1 /* ---------------------------------------------------------------------------
3 * (c) The GHC Team, 1998-2006
5 * The scheduler and thread-related functionality
7 * --------------------------------------------------------------------------*/
9 #include "PosixSource.h"
10 #define KEEP_LOCKCLOSURE
15 #include "OSThreads.h"
20 #include "StgMiscClosures.h"
21 #include "Interpreter.h"
23 #include "RtsSignals.h"
29 #include "ThreadLabels.h"
30 #include "LdvProfile.h"
32 #include "Proftimer.h"
36 /* PARALLEL_HASKELL includes go here */
39 #include "Capability.h"
41 #include "AwaitEvent.h"
42 #if defined(mingw32_HOST_OS)
43 #include "win32/IOManager.h"
46 #include "RaiseAsync.h"
48 #include "ThrIOManager.h"
50 #ifdef HAVE_SYS_TYPES_H
51 #include <sys/types.h>
65 // Turn off inlining when debugging - it obfuscates things
68 # define STATIC_INLINE static
71 /* -----------------------------------------------------------------------------
73 * -------------------------------------------------------------------------- */
75 #if !defined(THREADED_RTS)
76 // Blocked/sleeping thrads
77 StgTSO *blocked_queue_hd = NULL;
78 StgTSO *blocked_queue_tl = NULL;
79 StgTSO *sleeping_queue = NULL; // perhaps replace with a hash table?
82 /* Threads blocked on blackholes.
83 * LOCK: sched_mutex+capability, or all capabilities
85 StgTSO *blackhole_queue = NULL;
87 /* The blackhole_queue should be checked for threads to wake up. See
88 * Schedule.h for more thorough comment.
89 * LOCK: none (doesn't matter if we miss an update)
91 rtsBool blackholes_need_checking = rtsFalse;
93 /* flag that tracks whether we have done any execution in this time slice.
94 * LOCK: currently none, perhaps we should lock (but needs to be
95 * updated in the fast path of the scheduler).
97 * NB. must be StgWord, we do xchg() on it.
99 volatile StgWord recent_activity = ACTIVITY_YES;
101 /* if this flag is set as well, give up execution
102 * LOCK: none (changes monotonically)
104 volatile StgWord sched_state = SCHED_RUNNING;
106 /* This is used in `TSO.h' and gcc 2.96 insists that this variable actually
107 * exists - earlier gccs apparently didn't.
113 * Set to TRUE when entering a shutdown state (via shutdownHaskellAndExit()) --
114 * in an MT setting, needed to signal that a worker thread shouldn't hang around
115 * in the scheduler when it is out of work.
117 rtsBool shutting_down_scheduler = rtsFalse;
120 * This mutex protects most of the global scheduler data in
121 * the THREADED_RTS runtime.
123 #if defined(THREADED_RTS)
127 #if !defined(mingw32_HOST_OS)
128 #define FORKPROCESS_PRIMOP_SUPPORTED
131 /* -----------------------------------------------------------------------------
132 * static function prototypes
133 * -------------------------------------------------------------------------- */
135 static Capability *schedule (Capability *initialCapability, Task *task);
138 // These function all encapsulate parts of the scheduler loop, and are
139 // abstracted only to make the structure and control flow of the
140 // scheduler clearer.
142 static void schedulePreLoop (void);
143 static void scheduleFindWork (Capability *cap);
144 #if defined(THREADED_RTS)
145 static void scheduleYield (Capability **pcap, Task *task);
147 static void scheduleStartSignalHandlers (Capability *cap);
148 static void scheduleCheckBlockedThreads (Capability *cap);
149 static void scheduleCheckWakeupThreads(Capability *cap USED_IF_NOT_THREADS);
150 static void scheduleCheckBlackHoles (Capability *cap);
151 static void scheduleDetectDeadlock (Capability *cap, Task *task);
152 static void schedulePushWork(Capability *cap, Task *task);
153 #if defined(PARALLEL_HASKELL)
154 static rtsBool scheduleGetRemoteWork(Capability *cap);
155 static void scheduleSendPendingMessages(void);
157 #if defined(PARALLEL_HASKELL) || defined(THREADED_RTS)
158 static void scheduleActivateSpark(Capability *cap);
160 static void schedulePostRunThread(Capability *cap, StgTSO *t);
161 static rtsBool scheduleHandleHeapOverflow( Capability *cap, StgTSO *t );
162 static void scheduleHandleStackOverflow( Capability *cap, Task *task,
164 static rtsBool scheduleHandleYield( Capability *cap, StgTSO *t,
165 nat prev_what_next );
166 static void scheduleHandleThreadBlocked( StgTSO *t );
167 static rtsBool scheduleHandleThreadFinished( Capability *cap, Task *task,
169 static rtsBool scheduleNeedHeapProfile(rtsBool ready_to_gc);
170 static Capability *scheduleDoGC(Capability *cap, Task *task,
171 rtsBool force_major);
173 static rtsBool checkBlackHoles(Capability *cap);
175 static StgTSO *threadStackOverflow(Capability *cap, StgTSO *tso);
176 static StgTSO *threadStackUnderflow(Task *task, StgTSO *tso);
178 static void deleteThread (Capability *cap, StgTSO *tso);
179 static void deleteAllThreads (Capability *cap);
181 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
182 static void deleteThread_(Capability *cap, StgTSO *tso);
186 static char *whatNext_strs[] = {
196 /* -----------------------------------------------------------------------------
197 * Putting a thread on the run queue: different scheduling policies
198 * -------------------------------------------------------------------------- */
201 addToRunQueue( Capability *cap, StgTSO *t )
203 #if defined(PARALLEL_HASKELL)
204 if (RtsFlags.ParFlags.doFairScheduling) {
205 // this does round-robin scheduling; good for concurrency
206 appendToRunQueue(cap,t);
208 // this does unfair scheduling; good for parallelism
209 pushOnRunQueue(cap,t);
212 // this does round-robin scheduling; good for concurrency
213 appendToRunQueue(cap,t);
217 /* ---------------------------------------------------------------------------
218 Main scheduling loop.
220 We use round-robin scheduling, each thread returning to the
221 scheduler loop when one of these conditions is detected:
224 * timer expires (thread yields)
230 In a GranSim setup this loop iterates over the global event queue.
231 This revolves around the global event queue, which determines what
232 to do next. Therefore, it's more complicated than either the
233 concurrent or the parallel (GUM) setup.
234 This version has been entirely removed (JB 2008/08).
237 GUM iterates over incoming messages.
238 It starts with nothing to do (thus CurrentTSO == END_TSO_QUEUE),
239 and sends out a fish whenever it has nothing to do; in-between
240 doing the actual reductions (shared code below) it processes the
241 incoming messages and deals with delayed operations
242 (see PendingFetches).
243 This is not the ugliest code you could imagine, but it's bloody close.
245 (JB 2008/08) This version was formerly indicated by a PP-Flag PAR,
246 now by PP-flag PARALLEL_HASKELL. The Eden RTS (in GHC-6.x) uses it,
247 as well as future GUM versions. This file has been refurbished to
248 only contain valid code, which is however incomplete, refers to
249 invalid includes etc.
251 ------------------------------------------------------------------------ */
254 schedule (Capability *initialCapability, Task *task)
258 StgThreadReturnCode ret;
259 #if defined(PARALLEL_HASKELL)
260 rtsBool receivedFinish = rtsFalse;
264 #if defined(THREADED_RTS)
265 rtsBool first = rtsTrue;
268 cap = initialCapability;
270 // Pre-condition: this task owns initialCapability.
271 // The sched_mutex is *NOT* held
272 // NB. on return, we still hold a capability.
274 debugTrace (DEBUG_sched,
275 "### NEW SCHEDULER LOOP (task: %p, cap: %p)",
276 task, initialCapability);
280 // -----------------------------------------------------------
281 // Scheduler loop starts here:
283 #if defined(PARALLEL_HASKELL)
284 #define TERMINATION_CONDITION (!receivedFinish)
286 #define TERMINATION_CONDITION rtsTrue
289 while (TERMINATION_CONDITION) {
291 // Check whether we have re-entered the RTS from Haskell without
292 // going via suspendThread()/resumeThread (i.e. a 'safe' foreign
294 if (cap->in_haskell) {
295 errorBelch("schedule: re-entered unsafely.\n"
296 " Perhaps a 'foreign import unsafe' should be 'safe'?");
297 stg_exit(EXIT_FAILURE);
300 // The interruption / shutdown sequence.
302 // In order to cleanly shut down the runtime, we want to:
303 // * make sure that all main threads return to their callers
304 // with the state 'Interrupted'.
305 // * clean up all OS threads assocated with the runtime
306 // * free all memory etc.
308 // So the sequence for ^C goes like this:
310 // * ^C handler sets sched_state := SCHED_INTERRUPTING and
311 // arranges for some Capability to wake up
313 // * all threads in the system are halted, and the zombies are
314 // placed on the run queue for cleaning up. We acquire all
315 // the capabilities in order to delete the threads, this is
316 // done by scheduleDoGC() for convenience (because GC already
317 // needs to acquire all the capabilities). We can't kill
318 // threads involved in foreign calls.
320 // * somebody calls shutdownHaskell(), which calls exitScheduler()
322 // * sched_state := SCHED_SHUTTING_DOWN
324 // * all workers exit when the run queue on their capability
325 // drains. All main threads will also exit when their TSO
326 // reaches the head of the run queue and they can return.
