1 /* ---------------------------------------------------------------------------
2 * $Id: Schedule.c,v 1.80 2000/11/07 10:42:57 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 "Evaluator.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 ThreadEnterHugs:
893 IF_DEBUG(scheduler,sched_belch("entering Hugs"));
894 c = (StgClosure *)(cap->rCurrentTSO->sp[0]);
895 cap->rCurrentTSO->sp += 1;
900 barf("Panic: entered a BCO but no bytecode interpreter in this build");
903 barf("schedule: invalid what_next field");
905 /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
907 /* Costs for the scheduler are assigned to CCS_SYSTEM */
912 ACQUIRE_LOCK(&sched_mutex);
915 IF_DEBUG(scheduler,fprintf(stderr,"scheduler (task %ld): ", pthread_self()););
916 #elif !defined(GRAN) && !defined(PAR)
917 IF_DEBUG(scheduler,fprintf(stderr,"scheduler: "););
919 t = cap->rCurrentTSO;
922 /* HACK 675: if the last thread didn't yield, make sure to print a
923 SCHEDULE event to the log file when StgRunning the next thread, even
924 if it is the same one as before */
925 LastTSO = t; //(ret == ThreadBlocked) ? END_TSO_QUEUE : t;
926 TimeOfLastYield = CURRENT_TIME;
931 /* make all the running tasks block on a condition variable,
932 * maybe set context_switch and wait till they all pile in,
933 * then have them wait on a GC condition variable.
935 IF_DEBUG(scheduler,belch("--<< thread %ld (%p; %s) stopped: HeapOverflow",
936 t->id, t, whatNext_strs[t->what_next]));
939 ASSERT(!is_on_queue(t,CurrentProc));
942 ready_to_gc = rtsTrue;
943 context_switch = 1; /* stop other threads ASAP */
944 PUSH_ON_RUN_QUEUE(t);
945 /* actual GC is done at the end of the while loop */
949 IF_DEBUG(scheduler,belch("--<< thread %ld (%p; %s) stopped, StackOverflow",
950 t->id, t, whatNext_strs[t->what_next]));
951 /* just adjust the stack for this thread, then pop it back
957 /* enlarge the stack */
958 StgTSO *new_t = threadStackOverflow(t);
960 /* This TSO has moved, so update any pointers to it from the
961 * main thread stack. It better not be on any other queues...
964 for (m = main_threads; m != NULL; m = m->link) {
970 PUSH_ON_RUN_QUEUE(new_t);
977 DumpGranEvent(GR_DESCHEDULE, t));
978 globalGranStats.tot_yields++;
981 DumpGranEvent(GR_DESCHEDULE, t));
983 /* put the thread back on the run queue. Then, if we're ready to
984 * GC, check whether this is the last task to stop. If so, wake
985 * up the GC thread. getThread will block during a GC until the
989 if (t->what_next == ThreadEnterHugs) {
990 /* ToDo: or maybe a timer expired when we were in Hugs?
991 * or maybe someone hit ctrl-C
993 belch("--<< thread %ld (%p; %s) stopped to switch to Hugs",
994 t->id, t, whatNext_strs[t->what_next]);
996 belch("--<< thread %ld (%p; %s) stopped, yielding",
997 t->id, t, whatNext_strs[t->what_next]);
1004 //belch("&& Doing sanity check on yielding TSO %ld.", t->id);
1006 ASSERT(t->link == END_TSO_QUEUE);
1008 ASSERT(!is_on_queue(t,CurrentProc));
1011 //belch("&& Doing sanity check on all ThreadQueues (and their TSOs).");
1012 checkThreadQsSanity(rtsTrue));
1014 APPEND_TO_RUN_QUEUE(t);
1016 /* add a ContinueThread event to actually process the thread */
1017 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1019 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1021 belch("GRAN: eventq and runnableq after adding yielded thread to queue again:");
1030 belch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: ",
1031 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)));
1032 if (t->block_info.closure!=(StgClosure*)NULL) print_bq(t->block_info.closure));
1034 // ??? needed; should emit block before
1036 DumpGranEvent(GR_DESCHEDULE, t));
1037 prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
1040 ASSERT(procStatus[CurrentProc]==Busy ||
1041 ((procStatus[CurrentProc]==Fetching) &&
1042 (t->block_info.closure!=(StgClosure*)NULL)));
1043 if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
1044 !(!RtsFlags.GranFlags.DoAsyncFetch &&
1045 procStatus[CurrentProc]==Fetching))
1046 procStatus[CurrentProc] = Idle;
1050 DumpGranEvent(GR_DESCHEDULE, t));
1052 /* Send a fetch (if BlockedOnGA) and dump event to log file */
1056 belch("--<< thread %ld (%p; %s) stopped, blocking on node %p with BQ: ",
1057 t->id, t, whatNext_strs[t->what_next], t->block_info.closure);
1058 if (t->block_info.closure!=(StgClosure*)NULL) print_bq(t->block_info.closure));
1061 /* don't need to do anything. Either the thread is blocked on
1062 * I/O, in which case we'll have called addToBlockedQueue
1063 * previously, or it's blocked on an MVar or Blackhole, in which
1064 * case it'll be on the relevant queue already.
1067 fprintf(stderr, "--<< thread %d (%p) stopped: ", t->id, t);
1068 printThreadBlockage(t);
1069 fprintf(stderr, "\n"));
1071 /* Only for dumping event to log file
1072 ToDo: do I need this in GranSim, too?
1079 case ThreadFinished:
1080 /* Need to check whether this was a main thread, and if so, signal
1081 * the task that started it with the return value. If we have no
1082 * more main threads, we probably need to stop all the tasks until
1085 /* We also end up here if the thread kills itself with an
1086 * uncaught exception, see Exception.hc.
1088 IF_DEBUG(scheduler,belch("--++ thread %d (%p) finished", t->id, t));
1090 endThread(t, CurrentProc); // clean-up the thread
1092 advisory_thread_count--;
1093 if (RtsFlags.ParFlags.ParStats.Full)
1094 DumpEndEvent(CURRENT_PROC, t, rtsFalse /* not mandatory */);
1099 barf("schedule: invalid thread return code %d", (int)ret);
1103 cap->link = free_capabilities;
1104 free_capabilities = cap;
1105 n_free_capabilities++;
1109 if (ready_to_gc && n_free_capabilities == RtsFlags.ParFlags.nNodes)
1114 /* everybody back, start the GC.
1115 * Could do it in this thread, or signal a condition var
1116 * to do it in another thread. Either way, we need to
1117 * broadcast on gc_pending_cond afterward.
1120 IF_DEBUG(scheduler,sched_belch("doing GC"));
1122 GarbageCollect(GetRoots,rtsFalse);
1123 ready_to_gc = rtsFalse;
1125 pthread_cond_broadcast(&gc_pending_cond);
1128 /* add a ContinueThread event to continue execution of current thread */
1129 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1131 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1133 fprintf(stderr, "GRAN: eventq and runnableq after Garbage collection:\n");
1140 IF_GRAN_DEBUG(unused,
1141 print_eventq(EventHd));
1143 event = get_next_event();
1147 /* ToDo: wait for next message to arrive rather than busy wait */
1152 t = take_off_run_queue(END_TSO_QUEUE);
1155 } /* end of while(1) */
1158 /* ---------------------------------------------------------------------------
1159 * deleteAllThreads(): kill all the live threads.
1161 * This is used when we catch a user interrupt (^C), before performing
1162 * any necessary cleanups and running finalizers.
1163 * ------------------------------------------------------------------------- */
1165 void deleteAllThreads ( void )
1168 IF_DEBUG(scheduler,sched_belch("deleting all threads"));
1169 for (t = run_queue_hd; t != END_TSO_QUEUE; t = t->link) {
1172 for (t = blocked_queue_hd; t != END_TSO_QUEUE; t = t->link) {
1175 for (t = sleeping_queue; t != END_TSO_QUEUE; t = t->link) {
1178 run_queue_hd = run_queue_tl = END_TSO_QUEUE;
1179 blocked_queue_hd = blocked_queue_tl = END_TSO_QUEUE;
1180 sleeping_queue = END_TSO_QUEUE;
1183 /* startThread and insertThread are now in GranSim.c -- HWL */
1185 //@node Suspend and Resume, Run queue code, Main scheduling loop, Main scheduling code
1186 //@subsection Suspend and Resume
1188 /* ---------------------------------------------------------------------------
1189 * Suspending & resuming Haskell threads.
1191 * When making a "safe" call to C (aka _ccall_GC), the task gives back
1192 * its capability before calling the C function. This allows another
1193 * task to pick up the capability and carry on running Haskell
1194 * threads. It also means that if the C call blocks, it won't lock
1197 * The Haskell thread making the C call is put to sleep for the
1198 * duration of the call, on the susepended_ccalling_threads queue. We
1199 * give out a token to the task, which it can use to resume the thread
1200 * on return from the C function.
