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 scheduleCheckWakeupThreads(Capability *cap USED_IF_NOT_THREADS);
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 (task->tso == NULL && 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 Task *bound = t->bound;
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.
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->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 scheduleCheckWakeupThreads(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->tso == NULL
643 ? cap->run_queue_hd->bound != NULL
644 : cap->run_queue_hd->bound != task)));
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) ||
676 !emptyWakeupQueue(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) || cap->returning_tasks_hd != NULL) {
729 // it already has some work, we just grabbed it at
730 // the wrong moment. Or maybe it's deadlocked!
731 releaseCapability(cap0);
733 free_caps[n_free_caps++] = cap0;
738 // we now have n_free_caps free capabilities stashed in
739 // free_caps[]. Share our run queue equally with them. This is
740 // probably the simplest thing we could do; improvements we might
741 // want to do include:
743 // - giving high priority to moving relatively new threads, on
744 // the gournds that they haven't had time to build up a
745 // working set in the cache on this CPU/Capability.
747 // - giving low priority to moving long-lived threads
749 if (n_free_caps > 0) {
750 StgTSO *prev, *t, *next;
751 rtsBool pushed_to_all;
753 debugTrace(DEBUG_sched,
754 "cap %d: %s and %d free capabilities, sharing...",
756 (!emptyRunQueue(cap) && cap->run_queue_hd->_link != END_TSO_QUEUE)?
757 "excess threads on run queue":"sparks to share (>=2)",
761 pushed_to_all = rtsFalse;
763 if (cap->run_queue_hd != END_TSO_QUEUE) {
764 prev = cap->run_queue_hd;
766 prev->_link = END_TSO_QUEUE;
767 for (; t != END_TSO_QUEUE; t = next) {
769 t->_link = END_TSO_QUEUE;
770 if (t->what_next == ThreadRelocated
771 || t->bound == task // don't move my bound thread
772 || tsoLocked(t)) { // don't move a locked thread
773 setTSOLink(cap, prev, t);
775 } else if (i == n_free_caps) {
776 pushed_to_all = rtsTrue;
779 setTSOLink(cap, prev, t);
782 appendToRunQueue(free_caps[i],t);
784 traceEventMigrateThread (cap, t, free_caps[i]->no);
786 if (t->bound) { t->bound->cap = free_caps[i]; }
787 t->cap = free_caps[i];
791 cap->run_queue_tl = prev;
795 /* JB I left this code in place, it would work but is not necessary */
797 // If there are some free capabilities that we didn't push any
798 // threads to, then try to push a spark to each one.
799 if (!pushed_to_all) {
801 // i is the next free capability to push to
802 for (; i < n_free_caps; i++) {
803 if (emptySparkPoolCap(free_caps[i])) {
804 spark = tryStealSpark(cap->sparks);
806 debugTrace(DEBUG_sched, "pushing spark %p to capability %d", spark, free_caps[i]->no);
808 traceEventStealSpark(free_caps[i], t, cap->no);
810 newSpark(&(free_caps[i]->r), spark);
815 #endif /* SPARK_PUSHING */
817 // release the capabilities
818 for (i = 0; i < n_free_caps; i++) {
819 task->cap = free_caps[i];
820 releaseAndWakeupCapability(free_caps[i]);
823 task->cap = cap; // reset to point to our Capability.
825 #endif /* THREADED_RTS */
829 /* ----------------------------------------------------------------------------
830 * Start any pending signal handlers
831 * ------------------------------------------------------------------------- */
833 #if defined(RTS_USER_SIGNALS) && !defined(THREADED_RTS)
835 scheduleStartSignalHandlers(Capability *cap)
837 if (RtsFlags.MiscFlags.install_signal_handlers && signals_pending()) {
838 // safe outside the lock
839 startSignalHandlers(cap);
844 scheduleStartSignalHandlers(Capability *cap STG_UNUSED)
849 /* ----------------------------------------------------------------------------
850 * Check for blocked threads that can be woken up.
851 * ------------------------------------------------------------------------- */
854 scheduleCheckBlockedThreads(Capability *cap USED_IF_NOT_THREADS)
856 #if !defined(THREADED_RTS)
858 // Check whether any waiting threads need to be woken up. If the
859 // run queue is empty, and there are no other tasks running, we
860 // can wait indefinitely for something to happen.
862 if ( !emptyQueue(blocked_queue_hd) || !emptyQueue(sleeping_queue) )
864 awaitEvent( emptyRunQueue(cap) && !blackholes_need_checking );
870 /* ----------------------------------------------------------------------------
871 * Check for threads woken up by other Capabilities
872 * ------------------------------------------------------------------------- */
875 scheduleCheckWakeupThreads(Capability *cap USED_IF_THREADS)
877 #if defined(THREADED_RTS)
878 // Any threads that were woken up by other Capabilities get
879 // appended to our run queue.
880 if (!emptyWakeupQueue(cap)) {
881 ACQUIRE_LOCK(&cap->lock);
882 if (emptyRunQueue(cap)) {
883 cap->run_queue_hd = cap->wakeup_queue_hd;
884 cap->run_queue_tl = cap->wakeup_queue_tl;
886 setTSOLink(cap, cap->run_queue_tl, cap->wakeup_queue_hd);
887 cap->run_queue_tl = cap->wakeup_queue_tl;
889 cap->wakeup_queue_hd = cap->wakeup_queue_tl = END_TSO_QUEUE;
890 RELEASE_LOCK(&cap->lock);
895 /* ----------------------------------------------------------------------------
896 * Check for threads blocked on BLACKHOLEs that can be woken up
897 * ------------------------------------------------------------------------- */
899 scheduleCheckBlackHoles (Capability *cap)
901 if ( blackholes_need_checking ) // check without the lock first
903 ACQUIRE_LOCK(&sched_mutex);
904 if ( blackholes_need_checking ) {
905 blackholes_need_checking = rtsFalse;
906 // important that we reset the flag *before* checking the
907 // blackhole queue, otherwise we could get deadlock. This
908 // happens as follows: we wake up a thread that
909 // immediately runs on another Capability, blocks on a
910 // blackhole, and then we reset the blackholes_need_checking flag.
911 checkBlackHoles(cap);
913 RELEASE_LOCK(&sched_mutex);
917 /* ----------------------------------------------------------------------------
918 * Detect deadlock conditions and attempt to resolve them.
919 * ------------------------------------------------------------------------- */
922 scheduleDetectDeadlock (Capability *cap, Task *task)
925 * Detect deadlock: when we have no threads to run, there are no
926 * threads blocked, waiting for I/O, or sleeping, and all the
927 * other tasks are waiting for work, we must have a deadlock of
930 if ( emptyThreadQueues(cap) )
932 #if defined(THREADED_RTS)
934 * In the threaded RTS, we only check for deadlock if there
935 * has been no activity in a complete timeslice. This means
936 * we won't eagerly start a full GC just because we don't have
937 * any threads to run currently.
939 if (recent_activity != ACTIVITY_INACTIVE) return;
942 debugTrace(DEBUG_sched, "deadlocked, forcing major GC...");
944 // Garbage collection can release some new threads due to
945 // either (a) finalizers or (b) threads resurrected because
946 // they are unreachable and will therefore be sent an
947 // exception. Any threads thus released will be immediately
949 cap = scheduleDoGC (cap, task, rtsTrue/*force major GC*/);
950 // when force_major == rtsTrue. scheduleDoGC sets
951 // recent_activity to ACTIVITY_DONE_GC and turns off the timer
954 if ( !emptyRunQueue(cap) ) return;
956 #if defined(RTS_USER_SIGNALS) && !defined(THREADED_RTS)
957 /* If we have user-installed signal handlers, then wait
958 * for signals to arrive rather then bombing out with a
961 if ( RtsFlags.MiscFlags.install_signal_handlers && anyUserHandlers() ) {
962 debugTrace(DEBUG_sched,
963 "still deadlocked, waiting for signals...");
967 if (signals_pending()) {
968 startSignalHandlers(cap);
971 // either we have threads to run, or we were interrupted:
972 ASSERT(!emptyRunQueue(cap) || sched_state >= SCHED_INTERRUPTING);
978 #if !defined(THREADED_RTS)
979 /* Probably a real deadlock. Send the current main thread the
980 * Deadlock exception.
