1 /* ---------------------------------------------------------------------------
2 * $Id: Schedule.c,v 1.41 2000/01/13 14:34:05 hwloidl Exp $
4 * (c) The GHC Team, 1998-1999
8 * The main scheduling code in GranSim is quite different from that in std
9 * (concurrent) Haskell: while concurrent Haskell just iterates over the
10 * threads in the runnable queue, GranSim is event driven, i.e. it iterates
11 * over the events in the global event queue. -- HWL
12 * --------------------------------------------------------------------------*/
14 //@node Main scheduling code, , ,
15 //@section Main scheduling code
17 /* Version with scheduler monitor support for SMPs.
19 This design provides a high-level API to create and schedule threads etc.
20 as documented in the SMP design document.
22 It uses a monitor design controlled by a single mutex to exercise control
23 over accesses to shared data structures, and builds on the Posix threads
26 The majority of state is shared. In order to keep essential per-task state,
27 there is a Capability structure, which contains all the information
28 needed to run a thread: its STG registers, a pointer to its TSO, a
29 nursery etc. During STG execution, a pointer to the capability is
30 kept in a register (BaseReg).
32 In a non-SMP build, there is one global capability, namely MainRegTable.
39 //* Variables and Data structures::
41 //* Main scheduling loop::
42 //* Suspend and Resume::
44 //* Garbage Collextion Routines::
45 //* Blocking Queue Routines::
46 //* Exception Handling Routines::
47 //* Debugging Routines::
51 //@node Includes, Variables and Data structures, Main scheduling code, Main scheduling code
52 //@subsection Includes
60 #include "StgStartup.h"
64 #include "StgMiscClosures.h"
66 #include "Evaluator.h"
67 #include "Exception.h"
71 #include "Profiling.h"
75 #if defined(GRAN) || defined(PAR)
76 # include "GranSimRts.h"
78 # include "ParallelRts.h"
79 # include "Parallel.h"
80 # include "ParallelDebug.h"
87 //@node Variables and Data structures, Prototypes, Includes, Main scheduling code
88 //@subsection Variables and Data structures
92 * These are the threads which clients have requested that we run.
94 * In an SMP build, we might have several concurrent clients all
95 * waiting for results, and each one will wait on a condition variable
96 * until the result is available.
98 * In non-SMP, clients are strictly nested: the first client calls
99 * into the RTS, which might call out again to C with a _ccall_GC, and
100 * eventually re-enter the RTS.
102 * Main threads information is kept in a linked list:
104 //@cindex StgMainThread
105 typedef struct StgMainThread_ {
107 SchedulerStatus stat;
110 pthread_cond_t wakeup;
112 struct StgMainThread_ *link;
115 /* Main thread queue.
116 * Locks required: sched_mutex.
118 static StgMainThread *main_threads;
121 * Locks required: sched_mutex.
125 char *whatNext_strs[] = {
133 char *threadReturnCode_strs[] = {
134 "HeapOverflow", /* might also be StackOverflow */
144 StgTSO* ActiveTSO = NULL; /* for assigning system costs; GranSim-Light only */
145 // rtsTime TimeOfNextEvent, EndOfTimeSlice; now in GranSim.c
148 In GranSim we have a runable and a blocked queue for each processor.
149 In order to minimise code changes new arrays run_queue_hds/tls
150 are created. run_queue_hd is then a short cut (macro) for
151 run_queue_hds[CurrentProc] (see GranSim.h).
154 StgTSO *run_queue_hds[MAX_PROC], *run_queue_tls[MAX_PROC];
155 StgTSO *blocked_queue_hds[MAX_PROC], *blocked_queue_tls[MAX_PROC];
156 StgTSO *ccalling_threadss[MAX_PROC];
160 //@cindex run_queue_hd
161 //@cindex run_queue_tl
162 //@cindex blocked_queue_hd
163 //@cindex blocked_queue_tl
164 StgTSO *run_queue_hd, *run_queue_tl;
165 StgTSO *blocked_queue_hd, *blocked_queue_tl;
167 /* Threads suspended in _ccall_GC.
168 * Locks required: sched_mutex.
170 static StgTSO *suspended_ccalling_threads;
172 static void GetRoots(void);
173 static StgTSO *threadStackOverflow(StgTSO *tso);
176 /* KH: The following two flags are shared memory locations. There is no need
177 to lock them, since they are only unset at the end of a scheduler
181 /* flag set by signal handler to precipitate a context switch */
182 //@cindex context_switch
185 /* if this flag is set as well, give up execution */
186 //@cindex interrupted
189 /* Next thread ID to allocate.
190 * Locks required: sched_mutex
192 //@cindex next_thread_id
193 StgThreadID next_thread_id = 1;
196 * Pointers to the state of the current thread.
197 * Rule of thumb: if CurrentTSO != NULL, then we're running a Haskell
198 * thread. If CurrentTSO == NULL, then we're at the scheduler level.
201 /* The smallest stack size that makes any sense is:
202 * RESERVED_STACK_WORDS (so we can get back from the stack overflow)
203 * + sizeofW(StgStopFrame) (the stg_stop_thread_info frame)
204 * + 1 (the realworld token for an IO thread)
205 * + 1 (the closure to enter)
207 * A thread with this stack will bomb immediately with a stack
208 * overflow, which will increase its stack size.
211 #define MIN_STACK_WORDS (RESERVED_STACK_WORDS + sizeofW(StgStopFrame) + 2)
213 /* Free capability list.
214 * Locks required: sched_mutex.
217 //@cindex free_capabilities
218 //@cindex n_free_capabilities
219 Capability *free_capabilities; /* Available capabilities for running threads */
220 nat n_free_capabilities; /* total number of available capabilities */
222 //@cindex MainRegTable
223 Capability MainRegTable; /* for non-SMP, we have one global capability */
227 StgTSO *CurrentTSOs[MAX_PROC];
234 /* All our current task ids, saved in case we need to kill them later.
241 void addToBlockedQueue ( StgTSO *tso );
243 static void schedule ( void );
244 void interruptStgRts ( void );
245 static StgTSO * createThread_ ( nat size, rtsBool have_lock );
248 static void sched_belch(char *s, ...);
252 //@cindex sched_mutex
254 //@cindex thread_ready_cond
255 //@cindex gc_pending_cond
256 pthread_mutex_t sched_mutex = PTHREAD_MUTEX_INITIALIZER;
257 pthread_mutex_t term_mutex = PTHREAD_MUTEX_INITIALIZER;
258 pthread_cond_t thread_ready_cond = PTHREAD_COND_INITIALIZER;
259 pthread_cond_t gc_pending_cond = PTHREAD_COND_INITIALIZER;
266 rtsTime TimeOfLastYield;
270 * The thread state for the main thread.
271 // ToDo: check whether not needed any more
276 //@node Prototypes, Main scheduling loop, Variables and Data structures, Main scheduling code
277 //@subsection Prototypes
279 #if 0 && defined(GRAN)
280 // ToDo: replace these with macros
281 static /* inline */ void add_to_run_queue(StgTSO* tso);
282 static /* inline */ void push_on_run_queue(StgTSO* tso);
283 static /* inline */ StgTSO *take_off_run_queue(StgTSO *tso);
285 /* Thread management */
286 void initScheduler(void);
289 //@node Main scheduling loop, Suspend and Resume, Prototypes, Main scheduling code
290 //@subsection Main scheduling loop
292 /* ---------------------------------------------------------------------------
293 Main scheduling loop.
295 We use round-robin scheduling, each thread returning to the
296 scheduler loop when one of these conditions is detected:
299 * timer expires (thread yields)
304 Locking notes: we acquire the scheduler lock once at the beginning
305 of the scheduler loop, and release it when
307 * running a thread, or
308 * waiting for work, or
309 * waiting for a GC to complete.
311 ------------------------------------------------------------------------ */
318 StgThreadReturnCode ret;
327 ACQUIRE_LOCK(&sched_mutex);
330 # error ToDo: implement GranSim scheduler
332 while (!GlobalStopPending) { /* GlobalStopPending set in par_exit */
334 if (PendingFetches != END_BF_QUEUE) {
341 /* If we're interrupted (the user pressed ^C, or some other
342 * termination condition occurred), kill all the currently running
346 IF_DEBUG(scheduler, sched_belch("interrupted"));
347 for (t = run_queue_hd; t != END_TSO_QUEUE; t = t->link) {
350 for (t = blocked_queue_hd; t != END_TSO_QUEUE; t = t->link) {
353 run_queue_hd = run_queue_tl = END_TSO_QUEUE;
354 blocked_queue_hd = blocked_queue_tl = END_TSO_QUEUE;
357 /* Go through the list of main threads and wake up any
358 * clients whose computations have finished. ToDo: this
359 * should be done more efficiently without a linear scan
360 * of the main threads list, somehow...
364 StgMainThread *m, **prev;
365 prev = &main_threads;
366 for (m = main_threads; m != NULL; m = m->link) {
367 switch (m->tso->whatNext) {
370 *(m->ret) = (StgClosure *)m->tso->sp[0];
374 pthread_cond_broadcast(&m->wakeup);
379 pthread_cond_broadcast(&m->wakeup);
387 /* If our main thread has finished or been killed, return.
390 StgMainThread *m = main_threads;
391 if (m->tso->whatNext == ThreadComplete
392 || m->tso->whatNext == ThreadKilled) {
393 main_threads = main_threads->link;
394 if (m->tso->whatNext == ThreadComplete) {
395 /* we finished successfully, fill in the return value */
396 if (m->ret) { *(m->ret) = (StgClosure *)m->tso->sp[0]; };
407 /* Top up the run queue from our spark pool. We try to make the
408 * number of threads in the run queue equal to the number of
413 nat n = n_free_capabilities;
414 StgTSO *tso = run_queue_hd;
416 /* Count the run queue */
417 while (n > 0 && tso != END_TSO_QUEUE) {
426 break; /* no more sparks in the pool */
428 // I'd prefer this to be done in activateSpark -- HWL
430 tso = createThread_(RtsFlags.GcFlags.initialStkSize, rtsTrue);
431 pushClosure(tso,spark);
432 PUSH_ON_RUN_QUEUE(tso);
434 advisory_thread_count++;
438 sched_belch("turning spark of closure %p into a thread",
439 (StgClosure *)spark));
442 /* We need to wake up the other tasks if we just created some
445 if (n_free_capabilities - n > 1) {
446 pthread_cond_signal(&thread_ready_cond);
451 /* Check whether any waiting threads need to be woken up. If the
452 * run queue is empty, and there are no other tasks running, we
453 * can wait indefinitely for something to happen.
