1 /* ---------------------------------------------------------------------------
2 * $Id: Schedule.c,v 1.72 2000/06/20 15:15:44 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;
150 /* Linked list of all threads.
151 * Used for detecting garbage collected threads.
155 /* Threads suspended in _ccall_GC.
157 static StgTSO *suspended_ccalling_threads;
159 static void GetRoots(void);
160 static StgTSO *threadStackOverflow(StgTSO *tso);
162 /* KH: The following two flags are shared memory locations. There is no need
163 to lock them, since they are only unset at the end of a scheduler
167 /* flag set by signal handler to precipitate a context switch */
168 //@cindex context_switch
171 /* if this flag is set as well, give up execution */
172 //@cindex interrupted
175 /* Next thread ID to allocate.
176 * Locks required: sched_mutex
178 //@cindex next_thread_id
179 StgThreadID next_thread_id = 1;
182 * Pointers to the state of the current thread.
183 * Rule of thumb: if CurrentTSO != NULL, then we're running a Haskell
184 * thread. If CurrentTSO == NULL, then we're at the scheduler level.
187 /* The smallest stack size that makes any sense is:
188 * RESERVED_STACK_WORDS (so we can get back from the stack overflow)
189 * + sizeofW(StgStopFrame) (the stg_stop_thread_info frame)
190 * + 1 (the realworld token for an IO thread)
191 * + 1 (the closure to enter)
193 * A thread with this stack will bomb immediately with a stack
194 * overflow, which will increase its stack size.
197 #define MIN_STACK_WORDS (RESERVED_STACK_WORDS + sizeofW(StgStopFrame) + 2)
199 /* Free capability list.
200 * Locks required: sched_mutex.
203 //@cindex free_capabilities
204 //@cindex n_free_capabilities
205 Capability *free_capabilities; /* Available capabilities for running threads */
206 nat n_free_capabilities; /* total number of available capabilities */
208 //@cindex MainRegTable
209 Capability MainRegTable; /* for non-SMP, we have one global capability */
218 /* All our current task ids, saved in case we need to kill them later.
225 void addToBlockedQueue ( StgTSO *tso );
227 static void schedule ( void );
228 void interruptStgRts ( void );
230 static StgTSO * createThread_ ( nat size, rtsBool have_lock, StgInt pri );
232 static StgTSO * createThread_ ( nat size, rtsBool have_lock );
235 static void detectBlackHoles ( void );
238 static void sched_belch(char *s, ...);
242 //@cindex sched_mutex
244 //@cindex thread_ready_cond
245 //@cindex gc_pending_cond
246 pthread_mutex_t sched_mutex = PTHREAD_MUTEX_INITIALIZER;
247 pthread_mutex_t term_mutex = PTHREAD_MUTEX_INITIALIZER;
248 pthread_cond_t thread_ready_cond = PTHREAD_COND_INITIALIZER;
249 pthread_cond_t gc_pending_cond = PTHREAD_COND_INITIALIZER;
256 rtsTime TimeOfLastYield;
260 char *whatNext_strs[] = {
268 char *threadReturnCode_strs[] = {
269 "HeapOverflow", /* might also be StackOverflow */
278 * The thread state for the main thread.
279 // ToDo: check whether not needed any more
283 //@node Main scheduling loop, Suspend and Resume, Prototypes, Main scheduling code
284 //@subsection Main scheduling loop
286 /* ---------------------------------------------------------------------------
287 Main scheduling loop.
289 We use round-robin scheduling, each thread returning to the
290 scheduler loop when one of these conditions is detected:
293 * timer expires (thread yields)
298 Locking notes: we acquire the scheduler lock once at the beginning
299 of the scheduler loop, and release it when
301 * running a thread, or
302 * waiting for work, or
303 * waiting for a GC to complete.
306 In a GranSim setup this loop iterates over the global event queue.
307 This revolves around the global event queue, which determines what
308 to do next. Therefore, it's more complicated than either the
309 concurrent or the parallel (GUM) setup.
312 GUM iterates over incoming messages.
313 It starts with nothing to do (thus CurrentTSO == END_TSO_QUEUE),
314 and sends out a fish whenever it has nothing to do; in-between
315 doing the actual reductions (shared code below) it processes the
316 incoming messages and deals with delayed operations
317 (see PendingFetches).
318 This is not the ugliest code you could imagine, but it's bloody close.
320 ------------------------------------------------------------------------ */
327 StgThreadReturnCode ret;
336 rtsBool was_interrupted = rtsFalse;
338 ACQUIRE_LOCK(&sched_mutex);
342 /* set up first event to get things going */
343 /* ToDo: assign costs for system setup and init MainTSO ! */
344 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
346 CurrentTSO, (StgClosure*)NULL, (rtsSpark*)NULL);
349 fprintf(stderr, "GRAN: Init CurrentTSO (in schedule) = %p\n", CurrentTSO);
350 G_TSO(CurrentTSO, 5));
352 if (RtsFlags.GranFlags.Light) {
353 /* Save current time; GranSim Light only */
354 CurrentTSO->gran.clock = CurrentTime[CurrentProc];
357 event = get_next_event();
359 while (event!=(rtsEvent*)NULL) {
360 /* Choose the processor with the next event */
361 CurrentProc = event->proc;
362 CurrentTSO = event->tso;
366 while (!GlobalStopPending) { /* GlobalStopPending set in par_exit */
374 IF_DEBUG(scheduler, printAllThreads());
376 /* If we're interrupted (the user pressed ^C, or some other
377 * termination condition occurred), kill all the currently running
381 IF_DEBUG(scheduler, sched_belch("interrupted"));
382 for (t = run_queue_hd; t != END_TSO_QUEUE; t = t->link) {
385 for (t = blocked_queue_hd; t != END_TSO_QUEUE; t = t->link) {
388 run_queue_hd = run_queue_tl = END_TSO_QUEUE;
389 blocked_queue_hd = blocked_queue_tl = END_TSO_QUEUE;
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) {
511 (run_queue_hd == END_TSO_QUEUE)
513 && (n_free_capabilities == RtsFlags.ParFlags.nNodes)
518 /* check for signals each time around the scheduler */
520 if (signals_pending()) {
521 start_signal_handlers();
526 * Detect deadlock: when we have no threads to run, there are no
527 * threads waiting on I/O or sleeping, and all the other tasks are
528 * waiting for work, we must have a deadlock of some description.
530 * We first try to find threads blocked on themselves (ie. black
531 * holes), and generate NonTermination exceptions where necessary.
533 * If no threads are black holed, we have a deadlock situation, so
534 * inform all the main threads.
537 if (blocked_queue_hd == END_TSO_QUEUE
538 && run_queue_hd == END_TSO_QUEUE
539 && (n_free_capabilities == RtsFlags.ParFlags.nNodes))
541 IF_DEBUG(scheduler, sched_belch("deadlocked, checking for black holes..."));
543 if (run_queue_hd == END_TSO_QUEUE) {
545 for (m = main_threads; m != NULL; m = m->link) {
548 pthread_cond_broadcast(&m->wakeup);
554 if (blocked_queue_hd == END_TSO_QUEUE
555 && run_queue_hd == END_TSO_QUEUE)
557 IF_DEBUG(scheduler, sched_belch("deadlocked, checking for black holes..."));
559 if (run_queue_hd == END_TSO_QUEUE) {
560 StgMainThread *m = main_threads;
563 main_threads = m->link;
570 /* If there's a GC pending, don't do anything until it has
574 IF_DEBUG(scheduler,sched_belch("waiting for GC"));
575 pthread_cond_wait(&gc_pending_cond, &sched_mutex);
578 /* block until we've got a thread on the run queue and a free
581 while (run_queue_hd == END_TSO_QUEUE || free_capabilities == NULL) {
582 IF_DEBUG(scheduler, sched_belch("waiting for work"));
583 pthread_cond_wait(&thread_ready_cond, &sched_mutex);
584 IF_DEBUG(scheduler, sched_belch("work now available"));
590 if (RtsFlags.GranFlags.Light)
591 GranSimLight_enter_system(event, &ActiveTSO); // adjust ActiveTSO etc
593 /* adjust time based on time-stamp */
594 if (event->time > CurrentTime[CurrentProc] &&
595 event->evttype != ContinueThread)
596 CurrentTime[CurrentProc] = event->time;
598 /* Deal with the idle PEs (may issue FindWork or MoveSpark events) */
599 if (!RtsFlags.GranFlags.Light)
602 IF_DEBUG(gran, fprintf(stderr, "GRAN: switch by event-type\n"))
604 /* main event dispatcher in GranSim */
605 switch (event->evttype) {
606 /* Should just be continuing execution */
608 IF_DEBUG(gran, fprintf(stderr, "GRAN: doing ContinueThread\n"));
609 /* ToDo: check assertion
610 ASSERT(run_queue_hd != (StgTSO*)NULL &&
611 run_queue_hd != END_TSO_QUEUE);
613 /* Ignore ContinueThreads for fetching threads (if synchr comm) */
614 if (!RtsFlags.GranFlags.DoAsyncFetch &&
615 procStatus[CurrentProc]==Fetching) {
616 belch("ghuH: Spurious ContinueThread while Fetching ignored; TSO %d (%p) [PE %d]",
617 CurrentTSO->id, CurrentTSO, CurrentProc);
620 /* Ignore ContinueThreads for completed threads */
621 if (CurrentTSO->what_next == ThreadComplete) {
622 belch("ghuH: found a ContinueThread event for completed thread %d (%p) [PE %d] (ignoring ContinueThread)",
623 CurrentTSO->id, CurrentTSO, CurrentProc);
626 /* Ignore ContinueThreads for threads that are being migrated */
627 if (PROCS(CurrentTSO)==Nowhere) {
628 belch("ghuH: trying to run the migrating TSO %d (%p) [PE %d] (ignoring ContinueThread)",
629 CurrentTSO->id, CurrentTSO, CurrentProc);
632 /* The thread should be at the beginning of the run queue */
633 if (CurrentTSO!=run_queue_hds[CurrentProc]) {
634 belch("ghuH: TSO %d (%p) [PE %d] is not at the start of the run_queue when doing a ContinueThread",
635 CurrentTSO->id, CurrentTSO, CurrentProc);
636 break; // run the thread anyway
639 new_event(proc, proc, CurrentTime[proc],
641 (StgTSO*)NULL, (StgClosure*)NULL, (rtsSpark*)NULL);
643 */ /* Catches superfluous CONTINUEs -- should be unnecessary */
644 break; // now actually run the thread; DaH Qu'vam yImuHbej
647 do_the_fetchnode(event);
648 goto next_thread; /* handle next event in event queue */
651 do_the_globalblock(event);
652 goto next_thread; /* handle next event in event queue */
655 do_the_fetchreply(event);
656 goto next_thread; /* handle next event in event queue */
658 case UnblockThread: /* Move from the blocked queue to the tail of */
659 do_the_unblock(event);
660 goto next_thread; /* handle next event in event queue */
662 case ResumeThread: /* Move from the blocked queue to the tail of */
663 /* the runnable queue ( i.e. Qu' SImqa'lu') */
664 event->tso->gran.blocktime +=
665 CurrentTime[CurrentProc] - event->tso->gran.blockedat;
666 do_the_startthread(event);
667 goto next_thread; /* handle next event in event queue */
670 do_the_startthread(event);
671 goto next_thread; /* handle next event in event queue */
674 do_the_movethread(event);
675 goto next_thread; /* handle next event in event queue */
678 do_the_movespark(event);
679 goto next_thread; /* handle next event in event queue */
682 do_the_findwork(event);
683 goto next_thread; /* handle next event in event queue */
686 barf("Illegal event type %u\n", event->evttype);
689 /* This point was scheduler_loop in the old RTS */
691 IF_DEBUG(gran, belch("GRAN: after main switch"));
693 TimeOfLastEvent = CurrentTime[CurrentProc];
694 TimeOfNextEvent = get_time_of_next_event();
695 IgnoreEvents=(TimeOfNextEvent==0); // HWL HACK
696 // CurrentTSO = ThreadQueueHd;
698 IF_DEBUG(gran, belch("GRAN: time of next event is: %ld",
701 if (RtsFlags.GranFlags.Light)
702 GranSimLight_leave_system(event, &ActiveTSO);
704 EndOfTimeSlice = CurrentTime[CurrentProc]+RtsFlags.GranFlags.time_slice;
707 belch("GRAN: end of time-slice is %#lx", EndOfTimeSlice));
709 /* in a GranSim setup the TSO stays on the run queue */
711 /* Take a thread from the run queue. */
712 t = POP_RUN_QUEUE(); // take_off_run_queue(t);
715 fprintf(stderr, "GRAN: About to run current thread, which is\n");
718 context_switch = 0; // turned on via GranYield, checking events and time slice
721 DumpGranEvent(GR_SCHEDULE, t));
723 procStatus[CurrentProc] = Busy;
727 if (PendingFetches != END_BF_QUEUE) {
731 /* ToDo: phps merge with spark activation above */
732 /* check whether we have local work and send requests if we have none */
733 if (run_queue_hd == END_TSO_QUEUE) { /* no runnable threads */
734 /* :-[ no local threads => look out for local sparks */
735 /* the spark pool for the current PE */
736 pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable
737 if (advisory_thread_count < RtsFlags.ParFlags.maxThreads &&
738 pool->hd < pool->tl) {
740 * ToDo: add GC code check that we really have enough heap afterwards!!
