrentProc]==Busy ||
1038 ((procStatus[CurrentProc]==Fetching) &&
1039 (t->block_info.closure!=(StgClosure*)NULL)));
1040 if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
1041 !(!RtsFlags.GranFlags.DoAsyncFetch &&
1042 procStatus[CurrentProc]==Fetching))
1043 procStatus[CurrentProc] = Idle;
1047 DumpGranEvent(GR_DESCHEDULE, t));
1049 /* Send a fetch (if BlockedOnGA) and dump event to log file */
1053 belch("--<< thread %ld (%p; %s) stopped, blocking on node %p with BQ: ",
1054 t->id, t, whatNext_strs[t->what_next], t->block_info.closure);
1055 if (t->block_info.closure!=(StgClosure*)NULL) print_bq(t->block_info.closure));
1058 /* don't need to do anything. Either the thread is blocked on
1059 * I/O, in which case we'll have called addToBlockedQueue
1060 * previously, or it's blocked on an MVar or Blackhole, in which
1061 * case it'll be on the relevant queue already.
1064 fprintf(stderr, "--<< thread %d (%p) stopped: ", t->id, t);
1065 printThreadBlockage(t);
1066 fprintf(stderr, "\n"));
1068 /* Only for dumping event to log file
1069 ToDo: do I need this in GranSim, too?
1076 case ThreadFinished:
1077 /* Need to check whether this was a main thread, and if so, signal
1078 * the task that started it with the return value. If we have no
1079 * more main threads, we probably need to stop all the tasks until
1082 /* We also end up here if the thread kills itself with an
1083 * uncaught exception, see Exception.hc.
1085 IF_DEBUG(scheduler,belch("--++ thread %d (%p) finished", t->id, t));
1087 endThread(t, CurrentProc); // clean-up the thread
1089 advisory_thread_count--;
1090 if (RtsFlags.ParFlags.ParStats.Full)
1091 DumpEndEvent(CURRENT_PROC, t, rtsFalse /* not mandatory */);
1096 barf("schedule: invalid thread return code %d", (int)ret);
1100 cap->link = free_capabilities;
1101 free_capabilities = cap;
1102 n_free_capabilities++;
1106 if (ready_to_gc && n_free_capabilities == RtsFlags.ParFlags.nNodes)
1111 /* everybody back, start the GC.
1112 * Could do it in this thread, or signal a condition var
1113 * to do it in another thread. Either way, we need to
1114 * broadcast on gc_pending_cond afterward.
1117 IF_DEBUG(scheduler,sched_belch("doing GC"));
1119 GarbageCollect(GetRoots,rtsFalse);
1120 ready_to_gc = rtsFalse;
1122 pthread_cond_broadcast(&gc_pending_cond);
1125 /* add a ContinueThread event to continue execution of current thread */
1126 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1128 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1130 fprintf(stderr, "GRAN: eventq and runnableq after Garbage collection:\n");
1137 IF_GRAN_DEBUG(unused,
1138 print_eventq(EventHd));
1140 event = get_next_event();
1144 /* ToDo: wait for next message to arrive rather than busy wait */
1149 t = take_off_run_queue(END_TSO_QUEUE);
1152 } /* end of while(1) */
1155 /* ---------------------------------------------------------------------------
1156 * deleteAllThreads(): kill all the live threads.
1158 * This is used when we catch a user interrupt (^C), before performing
1159 * any necessary cleanups and running finalizers.
1160 * ------------------------------------------------------------------------- */
1162 void deleteAllThreads ( void )
1165 IF_DEBUG(scheduler,sched_belch("deleting all threads"));
1166 for (t = run_queue_hd; t != END_TSO_QUEUE; t = t->link) {
1169 for (t = blocked_queue_hd; t != END_TSO_QUEUE; t = t->link) {
1172 for (t = sleeping_queue; t != END_TSO_QUEUE; t = t->link) {
1175 run_queue_hd = run_queue_tl = END_TSO_QUEUE;
1176 blocked_queue_hd = blocked_queue_tl = END_TSO_QUEUE;
1177 sleeping_queue = END_TSO_QUEUE;
1180 /* startThread and insertThread are now in GranSim.c -- HWL */
1182 //@node Suspend and Resume, Run queue code, Main scheduling loop, Main scheduling code
1183 //@subsection Suspend and Resume
1185 /* ---------------------------------------------------------------------------
1186 * Suspending & resuming Haskell threads.
1188 * When making a "safe" call to C (aka _ccall_GC), the task gives back
1189 * its capability before calling the C function. This allows another
1190 * task to pick up the capability and carry on running Haskell
1191 * threads. It also means that if the C call blocks, it won't lock
1194 * The Haskell thread making the C call is put to sleep for the
1195 * duration of the call, on the susepended_ccalling_threads queue. We
1196 * give out a token to the task, which it can use to resume the thread
1197 * on return from the C function.
1198 * ------------------------------------------------------------------------- */
1201 suspendThread( Capability *cap )
1205 ACQUIRE_LOCK(&sched_mutex);
1208 sched_belch("thread %d did a _ccall_gc", cap->rCurrentTSO->id));
1210 threadPaused(cap->rCurrentTSO);
1211 cap->rCurrentTSO->link = suspended_ccalling_threads;
1212 suspended_ccalling_threads = cap->rCurrentTSO;
1214 /* Use the thread ID as the token; it should be unique */
1215 tok = cap->rCurrentTSO->id;
1218 cap->link = free_capabilities;
1219 free_capabilities = cap;
1220 n_free_capabilities++;
1223 RELEASE_LOCK(&sched_mutex);
1228 resumeThread( StgInt tok )
1230 StgTSO *tso, **prev;
1233 ACQUIRE_LOCK(&sched_mutex);
1235 prev = &suspended_ccalling_threads;
1236 for (tso = suspended_ccalling_threads;
1237 tso != END_TSO_QUEUE;
1238 prev = &tso->link, tso = tso->link) {
1239 if (tso->id == (StgThreadID)tok) {
1244 if (tso == END_TSO_QUEUE) {
1245 barf("resumeThread: thread not found");
1247 tso->link = END_TSO_QUEUE;
1250 while (free_capabilities == NULL) {
1251 IF_DEBUG(scheduler, sched_belch("waiting to resume"));
1252 pthread_cond_wait(&thread_ready_cond, &sched_mutex);
1253 IF_DEBUG(scheduler, sched_belch("resuming thread %d", tso->id));
1255 cap = free_capabilities;
1256 free_capabilities = cap->link;
1257 n_free_capabilities--;
1259 cap = &MainRegTable;
1262 cap->rCurrentTSO = tso;
1264 RELEASE_LOCK(&sched_mutex);
1269 /* ---------------------------------------------------------------------------
1271 * ------------------------------------------------------------------------ */
1272 static void unblockThread(StgTSO *tso);
1274 /* ---------------------------------------------------------------------------
1275 * Comparing Thread ids.
1277 * This is used from STG land in the implementation of the
1278 * instances of Eq/Ord for ThreadIds.
1279 * ------------------------------------------------------------------------ */
1281 int cmp_thread(const StgTSO *tso1, const StgTSO *tso2)
1283 StgThreadID id1 = tso1->id;
1284 StgThreadID id2 = tso2->id;
1286 if (id1 < id2) return (-1);
1287 if (id1 > id2) return 1;
1291 /* ---------------------------------------------------------------------------
1292 Create a new thread.
1294 The new thread starts with the given stack size. Before the
1295 scheduler can run, however, this thread needs to have a closure
1296 (and possibly some arguments) pushed on its stack. See
1297 pushClosure() in Schedule.h.
1299 createGenThread() and createIOThread() (in SchedAPI.h) are
1300 convenient packaged versions of this function.