454 * ToDo: what if another client comes along & requests another
457 if (blocked_queue_hd != END_TSO_QUEUE) {
459 (run_queue_hd == END_TSO_QUEUE)
461 && (n_free_capabilities == RtsFlags.ParFlags.nNodes)
466 /* check for signals each time around the scheduler */
468 if (signals_pending()) {
469 start_signal_handlers();
473 /* Detect deadlock: when we have no threads to run, there are
474 * no threads waiting on I/O or sleeping, and all the other
475 * tasks are waiting for work, we must have a deadlock. Inform
476 * all the main threads.
479 if (blocked_queue_hd == END_TSO_QUEUE
480 && run_queue_hd == END_TSO_QUEUE
481 && (n_free_capabilities == RtsFlags.ParFlags.nNodes)
484 for (m = main_threads; m != NULL; m = m->link) {
487 pthread_cond_broadcast(&m->wakeup);
492 if (blocked_queue_hd == END_TSO_QUEUE
493 && run_queue_hd == END_TSO_QUEUE) {
494 StgMainThread *m = main_threads;
497 main_threads = m->link;
503 /* If there's a GC pending, don't do anything until it has
507 IF_DEBUG(scheduler,sched_belch("waiting for GC"));
508 pthread_cond_wait(&gc_pending_cond, &sched_mutex);
511 /* block until we've got a thread on the run queue and a free
514 while (run_queue_hd == END_TSO_QUEUE || free_capabilities == NULL) {
515 IF_DEBUG(scheduler, sched_belch("waiting for work"));
516 pthread_cond_wait(&thread_ready_cond, &sched_mutex);
517 IF_DEBUG(scheduler, sched_belch("work now available"));
522 # error ToDo: implement GranSim scheduler
524 // ToDo: phps merge with spark activation above
525 /* check whether we have local work and send requests if we have none */
526 if (run_queue_hd == END_TSO_QUEUE) { /* no runnable threads */
527 /* :-[ no local threads => look out for local sparks */
528 if (advisory_thread_count < RtsFlags.ParFlags.maxThreads &&
529 (pending_sparks_hd[REQUIRED_POOL] < pending_sparks_tl[REQUIRED_POOL] ||
530 pending_sparks_hd[ADVISORY_POOL] < pending_sparks_tl[ADVISORY_POOL])) {
532 * ToDo: add GC code check that we really have enough heap afterwards!!
534 * If we're here (no runnable threads) and we have pending
535 * sparks, we must have a space problem. Get enough space
536 * to turn one of those pending sparks into a
540 spark = findSpark(); /* get a spark */
541 if (spark != (rtsSpark) NULL) {
542 tso = activateSpark(spark); /* turn the spark into a thread */
543 IF_PAR_DEBUG(verbose,
544 belch("== [%x] schedule: Created TSO %p (%d); %d threads active",
545 mytid, tso, tso->id, advisory_thread_count));
547 if (tso==END_TSO_QUEUE) { // failed to activate spark -> back to loop
548 belch("^^ failed to activate spark");
550 } // otherwise fall through & pick-up new tso
552 IF_PAR_DEBUG(verbose,
553 belch("^^ no local sparks (spark pool contains only NFs: %d)",
554 spark_queue_len(ADVISORY_POOL)));
558 /* =8-[ no local sparks => look for work on other PEs */
561 * We really have absolutely no work. Send out a fish
562 * (there may be some out there already), and wait for
563 * something to arrive. We clearly can't run any threads
564 * until a SCHEDULE or RESUME arrives, and so that's what
565 * we're hoping to see. (Of course, we still have to
566 * respond to other types of messages.)
569 outstandingFishes < RtsFlags.ParFlags.maxFishes ) { // &&
570 // (last_fish_arrived_at+FISH_DELAY < CURRENT_TIME)) {
571 /* fishing set in sendFish, processFish;
572 avoid flooding system with fishes via delay */
574 sendFish(pe, mytid, NEW_FISH_AGE, NEW_FISH_HISTORY,
582 } else if (PacketsWaiting()) { /* Look for incoming messages */
586 /* Now we are sure that we have some work available */
587 ASSERT(run_queue_hd != END_TSO_QUEUE);
588 /* Take a thread from the run queue, if we have work */
589 t = take_off_run_queue(END_TSO_QUEUE);
591 /* ToDo: write something to the log-file
592 if (RTSflags.ParFlags.granSimStats && !sameThread)
593 DumpGranEvent(GR_SCHEDULE, RunnableThreadsHd);
598 IF_DEBUG(scheduler, belch("--^^ %d sparks on [%#x] (hd=%x; tl=%x; lim=%x)",
599 spark_queue_len(ADVISORY_POOL), CURRENT_PROC,
600 pending_sparks_hd[ADVISORY_POOL],
601 pending_sparks_tl[ADVISORY_POOL],
602 pending_sparks_lim[ADVISORY_POOL]));
604 IF_DEBUG(scheduler, belch("--== %d threads on [%#x] (hd=%x; tl=%x)",
605 run_queue_len(), CURRENT_PROC,
606 run_queue_hd, run_queue_tl));
610 we are running a different TSO, so write a schedule event to log file
611 NB: If we use fair scheduling we also have to write a deschedule
612 event for LastTSO; with unfair scheduling we know that the
613 previous tso has blocked whenever we switch to another tso, so
614 we don't need it in GUM for now
616 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
617 GR_SCHEDULE, t, (StgClosure *)NULL, 0, 0);
621 #else /* !GRAN && !PAR */
623 /* grab a thread from the run queue
632 cap = free_capabilities;
633 free_capabilities = cap->link;
634 n_free_capabilities--;
639 cap->rCurrentTSO = t;
641 /* set the context_switch flag
643 if (run_queue_hd == END_TSO_QUEUE)
648 RELEASE_LOCK(&sched_mutex);
650 IF_DEBUG(scheduler,sched_belch("running thread %d", t->id));
652 /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
653 /* Run the current thread
655 switch (cap->rCurrentTSO->whatNext) {
658 /* Thread already finished, return to scheduler. */
659 ret = ThreadFinished;
662 ret = StgRun((StgFunPtr) stg_enterStackTop, cap);
665 ret = StgRun((StgFunPtr) stg_returnToStackTop, cap);
667 case ThreadEnterHugs:
671 IF_DEBUG(scheduler,sched_belch("entering Hugs"));
672 c = (StgClosure *)(cap->rCurrentTSO->sp[0]);
673 cap->rCurrentTSO->sp += 1;
678 barf("Panic: entered a BCO but no bytecode interpreter in this build");
681 barf("schedule: invalid whatNext field");
683 /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
685 /* Costs for the scheduler are assigned to CCS_SYSTEM */
690 ACQUIRE_LOCK(&sched_mutex);
693 IF_DEBUG(scheduler,fprintf(stderr,"scheduler (task %ld): ", pthread_self()););
695 IF_DEBUG(scheduler,fprintf(stderr,"scheduler: "););
697 t = cap->rCurrentTSO;
701 /* make all the running tasks block on a condition variable,
702 * maybe set context_switch and wait till they all pile in,
703 * then have them wait on a GC condition variable.
705 IF_DEBUG(scheduler,belch("thread %ld stopped: HeapOverflow", t->id));
708 ready_to_gc = rtsTrue;
709 context_switch = 1; /* stop other threads ASAP */
710 PUSH_ON_RUN_QUEUE(t);
714 /* just adjust the stack for this thread, then pop it back
717 IF_DEBUG(scheduler,belch("thread %ld stopped, StackOverflow", t->id));
721 /* enlarge the stack */
722 StgTSO *new_t = threadStackOverflow(t);
724 /* This TSO has moved, so update any pointers to it from the
725 * main thread stack. It better not be on any other queues...
728 for (m = main_threads; m != NULL; m = m->link) {
733 PUSH_ON_RUN_QUEUE(new_t);
740 DumpGranEvent(GR_DESCHEDULE, t));
741 globalGranStats.tot_yields++;
744 DumpGranEvent(GR_DESCHEDULE, t));
746 /* put the thread back on the run queue. Then, if we're ready to
747 * GC, check whether this is the last task to stop. If so, wake
748 * up the GC thread. getThread will block during a GC until the
752 if (t->whatNext == ThreadEnterHugs) {
753 /* ToDo: or maybe a timer expired when we were in Hugs?
754 * or maybe someone hit ctrl-C
756 belch("thread %ld stopped to switch to Hugs", t->id);
758 belch("thread %ld stopped, yielding", t->id);
762 APPEND_TO_RUN_QUEUE(t);
767 # error ToDo: implement GranSim scheduler
770 DumpGranEvent(GR_DESCHEDULE, t));
773 /* don't need to do anything. Either the thread is blocked on
774 * I/O, in which case we'll have called addToBlockedQueue
775 * previously, or it's blocked on an MVar or Blackhole, in which
776 * case it'll be on the relevant queue already.