742 * If we're here (no runnable threads) and we have pending
743 * sparks, we must have a space problem. Get enough space
744 * to turn one of those pending sparks into a
748 spark = findSpark(); /* get a spark */
749 if (spark != (rtsSpark) NULL) {
750 tso = activateSpark(spark); /* turn the spark into a thread */
751 IF_PAR_DEBUG(schedule,
752 belch("==== schedule: Created TSO %d (%p); %d threads active",
753 tso->id, tso, advisory_thread_count));
755 if (tso==END_TSO_QUEUE) { /* failed to activate spark->back to loop */
756 belch("==^^ failed to activate spark");
758 } /* otherwise fall through & pick-up new tso */
760 IF_PAR_DEBUG(verbose,
761 belch("==^^ no local sparks (spark pool contains only NFs: %d)",
762 spark_queue_len(pool)));
766 /* =8-[ no local sparks => look for work on other PEs */
769 * We really have absolutely no work. Send out a fish
770 * (there may be some out there already), and wait for
771 * something to arrive. We clearly can't run any threads
772 * until a SCHEDULE or RESUME arrives, and so that's what
773 * we're hoping to see. (Of course, we still have to
774 * respond to other types of messages.)
777 outstandingFishes < RtsFlags.ParFlags.maxFishes ) { // &&
778 // (last_fish_arrived_at+FISH_DELAY < CURRENT_TIME)) {
779 /* fishing set in sendFish, processFish;
780 avoid flooding system with fishes via delay */
782 sendFish(pe, mytid, NEW_FISH_AGE, NEW_FISH_HISTORY,
790 } else if (PacketsWaiting()) { /* Look for incoming messages */
794 /* Now we are sure that we have some work available */
795 ASSERT(run_queue_hd != END_TSO_QUEUE);
796 /* Take a thread from the run queue, if we have work */
797 t = POP_RUN_QUEUE(); // take_off_run_queue(END_TSO_QUEUE);
799 /* ToDo: write something to the log-file
800 if (RTSflags.ParFlags.granSimStats && !sameThread)
801 DumpGranEvent(GR_SCHEDULE, RunnableThreadsHd);
805 /* the spark pool for the current PE */
806 pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable
808 IF_DEBUG(scheduler, belch("--^^ %d sparks on [%#x] (hd=%x; tl=%x; base=%x, lim=%x)",
809 spark_queue_len(pool),
811 pool->hd, pool->tl, pool->base, pool->lim));
813 IF_DEBUG(scheduler, belch("--== %d threads on [%#x] (hd=%x; tl=%x)",
814 run_queue_len(), CURRENT_PROC,
815 run_queue_hd, run_queue_tl));
820 we are running a different TSO, so write a schedule event to log file
821 NB: If we use fair scheduling we also have to write a deschedule
822 event for LastTSO; with unfair scheduling we know that the
823 previous tso has blocked whenever we switch to another tso, so
824 we don't need it in GUM for now
826 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
827 GR_SCHEDULE, t, (StgClosure *)NULL, 0, 0);
831 #else /* !GRAN && !PAR */
833 /* grab a thread from the run queue
835 ASSERT(run_queue_hd != END_TSO_QUEUE);
837 IF_DEBUG(sanity,checkTSO(t));
844 cap = free_capabilities;
845 free_capabilities = cap->link;
846 n_free_capabilities--;
851 cap->rCurrentTSO = t;
853 /* context switches are now initiated by the timer signal */
856 RELEASE_LOCK(&sched_mutex);
858 IF_DEBUG(scheduler, sched_belch("-->> Running TSO %ld (%p) %s ...",
859 t->id, t, whatNext_strs[t->what_next]));
861 /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
862 /* Run the current thread
864 switch (cap->rCurrentTSO->what_next) {
867 /* Thread already finished, return to scheduler. */
868 ret = ThreadFinished;
871 ret = StgRun((StgFunPtr) stg_enterStackTop, cap);
874 ret = StgRun((StgFunPtr) stg_returnToStackTop, cap);
876 case ThreadEnterHugs:
880 IF_DEBUG(scheduler,sched_belch("entering Hugs"));
881 c = (StgClosure *)(cap->rCurrentTSO->sp[0]);
882 cap->rCurrentTSO->sp += 1;
887 barf("Panic: entered a BCO but no bytecode interpreter in this build");
890 barf("schedule: invalid what_next field");
892 /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
894 /* Costs for the scheduler are assigned to CCS_SYSTEM */
899 ACQUIRE_LOCK(&sched_mutex);
902 IF_DEBUG(scheduler,fprintf(stderr,"scheduler (task %ld): ", pthread_self()););
903 #elif !defined(GRAN) && !defined(PAR)
904 IF_DEBUG(scheduler,fprintf(stderr,"scheduler: "););
906 t = cap->rCurrentTSO;
909 /* HACK 675: if the last thread didn't yield, make sure to print a
910 SCHEDULE event to the log file when StgRunning the next thread, even
911 if it is the same one as before */
912 LastTSO = t; //(ret == ThreadBlocked) ? END_TSO_QUEUE : t;
913 TimeOfLastYield = CURRENT_TIME;
918 /* make all the running tasks block on a condition variable,
919 * maybe set context_switch and wait till they all pile in,
920 * then have them wait on a GC condition variable.
922 IF_DEBUG(scheduler,belch("--<< thread %ld (%p; %s) stopped: HeapOverflow",
923 t->id, t, whatNext_strs[t->what_next]));
926 ASSERT(!is_on_queue(t,CurrentProc));
929 ready_to_gc = rtsTrue;
930 context_switch = 1; /* stop other threads ASAP */
931 PUSH_ON_RUN_QUEUE(t);
932 /* actual GC is done at the end of the while loop */
936 IF_DEBUG(scheduler,belch("--<< thread %ld (%p; %s) stopped, StackOverflow",
937 t->id, t, whatNext_strs[t->what_next]));
938 /* just adjust the stack for this thread, then pop it back
944 /* enlarge the stack */
945 StgTSO *new_t = threadStackOverflow(t);
947 /* This TSO has moved, so update any pointers to it from the
948 * main thread stack. It better not be on any other queues...
951 for (m = main_threads; m != NULL; m = m->link) {
957 PUSH_ON_RUN_QUEUE(new_t);
964 DumpGranEvent(GR_DESCHEDULE, t));
965 globalGranStats.tot_yields++;
968 DumpGranEvent(GR_DESCHEDULE, t));
970 /* put the thread back on the run queue. Then, if we're ready to
971 * GC, check whether this is the last task to stop. If so, wake
972 * up the GC thread. getThread will block during a GC until the
976 if (t->what_next == ThreadEnterHugs) {
977 /* ToDo: or maybe a timer expired when we were in Hugs?
978 * or maybe someone hit ctrl-C
980 belch("--<< thread %ld (%p; %s) stopped to switch to Hugs",
981 t->id, t, whatNext_strs[t->what_next]);
983 belch("--<< thread %ld (%p; %s) stopped, yielding",
984 t->id, t, whatNext_strs[t->what_next]);
991 //belch("&& Doing sanity check on yielding TSO %ld.", t->id);
993 ASSERT(t->link == END_TSO_QUEUE);
995 ASSERT(!is_on_queue(t,CurrentProc));
998 //belch("&& Doing sanity check on all ThreadQueues (and their TSOs).");
999 checkThreadQsSanity(rtsTrue));
1001 APPEND_TO_RUN_QUEUE(t);
1003 /* add a ContinueThread event to actually process the thread */
1004 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1006 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1008 belch("GRAN: eventq and runnableq after adding yielded thread to queue again:");
1017 belch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: ",
1018 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)));
1019 if (t->block_info.closure!=(StgClosure*)NULL) print_bq(t->block_info.closure));
1021 // ??? needed; should emit block before
1023 DumpGranEvent(GR_DESCHEDULE, t));
1024 prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
1027 ASSERT(procStatus[CurrentProc]==Busy ||
1028 ((procStatus[CurrentProc]==Fetching) &&
1029 (t->block_info.closure!=(StgClosure*)NULL)));
1030 if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
1031 !(!RtsFlags.GranFlags.DoAsyncFetch &&
1032 procStatus[CurrentProc]==Fetching))
1033 procStatus[CurrentProc] = Idle;
1037 DumpGranEvent(GR_DESCHEDULE, t));
1039 /* Send a fetch (if BlockedOnGA) and dump event to log file */
1043 belch("--<< thread %ld (%p; %s) stopped, blocking on node %p with BQ: ",
1044 t->id, t, whatNext_strs[t->what_next], t->block_info.closure);
1045 if (t->block_info.closure!=(StgClosure*)NULL) print_bq(t->block_info.closure));
1048 /* don't need to do anything. Either the thread is blocked on
1049 * I/O, in which case we'll have called addToBlockedQueue
1050 * previously, or it's blocked on an MVar or Blackhole, in which
1051 * case it'll be on the relevant queue already.