1302 currently pri (priority) is only used in a GRAN setup -- HWL
1303 ------------------------------------------------------------------------ */
1304 //@cindex createThread
1306 /* currently pri (priority) is only used in a GRAN setup -- HWL */
1308 createThread(nat stack_size, StgInt pri)
1310 return createThread_(stack_size, rtsFalse, pri);
1314 createThread_(nat size, rtsBool have_lock, StgInt pri)
1318 createThread(nat stack_size)
1320 return createThread_(stack_size, rtsFalse);
1324 createThread_(nat size, rtsBool have_lock)
1331 /* First check whether we should create a thread at all */
1333 /* check that no more than RtsFlags.ParFlags.maxThreads threads are created */
1334 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads) {
1336 belch("{createThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
1337 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
1338 return END_TSO_QUEUE;
1344 ASSERT(!RtsFlags.GranFlags.Light || CurrentProc==0);
1347 // ToDo: check whether size = stack_size - TSO_STRUCT_SIZEW
1349 /* catch ridiculously small stack sizes */
1350 if (size < MIN_STACK_WORDS + TSO_STRUCT_SIZEW) {
1351 size = MIN_STACK_WORDS + TSO_STRUCT_SIZEW;
1354 stack_size = size - TSO_STRUCT_SIZEW;
1356 tso = (StgTSO *)allocate(size);
1357 TICK_ALLOC_TSO(size-TSO_STRUCT_SIZEW, 0);
1359 SET_HDR(tso, &stg_TSO_info, CCS_SYSTEM);
1361 SET_GRAN_HDR(tso, ThisPE);
1363 tso->what_next = ThreadEnterGHC;
1365 /* tso->id needs to be unique. For now we use a heavyweight mutex to
1366 * protect the increment operation on next_thread_id.
1367 * In future, we could use an atomic increment instead.
1369 if (!have_lock) { ACQUIRE_LOCK(&sched_mutex); }
1370 tso->id = next_thread_id++;
1371 if (!have_lock) { RELEASE_LOCK(&sched_mutex); }
1373 tso->why_blocked = NotBlocked;
1374 tso->blocked_exceptions = NULL;
1376 tso->stack_size = stack_size;
1377 tso->max_stack_size = round_to_mblocks(RtsFlags.GcFlags.maxStkSize)
1379 tso->sp = (P_)&(tso->stack) + stack_size;
1382 tso->prof.CCCS = CCS_MAIN;
1385 /* put a stop frame on the stack */
1386 tso->sp -= sizeofW(StgStopFrame);
1387 SET_HDR((StgClosure*)tso->sp,(StgInfoTable *)&stg_stop_thread_info,CCS_SYSTEM);
1388 tso->su = (StgUpdateFrame*)tso->sp;
1392 tso->link = END_TSO_QUEUE;
1393 /* uses more flexible routine in GranSim */
1394 insertThread(tso, CurrentProc);
1396 /* In a non-GranSim setup the pushing of a TSO onto the runq is separated
1401 #if defined(GRAN) || defined(PAR)
1402 DumpGranEvent(GR_START,tso);
1405 /* Link the new thread on the global thread list.
1407 tso->global_link = all_threads;
1411 tso->gran.pri = pri;
1413 tso->gran.magic = TSO_MAGIC; // debugging only
1415 tso->gran.sparkname = 0;
1416 tso->gran.startedat = CURRENT_TIME;
1417 tso->gran.exported = 0;
1418 tso->gran.basicblocks = 0;
1419 tso->gran.allocs = 0;
1420 tso->gran.exectime = 0;
1421 tso->gran.fetchtime = 0;
1422 tso->gran.fetchcount = 0;
1423 tso->gran.blocktime = 0;
1424 tso->gran.blockcount = 0;
1425 tso->gran.blockedat = 0;
1426 tso->gran.globalsparks = 0;
1427 tso->gran.localsparks = 0;
1428 if (RtsFlags.GranFlags.Light)
1429 tso->gran.clock = Now; /* local clock */
1431 tso->gran.clock = 0;
1433 IF_DEBUG(gran,printTSO(tso));
1436 tso->par.magic = TSO_MAGIC; // debugging only
1438 tso->par.sparkname = 0;
1439 tso->par.startedat = CURRENT_TIME;
1440 tso->par.exported = 0;
1441 tso->par.basicblocks = 0;
1442 tso->par.allocs = 0;
1443 tso->par.exectime = 0;
1444 tso->par.fetchtime = 0;
1445 tso->par.fetchcount = 0;
1446 tso->par.blocktime = 0;
1447 tso->par.blockcount = 0;
1448 tso->par.blockedat = 0;
1449 tso->par.globalsparks = 0;
1450 tso->par.localsparks = 0;
1454 globalGranStats.tot_threads_created++;
1455 globalGranStats.threads_created_on_PE[CurrentProc]++;
1456 globalGranStats.tot_sq_len += spark_queue_len(CurrentProc);
1457 globalGranStats.tot_sq_probes++;
1462 belch("==__ schedule: Created TSO %d (%p);",
1463 CurrentProc, tso, tso->id));
1465 IF_PAR_DEBUG(verbose,
1466 belch("==__ schedule: Created TSO %d (%p); %d threads active",
1467 tso->id, tso, advisory_thread_count));
1469 IF_DEBUG(scheduler,sched_belch("created thread %ld, stack size = %lx words",
1470 tso->id, tso->stack_size));
1476 Turn a spark into a thread.
1477 ToDo: fix for SMP (needs to acquire SCHED_MUTEX!)
1480 //@cindex activateSpark
1482 activateSpark (rtsSpark spark)
1486 ASSERT(spark != (rtsSpark)NULL);
1487 tso = createThread_(RtsFlags.GcFlags.initialStkSize, rtsTrue);
1488 if (tso!=END_TSO_QUEUE) {
1489 pushClosure(tso,spark);
1490 PUSH_ON_RUN_QUEUE(tso);
1491 advisory_thread_count++;
1493 if (RtsFlags.ParFlags.ParStats.Full) {
1494 //ASSERT(run_queue_hd == END_TSO_QUEUE); // I think ...
1495 IF_PAR_DEBUG(verbose,
1496 belch("==^^ activateSpark: turning spark of closure %p (%s) into a thread",
1497 (StgClosure *)spark, info_type((StgClosure *)spark)));
1500 barf("activateSpark: Cannot create TSO");
1502 // ToDo: fwd info on local/global spark to thread -- HWL
1503 // tso->gran.exported = spark->exported;
1504 // tso->gran.locked = !spark->global;
1505 // tso->gran.sparkname = spark->name;
1511 /* ---------------------------------------------------------------------------
1514 * scheduleThread puts a thread on the head of the runnable queue.
1515 * This will usually be done immediately after a thread is created.
1516 * The caller of scheduleThread must create the thread using e.g.
1517 * createThread and push an appropriate closure
1518 * on this thread's stack before the scheduler is invoked.
1519 * ------------------------------------------------------------------------ */
1522 scheduleThread(StgTSO *tso)
1524 if (tso==END_TSO_QUEUE){
1529 ACQUIRE_LOCK(&sched_mutex);
1531 /* Put the new thread on the head of the runnable queue. The caller
1532 * better push an appropriate closure on this thread's stack
1533 * beforehand. In the SMP case, the thread may start running as
1534 * soon as we release the scheduler lock below.
1536 PUSH_ON_RUN_QUEUE(tso);
1540 IF_DEBUG(scheduler,printTSO(tso));
1542 RELEASE_LOCK(&sched_mutex);
1545 /* ---------------------------------------------------------------------------
1548 * Start up Posix threads to run each of the scheduler tasks.
1549 * I believe the task ids are not needed in the system as defined.
