1 /* ---------------------------------------------------------------------------
3 * (c) The GHC Team, 1998-2006
5 * The scheduler and thread-related functionality
7 * --------------------------------------------------------------------------*/
9 #include "PosixSource.h"
10 #define KEEP_LOCKCLOSURE
13 #include "sm/Storage.h"
17 #include "Interpreter.h"
19 #include "RtsSignals.h"
20 #include "sm/Sanity.h"
24 #include "ThreadLabels.h"
26 #include "Proftimer.h"
29 #include "sm/GC.h" // waitForGcThreads, releaseGCThreads, N
31 #include "Capability.h"
33 #include "AwaitEvent.h"
34 #if defined(mingw32_HOST_OS)
35 #include "win32/IOManager.h"
38 #include "RaiseAsync.h"
41 #include "ThreadPaused.h"
43 #ifdef HAVE_SYS_TYPES_H
44 #include <sys/types.h>
58 /* -----------------------------------------------------------------------------
60 * -------------------------------------------------------------------------- */
62 #if !defined(THREADED_RTS)
63 // Blocked/sleeping thrads
64 StgTSO *blocked_queue_hd = NULL;
65 StgTSO *blocked_queue_tl = NULL;
66 StgTSO *sleeping_queue = NULL; // perhaps replace with a hash table?
69 /* Threads blocked on blackholes.
70 * LOCK: sched_mutex+capability, or all capabilities
72 StgTSO *blackhole_queue = NULL;
74 /* The blackhole_queue should be checked for threads to wake up. See
75 * Schedule.h for more thorough comment.
76 * LOCK: none (doesn't matter if we miss an update)
78 rtsBool blackholes_need_checking = rtsFalse;
80 /* Set to true when the latest garbage collection failed to reclaim
81 * enough space, and the runtime should proceed to shut itself down in
82 * an orderly fashion (emitting profiling info etc.)
84 rtsBool heap_overflow = rtsFalse;
86 /* flag that tracks whether we have done any execution in this time slice.
87 * LOCK: currently none, perhaps we should lock (but needs to be
88 * updated in the fast path of the scheduler).
90 * NB. must be StgWord, we do xchg() on it.
92 volatile StgWord recent_activity = ACTIVITY_YES;
94 /* if this flag is set as well, give up execution
95 * LOCK: none (changes monotonically)
97 volatile StgWord sched_state = SCHED_RUNNING;
99 /* This is used in `TSO.h' and gcc 2.96 insists that this variable actually
100 * exists - earlier gccs apparently didn't.
106 * Set to TRUE when entering a shutdown state (via shutdownHaskellAndExit()) --
107 * in an MT setting, needed to signal that a worker thread shouldn't hang around
108 * in the scheduler when it is out of work.
110 rtsBool shutting_down_scheduler = rtsFalse;
113 * This mutex protects most of the global scheduler data in
114 * the THREADED_RTS runtime.
116 #if defined(THREADED_RTS)
120 #if !defined(mingw32_HOST_OS)
121 #define FORKPROCESS_PRIMOP_SUPPORTED
124 /* -----------------------------------------------------------------------------
125 * static function prototypes
126 * -------------------------------------------------------------------------- */
128 static Capability *schedule (Capability *initialCapability, Task *task);
131 // These function all encapsulate parts of the scheduler loop, and are
132 // abstracted only to make the structure and control flow of the
133 // scheduler clearer.
135 static void schedulePreLoop (void);
136 static void scheduleFindWork (Capability *cap);
137 #if defined(THREADED_RTS)
138 static void scheduleYield (Capability **pcap, Task *task, rtsBool);
140 static void scheduleStartSignalHandlers (Capability *cap);
141 static void scheduleCheckBlockedThreads (Capability *cap);
142 static void scheduleProcessInbox(Capability *cap);
143 static void scheduleCheckBlackHoles (Capability *cap);
144 static void scheduleDetectDeadlock (Capability *cap, Task *task);
145 static void schedulePushWork(Capability *cap, Task *task);
146 #if defined(THREADED_RTS)
147 static void scheduleActivateSpark(Capability *cap);
149 static void schedulePostRunThread(Capability *cap, StgTSO *t);
150 static rtsBool scheduleHandleHeapOverflow( Capability *cap, StgTSO *t );
151 static void scheduleHandleStackOverflow( Capability *cap, Task *task,
153 static rtsBool scheduleHandleYield( Capability *cap, StgTSO *t,
154 nat prev_what_next );
155 static void scheduleHandleThreadBlocked( StgTSO *t );
156 static rtsBool scheduleHandleThreadFinished( Capability *cap, Task *task,
158 static rtsBool scheduleNeedHeapProfile(rtsBool ready_to_gc);
159 static Capability *scheduleDoGC(Capability *cap, Task *task,
160 rtsBool force_major);
162 static rtsBool checkBlackHoles(Capability *cap);
164 static StgTSO *threadStackOverflow(Capability *cap, StgTSO *tso);
165 static StgTSO *threadStackUnderflow(Capability *cap, Task *task, StgTSO *tso);
167 static void deleteThread (Capability *cap, StgTSO *tso);
168 static void deleteAllThreads (Capability *cap);
170 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
171 static void deleteThread_(Capability *cap, StgTSO *tso);
174 /* -----------------------------------------------------------------------------
175 * Putting a thread on the run queue: different scheduling policies
176 * -------------------------------------------------------------------------- */
179 addToRunQueue( Capability *cap, StgTSO *t )
181 // this does round-robin scheduling; good for concurrency
182 appendToRunQueue(cap,t);
185 /* ---------------------------------------------------------------------------
186 Main scheduling loop.
188 We use round-robin scheduling, each thread returning to the
189 scheduler loop when one of these conditions is detected:
192 * timer expires (thread yields)
198 In a GranSim setup this loop iterates over the global event queue.
199 This revolves around the global event queue, which determines what
200 to do next. Therefore, it's more complicated than either the
201 concurrent or the parallel (GUM) setup.
202 This version has been entirely removed (JB 2008/08).
205 GUM iterates over incoming messages.
206 It starts with nothing to do (thus CurrentTSO == END_TSO_QUEUE),
207 and sends out a fish whenever it has nothing to do; in-between
208 doing the actual reductions (shared code below) it processes the
209 incoming messages and deals with delayed operations
210 (see PendingFetches).
211 This is not the ugliest code you could imagine, but it's bloody close.
213 (JB 2008/08) This version was formerly indicated by a PP-Flag PAR,
214 now by PP-flag PARALLEL_HASKELL. The Eden RTS (in GHC-6.x) uses it,
215 as well as future GUM versions. This file has been refurbished to
216 only contain valid code, which is however incomplete, refers to
217 invalid includes etc.
219 ------------------------------------------------------------------------ */
222 schedule (Capability *initialCapability, Task *task)
226 StgThreadReturnCode ret;
229 #if defined(THREADED_RTS)
230 rtsBool first = rtsTrue;
231 rtsBool force_yield = rtsFalse;
234 cap = initialCapability;
236 // Pre-condition: this task owns initialCapability.
237 // The sched_mutex is *NOT* held
238 // NB. on return, we still hold a capability.
240 debugTrace (DEBUG_sched, "cap %d: schedule()", initialCapability->no);
244 // -----------------------------------------------------------
245 // Scheduler loop starts here:
249 // Check whether we have re-entered the RTS from Haskell without
250 // going via suspendThread()/resumeThread (i.e. a 'safe' foreign
252 if (cap->in_haskell) {
253 errorBelch("schedule: re-entered unsafely.\n"
254 " Perhaps a 'foreign import unsafe' should be 'safe'?");
255 stg_exit(EXIT_FAILURE);
258 // The interruption / shutdown sequence.
260 // In order to cleanly shut down the runtime, we want to:
261 // * make sure that all main threads return to their callers
262 // with the state 'Interrupted'.
263 // * clean up all OS threads assocated with the runtime
264 // * free all memory etc.
266 // So the sequence for ^C goes like this:
268 // * ^C handler sets sched_state := SCHED_INTERRUPTING and
269 // arranges for some Capability to wake up
271 // * all threads in the system are halted, and the zombies are
272 // placed on the run queue for cleaning up. We acquire all
273 // the capabilities in order to delete the threads, this is
274 // done by scheduleDoGC() for convenience (because GC already
275 // needs to acquire all the capabilities). We can't kill
276 // threads involved in foreign calls.
278 // * somebody calls shutdownHaskell(), which calls exitScheduler()
280 // * sched_state := SCHED_SHUTTING_DOWN
282 // * all workers exit when the run queue on their capability
283 // drains. All main threads will also exit when their TSO
284 // reaches the head of the run queue and they can return.
286 // * eventually all Capabilities will shut down, and the RTS can
289 // * We might be left with threads blocked in foreign calls,
290 // we should really attempt to kill these somehow (TODO);
292 switch (sched_state) {
295 case SCHED_INTERRUPTING:
296 debugTrace(DEBUG_sched, "SCHED_INTERRUPTING");
297 #if defined(THREADED_RTS)
298 discardSparksCap(cap);
300 /* scheduleDoGC() deletes all the threads */
301 cap = scheduleDoGC(cap,task,rtsFalse);
303 // after scheduleDoGC(), we must be shutting down. Either some
304 // other Capability did the final GC, or we did it above,
305 // either way we can fall through to the SCHED_SHUTTING_DOWN
307 ASSERT(sched_state == SCHED_SHUTTING_DOWN);
310 case SCHED_SHUTTING_DOWN:
311 debugTrace(DEBUG_sched, "SCHED_SHUTTING_DOWN");
312 // If we are a worker, just exit. If we're a bound thread
313 // then we will exit below when we've removed our TSO from
315 if (!isBoundTask(task) && emptyRunQueue(cap)) {
320 barf("sched_state: %d", sched_state);
323 scheduleFindWork(cap);
325 /* work pushing, currently relevant only for THREADED_RTS:
326 (pushes threads, wakes up idle capabilities for stealing) */
327 schedulePushWork(cap,task);
329 scheduleDetectDeadlock(cap,task);
331 #if defined(THREADED_RTS)
332 cap = task->cap; // reload cap, it might have changed
335 // Normally, the only way we can get here with no threads to
336 // run is if a keyboard interrupt received during
337 // scheduleCheckBlockedThreads() or scheduleDetectDeadlock().
