1 /* ---------------------------------------------------------------------------
2 * $Id: Schedule.c,v 1.94 2001/03/22 03:51:10 hwloidl Exp $
4 * (c) The GHC Team, 1998-2000
8 * Different GHC ways use this scheduler quite differently (see comments below)
9 * Here is the global picture:
11 * WAY Name CPP flag What's it for
12 * --------------------------------------
13 * mp GUM PAR Parallel execution on a distributed memory machine
14 * s SMP SMP Parallel execution on a shared memory machine
15 * mg GranSim GRAN Simulation of parallel execution
16 * md GUM/GdH DIST Distributed execution (based on GUM)
17 * --------------------------------------------------------------------------*/
19 //@node Main scheduling code, , ,
20 //@section Main scheduling code
23 * Version with scheduler monitor support for SMPs (WAY=s):
25 This design provides a high-level API to create and schedule threads etc.
26 as documented in the SMP design document.
28 It uses a monitor design controlled by a single mutex to exercise control
29 over accesses to shared data structures, and builds on the Posix threads
32 The majority of state is shared. In order to keep essential per-task state,
33 there is a Capability structure, which contains all the information
34 needed to run a thread: its STG registers, a pointer to its TSO, a
35 nursery etc. During STG execution, a pointer to the capability is
36 kept in a register (BaseReg).
38 In a non-SMP build, there is one global capability, namely MainRegTable.
42 * Version with support for distributed memory parallelism aka GUM (WAY=mp):
44 The main scheduling loop in GUM iterates until a finish message is received.
45 In that case a global flag @receivedFinish@ is set and this instance of
46 the RTS shuts down. See ghc/rts/parallel/HLComms.c:processMessages()
47 for the handling of incoming messages, such as PP_FINISH.
48 Note that in the parallel case we have a system manager that coordinates
49 different PEs, each of which are running one instance of the RTS.
50 See ghc/rts/parallel/SysMan.c for the main routine of the parallel program.
51 From this routine processes executing ghc/rts/Main.c are spawned. -- HWL
53 * Version with support for simulating parallel execution aka GranSim (WAY=mg):
55 The main scheduling code in GranSim is quite different from that in std
56 (concurrent) Haskell: while concurrent Haskell just iterates over the
57 threads in the runnable queue, GranSim is event driven, i.e. it iterates
58 over the events in the global event queue. -- HWL
63 //* Variables and Data structures::
64 //* Main scheduling loop::
65 //* Suspend and Resume::
67 //* Garbage Collextion Routines::
68 //* Blocking Queue Routines::
69 //* Exception Handling Routines::
70 //* Debugging Routines::
74 //@node Includes, Variables and Data structures, Main scheduling code, Main scheduling code
75 //@subsection Includes
83 #include "StgStartup.h"
87 #include "StgMiscClosures.h"
89 #include "Interpreter.h"
90 #include "Exception.h"
98 #if defined(GRAN) || defined(PAR)
99 # include "GranSimRts.h"
100 # include "GranSim.h"
101 # include "ParallelRts.h"
102 # include "Parallel.h"
103 # include "ParallelDebug.h"
104 # include "FetchMe.h"
111 //@node Variables and Data structures, Prototypes, Includes, Main scheduling code
112 //@subsection Variables and Data structures
116 * These are the threads which clients have requested that we run.
118 * In an SMP build, we might have several concurrent clients all
119 * waiting for results, and each one will wait on a condition variable
120 * until the result is available.
122 * In non-SMP, clients are strictly nested: the first client calls
123 * into the RTS, which might call out again to C with a _ccall_GC, and
124 * eventually re-enter the RTS.
126 * Main threads information is kept in a linked list:
128 //@cindex StgMainThread
129 typedef struct StgMainThread_ {
131 SchedulerStatus stat;
134 pthread_cond_t wakeup;
136 struct StgMainThread_ *link;
139 /* Main thread queue.
140 * Locks required: sched_mutex.
142 static StgMainThread *main_threads;
145 * Locks required: sched_mutex.
149 StgTSO* ActiveTSO = NULL; /* for assigning system costs; GranSim-Light only */
150 /* rtsTime TimeOfNextEvent, EndOfTimeSlice; now in GranSim.c */
153 In GranSim we have a runable and a blocked queue for each processor.
154 In order to minimise code changes new arrays run_queue_hds/tls
155 are created. run_queue_hd is then a short cut (macro) for
156 run_queue_hds[CurrentProc] (see GranSim.h).
159 StgTSO *run_queue_hds[MAX_PROC], *run_queue_tls[MAX_PROC];
160 StgTSO *blocked_queue_hds[MAX_PROC], *blocked_queue_tls[MAX_PROC];
161 StgTSO *ccalling_threadss[MAX_PROC];
162 /* We use the same global list of threads (all_threads) in GranSim as in
163 the std RTS (i.e. we are cheating). However, we don't use this list in
164 the GranSim specific code at the moment (so we are only potentially
169 StgTSO *run_queue_hd, *run_queue_tl;
170 StgTSO *blocked_queue_hd, *blocked_queue_tl;
171 StgTSO *sleeping_queue; /* perhaps replace with a hash table? */
175 /* Linked list of all threads.
176 * Used for detecting garbage collected threads.
180 /* Threads suspended in _ccall_GC.
182 static StgTSO *suspended_ccalling_threads;
184 static void GetRoots(void);
185 static StgTSO *threadStackOverflow(StgTSO *tso);
187 /* KH: The following two flags are shared memory locations. There is no need
188 to lock them, since they are only unset at the end of a scheduler
192 /* flag set by signal handler to precipitate a context switch */
193 //@cindex context_switch
196 /* if this flag is set as well, give up execution */
197 //@cindex interrupted
200 /* Next thread ID to allocate.
201 * Locks required: sched_mutex
203 //@cindex next_thread_id
204 StgThreadID next_thread_id = 1;
207 * Pointers to the state of the current thread.
208 * Rule of thumb: if CurrentTSO != NULL, then we're running a Haskell
209 * thread. If CurrentTSO == NULL, then we're at the scheduler level.
212 /* The smallest stack size that makes any sense is:
213 * RESERVED_STACK_WORDS (so we can get back from the stack overflow)
214 * + sizeofW(StgStopFrame) (the stg_stop_thread_info frame)
215 * + 1 (the realworld token for an IO thread)
216 * + 1 (the closure to enter)
218 * A thread with this stack will bomb immediately with a stack
219 * overflow, which will increase its stack size.
222 #define MIN_STACK_WORDS (RESERVED_STACK_WORDS + sizeofW(StgStopFrame) + 2)
224 /* Free capability list.
225 * Locks required: sched_mutex.
228 //@cindex free_capabilities
229 //@cindex n_free_capabilities
230 Capability *free_capabilities; /* Available capabilities for running threads */
231 nat n_free_capabilities; /* total number of available capabilities */
233 //@cindex MainRegTable
234 Capability MainRegTable; /* for non-SMP, we have one global capability */
241 /* This is used in `TSO.h' and gcc 2.96 insists that this variable actually
242 * exists - earlier gccs apparently didn't.
249 /* All our current task ids, saved in case we need to kill them later.
256 void addToBlockedQueue ( StgTSO *tso );
258 static void schedule ( void );
259 void interruptStgRts ( void );
261 static StgTSO * createThread_ ( nat size, rtsBool have_lock, StgInt pri );
263 static StgTSO * createThread_ ( nat size, rtsBool have_lock );
266 static void detectBlackHoles ( void );
269 static void sched_belch(char *s, ...);
273 //@cindex sched_mutex
275 //@cindex thread_ready_cond
276 //@cindex gc_pending_cond
277 pthread_mutex_t sched_mutex = PTHREAD_MUTEX_INITIALIZER;
278 pthread_mutex_t term_mutex = PTHREAD_MUTEX_INITIALIZER;
279 pthread_cond_t thread_ready_cond = PTHREAD_COND_INITIALIZER;
280 pthread_cond_t gc_pending_cond = PTHREAD_COND_INITIALIZER;
287 rtsTime TimeOfLastYield;
288 rtsBool emitSchedule = rtsTrue;
292 char *whatNext_strs[] = {
300 char *threadReturnCode_strs[] = {
301 "HeapOverflow", /* might also be StackOverflow */
310 StgTSO * createSparkThread(rtsSpark spark);
311 StgTSO * activateSpark (rtsSpark spark);
315 * The thread state for the main thread.
316 // ToDo: check whether not needed any more
320 //@node Main scheduling loop, Suspend and Resume, Prototypes, Main scheduling code
321 //@subsection Main scheduling loop
323 /* ---------------------------------------------------------------------------
324 Main scheduling loop.
326 We use round-robin scheduling, each thread returning to the
327 scheduler loop when one of these conditions is detected:
330 * timer expires (thread yields)
335 Locking notes: we acquire the scheduler lock once at the beginning
336 of the scheduler loop, and release it when
338 * running a thread, or
339 * waiting for work, or
340 * waiting for a GC to complete.
343 In a GranSim setup this loop iterates over the global event queue.
344 This revolves around the global event queue, which determines what
345 to do next. Therefore, it's more complicated than either the
346 concurrent or the parallel (GUM) setup.
349 GUM iterates over incoming messages.
350 It starts with nothing to do (thus CurrentTSO == END_TSO_QUEUE),
351 and sends out a fish whenever it has nothing to do; in-between
352 doing the actual reductions (shared code below) it processes the
353 incoming messages and deals with delayed operations
354 (see PendingFetches).
355 This is not the ugliest code you could imagine, but it's bloody close.
357 ------------------------------------------------------------------------ */
364 StgThreadReturnCode ret;
372 rtsBool receivedFinish = rtsFalse;
374 nat tp_size, sp_size; // stats only
377 rtsBool was_interrupted = rtsFalse;
379 ACQUIRE_LOCK(&sched_mutex);
383 /* set up first event to get things going */
384 /* ToDo: assign costs for system setup and init MainTSO ! */
385 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
387 CurrentTSO, (StgClosure*)NULL, (rtsSpark*)NULL);
390 fprintf(stderr, "GRAN: Init CurrentTSO (in schedule) = %p\n", CurrentTSO);
391 G_TSO(CurrentTSO, 5));
393 if (RtsFlags.GranFlags.Light) {
394 /* Save current time; GranSim Light only */
395 CurrentTSO->gran.clock = CurrentTime[CurrentProc];
398 event = get_next_event();
400 while (event!=(rtsEvent*)NULL) {
401 /* Choose the processor with the next event */
402 CurrentProc = event->proc;
403 CurrentTSO = event->tso;
407 while (!receivedFinish) { /* set by processMessages */
408 /* when receiving PP_FINISH message */
415 IF_DEBUG(scheduler, printAllThreads());
417 /* If we're interrupted (the user pressed ^C, or some other
418 * termination condition occurred), kill all the currently running
422 IF_DEBUG(scheduler, sched_belch("interrupted"));
424 interrupted = rtsFalse;
425 was_interrupted = rtsTrue;
428 /* Go through the list of main threads and wake up any
429 * clients whose computations have finished. ToDo: this
430 * should be done more efficiently without a linear scan
431 * of the main threads list, somehow...
435 StgMainThread *m, **prev;
436 prev = &main_threads;
437 for (m = main_threads; m != NULL; m = m->link) {
438 switch (m->tso->what_next) {
441 *(m->ret) = (StgClosure *)m->tso->sp[0];
445 pthread_cond_broadcast(&m->wakeup);
449 if (was_interrupted) {
450 m->stat = Interrupted;
454 pthread_cond_broadcast(&m->wakeup);
464 /* in GUM do this only on the Main PE */
467 /* If our main thread has finished or been killed, return.
470 StgMainThread *m = main_threads;
471 if (m->tso->what_next == ThreadComplete
472 || m->tso->what_next == ThreadKilled) {
473 main_threads = main_threads->link;
474 if (m->tso->what_next == ThreadComplete) {
475 /* we finished successfully, fill in the return value */
476 if (m->ret) { *(m->ret) = (StgClosure *)m->tso->sp[0]; };
480 if (was_interrupted) {
481 m->stat = Interrupted;
491 /* Top up the run queue from our spark pool. We try to make the
492 * number of threads in the run queue equal to the number of
497 nat n = n_free_capabilities;
498 StgTSO *tso = run_queue_hd;
500 /* Count the run queue */
501 while (n > 0 && tso != END_TSO_QUEUE) {
508 spark = findSpark(rtsFalse);
510 break; /* no more sparks in the pool */
512 /* I'd prefer this to be done in activateSpark -- HWL */
513 /* tricky - it needs to hold the scheduler lock and
514 * not try to re-acquire it -- SDM */
515 createSparkThread(spark);
517 sched_belch("==^^ turning spark of closure %p into a thread",
518 (StgClosure *)spark));
521 /* We need to wake up the other tasks if we just created some
524 if (n_free_capabilities - n > 1) {
525 pthread_cond_signal(&thread_ready_cond);
530 /* Check whether any waiting threads need to be woken up. If the
531 * run queue is empty, and there are no other tasks running, we
532 * can wait indefinitely for something to happen.
533 * ToDo: what if another client comes along & requests another
536 if (blocked_queue_hd != END_TSO_QUEUE || sleeping_queue != END_TSO_QUEUE) {
538 (run_queue_hd == END_TSO_QUEUE)
540 && (n_free_capabilities == RtsFlags.ParFlags.nNodes)
544 /* we can be interrupted while waiting for I/O... */
545 if (interrupted) continue;
547 /* check for signals each time around the scheduler */
548 #ifndef mingw32_TARGET_OS
549 if (signals_pending()) {
550 start_signal_handlers();
555 * Detect deadlock: when we have no threads to run, there are no
556 * threads waiting on I/O or sleeping, and all the other tasks are
557 * waiting for work, we must have a deadlock of some description.
559 * We first try to find threads blocked on themselves (ie. black
560 * holes), and generate NonTermination exceptions where necessary.
562 * If no threads are black holed, we have a deadlock situation, so
563 * inform all the main threads.
566 if (blocked_queue_hd == END_TSO_QUEUE
567 && run_queue_hd == END_TSO_QUEUE
568 && sleeping_queue == END_TSO_QUEUE
569 && (n_free_capabilities == RtsFlags.ParFlags.nNodes))
571 IF_DEBUG(scheduler, sched_belch("deadlocked, checking for black holes..."));
573 if (run_queue_hd == END_TSO_QUEUE) {
575 for (m = main_threads; m != NULL; m = m->link) {
578 pthread_cond_broadcast(&m->wakeup);
584 /* ToDo: add deadlock detection in GUM (similar to SMP) -- HWL */
586 if (blocked_queue_hd == END_TSO_QUEUE
587 && run_queue_hd == END_TSO_QUEUE
588 && sleeping_queue == END_TSO_QUEUE)
590 IF_DEBUG(scheduler, sched_belch("deadlocked, checking for black holes..."));
592 if (run_queue_hd == END_TSO_QUEUE) {
593 StgMainThread *m = main_threads;
596 main_threads = m->link;
603 /* If there's a GC pending, don't do anything until it has
607 IF_DEBUG(scheduler,sched_belch("waiting for GC"));
608 pthread_cond_wait(&gc_pending_cond, &sched_mutex);
611 /* block until we've got a thread on the run queue and a free
614 while (run_queue_hd == END_TSO_QUEUE || free_capabilities == NULL) {
615 IF_DEBUG(scheduler, sched_belch("waiting for work"));
616 pthread_cond_wait(&thread_ready_cond, &sched_mutex);
617 IF_DEBUG(scheduler, sched_belch("work now available"));
623 if (RtsFlags.GranFlags.Light)
624 GranSimLight_enter_system(event, &ActiveTSO); // adjust ActiveTSO etc
626 /* adjust time based on time-stamp */
627 if (event->time > CurrentTime[CurrentProc] &&
628 event->evttype != ContinueThread)
629 CurrentTime[CurrentProc] = event->time;
631 /* Deal with the idle PEs (may issue FindWork or MoveSpark events) */
632 if (!RtsFlags.GranFlags.Light)
635 IF_DEBUG(gran, fprintf(stderr, "GRAN: switch by event-type\n"));
637 /* main event dispatcher in GranSim */
638 switch (event->evttype) {
639 /* Should just be continuing execution */
641 IF_DEBUG(gran, fprintf(stderr, "GRAN: doing ContinueThread\n"));
642 /* ToDo: check assertion
643 ASSERT(run_queue_hd != (StgTSO*)NULL &&
644 run_queue_hd != END_TSO_QUEUE);
646 /* Ignore ContinueThreads for fetching threads (if synchr comm) */
647 if (!RtsFlags.GranFlags.DoAsyncFetch &&
648 procStatus[CurrentProc]==Fetching) {
649 belch("ghuH: Spurious ContinueThread while Fetching ignored; TSO %d (%p) [PE %d]",
650 CurrentTSO->id, CurrentTSO, CurrentProc);
653 /* Ignore ContinueThreads for completed threads */
654 if (CurrentTSO->what_next == ThreadComplete) {
655 belch("ghuH: found a ContinueThread event for completed thread %d (%p) [PE %d] (ignoring ContinueThread)",
656 CurrentTSO->id, CurrentTSO, CurrentProc);
659 /* Ignore ContinueThreads for threads that are being migrated */
660 if (PROCS(CurrentTSO)==Nowhere) {
661 belch("ghuH: trying to run the migrating TSO %d (%p) [PE %d] (ignoring ContinueThread)",
662 CurrentTSO->id, CurrentTSO, CurrentProc);
665 /* The thread should be at the beginning of the run queue */
666 if (CurrentTSO!=run_queue_hds[CurrentProc]) {
667 belch("ghuH: TSO %d (%p) [PE %d] is not at the start of the run_queue when doing a ContinueThread",
668 CurrentTSO->id, CurrentTSO, CurrentProc);
669 break; // run the thread anyway
672 new_event(proc, proc, CurrentTime[proc],
674 (StgTSO*)NULL, (StgClosure*)NULL, (rtsSpark*)NULL);
676 */ /* Catches superfluous CONTINUEs -- should be unnecessary */
677 break; // now actually run the thread; DaH Qu'vam yImuHbej
680 do_the_fetchnode(event);
681 goto next_thread; /* handle next event in event queue */
684 do_the_globalblock(event);
685 goto next_thread; /* handle next event in event queue */
688 do_the_fetchreply(event);
689 goto next_thread; /* handle next event in event queue */
691 case UnblockThread: /* Move from the blocked queue to the tail of */
692 do_the_unblock(event);
693 goto next_thread; /* handle next event in event queue */
695 case ResumeThread: /* Move from the blocked queue to the tail of */
696 /* the runnable queue ( i.e. Qu' SImqa'lu') */
697 event->tso->gran.blocktime +=
698 CurrentTime[CurrentProc] - event->tso->gran.blockedat;
699 do_the_startthread(event);
700 goto next_thread; /* handle next event in event queue */
703 do_the_startthread(event);
704 goto next_thread; /* handle next event in event queue */
707 do_the_movethread(event);
708 goto next_thread; /* handle next event in event queue */
711 do_the_movespark(event);
712 goto next_thread; /* handle next event in event queue */
715 do_the_findwork(event);
716 goto next_thread; /* handle next event in event queue */
719 barf("Illegal event type %u\n", event->evttype);
722 /* This point was scheduler_loop in the old RTS */
724 IF_DEBUG(gran, belch("GRAN: after main switch"));
726 TimeOfLastEvent = CurrentTime[CurrentProc];
727 TimeOfNextEvent = get_time_of_next_event();
728 IgnoreEvents=(TimeOfNextEvent==0); // HWL HACK
729 // CurrentTSO = ThreadQueueHd;
731 IF_DEBUG(gran, belch("GRAN: time of next event is: %ld",
734 if (RtsFlags.GranFlags.Light)
735 GranSimLight_leave_system(event, &ActiveTSO);
737 EndOfTimeSlice = CurrentTime[CurrentProc]+RtsFlags.GranFlags.time_slice;
740 belch("GRAN: end of time-slice is %#lx", EndOfTimeSlice));
742 /* in a GranSim setup the TSO stays on the run queue */
744 /* Take a thread from the run queue. */
745 t = POP_RUN_QUEUE(); // take_off_run_queue(t);
748 fprintf(stderr, "GRAN: About to run current thread, which is\n");
751 context_switch = 0; // turned on via GranYield, checking events and time slice
754 DumpGranEvent(GR_SCHEDULE, t));
756 procStatus[CurrentProc] = Busy;
759 if (PendingFetches != END_BF_QUEUE) {
763 /* ToDo: phps merge with spark activation above */
764 /* check whether we have local work and send requests if we have none */
765 if (EMPTY_RUN_QUEUE()) { /* no runnable threads */
766 /* :-[ no local threads => look out for local sparks */
767 /* the spark pool for the current PE */
768 pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable
769 if (advisory_thread_count < RtsFlags.ParFlags.maxThreads &&
770 pool->hd < pool->tl) {
772 * ToDo: add GC code check that we really have enough heap afterwards!!
774 * If we're here (no runnable threads) and we have pending
775 * sparks, we must have a space problem. Get enough space
776 * to turn one of those pending sparks into a
780 spark = findSpark(rtsFalse); /* get a spark */
781 if (spark != (rtsSpark) NULL) {
782 tso = activateSpark(spark); /* turn the spark into a thread */
783 IF_PAR_DEBUG(schedule,
784 belch("==== schedule: Created TSO %d (%p); %d threads active",
785 tso->id, tso, advisory_thread_count));
787 if (tso==END_TSO_QUEUE) { /* failed to activate spark->back to loop */
788 belch("==^^ failed to activate spark");
790 } /* otherwise fall through & pick-up new tso */
792 IF_PAR_DEBUG(verbose,
793 belch("==^^ no local sparks (spark pool contains only NFs: %d)",
794 spark_queue_len(pool)));
799 /* If we still have no work we need to send a FISH to get a spark
802 if (EMPTY_RUN_QUEUE()) {
803 /* =8-[ no local sparks => look for work on other PEs */
805 * We really have absolutely no work. Send out a fish
806 * (there may be some out there already), and wait for
807 * something to arrive. We clearly can't run any threads
808 * until a SCHEDULE or RESUME arrives, and so that's what
809 * we're hoping to see. (Of course, we still have to
810 * respond to other types of messages.)
812 TIME now = msTime() /*CURRENT_TIME*/;
813 IF_PAR_DEBUG(verbose,
814 belch("-- now=%ld", now));
815 IF_PAR_DEBUG(verbose,
816 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
817 (last_fish_arrived_at!=0 &&
818 last_fish_arrived_at+RtsFlags.ParFlags.fishDelay > now)) {
819 belch("--$$ delaying FISH until %ld (last fish %ld, delay %ld, now %ld)",
820 last_fish_arrived_at+RtsFlags.ParFlags.fishDelay,
821 last_fish_arrived_at,
822 RtsFlags.ParFlags.fishDelay, now);
825 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
826 (last_fish_arrived_at==0 ||
827 (last_fish_arrived_at+RtsFlags.ParFlags.fishDelay <= now))) {
828 /* outstandingFishes is set in sendFish, processFish;
829 avoid flooding system with fishes via delay */
831 sendFish(pe, mytid, NEW_FISH_AGE, NEW_FISH_HISTORY,
834 // Global statistics: count no. of fishes
835 if (RtsFlags.ParFlags.ParStats.Global &&
836 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
837 globalParStats.tot_fish_mess++;
841 receivedFinish = processMessages();
844 } else if (PacketsWaiting()) { /* Look for incoming messages */
845 receivedFinish = processMessages();
848 /* Now we are sure that we have some work available */
849 ASSERT(run_queue_hd != END_TSO_QUEUE);
851 /* Take a thread from the run queue, if we have work */
852 t = POP_RUN_QUEUE(); // take_off_run_queue(END_TSO_QUEUE);
853 IF_DEBUG(sanity,checkTSO(t));
855 /* ToDo: write something to the log-file
856 if (RTSflags.ParFlags.granSimStats && !sameThread)
857 DumpGranEvent(GR_SCHEDULE, RunnableThreadsHd);
861 /* the spark pool for the current PE */
862 pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable
865 belch("--=^ %d threads, %d sparks on [%#x]",
866 run_queue_len(), spark_queue_len(pool), CURRENT_PROC));
869 if (0 && RtsFlags.ParFlags.ParStats.Full &&
870 t && LastTSO && t->id != LastTSO->id &&
871 LastTSO->why_blocked == NotBlocked &&
872 LastTSO->what_next != ThreadComplete) {
873 // if previously scheduled TSO not blocked we have to record the context switch
874 DumpVeryRawGranEvent(TimeOfLastYield, CURRENT_PROC, CURRENT_PROC,
875 GR_DESCHEDULE, LastTSO, (StgClosure *)NULL, 0, 0);
878 if (RtsFlags.ParFlags.ParStats.Full &&
879 (emitSchedule /* forced emit */ ||
880 (t && LastTSO && t->id != LastTSO->id))) {
882 we are running a different TSO, so write a schedule event to log file
883 NB: If we use fair scheduling we also have to write a deschedule
884 event for LastTSO; with unfair scheduling we know that the
885 previous tso has blocked whenever we switch to another tso, so
886 we don't need it in GUM for now
888 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
889 GR_SCHEDULE, t, (StgClosure *)NULL, 0, 0);
890 emitSchedule = rtsFalse;
894 #else /* !GRAN && !PAR */
896 /* grab a thread from the run queue
898 ASSERT(run_queue_hd != END_TSO_QUEUE);
900 IF_DEBUG(sanity,checkTSO(t));
907 cap = free_capabilities;
908 free_capabilities = cap->link;
909 n_free_capabilities--;
914 cap->rCurrentTSO = t;
916 /* context switches are now initiated by the timer signal, unless
917 * the user specified "context switch as often as possible", with
920 if (RtsFlags.ConcFlags.ctxtSwitchTicks == 0
921 && (run_queue_hd != END_TSO_QUEUE
922 || blocked_queue_hd != END_TSO_QUEUE
923 || sleeping_queue != END_TSO_QUEUE))
928 RELEASE_LOCK(&sched_mutex);
930 IF_DEBUG(scheduler, sched_belch("-->> Running TSO %ld (%p) %s ...",
931 t->id, t, whatNext_strs[t->what_next]));
933 /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
934 /* Run the current thread
936 switch (cap->rCurrentTSO->what_next) {
939 /* Thread already finished, return to scheduler. */
940 ret = ThreadFinished;
943 ret = StgRun((StgFunPtr) stg_enterStackTop, cap);
946 ret = StgRun((StgFunPtr) stg_returnToStackTop, cap);
948 case ThreadEnterInterp:
949 ret = interpretBCO(cap);
952 barf("schedule: invalid what_next field");
954 /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
956 /* Costs for the scheduler are assigned to CCS_SYSTEM */
961 ACQUIRE_LOCK(&sched_mutex);
964 IF_DEBUG(scheduler,fprintf(stderr,"scheduler (task %ld): ", pthread_self()););
965 #elif !defined(GRAN) && !defined(PAR)
966 IF_DEBUG(scheduler,fprintf(stderr,"scheduler: "););
968 t = cap->rCurrentTSO;
971 /* HACK 675: if the last thread didn't yield, make sure to print a
972 SCHEDULE event to the log file when StgRunning the next thread, even
973 if it is the same one as before */
975 TimeOfLastYield = CURRENT_TIME;
982 DumpGranEvent(GR_DESCHEDULE, t));
983 globalGranStats.tot_heapover++;
986 //DumpGranEvent(GR_DESCHEDULE, t);
987 globalParStats.tot_heapover++;
989 /* make all the running tasks block on a condition variable,
990 * maybe set context_switch and wait till they all pile in,
991 * then have them wait on a GC condition variable.
993 IF_DEBUG(scheduler,belch("--<< thread %ld (%p; %s) stopped: HeapOverflow",
994 t->id, t, whatNext_strs[t->what_next]));
997 ASSERT(!is_on_queue(t,CurrentProc));
999 /* Currently we emit a DESCHEDULE event before GC in GUM.
1000 ToDo: either add separate event to distinguish SYSTEM time from rest
1001 or just nuke this DESCHEDULE (and the following SCHEDULE) */
1002 if (0 && RtsFlags.ParFlags.ParStats.Full) {
1003 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
1004 GR_DESCHEDULE, t, (StgClosure *)NULL, 0, 0);
1005 emitSchedule = rtsTrue;
1009 ready_to_gc = rtsTrue;
1010 context_switch = 1; /* stop other threads ASAP */
1011 PUSH_ON_RUN_QUEUE(t);
1012 /* actual GC is done at the end of the while loop */
1018 DumpGranEvent(GR_DESCHEDULE, t));
1019 globalGranStats.tot_stackover++;
1022 // DumpGranEvent(GR_DESCHEDULE, t);
1023 globalParStats.tot_stackover++;
1025 IF_DEBUG(scheduler,belch("--<< thread %ld (%p; %s) stopped, StackOverflow",
1026 t->id, t, whatNext_strs[t->what_next]));
1027 /* just adjust the stack for this thread, then pop it back
1033 /* enlarge the stack */
1034 StgTSO *new_t = threadStackOverflow(t);
1036 /* This TSO has moved, so update any pointers to it from the
1037 * main thread stack. It better not be on any other queues...
1038 * (it shouldn't be).
1040 for (m = main_threads; m != NULL; m = m->link) {
1045 threadPaused(new_t);
1046 PUSH_ON_RUN_QUEUE(new_t);
1050 case ThreadYielding:
1053 DumpGranEvent(GR_DESCHEDULE, t));
1054 globalGranStats.tot_yields++;
1057 // DumpGranEvent(GR_DESCHEDULE, t);
1058 globalParStats.tot_yields++;
1060 /* put the thread back on the run queue. Then, if we're ready to
1061 * GC, check whether this is the last task to stop. If so, wake
1062 * up the GC thread. getThread will block during a GC until the
1066 if (t->what_next == ThreadEnterInterp) {
1067 /* ToDo: or maybe a timer expired when we were in Hugs?
1068 * or maybe someone hit ctrl-C
1070 belch("--<< thread %ld (%p; %s) stopped to switch to Hugs",
1071 t->id, t, whatNext_strs[t->what_next]);
1073 belch("--<< thread %ld (%p; %s) stopped, yielding",
1074 t->id, t, whatNext_strs[t->what_next]);
1081 //belch("&& Doing sanity check on yielding TSO %ld.", t->id);
1083 ASSERT(t->link == END_TSO_QUEUE);
1085 ASSERT(!is_on_queue(t,CurrentProc));
1088 //belch("&& Doing sanity check on all ThreadQueues (and their TSOs).");
1089 checkThreadQsSanity(rtsTrue));
1092 if (RtsFlags.ParFlags.doFairScheduling) {
1093 /* this does round-robin scheduling; good for concurrency */
1094 APPEND_TO_RUN_QUEUE(t);
1096 /* this does unfair scheduling; good for parallelism */
1097 PUSH_ON_RUN_QUEUE(t);
1100 /* this does round-robin scheduling; good for concurrency */
1101 APPEND_TO_RUN_QUEUE(t);
1104 /* add a ContinueThread event to actually process the thread */
1105 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1107 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1109 belch("GRAN: eventq and runnableq after adding yielded thread to queue again:");
1118 belch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: ",
1119 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)));
1120 if (t->block_info.closure!=(StgClosure*)NULL) print_bq(t->block_info.closure));
1122 // ??? needed; should emit block before
1124 DumpGranEvent(GR_DESCHEDULE, t));
1125 prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
1128 ASSERT(procStatus[CurrentProc]==Busy ||
1129 ((procStatus[CurrentProc]==Fetching) &&
1130 (t->block_info.closure!=(StgClosure*)NULL)));
1131 if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
1132 !(!RtsFlags.GranFlags.DoAsyncFetch &&
1133 procStatus[CurrentProc]==Fetching))
1134 procStatus[CurrentProc] = Idle;
1138 belch("--<< thread %ld (%p; %s) stopped, blocking on node %p with BQ: ",
1139 t->id, t, whatNext_strs[t->what_next], t->block_info.closure));
1142 if (t->block_info.closure!=(StgClosure*)NULL)
1143 print_bq(t->block_info.closure));
1145 /* Send a fetch (if BlockedOnGA) and dump event to log file */
1148 /* whatever we schedule next, we must log that schedule */
1149 emitSchedule = rtsTrue;
1152 /* don't need to do anything. Either the thread is blocked on
1153 * I/O, in which case we'll have called addToBlockedQueue
1154 * previously, or it's blocked on an MVar or Blackhole, in which
1155 * case it'll be on the relevant queue already.
1158 fprintf(stderr, "--<< thread %d (%p) stopped: ", t->id, t);
1159 printThreadBlockage(t);
1160 fprintf(stderr, "\n"));
1162 /* Only for dumping event to log file
1163 ToDo: do I need this in GranSim, too?
1170 case ThreadFinished:
1171 /* Need to check whether this was a main thread, and if so, signal
1172 * the task that started it with the return value. If we have no
1173 * more main threads, we probably need to stop all the tasks until
1176 /* We also end up here if the thread kills itself with an
1177 * uncaught exception, see Exception.hc.
1179 IF_DEBUG(scheduler,belch("--++ thread %d (%p) finished", t->id, t));
1181 endThread(t, CurrentProc); // clean-up the thread
1183 /* For now all are advisory -- HWL */
1184 //if(t->priority==AdvisoryPriority) ??
1185 advisory_thread_count--;
1188 if(t->dist.priority==RevalPriority)
1192 if (RtsFlags.ParFlags.ParStats.Full &&
1193 !RtsFlags.ParFlags.ParStats.Suppressed)
1194 DumpEndEvent(CURRENT_PROC, t, rtsFalse /* not mandatory */);
1199 barf("schedule: invalid thread return code %d", (int)ret);
1203 cap->link = free_capabilities;
1204 free_capabilities = cap;
1205 n_free_capabilities++;
1209 if (ready_to_gc && n_free_capabilities == RtsFlags.ParFlags.nNodes)
1214 /* everybody back, start the GC.
1215 * Could do it in this thread, or signal a condition var
1216 * to do it in another thread. Either way, we need to
1217 * broadcast on gc_pending_cond afterward.
1220 IF_DEBUG(scheduler,sched_belch("doing GC"));
1222 GarbageCollect(GetRoots,rtsFalse);
1223 ready_to_gc = rtsFalse;
1225 pthread_cond_broadcast(&gc_pending_cond);
1228 /* add a ContinueThread event to continue execution of current thread */
1229 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1231 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1233 fprintf(stderr, "GRAN: eventq and runnableq after Garbage collection:\n");
1240 IF_GRAN_DEBUG(unused,
1241 print_eventq(EventHd));
1243 event = get_next_event();
1247 /* ToDo: wait for next message to arrive rather than busy wait */
1252 t = take_off_run_queue(END_TSO_QUEUE);
1255 } /* end of while(1) */
1256 IF_PAR_DEBUG(verbose,
1257 belch("== Leaving schedule() after having received Finish"));
1260 /* ---------------------------------------------------------------------------
1261 * deleteAllThreads(): kill all the live threads.
1263 * This is used when we catch a user interrupt (^C), before performing
1264 * any necessary cleanups and running finalizers.
1265 * ------------------------------------------------------------------------- */
1267 void deleteAllThreads ( void )
1270 IF_DEBUG(scheduler,sched_belch("deleting all threads"));
1271 for (t = run_queue_hd; t != END_TSO_QUEUE; t = t->link) {
1274 for (t = blocked_queue_hd; t != END_TSO_QUEUE; t = t->link) {
1277 for (t = sleeping_queue; t != END_TSO_QUEUE; t = t->link) {
1280 run_queue_hd = run_queue_tl = END_TSO_QUEUE;
1281 blocked_queue_hd = blocked_queue_tl = END_TSO_QUEUE;
1282 sleeping_queue = END_TSO_QUEUE;
1285 /* startThread and insertThread are now in GranSim.c -- HWL */
1287 //@node Suspend and Resume, Run queue code, Main scheduling loop, Main scheduling code
1288 //@subsection Suspend and Resume
1290 /* ---------------------------------------------------------------------------
1291 * Suspending & resuming Haskell threads.
1293 * When making a "safe" call to C (aka _ccall_GC), the task gives back
1294 * its capability before calling the C function. This allows another
1295 * task to pick up the capability and carry on running Haskell
1296 * threads. It also means that if the C call blocks, it won't lock
1299 * The Haskell thread making the C call is put to sleep for the
1300 * duration of the call, on the susepended_ccalling_threads queue. We
1301 * give out a token to the task, which it can use to resume the thread
1302 * on return from the C function.
1303 * ------------------------------------------------------------------------- */
1306 suspendThread( Capability *cap )
1310 ACQUIRE_LOCK(&sched_mutex);
1313 sched_belch("thread %d did a _ccall_gc", cap->rCurrentTSO->id));
1315 threadPaused(cap->rCurrentTSO);
1316 cap->rCurrentTSO->link = suspended_ccalling_threads;
1317 suspended_ccalling_threads = cap->rCurrentTSO;
1319 /* Use the thread ID as the token; it should be unique */
1320 tok = cap->rCurrentTSO->id;
1323 cap->link = free_capabilities;
1324 free_capabilities = cap;
1325 n_free_capabilities++;
1328 RELEASE_LOCK(&sched_mutex);
1333 resumeThread( StgInt tok )
1335 StgTSO *tso, **prev;
1338 ACQUIRE_LOCK(&sched_mutex);
1340 prev = &suspended_ccalling_threads;
1341 for (tso = suspended_ccalling_threads;
1342 tso != END_TSO_QUEUE;
1343 prev = &tso->link, tso = tso->link) {
1344 if (tso->id == (StgThreadID)tok) {
1349 if (tso == END_TSO_QUEUE) {
1350 barf("resumeThread: thread not found");
1352 tso->link = END_TSO_QUEUE;
1355 while (free_capabilities == NULL) {
1356 IF_DEBUG(scheduler, sched_belch("waiting to resume"));
1357 pthread_cond_wait(&thread_ready_cond, &sched_mutex);
1358 IF_DEBUG(scheduler, sched_belch("resuming thread %d", tso->id));
1360 cap = free_capabilities;
1361 free_capabilities = cap->link;
1362 n_free_capabilities--;
1364 cap = &MainRegTable;
1367 cap->rCurrentTSO = tso;
1369 RELEASE_LOCK(&sched_mutex);
1374 /* ---------------------------------------------------------------------------
1376 * ------------------------------------------------------------------------ */
1377 static void unblockThread(StgTSO *tso);
1379 /* ---------------------------------------------------------------------------
1380 * Comparing Thread ids.
1382 * This is used from STG land in the implementation of the
1383 * instances of Eq/Ord for ThreadIds.
1384 * ------------------------------------------------------------------------ */
1386 int cmp_thread(const StgTSO *tso1, const StgTSO *tso2)
1388 StgThreadID id1 = tso1->id;
1389 StgThreadID id2 = tso2->id;
1391 if (id1 < id2) return (-1);
1392 if (id1 > id2) return 1;
1396 /* ---------------------------------------------------------------------------
1397 Create a new thread.
1399 The new thread starts with the given stack size. Before the
1400 scheduler can run, however, this thread needs to have a closure
1401 (and possibly some arguments) pushed on its stack. See
1402 pushClosure() in Schedule.h.
1404 createGenThread() and createIOThread() (in SchedAPI.h) are
1405 convenient packaged versions of this function.
1407 currently pri (priority) is only used in a GRAN setup -- HWL
1408 ------------------------------------------------------------------------ */
1409 //@cindex createThread
1411 /* currently pri (priority) is only used in a GRAN setup -- HWL */
1413 createThread(nat stack_size, StgInt pri)
1415 return createThread_(stack_size, rtsFalse, pri);
1419 createThread_(nat size, rtsBool have_lock, StgInt pri)
1423 createThread(nat stack_size)
1425 return createThread_(stack_size, rtsFalse);
1429 createThread_(nat size, rtsBool have_lock)
1436 /* First check whether we should create a thread at all */
1438 /* check that no more than RtsFlags.ParFlags.maxThreads threads are created */
1439 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads) {
1441 belch("{createThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
1442 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
1443 return END_TSO_QUEUE;
1449 ASSERT(!RtsFlags.GranFlags.Light || CurrentProc==0);
1452 // ToDo: check whether size = stack_size - TSO_STRUCT_SIZEW
1454 /* catch ridiculously small stack sizes */
1455 if (size < MIN_STACK_WORDS + TSO_STRUCT_SIZEW) {
1456 size = MIN_STACK_WORDS + TSO_STRUCT_SIZEW;
1459 stack_size = size - TSO_STRUCT_SIZEW;
1461 tso = (StgTSO *)allocate(size);
1462 TICK_ALLOC_TSO(size-TSO_STRUCT_SIZEW, 0);
1464 SET_HDR(tso, &stg_TSO_info, CCS_SYSTEM);
1466 SET_GRAN_HDR(tso, ThisPE);
1468 tso->what_next = ThreadEnterGHC;
1470 /* tso->id needs to be unique. For now we use a heavyweight mutex to
1471 * protect the increment operation on next_thread_id.
1472 * In future, we could use an atomic increment instead.
1474 if (!have_lock) { ACQUIRE_LOCK(&sched_mutex); }
1475 tso->id = next_thread_id++;
1476 if (!have_lock) { RELEASE_LOCK(&sched_mutex); }
1478 tso->why_blocked = NotBlocked;
1479 tso->blocked_exceptions = NULL;
1481 //tso->splim = (P_)&(tso->stack) + RESERVED_STACK_WORDS;
1482 tso->stack_size = stack_size;
1483 tso->max_stack_size = round_to_mblocks(RtsFlags.GcFlags.maxStkSize)
1485 tso->sp = (P_)&(tso->stack) + stack_size;
1488 tso->prof.CCCS = CCS_MAIN;
1491 /* put a stop frame on the stack */
1492 tso->sp -= sizeofW(StgStopFrame);
1493 SET_HDR((StgClosure*)tso->sp,(StgInfoTable *)&stg_stop_thread_info,CCS_SYSTEM);
1494 tso->su = (StgUpdateFrame*)tso->sp;
1498 tso->link = END_TSO_QUEUE;
1499 /* uses more flexible routine in GranSim */
1500 insertThread(tso, CurrentProc);
1502 /* In a non-GranSim setup the pushing of a TSO onto the runq is separated
1508 if (RtsFlags.GranFlags.GranSimStats.Full)
1509 DumpGranEvent(GR_START,tso);
1511 if (RtsFlags.ParFlags.ParStats.Full)
1512 DumpGranEvent(GR_STARTQ,tso);
1513 /* HACk to avoid SCHEDULE
1517 /* Link the new thread on the global thread list.
1519 tso->global_link = all_threads;
1523 tso->dist.priority = MandatoryPriority; //by default that is...
1527 tso->gran.pri = pri;
1529 tso->gran.magic = TSO_MAGIC; // debugging only
1531 tso->gran.sparkname = 0;
1532 tso->gran.startedat = CURRENT_TIME;
1533 tso->gran.exported = 0;
1534 tso->gran.basicblocks = 0;
1535 tso->gran.allocs = 0;
1536 tso->gran.exectime = 0;
1537 tso->gran.fetchtime = 0;
1538 tso->gran.fetchcount = 0;
1539 tso->gran.blocktime = 0;
1540 tso->gran.blockcount = 0;
1541 tso->gran.blockedat = 0;
1542 tso->gran.globalsparks = 0;
1543 tso->gran.localsparks = 0;
1544 if (RtsFlags.GranFlags.Light)
1545 tso->gran.clock = Now; /* local clock */
1547 tso->gran.clock = 0;
1549 IF_DEBUG(gran,printTSO(tso));
1552 tso->par.magic = TSO_MAGIC; // debugging only
1554 tso->par.sparkname = 0;
1555 tso->par.startedat = CURRENT_TIME;
1556 tso->par.exported = 0;
1557 tso->par.basicblocks = 0;
1558 tso->par.allocs = 0;
1559 tso->par.exectime = 0;
1560 tso->par.fetchtime = 0;
1561 tso->par.fetchcount = 0;
1562 tso->par.blocktime = 0;
1563 tso->par.blockcount = 0;
1564 tso->par.blockedat = 0;
1565 tso->par.globalsparks = 0;
1566 tso->par.localsparks = 0;
1570 globalGranStats.tot_threads_created++;
1571 globalGranStats.threads_created_on_PE[CurrentProc]++;
1572 globalGranStats.tot_sq_len += spark_queue_len(CurrentProc);
1573 globalGranStats.tot_sq_probes++;
1575 // collect parallel global statistics (currently done together with GC stats)
1576 if (RtsFlags.ParFlags.ParStats.Global &&
1577 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
1578 //fprintf(stderr, "Creating thread %d @ %11.2f\n", tso->id, usertime());
1579 globalParStats.tot_threads_created++;
1585 belch("==__ schedule: Created TSO %d (%p);",
1586 CurrentProc, tso, tso->id));
1588 IF_PAR_DEBUG(verbose,
1589 belch("==__ schedule: Created TSO %d (%p); %d threads active",
1590 tso->id, tso, advisory_thread_count));
1592 IF_DEBUG(scheduler,sched_belch("created thread %ld, stack size = %lx words",
1593 tso->id, tso->stack_size));
1600 all parallel thread creation calls should fall through the following routine.
1603 createSparkThread(rtsSpark spark)
1605 ASSERT(spark != (rtsSpark)NULL);
1606 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads)
1608 barf("{createSparkThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
1609 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
1610 return END_TSO_QUEUE;
1614 tso = createThread_(RtsFlags.GcFlags.initialStkSize, rtsTrue);
1615 if (tso==END_TSO_QUEUE)
1616 barf("createSparkThread: Cannot create TSO");
1618 tso->priority = AdvisoryPriority;
1620 pushClosure(tso,spark);
1621 PUSH_ON_RUN_QUEUE(tso);
1622 advisory_thread_count++;
1629 Turn a spark into a thread.
1630 ToDo: fix for SMP (needs to acquire SCHED_MUTEX!)
1633 //@cindex activateSpark
1635 activateSpark (rtsSpark spark)
1639 tso = createSparkThread(spark);
1640 if (RtsFlags.ParFlags.ParStats.Full) {
1641 //ASSERT(run_queue_hd == END_TSO_QUEUE); // I think ...
1642 IF_PAR_DEBUG(verbose,
1643 belch("==^^ activateSpark: turning spark of closure %p (%s) into a thread",
1644 (StgClosure *)spark, info_type((StgClosure *)spark)));
1646 // ToDo: fwd info on local/global spark to thread -- HWL
1647 // tso->gran.exported = spark->exported;
1648 // tso->gran.locked = !spark->global;
1649 // tso->gran.sparkname = spark->name;
1655 /* ---------------------------------------------------------------------------
1658 * scheduleThread puts a thread on the head of the runnable queue.
1659 * This will usually be done immediately after a thread is created.
1660 * The caller of scheduleThread must create the thread using e.g.
1661 * createThread and push an appropriate closure
1662 * on this thread's stack before the scheduler is invoked.
1663 * ------------------------------------------------------------------------ */
1666 scheduleThread(StgTSO *tso)
1668 if (tso==END_TSO_QUEUE){
1673 ACQUIRE_LOCK(&sched_mutex);
1675 /* Put the new thread on the head of the runnable queue. The caller
1676 * better push an appropriate closure on this thread's stack
1677 * beforehand. In the SMP case, the thread may start running as
1678 * soon as we release the scheduler lock below.
1680 PUSH_ON_RUN_QUEUE(tso);
1684 IF_DEBUG(scheduler,printTSO(tso));
1686 RELEASE_LOCK(&sched_mutex);
1689 /* ---------------------------------------------------------------------------
1692 * Start up Posix threads to run each of the scheduler tasks.
1693 * I believe the task ids are not needed in the system as defined.
1695 * ------------------------------------------------------------------------ */
1697 #if defined(PAR) || defined(SMP)
1699 taskStart(void) /* ( void *arg STG_UNUSED) */
1701 scheduleThread(END_TSO_QUEUE);
1705 /* ---------------------------------------------------------------------------
1708 * Initialise the scheduler. This resets all the queues - if the
1709 * queues contained any threads, they'll be garbage collected at the
1712 * This now calls startTasks(), so should only be called once! KH @ 25/10/99
1713 * ------------------------------------------------------------------------ */
1717 term_handler(int sig STG_UNUSED)
1720 ACQUIRE_LOCK(&term_mutex);
1722 RELEASE_LOCK(&term_mutex);
1727 //@cindex initScheduler
1734 for (i=0; i<=MAX_PROC; i++) {
1735 run_queue_hds[i] = END_TSO_QUEUE;
1736 run_queue_tls[i] = END_TSO_QUEUE;
1737 blocked_queue_hds[i] = END_TSO_QUEUE;
1738 blocked_queue_tls[i] = END_TSO_QUEUE;
1739 ccalling_threadss[i] = END_TSO_QUEUE;
1740 sleeping_queue = END_TSO_QUEUE;
1743 run_queue_hd = END_TSO_QUEUE;
1744 run_queue_tl = END_TSO_QUEUE;
1745 blocked_queue_hd = END_TSO_QUEUE;
1746 blocked_queue_tl = END_TSO_QUEUE;
1747 sleeping_queue = END_TSO_QUEUE;
1750 suspended_ccalling_threads = END_TSO_QUEUE;
1752 main_threads = NULL;
1753 all_threads = END_TSO_QUEUE;
1758 RtsFlags.ConcFlags.ctxtSwitchTicks =
1759 RtsFlags.ConcFlags.ctxtSwitchTime / TICK_MILLISECS;
1761 /* Install the SIGHUP handler */
1764 struct sigaction action,oact;
1766 action.sa_handler = term_handler;
1767 sigemptyset(&action.sa_mask);
1768 action.sa_flags = 0;
1769 if (sigaction(SIGTERM, &action, &oact) != 0) {
1770 barf("can't install TERM handler");
1776 /* Allocate N Capabilities */
1779 Capability *cap, *prev;
1782 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
1783 cap = stgMallocBytes(sizeof(Capability), "initScheduler:capabilities");
1787 free_capabilities = cap;
1788 n_free_capabilities = RtsFlags.ParFlags.nNodes;
1790 IF_DEBUG(scheduler,fprintf(stderr,"scheduler: Allocated %d capabilities\n",
1791 n_free_capabilities););
1794 #if defined(SMP) || defined(PAR)
1807 /* make some space for saving all the thread ids */
1808 task_ids = stgMallocBytes(RtsFlags.ParFlags.nNodes * sizeof(task_info),
1809 "initScheduler:task_ids");
1811 /* and create all the threads */
1812 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
1813 r = pthread_create(&tid,NULL,taskStart,NULL);
1815 barf("startTasks: Can't create new Posix thread");
1817 task_ids[i].id = tid;
1818 task_ids[i].mut_time = 0.0;
1819 task_ids[i].mut_etime = 0.0;
1820 task_ids[i].gc_time = 0.0;
1821 task_ids[i].gc_etime = 0.0;
1822 task_ids[i].elapsedtimestart = elapsedtime();
1823 IF_DEBUG(scheduler,fprintf(stderr,"scheduler: Started task: %ld\n",tid););
1829 exitScheduler( void )
1834 /* Don't want to use pthread_cancel, since we'd have to install
1835 * these silly exception handlers (pthread_cleanup_{push,pop}) around
1839 /* Cancel all our tasks */
1840 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
1841 pthread_cancel(task_ids[i].id);
1844 /* Wait for all the tasks to terminate */
1845 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
1846 IF_DEBUG(scheduler,fprintf(stderr,"scheduler: waiting for task %ld\n",
1848 pthread_join(task_ids[i].id, NULL);
1852 /* Send 'em all a SIGHUP. That should shut 'em up.
1854 await_death = RtsFlags.ParFlags.nNodes;
1855 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
1856 pthread_kill(task_ids[i].id,SIGTERM);
1858 while (await_death > 0) {
1864 /* -----------------------------------------------------------------------------
1865 Managing the per-task allocation areas.
1867 Each capability comes with an allocation area. These are
1868 fixed-length block lists into which allocation can be done.
1870 ToDo: no support for two-space collection at the moment???
1871 -------------------------------------------------------------------------- */
1873 /* -----------------------------------------------------------------------------
1874 * waitThread is the external interface for running a new computation
1875 * and waiting for the result.
1877 * In the non-SMP case, we create a new main thread, push it on the
1878 * main-thread stack, and invoke the scheduler to run it. The
1879 * scheduler will return when the top main thread on the stack has
1880 * completed or died, and fill in the necessary fields of the
1881 * main_thread structure.
1883 * In the SMP case, we create a main thread as before, but we then
1884 * create a new condition variable and sleep on it. When our new
1885 * main thread has completed, we'll be woken up and the status/result
1886 * will be in the main_thread struct.
1887 * -------------------------------------------------------------------------- */
1890 howManyThreadsAvail ( void )
1894 for (q = run_queue_hd; q != END_TSO_QUEUE; q = q->link)
1896 for (q = blocked_queue_hd; q != END_TSO_QUEUE; q = q->link)
1898 for (q = sleeping_queue; q != END_TSO_QUEUE; q = q->link)
1904 finishAllThreads ( void )
1907 while (run_queue_hd != END_TSO_QUEUE) {
1908 waitThread ( run_queue_hd, NULL );
1910 while (blocked_queue_hd != END_TSO_QUEUE) {
1911 waitThread ( blocked_queue_hd, NULL );
1913 while (sleeping_queue != END_TSO_QUEUE) {
1914 waitThread ( blocked_queue_hd, NULL );
1917 (blocked_queue_hd != END_TSO_QUEUE ||
1918 run_queue_hd != END_TSO_QUEUE ||
1919 sleeping_queue != END_TSO_QUEUE);
1923 waitThread(StgTSO *tso, /*out*/StgClosure **ret)
1926 SchedulerStatus stat;
1928 ACQUIRE_LOCK(&sched_mutex);
1930 m = stgMallocBytes(sizeof(StgMainThread), "waitThread");
1936 pthread_cond_init(&m->wakeup, NULL);
1939 m->link = main_threads;
1942 IF_DEBUG(scheduler, fprintf(stderr, "== scheduler: new main thread (%d)\n",
1947 pthread_cond_wait(&m->wakeup, &sched_mutex);
1948 } while (m->stat == NoStatus);
1950 /* GranSim specific init */
1951 CurrentTSO = m->tso; // the TSO to run
1952 procStatus[MainProc] = Busy; // status of main PE
1953 CurrentProc = MainProc; // PE to run it on
1958 ASSERT(m->stat != NoStatus);
1964 pthread_cond_destroy(&m->wakeup);
1967 IF_DEBUG(scheduler, fprintf(stderr, "== scheduler: main thread (%d) finished\n",
1971 RELEASE_LOCK(&sched_mutex);
1976 //@node Run queue code, Garbage Collextion Routines, Suspend and Resume, Main scheduling code
1977 //@subsection Run queue code
1981 NB: In GranSim we have many run queues; run_queue_hd is actually a macro
1982 unfolding to run_queue_hds[CurrentProc], thus CurrentProc is an
1983 implicit global variable that has to be correct when calling these
1987 /* Put the new thread on the head of the runnable queue.
1988 * The caller of createThread better push an appropriate closure
1989 * on this thread's stack before the scheduler is invoked.
1991 static /* inline */ void
1992 add_to_run_queue(tso)
1995 ASSERT(tso!=run_queue_hd && tso!=run_queue_tl);
1996 tso->link = run_queue_hd;
1998 if (run_queue_tl == END_TSO_QUEUE) {
2003 /* Put the new thread at the end of the runnable queue. */
2004 static /* inline */ void
2005 push_on_run_queue(tso)
2008 ASSERT(get_itbl((StgClosure *)tso)->type == TSO);
2009 ASSERT(run_queue_hd!=NULL && run_queue_tl!=NULL);
2010 ASSERT(tso!=run_queue_hd && tso!=run_queue_tl);
2011 if (run_queue_hd == END_TSO_QUEUE) {
2014 run_queue_tl->link = tso;
2020 Should be inlined because it's used very often in schedule. The tso
2021 argument is actually only needed in GranSim, where we want to have the
2022 possibility to schedule *any* TSO on the run queue, irrespective of the
2023 actual ordering. Therefore, if tso is not the nil TSO then we traverse
2024 the run queue and dequeue the tso, adjusting the links in the queue.
2026 //@cindex take_off_run_queue
2027 static /* inline */ StgTSO*
2028 take_off_run_queue(StgTSO *tso) {
2032 qetlaHbogh Qu' ngaSbogh ghomDaQ {tso} yIteq!
2034 if tso is specified, unlink that tso from the run_queue (doesn't have
2035 to be at the beginning of the queue); GranSim only
2037 if (tso!=END_TSO_QUEUE) {
2038 /* find tso in queue */
2039 for (t=run_queue_hd, prev=END_TSO_QUEUE;
2040 t!=END_TSO_QUEUE && t!=tso;
2044 /* now actually dequeue the tso */
2045 if (prev!=END_TSO_QUEUE) {
2046 ASSERT(run_queue_hd!=t);
2047 prev->link = t->link;
2049 /* t is at beginning of thread queue */
2050 ASSERT(run_queue_hd==t);
2051 run_queue_hd = t->link;
2053 /* t is at end of thread queue */
2054 if (t->link==END_TSO_QUEUE) {
2055 ASSERT(t==run_queue_tl);
2056 run_queue_tl = prev;
2058 ASSERT(run_queue_tl!=t);
2060 t->link = END_TSO_QUEUE;
2062 /* take tso from the beginning of the queue; std concurrent code */
2064 if (t != END_TSO_QUEUE) {
2065 run_queue_hd = t->link;
2066 t->link = END_TSO_QUEUE;
2067 if (run_queue_hd == END_TSO_QUEUE) {
2068 run_queue_tl = END_TSO_QUEUE;
2077 //@node Garbage Collextion Routines, Blocking Queue Routines, Run queue code, Main scheduling code
2078 //@subsection Garbage Collextion Routines
2080 /* ---------------------------------------------------------------------------
2081 Where are the roots that we know about?
2083 - all the threads on the runnable queue
2084 - all the threads on the blocked queue
2085 - all the threads on the sleeping queue
2086 - all the thread currently executing a _ccall_GC
2087 - all the "main threads"
2089 ------------------------------------------------------------------------ */
2091 /* This has to be protected either by the scheduler monitor, or by the
2092 garbage collection monitor (probably the latter).
2096 static void GetRoots(void)
2103 for (i=0; i<=RtsFlags.GranFlags.proc; i++) {
2104 if ((run_queue_hds[i] != END_TSO_QUEUE) && ((run_queue_hds[i] != NULL)))
2105 run_queue_hds[i] = (StgTSO *)MarkRoot((StgClosure *)run_queue_hds[i]);
2106 if ((run_queue_tls[i] != END_TSO_QUEUE) && ((run_queue_tls[i] != NULL)))
2107 run_queue_tls[i] = (StgTSO *)MarkRoot((StgClosure *)run_queue_tls[i]);
2109 if ((blocked_queue_hds[i] != END_TSO_QUEUE) && ((blocked_queue_hds[i] != NULL)))
2110 blocked_queue_hds[i] = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_hds[i]);
2111 if ((blocked_queue_tls[i] != END_TSO_QUEUE) && ((blocked_queue_tls[i] != NULL)))
2112 blocked_queue_tls[i] = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_tls[i]);
2113 if ((ccalling_threadss[i] != END_TSO_QUEUE) && ((ccalling_threadss[i] != NULL)))
2114 ccalling_threadss[i] = (StgTSO *)MarkRoot((StgClosure *)ccalling_threadss[i]);
2121 if (run_queue_hd != END_TSO_QUEUE) {
2122 ASSERT(run_queue_tl != END_TSO_QUEUE);
2123 run_queue_hd = (StgTSO *)MarkRoot((StgClosure *)run_queue_hd);
2124 run_queue_tl = (StgTSO *)MarkRoot((StgClosure *)run_queue_tl);
2127 if (blocked_queue_hd != END_TSO_QUEUE) {
2128 ASSERT(blocked_queue_tl != END_TSO_QUEUE);
2129 blocked_queue_hd = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_hd);
2130 blocked_queue_tl = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_tl);
2133 if (sleeping_queue != END_TSO_QUEUE) {
2134 sleeping_queue = (StgTSO *)MarkRoot((StgClosure *)sleeping_queue);
2138 for (m = main_threads; m != NULL; m = m->link) {
2139 m->tso = (StgTSO *)MarkRoot((StgClosure *)m->tso);
2141 if (suspended_ccalling_threads != END_TSO_QUEUE)
2142 suspended_ccalling_threads =
2143 (StgTSO *)MarkRoot((StgClosure *)suspended_ccalling_threads);
2145 #if defined(SMP) || defined(PAR) || defined(GRAN)
2150 /* -----------------------------------------------------------------------------
2153 This is the interface to the garbage collector from Haskell land.
2154 We provide this so that external C code can allocate and garbage
2155 collect when called from Haskell via _ccall_GC.
2157 It might be useful to provide an interface whereby the programmer
2158 can specify more roots (ToDo).
2160 This needs to be protected by the GC condition variable above. KH.
2161 -------------------------------------------------------------------------- */
2163 void (*extra_roots)(void);
2168 GarbageCollect(GetRoots,rtsFalse);
2172 performMajorGC(void)
2174 GarbageCollect(GetRoots,rtsTrue);
2180 GetRoots(); /* the scheduler's roots */
2181 extra_roots(); /* the user's roots */
2185 performGCWithRoots(void (*get_roots)(void))
2187 extra_roots = get_roots;
2189 GarbageCollect(AllRoots,rtsFalse);
2192 /* -----------------------------------------------------------------------------
2195 If the thread has reached its maximum stack size, then raise the
2196 StackOverflow exception in the offending thread. Otherwise
2197 relocate the TSO into a larger chunk of memory and adjust its stack
2199 -------------------------------------------------------------------------- */
2202 threadStackOverflow(StgTSO *tso)
2204 nat new_stack_size, new_tso_size, diff, stack_words;
2208 IF_DEBUG(sanity,checkTSO(tso));
2209 if (tso->stack_size >= tso->max_stack_size) {
2212 belch("@@ threadStackOverflow of TSO %d (%p): stack too large (now %ld; max is %ld",
2213 tso->id, tso, tso->stack_size, tso->max_stack_size);
2214 /* If we're debugging, just print out the top of the stack */
2215 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2218 /* Send this thread the StackOverflow exception */
2219 raiseAsync(tso, (StgClosure *)stackOverflow_closure);
2223 /* Try to double the current stack size. If that takes us over the
2224 * maximum stack size for this thread, then use the maximum instead.
2225 * Finally round up so the TSO ends up as a whole number of blocks.
2227 new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
2228 new_tso_size = (nat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) +
2229 TSO_STRUCT_SIZE)/sizeof(W_);
2230 new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */
2231 new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
2233 IF_DEBUG(scheduler, fprintf(stderr,"== scheduler: increasing stack size from %d words to %d.\n", tso->stack_size, new_stack_size));
2235 dest = (StgTSO *)allocate(new_tso_size);
2236 TICK_ALLOC_TSO(new_tso_size-sizeofW(StgTSO),0);
2238 /* copy the TSO block and the old stack into the new area */
2239 memcpy(dest,tso,TSO_STRUCT_SIZE);
2240 stack_words = tso->stack + tso->stack_size - tso->sp;
2241 new_sp = (P_)dest + new_tso_size - stack_words;
2242 memcpy(new_sp, tso->sp, stack_words * sizeof(W_));
2244 /* relocate the stack pointers... */
2245 diff = (P_)new_sp - (P_)tso->sp; /* In *words* */
2246 dest->su = (StgUpdateFrame *) ((P_)dest->su + diff);
2248 //dest->splim = (P_)dest->splim + (nat)((P_)dest - (P_)tso);
2249 dest->stack_size = new_stack_size;
2251 /* and relocate the update frame list */
2252 relocate_TSO(tso, dest);
2254 /* Mark the old TSO as relocated. We have to check for relocated
2255 * TSOs in the garbage collector and any primops that deal with TSOs.
2257 * It's important to set the sp and su values to just beyond the end
2258 * of the stack, so we don't attempt to scavenge any part of the
2261 tso->what_next = ThreadRelocated;
2263 tso->sp = (P_)&(tso->stack[tso->stack_size]);
2264 tso->su = (StgUpdateFrame *)tso->sp;
2265 tso->why_blocked = NotBlocked;
2266 dest->mut_link = NULL;
2268 IF_PAR_DEBUG(verbose,
2269 belch("@@ threadStackOverflow of TSO %d (now at %p): stack size increased to %ld",
2270 tso->id, tso, tso->stack_size);
2271 /* If we're debugging, just print out the top of the stack */
2272 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2275 IF_DEBUG(sanity,checkTSO(tso));
2277 IF_DEBUG(scheduler,printTSO(dest));
2283 //@node Blocking Queue Routines, Exception Handling Routines, Garbage Collextion Routines, Main scheduling code
2284 //@subsection Blocking Queue Routines
2286 /* ---------------------------------------------------------------------------
2287 Wake up a queue that was blocked on some resource.
2288 ------------------------------------------------------------------------ */
2292 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2297 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2299 /* write RESUME events to log file and
2300 update blocked and fetch time (depending on type of the orig closure) */
2301 if (RtsFlags.ParFlags.ParStats.Full) {
2302 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
2303 GR_RESUMEQ, ((StgTSO *)bqe), ((StgTSO *)bqe)->block_info.closure,
2304 0, 0 /* spark_queue_len(ADVISORY_POOL) */);
2305 if (EMPTY_RUN_QUEUE())
2306 emitSchedule = rtsTrue;
2308 switch (get_itbl(node)->type) {
2310 ((StgTSO *)bqe)->par.fetchtime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2315 ((StgTSO *)bqe)->par.blocktime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2322 barf("{unblockOneLocked}Daq Qagh: unexpected closure in blocking queue");
2329 static StgBlockingQueueElement *
2330 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2333 PEs node_loc, tso_loc;
2335 node_loc = where_is(node); // should be lifted out of loop
2336 tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2337 tso_loc = where_is((StgClosure *)tso);
2338 if (IS_LOCAL_TO(PROCS(node),tso_loc)) { // TSO is local
2339 /* !fake_fetch => TSO is on CurrentProc is same as IS_LOCAL_TO */
2340 ASSERT(CurrentProc!=node_loc || tso_loc==CurrentProc);
2341 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.lunblocktime;
2342 // insertThread(tso, node_loc);
2343 new_event(tso_loc, tso_loc, CurrentTime[CurrentProc],
2345 tso, node, (rtsSpark*)NULL);
2346 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2349 } else { // TSO is remote (actually should be FMBQ)
2350 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.mpacktime +
2351 RtsFlags.GranFlags.Costs.gunblocktime +
2352 RtsFlags.GranFlags.Costs.latency;
2353 new_event(tso_loc, CurrentProc, CurrentTime[CurrentProc],
2355 tso, node, (rtsSpark*)NULL);
2356 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2359 /* the thread-queue-overhead is accounted for in either Resume or UnblockThread */
2361 fprintf(stderr," %s TSO %d (%p) [PE %d] (block_info.closure=%p) (next=%p) ,",
2362 (node_loc==tso_loc ? "Local" : "Global"),
2363 tso->id, tso, CurrentProc, tso->block_info.closure, tso->link));
2364 tso->block_info.closure = NULL;
2365 IF_DEBUG(scheduler,belch("-- Waking up thread %ld (%p)",
2369 static StgBlockingQueueElement *
2370 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2372 StgBlockingQueueElement *next;
2374 switch (get_itbl(bqe)->type) {
2376 ASSERT(((StgTSO *)bqe)->why_blocked != NotBlocked);
2377 /* if it's a TSO just push it onto the run_queue */
2379 // ((StgTSO *)bqe)->link = END_TSO_QUEUE; // debugging?
2380 PUSH_ON_RUN_QUEUE((StgTSO *)bqe);
2382 unblockCount(bqe, node);
2383 /* reset blocking status after dumping event */
2384 ((StgTSO *)bqe)->why_blocked = NotBlocked;
2388 /* if it's a BLOCKED_FETCH put it on the PendingFetches list */
2390 bqe->link = (StgBlockingQueueElement *)PendingFetches;
2391 PendingFetches = (StgBlockedFetch *)bqe;
2395 /* can ignore this case in a non-debugging setup;
2396 see comments on RBHSave closures above */
2398 /* check that the closure is an RBHSave closure */
2399 ASSERT(get_itbl((StgClosure *)bqe) == &stg_RBH_Save_0_info ||
2400 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_1_info ||
2401 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_2_info);
2405 barf("{unblockOneLocked}Daq Qagh: Unexpected IP (%#lx; %s) in blocking queue at %#lx\n",
2406 get_itbl((StgClosure *)bqe), info_type((StgClosure *)bqe),
2410 IF_PAR_DEBUG(bq, fprintf(stderr, ", %p (%s)", bqe, info_type((StgClosure*)bqe)));
2414 #else /* !GRAN && !PAR */
2416 unblockOneLocked(StgTSO *tso)
2420 ASSERT(get_itbl(tso)->type == TSO);
2421 ASSERT(tso->why_blocked != NotBlocked);
2422 tso->why_blocked = NotBlocked;
2424 PUSH_ON_RUN_QUEUE(tso);
2426 IF_DEBUG(scheduler,sched_belch("waking up thread %ld", tso->id));
2431 #if defined(GRAN) || defined(PAR)
2432 inline StgBlockingQueueElement *
2433 unblockOne(StgBlockingQueueElement *bqe, StgClosure *node)
2435 ACQUIRE_LOCK(&sched_mutex);
2436 bqe = unblockOneLocked(bqe, node);
2437 RELEASE_LOCK(&sched_mutex);
2442 unblockOne(StgTSO *tso)
2444 ACQUIRE_LOCK(&sched_mutex);
2445 tso = unblockOneLocked(tso);
2446 RELEASE_LOCK(&sched_mutex);
2453 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2455 StgBlockingQueueElement *bqe;
2460 belch("##-_ AwBQ for node %p on PE %d @ %ld by TSO %d (%p): ", \
2461 node, CurrentProc, CurrentTime[CurrentProc],
2462 CurrentTSO->id, CurrentTSO));
2464 node_loc = where_is(node);
2466 ASSERT(q == END_BQ_QUEUE ||
2467 get_itbl(q)->type == TSO || // q is either a TSO or an RBHSave
2468 get_itbl(q)->type == CONSTR); // closure (type constructor)
2469 ASSERT(is_unique(node));
2471 /* FAKE FETCH: magically copy the node to the tso's proc;
2472 no Fetch necessary because in reality the node should not have been
2473 moved to the other PE in the first place
2475 if (CurrentProc!=node_loc) {
2477 belch("## node %p is on PE %d but CurrentProc is %d (TSO %d); assuming fake fetch and adjusting bitmask (old: %#x)",
2478 node, node_loc, CurrentProc, CurrentTSO->id,
2479 // CurrentTSO, where_is(CurrentTSO),
2480 node->header.gran.procs));
2481 node->header.gran.procs = (node->header.gran.procs) | PE_NUMBER(CurrentProc);
2483 belch("## new bitmask of node %p is %#x",
2484 node, node->header.gran.procs));
2485 if (RtsFlags.GranFlags.GranSimStats.Global) {
2486 globalGranStats.tot_fake_fetches++;
2491 // ToDo: check: ASSERT(CurrentProc==node_loc);
2492 while (get_itbl(bqe)->type==TSO) { // q != END_TSO_QUEUE) {
2495 bqe points to the current element in the queue
2496 next points to the next element in the queue
2498 //tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2499 //tso_loc = where_is(tso);
2501 bqe = unblockOneLocked(bqe, node);
2504 /* if this is the BQ of an RBH, we have to put back the info ripped out of
2505 the closure to make room for the anchor of the BQ */
2506 if (bqe!=END_BQ_QUEUE) {
2507 ASSERT(get_itbl(node)->type == RBH && get_itbl(bqe)->type == CONSTR);
2509 ASSERT((info_ptr==&RBH_Save_0_info) ||
2510 (info_ptr==&RBH_Save_1_info) ||
2511 (info_ptr==&RBH_Save_2_info));
2513 /* cf. convertToRBH in RBH.c for writing the RBHSave closure */
2514 ((StgRBH *)node)->blocking_queue = (StgBlockingQueueElement *)((StgRBHSave *)bqe)->payload[0];
2515 ((StgRBH *)node)->mut_link = (StgMutClosure *)((StgRBHSave *)bqe)->payload[1];
2518 belch("## Filled in RBH_Save for %p (%s) at end of AwBQ",
2519 node, info_type(node)));
2522 /* statistics gathering */
2523 if (RtsFlags.GranFlags.GranSimStats.Global) {
2524 // globalGranStats.tot_bq_processing_time += bq_processing_time;
2525 globalGranStats.tot_bq_len += len; // total length of all bqs awakened
2526 // globalGranStats.tot_bq_len_local += len_local; // same for local TSOs only
2527 globalGranStats.tot_awbq++; // total no. of bqs awakened
2530 fprintf(stderr,"## BQ Stats of %p: [%d entries] %s\n",
2531 node, len, (bqe!=END_BQ_QUEUE) ? "RBH" : ""));
2535 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2537 StgBlockingQueueElement *bqe;
2539 ACQUIRE_LOCK(&sched_mutex);
2541 IF_PAR_DEBUG(verbose,
2542 belch("##-_ AwBQ for node %p on [%x]: ",
2546 if(get_itbl(q)->type == CONSTR || q==END_BQ_QUEUE) {
2547 IF_PAR_DEBUG(verbose, belch("## ... nothing to unblock so lets just return. RFP (BUG?)"));
2552 ASSERT(q == END_BQ_QUEUE ||
2553 get_itbl(q)->type == TSO ||
2554 get_itbl(q)->type == BLOCKED_FETCH ||
2555 get_itbl(q)->type == CONSTR);
2558 while (get_itbl(bqe)->type==TSO ||
2559 get_itbl(bqe)->type==BLOCKED_FETCH) {
2560 bqe = unblockOneLocked(bqe, node);
2562 RELEASE_LOCK(&sched_mutex);
2565 #else /* !GRAN && !PAR */
2567 awakenBlockedQueue(StgTSO *tso)
2569 ACQUIRE_LOCK(&sched_mutex);
2570 while (tso != END_TSO_QUEUE) {
2571 tso = unblockOneLocked(tso);
2573 RELEASE_LOCK(&sched_mutex);
2577 //@node Exception Handling Routines, Debugging Routines, Blocking Queue Routines, Main scheduling code
2578 //@subsection Exception Handling Routines
2580 /* ---------------------------------------------------------------------------
2582 - usually called inside a signal handler so it mustn't do anything fancy.
2583 ------------------------------------------------------------------------ */
2586 interruptStgRts(void)
2592 /* -----------------------------------------------------------------------------
2595 This is for use when we raise an exception in another thread, which
2597 This has nothing to do with the UnblockThread event in GranSim. -- HWL
2598 -------------------------------------------------------------------------- */
2600 #if defined(GRAN) || defined(PAR)
2602 NB: only the type of the blocking queue is different in GranSim and GUM
2603 the operations on the queue-elements are the same
2604 long live polymorphism!
2607 unblockThread(StgTSO *tso)
2609 StgBlockingQueueElement *t, **last;
2611 ACQUIRE_LOCK(&sched_mutex);
2612 switch (tso->why_blocked) {
2615 return; /* not blocked */
2618 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
2620 StgBlockingQueueElement *last_tso = END_BQ_QUEUE;
2621 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
2623 last = (StgBlockingQueueElement **)&mvar->head;
2624 for (t = (StgBlockingQueueElement *)mvar->head;
2626 last = &t->link, last_tso = t, t = t->link) {
2627 if (t == (StgBlockingQueueElement *)tso) {
2628 *last = (StgBlockingQueueElement *)tso->link;
2629 if (mvar->tail == tso) {
2630 mvar->tail = (StgTSO *)last_tso;
2635 barf("unblockThread (MVAR): TSO not found");
2638 case BlockedOnBlackHole:
2639 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
2641 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
2643 last = &bq->blocking_queue;
2644 for (t = bq->blocking_queue;
2646 last = &t->link, t = t->link) {
2647 if (t == (StgBlockingQueueElement *)tso) {
2648 *last = (StgBlockingQueueElement *)tso->link;
2652 barf("unblockThread (BLACKHOLE): TSO not found");
2655 case BlockedOnException:
2657 StgTSO *target = tso->block_info.tso;
2659 ASSERT(get_itbl(target)->type == TSO);
2661 if (target->what_next == ThreadRelocated) {
2662 target = target->link;
2663 ASSERT(get_itbl(target)->type == TSO);
2666 ASSERT(target->blocked_exceptions != NULL);
2668 last = (StgBlockingQueueElement **)&target->blocked_exceptions;
2669 for (t = (StgBlockingQueueElement *)target->blocked_exceptions;
2671 last = &t->link, t = t->link) {
2672 ASSERT(get_itbl(t)->type == TSO);
2673 if (t == (StgBlockingQueueElement *)tso) {
2674 *last = (StgBlockingQueueElement *)tso->link;
2678 barf("unblockThread (Exception): TSO not found");
2682 case BlockedOnWrite:
2684 /* take TSO off blocked_queue */
2685 StgBlockingQueueElement *prev = NULL;
2686 for (t = (StgBlockingQueueElement *)blocked_queue_hd; t != END_BQ_QUEUE;
2687 prev = t, t = t->link) {
2688 if (t == (StgBlockingQueueElement *)tso) {
2690 blocked_queue_hd = (StgTSO *)t->link;
2691 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
2692 blocked_queue_tl = END_TSO_QUEUE;
2695 prev->link = t->link;
2696 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
2697 blocked_queue_tl = (StgTSO *)prev;
2703 barf("unblockThread (I/O): TSO not found");
2706 case BlockedOnDelay:
2708 /* take TSO off sleeping_queue */
2709 StgBlockingQueueElement *prev = NULL;
2710 for (t = (StgBlockingQueueElement *)sleeping_queue; t != END_BQ_QUEUE;
2711 prev = t, t = t->link) {
2712 if (t == (StgBlockingQueueElement *)tso) {
2714 sleeping_queue = (StgTSO *)t->link;
2716 prev->link = t->link;
2721 barf("unblockThread (I/O): TSO not found");
2725 barf("unblockThread");
2729 tso->link = END_TSO_QUEUE;
2730 tso->why_blocked = NotBlocked;
2731 tso->block_info.closure = NULL;
2732 PUSH_ON_RUN_QUEUE(tso);
2733 RELEASE_LOCK(&sched_mutex);
2737 unblockThread(StgTSO *tso)
2741 ACQUIRE_LOCK(&sched_mutex);
2742 switch (tso->why_blocked) {
2745 return; /* not blocked */
2748 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
2750 StgTSO *last_tso = END_TSO_QUEUE;
2751 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
2754 for (t = mvar->head; t != END_TSO_QUEUE;
2755 last = &t->link, last_tso = t, t = t->link) {
2758 if (mvar->tail == tso) {
2759 mvar->tail = last_tso;
2764 barf("unblockThread (MVAR): TSO not found");
2767 case BlockedOnBlackHole:
2768 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
2770 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
2772 last = &bq->blocking_queue;
2773 for (t = bq->blocking_queue; t != END_TSO_QUEUE;
2774 last = &t->link, t = t->link) {
2780 barf("unblockThread (BLACKHOLE): TSO not found");
2783 case BlockedOnException:
2785 StgTSO *target = tso->block_info.tso;
2787 ASSERT(get_itbl(target)->type == TSO);
2789 while (target->what_next == ThreadRelocated) {
2790 target = target->link;
2791 ASSERT(get_itbl(target)->type == TSO);
2794 ASSERT(target->blocked_exceptions != NULL);
2796 last = &target->blocked_exceptions;
2797 for (t = target->blocked_exceptions; t != END_TSO_QUEUE;
2798 last = &t->link, t = t->link) {
2799 ASSERT(get_itbl(t)->type == TSO);
2805 barf("unblockThread (Exception): TSO not found");
2809 case BlockedOnWrite:
2811 StgTSO *prev = NULL;
2812 for (t = blocked_queue_hd; t != END_TSO_QUEUE;
2813 prev = t, t = t->link) {
2816 blocked_queue_hd = t->link;
2817 if (blocked_queue_tl == t) {
2818 blocked_queue_tl = END_TSO_QUEUE;
2821 prev->link = t->link;
2822 if (blocked_queue_tl == t) {
2823 blocked_queue_tl = prev;
2829 barf("unblockThread (I/O): TSO not found");
2832 case BlockedOnDelay:
2834 StgTSO *prev = NULL;
2835 for (t = sleeping_queue; t != END_TSO_QUEUE;
2836 prev = t, t = t->link) {
2839 sleeping_queue = t->link;
2841 prev->link = t->link;
2846 barf("unblockThread (I/O): TSO not found");
2850 barf("unblockThread");
2854 tso->link = END_TSO_QUEUE;
2855 tso->why_blocked = NotBlocked;
2856 tso->block_info.closure = NULL;
2857 PUSH_ON_RUN_QUEUE(tso);
2858 RELEASE_LOCK(&sched_mutex);
2862 /* -----------------------------------------------------------------------------
2865 * The following function implements the magic for raising an
2866 * asynchronous exception in an existing thread.
2868 * We first remove the thread from any queue on which it might be
2869 * blocked. The possible blockages are MVARs and BLACKHOLE_BQs.
2871 * We strip the stack down to the innermost CATCH_FRAME, building
2872 * thunks in the heap for all the active computations, so they can
2873 * be restarted if necessary. When we reach a CATCH_FRAME, we build
2874 * an application of the handler to the exception, and push it on
2875 * the top of the stack.
2877 * How exactly do we save all the active computations? We create an
2878 * AP_UPD for every UpdateFrame on the stack. Entering one of these
2879 * AP_UPDs pushes everything from the corresponding update frame
2880 * upwards onto the stack. (Actually, it pushes everything up to the
2881 * next update frame plus a pointer to the next AP_UPD object.
2882 * Entering the next AP_UPD object pushes more onto the stack until we
2883 * reach the last AP_UPD object - at which point the stack should look
2884 * exactly as it did when we killed the TSO and we can continue
2885 * execution by entering the closure on top of the stack.
2887 * We can also kill a thread entirely - this happens if either (a) the
2888 * exception passed to raiseAsync is NULL, or (b) there's no
2889 * CATCH_FRAME on the stack. In either case, we strip the entire
2890 * stack and replace the thread with a zombie.
2892 * -------------------------------------------------------------------------- */
2895 deleteThread(StgTSO *tso)
2897 raiseAsync(tso,NULL);
2901 raiseAsync(StgTSO *tso, StgClosure *exception)
2903 StgUpdateFrame* su = tso->su;
2904 StgPtr sp = tso->sp;
2906 /* Thread already dead? */
2907 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
2911 IF_DEBUG(scheduler, sched_belch("raising exception in thread %ld.", tso->id));
2913 /* Remove it from any blocking queues */
2916 /* The stack freezing code assumes there's a closure pointer on
2917 * the top of the stack. This isn't always the case with compiled
2918 * code, so we have to push a dummy closure on the top which just
2919 * returns to the next return address on the stack.
2921 if ( LOOKS_LIKE_GHC_INFO((void*)*sp) ) {
2922 *(--sp) = (W_)&stg_dummy_ret_closure;
2926 int words = ((P_)su - (P_)sp) - 1;
2930 /* If we find a CATCH_FRAME, and we've got an exception to raise,
2931 * then build PAP(handler,exception,realworld#), and leave it on
2932 * top of the stack ready to enter.
2934 if (get_itbl(su)->type == CATCH_FRAME && exception != NULL) {
2935 StgCatchFrame *cf = (StgCatchFrame *)su;
2936 /* we've got an exception to raise, so let's pass it to the
2937 * handler in this frame.
2939 ap = (StgAP_UPD *)allocate(sizeofW(StgPAP) + 2);
2940 TICK_ALLOC_UPD_PAP(3,0);
2941 SET_HDR(ap,&stg_PAP_info,cf->header.prof.ccs);
2944 ap->fun = cf->handler; /* :: Exception -> IO a */
2945 ap->payload[0] = exception;
2946 ap->payload[1] = ARG_TAG(0); /* realworld token */
2948 /* throw away the stack from Sp up to and including the
2951 sp = (P_)su + sizeofW(StgCatchFrame) - 1;
2954 /* Restore the blocked/unblocked state for asynchronous exceptions
2955 * at the CATCH_FRAME.
2957 * If exceptions were unblocked at the catch, arrange that they
2958 * are unblocked again after executing the handler by pushing an
2959 * unblockAsyncExceptions_ret stack frame.
2961 if (!cf->exceptions_blocked) {
2962 *(sp--) = (W_)&stg_unblockAsyncExceptionszh_ret_info;
2965 /* Ensure that async exceptions are blocked when running the handler.
2967 if (tso->blocked_exceptions == NULL) {
2968 tso->blocked_exceptions = END_TSO_QUEUE;
2971 /* Put the newly-built PAP on top of the stack, ready to execute
2972 * when the thread restarts.
2976 tso->what_next = ThreadEnterGHC;
2977 IF_DEBUG(sanity, checkTSO(tso));
2981 /* First build an AP_UPD consisting of the stack chunk above the
2982 * current update frame, with the top word on the stack as the
2985 ap = (StgAP_UPD *)allocate(AP_sizeW(words));
2990 ap->fun = (StgClosure *)sp[0];
2992 for(i=0; i < (nat)words; ++i) {
2993 ap->payload[i] = (StgClosure *)*sp++;
2996 switch (get_itbl(su)->type) {
3000 SET_HDR(ap,&stg_AP_UPD_info,su->header.prof.ccs /* ToDo */);
3001 TICK_ALLOC_UP_THK(words+1,0);
3004 fprintf(stderr, "scheduler: Updating ");
3005 printPtr((P_)su->updatee);
3006 fprintf(stderr, " with ");
3007 printObj((StgClosure *)ap);
3010 /* Replace the updatee with an indirection - happily
3011 * this will also wake up any threads currently
3012 * waiting on the result.
3014 * Warning: if we're in a loop, more than one update frame on
3015 * the stack may point to the same object. Be careful not to
3016 * overwrite an IND_OLDGEN in this case, because we'll screw
3017 * up the mutable lists. To be on the safe side, don't
3018 * overwrite any kind of indirection at all. See also
3019 * threadSqueezeStack in GC.c, where we have to make a similar
3022 if (!closure_IND(su->updatee)) {
3023 UPD_IND_NOLOCK(su->updatee,ap); /* revert the black hole */
3026 sp += sizeofW(StgUpdateFrame) -1;
3027 sp[0] = (W_)ap; /* push onto stack */
3033 StgCatchFrame *cf = (StgCatchFrame *)su;
3036 /* We want a PAP, not an AP_UPD. Fortunately, the
3037 * layout's the same.
3039 SET_HDR(ap,&stg_PAP_info,su->header.prof.ccs /* ToDo */);
3040 TICK_ALLOC_UPD_PAP(words+1,0);
3042 /* now build o = FUN(catch,ap,handler) */
3043 o = (StgClosure *)allocate(sizeofW(StgClosure)+2);
3044 TICK_ALLOC_FUN(2,0);
3045 SET_HDR(o,&stg_catch_info,su->header.prof.ccs /* ToDo */);
3046 o->payload[0] = (StgClosure *)ap;
3047 o->payload[1] = cf->handler;
3050 fprintf(stderr, "scheduler: Built ");
3051 printObj((StgClosure *)o);
3054 /* pop the old handler and put o on the stack */
3056 sp += sizeofW(StgCatchFrame) - 1;
3063 StgSeqFrame *sf = (StgSeqFrame *)su;
3066 SET_HDR(ap,&stg_PAP_info,su->header.prof.ccs /* ToDo */);
3067 TICK_ALLOC_UPD_PAP(words+1,0);
3069 /* now build o = FUN(seq,ap) */
3070 o = (StgClosure *)allocate(sizeofW(StgClosure)+1);
3071 TICK_ALLOC_SE_THK(1,0);
3072 SET_HDR(o,&stg_seq_info,su->header.prof.ccs /* ToDo */);
3073 o->payload[0] = (StgClosure *)ap;
3076 fprintf(stderr, "scheduler: Built ");
3077 printObj((StgClosure *)o);
3080 /* pop the old handler and put o on the stack */
3082 sp += sizeofW(StgSeqFrame) - 1;
3088 /* We've stripped the entire stack, the thread is now dead. */
3089 sp += sizeofW(StgStopFrame) - 1;
3090 sp[0] = (W_)exception; /* save the exception */
3091 tso->what_next = ThreadKilled;
3092 tso->su = (StgUpdateFrame *)(sp+1);
3103 /* -----------------------------------------------------------------------------
3104 resurrectThreads is called after garbage collection on the list of
3105 threads found to be garbage. Each of these threads will be woken
3106 up and sent a signal: BlockedOnDeadMVar if the thread was blocked
3107 on an MVar, or NonTermination if the thread was blocked on a Black
3109 -------------------------------------------------------------------------- */
3112 resurrectThreads( StgTSO *threads )
3116 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
3117 next = tso->global_link;
3118 tso->global_link = all_threads;
3120 IF_DEBUG(scheduler, sched_belch("resurrecting thread %d", tso->id));
3122 switch (tso->why_blocked) {
3124 case BlockedOnException:
3125 raiseAsync(tso,(StgClosure *)BlockedOnDeadMVar_closure);
3127 case BlockedOnBlackHole:
3128 raiseAsync(tso,(StgClosure *)NonTermination_closure);
3131 /* This might happen if the thread was blocked on a black hole
3132 * belonging to a thread that we've just woken up (raiseAsync
3133 * can wake up threads, remember...).
3137 barf("resurrectThreads: thread blocked in a strange way");
3142 /* -----------------------------------------------------------------------------
3143 * Blackhole detection: if we reach a deadlock, test whether any
3144 * threads are blocked on themselves. Any threads which are found to
3145 * be self-blocked get sent a NonTermination exception.
3147 * This is only done in a deadlock situation in order to avoid
3148 * performance overhead in the normal case.
3149 * -------------------------------------------------------------------------- */
3152 detectBlackHoles( void )
3154 StgTSO *t = all_threads;
3155 StgUpdateFrame *frame;
3156 StgClosure *blocked_on;
3158 for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
3160 while (t->what_next == ThreadRelocated) {
3162 ASSERT(get_itbl(t)->type == TSO);
3165 if (t->why_blocked != BlockedOnBlackHole) {
3169 blocked_on = t->block_info.closure;
3171 for (frame = t->su; ; frame = frame->link) {
3172 switch (get_itbl(frame)->type) {
3175 if (frame->updatee == blocked_on) {
3176 /* We are blocking on one of our own computations, so
3177 * send this thread the NonTermination exception.
3180 sched_belch("thread %d is blocked on itself", t->id));
3181 raiseAsync(t, (StgClosure *)NonTermination_closure);
3202 //@node Debugging Routines, Index, Exception Handling Routines, Main scheduling code
3203 //@subsection Debugging Routines
3205 /* -----------------------------------------------------------------------------
3206 Debugging: why is a thread blocked
3207 -------------------------------------------------------------------------- */
3212 printThreadBlockage(StgTSO *tso)
3214 switch (tso->why_blocked) {
3216 fprintf(stderr,"is blocked on read from fd %d", tso->block_info.fd);
3218 case BlockedOnWrite:
3219 fprintf(stderr,"is blocked on write to fd %d", tso->block_info.fd);
3221 case BlockedOnDelay:
3222 fprintf(stderr,"is blocked until %d", tso->block_info.target);
3225 fprintf(stderr,"is blocked on an MVar");
3227 case BlockedOnException:
3228 fprintf(stderr,"is blocked on delivering an exception to thread %d",
3229 tso->block_info.tso->id);
3231 case BlockedOnBlackHole:
3232 fprintf(stderr,"is blocked on a black hole");
3235 fprintf(stderr,"is not blocked");
3239 fprintf(stderr,"is blocked on global address; local FM_BQ is %p (%s)",
3240 tso->block_info.closure, info_type(tso->block_info.closure));
3242 case BlockedOnGA_NoSend:
3243 fprintf(stderr,"is blocked on global address (no send); local FM_BQ is %p (%s)",
3244 tso->block_info.closure, info_type(tso->block_info.closure));
3248 barf("printThreadBlockage: strange tso->why_blocked: %d for TSO %d (%d)",
3249 tso->why_blocked, tso->id, tso);
3254 printThreadStatus(StgTSO *tso)
3256 switch (tso->what_next) {
3258 fprintf(stderr,"has been killed");
3260 case ThreadComplete:
3261 fprintf(stderr,"has completed");
3264 printThreadBlockage(tso);
3269 printAllThreads(void)
3274 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
3275 ullong_format_string(TIME_ON_PROC(CurrentProc),
3276 time_string, rtsFalse/*no commas!*/);
3278 sched_belch("all threads at [%s]:", time_string);
3280 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
3281 ullong_format_string(CURRENT_TIME,
3282 time_string, rtsFalse/*no commas!*/);
3284 sched_belch("all threads at [%s]:", time_string);
3286 sched_belch("all threads:");
3289 for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
3290 fprintf(stderr, "\tthread %d ", t->id);
3291 printThreadStatus(t);
3292 fprintf(stderr,"\n");
3297 Print a whole blocking queue attached to node (debugging only).
3302 print_bq (StgClosure *node)
3304 StgBlockingQueueElement *bqe;
3308 fprintf(stderr,"## BQ of closure %p (%s): ",
3309 node, info_type(node));
3311 /* should cover all closures that may have a blocking queue */
3312 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
3313 get_itbl(node)->type == FETCH_ME_BQ ||
3314 get_itbl(node)->type == RBH ||
3315 get_itbl(node)->type == MVAR);
3317 ASSERT(node!=(StgClosure*)NULL); // sanity check
3319 print_bqe(((StgBlockingQueue*)node)->blocking_queue);
3323 Print a whole blocking queue starting with the element bqe.
3326 print_bqe (StgBlockingQueueElement *bqe)
3331 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
3333 for (end = (bqe==END_BQ_QUEUE);
3334 !end; // iterate until bqe points to a CONSTR
3335 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE),
3336 bqe = end ? END_BQ_QUEUE : bqe->link) {
3337 ASSERT(bqe != END_BQ_QUEUE); // sanity check
3338 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
3339 /* types of closures that may appear in a blocking queue */
3340 ASSERT(get_itbl(bqe)->type == TSO ||
3341 get_itbl(bqe)->type == BLOCKED_FETCH ||
3342 get_itbl(bqe)->type == CONSTR);
3343 /* only BQs of an RBH end with an RBH_Save closure */
3344 //ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
3346 switch (get_itbl(bqe)->type) {
3348 fprintf(stderr," TSO %u (%x),",
3349 ((StgTSO *)bqe)->id, ((StgTSO *)bqe));
3352 fprintf(stderr," BF (node=%p, ga=((%x, %d, %x)),",
3353 ((StgBlockedFetch *)bqe)->node,
3354 ((StgBlockedFetch *)bqe)->ga.payload.gc.gtid,
3355 ((StgBlockedFetch *)bqe)->ga.payload.gc.slot,
3356 ((StgBlockedFetch *)bqe)->ga.weight);
3359 fprintf(stderr," %s (IP %p),",
3360 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
3361 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
3362 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
3363 "RBH_Save_?"), get_itbl(bqe));
3366 barf("Unexpected closure type %s in blocking queue", // of %p (%s)",
3367 info_type((StgClosure *)bqe)); // , node, info_type(node));
3371 fputc('\n', stderr);
3373 # elif defined(GRAN)
3375 print_bq (StgClosure *node)
3377 StgBlockingQueueElement *bqe;
3378 PEs node_loc, tso_loc;
3381 /* should cover all closures that may have a blocking queue */
3382 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
3383 get_itbl(node)->type == FETCH_ME_BQ ||
3384 get_itbl(node)->type == RBH);
3386 ASSERT(node!=(StgClosure*)NULL); // sanity check
3387 node_loc = where_is(node);
3389 fprintf(stderr,"## BQ of closure %p (%s) on [PE %d]: ",
3390 node, info_type(node), node_loc);
3393 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
3395 for (bqe = ((StgBlockingQueue*)node)->blocking_queue, end = (bqe==END_BQ_QUEUE);
3396 !end; // iterate until bqe points to a CONSTR
3397 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) {
3398 ASSERT(bqe != END_BQ_QUEUE); // sanity check
3399 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
3400 /* types of closures that may appear in a blocking queue */
3401 ASSERT(get_itbl(bqe)->type == TSO ||
3402 get_itbl(bqe)->type == CONSTR);
3403 /* only BQs of an RBH end with an RBH_Save closure */
3404 ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
3406 tso_loc = where_is((StgClosure *)bqe);
3407 switch (get_itbl(bqe)->type) {
3409 fprintf(stderr," TSO %d (%p) on [PE %d],",
3410 ((StgTSO *)bqe)->id, (StgTSO *)bqe, tso_loc);
3413 fprintf(stderr," %s (IP %p),",
3414 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
3415 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
3416 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
3417 "RBH_Save_?"), get_itbl(bqe));
3420 barf("Unexpected closure type %s in blocking queue of %p (%s)",
3421 info_type((StgClosure *)bqe), node, info_type(node));
3425 fputc('\n', stderr);
3429 Nice and easy: only TSOs on the blocking queue
3432 print_bq (StgClosure *node)
3436 ASSERT(node!=(StgClosure*)NULL); // sanity check
3437 for (tso = ((StgBlockingQueue*)node)->blocking_queue;
3438 tso != END_TSO_QUEUE;
3440 ASSERT(tso!=NULL && tso!=END_TSO_QUEUE); // sanity check
3441 ASSERT(get_itbl(tso)->type == TSO); // guess what, sanity check
3442 fprintf(stderr," TSO %d (%p),", tso->id, tso);
3444 fputc('\n', stderr);
3455 for (i=0, tso=run_queue_hd;
3456 tso != END_TSO_QUEUE;
3465 sched_belch(char *s, ...)
3470 fprintf(stderr, "scheduler (task %ld): ", pthread_self());
3472 fprintf(stderr, "== ");
3474 fprintf(stderr, "scheduler: ");
3476 vfprintf(stderr, s, ap);
3477 fprintf(stderr, "\n");
3483 //@node Index, , Debugging Routines, Main scheduling code
3487 //* MainRegTable:: @cindex\s-+MainRegTable
3488 //* StgMainThread:: @cindex\s-+StgMainThread
3489 //* awaken_blocked_queue:: @cindex\s-+awaken_blocked_queue
3490 //* blocked_queue_hd:: @cindex\s-+blocked_queue_hd
3491 //* blocked_queue_tl:: @cindex\s-+blocked_queue_tl
3492 //* context_switch:: @cindex\s-+context_switch
3493 //* createThread:: @cindex\s-+createThread
3494 //* free_capabilities:: @cindex\s-+free_capabilities
3495 //* gc_pending_cond:: @cindex\s-+gc_pending_cond
3496 //* initScheduler:: @cindex\s-+initScheduler
3497 //* interrupted:: @cindex\s-+interrupted
3498 //* n_free_capabilities:: @cindex\s-+n_free_capabilities
3499 //* next_thread_id:: @cindex\s-+next_thread_id
3500 //* print_bq:: @cindex\s-+print_bq
3501 //* run_queue_hd:: @cindex\s-+run_queue_hd
3502 //* run_queue_tl:: @cindex\s-+run_queue_tl
3503 //* sched_mutex:: @cindex\s-+sched_mutex
3504 //* schedule:: @cindex\s-+schedule
3505 //* take_off_run_queue:: @cindex\s-+take_off_run_queue
3506 //* task_ids:: @cindex\s-+task_ids
3507 //* term_mutex:: @cindex\s-+term_mutex
3508 //* thread_ready_cond:: @cindex\s-+thread_ready_cond