1054 fprintf(stderr, "--<< thread %d (%p) stopped: ", t->id, t);
1055 printThreadBlockage(t);
1056 fprintf(stderr, "\n"));
1058 /* Only for dumping event to log file
1059 ToDo: do I need this in GranSim, too?
1066 case ThreadFinished:
1067 /* Need to check whether this was a main thread, and if so, signal
1068 * the task that started it with the return value. If we have no
1069 * more main threads, we probably need to stop all the tasks until
1072 /* We also end up here if the thread kills itself with an
1073 * uncaught exception, see Exception.hc.
1075 IF_DEBUG(scheduler,belch("--++ thread %d (%p) finished", t->id, t));
1077 endThread(t, CurrentProc); // clean-up the thread
1079 advisory_thread_count--;
1080 if (RtsFlags.ParFlags.ParStats.Full)
1081 DumpEndEvent(CURRENT_PROC, t, rtsFalse /* not mandatory */);
1086 barf("schedule: invalid thread return code %d", (int)ret);
1090 cap->link = free_capabilities;
1091 free_capabilities = cap;
1092 n_free_capabilities++;
1096 if (ready_to_gc && n_free_capabilities == RtsFlags.ParFlags.nNodes)
1101 /* everybody back, start the GC.
1102 * Could do it in this thread, or signal a condition var
1103 * to do it in another thread. Either way, we need to
1104 * broadcast on gc_pending_cond afterward.
1107 IF_DEBUG(scheduler,sched_belch("doing GC"));
1109 GarbageCollect(GetRoots,rtsFalse);
1110 ready_to_gc = rtsFalse;
1112 pthread_cond_broadcast(&gc_pending_cond);
1115 /* add a ContinueThread event to continue execution of current thread */
1116 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1118 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1120 fprintf(stderr, "GRAN: eventq and runnableq after Garbage collection:\n");
1127 IF_GRAN_DEBUG(unused,
1128 print_eventq(EventHd));
1130 event = get_next_event();
1134 /* ToDo: wait for next message to arrive rather than busy wait */
1139 t = take_off_run_queue(END_TSO_QUEUE);
1142 } /* end of while(1) */
1145 /* A hack for Hugs concurrency support. Needs sanitisation (?) */
1146 void deleteAllThreads ( void )
1149 IF_DEBUG(scheduler,sched_belch("deleteAllThreads()"));
1150 for (t = run_queue_hd; t != END_TSO_QUEUE; t = t->link) {
1153 for (t = blocked_queue_hd; t != END_TSO_QUEUE; t = t->link) {
1156 run_queue_hd = run_queue_tl = END_TSO_QUEUE;
1157 blocked_queue_hd = blocked_queue_tl = END_TSO_QUEUE;
1160 /* startThread and insertThread are now in GranSim.c -- HWL */
1162 //@node Suspend and Resume, Run queue code, Main scheduling loop, Main scheduling code
1163 //@subsection Suspend and Resume
1165 /* ---------------------------------------------------------------------------
1166 * Suspending & resuming Haskell threads.
1168 * When making a "safe" call to C (aka _ccall_GC), the task gives back
1169 * its capability before calling the C function. This allows another
1170 * task to pick up the capability and carry on running Haskell
1171 * threads. It also means that if the C call blocks, it won't lock
1174 * The Haskell thread making the C call is put to sleep for the
1175 * duration of the call, on the susepended_ccalling_threads queue. We
1176 * give out a token to the task, which it can use to resume the thread
1177 * on return from the C function.
1178 * ------------------------------------------------------------------------- */
1181 suspendThread( Capability *cap )
1185 ACQUIRE_LOCK(&sched_mutex);
1188 sched_belch("thread %d did a _ccall_gc", cap->rCurrentTSO->id));
1190 threadPaused(cap->rCurrentTSO);
1191 cap->rCurrentTSO->link = suspended_ccalling_threads;
1192 suspended_ccalling_threads = cap->rCurrentTSO;
1194 /* Use the thread ID as the token; it should be unique */
1195 tok = cap->rCurrentTSO->id;
1198 cap->link = free_capabilities;
1199 free_capabilities = cap;
1200 n_free_capabilities++;
1203 RELEASE_LOCK(&sched_mutex);
1208 resumeThread( StgInt tok )
1210 StgTSO *tso, **prev;
1213 ACQUIRE_LOCK(&sched_mutex);
1215 prev = &suspended_ccalling_threads;
1216 for (tso = suspended_ccalling_threads;
1217 tso != END_TSO_QUEUE;
1218 prev = &tso->link, tso = tso->link) {
1219 if (tso->id == (StgThreadID)tok) {
1224 if (tso == END_TSO_QUEUE) {
1225 barf("resumeThread: thread not found");
1229 while (free_capabilities == NULL) {
1230 IF_DEBUG(scheduler, sched_belch("waiting to resume"));
1231 pthread_cond_wait(&thread_ready_cond, &sched_mutex);
1232 IF_DEBUG(scheduler, sched_belch("resuming thread %d", tso->id));
1234 cap = free_capabilities;
1235 free_capabilities = cap->link;
1236 n_free_capabilities--;
1238 cap = &MainRegTable;
1241 cap->rCurrentTSO = tso;
1243 RELEASE_LOCK(&sched_mutex);
1248 /* ---------------------------------------------------------------------------
1250 * ------------------------------------------------------------------------ */
1251 static void unblockThread(StgTSO *tso);
1253 /* ---------------------------------------------------------------------------
1254 * Comparing Thread ids.
1256 * This is used from STG land in the implementation of the
1257 * instances of Eq/Ord for ThreadIds.
1258 * ------------------------------------------------------------------------ */
1260 int cmp_thread(const StgTSO *tso1, const StgTSO *tso2)
1262 StgThreadID id1 = tso1->id;
1263 StgThreadID id2 = tso2->id;
1265 if (id1 < id2) return (-1);
1266 if (id1 > id2) return 1;
1270 /* ---------------------------------------------------------------------------
1271 Create a new thread.
1273 The new thread starts with the given stack size. Before the
1274 scheduler can run, however, this thread needs to have a closure
1275 (and possibly some arguments) pushed on its stack. See
1276 pushClosure() in Schedule.h.
1278 createGenThread() and createIOThread() (in SchedAPI.h) are
1279 convenient packaged versions of this function.
1281 currently pri (priority) is only used in a GRAN setup -- HWL
1282 ------------------------------------------------------------------------ */
1283 //@cindex createThread
1285 /* currently pri (priority) is only used in a GRAN setup -- HWL */
1287 createThread(nat stack_size, StgInt pri)
1289 return createThread_(stack_size, rtsFalse, pri);
1293 createThread_(nat size, rtsBool have_lock, StgInt pri)
1297 createThread(nat stack_size)
1299 return createThread_(stack_size, rtsFalse);
1303 createThread_(nat size, rtsBool have_lock)
1310 /* First check whether we should create a thread at all */
1312 /* check that no more than RtsFlags.ParFlags.maxThreads threads are created */
1313 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads) {
1315 belch("{createThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
1316 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
1317 return END_TSO_QUEUE;
1323 ASSERT(!RtsFlags.GranFlags.Light || CurrentProc==0);
1326 // ToDo: check whether size = stack_size - TSO_STRUCT_SIZEW
1328 /* catch ridiculously small stack sizes */
1329 if (size < MIN_STACK_WORDS + TSO_STRUCT_SIZEW) {
1330 size = MIN_STACK_WORDS + TSO_STRUCT_SIZEW;
1333 stack_size = size - TSO_STRUCT_SIZEW;
1335 tso = (StgTSO *)allocate(size);
1336 TICK_ALLOC_TSO(size-TSO_STRUCT_SIZEW, 0);
1338 SET_HDR(tso, &TSO_info, CCS_SYSTEM);
1340 SET_GRAN_HDR(tso, ThisPE);
1342 tso->what_next = ThreadEnterGHC;
1344 /* tso->id needs to be unique. For now we use a heavyweight mutex to
1345 * protect the increment operation on next_thread_id.
1346 * In future, we could use an atomic increment instead.
1348 if (!have_lock) { ACQUIRE_LOCK(&sched_mutex); }
1349 tso->id = next_thread_id++;
1350 if (!have_lock) { RELEASE_LOCK(&sched_mutex); }
1352 tso->why_blocked = NotBlocked;
1353 tso->blocked_exceptions = NULL;
1355 tso->splim = (P_)&(tso->stack) + RESERVED_STACK_WORDS;
1356 tso->stack_size = stack_size;
1357 tso->max_stack_size = round_to_mblocks(RtsFlags.GcFlags.maxStkSize)
1359 tso->sp = (P_)&(tso->stack) + stack_size;
1362 tso->prof.CCCS = CCS_MAIN;
1365 /* put a stop frame on the stack */
1366 tso->sp -= sizeofW(StgStopFrame);
1367 SET_HDR((StgClosure*)tso->sp,(StgInfoTable *)&stg_stop_thread_info,CCS_SYSTEM);
1368 tso->su = (StgUpdateFrame*)tso->sp;
1372 tso->link = END_TSO_QUEUE;
1373 /* uses more flexible routine in GranSim */
1374 insertThread(tso, CurrentProc);
1376 /* In a non-GranSim setup the pushing of a TSO onto the runq is separated
1381 #if defined(GRAN) || defined(PAR)
1382 DumpGranEvent(GR_START,tso);
1385 /* Link the new thread on the global thread list.