338 // Additionally, it is not fatal for the
339 // threaded RTS to reach here with no threads to run.
341 // win32: might be here due to awaitEvent() being abandoned
342 // as a result of a console event having been delivered.
344 #if defined(THREADED_RTS)
348 // // don't yield the first time, we want a chance to run this
349 // // thread for a bit, even if there are others banging at the
352 // ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
356 scheduleYield(&cap,task,force_yield);
357 force_yield = rtsFalse;
359 if (emptyRunQueue(cap)) continue; // look for work again
362 #if !defined(THREADED_RTS) && !defined(mingw32_HOST_OS)
363 if ( emptyRunQueue(cap) ) {
364 ASSERT(sched_state >= SCHED_INTERRUPTING);
369 // Get a thread to run
371 t = popRunQueue(cap);
373 // Sanity check the thread we're about to run. This can be
374 // expensive if there is lots of thread switching going on...
375 IF_DEBUG(sanity,checkTSO(t));
377 #if defined(THREADED_RTS)
378 // Check whether we can run this thread in the current task.
379 // If not, we have to pass our capability to the right task.
381 InCall *bound = t->bound;
384 if (bound->task == task) {
385 // yes, the Haskell thread is bound to the current native thread
387 debugTrace(DEBUG_sched,
388 "thread %lu bound to another OS thread",
389 (unsigned long)t->id);
390 // no, bound to a different Haskell thread: pass to that thread
391 pushOnRunQueue(cap,t);
395 // The thread we want to run is unbound.
396 if (task->incall->tso) {
397 debugTrace(DEBUG_sched,
398 "this OS thread cannot run thread %lu",
399 (unsigned long)t->id);
400 // no, the current native thread is bound to a different
401 // Haskell thread, so pass it to any worker thread
402 pushOnRunQueue(cap,t);
409 // If we're shutting down, and this thread has not yet been
410 // killed, kill it now. This sometimes happens when a finalizer
411 // thread is created by the final GC, or a thread previously
412 // in a foreign call returns.
413 if (sched_state >= SCHED_INTERRUPTING &&
414 !(t->what_next == ThreadComplete || t->what_next == ThreadKilled)) {
418 /* context switches are initiated by the timer signal, unless
419 * the user specified "context switch as often as possible", with
422 if (RtsFlags.ConcFlags.ctxtSwitchTicks == 0
423 && !emptyThreadQueues(cap)) {
424 cap->context_switch = 1;
429 // CurrentTSO is the thread to run. t might be different if we
430 // loop back to run_thread, so make sure to set CurrentTSO after
432 cap->r.rCurrentTSO = t;
434 startHeapProfTimer();
436 // Check for exceptions blocked on this thread
437 maybePerformBlockedException (cap, t);
439 // ----------------------------------------------------------------------
440 // Run the current thread
442 ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
443 ASSERT(t->cap == cap);
444 ASSERT(t->bound ? t->bound->task->cap == cap : 1);
446 prev_what_next = t->what_next;
448 errno = t->saved_errno;
450 SetLastError(t->saved_winerror);
453 cap->in_haskell = rtsTrue;
457 #if defined(THREADED_RTS)
458 if (recent_activity == ACTIVITY_DONE_GC) {
459 // ACTIVITY_DONE_GC means we turned off the timer signal to
460 // conserve power (see #1623). Re-enable it here.
462 prev = xchg((P_)&recent_activity, ACTIVITY_YES);
463 if (prev == ACTIVITY_DONE_GC) {
466 } else if (recent_activity != ACTIVITY_INACTIVE) {
467 // If we reached ACTIVITY_INACTIVE, then don't reset it until
468 // we've done the GC. The thread running here might just be
469 // the IO manager thread that handle_tick() woke up via
471 recent_activity = ACTIVITY_YES;
475 traceEventRunThread(cap, t);
477 switch (prev_what_next) {
481 /* Thread already finished, return to scheduler. */
482 ret = ThreadFinished;
488 r = StgRun((StgFunPtr) stg_returnToStackTop, &cap->r);
489 cap = regTableToCapability(r);
494 case ThreadInterpret:
495 cap = interpretBCO(cap);
500 barf("schedule: invalid what_next field");
503 cap->in_haskell = rtsFalse;
505 // The TSO might have moved, eg. if it re-entered the RTS and a GC
506 // happened. So find the new location:
507 t = cap->r.rCurrentTSO;
509 // We have run some Haskell code: there might be blackhole-blocked
510 // threads to wake up now.
511 // Lock-free test here should be ok, we're just setting a flag.
512 if ( blackhole_queue != END_TSO_QUEUE ) {
513 blackholes_need_checking = rtsTrue;
516 // And save the current errno in this thread.
517 // XXX: possibly bogus for SMP because this thread might already
518 // be running again, see code below.
519 t->saved_errno = errno;
521 // Similarly for Windows error code
522 t->saved_winerror = GetLastError();
525 traceEventStopThread(cap, t, ret);
527 #if defined(THREADED_RTS)
528 // If ret is ThreadBlocked, and this Task is bound to the TSO that
529 // blocked, we are in limbo - the TSO is now owned by whatever it
530 // is blocked on, and may in fact already have been woken up,
531 // perhaps even on a different Capability. It may be the case
532 // that task->cap != cap. We better yield this Capability
533 // immediately and return to normaility.
534 if (ret == ThreadBlocked) {
535 force_yield = rtsTrue;
540 ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
541 ASSERT(t->cap == cap);
543 // ----------------------------------------------------------------------
545 // Costs for the scheduler are assigned to CCS_SYSTEM
547 #if defined(PROFILING)
551 schedulePostRunThread(cap,t);
553 if (ret != StackOverflow) {
554 t = threadStackUnderflow(cap,task,t);
557 ready_to_gc = rtsFalse;
561 ready_to_gc = scheduleHandleHeapOverflow(cap,t);
565 scheduleHandleStackOverflow(cap,task,t);
569 if (scheduleHandleYield(cap, t, prev_what_next)) {
570 // shortcut for switching between compiler/interpreter:
576 scheduleHandleThreadBlocked(t);
580 if (scheduleHandleThreadFinished(cap, task, t)) return cap;
581 ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
585 barf("schedule: invalid thread return code %d", (int)ret);
588 if (ready_to_gc || scheduleNeedHeapProfile(ready_to_gc)) {
589 cap = scheduleDoGC(cap,task,rtsFalse);
591 } /* end of while() */
594 /* ----------------------------------------------------------------------------
595 * Setting up the scheduler loop
596 * ------------------------------------------------------------------------- */
599 schedulePreLoop(void)
601 // initialisation for scheduler - what cannot go into initScheduler()
604 /* -----------------------------------------------------------------------------
607 * Search for work to do, and handle messages from elsewhere.
608 * -------------------------------------------------------------------------- */
611 scheduleFindWork (Capability *cap)
613 scheduleStartSignalHandlers(cap);
615 // Only check the black holes here if we've nothing else to do.
616 // During normal execution, the black hole list only gets checked
617 // at GC time, to avoid repeatedly traversing this possibly long
618 // list each time around the scheduler.
619 if (emptyRunQueue(cap)) { scheduleCheckBlackHoles(cap); }
621 scheduleProcessInbox(cap);
623 scheduleCheckBlockedThreads(cap);
625 #if defined(THREADED_RTS)
626 if (emptyRunQueue(cap)) { scheduleActivateSpark(cap); }
630 #if defined(THREADED_RTS)
631 STATIC_INLINE rtsBool
632 shouldYieldCapability (Capability *cap, Task *task)
634 // we need to yield this capability to someone else if..
635 // - another thread is initiating a GC
636 // - another Task is returning from a foreign call
637 // - the thread at the head of the run queue cannot be run
638 // by this Task (it is bound to another Task, or it is unbound
639 // and this task it bound).
640 return (waiting_for_gc ||
641 cap->returning_tasks_hd != NULL ||
642 (!emptyRunQueue(cap) && (task->incall->tso == NULL
643 ? cap->run_queue_hd->bound != NULL
644 : cap->run_queue_hd->bound != task->incall)));
647 // This is the single place where a Task goes to sleep. There are
648 // two reasons it might need to sleep:
649 // - there are no threads to run
650 // - we need to yield this Capability to someone else
651 // (see shouldYieldCapability())
653 // Careful: the scheduler loop is quite delicate. Make sure you run
654 // the tests in testsuite/concurrent (all ways) after modifying this,
655 // and also check the benchmarks in nofib/parallel for regressions.
658 scheduleYield (Capability **pcap, Task *task, rtsBool force_yield)
660 Capability *cap = *pcap;
662 // if we have work, and we don't need to give up the Capability, continue.
664 // The force_yield flag is used when a bound thread blocks. This
665 // is a particularly tricky situation: the current Task does not
666 // own the TSO any more, since it is on some queue somewhere, and
667 // might be woken up or manipulated by another thread at any time.
668 // The TSO and Task might be migrated to another Capability.
669 // Certain invariants might be in doubt, such as task->bound->cap
670 // == cap. We have to yield the current Capability immediately,
671 // no messing around.
674 !shouldYieldCapability(cap,task) &&
675 (!emptyRunQueue(cap) ||
677 blackholes_need_checking ||
678 sched_state >= SCHED_INTERRUPTING))
681 // otherwise yield (sleep), and keep yielding if necessary.
683 yieldCapability(&cap,task);
685 while (shouldYieldCapability(cap,task));
687 // note there may still be no threads on the run queue at this
688 // point, the caller has to check.
695 /* -----------------------------------------------------------------------------
698 * Push work to other Capabilities if we have some.