779 fprintf(stderr, "thread %d stopped, ", t->id);
780 printThreadBlockage(t);
781 fprintf(stderr, "\n"));
786 /* Need to check whether this was a main thread, and if so, signal
787 * the task that started it with the return value. If we have no
788 * more main threads, we probably need to stop all the tasks until
791 IF_DEBUG(scheduler,belch("thread %ld finished", t->id));
792 t->whatNext = ThreadComplete;
794 // ToDo: endThread(t, CurrentProc); // clean-up the thread
796 advisory_thread_count--;
797 if (RtsFlags.ParFlags.ParStats.Full)
798 DumpEndEvent(CURRENT_PROC, t, rtsFalse /* not mandatory */);
803 barf("doneThread: invalid thread return code");
807 cap->link = free_capabilities;
808 free_capabilities = cap;
809 n_free_capabilities++;
813 if (ready_to_gc && n_free_capabilities == RtsFlags.ParFlags.nNodes)
818 /* everybody back, start the GC.
819 * Could do it in this thread, or signal a condition var
820 * to do it in another thread. Either way, we need to
821 * broadcast on gc_pending_cond afterward.
824 IF_DEBUG(scheduler,sched_belch("doing GC"));
826 GarbageCollect(GetRoots);
827 ready_to_gc = rtsFalse;
829 pthread_cond_broadcast(&gc_pending_cond);
834 IF_GRAN_DEBUG(unused,
835 print_eventq(EventHd));
837 event = get_next_event();
841 /* ToDo: wait for next message to arrive rather than busy wait */
846 t = take_off_run_queue(END_TSO_QUEUE);
849 } /* end of while(1) */
852 /* A hack for Hugs concurrency support. Needs sanitisation (?) */
853 void deleteAllThreads ( void )
856 IF_DEBUG(scheduler,sched_belch("deleteAllThreads()"));
857 for (t = run_queue_hd; t != END_TSO_QUEUE; t = t->link) {
860 for (t = blocked_queue_hd; t != END_TSO_QUEUE; t = t->link) {
863 run_queue_hd = run_queue_tl = END_TSO_QUEUE;
864 blocked_queue_hd = blocked_queue_tl = END_TSO_QUEUE;
867 /* startThread and insertThread are now in GranSim.c -- HWL */
869 //@node Suspend and Resume, Run queue code, Main scheduling loop, Main scheduling code
870 //@subsection Suspend and Resume
872 /* ---------------------------------------------------------------------------
873 * Suspending & resuming Haskell threads.
875 * When making a "safe" call to C (aka _ccall_GC), the task gives back
876 * its capability before calling the C function. This allows another
877 * task to pick up the capability and carry on running Haskell
878 * threads. It also means that if the C call blocks, it won't lock
881 * The Haskell thread making the C call is put to sleep for the
882 * duration of the call, on the susepended_ccalling_threads queue. We
883 * give out a token to the task, which it can use to resume the thread
884 * on return from the C function.
885 * ------------------------------------------------------------------------- */
888 suspendThread( Capability *cap )
892 ACQUIRE_LOCK(&sched_mutex);
895 sched_belch("thread %d did a _ccall_gc\n", cap->rCurrentTSO->id));
897 threadPaused(cap->rCurrentTSO);
898 cap->rCurrentTSO->link = suspended_ccalling_threads;
899 suspended_ccalling_threads = cap->rCurrentTSO;
901 /* Use the thread ID as the token; it should be unique */
902 tok = cap->rCurrentTSO->id;
905 cap->link = free_capabilities;
906 free_capabilities = cap;
907 n_free_capabilities++;
910 RELEASE_LOCK(&sched_mutex);
915 resumeThread( StgInt tok )
920 ACQUIRE_LOCK(&sched_mutex);
922 prev = &suspended_ccalling_threads;
923 for (tso = suspended_ccalling_threads;
924 tso != END_TSO_QUEUE;
925 prev = &tso->link, tso = tso->link) {
926 if (tso->id == (StgThreadID)tok) {
931 if (tso == END_TSO_QUEUE) {
932 barf("resumeThread: thread not found");
936 while (free_capabilities == NULL) {
937 IF_DEBUG(scheduler, sched_belch("waiting to resume"));
938 pthread_cond_wait(&thread_ready_cond, &sched_mutex);
939 IF_DEBUG(scheduler, sched_belch("resuming thread %d", tso->id));
941 cap = free_capabilities;
942 free_capabilities = cap->link;
943 n_free_capabilities--;
948 cap->rCurrentTSO = tso;
950 RELEASE_LOCK(&sched_mutex);
955 /* ---------------------------------------------------------------------------
957 * ------------------------------------------------------------------------ */
958 static void unblockThread(StgTSO *tso);
960 /* ---------------------------------------------------------------------------
961 * Comparing Thread ids.
963 * This is used from STG land in the implementation of the
964 * instances of Eq/Ord for ThreadIds.
965 * ------------------------------------------------------------------------ */
967 int cmp_thread(const StgTSO *tso1, const StgTSO *tso2)
969 StgThreadID id1 = tso1->id;
970 StgThreadID id2 = tso2->id;
972 if (id1 < id2) return (-1);
973 if (id1 > id2) return 1;
977 /* ---------------------------------------------------------------------------
980 The new thread starts with the given stack size. Before the
981 scheduler can run, however, this thread needs to have a closure
982 (and possibly some arguments) pushed on its stack. See
983 pushClosure() in Schedule.h.
985 createGenThread() and createIOThread() (in SchedAPI.h) are
986 convenient packaged versions of this function.
987 ------------------------------------------------------------------------ */
988 //@cindex createThread
990 /* currently pri (priority) is only used in a GRAN setup -- HWL */
992 createThread(nat stack_size, StgInt pri)
994 return createThread_(stack_size, rtsFalse, pri);
998 createThread_(nat size, rtsBool have_lock, StgInt pri)
1002 createThread(nat stack_size)
1004 return createThread_(stack_size, rtsFalse);
1008 createThread_(nat size, rtsBool have_lock)
1014 /* First check whether we should create a thread at all */
1016 /* check that no more than RtsFlags.ParFlags.maxThreads threads are created */
1017 if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads) {
1019 belch("{createThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
1020 RtsFlags.ParFlags.maxThreads, advisory_thread_count);
1021 return END_TSO_QUEUE;
1027 ASSERT(!RtsFlags.GranFlags.Light || CurrentProc==0);
1030 // ToDo: check whether size = stack_size - TSO_STRUCT_SIZEW
1032 /* catch ridiculously small stack sizes */
1033 if (size < MIN_STACK_WORDS + TSO_STRUCT_SIZEW) {
1034 size = MIN_STACK_WORDS + TSO_STRUCT_SIZEW;
1037 tso = (StgTSO *)allocate(size);
1038 TICK_ALLOC_TSO(size-sizeofW(StgTSO),0);
1040 stack_size = size - TSO_STRUCT_SIZEW;
1042 // Hmm, this CCS_MAIN is not protected by a PROFILING cpp var;
1043 SET_HDR(tso, &TSO_info, CCS_MAIN);
1045 SET_GRAN_HDR(tso, ThisPE);
1047 tso->whatNext = ThreadEnterGHC;
1049 /* tso->id needs to be unique. For now we use a heavyweight mutex to
1050 protect the increment operation on next_thread_id.
1051 In future, we could use an atomic increment instead.
1054 if (!have_lock) { ACQUIRE_LOCK(&sched_mutex); }
1055 tso->id = next_thread_id++;
1056 if (!have_lock) { RELEASE_LOCK(&sched_mutex); }
1058 tso->why_blocked = NotBlocked;
1059 tso->blocked_exceptions = NULL;
1061 tso->splim = (P_)&(tso->stack) + RESERVED_STACK_WORDS;
1062 tso->stack_size = stack_size;
1063 tso->max_stack_size = round_to_mblocks(RtsFlags.GcFlags.maxStkSize)
1065 tso->sp = (P_)&(tso->stack) + stack_size;
1068 tso->prof.CCCS = CCS_MAIN;
1071 /* put a stop frame on the stack */
1072 tso->sp -= sizeofW(StgStopFrame);
1073 SET_HDR((StgClosure*)tso->sp,(StgInfoTable *)&stg_stop_thread_info,CCS_MAIN);
1074 tso->su = (StgUpdateFrame*)tso->sp;
1076 IF_DEBUG(scheduler,belch("---- Initialised TSO %ld (%p), stack size = %lx words",
1077 tso->id, tso, tso->stack_size));
1081 tso->link = END_TSO_QUEUE;
1082 /* uses more flexible routine in GranSim */
1083 insertThread(tso, CurrentProc);
1085 add_to_run_queue(tso);
1089 tso->gran.pri = pri;
1090 tso->gran.magic = TSO_MAGIC; // debugging only
1091 tso->gran.sparkname = 0;
1092 tso->gran.startedat = CURRENT_TIME;
1093 tso->gran.exported = 0;
1094 tso->gran.basicblocks = 0;
1095 tso->gran.allocs = 0;
1096 tso->gran.exectime = 0;
1097 tso->gran.fetchtime = 0;
1098 tso->gran.fetchcount = 0;
1099 tso->gran.blocktime = 0;
1100 tso->gran.blockcount = 0;
1101 tso->gran.blockedat = 0;
1102 tso->gran.globalsparks = 0;
1103 tso->gran.localsparks = 0;
1104 if (RtsFlags.GranFlags.Light)
1105 tso->gran.clock = Now; /* local clock */
1107 tso->gran.clock = 0;
1109 IF_DEBUG(gran,printTSO(tso));
1111 tso->par.sparkname = 0;
1112 tso->par.startedat = CURRENT_TIME;
1113 tso->par.exported = 0;
1114 tso->par.basicblocks = 0;
1115 tso->par.allocs = 0;
1116 tso->par.exectime = 0;
1117 tso->par.fetchtime = 0;
1118 tso->par.fetchcount = 0;
1119 tso->par.blocktime = 0;
1120 tso->par.blockcount = 0;
1121 tso->par.blockedat = 0;
1122 tso->par.globalsparks = 0;
1123 tso->par.localsparks = 0;
1127 globalGranStats.tot_threads_created++;
1128 globalGranStats.threads_created_on_PE[CurrentProc]++;
1129 globalGranStats.tot_sq_len += spark_queue_len(CurrentProc);
1130 globalGranStats.tot_sq_probes++;
1133 IF_DEBUG(scheduler,sched_belch("created thread %ld, stack size = %lx words",
1134 tso->id, tso->stack_size));
1138 /* ---------------------------------------------------------------------------
1141 * scheduleThread puts a thread on the head of the runnable queue.