1551 * ------------------------------------------------------------------------ */
1553 #if defined(PAR) || defined(SMP)
1555 taskStart( void *arg STG_UNUSED )
1557 rts_evalNothing(NULL);
1561 /* ---------------------------------------------------------------------------
1564 * Initialise the scheduler. This resets all the queues - if the
1565 * queues contained any threads, they'll be garbage collected at the
1568 * This now calls startTasks(), so should only be called once! KH @ 25/10/99
1569 * ------------------------------------------------------------------------ */
1573 term_handler(int sig STG_UNUSED)
1576 ACQUIRE_LOCK(&term_mutex);
1578 RELEASE_LOCK(&term_mutex);
1583 //@cindex initScheduler
1590 for (i=0; i<=MAX_PROC; i++) {
1591 run_queue_hds[i] = END_TSO_QUEUE;
1592 run_queue_tls[i] = END_TSO_QUEUE;
1593 blocked_queue_hds[i] = END_TSO_QUEUE;
1594 blocked_queue_tls[i] = END_TSO_QUEUE;
1595 ccalling_threadss[i] = END_TSO_QUEUE;
1596 sleeping_queue = END_TSO_QUEUE;
1599 run_queue_hd = END_TSO_QUEUE;
1600 run_queue_tl = END_TSO_QUEUE;
1601 blocked_queue_hd = END_TSO_QUEUE;
1602 blocked_queue_tl = END_TSO_QUEUE;
1603 sleeping_queue = END_TSO_QUEUE;
1606 suspended_ccalling_threads = END_TSO_QUEUE;
1608 main_threads = NULL;
1609 all_threads = END_TSO_QUEUE;
1614 RtsFlags.ConcFlags.ctxtSwitchTicks =
1615 RtsFlags.ConcFlags.ctxtSwitchTime / TICK_MILLISECS;
1617 /* Install the SIGHUP handler */
1620 struct sigaction action,oact;
1622 action.sa_handler = term_handler;
1623 sigemptyset(&action.sa_mask);
1624 action.sa_flags = 0;
1625 if (sigaction(SIGTERM, &action, &oact) != 0) {
1626 barf("can't install TERM handler");
1632 /* Allocate N Capabilities */
1635 Capability *cap, *prev;
1638 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
1639 cap = stgMallocBytes(sizeof(Capability), "initScheduler:capabilities");
1643 free_capabilities = cap;
1644 n_free_capabilities = RtsFlags.ParFlags.nNodes;
1646 IF_DEBUG(scheduler,fprintf(stderr,"scheduler: Allocated %d capabilities\n",
1647 n_free_capabilities););
1650 #if defined(SMP) || defined(PAR)
1663 /* make some space for saving all the thread ids */
1664 task_ids = stgMallocBytes(RtsFlags.ParFlags.nNodes * sizeof(task_info),
1665 "initScheduler:task_ids");
1667 /* and create all the threads */
1668 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
1669 r = pthread_create(&tid,NULL,taskStart,NULL);
1671 barf("startTasks: Can't create new Posix thread");
1673 task_ids[i].id = tid;
1674 task_ids[i].mut_time = 0.0;
1675 task_ids[i].mut_etime = 0.0;
1676 task_ids[i].gc_time = 0.0;
1677 task_ids[i].gc_etime = 0.0;
1678 task_ids[i].elapsedtimestart = elapsedtime();
1679 IF_DEBUG(scheduler,fprintf(stderr,"scheduler: Started task: %ld\n",tid););
1685 exitScheduler( void )
1690 /* Don't want to use pthread_cancel, since we'd have to install
1691 * these silly exception handlers (pthread_cleanup_{push,pop}) around
1695 /* Cancel all our tasks */
1696 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
1697 pthread_cancel(task_ids[i].id);
1700 /* Wait for all the tasks to terminate */
1701 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
1702 IF_DEBUG(scheduler,fprintf(stderr,"scheduler: waiting for task %ld\n",
1704 pthread_join(task_ids[i].id, NULL);
1708 /* Send 'em all a SIGHUP. That should shut 'em up.
1710 await_death = RtsFlags.ParFlags.nNodes;
1711 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
1712 pthread_kill(task_ids[i].id,SIGTERM);
1714 while (await_death > 0) {
1720 /* -----------------------------------------------------------------------------
1721 Managing the per-task allocation areas.
1723 Each capability comes with an allocation area. These are
1724 fixed-length block lists into which allocation can be done.
1726 ToDo: no support for two-space collection at the moment???
1727 -------------------------------------------------------------------------- */
1729 /* -----------------------------------------------------------------------------
1730 * waitThread is the external interface for running a new computation
1731 * and waiting for the result.
1733 * In the non-SMP case, we create a new main thread, push it on the
1734 * main-thread stack, and invoke the scheduler to run it. The
1735 * scheduler will return when the top main thread on the stack has
1736 * completed or died, and fill in the necessary fields of the
1737 * main_thread structure.
1739 * In the SMP case, we create a main thread as before, but we then
1740 * create a new condition variable and sleep on it. When our new
1741 * main thread has completed, we'll be woken up and the status/result
1742 * will be in the main_thread struct.
1743 * -------------------------------------------------------------------------- */
1746 howManyThreadsAvail ( void )
1750 for (q = run_queue_hd; q != END_TSO_QUEUE; q = q->link)
1752 for (q = blocked_queue_hd; q != END_TSO_QUEUE; q = q->link)
1754 for (q = sleeping_queue; q != END_TSO_QUEUE; q = q->link)
1760 finishAllThreads ( void )
1763 while (run_queue_hd != END_TSO_QUEUE) {
1764 waitThread ( run_queue_hd, NULL );
1766 while (blocked_queue_hd != END_TSO_QUEUE) {
1767 waitThread ( blocked_queue_hd, NULL );
1769 while (sleeping_queue != END_TSO_QUEUE) {
1770 waitThread ( blocked_queue_hd, NULL );
1773 (blocked_queue_hd != END_TSO_QUEUE ||
1774 run_queue_hd != END_TSO_QUEUE ||
1775 sleeping_queue != END_TSO_QUEUE);
1779 waitThread(StgTSO *tso, /*out*/StgClosure **ret)
1782 SchedulerStatus stat;
1784 ACQUIRE_LOCK(&sched_mutex);
1786 m = stgMallocBytes(sizeof(StgMainThread), "waitThread");
1792 pthread_cond_init(&m->wakeup, NULL);
1795 m->link = main_threads;
1798 IF_DEBUG(scheduler, fprintf(stderr, "scheduler: new main thread (%d)\n",
1803 pthread_cond_wait(&m->wakeup, &sched_mutex);
1804 } while (m->stat == NoStatus);
1806 /* GranSim specific init */
1807 CurrentTSO = m->tso; // the TSO to run
1808 procStatus[MainProc] = Busy; // status of main PE
1809 CurrentProc = MainProc; // PE to run it on
1814 ASSERT(m->stat != NoStatus);
1820 pthread_cond_destroy(&m->wakeup);
1823 IF_DEBUG(scheduler, fprintf(stderr, "scheduler: main thread (%d) finished\n",
1827 RELEASE_LOCK(&sched_mutex);
1832 //@node Run queue code, Garbage Collextion Routines, Suspend and Resume, Main scheduling code
1833 //@subsection Run queue code
1837 NB: In GranSim we have many run queues; run_queue_hd is actually a macro
1838 unfolding to run_queue_hds[CurrentProc], thus CurrentProc is an
1839 implicit global variable that has to be correct when calling these
1843 /* Put the new thread on the head of the runnable queue.
1844 * The caller of createThread better push an appropriate closure
1845 * on this thread's stack before the scheduler is invoked.
1847 static /* inline */ void
1848 add_to_run_queue(tso)
1851 ASSERT(tso!=run_queue_hd && tso!=run_queue_tl);
1852 tso->link = run_queue_hd;
1854 if (run_queue_tl == END_TSO_QUEUE) {
1859 /* Put the new thread at the end of the runnable queue. */
1860 static /* inline */ void
1861 push_on_run_queue(tso)
1864 ASSERT(get_itbl((StgClosure *)tso)->type == TSO);
1865 ASSERT(run_queue_hd!=NULL && run_queue_tl!=NULL);
1866 ASSERT(tso!=run_queue_hd && tso!=run_queue_tl);
1867 if (run_queue_hd == END_TSO_QUEUE) {
1870 run_queue_tl->link = tso;
1876 Should be inlined because it's used very often in schedule. The tso
1877 argument is actually only needed in GranSim, where we want to have the
1878 possibility to schedule *any* TSO on the run queue, irrespective of the
1879 actual ordering. Therefore, if tso is not the nil TSO then we traverse
1880 the run queue and dequeue the tso, adjusting the links in the queue.
1882 //@cindex take_off_run_queue
1883 static /* inline */ StgTSO*
1884 take_off_run_queue(StgTSO *tso) {
1888 qetlaHbogh Qu' ngaSbogh ghomDaQ {tso} yIteq!