983 switch (task->tso->why_blocked) {
985 case BlockedOnBlackHole:
986 case BlockedOnException:
988 throwToSingleThreaded(cap, task->tso,
989 (StgClosure *)nonTermination_closure);
992 barf("deadlock: main thread blocked in a strange way");
1001 /* ----------------------------------------------------------------------------
1002 * Send pending messages (PARALLEL_HASKELL only)
1003 * ------------------------------------------------------------------------- */
1005 #if defined(PARALLEL_HASKELL)
1007 scheduleSendPendingMessages(void)
1010 # if defined(PAR) // global Mem.Mgmt., omit for now
1011 if (PendingFetches != END_BF_QUEUE) {
1016 if (RtsFlags.ParFlags.BufferTime) {
1017 // if we use message buffering, we must send away all message
1018 // packets which have become too old...
1024 /* ----------------------------------------------------------------------------
1025 * Activate spark threads (PARALLEL_HASKELL and THREADED_RTS)
1026 * ------------------------------------------------------------------------- */
1028 #if defined(THREADED_RTS)
1030 scheduleActivateSpark(Capability *cap)
1034 createSparkThread(cap);
1035 debugTrace(DEBUG_sched, "creating a spark thread");
1038 #endif // PARALLEL_HASKELL || THREADED_RTS
1040 /* ----------------------------------------------------------------------------
1041 * After running a thread...
1042 * ------------------------------------------------------------------------- */
1045 schedulePostRunThread (Capability *cap, StgTSO *t)
1047 // We have to be able to catch transactions that are in an
1048 // infinite loop as a result of seeing an inconsistent view of
1052 // [a,b] <- mapM readTVar [ta,tb]
1053 // when (a == b) loop
1055 // and a is never equal to b given a consistent view of memory.
1057 if (t -> trec != NO_TREC && t -> why_blocked == NotBlocked) {
1058 if (!stmValidateNestOfTransactions (t -> trec)) {
1059 debugTrace(DEBUG_sched | DEBUG_stm,
1060 "trec %p found wasting its time", t);
1062 // strip the stack back to the
1063 // ATOMICALLY_FRAME, aborting the (nested)
1064 // transaction, and saving the stack of any
1065 // partially-evaluated thunks on the heap.
1066 throwToSingleThreaded_(cap, t, NULL, rtsTrue);
1068 // ASSERT(get_itbl((StgClosure *)t->sp)->type == ATOMICALLY_FRAME);
1072 /* some statistics gathering in the parallel case */
1075 /* -----------------------------------------------------------------------------
1076 * Handle a thread that returned to the scheduler with ThreadHeepOverflow
1077 * -------------------------------------------------------------------------- */
1080 scheduleHandleHeapOverflow( Capability *cap, StgTSO *t )
1082 // did the task ask for a large block?
1083 if (cap->r.rHpAlloc > BLOCK_SIZE) {
1084 // if so, get one and push it on the front of the nursery.
1088 blocks = (lnat)BLOCK_ROUND_UP(cap->r.rHpAlloc) / BLOCK_SIZE;
1090 debugTrace(DEBUG_sched,
1091 "--<< thread %ld (%s) stopped: requesting a large block (size %ld)\n",
1092 (long)t->id, what_next_strs[t->what_next], blocks);
1094 // don't do this if the nursery is (nearly) full, we'll GC first.
1095 if (cap->r.rCurrentNursery->link != NULL ||
1096 cap->r.rNursery->n_blocks == 1) { // paranoia to prevent infinite loop
1097 // if the nursery has only one block.
1100 bd = allocGroup( blocks );
1102 cap->r.rNursery->n_blocks += blocks;
1104 // link the new group into the list
1105 bd->link = cap->r.rCurrentNursery;
1106 bd->u.back = cap->r.rCurrentNursery->u.back;
1107 if (cap->r.rCurrentNursery->u.back != NULL) {
1108 cap->r.rCurrentNursery->u.back->link = bd;
1110 cap->r.rNursery->blocks = bd;
1112 cap->r.rCurrentNursery->u.back = bd;
1114 // initialise it as a nursery block. We initialise the
1115 // step, gen_no, and flags field of *every* sub-block in
1116 // this large block, because this is easier than making
1117 // sure that we always find the block head of a large
1118 // block whenever we call Bdescr() (eg. evacuate() and
1119 // isAlive() in the GC would both have to do this, at
1123 for (x = bd; x < bd + blocks; x++) {
1124 initBdescr(x,g0,g0);
1130 // This assert can be a killer if the app is doing lots
1131 // of large block allocations.
1132 IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
1134 // now update the nursery to point to the new block
1135 cap->r.rCurrentNursery = bd;
1137 // we might be unlucky and have another thread get on the
1138 // run queue before us and steal the large block, but in that
1139 // case the thread will just end up requesting another large
1141 pushOnRunQueue(cap,t);
1142 return rtsFalse; /* not actually GC'ing */
1146 if (cap->r.rHpLim == NULL || cap->context_switch) {
1147 // Sometimes we miss a context switch, e.g. when calling
1148 // primitives in a tight loop, MAYBE_GC() doesn't check the
1149 // context switch flag, and we end up waiting for a GC.
1150 // See #1984, and concurrent/should_run/1984
1151 cap->context_switch = 0;
1152 addToRunQueue(cap,t);
1154 pushOnRunQueue(cap,t);
1157 /* actual GC is done at the end of the while loop in schedule() */
1160 /* -----------------------------------------------------------------------------
1161 * Handle a thread that returned to the scheduler with ThreadStackOverflow
1162 * -------------------------------------------------------------------------- */
1165 scheduleHandleStackOverflow (Capability *cap, Task *task, StgTSO *t)
1167 /* just adjust the stack for this thread, then pop it back
1171 /* enlarge the stack */
1172 StgTSO *new_t = threadStackOverflow(cap, t);
1174 /* The TSO attached to this Task may have moved, so update the
1177 if (task->tso == t) {
1180 pushOnRunQueue(cap,new_t);
1184 /* -----------------------------------------------------------------------------
1185 * Handle a thread that returned to the scheduler with ThreadYielding
1186 * -------------------------------------------------------------------------- */
1189 scheduleHandleYield( Capability *cap, StgTSO *t, nat prev_what_next )
1191 // Reset the context switch flag. We don't do this just before
1192 // running the thread, because that would mean we would lose ticks
1193 // during GC, which can lead to unfair scheduling (a thread hogs
1194 // the CPU because the tick always arrives during GC). This way
1195 // penalises threads that do a lot of allocation, but that seems
1196 // better than the alternative.