328 // * eventually all Capabilities will shut down, and the RTS can
331 // * We might be left with threads blocked in foreign calls,
332 // we should really attempt to kill these somehow (TODO);
334 switch (sched_state) {
337 case SCHED_INTERRUPTING:
338 debugTrace(DEBUG_sched, "SCHED_INTERRUPTING");
339 #if defined(THREADED_RTS)
340 discardSparksCap(cap);
342 /* scheduleDoGC() deletes all the threads */
343 cap = scheduleDoGC(cap,task,rtsFalse);
345 // after scheduleDoGC(), we must be shutting down. Either some
346 // other Capability did the final GC, or we did it above,
347 // either way we can fall through to the SCHED_SHUTTING_DOWN
349 ASSERT(sched_state == SCHED_SHUTTING_DOWN);
352 case SCHED_SHUTTING_DOWN:
353 debugTrace(DEBUG_sched, "SCHED_SHUTTING_DOWN");
354 // If we are a worker, just exit. If we're a bound thread
355 // then we will exit below when we've removed our TSO from
357 if (task->tso == NULL && emptyRunQueue(cap)) {
362 barf("sched_state: %d", sched_state);
365 scheduleFindWork(cap);
367 /* work pushing, currently relevant only for THREADED_RTS:
368 (pushes threads, wakes up idle capabilities for stealing) */
369 schedulePushWork(cap,task);
371 #if defined(PARALLEL_HASKELL)
372 /* since we perform a blocking receive and continue otherwise,
373 either we never reach here or we definitely have work! */
374 // from here: non-empty run queue
375 ASSERT(!emptyRunQueue(cap));
377 if (PacketsWaiting()) { /* now process incoming messages, if any
380 CAUTION: scheduleGetRemoteWork called
381 above, waits for messages as well! */
382 processMessages(cap, &receivedFinish);
384 #endif // PARALLEL_HASKELL: non-empty run queue!
386 scheduleDetectDeadlock(cap,task);
388 #if defined(THREADED_RTS)
389 cap = task->cap; // reload cap, it might have changed
392 // Normally, the only way we can get here with no threads to
393 // run is if a keyboard interrupt received during
394 // scheduleCheckBlockedThreads() or scheduleDetectDeadlock().
395 // Additionally, it is not fatal for the
396 // threaded RTS to reach here with no threads to run.
398 // win32: might be here due to awaitEvent() being abandoned
399 // as a result of a console event having been delivered.
401 #if defined(THREADED_RTS)
405 // // don't yield the first time, we want a chance to run this
406 // // thread for a bit, even if there are others banging at the
409 // ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
413 scheduleYield(&cap,task);
414 if (emptyRunQueue(cap)) continue; // look for work again
417 #if !defined(THREADED_RTS) && !defined(mingw32_HOST_OS)
418 if ( emptyRunQueue(cap) ) {
419 ASSERT(sched_state >= SCHED_INTERRUPTING);
424 // Get a thread to run
426 t = popRunQueue(cap);
428 // Sanity check the thread we're about to run. This can be
429 // expensive if there is lots of thread switching going on...
430 IF_DEBUG(sanity,checkTSO(t));
432 #if defined(THREADED_RTS)
433 // Check whether we can run this thread in the current task.
434 // If not, we have to pass our capability to the right task.
436 Task *bound = t->bound;
440 debugTrace(DEBUG_sched,
441 "### Running thread %lu in bound thread", (unsigned long)t->id);
442 // yes, the Haskell thread is bound to the current native thread
444 debugTrace(DEBUG_sched,
445 "### thread %lu bound to another OS thread", (unsigned long)t->id);
446 // no, bound to a different Haskell thread: pass to that thread
447 pushOnRunQueue(cap,t);
451 // The thread we want to run is unbound.
453 debugTrace(DEBUG_sched,
454 "### this OS thread cannot run thread %lu", (unsigned long)t->id);
455 // no, the current native thread is bound to a different
456 // Haskell thread, so pass it to any worker thread
457 pushOnRunQueue(cap,t);
464 // If we're shutting down, and this thread has not yet been
465 // killed, kill it now. This sometimes happens when a finalizer
466 // thread is created by the final GC, or a thread previously
467 // in a foreign call returns.
468 if (sched_state >= SCHED_INTERRUPTING &&
469 !(t->what_next == ThreadComplete || t->what_next == ThreadKilled)) {
473 /* context switches are initiated by the timer signal, unless
474 * the user specified "context switch as often as possible", with
477 if (RtsFlags.ConcFlags.ctxtSwitchTicks == 0
478 && !emptyThreadQueues(cap)) {
479 cap->context_switch = 1;
484 // CurrentTSO is the thread to run. t might be different if we
485 // loop back to run_thread, so make sure to set CurrentTSO after
487 cap->r.rCurrentTSO = t;
489 debugTrace(DEBUG_sched, "-->> running thread %ld %s ...",
490 (long)t->id, whatNext_strs[t->what_next]);
492 startHeapProfTimer();
494 // Check for exceptions blocked on this thread
495 maybePerformBlockedException (cap, t);
497 // ----------------------------------------------------------------------
498 // Run the current thread
500 ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
501 ASSERT(t->cap == cap);
502 ASSERT(t->bound ? t->bound->cap == cap : 1);
504 prev_what_next = t->what_next;
506 errno = t->saved_errno;
508 SetLastError(t->saved_winerror);
511 cap->in_haskell = rtsTrue;
515 #if defined(THREADED_RTS)
516 if (recent_activity == ACTIVITY_DONE_GC) {
517 // ACTIVITY_DONE_GC means we turned off the timer signal to
518 // conserve power (see #1623). Re-enable it here.
520 prev = xchg((P_)&recent_activity, ACTIVITY_YES);
521 if (prev == ACTIVITY_DONE_GC) {
525 recent_activity = ACTIVITY_YES;
529 switch (prev_what_next) {
533 /* Thread already finished, return to scheduler. */
534 ret = ThreadFinished;
540 r = StgRun((StgFunPtr) stg_returnToStackTop, &cap->r);
541 cap = regTableToCapability(r);
546 case ThreadInterpret:
547 cap = interpretBCO(cap);
552 barf("schedule: invalid what_next field");
555 cap->in_haskell = rtsFalse;
557 // The TSO might have moved, eg. if it re-entered the RTS and a GC
558 // happened. So find the new location:
559 t = cap->r.rCurrentTSO;
561 // We have run some Haskell code: there might be blackhole-blocked
562 // threads to wake up now.
563 // Lock-free test here should be ok, we're just setting a flag.
564 if ( blackhole_queue != END_TSO_QUEUE ) {
565 blackholes_need_checking = rtsTrue;
568 // And save the current errno in this thread.
569 // XXX: possibly bogus for SMP because this thread might already
570 // be running again, see code below.
571 t->saved_errno = errno;
573 // Similarly for Windows error code
574 t->saved_winerror = GetLastError();
577 #if defined(THREADED_RTS)
578 // If ret is ThreadBlocked, and this Task is bound to the TSO that
579 // blocked, we are in limbo - the TSO is now owned by whatever it
580 // is blocked on, and may in fact already have been woken up,
581 // perhaps even on a different Capability. It may be the case
582 // that task->cap != cap. We better yield this Capability
583 // immediately and return to normaility.
584 if (ret == ThreadBlocked) {
585 debugTrace(DEBUG_sched,
586 "--<< thread %lu (%s) stopped: blocked",
587 (unsigned long)t->id, whatNext_strs[t->what_next]);
592 ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
593 ASSERT(t->cap == cap);
595 // ----------------------------------------------------------------------
597 // Costs for the scheduler are assigned to CCS_SYSTEM
599 #if defined(PROFILING)
603 schedulePostRunThread(cap,t);
605 t = threadStackUnderflow(task,t);
607 ready_to_gc = rtsFalse;
611 ready_to_gc = scheduleHandleHeapOverflow(cap,t);
615 scheduleHandleStackOverflow(cap,task,t);
619 if (scheduleHandleYield(cap, t, prev_what_next)) {
620 // shortcut for switching between compiler/interpreter:
626 scheduleHandleThreadBlocked(t);
630 if (scheduleHandleThreadFinished(cap, task, t)) return cap;
631 ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
635 barf("schedule: invalid thread return code %d", (int)ret);
638 if (ready_to_gc || scheduleNeedHeapProfile(ready_to_gc)) {
639 cap = scheduleDoGC(cap,task,rtsFalse);
641 } /* end of while() */
644 /* ----------------------------------------------------------------------------
645 * Setting up the scheduler loop
646 * ------------------------------------------------------------------------- */
649 schedulePreLoop(void)
651 // initialisation for scheduler - what cannot go into initScheduler()
654 /* -----------------------------------------------------------------------------
657 * Search for work to do, and handle messages from elsewhere.
658 * -------------------------------------------------------------------------- */
661 scheduleFindWork (Capability *cap)
663 scheduleStartSignalHandlers(cap);
665 // Only check the black holes here if we've nothing else to do.
666 // During normal execution, the black hole list only gets checked
667 // at GC time, to avoid repeatedly traversing this possibly long
668 // list each time around the scheduler.
669 if (emptyRunQueue(cap)) { scheduleCheckBlackHoles(cap); }
671 scheduleCheckWakeupThreads(cap);
673 scheduleCheckBlockedThreads(cap);
675 #if defined(THREADED_RTS) || defined(PARALLEL_HASKELL)
676 if (emptyRunQueue(cap)) { scheduleActivateSpark(cap); }
679 #if defined(PARALLEL_HASKELL)
680 // if messages have been buffered...
681 scheduleSendPendingMessages();
684 #if defined(PARALLEL_HASKELL)
685 if (emptyRunQueue(cap)) {
686 receivedFinish = scheduleGetRemoteWork(cap);
687 continue; // a new round, (hopefully) with new work
689 in GUM, this a) sends out a FISH and returns IF no fish is
691 b) (blocking) awaits and receives messages
693 in Eden, this is only the blocking receive, as b) in GUM.
699 #if defined(THREADED_RTS)
700 STATIC_INLINE rtsBool
701 shouldYieldCapability (Capability *cap, Task *task)
703 // we need to yield this capability to someone else if..
704 // - another thread is initiating a GC
705 // - another Task is returning from a foreign call
706 // - the thread at the head of the run queue cannot be run
707 // by this Task (it is bound to another Task, or it is unbound
708 // and this task it bound).
709 return (waiting_for_gc ||
710 cap->returning_tasks_hd != NULL ||
711 (!emptyRunQueue(cap) && (task->tso == NULL
712 ? cap->run_queue_hd->bound != NULL
713 : cap->run_queue_hd->bound != task)));
716 // This is the single place where a Task goes to sleep. There are
717 // two reasons it might need to sleep:
718 // - there are no threads to run
719 // - we need to yield this Capability to someone else
720 // (see shouldYieldCapability())
722 // Careful: the scheduler loop is quite delicate. Make sure you run
723 // the tests in testsuite/concurrent (all ways) after modifying this,
724 // and also check the benchmarks in nofib/parallel for regressions.