1890 if tso is specified, unlink that tso from the run_queue (doesn't have
1891 to be at the beginning of the queue); GranSim only
1893 if (tso!=END_TSO_QUEUE) {
1894 /* find tso in queue */
1895 for (t=run_queue_hd, prev=END_TSO_QUEUE;
1896 t!=END_TSO_QUEUE && t!=tso;
1900 /* now actually dequeue the tso */
1901 if (prev!=END_TSO_QUEUE) {
1902 ASSERT(run_queue_hd!=t);
1903 prev->link = t->link;
1905 /* t is at beginning of thread queue */
1906 ASSERT(run_queue_hd==t);
1907 run_queue_hd = t->link;
1909 /* t is at end of thread queue */
1910 if (t->link==END_TSO_QUEUE) {
1911 ASSERT(t==run_queue_tl);
1912 run_queue_tl = prev;
1914 ASSERT(run_queue_tl!=t);
1916 t->link = END_TSO_QUEUE;
1918 /* take tso from the beginning of the queue; std concurrent code */
1920 if (t != END_TSO_QUEUE) {
1921 run_queue_hd = t->link;
1922 t->link = END_TSO_QUEUE;
1923 if (run_queue_hd == END_TSO_QUEUE) {
1924 run_queue_tl = END_TSO_QUEUE;
1933 //@node Garbage Collextion Routines, Blocking Queue Routines, Run queue code, Main scheduling code
1934 //@subsection Garbage Collextion Routines
1936 /* ---------------------------------------------------------------------------
1937 Where are the roots that we know about?
1939 - all the threads on the runnable queue
1940 - all the threads on the blocked queue
1941 - all the threads on the sleeping queue
1942 - all the thread currently executing a _ccall_GC
1943 - all the "main threads"
1945 ------------------------------------------------------------------------ */
1947 /* This has to be protected either by the scheduler monitor, or by the
1948 garbage collection monitor (probably the latter).
1952 static void GetRoots(void)
1959 for (i=0; i<=RtsFlags.GranFlags.proc; i++) {
1960 if ((run_queue_hds[i] != END_TSO_QUEUE) && ((run_queue_hds[i] != NULL)))
1961 run_queue_hds[i] = (StgTSO *)MarkRoot((StgClosure *)run_queue_hds[i]);
1962 if ((run_queue_tls[i] != END_TSO_QUEUE) && ((run_queue_tls[i] != NULL)))
1963 run_queue_tls[i] = (StgTSO *)MarkRoot((StgClosure *)run_queue_tls[i]);
1965 if ((blocked_queue_hds[i] != END_TSO_QUEUE) && ((blocked_queue_hds[i] != NULL)))
1966 blocked_queue_hds[i] = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_hds[i]);
1967 if ((blocked_queue_tls[i] != END_TSO_QUEUE) && ((blocked_queue_tls[i] != NULL)))
1968 blocked_queue_tls[i] = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_tls[i]);
1969 if ((ccalling_threadss[i] != END_TSO_QUEUE) && ((ccalling_threadss[i] != NULL)))
1970 ccalling_threadss[i] = (StgTSO *)MarkRoot((StgClosure *)ccalling_threadss[i]);
1977 if (run_queue_hd != END_TSO_QUEUE) {
1978 ASSERT(run_queue_tl != END_TSO_QUEUE);
1979 run_queue_hd = (StgTSO *)MarkRoot((StgClosure *)run_queue_hd);
1980 run_queue_tl = (StgTSO *)MarkRoot((StgClosure *)run_queue_tl);
1983 if (blocked_queue_hd != END_TSO_QUEUE) {
1984 ASSERT(blocked_queue_tl != END_TSO_QUEUE);
1985 blocked_queue_hd = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_hd);
1986 blocked_queue_tl = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_tl);
1989 if (sleeping_queue != END_TSO_QUEUE) {
1990 sleeping_queue = (StgTSO *)MarkRoot((StgClosure *)sleeping_queue);
1994 for (m = main_threads; m != NULL; m = m->link) {
1995 m->tso = (StgTSO *)MarkRoot((StgClosure *)m->tso);
1997 if (suspended_ccalling_threads != END_TSO_QUEUE)
1998 suspended_ccalling_threads =
1999 (StgTSO *)MarkRoot((StgClosure *)suspended_ccalling_threads);
2001 #if defined(SMP) || defined(PAR) || defined(GRAN)
2006 /* -----------------------------------------------------------------------------
2009 This is the interface to the garbage collector from Haskell land.
2010 We provide this so that external C code can allocate and garbage
2011 collect when called from Haskell via _ccall_GC.
2013 It might be useful to provide an interface whereby the programmer
2014 can specify more roots (ToDo).
2016 This needs to be protected by the GC condition variable above. KH.
2017 -------------------------------------------------------------------------- */
2019 void (*extra_roots)(void);
2024 GarbageCollect(GetRoots,rtsFalse);
2028 performMajorGC(void)
2030 GarbageCollect(GetRoots,rtsTrue);
2036 GetRoots(); /* the scheduler's roots */
2037 extra_roots(); /* the user's roots */
2041 performGCWithRoots(void (*get_roots)(void))
2043 extra_roots = get_roots;
2045 GarbageCollect(AllRoots,rtsFalse);
2048 /* -----------------------------------------------------------------------------
2051 If the thread has reached its maximum stack size, then raise the
2052 StackOverflow exception in the offending thread. Otherwise
2053 relocate the TSO into a larger chunk of memory and adjust its stack
2055 -------------------------------------------------------------------------- */
2058 threadStackOverflow(StgTSO *tso)
2060 nat new_stack_size, new_tso_size, diff, stack_words;
2064 IF_DEBUG(sanity,checkTSO(tso));
2065 if (tso->stack_size >= tso->max_stack_size) {
2068 belch("@@ threadStackOverflow of TSO %d (%p): stack too large (now %ld; max is %ld",
2069 tso->id, tso, tso->stack_size, tso->max_stack_size);
2070 /* If we're debugging, just print out the top of the stack */
2071 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2074 /* Send this thread the StackOverflow exception */
2075 raiseAsync(tso, (StgClosure *)stackOverflow_closure);
2079 /* Try to double the current stack size. If that takes us over the
2080 * maximum stack size for this thread, then use the maximum instead.
2081 * Finally round up so the TSO ends up as a whole number of blocks.
2083 new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
2084 new_tso_size = (nat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) +
2085 TSO_STRUCT_SIZE)/sizeof(W_);
2086 new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */
2087 new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
2089 IF_DEBUG(scheduler, fprintf(stderr,"scheduler: increasing stack size from %d words to %d.\n", tso->stack_size, new_stack_size));
2091 dest = (StgTSO *)allocate(new_tso_size);
2092 TICK_ALLOC_TSO(new_tso_size-sizeofW(StgTSO),0);
2094 /* copy the TSO block and the old stack into the new area */
2095 memcpy(dest,tso,TSO_STRUCT_SIZE);
2096 stack_words = tso->stack + tso->stack_size - tso->sp;
2097 new_sp = (P_)dest + new_tso_size - stack_words;
2098 memcpy(new_sp, tso->sp, stack_words * sizeof(W_));
2100 /* relocate the stack pointers... */
2101 diff = (P_)new_sp - (P_)tso->sp; /* In *words* */
2102 dest->su = (StgUpdateFrame *) ((P_)dest->su + diff);
2104 dest->stack_size = new_stack_size;
2106 /* and relocate the update frame list */
2107 relocate_TSO(tso, dest);
2109 /* Mark the old TSO as relocated. We have to check for relocated
2110 * TSOs in the garbage collector and any primops that deal with TSOs.
2112 * It's important to set the sp and su values to just beyond the end
2113 * of the stack, so we don't attempt to scavenge any part of the
2116 tso->what_next = ThreadRelocated;
2118 tso->sp = (P_)&(tso->stack[tso->stack_size]);
2119 tso->su = (StgUpdateFrame *)tso->sp;
2120 tso->why_blocked = NotBlocked;
2121 dest->mut_link = NULL;
2123 IF_PAR_DEBUG(verbose,
2124 belch("@@ threadStackOverflow of TSO %d (now at %p): stack size increased to %ld",
2125 tso->id, tso, tso->stack_size);
2126 /* If we're debugging, just print out the top of the stack */
2127 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2130 IF_DEBUG(sanity,checkTSO(tso));
2132 IF_DEBUG(scheduler,printTSO(dest));
2138 //@node Blocking Queue Routines, Exception Handling Routines, Garbage Collextion Routines, Main scheduling code
2139 //@subsection Blocking Queue Routines
2141 /* ---------------------------------------------------------------------------
