1 /* ---------------------------------------------------------------------------
2 * $Id: Schedule.c,v 1.111 2002/01/22 13:54:22 simonmar 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
77 #include "PosixSource.h"
84 #include "StgStartup.h"
87 #include "StgMiscClosures.h"
89 #include "Interpreter.h"
90 #include "Exception.h"
99 #include "Proftimer.h"
100 #include "ProfHeap.h"
102 #if defined(GRAN) || defined(PAR)
103 # include "GranSimRts.h"
104 # include "GranSim.h"
105 # include "ParallelRts.h"
106 # include "Parallel.h"
107 # include "ParallelDebug.h"
108 # include "FetchMe.h"
115 //@node Variables and Data structures, Prototypes, Includes, Main scheduling code
116 //@subsection Variables and Data structures
120 * These are the threads which clients have requested that we run.
122 * In an SMP build, we might have several concurrent clients all
123 * waiting for results, and each one will wait on a condition variable
124 * until the result is available.
126 * In non-SMP, clients are strictly nested: the first client calls
127 * into the RTS, which might call out again to C with a _ccall_GC, and
128 * eventually re-enter the RTS.
130 * Main threads information is kept in a linked list:
132 //@cindex StgMainThread
133 typedef struct StgMainThread_ {
135 SchedulerStatus stat;
138 pthread_cond_t wakeup;
140 struct StgMainThread_ *link;
143 /* Main thread queue.
144 * Locks required: sched_mutex.
146 static StgMainThread *main_threads;
149 * Locks required: sched_mutex.
153 StgTSO* ActiveTSO = NULL; /* for assigning system costs; GranSim-Light only */
154 /* rtsTime TimeOfNextEvent, EndOfTimeSlice; now in GranSim.c */
157 In GranSim we have a runable and a blocked queue for each processor.
158 In order to minimise code changes new arrays run_queue_hds/tls
159 are created. run_queue_hd is then a short cut (macro) for
160 run_queue_hds[CurrentProc] (see GranSim.h).
163 StgTSO *run_queue_hds[MAX_PROC], *run_queue_tls[MAX_PROC];
164 StgTSO *blocked_queue_hds[MAX_PROC], *blocked_queue_tls[MAX_PROC];
165 StgTSO *ccalling_threadss[MAX_PROC];
166 /* We use the same global list of threads (all_threads) in GranSim as in
167 the std RTS (i.e. we are cheating). However, we don't use this list in
168 the GranSim specific code at the moment (so we are only potentially
173 StgTSO *run_queue_hd, *run_queue_tl;
174 StgTSO *blocked_queue_hd, *blocked_queue_tl;
175 StgTSO *sleeping_queue; /* perhaps replace with a hash table? */
179 /* Linked list of all threads.
180 * Used for detecting garbage collected threads.
184 /* Threads suspended in _ccall_GC.
186 static StgTSO *suspended_ccalling_threads;
188 static StgTSO *threadStackOverflow(StgTSO *tso);
190 /* KH: The following two flags are shared memory locations. There is no need
191 to lock them, since they are only unset at the end of a scheduler
195 /* flag set by signal handler to precipitate a context switch */
196 //@cindex context_switch
199 /* if this flag is set as well, give up execution */
200 //@cindex interrupted
203 /* Next thread ID to allocate.
204 * Locks required: sched_mutex
206 //@cindex next_thread_id
207 StgThreadID next_thread_id = 1;
210 * Pointers to the state of the current thread.
211 * Rule of thumb: if CurrentTSO != NULL, then we're running a Haskell
212 * thread. If CurrentTSO == NULL, then we're at the scheduler level.
215 /* The smallest stack size that makes any sense is:
216 * RESERVED_STACK_WORDS (so we can get back from the stack overflow)
217 * + sizeofW(StgStopFrame) (the stg_stop_thread_info frame)
218 * + 1 (the realworld token for an IO thread)
219 * + 1 (the closure to enter)
221 * A thread with this stack will bomb immediately with a stack
222 * overflow, which will increase its stack size.
225 #define MIN_STACK_WORDS (RESERVED_STACK_WORDS + sizeofW(StgStopFrame) + 2)
227 /* Free capability list.
228 * Locks required: sched_mutex.
231 Capability *free_capabilities; /* Available capabilities for running threads */
232 nat n_free_capabilities; /* total number of available capabilities */
234 Capability MainCapability; /* 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);
448 if (m->ret) *(m->ret) = NULL;
450 if (was_interrupted) {
451 m->stat = Interrupted;
455 pthread_cond_broadcast(&m->wakeup);
466 /* in GUM do this only on the Main PE */
469 /* If our main thread has finished or been killed, return.
472 StgMainThread *m = main_threads;
473 if (m->tso->what_next == ThreadComplete
474 || m->tso->what_next == ThreadKilled) {
475 main_threads = main_threads->link;
476 if (m->tso->what_next == ThreadComplete) {
477 /* we finished successfully, fill in the return value */
478 if (m->ret) { *(m->ret) = (StgClosure *)m->tso->sp[0]; };
482 if (m->ret) { *(m->ret) = NULL; };
483 if (was_interrupted) {
484 m->stat = Interrupted;
494 /* Top up the run queue from our spark pool. We try to make the
495 * number of threads in the run queue equal to the number of
500 nat n = n_free_capabilities;
501 StgTSO *tso = run_queue_hd;
503 /* Count the run queue */
504 while (n > 0 && tso != END_TSO_QUEUE) {
511 spark = findSpark(rtsFalse);
513 break; /* no more sparks in the pool */
515 /* I'd prefer this to be done in activateSpark -- HWL */
516 /* tricky - it needs to hold the scheduler lock and
517 * not try to re-acquire it -- SDM */
518 createSparkThread(spark);
520 sched_belch("==^^ turning spark of closure %p into a thread",
521 (StgClosure *)spark));
524 /* We need to wake up the other tasks if we just created some
527 if (n_free_capabilities - n > 1) {
528 pthread_cond_signal(&thread_ready_cond);
533 /* check for signals each time around the scheduler */
534 #ifndef mingw32_TARGET_OS
535 if (signals_pending()) {
536 startSignalHandlers();
540 /* Check whether any waiting threads need to be woken up. If the
541 * run queue is empty, and there are no other tasks running, we
542 * can wait indefinitely for something to happen.
543 * ToDo: what if another client comes along & requests another
546 if (blocked_queue_hd != END_TSO_QUEUE || sleeping_queue != END_TSO_QUEUE) {
548 (run_queue_hd == END_TSO_QUEUE)
550 && (n_free_capabilities == RtsFlags.ParFlags.nNodes)
554 /* we can be interrupted while waiting for I/O... */
555 if (interrupted) continue;
558 * Detect deadlock: when we have no threads to run, there are no
559 * threads waiting on I/O or sleeping, and all the other tasks are
560 * waiting for work, we must have a deadlock of some description.
562 * We first try to find threads blocked on themselves (ie. black
563 * holes), and generate NonTermination exceptions where necessary.
565 * If no threads are black holed, we have a deadlock situation, so
566 * inform all the main threads.
569 if (blocked_queue_hd == END_TSO_QUEUE
570 && run_queue_hd == END_TSO_QUEUE
571 && sleeping_queue == END_TSO_QUEUE
573 && (n_free_capabilities == RtsFlags.ParFlags.nNodes)
577 IF_DEBUG(scheduler, sched_belch("deadlocked, forcing major GC..."));
578 GarbageCollect(GetRoots,rtsTrue);
579 if (blocked_queue_hd == END_TSO_QUEUE
580 && run_queue_hd == END_TSO_QUEUE
581 && sleeping_queue == END_TSO_QUEUE) {
583 IF_DEBUG(scheduler, sched_belch("still deadlocked, checking for black holes..."));
586 // No black holes, so probably a real deadlock. Send the
587 // current main thread the Deadlock exception (or in the SMP
588 // build, send *all* main threads the deadlock exception,
589 // since none of them can make progress).
590 if (run_queue_hd == END_TSO_QUEUE) {
593 for (m = main_threads; m != NULL; m = m->link) {
594 switch (m->tso->why_blocked) {
595 case BlockedOnBlackHole:
596 raiseAsync(m->tso, (StgClosure *)NonTermination_closure);
598 case BlockedOnException:
600 raiseAsync(m->tso, (StgClosure *)Deadlock_closure);
603 barf("deadlock: main thread blocked in a strange way");
608 switch (m->tso->why_blocked) {
609 case BlockedOnBlackHole:
610 raiseAsync(m->tso, (StgClosure *)NonTermination_closure);
612 case BlockedOnException:
614 raiseAsync(m->tso, (StgClosure *)Deadlock_closure);
617 barf("deadlock: main thread blocked in a strange way");
621 ASSERT( run_queue_hd != END_TSO_QUEUE );
625 /* ToDo: add deadlock detection in GUM (similar to SMP) -- HWL */
629 /* If there's a GC pending, don't do anything until it has
633 IF_DEBUG(scheduler,sched_belch("waiting for GC"));
634 pthread_cond_wait(&gc_pending_cond, &sched_mutex);
637 /* block until we've got a thread on the run queue and a free
640 while (run_queue_hd == END_TSO_QUEUE || free_capabilities == NULL) {
641 IF_DEBUG(scheduler, sched_belch("waiting for work"));
642 pthread_cond_wait(&thread_ready_cond, &sched_mutex);
643 IF_DEBUG(scheduler, sched_belch("work now available"));
649 if (RtsFlags.GranFlags.Light)
650 GranSimLight_enter_system(event, &ActiveTSO); // adjust ActiveTSO etc
652 /* adjust time based on time-stamp */
653 if (event->time > CurrentTime[CurrentProc] &&
654 event->evttype != ContinueThread)
655 CurrentTime[CurrentProc] = event->time;
657 /* Deal with the idle PEs (may issue FindWork or MoveSpark events) */
658 if (!RtsFlags.GranFlags.Light)
661 IF_DEBUG(gran, fprintf(stderr, "GRAN: switch by event-type\n"));
663 /* main event dispatcher in GranSim */
664 switch (event->evttype) {
665 /* Should just be continuing execution */
667 IF_DEBUG(gran, fprintf(stderr, "GRAN: doing ContinueThread\n"));
668 /* ToDo: check assertion
669 ASSERT(run_queue_hd != (StgTSO*)NULL &&
670 run_queue_hd != END_TSO_QUEUE);
672 /* Ignore ContinueThreads for fetching threads (if synchr comm) */
673 if (!RtsFlags.GranFlags.DoAsyncFetch &&
674 procStatus[CurrentProc]==Fetching) {
675 belch("ghuH: Spurious ContinueThread while Fetching ignored; TSO %d (%p) [PE %d]",
676 CurrentTSO->id, CurrentTSO, CurrentProc);
679 /* Ignore ContinueThreads for completed threads */
680 if (CurrentTSO->what_next == ThreadComplete) {
681 belch("ghuH: found a ContinueThread event for completed thread %d (%p) [PE %d] (ignoring ContinueThread)",
682 CurrentTSO->id, CurrentTSO, CurrentProc);
685 /* Ignore ContinueThreads for threads that are being migrated */
686 if (PROCS(CurrentTSO)==Nowhere) {
687 belch("ghuH: trying to run the migrating TSO %d (%p) [PE %d] (ignoring ContinueThread)",
688 CurrentTSO->id, CurrentTSO, CurrentProc);
691 /* The thread should be at the beginning of the run queue */
692 if (CurrentTSO!=run_queue_hds[CurrentProc]) {
693 belch("ghuH: TSO %d (%p) [PE %d] is not at the start of the run_queue when doing a ContinueThread",
694 CurrentTSO->id, CurrentTSO, CurrentProc);
695 break; // run the thread anyway
698 new_event(proc, proc, CurrentTime[proc],
700 (StgTSO*)NULL, (StgClosure*)NULL, (rtsSpark*)NULL);
702 */ /* Catches superfluous CONTINUEs -- should be unnecessary */
703 break; // now actually run the thread; DaH Qu'vam yImuHbej
706 do_the_fetchnode(event);
707 goto next_thread; /* handle next event in event queue */
710 do_the_globalblock(event);
711 goto next_thread; /* handle next event in event queue */
714 do_the_fetchreply(event);
715 goto next_thread; /* handle next event in event queue */
717 case UnblockThread: /* Move from the blocked queue to the tail of */
718 do_the_unblock(event);
719 goto next_thread; /* handle next event in event queue */
721 case ResumeThread: /* Move from the blocked queue to the tail of */
722 /* the runnable queue ( i.e. Qu' SImqa'lu') */
723 event->tso->gran.blocktime +=
724 CurrentTime[CurrentProc] - event->tso->gran.blockedat;
725 do_the_startthread(event);
726 goto next_thread; /* handle next event in event queue */
729 do_the_startthread(event);
730 goto next_thread; /* handle next event in event queue */
733 do_the_movethread(event);
734 goto next_thread; /* handle next event in event queue */
737 do_the_movespark(event);
738 goto next_thread; /* handle next event in event queue */
741 do_the_findwork(event);
742 goto next_thread; /* handle next event in event queue */
745 barf("Illegal event type %u\n", event->evttype);
748 /* This point was scheduler_loop in the old RTS */
750 IF_DEBUG(gran, belch("GRAN: after main switch"));
752 TimeOfLastEvent = CurrentTime[CurrentProc];
753 TimeOfNextEvent = get_time_of_next_event();
754 IgnoreEvents=(TimeOfNextEvent==0); // HWL HACK
755 // CurrentTSO = ThreadQueueHd;
757 IF_DEBUG(gran, belch("GRAN: time of next event is: %ld",
760 if (RtsFlags.GranFlags.Light)
761 GranSimLight_leave_system(event, &ActiveTSO);
763 EndOfTimeSlice = CurrentTime[CurrentProc]+RtsFlags.GranFlags.time_slice;
766 belch("GRAN: end of time-slice is %#lx", EndOfTimeSlice));
768 /* in a GranSim setup the TSO stays on the run queue */
770 /* Take a thread from the run queue. */
771 t = POP_RUN_QUEUE(); // take_off_run_queue(t);
774 fprintf(stderr, "GRAN: About to run current thread, which is\n");
777 context_switch = 0; // turned on via GranYield, checking events and time slice
780 DumpGranEvent(GR_SCHEDULE, t));
782 procStatus[CurrentProc] = Busy;
785 if (PendingFetches != END_BF_QUEUE) {
789 /* ToDo: phps merge with spark activation above */
790 /* check whether we have local work and send requests if we have none */
791 if (EMPTY_RUN_QUEUE()) { /* no runnable threads */
792 /* :-[ no local threads => look out for local sparks */
793 /* the spark pool for the current PE */
794 pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable
795 if (advisory_thread_count < RtsFlags.ParFlags.maxThreads &&
796 pool->hd < pool->tl) {
798 * ToDo: add GC code check that we really have enough heap afterwards!!