1142 * This will usually be done immediately after a thread is created.
1143 * The caller of scheduleThread must create the thread using e.g.
1144 * createThread and push an appropriate closure
1145 * on this thread's stack before the scheduler is invoked.
1146 * ------------------------------------------------------------------------ */
1149 scheduleThread(StgTSO *tso)
1151 ACQUIRE_LOCK(&sched_mutex);
1153 /* Put the new thread on the head of the runnable queue. The caller
1154 * better push an appropriate closure on this thread's stack
1155 * beforehand. In the SMP case, the thread may start running as
1156 * soon as we release the scheduler lock below.
1158 PUSH_ON_RUN_QUEUE(tso);
1161 IF_DEBUG(scheduler,printTSO(tso));
1162 RELEASE_LOCK(&sched_mutex);
1165 /* ---------------------------------------------------------------------------
1168 * Start up Posix threads to run each of the scheduler tasks.
1169 * I believe the task ids are not needed in the system as defined.
1171 * ------------------------------------------------------------------------ */
1175 taskStart( void *arg STG_UNUSED )
1182 /* ---------------------------------------------------------------------------
1185 * Initialise the scheduler. This resets all the queues - if the
1186 * queues contained any threads, they'll be garbage collected at the
1189 * This now calls startTasks(), so should only be called once! KH @ 25/10/99
1190 * ------------------------------------------------------------------------ */
1194 term_handler(int sig STG_UNUSED)
1197 ACQUIRE_LOCK(&term_mutex);
1199 RELEASE_LOCK(&term_mutex);
1204 //@cindex initScheduler
1211 for (i=0; i<=MAX_PROC; i++) {
1212 run_queue_hds[i] = END_TSO_QUEUE;
1213 run_queue_tls[i] = END_TSO_QUEUE;
1214 blocked_queue_hds[i] = END_TSO_QUEUE;
1215 blocked_queue_tls[i] = END_TSO_QUEUE;
1216 ccalling_threadss[i] = END_TSO_QUEUE;
1219 run_queue_hd = END_TSO_QUEUE;
1220 run_queue_tl = END_TSO_QUEUE;
1221 blocked_queue_hd = END_TSO_QUEUE;
1222 blocked_queue_tl = END_TSO_QUEUE;
1225 suspended_ccalling_threads = END_TSO_QUEUE;
1227 main_threads = NULL;
1232 enteredCAFs = END_CAF_LIST;
1234 /* Install the SIGHUP handler */
1237 struct sigaction action,oact;
1239 action.sa_handler = term_handler;
1240 sigemptyset(&action.sa_mask);
1241 action.sa_flags = 0;
1242 if (sigaction(SIGTERM, &action, &oact) != 0) {
1243 barf("can't install TERM handler");
1249 /* Allocate N Capabilities */
1252 Capability *cap, *prev;
1255 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
1256 cap = stgMallocBytes(sizeof(Capability), "initScheduler:capabilities");
1260 free_capabilities = cap;
1261 n_free_capabilities = RtsFlags.ParFlags.nNodes;
1263 IF_DEBUG(scheduler,fprintf(stderr,"scheduler: Allocated %d capabilities\n",
1264 n_free_capabilities););
1267 #if defined(SMP) || defined(PAR)
1280 /* make some space for saving all the thread ids */
1281 task_ids = stgMallocBytes(RtsFlags.ParFlags.nNodes * sizeof(task_info),
1282 "initScheduler:task_ids");
1284 /* and create all the threads */
1285 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
1286 r = pthread_create(&tid,NULL,taskStart,NULL);
1288 barf("startTasks: Can't create new Posix thread");
1290 task_ids[i].id = tid;
1291 task_ids[i].mut_time = 0.0;
1292 task_ids[i].mut_etime = 0.0;
1293 task_ids[i].gc_time = 0.0;
1294 task_ids[i].gc_etime = 0.0;
1295 task_ids[i].elapsedtimestart = elapsedtime();
1296 IF_DEBUG(scheduler,fprintf(stderr,"scheduler: Started task: %ld\n",tid););
1302 exitScheduler( void )
1307 /* Don't want to use pthread_cancel, since we'd have to install
1308 * these silly exception handlers (pthread_cleanup_{push,pop}) around
1312 /* Cancel all our tasks */
1313 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
1314 pthread_cancel(task_ids[i].id);
1317 /* Wait for all the tasks to terminate */
1318 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
1319 IF_DEBUG(scheduler,fprintf(stderr,"scheduler: waiting for task %ld\n",
1321 pthread_join(task_ids[i].id, NULL);
1325 /* Send 'em all a SIGHUP. That should shut 'em up.
1327 await_death = RtsFlags.ParFlags.nNodes;
1328 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
1329 pthread_kill(task_ids[i].id,SIGTERM);
1331 while (await_death > 0) {
1337 /* -----------------------------------------------------------------------------
1338 Managing the per-task allocation areas.
1340 Each capability comes with an allocation area. These are
1341 fixed-length block lists into which allocation can be done.
1343 ToDo: no support for two-space collection at the moment???
1344 -------------------------------------------------------------------------- */
1346 /* -----------------------------------------------------------------------------
1347 * waitThread is the external interface for running a new computataion
1348 * and waiting for the result.
1350 * In the non-SMP case, we create a new main thread, push it on the
1351 * main-thread stack, and invoke the scheduler to run it. The
1352 * scheduler will return when the top main thread on the stack has
1353 * completed or died, and fill in the necessary fields of the
1354 * main_thread structure.
1356 * In the SMP case, we create a main thread as before, but we then
1357 * create a new condition variable and sleep on it. When our new
1358 * main thread has completed, we'll be woken up and the status/result
1359 * will be in the main_thread struct.
1360 * -------------------------------------------------------------------------- */
1363 waitThread(StgTSO *tso, /*out*/StgClosure **ret)
1366 SchedulerStatus stat;
1368 ACQUIRE_LOCK(&sched_mutex);
1370 m = stgMallocBytes(sizeof(StgMainThread), "waitThread");
1376 pthread_cond_init(&m->wakeup, NULL);
1379 m->link = main_threads;
1382 IF_DEBUG(scheduler, fprintf(stderr, "scheduler: new main thread (%d)\n",
1387 pthread_cond_wait(&m->wakeup, &sched_mutex);
1388 } while (m->stat == NoStatus);
1391 ASSERT(m->stat != NoStatus);
1397 pthread_cond_destroy(&m->wakeup);
1400 IF_DEBUG(scheduler, fprintf(stderr, "scheduler: main thread (%d) finished\n",
1404 RELEASE_LOCK(&sched_mutex);
1409 //@node Run queue code, Garbage Collextion Routines, Suspend and Resume, Main scheduling code
1410 //@subsection Run queue code
1414 NB: In GranSim we have many run queues; run_queue_hd is actually a macro
1415 unfolding to run_queue_hds[CurrentProc], thus CurrentProc is an
1416 implicit global variable that has to be correct when calling these
1420 /* Put the new thread on the head of the runnable queue.
1421 * The caller of createThread better push an appropriate closure
1422 * on this thread's stack before the scheduler is invoked.
1424 static /* inline */ void
1425 add_to_run_queue(tso)
1428 ASSERT(tso!=run_queue_hd && tso!=run_queue_tl);
1429 tso->link = run_queue_hd;
1431 if (run_queue_tl == END_TSO_QUEUE) {
1436 /* Put the new thread at the end of the runnable queue. */
1437 static /* inline */ void
1438 push_on_run_queue(tso)
1441 ASSERT(get_itbl((StgClosure *)tso)->type == TSO);
1442 ASSERT(run_queue_hd!=NULL && run_queue_tl!=NULL);
1443 ASSERT(tso!=run_queue_hd && tso!=run_queue_tl);
1444 if (run_queue_hd == END_TSO_QUEUE) {
1447 run_queue_tl->link = tso;
1453 Should be inlined because it's used very often in schedule. The tso
1454 argument is actually only needed in GranSim, where we want to have the
1455 possibility to schedule *any* TSO on the run queue, irrespective of the
1456 actual ordering. Therefore, if tso is not the nil TSO then we traverse
1457 the run queue and dequeue the tso, adjusting the links in the queue.
1459 //@cindex take_off_run_queue
1460 static /* inline */ StgTSO*
1461 take_off_run_queue(StgTSO *tso) {
1465 qetlaHbogh Qu' ngaSbogh ghomDaQ {tso} yIteq!
1467 if tso is specified, unlink that tso from the run_queue (doesn't have
1468 to be at the beginning of the queue); GranSim only
1470 if (tso!=END_TSO_QUEUE) {
1471 /* find tso in queue */
1472 for (t=run_queue_hd, prev=END_TSO_QUEUE;
1473 t!=END_TSO_QUEUE && t!=tso;
1477 /* now actually dequeue the tso */
1478 if (prev!=END_TSO_QUEUE) {
1479 ASSERT(run_queue_hd!=t);
1480 prev->link = t->link;
1482 /* t is at beginning of thread queue */
1483 ASSERT(run_queue_hd==t);
1484 run_queue_hd = t->link;
1486 /* t is at end of thread queue */
1487 if (t->link==END_TSO_QUEUE) {
1488 ASSERT(t==run_queue_tl);
1489 run_queue_tl = prev;
1491 ASSERT(run_queue_tl!=t);
1493 t->link = END_TSO_QUEUE;
1495 /* take tso from the beginning of the queue; std concurrent code */
1497 if (t != END_TSO_QUEUE) {
1498 run_queue_hd = t->link;
1499 t->link = END_TSO_QUEUE;
1500 if (run_queue_hd == END_TSO_QUEUE) {
1501 run_queue_tl = END_TSO_QUEUE;
1516 for (i=0, tso=run_queue_hd;
1517 tso != END_TSO_QUEUE;
1525 //@node Garbage Collextion Routines, Blocking Queue Routines, Run queue code, Main scheduling code
1526 //@subsection Garbage Collextion Routines
1528 /* ---------------------------------------------------------------------------
1529 Where are the roots that we know about?
1531 - all the threads on the runnable queue
1532 - all the threads on the blocked queue
1533 - all the thread currently executing a _ccall_GC
1534 - all the "main threads"
1536 ------------------------------------------------------------------------ */
1538 /* This has to be protected either by the scheduler monitor, or by the
1539 garbage collection monitor (probably the latter).