1197 cap->context_switch = 0;
1199 /* put the thread back on the run queue. Then, if we're ready to
1200 * GC, check whether this is the last task to stop. If so, wake
1201 * up the GC thread. getThread will block during a GC until the
1205 if (t->what_next != prev_what_next) {
1206 debugTrace(DEBUG_sched,
1207 "--<< thread %ld (%s) stopped to switch evaluators",
1208 (long)t->id, what_next_strs[t->what_next]);
1212 ASSERT(t->_link == END_TSO_QUEUE);
1214 // Shortcut if we're just switching evaluators: don't bother
1215 // doing stack squeezing (which can be expensive), just run the
1217 if (t->what_next != prev_what_next) {
1222 //debugBelch("&& Doing sanity check on yielding TSO %ld.", t->id);
1225 addToRunQueue(cap,t);
1230 /* -----------------------------------------------------------------------------
1231 * Handle a thread that returned to the scheduler with ThreadBlocked
1232 * -------------------------------------------------------------------------- */
1235 scheduleHandleThreadBlocked( StgTSO *t
1242 // We don't need to do anything. The thread is blocked, and it
1243 // has tidied up its stack and placed itself on whatever queue
1244 // it needs to be on.
1246 // ASSERT(t->why_blocked != NotBlocked);
1247 // Not true: for example,
1248 // - in THREADED_RTS, the thread may already have been woken
1249 // up by another Capability. This actually happens: try
1250 // conc023 +RTS -N2.
1251 // - the thread may have woken itself up already, because
1252 // threadPaused() might have raised a blocked throwTo
1253 // exception, see maybePerformBlockedException().
1256 traceThreadStatus(DEBUG_sched, t);
1260 /* -----------------------------------------------------------------------------
1261 * Handle a thread that returned to the scheduler with ThreadFinished
1262 * -------------------------------------------------------------------------- */
1265 scheduleHandleThreadFinished (Capability *cap STG_UNUSED, Task *task, StgTSO *t)
1267 /* Need to check whether this was a main thread, and if so,
1268 * return with the return value.
1270 * We also end up here if the thread kills itself with an
1271 * uncaught exception, see Exception.cmm.
1274 // blocked exceptions can now complete, even if the thread was in
1275 // blocked mode (see #2910). This unconditionally calls
1276 // lockTSO(), which ensures that we don't miss any threads that
1277 // are engaged in throwTo() with this thread as a target.
1278 awakenBlockedExceptionQueue (cap, t);
1281 // Check whether the thread that just completed was a bound
1282 // thread, and if so return with the result.
1284 // There is an assumption here that all thread completion goes
1285 // through this point; we need to make sure that if a thread
1286 // ends up in the ThreadKilled state, that it stays on the run
1287 // queue so it can be dealt with here.
1292 if (t->bound != task) {
1293 #if !defined(THREADED_RTS)
1294 // Must be a bound thread that is not the topmost one. Leave
1295 // it on the run queue until the stack has unwound to the
1296 // point where we can deal with this. Leaving it on the run
1297 // queue also ensures that the garbage collector knows about
1298 // this thread and its return value (it gets dropped from the
1299 // step->threads list so there's no other way to find it).
1300 appendToRunQueue(cap,t);
1303 // this cannot happen in the threaded RTS, because a
1304 // bound thread can only be run by the appropriate Task.
1305 barf("finished bound thread that isn't mine");
1309 ASSERT(task->tso == t);
1311 if (t->what_next == ThreadComplete) {
1313 // NOTE: return val is tso->sp[1] (see StgStartup.hc)
1314 *(task->ret) = (StgClosure *)task->tso->sp[1];
1316 task->stat = Success;
1319 *(task->ret) = NULL;
1321 if (sched_state >= SCHED_INTERRUPTING) {
1322 if (heap_overflow) {
1323 task->stat = HeapExhausted;
1325 task->stat = Interrupted;
1328 task->stat = Killed;
1332 removeThreadLabel((StgWord)task->tso->id);
1335 // We no longer consider this thread and task to be bound to
1336 // each other. The TSO lives on until it is GC'd, but the
1337 // task is about to be released by the caller, and we don't
1338 // want anyone following the pointer from the TSO to the
1339 // defunct task (which might have already been
1340 // re-used). This was a real bug: the GC updated
1341 // tso->bound->tso which lead to a deadlock.
1345 return rtsTrue; // tells schedule() to return
1351 /* -----------------------------------------------------------------------------
1352 * Perform a heap census
1353 * -------------------------------------------------------------------------- */
1356 scheduleNeedHeapProfile( rtsBool ready_to_gc STG_UNUSED )
1358 // When we have +RTS -i0 and we're heap profiling, do a census at
1359 // every GC. This lets us get repeatable runs for debugging.
1360 if (performHeapProfile ||
1361 (RtsFlags.ProfFlags.profileInterval==0 &&
1362 RtsFlags.ProfFlags.doHeapProfile && ready_to_gc)) {
1369 /* -----------------------------------------------------------------------------
1370 * Perform a garbage collection if necessary
1371 * -------------------------------------------------------------------------- */
1374 scheduleDoGC (Capability *cap, Task *task USED_IF_THREADS, rtsBool force_major)
1376 rtsBool heap_census;
1378 /* extern static volatile StgWord waiting_for_gc;
1379 lives inside capability.c */
1380 rtsBool gc_type, prev_pending_gc;
1384 if (sched_state == SCHED_SHUTTING_DOWN) {
1385 // The final GC has already been done, and the system is
1386 // shutting down. We'll probably deadlock if we try to GC
1392 if (sched_state < SCHED_INTERRUPTING
1393 && RtsFlags.ParFlags.parGcEnabled
1394 && N >= RtsFlags.ParFlags.parGcGen
1395 && ! oldest_gen->mark)
1397 gc_type = PENDING_GC_PAR;
1399 gc_type = PENDING_GC_SEQ;
1402 // In order to GC, there must be no threads running Haskell code.
1403 // Therefore, the GC thread needs to hold *all* the capabilities,
1404 // and release them after the GC has completed.
1406 // This seems to be the simplest way: previous attempts involved
1407 // making all the threads with capabilities give up their
1408 // capabilities and sleep except for the *last* one, which
1409 // actually did the GC. But it's quite hard to arrange for all
1410 // the other tasks to sleep and stay asleep.
1413 /* Other capabilities are prevented from running yet more Haskell
1414 threads if waiting_for_gc is set. Tested inside
1415 yieldCapability() and releaseCapability() in Capability.c */
1417 prev_pending_gc = cas(&waiting_for_gc, 0, gc_type);
1418 if (prev_pending_gc) {
1420 debugTrace(DEBUG_sched, "someone else is trying to GC (%d)...",
1423 yieldCapability(&cap,task);
1424 } while (waiting_for_gc);
1425 return cap; // NOTE: task->cap might have changed here
1428 setContextSwitches();
1430 // The final shutdown GC is always single-threaded, because it's
1431 // possible that some of the Capabilities have no worker threads.
1433 if (gc_type == PENDING_GC_SEQ)
1435 traceEventRequestSeqGc(cap);
1439 traceEventRequestParGc(cap);
1440 debugTrace(DEBUG_sched, "ready_to_gc, grabbing GC threads");
1443 // do this while the other Capabilities stop:
1444 if (cap) scheduleCheckBlackHoles(cap);
1446 if (gc_type == PENDING_GC_SEQ)
1448 // single-threaded GC: grab all the capabilities
1449 for (i=0; i < n_capabilities; i++) {
1450 debugTrace(DEBUG_sched, "ready_to_gc, grabbing all the capabilies (%d/%d)", i, n_capabilities);
1451 if (cap != &capabilities[i]) {
1452 Capability *pcap = &capabilities[i];
1453 // we better hope this task doesn't get migrated to
1454 // another Capability while we're waiting for this one.
1455 // It won't, because load balancing happens while we have
1456 // all the Capabilities, but even so it's a slightly
1457 // unsavoury invariant.