800 * If we're here (no runnable threads) and we have pending
801 * sparks, we must have a space problem. Get enough space
802 * to turn one of those pending sparks into a
806 spark = findSpark(rtsFalse); /* get a spark */
807 if (spark != (rtsSpark) NULL) {
808 tso = activateSpark(spark); /* turn the spark into a thread */
809 IF_PAR_DEBUG(schedule,
810 belch("==== schedule: Created TSO %d (%p); %d threads active",
811 tso->id, tso, advisory_thread_count));
813 if (tso==END_TSO_QUEUE) { /* failed to activate spark->back to loop */
814 belch("==^^ failed to activate spark");
816 } /* otherwise fall through & pick-up new tso */
818 IF_PAR_DEBUG(verbose,
819 belch("==^^ no local sparks (spark pool contains only NFs: %d)",
820 spark_queue_len(pool)));
825 /* If we still have no work we need to send a FISH to get a spark
828 if (EMPTY_RUN_QUEUE()) {
829 /* =8-[ no local sparks => look for work on other PEs */
831 * We really have absolutely no work. Send out a fish
832 * (there may be some out there already), and wait for
833 * something to arrive. We clearly can't run any threads
834 * until a SCHEDULE or RESUME arrives, and so that's what
835 * we're hoping to see. (Of course, we still have to
836 * respond to other types of messages.)
838 TIME now = msTime() /*CURRENT_TIME*/;
839 IF_PAR_DEBUG(verbose,
840 belch("-- now=%ld", now));
841 IF_PAR_DEBUG(verbose,
842 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
843 (last_fish_arrived_at!=0 &&
844 last_fish_arrived_at+RtsFlags.ParFlags.fishDelay > now)) {
845 belch("--$$ delaying FISH until %ld (last fish %ld, delay %ld, now %ld)",
846 last_fish_arrived_at+RtsFlags.ParFlags.fishDelay,
847 last_fish_arrived_at,
848 RtsFlags.ParFlags.fishDelay, now);
851 if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
852 (last_fish_arrived_at==0 ||
853 (last_fish_arrived_at+RtsFlags.ParFlags.fishDelay <= now))) {
854 /* outstandingFishes is set in sendFish, processFish;
855 avoid flooding system with fishes via delay */
857 sendFish(pe, mytid, NEW_FISH_AGE, NEW_FISH_HISTORY,
860 // Global statistics: count no. of fishes
861 if (RtsFlags.ParFlags.ParStats.Global &&
862 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
863 globalParStats.tot_fish_mess++;
867 receivedFinish = processMessages();
870 } else if (PacketsWaiting()) { /* Look for incoming messages */
871 receivedFinish = processMessages();
874 /* Now we are sure that we have some work available */
875 ASSERT(run_queue_hd != END_TSO_QUEUE);
877 /* Take a thread from the run queue, if we have work */
878 t = POP_RUN_QUEUE(); // take_off_run_queue(END_TSO_QUEUE);
879 IF_DEBUG(sanity,checkTSO(t));
881 /* ToDo: write something to the log-file
882 if (RTSflags.ParFlags.granSimStats && !sameThread)
883 DumpGranEvent(GR_SCHEDULE, RunnableThreadsHd);
887 /* the spark pool for the current PE */
888 pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable
891 belch("--=^ %d threads, %d sparks on [%#x]",
892 run_queue_len(), spark_queue_len(pool), CURRENT_PROC));
895 if (0 && RtsFlags.ParFlags.ParStats.Full &&
896 t && LastTSO && t->id != LastTSO->id &&
897 LastTSO->why_blocked == NotBlocked &&
898 LastTSO->what_next != ThreadComplete) {
899 // if previously scheduled TSO not blocked we have to record the context switch
900 DumpVeryRawGranEvent(TimeOfLastYield, CURRENT_PROC, CURRENT_PROC,
901 GR_DESCHEDULE, LastTSO, (StgClosure *)NULL, 0, 0);
904 if (RtsFlags.ParFlags.ParStats.Full &&
905 (emitSchedule /* forced emit */ ||
906 (t && LastTSO && t->id != LastTSO->id))) {
908 we are running a different TSO, so write a schedule event to log file
909 NB: If we use fair scheduling we also have to write a deschedule
910 event for LastTSO; with unfair scheduling we know that the
911 previous tso has blocked whenever we switch to another tso, so
912 we don't need it in GUM for now
914 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
915 GR_SCHEDULE, t, (StgClosure *)NULL, 0, 0);
916 emitSchedule = rtsFalse;
920 #else /* !GRAN && !PAR */
922 /* grab a thread from the run queue
924 ASSERT(run_queue_hd != END_TSO_QUEUE);
927 // Sanity check the thread we're about to run. This can be
928 // expensive if there is lots of thread switching going on...
929 IF_DEBUG(sanity,checkTSO(t));
936 cap = free_capabilities;
937 free_capabilities = cap->link;
938 n_free_capabilities--;
940 cap = &MainCapability;
943 cap->r.rCurrentTSO = t;
945 /* context switches are now initiated by the timer signal, unless
946 * the user specified "context switch as often as possible", with
951 RtsFlags.ProfFlags.profileInterval == 0 ||
953 (RtsFlags.ConcFlags.ctxtSwitchTicks == 0
954 && (run_queue_hd != END_TSO_QUEUE
955 || blocked_queue_hd != END_TSO_QUEUE
956 || sleeping_queue != END_TSO_QUEUE)))
961 RELEASE_LOCK(&sched_mutex);
963 IF_DEBUG(scheduler, sched_belch("-->> Running TSO %ld (%p) %s ...",
964 t->id, t, whatNext_strs[t->what_next]));
967 startHeapProfTimer();
970 /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
971 /* Run the current thread
973 switch (cap->r.rCurrentTSO->what_next) {
976 /* Thread already finished, return to scheduler. */
977 ret = ThreadFinished;
980 ret = StgRun((StgFunPtr) stg_enterStackTop, &cap->r);
983 ret = StgRun((StgFunPtr) stg_returnToStackTop, &cap->r);
985 case ThreadEnterInterp:
986 ret = interpretBCO(cap);
989 barf("schedule: invalid what_next field");
991 /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
993 /* Costs for the scheduler are assigned to CCS_SYSTEM */
999 ACQUIRE_LOCK(&sched_mutex);
1002 IF_DEBUG(scheduler,fprintf(stderr,"scheduler (task %ld): ", pthread_self()););
1003 #elif !defined(GRAN) && !defined(PAR)
1004 IF_DEBUG(scheduler,fprintf(stderr,"scheduler: "););
1006 t = cap->r.rCurrentTSO;
1009 /* HACK 675: if the last thread didn't yield, make sure to print a
1010 SCHEDULE event to the log file when StgRunning the next thread, even
1011 if it is the same one as before */
1013 TimeOfLastYield = CURRENT_TIME;
1019 IF_DEBUG(gran, DumpGranEvent(GR_DESCHEDULE, t));
1020 globalGranStats.tot_heapover++;
1022 globalParStats.tot_heapover++;
1025 // did the task ask for a large block?
1026 if (cap->r.rHpAlloc > BLOCK_SIZE_W) {
1027 // if so, get one and push it on the front of the nursery.
1031 blocks = (nat)BLOCK_ROUND_UP(cap->r.rHpAlloc * sizeof(W_)) / BLOCK_SIZE;
1033 IF_DEBUG(scheduler,belch("--<< thread %ld (%p; %s) stopped: requesting a large block (size %d)",
1035 whatNext_strs[t->what_next], blocks));
1037 // don't do this if it would push us over the
1038 // alloc_blocks_lim limit; we'll GC first.
1039 if (alloc_blocks + blocks < alloc_blocks_lim) {
1041 alloc_blocks += blocks;
1042 bd = allocGroup( blocks );
1044 // link the new group into the list
1045 bd->link = cap->r.rCurrentNursery;
1046 bd->u.back = cap->r.rCurrentNursery->u.back;
1047 if (cap->r.rCurrentNursery->u.back != NULL) {
1048 cap->r.rCurrentNursery->u.back->link = bd;
1050 ASSERT(g0s0->blocks == cap->r.rCurrentNursery &&
1051 g0s0->blocks == cap->r.rNursery);
1052 cap->r.rNursery = g0s0->blocks = bd;
1054 cap->r.rCurrentNursery->u.back = bd;
1056 // initialise it as a nursery block
1060 bd->free = bd->start;
1062 // don't forget to update the block count in g0s0.
1063 g0s0->n_blocks += blocks;
1064 ASSERT(countBlocks(g0s0->blocks) == g0s0->n_blocks);
1066 // now update the nursery to point to the new block
1067 cap->r.rCurrentNursery = bd;
1069 // we might be unlucky and have another thread get on the
1070 // run queue before us and steal the large block, but in that
1071 // case the thread will just end up requesting another large
1073 PUSH_ON_RUN_QUEUE(t);
1078 /* make all the running tasks block on a condition variable,
1079 * maybe set context_switch and wait till they all pile in,
1080 * then have them wait on a GC condition variable.
1082 IF_DEBUG(scheduler,belch("--<< thread %ld (%p; %s) stopped: HeapOverflow",
1083 t->id, t, whatNext_strs[t->what_next]));
1086 ASSERT(!is_on_queue(t,CurrentProc));
1088 /* Currently we emit a DESCHEDULE event before GC in GUM.