727 scheduleYield (Capability **pcap, Task *task)
729 Capability *cap = *pcap;
731 // if we have work, and we don't need to give up the Capability, continue.
732 if (!shouldYieldCapability(cap,task) &&
733 (!emptyRunQueue(cap) ||
734 blackholes_need_checking ||
735 sched_state >= SCHED_INTERRUPTING))
738 // otherwise yield (sleep), and keep yielding if necessary.
740 yieldCapability(&cap,task);
742 while (shouldYieldCapability(cap,task));
744 // note there may still be no threads on the run queue at this
745 // point, the caller has to check.
752 /* -----------------------------------------------------------------------------
755 * Push work to other Capabilities if we have some.
756 * -------------------------------------------------------------------------- */
759 schedulePushWork(Capability *cap USED_IF_THREADS,
760 Task *task USED_IF_THREADS)
762 /* following code not for PARALLEL_HASKELL. I kept the call general,
763 future GUM versions might use pushing in a distributed setup */
764 #if defined(THREADED_RTS)
766 Capability *free_caps[n_capabilities], *cap0;
769 // migration can be turned off with +RTS -qg
770 if (!RtsFlags.ParFlags.migrate) return;
772 // Check whether we have more threads on our run queue, or sparks
773 // in our pool, that we could hand to another Capability.
774 if ((emptyRunQueue(cap) || cap->run_queue_hd->_link == END_TSO_QUEUE)
775 && sparkPoolSizeCap(cap) < 2) {
779 // First grab as many free Capabilities as we can.
780 for (i=0, n_free_caps=0; i < n_capabilities; i++) {
781 cap0 = &capabilities[i];
782 if (cap != cap0 && tryGrabCapability(cap0,task)) {
783 if (!emptyRunQueue(cap0) || cap->returning_tasks_hd != NULL) {
784 // it already has some work, we just grabbed it at
785 // the wrong moment. Or maybe it's deadlocked!
786 releaseCapability(cap0);
788 free_caps[n_free_caps++] = cap0;
793 // we now have n_free_caps free capabilities stashed in
794 // free_caps[]. Share our run queue equally with them. This is
795 // probably the simplest thing we could do; improvements we might
796 // want to do include:
798 // - giving high priority to moving relatively new threads, on
799 // the gournds that they haven't had time to build up a
800 // working set in the cache on this CPU/Capability.
802 // - giving low priority to moving long-lived threads
804 if (n_free_caps > 0) {
805 StgTSO *prev, *t, *next;
806 rtsBool pushed_to_all;
808 debugTrace(DEBUG_sched,
809 "cap %d: %s and %d free capabilities, sharing...",
811 (!emptyRunQueue(cap) && cap->run_queue_hd->_link != END_TSO_QUEUE)?
812 "excess threads on run queue":"sparks to share (>=2)",
816 pushed_to_all = rtsFalse;
818 if (cap->run_queue_hd != END_TSO_QUEUE) {
819 prev = cap->run_queue_hd;
821 prev->_link = END_TSO_QUEUE;
822 for (; t != END_TSO_QUEUE; t = next) {
824 t->_link = END_TSO_QUEUE;
825 if (t->what_next == ThreadRelocated
826 || t->bound == task // don't move my bound thread
827 || tsoLocked(t)) { // don't move a locked thread
828 setTSOLink(cap, prev, t);
830 } else if (i == n_free_caps) {
831 pushed_to_all = rtsTrue;
834 setTSOLink(cap, prev, t);
837 debugTrace(DEBUG_sched, "pushing thread %lu to capability %d", (unsigned long)t->id, free_caps[i]->no);
838 appendToRunQueue(free_caps[i],t);
839 if (t->bound) { t->bound->cap = free_caps[i]; }
840 t->cap = free_caps[i];
844 cap->run_queue_tl = prev;
848 /* JB I left this code in place, it would work but is not necessary */
850 // If there are some free capabilities that we didn't push any
851 // threads to, then try to push a spark to each one.
852 if (!pushed_to_all) {
854 // i is the next free capability to push to
855 for (; i < n_free_caps; i++) {
856 if (emptySparkPoolCap(free_caps[i])) {
857 spark = tryStealSpark(cap->sparks);
859 debugTrace(DEBUG_sched, "pushing spark %p to capability %d", spark, free_caps[i]->no);
860 newSpark(&(free_caps[i]->r), spark);
865 #endif /* SPARK_PUSHING */
867 // release the capabilities
868 for (i = 0; i < n_free_caps; i++) {
869 task->cap = free_caps[i];
870 releaseAndWakeupCapability(free_caps[i]);
873 task->cap = cap; // reset to point to our Capability.
875 #endif /* THREADED_RTS */
879 /* ----------------------------------------------------------------------------
880 * Start any pending signal handlers
881 * ------------------------------------------------------------------------- */
883 #if defined(RTS_USER_SIGNALS) && !defined(THREADED_RTS)
885 scheduleStartSignalHandlers(Capability *cap)
887 if (RtsFlags.MiscFlags.install_signal_handlers && signals_pending()) {
888 // safe outside the lock
889 startSignalHandlers(cap);
894 scheduleStartSignalHandlers(Capability *cap STG_UNUSED)
899 /* ----------------------------------------------------------------------------
900 * Check for blocked threads that can be woken up.
901 * ------------------------------------------------------------------------- */
904 scheduleCheckBlockedThreads(Capability *cap USED_IF_NOT_THREADS)
906 #if !defined(THREADED_RTS)
908 // Check whether any waiting threads need to be woken up. If the
909 // run queue is empty, and there are no other tasks running, we
910 // can wait indefinitely for something to happen.
912 if ( !emptyQueue(blocked_queue_hd) || !emptyQueue(sleeping_queue) )
914 awaitEvent( emptyRunQueue(cap) && !blackholes_need_checking );
920 /* ----------------------------------------------------------------------------
921 * Check for threads woken up by other Capabilities
922 * ------------------------------------------------------------------------- */
925 scheduleCheckWakeupThreads(Capability *cap USED_IF_THREADS)
927 #if defined(THREADED_RTS)
928 // Any threads that were woken up by other Capabilities get
929 // appended to our run queue.
930 if (!emptyWakeupQueue(cap)) {
931 ACQUIRE_LOCK(&cap->lock);
932 if (emptyRunQueue(cap)) {
933 cap->run_queue_hd = cap->wakeup_queue_hd;
934 cap->run_queue_tl = cap->wakeup_queue_tl;
936 setTSOLink(cap, cap->run_queue_tl, cap->wakeup_queue_hd);
937 cap->run_queue_tl = cap->wakeup_queue_tl;
939 cap->wakeup_queue_hd = cap->wakeup_queue_tl = END_TSO_QUEUE;
940 RELEASE_LOCK(&cap->lock);
945 /* ----------------------------------------------------------------------------
946 * Check for threads blocked on BLACKHOLEs that can be woken up
947 * ------------------------------------------------------------------------- */
949 scheduleCheckBlackHoles (Capability *cap)
951 if ( blackholes_need_checking ) // check without the lock first
953 ACQUIRE_LOCK(&sched_mutex);
954 if ( blackholes_need_checking ) {
955 blackholes_need_checking = rtsFalse;
956 // important that we reset the flag *before* checking the
957 // blackhole queue, otherwise we could get deadlock. This
958 // happens as follows: we wake up a thread that
959 // immediately runs on another Capability, blocks on a
960 // blackhole, and then we reset the blackholes_need_checking flag.
961 checkBlackHoles(cap);
963 RELEASE_LOCK(&sched_mutex);
967 /* ----------------------------------------------------------------------------
968 * Detect deadlock conditions and attempt to resolve them.
969 * ------------------------------------------------------------------------- */
972 scheduleDetectDeadlock (Capability *cap, Task *task)
975 #if defined(PARALLEL_HASKELL)
976 // ToDo: add deadlock detection in GUM (similar to THREADED_RTS) -- HWL
981 * Detect deadlock: when we have no threads to run, there are no
982 * threads blocked, waiting for I/O, or sleeping, and all the
983 * other tasks are waiting for work, we must have a deadlock of
986 if ( emptyThreadQueues(cap) )
988 #if defined(THREADED_RTS)
990 * In the threaded RTS, we only check for deadlock if there
991 * has been no activity in a complete timeslice. This means
992 * we won't eagerly start a full GC just because we don't have
993 * any threads to run currently.
995 if (recent_activity != ACTIVITY_INACTIVE) return;
998 debugTrace(DEBUG_sched, "deadlocked, forcing major GC...");
1000 // Garbage collection can release some new threads due to
1001 // either (a) finalizers or (b) threads resurrected because
1002 // they are unreachable and will therefore be sent an
1003 // exception. Any threads thus released will be immediately
1005 cap = scheduleDoGC (cap, task, rtsTrue/*force major GC*/);
1006 // when force_major == rtsTrue. scheduleDoGC sets
1007 // recent_activity to ACTIVITY_DONE_GC and turns off the timer
1010 if ( !emptyRunQueue(cap) ) return;
1012 #if defined(RTS_USER_SIGNALS) && !defined(THREADED_RTS)
1013 /* If we have user-installed signal handlers, then wait
1014 * for signals to arrive rather then bombing out with a
1017 if ( RtsFlags.MiscFlags.install_signal_handlers && anyUserHandlers() ) {
1018 debugTrace(DEBUG_sched,
1019 "still deadlocked, waiting for signals...");
1023 if (signals_pending()) {
1024 startSignalHandlers(cap);
1027 // either we have threads to run, or we were interrupted:
1028 ASSERT(!emptyRunQueue(cap) || sched_state >= SCHED_INTERRUPTING);
1034 #if !defined(THREADED_RTS)
1035 /* Probably a real deadlock. Send the current main thread the
1036 * Deadlock exception.