1543 static void GetRoots(void)
1549 for (i=0; i<=RtsFlags.GranFlags.proc; i++) {
1550 if ((run_queue_hds[i] != END_TSO_QUEUE) && ((run_queue_hds[i] != NULL)))
1551 run_queue_hds[i] = (StgTSO *)MarkRoot((StgClosure *)run_queue_hds[i]);
1552 if ((run_queue_tls[i] != END_TSO_QUEUE) && ((run_queue_tls[i] != NULL)))
1553 run_queue_tls[i] = (StgTSO *)MarkRoot((StgClosure *)run_queue_tls[i]);
1555 if ((blocked_queue_hds[i] != END_TSO_QUEUE) && ((blocked_queue_hds[i] != NULL)))
1556 blocked_queue_hds[i] = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_hds[i]);
1557 if ((blocked_queue_tls[i] != END_TSO_QUEUE) && ((blocked_queue_tls[i] != NULL)))
1558 blocked_queue_tls[i] = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_tls[i]);
1559 if ((ccalling_threadss[i] != END_TSO_QUEUE) && ((ccalling_threadss[i] != NULL)))
1560 ccalling_threadss[i] = (StgTSO *)MarkRoot((StgClosure *)ccalling_threadss[i]);
1565 run_queue_hd = (StgTSO *)MarkRoot((StgClosure *)run_queue_hd);
1566 run_queue_tl = (StgTSO *)MarkRoot((StgClosure *)run_queue_tl);
1567 blocked_queue_hd = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_hd);
1568 blocked_queue_tl = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_tl);
1570 run_queue_hd = (StgTSO *)MarkRoot((StgClosure *)run_queue_hd);
1571 run_queue_tl = (StgTSO *)MarkRoot((StgClosure *)run_queue_tl);
1573 blocked_queue_hd = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_hd);
1574 blocked_queue_tl = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_tl);
1577 for (m = main_threads; m != NULL; m = m->link) {
1578 m->tso = (StgTSO *)MarkRoot((StgClosure *)m->tso);
1580 suspended_ccalling_threads =
1581 (StgTSO *)MarkRoot((StgClosure *)suspended_ccalling_threads);
1583 #if defined(SMP) || defined(PAR) || defined(GRAN)
1588 /* -----------------------------------------------------------------------------
1591 This is the interface to the garbage collector from Haskell land.
1592 We provide this so that external C code can allocate and garbage
1593 collect when called from Haskell via _ccall_GC.
1595 It might be useful to provide an interface whereby the programmer
1596 can specify more roots (ToDo).
1598 This needs to be protected by the GC condition variable above. KH.
1599 -------------------------------------------------------------------------- */
1601 void (*extra_roots)(void);
1606 GarbageCollect(GetRoots);
1612 GetRoots(); /* the scheduler's roots */
1613 extra_roots(); /* the user's roots */
1617 performGCWithRoots(void (*get_roots)(void))
1619 extra_roots = get_roots;
1621 GarbageCollect(AllRoots);
1624 /* -----------------------------------------------------------------------------
1627 If the thread has reached its maximum stack size,
1628 then bomb out. Otherwise relocate the TSO into a larger chunk of
1629 memory and adjust its stack size appropriately.
1630 -------------------------------------------------------------------------- */
1633 threadStackOverflow(StgTSO *tso)
1635 nat new_stack_size, new_tso_size, diff, stack_words;
1639 if (tso->stack_size >= tso->max_stack_size) {
1641 /* If we're debugging, just print out the top of the stack */
1642 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
1646 fprintf(stderr, "fatal: stack overflow in Hugs; aborting\n" );
1649 /* Send this thread the StackOverflow exception */
1650 raiseAsync(tso, (StgClosure *)&stackOverflow_closure);
1655 /* Try to double the current stack size. If that takes us over the
1656 * maximum stack size for this thread, then use the maximum instead.
1657 * Finally round up so the TSO ends up as a whole number of blocks.
1659 new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
1660 new_tso_size = (nat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) +
1661 TSO_STRUCT_SIZE)/sizeof(W_);
1662 new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */
1663 new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
1665 IF_DEBUG(scheduler, fprintf(stderr,"scheduler: increasing stack size from %d words to %d.\n", tso->stack_size, new_stack_size));
1667 dest = (StgTSO *)allocate(new_tso_size);
1668 TICK_ALLOC_TSO(new_tso_size-sizeofW(StgTSO),0);
1670 /* copy the TSO block and the old stack into the new area */
1671 memcpy(dest,tso,TSO_STRUCT_SIZE);
1672 stack_words = tso->stack + tso->stack_size - tso->sp;
1673 new_sp = (P_)dest + new_tso_size - stack_words;
1674 memcpy(new_sp, tso->sp, stack_words * sizeof(W_));
1676 /* relocate the stack pointers... */
1677 diff = (P_)new_sp - (P_)tso->sp; /* In *words* */
1678 dest->su = (StgUpdateFrame *) ((P_)dest->su + diff);
1680 dest->splim = (P_)dest->splim + (nat)((P_)dest - (P_)tso);
1681 dest->stack_size = new_stack_size;
1683 /* and relocate the update frame list */
1684 relocate_TSO(tso, dest);
1686 /* Mark the old one as dead so we don't try to scavenge it during
1687 * garbage collection (the TSO will likely be on a mutables list in
1688 * some generation, but it'll get collected soon enough). It's
1689 * important to set the sp and su values to just beyond the end of
1690 * the stack, so we don't attempt to scavenge any part of the dead
1693 tso->whatNext = ThreadKilled;
1694 tso->sp = (P_)&(tso->stack[tso->stack_size]);
1695 tso->su = (StgUpdateFrame *)tso->sp;
1696 tso->why_blocked = NotBlocked;
1697 dest->mut_link = NULL;
1699 IF_DEBUG(sanity,checkTSO(tso));
1701 IF_DEBUG(scheduler,printTSO(dest));
1705 /* This will no longer work: KH */
1706 if (tso == MainTSO) { /* hack */
1713 //@node Blocking Queue Routines, Exception Handling Routines, Garbage Collextion Routines, Main scheduling code
1714 //@subsection Blocking Queue Routines
1716 /* ---------------------------------------------------------------------------
1717 Wake up a queue that was blocked on some resource.
1718 ------------------------------------------------------------------------ */
1720 // ToDo: check push_on_run_queue vs. PUSH_ON_RUN_QUEUE
1726 unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
1728 /* write RESUME events to log file and
1729 update blocked and fetch time (depending on type of the orig closure) */
1730 if (RtsFlags.ParFlags.ParStats.Full) {
1731 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
1732 GR_RESUMEQ, ((StgTSO *)bqe), ((StgTSO *)bqe)->block_info.closure,
1733 0, 0 /* spark_queue_len(ADVISORY_POOL) */);
1735 switch (get_itbl(node)->type) {
1737 ((StgTSO *)bqe)->par.fetchtime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
1742 ((StgTSO *)bqe)->par.blocktime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
1745 barf("{unblockOneLocked}Daq Qagh: unexpected closure in blocking queue");
1754 static StgBlockingQueueElement *
1755 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
1757 StgBlockingQueueElement *next;
1759 switch (get_itbl(bqe)->type) {
1761 ASSERT(((StgTSO *)bqe)->why_blocked != NotBlocked);
1762 /* if it's a TSO just push it onto the run_queue */
1764 // ((StgTSO *)bqe)->link = END_TSO_QUEUE; // debugging?
1765 PUSH_ON_RUN_QUEUE((StgTSO *)bqe);
1767 unblockCount(bqe, node);
1768 /* reset blocking status after dumping event */
1769 ((StgTSO *)bqe)->why_blocked = NotBlocked;
1773 /* if it's a BLOCKED_FETCH put it on the PendingFetches list */
1775 bqe->link = PendingFetches;
1776 PendingFetches = bqe;
1780 /* can ignore this case in a non-debugging setup;
1781 see comments on RBHSave closures above */
1783 /* check that the closure is an RBHSave closure */
1784 ASSERT(get_itbl((StgClosure *)bqe) == &RBH_Save_0_info ||
1785 get_itbl((StgClosure *)bqe) == &RBH_Save_1_info ||
1786 get_itbl((StgClosure *)bqe) == &RBH_Save_2_info);
1790 barf("{unblockOneLocked}Daq Qagh: Unexpected IP (%#lx; %s) in blocking queue at %#lx\n",
1791 get_itbl((StgClosure *)bqe), info_type((StgClosure *)bqe),
1795 // IF_DEBUG(scheduler,sched_belch("waking up thread %ld", tso->id));
1799 #else /* !GRAN && !PAR */
1801 unblockOneLocked(StgTSO *tso)
1805 ASSERT(get_itbl(tso)->type == TSO);
1806 ASSERT(tso->why_blocked != NotBlocked);
1807 tso->why_blocked = NotBlocked;
1809 PUSH_ON_RUN_QUEUE(tso);
1811 IF_DEBUG(scheduler,sched_belch("waking up thread %ld", tso->id));
1820 unblockOne(StgTSO *tso, StgClosure *node)
1822 ACQUIRE_LOCK(&sched_mutex);
1823 tso = unblockOneLocked(tso, node);
1824 RELEASE_LOCK(&sched_mutex);
1829 unblockOne(StgTSO *tso)
1831 ACQUIRE_LOCK(&sched_mutex);
1832 tso = unblockOneLocked(tso);
1833 RELEASE_LOCK(&sched_mutex);
1842 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
1844 StgBlockingQueueElement *bqe, *next;
1846 ACQUIRE_LOCK(&sched_mutex);
1848 IF_PAR_DEBUG(verbose,
1849 belch("## AwBQ for node %p on [%x]: ",
1852 ASSERT(get_itbl(q)->type == TSO ||
1853 get_itbl(q)->type == BLOCKED_FETCH ||
1854 get_itbl(q)->type == CONSTR);
1857 while (get_itbl(bqe)->type==TSO ||
1858 get_itbl(bqe)->type==BLOCKED_FETCH) {
1859 bqe = unblockOneLocked(bqe, node);
1861 RELEASE_LOCK(&sched_mutex);
1864 #else /* !GRAN && !PAR */
1866 awakenBlockedQueue(StgTSO *tso)
1868 ACQUIRE_LOCK(&sched_mutex);
1869 while (tso != END_TSO_QUEUE) {
1870 tso = unblockOneLocked(tso);
1872 RELEASE_LOCK(&sched_mutex);
1881 Awakening a blocking queue in GranSim means checking for each of the
1882 TSOs in the queue whether they are local or not, issuing a ResumeThread
1883 or an UnblockThread event, respectively. The basic iteration over the
1884 blocking queue is the same as in the standard setup.