699 * -------------------------------------------------------------------------- */
702 schedulePushWork(Capability *cap USED_IF_THREADS,
703 Task *task USED_IF_THREADS)
705 /* following code not for PARALLEL_HASKELL. I kept the call general,
706 future GUM versions might use pushing in a distributed setup */
707 #if defined(THREADED_RTS)
709 Capability *free_caps[n_capabilities], *cap0;
712 // migration can be turned off with +RTS -qm
713 if (!RtsFlags.ParFlags.migrate) return;
715 // Check whether we have more threads on our run queue, or sparks
716 // in our pool, that we could hand to another Capability.
717 if (cap->run_queue_hd == END_TSO_QUEUE) {
718 if (sparkPoolSizeCap(cap) < 2) return;
720 if (cap->run_queue_hd->_link == END_TSO_QUEUE &&
721 sparkPoolSizeCap(cap) < 1) return;
724 // First grab as many free Capabilities as we can.
725 for (i=0, n_free_caps=0; i < n_capabilities; i++) {
726 cap0 = &capabilities[i];
727 if (cap != cap0 && tryGrabCapability(cap0,task)) {
728 if (!emptyRunQueue(cap0)
729 || cap->returning_tasks_hd != NULL
730 || cap->inbox != (Message*)END_TSO_QUEUE) {
731 // it already has some work, we just grabbed it at
732 // the wrong moment. Or maybe it's deadlocked!
733 releaseCapability(cap0);
735 free_caps[n_free_caps++] = cap0;
740 // we now have n_free_caps free capabilities stashed in
741 // free_caps[]. Share our run queue equally with them. This is
742 // probably the simplest thing we could do; improvements we might
743 // want to do include:
745 // - giving high priority to moving relatively new threads, on
746 // the gournds that they haven't had time to build up a
747 // working set in the cache on this CPU/Capability.
749 // - giving low priority to moving long-lived threads
751 if (n_free_caps > 0) {
752 StgTSO *prev, *t, *next;
753 rtsBool pushed_to_all;
755 debugTrace(DEBUG_sched,
756 "cap %d: %s and %d free capabilities, sharing...",
758 (!emptyRunQueue(cap) && cap->run_queue_hd->_link != END_TSO_QUEUE)?
759 "excess threads on run queue":"sparks to share (>=2)",
763 pushed_to_all = rtsFalse;
765 if (cap->run_queue_hd != END_TSO_QUEUE) {
766 prev = cap->run_queue_hd;
768 prev->_link = END_TSO_QUEUE;
769 for (; t != END_TSO_QUEUE; t = next) {
771 t->_link = END_TSO_QUEUE;
772 if (t->what_next == ThreadRelocated
773 || t->bound == task->incall // don't move my bound thread
774 || tsoLocked(t)) { // don't move a locked thread
775 setTSOLink(cap, prev, t);
777 } else if (i == n_free_caps) {
778 pushed_to_all = rtsTrue;
781 setTSOLink(cap, prev, t);
784 appendToRunQueue(free_caps[i],t);
786 traceEventMigrateThread (cap, t, free_caps[i]->no);
788 if (t->bound) { t->bound->task->cap = free_caps[i]; }
789 t->cap = free_caps[i];
793 cap->run_queue_tl = prev;
797 /* JB I left this code in place, it would work but is not necessary */
799 // If there are some free capabilities that we didn't push any
800 // threads to, then try to push a spark to each one.
801 if (!pushed_to_all) {
803 // i is the next free capability to push to
804 for (; i < n_free_caps; i++) {
805 if (emptySparkPoolCap(free_caps[i])) {
806 spark = tryStealSpark(cap->sparks);
808 debugTrace(DEBUG_sched, "pushing spark %p to capability %d", spark, free_caps[i]->no);
810 traceEventStealSpark(free_caps[i], t, cap->no);
812 newSpark(&(free_caps[i]->r), spark);
817 #endif /* SPARK_PUSHING */
819 // release the capabilities
820 for (i = 0; i < n_free_caps; i++) {
821 task->cap = free_caps[i];
822 releaseAndWakeupCapability(free_caps[i]);
825 task->cap = cap; // reset to point to our Capability.
827 #endif /* THREADED_RTS */
831 /* ----------------------------------------------------------------------------
832 * Start any pending signal handlers
833 * ------------------------------------------------------------------------- */
835 #if defined(RTS_USER_SIGNALS) && !defined(THREADED_RTS)
837 scheduleStartSignalHandlers(Capability *cap)
839 if (RtsFlags.MiscFlags.install_signal_handlers && signals_pending()) {
840 // safe outside the lock
841 startSignalHandlers(cap);
846 scheduleStartSignalHandlers(Capability *cap STG_UNUSED)
851 /* ----------------------------------------------------------------------------
852 * Check for blocked threads that can be woken up.
853 * ------------------------------------------------------------------------- */
856 scheduleCheckBlockedThreads(Capability *cap USED_IF_NOT_THREADS)
858 #if !defined(THREADED_RTS)
860 // Check whether any waiting threads need to be woken up. If the
861 // run queue is empty, and there are no other tasks running, we
862 // can wait indefinitely for something to happen.
864 if ( !emptyQueue(blocked_queue_hd) || !emptyQueue(sleeping_queue) )
866 awaitEvent( emptyRunQueue(cap) && !blackholes_need_checking );
872 /* ----------------------------------------------------------------------------
873 * Check for threads woken up by other Capabilities
874 * ------------------------------------------------------------------------- */
876 #if defined(THREADED_RTS)
878 executeMessage (Capability *cap, Message *m)
880 const StgInfoTable *i;
883 write_barrier(); // allow m->header to be modified by another thread
885 if (i == &stg_MSG_WAKEUP_info)
887 MessageWakeup *w = (MessageWakeup *)m;
888 StgTSO *tso = w->tso;
889 debugTraceCap(DEBUG_sched, cap, "message: wakeup thread %ld",
891 ASSERT(tso->cap == cap);
892 ASSERT(tso->why_blocked == BlockedOnMsgWakeup);
893 ASSERT(tso->block_info.closure == (StgClosure *)m);
894 tso->why_blocked = NotBlocked;
895 appendToRunQueue(cap, tso);
897 else if (i == &stg_MSG_THROWTO_info)
899 MessageThrowTo *t = (MessageThrowTo *)m;
901 const StgInfoTable *i;
903 i = lockClosure((StgClosure*)m);
904 if (i != &stg_MSG_THROWTO_info) {
905 unlockClosure((StgClosure*)m, i);
909 debugTraceCap(DEBUG_sched, cap, "message: throwTo %ld -> %ld",
910 (lnat)t->source->id, (lnat)t->target->id);
912 ASSERT(t->source->why_blocked == BlockedOnMsgThrowTo);
913 ASSERT(t->source->block_info.closure == (StgClosure *)m);
915 r = throwToMsg(cap, t);
918 case THROWTO_SUCCESS:
919 ASSERT(t->source->sp[0] == (StgWord)&stg_block_throwto_info);
921 unblockOne(cap, t->source);
922 // this message is done
923 unlockClosure((StgClosure*)m, &stg_IND_info);
925 case THROWTO_BLOCKED:
926 // unlock the message
927 unlockClosure((StgClosure*)m, &stg_MSG_THROWTO_info);
931 else if (i == &stg_IND_info)
933 // message was revoked
936 else if (i == &stg_WHITEHOLE_info)
942 barf("executeMessage: %p", i);
948 scheduleProcessInbox (Capability *cap USED_IF_THREADS)
950 #if defined(THREADED_RTS)
953 while (!emptyInbox(cap)) {
954 ACQUIRE_LOCK(&cap->lock);
956 cap->inbox = m->link;
957 RELEASE_LOCK(&cap->lock);
958 executeMessage(cap, (Message *)m);
963 /* ----------------------------------------------------------------------------
964 * Check for threads blocked on BLACKHOLEs that can be woken up
965 * ------------------------------------------------------------------------- */
967 scheduleCheckBlackHoles (Capability *cap)
969 if ( blackholes_need_checking ) // check without the lock first
971 ACQUIRE_LOCK(&sched_mutex);
972 if ( blackholes_need_checking ) {
973 blackholes_need_checking = rtsFalse;
974 // important that we reset the flag *before* checking the
975 // blackhole queue, otherwise we could get deadlock. This
976 // happens as follows: we wake up a thread that
977 // immediately runs on another Capability, blocks on a
978 // blackhole, and then we reset the blackholes_need_checking flag.
979 checkBlackHoles(cap);
981 RELEASE_LOCK(&sched_mutex);
985 /* ----------------------------------------------------------------------------
986 * Detect deadlock conditions and attempt to resolve them.
987 * ------------------------------------------------------------------------- */
990 scheduleDetectDeadlock (Capability *cap, Task *task)
993 * Detect deadlock: when we have no threads to run, there are no
994 * threads blocked, waiting for I/O, or sleeping, and all the
995 * other tasks are waiting for work, we must have a deadlock of
998 if ( emptyThreadQueues(cap) )
1000 #if defined(THREADED_RTS)
1002 * In the threaded RTS, we only check for deadlock if there
1003 * has been no activity in a complete timeslice. This means
1004 * we won't eagerly start a full GC just because we don't have
1005 * any threads to run currently.
1007 if (recent_activity != ACTIVITY_INACTIVE) return;
1010 debugTrace(DEBUG_sched, "deadlocked, forcing major GC...");
1012 // Garbage collection can release some new threads due to
1013 // either (a) finalizers or (b) threads resurrected because
1014 // they are unreachable and will therefore be sent an
1015 // exception. Any threads thus released will be immediately
1017 cap = scheduleDoGC (cap, task, rtsTrue/*force major GC*/);
1018 // when force_major == rtsTrue. scheduleDoGC sets
1019 // recent_activity to ACTIVITY_DONE_GC and turns off the timer
1022 if ( !emptyRunQueue(cap) ) return;
1024 #if defined(RTS_USER_SIGNALS) && !defined(THREADED_RTS)
1025 /* If we have user-installed signal handlers, then wait
1026 * for signals to arrive rather then bombing out with a
1029 if ( RtsFlags.MiscFlags.install_signal_handlers && anyUserHandlers() ) {
1030 debugTrace(DEBUG_sched,
1031 "still deadlocked, waiting for signals...");
1035 if (signals_pending()) {
1036 startSignalHandlers(cap);
1039 // either we have threads to run, or we were interrupted:
1040 ASSERT(!emptyRunQueue(cap) || sched_state >= SCHED_INTERRUPTING);
1046 #if !defined(THREADED_RTS)
1047 /* Probably a real deadlock. Send the current main thread the
1048 * Deadlock exception.