2142 Wake up a queue that was blocked on some resource.
2143 ------------------------------------------------------------------------ */
2147 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2152 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2154 /* write RESUME events to log file and
2155 update blocked and fetch time (depending on type of the orig closure) */
2156 if (RtsFlags.ParFlags.ParStats.Full) {
2157 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
2158 GR_RESUME, ((StgTSO *)bqe), ((StgTSO *)bqe)->block_info.closure,
2159 0, 0 /* spark_queue_len(ADVISORY_POOL) */);
2161 switch (get_itbl(node)->type) {
2163 ((StgTSO *)bqe)->par.fetchtime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2168 ((StgTSO *)bqe)->par.blocktime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2171 barf("{unblockOneLocked}Daq Qagh: unexpected closure in blocking queue");
2178 static StgBlockingQueueElement *
2179 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2182 PEs node_loc, tso_loc;
2184 node_loc = where_is(node); // should be lifted out of loop
2185 tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2186 tso_loc = where_is((StgClosure *)tso);
2187 if (IS_LOCAL_TO(PROCS(node),tso_loc)) { // TSO is local
2188 /* !fake_fetch => TSO is on CurrentProc is same as IS_LOCAL_TO */
2189 ASSERT(CurrentProc!=node_loc || tso_loc==CurrentProc);
2190 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.lunblocktime;
2191 // insertThread(tso, node_loc);
2192 new_event(tso_loc, tso_loc, CurrentTime[CurrentProc],
2194 tso, node, (rtsSpark*)NULL);
2195 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2198 } else { // TSO is remote (actually should be FMBQ)
2199 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.mpacktime +
2200 RtsFlags.GranFlags.Costs.gunblocktime +
2201 RtsFlags.GranFlags.Costs.latency;
2202 new_event(tso_loc, CurrentProc, CurrentTime[CurrentProc],
2204 tso, node, (rtsSpark*)NULL);
2205 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2208 /* the thread-queue-overhead is accounted for in either Resume or UnblockThread */
2210 fprintf(stderr," %s TSO %d (%p) [PE %d] (block_info.closure=%p) (next=%p) ,",
2211 (node_loc==tso_loc ? "Local" : "Global"),
2212 tso->id, tso, CurrentProc, tso->block_info.closure, tso->link));
2213 tso->block_info.closure = NULL;
2214 IF_DEBUG(scheduler,belch("-- Waking up thread %ld (%p)",
2218 static StgBlockingQueueElement *
2219 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2221 StgBlockingQueueElement *next;
2223 switch (get_itbl(bqe)->type) {
2225 ASSERT(((StgTSO *)bqe)->why_blocked != NotBlocked);
2226 /* if it's a TSO just push it onto the run_queue */
2228 // ((StgTSO *)bqe)->link = END_TSO_QUEUE; // debugging?
2229 PUSH_ON_RUN_QUEUE((StgTSO *)bqe);
2231 unblockCount(bqe, node);
2232 /* reset blocking status after dumping event */
2233 ((StgTSO *)bqe)->why_blocked = NotBlocked;
2237 /* if it's a BLOCKED_FETCH put it on the PendingFetches list */
2239 bqe->link = PendingFetches;
2240 PendingFetches = bqe;
2244 /* can ignore this case in a non-debugging setup;
2245 see comments on RBHSave closures above */
2247 /* check that the closure is an RBHSave closure */
2248 ASSERT(get_itbl((StgClosure *)bqe) == &stg_RBH_Save_0_info ||
2249 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_1_info ||
2250 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_2_info);
2254 barf("{unblockOneLocked}Daq Qagh: Unexpected IP (%#lx; %s) in blocking queue at %#lx\n",
2255 get_itbl((StgClosure *)bqe), info_type((StgClosure *)bqe),
2259 // IF_DEBUG(scheduler,sched_belch("waking up thread %ld", tso->id));
2263 #else /* !GRAN && !PAR */
2265 unblockOneLocked(StgTSO *tso)
2269 ASSERT(get_itbl(tso)->type == TSO);
2270 ASSERT(tso->why_blocked != NotBlocked);
2271 tso->why_blocked = NotBlocked;
2273 PUSH_ON_RUN_QUEUE(tso);
2275 IF_DEBUG(scheduler,sched_belch("waking up thread %ld", tso->id));
2280 #if defined(GRAN) || defined(PAR)
2281 inline StgBlockingQueueElement *
2282 unblockOne(StgBlockingQueueElement *bqe, StgClosure *node)
2284 ACQUIRE_LOCK(&sched_mutex);
2285 bqe = unblockOneLocked(bqe, node);
2286 RELEASE_LOCK(&sched_mutex);
2291 unblockOne(StgTSO *tso)
2293 ACQUIRE_LOCK(&sched_mutex);
2294 tso = unblockOneLocked(tso);
2295 RELEASE_LOCK(&sched_mutex);
2302 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2304 StgBlockingQueueElement *bqe;
2309 belch("## AwBQ for node %p on PE %d @ %ld by TSO %d (%p): ", \
2310 node, CurrentProc, CurrentTime[CurrentProc],
2311 CurrentTSO->id, CurrentTSO));
2313 node_loc = where_is(node);
2315 ASSERT(get_itbl(q)->type == TSO || // q is either a TSO or an RBHSave
2316 get_itbl(q)->type == CONSTR); // closure (type constructor)
2317 ASSERT(is_unique(node));
2319 /* FAKE FETCH: magically copy the node to the tso's proc;
2320 no Fetch necessary because in reality the node should not have been
2321 moved to the other PE in the first place
2323 if (CurrentProc!=node_loc) {
2325 belch("## node %p is on PE %d but CurrentProc is %d (TSO %d); assuming fake fetch and adjusting bitmask (old: %#x)",
2326 node, node_loc, CurrentProc, CurrentTSO->id,
2327 // CurrentTSO, where_is(CurrentTSO),
2328 node->header.gran.procs));
2329 node->header.gran.procs = (node->header.gran.procs) | PE_NUMBER(CurrentProc);
2331 belch("## new bitmask of node %p is %#x",
2332 node, node->header.gran.procs));
2333 if (RtsFlags.GranFlags.GranSimStats.Global) {
2334 globalGranStats.tot_fake_fetches++;
2339 // ToDo: check: ASSERT(CurrentProc==node_loc);
2340 while (get_itbl(bqe)->type==TSO) { // q != END_TSO_QUEUE) {
2343 bqe points to the current element in the queue
2344 next points to the next element in the queue
2346 //tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2347 //tso_loc = where_is(tso);
2349 bqe = unblockOneLocked(bqe, node);
2352 /* if this is the BQ of an RBH, we have to put back the info ripped out of
2353 the closure to make room for the anchor of the BQ */
2354 if (bqe!=END_BQ_QUEUE) {
2355 ASSERT(get_itbl(node)->type == RBH && get_itbl(bqe)->type == CONSTR);
2357 ASSERT((info_ptr==&RBH_Save_0_info) ||
2358 (info_ptr==&RBH_Save_1_info) ||
2359 (info_ptr==&RBH_Save_2_info));
2361 /* cf. convertToRBH in RBH.c for writing the RBHSave closure */
2362 ((StgRBH *)node)->blocking_queue = (StgBlockingQueueElement *)((StgRBHSave *)bqe)->payload[0];
2363 ((StgRBH *)node)->mut_link = (StgMutClosure *)((StgRBHSave *)bqe)->payload[1];
2366 belch("## Filled in RBH_Save for %p (%s) at end of AwBQ",
2367 node, info_type(node)));
2370 /* statistics gathering */
2371 if (RtsFlags.GranFlags.GranSimStats.Global) {
2372 // globalGranStats.tot_bq_processing_time += bq_processing_time;
2373 globalGranStats.tot_bq_len += len; // total length of all bqs awakened
2374 // globalGranStats.tot_bq_len_local += len_local; // same for local TSOs only
2375 globalGranStats.tot_awbq++; // total no. of bqs awakened
2378 fprintf(stderr,"## BQ Stats of %p: [%d entries] %s\n",
2379 node, len, (bqe!=END_BQ_QUEUE) ? "RBH" : ""));
2383 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2385 StgBlockingQueueElement *bqe, *next;
2387 ACQUIRE_LOCK(&sched_mutex);
2389 IF_PAR_DEBUG(verbose,
2390 belch("## AwBQ for node %p on [%x]: ",
2393 ASSERT(get_itbl(q)->type == TSO ||
2394 get_itbl(q)->type == BLOCKED_FETCH ||
2395 get_itbl(q)->type == CONSTR);
2398 while (get_itbl(bqe)->type==TSO ||
2399 get_itbl(bqe)->type==BLOCKED_FETCH) {
2400 bqe = unblockOneLocked(bqe, node);
2402 RELEASE_LOCK(&sched_mutex);
2405 #else /* !GRAN && !PAR */
2407 awakenBlockedQueue(StgTSO *tso)
2409 ACQUIRE_LOCK(&sched_mutex);
2410 while (tso != END_TSO_QUEUE) {
2411 tso = unblockOneLocked(tso);
2413 RELEASE_LOCK(&sched_mutex);
2417 //@node Exception Handling Routines, Debugging Routines, Blocking Queue Routines, Main scheduling code
2418 //@subsection Exception Handling Routines
2420 /* ---------------------------------------------------------------------------
2422 - usually called inside a signal handler so it mustn't do anything fancy.