1887 awaken_blocked_queue(StgBlockingQueueElement *q, StgClosure *node)
1889 StgBlockingQueueElement *bqe, *next;
1891 PEs node_loc, tso_loc;
1892 rtsTime bq_processing_time = 0;
1893 nat len = 0, len_local = 0;
1896 belch("## AwBQ for node %p on PE %d @ %ld by TSO %d (%p): ", \
1897 node, CurrentProc, CurrentTime[CurrentProc],
1898 CurrentTSO->id, CurrentTSO));
1900 node_loc = where_is(node);
1902 ASSERT(get_itbl(q)->type == TSO || // q is either a TSO or an RBHSave
1903 get_itbl(q)->type == CONSTR); // closure (type constructor)
1904 ASSERT(is_unique(node));
1906 /* FAKE FETCH: magically copy the node to the tso's proc;
1907 no Fetch necessary because in reality the node should not have been
1908 moved to the other PE in the first place
1910 if (CurrentProc!=node_loc) {
1912 belch("## node %p is on PE %d but CurrentProc is %d (TSO %d); assuming fake fetch and adjusting bitmask (old: %#x)",
1913 node, node_loc, CurrentProc, CurrentTSO->id,
1914 // CurrentTSO, where_is(CurrentTSO),
1915 node->header.gran.procs));
1916 node->header.gran.procs = (node->header.gran.procs) | PE_NUMBER(CurrentProc);
1918 belch("## new bitmask of node %p is %#x",
1919 node, node->header.gran.procs));
1920 if (RtsFlags.GranFlags.GranSimStats.Global) {
1921 globalGranStats.tot_fake_fetches++;
1926 // ToDo: check: ASSERT(CurrentProc==node_loc);
1927 while (get_itbl(next)->type==TSO) { // q != END_TSO_QUEUE) {
1931 bqe points to the current element in the queue
1932 next points to the next element in the queue
1934 tso = (StgTSO *)bqe; // wastes an assignment to get the type right
1935 tso_loc = where_is(tso);
1936 if (IS_LOCAL_TO(PROCS(node),tso_loc)) { // TSO is local
1937 /* !fake_fetch => TSO is on CurrentProc is same as IS_LOCAL_TO */
1938 ASSERT(CurrentProc!=node_loc || tso_loc==CurrentProc);
1939 bq_processing_time += RtsFlags.GranFlags.Costs.lunblocktime;
1940 // insertThread(tso, node_loc);
1941 new_event(tso_loc, tso_loc,
1942 CurrentTime[CurrentProc]+bq_processing_time,
1944 tso, node, (rtsSpark*)NULL);
1945 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
1948 } else { // TSO is remote (actually should be FMBQ)
1949 bq_processing_time += RtsFlags.GranFlags.Costs.mpacktime;
1950 bq_processing_time += RtsFlags.GranFlags.Costs.gunblocktime;
1951 new_event(tso_loc, CurrentProc,
1952 CurrentTime[CurrentProc]+bq_processing_time+
1953 RtsFlags.GranFlags.Costs.latency,
1955 tso, node, (rtsSpark*)NULL);
1956 tso->link = END_TSO_QUEUE; // overwrite link just to be sure
1957 bq_processing_time += RtsFlags.GranFlags.Costs.mtidytime;
1960 /* the thread-queue-overhead is accounted for in either Resume or UnblockThread */
1962 fprintf(stderr," %s TSO %d (%p) [PE %d] (blocked_on=%p) (next=%p) ,",
1963 (node_loc==tso_loc ? "Local" : "Global"),
1964 tso->id, tso, CurrentProc, tso->block_info.closure, tso->link))
1965 tso->block_info.closure = NULL;
1966 IF_DEBUG(scheduler,belch("-- Waking up thread %ld (%p)",
1970 /* if this is the BQ of an RBH, we have to put back the info ripped out of
1971 the closure to make room for the anchor of the BQ */
1972 if (next!=END_BQ_QUEUE) {
1973 ASSERT(get_itbl(node)->type == RBH && get_itbl(next)->type == CONSTR);
1975 ASSERT((info_ptr==&RBH_Save_0_info) ||
1976 (info_ptr==&RBH_Save_1_info) ||
1977 (info_ptr==&RBH_Save_2_info));
1979 /* cf. convertToRBH in RBH.c for writing the RBHSave closure */
1980 ((StgRBH *)node)->blocking_queue = ((StgRBHSave *)next)->payload[0];
1981 ((StgRBH *)node)->mut_link = ((StgRBHSave *)next)->payload[1];
1984 belch("## Filled in RBH_Save for %p (%s) at end of AwBQ",
1985 node, info_type(node)));
1988 /* statistics gathering */
1989 if (RtsFlags.GranFlags.GranSimStats.Global) {
1990 globalGranStats.tot_bq_processing_time += bq_processing_time;
1991 globalGranStats.tot_bq_len += len; // total length of all bqs awakened
1992 globalGranStats.tot_bq_len_local += len_local; // same for local TSOs only
1993 globalGranStats.tot_awbq++; // total no. of bqs awakened
1996 fprintf(stderr,"## BQ Stats of %p: [%d entries, %d local] %s\n",
1997 node, len, len_local, (next!=END_TSO_QUEUE) ? "RBH" : ""));
2003 Awakening a blocking queue in GUM has to check whether an entry in the
2004 queue is a normal TSO or a BLOCKED_FETCH. The later indicates that a TSO is
2005 waiting for the result of this computation on another PE. Thus, when
2006 finding a BLOCKED_FETCH we have to send off a message to that PE.
2007 Actually, we defer sending off a message, by just putting the BLOCKED_FETCH
2008 onto the PendingFetches queue, which will be later traversed by
2009 processFetches, sending off a RESUME message for each BLOCKED_FETCH.
2011 NB: There is no check for an RBHSave closure (type CONSTR) in the code
2012 below. The reason is, if we awaken the BQ of an RBH closure (RBHSaves
2013 only exist at the end of such BQs) we know that the closure has been
2014 unpacked successfully on the other PE, and we can discard the info
2015 contained in the RBHSave closure. The current closure will be turned
2016 into a FetchMe closure anyway.
2019 awaken_blocked_queue(StgBlockingQueueElement *q, StgClosure *node)
2021 StgBlockingQueueElement *bqe, *next;
2023 IF_PAR_DEBUG(verbose,
2024 belch("## AwBQ for node %p on [%x]: ",
2027 ASSERT(get_itbl(q)->type == TSO ||
2028 get_itbl(q)->type == BLOCKED_FETCH ||
2029 get_itbl(q)->type == CONSTR);
2032 while (get_itbl(next)->type==TSO ||
2033 get_itbl(next)->type==BLOCKED_FETCH) {
2035 switch (get_itbl(bqe)->type) {
2037 /* if it's a TSO just push it onto the run_queue */
2040 ((StgTSO *)bqe)->link = END_TSO_QUEUE; // debugging only
2042 push_on_run_queue((StgTSO *)bqe); // HWL: was: PUSH_ON_RUN_QUEUE(tso);
2044 /* write RESUME events to log file and
2045 update blocked and fetch time (depending on type of the orig closure) */
2046 if (RtsFlags.ParFlags.ParStats.Full) {
2047 DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
2048 GR_RESUMEQ, ((StgTSO *)bqe), ((StgTSO *)bqe)->block_info.closure,
2049 0, spark_queue_len(ADVISORY_POOL));
2051 switch (get_itbl(node)->type) {
2053 ((StgTSO *)bqe)->par.fetchtime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2058 ((StgTSO *)bqe)->par.blocktime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
2061 barf("{awaken_blocked_queue}Daq Qagh: unexpected closure %p (%s) with blocking queue",
2062 node, info_type(node));
2065 /* reset block_info.closure field after dumping event */
2066 ((StgTSO *)bqe)->block_info.closure = NULL;
2068 /* rest of this branch is debugging only */
2069 IF_PAR_DEBUG(verbose,
2070 fprintf(stderr," TSO %d (%p) [PE %lx] (block_info.closure=%p) (next=%p) ,",
2071 ((StgTSO *)bqe)->id, (StgTSO *)bqe,
2072 mytid, ((StgTSO *)bqe)->block_info.closure, ((StgTSO *)bqe)->link));
2075 if (!RtsFlags.ParFlags.Debug.verbose)
2076 belch("-- Waking up thread %ld (%p)",
2077 ((StgTSO *)bqe)->id, (StgTSO *)bqe));
2081 /* if it's a BLOCKED_FETCH put it on the PendingFetches list */
2083 bqe->link = PendingFetches;
2084 PendingFetches = bqe;
2085 // bqe.tso->block_info.closure = NULL;
2087 /* rest of this branch is debugging only */
2088 IF_PAR_DEBUG(verbose,
2089 fprintf(stderr," BLOCKED_FETCH (%p) on node %p [PE %lx] (next=%p) ,",
2090 ((StgBlockedFetch *)bqe),
2091 ((StgBlockedFetch *)bqe)->node,
2092 mytid, ((StgBlockedFetch *)bqe)->link));
2096 /* can ignore this case in a non-debugging setup;
2097 see comments on RBHSave closures above */
2099 /* check that the closure is an RBHSave closure */
2100 ASSERT(get_itbl((StgClosure *)bqe) == &RBH_Save_0_info ||
2101 get_itbl((StgClosure *)bqe) == &RBH_Save_1_info ||
2102 get_itbl((StgClosure *)bqe) == &RBH_Save_2_info);
2106 barf("{awaken_blocked_queue}Daq Qagh: Unexpected IP (%#lx; %s) in blocking queue at %#lx\n",
2107 get_itbl((StgClosure *)bqe), info_type((StgClosure *)bqe),
2114 #else /* !GRAN && !PAR */
2117 awaken_blocked_queue(StgTSO *q) { awakenBlockedQueue(q); }
2123 while (q != END_TSO_QUEUE) {
2124 ASSERT(get_itbl(q)->type == TSO);
2127 push_on_run_queue(tso); // HWL: was: PUSH_ON_RUN_QUEUE(tso);
2128 //tso->block_info.closure = NULL;
2129 IF_DEBUG(scheduler, belch("-- Waking up thread %ld (%p)", tso->id, tso));
2136 //@node Exception Handling Routines, Debugging Routines, Blocking Queue Routines, Main scheduling code
2137 //@subsection Exception Handling Routines
2139 /* ---------------------------------------------------------------------------
2141 - usually called inside a signal handler so it mustn't do anything fancy.