1050 if (task->incall->tso) {
1051 switch (task->incall->tso->why_blocked) {
1053 case BlockedOnBlackHole:
1054 case BlockedOnMsgThrowTo:
1056 throwToSingleThreaded(cap, task->incall->tso,
1057 (StgClosure *)nonTermination_closure);
1060 barf("deadlock: main thread blocked in a strange way");
1069 /* ----------------------------------------------------------------------------
1070 * Send pending messages (PARALLEL_HASKELL only)
1071 * ------------------------------------------------------------------------- */
1073 #if defined(PARALLEL_HASKELL)
1075 scheduleSendPendingMessages(void)
1078 # if defined(PAR) // global Mem.Mgmt., omit for now
1079 if (PendingFetches != END_BF_QUEUE) {
1084 if (RtsFlags.ParFlags.BufferTime) {
1085 // if we use message buffering, we must send away all message
1086 // packets which have become too old...
1092 /* ----------------------------------------------------------------------------
1093 * Activate spark threads (PARALLEL_HASKELL and THREADED_RTS)
1094 * ------------------------------------------------------------------------- */
1096 #if defined(THREADED_RTS)
1098 scheduleActivateSpark(Capability *cap)
1102 createSparkThread(cap);
1103 debugTrace(DEBUG_sched, "creating a spark thread");
1106 #endif // PARALLEL_HASKELL || THREADED_RTS
1108 /* ----------------------------------------------------------------------------
1109 * After running a thread...
1110 * ------------------------------------------------------------------------- */
1113 schedulePostRunThread (Capability *cap, StgTSO *t)
1115 // We have to be able to catch transactions that are in an
1116 // infinite loop as a result of seeing an inconsistent view of
1120 // [a,b] <- mapM readTVar [ta,tb]
1121 // when (a == b) loop
1123 // and a is never equal to b given a consistent view of memory.
1125 if (t -> trec != NO_TREC && t -> why_blocked == NotBlocked) {
1126 if (!stmValidateNestOfTransactions (t -> trec)) {
1127 debugTrace(DEBUG_sched | DEBUG_stm,
1128 "trec %p found wasting its time", t);
1130 // strip the stack back to the
1131 // ATOMICALLY_FRAME, aborting the (nested)
1132 // transaction, and saving the stack of any
1133 // partially-evaluated thunks on the heap.
1134 throwToSingleThreaded_(cap, t, NULL, rtsTrue);
1136 // ASSERT(get_itbl((StgClosure *)t->sp)->type == ATOMICALLY_FRAME);
1140 /* some statistics gathering in the parallel case */
1143 /* -----------------------------------------------------------------------------
1144 * Handle a thread that returned to the scheduler with ThreadHeepOverflow
1145 * -------------------------------------------------------------------------- */
1148 scheduleHandleHeapOverflow( Capability *cap, StgTSO *t )
1150 // did the task ask for a large block?
1151 if (cap->r.rHpAlloc > BLOCK_SIZE) {
1152 // if so, get one and push it on the front of the nursery.
1156 blocks = (lnat)BLOCK_ROUND_UP(cap->r.rHpAlloc) / BLOCK_SIZE;
1158 debugTrace(DEBUG_sched,
1159 "--<< thread %ld (%s) stopped: requesting a large block (size %ld)\n",
1160 (long)t->id, what_next_strs[t->what_next], blocks);
1162 // don't do this if the nursery is (nearly) full, we'll GC first.
1163 if (cap->r.rCurrentNursery->link != NULL ||
1164 cap->r.rNursery->n_blocks == 1) { // paranoia to prevent infinite loop
1165 // if the nursery has only one block.
1168 bd = allocGroup( blocks );
1170 cap->r.rNursery->n_blocks += blocks;
1172 // link the new group into the list
1173 bd->link = cap->r.rCurrentNursery;
1174 bd->u.back = cap->r.rCurrentNursery->u.back;
1175 if (cap->r.rCurrentNursery->u.back != NULL) {
1176 cap->r.rCurrentNursery->u.back->link = bd;
1178 cap->r.rNursery->blocks = bd;
1180 cap->r.rCurrentNursery->u.back = bd;
1182 // initialise it as a nursery block. We initialise the
1183 // step, gen_no, and flags field of *every* sub-block in
1184 // this large block, because this is easier than making
1185 // sure that we always find the block head of a large
1186 // block whenever we call Bdescr() (eg. evacuate() and
1187 // isAlive() in the GC would both have to do this, at
1191 for (x = bd; x < bd + blocks; x++) {
1192 initBdescr(x,g0,g0);
1198 // This assert can be a killer if the app is doing lots
1199 // of large block allocations.
1200 IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
1202 // now update the nursery to point to the new block
1203 cap->r.rCurrentNursery = bd;
1205 // we might be unlucky and have another thread get on the
1206 // run queue before us and steal the large block, but in that
1207 // case the thread will just end up requesting another large
1209 pushOnRunQueue(cap,t);
1210 return rtsFalse; /* not actually GC'ing */
1214 if (cap->r.rHpLim == NULL || cap->context_switch) {
1215 // Sometimes we miss a context switch, e.g. when calling
1216 // primitives in a tight loop, MAYBE_GC() doesn't check the
1217 // context switch flag, and we end up waiting for a GC.
1218 // See #1984, and concurrent/should_run/1984
1219 cap->context_switch = 0;
1220 addToRunQueue(cap,t);
1222 pushOnRunQueue(cap,t);
1225 /* actual GC is done at the end of the while loop in schedule() */
1228 /* -----------------------------------------------------------------------------
1229 * Handle a thread that returned to the scheduler with ThreadStackOverflow
1230 * -------------------------------------------------------------------------- */
1233 scheduleHandleStackOverflow (Capability *cap, Task *task, StgTSO *t)
1235 /* just adjust the stack for this thread, then pop it back
1239 /* enlarge the stack */
1240 StgTSO *new_t = threadStackOverflow(cap, t);
1242 /* The TSO attached to this Task may have moved, so update the
1245 if (task->incall->tso == t) {
1246 task->incall->tso = new_t;
1248 pushOnRunQueue(cap,new_t);
1252 /* -----------------------------------------------------------------------------
1253 * Handle a thread that returned to the scheduler with ThreadYielding
1254 * -------------------------------------------------------------------------- */
1257 scheduleHandleYield( Capability *cap, StgTSO *t, nat prev_what_next )
1259 /* put the thread back on the run queue. Then, if we're ready to
1260 * GC, check whether this is the last task to stop. If so, wake
1261 * up the GC thread. getThread will block during a GC until the
1265 ASSERT(t->_link == END_TSO_QUEUE);
1267 // Shortcut if we're just switching evaluators: don't bother
1268 // doing stack squeezing (which can be expensive), just run the
1270 if (cap->context_switch == 0 && t->what_next != prev_what_next) {
1271 debugTrace(DEBUG_sched,
1272 "--<< thread %ld (%s) stopped to switch evaluators",
1273 (long)t->id, what_next_strs[t->what_next]);
1277 // Reset the context switch flag. We don't do this just before
1278 // running the thread, because that would mean we would lose ticks
1279 // during GC, which can lead to unfair scheduling (a thread hogs
1280 // the CPU because the tick always arrives during GC). This way
1281 // penalises threads that do a lot of allocation, but that seems
1282 // better than the alternative.
1283 cap->context_switch = 0;
1286 //debugBelch("&& Doing sanity check on yielding TSO %ld.", t->id);
1289 addToRunQueue(cap,t);
1294 /* -----------------------------------------------------------------------------
1295 * Handle a thread that returned to the scheduler with ThreadBlocked
1296 * -------------------------------------------------------------------------- */
1299 scheduleHandleThreadBlocked( StgTSO *t
1306 // We don't need to do anything. The thread is blocked, and it
1307 // has tidied up its stack and placed itself on whatever queue
1308 // it needs to be on.
1310 // ASSERT(t->why_blocked != NotBlocked);
1311 // Not true: for example,
1312 // - in THREADED_RTS, the thread may already have been woken
1313 // up by another Capability. This actually happens: try
1314 // conc023 +RTS -N2.
1315 // - the thread may have woken itself up already, because
1316 // threadPaused() might have raised a blocked throwTo
1317 // exception, see maybePerformBlockedException().
1320 traceThreadStatus(DEBUG_sched, t);
1324 /* -----------------------------------------------------------------------------
1325 * Handle a thread that returned to the scheduler with ThreadFinished
1326 * -------------------------------------------------------------------------- */
1329 scheduleHandleThreadFinished (Capability *cap STG_UNUSED, Task *task, StgTSO *t)
1331 /* Need to check whether this was a main thread, and if so,
1332 * return with the return value.
1334 * We also end up here if the thread kills itself with an
1335 * uncaught exception, see Exception.cmm.
1338 // blocked exceptions can now complete, even if the thread was in
1339 // blocked mode (see #2910).
1340 awakenBlockedExceptionQueue (cap, t);
1343 // Check whether the thread that just completed was a bound
1344 // thread, and if so return with the result.
1346 // There is an assumption here that all thread completion goes
1347 // through this point; we need to make sure that if a thread
1348 // ends up in the ThreadKilled state, that it stays on the run
1349 // queue so it can be dealt with here.
1354 if (t->bound != task->incall) {
1355 #if !defined(THREADED_RTS)
1356 // Must be a bound thread that is not the topmost one. Leave
1357 // it on the run queue until the stack has unwound to the
1358 // point where we can deal with this. Leaving it on the run
1359 // queue also ensures that the garbage collector knows about
1360 // this thread and its return value (it gets dropped from the
1361 // step->threads list so there's no other way to find it).