1039 switch (task->tso->why_blocked) {
1041 case BlockedOnBlackHole:
1042 case BlockedOnException:
1044 throwToSingleThreaded(cap, task->tso,
1045 (StgClosure *)nonTermination_closure);
1048 barf("deadlock: main thread blocked in a strange way");
1057 /* ----------------------------------------------------------------------------
1058 * Send pending messages (PARALLEL_HASKELL only)
1059 * ------------------------------------------------------------------------- */
1061 #if defined(PARALLEL_HASKELL)
1063 scheduleSendPendingMessages(void)
1066 # if defined(PAR) // global Mem.Mgmt., omit for now
1067 if (PendingFetches != END_BF_QUEUE) {
1072 if (RtsFlags.ParFlags.BufferTime) {
1073 // if we use message buffering, we must send away all message
1074 // packets which have become too old...
1080 /* ----------------------------------------------------------------------------
1081 * Activate spark threads (PARALLEL_HASKELL and THREADED_RTS)
1082 * ------------------------------------------------------------------------- */
1084 #if defined(PARALLEL_HASKELL) || defined(THREADED_RTS)
1086 scheduleActivateSpark(Capability *cap)
1090 createSparkThread(cap);
1091 debugTrace(DEBUG_sched, "creating a spark thread");
1094 #endif // PARALLEL_HASKELL || THREADED_RTS
1096 /* ----------------------------------------------------------------------------
1097 * Get work from a remote node (PARALLEL_HASKELL only)
1098 * ------------------------------------------------------------------------- */
1100 #if defined(PARALLEL_HASKELL)
1101 static rtsBool /* return value used in PARALLEL_HASKELL only */
1102 scheduleGetRemoteWork (Capability *cap STG_UNUSED)
1104 #if defined(PARALLEL_HASKELL)
1105 rtsBool receivedFinish = rtsFalse;
1107 // idle() , i.e. send all buffers, wait for work
1108 if (RtsFlags.ParFlags.BufferTime) {
1109 IF_PAR_DEBUG(verbose,
1110 debugBelch("...send all pending data,"));
1113 for (i=1; i<=nPEs; i++)
1114 sendImmediately(i); // send all messages away immediately
1118 /* this would be the place for fishing in GUM...
1120 if (no-earlier-fish-around)
1121 sendFish(choosePe());
1124 // Eden:just look for incoming messages (blocking receive)
1125 IF_PAR_DEBUG(verbose,
1126 debugBelch("...wait for incoming messages...\n"));
1127 processMessages(cap, &receivedFinish); // blocking receive...
1130 return receivedFinish;
1131 // reenter scheduling look after having received something
1133 #else /* !PARALLEL_HASKELL, i.e. THREADED_RTS */
1135 return rtsFalse; /* return value unused in THREADED_RTS */
1137 #endif /* PARALLEL_HASKELL */
1139 #endif // PARALLEL_HASKELL || THREADED_RTS
1141 /* ----------------------------------------------------------------------------
1142 * After running a thread...
1143 * ------------------------------------------------------------------------- */
1146 schedulePostRunThread (Capability *cap, StgTSO *t)
1148 // We have to be able to catch transactions that are in an
1149 // infinite loop as a result of seeing an inconsistent view of
1153 // [a,b] <- mapM readTVar [ta,tb]
1154 // when (a == b) loop
1156 // and a is never equal to b given a consistent view of memory.
1158 if (t -> trec != NO_TREC && t -> why_blocked == NotBlocked) {
1159 if (!stmValidateNestOfTransactions (t -> trec)) {
1160 debugTrace(DEBUG_sched | DEBUG_stm,
1161 "trec %p found wasting its time", t);
1163 // strip the stack back to the
1164 // ATOMICALLY_FRAME, aborting the (nested)
1165 // transaction, and saving the stack of any
1166 // partially-evaluated thunks on the heap.
1167 throwToSingleThreaded_(cap, t, NULL, rtsTrue);
1169 ASSERT(get_itbl((StgClosure *)t->sp)->type == ATOMICALLY_FRAME);
1173 /* some statistics gathering in the parallel case */
1176 /* -----------------------------------------------------------------------------
1177 * Handle a thread that returned to the scheduler with ThreadHeepOverflow
1178 * -------------------------------------------------------------------------- */
1181 scheduleHandleHeapOverflow( Capability *cap, StgTSO *t )
1183 // did the task ask for a large block?
1184 if (cap->r.rHpAlloc > BLOCK_SIZE) {
1185 // if so, get one and push it on the front of the nursery.
1189 blocks = (lnat)BLOCK_ROUND_UP(cap->r.rHpAlloc) / BLOCK_SIZE;
1191 debugTrace(DEBUG_sched,
1192 "--<< thread %ld (%s) stopped: requesting a large block (size %ld)\n",
1193 (long)t->id, whatNext_strs[t->what_next], blocks);
1195 // don't do this if the nursery is (nearly) full, we'll GC first.
1196 if (cap->r.rCurrentNursery->link != NULL ||
1197 cap->r.rNursery->n_blocks == 1) { // paranoia to prevent infinite loop
1198 // if the nursery has only one block.
1201 bd = allocGroup( blocks );
1203 cap->r.rNursery->n_blocks += blocks;
1205 // link the new group into the list
1206 bd->link = cap->r.rCurrentNursery;
1207 bd->u.back = cap->r.rCurrentNursery->u.back;
1208 if (cap->r.rCurrentNursery->u.back != NULL) {
1209 cap->r.rCurrentNursery->u.back->link = bd;
1211 #if !defined(THREADED_RTS)
1212 ASSERT(g0s0->blocks == cap->r.rCurrentNursery &&
1213 g0s0 == cap->r.rNursery);
1215 cap->r.rNursery->blocks = bd;
1217 cap->r.rCurrentNursery->u.back = bd;
1219 // initialise it as a nursery block. We initialise the
1220 // step, gen_no, and flags field of *every* sub-block in
1221 // this large block, because this is easier than making
1222 // sure that we always find the block head of a large
1223 // block whenever we call Bdescr() (eg. evacuate() and
1224 // isAlive() in the GC would both have to do this, at
1228 for (x = bd; x < bd + blocks; x++) {
1229 x->step = cap->r.rNursery;
1235 // This assert can be a killer if the app is doing lots
1236 // of large block allocations.
1237 IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
1239 // now update the nursery to point to the new block
1240 cap->r.rCurrentNursery = bd;
1242 // we might be unlucky and have another thread get on the
1243 // run queue before us and steal the large block, but in that
1244 // case the thread will just end up requesting another large
1246 pushOnRunQueue(cap,t);
1247 return rtsFalse; /* not actually GC'ing */
1251 debugTrace(DEBUG_sched,
1252 "--<< thread %ld (%s) stopped: HeapOverflow",
1253 (long)t->id, whatNext_strs[t->what_next]);
1255 if (cap->context_switch) {
1256 // Sometimes we miss a context switch, e.g. when calling
1257 // primitives in a tight loop, MAYBE_GC() doesn't check the
1258 // context switch flag, and we end up waiting for a GC.
1259 // See #1984, and concurrent/should_run/1984
1260 cap->context_switch = 0;
1261 addToRunQueue(cap,t);
1263 pushOnRunQueue(cap,t);
1266 /* actual GC is done at the end of the while loop in schedule() */
1269 /* -----------------------------------------------------------------------------
1270 * Handle a thread that returned to the scheduler with ThreadStackOverflow
1271 * -------------------------------------------------------------------------- */
1274 scheduleHandleStackOverflow (Capability *cap, Task *task, StgTSO *t)
1276 debugTrace (DEBUG_sched,
1277 "--<< thread %ld (%s) stopped, StackOverflow",
1278 (long)t->id, whatNext_strs[t->what_next]);
1280 /* just adjust the stack for this thread, then pop it back
1284 /* enlarge the stack */
1285 StgTSO *new_t = threadStackOverflow(cap, t);
1287 /* The TSO attached to this Task may have moved, so update the
1290 if (task->tso == t) {
1293 pushOnRunQueue(cap,new_t);
1297 /* -----------------------------------------------------------------------------
1298 * Handle a thread that returned to the scheduler with ThreadYielding
1299 * -------------------------------------------------------------------------- */
1302 scheduleHandleYield( Capability *cap, StgTSO *t, nat prev_what_next )
1304 // Reset the context switch flag. We don't do this just before
1305 // running the thread, because that would mean we would lose ticks
1306 // during GC, which can lead to unfair scheduling (a thread hogs
1307 // the CPU because the tick always arrives during GC). This way
1308 // penalises threads that do a lot of allocation, but that seems
1309 // better than the alternative.
1310 cap->context_switch = 0;
1312 /* put the thread back on the run queue. Then, if we're ready to
1313 * GC, check whether this is the last task to stop. If so, wake
1314 * up the GC thread. getThread will block during a GC until the
1318 if (t->what_next != prev_what_next) {
1319 debugTrace(DEBUG_sched,
1320 "--<< thread %ld (%s) stopped to switch evaluators",
1321 (long)t->id, whatNext_strs[t->what_next]);
1323 debugTrace(DEBUG_sched,
1324 "--<< thread %ld (%s) stopped, yielding",
1325 (long)t->id, whatNext_strs[t->what_next]);
1330 //debugBelch("&& Doing sanity check on yielding TSO %ld.", t->id);
1332 ASSERT(t->_link == END_TSO_QUEUE);
1334 // Shortcut if we're just switching evaluators: don't bother
1335 // doing stack squeezing (which can be expensive), just run the
1337 if (t->what_next != prev_what_next) {
1341 addToRunQueue(cap,t);
1346 /* -----------------------------------------------------------------------------
1347 * Handle a thread that returned to the scheduler with ThreadBlocked
1348 * -------------------------------------------------------------------------- */
1351 scheduleHandleThreadBlocked( StgTSO *t
1352 #if !defined(GRAN) && !defined(DEBUG)
1358 // We don't need to do anything. The thread is blocked, and it
1359 // has tidied up its stack and placed itself on whatever queue
1360 // it needs to be on.
1362 // ASSERT(t->why_blocked != NotBlocked);
1363 // Not true: for example,
1364 // - in THREADED_RTS, the thread may already have been woken
1365 // up by another Capability. This actually happens: try
1366 // conc023 +RTS -N2.
1367 // - the thread may have woken itself up already, because
1368 // threadPaused() might have raised a blocked throwTo
1369 // exception, see maybePerformBlockedException().