1387 tso->global_link = all_threads;
1391 tso->gran.pri = pri;
1393 tso->gran.magic = TSO_MAGIC; // debugging only
1395 tso->gran.sparkname = 0;
1396 tso->gran.startedat = CURRENT_TIME;
1397 tso->gran.exported = 0;
1398 tso->gran.basicblocks = 0;
1399 tso->gran.allocs = 0;
1400 tso->gran.exectime = 0;
1401 tso->gran.fetchtime = 0;
1402 tso->gran.fetchcount = 0;
1403 tso->gran.blocktime = 0;
1404 tso->gran.blockcount = 0;
1405 tso->gran.blockedat = 0;
1406 tso->gran.globalsparks = 0;
1407 tso->gran.localsparks = 0;
1408 if (RtsFlags.GranFlags.Light)
1409 tso->gran.clock = Now; /* local clock */
1411 tso->gran.clock = 0;
1413 IF_DEBUG(gran,printTSO(tso));
1416 tso->par.magic = TSO_MAGIC; // debugging only
1418 tso->par.sparkname = 0;
1419 tso->par.startedat = CURRENT_TIME;
1420 tso->par.exported = 0;
1421 tso->par.basicblocks = 0;
1422 tso->par.allocs = 0;
1423 tso->par.exectime = 0;
1424 tso->par.fetchtime = 0;
1425 tso->par.fetchcount = 0;
1426 tso->par.blocktime = 0;
1427 tso->par.blockcount = 0;
1428 tso->par.blockedat = 0;
1429 tso->par.globalsparks = 0;
1430 tso->par.localsparks = 0;
1434 globalGranStats.tot_threads_created++;
1435 globalGranStats.threads_created_on_PE[CurrentProc]++;
1436 globalGranStats.tot_sq_len += spark_queue_len(CurrentProc);
1437 globalGranStats.tot_sq_probes++;
1442 belch("==__ schedule: Created TSO %d (%p);",
1443 CurrentProc, tso, tso->id));
1445 IF_PAR_DEBUG(verbose,
1446 belch("==__ schedule: Created TSO %d (%p); %d threads active",
1447 tso->id, tso, advisory_thread_count));
1449 IF_DEBUG(scheduler,sched_belch("created thread %ld, stack size = %lx words",
1450 tso->id, tso->stack_size));
1456 Turn a spark into a thread.
1457 ToDo: fix for SMP (needs to acquire SCHED_MUTEX!)
1460 //@cindex activateSpark
1462 activateSpark (rtsSpark spark)
1466 ASSERT(spark != (rtsSpark)NULL);
1467 tso = createThread_(RtsFlags.GcFlags.initialStkSize, rtsTrue);
1468 if (tso!=END_TSO_QUEUE) {
1469 pushClosure(tso,spark);
1470 PUSH_ON_RUN_QUEUE(tso);
1471 advisory_thread_count++;
1473 if (RtsFlags.ParFlags.ParStats.Full) {
1474 //ASSERT(run_queue_hd == END_TSO_QUEUE); // I think ...
1475 IF_PAR_DEBUG(verbose,
1476 belch("==^^ activateSpark: turning spark of closure %p (%s) into a thread",
1477 (StgClosure *)spark, info_type((StgClosure *)spark)));
1480 barf("activateSpark: Cannot create TSO");
1482 // ToDo: fwd info on local/global spark to thread -- HWL
1483 // tso->gran.exported = spark->exported;
1484 // tso->gran.locked = !spark->global;
1485 // tso->gran.sparkname = spark->name;
1491 /* ---------------------------------------------------------------------------
1494 * scheduleThread puts a thread on the head of the runnable queue.
1495 * This will usually be done immediately after a thread is created.
1496 * The caller of scheduleThread must create the thread using e.g.
1497 * createThread and push an appropriate closure
1498 * on this thread's stack before the scheduler is invoked.
1499 * ------------------------------------------------------------------------ */
1502 scheduleThread(StgTSO *tso)
1504 if (tso==END_TSO_QUEUE){
1509 ACQUIRE_LOCK(&sched_mutex);
1511 /* Put the new thread on the head of the runnable queue. The caller
1512 * better push an appropriate closure on this thread's stack
1513 * beforehand. In the SMP case, the thread may start running as
1514 * soon as we release the scheduler lock below.
1516 PUSH_ON_RUN_QUEUE(tso);
1520 IF_DEBUG(scheduler,printTSO(tso));
1522 RELEASE_LOCK(&sched_mutex);
1525 /* ---------------------------------------------------------------------------
1528 * Start up Posix threads to run each of the scheduler tasks.
1529 * I believe the task ids are not needed in the system as defined.
1531 * ------------------------------------------------------------------------ */
1533 #if defined(PAR) || defined(SMP)
1535 taskStart( void *arg STG_UNUSED )
1537 rts_evalNothing(NULL);
1541 /* ---------------------------------------------------------------------------
1544 * Initialise the scheduler. This resets all the queues - if the
1545 * queues contained any threads, they'll be garbage collected at the
1548 * This now calls startTasks(), so should only be called once! KH @ 25/10/99
1549 * ------------------------------------------------------------------------ */
1553 term_handler(int sig STG_UNUSED)
1556 ACQUIRE_LOCK(&term_mutex);
1558 RELEASE_LOCK(&term_mutex);
1563 //@cindex initScheduler
1570 for (i=0; i<=MAX_PROC; i++) {
1571 run_queue_hds[i] = END_TSO_QUEUE;
1572 run_queue_tls[i] = END_TSO_QUEUE;
1573 blocked_queue_hds[i] = END_TSO_QUEUE;
1574 blocked_queue_tls[i] = END_TSO_QUEUE;
1575 ccalling_threadss[i] = END_TSO_QUEUE;
1578 run_queue_hd = END_TSO_QUEUE;
1579 run_queue_tl = END_TSO_QUEUE;
1580 blocked_queue_hd = END_TSO_QUEUE;
1581 blocked_queue_tl = END_TSO_QUEUE;
1584 suspended_ccalling_threads = END_TSO_QUEUE;
1586 main_threads = NULL;
1587 all_threads = END_TSO_QUEUE;
1593 ecafList = END_ECAF_LIST;
1597 /* Install the SIGHUP handler */
1600 struct sigaction action,oact;
1602 action.sa_handler = term_handler;
1603 sigemptyset(&action.sa_mask);
1604 action.sa_flags = 0;
1605 if (sigaction(SIGTERM, &action, &oact) != 0) {
1606 barf("can't install TERM handler");
1612 /* Allocate N Capabilities */
1615 Capability *cap, *prev;
1618 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
1619 cap = stgMallocBytes(sizeof(Capability), "initScheduler:capabilities");
1623 free_capabilities = cap;
1624 n_free_capabilities = RtsFlags.ParFlags.nNodes;
1626 IF_DEBUG(scheduler,fprintf(stderr,"scheduler: Allocated %d capabilities\n",
1627 n_free_capabilities););
1630 #if defined(SMP) || defined(PAR)
1643 /* make some space for saving all the thread ids */
1644 task_ids = stgMallocBytes(RtsFlags.ParFlags.nNodes * sizeof(task_info),
1645 "initScheduler:task_ids");
1647 /* and create all the threads */
1648 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
1649 r = pthread_create(&tid,NULL,taskStart,NULL);
1651 barf("startTasks: Can't create new Posix thread");
1653 task_ids[i].id = tid;
1654 task_ids[i].mut_time = 0.0;
1655 task_ids[i].mut_etime = 0.0;
1656 task_ids[i].gc_time = 0.0;
1657 task_ids[i].gc_etime = 0.0;
1658 task_ids[i].elapsedtimestart = elapsedtime();
1659 IF_DEBUG(scheduler,fprintf(stderr,"scheduler: Started task: %ld\n",tid););
1665 exitScheduler( void )
1670 /* Don't want to use pthread_cancel, since we'd have to install
1671 * these silly exception handlers (pthread_cleanup_{push,pop}) around
1675 /* Cancel all our tasks */
1676 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
1677 pthread_cancel(task_ids[i].id);
1680 /* Wait for all the tasks to terminate */
1681 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
1682 IF_DEBUG(scheduler,fprintf(stderr,"scheduler: waiting for task %ld\n",
1684 pthread_join(task_ids[i].id, NULL);
1688 /* Send 'em all a SIGHUP. That should shut 'em up.
1690 await_death = RtsFlags.ParFlags.nNodes;
1691 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
1692 pthread_kill(task_ids[i].id,SIGTERM);
1694 while (await_death > 0) {
1700 /* -----------------------------------------------------------------------------
1701 Managing the per-task allocation areas.
1703 Each capability comes with an allocation area. These are
1704 fixed-length block lists into which allocation can be done.
1706 ToDo: no support for two-space collection at the moment???
1707 -------------------------------------------------------------------------- */
1709 /* -----------------------------------------------------------------------------
1710 * waitThread is the external interface for running a new computation
1711 * and waiting for the result.
1713 * In the non-SMP case, we create a new main thread, push it on the
1714 * main-thread stack, and invoke the scheduler to run it. The
1715 * scheduler will return when the top main thread on the stack has
1716 * completed or died, and fill in the necessary fields of the
1717 * main_thread structure.
1719 * In the SMP case, we create a main thread as before, but we then
1720 * create a new condition variable and sleep on it. When our new
1721 * main thread has completed, we'll be woken up and the status/result
1722 * will be in the main_thread struct.
1723 * -------------------------------------------------------------------------- */
1726 howManyThreadsAvail ( void )
1730 for (q = run_queue_hd; q != END_TSO_QUEUE; q = q->link)
1732 for (q = blocked_queue_hd; q != END_TSO_QUEUE; q = q->link)
1738 finishAllThreads ( void )
1741 while (run_queue_hd != END_TSO_QUEUE) {
1742 waitThread ( run_queue_hd, NULL );
1744 while (blocked_queue_hd != END_TSO_QUEUE) {
1745 waitThread ( blocked_queue_hd, NULL );
1748 (blocked_queue_hd != END_TSO_QUEUE ||
1749 run_queue_hd != END_TSO_QUEUE);
1753 waitThread(StgTSO *tso, /*out*/StgClosure **ret)
1756 SchedulerStatus stat;
1758 ACQUIRE_LOCK(&sched_mutex);
1760 m = stgMallocBytes(sizeof(StgMainThread), "waitThread");
1766 pthread_cond_init(&m->wakeup, NULL);
1769 m->link = main_threads;
1772 IF_DEBUG(scheduler, fprintf(stderr, "scheduler: new main thread (%d)\n",
1777 pthread_cond_wait(&m->wakeup, &sched_mutex);
1778 } while (m->stat == NoStatus);
1780 /* GranSim specific init */
1781 CurrentTSO = m->tso; // the TSO to run
1782 procStatus[MainProc] = Busy; // status of main PE
1783 CurrentProc = MainProc; // PE to run it on
1788 ASSERT(m->stat != NoStatus);
1794 pthread_cond_destroy(&m->wakeup);
1797 IF_DEBUG(scheduler, fprintf(stderr, "scheduler: main thread (%d) finished\n",
1801 RELEASE_LOCK(&sched_mutex);
1806 //@node Run queue code, Garbage Collextion Routines, Suspend and Resume, Main scheduling code
1807 //@subsection Run queue code
1811 NB: In GranSim we have many run queues; run_queue_hd is actually a macro
1812 unfolding to run_queue_hds[CurrentProc], thus CurrentProc is an
1813 implicit global variable that has to be correct when calling these
1817 /* Put the new thread on the head of the runnable queue.
1818 * The caller of createThread better push an appropriate closure
1819 * on this thread's stack before the scheduler is invoked.