1362 appendToRunQueue(cap,t);
1365 // this cannot happen in the threaded RTS, because a
1366 // bound thread can only be run by the appropriate Task.
1367 barf("finished bound thread that isn't mine");
1371 ASSERT(task->incall->tso == t);
1373 if (t->what_next == ThreadComplete) {
1375 // NOTE: return val is tso->sp[1] (see StgStartup.hc)
1376 *(task->ret) = (StgClosure *)task->incall->tso->sp[1];
1378 task->stat = Success;
1381 *(task->ret) = NULL;
1383 if (sched_state >= SCHED_INTERRUPTING) {
1384 if (heap_overflow) {
1385 task->stat = HeapExhausted;
1387 task->stat = Interrupted;
1390 task->stat = Killed;
1394 removeThreadLabel((StgWord)task->incall->tso->id);
1397 // We no longer consider this thread and task to be bound to
1398 // each other. The TSO lives on until it is GC'd, but the
1399 // task is about to be released by the caller, and we don't
1400 // want anyone following the pointer from the TSO to the
1401 // defunct task (which might have already been
1402 // re-used). This was a real bug: the GC updated
1403 // tso->bound->tso which lead to a deadlock.
1405 task->incall->tso = NULL;
1407 return rtsTrue; // tells schedule() to return
1413 /* -----------------------------------------------------------------------------
1414 * Perform a heap census
1415 * -------------------------------------------------------------------------- */
1418 scheduleNeedHeapProfile( rtsBool ready_to_gc STG_UNUSED )
1420 // When we have +RTS -i0 and we're heap profiling, do a census at
1421 // every GC. This lets us get repeatable runs for debugging.
1422 if (performHeapProfile ||
1423 (RtsFlags.ProfFlags.profileInterval==0 &&
1424 RtsFlags.ProfFlags.doHeapProfile && ready_to_gc)) {
1431 /* -----------------------------------------------------------------------------
1432 * Perform a garbage collection if necessary
1433 * -------------------------------------------------------------------------- */
1436 scheduleDoGC (Capability *cap, Task *task USED_IF_THREADS, rtsBool force_major)
1438 rtsBool heap_census;
1440 /* extern static volatile StgWord waiting_for_gc;
1441 lives inside capability.c */
1442 rtsBool gc_type, prev_pending_gc;
1446 if (sched_state == SCHED_SHUTTING_DOWN) {
1447 // The final GC has already been done, and the system is
1448 // shutting down. We'll probably deadlock if we try to GC
1454 if (sched_state < SCHED_INTERRUPTING
1455 && RtsFlags.ParFlags.parGcEnabled
1456 && N >= RtsFlags.ParFlags.parGcGen
1457 && ! oldest_gen->mark)
1459 gc_type = PENDING_GC_PAR;
1461 gc_type = PENDING_GC_SEQ;
1464 // In order to GC, there must be no threads running Haskell code.
1465 // Therefore, the GC thread needs to hold *all* the capabilities,
1466 // and release them after the GC has completed.
1468 // This seems to be the simplest way: previous attempts involved
1469 // making all the threads with capabilities give up their
1470 // capabilities and sleep except for the *last* one, which
1471 // actually did the GC. But it's quite hard to arrange for all
1472 // the other tasks to sleep and stay asleep.
1475 /* Other capabilities are prevented from running yet more Haskell
1476 threads if waiting_for_gc is set. Tested inside
1477 yieldCapability() and releaseCapability() in Capability.c */
1479 prev_pending_gc = cas(&waiting_for_gc, 0, gc_type);
1480 if (prev_pending_gc) {
1482 debugTrace(DEBUG_sched, "someone else is trying to GC (%d)...",
1485 yieldCapability(&cap,task);
1486 } while (waiting_for_gc);
1487 return cap; // NOTE: task->cap might have changed here
1490 setContextSwitches();
1492 // The final shutdown GC is always single-threaded, because it's
1493 // possible that some of the Capabilities have no worker threads.
1495 if (gc_type == PENDING_GC_SEQ)
1497 traceEventRequestSeqGc(cap);
1501 traceEventRequestParGc(cap);
1502 debugTrace(DEBUG_sched, "ready_to_gc, grabbing GC threads");
1505 // do this while the other Capabilities stop:
1506 if (cap) scheduleCheckBlackHoles(cap);
1508 if (gc_type == PENDING_GC_SEQ)
1510 // single-threaded GC: grab all the capabilities
1511 for (i=0; i < n_capabilities; i++) {
1512 debugTrace(DEBUG_sched, "ready_to_gc, grabbing all the capabilies (%d/%d)", i, n_capabilities);
1513 if (cap != &capabilities[i]) {
1514 Capability *pcap = &capabilities[i];
1515 // we better hope this task doesn't get migrated to
1516 // another Capability while we're waiting for this one.
1517 // It won't, because load balancing happens while we have
1518 // all the Capabilities, but even so it's a slightly
1519 // unsavoury invariant.
1521 waitForReturnCapability(&pcap, task);
1522 if (pcap != &capabilities[i]) {
1523 barf("scheduleDoGC: got the wrong capability");
1530 // multi-threaded GC: make sure all the Capabilities donate one
1532 waitForGcThreads(cap);
1535 #else /* !THREADED_RTS */
1537 // do this while the other Capabilities stop:
1538 if (cap) scheduleCheckBlackHoles(cap);
1542 IF_DEBUG(scheduler, printAllThreads());
1544 delete_threads_and_gc:
1546 * We now have all the capabilities; if we're in an interrupting
1547 * state, then we should take the opportunity to delete all the
1548 * threads in the system.
1550 if (sched_state == SCHED_INTERRUPTING) {
1551 deleteAllThreads(cap);
1552 sched_state = SCHED_SHUTTING_DOWN;
1555 heap_census = scheduleNeedHeapProfile(rtsTrue);
1557 traceEventGcStart(cap);
1558 #if defined(THREADED_RTS)
1559 // reset waiting_for_gc *before* GC, so that when the GC threads
1560 // emerge they don't immediately re-enter the GC.
1562 GarbageCollect(force_major || heap_census, gc_type, cap);
1564 GarbageCollect(force_major || heap_census, 0, cap);
1566 traceEventGcEnd(cap);
1568 if (recent_activity == ACTIVITY_INACTIVE && force_major)
1570 // We are doing a GC because the system has been idle for a
1571 // timeslice and we need to check for deadlock. Record the
1572 // fact that we've done a GC and turn off the timer signal;
1573 // it will get re-enabled if we run any threads after the GC.
1574 recent_activity = ACTIVITY_DONE_GC;
1579 // the GC might have taken long enough for the timer to set
1580 // recent_activity = ACTIVITY_INACTIVE, but we aren't
1581 // necessarily deadlocked:
1582 recent_activity = ACTIVITY_YES;
1585 #if defined(THREADED_RTS)
1586 if (gc_type == PENDING_GC_PAR)
1588 releaseGCThreads(cap);
1593 debugTrace(DEBUG_sched, "performing heap census");
1595 performHeapProfile = rtsFalse;
1598 if (heap_overflow && sched_state < SCHED_INTERRUPTING) {
1599 // GC set the heap_overflow flag, so we should proceed with
1600 // an orderly shutdown now. Ultimately we want the main
1601 // thread to return to its caller with HeapExhausted, at which
1602 // point the caller should call hs_exit(). The first step is
1603 // to delete all the threads.
1605 // Another way to do this would be to raise an exception in
1606 // the main thread, which we really should do because it gives
1607 // the program a chance to clean up. But how do we find the
1608 // main thread? It should presumably be the same one that
1609 // gets ^C exceptions, but that's all done on the Haskell side
1610 // (GHC.TopHandler).
1611 sched_state = SCHED_INTERRUPTING;
1612 goto delete_threads_and_gc;
1617 Once we are all together... this would be the place to balance all
1618 spark pools. No concurrent stealing or adding of new sparks can
1619 occur. Should be defined in Sparks.c. */
1620 balanceSparkPoolsCaps(n_capabilities, capabilities);
1623 #if defined(THREADED_RTS)
1624 if (gc_type == PENDING_GC_SEQ) {
1625 // release our stash of capabilities.
1626 for (i = 0; i < n_capabilities; i++) {
1627 if (cap != &capabilities[i]) {
1628 task->cap = &capabilities[i];
1629 releaseCapability(&capabilities[i]);
1643 /* ---------------------------------------------------------------------------
1644 * Singleton fork(). Do not copy any running threads.
1645 * ------------------------------------------------------------------------- */
1648 forkProcess(HsStablePtr *entry
1649 #ifndef FORKPROCESS_PRIMOP_SUPPORTED
1654 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
1660 #if defined(THREADED_RTS)
1661 if (RtsFlags.ParFlags.nNodes > 1) {
1662 errorBelch("forking not supported with +RTS -N<n> greater than 1");
1663 stg_exit(EXIT_FAILURE);
1667 debugTrace(DEBUG_sched, "forking!");
1669 // ToDo: for SMP, we should probably acquire *all* the capabilities
1672 // no funny business: hold locks while we fork, otherwise if some
1673 // other thread is holding a lock when the fork happens, the data
1674 // structure protected by the lock will forever be in an
1675 // inconsistent state in the child. See also #1391.
1676 ACQUIRE_LOCK(&sched_mutex);
1677 ACQUIRE_LOCK(&cap->lock);
1678 ACQUIRE_LOCK(&cap->running_task->lock);
1682 if (pid) { // parent
1684 RELEASE_LOCK(&sched_mutex);
1685 RELEASE_LOCK(&cap->lock);
1686 RELEASE_LOCK(&cap->running_task->lock);
1688 // just return the pid
1694 #if defined(THREADED_RTS)
1695 initMutex(&sched_mutex);
1696 initMutex(&cap->lock);
1697 initMutex(&cap->running_task->lock);
1700 // Now, all OS threads except the thread that forked are
1701 // stopped. We need to stop all Haskell threads, including
1702 // those involved in foreign calls. Also we need to delete
1703 // all Tasks, because they correspond to OS threads that are
1706 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1707 for (t = generations[g].threads; t != END_TSO_QUEUE; t = next) {
1708 if (t->what_next == ThreadRelocated) {
1711 next = t->global_link;
1712 // don't allow threads to catch the ThreadKilled
1713 // exception, but we do want to raiseAsync() because these
1714 // threads may be evaluating thunks that we need later.