2423 ------------------------------------------------------------------------ */
2426 interruptStgRts(void)
2432 /* -----------------------------------------------------------------------------
2435 This is for use when we raise an exception in another thread, which
2437 This has nothing to do with the UnblockThread event in GranSim. -- HWL
2438 -------------------------------------------------------------------------- */
2440 #if defined(GRAN) || defined(PAR)
2442 NB: only the type of the blocking queue is different in GranSim and GUM
2443 the operations on the queue-elements are the same
2444 long live polymorphism!
2447 unblockThread(StgTSO *tso)
2449 StgBlockingQueueElement *t, **last;
2451 ACQUIRE_LOCK(&sched_mutex);
2452 switch (tso->why_blocked) {
2455 return; /* not blocked */
2458 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
2460 StgBlockingQueueElement *last_tso = END_BQ_QUEUE;
2461 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
2463 last = (StgBlockingQueueElement **)&mvar->head;
2464 for (t = (StgBlockingQueueElement *)mvar->head;
2466 last = &t->link, last_tso = t, t = t->link) {
2467 if (t == (StgBlockingQueueElement *)tso) {
2468 *last = (StgBlockingQueueElement *)tso->link;
2469 if (mvar->tail == tso) {
2470 mvar->tail = (StgTSO *)last_tso;
2475 barf("unblockThread (MVAR): TSO not found");
2478 case BlockedOnBlackHole:
2479 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
2481 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
2483 last = &bq->blocking_queue;
2484 for (t = bq->blocking_queue;
2486 last = &t->link, t = t->link) {
2487 if (t == (StgBlockingQueueElement *)tso) {
2488 *last = (StgBlockingQueueElement *)tso->link;
2492 barf("unblockThread (BLACKHOLE): TSO not found");
2495 case BlockedOnException:
2497 StgTSO *target = tso->block_info.tso;
2499 ASSERT(get_itbl(target)->type == TSO);
2500 ASSERT(target->blocked_exceptions != NULL);
2502 last = (StgBlockingQueueElement **)&target->blocked_exceptions;
2503 for (t = (StgBlockingQueueElement *)target->blocked_exceptions;
2505 last = &t->link, t = t->link) {
2506 ASSERT(get_itbl(t)->type == TSO);
2507 if (t == (StgBlockingQueueElement *)tso) {
2508 *last = (StgBlockingQueueElement *)tso->link;
2512 barf("unblockThread (Exception): TSO not found");
2516 case BlockedOnWrite:
2518 StgBlockingQueueElement *prev = NULL;
2519 for (t = (StgBlockingQueueElement *)blocked_queue_hd; t != END_BQ_QUEUE;
2520 prev = t, t = t->link) {
2521 if (t == (StgBlockingQueueElement *)tso) {
2523 blocked_queue_hd = (StgTSO *)t->link;
2524 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
2525 blocked_queue_tl = END_TSO_QUEUE;
2528 prev->link = t->link;
2529 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
2530 blocked_queue_tl = (StgTSO *)prev;
2536 barf("unblockThread (I/O): TSO not found");
2539 case BlockedOnDelay:
2541 StgBlockingQueueElement *prev = NULL;
2542 for (t = (StgBlockingQueueElement *)sleeping_queue; t != END_BQ_QUEUE;
2543 prev = t, t = t->link) {
2544 if (t == (StgBlockingQueueElement *)tso) {
2546 sleeping_queue = (StgTSO *)t->link;
2548 prev->link = t->link;
2553 barf("unblockThread (I/O): TSO not found");
2557 barf("unblockThread");
2561 tso->link = END_TSO_QUEUE;
2562 tso->why_blocked = NotBlocked;
2563 tso->block_info.closure = NULL;
2564 PUSH_ON_RUN_QUEUE(tso);
2565 RELEASE_LOCK(&sched_mutex);
2569 unblockThread(StgTSO *tso)
2573 ACQUIRE_LOCK(&sched_mutex);
2574 switch (tso->why_blocked) {
2577 return; /* not blocked */
2580 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
2582 StgTSO *last_tso = END_TSO_QUEUE;
2583 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
2586 for (t = mvar->head; t != END_TSO_QUEUE;
2587 last = &t->link, last_tso = t, t = t->link) {
2590 if (mvar->tail == tso) {
2591 mvar->tail = last_tso;
2596 barf("unblockThread (MVAR): TSO not found");
2599 case BlockedOnBlackHole:
2600 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
2602 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
2604 last = &bq->blocking_queue;
2605 for (t = bq->blocking_queue; t != END_TSO_QUEUE;
2606 last = &t->link, t = t->link) {
2612 barf("unblockThread (BLACKHOLE): TSO not found");
2615 case BlockedOnException:
2617 StgTSO *target = tso->block_info.tso;
2619 ASSERT(get_itbl(target)->type == TSO);
2620 ASSERT(target->blocked_exceptions != NULL);
2622 last = &target->blocked_exceptions;
2623 for (t = target->blocked_exceptions; t != END_TSO_QUEUE;
2624 last = &t->link, t = t->link) {
2625 ASSERT(get_itbl(t)->type == TSO);
2631 barf("unblockThread (Exception): TSO not found");
2635 case BlockedOnWrite:
2637 StgTSO *prev = NULL;
2638 for (t = blocked_queue_hd; t != END_TSO_QUEUE;
2639 prev = t, t = t->link) {
2642 blocked_queue_hd = t->link;
2643 if (blocked_queue_tl == t) {
2644 blocked_queue_tl = END_TSO_QUEUE;
2647 prev->link = t->link;
2648 if (blocked_queue_tl == t) {
2649 blocked_queue_tl = prev;
2655 barf("unblockThread (I/O): TSO not found");
2658 case BlockedOnDelay:
2660 StgTSO *prev = NULL;
2661 for (t = sleeping_queue; t != END_TSO_QUEUE;
2662 prev = t, t = t->link) {
2665 sleeping_queue = t->link;
2667 prev->link = t->link;
2672 barf("unblockThread (I/O): TSO not found");
2676 barf("unblockThread");
2680 tso->link = END_TSO_QUEUE;
2681 tso->why_blocked = NotBlocked;
2682 tso->block_info.closure = NULL;
2683 PUSH_ON_RUN_QUEUE(tso);
2684 RELEASE_LOCK(&sched_mutex);
2688 /* -----------------------------------------------------------------------------
2691 * The following function implements the magic for raising an
2692 * asynchronous exception in an existing thread.
2694 * We first remove the thread from any queue on which it might be
2695 * blocked. The possible blockages are MVARs and BLACKHOLE_BQs.
2697 * We strip the stack down to the innermost CATCH_FRAME, building
2698 * thunks in the heap for all the active computations, so they can
2699 * be restarted if necessary. When we reach a CATCH_FRAME, we build
2700 * an application of the handler to the exception, and push it on
2701 * the top of the stack.
2703 * How exactly do we save all the active computations? We create an
2704 * AP_UPD for every UpdateFrame on the stack. Entering one of these
2705 * AP_UPDs pushes everything from the corresponding update frame
2706 * upwards onto the stack. (Actually, it pushes everything up to the
2707 * next update frame plus a pointer to the next AP_UPD object.
2708 * Entering the next AP_UPD object pushes more onto the stack until we
2709 * reach the last AP_UPD object - at which point the stack should look
2710 * exactly as it did when we killed the TSO and we can continue
2711 * execution by entering the closure on top of the stack.
2713 * We can also kill a thread entirely - this happens if either (a) the
2714 * exception passed to raiseAsync is NULL, or (b) there's no
2715 * CATCH_FRAME on the stack. In either case, we strip the entire
2716 * stack and replace the thread with a zombie.