1372 if (traceClass(DEBUG_sched)) {
1373 debugTraceBegin("--<< thread %lu (%s) stopped: ",
1374 (unsigned long)t->id, whatNext_strs[t->what_next]);
1375 printThreadBlockage(t);
1381 /* -----------------------------------------------------------------------------
1382 * Handle a thread that returned to the scheduler with ThreadFinished
1383 * -------------------------------------------------------------------------- */
1386 scheduleHandleThreadFinished (Capability *cap STG_UNUSED, Task *task, StgTSO *t)
1388 /* Need to check whether this was a main thread, and if so,
1389 * return with the return value.
1391 * We also end up here if the thread kills itself with an
1392 * uncaught exception, see Exception.cmm.
1394 debugTrace(DEBUG_sched, "--++ thread %lu (%s) finished",
1395 (unsigned long)t->id, whatNext_strs[t->what_next]);
1398 // Check whether the thread that just completed was a bound
1399 // thread, and if so return with the result.
1401 // There is an assumption here that all thread completion goes
1402 // through this point; we need to make sure that if a thread
1403 // ends up in the ThreadKilled state, that it stays on the run
1404 // queue so it can be dealt with here.
1409 if (t->bound != task) {
1410 #if !defined(THREADED_RTS)
1411 // Must be a bound thread that is not the topmost one. Leave
1412 // it on the run queue until the stack has unwound to the
1413 // point where we can deal with this. Leaving it on the run
1414 // queue also ensures that the garbage collector knows about
1415 // this thread and its return value (it gets dropped from the
1416 // step->threads list so there's no other way to find it).
1417 appendToRunQueue(cap,t);
1420 // this cannot happen in the threaded RTS, because a
1421 // bound thread can only be run by the appropriate Task.
1422 barf("finished bound thread that isn't mine");
1426 ASSERT(task->tso == t);
1428 if (t->what_next == ThreadComplete) {
1430 // NOTE: return val is tso->sp[1] (see StgStartup.hc)
1431 *(task->ret) = (StgClosure *)task->tso->sp[1];
1433 task->stat = Success;
1436 *(task->ret) = NULL;
1438 if (sched_state >= SCHED_INTERRUPTING) {
1439 task->stat = Interrupted;
1441 task->stat = Killed;
1445 removeThreadLabel((StgWord)task->tso->id);
1447 return rtsTrue; // tells schedule() to return
1453 /* -----------------------------------------------------------------------------
1454 * Perform a heap census
1455 * -------------------------------------------------------------------------- */
1458 scheduleNeedHeapProfile( rtsBool ready_to_gc STG_UNUSED )
1460 // When we have +RTS -i0 and we're heap profiling, do a census at
1461 // every GC. This lets us get repeatable runs for debugging.
1462 if (performHeapProfile ||
1463 (RtsFlags.ProfFlags.profileInterval==0 &&
1464 RtsFlags.ProfFlags.doHeapProfile && ready_to_gc)) {
1471 /* -----------------------------------------------------------------------------
1472 * Perform a garbage collection if necessary
1473 * -------------------------------------------------------------------------- */
1476 scheduleDoGC (Capability *cap, Task *task USED_IF_THREADS, rtsBool force_major)
1478 rtsBool heap_census;
1480 /* extern static volatile StgWord waiting_for_gc;
1481 lives inside capability.c */
1482 rtsBool gc_type, prev_pending_gc;
1486 if (sched_state == SCHED_SHUTTING_DOWN) {
1487 // The final GC has already been done, and the system is
1488 // shutting down. We'll probably deadlock if we try to GC
1494 if (sched_state < SCHED_INTERRUPTING
1495 && RtsFlags.ParFlags.parGcEnabled
1496 && N >= RtsFlags.ParFlags.parGcGen
1497 && ! oldest_gen->steps[0].mark)
1499 gc_type = PENDING_GC_PAR;
1501 gc_type = PENDING_GC_SEQ;
1504 // In order to GC, there must be no threads running Haskell code.
1505 // Therefore, the GC thread needs to hold *all* the capabilities,
1506 // and release them after the GC has completed.
1508 // This seems to be the simplest way: previous attempts involved
1509 // making all the threads with capabilities give up their
1510 // capabilities and sleep except for the *last* one, which
1511 // actually did the GC. But it's quite hard to arrange for all
1512 // the other tasks to sleep and stay asleep.
1515 /* Other capabilities are prevented from running yet more Haskell
1516 threads if waiting_for_gc is set. Tested inside
1517 yieldCapability() and releaseCapability() in Capability.c */
1519 prev_pending_gc = cas(&waiting_for_gc, 0, gc_type);
1520 if (prev_pending_gc) {
1522 debugTrace(DEBUG_sched, "someone else is trying to GC (%d)...",
1525 yieldCapability(&cap,task);
1526 } while (waiting_for_gc);
1527 return cap; // NOTE: task->cap might have changed here
1530 setContextSwitches();
1532 // The final shutdown GC is always single-threaded, because it's
1533 // possible that some of the Capabilities have no worker threads.
1535 if (gc_type == PENDING_GC_SEQ)
1537 // single-threaded GC: grab all the capabilities
1538 for (i=0; i < n_capabilities; i++) {
1539 debugTrace(DEBUG_sched, "ready_to_gc, grabbing all the capabilies (%d/%d)", i, n_capabilities);
1540 if (cap != &capabilities[i]) {
1541 Capability *pcap = &capabilities[i];
1542 // we better hope this task doesn't get migrated to
1543 // another Capability while we're waiting for this one.
1544 // It won't, because load balancing happens while we have
1545 // all the Capabilities, but even so it's a slightly
1546 // unsavoury invariant.
1548 waitForReturnCapability(&pcap, task);
1549 if (pcap != &capabilities[i]) {
1550 barf("scheduleDoGC: got the wrong capability");
1557 // multi-threaded GC: make sure all the Capabilities donate one
1559 debugTrace(DEBUG_sched, "ready_to_gc, grabbing GC threads");
1561 waitForGcThreads(cap);
1565 // so this happens periodically:
1566 if (cap) scheduleCheckBlackHoles(cap);
1568 IF_DEBUG(scheduler, printAllThreads());
1571 * We now have all the capabilities; if we're in an interrupting
1572 * state, then we should take the opportunity to delete all the
1573 * threads in the system.
1575 if (sched_state == SCHED_INTERRUPTING) {
1576 deleteAllThreads(cap);
1577 sched_state = SCHED_SHUTTING_DOWN;
1580 heap_census = scheduleNeedHeapProfile(rtsTrue);
1582 #if defined(THREADED_RTS)
1583 debugTrace(DEBUG_sched, "doing GC");
1584 // reset waiting_for_gc *before* GC, so that when the GC threads
1585 // emerge they don't immediately re-enter the GC.
1587 GarbageCollect(force_major || heap_census, gc_type, cap);
1589 GarbageCollect(force_major || heap_census, 0, cap);
1593 debugTrace(DEBUG_sched, "performing heap census");
1595 performHeapProfile = rtsFalse;
1600 Once we are all together... this would be the place to balance all
1601 spark pools. No concurrent stealing or adding of new sparks can
1602 occur. Should be defined in Sparks.c. */
1603 balanceSparkPoolsCaps(n_capabilities, capabilities);
1608 // We've just done a major GC and we don't need the timer
1609 // signal turned on any more (#1623).
1610 // NB. do this *before* releasing the Capabilities, to avoid
1612 recent_activity = ACTIVITY_DONE_GC;
1616 #if defined(THREADED_RTS)
1617 if (gc_type == PENDING_GC_SEQ) {
1618 // release our stash of capabilities.
1619 for (i = 0; i < n_capabilities; i++) {
1620 if (cap != &capabilities[i]) {
1621 task->cap = &capabilities[i];
1622 releaseCapability(&capabilities[i]);
1636 /* ---------------------------------------------------------------------------
1637 * Singleton fork(). Do not copy any running threads.
1638 * ------------------------------------------------------------------------- */
1641 forkProcess(HsStablePtr *entry
1642 #ifndef FORKPROCESS_PRIMOP_SUPPORTED
1647 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
1654 #if defined(THREADED_RTS)
1655 if (RtsFlags.ParFlags.nNodes > 1) {
1656 errorBelch("forking not supported with +RTS -N<n> greater than 1");
1657 stg_exit(EXIT_FAILURE);
1661 debugTrace(DEBUG_sched, "forking!");
1663 // ToDo: for SMP, we should probably acquire *all* the capabilities
1666 // no funny business: hold locks while we fork, otherwise if some
1667 // other thread is holding a lock when the fork happens, the data
1668 // structure protected by the lock will forever be in an
1669 // inconsistent state in the child. See also #1391.
1670 ACQUIRE_LOCK(&sched_mutex);
1671 ACQUIRE_LOCK(&cap->lock);
1672 ACQUIRE_LOCK(&cap->running_task->lock);
1676 if (pid) { // parent
1678 RELEASE_LOCK(&sched_mutex);
1679 RELEASE_LOCK(&cap->lock);
1680 RELEASE_LOCK(&cap->running_task->lock);
1682 // just return the pid
1688 #if defined(THREADED_RTS)
1689 initMutex(&sched_mutex);
1690 initMutex(&cap->lock);
1691 initMutex(&cap->running_task->lock);
1694 // Now, all OS threads except the thread that forked are
1695 // stopped. We need to stop all Haskell threads, including
1696 // those involved in foreign calls. Also we need to delete
1697 // all Tasks, because they correspond to OS threads that are
1700 for (s = 0; s < total_steps; s++) {
1701 for (t = all_steps[s].threads; t != END_TSO_QUEUE; t = next) {
1702 if (t->what_next == ThreadRelocated) {
1705 next = t->global_link;
1706 // don't allow threads to catch the ThreadKilled
1707 // exception, but we do want to raiseAsync() because these
1708 // threads may be evaluating thunks that we need later.
1709 deleteThread_(cap,t);
1714 // Empty the run queue. It seems tempting to let all the
1715 // killed threads stay on the run queue as zombies to be
1716 // cleaned up later, but some of them correspond to bound
1717 // threads for which the corresponding Task does not exist.