2142 ------------------------------------------------------------------------ */
2145 interruptStgRts(void)
2151 /* -----------------------------------------------------------------------------
2154 This is for use when we raise an exception in another thread, which
2156 This has nothing to do with the UnblockThread event in GranSim. -- HWL
2157 -------------------------------------------------------------------------- */
2160 unblockThread(StgTSO *tso)
2164 ACQUIRE_LOCK(&sched_mutex);
2165 switch (tso->why_blocked) {
2168 return; /* not blocked */
2171 ASSERT(get_itbl(tso->block_info.closure)->type == MVAR);
2173 StgTSO *last_tso = END_TSO_QUEUE;
2174 StgMVar *mvar = (StgMVar *)(tso->block_info.closure);
2177 for (t = mvar->head; t != END_TSO_QUEUE;
2178 last = &t->link, last_tso = t, t = t->link) {
2181 if (mvar->tail == tso) {
2182 mvar->tail = last_tso;
2187 barf("unblockThread (MVAR): TSO not found");
2190 case BlockedOnBlackHole:
2191 ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ);
2193 StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure);
2195 last = &bq->blocking_queue;
2196 for (t = bq->blocking_queue; t != END_TSO_QUEUE;
2197 last = &t->link, t = t->link) {
2203 barf("unblockThread (BLACKHOLE): TSO not found");
2206 case BlockedOnException:
2208 StgTSO *target = tso->block_info.tso;
2210 ASSERT(get_itbl(target)->type == TSO);
2211 ASSERT(target->blocked_exceptions != NULL);
2213 last = &target->blocked_exceptions;
2214 for (t = target->blocked_exceptions; t != END_TSO_QUEUE;
2215 last = &t->link, t = t->link) {
2216 ASSERT(get_itbl(t)->type == TSO);
2222 barf("unblockThread (Exception): TSO not found");
2225 case BlockedOnDelay:
2227 case BlockedOnWrite:
2229 StgTSO *prev = NULL;
2230 for (t = blocked_queue_hd; t != END_TSO_QUEUE;
2231 prev = t, t = t->link) {
2234 blocked_queue_hd = t->link;
2235 if (blocked_queue_tl == t) {
2236 blocked_queue_tl = END_TSO_QUEUE;
2239 prev->link = t->link;
2240 if (blocked_queue_tl == t) {
2241 blocked_queue_tl = prev;
2247 barf("unblockThread (I/O): TSO not found");
2251 barf("unblockThread");
2255 tso->link = END_TSO_QUEUE;
2256 tso->why_blocked = NotBlocked;
2257 tso->block_info.closure = NULL;
2258 PUSH_ON_RUN_QUEUE(tso);
2259 RELEASE_LOCK(&sched_mutex);
2262 /* -----------------------------------------------------------------------------
2265 * The following function implements the magic for raising an
2266 * asynchronous exception in an existing thread.
2268 * We first remove the thread from any queue on which it might be
2269 * blocked. The possible blockages are MVARs and BLACKHOLE_BQs.
2271 * We strip the stack down to the innermost CATCH_FRAME, building
2272 * thunks in the heap for all the active computations, so they can
2273 * be restarted if necessary. When we reach a CATCH_FRAME, we build
2274 * an application of the handler to the exception, and push it on
2275 * the top of the stack.
2277 * How exactly do we save all the active computations? We create an
2278 * AP_UPD for every UpdateFrame on the stack. Entering one of these
2279 * AP_UPDs pushes everything from the corresponding update frame
2280 * upwards onto the stack. (Actually, it pushes everything up to the
2281 * next update frame plus a pointer to the next AP_UPD object.
2282 * Entering the next AP_UPD object pushes more onto the stack until we
2283 * reach the last AP_UPD object - at which point the stack should look
2284 * exactly as it did when we killed the TSO and we can continue
2285 * execution by entering the closure on top of the stack.
2287 * We can also kill a thread entirely - this happens if either (a) the
2288 * exception passed to raiseAsync is NULL, or (b) there's no
2289 * CATCH_FRAME on the stack. In either case, we strip the entire
2290 * stack and replace the thread with a zombie.
2292 * -------------------------------------------------------------------------- */
2295 deleteThread(StgTSO *tso)
2297 raiseAsync(tso,NULL);
2301 raiseAsync(StgTSO *tso, StgClosure *exception)
2303 StgUpdateFrame* su = tso->su;
2304 StgPtr sp = tso->sp;
2306 /* Thread already dead? */
2307 if (tso->whatNext == ThreadComplete || tso->whatNext == ThreadKilled) {
2311 IF_DEBUG(scheduler, sched_belch("raising exception in thread %ld.", tso->id));
2313 /* Remove it from any blocking queues */
2316 /* The stack freezing code assumes there's a closure pointer on
2317 * the top of the stack. This isn't always the case with compiled
2318 * code, so we have to push a dummy closure on the top which just
2319 * returns to the next return address on the stack.
2321 if ( LOOKS_LIKE_GHC_INFO((void*)*sp) ) {
2322 *(--sp) = (W_)&dummy_ret_closure;
2326 int words = ((P_)su - (P_)sp) - 1;
2330 /* If we find a CATCH_FRAME, and we've got an exception to raise,
2331 * then build PAP(handler,exception), and leave it on top of
2332 * the stack ready to enter.
2334 if (get_itbl(su)->type == CATCH_FRAME && exception != NULL) {
2335 StgCatchFrame *cf = (StgCatchFrame *)su;
2336 /* we've got an exception to raise, so let's pass it to the
2337 * handler in this frame.
2339 ap = (StgAP_UPD *)allocate(sizeofW(StgPAP) + 1);
2340 TICK_ALLOC_UPD_PAP(2,0);
2341 SET_HDR(ap,&PAP_info,cf->header.prof.ccs);
2344 ap->fun = cf->handler;
2345 ap->payload[0] = (P_)exception;
2347 /* sp currently points to the word above the CATCH_FRAME on the stack.
2349 sp += sizeofW(StgCatchFrame);
2352 /* Restore the blocked/unblocked state for asynchronous exceptions
2353 * at the CATCH_FRAME.
2355 * If exceptions were unblocked at the catch, arrange that they
2356 * are unblocked again after executing the handler by pushing an
2357 * unblockAsyncExceptions_ret stack frame.
2359 if (!cf->exceptions_blocked) {
2360 *(sp--) = (W_)&unblockAsyncExceptionszh_ret_info;
2363 /* Ensure that async exceptions are blocked when running the handler.
2365 if (tso->blocked_exceptions == NULL) {
2366 tso->blocked_exceptions = END_TSO_QUEUE;
2369 /* Put the newly-built PAP on top of the stack, ready to execute
2370 * when the thread restarts.
2374 tso->whatNext = ThreadEnterGHC;
2378 /* First build an AP_UPD consisting of the stack chunk above the
2379 * current update frame, with the top word on the stack as the
2382 ap = (StgAP_UPD *)allocate(AP_sizeW(words));
2387 ap->fun = (StgClosure *)sp[0];
2389 for(i=0; i < (nat)words; ++i) {
2390 ap->payload[i] = (P_)*sp++;
2393 switch (get_itbl(su)->type) {
2397 SET_HDR(ap,&AP_UPD_info,su->header.prof.ccs /* ToDo */);
2398 TICK_ALLOC_UP_THK(words+1,0);
2401 fprintf(stderr, "scheduler: Updating ");
2402 printPtr((P_)su->updatee);
2403 fprintf(stderr, " with ");
2404 printObj((StgClosure *)ap);
2407 /* Replace the updatee with an indirection - happily
2408 * this will also wake up any threads currently
2409 * waiting on the result.
2411 UPD_IND_NOLOCK(su->updatee,ap); /* revert the black hole */
2413 sp += sizeofW(StgUpdateFrame) -1;
2414 sp[0] = (W_)ap; /* push onto stack */
2420 StgCatchFrame *cf = (StgCatchFrame *)su;
2423 /* We want a PAP, not an AP_UPD. Fortunately, the
2424 * layout's the same.