1459 waitForReturnCapability(&pcap, task);
1460 if (pcap != &capabilities[i]) {
1461 barf("scheduleDoGC: got the wrong capability");
1468 // multi-threaded GC: make sure all the Capabilities donate one
1470 waitForGcThreads(cap);
1473 #else /* !THREADED_RTS */
1475 // do this while the other Capabilities stop:
1476 if (cap) scheduleCheckBlackHoles(cap);
1480 IF_DEBUG(scheduler, printAllThreads());
1482 delete_threads_and_gc:
1484 * We now have all the capabilities; if we're in an interrupting
1485 * state, then we should take the opportunity to delete all the
1486 * threads in the system.
1488 if (sched_state == SCHED_INTERRUPTING) {
1489 deleteAllThreads(cap);
1490 sched_state = SCHED_SHUTTING_DOWN;
1493 heap_census = scheduleNeedHeapProfile(rtsTrue);
1495 traceEventGcStart(cap);
1496 #if defined(THREADED_RTS)
1497 // reset waiting_for_gc *before* GC, so that when the GC threads
1498 // emerge they don't immediately re-enter the GC.
1500 GarbageCollect(force_major || heap_census, gc_type, cap);
1502 GarbageCollect(force_major || heap_census, 0, cap);
1504 traceEventGcEnd(cap);
1506 if (recent_activity == ACTIVITY_INACTIVE && force_major)
1508 // We are doing a GC because the system has been idle for a
1509 // timeslice and we need to check for deadlock. Record the
1510 // fact that we've done a GC and turn off the timer signal;
1511 // it will get re-enabled if we run any threads after the GC.
1512 recent_activity = ACTIVITY_DONE_GC;
1517 // the GC might have taken long enough for the timer to set
1518 // recent_activity = ACTIVITY_INACTIVE, but we aren't
1519 // necessarily deadlocked:
1520 recent_activity = ACTIVITY_YES;
1523 #if defined(THREADED_RTS)
1524 if (gc_type == PENDING_GC_PAR)
1526 releaseGCThreads(cap);
1531 debugTrace(DEBUG_sched, "performing heap census");
1533 performHeapProfile = rtsFalse;
1536 if (heap_overflow && sched_state < SCHED_INTERRUPTING) {
1537 // GC set the heap_overflow flag, so we should proceed with
1538 // an orderly shutdown now. Ultimately we want the main
1539 // thread to return to its caller with HeapExhausted, at which
1540 // point the caller should call hs_exit(). The first step is
1541 // to delete all the threads.
1543 // Another way to do this would be to raise an exception in
1544 // the main thread, which we really should do because it gives
1545 // the program a chance to clean up. But how do we find the
1546 // main thread? It should presumably be the same one that
1547 // gets ^C exceptions, but that's all done on the Haskell side
1548 // (GHC.TopHandler).
1549 sched_state = SCHED_INTERRUPTING;
1550 goto delete_threads_and_gc;
1555 Once we are all together... this would be the place to balance all
1556 spark pools. No concurrent stealing or adding of new sparks can
1557 occur. Should be defined in Sparks.c. */
1558 balanceSparkPoolsCaps(n_capabilities, capabilities);
1561 #if defined(THREADED_RTS)
1562 if (gc_type == PENDING_GC_SEQ) {
1563 // release our stash of capabilities.
1564 for (i = 0; i < n_capabilities; i++) {
1565 if (cap != &capabilities[i]) {
1566 task->cap = &capabilities[i];
1567 releaseCapability(&capabilities[i]);
1581 /* ---------------------------------------------------------------------------
1582 * Singleton fork(). Do not copy any running threads.
1583 * ------------------------------------------------------------------------- */
1586 forkProcess(HsStablePtr *entry
1587 #ifndef FORKPROCESS_PRIMOP_SUPPORTED
1592 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
1599 #if defined(THREADED_RTS)
1600 if (RtsFlags.ParFlags.nNodes > 1) {
1601 errorBelch("forking not supported with +RTS -N<n> greater than 1");
1602 stg_exit(EXIT_FAILURE);
1606 debugTrace(DEBUG_sched, "forking!");
1608 // ToDo: for SMP, we should probably acquire *all* the capabilities
1611 // no funny business: hold locks while we fork, otherwise if some
1612 // other thread is holding a lock when the fork happens, the data
1613 // structure protected by the lock will forever be in an
1614 // inconsistent state in the child. See also #1391.
1615 ACQUIRE_LOCK(&sched_mutex);
1616 ACQUIRE_LOCK(&cap->lock);
1617 ACQUIRE_LOCK(&cap->running_task->lock);
1621 if (pid) { // parent
1623 RELEASE_LOCK(&sched_mutex);
1624 RELEASE_LOCK(&cap->lock);
1625 RELEASE_LOCK(&cap->running_task->lock);
1627 // just return the pid
1633 #if defined(THREADED_RTS)
1634 initMutex(&sched_mutex);
1635 initMutex(&cap->lock);
1636 initMutex(&cap->running_task->lock);
1639 // Now, all OS threads except the thread that forked are
1640 // stopped. We need to stop all Haskell threads, including
1641 // those involved in foreign calls. Also we need to delete
1642 // all Tasks, because they correspond to OS threads that are
1645 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1646 for (t = generations[g].threads; t != END_TSO_QUEUE; t = next) {
1647 if (t->what_next == ThreadRelocated) {
1650 next = t->global_link;
1651 // don't allow threads to catch the ThreadKilled
1652 // exception, but we do want to raiseAsync() because these
1653 // threads may be evaluating thunks that we need later.
1654 deleteThread_(cap,t);
1659 // Empty the run queue. It seems tempting to let all the
1660 // killed threads stay on the run queue as zombies to be
1661 // cleaned up later, but some of them correspond to bound
1662 // threads for which the corresponding Task does not exist.
1663 cap->run_queue_hd = END_TSO_QUEUE;
1664 cap->run_queue_tl = END_TSO_QUEUE;
1666 // Any suspended C-calling Tasks are no more, their OS threads
1668 cap->suspended_ccalling_tasks = NULL;
1670 // Empty the threads lists. Otherwise, the garbage
1671 // collector may attempt to resurrect some of these threads.
1672 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1673 generations[g].threads = END_TSO_QUEUE;
1676 // Wipe the task list, except the current Task.
1677 ACQUIRE_LOCK(&sched_mutex);
1678 for (task = all_tasks; task != NULL; task=task->all_link) {
1679 if (task != cap->running_task) {
1680 #if defined(THREADED_RTS)
1681 initMutex(&task->lock); // see #1391
1686 RELEASE_LOCK(&sched_mutex);
1688 #if defined(THREADED_RTS)
1689 // Wipe our spare workers list, they no longer exist. New
1690 // workers will be created if necessary.
1691 cap->spare_workers = NULL;
1692 cap->returning_tasks_hd = NULL;
1693 cap->returning_tasks_tl = NULL;
1696 // On Unix, all timers are reset in the child, so we need to start
1701 #if defined(THREADED_RTS)
1702 cap = ioManagerStartCap(cap);
1705 cap = rts_evalStableIO(cap, entry, NULL); // run the action
1706 rts_checkSchedStatus("forkProcess",cap);
1709 hs_exit(); // clean up and exit
1710 stg_exit(EXIT_SUCCESS);
1712 #else /* !FORKPROCESS_PRIMOP_SUPPORTED */
1713 barf("forkProcess#: primop not supported on this platform, sorry!\n");
1717 /* ---------------------------------------------------------------------------
1718 * Delete all the threads in the system
1719 * ------------------------------------------------------------------------- */
1722 deleteAllThreads ( Capability *cap )
1724 // NOTE: only safe to call if we own all capabilities.
1729 debugTrace(DEBUG_sched,"deleting all threads");
1730 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1731 for (t = generations[g].threads; t != END_TSO_QUEUE; t = next) {
1732 if (t->what_next == ThreadRelocated) {
1735 next = t->global_link;
1736 deleteThread(cap,t);
1741 // The run queue now contains a bunch of ThreadKilled threads. We
1742 // must not throw these away: the main thread(s) will be in there
1743 // somewhere, and the main scheduler loop has to deal with it.