1718 cap->run_queue_hd = END_TSO_QUEUE;
1719 cap->run_queue_tl = END_TSO_QUEUE;
1721 // Any suspended C-calling Tasks are no more, their OS threads
1723 cap->suspended_ccalling_tasks = NULL;
1725 // Empty the threads lists. Otherwise, the garbage
1726 // collector may attempt to resurrect some of these threads.
1727 for (s = 0; s < total_steps; s++) {
1728 all_steps[s].threads = END_TSO_QUEUE;
1731 // Wipe the task list, except the current Task.
1732 ACQUIRE_LOCK(&sched_mutex);
1733 for (task = all_tasks; task != NULL; task=task->all_link) {
1734 if (task != cap->running_task) {
1735 #if defined(THREADED_RTS)
1736 initMutex(&task->lock); // see #1391
1741 RELEASE_LOCK(&sched_mutex);
1743 #if defined(THREADED_RTS)
1744 // Wipe our spare workers list, they no longer exist. New
1745 // workers will be created if necessary.
1746 cap->spare_workers = NULL;
1747 cap->returning_tasks_hd = NULL;
1748 cap->returning_tasks_tl = NULL;
1751 // On Unix, all timers are reset in the child, so we need to start
1756 cap = rts_evalStableIO(cap, entry, NULL); // run the action
1757 rts_checkSchedStatus("forkProcess",cap);
1760 hs_exit(); // clean up and exit
1761 stg_exit(EXIT_SUCCESS);
1763 #else /* !FORKPROCESS_PRIMOP_SUPPORTED */
1764 barf("forkProcess#: primop not supported on this platform, sorry!\n");
1769 /* ---------------------------------------------------------------------------
1770 * Delete all the threads in the system
1771 * ------------------------------------------------------------------------- */
1774 deleteAllThreads ( Capability *cap )
1776 // NOTE: only safe to call if we own all capabilities.
1781 debugTrace(DEBUG_sched,"deleting all threads");
1782 for (s = 0; s < total_steps; s++) {
1783 for (t = all_steps[s].threads; t != END_TSO_QUEUE; t = next) {
1784 if (t->what_next == ThreadRelocated) {
1787 next = t->global_link;
1788 deleteThread(cap,t);
1793 // The run queue now contains a bunch of ThreadKilled threads. We
1794 // must not throw these away: the main thread(s) will be in there
1795 // somewhere, and the main scheduler loop has to deal with it.
1796 // Also, the run queue is the only thing keeping these threads from
1797 // being GC'd, and we don't want the "main thread has been GC'd" panic.
1799 #if !defined(THREADED_RTS)
1800 ASSERT(blocked_queue_hd == END_TSO_QUEUE);
1801 ASSERT(sleeping_queue == END_TSO_QUEUE);
1805 /* -----------------------------------------------------------------------------
1806 Managing the suspended_ccalling_tasks list.
1807 Locks required: sched_mutex
1808 -------------------------------------------------------------------------- */
1811 suspendTask (Capability *cap, Task *task)
1813 ASSERT(task->next == NULL && task->prev == NULL);
1814 task->next = cap->suspended_ccalling_tasks;
1816 if (cap->suspended_ccalling_tasks) {
1817 cap->suspended_ccalling_tasks->prev = task;
1819 cap->suspended_ccalling_tasks = task;
1823 recoverSuspendedTask (Capability *cap, Task *task)
1826 task->prev->next = task->next;
1828 ASSERT(cap->suspended_ccalling_tasks == task);
1829 cap->suspended_ccalling_tasks = task->next;
1832 task->next->prev = task->prev;
1834 task->next = task->prev = NULL;
1837 /* ---------------------------------------------------------------------------
1838 * Suspending & resuming Haskell threads.
1840 * When making a "safe" call to C (aka _ccall_GC), the task gives back
1841 * its capability before calling the C function. This allows another
1842 * task to pick up the capability and carry on running Haskell
1843 * threads. It also means that if the C call blocks, it won't lock
1846 * The Haskell thread making the C call is put to sleep for the
1847 * duration of the call, on the susepended_ccalling_threads queue. We
1848 * give out a token to the task, which it can use to resume the thread
1849 * on return from the C function.
1850 * ------------------------------------------------------------------------- */
1853 suspendThread (StgRegTable *reg)
1860 StgWord32 saved_winerror;
1863 saved_errno = errno;
1865 saved_winerror = GetLastError();
1868 /* assume that *reg is a pointer to the StgRegTable part of a Capability.
1870 cap = regTableToCapability(reg);
1872 task = cap->running_task;
1873 tso = cap->r.rCurrentTSO;
1875 debugTrace(DEBUG_sched,
1876 "thread %lu did a safe foreign call",
1877 (unsigned long)cap->r.rCurrentTSO->id);
1879 // XXX this might not be necessary --SDM
1880 tso->what_next = ThreadRunGHC;
1882 threadPaused(cap,tso);
1884 if ((tso->flags & TSO_BLOCKEX) == 0) {
1885 tso->why_blocked = BlockedOnCCall;
1886 tso->flags |= TSO_BLOCKEX;
1887 tso->flags &= ~TSO_INTERRUPTIBLE;
1889 tso->why_blocked = BlockedOnCCall_NoUnblockExc;
1892 // Hand back capability
1893 task->suspended_tso = tso;
1895 ACQUIRE_LOCK(&cap->lock);
1897 suspendTask(cap,task);
1898 cap->in_haskell = rtsFalse;
1899 releaseCapability_(cap,rtsFalse);
1901 RELEASE_LOCK(&cap->lock);
1903 #if defined(THREADED_RTS)
1904 /* Preparing to leave the RTS, so ensure there's a native thread/task
1905 waiting to take over.
1907 debugTrace(DEBUG_sched, "thread %lu: leaving RTS", (unsigned long)tso->id);
1910 errno = saved_errno;
1912 SetLastError(saved_winerror);
1918 resumeThread (void *task_)
1925 StgWord32 saved_winerror;
1928 saved_errno = errno;
1930 saved_winerror = GetLastError();
1934 // Wait for permission to re-enter the RTS with the result.
1935 waitForReturnCapability(&cap,task);
1936 // we might be on a different capability now... but if so, our
1937 // entry on the suspended_ccalling_tasks list will also have been
1940 // Remove the thread from the suspended list
1941 recoverSuspendedTask(cap,task);
1943 tso = task->suspended_tso;
1944 task->suspended_tso = NULL;
1945 tso->_link = END_TSO_QUEUE; // no write barrier reqd
1946 debugTrace(DEBUG_sched, "thread %lu: re-entering RTS", (unsigned long)tso->id);
1948 if (tso->why_blocked == BlockedOnCCall) {
1949 awakenBlockedExceptionQueue(cap,tso);
1950 tso->flags &= ~(TSO_BLOCKEX | TSO_INTERRUPTIBLE);
1953 /* Reset blocking status */
1954 tso->why_blocked = NotBlocked;
1956 cap->r.rCurrentTSO = tso;
1957 cap->in_haskell = rtsTrue;
1958 errno = saved_errno;
1960 SetLastError(saved_winerror);
1963 /* We might have GC'd, mark the TSO dirty again */
1966 IF_DEBUG(sanity, checkTSO(tso));
1971 /* ---------------------------------------------------------------------------
1974 * scheduleThread puts a thread on the end of the runnable queue.
1975 * This will usually be done immediately after a thread is created.
1976 * The caller of scheduleThread must create the thread using e.g.
1977 * createThread and push an appropriate closure
1978 * on this thread's stack before the scheduler is invoked.
1979 * ------------------------------------------------------------------------ */
1982 scheduleThread(Capability *cap, StgTSO *tso)
1984 // The thread goes at the *end* of the run-queue, to avoid possible
1985 // starvation of any threads already on the queue.
1986 appendToRunQueue(cap,tso);
1990 scheduleThreadOn(Capability *cap, StgWord cpu USED_IF_THREADS, StgTSO *tso)
1992 #if defined(THREADED_RTS)
1993 tso->flags |= TSO_LOCKED; // we requested explicit affinity; don't
1994 // move this thread from now on.
1995 cpu %= RtsFlags.ParFlags.nNodes;
1996 if (cpu == cap->no) {
1997 appendToRunQueue(cap,tso);
1999 wakeupThreadOnCapability(cap, &capabilities[cpu], tso);
2002 appendToRunQueue(cap,tso);
2007 scheduleWaitThread (StgTSO* tso, /*[out]*/HaskellObj* ret, Capability *cap)
2011 // We already created/initialised the Task
2012 task = cap->running_task;
2014 // This TSO is now a bound thread; make the Task and TSO
2015 // point to each other.
2021 task->stat = NoStatus;
2023 appendToRunQueue(cap,tso);
2025 debugTrace(DEBUG_sched, "new bound thread (%lu)", (unsigned long)tso->id);
2027 cap = schedule(cap,task);
2029 ASSERT(task->stat != NoStatus);
2030 ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
2032 debugTrace(DEBUG_sched, "bound thread (%lu) finished", (unsigned long)task->tso->id);
2036 /* ----------------------------------------------------------------------------
2038 * ------------------------------------------------------------------------- */
2040 #if defined(THREADED_RTS)
2041 void OSThreadProcAttr
2042 workerStart(Task *task)
2046 // See startWorkerTask().
2047 ACQUIRE_LOCK(&task->lock);
2049 RELEASE_LOCK(&task->lock);
2051 // set the thread-local pointer to the Task:
2054 // schedule() runs without a lock.
2055 cap = schedule(cap,task);
2057 // On exit from schedule(), we have a Capability, but possibly not
2058 // the same one we started with.
2060 // During shutdown, the requirement is that after all the
2061 // Capabilities are shut down, all workers that are shutting down
2062 // have finished workerTaskStop(). This is why we hold on to
2063 // cap->lock until we've finished workerTaskStop() below.
2065 // There may be workers still involved in foreign calls; those
2066 // will just block in waitForReturnCapability() because the
2067 // Capability has been shut down.