1089 ToDo: either add separate event to distinguish SYSTEM time from rest
1090 or just nuke this DESCHEDULE (and the following SCHEDULE) */
1091 if (0 && RtsFlags.ParFlags.ParStats.Full) {
1092 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
1093 GR_DESCHEDULE, t, (StgClosure *)NULL, 0, 0);
1094 emitSchedule = rtsTrue;
1098 ready_to_gc = rtsTrue;
1099 context_switch = 1; /* stop other threads ASAP */
1100 PUSH_ON_RUN_QUEUE(t);
1101 /* actual GC is done at the end of the while loop */
1107 DumpGranEvent(GR_DESCHEDULE, t));
1108 globalGranStats.tot_stackover++;
1111 // DumpGranEvent(GR_DESCHEDULE, t);
1112 globalParStats.tot_stackover++;
1114 IF_DEBUG(scheduler,belch("--<< thread %ld (%p; %s) stopped, StackOverflow",
1115 t->id, t, whatNext_strs[t->what_next]));
1116 /* just adjust the stack for this thread, then pop it back
1122 /* enlarge the stack */
1123 StgTSO *new_t = threadStackOverflow(t);
1125 /* This TSO has moved, so update any pointers to it from the
1126 * main thread stack. It better not be on any other queues...
1127 * (it shouldn't be).
1129 for (m = main_threads; m != NULL; m = m->link) {
1134 threadPaused(new_t);
1135 PUSH_ON_RUN_QUEUE(new_t);
1139 case ThreadYielding:
1142 DumpGranEvent(GR_DESCHEDULE, t));
1143 globalGranStats.tot_yields++;
1146 // DumpGranEvent(GR_DESCHEDULE, t);
1147 globalParStats.tot_yields++;
1149 /* put the thread back on the run queue. Then, if we're ready to
1150 * GC, check whether this is the last task to stop. If so, wake
1151 * up the GC thread. getThread will block during a GC until the
1155 if (t->what_next == ThreadEnterInterp) {
1156 /* ToDo: or maybe a timer expired when we were in Hugs?
1157 * or maybe someone hit ctrl-C
1159 belch("--<< thread %ld (%p; %s) stopped to switch to Hugs",
1160 t->id, t, whatNext_strs[t->what_next]);
1162 belch("--<< thread %ld (%p; %s) stopped, yielding",
1163 t->id, t, whatNext_strs[t->what_next]);
1170 //belch("&& Doing sanity check on yielding TSO %ld.", t->id);
1172 ASSERT(t->link == END_TSO_QUEUE);
1174 ASSERT(!is_on_queue(t,CurrentProc));
1177 //belch("&& Doing sanity check on all ThreadQueues (and their TSOs).");
1178 checkThreadQsSanity(rtsTrue));
1181 if (RtsFlags.ParFlags.doFairScheduling) {
1182 /* this does round-robin scheduling; good for concurrency */
1183 APPEND_TO_RUN_QUEUE(t);
1185 /* this does unfair scheduling; good for parallelism */
1186 PUSH_ON_RUN_QUEUE(t);
1189 /* this does round-robin scheduling; good for concurrency */
1190 APPEND_TO_RUN_QUEUE(t);
1193 /* add a ContinueThread event to actually process the thread */
1194 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1196 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1198 belch("GRAN: eventq and runnableq after adding yielded thread to queue again:");
1207 belch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: ",
1208 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)));
1209 if (t->block_info.closure!=(StgClosure*)NULL) print_bq(t->block_info.closure));
1211 // ??? needed; should emit block before
1213 DumpGranEvent(GR_DESCHEDULE, t));
1214 prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
1217 ASSERT(procStatus[CurrentProc]==Busy ||
1218 ((procStatus[CurrentProc]==Fetching) &&
1219 (t->block_info.closure!=(StgClosure*)NULL)));
1220 if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
1221 !(!RtsFlags.GranFlags.DoAsyncFetch &&
1222 procStatus[CurrentProc]==Fetching))
1223 procStatus[CurrentProc] = Idle;
1227 belch("--<< thread %ld (%p; %s) stopped, blocking on node %p with BQ: ",
1228 t->id, t, whatNext_strs[t->what_next], t->block_info.closure));
1231 if (t->block_info.closure!=(StgClosure*)NULL)
1232 print_bq(t->block_info.closure));
1234 /* Send a fetch (if BlockedOnGA) and dump event to log file */
1237 /* whatever we schedule next, we must log that schedule */
1238 emitSchedule = rtsTrue;
1241 /* don't need to do anything. Either the thread is blocked on
1242 * I/O, in which case we'll have called addToBlockedQueue
1243 * previously, or it's blocked on an MVar or Blackhole, in which
1244 * case it'll be on the relevant queue already.
1247 fprintf(stderr, "--<< thread %d (%p) stopped: ", t->id, t);
1248 printThreadBlockage(t);
1249 fprintf(stderr, "\n"));
1251 /* Only for dumping event to log file
1252 ToDo: do I need this in GranSim, too?
1259 case ThreadFinished:
1260 /* Need to check whether this was a main thread, and if so, signal
1261 * the task that started it with the return value. If we have no
1262 * more main threads, we probably need to stop all the tasks until
1265 /* We also end up here if the thread kills itself with an
1266 * uncaught exception, see Exception.hc.
1268 IF_DEBUG(scheduler,belch("--++ thread %d (%p) finished", t->id, t));
1270 endThread(t, CurrentProc); // clean-up the thread
1272 /* For now all are advisory -- HWL */
1273 //if(t->priority==AdvisoryPriority) ??
1274 advisory_thread_count--;
1277 if(t->dist.priority==RevalPriority)
1281 if (RtsFlags.ParFlags.ParStats.Full &&
1282 !RtsFlags.ParFlags.ParStats.Suppressed)
1283 DumpEndEvent(CURRENT_PROC, t, rtsFalse /* not mandatory */);
1288 barf("schedule: invalid thread return code %d", (int)ret);
1292 cap->link = free_capabilities;
1293 free_capabilities = cap;
1294 n_free_capabilities++;
1298 if (RtsFlags.ProfFlags.profileInterval==0 || performHeapProfile) {
1299 GarbageCollect(GetRoots, rtsTrue);
1301 performHeapProfile = rtsFalse;
1302 ready_to_gc = rtsFalse; // we already GC'd
1307 if (ready_to_gc && n_free_capabilities == RtsFlags.ParFlags.nNodes)
1312 /* everybody back, start the GC.
1313 * Could do it in this thread, or signal a condition var
1314 * to do it in another thread. Either way, we need to
1315 * broadcast on gc_pending_cond afterward.
1318 IF_DEBUG(scheduler,sched_belch("doing GC"));
1320 GarbageCollect(GetRoots,rtsFalse);
1321 ready_to_gc = rtsFalse;
1323 pthread_cond_broadcast(&gc_pending_cond);
1326 /* add a ContinueThread event to continue execution of current thread */
1327 new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
1329 t, (StgClosure*)NULL, (rtsSpark*)NULL);
1331 fprintf(stderr, "GRAN: eventq and runnableq after Garbage collection:\n");
1339 IF_GRAN_DEBUG(unused,
1340 print_eventq(EventHd));
1342 event = get_next_event();
1345 /* ToDo: wait for next message to arrive rather than busy wait */
1348 } /* end of while(1) */
1350 IF_PAR_DEBUG(verbose,
1351 belch("== Leaving schedule() after having received Finish"));
1354 /* ---------------------------------------------------------------------------
1355 * deleteAllThreads(): kill all the live threads.
1357 * This is used when we catch a user interrupt (^C), before performing
1358 * any necessary cleanups and running finalizers.
1359 * ------------------------------------------------------------------------- */
1361 void deleteAllThreads ( void )
1364 IF_DEBUG(scheduler,sched_belch("deleting all threads"));
1365 for (t = run_queue_hd; t != END_TSO_QUEUE; t = t->link) {
1368 for (t = blocked_queue_hd; t != END_TSO_QUEUE; t = t->link) {
1371 for (t = sleeping_queue; t != END_TSO_QUEUE; t = t->link) {
1374 run_queue_hd = run_queue_tl = END_TSO_QUEUE;
1375 blocked_queue_hd = blocked_queue_tl = END_TSO_QUEUE;
1376 sleeping_queue = END_TSO_QUEUE;
1379 /* startThread and insertThread are now in GranSim.c -- HWL */
1381 //@node Suspend and Resume, Run queue code, Main scheduling loop, Main scheduling code
1382 //@subsection Suspend and Resume
1384 /* ---------------------------------------------------------------------------
1385 * Suspending & resuming Haskell threads.
1387 * When making a "safe" call to C (aka _ccall_GC), the task gives back
1388 * its capability before calling the C function. This allows another
1389 * task to pick up the capability and carry on running Haskell
1390 * threads. It also means that if the C call blocks, it won't lock
1393 * The Haskell thread making the C call is put to sleep for the
1394 * duration of the call, on the susepended_ccalling_threads queue. We
1395 * give out a token to the task, which it can use to resume the thread
1396 * on return from the C function.
1397 * ------------------------------------------------------------------------- */
1400 suspendThread( StgRegTable *reg )
1405 // assume that *reg is a pointer to the StgRegTable part of a Capability
1406 cap = (Capability *)((void *)reg - sizeof(StgFunTable));
1408 ACQUIRE_LOCK(&sched_mutex);
1411 sched_belch("thread %d did a _ccall_gc", cap->r.rCurrentTSO->id));
1413 threadPaused(cap->r.rCurrentTSO);
1414 cap->r.rCurrentTSO->link = suspended_ccalling_threads;
1415 suspended_ccalling_threads = cap->r.rCurrentTSO;
1417 /* Use the thread ID as the token; it should be unique */
1418 tok = cap->r.rCurrentTSO->id;
1421 cap->link = free_capabilities;
1422 free_capabilities = cap;
1423 n_free_capabilities++;
1426 RELEASE_LOCK(&sched_mutex);
1431 resumeThread( StgInt tok )
1433 StgTSO *tso, **prev;
1436 ACQUIRE_LOCK(&sched_mutex);
1438 prev = &suspended_ccalling_threads;
1439 for (tso = suspended_ccalling_threads;
1440 tso != END_TSO_QUEUE;
1441 prev = &tso->link, tso = tso->link) {
1442 if (tso->id == (StgThreadID)tok) {
1447 if (tso == END_TSO_QUEUE) {
1448 barf("resumeThread: thread not found");
1450 tso->link = END_TSO_QUEUE;
1453 while (free_capabilities == NULL) {
1454 IF_DEBUG(scheduler, sched_belch("waiting to resume"));
1455 pthread_cond_wait(&thread_ready_cond, &sched_mutex);
1456 IF_DEBUG(scheduler, sched_belch("resuming thread %d", tso->id));
1458 cap = free_capabilities;
1459 free_capabilities = cap->link;
1460 n_free_capabilities--;
1462 cap = &MainCapability;
1465 cap->r.rCurrentTSO = tso;
1467 RELEASE_LOCK(&sched_mutex);
1472 /* ---------------------------------------------------------------------------
1474 * ------------------------------------------------------------------------ */
1475 static void unblockThread(StgTSO *tso);
1477 /* ---------------------------------------------------------------------------
1478 * Comparing Thread ids.
1480 * This is used from STG land in the implementation of the
1481 * instances of Eq/Ord for ThreadIds.
1482 * ------------------------------------------------------------------------ */
1484 int cmp_thread(const StgTSO *tso1, const StgTSO *tso2)
1486 StgThreadID id1 = tso1->id;
1487 StgThreadID id2 = tso2->id;
1489 if (id1 < id2) return (-1);
1490 if (id1 > id2) return 1;
1494 /* ---------------------------------------------------------------------------
1495 * Fetching the ThreadID from an StgTSO.
1497 * This is used in the implementation of Show for ThreadIds.
1498 * ------------------------------------------------------------------------ */
1499 int rts_getThreadId(const StgTSO *tso)
1504 /* ---------------------------------------------------------------------------
1505 Create a new thread.
1507 The new thread starts with the given stack size. Before the
1508 scheduler can run, however, this thread needs to have a closure
1509 (and possibly some arguments) pushed on its stack. See
1510 pushClosure() in Schedule.h.
1512 createGenThread() and createIOThread() (in SchedAPI.h) are
1513 convenient packaged versions of this function.