1715 deleteThread_(cap,t);
1720 // Empty the run queue. It seems tempting to let all the
1721 // killed threads stay on the run queue as zombies to be
1722 // cleaned up later, but some of them correspond to bound
1723 // threads for which the corresponding Task does not exist.
1724 cap->run_queue_hd = END_TSO_QUEUE;
1725 cap->run_queue_tl = END_TSO_QUEUE;
1727 // Any suspended C-calling Tasks are no more, their OS threads
1729 cap->suspended_ccalls = NULL;
1731 // Empty the threads lists. Otherwise, the garbage
1732 // collector may attempt to resurrect some of these threads.
1733 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1734 generations[g].threads = END_TSO_QUEUE;
1737 discardTasksExcept(cap->running_task);
1739 #if defined(THREADED_RTS)
1740 // Wipe our spare workers list, they no longer exist. New
1741 // workers will be created if necessary.
1742 cap->spare_workers = NULL;
1743 cap->returning_tasks_hd = NULL;
1744 cap->returning_tasks_tl = NULL;
1747 // On Unix, all timers are reset in the child, so we need to start
1752 #if defined(THREADED_RTS)
1753 cap = ioManagerStartCap(cap);
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");
1768 /* ---------------------------------------------------------------------------
1769 * Delete all the threads in the system
1770 * ------------------------------------------------------------------------- */
1773 deleteAllThreads ( Capability *cap )
1775 // NOTE: only safe to call if we own all capabilities.
1780 debugTrace(DEBUG_sched,"deleting all threads");
1781 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1782 for (t = generations[g].threads; t != END_TSO_QUEUE; t = next) {
1783 if (t->what_next == ThreadRelocated) {
1786 next = t->global_link;
1787 deleteThread(cap,t);
1792 // The run queue now contains a bunch of ThreadKilled threads. We
1793 // must not throw these away: the main thread(s) will be in there
1794 // somewhere, and the main scheduler loop has to deal with it.
1795 // Also, the run queue is the only thing keeping these threads from
1796 // being GC'd, and we don't want the "main thread has been GC'd" panic.
1798 #if !defined(THREADED_RTS)
1799 ASSERT(blocked_queue_hd == END_TSO_QUEUE);
1800 ASSERT(sleeping_queue == END_TSO_QUEUE);
1804 /* -----------------------------------------------------------------------------
1805 Managing the suspended_ccalls list.
1806 Locks required: sched_mutex
1807 -------------------------------------------------------------------------- */
1810 suspendTask (Capability *cap, Task *task)
1814 incall = task->incall;
1815 ASSERT(incall->next == NULL && incall->prev == NULL);
1816 incall->next = cap->suspended_ccalls;
1817 incall->prev = NULL;
1818 if (cap->suspended_ccalls) {
1819 cap->suspended_ccalls->prev = incall;
1821 cap->suspended_ccalls = incall;
1825 recoverSuspendedTask (Capability *cap, Task *task)
1829 incall = task->incall;
1831 incall->prev->next = incall->next;
1833 ASSERT(cap->suspended_ccalls == incall);
1834 cap->suspended_ccalls = incall->next;
1837 incall->next->prev = incall->prev;
1839 incall->next = incall->prev = NULL;
1842 /* ---------------------------------------------------------------------------
1843 * Suspending & resuming Haskell threads.
1845 * When making a "safe" call to C (aka _ccall_GC), the task gives back
1846 * its capability before calling the C function. This allows another
1847 * task to pick up the capability and carry on running Haskell
1848 * threads. It also means that if the C call blocks, it won't lock
1851 * The Haskell thread making the C call is put to sleep for the
1852 * duration of the call, on the susepended_ccalling_threads queue. We
1853 * give out a token to the task, which it can use to resume the thread
1854 * on return from the C function.
1855 * ------------------------------------------------------------------------- */
1858 suspendThread (StgRegTable *reg)
1865 StgWord32 saved_winerror;
1868 saved_errno = errno;
1870 saved_winerror = GetLastError();
1873 /* assume that *reg is a pointer to the StgRegTable part of a Capability.
1875 cap = regTableToCapability(reg);
1877 task = cap->running_task;
1878 tso = cap->r.rCurrentTSO;
1880 traceEventStopThread(cap, tso, THREAD_SUSPENDED_FOREIGN_CALL);
1882 // XXX this might not be necessary --SDM
1883 tso->what_next = ThreadRunGHC;
1885 threadPaused(cap,tso);
1887 if ((tso->flags & TSO_BLOCKEX) == 0) {
1888 tso->why_blocked = BlockedOnCCall;
1889 tso->flags |= TSO_BLOCKEX;
1890 tso->flags &= ~TSO_INTERRUPTIBLE;
1892 tso->why_blocked = BlockedOnCCall_NoUnblockExc;
1895 // Hand back capability
1896 task->incall->suspended_tso = tso;
1897 task->incall->suspended_cap = cap;
1899 ACQUIRE_LOCK(&cap->lock);
1901 suspendTask(cap,task);
1902 cap->in_haskell = rtsFalse;
1903 releaseCapability_(cap,rtsFalse);
1905 RELEASE_LOCK(&cap->lock);
1907 errno = saved_errno;
1909 SetLastError(saved_winerror);
1915 resumeThread (void *task_)
1923 StgWord32 saved_winerror;
1926 saved_errno = errno;
1928 saved_winerror = GetLastError();
1931 incall = task->incall;
1932 cap = incall->suspended_cap;
1935 // Wait for permission to re-enter the RTS with the result.
1936 waitForReturnCapability(&cap,task);
1937 // we might be on a different capability now... but if so, our
1938 // entry on the suspended_ccalls list will also have been
1941 // Remove the thread from the suspended list
1942 recoverSuspendedTask(cap,task);
1944 tso = incall->suspended_tso;
1945 incall->suspended_tso = NULL;
1946 incall->suspended_cap = NULL;
1947 tso->_link = END_TSO_QUEUE; // no write barrier reqd
1949 traceEventRunThread(cap, tso);
1951 if (tso->why_blocked == BlockedOnCCall) {
1952 // avoid locking the TSO if we don't have to
1953 if (tso->blocked_exceptions != END_BLOCKED_EXCEPTIONS_QUEUE) {
1954 awakenBlockedExceptionQueue(cap,tso);
1956 tso->flags &= ~(TSO_BLOCKEX | TSO_INTERRUPTIBLE);
1959 /* Reset blocking status */
1960 tso->why_blocked = NotBlocked;
1962 cap->r.rCurrentTSO = tso;
1963 cap->in_haskell = rtsTrue;
1964 errno = saved_errno;
1966 SetLastError(saved_winerror);
1969 /* We might have GC'd, mark the TSO dirty again */
1972 IF_DEBUG(sanity, checkTSO(tso));
1977 /* ---------------------------------------------------------------------------
1980 * scheduleThread puts a thread on the end of the runnable queue.
1981 * This will usually be done immediately after a thread is created.
1982 * The caller of scheduleThread must create the thread using e.g.
1983 * createThread and push an appropriate closure
1984 * on this thread's stack before the scheduler is invoked.
1985 * ------------------------------------------------------------------------ */
1988 scheduleThread(Capability *cap, StgTSO *tso)
1990 // The thread goes at the *end* of the run-queue, to avoid possible
1991 // starvation of any threads already on the queue.
1992 appendToRunQueue(cap,tso);
1996 scheduleThreadOn(Capability *cap, StgWord cpu USED_IF_THREADS, StgTSO *tso)
1998 #if defined(THREADED_RTS)
1999 tso->flags |= TSO_LOCKED; // we requested explicit affinity; don't
2000 // move this thread from now on.
2001 cpu %= RtsFlags.ParFlags.nNodes;
2002 if (cpu == cap->no) {
2003 appendToRunQueue(cap,tso);
2005 traceEventMigrateThread (cap, tso, capabilities[cpu].no);
2006 wakeupThreadOnCapability(cap, &capabilities[cpu], tso);
2009 appendToRunQueue(cap,tso);
2014 scheduleWaitThread (StgTSO* tso, /*[out]*/HaskellObj* ret, Capability *cap)
2019 // We already created/initialised the Task
2020 task = cap->running_task;
2022 // This TSO is now a bound thread; make the Task and TSO
2023 // point to each other.
2024 tso->bound = task->incall;
2027 task->incall->tso = tso;
2029 task->stat = NoStatus;
2031 appendToRunQueue(cap,tso);
2034 debugTrace(DEBUG_sched, "new bound thread (%lu)", (unsigned long)id);
2036 cap = schedule(cap,task);
2038 ASSERT(task->stat != NoStatus);
2039 ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
2041 debugTrace(DEBUG_sched, "bound thread (%lu) finished", (unsigned long)id);
2045 /* ----------------------------------------------------------------------------
2047 * ------------------------------------------------------------------------- */
2049 #if defined(THREADED_RTS)
2050 void scheduleWorker (Capability *cap, Task *task)
2052 // schedule() runs without a lock.
2053 cap = schedule(cap,task);
2055 // On exit from schedule(), we have a Capability, but possibly not
2056 // the same one we started with.
2058 // During shutdown, the requirement is that after all the
2059 // Capabilities are shut down, all workers that are shutting down
2060 // have finished workerTaskStop(). This is why we hold on to
2061 // cap->lock until we've finished workerTaskStop() below.
2063 // There may be workers still involved in foreign calls; those
2064 // will just block in waitForReturnCapability() because the
2065 // Capability has been shut down.