1744 // Also, the run queue is the only thing keeping these threads from
1745 // being GC'd, and we don't want the "main thread has been GC'd" panic.
1747 #if !defined(THREADED_RTS)
1748 ASSERT(blocked_queue_hd == END_TSO_QUEUE);
1749 ASSERT(sleeping_queue == END_TSO_QUEUE);
1753 /* -----------------------------------------------------------------------------
1754 Managing the suspended_ccalling_tasks list.
1755 Locks required: sched_mutex
1756 -------------------------------------------------------------------------- */
1759 suspendTask (Capability *cap, Task *task)
1761 ASSERT(task->next == NULL && task->prev == NULL);
1762 task->next = cap->suspended_ccalling_tasks;
1764 if (cap->suspended_ccalling_tasks) {
1765 cap->suspended_ccalling_tasks->prev = task;
1767 cap->suspended_ccalling_tasks = task;
1771 recoverSuspendedTask (Capability *cap, Task *task)
1774 task->prev->next = task->next;
1776 ASSERT(cap->suspended_ccalling_tasks == task);
1777 cap->suspended_ccalling_tasks = task->next;
1780 task->next->prev = task->prev;
1782 task->next = task->prev = NULL;
1785 /* ---------------------------------------------------------------------------
1786 * Suspending & resuming Haskell threads.
1788 * When making a "safe" call to C (aka _ccall_GC), the task gives back
1789 * its capability before calling the C function. This allows another
1790 * task to pick up the capability and carry on running Haskell
1791 * threads. It also means that if the C call blocks, it won't lock
1794 * The Haskell thread making the C call is put to sleep for the
1795 * duration of the call, on the susepended_ccalling_threads queue. We
1796 * give out a token to the task, which it can use to resume the thread
1797 * on return from the C function.
1798 * ------------------------------------------------------------------------- */
1801 suspendThread (StgRegTable *reg)
1808 StgWord32 saved_winerror;
1811 saved_errno = errno;
1813 saved_winerror = GetLastError();
1816 /* assume that *reg is a pointer to the StgRegTable part of a Capability.
1818 cap = regTableToCapability(reg);
1820 task = cap->running_task;
1821 tso = cap->r.rCurrentTSO;
1823 traceEventStopThread(cap, tso, THREAD_SUSPENDED_FOREIGN_CALL);
1825 // XXX this might not be necessary --SDM
1826 tso->what_next = ThreadRunGHC;
1828 threadPaused(cap,tso);
1830 if ((tso->flags & TSO_BLOCKEX) == 0) {
1831 tso->why_blocked = BlockedOnCCall;
1832 tso->flags |= TSO_BLOCKEX;
1833 tso->flags &= ~TSO_INTERRUPTIBLE;
1835 tso->why_blocked = BlockedOnCCall_NoUnblockExc;
1838 // Hand back capability
1839 task->suspended_tso = tso;
1841 ACQUIRE_LOCK(&cap->lock);
1843 suspendTask(cap,task);
1844 cap->in_haskell = rtsFalse;
1845 releaseCapability_(cap,rtsFalse);
1847 RELEASE_LOCK(&cap->lock);
1849 errno = saved_errno;
1851 SetLastError(saved_winerror);
1857 resumeThread (void *task_)
1864 StgWord32 saved_winerror;
1867 saved_errno = errno;
1869 saved_winerror = GetLastError();
1873 // Wait for permission to re-enter the RTS with the result.
1874 waitForReturnCapability(&cap,task);
1875 // we might be on a different capability now... but if so, our
1876 // entry on the suspended_ccalling_tasks list will also have been
1879 // Remove the thread from the suspended list
1880 recoverSuspendedTask(cap,task);
1882 tso = task->suspended_tso;
1883 task->suspended_tso = NULL;
1884 tso->_link = END_TSO_QUEUE; // no write barrier reqd
1886 traceEventRunThread(cap, tso);
1888 if (tso->why_blocked == BlockedOnCCall) {
1889 // avoid locking the TSO if we don't have to
1890 if (tso->blocked_exceptions != END_TSO_QUEUE) {
1891 awakenBlockedExceptionQueue(cap,tso);
1893 tso->flags &= ~(TSO_BLOCKEX | TSO_INTERRUPTIBLE);
1896 /* Reset blocking status */
1897 tso->why_blocked = NotBlocked;
1899 cap->r.rCurrentTSO = tso;
1900 cap->in_haskell = rtsTrue;
1901 errno = saved_errno;
1903 SetLastError(saved_winerror);
1906 /* We might have GC'd, mark the TSO dirty again */
1909 IF_DEBUG(sanity, checkTSO(tso));
1914 /* ---------------------------------------------------------------------------
1917 * scheduleThread puts a thread on the end of the runnable queue.
1918 * This will usually be done immediately after a thread is created.
1919 * The caller of scheduleThread must create the thread using e.g.
1920 * createThread and push an appropriate closure
1921 * on this thread's stack before the scheduler is invoked.
1922 * ------------------------------------------------------------------------ */
1925 scheduleThread(Capability *cap, StgTSO *tso)
1927 // The thread goes at the *end* of the run-queue, to avoid possible
1928 // starvation of any threads already on the queue.
1929 appendToRunQueue(cap,tso);
1933 scheduleThreadOn(Capability *cap, StgWord cpu USED_IF_THREADS, StgTSO *tso)
1935 #if defined(THREADED_RTS)
1936 tso->flags |= TSO_LOCKED; // we requested explicit affinity; don't
1937 // move this thread from now on.
1938 cpu %= RtsFlags.ParFlags.nNodes;
1939 if (cpu == cap->no) {
1940 appendToRunQueue(cap,tso);
1942 traceEventMigrateThread (cap, tso, capabilities[cpu].no);
1943 wakeupThreadOnCapability(cap, &capabilities[cpu], tso);
1946 appendToRunQueue(cap,tso);
1951 scheduleWaitThread (StgTSO* tso, /*[out]*/HaskellObj* ret, Capability *cap)
1956 // We already created/initialised the Task
1957 task = cap->running_task;
1959 // This TSO is now a bound thread; make the Task and TSO
1960 // point to each other.
1966 task->stat = NoStatus;
1968 appendToRunQueue(cap,tso);
1971 debugTrace(DEBUG_sched, "new bound thread (%lu)", (unsigned long)id);
1973 cap = schedule(cap,task);
1975 ASSERT(task->stat != NoStatus);
1976 ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
1978 debugTrace(DEBUG_sched, "bound thread (%lu) finished", (unsigned long)id);
1982 /* ----------------------------------------------------------------------------
1984 * ------------------------------------------------------------------------- */
1986 #if defined(THREADED_RTS)
1987 void OSThreadProcAttr
1988 workerStart(Task *task)
1992 // See startWorkerTask().
1993 ACQUIRE_LOCK(&task->lock);
1995 RELEASE_LOCK(&task->lock);
1997 if (RtsFlags.ParFlags.setAffinity) {
1998 setThreadAffinity(cap->no, n_capabilities);
2001 // set the thread-local pointer to the Task:
2004 // schedule() runs without a lock.
2005 cap = schedule(cap,task);
2007 // On exit from schedule(), we have a Capability, but possibly not
2008 // the same one we started with.
2010 // During shutdown, the requirement is that after all the
2011 // Capabilities are shut down, all workers that are shutting down
2012 // have finished workerTaskStop(). This is why we hold on to
2013 // cap->lock until we've finished workerTaskStop() below.