1515 currently pri (priority) is only used in a GRAN setup -- HWL
1516 ------------------------------------------------------------------------ */
1517 //@cindex createThread
1519 /* currently pri (priority) is only used in a GRAN setup -- HWL */
1521 createThread(nat stack_size, StgInt pri)
1523 return createThread_(stack_size, rtsFalse, pri);
1527 createThread_(nat size, rtsBool have_lock, StgInt pri)
1531 createThread(nat stack_size)
1533 return createThread_(stack_size, rtsFalse);
1537 createThread_(nat size, rtsBool have_lock)
1544 /* First check whether we should create a thread at all */
1546 /* check that no more than RtsFlags.ParFlags.maxThreads threads are created */
1547 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads) {
1549 belch("{createThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
1550 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
1551 return END_TSO_QUEUE;
1557 ASSERT(!RtsFlags.GranFlags.Light || CurrentProc==0);
1560 // ToDo: check whether size = stack_size - TSO_STRUCT_SIZEW
1562 /* catch ridiculously small stack sizes */
1563 if (size < MIN_STACK_WORDS + TSO_STRUCT_SIZEW) {
1564 size = MIN_STACK_WORDS + TSO_STRUCT_SIZEW;
1567 stack_size = size - TSO_STRUCT_SIZEW;
1569 tso = (StgTSO *)allocate(size);
1570 TICK_ALLOC_TSO(stack_size, 0);
1572 SET_HDR(tso, &stg_TSO_info, CCS_SYSTEM);
1574 SET_GRAN_HDR(tso, ThisPE);
1576 tso->what_next = ThreadEnterGHC;
1578 /* tso->id needs to be unique. For now we use a heavyweight mutex to
1579 * protect the increment operation on next_thread_id.
1580 * In future, we could use an atomic increment instead.
1582 if (!have_lock) { ACQUIRE_LOCK(&sched_mutex); }
1583 tso->id = next_thread_id++;
1584 if (!have_lock) { RELEASE_LOCK(&sched_mutex); }
1586 tso->why_blocked = NotBlocked;
1587 tso->blocked_exceptions = NULL;
1589 tso->stack_size = stack_size;
1590 tso->max_stack_size = round_to_mblocks(RtsFlags.GcFlags.maxStkSize)
1592 tso->sp = (P_)&(tso->stack) + stack_size;
1595 tso->prof.CCCS = CCS_MAIN;
1598 /* put a stop frame on the stack */
1599 tso->sp -= sizeofW(StgStopFrame);
1600 SET_HDR((StgClosure*)tso->sp,(StgInfoTable *)&stg_stop_thread_info,CCS_SYSTEM);
1601 tso->su = (StgUpdateFrame*)tso->sp;
1605 tso->link = END_TSO_QUEUE;
1606 /* uses more flexible routine in GranSim */
1607 insertThread(tso, CurrentProc);
1609 /* In a non-GranSim setup the pushing of a TSO onto the runq is separated
1615 if (RtsFlags.GranFlags.GranSimStats.Full)
1616 DumpGranEvent(GR_START,tso);
1618 if (RtsFlags.ParFlags.ParStats.Full)
1619 DumpGranEvent(GR_STARTQ,tso);
1620 /* HACk to avoid SCHEDULE
1624 /* Link the new thread on the global thread list.
1626 tso->global_link = all_threads;
1630 tso->dist.priority = MandatoryPriority; //by default that is...
1634 tso->gran.pri = pri;
1636 tso->gran.magic = TSO_MAGIC; // debugging only
1638 tso->gran.sparkname = 0;
1639 tso->gran.startedat = CURRENT_TIME;
1640 tso->gran.exported = 0;
1641 tso->gran.basicblocks = 0;
1642 tso->gran.allocs = 0;
1643 tso->gran.exectime = 0;
1644 tso->gran.fetchtime = 0;
1645 tso->gran.fetchcount = 0;
1646 tso->gran.blocktime = 0;
1647 tso->gran.blockcount = 0;
1648 tso->gran.blockedat = 0;
1649 tso->gran.globalsparks = 0;
1650 tso->gran.localsparks = 0;
1651 if (RtsFlags.GranFlags.Light)
1652 tso->gran.clock = Now; /* local clock */
1654 tso->gran.clock = 0;
1656 IF_DEBUG(gran,printTSO(tso));
1659 tso->par.magic = TSO_MAGIC; // debugging only
1661 tso->par.sparkname = 0;
1662 tso->par.startedat = CURRENT_TIME;
1663 tso->par.exported = 0;
1664 tso->par.basicblocks = 0;
1665 tso->par.allocs = 0;
1666 tso->par.exectime = 0;
1667 tso->par.fetchtime = 0;
1668 tso->par.fetchcount = 0;
1669 tso->par.blocktime = 0;
1670 tso->par.blockcount = 0;
1671 tso->par.blockedat = 0;
1672 tso->par.globalsparks = 0;
1673 tso->par.localsparks = 0;
1677 globalGranStats.tot_threads_created++;
1678 globalGranStats.threads_created_on_PE[CurrentProc]++;
1679 globalGranStats.tot_sq_len += spark_queue_len(CurrentProc);
1680 globalGranStats.tot_sq_probes++;
1682 // collect parallel global statistics (currently done together with GC stats)
1683 if (RtsFlags.ParFlags.ParStats.Global &&
1684 RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
1685 //fprintf(stderr, "Creating thread %d @ %11.2f\n", tso->id, usertime());
1686 globalParStats.tot_threads_created++;
1692 belch("==__ schedule: Created TSO %d (%p);",
1693 CurrentProc, tso, tso->id));
1695 IF_PAR_DEBUG(verbose,
1696 belch("==__ schedule: Created TSO %d (%p); %d threads active",
1697 tso->id, tso, advisory_thread_count));
1699 IF_DEBUG(scheduler,sched_belch("created thread %ld, stack size = %lx words",
1700 tso->id, tso->stack_size));
1707 all parallel thread creation calls should fall through the following routine.
1710 createSparkThread(rtsSpark spark)
1712 ASSERT(spark != (rtsSpark)NULL);
1713 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads)
1715 barf("{createSparkThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
1716 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
1717 return END_TSO_QUEUE;
1721 tso = createThread_(RtsFlags.GcFlags.initialStkSize, rtsTrue);
1722 if (tso==END_TSO_QUEUE)
1723 barf("createSparkThread: Cannot create TSO");
1725 tso->priority = AdvisoryPriority;
1727 pushClosure(tso,spark);
1728 PUSH_ON_RUN_QUEUE(tso);
1729 advisory_thread_count++;
1736 Turn a spark into a thread.
1737 ToDo: fix for SMP (needs to acquire SCHED_MUTEX!)
1740 //@cindex activateSpark
1742 activateSpark (rtsSpark spark)
1746 tso = createSparkThread(spark);
1747 if (RtsFlags.ParFlags.ParStats.Full) {
1748 //ASSERT(run_queue_hd == END_TSO_QUEUE); // I think ...
1749 IF_PAR_DEBUG(verbose,
1750 belch("==^^ activateSpark: turning spark of closure %p (%s) into a thread",
1751 (StgClosure *)spark, info_type((StgClosure *)spark)));
1753 // ToDo: fwd info on local/global spark to thread -- HWL
1754 // tso->gran.exported = spark->exported;
1755 // tso->gran.locked = !spark->global;
1756 // tso->gran.sparkname = spark->name;
1762 /* ---------------------------------------------------------------------------
1765 * scheduleThread puts a thread on the head of the runnable queue.
1766 * This will usually be done immediately after a thread is created.
1767 * The caller of scheduleThread must create the thread using e.g.
1768 * createThread and push an appropriate closure
1769 * on this thread's stack before the scheduler is invoked.
1770 * ------------------------------------------------------------------------ */
1773 scheduleThread(StgTSO *tso)
1775 ACQUIRE_LOCK(&sched_mutex);
1777 /* Put the new thread on the head of the runnable queue. The caller
1778 * better push an appropriate closure on this thread's stack
1779 * beforehand. In the SMP case, the thread may start running as
1780 * soon as we release the scheduler lock below.
1782 PUSH_ON_RUN_QUEUE(tso);
1786 IF_DEBUG(scheduler,printTSO(tso));
1788 RELEASE_LOCK(&sched_mutex);
1791 /* ---------------------------------------------------------------------------
1794 * Start up Posix threads to run each of the scheduler tasks.
1795 * I believe the task ids are not needed in the system as defined.
1797 * ------------------------------------------------------------------------ */
1799 #if defined(PAR) || defined(SMP)
1801 taskStart(void) /* ( void *arg STG_UNUSED) */
1807 /* ---------------------------------------------------------------------------
1810 * Initialise the scheduler. This resets all the queues - if the
1811 * queues contained any threads, they'll be garbage collected at the
1814 * This now calls startTasks(), so should only be called once! KH @ 25/10/99
1815 * ------------------------------------------------------------------------ */
1819 term_handler(int sig STG_UNUSED)
1822 ACQUIRE_LOCK(&term_mutex);
1824 RELEASE_LOCK(&term_mutex);
1830 initCapability( Capability *cap )
1832 cap->f.stgChk0 = (F_)__stg_chk_0;
1833 cap->f.stgChk1 = (F_)__stg_chk_1;
1834 cap->f.stgGCEnter1 = (F_)__stg_gc_enter_1;
1835 cap->f.stgUpdatePAP = (F_)__stg_update_PAP;
1844 for (i=0; i<=MAX_PROC; i++) {
1845 run_queue_hds[i] = END_TSO_QUEUE;
1846 run_queue_tls[i] = END_TSO_QUEUE;
1847 blocked_queue_hds[i] = END_TSO_QUEUE;
1848 blocked_queue_tls[i] = END_TSO_QUEUE;
1849 ccalling_threadss[i] = END_TSO_QUEUE;
1850 sleeping_queue = END_TSO_QUEUE;
1853 run_queue_hd = END_TSO_QUEUE;
1854 run_queue_tl = END_TSO_QUEUE;
1855 blocked_queue_hd = END_TSO_QUEUE;
1856 blocked_queue_tl = END_TSO_QUEUE;
1857 sleeping_queue = END_TSO_QUEUE;
1860 suspended_ccalling_threads = END_TSO_QUEUE;
1862 main_threads = NULL;
1863 all_threads = END_TSO_QUEUE;
1868 RtsFlags.ConcFlags.ctxtSwitchTicks =
1869 RtsFlags.ConcFlags.ctxtSwitchTime / TICK_MILLISECS;
1871 /* Install the SIGHUP handler */
1874 struct sigaction action,oact;
1876 action.sa_handler = term_handler;
1877 sigemptyset(&action.sa_mask);
1878 action.sa_flags = 0;
1879 if (sigaction(SIGTERM, &action, &oact) != 0) {
1880 barf("can't install TERM handler");
1886 /* Allocate N Capabilities */
1889 Capability *cap, *prev;
1892 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
1893 cap = stgMallocBytes(sizeof(Capability), "initScheduler:capabilities");
1894 initCapability(cap);
1898 free_capabilities = cap;
1899 n_free_capabilities = RtsFlags.ParFlags.nNodes;
1901 IF_DEBUG(scheduler,fprintf(stderr,"scheduler: Allocated %d capabilities\n",
1902 n_free_capabilities););
1904 initCapability(&MainCapability);
1907 #if defined(SMP) || defined(PAR)
1920 /* make some space for saving all the thread ids */
1921 task_ids = stgMallocBytes(RtsFlags.ParFlags.nNodes * sizeof(task_info),
1922 "initScheduler:task_ids");
1924 /* and create all the threads */
1925 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
1926 r = pthread_create(&tid,NULL,taskStart,NULL);
1928 barf("startTasks: Can't create new Posix thread");
1930 task_ids[i].id = tid;
1931 task_ids[i].mut_time = 0.0;
1932 task_ids[i].mut_etime = 0.0;
1933 task_ids[i].gc_time = 0.0;
1934 task_ids[i].gc_etime = 0.0;
1935 task_ids[i].elapsedtimestart = elapsedtime();
1936 IF_DEBUG(scheduler,fprintf(stderr,"scheduler: Started task: %ld\n",tid););
1942 exitScheduler( void )
1947 /* Don't want to use pthread_cancel, since we'd have to install
1948 * these silly exception handlers (pthread_cleanup_{push,pop}) around
1952 /* Cancel all our tasks */
1953 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
1954 pthread_cancel(task_ids[i].id);
1957 /* Wait for all the tasks to terminate */
1958 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
1959 IF_DEBUG(scheduler,fprintf(stderr,"scheduler: waiting for task %ld\n",
1961 pthread_join(task_ids[i].id, NULL);
1965 /* Send 'em all a SIGHUP. That should shut 'em up.
1967 await_death = RtsFlags.ParFlags.nNodes;
1968 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
1969 pthread_kill(task_ids[i].id,SIGTERM);
1971 while (await_death > 0) {
1977 /* -----------------------------------------------------------------------------
1978 Managing the per-task allocation areas.
1980 Each capability comes with an allocation area. These are
1981 fixed-length block lists into which allocation can be done.
1983 ToDo: no support for two-space collection at the moment???
1984 -------------------------------------------------------------------------- */
1986 /* -----------------------------------------------------------------------------
1987 * waitThread is the external interface for running a new computation
1988 * and waiting for the result.
1990 * In the non-SMP case, we create a new main thread, push it on the
1991 * main-thread stack, and invoke the scheduler to run it. The
1992 * scheduler will return when the top main thread on the stack has
1993 * completed or died, and fill in the necessary fields of the
1994 * main_thread structure.
1996 * In the SMP case, we create a main thread as before, but we then
1997 * create a new condition variable and sleep on it. When our new
1998 * main thread has completed, we'll be woken up and the status/result
1999 * will be in the main_thread struct.
2000 * -------------------------------------------------------------------------- */
2003 howManyThreadsAvail ( void )
2007 for (q = run_queue_hd; q != END_TSO_QUEUE; q = q->link)
2009 for (q = blocked_queue_hd; q != END_TSO_QUEUE; q = q->link)
2011 for (q = sleeping_queue; q != END_TSO_QUEUE; q = q->link)
2017 finishAllThreads ( void )
2020 while (run_queue_hd != END_TSO_QUEUE) {
2021 waitThread ( run_queue_hd, NULL );
2023 while (blocked_queue_hd != END_TSO_QUEUE) {
2024 waitThread ( blocked_queue_hd, NULL );
2026 while (sleeping_queue != END_TSO_QUEUE) {
2027 waitThread ( blocked_queue_hd, NULL );
2030 (blocked_queue_hd != END_TSO_QUEUE ||
2031 run_queue_hd != END_TSO_QUEUE ||
2032 sleeping_queue != END_TSO_QUEUE);
2036 waitThread(StgTSO *tso, /*out*/StgClosure **ret)
2039 SchedulerStatus stat;
2041 ACQUIRE_LOCK(&sched_mutex);
2043 m = stgMallocBytes(sizeof(StgMainThread), "waitThread");
2049 pthread_cond_init(&m->wakeup, NULL);
2052 m->link = main_threads;
2055 IF_DEBUG(scheduler, fprintf(stderr, "== scheduler: new main thread (%d)\n",
2060 pthread_cond_wait(&m->wakeup, &sched_mutex);
2061 } while (m->stat == NoStatus);
2063 /* GranSim specific init */
2064 CurrentTSO = m->tso; // the TSO to run
2065 procStatus[MainProc] = Busy; // status of main PE
2066 CurrentProc = MainProc; // PE to run it on
2071 ASSERT(m->stat != NoStatus);
2077 pthread_cond_destroy(&m->wakeup);
2080 IF_DEBUG(scheduler, fprintf(stderr, "== scheduler: main thread (%d) finished\n",
2084 RELEASE_LOCK(&sched_mutex);
2089 //@node Run queue code, Garbage Collextion Routines, Suspend and Resume, Main scheduling code
2090 //@subsection Run queue code
2094 NB: In GranSim we have many run queues; run_queue_hd is actually a macro
2095 unfolding to run_queue_hds[CurrentProc], thus CurrentProc is an
2096 implicit global variable that has to be correct when calling these
2100 /* Put the new thread on the head of the runnable queue.
2101 * The caller of createThread better push an appropriate closure
2102 * on this thread's stack before the scheduler is invoked.
2104 static /* inline */ void
2105 add_to_run_queue(tso)
2108 ASSERT(tso!=run_queue_hd && tso!=run_queue_tl);
2109 tso->link = run_queue_hd;
2111 if (run_queue_tl == END_TSO_QUEUE) {
2116 /* Put the new thread at the end of the runnable queue. */
2117 static /* inline */ void
2118 push_on_run_queue(tso)
2121 ASSERT(get_itbl((StgClosure *)tso)->type == TSO);
2122 ASSERT(run_queue_hd!=NULL && run_queue_tl!=NULL);
2123 ASSERT(tso!=run_queue_hd && tso!=run_queue_tl);
2124 if (run_queue_hd == END_TSO_QUEUE) {
2127 run_queue_tl->link = tso;
2133 Should be inlined because it's used very often in schedule. The tso
2134 argument is actually only needed in GranSim, where we want to have the
2135 possibility to schedule *any* TSO on the run queue, irrespective of the
2136 actual ordering. Therefore, if tso is not the nil TSO then we traverse
2137 the run queue and dequeue the tso, adjusting the links in the queue.
2139 //@cindex take_off_run_queue
2140 static /* inline */ StgTSO*
2141 take_off_run_queue(StgTSO *tso) {
2145 qetlaHbogh Qu' ngaSbogh ghomDaQ {tso} yIteq!
2147 if tso is specified, unlink that tso from the run_queue (doesn't have
2148 to be at the beginning of the queue); GranSim only
2150 if (tso!=END_TSO_QUEUE) {
2151 /* find tso in queue */
2152 for (t=run_queue_hd, prev=END_TSO_QUEUE;
2153 t!=END_TSO_QUEUE && t!=tso;
2157 /* now actually dequeue the tso */
2158 if (prev!=END_TSO_QUEUE) {
2159 ASSERT(run_queue_hd!=t);
2160 prev->link = t->link;
2162 /* t is at beginning of thread queue */
2163 ASSERT(run_queue_hd==t);
2164 run_queue_hd = t->link;
2166 /* t is at end of thread queue */
2167 if (t->link==END_TSO_QUEUE) {
2168 ASSERT(t==run_queue_tl);
2169 run_queue_tl = prev;
2171 ASSERT(run_queue_tl!=t);
2173 t->link = END_TSO_QUEUE;
2175 /* take tso from the beginning of the queue; std concurrent code */
2177 if (t != END_TSO_QUEUE) {
2178 run_queue_hd = t->link;
2179 t->link = END_TSO_QUEUE;
2180 if (run_queue_hd == END_TSO_QUEUE) {
2181 run_queue_tl = END_TSO_QUEUE;
2190 //@node Garbage Collextion Routines, Blocking Queue Routines, Run queue code, Main scheduling code
2191 //@subsection Garbage Collextion Routines
2193 /* ---------------------------------------------------------------------------
2194 Where are the roots that we know about?
2196 - all the threads on the runnable queue
2197 - all the threads on the blocked queue
2198 - all the threads on the sleeping queue
2199 - all the thread currently executing a _ccall_GC
2200 - all the "main threads"
2202 ------------------------------------------------------------------------ */
2204 /* This has to be protected either by the scheduler monitor, or by the
2205 garbage collection monitor (probably the latter).
2210 GetRoots(evac_fn evac)
2217 for (i=0; i<=RtsFlags.GranFlags.proc; i++) {
2218 if ((run_queue_hds[i] != END_TSO_QUEUE) && ((run_queue_hds[i] != NULL)))
2219 evac((StgClosure **)&run_queue_hds[i]);
2220 if ((run_queue_tls[i] != END_TSO_QUEUE) && ((run_queue_tls[i] != NULL)))
2221 evac((StgClosure **)&run_queue_tls[i]);
2223 if ((blocked_queue_hds[i] != END_TSO_QUEUE) && ((blocked_queue_hds[i] != NULL)))
2224 evac((StgClosure **)&blocked_queue_hds[i]);
2225 if ((blocked_queue_tls[i] != END_TSO_QUEUE) && ((blocked_queue_tls[i] != NULL)))
2226 evac((StgClosure **)&blocked_queue_tls[i]);
2227 if ((ccalling_threadss[i] != END_TSO_QUEUE) && ((ccalling_threadss[i] != NULL)))
2228 evac((StgClosure **)&ccalling_threads[i]);
2235 if (run_queue_hd != END_TSO_QUEUE) {
2236 ASSERT(run_queue_tl != END_TSO_QUEUE);
2237 evac((StgClosure **)&run_queue_hd);
2238 evac((StgClosure **)&run_queue_tl);
2241 if (blocked_queue_hd != END_TSO_QUEUE) {
2242 ASSERT(blocked_queue_tl != END_TSO_QUEUE);
2243 evac((StgClosure **)&blocked_queue_hd);
2244 evac((StgClosure **)&blocked_queue_tl);
2247 if (sleeping_queue != END_TSO_QUEUE) {
2248 evac((StgClosure **)&sleeping_queue);
2252 for (m = main_threads; m != NULL; m = m->link) {
2253 evac((StgClosure **)&m->tso);
2255 if (suspended_ccalling_threads != END_TSO_QUEUE) {
2256 evac((StgClosure **)&suspended_ccalling_threads);
2259 #if defined(SMP) || defined(PAR) || defined(GRAN)
2260 markSparkQueue(evac);
2264 /* -----------------------------------------------------------------------------
2267 This is the interface to the garbage collector from Haskell land.
2268 We provide this so that external C code can allocate and garbage
2269 collect when called from Haskell via _ccall_GC.
2271 It might be useful to provide an interface whereby the programmer
2272 can specify more roots (ToDo).
2274 This needs to be protected by the GC condition variable above. KH.
2275 -------------------------------------------------------------------------- */
2277 void (*extra_roots)(evac_fn);
2282 GarbageCollect(GetRoots,rtsFalse);
2286 performMajorGC(void)
2288 GarbageCollect(GetRoots,rtsTrue);
2292 AllRoots(evac_fn evac)
2294 GetRoots(evac); // the scheduler's roots
2295 extra_roots(evac); // the user's roots
2299 performGCWithRoots(void (*get_roots)(evac_fn))
2301 extra_roots = get_roots;
2302 GarbageCollect(AllRoots,rtsFalse);
2305 /* -----------------------------------------------------------------------------
2308 If the thread has reached its maximum stack size, then raise the
2309 StackOverflow exception in the offending thread. Otherwise
2310 relocate the TSO into a larger chunk of memory and adjust its stack
2312 -------------------------------------------------------------------------- */
2315 threadStackOverflow(StgTSO *tso)
2317 nat new_stack_size, new_tso_size, diff, stack_words;
2321 IF_DEBUG(sanity,checkTSO(tso));
2322 if (tso->stack_size >= tso->max_stack_size) {
2325 belch("@@ threadStackOverflow of TSO %d (%p): stack too large (now %ld; max is %ld",
2326 tso->id, tso, tso->stack_size, tso->max_stack_size);
2327 /* If we're debugging, just print out the top of the stack */
2328 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2331 /* Send this thread the StackOverflow exception */
2332 raiseAsync(tso, (StgClosure *)stackOverflow_closure);
2336 /* Try to double the current stack size. If that takes us over the
2337 * maximum stack size for this thread, then use the maximum instead.
2338 * Finally round up so the TSO ends up as a whole number of blocks.
2340 new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
2341 new_tso_size = (nat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) +
2342 TSO_STRUCT_SIZE)/sizeof(W_);
2343 new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */
2344 new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
2346 IF_DEBUG(scheduler, fprintf(stderr,"== scheduler: increasing stack size from %d words to %d.\n", tso->stack_size, new_stack_size));
2348 dest = (StgTSO *)allocate(new_tso_size);
2349 TICK_ALLOC_TSO(new_stack_size,0);
2351 /* copy the TSO block and the old stack into the new area */
2352 memcpy(dest,tso,TSO_STRUCT_SIZE);
2353 stack_words = tso->stack + tso->stack_size - tso->sp;
2354 new_sp = (P_)dest + new_tso_size - stack_words;
2355 memcpy(new_sp, tso->sp, stack_words * sizeof(W_));
2357 /* relocate the stack pointers... */
2358 diff = (P_)new_sp - (P_)tso->sp; /* In *words* */
2359 dest->su = (StgUpdateFrame *) ((P_)dest->su + diff);
2361 dest->stack_size = new_stack_size;
2363 /* and relocate the update frame list */
2364 relocate_stack(dest, diff);
2366 /* Mark the old TSO as relocated. We have to check for relocated
2367 * TSOs in the garbage collector and any primops that deal with TSOs.
2369 * It's important to set the sp and su values to just beyond the end
2370 * of the stack, so we don't attempt to scavenge any part of the
2373 tso->what_next = ThreadRelocated;
2375 tso->sp = (P_)&(tso->stack[tso->stack_size]);
2376 tso->su = (StgUpdateFrame *)tso->sp;
2377 tso->why_blocked = NotBlocked;
2378 dest->mut_link = NULL;
2380 IF_PAR_DEBUG(verbose,
2381 belch("@@ threadStackOverflow of TSO %d (now at %p): stack size increased to %ld",
2382 tso->id, tso, tso->stack_size);
2383 /* If we're debugging, just print out the top of the stack */
2384 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2387 IF_DEBUG(sanity,checkTSO(tso));
2389 IF_DEBUG(scheduler,printTSO(dest));
2395 //@node Blocking Queue Routines, Exception Handling Routines, Garbage Collextion Routines, Main scheduling code
2396 //@subsection Blocking Queue Routines
2398 /* ---------------------------------------------------------------------------
2399 Wake up a queue that was blocked on some resource.
2400 ------------------------------------------------------------------------ */
2404 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2409 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
2411 /* write RESUME events to log file and
2412 update blocked and fetch time (depending on type of the orig closure) */
2413 if (RtsFlags.ParFlags.ParStats.Full) {
2414 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
2415 GR_RESUMEQ, ((StgTSO *)bqe), ((StgTSO *)bqe)->block_info.closure,
2416 0, 0 /* spark_queue_len(ADVISORY_POOL) */);
2417 if (EMPTY_RUN_QUEUE())
2418 emitSchedule = rtsTrue;
2420 switch (get_itbl(node)->type) {
2422 ((StgTSO *)bqe)->par.fetchtime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2427 ((StgTSO *)bqe)->par.blocktime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2434 barf("{unblockOneLocked}Daq Qagh: unexpected closure in blocking queue");
2441 static StgBlockingQueueElement *
2442 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2445 PEs node_loc, tso_loc;
2447 node_loc = where_is(node); // should be lifted out of loop
2448 tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2449 tso_loc = where_is((StgClosure *)tso);
2450 if (IS_LOCAL_TO(PROCS(node),tso_loc)) { // TSO is local
2451 /* !fake_fetch => TSO is on CurrentProc is same as IS_LOCAL_TO */
2452 ASSERT(CurrentProc!=node_loc || tso_loc==CurrentProc);
2453 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.lunblocktime;
2454 // insertThread(tso, node_loc);
2455 new_event(tso_loc, tso_loc, CurrentTime[CurrentProc],
2457 tso, node, (rtsSpark*)NULL);
2458 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2461 } else { // TSO is remote (actually should be FMBQ)
2462 CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.mpacktime +
2463 RtsFlags.GranFlags.Costs.gunblocktime +
2464 RtsFlags.GranFlags.Costs.latency;
2465 new_event(tso_loc, CurrentProc, CurrentTime[CurrentProc],
2467 tso, node, (rtsSpark*)NULL);
2468 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
2471 /* the thread-queue-overhead is accounted for in either Resume or UnblockThread */
2473 fprintf(stderr," %s TSO %d (%p) [PE %d] (block_info.closure=%p) (next=%p) ,",
2474 (node_loc==tso_loc ? "Local" : "Global"),
2475 tso->id, tso, CurrentProc, tso->block_info.closure, tso->link));
2476 tso->block_info.closure = NULL;
2477 IF_DEBUG(scheduler,belch("-- Waking up thread %ld (%p)",
2481 static StgBlockingQueueElement *
2482 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
2484 StgBlockingQueueElement *next;
2486 switch (get_itbl(bqe)->type) {
2488 ASSERT(((StgTSO *)bqe)->why_blocked != NotBlocked);
2489 /* if it's a TSO just push it onto the run_queue */
2491 // ((StgTSO *)bqe)->link = END_TSO_QUEUE; // debugging?
2492 PUSH_ON_RUN_QUEUE((StgTSO *)bqe);
2494 unblockCount(bqe, node);
2495 /* reset blocking status after dumping event */
2496 ((StgTSO *)bqe)->why_blocked = NotBlocked;
2500 /* if it's a BLOCKED_FETCH put it on the PendingFetches list */
2502 bqe->link = (StgBlockingQueueElement *)PendingFetches;
2503 PendingFetches = (StgBlockedFetch *)bqe;
2507 /* can ignore this case in a non-debugging setup;
2508 see comments on RBHSave closures above */
2510 /* check that the closure is an RBHSave closure */
2511 ASSERT(get_itbl((StgClosure *)bqe) == &stg_RBH_Save_0_info ||
2512 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_1_info ||
2513 get_itbl((StgClosure *)bqe) == &stg_RBH_Save_2_info);
2517 barf("{unblockOneLocked}Daq Qagh: Unexpected IP (%#lx; %s) in blocking queue at %#lx\n",
2518 get_itbl((StgClosure *)bqe), info_type((StgClosure *)bqe),
2522 IF_PAR_DEBUG(bq, fprintf(stderr, ", %p (%s)", bqe, info_type((StgClosure*)bqe)));
2526 #else /* !GRAN && !PAR */
2528 unblockOneLocked(StgTSO *tso)
2532 ASSERT(get_itbl(tso)->type == TSO);
2533 ASSERT(tso->why_blocked != NotBlocked);
2534 tso->why_blocked = NotBlocked;
2536 PUSH_ON_RUN_QUEUE(tso);
2538 IF_DEBUG(scheduler,sched_belch("waking up thread %ld", tso->id));
2543 #if defined(GRAN) || defined(PAR)
2544 inline StgBlockingQueueElement *
2545 unblockOne(StgBlockingQueueElement *bqe, StgClosure *node)
2547 ACQUIRE_LOCK(&sched_mutex);
2548 bqe = unblockOneLocked(bqe, node);
2549 RELEASE_LOCK(&sched_mutex);
2554 unblockOne(StgTSO *tso)
2556 ACQUIRE_LOCK(&sched_mutex);
2557 tso = unblockOneLocked(tso);
2558 RELEASE_LOCK(&sched_mutex);
2565 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2567 StgBlockingQueueElement *bqe;
2572 belch("##-_ AwBQ for node %p on PE %d @ %ld by TSO %d (%p): ", \
2573 node, CurrentProc, CurrentTime[CurrentProc],
2574 CurrentTSO->id, CurrentTSO));
2576 node_loc = where_is(node);
2578 ASSERT(q == END_BQ_QUEUE ||
2579 get_itbl(q)->type == TSO || // q is either a TSO or an RBHSave
2580 get_itbl(q)->type == CONSTR); // closure (type constructor)
2581 ASSERT(is_unique(node));
2583 /* FAKE FETCH: magically copy the node to the tso's proc;
2584 no Fetch necessary because in reality the node should not have been
2585 moved to the other PE in the first place
2587 if (CurrentProc!=node_loc) {
2589 belch("## node %p is on PE %d but CurrentProc is %d (TSO %d); assuming fake fetch and adjusting bitmask (old: %#x)",
2590 node, node_loc, CurrentProc, CurrentTSO->id,
2591 // CurrentTSO, where_is(CurrentTSO),
2592 node->header.gran.procs));
2593 node->header.gran.procs = (node->header.gran.procs) | PE_NUMBER(CurrentProc);
2595 belch("## new bitmask of node %p is %#x",
2596 node, node->header.gran.procs));
2597 if (RtsFlags.GranFlags.GranSimStats.Global) {
2598 globalGranStats.tot_fake_fetches++;
2603 // ToDo: check: ASSERT(CurrentProc==node_loc);
2604 while (get_itbl(bqe)->type==TSO) { // q != END_TSO_QUEUE) {
2607 bqe points to the current element in the queue
2608 next points to the next element in the queue
2610 //tso = (StgTSO *)bqe; // wastes an assignment to get the type right
2611 //tso_loc = where_is(tso);
2613 bqe = unblockOneLocked(bqe, node);
2616 /* if this is the BQ of an RBH, we have to put back the info ripped out of
2617 the closure to make room for the anchor of the BQ */
2618 if (bqe!=END_BQ_QUEUE) {
2619 ASSERT(get_itbl(node)->type == RBH && get_itbl(bqe)->type == CONSTR);
2621 ASSERT((info_ptr==&RBH_Save_0_info) ||
2622 (info_ptr==&RBH_Save_1_info) ||
2623 (info_ptr==&RBH_Save_2_info));
2625 /* cf. convertToRBH in RBH.c for writing the RBHSave closure */
2626 ((StgRBH *)node)->blocking_queue = (StgBlockingQueueElement *)((StgRBHSave *)bqe)->payload[0];
2627 ((StgRBH *)node)->mut_link = (StgMutClosure *)((StgRBHSave *)bqe)->payload[1];
2630 belch("## Filled in RBH_Save for %p (%s) at end of AwBQ",
2631 node, info_type(node)));
2634 /* statistics gathering */
2635 if (RtsFlags.GranFlags.GranSimStats.Global) {
2636 // globalGranStats.tot_bq_processing_time += bq_processing_time;
2637 globalGranStats.tot_bq_len += len; // total length of all bqs awakened
2638 // globalGranStats.tot_bq_len_local += len_local; // same for local TSOs only
2639 globalGranStats.tot_awbq++; // total no. of bqs awakened
2642 fprintf(stderr,"## BQ Stats of %p: [%d entries] %s\n",
2643 node, len, (bqe!=END_BQ_QUEUE) ? "RBH" : ""));
2647 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
2649 StgBlockingQueueElement *bqe;
2651 ACQUIRE_LOCK(&sched_mutex);
2653 IF_PAR_DEBUG(verbose,
2654 belch("##-_ AwBQ for node %p on [%x]: ",
2658 if(get_itbl(q)->type == CONSTR || q==END_BQ_QUEUE) {
2659 IF_PAR_DEBUG(verbose, belch("## ... nothing to unblock so lets just return. RFP (BUG?)"));
2664 ASSERT(q == END_BQ_QUEUE ||
2665 get_itbl(q)->type == TSO ||
2666 get_itbl(q)->type == BLOCKED_FETCH ||
2667 get_itbl(q)->type == CONSTR);
2670 while (get_itbl(bqe)->type==TSO ||
2671 get_itbl(bqe)->type==BLOCKED_FETCH) {
2672 bqe = unblockOneLocked(bqe, node);
2674 RELEASE_LOCK(&sched_mutex);
2677 #else /* !GRAN && !PAR */
2679 awakenBlockedQueue(StgTSO *tso)
2681 ACQUIRE_LOCK(&sched_mutex);
2682 while (tso != END_TSO_QUEUE) {
2683 tso = unblockOneLocked(tso);
2685 RELEASE_LOCK(&sched_mutex);
2689 //@node Exception Handling Routines, Debugging Routines, Blocking Queue Routines, Main scheduling code
2690 //@subsection Exception Handling Routines
2692 /* ---------------------------------------------------------------------------
2694 - usually called inside a signal handler so it mustn't do anything fancy.
2695 ------------------------------------------------------------------------ */
2698 interruptStgRts(void)
2704 /* -----------------------------------------------------------------------------
2707 This is for use when we raise an exception in another thread, which
2709 This has nothing to do with the UnblockThread event in GranSim. -- HWL
2710 -------------------------------------------------------------------------- */
2712 #if defined(GRAN) || defined(PAR)
2714 NB: only the type of the blocking queue is different in GranSim and GUM
2715 the operations on the queue-elements are the same
2716 long live polymorphism!