2718 * -------------------------------------------------------------------------- */
2721 deleteThread(StgTSO *tso)
2723 raiseAsync(tso,NULL);
2727 raiseAsync(StgTSO *tso, StgClosure *exception)
2729 StgUpdateFrame* su = tso->su;
2730 StgPtr sp = tso->sp;
2732 /* Thread already dead? */
2733 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
2737 IF_DEBUG(scheduler, sched_belch("raising exception in thread %ld.", tso->id));
2739 /* Remove it from any blocking queues */
2742 /* The stack freezing code assumes there's a closure pointer on
2743 * the top of the stack. This isn't always the case with compiled
2744 * code, so we have to push a dummy closure on the top which just
2745 * returns to the next return address on the stack.
2747 if ( LOOKS_LIKE_GHC_INFO((void*)*sp) ) {
2748 *(--sp) = (W_)&stg_dummy_ret_closure;
2752 int words = ((P_)su - (P_)sp) - 1;
2756 /* If we find a CATCH_FRAME, and we've got an exception to raise,
2757 * then build PAP(handler,exception,realworld#), and leave it on
2758 * top of the stack ready to enter.
2760 if (get_itbl(su)->type == CATCH_FRAME && exception != NULL) {
2761 StgCatchFrame *cf = (StgCatchFrame *)su;
2762 /* we've got an exception to raise, so let's pass it to the
2763 * handler in this frame.
2765 ap = (StgAP_UPD *)allocate(sizeofW(StgPAP) + 2);
2766 TICK_ALLOC_UPD_PAP(3,0);
2767 SET_HDR(ap,&stg_PAP_info,cf->header.prof.ccs);
2770 ap->fun = cf->handler; /* :: Exception -> IO a */
2771 ap->payload[0] = exception;
2772 ap->payload[1] = ARG_TAG(0); /* realworld token */
2774 /* throw away the stack from Sp up to and including the
2777 sp = (P_)su + sizeofW(StgCatchFrame) - 1;
2780 /* Restore the blocked/unblocked state for asynchronous exceptions
2781 * at the CATCH_FRAME.
2783 * If exceptions were unblocked at the catch, arrange that they
2784 * are unblocked again after executing the handler by pushing an
2785 * unblockAsyncExceptions_ret stack frame.
2787 if (!cf->exceptions_blocked) {
2788 *(sp--) = (W_)&stg_unblockAsyncExceptionszh_ret_info;
2791 /* Ensure that async exceptions are blocked when running the handler.
2793 if (tso->blocked_exceptions == NULL) {
2794 tso->blocked_exceptions = END_TSO_QUEUE;
2797 /* Put the newly-built PAP on top of the stack, ready to execute
2798 * when the thread restarts.
2802 tso->what_next = ThreadEnterGHC;
2803 IF_DEBUG(sanity, checkTSO(tso));
2807 /* First build an AP_UPD consisting of the stack chunk above the
2808 * current update frame, with the top word on the stack as the
2811 ap = (StgAP_UPD *)allocate(AP_sizeW(words));
2816 ap->fun = (StgClosure *)sp[0];
2818 for(i=0; i < (nat)words; ++i) {
2819 ap->payload[i] = (StgClosure *)*sp++;
2822 switch (get_itbl(su)->type) {
2826 SET_HDR(ap,&stg_AP_UPD_info,su->header.prof.ccs /* ToDo */);
2827 TICK_ALLOC_UP_THK(words+1,0);
2830 fprintf(stderr, "scheduler: Updating ");
2831 printPtr((P_)su->updatee);
2832 fprintf(stderr, " with ");
2833 printObj((StgClosure *)ap);
2836 /* Replace the updatee with an indirection - happily
2837 * this will also wake up any threads currently
2838 * waiting on the result.
2840 UPD_IND_NOLOCK(su->updatee,ap); /* revert the black hole */
2842 sp += sizeofW(StgUpdateFrame) -1;
2843 sp[0] = (W_)ap; /* push onto stack */
2849 StgCatchFrame *cf = (StgCatchFrame *)su;
2852 /* We want a PAP, not an AP_UPD. Fortunately, the
2853 * layout's the same.
2855 SET_HDR(ap,&stg_PAP_info,su->header.prof.ccs /* ToDo */);
2856 TICK_ALLOC_UPD_PAP(words+1,0);
2858 /* now build o = FUN(catch,ap,handler) */
2859 o = (StgClosure *)allocate(sizeofW(StgClosure)+2);
2860 TICK_ALLOC_FUN(2,0);
2861 SET_HDR(o,&stg_catch_info,su->header.prof.ccs /* ToDo */);
2862 o->payload[0] = (StgClosure *)ap;
2863 o->payload[1] = cf->handler;
2866 fprintf(stderr, "scheduler: Built ");
2867 printObj((StgClosure *)o);
2870 /* pop the old handler and put o on the stack */
2872 sp += sizeofW(StgCatchFrame) - 1;
2879 StgSeqFrame *sf = (StgSeqFrame *)su;
2882 SET_HDR(ap,&stg_PAP_info,su->header.prof.ccs /* ToDo */);
2883 TICK_ALLOC_UPD_PAP(words+1,0);
2885 /* now build o = FUN(seq,ap) */
2886 o = (StgClosure *)allocate(sizeofW(StgClosure)+1);
2887 TICK_ALLOC_SE_THK(1,0);
2888 SET_HDR(o,&stg_seq_info,su->header.prof.ccs /* ToDo */);
2889 o->payload[0] = (StgClosure *)ap;
2892 fprintf(stderr, "scheduler: Built ");
2893 printObj((StgClosure *)o);
2896 /* pop the old handler and put o on the stack */
2898 sp += sizeofW(StgSeqFrame) - 1;
2904 /* We've stripped the entire stack, the thread is now dead. */
2905 sp += sizeofW(StgStopFrame) - 1;
2906 sp[0] = (W_)exception; /* save the exception */
2907 tso->what_next = ThreadKilled;
2908 tso->su = (StgUpdateFrame *)(sp+1);
2919 /* -----------------------------------------------------------------------------
2920 resurrectThreads is called after garbage collection on the list of
2921 threads found to be garbage. Each of these threads will be woken
2922 up and sent a signal: BlockedOnDeadMVar if the thread was blocked
2923 on an MVar, or NonTermination if the thread was blocked on a Black
2925 -------------------------------------------------------------------------- */
2928 resurrectThreads( StgTSO *threads )
2932 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
2933 next = tso->global_link;
2934 tso->global_link = all_threads;
2936 IF_DEBUG(scheduler, sched_belch("resurrecting thread %d", tso->id));
2938 switch (tso->why_blocked) {
2940 case BlockedOnException:
2941 raiseAsync(tso,(StgClosure *)BlockedOnDeadMVar_closure);
2943 case BlockedOnBlackHole:
2944 raiseAsync(tso,(StgClosure *)NonTermination_closure);
2947 /* This might happen if the thread was blocked on a black hole
2948 * belonging to a thread that we've just woken up (raiseAsync
2949 * can wake up threads, remember...).
2953 barf("resurrectThreads: thread blocked in a strange way");
2958 /* -----------------------------------------------------------------------------
2959 * Blackhole detection: if we reach a deadlock, test whether any
2960 * threads are blocked on themselves. Any threads which are found to
2961 * be self-blocked get sent a NonTermination exception.
2963 * This is only done in a deadlock situation in order to avoid
2964 * performance overhead in the normal case.
2965 * -------------------------------------------------------------------------- */
2968 detectBlackHoles( void )
2970 StgTSO *t = all_threads;
2971 StgUpdateFrame *frame;
2972 StgClosure *blocked_on;
2974 for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
2976 if (t->why_blocked != BlockedOnBlackHole) {
2980 blocked_on = t->block_info.closure;
2982 for (frame = t->su; ; frame = frame->link) {
2983 switch (get_itbl(frame)->type) {
2986 if (frame->updatee == blocked_on) {
2987 /* We are blocking on one of our own computations, so
2988 * send this thread the NonTermination exception.
2991 sched_belch("thread %d is blocked on itself", t->id));
2992 raiseAsync(t, (StgClosure *)NonTermination_closure);
3013 //@node Debugging Routines, Index, Exception Handling Routines, Main scheduling code
3014 //@subsection Debugging Routines
3016 /* -----------------------------------------------------------------------------
3017 Debugging: why is a thread blocked
3018 -------------------------------------------------------------------------- */
3023 printThreadBlockage(StgTSO *tso)
3025 switch (tso->why_blocked) {
3027 fprintf(stderr,"is blocked on read from fd %d", tso->block_info.fd);
3029 case BlockedOnWrite:
3030 fprintf(stderr,"is blocked on write to fd %d", tso->block_info.fd);
3032 case BlockedOnDelay:
3033 fprintf(stderr,"is blocked until %d", tso->block_info.target);
3036 fprintf(stderr,"is blocked on an MVar");
3038 case BlockedOnException:
3039 fprintf(stderr,"is blocked on delivering an exception to thread %d",
3040 tso->block_info.tso->id);
3042 case BlockedOnBlackHole:
3043 fprintf(stderr,"is blocked on a black hole");
3046 fprintf(stderr,"is not blocked");
3050 fprintf(stderr,"is blocked on global address; local FM_BQ is %p (%s)",
3051 tso->block_info.closure, info_type(tso->block_info.closure));
3053 case BlockedOnGA_NoSend:
3054 fprintf(stderr,"is blocked on global address (no send); local FM_BQ is %p (%s)",
3055 tso->block_info.closure, info_type(tso->block_info.closure));
3059 barf("printThreadBlockage: strange tso->why_blocked: %d for TSO %d (%d)",
3060 tso->why_blocked, tso->id, tso);
3065 printThreadStatus(StgTSO *tso)
3067 switch (tso->what_next) {
3069 fprintf(stderr,"has been killed");
3071 case ThreadComplete:
3072 fprintf(stderr,"has completed");
3075 printThreadBlockage(tso);
3080 printAllThreads(void)
3084 sched_belch("all threads:");
3085 for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
3086 fprintf(stderr, "\tthread %d ", t->id);
3087 printThreadStatus(t);
3088 fprintf(stderr,"\n");
3093 Print a whole blocking queue attached to node (debugging only).
3098 print_bq (StgClosure *node)
3100 StgBlockingQueueElement *bqe;
3104 fprintf(stderr,"## BQ of closure %p (%s): ",
3105 node, info_type(node));
3107 /* should cover all closures that may have a blocking queue */
3108 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
3109 get_itbl(node)->type == FETCH_ME_BQ ||
3110 get_itbl(node)->type == RBH);
3112 ASSERT(node!=(StgClosure*)NULL); // sanity check
3114 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
3116 for (bqe = ((StgBlockingQueue*)node)->blocking_queue, end = (bqe==END_BQ_QUEUE);
3117 !end; // iterate until bqe points to a CONSTR
3118 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) {
3119 ASSERT(bqe != END_BQ_QUEUE); // sanity check
3120 ASSERT(bqe != (StgTSO*)NULL); // sanity check
3121 /* types of closures that may appear in a blocking queue */
3122 ASSERT(get_itbl(bqe)->type == TSO ||
3123 get_itbl(bqe)->type == BLOCKED_FETCH ||
3124 get_itbl(bqe)->type == CONSTR);
3125 /* only BQs of an RBH end with an RBH_Save closure */
3126 ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
3128 switch (get_itbl(bqe)->type) {
3130 fprintf(stderr," TSO %d (%x),",
3131 ((StgTSO *)bqe)->id, ((StgTSO *)bqe));
3134 fprintf(stderr," BF (node=%p, ga=((%x, %d, %x)),",
3135 ((StgBlockedFetch *)bqe)->node,
3136 ((StgBlockedFetch *)bqe)->ga.payload.gc.gtid,
3137 ((StgBlockedFetch *)bqe)->ga.payload.gc.slot,
3138 ((StgBlockedFetch *)bqe)->ga.weight);
3141 fprintf(stderr," %s (IP %p),",
3142 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
3143 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
3144 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
3145 "RBH_Save_?"), get_itbl(bqe));
3148 barf("Unexpected closure type %s in blocking queue of %p (%s)",
3149 info_type(bqe), node, info_type(node));
3153 fputc('\n', stderr);
3155 # elif defined(GRAN)
3157 print_bq (StgClosure *node)
3159 StgBlockingQueueElement *bqe;
3160 PEs node_loc, tso_loc;
3163 /* should cover all closures that may have a blocking queue */
3164 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
3165 get_itbl(node)->type == FETCH_ME_BQ ||
3166 get_itbl(node)->type == RBH);
3168 ASSERT(node!=(StgClosure*)NULL); // sanity check
3169 node_loc = where_is(node);
3171 fprintf(stderr,"## BQ of closure %p (%s) on [PE %d]: ",
3172 node, info_type(node), node_loc);
3175 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
3177 for (bqe = ((StgBlockingQueue*)node)->blocking_queue, end = (bqe==END_BQ_QUEUE);
3178 !end; // iterate until bqe points to a CONSTR
3179 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) {
3180 ASSERT(bqe != END_BQ_QUEUE); // sanity check
3181 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
3182 /* types of closures that may appear in a blocking queue */
3183 ASSERT(get_itbl(bqe)->type == TSO ||
3184 get_itbl(bqe)->type == CONSTR);
3185 /* only BQs of an RBH end with an RBH_Save closure */
3186 ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
3188 tso_loc = where_is((StgClosure *)bqe);
3189 switch (get_itbl(bqe)->type) {
3191 fprintf(stderr," TSO %d (%p) on [PE %d],",
3192 ((StgTSO *)bqe)->id, (StgTSO *)bqe, tso_loc);
3195 fprintf(stderr," %s (IP %p),",
3196 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
3197 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
3198 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
3199 "RBH_Save_?"), get_itbl(bqe));
3202 barf("Unexpected closure type %s in blocking queue of %p (%s)",
3203 info_type((StgClosure *)bqe), node, info_type(node));
3207 fputc('\n', stderr);
3211 Nice and easy: only TSOs on the blocking queue
3214 print_bq (StgClosure *node)
3218 ASSERT(node!=(StgClosure*)NULL); // sanity check
3219 for (tso = ((StgBlockingQueue*)node)->blocking_queue;
3220 tso != END_TSO_QUEUE;
3222 ASSERT(tso!=NULL && tso!=END_TSO_QUEUE); // sanity check
3223 ASSERT(get_itbl(tso)->type == TSO); // guess what, sanity check
3224 fprintf(stderr," TSO %d (%p),", tso->id, tso);
3226 fputc('\n', stderr);
3237 for (i=0, tso=run_queue_hd;
3238 tso != END_TSO_QUEUE;
3247 sched_belch(char *s, ...)
3252 fprintf(stderr, "scheduler (task %ld): ", pthread_self());
3254 fprintf(stderr, "scheduler: ");
3256 vfprintf(stderr, s, ap);
3257 fprintf(stderr, "\n");
3263 //@node Index, , Debugging Routines, Main scheduling code
3267 //* MainRegTable:: @cindex\s-+MainRegTable
3268 //* StgMainThread:: @cindex\s-+StgMainThread
3269 //* awaken_blocked_queue:: @cindex\s-+awaken_blocked_queue
3270 //* blocked_queue_hd:: @cindex\s-+blocked_queue_hd
3271 //* blocked_queue_tl:: @cindex\s-+blocked_queue_tl
3272 //* context_switch:: @cindex\s-+context_switch
3273 //* createThread:: @cindex\s-+createThread
3274 //* free_capabilities:: @cindex\s-+free_capabilities
3275 //* gc_pending_cond:: @cindex\s-+gc_pending_cond
3276 //* initScheduler:: @cindex\s-+initScheduler
3277 //* interrupted:: @cindex\s-+interrupted
3278 //* n_free_capabilities:: @cindex\s-+n_free_capabilities
3279 //* next_thread_id:: @cindex\s-+next_thread_id
3280 //* print_bq:: @cindex\s-+print_bq
3281 //* run_queue_hd:: @cindex\s-+run_queue_hd
3282 //* run_queue_tl:: @cindex\s-+run_queue_tl
3283 //* sched_mutex:: @cindex\s-+sched_mutex
3284 //* schedule:: @cindex\s-+schedule
3285 //* take_off_run_queue:: @cindex\s-+take_off_run_queue
3286 //* task_ids:: @cindex\s-+task_ids
3287 //* term_mutex:: @cindex\s-+term_mutex
3288 //* thread_ready_cond:: @cindex\s-+thread_ready_cond