2015 // There may be workers still involved in foreign calls; those
2016 // will just block in waitForReturnCapability() because the
2017 // Capability has been shut down.
2019 ACQUIRE_LOCK(&cap->lock);
2020 releaseCapability_(cap,rtsFalse);
2021 workerTaskStop(task);
2022 RELEASE_LOCK(&cap->lock);
2026 /* ---------------------------------------------------------------------------
2029 * Initialise the scheduler. This resets all the queues - if the
2030 * queues contained any threads, they'll be garbage collected at the
2033 * ------------------------------------------------------------------------ */
2038 #if !defined(THREADED_RTS)
2039 blocked_queue_hd = END_TSO_QUEUE;
2040 blocked_queue_tl = END_TSO_QUEUE;
2041 sleeping_queue = END_TSO_QUEUE;
2044 blackhole_queue = END_TSO_QUEUE;
2046 sched_state = SCHED_RUNNING;
2047 recent_activity = ACTIVITY_YES;
2049 #if defined(THREADED_RTS)
2050 /* Initialise the mutex and condition variables used by
2052 initMutex(&sched_mutex);
2055 ACQUIRE_LOCK(&sched_mutex);
2057 /* A capability holds the state a native thread needs in
2058 * order to execute STG code. At least one capability is
2059 * floating around (only THREADED_RTS builds have more than one).
2065 #if defined(THREADED_RTS)
2069 #if defined(THREADED_RTS)
2071 * Eagerly start one worker to run each Capability, except for
2072 * Capability 0. The idea is that we're probably going to start a
2073 * bound thread on Capability 0 pretty soon, so we don't want a
2074 * worker task hogging it.
2079 for (i = 1; i < n_capabilities; i++) {
2080 cap = &capabilities[i];
2081 ACQUIRE_LOCK(&cap->lock);
2082 startWorkerTask(cap, workerStart);
2083 RELEASE_LOCK(&cap->lock);
2088 RELEASE_LOCK(&sched_mutex);
2093 rtsBool wait_foreign
2094 #if !defined(THREADED_RTS)
2095 __attribute__((unused))
2098 /* see Capability.c, shutdownCapability() */
2102 task = newBoundTask();
2104 // If we haven't killed all the threads yet, do it now.
2105 if (sched_state < SCHED_SHUTTING_DOWN) {
2106 sched_state = SCHED_INTERRUPTING;
2107 waitForReturnCapability(&task->cap,task);
2108 scheduleDoGC(task->cap,task,rtsFalse);
2109 ASSERT(task->tso == NULL);
2110 releaseCapability(task->cap);
2112 sched_state = SCHED_SHUTTING_DOWN;
2114 #if defined(THREADED_RTS)
2118 for (i = 0; i < n_capabilities; i++) {
2119 ASSERT(task->tso == NULL);
2120 shutdownCapability(&capabilities[i], task, wait_foreign);
2125 boundTaskExiting(task);
2129 freeScheduler( void )
2133 ACQUIRE_LOCK(&sched_mutex);
2134 still_running = freeTaskManager();
2135 // We can only free the Capabilities if there are no Tasks still
2136 // running. We might have a Task about to return from a foreign
2137 // call into waitForReturnCapability(), for example (actually,
2138 // this should be the *only* thing that a still-running Task can
2139 // do at this point, and it will block waiting for the
2141 if (still_running == 0) {
2143 if (n_capabilities != 1) {
2144 stgFree(capabilities);
2147 RELEASE_LOCK(&sched_mutex);
2148 #if defined(THREADED_RTS)
2149 closeMutex(&sched_mutex);
2153 /* -----------------------------------------------------------------------------
2156 This is the interface to the garbage collector from Haskell land.
2157 We provide this so that external C code can allocate and garbage
2158 collect when called from Haskell via _ccall_GC.
2159 -------------------------------------------------------------------------- */
2162 performGC_(rtsBool force_major)
2166 // We must grab a new Task here, because the existing Task may be
2167 // associated with a particular Capability, and chained onto the
2168 // suspended_ccalling_tasks queue.
2169 task = newBoundTask();
2171 waitForReturnCapability(&task->cap,task);
2172 scheduleDoGC(task->cap,task,force_major);
2173 releaseCapability(task->cap);
2174 boundTaskExiting(task);
2180 performGC_(rtsFalse);
2184 performMajorGC(void)
2186 performGC_(rtsTrue);
2189 /* -----------------------------------------------------------------------------
2192 If the thread has reached its maximum stack size, then raise the
2193 StackOverflow exception in the offending thread. Otherwise
2194 relocate the TSO into a larger chunk of memory and adjust its stack
2196 -------------------------------------------------------------------------- */
2199 threadStackOverflow(Capability *cap, StgTSO *tso)
2201 nat new_stack_size, stack_words;
2206 IF_DEBUG(sanity,checkTSO(tso));
2208 // don't allow throwTo() to modify the blocked_exceptions queue
2209 // while we are moving the TSO:
2210 lockClosure((StgClosure *)tso);
2212 if (tso->stack_size >= tso->max_stack_size
2213 && !(tso->flags & TSO_BLOCKEX)) {
2214 // NB. never raise a StackOverflow exception if the thread is
2215 // inside Control.Exceptino.block. It is impractical to protect
2216 // against stack overflow exceptions, since virtually anything
2217 // can raise one (even 'catch'), so this is the only sensible
2218 // thing to do here. See bug #767.
2221 if (tso->flags & TSO_SQUEEZED) {
2225 // #3677: In a stack overflow situation, stack squeezing may
2226 // reduce the stack size, but we don't know whether it has been
2227 // reduced enough for the stack check to succeed if we try
2228 // again. Fortunately stack squeezing is idempotent, so all we
2229 // need to do is record whether *any* squeezing happened. If we
2230 // are at the stack's absolute -K limit, and stack squeezing
2231 // happened, then we try running the thread again. The
2232 // TSO_SQUEEZED flag is set by threadPaused() to tell us whether
2233 // squeezing happened or not.
2235 debugTrace(DEBUG_gc,
2236 "threadStackOverflow of TSO %ld (%p): stack too large (now %ld; max is %ld)",
2237 (long)tso->id, tso, (long)tso->stack_size, (long)tso->max_stack_size);
2239 /* If we're debugging, just print out the top of the stack */
2240 printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size,
2243 // Send this thread the StackOverflow exception
2245 throwToSingleThreaded(cap, tso, (StgClosure *)stackOverflow_closure);
2250 // We also want to avoid enlarging the stack if squeezing has
2251 // already released some of it. However, we don't want to get into
2252 // a pathalogical situation where a thread has a nearly full stack
2253 // (near its current limit, but not near the absolute -K limit),
2254 // keeps allocating a little bit, squeezing removes a little bit,
2255 // and then it runs again. So to avoid this, if we squeezed *and*
2256 // there is still less than BLOCK_SIZE_W words free, then we enlarge
2257 // the stack anyway.
2258 if ((tso->flags & TSO_SQUEEZED) &&
2259 ((W_)(tso->sp - tso->stack) >= BLOCK_SIZE_W)) {
2264 /* Try to double the current stack size. If that takes us over the
2265 * maximum stack size for this thread, then use the maximum instead
2266 * (that is, unless we're already at or over the max size and we
2267 * can't raise the StackOverflow exception (see above), in which
2268 * case just double the size). Finally round up so the TSO ends up as
2269 * a whole number of blocks.