2719 unblockThread(StgTSO *tso)
2721 StgBlockingQueueElement *t, **last;
2723 ACQUIRE_LOCK(&sched_mutex);
2724 switch (tso->why_blocked) {
2727 return; /* not blocked */
2730 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
2732 StgBlockingQueueElement *last_tso = END_BQ_QUEUE;
2733 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
2735 last = (StgBlockingQueueElement **)&mvar->head;
2736 for (t = (StgBlockingQueueElement *)mvar->head;
2738 last = &t->link, last_tso = t, t = t->link) {
2739 if (t == (StgBlockingQueueElement *)tso) {
2740 *last = (StgBlockingQueueElement *)tso->link;
2741 if (mvar->tail == tso) {
2742 mvar->tail = (StgTSO *)last_tso;
2747 barf("unblockThread (MVAR): TSO not found");
2750 case BlockedOnBlackHole:
2751 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
2753 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
2755 last = &bq->blocking_queue;
2756 for (t = bq->blocking_queue;
2758 last = &t->link, t = t->link) {
2759 if (t == (StgBlockingQueueElement *)tso) {
2760 *last = (StgBlockingQueueElement *)tso->link;
2764 barf("unblockThread (BLACKHOLE): TSO not found");
2767 case BlockedOnException:
2769 StgTSO *target = tso->block_info.tso;
2771 ASSERT(get_itbl(target)->type == TSO);
2773 if (target->what_next == ThreadRelocated) {
2774 target = target->link;
2775 ASSERT(get_itbl(target)->type == TSO);
2778 ASSERT(target->blocked_exceptions != NULL);
2780 last = (StgBlockingQueueElement **)&target->blocked_exceptions;
2781 for (t = (StgBlockingQueueElement *)target->blocked_exceptions;
2783 last = &t->link, t = t->link) {
2784 ASSERT(get_itbl(t)->type == TSO);
2785 if (t == (StgBlockingQueueElement *)tso) {
2786 *last = (StgBlockingQueueElement *)tso->link;
2790 barf("unblockThread (Exception): TSO not found");
2794 case BlockedOnWrite:
2796 /* take TSO off blocked_queue */
2797 StgBlockingQueueElement *prev = NULL;
2798 for (t = (StgBlockingQueueElement *)blocked_queue_hd; t != END_BQ_QUEUE;
2799 prev = t, t = t->link) {
2800 if (t == (StgBlockingQueueElement *)tso) {
2802 blocked_queue_hd = (StgTSO *)t->link;
2803 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
2804 blocked_queue_tl = END_TSO_QUEUE;
2807 prev->link = t->link;
2808 if ((StgBlockingQueueElement *)blocked_queue_tl == t) {
2809 blocked_queue_tl = (StgTSO *)prev;
2815 barf("unblockThread (I/O): TSO not found");
2818 case BlockedOnDelay:
2820 /* take TSO off sleeping_queue */
2821 StgBlockingQueueElement *prev = NULL;
2822 for (t = (StgBlockingQueueElement *)sleeping_queue; t != END_BQ_QUEUE;
2823 prev = t, t = t->link) {
2824 if (t == (StgBlockingQueueElement *)tso) {
2826 sleeping_queue = (StgTSO *)t->link;
2828 prev->link = t->link;
2833 barf("unblockThread (I/O): TSO not found");
2837 barf("unblockThread");
2841 tso->link = END_TSO_QUEUE;
2842 tso->why_blocked = NotBlocked;
2843 tso->block_info.closure = NULL;
2844 PUSH_ON_RUN_QUEUE(tso);
2845 RELEASE_LOCK(&sched_mutex);
2849 unblockThread(StgTSO *tso)
2853 ACQUIRE_LOCK(&sched_mutex);
2854 switch (tso->why_blocked) {
2857 return; /* not blocked */
2860 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
2862 StgTSO *last_tso = END_TSO_QUEUE;
2863 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
2866 for (t = mvar->head; t != END_TSO_QUEUE;
2867 last = &t->link, last_tso = t, t = t->link) {
2870 if (mvar->tail == tso) {
2871 mvar->tail = last_tso;
2876 barf("unblockThread (MVAR): TSO not found");
2879 case BlockedOnBlackHole:
2880 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
2882 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
2884 last = &bq->blocking_queue;
2885 for (t = bq->blocking_queue; t != END_TSO_QUEUE;
2886 last = &t->link, t = t->link) {
2892 barf("unblockThread (BLACKHOLE): TSO not found");
2895 case BlockedOnException:
2897 StgTSO *target = tso->block_info.tso;
2899 ASSERT(get_itbl(target)->type == TSO);
2901 while (target->what_next == ThreadRelocated) {
2902 target = target->link;
2903 ASSERT(get_itbl(target)->type == TSO);
2906 ASSERT(target->blocked_exceptions != NULL);
2908 last = &target->blocked_exceptions;
2909 for (t = target->blocked_exceptions; t != END_TSO_QUEUE;
2910 last = &t->link, t = t->link) {
2911 ASSERT(get_itbl(t)->type == TSO);
2917 barf("unblockThread (Exception): TSO not found");
2921 case BlockedOnWrite:
2923 StgTSO *prev = NULL;
2924 for (t = blocked_queue_hd; t != END_TSO_QUEUE;
2925 prev = t, t = t->link) {
2928 blocked_queue_hd = t->link;
2929 if (blocked_queue_tl == t) {
2930 blocked_queue_tl = END_TSO_QUEUE;
2933 prev->link = t->link;
2934 if (blocked_queue_tl == t) {
2935 blocked_queue_tl = prev;
2941 barf("unblockThread (I/O): TSO not found");
2944 case BlockedOnDelay:
2946 StgTSO *prev = NULL;
2947 for (t = sleeping_queue; t != END_TSO_QUEUE;
2948 prev = t, t = t->link) {
2951 sleeping_queue = t->link;
2953 prev->link = t->link;
2958 barf("unblockThread (I/O): TSO not found");
2962 barf("unblockThread");
2966 tso->link = END_TSO_QUEUE;
2967 tso->why_blocked = NotBlocked;
2968 tso->block_info.closure = NULL;
2969 PUSH_ON_RUN_QUEUE(tso);
2970 RELEASE_LOCK(&sched_mutex);
2974 /* -----------------------------------------------------------------------------
2977 * The following function implements the magic for raising an
2978 * asynchronous exception in an existing thread.
2980 * We first remove the thread from any queue on which it might be
2981 * blocked. The possible blockages are MVARs and BLACKHOLE_BQs.
2983 * We strip the stack down to the innermost CATCH_FRAME, building
2984 * thunks in the heap for all the active computations, so they can
2985 * be restarted if necessary. When we reach a CATCH_FRAME, we build
2986 * an application of the handler to the exception, and push it on
2987 * the top of the stack.
2989 * How exactly do we save all the active computations? We create an
2990 * AP_UPD for every UpdateFrame on the stack. Entering one of these
2991 * AP_UPDs pushes everything from the corresponding update frame
2992 * upwards onto the stack. (Actually, it pushes everything up to the
2993 * next update frame plus a pointer to the next AP_UPD object.
2994 * Entering the next AP_UPD object pushes more onto the stack until we
2995 * reach the last AP_UPD object - at which point the stack should look
2996 * exactly as it did when we killed the TSO and we can continue
2997 * execution by entering the closure on top of the stack.
2999 * We can also kill a thread entirely - this happens if either (a) the
3000 * exception passed to raiseAsync is NULL, or (b) there's no
3001 * CATCH_FRAME on the stack. In either case, we strip the entire
3002 * stack and replace the thread with a zombie.
3004 * -------------------------------------------------------------------------- */
3007 deleteThread(StgTSO *tso)
3009 raiseAsync(tso,NULL);
3013 raiseAsync(StgTSO *tso, StgClosure *exception)
3015 StgUpdateFrame* su = tso->su;
3016 StgPtr sp = tso->sp;
3018 /* Thread already dead? */
3019 if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
3023 IF_DEBUG(scheduler, sched_belch("raising exception in thread %ld.", tso->id));
3025 /* Remove it from any blocking queues */
3028 /* The stack freezing code assumes there's a closure pointer on
3029 * the top of the stack. This isn't always the case with compiled
3030 * code, so we have to push a dummy closure on the top which just
3031 * returns to the next return address on the stack.
3033 if ( LOOKS_LIKE_GHC_INFO((void*)*sp) ) {
3034 *(--sp) = (W_)&stg_dummy_ret_closure;
3038 nat words = ((P_)su - (P_)sp) - 1;
3042 /* If we find a CATCH_FRAME, and we've got an exception to raise,
3043 * then build PAP(handler,exception,realworld#), and leave it on
3044 * top of the stack ready to enter.
3046 if (get_itbl(su)->type == CATCH_FRAME && exception != NULL) {
3047 StgCatchFrame *cf = (StgCatchFrame *)su;
3048 /* we've got an exception to raise, so let's pass it to the
3049 * handler in this frame.
3051 ap = (StgAP_UPD *)allocate(sizeofW(StgPAP) + 2);
3052 TICK_ALLOC_UPD_PAP(3,0);
3053 SET_HDR(ap,&stg_PAP_info,cf->header.prof.ccs);
3056 ap->fun = cf->handler; /* :: Exception -> IO a */
3057 ap->payload[0] = exception;
3058 ap->payload[1] = ARG_TAG(0); /* realworld token */
3060 /* throw away the stack from Sp up to and including the
3063 sp = (P_)su + sizeofW(StgCatchFrame) - 1;
3066 /* Restore the blocked/unblocked state for asynchronous exceptions
3067 * at the CATCH_FRAME.
3069 * If exceptions were unblocked at the catch, arrange that they
3070 * are unblocked again after executing the handler by pushing an
3071 * unblockAsyncExceptions_ret stack frame.
3073 if (!cf->exceptions_blocked) {
3074 *(sp--) = (W_)&stg_unblockAsyncExceptionszh_ret_info;
3077 /* Ensure that async exceptions are blocked when running the handler.
3079 if (tso->blocked_exceptions == NULL) {
3080 tso->blocked_exceptions = END_TSO_QUEUE;
3083 /* Put the newly-built PAP on top of the stack, ready to execute
3084 * when the thread restarts.
3088 tso->what_next = ThreadEnterGHC;
3089 IF_DEBUG(sanity, checkTSO(tso));
3093 /* First build an AP_UPD consisting of the stack chunk above the
3094 * current update frame, with the top word on the stack as the
3097 ap = (StgAP_UPD *)allocate(AP_sizeW(words));
3102 ap->fun = (StgClosure *)sp[0];
3104 for(i=0; i < (nat)words; ++i) {
3105 ap->payload[i] = (StgClosure *)*sp++;
3108 switch (get_itbl(su)->type) {
3112 SET_HDR(ap,&stg_AP_UPD_info,su->header.prof.ccs /* ToDo */);
3113 TICK_ALLOC_UP_THK(words+1,0);
3116 fprintf(stderr, "scheduler: Updating ");
3117 printPtr((P_)su->updatee);
3118 fprintf(stderr, " with ");
3119 printObj((StgClosure *)ap);
3122 /* Replace the updatee with an indirection - happily
3123 * this will also wake up any threads currently
3124 * waiting on the result.
3126 * Warning: if we're in a loop, more than one update frame on
3127 * the stack may point to the same object. Be careful not to
3128 * overwrite an IND_OLDGEN in this case, because we'll screw
3129 * up the mutable lists. To be on the safe side, don't
3130 * overwrite any kind of indirection at all. See also
3131 * threadSqueezeStack in GC.c, where we have to make a similar
3134 if (!closure_IND(su->updatee)) {
3135 UPD_IND_NOLOCK(su->updatee,ap); /* revert the black hole */
3138 sp += sizeofW(StgUpdateFrame) -1;
3139 sp[0] = (W_)ap; /* push onto stack */
3145 StgCatchFrame *cf = (StgCatchFrame *)su;
3148 /* We want a PAP, not an AP_UPD. Fortunately, the
3149 * layout's the same.
3151 SET_HDR(ap,&stg_PAP_info,su->header.prof.ccs /* ToDo */);
3152 TICK_ALLOC_UPD_PAP(words+1,0);
3154 /* now build o = FUN(catch,ap,handler) */
3155 o = (StgClosure *)allocate(sizeofW(StgClosure)+2);
3156 TICK_ALLOC_FUN(2,0);
3157 SET_HDR(o,&stg_catch_info,su->header.prof.ccs /* ToDo */);
3158 o->payload[0] = (StgClosure *)ap;
3159 o->payload[1] = cf->handler;
3162 fprintf(stderr, "scheduler: Built ");
3163 printObj((StgClosure *)o);
3166 /* pop the old handler and put o on the stack */
3168 sp += sizeofW(StgCatchFrame) - 1;
3175 StgSeqFrame *sf = (StgSeqFrame *)su;
3178 SET_HDR(ap,&stg_PAP_info,su->header.prof.ccs /* ToDo */);
3179 TICK_ALLOC_UPD_PAP(words+1,0);
3181 /* now build o = FUN(seq,ap) */
3182 o = (StgClosure *)allocate(sizeofW(StgClosure)+1);
3183 TICK_ALLOC_SE_THK(1,0);
3184 SET_HDR(o,&stg_seq_info,su->header.prof.ccs /* ToDo */);
3185 o->payload[0] = (StgClosure *)ap;
3188 fprintf(stderr, "scheduler: Built ");
3189 printObj((StgClosure *)o);
3192 /* pop the old handler and put o on the stack */
3194 sp += sizeofW(StgSeqFrame) - 1;
3200 /* We've stripped the entire stack, the thread is now dead. */
3201 sp += sizeofW(StgStopFrame) - 1;
3202 sp[0] = (W_)exception; /* save the exception */
3203 tso->what_next = ThreadKilled;
3204 tso->su = (StgUpdateFrame *)(sp+1);
3215 /* -----------------------------------------------------------------------------
3216 resurrectThreads is called after garbage collection on the list of
3217 threads found to be garbage. Each of these threads will be woken
3218 up and sent a signal: BlockedOnDeadMVar if the thread was blocked
3219 on an MVar, or NonTermination if the thread was blocked on a Black
3221 -------------------------------------------------------------------------- */
3224 resurrectThreads( StgTSO *threads )
3228 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
3229 next = tso->global_link;
3230 tso->global_link = all_threads;
3232 IF_DEBUG(scheduler, sched_belch("resurrecting thread %d", tso->id));
3234 switch (tso->why_blocked) {
3236 case BlockedOnException:
3237 raiseAsync(tso,(StgClosure *)BlockedOnDeadMVar_closure);
3239 case BlockedOnBlackHole:
3240 raiseAsync(tso,(StgClosure *)NonTermination_closure);
3243 /* This might happen if the thread was blocked on a black hole
3244 * belonging to a thread that we've just woken up (raiseAsync
3245 * can wake up threads, remember...).
3249 barf("resurrectThreads: thread blocked in a strange way");
3254 /* -----------------------------------------------------------------------------
3255 * Blackhole detection: if we reach a deadlock, test whether any
3256 * threads are blocked on themselves. Any threads which are found to
3257 * be self-blocked get sent a NonTermination exception.
3259 * This is only done in a deadlock situation in order to avoid
3260 * performance overhead in the normal case.
3261 * -------------------------------------------------------------------------- */
3264 detectBlackHoles( void )
3266 StgTSO *t = all_threads;
3267 StgUpdateFrame *frame;
3268 StgClosure *blocked_on;
3270 for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
3272 while (t->what_next == ThreadRelocated) {
3274 ASSERT(get_itbl(t)->type == TSO);
3277 if (t->why_blocked != BlockedOnBlackHole) {
3281 blocked_on = t->block_info.closure;
3283 for (frame = t->su; ; frame = frame->link) {
3284 switch (get_itbl(frame)->type) {
3287 if (frame->updatee == blocked_on) {
3288 /* We are blocking on one of our own computations, so
3289 * send this thread the NonTermination exception.
3292 sched_belch("thread %d is blocked on itself", t->id));
3293 raiseAsync(t, (StgClosure *)NonTermination_closure);
3314 //@node Debugging Routines, Index, Exception Handling Routines, Main scheduling code
3315 //@subsection Debugging Routines
3317 /* -----------------------------------------------------------------------------
3318 Debugging: why is a thread blocked
3319 -------------------------------------------------------------------------- */
3324 printThreadBlockage(StgTSO *tso)
3326 switch (tso->why_blocked) {
3328 fprintf(stderr,"is blocked on read from fd %d", tso->block_info.fd);
3330 case BlockedOnWrite:
3331 fprintf(stderr,"is blocked on write to fd %d", tso->block_info.fd);
3333 case BlockedOnDelay:
3334 fprintf(stderr,"is blocked until %d", tso->block_info.target);
3337 fprintf(stderr,"is blocked on an MVar");
3339 case BlockedOnException:
3340 fprintf(stderr,"is blocked on delivering an exception to thread %d",
3341 tso->block_info.tso->id);
3343 case BlockedOnBlackHole:
3344 fprintf(stderr,"is blocked on a black hole");
3347 fprintf(stderr,"is not blocked");
3351 fprintf(stderr,"is blocked on global address; local FM_BQ is %p (%s)",
3352 tso->block_info.closure, info_type(tso->block_info.closure));
3354 case BlockedOnGA_NoSend:
3355 fprintf(stderr,"is blocked on global address (no send); local FM_BQ is %p (%s)",
3356 tso->block_info.closure, info_type(tso->block_info.closure));
3360 barf("printThreadBlockage: strange tso->why_blocked: %d for TSO %d (%d)",
3361 tso->why_blocked, tso->id, tso);
3366 printThreadStatus(StgTSO *tso)
3368 switch (tso->what_next) {
3370 fprintf(stderr,"has been killed");
3372 case ThreadComplete:
3373 fprintf(stderr,"has completed");
3376 printThreadBlockage(tso);
3381 printAllThreads(void)
3386 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
3387 ullong_format_string(TIME_ON_PROC(CurrentProc),
3388 time_string, rtsFalse/*no commas!*/);
3390 sched_belch("all threads at [%s]:", time_string);
3392 char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
3393 ullong_format_string(CURRENT_TIME,
3394 time_string, rtsFalse/*no commas!*/);
3396 sched_belch("all threads at [%s]:", time_string);
3398 sched_belch("all threads:");
3401 for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
3402 fprintf(stderr, "\tthread %d ", t->id);
3403 printThreadStatus(t);
3404 fprintf(stderr,"\n");
3409 Print a whole blocking queue attached to node (debugging only).
3414 print_bq (StgClosure *node)
3416 StgBlockingQueueElement *bqe;
3420 fprintf(stderr,"## BQ of closure %p (%s): ",
3421 node, info_type(node));
3423 /* should cover all closures that may have a blocking queue */
3424 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
3425 get_itbl(node)->type == FETCH_ME_BQ ||
3426 get_itbl(node)->type == RBH ||
3427 get_itbl(node)->type == MVAR);
3429 ASSERT(node!=(StgClosure*)NULL); // sanity check
3431 print_bqe(((StgBlockingQueue*)node)->blocking_queue);
3435 Print a whole blocking queue starting with the element bqe.
3438 print_bqe (StgBlockingQueueElement *bqe)
3443 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
3445 for (end = (bqe==END_BQ_QUEUE);
3446 !end; // iterate until bqe points to a CONSTR
3447 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE),
3448 bqe = end ? END_BQ_QUEUE : bqe->link) {
3449 ASSERT(bqe != END_BQ_QUEUE); // sanity check
3450 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
3451 /* types of closures that may appear in a blocking queue */
3452 ASSERT(get_itbl(bqe)->type == TSO ||
3453 get_itbl(bqe)->type == BLOCKED_FETCH ||
3454 get_itbl(bqe)->type == CONSTR);
3455 /* only BQs of an RBH end with an RBH_Save closure */
3456 //ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
3458 switch (get_itbl(bqe)->type) {
3460 fprintf(stderr," TSO %u (%x),",
3461 ((StgTSO *)bqe)->id, ((StgTSO *)bqe));
3464 fprintf(stderr," BF (node=%p, ga=((%x, %d, %x)),",
3465 ((StgBlockedFetch *)bqe)->node,
3466 ((StgBlockedFetch *)bqe)->ga.payload.gc.gtid,
3467 ((StgBlockedFetch *)bqe)->ga.payload.gc.slot,
3468 ((StgBlockedFetch *)bqe)->ga.weight);
3471 fprintf(stderr," %s (IP %p),",
3472 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
3473 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
3474 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
3475 "RBH_Save_?"), get_itbl(bqe));
3478 barf("Unexpected closure type %s in blocking queue", // of %p (%s)",
3479 info_type((StgClosure *)bqe)); // , node, info_type(node));
3483 fputc('\n', stderr);
3485 # elif defined(GRAN)
3487 print_bq (StgClosure *node)
3489 StgBlockingQueueElement *bqe;
3490 PEs node_loc, tso_loc;
3493 /* should cover all closures that may have a blocking queue */
3494 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
3495 get_itbl(node)->type == FETCH_ME_BQ ||
3496 get_itbl(node)->type == RBH);
3498 ASSERT(node!=(StgClosure*)NULL); // sanity check
3499 node_loc = where_is(node);
3501 fprintf(stderr,"## BQ of closure %p (%s) on [PE %d]: ",
3502 node, info_type(node), node_loc);
3505 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
3507 for (bqe = ((StgBlockingQueue*)node)->blocking_queue, end = (bqe==END_BQ_QUEUE);
3508 !end; // iterate until bqe points to a CONSTR
3509 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) {
3510 ASSERT(bqe != END_BQ_QUEUE); // sanity check
3511 ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check
3512 /* types of closures that may appear in a blocking queue */
3513 ASSERT(get_itbl(bqe)->type == TSO ||
3514 get_itbl(bqe)->type == CONSTR);
3515 /* only BQs of an RBH end with an RBH_Save closure */
3516 ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
3518 tso_loc = where_is((StgClosure *)bqe);
3519 switch (get_itbl(bqe)->type) {
3521 fprintf(stderr," TSO %d (%p) on [PE %d],",
3522 ((StgTSO *)bqe)->id, (StgTSO *)bqe, tso_loc);
3525 fprintf(stderr," %s (IP %p),",
3526 (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" :
3527 get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" :
3528 get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" :
3529 "RBH_Save_?"), get_itbl(bqe));
3532 barf("Unexpected closure type %s in blocking queue of %p (%s)",
3533 info_type((StgClosure *)bqe), node, info_type(node));
3537 fputc('\n', stderr);
3541 Nice and easy: only TSOs on the blocking queue
3544 print_bq (StgClosure *node)
3548 ASSERT(node!=(StgClosure*)NULL); // sanity check
3549 for (tso = ((StgBlockingQueue*)node)->blocking_queue;
3550 tso != END_TSO_QUEUE;
3552 ASSERT(tso!=NULL && tso!=END_TSO_QUEUE); // sanity check
3553 ASSERT(get_itbl(tso)->type == TSO); // guess what, sanity check
3554 fprintf(stderr," TSO %d (%p),", tso->id, tso);
3556 fputc('\n', stderr);
3567 for (i=0, tso=run_queue_hd;
3568 tso != END_TSO_QUEUE;
3577 sched_belch(char *s, ...)
3582 fprintf(stderr, "scheduler (task %ld): ", pthread_self());
3584 fprintf(stderr, "== ");
3586 fprintf(stderr, "scheduler: ");
3588 vfprintf(stderr, s, ap);
3589 fprintf(stderr, "\n");
3595 //@node Index, , Debugging Routines, Main scheduling code
3599 //* MainRegTable:: @cindex\s-+MainRegTable
3600 //* StgMainThread:: @cindex\s-+StgMainThread
3601 //* awaken_blocked_queue:: @cindex\s-+awaken_blocked_queue
3602 //* blocked_queue_hd:: @cindex\s-+blocked_queue_hd
3603 //* blocked_queue_tl:: @cindex\s-+blocked_queue_tl
3604 //* context_switch:: @cindex\s-+context_switch
3605 //* createThread:: @cindex\s-+createThread
3606 //* free_capabilities:: @cindex\s-+free_capabilities
3607 //* gc_pending_cond:: @cindex\s-+gc_pending_cond
3608 //* initScheduler:: @cindex\s-+initScheduler
3609 //* interrupted:: @cindex\s-+interrupted
3610 //* n_free_capabilities:: @cindex\s-+n_free_capabilities
3611 //* next_thread_id:: @cindex\s-+next_thread_id
3612 //* print_bq:: @cindex\s-+print_bq
3613 //* run_queue_hd:: @cindex\s-+run_queue_hd
3614 //* run_queue_tl:: @cindex\s-+run_queue_tl
3615 //* sched_mutex:: @cindex\s-+sched_mutex
3616 //* schedule:: @cindex\s-+schedule
3617 //* take_off_run_queue:: @cindex\s-+take_off_run_queue
3618 //* task_ids:: @cindex\s-+task_ids
3619 //* term_mutex:: @cindex\s-+term_mutex
3620 //* thread_ready_cond:: @cindex\s-+thread_ready_cond