1821 static /* inline */ void
1822 add_to_run_queue(tso)
1825 ASSERT(tso!=run_queue_hd && tso!=run_queue_tl);
1826 tso->link = run_queue_hd;
1828 if (run_queue_tl == END_TSO_QUEUE) {
1833 /* Put the new thread at the end of the runnable queue. */
1834 static /* inline */ void
1835 push_on_run_queue(tso)
1838 ASSERT(get_itbl((StgClosure *)tso)->type == TSO);
1839 ASSERT(run_queue_hd!=NULL && run_queue_tl!=NULL);
1840 ASSERT(tso!=run_queue_hd && tso!=run_queue_tl);
1841 if (run_queue_hd == END_TSO_QUEUE) {
1844 run_queue_tl->link = tso;
1850 Should be inlined because it's used very often in schedule. The tso
1851 argument is actually only needed in GranSim, where we want to have the
1852 possibility to schedule *any* TSO on the run queue, irrespective of the
1853 actual ordering. Therefore, if tso is not the nil TSO then we traverse
1854 the run queue and dequeue the tso, adjusting the links in the queue.
1856 //@cindex take_off_run_queue
1857 static /* inline */ StgTSO*
1858 take_off_run_queue(StgTSO *tso) {
1862 qetlaHbogh Qu' ngaSbogh ghomDaQ {tso} yIteq!
1864 if tso is specified, unlink that tso from the run_queue (doesn't have
1865 to be at the beginning of the queue); GranSim only
1867 if (tso!=END_TSO_QUEUE) {
1868 /* find tso in queue */
1869 for (t=run_queue_hd, prev=END_TSO_QUEUE;
1870 t!=END_TSO_QUEUE && t!=tso;
1874 /* now actually dequeue the tso */
1875 if (prev!=END_TSO_QUEUE) {
1876 ASSERT(run_queue_hd!=t);
1877 prev->link = t->link;
1879 /* t is at beginning of thread queue */
1880 ASSERT(run_queue_hd==t);
1881 run_queue_hd = t->link;
1883 /* t is at end of thread queue */
1884 if (t->link==END_TSO_QUEUE) {
1885 ASSERT(t==run_queue_tl);
1886 run_queue_tl = prev;
1888 ASSERT(run_queue_tl!=t);
1890 t->link = END_TSO_QUEUE;
1892 /* take tso from the beginning of the queue; std concurrent code */
1894 if (t != END_TSO_QUEUE) {
1895 run_queue_hd = t->link;
1896 t->link = END_TSO_QUEUE;
1897 if (run_queue_hd == END_TSO_QUEUE) {
1898 run_queue_tl = END_TSO_QUEUE;
1907 //@node Garbage Collextion Routines, Blocking Queue Routines, Run queue code, Main scheduling code
1908 //@subsection Garbage Collextion Routines
1910 /* ---------------------------------------------------------------------------
1911 Where are the roots that we know about?
1913 - all the threads on the runnable queue
1914 - all the threads on the blocked queue
1915 - all the thread currently executing a _ccall_GC
1916 - all the "main threads"
1918 ------------------------------------------------------------------------ */
1920 /* This has to be protected either by the scheduler monitor, or by the
1921 garbage collection monitor (probably the latter).
1925 static void GetRoots(void)
1932 for (i=0; i<=RtsFlags.GranFlags.proc; i++) {
1933 if ((run_queue_hds[i] != END_TSO_QUEUE) && ((run_queue_hds[i] != NULL)))
1934 run_queue_hds[i] = (StgTSO *)MarkRoot((StgClosure *)run_queue_hds[i]);
1935 if ((run_queue_tls[i] != END_TSO_QUEUE) && ((run_queue_tls[i] != NULL)))
1936 run_queue_tls[i] = (StgTSO *)MarkRoot((StgClosure *)run_queue_tls[i]);
1938 if ((blocked_queue_hds[i] != END_TSO_QUEUE) && ((blocked_queue_hds[i] != NULL)))
1939 blocked_queue_hds[i] = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_hds[i]);
1940 if ((blocked_queue_tls[i] != END_TSO_QUEUE) && ((blocked_queue_tls[i] != NULL)))
1941 blocked_queue_tls[i] = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_tls[i]);
1942 if ((ccalling_threadss[i] != END_TSO_QUEUE) && ((ccalling_threadss[i] != NULL)))
1943 ccalling_threadss[i] = (StgTSO *)MarkRoot((StgClosure *)ccalling_threadss[i]);
1950 if (run_queue_hd != END_TSO_QUEUE) {
1951 ASSERT(run_queue_tl != END_TSO_QUEUE);
1952 run_queue_hd = (StgTSO *)MarkRoot((StgClosure *)run_queue_hd);
1953 run_queue_tl = (StgTSO *)MarkRoot((StgClosure *)run_queue_tl);
1956 if (blocked_queue_hd != END_TSO_QUEUE) {
1957 ASSERT(blocked_queue_tl != END_TSO_QUEUE);
1958 blocked_queue_hd = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_hd);
1959 blocked_queue_tl = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_tl);
1963 for (m = main_threads; m != NULL; m = m->link) {
1964 m->tso = (StgTSO *)MarkRoot((StgClosure *)m->tso);
1966 if (suspended_ccalling_threads != END_TSO_QUEUE)
1967 suspended_ccalling_threads =
1968 (StgTSO *)MarkRoot((StgClosure *)suspended_ccalling_threads);
1970 #if defined(SMP) || defined(PAR) || defined(GRAN)
1975 /* -----------------------------------------------------------------------------
1978 This is the interface to the garbage collector from Haskell land.
1979 We provide this so that external C code can allocate and garbage
1980 collect when called from Haskell via _ccall_GC.
1982 It might be useful to provide an interface whereby the programmer
1983 can specify more roots (ToDo).
1985 This needs to be protected by the GC condition variable above. KH.
1986 -------------------------------------------------------------------------- */
1988 void (*extra_roots)(void);
1993 GarbageCollect(GetRoots,rtsFalse);
1997 performMajorGC(void)
1999 GarbageCollect(GetRoots,rtsTrue);
2005 GetRoots(); /* the scheduler's roots */
2006 extra_roots(); /* the user's roots */
2010 performGCWithRoots(void (*get_roots)(void))
2012 extra_roots = get_roots;
2014 GarbageCollect(AllRoots,rtsFalse);
2017 /* -----------------------------------------------------------------------------
2020 If the thread has reached its maximum stack size, then raise the
2021 StackOverflow exception in the offending thread. Otherwise
2022 relocate the TSO into a larger chunk of memory and adjust its stack
2024 -------------------------------------------------------------------------- */
2027 threadStackOverflow(StgTSO *tso)
2029 nat new_stack_size, new_tso_size, diff, stack_words;
2033 IF_DEBUG(sanity,checkTSO(tso));
2034 if (tso->stack_size >= tso->max_stack_size) {
2037 belch("@@ threadStackOverflow of TSO %d (%p): stack too large (now %ld; max is %ld",
2038 tso->id, tso, tso->stack_size, tso->max_stack_size);
2039 /* If we're debugging, just print out the top of the stack */
2040 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2044 fprintf(stderr, "fatal: stack overflow in Hugs; aborting\n" );
2047 /* Send this thread the StackOverflow exception */
2048 raiseAsync(tso, (StgClosure *)stackOverflow_closure);
2053 /* Try to double the current stack size. If that takes us over the
2054 * maximum stack size for this thread, then use the maximum instead.
2055 * Finally round up so the TSO ends up as a whole number of blocks.
2057 new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
2058 new_tso_size = (nat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) +
2059 TSO_STRUCT_SIZE)/sizeof(W_);
2060 new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */
2061 new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
2063 IF_DEBUG(scheduler, fprintf(stderr,"scheduler: increasing stack size from %d words to %d.\n", tso->stack_size, new_stack_size));
2065 dest = (StgTSO *)allocate(new_tso_size);
2066 TICK_ALLOC_TSO(new_tso_size-sizeofW(StgTSO),0);
2068 /* copy the TSO block and the old stack into the new area */
2069 memcpy(dest,tso,TSO_STRUCT_SIZE);
2070 stack_words = tso->stack + tso->stack_size - tso->sp;
2071 new_sp = (P_)dest + new_tso_size - stack_words;
2072 memcpy(new_sp, tso->sp, stack_words * sizeof(W_));
2074 /* relocate the stack pointers... */
2075 diff = (P_)new_sp - (P_)tso->sp; /* In *words* */
2076 dest->su = (StgUpdateFrame *) ((P_)dest->su + diff);
2078 dest->splim = (P_)dest->splim + (nat)((P_)dest - (P_)tso);
2079 dest->stack_size = new_stack_size;
2081 /* and relocate the update frame list */
2082 relocate_TSO(tso, dest);
2084 /* Mark the old TSO as relocated. We have to check for relocated
2085 * TSOs in the garbage collector and any primops that deal with TSOs.
2087 * It's important to set the sp and su values to just beyond the end
2088 * of the stack, so we don't attempt to scavenge any part of the
2091 tso->what_next = ThreadRelocated;
2093 tso->sp = (P_)&(tso->stack[tso->stack_size]);
2094 tso->su = (StgUpdateFrame *)tso->sp;
2095 tso->why_blocked = NotBlocked;
2096 dest->mut_link = NULL;
2098 IF_PAR_DEBUG(verbose,
2099 belch("@@ threadStackOverflow of TSO %d (now at %p): stack size increased to %ld",
2100 tso->id, tso, tso->stack_size);
2101 /* If we're debugging, just print out the top of the stack */
2102 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2105 IF_DEBUG(sanity,checkTSO(tso));
2107 IF_DEBUG(scheduler,printTSO(dest));
2113 //@node Blocking Queue Routines, Exception Handling Routines, Garbage Collextion Routines, Main scheduling code
2114 //@subsection Blocking Queue Routines
2116 /* ---------------------------------------------------------------------------