2069 ACQUIRE_LOCK(&cap->lock);
2070 releaseCapability_(cap,rtsFalse);
2071 workerTaskStop(task);
2072 RELEASE_LOCK(&cap->lock);
2076 /* ---------------------------------------------------------------------------
2079 * Initialise the scheduler. This resets all the queues - if the
2080 * queues contained any threads, they'll be garbage collected at the
2083 * ------------------------------------------------------------------------ */
2088 #if !defined(THREADED_RTS)
2089 blocked_queue_hd = END_TSO_QUEUE;
2090 blocked_queue_tl = END_TSO_QUEUE;
2091 sleeping_queue = END_TSO_QUEUE;
2094 blackhole_queue = END_TSO_QUEUE;
2096 sched_state = SCHED_RUNNING;
2097 recent_activity = ACTIVITY_YES;
2099 #if defined(THREADED_RTS)
2100 /* Initialise the mutex and condition variables used by
2102 initMutex(&sched_mutex);
2105 ACQUIRE_LOCK(&sched_mutex);
2107 /* A capability holds the state a native thread needs in
2108 * order to execute STG code. At least one capability is
2109 * floating around (only THREADED_RTS builds have more than one).
2115 #if defined(THREADED_RTS) || defined(PARALLEL_HASKELL)
2119 #if defined(THREADED_RTS)
2121 * Eagerly start one worker to run each Capability, except for
2122 * Capability 0. The idea is that we're probably going to start a
2123 * bound thread on Capability 0 pretty soon, so we don't want a
2124 * worker task hogging it.
2129 for (i = 1; i < n_capabilities; i++) {
2130 cap = &capabilities[i];
2131 ACQUIRE_LOCK(&cap->lock);
2132 startWorkerTask(cap, workerStart);
2133 RELEASE_LOCK(&cap->lock);
2138 trace(TRACE_sched, "start: %d capabilities", n_capabilities);
2140 RELEASE_LOCK(&sched_mutex);
2145 rtsBool wait_foreign
2146 #if !defined(THREADED_RTS)
2147 __attribute__((unused))
2150 /* see Capability.c, shutdownCapability() */
2154 #if defined(THREADED_RTS)
2155 ACQUIRE_LOCK(&sched_mutex);
2156 task = newBoundTask();
2157 RELEASE_LOCK(&sched_mutex);
2160 // If we haven't killed all the threads yet, do it now.
2161 if (sched_state < SCHED_SHUTTING_DOWN) {
2162 sched_state = SCHED_INTERRUPTING;
2163 #if defined(THREADED_RTS)
2164 waitForReturnCapability(&task->cap,task);
2165 scheduleDoGC(task->cap,task,rtsFalse);
2166 releaseCapability(task->cap);
2168 scheduleDoGC(&MainCapability,task,rtsFalse);
2171 sched_state = SCHED_SHUTTING_DOWN;
2173 #if defined(THREADED_RTS)
2177 for (i = 0; i < n_capabilities; i++) {
2178 shutdownCapability(&capabilities[i], task, wait_foreign);
2180 boundTaskExiting(task);
2186 freeScheduler( void )
2190 ACQUIRE_LOCK(&sched_mutex);
2191 still_running = freeTaskManager();
2192 // We can only free the Capabilities if there are no Tasks still
2193 // running. We might have a Task about to return from a foreign
2194 // call into waitForReturnCapability(), for example (actually,
2195 // this should be the *only* thing that a still-running Task can
2196 // do at this point, and it will block waiting for the
2198 if (still_running == 0) {
2200 if (n_capabilities != 1) {
2201 stgFree(capabilities);
2204 RELEASE_LOCK(&sched_mutex);
2205 #if defined(THREADED_RTS)
2206 closeMutex(&sched_mutex);
2210 /* -----------------------------------------------------------------------------
2213 This is the interface to the garbage collector from Haskell land.
2214 We provide this so that external C code can allocate and garbage
2215 collect when called from Haskell via _ccall_GC.
2216 -------------------------------------------------------------------------- */
2219 performGC_(rtsBool force_major)
2223 // We must grab a new Task here, because the existing Task may be
2224 // associated with a particular Capability, and chained onto the
2225 // suspended_ccalling_tasks queue.
2226 ACQUIRE_LOCK(&sched_mutex);
2227 task = newBoundTask();
2228 RELEASE_LOCK(&sched_mutex);
2230 waitForReturnCapability(&task->cap,task);
2231 scheduleDoGC(task->cap,task,force_major);
2232 releaseCapability(task->cap);
2233 boundTaskExiting(task);
2239 performGC_(rtsFalse);
2243 performMajorGC(void)
2245 performGC_(rtsTrue);
2248 /* -----------------------------------------------------------------------------
2251 If the thread has reached its maximum stack size, then raise the
2252 StackOverflow exception in the offending thread. Otherwise
2253 relocate the TSO into a larger chunk of memory and adjust its stack
2255 -------------------------------------------------------------------------- */
2258 threadStackOverflow(Capability *cap, StgTSO *tso)
2260 nat new_stack_size, stack_words;
2265 IF_DEBUG(sanity,checkTSO(tso));
2267 // don't allow throwTo() to modify the blocked_exceptions queue
2268 // while we are moving the TSO:
2269 lockClosure((StgClosure *)tso);
2271 if (tso->stack_size >= tso->max_stack_size && !(tso->flags & TSO_BLOCKEX)) {
2272 // NB. never raise a StackOverflow exception if the thread is
2273 // inside Control.Exceptino.block. It is impractical to protect
2274 // against stack overflow exceptions, since virtually anything
2275 // can raise one (even 'catch'), so this is the only sensible
2276 // thing to do here. See bug #767.
2278 debugTrace(DEBUG_gc,
2279 "threadStackOverflow of TSO %ld (%p): stack too large (now %ld; max is %ld)",
2280 (long)tso->id, tso, (long)tso->stack_size, (long)tso->max_stack_size);
2282 /* If we're debugging, just print out the top of the stack */
2283 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2286 // Send this thread the StackOverflow exception
2288 throwToSingleThreaded(cap, tso, (StgClosure *)stackOverflow_closure);
2292 /* Try to double the current stack size. If that takes us over the
2293 * maximum stack size for this thread, then use the maximum instead
2294 * (that is, unless we're already at or over the max size and we
2295 * can't raise the StackOverflow exception (see above), in which
2296 * case just double the size). Finally round up so the TSO ends up as
2297 * a whole number of blocks.
2299 if (tso->stack_size >= tso->max_stack_size) {
2300 new_stack_size = tso->stack_size * 2;
2302 new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
2304 new_tso_size = (lnat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) +
2305 TSO_STRUCT_SIZE)/sizeof(W_);
2306 new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */
2307 new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
2309 debugTrace(DEBUG_sched,
2310 "increasing stack size from %ld words to %d.",
2311 (long)tso->stack_size, new_stack_size);
2313 dest = (StgTSO *)allocateLocal(cap,new_tso_size);
2314 TICK_ALLOC_TSO(new_stack_size,0);
2316 /* copy the TSO block and the old stack into the new area */
2317 memcpy(dest,tso,TSO_STRUCT_SIZE);
2318 stack_words = tso->stack + tso->stack_size - tso->sp;
2319 new_sp = (P_)dest + new_tso_size - stack_words;
2320 memcpy(new_sp, tso->sp, stack_words * sizeof(W_));
2322 /* relocate the stack pointers... */
2324 dest->stack_size = new_stack_size;
2326 /* Mark the old TSO as relocated. We have to check for relocated
2327 * TSOs in the garbage collector and any primops that deal with TSOs.
2329 * It's important to set the sp value to just beyond the end
2330 * of the stack, so we don't attempt to scavenge any part of the
2333 tso->what_next = ThreadRelocated;
2334 setTSOLink(cap,tso,dest);
2335 tso->sp = (P_)&(tso->stack[tso->stack_size]);
2336 tso->why_blocked = NotBlocked;
2338 IF_PAR_DEBUG(verbose,
2339 debugBelch("@@ threadStackOverflow of TSO %d (now at %p): stack size increased to %ld\n",
2340 tso->id, tso, tso->stack_size);
2341 /* If we're debugging, just print out the top of the stack */
2342 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2348 IF_DEBUG(sanity,checkTSO(dest));
2350 IF_DEBUG(scheduler,printTSO(dest));
2357 threadStackUnderflow (Task *task STG_UNUSED, StgTSO *tso)
2359 bdescr *bd, *new_bd;
2360 lnat free_w, tso_size_w;
2363 tso_size_w = tso_sizeW(tso);
2365 if (tso_size_w < MBLOCK_SIZE_W ||
2366 (nat)(tso->stack + tso->stack_size - tso->sp) > tso->stack_size / 4)
2371 // don't allow throwTo() to modify the blocked_exceptions queue
2372 // while we are moving the TSO:
2373 lockClosure((StgClosure *)tso);
2375 // this is the number of words we'll free
2376 free_w = round_to_mblocks(tso_size_w/2);
2378 bd = Bdescr((StgPtr)tso);
2379 new_bd = splitLargeBlock(bd, free_w / BLOCK_SIZE_W);
2380 bd->free = bd->start + TSO_STRUCT_SIZEW;
2382 new_tso = (StgTSO *)new_bd->start;
2383 memcpy(new_tso,tso,TSO_STRUCT_SIZE);
2384 new_tso->stack_size = new_bd->free - new_tso->stack;
2386 debugTrace(DEBUG_sched, "thread %ld: reducing TSO size from %lu words to %lu",
2387 (long)tso->id, tso_size_w, tso_sizeW(new_tso));
2389 tso->what_next = ThreadRelocated;
2390 tso->_link = new_tso; // no write barrier reqd: same generation
2392 // The TSO attached to this Task may have moved, so update the
2394 if (task->tso == tso) {
2395 task->tso = new_tso;
2401 IF_DEBUG(sanity,checkTSO(new_tso));
2406 /* ---------------------------------------------------------------------------
2408 - usually called inside a signal handler so it mustn't do anything fancy.
2409 ------------------------------------------------------------------------ */
2412 interruptStgRts(void)
2414 sched_state = SCHED_INTERRUPTING;
2415 setContextSwitches();
2419 /* -----------------------------------------------------------------------------
2422 This function causes at least one OS thread to wake up and run the
2423 scheduler loop. It is invoked when the RTS might be deadlocked, or
2424 an external event has arrived that may need servicing (eg. a
2425 keyboard interrupt).
2427 In the single-threaded RTS we don't do anything here; we only have
2428 one thread anyway, and the event that caused us to want to wake up
2429 will have interrupted any blocking system call in progress anyway.
2430 -------------------------------------------------------------------------- */
2435 #if defined(THREADED_RTS)
2436 // This forces the IO Manager thread to wakeup, which will
2437 // in turn ensure that some OS thread wakes up and runs the
2438 // scheduler loop, which will cause a GC and deadlock check.
2443 /* -----------------------------------------------------------------------------
2446 * Check the blackhole_queue for threads that can be woken up. We do
2447 * this periodically: before every GC, and whenever the run queue is
2450 * An elegant solution might be to just wake up all the blocked
2451 * threads with awakenBlockedQueue occasionally: they'll go back to
2452 * sleep again if the object is still a BLACKHOLE. Unfortunately this
2453 * doesn't give us a way to tell whether we've actually managed to
2454 * wake up any threads, so we would be busy-waiting.
2456 * -------------------------------------------------------------------------- */
2459 checkBlackHoles (Capability *cap)
2462 rtsBool any_woke_up = rtsFalse;
2465 // blackhole_queue is global:
2466 ASSERT_LOCK_HELD(&sched_mutex);
2468 debugTrace(DEBUG_sched, "checking threads blocked on black holes");
2470 // ASSUMES: sched_mutex
2471 prev = &blackhole_queue;
2472 t = blackhole_queue;
2473 while (t != END_TSO_QUEUE) {
2474 ASSERT(t->why_blocked == BlockedOnBlackHole);
2475 type = get_itbl(UNTAG_CLOSURE(t->block_info.closure))->type;
2476 if (type != BLACKHOLE && type != CAF_BLACKHOLE) {
2477 IF_DEBUG(sanity,checkTSO(t));
2478 t = unblockOne(cap, t);
2480 any_woke_up = rtsTrue;
2490 /* -----------------------------------------------------------------------------
2493 This is used for interruption (^C) and forking, and corresponds to
2494 raising an exception but without letting the thread catch the
2496 -------------------------------------------------------------------------- */
2499 deleteThread (Capability *cap, StgTSO *tso)
2501 // NOTE: must only be called on a TSO that we have exclusive
2502 // access to, because we will call throwToSingleThreaded() below.
2503 // The TSO must be on the run queue of the Capability we own, or
2504 // we must own all Capabilities.
2506 if (tso->why_blocked != BlockedOnCCall &&
2507 tso->why_blocked != BlockedOnCCall_NoUnblockExc) {
2508 throwToSingleThreaded(cap,tso,NULL);
2512 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
2514 deleteThread_(Capability *cap, StgTSO *tso)
2515 { // for forkProcess only:
2516 // like deleteThread(), but we delete threads in foreign calls, too.
2518 if (tso->why_blocked == BlockedOnCCall ||
2519 tso->why_blocked == BlockedOnCCall_NoUnblockExc) {
2520 unblockOne(cap,tso);
2521 tso->what_next = ThreadKilled;
2523 deleteThread(cap,tso);
2528 /* -----------------------------------------------------------------------------
2529 raiseExceptionHelper
2531 This function is called by the raise# primitve, just so that we can
2532 move some of the tricky bits of raising an exception from C-- into
2533 C. Who knows, it might be a useful re-useable thing here too.
2534 -------------------------------------------------------------------------- */
2537 raiseExceptionHelper (StgRegTable *reg, StgTSO *tso, StgClosure *exception)
2539 Capability *cap = regTableToCapability(reg);
2540 StgThunk *raise_closure = NULL;
2542 StgRetInfoTable *info;
2544 // This closure represents the expression 'raise# E' where E
2545 // is the exception raise. It is used to overwrite all the
2546 // thunks which are currently under evaluataion.
2549 // OLD COMMENT (we don't have MIN_UPD_SIZE now):
2550 // LDV profiling: stg_raise_info has THUNK as its closure
2551 // type. Since a THUNK takes at least MIN_UPD_SIZE words in its
2552 // payload, MIN_UPD_SIZE is more approprate than 1. It seems that
2553 // 1 does not cause any problem unless profiling is performed.
2554 // However, when LDV profiling goes on, we need to linearly scan
2555 // small object pool, where raise_closure is stored, so we should
2556 // use MIN_UPD_SIZE.
2558 // raise_closure = (StgClosure *)RET_STGCALL1(P_,allocate,
2559 // sizeofW(StgClosure)+1);
2563 // Walk up the stack, looking for the catch frame. On the way,
2564 // we update any closures pointed to from update frames with the
2565 // raise closure that we just built.
2569 info = get_ret_itbl((StgClosure *)p);
2570 next = p + stack_frame_sizeW((StgClosure *)p);
2571 switch (info->i.type) {
2574 // Only create raise_closure if we need to.
2575 if (raise_closure == NULL) {
2577 (StgThunk *)allocateLocal(cap,sizeofW(StgThunk)+1);
2578 SET_HDR(raise_closure, &stg_raise_info, CCCS);
2579 raise_closure->payload[0] = exception;
2581 UPD_IND(((StgUpdateFrame *)p)->updatee,(StgClosure *)raise_closure);
2585 case ATOMICALLY_FRAME:
2586 debugTrace(DEBUG_stm, "found ATOMICALLY_FRAME at %p", p);
2588 return ATOMICALLY_FRAME;
2594 case CATCH_STM_FRAME:
2595 debugTrace(DEBUG_stm, "found CATCH_STM_FRAME at %p", p);
2597 return CATCH_STM_FRAME;
2603 case CATCH_RETRY_FRAME:
2612 /* -----------------------------------------------------------------------------
2613 findRetryFrameHelper
2615 This function is called by the retry# primitive. It traverses the stack
2616 leaving tso->sp referring to the frame which should handle the retry.
2618 This should either be a CATCH_RETRY_FRAME (if the retry# is within an orElse#)
2619 or should be a ATOMICALLY_FRAME (if the retry# reaches the top level).
2621 We skip CATCH_STM_FRAMEs (aborting and rolling back the nested tx that they
2622 create) because retries are not considered to be exceptions, despite the
2623 similar implementation.
2625 We should not expect to see CATCH_FRAME or STOP_FRAME because those should
2626 not be created within memory transactions.
2627 -------------------------------------------------------------------------- */
2630 findRetryFrameHelper (StgTSO *tso)
2633 StgRetInfoTable *info;
2637 info = get_ret_itbl((StgClosure *)p);
2638 next = p + stack_frame_sizeW((StgClosure *)p);
2639 switch (info->i.type) {
2641 case ATOMICALLY_FRAME:
2642 debugTrace(DEBUG_stm,
2643 "found ATOMICALLY_FRAME at %p during retry", p);
2645 return ATOMICALLY_FRAME;
2647 case CATCH_RETRY_FRAME:
2648 debugTrace(DEBUG_stm,
2649 "found CATCH_RETRY_FRAME at %p during retrry", p);
2651 return CATCH_RETRY_FRAME;
2653 case CATCH_STM_FRAME: {
2654 StgTRecHeader *trec = tso -> trec;
2655 StgTRecHeader *outer = stmGetEnclosingTRec(trec);
2656 debugTrace(DEBUG_stm,
2657 "found CATCH_STM_FRAME at %p during retry", p);
2658 debugTrace(DEBUG_stm, "trec=%p outer=%p", trec, outer);
2659 stmAbortTransaction(tso -> cap, trec);
2660 stmFreeAbortedTRec(tso -> cap, trec);
2661 tso -> trec = outer;
2668 ASSERT(info->i.type != CATCH_FRAME);
2669 ASSERT(info->i.type != STOP_FRAME);
2676 /* -----------------------------------------------------------------------------
2677 resurrectThreads is called after garbage collection on the list of
2678 threads found to be garbage. Each of these threads will be woken
2679 up and sent a signal: BlockedOnDeadMVar if the thread was blocked
2680 on an MVar, or NonTermination if the thread was blocked on a Black
2683 Locks: assumes we hold *all* the capabilities.
2684 -------------------------------------------------------------------------- */
2687 resurrectThreads (StgTSO *threads)
2693 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
2694 next = tso->global_link;
2696 step = Bdescr((P_)tso)->step;
2697 tso->global_link = step->threads;
2698 step->threads = tso;
2700 debugTrace(DEBUG_sched, "resurrecting thread %lu", (unsigned long)tso->id);
2702 // Wake up the thread on the Capability it was last on
2705 switch (tso->why_blocked) {
2707 case BlockedOnException:
2708 /* Called by GC - sched_mutex lock is currently held. */
2709 throwToSingleThreaded(cap, tso,
2710 (StgClosure *)blockedOnDeadMVar_closure);
2712 case BlockedOnBlackHole:
2713 throwToSingleThreaded(cap, tso,
2714 (StgClosure *)nonTermination_closure);
2717 throwToSingleThreaded(cap, tso,
2718 (StgClosure *)blockedIndefinitely_closure);
2721 /* This might happen if the thread was blocked on a black hole
2722 * belonging to a thread that we've just woken up (raiseAsync
2723 * can wake up threads, remember...).
2727 barf("resurrectThreads: thread blocked in a strange way");
2732 /* -----------------------------------------------------------------------------
2733 performPendingThrowTos is called after garbage collection, and
2734 passed a list of threads that were found to have pending throwTos
2735 (tso->blocked_exceptions was not empty), and were blocked.
2736 Normally this doesn't happen, because we would deliver the
2737 exception directly if the target thread is blocked, but there are
2738 small windows where it might occur on a multiprocessor (see
2741 NB. we must be holding all the capabilities at this point, just
2742 like resurrectThreads().
2743 -------------------------------------------------------------------------- */
2746 performPendingThrowTos (StgTSO *threads)
2752 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
2753 next = tso->global_link;
2755 step = Bdescr((P_)tso)->step;
2756 tso->global_link = step->threads;
2757 step->threads = tso;
2759 debugTrace(DEBUG_sched, "performing blocked throwTo to thread %lu", (unsigned long)tso->id);
2762 maybePerformBlockedException(cap, tso);