2067 ACQUIRE_LOCK(&cap->lock);
2068 releaseCapability_(cap,rtsFalse);
2069 workerTaskStop(task);
2070 RELEASE_LOCK(&cap->lock);
2074 /* ---------------------------------------------------------------------------
2077 * Initialise the scheduler. This resets all the queues - if the
2078 * queues contained any threads, they'll be garbage collected at the
2081 * ------------------------------------------------------------------------ */
2086 #if !defined(THREADED_RTS)
2087 blocked_queue_hd = END_TSO_QUEUE;
2088 blocked_queue_tl = END_TSO_QUEUE;
2089 sleeping_queue = END_TSO_QUEUE;
2092 blackhole_queue = END_TSO_QUEUE;
2094 sched_state = SCHED_RUNNING;
2095 recent_activity = ACTIVITY_YES;
2097 #if defined(THREADED_RTS)
2098 /* Initialise the mutex and condition variables used by
2100 initMutex(&sched_mutex);
2103 ACQUIRE_LOCK(&sched_mutex);
2105 /* A capability holds the state a native thread needs in
2106 * order to execute STG code. At least one capability is
2107 * floating around (only THREADED_RTS builds have more than one).
2113 #if defined(THREADED_RTS)
2117 RELEASE_LOCK(&sched_mutex);
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);
2133 RELEASE_LOCK(&cap->lock);
2141 rtsBool wait_foreign
2142 #if !defined(THREADED_RTS)
2143 __attribute__((unused))
2146 /* see Capability.c, shutdownCapability() */
2150 task = newBoundTask();
2152 // If we haven't killed all the threads yet, do it now.
2153 if (sched_state < SCHED_SHUTTING_DOWN) {
2154 sched_state = SCHED_INTERRUPTING;
2155 waitForReturnCapability(&task->cap,task);
2156 scheduleDoGC(task->cap,task,rtsFalse);
2157 ASSERT(task->incall->tso == NULL);
2158 releaseCapability(task->cap);
2160 sched_state = SCHED_SHUTTING_DOWN;
2162 #if defined(THREADED_RTS)
2166 for (i = 0; i < n_capabilities; i++) {
2167 ASSERT(task->incall->tso == NULL);
2168 shutdownCapability(&capabilities[i], task, wait_foreign);
2173 boundTaskExiting(task);
2177 freeScheduler( void )
2181 ACQUIRE_LOCK(&sched_mutex);
2182 still_running = freeTaskManager();
2183 // We can only free the Capabilities if there are no Tasks still
2184 // running. We might have a Task about to return from a foreign
2185 // call into waitForReturnCapability(), for example (actually,
2186 // this should be the *only* thing that a still-running Task can
2187 // do at this point, and it will block waiting for the
2189 if (still_running == 0) {
2191 if (n_capabilities != 1) {
2192 stgFree(capabilities);
2195 RELEASE_LOCK(&sched_mutex);
2196 #if defined(THREADED_RTS)
2197 closeMutex(&sched_mutex);
2201 /* -----------------------------------------------------------------------------
2204 This is the interface to the garbage collector from Haskell land.
2205 We provide this so that external C code can allocate and garbage
2206 collect when called from Haskell via _ccall_GC.
2207 -------------------------------------------------------------------------- */
2210 performGC_(rtsBool force_major)
2214 // We must grab a new Task here, because the existing Task may be
2215 // associated with a particular Capability, and chained onto the
2216 // suspended_ccalls queue.
2217 task = newBoundTask();
2219 waitForReturnCapability(&task->cap,task);
2220 scheduleDoGC(task->cap,task,force_major);
2221 releaseCapability(task->cap);
2222 boundTaskExiting(task);
2228 performGC_(rtsFalse);
2232 performMajorGC(void)
2234 performGC_(rtsTrue);
2237 /* -----------------------------------------------------------------------------
2240 If the thread has reached its maximum stack size, then raise the
2241 StackOverflow exception in the offending thread. Otherwise
2242 relocate the TSO into a larger chunk of memory and adjust its stack
2244 -------------------------------------------------------------------------- */
2247 threadStackOverflow(Capability *cap, StgTSO *tso)
2249 nat new_stack_size, stack_words;
2254 IF_DEBUG(sanity,checkTSO(tso));
2256 if (tso->stack_size >= tso->max_stack_size
2257 && !(tso->flags & TSO_BLOCKEX)) {
2258 // NB. never raise a StackOverflow exception if the thread is
2259 // inside Control.Exceptino.block. It is impractical to protect
2260 // against stack overflow exceptions, since virtually anything
2261 // can raise one (even 'catch'), so this is the only sensible
2262 // thing to do here. See bug #767.
2265 if (tso->flags & TSO_SQUEEZED) {
2268 // #3677: In a stack overflow situation, stack squeezing may
2269 // reduce the stack size, but we don't know whether it has been
2270 // reduced enough for the stack check to succeed if we try
2271 // again. Fortunately stack squeezing is idempotent, so all we
2272 // need to do is record whether *any* squeezing happened. If we
2273 // are at the stack's absolute -K limit, and stack squeezing
2274 // happened, then we try running the thread again. The
2275 // TSO_SQUEEZED flag is set by threadPaused() to tell us whether
2276 // squeezing happened or not.
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
2287 throwToSingleThreaded(cap, tso, (StgClosure *)stackOverflow_closure);
2292 // We also want to avoid enlarging the stack if squeezing has
2293 // already released some of it. However, we don't want to get into
2294 // a pathalogical situation where a thread has a nearly full stack
2295 // (near its current limit, but not near the absolute -K limit),
2296 // keeps allocating a little bit, squeezing removes a little bit,
2297 // and then it runs again. So to avoid this, if we squeezed *and*
2298 // there is still less than BLOCK_SIZE_W words free, then we enlarge
2299 // the stack anyway.
2300 if ((tso->flags & TSO_SQUEEZED) &&
2301 ((W_)(tso->sp - tso->stack) >= BLOCK_SIZE_W)) {
2305 /* Try to double the current stack size. If that takes us over the
2306 * maximum stack size for this thread, then use the maximum instead
2307 * (that is, unless we're already at or over the max size and we
2308 * can't raise the StackOverflow exception (see above), in which
2309 * case just double the size). Finally round up so the TSO ends up as
2310 * a whole number of blocks.
2312 if (tso->stack_size >= tso->max_stack_size) {
2313 new_stack_size = tso->stack_size * 2;
2315 new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
2317 new_tso_size = (lnat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) +
2318 TSO_STRUCT_SIZE)/sizeof(W_);
2319 new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */
2320 new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
2322 debugTrace(DEBUG_sched,
2323 "increasing stack size from %ld words to %d.",
2324 (long)tso->stack_size, new_stack_size);
2326 dest = (StgTSO *)allocate(cap,new_tso_size);
2327 TICK_ALLOC_TSO(new_stack_size,0);
2329 /* copy the TSO block and the old stack into the new area */
2330 memcpy(dest,tso,TSO_STRUCT_SIZE);
2331 stack_words = tso->stack + tso->stack_size - tso->sp;
2332 new_sp = (P_)dest + new_tso_size - stack_words;
2333 memcpy(new_sp, tso->sp, stack_words * sizeof(W_));
2335 /* relocate the stack pointers... */
2337 dest->stack_size = new_stack_size;
2339 /* Mark the old TSO as relocated. We have to check for relocated
2340 * TSOs in the garbage collector and any primops that deal with TSOs.
2342 * It's important to set the sp value to just beyond the end
2343 * of the stack, so we don't attempt to scavenge any part of the
2346 tso->what_next = ThreadRelocated;
2347 setTSOLink(cap,tso,dest);
2348 tso->sp = (P_)&(tso->stack[tso->stack_size]);
2349 tso->why_blocked = NotBlocked;
2351 IF_DEBUG(sanity,checkTSO(dest));
2353 IF_DEBUG(scheduler,printTSO(dest));
2360 threadStackUnderflow (Capability *cap, Task *task, StgTSO *tso)
2362 bdescr *bd, *new_bd;
2363 lnat free_w, tso_size_w;
2366 tso_size_w = tso_sizeW(tso);
2368 if (tso_size_w < MBLOCK_SIZE_W ||
2369 // TSO is less than 2 mblocks (since the first mblock is
2370 // shorter than MBLOCK_SIZE_W)
2371 (tso_size_w - BLOCKS_PER_MBLOCK*BLOCK_SIZE_W) % MBLOCK_SIZE_W != 0 ||
2372 // or TSO is not a whole number of megablocks (ensuring
2373 // precondition of splitLargeBlock() below)
2374 (tso_size_w <= round_up_to_mblocks(RtsFlags.GcFlags.initialStkSize)) ||
2375 // or TSO is smaller than the minimum stack size (rounded up)
2376 (nat)(tso->stack + tso->stack_size - tso->sp) > tso->stack_size / 4)
2377 // or stack is using more than 1/4 of the available space
2383 // this is the number of words we'll free
2384 free_w = round_to_mblocks(tso_size_w/2);
2386 bd = Bdescr((StgPtr)tso);
2387 new_bd = splitLargeBlock(bd, free_w / BLOCK_SIZE_W);
2388 bd->free = bd->start + TSO_STRUCT_SIZEW;
2390 new_tso = (StgTSO *)new_bd->start;
2391 memcpy(new_tso,tso,TSO_STRUCT_SIZE);
2392 new_tso->stack_size = new_bd->free - new_tso->stack;
2394 // The original TSO was dirty and probably on the mutable
2395 // list. The new TSO is not yet on the mutable list, so we better
2398 new_tso->flags &= ~TSO_LINK_DIRTY;
2399 dirty_TSO(cap, new_tso);
2401 debugTrace(DEBUG_sched, "thread %ld: reducing TSO size from %lu words to %lu",
2402 (long)tso->id, tso_size_w, tso_sizeW(new_tso));
2404 tso->what_next = ThreadRelocated;
2405 tso->_link = new_tso; // no write barrier reqd: same generation
2407 // The TSO attached to this Task may have moved, so update the
2409 if (task->incall->tso == tso) {
2410 task->incall->tso = new_tso;
2413 IF_DEBUG(sanity,checkTSO(new_tso));
2418 /* ---------------------------------------------------------------------------
2420 - usually called inside a signal handler so it mustn't do anything fancy.
2421 ------------------------------------------------------------------------ */
2424 interruptStgRts(void)
2426 sched_state = SCHED_INTERRUPTING;
2427 setContextSwitches();
2428 #if defined(THREADED_RTS)
2433 /* -----------------------------------------------------------------------------
2436 This function causes at least one OS thread to wake up and run the
2437 scheduler loop. It is invoked when the RTS might be deadlocked, or
2438 an external event has arrived that may need servicing (eg. a
2439 keyboard interrupt).
2441 In the single-threaded RTS we don't do anything here; we only have
2442 one thread anyway, and the event that caused us to want to wake up
2443 will have interrupted any blocking system call in progress anyway.
2444 -------------------------------------------------------------------------- */
2446 #if defined(THREADED_RTS)
2447 void wakeUpRts(void)
2449 // This forces the IO Manager thread to wakeup, which will
2450 // in turn ensure that some OS thread wakes up and runs the
2451 // scheduler loop, which will cause a GC and deadlock check.
2456 /* -----------------------------------------------------------------------------
2459 * Check the blackhole_queue for threads that can be woken up. We do
2460 * this periodically: before every GC, and whenever the run queue is
2463 * An elegant solution might be to just wake up all the blocked
2464 * threads with awakenBlockedQueue occasionally: they'll go back to
2465 * sleep again if the object is still a BLACKHOLE. Unfortunately this
2466 * doesn't give us a way to tell whether we've actually managed to
2467 * wake up any threads, so we would be busy-waiting.
2469 * -------------------------------------------------------------------------- */
2472 checkBlackHoles (Capability *cap)
2475 rtsBool any_woke_up = rtsFalse;
2478 // blackhole_queue is global:
2479 ASSERT_LOCK_HELD(&sched_mutex);
2481 debugTrace(DEBUG_sched, "checking threads blocked on black holes");
2483 // ASSUMES: sched_mutex
2484 prev = &blackhole_queue;
2485 t = blackhole_queue;
2486 while (t != END_TSO_QUEUE) {
2487 if (t->what_next == ThreadRelocated) {
2491 ASSERT(t->why_blocked == BlockedOnBlackHole);
2492 type = get_itbl(UNTAG_CLOSURE(t->block_info.closure))->type;
2493 if (type != BLACKHOLE && type != CAF_BLACKHOLE) {
2494 IF_DEBUG(sanity,checkTSO(t));
2495 t = unblockOne(cap, t);
2497 any_woke_up = rtsTrue;
2507 /* -----------------------------------------------------------------------------
2510 This is used for interruption (^C) and forking, and corresponds to
2511 raising an exception but without letting the thread catch the
2513 -------------------------------------------------------------------------- */
2516 deleteThread (Capability *cap, StgTSO *tso)
2518 // NOTE: must only be called on a TSO that we have exclusive
2519 // access to, because we will call throwToSingleThreaded() below.
2520 // The TSO must be on the run queue of the Capability we own, or
2521 // we must own all Capabilities.
2523 if (tso->why_blocked != BlockedOnCCall &&
2524 tso->why_blocked != BlockedOnCCall_NoUnblockExc) {
2525 throwToSingleThreaded(cap,tso,NULL);
2529 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
2531 deleteThread_(Capability *cap, StgTSO *tso)
2532 { // for forkProcess only:
2533 // like deleteThread(), but we delete threads in foreign calls, too.
2535 if (tso->why_blocked == BlockedOnCCall ||
2536 tso->why_blocked == BlockedOnCCall_NoUnblockExc) {
2537 unblockOne(cap,tso);
2538 tso->what_next = ThreadKilled;
2540 deleteThread(cap,tso);
2545 /* -----------------------------------------------------------------------------
2546 raiseExceptionHelper
2548 This function is called by the raise# primitve, just so that we can
2549 move some of the tricky bits of raising an exception from C-- into
2550 C. Who knows, it might be a useful re-useable thing here too.
2551 -------------------------------------------------------------------------- */
2554 raiseExceptionHelper (StgRegTable *reg, StgTSO *tso, StgClosure *exception)
2556 Capability *cap = regTableToCapability(reg);
2557 StgThunk *raise_closure = NULL;
2559 StgRetInfoTable *info;
2561 // This closure represents the expression 'raise# E' where E
2562 // is the exception raise. It is used to overwrite all the
2563 // thunks which are currently under evaluataion.
2566 // OLD COMMENT (we don't have MIN_UPD_SIZE now):
2567 // LDV profiling: stg_raise_info has THUNK as its closure
2568 // type. Since a THUNK takes at least MIN_UPD_SIZE words in its
2569 // payload, MIN_UPD_SIZE is more approprate than 1. It seems that
2570 // 1 does not cause any problem unless profiling is performed.
2571 // However, when LDV profiling goes on, we need to linearly scan
2572 // small object pool, where raise_closure is stored, so we should
2573 // use MIN_UPD_SIZE.
2575 // raise_closure = (StgClosure *)RET_STGCALL1(P_,allocate,
2576 // sizeofW(StgClosure)+1);
2580 // Walk up the stack, looking for the catch frame. On the way,
2581 // we update any closures pointed to from update frames with the
2582 // raise closure that we just built.
2586 info = get_ret_itbl((StgClosure *)p);
2587 next = p + stack_frame_sizeW((StgClosure *)p);
2588 switch (info->i.type) {
2591 // Only create raise_closure if we need to.
2592 if (raise_closure == NULL) {
2594 (StgThunk *)allocate(cap,sizeofW(StgThunk)+1);
2595 SET_HDR(raise_closure, &stg_raise_info, CCCS);
2596 raise_closure->payload[0] = exception;
2598 UPD_IND(cap, ((StgUpdateFrame *)p)->updatee,
2599 (StgClosure *)raise_closure);
2603 case ATOMICALLY_FRAME:
2604 debugTrace(DEBUG_stm, "found ATOMICALLY_FRAME at %p", p);
2606 return ATOMICALLY_FRAME;
2612 case CATCH_STM_FRAME:
2613 debugTrace(DEBUG_stm, "found CATCH_STM_FRAME at %p", p);
2615 return CATCH_STM_FRAME;
2621 case CATCH_RETRY_FRAME:
2630 /* -----------------------------------------------------------------------------
2631 findRetryFrameHelper
2633 This function is called by the retry# primitive. It traverses the stack
2634 leaving tso->sp referring to the frame which should handle the retry.
2636 This should either be a CATCH_RETRY_FRAME (if the retry# is within an orElse#)
2637 or should be a ATOMICALLY_FRAME (if the retry# reaches the top level).
2639 We skip CATCH_STM_FRAMEs (aborting and rolling back the nested tx that they
2640 create) because retries are not considered to be exceptions, despite the
2641 similar implementation.
2643 We should not expect to see CATCH_FRAME or STOP_FRAME because those should
2644 not be created within memory transactions.
2645 -------------------------------------------------------------------------- */
2648 findRetryFrameHelper (StgTSO *tso)
2651 StgRetInfoTable *info;
2655 info = get_ret_itbl((StgClosure *)p);
2656 next = p + stack_frame_sizeW((StgClosure *)p);
2657 switch (info->i.type) {
2659 case ATOMICALLY_FRAME:
2660 debugTrace(DEBUG_stm,
2661 "found ATOMICALLY_FRAME at %p during retry", p);
2663 return ATOMICALLY_FRAME;
2665 case CATCH_RETRY_FRAME:
2666 debugTrace(DEBUG_stm,
2667 "found CATCH_RETRY_FRAME at %p during retrry", p);
2669 return CATCH_RETRY_FRAME;
2671 case CATCH_STM_FRAME: {
2672 StgTRecHeader *trec = tso -> trec;
2673 StgTRecHeader *outer = trec -> enclosing_trec;
2674 debugTrace(DEBUG_stm,
2675 "found CATCH_STM_FRAME at %p during retry", p);
2676 debugTrace(DEBUG_stm, "trec=%p outer=%p", trec, outer);
2677 stmAbortTransaction(tso -> cap, trec);
2678 stmFreeAbortedTRec(tso -> cap, trec);
2679 tso -> trec = outer;
2686 ASSERT(info->i.type != CATCH_FRAME);
2687 ASSERT(info->i.type != STOP_FRAME);
2694 /* -----------------------------------------------------------------------------
2695 resurrectThreads is called after garbage collection on the list of
2696 threads found to be garbage. Each of these threads will be woken
2697 up and sent a signal: BlockedOnDeadMVar if the thread was blocked
2698 on an MVar, or NonTermination if the thread was blocked on a Black
2701 Locks: assumes we hold *all* the capabilities.
2702 -------------------------------------------------------------------------- */
2705 resurrectThreads (StgTSO *threads)
2711 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
2712 next = tso->global_link;
2714 gen = Bdescr((P_)tso)->gen;
2715 tso->global_link = gen->threads;
2718 debugTrace(DEBUG_sched, "resurrecting thread %lu", (unsigned long)tso->id);
2720 // Wake up the thread on the Capability it was last on
2723 switch (tso->why_blocked) {
2725 /* Called by GC - sched_mutex lock is currently held. */
2726 throwToSingleThreaded(cap, tso,
2727 (StgClosure *)blockedIndefinitelyOnMVar_closure);
2729 case BlockedOnBlackHole:
2730 throwToSingleThreaded(cap, tso,
2731 (StgClosure *)nonTermination_closure);
2734 throwToSingleThreaded(cap, tso,
2735 (StgClosure *)blockedIndefinitelyOnSTM_closure);
2738 /* This might happen if the thread was blocked on a black hole
2739 * belonging to a thread that we've just woken up (raiseAsync
2740 * can wake up threads, remember...).
2744 barf("resurrectThreads: thread blocked in a strange way: %d",