2271 if (tso->stack_size >= tso->max_stack_size) {
2272 new_stack_size = tso->stack_size * 2;
2274 new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size);
2276 new_tso_size = (lnat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) +
2277 TSO_STRUCT_SIZE)/sizeof(W_);
2278 new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */
2279 new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
2281 debugTrace(DEBUG_sched,
2282 "increasing stack size from %ld words to %d.",
2283 (long)tso->stack_size, new_stack_size);
2285 dest = (StgTSO *)allocate(cap,new_tso_size);
2286 TICK_ALLOC_TSO(new_stack_size,0);
2288 /* copy the TSO block and the old stack into the new area */
2289 memcpy(dest,tso,TSO_STRUCT_SIZE);
2290 stack_words = tso->stack + tso->stack_size - tso->sp;
2291 new_sp = (P_)dest + new_tso_size - stack_words;
2292 memcpy(new_sp, tso->sp, stack_words * sizeof(W_));
2294 /* relocate the stack pointers... */
2296 dest->stack_size = new_stack_size;
2298 /* Mark the old TSO as relocated. We have to check for relocated
2299 * TSOs in the garbage collector and any primops that deal with TSOs.
2301 * It's important to set the sp value to just beyond the end
2302 * of the stack, so we don't attempt to scavenge any part of the
2305 tso->what_next = ThreadRelocated;
2306 setTSOLink(cap,tso,dest);
2307 tso->sp = (P_)&(tso->stack[tso->stack_size]);
2308 tso->why_blocked = NotBlocked;
2313 IF_DEBUG(sanity,checkTSO(dest));
2315 IF_DEBUG(scheduler,printTSO(dest));
2322 threadStackUnderflow (Capability *cap, Task *task, StgTSO *tso)
2324 bdescr *bd, *new_bd;
2325 lnat free_w, tso_size_w;
2328 tso_size_w = tso_sizeW(tso);
2330 if (tso_size_w < MBLOCK_SIZE_W ||
2331 // TSO is less than 2 mblocks (since the first mblock is
2332 // shorter than MBLOCK_SIZE_W)
2333 (tso_size_w - BLOCKS_PER_MBLOCK*BLOCK_SIZE_W) % MBLOCK_SIZE_W != 0 ||
2334 // or TSO is not a whole number of megablocks (ensuring
2335 // precondition of splitLargeBlock() below)
2336 (tso_size_w <= round_up_to_mblocks(RtsFlags.GcFlags.initialStkSize)) ||
2337 // or TSO is smaller than the minimum stack size (rounded up)
2338 (nat)(tso->stack + tso->stack_size - tso->sp) > tso->stack_size / 4)
2339 // or stack is using more than 1/4 of the available space
2345 // don't allow throwTo() to modify the blocked_exceptions queue
2346 // while we are moving the TSO:
2347 lockClosure((StgClosure *)tso);
2349 // this is the number of words we'll free
2350 free_w = round_to_mblocks(tso_size_w/2);
2352 bd = Bdescr((StgPtr)tso);
2353 new_bd = splitLargeBlock(bd, free_w / BLOCK_SIZE_W);
2354 bd->free = bd->start + TSO_STRUCT_SIZEW;
2356 new_tso = (StgTSO *)new_bd->start;
2357 memcpy(new_tso,tso,TSO_STRUCT_SIZE);
2358 new_tso->stack_size = new_bd->free - new_tso->stack;
2360 // The original TSO was dirty and probably on the mutable
2361 // list. The new TSO is not yet on the mutable list, so we better
2364 new_tso->flags &= ~TSO_LINK_DIRTY;
2365 dirty_TSO(cap, new_tso);
2367 debugTrace(DEBUG_sched, "thread %ld: reducing TSO size from %lu words to %lu",
2368 (long)tso->id, tso_size_w, tso_sizeW(new_tso));
2370 tso->what_next = ThreadRelocated;
2371 tso->_link = new_tso; // no write barrier reqd: same generation
2373 // The TSO attached to this Task may have moved, so update the
2375 if (task->tso == tso) {
2376 task->tso = new_tso;
2382 IF_DEBUG(sanity,checkTSO(new_tso));
2387 /* ---------------------------------------------------------------------------
2389 - usually called inside a signal handler so it mustn't do anything fancy.
2390 ------------------------------------------------------------------------ */
2393 interruptStgRts(void)
2395 sched_state = SCHED_INTERRUPTING;
2396 setContextSwitches();
2397 #if defined(THREADED_RTS)
2402 /* -----------------------------------------------------------------------------
2405 This function causes at least one OS thread to wake up and run the
2406 scheduler loop. It is invoked when the RTS might be deadlocked, or
2407 an external event has arrived that may need servicing (eg. a
2408 keyboard interrupt).
2410 In the single-threaded RTS we don't do anything here; we only have
2411 one thread anyway, and the event that caused us to want to wake up
2412 will have interrupted any blocking system call in progress anyway.
2413 -------------------------------------------------------------------------- */
2415 #if defined(THREADED_RTS)
2416 void wakeUpRts(void)
2418 // This forces the IO Manager thread to wakeup, which will
2419 // in turn ensure that some OS thread wakes up and runs the
2420 // scheduler loop, which will cause a GC and deadlock check.
2425 /* -----------------------------------------------------------------------------
2428 * Check the blackhole_queue for threads that can be woken up. We do
2429 * this periodically: before every GC, and whenever the run queue is
2432 * An elegant solution might be to just wake up all the blocked
2433 * threads with awakenBlockedQueue occasionally: they'll go back to
2434 * sleep again if the object is still a BLACKHOLE. Unfortunately this
2435 * doesn't give us a way to tell whether we've actually managed to
2436 * wake up any threads, so we would be busy-waiting.
2438 * -------------------------------------------------------------------------- */
2441 checkBlackHoles (Capability *cap)
2444 rtsBool any_woke_up = rtsFalse;
2447 // blackhole_queue is global:
2448 ASSERT_LOCK_HELD(&sched_mutex);
2450 debugTrace(DEBUG_sched, "checking threads blocked on black holes");
2452 // ASSUMES: sched_mutex
2453 prev = &blackhole_queue;
2454 t = blackhole_queue;
2455 while (t != END_TSO_QUEUE) {
2456 if (t->what_next == ThreadRelocated) {
2460 ASSERT(t->why_blocked == BlockedOnBlackHole);
2461 type = get_itbl(UNTAG_CLOSURE(t->block_info.closure))->type;
2462 if (type != BLACKHOLE && type != CAF_BLACKHOLE) {
2463 IF_DEBUG(sanity,checkTSO(t));
2464 t = unblockOne(cap, t);
2466 any_woke_up = rtsTrue;
2476 /* -----------------------------------------------------------------------------
2479 This is used for interruption (^C) and forking, and corresponds to
2480 raising an exception but without letting the thread catch the
2482 -------------------------------------------------------------------------- */
2485 deleteThread (Capability *cap, StgTSO *tso)
2487 // NOTE: must only be called on a TSO that we have exclusive
2488 // access to, because we will call throwToSingleThreaded() below.
2489 // The TSO must be on the run queue of the Capability we own, or
2490 // we must own all Capabilities.
2492 if (tso->why_blocked != BlockedOnCCall &&
2493 tso->why_blocked != BlockedOnCCall_NoUnblockExc) {
2494 throwToSingleThreaded(cap,tso,NULL);
2498 #ifdef FORKPROCESS_PRIMOP_SUPPORTED
2500 deleteThread_(Capability *cap, StgTSO *tso)
2501 { // for forkProcess only:
2502 // like deleteThread(), but we delete threads in foreign calls, too.