2426 SET_HDR(ap,&PAP_info,su->header.prof.ccs /* ToDo */);
2427 TICK_ALLOC_UPD_PAP(words+1,0);
2429 /* now build o = FUN(catch,ap,handler) */
2430 o = (StgClosure *)allocate(sizeofW(StgClosure)+2);
2431 TICK_ALLOC_FUN(2,0);
2432 SET_HDR(o,&catch_info,su->header.prof.ccs /* ToDo */);
2433 o->payload[0] = (StgClosure *)ap;
2434 o->payload[1] = cf->handler;
2437 fprintf(stderr, "scheduler: Built ");
2438 printObj((StgClosure *)o);
2441 /* pop the old handler and put o on the stack */
2443 sp += sizeofW(StgCatchFrame) - 1;
2450 StgSeqFrame *sf = (StgSeqFrame *)su;
2453 SET_HDR(ap,&PAP_info,su->header.prof.ccs /* ToDo */);
2454 TICK_ALLOC_UPD_PAP(words+1,0);
2456 /* now build o = FUN(seq,ap) */
2457 o = (StgClosure *)allocate(sizeofW(StgClosure)+1);
2458 TICK_ALLOC_SE_THK(1,0);
2459 SET_HDR(o,&seq_info,su->header.prof.ccs /* ToDo */);
2460 payloadCPtr(o,0) = (StgClosure *)ap;
2463 fprintf(stderr, "scheduler: Built ");
2464 printObj((StgClosure *)o);
2467 /* pop the old handler and put o on the stack */
2469 sp += sizeofW(StgSeqFrame) - 1;
2475 /* We've stripped the entire stack, the thread is now dead. */
2476 sp += sizeofW(StgStopFrame) - 1;
2477 sp[0] = (W_)exception; /* save the exception */
2478 tso->whatNext = ThreadKilled;
2479 tso->su = (StgUpdateFrame *)(sp+1);
2490 //@node Debugging Routines, Index, Exception Handling Routines, Main scheduling code
2491 //@subsection Debugging Routines
2493 /* -----------------------------------------------------------------------------
2494 Debugging: why is a thread blocked
2495 -------------------------------------------------------------------------- */
2499 void printThreadBlockage(StgTSO *tso)
2501 switch (tso->why_blocked) {
2503 fprintf(stderr,"blocked on read from fd %d", tso->block_info.fd);
2505 case BlockedOnWrite:
2506 fprintf(stderr,"blocked on write to fd %d", tso->block_info.fd);
2508 case BlockedOnDelay:
2509 fprintf(stderr,"blocked on delay of %d ms", tso->block_info.delay);
2512 fprintf(stderr,"blocked on an MVar");
2514 case BlockedOnException:
2515 fprintf(stderr,"blocked on delivering an exception to thread %d",
2516 tso->block_info.tso->id);
2518 case BlockedOnBlackHole:
2519 fprintf(stderr,"blocked on a black hole");
2522 fprintf(stderr,"not blocked");
2526 fprintf(stderr,"blocked on global address");
2533 Print a whole blocking queue attached to node (debugging only).
2538 print_bq (StgClosure *node)
2540 StgBlockingQueueElement *bqe;
2544 fprintf(stderr,"## BQ of closure %p (%s): ",
2545 node, info_type(node));
2547 /* should cover all closures that may have a blocking queue */
2548 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
2549 get_itbl(node)->type == FETCH_ME_BQ ||
2550 get_itbl(node)->type == RBH);
2552 ASSERT(node!=(StgClosure*)NULL); // sanity check
2554 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
2556 for (bqe = ((StgBlockingQueue*)node)->blocking_queue, end = (bqe==END_BQ_QUEUE);
2557 !end; // iterate until bqe points to a CONSTR
2558 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) {
2559 ASSERT(bqe != END_BQ_QUEUE); // sanity check
2560 ASSERT(bqe != (StgTSO*)NULL); // sanity check
2561 /* types of closures that may appear in a blocking queue */
2562 ASSERT(get_itbl(bqe)->type == TSO ||
2563 get_itbl(bqe)->type == BLOCKED_FETCH ||
2564 get_itbl(bqe)->type == CONSTR);
2565 /* only BQs of an RBH end with an RBH_Save closure */
2566 ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
2568 switch (get_itbl(bqe)->type) {
2570 fprintf(stderr," TSO %d (%x),",
2571 ((StgTSO *)bqe)->id, ((StgTSO *)bqe));
2574 fprintf(stderr," BF (node=%p, ga=((%x, %d, %x)),",
2575 ((StgBlockedFetch *)bqe)->node,
2576 ((StgBlockedFetch *)bqe)->ga.payload.gc.gtid,
2577 ((StgBlockedFetch *)bqe)->ga.payload.gc.slot,
2578 ((StgBlockedFetch *)bqe)->ga.weight);
2581 fprintf(stderr," %s (IP %p),",
2582 (get_itbl(bqe) == &RBH_Save_0_info ? "RBH_Save_0" :
2583 get_itbl(bqe) == &RBH_Save_1_info ? "RBH_Save_1" :
2584 get_itbl(bqe) == &RBH_Save_2_info ? "RBH_Save_2" :
2585 "RBH_Save_?"), get_itbl(bqe));
2588 barf("Unexpected closure type %s in blocking queue of %p (%s)",
2589 info_type(bqe), node, info_type(node));
2593 fputc('\n', stderr);
2595 # elif defined(GRAN)
2597 print_bq (StgClosure *node)
2599 StgBlockingQueueElement *bqe;
2601 PEs node_loc, tso_loc;
2604 /* should cover all closures that may have a blocking queue */
2605 ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
2606 get_itbl(node)->type == FETCH_ME_BQ ||
2607 get_itbl(node)->type == RBH);
2609 ASSERT(node!=(StgClosure*)NULL); // sanity check
2610 node_loc = where_is(node);
2612 fprintf(stderr,"## BQ of closure %p (%s) on [PE %d]: ",
2613 node, info_type(node), node_loc);
2616 NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
2618 for (bqe = ((StgBlockingQueue*)node)->blocking_queue, end = (bqe==END_BQ_QUEUE);
2619 !end; // iterate until bqe points to a CONSTR
2620 end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) {
2621 ASSERT(bqe != END_BQ_QUEUE); // sanity check
2622 ASSERT(bqe != (StgTSO*)NULL); // sanity check
2623 /* types of closures that may appear in a blocking queue */
2624 ASSERT(get_itbl(bqe)->type == TSO ||
2625 get_itbl(bqe)->type == CONSTR);
2626 /* only BQs of an RBH end with an RBH_Save closure */
2627 ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
2629 tso_loc = where_is((StgClosure *)bqe);
2630 switch (get_itbl(bqe)->type) {
2632 fprintf(stderr," TSO %d (%x) on [PE %d],",
2633 ((StgTSO *)bqe)->id, ((StgTSO *)bqe), tso_loc);
2636 fprintf(stderr," %s (IP %p),",
2637 (get_itbl(bqe) == &RBH_Save_0_info ? "RBH_Save_0" :
2638 get_itbl(bqe) == &RBH_Save_1_info ? "RBH_Save_1" :
2639 get_itbl(bqe) == &RBH_Save_2_info ? "RBH_Save_2" :
2640 "RBH_Save_?"), get_itbl(bqe));
2643 barf("Unexpected closure type %s in blocking queue of %p (%s)",
2644 info_type(bqe), node, info_type(node));
2648 fputc('\n', stderr);
2652 Nice and easy: only TSOs on the blocking queue
2655 print_bq (StgClosure *node)
2659 ASSERT(node!=(StgClosure*)NULL); // sanity check
2660 for (tso = ((StgBlockingQueue*)node)->blocking_queue;
2661 tso != END_TSO_QUEUE;
2663 ASSERT(tso!=(StgTSO*)NULL && tso!=END_TSO_QUEUE); // sanity check
2664 ASSERT(get_itbl(tso)->type == TSO); // guess what, sanity check
2665 fprintf(stderr," TSO %d (%x),", tso->id, tso);
2667 fputc('\n', stderr);
2671 /* A debugging function used all over the place in GranSim and GUM.
2672 Dummy function in other setups.
2674 # if !defined(GRAN) && !defined(PAR)
2676 info_type(StgClosure *closure){
2681 info_type_by_ip(StgInfoTable *ip){
2687 sched_belch(char *s, ...)
2692 fprintf(stderr, "scheduler (task %ld): ", pthread_self());
2694 fprintf(stderr, "scheduler: ");
2696 vfprintf(stderr, s, ap);
2697 fprintf(stderr, "\n");
2702 //@node Index, , Debugging Routines, Main scheduling code
2706 //* MainRegTable:: @cindex\s-+MainRegTable
2707 //* StgMainThread:: @cindex\s-+StgMainThread
2708 //* awaken_blocked_queue:: @cindex\s-+awaken_blocked_queue
2709 //* blocked_queue_hd:: @cindex\s-+blocked_queue_hd
2710 //* blocked_queue_tl:: @cindex\s-+blocked_queue_tl
2711 //* context_switch:: @cindex\s-+context_switch
2712 //* createThread:: @cindex\s-+createThread
2713 //* free_capabilities:: @cindex\s-+free_capabilities
2714 //* gc_pending_cond:: @cindex\s-+gc_pending_cond
2715 //* initScheduler:: @cindex\s-+initScheduler
2716 //* interrupted:: @cindex\s-+interrupted
2717 //* n_free_capabilities:: @cindex\s-+n_free_capabilities
2718 //* next_thread_id:: @cindex\s-+next_thread_id
2719 //* print_bq:: @cindex\s-+print_bq
2720 //* run_queue_hd:: @cindex\s-+run_queue_hd
2721 //* run_queue_tl:: @cindex\s-+run_queue_tl
2722 //* sched_mutex:: @cindex\s-+sched_mutex
2723 //* schedule:: @cindex\s-+schedule
2724 //* take_off_run_queue:: @cindex\s-+take_off_run_queue
2725 //* task_ids:: @cindex\s-+task_ids
2726 //* term_mutex:: @cindex\s-+term_mutex
2727 //* thread_ready_cond:: @cindex\s-+thread_ready_cond