1201 * ------------------------------------------------------------------------- */
1204 suspendThread( Capability *cap )
1208 ACQUIRE_LOCK(&sched_mutex);
1211 sched_belch("thread %d did a _ccall_gc", cap->rCurrentTSO->id));
1213 threadPaused(cap->rCurrentTSO);
1214 cap->rCurrentTSO->link = suspended_ccalling_threads;
1215 suspended_ccalling_threads = cap->rCurrentTSO;
1217 /* Use the thread ID as the token; it should be unique */
1218 tok = cap->rCurrentTSO->id;
1221 cap->link = free_capabilities;
1222 free_capabilities = cap;
1223 n_free_capabilities++;
1226 RELEASE_LOCK(&sched_mutex);
1231 resumeThread( StgInt tok )
1233 StgTSO *tso, **prev;
1236 ACQUIRE_LOCK(&sched_mutex);
1238 prev = &suspended_ccalling_threads;
1239 for (tso = suspended_ccalling_threads;
1240 tso != END_TSO_QUEUE;
1241 prev = &tso->link, tso = tso->link) {
1242 if (tso->id == (StgThreadID)tok) {
1247 if (tso == END_TSO_QUEUE) {
1248 barf("resumeThread: thread not found");
1252 while (free_capabilities == NULL) {
1253 IF_DEBUG(scheduler, sched_belch("waiting to resume"));
1254 pthread_cond_wait(&thread_ready_cond, &sched_mutex);
1255 IF_DEBUG(scheduler, sched_belch("resuming thread %d", tso->id));
1257 cap = free_capabilities;
1258 free_capabilities = cap->link;
1259 n_free_capabilities--;
1261 cap = &MainRegTable;
1264 cap->rCurrentTSO = tso;
1266 RELEASE_LOCK(&sched_mutex);
1271 /* ---------------------------------------------------------------------------
1273 * ------------------------------------------------------------------------ */
1274 static void unblockThread(StgTSO *tso);
1276 /* ---------------------------------------------------------------------------
1277 * Comparing Thread ids.
1279 * This is used from STG land in the implementation of the
1280 * instances of Eq/Ord for ThreadIds.
1281 * ------------------------------------------------------------------------ */
1283 int cmp_thread(const StgTSO *tso1, const StgTSO *tso2)
1285 StgThreadID id1 = tso1->id;
1286 StgThreadID id2 = tso2->id;
1288 if (id1 < id2) return (-1);
1289 if (id1 > id2) return 1;
1293 /* ---------------------------------------------------------------------------
1294 Create a new thread.
1296 The new thread starts with the given stack size. Before the
1297 scheduler can run, however, this thread needs to have a closure
1298 (and possibly some arguments) pushed on its stack. See
1299 pushClosure() in Schedule.h.
1301 createGenThread() and createIOThread() (in SchedAPI.h) are
1302 convenient packaged versions of this function.
1304 currently pri (priority) is only used in a GRAN setup -- HWL
1305 ------------------------------------------------------------------------ */
1306 //@cindex createThread
1308 /* currently pri (priority) is only used in a GRAN setup -- HWL */
1310 createThread(nat stack_size, StgInt pri)
1312 return createThread_(stack_size, rtsFalse, pri);
1316 createThread_(nat size, rtsBool have_lock, StgInt pri)
1320 createThread(nat stack_size)
1322 return createThread_(stack_size, rtsFalse);
1326 createThread_(nat size, rtsBool have_lock)
1333 /* First check whether we should create a thread at all */
1335 /* check that no more than RtsFlags.ParFlags.maxThreads threads are created */
1336 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads) {
1338 belch("{createThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
1339 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
1340 return END_TSO_QUEUE;
1346 ASSERT(!RtsFlags.GranFlags.Light || CurrentProc==0);
1349 // ToDo: check whether size = stack_size - TSO_STRUCT_SIZEW
1351 /* catch ridiculously small stack sizes */
1352 if (size < MIN_STACK_WORDS + TSO_STRUCT_SIZEW) {
1353 size = MIN_STACK_WORDS + TSO_STRUCT_SIZEW;
1356 stack_size = size - TSO_STRUCT_SIZEW;
1358 tso = (StgTSO *)allocate(size);
1359 TICK_ALLOC_TSO(size-TSO_STRUCT_SIZEW, 0);
1361 SET_HDR(tso, &TSO_info, CCS_SYSTEM);
1363 SET_GRAN_HDR(tso, ThisPE);
1365 tso->what_next = ThreadEnterGHC;
1367 /* tso->id needs to be unique. For now we use a heavyweight mutex to
1368 * protect the increment operation on next_thread_id.
1369 * In future, we could use an atomic increment instead.
1371 if (!have_lock) { ACQUIRE_LOCK(&sched_mutex); }
1372 tso->id = next_thread_id++;
1373 if (!have_lock) { RELEASE_LOCK(&sched_mutex); }
1375 tso->why_blocked = NotBlocked;
1376 tso->blocked_exceptions = NULL;
1378 tso->stack_size = stack_size;
1379 tso->max_stack_size = round_to_mblocks(RtsFlags.GcFlags.maxStkSize)
1381 tso->sp = (P_)&(tso->stack) + stack_size;
1384 tso->prof.CCCS = CCS_MAIN;
1387 /* put a stop frame on the stack */
1388 tso->sp -= sizeofW(StgStopFrame);
1389 SET_HDR((StgClosure*)tso->sp,(StgInfoTable *)&stg_stop_thread_info,CCS_SYSTEM);
1390 tso->su = (StgUpdateFrame*)tso->sp;
1394 tso->link = END_TSO_QUEUE;
1395 /* uses more flexible routine in GranSim */
1396 insertThread(tso, CurrentProc);
1398 /* In a non-GranSim setup the pushing of a TSO onto the runq is separated
1403 #if defined(GRAN) || defined(PAR)
1404 DumpGranEvent(GR_START,tso);
1407 /* Link the new thread on the global thread list.
1409 tso->global_link = all_threads;
1413 tso->gran.pri = pri;
1415 tso->gran.magic = TSO_MAGIC; // debugging only
1417 tso->gran.sparkname = 0;
1418 tso->gran.startedat = CURRENT_TIME;
1419 tso->gran.exported = 0;
1420 tso->gran.basicblocks = 0;
1421 tso->gran.allocs = 0;
1422 tso->gran.exectime = 0;
1423 tso->gran.fetchtime = 0;
1424 tso->gran.fetchcount = 0;
1425 tso->gran.blocktime = 0;
1426 tso->gran.blockcount = 0;
1427 tso->gran.blockedat = 0;
1428 tso->gran.globalsparks = 0;
1429 tso->gran.localsparks = 0;
1430 if (RtsFlags.GranFlags.Light)
1431 tso->gran.clock = Now; /* local clock */
1433 tso->gran.clock = 0;
1435 IF_DEBUG(gran,printTSO(tso));
1438 tso->par.magic = TSO_MAGIC; // debugging only
1440 tso->par.sparkname = 0;
1441 tso->par.startedat = CURRENT_TIME;
1442 tso->par.exported = 0;
1443 tso->par.basicblocks = 0;
1444 tso->par.allocs = 0;
1445 tso->par.exectime = 0;
1446 tso->par.fetchtime = 0;
1447 tso->par.fetchcount = 0;
1448 tso->par.blocktime = 0;
1449 tso->par.blockcount = 0;
1450 tso->par.blockedat = 0;
1451 tso->par.globalsparks = 0;
1452 tso->par.localsparks = 0;
1456 globalGranStats.tot_threads_created++;
1457 globalGranStats.threads_created_on_PE[CurrentProc]++;
1458 globalGranStats.tot_sq_len += spark_queue_len(CurrentProc);
1459 globalGranStats.tot_sq_probes++;
1464 belch("==__ schedule: Created TSO %d (%p);",
1465 CurrentProc, tso, tso->id));
1467 IF_PAR_DEBUG(verbose,
1468 belch("==__ schedule: Created TSO %d (%p); %d threads active",
1469 tso->id, tso, advisory_thread_count));
1471 IF_DEBUG(scheduler,sched_belch("created thread %ld, stack size = %lx words",
1472 tso->id, tso->stack_size));
1478 Turn a spark into a thread.
1479 ToDo: fix for SMP (needs to acquire SCHED_MUTEX!)
1482 //@cindex activateSpark
1484 activateSpark (rtsSpark spark)
1488 ASSERT(spark != (rtsSpark)NULL);
1489 tso = createThread_(RtsFlags.GcFlags.initialStkSize, rtsTrue);
1490 if (tso!=END_TSO_QUEUE) {
1491 pushClosure(tso,spark);
1492 PUSH_ON_RUN_QUEUE(tso);
1493 advisory_thread_count++;
1495 if (RtsFlags.ParFlags.ParStats.Full) {
1496 //ASSERT(run_queue_hd == END_TSO_QUEUE); // I think ...
1497 IF_PAR_DEBUG(verbose,
1498 belch("==^^ activateSpark: turning spark of closure %p (%s) into a thread",
1499 (StgClosure *)spark, info_type((StgClosure *)spark)));
1502 barf("activateSpark: Cannot create TSO");
1504 // ToDo: fwd info on local/global spark to thread -- HWL
1505 // tso->gran.exported = spark->exported;
1506 // tso->gran.locked = !spark->global;
1507 // tso->gran.sparkname = spark->name;
1513 /* ---------------------------------------------------------------------------
1516 * scheduleThread puts a thread on the head of the runnable queue.
1517 * This will usually be done immediately after a thread is created.
1518 * The caller of scheduleThread must create the thread using e.g.
1519 * createThread and push an appropriate closure
1520 * on this thread's stack before the scheduler is invoked.
1521 * ------------------------------------------------------------------------ */
1524 scheduleThread(StgTSO *tso)
1526 if (tso==END_TSO_QUEUE){
1531 ACQUIRE_LOCK(&sched_mutex);
1533 /* Put the new thread on the head of the runnable queue. The caller
1534 * better push an appropriate closure on this thread's stack
1535 * beforehand. In the SMP case, the thread may start running as
1536 * soon as we release the scheduler lock below.
1538 PUSH_ON_RUN_QUEUE(tso);
1542 IF_DEBUG(scheduler,printTSO(tso));
1544 RELEASE_LOCK(&sched_mutex);
1547 /* ---------------------------------------------------------------------------
1550 * Start up Posix threads to run each of the scheduler tasks.
1551 * I believe the task ids are not needed in the system as defined.
1553 * ------------------------------------------------------------------------ */
1555 #if defined(PAR) || defined(SMP)
1557 taskStart( void *arg STG_UNUSED )
1559 rts_evalNothing(NULL);
1563 /* ---------------------------------------------------------------------------
1566 * Initialise the scheduler. This resets all the queues - if the
1567 * queues contained any threads, they'll be garbage collected at the
1570 * This now calls startTasks(), so should only be called once! KH @ 25/10/99
1571 * ------------------------------------------------------------------------ */
1575 term_handler(int sig STG_UNUSED)
1578 ACQUIRE_LOCK(&term_mutex);
1580 RELEASE_LOCK(&term_mutex);
1585 //@cindex initScheduler
1592 for (i=0; i<=MAX_PROC; i++) {
1593 run_queue_hds[i] = END_TSO_QUEUE;
1594 run_queue_tls[i] = END_TSO_QUEUE;
1595 blocked_queue_hds[i] = END_TSO_QUEUE;
1596 blocked_queue_tls[i] = END_TSO_QUEUE;
1597 ccalling_threadss[i] = END_TSO_QUEUE;
1598 sleeping_queue = END_TSO_QUEUE;
1601 run_queue_hd = END_TSO_QUEUE;
1602 run_queue_tl = END_TSO_QUEUE;
1603 blocked_queue_hd = END_TSO_QUEUE;
1604 blocked_queue_tl = END_TSO_QUEUE;
1605 sleeping_queue = END_TSO_QUEUE;
1608 suspended_ccalling_threads = END_TSO_QUEUE;
1610 main_threads = NULL;
1611 all_threads = END_TSO_QUEUE;
1616 RtsFlags.ConcFlags.ctxtSwitchTicks =
1617 RtsFlags.ConcFlags.ctxtSwitchTime / TICK_MILLISECS;
1620 ecafList = END_ECAF_LIST;
1624 /* Install the SIGHUP handler */
1627 struct sigaction action,oact;
1629 action.sa_handler = term_handler;
1630 sigemptyset(&action.sa_mask);
1631 action.sa_flags = 0;
1632 if (sigaction(SIGTERM, &action, &oact) != 0) {
1633 barf("can't install TERM handler");
1639 /* Allocate N Capabilities */
1642 Capability *cap, *prev;
1645 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
1646 cap = stgMallocBytes(sizeof(Capability), "initScheduler:capabilities");
1650 free_capabilities = cap;
1651 n_free_capabilities = RtsFlags.ParFlags.nNodes;
1653 IF_DEBUG(scheduler,fprintf(stderr,"scheduler: Allocated %d capabilities\n",
1654 n_free_capabilities););
1657 #if defined(SMP) || defined(PAR)
1670 /* make some space for saving all the thread ids */
1671 task_ids = stgMallocBytes(RtsFlags.ParFlags.nNodes * sizeof(task_info),
1672 "initScheduler:task_ids");
1674 /* and create all the threads */
1675 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
1676 r = pthread_create(&tid,NULL,taskStart,NULL);
1678 barf("startTasks: Can't create new Posix thread");
1680 task_ids[i].id = tid;
1681 task_ids[i].mut_time = 0.0;
1682 task_ids[i].mut_etime = 0.0;
1683 task_ids[i].gc_time = 0.0;
1684 task_ids[i].gc_etime = 0.0;
1685 task_ids[i].elapsedtimestart = elapsedtime();
1686 IF_DEBUG(scheduler,fprintf(stderr,"scheduler: Started task: %ld\n",tid););
1692 exitScheduler( void )
1697 /* Don't want to use pthread_cancel, since we'd have to install
1698 * these silly exception handlers (pthread_cleanup_{push,pop}) around
1702 /* Cancel all our tasks */
1703 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
1704 pthread_cancel(task_ids[i].id);
1707 /* Wait for all the tasks to terminate */
1708 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
1709 IF_DEBUG(scheduler,fprintf(stderr,"scheduler: waiting for task %ld\n",
1711 pthread_join(task_ids[i].id, NULL);
1715 /* Send 'em all a SIGHUP. That should shut 'em up.
1717 await_death = RtsFlags.ParFlags.nNodes;
1718 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
1719 pthread_kill(task_ids[i].id,SIGTERM);
1721 while (await_death > 0) {
1727 /* -----------------------------------------------------------------------------
1728 Managing the per-task allocation areas.
1730 Each capability comes with an allocation area. These are
1731 fixed-length block lists into which allocation can be done.
1733 ToDo: no support for two-space collection at the moment???
1734 -------------------------------------------------------------------------- */
1736 /* -----------------------------------------------------------------------------
1737 * waitThread is the external interface for running a new computation
1738 * and waiting for the result.
1740 * In the non-SMP case, we create a new main thread, push it on the
1741 * main-thread stack, and invoke the scheduler to run it. The
1742 * scheduler will return when the top main thread on the stack has
1743 * completed or died, and fill in the necessary fields of the
1744 * main_thread structure.
1746 * In the SMP case, we create a main thread as before, but we then
1747 * create a new condition variable and sleep on it. When our new
1748 * main thread has completed, we'll be woken up and the status/result
1749 * will be in the main_thread struct.
1750 * -------------------------------------------------------------------------- */
1753 howManyThreadsAvail ( void )
1757 for (q = run_queue_hd; q != END_TSO_QUEUE; q = q->link)
1759 for (q = blocked_queue_hd; q != END_TSO_QUEUE; q = q->link)
1761 for (q = sleeping_queue; q != END_TSO_QUEUE; q = q->link)
1767 finishAllThreads ( void )
1770 while (run_queue_hd != END_TSO_QUEUE) {
1771 waitThread ( run_queue_hd, NULL );
1773 while (blocked_queue_hd != END_TSO_QUEUE) {
1774 waitThread ( blocked_queue_hd, NULL );
1776 while (sleeping_queue != END_TSO_QUEUE) {
1777 waitThread ( blocked_queue_hd, NULL );
1780 (blocked_queue_hd != END_TSO_QUEUE ||
1781 run_queue_hd != END_TSO_QUEUE ||
1782 sleeping_queue != END_TSO_QUEUE);
1786 waitThread(StgTSO *tso, /*out*/StgClosure **ret)
1789 SchedulerStatus stat;
1791 ACQUIRE_LOCK(&sched_mutex);
1793 m = stgMallocBytes(sizeof(StgMainThread), "waitThread");
1799 pthread_cond_init(&m->wakeup, NULL);
1802 m->link = main_threads;
1805 IF_DEBUG(scheduler, fprintf(stderr, "scheduler: new main thread (%d)\n",
1810 pthread_cond_wait(&m->wakeup, &sched_mutex);
1811 } while (m->stat == NoStatus);
1813 /* GranSim specific init */
1814 CurrentTSO = m->tso; // the TSO to run
1815 procStatus[MainProc] = Busy; // status of main PE
1816 CurrentProc = MainProc; // PE to run it on
1821 ASSERT(m->stat != NoStatus);
1827 pthread_cond_destroy(&m->wakeup);
1830 IF_DEBUG(scheduler, fprintf(stderr, "scheduler: main thread (%d) finished\n",
1834 RELEASE_LOCK(&sched_mutex);
1839 //@node Run queue code, Garbage Collextion Routines, Suspend and Resume, Main scheduling code
1840 //@subsection Run queue code
1844 NB: In GranSim we have many run queues; run_queue_hd is actually a macro
1845 unfolding to run_queue_hds[CurrentProc], thus CurrentProc is an
1846 implicit global variable that has to be correct when calling these
1850 /* Put the new thread on the head of the runnable queue.
1851 * The caller of createThread better push an appropriate closure
1852 * on this thread's stack before the scheduler is invoked.
1854 static /* inline */ void
1855 add_to_run_queue(tso)
1858 ASSERT(tso!=run_queue_hd && tso!=run_queue_tl);
1859 tso->link = run_queue_hd;
1861 if (run_queue_tl == END_TSO_QUEUE) {
1866 /* Put the new thread at the end of the runnable queue. */
1867 static /* inline */ void
1868 push_on_run_queue(tso)
1871 ASSERT(get_itbl((StgClosure *)tso)->type == TSO);
1872 ASSERT(run_queue_hd!=NULL && run_queue_tl!=NULL);
1873 ASSERT(tso!=run_queue_hd && tso!=run_queue_tl);
1874 if (run_queue_hd == END_TSO_QUEUE) {
1877 run_queue_tl->link = tso;
1883 Should be inlined because it's used very often in schedule. The tso
1884 argument is actually only needed in GranSim, where we want to have the
1885 possibility to schedule *any* TSO on the run queue, irrespective of the
1886 actual ordering. Therefore, if tso is not the nil TSO then we traverse
1887 the run queue and dequeue the tso, adjusting the links in the queue.
1889 //@cindex take_off_run_queue
1890 static /* inline */ StgTSO*
1891 take_off_run_queue(StgTSO *tso) {
1895 qetlaHbogh Qu' ngaSbogh ghomDaQ {tso} yIteq!
1897 if tso is specified, unlink that tso from the run_queue (doesn't have
1898 to be at the beginning of the queue); GranSim only
1900 if (tso!=END_TSO_QUEUE) {
1901 /* find tso in queue */
1902 for (t=run_queue_hd, prev=END_TSO_QUEUE;
1903 t!=END_TSO_QUEUE && t!=tso;
1907 /* now actually dequeue the tso */
1908 if (prev!=END_TSO_QUEUE) {
1909 ASSERT(run_queue_hd!=t);
1910 prev->link = t->link;
1912 /* t is at beginning of thread queue */
1913 ASSERT(run_queue_hd==t);
1914 run_queue_hd = t->link;
1916 /* t is at end of thread queue */
1917 if (t->link==END_TSO_QUEUE) {
1918 ASSERT(t==run_queue_tl);
1919 run_queue_tl = prev;
1921 ASSERT(run_queue_tl!=t);
1923 t->link = END_TSO_QUEUE;
1925 /* take tso from the beginning of the queue; std concurrent code */
1927 if (t != END_TSO_QUEUE) {
1928 run_queue_hd = t->link;
1929 t->link = END_TSO_QUEUE;
1930 if (run_queue_hd == END_TSO_QUEUE) {
1931 run_queue_tl = END_TSO_QUEUE;
1940 //@node Garbage Collextion Routines, Blocking Queue Routines, Run queue code, Main scheduling code
1941 //@subsection Garbage Collextion Routines
1943 /* ---------------------------------------------------------------------------
1944 Where are the roots that we know about?
1946 - all the threads on the runnable queue
1947 - all the threads on the blocked queue
1948 - all the threads on the sleeping queue
1949 - all the thread currently executing a _ccall_GC
1950 - all the "main threads"
1952 ------------------------------------------------------------------------ */
1954 /* This has to be protected either by the scheduler monitor, or by the
1955 garbage collection monitor (probably the latter).
1959 static void GetRoots(void)
1966 for (i=0; i<=RtsFlags.GranFlags.proc; i++) {
1967 if ((run_queue_hds[i] != END_TSO_QUEUE) && ((run_queue_hds[i] != NULL)))
1968 run_queue_hds[i] = (StgTSO *)MarkRoot((StgClosure *)run_queue_hds[i]);
1969 if ((run_queue_tls[i] != END_TSO_QUEUE) && ((run_queue_tls[i] != NULL)))
1970 run_queue_tls[i] = (StgTSO *)MarkRoot((StgClosure *)run_queue_tls[i]);
1972 if ((blocked_queue_hds[i] != END_TSO_QUEUE) && ((blocked_queue_hds[i] != NULL)))
1973 blocked_queue_hds[i] = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_hds[i]);
1974 if ((blocked_queue_tls[i] != END_TSO_QUEUE) && ((blocked_queue_tls[i] != NULL)))
1975 blocked_queue_tls[i] = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_tls[i]);
1976 if ((ccalling_threadss[i] != END_TSO_QUEUE) && ((ccalling_threadss[i] != NULL)))
1977 ccalling_threadss[i] = (StgTSO *)MarkRoot((StgClosure *)ccalling_threadss[i]);
1984 if (run_queue_hd != END_TSO_QUEUE) {
1985 ASSERT(run_queue_tl != END_TSO_QUEUE);
1986 run_queue_hd = (StgTSO *)MarkRoot((StgClosure *)run_queue_hd);
1987 run_queue_tl = (StgTSO *)MarkRoot((StgClosure *)run_queue_tl);
1990 if (blocked_queue_hd != END_TSO_QUEUE) {
1991 ASSERT(blocked_queue_tl != END_TSO_QUEUE);
1992 blocked_queue_hd = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_hd);
1993 blocked_queue_tl = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_tl);
1996 if (sleeping_queue != END_TSO_QUEUE) {
1997 sleeping_queue = (StgTSO *)MarkRoot((StgClosure *)sleeping_queue);
2001 for (m = main_threads; m != NULL; m = m->link) {
2002 m->tso = (StgTSO *)MarkRoot((StgClosure *)m->tso);
2004 if (suspended_ccalling_threads != END_TSO_QUEUE)
2005 suspended_ccalling_threads =
2006 (StgTSO *)MarkRoot((StgClosure *)suspended_ccalling_threads);
2008 #if defined(SMP) || defined(PAR) || defined(GRAN)
2013 /* -----------------------------------------------------------------------------
2016 This is the interface to the garbage collector from Haskell land.
2017 We provide this so that external C code can allocate and garbage
2018 collect when called from Haskell via _ccall_GC.
2020 It might be useful to provide an interface whereby the programmer
2021 can specify more roots (ToDo).
2023 This needs to be protected by the GC condition variable above. KH.
2024 -------------------------------------------------------------------------- */
2026 void (*extra_roots)(void);
2031 GarbageCollect(GetRoots,rtsFalse);
2035 performMajorGC(void)
2037 GarbageCollect(GetRoots,rtsTrue);
2043 GetRoots(); /* the scheduler's roots */
2044 extra_roots(); /* the user's roots */
2048 performGCWithRoots(void (*get_roots)(void))
2050 extra_roots = get_roots;
2052 GarbageCollect(AllRoots,rtsFalse);
2055 /* -----------------------------------------------------------------------------
2058 If the thread has reached its maximum stack size, then raise the
2059 StackOverflow exception in the offending thread. Otherwise
2060 relocate the TSO into a larger chunk of memory and adjust its stack
2062 -------------------------------------------------------------------------- */
2065 threadStackOverflow(StgTSO *tso)
2067 nat new_stack_size, new_tso_size, diff, stack_words;
2071 IF_DEBUG(sanity,checkTSO(tso));
2072 if (tso->stack_size >= tso->max_stack_size) {
2075 belch("@@ threadStackOverflow of TSO %d (%p): stack too large (now %ld; max is %ld",
2076 tso->id, tso, tso->stack_size, tso->max_stack_size);
2077 /* If we're debugging, just print out the top of the stack */
2078 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2082 fprintf(stderr, "fatal: stack overflow in Hugs; aborting\n" );
2085 /* Send this thread the StackOverflow exception */
2086 raiseAsync(tso, (StgClosure *)stackOverflow_closure);
2091 /* Try to double the current stack size. If that takes us over the
2092 * maximum stack size for this thread, then use the maximum instead.
2093 * Finally round up so the TSO ends up as a whole number of blocks.
2095 new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
2096 new_tso_size = (nat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) +
2097 TSO_STRUCT_SIZE)/sizeof(W_);
2098 new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */
2099 new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
2101 IF_DEBUG(scheduler, fprintf(stderr,"scheduler: increasing stack size from %d words to %d.\n", tso->stack_size, new_stack_size));
2103 dest = (StgTSO *)allocate(new_tso_size);
2104 TICK_ALLOC_TSO(new_tso_size-sizeofW(StgTSO),0);
2106 /* copy the TSO block and the old stack into the new area */
2107 memcpy(dest,tso,TSO_STRUCT_SIZE);
2108 stack_words = tso->stack + tso->stack_size - tso->sp;
2109 new_sp = (P_)dest + new_tso_size - stack_words;
2110 memcpy(new_sp, tso->sp, stack_words * sizeof(W_));
2112 /* relocate the stack pointers... */
2113 diff = (P_)new_sp - (P_)tso->sp; /* In *words* */
2114 dest->su = (StgUpdateFrame *) ((P_)dest->su + diff);
2116 dest->stack_size = new_stack_size;
2118 /* and relocate the update frame list */
2119 relocate_TSO(tso, dest);
2121 /* Mark the old TSO as relocated. We have to check for relocated
2122 * TSOs in the garbage collector and any primops that deal with TSOs.
2124 * It's important to set the sp and su values to just beyond the end
2125 * of the stack, so we don't attempt to scavenge any part of the
2128 tso->what_next = ThreadRelocated;
2130 tso->sp = (P_)&(tso->stack[tso->stack_size]);
2131 tso->su = (StgUpdateFrame *)tso->sp;
2132 tso->why_blocked = NotBlocked;
2133 dest->mut_link = NULL;
2135 IF_PAR_DEBUG(verbose,
2136 belch("@@ threadStackOverflow of TSO %d (now at %p): stack size increased to %ld",
2137 tso->id, tso, tso->stack_size);
2138 /* If we're debugging, just print out the top of the stack */
2139 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2142 IF_DEBUG(sanity,checkTSO(tso));
2144 IF_DEBUG(scheduler,printTSO(dest));
2150 //@node Blocking Queue Routines, Exception Handling Routines, Garbage Collextion Routines, Main scheduling code
2151 //@subsection Blocking Queue Routines
2153 /* ---------------------------------------------------------------------------
2154 Wake up a queue that was blocked on some resource.
2155 ------------------------------------------------------------------------ */
2159 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2164 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2166 /* write RESUME events to log file and
2167 update blocked and fetch time (depending on type of the orig closure) */
2168 if (RtsFlags.ParFlags.ParStats.Full) {
2169 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
2170 GR_RESUME, ((StgTSO *)bqe), ((StgTSO *)bqe)->block_info.closure,
2171 0, 0 /* spark_queue_len(ADVISORY_POOL) */);
2173 switch (get_itbl(node)->type) {
2175 ((StgTSO *)bqe)->par.fetchtime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2180 ((StgTSO *)bqe)->par.blocktime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2183 barf("{unblockOneLocked}Daq Qagh: unexpected closure in blocking queue");
2190 static StgBlockingQueueElement *
2191 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2194 PEs node_loc, tso_loc;
2196 node_loc = where_is(node); // should be lifted out of loop
2197 tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2198 tso_loc = where_is((StgClosure *)tso);
2199 if (IS_LOCAL_TO(PROCS(node),tso_loc)) { // TSO is local
2200 /* !fake_fetch => TSO is on CurrentProc is same as IS_LOCAL_TO */
2201 ASSERT(CurrentProc!=node_loc || tso_loc==CurrentProc);
2202 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.lunblocktime;
2203 // insertThread(tso, node_loc);
2204 new_event(tso_loc, tso_loc, CurrentTime[CurrentProc],
2206 tso, node, (rtsSpark*)NULL);
2207 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2210 } else { // TSO is remote (actually should be FMBQ)
2211 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.mpacktime +
2212 RtsFlags.GranFlags.Costs.gunblocktime +
2213 RtsFlags.GranFlags.Costs.latency;
2214 new_event(tso_loc, CurrentProc, CurrentTime[CurrentProc],
2216 tso, node, (rtsSpark*)NULL);
2217 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2220 /* the thread-queue-overhead is accounted for in either Resume or UnblockThread */
2222 fprintf(stderr," %s TSO %d (%p) [PE %d] (block_info.closure=%p) (next=%p) ,",
2223 (node_loc==tso_loc ? "Local" : "Global"),
2224 tso->id, tso, CurrentProc, tso->block_info.closure, tso->link));
2225 tso->block_info.closure = NULL;
2226 IF_DEBUG(scheduler,belch("-- Waking up thread %ld (%p)",
2230 static StgBlockingQueueElement *
2231 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2233 StgBlockingQueueElement *next;
2235 switch (get_itbl(bqe)->type) {
2237 ASSERT(((StgTSO *)bqe)->why_blocked != NotBlocked);
2238 /* if it's a TSO just push it onto the run_queue */
2240 // ((StgTSO *)bqe)->link = END_TSO_QUEUE; // debugging?
2241 PUSH_ON_RUN_QUEUE((StgTSO *)bqe);
2243 unblockCount(bqe, node);
2244 /* reset blocking status after dumping event */
2245 ((StgTSO *)bqe)->why_blocked = NotBlocked;
2249 /* if it's a BLOCKED_FETCH put it on the PendingFetches list */
2251 bqe->link = PendingFetches;
2252 PendingFetches = bqe;
2256 /* can ignore this case in a non-debugging setup;
2257 see comments on RBHSave closures above */
2259 /* check that the closure is an RBHSave closure */
2260 ASSERT(get_itbl((StgClosure *)bqe) == &RBH_Save_0_info ||
2261 get_itbl((StgClosure *)bqe) == &RBH_Save_1_info ||
2262 get_itbl((StgClosure *)bqe) == &RBH_Save_2_info);
2266 barf("{unblockOneLocked}Daq Qagh: Unexpected IP (%#lx; %s) in blocking queue at %#lx\n",
2267 get_itbl((StgClosure *)bqe), info_type((StgClosure *)bqe),
2271 // IF_DEBUG(scheduler,sched_belch("waking up thread %ld", tso->id));
2275 #else /* !GRAN && !PAR */
2277 unblockOneLocked(StgTSO *tso)
2281 ASSERT(get_itbl(tso)->type == TSO);
2282 ASSERT(tso->why_blocked != NotBlocked);
2283 tso->why_blocked = NotBlocked;
2285 PUSH_ON_RUN_QUEUE(tso);
2287 IF_DEBUG(scheduler,sched_belch("waking up thread %ld", tso->id));
2292 #if defined(GRAN) || defined(PAR)
2293 inline StgBlockingQueueElement *
2294 unblockOne(StgBlockingQueueElement *bqe, StgClosure *node)
2296 ACQUIRE_LOCK(&sched_mutex);
2297 bqe = unblockOneLocked(bqe, node);
2298 RELEASE_LOCK(&sched_mutex);
2303 unblockOne(StgTSO *tso)
2305 ACQUIRE_LOCK(&sched_mutex);
2306 tso = unblockOneLocked(tso);
2307 RELEASE_LOCK(&sched_mutex);
2314 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2316 StgBlockingQueueElement *bqe;
2321 belch("## AwBQ for node %p on PE %d @ %ld by TSO %d (%p): ", \
2322 node, CurrentProc, CurrentTime[CurrentProc],
2323 CurrentTSO->id, CurrentTSO));
2325 node_loc = where_is(node);
2327 ASSERT(get_itbl(q)->type == TSO || // q is either a TSO or an RBHSave
2328 get_itbl(q)->type == CONSTR); // closure (type constructor)
2329 ASSERT(is_unique(node));
2331 /* FAKE FETCH: magically copy the node to the tso's proc;
2332 no Fetch necessary because in reality the node should not have been
2333 moved to the other PE in the first place
2335 if (CurrentProc!=node_loc) {
2337 belch("## node %p is on PE %d but CurrentProc is %d (TSO %d); assuming fake fetch and adjusting bitmask (old: %#x)",
2338 node, node_loc, CurrentProc, CurrentTSO->id,
2339 // CurrentTSO, where_is(CurrentTSO),
2340 node->header.gran.procs));
2341 node->header.gran.procs = (node->header.gran.procs) | PE_NUMBER(CurrentProc);
2343 belch("## new bitmask of node %p is %#x",
2344 node, node->header.gran.procs));
2345 if (RtsFlags.GranFlags.GranSimStats.Global) {
2346 globalGranStats.tot_fake_fetches++;
2351 // ToDo: check: ASSERT(CurrentProc==node_loc);
2352 while (get_itbl(bqe)->type==TSO) { // q != END_TSO_QUEUE) {
2355 bqe points to the current element in the queue
2356 next points to the next element in the queue
2358 //tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2359 //tso_loc = where_is(tso);
2361 bqe = unblockOneLocked(bqe, node);
2364 /* if this is the BQ of an RBH, we have to put back the info ripped out of
2365 the closure to make room for the anchor of the BQ */
2366 if (bqe!=END_BQ_QUEUE) {
2367 ASSERT(get_itbl(node)->type == RBH && get_itbl(bqe)->type == CONSTR);
2369 ASSERT((info_ptr==&RBH_Save_0_info) ||
2370 (info_ptr==&RBH_Save_1_info) ||
2371 (info_ptr==&RBH_Save_2_info));
2373 /* cf. convertToRBH in RBH.c for writing the RBHSave closure */
2374 ((StgRBH *)node)->blocking_queue = (StgBlockingQueueElement *)((StgRBHSave *)bqe)->payload[0];
2375 ((StgRBH *)node)->mut_link = (StgMutClosure *)((StgRBHSave *)bqe)->payload[1];
2378 belch("## Filled in RBH_Save for %p (%s) at end of AwBQ",
2379 node, info_type(node)));
2382 /* statistics gathering */
2383 if (RtsFlags.GranFlags.GranSimStats.Global) {
2384 // globalGranStats.tot_bq_processing_time += bq_processing_time;
2385 globalGranStats.tot_bq_len += len; // total length of all bqs awakened
2386 // globalGranStats.tot_bq_len_local += len_local; // same for local TSOs only
2387 globalGranStats.tot_awbq++; // total no. of bqs awakened
2390 fprintf(stderr,"## BQ Stats of %p: [%d entries] %s\n",
2391 node, len, (bqe!=END_BQ_QUEUE) ? "RBH" : ""));
2395 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2397 StgBlockingQueueElement *bqe, *next;
2399 ACQUIRE_LOCK(&sched_mutex);
2401 IF_PAR_DEBUG(verbose,
2402 belch("## AwBQ for node %p on [%x]: ",
2405 ASSERT(get_itbl(q)->type == TSO ||
2406 get_itbl(q)->type == BLOCKED_FETCH ||
2407 get_itbl(q)->type == CONSTR);
2410 while (get_itbl(bqe)->type==TSO ||
2411 get_itbl(bqe)->type==BLOCKED_FETCH) {
2412 bqe = unblockOneLocked(bqe, node);
2414 RELEASE_LOCK(&sched_mutex);
2417 #else /* !GRAN && !PAR */
2419 awakenBlockedQueue(StgTSO *tso)
2421 ACQUIRE_LOCK(&sched_mutex);
2422 while (tso != END_TSO_QUEUE) {
2423 tso = unblockOneLocked(tso);
2425 RELEASE_LOCK(&sched_mutex);
2429 //@node Exception Handling Routines, Debugging Routines, Blocking Queue Routines, Main scheduling code
2430 //@subsection Exception Handling Routines
2432 /* ---------------------------------------------------------------------------
2434 - usually called inside a signal handler so it mustn't do anything fancy.
2435 ------------------------------------------------------------------------ */
2438 interruptStgRts(void)
2444 /* -----------------------------------------------------------------------------
2447 This is for use when we raise an exception in another thread, which
2449 This has nothing to do with the UnblockThread event in GranSim. -- HWL
2450 -------------------------------------------------------------------------- */
2452 #if defined(GRAN) || defined(PAR)
2454 NB: only the type of the blocking queue is different in GranSim and GUM
2455 the operations on the queue-elements are the same
2456 long live polymorphism!
2459 unblockThread(StgTSO *tso)
2461 StgBlockingQueueElement *t, **last;
2463 ACQUIRE_LOCK(&sched_mutex);
2464 switch (tso->why_blocked) {
2467 return; /* not blocked */
2470 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
2472 StgBlockingQueueElement *last_tso = END_BQ_QUEUE;
2473 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
2475 last = (StgBlockingQueueElement **)&mvar->head;
2476 for (t = (StgBlockingQueueElement *)mvar->head;
2478 last = &t->link, last_tso = t, t = t->link) {
2479 if (t == (StgBlockingQueueElement *)tso) {
2480 *last = (StgBlockingQueueElement *)tso->link;
2481 if (mvar->tail == tso) {
2482 mvar->tail = (StgTSO *)last_tso;
2487 barf("unblockThread (MVAR): TSO not found");
2490 case BlockedOnBlackHole:
2491 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
2493 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
2495 last = &bq->blocking_queue;
2496 for (t = bq->blocking_queue;
2498 last = &t->link, t = t->link) {
2499 if (t == (StgBlockingQueueElement *)tso) {
2500 *last = (StgBlockingQueueElement *)tso->link;
2504 barf("unblockThread (BLACKHOLE): TSO not found");
2507 case BlockedOnException:
2509 StgTSO *target = tso->block_info.tso;
2511 ASSERT(get_itbl(target)->type == TSO);
2512 ASSERT(target->blocked_exceptions != NULL);
2514 last = (StgBlockingQueueElement **)&target->blocked_exceptions;
2515 for (t = (StgBlockingQueueElement *)target->blocked_exceptions;
2517 last = &t->link, t = t->link) {
2518 ASSERT(get_itbl(t)->type == TSO);
2519 if (t == (StgBlockingQueueElement *)tso) {
2520 *last = (StgBlockingQueueElement *)tso->link;
2524 barf("unblockThread (Exception): TSO not found");
2528 case BlockedOnWrite:
2530 StgBlockingQueueElement *prev = NULL;
2531 for (t = (StgBlockingQueueElement *)blocked_queue_hd; t != END_BQ_QUEUE;
2532 prev = t, t = t->link) {
2533 if (t == (StgBlockingQueueElement *)tso) {
2535 blocked_queue_hd = (StgTSO *)t->link;
2536 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
2537 blocked_queue_tl = END_TSO_QUEUE;
2540 prev->link = t->link;
2541 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
2542 blocked_queue_tl = (StgTSO *)prev;
2548 barf("unblockThread (I/O): TSO not found");
2551 case BlockedOnDelay:
2553 StgBlockingQueueElement *prev = NULL;
2554 for (t = (StgBlockingQueueElement *)sleeping_queue; t != END_BQ_QUEUE;
2555 prev = t, t = t->link) {
2556 if (t == (StgBlockingQueueElement *)tso) {
2558 sleeping_queue = (StgTSO *)t->link;
2560 prev->link = t->link;
2565 barf("unblockThread (I/O): TSO not found");
2569 barf("unblockThread");
2573 tso->link = END_TSO_QUEUE;
2574 tso->why_blocked = NotBlocked;
2575 tso->block_info.closure = NULL;
2576 PUSH_ON_RUN_QUEUE(tso);
2577 RELEASE_LOCK(&sched_mutex);
2581 unblockThread(StgTSO *tso)
2585 ACQUIRE_LOCK(&sched_mutex);
2586 switch (tso->why_blocked) {
2589 return; /* not blocked */
2592 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
2594 StgTSO *last_tso = END_TSO_QUEUE;
2595 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
2598 for (t = mvar->head; t != END_TSO_QUEUE;
2599 last = &t->link, last_tso = t, t = t->link) {
2602 if (mvar->tail == tso) {
2603 mvar->tail = last_tso;
2608 barf("unblockThread (MVAR): TSO not found");
2611 case BlockedOnBlackHole:
2612 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
2614 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
2616 last = &bq->blocking_queue;
2617 for (t = bq->blocking_queue; t != END_TSO_QUEUE;
2618 last = &t->link, t = t->link) {
2624 barf("unblockThread (BLACKHOLE): TSO not found");
2627 case BlockedOnException:
2629 StgTSO *target = tso->block_info.tso;
2631 ASSERT(get_itbl(target)->type == TSO);
2632 ASSERT(target->blocked_exceptions != NULL);
2634 last = &target->blocked_exceptions;
2635 for (t = target->blocked_exceptions; t != END_TSO_QUEUE;
2636 last = &t->link, t = t->link) {
2637 ASSERT(get_itbl(t)->type == TSO);
2643 barf("unblockThread (Exception): TSO not found");
2647 case BlockedOnWrite:
2649 StgTSO *prev = NULL;
2650 for (t = blocked_queue_hd; t != END_TSO_QUEUE;
2651 prev = t, t = t->link) {
2654 blocked_queue_hd = t->link;
2655 if (blocked_queue_tl == t) {
2656 blocked_queue_tl = END_TSO_QUEUE;
2659 prev->link = t->link;
2660 if (blocked_queue_tl == t) {
2661 blocked_queue_tl = prev;
2667 barf("unblockThread (I/O): TSO not found");
2670 case BlockedOnDelay:
2672 StgTSO *prev = NULL;
2673 for (t = sleeping_queue; t != END_TSO_QUEUE;
2674 prev = t, t = t->link) {
2677 sleeping_queue = t->link;
2679 prev->link = t->link;
2684 barf("unblockThread (I/O): TSO not found");
2688 barf("unblockThread");
2692 tso->link = END_TSO_QUEUE;
2693 tso->why_blocked = NotBlocked;
2694 tso->block_info.closure = NULL;
2695 PUSH_ON_RUN_QUEUE(tso);
2696 RELEASE_LOCK(&sched_mutex);
2700 /* -----------------------------------------------------------------------------
2703 * The following function implements the magic for raising an
2704 * asynchronous exception in an existing thread.
2706 * We first remove the thread from any queue on which it might be
2707 * blocked. The possible blockages are MVARs and BLACKHOLE_BQs.
2709 * We strip the stack down to the innermost CATCH_FRAME, building
2710 * thunks in the heap for all the active computations, so they can
2711 * be restarted if necessary. When we reach a CATCH_FRAME, we build
2712 * an application of the handler to the exception, and push it on
2713 * the top of the stack.
2715 * How exactly do we save all the active computations? We create an
2716 * AP_UPD for every UpdateFrame on the stack. Entering one of these
2717 * AP_UPDs pushes everything from the corresponding update frame
2718 * upwards onto the stack. (Actually, it pushes everything up to the
2719 * next update frame plus a pointer to the next AP_UPD object.
2720 * Entering the next AP_UPD object pushes more onto the stack until we
2721 * reach the last AP_UPD object - at which point the stack should look
2722 * exactly as it did when we killed the TSO and we can continue
2723 * execution by entering the closure on top of the stack.
2725 * We can also kill a thread entirely - this happens if either (a) the
2726 * exception passed to raiseAsync is NULL, or (b) there's no
2727 * CATCH_FRAME on the stack. In either case, we strip the entire
2728 * stack and replace the thread with a zombie.
2730 * -------------------------------------------------------------------------- */
2733 deleteThread(StgTSO *tso)
2735 raiseAsync(tso,NULL);
2739 raiseAsync(StgTSO *tso, StgClosure *exception)
2741 StgUpdateFrame* su = tso->su;
2742 StgPtr sp = tso->sp;
2744 /* Thread already dead? */
2745 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
2749 IF_DEBUG(scheduler, sched_belch("raising exception in thread %ld.", tso->id));
2751 /* Remove it from any blocking queues */
2754 /* The stack freezing code assumes there's a closure pointer on
2755 * the top of the stack. This isn't always the case with compiled
2756 * code, so we have to push a dummy closure on the top which just
2757 * returns to the next return address on the stack.
2759 if ( LOOKS_LIKE_GHC_INFO((void*)*sp) ) {
2760 *(--sp) = (W_)&dummy_ret_closure;
2764 int words = ((P_)su - (P_)sp) - 1;
2768 /* If we find a CATCH_FRAME, and we've got an exception to raise,
2769 * then build PAP(handler,exception,realworld#), and leave it on
2770 * top of the stack ready to enter.
2772 if (get_itbl(su)->type == CATCH_FRAME && exception != NULL) {
2773 StgCatchFrame *cf = (StgCatchFrame *)su;
2774 /* we've got an exception to raise, so let's pass it to the
2775 * handler in this frame.
2777 ap = (StgAP_UPD *)allocate(sizeofW(StgPAP) + 2);
2778 TICK_ALLOC_UPD_PAP(3,0);
2779 SET_HDR(ap,&PAP_info,cf->header.prof.ccs);
2782 ap->fun = cf->handler; /* :: Exception -> IO a */
2783 ap->payload[0] = exception;
2784 ap->payload[1] = ARG_TAG(0); /* realworld token */
2786 /* throw away the stack from Sp up to and including the
2789 sp = (P_)su + sizeofW(StgCatchFrame) - 1;
2792 /* Restore the blocked/unblocked state for asynchronous exceptions
2793 * at the CATCH_FRAME.
2795 * If exceptions were unblocked at the catch, arrange that they
2796 * are unblocked again after executing the handler by pushing an
2797 * unblockAsyncExceptions_ret stack frame.
2799 if (!cf->exceptions_blocked) {
2800 *(sp--) = (W_)&unblockAsyncExceptionszh_ret_info;
2803 /* Ensure that async exceptions are blocked when running the handler.
2805 if (tso->blocked_exceptions == NULL) {
2806 tso->blocked_exceptions = END_TSO_QUEUE;
2809 /* Put the newly-built PAP on top of the stack, ready to execute
2810 * when the thread restarts.
2814 tso->what_next = ThreadEnterGHC;
2815 IF_DEBUG(sanity, checkTSO(tso));
2819 /* First build an AP_UPD consisting of the stack chunk above the
2820 * current update frame, with the top word on the stack as the
2823 ap = (StgAP_UPD *)allocate(AP_sizeW(words));
2828 ap->fun = (StgClosure *)sp[0];
2830 for(i=0; i < (nat)words; ++i) {
2831 ap->payload[i] = (StgClosure *)*sp++;
2834 switch (get_itbl(su)->type) {
2838 SET_HDR(ap,&AP_UPD_info,su->header.prof.ccs /* ToDo */);
2839 TICK_ALLOC_UP_THK(words+1,0);
2842 fprintf(stderr, "scheduler: Updating ");
2843 printPtr((P_)su->updatee);
2844 fprintf(stderr, " with ");
2845 printObj((StgClosure *)ap);
2848 /* Replace the updatee with an indirection - happily
2849 * this will also wake up any threads currently
2850 * waiting on the result.
2852 UPD_IND_NOLOCK(su->updatee,ap); /* revert the black hole */
2854 sp += sizeofW(StgUpdateFrame) -1;
2855 sp[0] = (W_)ap; /* push onto stack */
2861 StgCatchFrame *cf = (StgCatchFrame *)su;
2864 /* We want a PAP, not an AP_UPD. Fortunately, the
2865 * layout's the same.
2867 SET_HDR(ap,&PAP_info,su->header.prof.ccs /* ToDo */);
2868 TICK_ALLOC_UPD_PAP(words+1,0);
2870 /* now build o = FUN(catch,ap,handler) */
2871 o = (StgClosure *)allocate(sizeofW(StgClosure)+2);
2872 TICK_ALLOC_FUN(2,0);
2873 SET_HDR(o,&catch_info,su->header.prof.ccs /* ToDo */);
2874 o->payload[0] = (StgClosure *)ap;
2875 o->payload[1] = cf->handler;
2878 fprintf(stderr, "scheduler: Built ");
2879 printObj((StgClosure *)o);
2882 /* pop the old handler and put o on the stack */
2884 sp += sizeofW(StgCatchFrame) - 1;
2891 StgSeqFrame *sf = (StgSeqFrame *)su;
2894 SET_HDR(ap,&PAP_info,su->header.prof.ccs /* ToDo */);
2895 TICK_ALLOC_UPD_PAP(words+1,0);
2897 /* now build o = FUN(seq,ap) */
2898 o = (StgClosure *)allocate(sizeofW(StgClosure)+1);
2899 TICK_ALLOC_SE_THK(1,0);
2900 SET_HDR(o,&seq_info,su->header.prof.ccs /* ToDo */);
2901 o->payload[0] = (StgClosure *)ap;
2904 fprintf(stderr, "scheduler: Built ");
2905 printObj((StgClosure *)o);
2908 /* pop the old handler and put o on the stack */
2910 sp += sizeofW(StgSeqFrame) - 1;
2916 /* We've stripped the entire stack, the thread is now dead. */
2917 sp += sizeofW(StgStopFrame) - 1;
2918 sp[0] = (W_)exception; /* save the exception */
2919 tso->what_next = ThreadKilled;
2920 tso->su = (StgUpdateFrame *)(sp+1);
2931 /* -----------------------------------------------------------------------------
2932 resurrectThreads is called after garbage collection on the list of
2933 threads found to be garbage. Each of these threads will be woken
2934 up and sent a signal: BlockedOnDeadMVar if the thread was blocked
2935 on an MVar, or NonTermination if the thread was blocked on a Black
2937 -------------------------------------------------------------------------- */
2940 resurrectThreads( StgTSO *threads )
2944 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
2945 next = tso->global_link;
2946 tso->global_link = all_threads;
2948 IF_DEBUG(scheduler, sched_belch("resurrecting thread %d", tso->id));
2950 switch (tso->why_blocked) {
2952 case BlockedOnException:
2953 raiseAsync(tso,(StgClosure *)BlockedOnDeadMVar_closure);
2955 case BlockedOnBlackHole:
2956 raiseAsync(tso,(StgClosure *)NonTermination_closure);
2959 /* This might happen if the thread was blocked on a black hole
2960 * belonging to a thread that we've just woken up (raiseAsync
2961 * can wake up threads, remember...).
2965 barf("resurrectThreads: thread blocked in a strange way");
2970 /* -----------------------------------------------------------------------------
2971 * Blackhole detection: if we reach a deadlock, test whether any
2972 * threads are blocked on themselves. Any threads which are found to
2973 * be self-blocked get sent a NonTermination exception.
2975 * This is only done in a deadlock situation in order to avoid
2976 * performance overhead in the normal case.
2977 * -------------------------------------------------------------------------- */
2980 detectBlackHoles( void )
2982 StgTSO *t = all_threads;
2983 StgUpdateFrame *frame;
2984 StgClosure *blocked_on;
2986 for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
2988 if (t->why_blocked != BlockedOnBlackHole) {
2992 blocked_on = t->block_info.closure;
2994 for (frame = t->su; ; frame = frame->link) {
2995 switch (get_itbl(frame)->type) {
2998 if (frame->updatee == blocked_on) {
2999 /* We are blocking on one of our own computations, so
3000 * send this thread the NonTermination exception.
3003 sched_belch("thread %d is blocked on itself", t->id));
3004 raiseAsync(t, (StgClosure *)NonTermination_closure);
3025 //@node Debugging Routines, Index, Exception Handling Routines, Main scheduling code
3026 //@subsection Debugging Routines
3028 /* -----------------------------------------------------------------------------
3029 Debugging: why is a thread blocked
3030 -------------------------------------------------------------------------- */
3035 printThreadBlockage(StgTSO *tso)
3037 switch (tso->why_blocked) {
3039 fprintf(stderr,"is blocked on read from fd %d", tso->block_info.fd);
3041 case BlockedOnWrite:
3042 fprintf(stderr,"is blocked on write to fd %d", tso->block_info.fd);
3044 case BlockedOnDelay:
3045 fprintf(stderr,"is blocked until %d", tso->block_info.target);
3048 fprintf(stderr,"is blocked on an MVar");
3050 case BlockedOnException:
3051 fprintf(stderr,"is blocked on delivering an exception to thread %d",
3052 tso->block_info.tso->id);
3054 case BlockedOnBlackHole:
3055 fprintf(stderr,"is blocked on a black hole");
3058 fprintf(stderr,"is not blocked");
3062 fprintf(stderr,"is blocked on global address; local FM_BQ is %p (%s)",
3063 tso->block_info.closure, info_type(tso->block_info.closure));
3065 case BlockedOnGA_NoSend:
3066 fprintf(stderr,"is blocked on global address (no send); local FM_BQ is %p (%s)",
3067 tso->block_info.closure, info_type(tso->block_info.closure));
3071 barf("printThreadBlockage: strange tso->why_blocked: %d for TSO %d (%d)",
3072 tso->why_blocked, tso->id, tso);
3077 printThreadStatus(StgTSO *tso)
3079 switch (tso->what_next) {
3081 fprintf(stderr,"has been killed");
3083 case ThreadComplete:
3084 fprintf(stderr,"has completed");
3087 printThreadBlockage(tso);
3092 printAllThreads(void)
3096 sched_belch("all threads:");
3097 for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
3098 fprintf(stderr, "\tthread %d ", t->id);
3099 printThreadStatus(t);
3100 fprintf(stderr,"\n");
3105 Print a whole blocking queue attached to node (debugging only).
3110 print_bq (StgClosure *node)
3112 StgBlockingQueueElement *bqe;
3116 fprintf(stderr,"## BQ of closure %p (%s): ",
3117 node, info_type(node));
3119 /* should cover all closures that may have a blocking queue */
3120 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
3121 get_itbl(node)->type == FETCH_ME_BQ ||
3122 get_itbl(node)->type == RBH);
3124 ASSERT(node!=(StgClosure*)NULL); // sanity check
3126 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
3128 for (bqe = ((StgBlockingQueue*)node)->blocking_queue, end = (bqe==END_BQ_QUEUE);
3129 !end; // iterate until bqe points to a CONSTR
3130 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) {
3131 ASSERT(bqe != END_BQ_QUEUE); // sanity check
3132 ASSERT(bqe != (StgTSO*)NULL); // sanity check
3133 /* types of closures that may appear in a blocking queue */
3134 ASSERT(get_itbl(bqe)->type == TSO ||
3135 get_itbl(bqe)->type == BLOCKED_FETCH ||
3136 get_itbl(bqe)->type == CONSTR);
3137 /* only BQs of an RBH end with an RBH_Save closure */
3138 ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
3140 switch (get_itbl(bqe)->type) {
3142 fprintf(stderr," TSO %d (%x),",
3143 ((StgTSO *)bqe)->id, ((StgTSO *)bqe));
3146 fprintf(stderr," BF (node=%p, ga=((%x, %d, %x)),",
3147 ((StgBlockedFetch *)bqe)->node,
3148 ((StgBlockedFetch *)bqe)->ga.payload.gc.gtid,
3149 ((StgBlockedFetch *)bqe)->ga.payload.gc.slot,
3150 ((StgBlockedFetch *)bqe)->ga.weight);
3153 fprintf(stderr," %s (IP %p),",
3154 (get_itbl(bqe) == &RBH_Save_0_info ? "RBH_Save_0" :
3155 get_itbl(bqe) == &RBH_Save_1_info ? "RBH_Save_1" :
3156 get_itbl(bqe) == &RBH_Save_2_info ? "RBH_Save_2" :
3157 "RBH_Save_?"), get_itbl(bqe));
3160 barf("Unexpected closure type %s in blocking queue of %p (%s)",
3161 info_type(bqe), node, info_type(node));
3165 fputc('\n', stderr);
3167 # elif defined(GRAN)
3169 print_bq (StgClosure *node)
3171 StgBlockingQueueElement *bqe;
3172 PEs node_loc, tso_loc;
3175 /* should cover all closures that may have a blocking queue */
3176 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
3177 get_itbl(node)->type == FETCH_ME_BQ ||
3178 get_itbl(node)->type == RBH);
3180 ASSERT(node!=(StgClosure*)NULL); // sanity check
3181 node_loc = where_is(node);
3183 fprintf(stderr,"## BQ of closure %p (%s) on [PE %d]: ",
3184 node, info_type(node), node_loc);
3187 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
3189 for (bqe = ((StgBlockingQueue*)node)->blocking_queue, end = (bqe==END_BQ_QUEUE);
3190 !end; // iterate until bqe points to a CONSTR
3191 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) {
3192 ASSERT(bqe != END_BQ_QUEUE); // sanity check
3193 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
3194 /* types of closures that may appear in a blocking queue */
3195 ASSERT(get_itbl(bqe)->type == TSO ||
3196 get_itbl(bqe)->type == CONSTR);
3197 /* only BQs of an RBH end with an RBH_Save closure */
3198 ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
3200 tso_loc = where_is((StgClosure *)bqe);
3201 switch (get_itbl(bqe)->type) {
3203 fprintf(stderr," TSO %d (%p) on [PE %d],",
3204 ((StgTSO *)bqe)->id, (StgTSO *)bqe, tso_loc);
3207 fprintf(stderr," %s (IP %p),",
3208 (get_itbl(bqe) == &RBH_Save_0_info ? "RBH_Save_0" :
3209 get_itbl(bqe) == &RBH_Save_1_info ? "RBH_Save_1" :
3210 get_itbl(bqe) == &RBH_Save_2_info ? "RBH_Save_2" :
3211 "RBH_Save_?"), get_itbl(bqe));
3214 barf("Unexpected closure type %s in blocking queue of %p (%s)",
3215 info_type((StgClosure *)bqe), node, info_type(node));
3219 fputc('\n', stderr);
3223 Nice and easy: only TSOs on the blocking queue
3226 print_bq (StgClosure *node)
3230 ASSERT(node!=(StgClosure*)NULL); // sanity check
3231 for (tso = ((StgBlockingQueue*)node)->blocking_queue;
3232 tso != END_TSO_QUEUE;
3234 ASSERT(tso!=NULL && tso!=END_TSO_QUEUE); // sanity check
3235 ASSERT(get_itbl(tso)->type == TSO); // guess what, sanity check
3236 fprintf(stderr," TSO %d (%p),", tso->id, tso);
3238 fputc('\n', stderr);
3249 for (i=0, tso=run_queue_hd;
3250 tso != END_TSO_QUEUE;
3259 sched_belch(char *s, ...)
3264 fprintf(stderr, "scheduler (task %ld): ", pthread_self());
3266 fprintf(stderr, "scheduler: ");
3268 vfprintf(stderr, s, ap);
3269 fprintf(stderr, "\n");
3275 //@node Index, , Debugging Routines, Main scheduling code
3279 //* MainRegTable:: @cindex\s-+MainRegTable
3280 //* StgMainThread:: @cindex\s-+StgMainThread
3281 //* awaken_blocked_queue:: @cindex\s-+awaken_blocked_queue
3282 //* blocked_queue_hd:: @cindex\s-+blocked_queue_hd
3283 //* blocked_queue_tl:: @cindex\s-+blocked_queue_tl
3284 //* context_switch:: @cindex\s-+context_switch
3285 //* createThread:: @cindex\s-+createThread
3286 //* free_capabilities:: @cindex\s-+free_capabilities
3287 //* gc_pending_cond:: @cindex\s-+gc_pending_cond
3288 //* initScheduler:: @cindex\s-+initScheduler
3289 //* interrupted:: @cindex\s-+interrupted
3290 //* n_free_capabilities:: @cindex\s-+n_free_capabilities
3291 //* next_thread_id:: @cindex\s-+next_thread_id
3292 //* print_bq:: @cindex\s-+print_bq
3293 //* run_queue_hd:: @cindex\s-+run_queue_hd
3294 //* run_queue_tl:: @cindex\s-+run_queue_tl
3295 //* sched_mutex:: @cindex\s-+sched_mutex
3296 //* schedule:: @cindex\s-+schedule
3297 //* take_off_run_queue:: @cindex\s-+take_off_run_queue
3298 //* task_ids:: @cindex\s-+task_ids
3299 //* term_mutex:: @cindex\s-+term_mutex
3300 //* thread_ready_cond:: @cindex\s-+thread_ready_cond