2117 Wake up a queue that was blocked on some resource.
2118 ------------------------------------------------------------------------ */
2120 /* ToDo: check push_on_run_queue vs. PUSH_ON_RUN_QUEUE */
2124 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2129 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2131 /* write RESUME events to log file and
2132 update blocked and fetch time (depending on type of the orig closure) */
2133 if (RtsFlags.ParFlags.ParStats.Full) {
2134 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
2135 GR_RESUME, ((StgTSO *)bqe), ((StgTSO *)bqe)->block_info.closure,
2136 0, 0 /* spark_queue_len(ADVISORY_POOL) */);
2138 switch (get_itbl(node)->type) {
2140 ((StgTSO *)bqe)->par.fetchtime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2145 ((StgTSO *)bqe)->par.blocktime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2148 barf("{unblockOneLocked}Daq Qagh: unexpected closure in blocking queue");
2155 static StgBlockingQueueElement *
2156 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2159 PEs node_loc, tso_loc;
2161 node_loc = where_is(node); // should be lifted out of loop
2162 tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2163 tso_loc = where_is((StgClosure *)tso);
2164 if (IS_LOCAL_TO(PROCS(node),tso_loc)) { // TSO is local
2165 /* !fake_fetch => TSO is on CurrentProc is same as IS_LOCAL_TO */
2166 ASSERT(CurrentProc!=node_loc || tso_loc==CurrentProc);
2167 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.lunblocktime;
2168 // insertThread(tso, node_loc);
2169 new_event(tso_loc, tso_loc, CurrentTime[CurrentProc],
2171 tso, node, (rtsSpark*)NULL);
2172 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2175 } else { // TSO is remote (actually should be FMBQ)
2176 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.mpacktime +
2177 RtsFlags.GranFlags.Costs.gunblocktime +
2178 RtsFlags.GranFlags.Costs.latency;
2179 new_event(tso_loc, CurrentProc, CurrentTime[CurrentProc],
2181 tso, node, (rtsSpark*)NULL);
2182 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2185 /* the thread-queue-overhead is accounted for in either Resume or UnblockThread */
2187 fprintf(stderr," %s TSO %d (%p) [PE %d] (block_info.closure=%p) (next=%p) ,",
2188 (node_loc==tso_loc ? "Local" : "Global"),
2189 tso->id, tso, CurrentProc, tso->block_info.closure, tso->link));
2190 tso->block_info.closure = NULL;
2191 IF_DEBUG(scheduler,belch("-- Waking up thread %ld (%p)",
2195 static StgBlockingQueueElement *
2196 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2198 StgBlockingQueueElement *next;
2200 switch (get_itbl(bqe)->type) {
2202 ASSERT(((StgTSO *)bqe)->why_blocked != NotBlocked);
2203 /* if it's a TSO just push it onto the run_queue */
2205 // ((StgTSO *)bqe)->link = END_TSO_QUEUE; // debugging?
2206 PUSH_ON_RUN_QUEUE((StgTSO *)bqe);
2208 unblockCount(bqe, node);
2209 /* reset blocking status after dumping event */
2210 ((StgTSO *)bqe)->why_blocked = NotBlocked;
2214 /* if it's a BLOCKED_FETCH put it on the PendingFetches list */
2216 bqe->link = PendingFetches;
2217 PendingFetches = bqe;
2221 /* can ignore this case in a non-debugging setup;
2222 see comments on RBHSave closures above */
2224 /* check that the closure is an RBHSave closure */
2225 ASSERT(get_itbl((StgClosure *)bqe) == &RBH_Save_0_info ||
2226 get_itbl((StgClosure *)bqe) == &RBH_Save_1_info ||
2227 get_itbl((StgClosure *)bqe) == &RBH_Save_2_info);
2231 barf("{unblockOneLocked}Daq Qagh: Unexpected IP (%#lx; %s) in blocking queue at %#lx\n",
2232 get_itbl((StgClosure *)bqe), info_type((StgClosure *)bqe),
2236 // IF_DEBUG(scheduler,sched_belch("waking up thread %ld", tso->id));
2240 #else /* !GRAN && !PAR */
2242 unblockOneLocked(StgTSO *tso)
2246 ASSERT(get_itbl(tso)->type == TSO);
2247 ASSERT(tso->why_blocked != NotBlocked);
2248 tso->why_blocked = NotBlocked;
2250 PUSH_ON_RUN_QUEUE(tso);
2252 IF_DEBUG(scheduler,sched_belch("waking up thread %ld", tso->id));
2257 #if defined(GRAN) || defined(PAR)
2258 inline StgBlockingQueueElement *
2259 unblockOne(StgBlockingQueueElement *bqe, StgClosure *node)
2261 ACQUIRE_LOCK(&sched_mutex);
2262 bqe = unblockOneLocked(bqe, node);
2263 RELEASE_LOCK(&sched_mutex);
2268 unblockOne(StgTSO *tso)
2270 ACQUIRE_LOCK(&sched_mutex);
2271 tso = unblockOneLocked(tso);
2272 RELEASE_LOCK(&sched_mutex);
2279 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2281 StgBlockingQueueElement *bqe;
2286 belch("## AwBQ for node %p on PE %d @ %ld by TSO %d (%p): ", \
2287 node, CurrentProc, CurrentTime[CurrentProc],
2288 CurrentTSO->id, CurrentTSO));
2290 node_loc = where_is(node);
2292 ASSERT(get_itbl(q)->type == TSO || // q is either a TSO or an RBHSave
2293 get_itbl(q)->type == CONSTR); // closure (type constructor)
2294 ASSERT(is_unique(node));
2296 /* FAKE FETCH: magically copy the node to the tso's proc;
2297 no Fetch necessary because in reality the node should not have been
2298 moved to the other PE in the first place
2300 if (CurrentProc!=node_loc) {
2302 belch("## node %p is on PE %d but CurrentProc is %d (TSO %d); assuming fake fetch and adjusting bitmask (old: %#x)",
2303 node, node_loc, CurrentProc, CurrentTSO->id,
2304 // CurrentTSO, where_is(CurrentTSO),
2305 node->header.gran.procs));
2306 node->header.gran.procs = (node->header.gran.procs) | PE_NUMBER(CurrentProc);
2308 belch("## new bitmask of node %p is %#x",
2309 node, node->header.gran.procs));
2310 if (RtsFlags.GranFlags.GranSimStats.Global) {
2311 globalGranStats.tot_fake_fetches++;
2316 // ToDo: check: ASSERT(CurrentProc==node_loc);
2317 while (get_itbl(bqe)->type==TSO) { // q != END_TSO_QUEUE) {
2320 bqe points to the current element in the queue
2321 next points to the next element in the queue
2323 //tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2324 //tso_loc = where_is(tso);
2326 bqe = unblockOneLocked(bqe, node);
2329 /* if this is the BQ of an RBH, we have to put back the info ripped out of
2330 the closure to make room for the anchor of the BQ */
2331 if (bqe!=END_BQ_QUEUE) {
2332 ASSERT(get_itbl(node)->type == RBH && get_itbl(bqe)->type == CONSTR);
2334 ASSERT((info_ptr==&RBH_Save_0_info) ||
2335 (info_ptr==&RBH_Save_1_info) ||
2336 (info_ptr==&RBH_Save_2_info));
2338 /* cf. convertToRBH in RBH.c for writing the RBHSave closure */
2339 ((StgRBH *)node)->blocking_queue = (StgBlockingQueueElement *)((StgRBHSave *)bqe)->payload[0];
2340 ((StgRBH *)node)->mut_link = (StgMutClosure *)((StgRBHSave *)bqe)->payload[1];
2343 belch("## Filled in RBH_Save for %p (%s) at end of AwBQ",
2344 node, info_type(node)));
2347 /* statistics gathering */
2348 if (RtsFlags.GranFlags.GranSimStats.Global) {
2349 // globalGranStats.tot_bq_processing_time += bq_processing_time;
2350 globalGranStats.tot_bq_len += len; // total length of all bqs awakened
2351 // globalGranStats.tot_bq_len_local += len_local; // same for local TSOs only
2352 globalGranStats.tot_awbq++; // total no. of bqs awakened
2355 fprintf(stderr,"## BQ Stats of %p: [%d entries] %s\n",
2356 node, len, (bqe!=END_BQ_QUEUE) ? "RBH" : ""));
2360 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2362 StgBlockingQueueElement *bqe, *next;
2364 ACQUIRE_LOCK(&sched_mutex);
2366 IF_PAR_DEBUG(verbose,
2367 belch("## AwBQ for node %p on [%x]: ",
2370 ASSERT(get_itbl(q)->type == TSO ||
2371 get_itbl(q)->type == BLOCKED_FETCH ||
2372 get_itbl(q)->type == CONSTR);
2375 while (get_itbl(bqe)->type==TSO ||
2376 get_itbl(bqe)->type==BLOCKED_FETCH) {
2377 bqe = unblockOneLocked(bqe, node);
2379 RELEASE_LOCK(&sched_mutex);
2382 #else /* !GRAN && !PAR */
2384 awakenBlockedQueue(StgTSO *tso)
2386 ACQUIRE_LOCK(&sched_mutex);
2387 while (tso != END_TSO_QUEUE) {
2388 tso = unblockOneLocked(tso);
2390 RELEASE_LOCK(&sched_mutex);
2394 //@node Exception Handling Routines, Debugging Routines, Blocking Queue Routines, Main scheduling code
2395 //@subsection Exception Handling Routines
2397 /* ---------------------------------------------------------------------------
2399 - usually called inside a signal handler so it mustn't do anything fancy.
2400 ------------------------------------------------------------------------ */
2403 interruptStgRts(void)
2409 /* -----------------------------------------------------------------------------
2412 This is for use when we raise an exception in another thread, which
2414 This has nothing to do with the UnblockThread event in GranSim. -- HWL
2415 -------------------------------------------------------------------------- */
2417 #if defined(GRAN) || defined(PAR)
2419 NB: only the type of the blocking queue is different in GranSim and GUM
2420 the operations on the queue-elements are the same
2421 long live polymorphism!
2424 unblockThread(StgTSO *tso)
2426 StgBlockingQueueElement *t, **last;
2428 ACQUIRE_LOCK(&sched_mutex);
2429 switch (tso->why_blocked) {
2432 return; /* not blocked */
2435 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
2437 StgBlockingQueueElement *last_tso = END_BQ_QUEUE;
2438 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
2440 last = (StgBlockingQueueElement **)&mvar->head;
2441 for (t = (StgBlockingQueueElement *)mvar->head;
2443 last = &t->link, last_tso = t, t = t->link) {
2444 if (t == (StgBlockingQueueElement *)tso) {
2445 *last = (StgBlockingQueueElement *)tso->link;
2446 if (mvar->tail == tso) {
2447 mvar->tail = (StgTSO *)last_tso;
2452 barf("unblockThread (MVAR): TSO not found");
2455 case BlockedOnBlackHole:
2456 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
2458 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
2460 last = &bq->blocking_queue;
2461 for (t = bq->blocking_queue;
2463 last = &t->link, t = t->link) {
2464 if (t == (StgBlockingQueueElement *)tso) {
2465 *last = (StgBlockingQueueElement *)tso->link;
2469 barf("unblockThread (BLACKHOLE): TSO not found");
2472 case BlockedOnException:
2474 StgTSO *target = tso->block_info.tso;
2476 ASSERT(get_itbl(target)->type == TSO);
2477 ASSERT(target->blocked_exceptions != NULL);
2479 last = (StgBlockingQueueElement **)&target->blocked_exceptions;
2480 for (t = (StgBlockingQueueElement *)target->blocked_exceptions;
2482 last = &t->link, t = t->link) {
2483 ASSERT(get_itbl(t)->type == TSO);
2484 if (t == (StgBlockingQueueElement *)tso) {
2485 *last = (StgBlockingQueueElement *)tso->link;
2489 barf("unblockThread (Exception): TSO not found");
2492 case BlockedOnDelay:
2494 case BlockedOnWrite:
2496 StgBlockingQueueElement *prev = NULL;
2497 for (t = (StgBlockingQueueElement *)blocked_queue_hd; t != END_BQ_QUEUE;
2498 prev = t, t = t->link) {
2499 if (t == (StgBlockingQueueElement *)tso) {
2501 blocked_queue_hd = (StgTSO *)t->link;
2502 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
2503 blocked_queue_tl = END_TSO_QUEUE;
2506 prev->link = t->link;
2507 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
2508 blocked_queue_tl = (StgTSO *)prev;
2514 barf("unblockThread (I/O): TSO not found");
2518 barf("unblockThread");
2522 tso->link = END_TSO_QUEUE;
2523 tso->why_blocked = NotBlocked;
2524 tso->block_info.closure = NULL;
2525 PUSH_ON_RUN_QUEUE(tso);
2526 RELEASE_LOCK(&sched_mutex);
2530 unblockThread(StgTSO *tso)
2534 ACQUIRE_LOCK(&sched_mutex);
2535 switch (tso->why_blocked) {
2538 return; /* not blocked */
2541 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
2543 StgTSO *last_tso = END_TSO_QUEUE;
2544 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
2547 for (t = mvar->head; t != END_TSO_QUEUE;
2548 last = &t->link, last_tso = t, t = t->link) {
2551 if (mvar->tail == tso) {
2552 mvar->tail = last_tso;
2557 barf("unblockThread (MVAR): TSO not found");
2560 case BlockedOnBlackHole:
2561 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
2563 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
2565 last = &bq->blocking_queue;
2566 for (t = bq->blocking_queue; t != END_TSO_QUEUE;
2567 last = &t->link, t = t->link) {
2573 barf("unblockThread (BLACKHOLE): TSO not found");
2576 case BlockedOnException:
2578 StgTSO *target = tso->block_info.tso;
2580 ASSERT(get_itbl(target)->type == TSO);
2581 ASSERT(target->blocked_exceptions != NULL);
2583 last = &target->blocked_exceptions;
2584 for (t = target->blocked_exceptions; t != END_TSO_QUEUE;
2585 last = &t->link, t = t->link) {
2586 ASSERT(get_itbl(t)->type == TSO);
2592 barf("unblockThread (Exception): TSO not found");
2595 case BlockedOnDelay:
2597 case BlockedOnWrite:
2599 StgTSO *prev = NULL;
2600 for (t = blocked_queue_hd; t != END_TSO_QUEUE;
2601 prev = t, t = t->link) {
2604 blocked_queue_hd = t->link;
2605 if (blocked_queue_tl == t) {
2606 blocked_queue_tl = END_TSO_QUEUE;
2609 prev->link = t->link;
2610 if (blocked_queue_tl == t) {
2611 blocked_queue_tl = prev;
2617 barf("unblockThread (I/O): TSO not found");
2621 barf("unblockThread");
2625 tso->link = END_TSO_QUEUE;
2626 tso->why_blocked = NotBlocked;
2627 tso->block_info.closure = NULL;
2628 PUSH_ON_RUN_QUEUE(tso);
2629 RELEASE_LOCK(&sched_mutex);
2633 /* -----------------------------------------------------------------------------
2636 * The following function implements the magic for raising an
2637 * asynchronous exception in an existing thread.
2639 * We first remove the thread from any queue on which it might be
2640 * blocked. The possible blockages are MVARs and BLACKHOLE_BQs.
2642 * We strip the stack down to the innermost CATCH_FRAME, building
2643 * thunks in the heap for all the active computations, so they can
2644 * be restarted if necessary. When we reach a CATCH_FRAME, we build
2645 * an application of the handler to the exception, and push it on
2646 * the top of the stack.
2648 * How exactly do we save all the active computations? We create an
2649 * AP_UPD for every UpdateFrame on the stack. Entering one of these
2650 * AP_UPDs pushes everything from the corresponding update frame
2651 * upwards onto the stack. (Actually, it pushes everything up to the
2652 * next update frame plus a pointer to the next AP_UPD object.
2653 * Entering the next AP_UPD object pushes more onto the stack until we
2654 * reach the last AP_UPD object - at which point the stack should look
2655 * exactly as it did when we killed the TSO and we can continue
2656 * execution by entering the closure on top of the stack.
2658 * We can also kill a thread entirely - this happens if either (a) the
2659 * exception passed to raiseAsync is NULL, or (b) there's no
2660 * CATCH_FRAME on the stack. In either case, we strip the entire
2661 * stack and replace the thread with a zombie.
2663 * -------------------------------------------------------------------------- */
2666 deleteThread(StgTSO *tso)
2668 raiseAsync(tso,NULL);
2672 raiseAsync(StgTSO *tso, StgClosure *exception)
2674 StgUpdateFrame* su = tso->su;
2675 StgPtr sp = tso->sp;
2677 /* Thread already dead? */
2678 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
2682 IF_DEBUG(scheduler, sched_belch("raising exception in thread %ld.", tso->id));
2684 /* Remove it from any blocking queues */
2687 /* The stack freezing code assumes there's a closure pointer on
2688 * the top of the stack. This isn't always the case with compiled
2689 * code, so we have to push a dummy closure on the top which just
2690 * returns to the next return address on the stack.
2692 if ( LOOKS_LIKE_GHC_INFO((void*)*sp) ) {
2693 *(--sp) = (W_)&dummy_ret_closure;
2697 int words = ((P_)su - (P_)sp) - 1;
2701 /* If we find a CATCH_FRAME, and we've got an exception to raise,
2702 * then build PAP(handler,exception,realworld#), and leave it on
2703 * top of the stack ready to enter.
2705 if (get_itbl(su)->type == CATCH_FRAME && exception != NULL) {
2706 StgCatchFrame *cf = (StgCatchFrame *)su;
2707 /* we've got an exception to raise, so let's pass it to the
2708 * handler in this frame.
2710 ap = (StgAP_UPD *)allocate(sizeofW(StgPAP) + 2);
2711 TICK_ALLOC_UPD_PAP(3,0);
2712 SET_HDR(ap,&PAP_info,cf->header.prof.ccs);
2715 ap->fun = cf->handler; /* :: Exception -> IO a */
2716 ap->payload[0] = (P_)exception;
2717 ap->payload[1] = ARG_TAG(0); /* realworld token */
2719 /* throw away the stack from Sp up to and including the
2722 sp = (P_)su + sizeofW(StgCatchFrame) - 1;
2725 /* Restore the blocked/unblocked state for asynchronous exceptions
2726 * at the CATCH_FRAME.
2728 * If exceptions were unblocked at the catch, arrange that they
2729 * are unblocked again after executing the handler by pushing an
2730 * unblockAsyncExceptions_ret stack frame.
2732 if (!cf->exceptions_blocked) {
2733 *(sp--) = (W_)&unblockAsyncExceptionszh_ret_info;
2736 /* Ensure that async exceptions are blocked when running the handler.
2738 if (tso->blocked_exceptions == NULL) {
2739 tso->blocked_exceptions = END_TSO_QUEUE;
2742 /* Put the newly-built PAP on top of the stack, ready to execute
2743 * when the thread restarts.
2747 tso->what_next = ThreadEnterGHC;
2748 IF_DEBUG(sanity, checkTSO(tso));
2752 /* First build an AP_UPD consisting of the stack chunk above the
2753 * current update frame, with the top word on the stack as the
2756 ap = (StgAP_UPD *)allocate(AP_sizeW(words));
2761 ap->fun = (StgClosure *)sp[0];
2763 for(i=0; i < (nat)words; ++i) {
2764 ap->payload[i] = (P_)*sp++;
2767 switch (get_itbl(su)->type) {
2771 SET_HDR(ap,&AP_UPD_info,su->header.prof.ccs /* ToDo */);
2772 TICK_ALLOC_UP_THK(words+1,0);
2775 fprintf(stderr, "scheduler: Updating ");
2776 printPtr((P_)su->updatee);
2777 fprintf(stderr, " with ");
2778 printObj((StgClosure *)ap);
2781 /* Replace the updatee with an indirection - happily
2782 * this will also wake up any threads currently
2783 * waiting on the result.
2785 UPD_IND_NOLOCK(su->updatee,ap); /* revert the black hole */
2787 sp += sizeofW(StgUpdateFrame) -1;
2788 sp[0] = (W_)ap; /* push onto stack */
2794 StgCatchFrame *cf = (StgCatchFrame *)su;
2797 /* We want a PAP, not an AP_UPD. Fortunately, the
2798 * layout's the same.
2800 SET_HDR(ap,&PAP_info,su->header.prof.ccs /* ToDo */);
2801 TICK_ALLOC_UPD_PAP(words+1,0);
2803 /* now build o = FUN(catch,ap,handler) */
2804 o = (StgClosure *)allocate(sizeofW(StgClosure)+2);
2805 TICK_ALLOC_FUN(2,0);
2806 SET_HDR(o,&catch_info,su->header.prof.ccs /* ToDo */);
2807 o->payload[0] = (StgClosure *)ap;
2808 o->payload[1] = cf->handler;
2811 fprintf(stderr, "scheduler: Built ");
2812 printObj((StgClosure *)o);
2815 /* pop the old handler and put o on the stack */
2817 sp += sizeofW(StgCatchFrame) - 1;
2824 StgSeqFrame *sf = (StgSeqFrame *)su;
2827 SET_HDR(ap,&PAP_info,su->header.prof.ccs /* ToDo */);
2828 TICK_ALLOC_UPD_PAP(words+1,0);
2830 /* now build o = FUN(seq,ap) */
2831 o = (StgClosure *)allocate(sizeofW(StgClosure)+1);
2832 TICK_ALLOC_SE_THK(1,0);
2833 SET_HDR(o,&seq_info,su->header.prof.ccs /* ToDo */);
2834 o->payload[0] = (StgClosure *)ap;
2837 fprintf(stderr, "scheduler: Built ");
2838 printObj((StgClosure *)o);
2841 /* pop the old handler and put o on the stack */
2843 sp += sizeofW(StgSeqFrame) - 1;
2849 /* We've stripped the entire stack, the thread is now dead. */
2850 sp += sizeofW(StgStopFrame) - 1;
2851 sp[0] = (W_)exception; /* save the exception */
2852 tso->what_next = ThreadKilled;
2853 tso->su = (StgUpdateFrame *)(sp+1);
2864 /* -----------------------------------------------------------------------------
2865 resurrectThreads is called after garbage collection on the list of
2866 threads found to be garbage. Each of these threads will be woken
2867 up and sent a signal: BlockedOnDeadMVar if the thread was blocked
2868 on an MVar, or NonTermination if the thread was blocked on a Black
2870 -------------------------------------------------------------------------- */
2873 resurrectThreads( StgTSO *threads )
2877 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
2878 next = tso->global_link;
2879 tso->global_link = all_threads;
2881 IF_DEBUG(scheduler, sched_belch("resurrecting thread %d", tso->id));
2883 switch (tso->why_blocked) {
2885 case BlockedOnException:
2886 raiseAsync(tso,(StgClosure *)BlockedOnDeadMVar_closure);
2888 case BlockedOnBlackHole:
2889 raiseAsync(tso,(StgClosure *)NonTermination_closure);
2892 /* This might happen if the thread was blocked on a black hole
2893 * belonging to a thread that we've just woken up (raiseAsync
2894 * can wake up threads, remember...).
2898 barf("resurrectThreads: thread blocked in a strange way");
2903 /* -----------------------------------------------------------------------------
2904 * Blackhole detection: if we reach a deadlock, test whether any
2905 * threads are blocked on themselves. Any threads which are found to
2906 * be self-blocked get sent a NonTermination exception.
2908 * This is only done in a deadlock situation in order to avoid
2909 * performance overhead in the normal case.
2910 * -------------------------------------------------------------------------- */
2913 detectBlackHoles( void )
2915 StgTSO *t = all_threads;
2916 StgUpdateFrame *frame;
2917 StgClosure *blocked_on;
2919 for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
2921 if (t->why_blocked != BlockedOnBlackHole) {
2925 blocked_on = t->block_info.closure;
2927 for (frame = t->su; ; frame = frame->link) {
2928 switch (get_itbl(frame)->type) {
2931 if (frame->updatee == blocked_on) {
2932 /* We are blocking on one of our own computations, so
2933 * send this thread the NonTermination exception.
2936 sched_belch("thread %d is blocked on itself", t->id));
2937 raiseAsync(t, (StgClosure *)NonTermination_closure);
2958 //@node Debugging Routines, Index, Exception Handling Routines, Main scheduling code
2959 //@subsection Debugging Routines
2961 /* -----------------------------------------------------------------------------
2962 Debugging: why is a thread blocked
2963 -------------------------------------------------------------------------- */
2968 printThreadBlockage(StgTSO *tso)
2970 switch (tso->why_blocked) {
2972 fprintf(stderr,"blocked on read from fd %d", tso->block_info.fd);
2974 case BlockedOnWrite:
2975 fprintf(stderr,"blocked on write to fd %d", tso->block_info.fd);
2977 case BlockedOnDelay:
2978 #if defined(HAVE_SETITIMER) || defined(mingw32_TARGET_OS)
2979 fprintf(stderr,"blocked on delay of %d ms", tso->block_info.delay);
2981 fprintf(stderr,"blocked on delay of %d ms",
2982 tso->block_info.target - getourtimeofday());
2986 fprintf(stderr,"blocked on an MVar");
2988 case BlockedOnException:
2989 fprintf(stderr,"blocked on delivering an exception to thread %d",
2990 tso->block_info.tso->id);
2992 case BlockedOnBlackHole:
2993 fprintf(stderr,"blocked on a black hole");
2996 fprintf(stderr,"not blocked");
3000 fprintf(stderr,"blocked on global address; local FM_BQ is %p (%s)",
3001 tso->block_info.closure, info_type(tso->block_info.closure));
3003 case BlockedOnGA_NoSend:
3004 fprintf(stderr,"blocked on global address (no send); local FM_BQ is %p (%s)",
3005 tso->block_info.closure, info_type(tso->block_info.closure));
3009 barf("printThreadBlockage: strange tso->why_blocked: %d for TSO %d (%d)",
3010 tso->why_blocked, tso->id, tso);
3015 printThreadStatus(StgTSO *tso)
3017 switch (tso->what_next) {
3019 fprintf(stderr,"has been killed");
3021 case ThreadComplete:
3022 fprintf(stderr,"has completed");
3025 printThreadBlockage(tso);
3030 printAllThreads(void)
3034 sched_belch("all threads:");
3035 for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
3036 fprintf(stderr, "\tthread %d is ", t->id);
3037 printThreadStatus(t);
3038 fprintf(stderr,"\n");
3043 Print a whole blocking queue attached to node (debugging only).
3048 print_bq (StgClosure *node)
3050 StgBlockingQueueElement *bqe;
3054 fprintf(stderr,"## BQ of closure %p (%s): ",
3055 node, info_type(node));
3057 /* should cover all closures that may have a blocking queue */
3058 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
3059 get_itbl(node)->type == FETCH_ME_BQ ||
3060 get_itbl(node)->type == RBH);
3062 ASSERT(node!=(StgClosure*)NULL); // sanity check
3064 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
3066 for (bqe = ((StgBlockingQueue*)node)->blocking_queue, end = (bqe==END_BQ_QUEUE);
3067 !end; // iterate until bqe points to a CONSTR
3068 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) {
3069 ASSERT(bqe != END_BQ_QUEUE); // sanity check
3070 ASSERT(bqe != (StgTSO*)NULL); // sanity check
3071 /* types of closures that may appear in a blocking queue */
3072 ASSERT(get_itbl(bqe)->type == TSO ||
3073 get_itbl(bqe)->type == BLOCKED_FETCH ||
3074 get_itbl(bqe)->type == CONSTR);
3075 /* only BQs of an RBH end with an RBH_Save closure */
3076 ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
3078 switch (get_itbl(bqe)->type) {
3080 fprintf(stderr," TSO %d (%x),",
3081 ((StgTSO *)bqe)->id, ((StgTSO *)bqe));
3084 fprintf(stderr," BF (node=%p, ga=((%x, %d, %x)),",
3085 ((StgBlockedFetch *)bqe)->node,
3086 ((StgBlockedFetch *)bqe)->ga.payload.gc.gtid,
3087 ((StgBlockedFetch *)bqe)->ga.payload.gc.slot,
3088 ((StgBlockedFetch *)bqe)->ga.weight);
3091 fprintf(stderr," %s (IP %p),",
3092 (get_itbl(bqe) == &RBH_Save_0_info ? "RBH_Save_0" :
3093 get_itbl(bqe) == &RBH_Save_1_info ? "RBH_Save_1" :
3094 get_itbl(bqe) == &RBH_Save_2_info ? "RBH_Save_2" :
3095 "RBH_Save_?"), get_itbl(bqe));
3098 barf("Unexpected closure type %s in blocking queue of %p (%s)",
3099 info_type(bqe), node, info_type(node));
3103 fputc('\n', stderr);
3105 # elif defined(GRAN)
3107 print_bq (StgClosure *node)
3109 StgBlockingQueueElement *bqe;
3110 PEs node_loc, tso_loc;
3113 /* should cover all closures that may have a blocking queue */
3114 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
3115 get_itbl(node)->type == FETCH_ME_BQ ||
3116 get_itbl(node)->type == RBH);
3118 ASSERT(node!=(StgClosure*)NULL); // sanity check
3119 node_loc = where_is(node);
3121 fprintf(stderr,"## BQ of closure %p (%s) on [PE %d]: ",
3122 node, info_type(node), node_loc);
3125 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
3127 for (bqe = ((StgBlockingQueue*)node)->blocking_queue, end = (bqe==END_BQ_QUEUE);
3128 !end; // iterate until bqe points to a CONSTR
3129 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) {
3130 ASSERT(bqe != END_BQ_QUEUE); // sanity check
3131 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
3132 /* types of closures that may appear in a blocking queue */
3133 ASSERT(get_itbl(bqe)->type == TSO ||
3134 get_itbl(bqe)->type == CONSTR);
3135 /* only BQs of an RBH end with an RBH_Save closure */
3136 ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
3138 tso_loc = where_is((StgClosure *)bqe);
3139 switch (get_itbl(bqe)->type) {
3141 fprintf(stderr," TSO %d (%p) on [PE %d],",
3142 ((StgTSO *)bqe)->id, (StgTSO *)bqe, tso_loc);
3145 fprintf(stderr," %s (IP %p),",
3146 (get_itbl(bqe) == &RBH_Save_0_info ? "RBH_Save_0" :
3147 get_itbl(bqe) == &RBH_Save_1_info ? "RBH_Save_1" :
3148 get_itbl(bqe) == &RBH_Save_2_info ? "RBH_Save_2" :
3149 "RBH_Save_?"), get_itbl(bqe));
3152 barf("Unexpected closure type %s in blocking queue of %p (%s)",
3153 info_type((StgClosure *)bqe), node, info_type(node));
3157 fputc('\n', stderr);
3161 Nice and easy: only TSOs on the blocking queue
3164 print_bq (StgClosure *node)
3168 ASSERT(node!=(StgClosure*)NULL); // sanity check
3169 for (tso = ((StgBlockingQueue*)node)->blocking_queue;
3170 tso != END_TSO_QUEUE;
3172 ASSERT(tso!=NULL && tso!=END_TSO_QUEUE); // sanity check
3173 ASSERT(get_itbl(tso)->type == TSO); // guess what, sanity check
3174 fprintf(stderr," TSO %d (%p),", tso->id, tso);
3176 fputc('\n', stderr);
3187 for (i=0, tso=run_queue_hd;
3188 tso != END_TSO_QUEUE;
3197 sched_belch(char *s, ...)
3202 fprintf(stderr, "scheduler (task %ld): ", pthread_self());
3204 fprintf(stderr, "scheduler: ");
3206 vfprintf(stderr, s, ap);
3207 fprintf(stderr, "\n");
3213 //@node Index, , Debugging Routines, Main scheduling code
3217 //* MainRegTable:: @cindex\s-+MainRegTable
3218 //* StgMainThread:: @cindex\s-+StgMainThread
3219 //* awaken_blocked_queue:: @cindex\s-+awaken_blocked_queue
3220 //* blocked_queue_hd:: @cindex\s-+blocked_queue_hd
3221 //* blocked_queue_tl:: @cindex\s-+blocked_queue_tl
3222 //* context_switch:: @cindex\s-+context_switch
3223 //* createThread:: @cindex\s-+createThread
3224 //* free_capabilities:: @cindex\s-+free_capabilities
3225 //* gc_pending_cond:: @cindex\s-+gc_pending_cond
3226 //* initScheduler:: @cindex\s-+initScheduler
3227 //* interrupted:: @cindex\s-+interrupted
3228 //* n_free_capabilities:: @cindex\s-+n_free_capabilities
3229 //* next_thread_id:: @cindex\s-+next_thread_id
3230 //* print_bq:: @cindex\s-+print_bq
3231 //* run_queue_hd:: @cindex\s-+run_queue_hd
3232 //* run_queue_tl:: @cindex\s-+run_queue_tl
3233 //* sched_mutex:: @cindex\s-+sched_mutex
3234 //* schedule:: @cindex\s-+schedule
3235 //* take_off_run_queue:: @cindex\s-+take_off_run_queue
3236 //* task_ids:: @cindex\s-+task_ids
3237 //* term_mutex:: @cindex\s-+term_mutex
3238 //* thread_ready_cond:: @cindex\s-+thread_ready_cond