2504 if (tso->why_blocked == BlockedOnCCall ||
2505 tso->why_blocked == BlockedOnCCall_NoUnblockExc) {
2506 unblockOne(cap,tso);
2507 tso->what_next = ThreadKilled;
2509 deleteThread(cap,tso);
2514 /* -----------------------------------------------------------------------------
2515 raiseExceptionHelper
2517 This function is called by the raise# primitve, just so that we can
2518 move some of the tricky bits of raising an exception from C-- into
2519 C. Who knows, it might be a useful re-useable thing here too.
2520 -------------------------------------------------------------------------- */
2523 raiseExceptionHelper (StgRegTable *reg, StgTSO *tso, StgClosure *exception)
2525 Capability *cap = regTableToCapability(reg);
2526 StgThunk *raise_closure = NULL;
2528 StgRetInfoTable *info;
2530 // This closure represents the expression 'raise# E' where E
2531 // is the exception raise. It is used to overwrite all the
2532 // thunks which are currently under evaluataion.
2535 // OLD COMMENT (we don't have MIN_UPD_SIZE now):
2536 // LDV profiling: stg_raise_info has THUNK as its closure
2537 // type. Since a THUNK takes at least MIN_UPD_SIZE words in its
2538 // payload, MIN_UPD_SIZE is more approprate than 1. It seems that
2539 // 1 does not cause any problem unless profiling is performed.
2540 // However, when LDV profiling goes on, we need to linearly scan
2541 // small object pool, where raise_closure is stored, so we should
2542 // use MIN_UPD_SIZE.
2544 // raise_closure = (StgClosure *)RET_STGCALL1(P_,allocate,
2545 // sizeofW(StgClosure)+1);
2549 // Walk up the stack, looking for the catch frame. On the way,
2550 // we update any closures pointed to from update frames with the
2551 // raise closure that we just built.
2555 info = get_ret_itbl((StgClosure *)p);
2556 next = p + stack_frame_sizeW((StgClosure *)p);
2557 switch (info->i.type) {
2560 // Only create raise_closure if we need to.
2561 if (raise_closure == NULL) {
2563 (StgThunk *)allocate(cap,sizeofW(StgThunk)+1);
2564 SET_HDR(raise_closure, &stg_raise_info, CCCS);
2565 raise_closure->payload[0] = exception;
2567 UPD_IND(cap, ((StgUpdateFrame *)p)->updatee,
2568 (StgClosure *)raise_closure);
2572 case ATOMICALLY_FRAME:
2573 debugTrace(DEBUG_stm, "found ATOMICALLY_FRAME at %p", p);
2575 return ATOMICALLY_FRAME;
2581 case CATCH_STM_FRAME:
2582 debugTrace(DEBUG_stm, "found CATCH_STM_FRAME at %p", p);
2584 return CATCH_STM_FRAME;
2590 case CATCH_RETRY_FRAME:
2599 /* -----------------------------------------------------------------------------
2600 findRetryFrameHelper
2602 This function is called by the retry# primitive. It traverses the stack
2603 leaving tso->sp referring to the frame which should handle the retry.
2605 This should either be a CATCH_RETRY_FRAME (if the retry# is within an orElse#)
2606 or should be a ATOMICALLY_FRAME (if the retry# reaches the top level).
2608 We skip CATCH_STM_FRAMEs (aborting and rolling back the nested tx that they
2609 create) because retries are not considered to be exceptions, despite the
2610 similar implementation.
2612 We should not expect to see CATCH_FRAME or STOP_FRAME because those should
2613 not be created within memory transactions.
2614 -------------------------------------------------------------------------- */
2617 findRetryFrameHelper (StgTSO *tso)
2620 StgRetInfoTable *info;
2624 info = get_ret_itbl((StgClosure *)p);
2625 next = p + stack_frame_sizeW((StgClosure *)p);
2626 switch (info->i.type) {
2628 case ATOMICALLY_FRAME:
2629 debugTrace(DEBUG_stm,
2630 "found ATOMICALLY_FRAME at %p during retry", p);
2632 return ATOMICALLY_FRAME;
2634 case CATCH_RETRY_FRAME:
2635 debugTrace(DEBUG_stm,
2636 "found CATCH_RETRY_FRAME at %p during retrry", p);
2638 return CATCH_RETRY_FRAME;
2640 case CATCH_STM_FRAME: {
2641 StgTRecHeader *trec = tso -> trec;
2642 StgTRecHeader *outer = trec -> enclosing_trec;
2643 debugTrace(DEBUG_stm,
2644 "found CATCH_STM_FRAME at %p during retry", p);
2645 debugTrace(DEBUG_stm, "trec=%p outer=%p", trec, outer);
2646 stmAbortTransaction(tso -> cap, trec);
2647 stmFreeAbortedTRec(tso -> cap, trec);
2648 tso -> trec = outer;
2655 ASSERT(info->i.type != CATCH_FRAME);
2656 ASSERT(info->i.type != STOP_FRAME);
2663 /* -----------------------------------------------------------------------------
2664 resurrectThreads is called after garbage collection on the list of
2665 threads found to be garbage. Each of these threads will be woken
2666 up and sent a signal: BlockedOnDeadMVar if the thread was blocked
2667 on an MVar, or NonTermination if the thread was blocked on a Black
2670 Locks: assumes we hold *all* the capabilities.
2671 -------------------------------------------------------------------------- */
2674 resurrectThreads (StgTSO *threads)
2680 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
2681 next = tso->global_link;
2683 gen = Bdescr((P_)tso)->gen;
2684 tso->global_link = gen->threads;
2687 debugTrace(DEBUG_sched, "resurrecting thread %lu", (unsigned long)tso->id);
2689 // Wake up the thread on the Capability it was last on
2692 switch (tso->why_blocked) {
2694 /* Called by GC - sched_mutex lock is currently held. */
2695 throwToSingleThreaded(cap, tso,
2696 (StgClosure *)blockedIndefinitelyOnMVar_closure);
2698 case BlockedOnBlackHole:
2699 throwToSingleThreaded(cap, tso,
2700 (StgClosure *)nonTermination_closure);
2703 throwToSingleThreaded(cap, tso,
2704 (StgClosure *)blockedIndefinitelyOnSTM_closure);
2707 /* This might happen if the thread was blocked on a black hole
2708 * belonging to a thread that we've just woken up (raiseAsync
2709 * can wake up threads, remember...).
2712 case BlockedOnException:
2713 // throwTo should never block indefinitely: if the target
2714 // thread dies or completes, throwTo returns.
2715 barf("resurrectThreads: thread BlockedOnException");
2718 barf("resurrectThreads: thread blocked in a strange way");
2723 /* -----------------------------------------------------------------------------
2724 performPendingThrowTos is called after garbage collection, and
2725 passed a list of threads that were found to have pending throwTos
2726 (tso->blocked_exceptions was not empty), and were blocked.
2727 Normally this doesn't happen, because we would deliver the
2728 exception directly if the target thread is blocked, but there are
2729 small windows where it might occur on a multiprocessor (see
2732 NB. we must be holding all the capabilities at this point, just
2733 like resurrectThreads().
2734 -------------------------------------------------------------------------- */
2737 performPendingThrowTos (StgTSO *threads)
2741 Task *task, *saved_task;;
2747 for (tso = threads; tso != END_TSO_QUEUE; tso = next) {
2748 next = tso->global_link;
2750 gen = Bdescr((P_)tso)->gen;
2751 tso->global_link = gen->threads;
2754 debugTrace(DEBUG_sched, "performing blocked throwTo to thread %lu", (unsigned long)tso->id);
2756 // We must pretend this Capability belongs to the current Task
2757 // for the time being, as invariants will be broken otherwise.
2758 // In fact the current Task has exclusive access to the systme
2759 // at this point, so this is just bookkeeping:
2760 task->cap = tso->cap;
2761 saved_task = tso->cap->running_task;
2762 tso->cap->running_task = task;
2763 maybePerformBlockedException(tso->cap, tso);
2764 tso->cap->running_task = saved_task;
2767 // Restore our original Capability: