Use work-stealing for load-balancing in the GC

[ghc-hetmet.git] / rts / Schedule.c

diff --git a/rts/Schedule.c b/rts/Schedule.c
index 6c46c09..040d16f 100644 (file)
--- a/rts/Schedule.c
+++ b/rts/Schedule.c
@@ -1268,7 +1268,7 @@ scheduleHandleHeapOverflow( Capability *cap, StgTSO *t )
               "--<< thread %ld (%s) stopped: HeapOverflow",
               (long)t->id, whatNext_strs[t->what_next]);
 
-    if (cap->context_switch) {
+    if (cap->r.rHpLim == NULL || cap->context_switch) {
         // Sometimes we miss a context switch, e.g. when calling
         // primitives in a tight loop, MAYBE_GC() doesn't check the
         // context switch flag, and we end up waiting for a GC.
@@ -1606,29 +1606,38 @@ delete_threads_and_gc:
     
     heap_census = scheduleNeedHeapProfile(rtsTrue);
 
+#if defined(THREADED_RTS)
+    debugTrace(DEBUG_sched, "doing GC");
+    // reset waiting_for_gc *before* GC, so that when the GC threads
+    // emerge they don't immediately re-enter the GC.
+    waiting_for_gc = 0;
+    GarbageCollect(force_major || heap_census, gc_type, cap);
+#else
+    GarbageCollect(force_major || heap_census, 0, cap);
+#endif
+
     if (recent_activity == ACTIVITY_INACTIVE && force_major)
     {
         // We are doing a GC because the system has been idle for a
         // timeslice and we need to check for deadlock.  Record the
         // fact that we've done a GC and turn off the timer signal;
         // it will get re-enabled if we run any threads after the GC.
-        //
-        // Note: this is done before GC, because after GC there might
-        // be threads already running (GarbageCollect() releases the
-        // GC threads when it completes), so we risk turning off the
-        // timer signal when it should really be on.
         recent_activity = ACTIVITY_DONE_GC;
         stopTimer();
     }
+    else
+    {
+        // the GC might have taken long enough for the timer to set
+        // recent_activity = ACTIVITY_INACTIVE, but we aren't
+        // necessarily deadlocked:
+        recent_activity = ACTIVITY_YES;
+    }
 
 #if defined(THREADED_RTS)
-    debugTrace(DEBUG_sched, "doing GC");
-    // reset waiting_for_gc *before* GC, so that when the GC threads
-    // emerge they don't immediately re-enter the GC.
-    waiting_for_gc = 0;
-    GarbageCollect(force_major || heap_census, gc_type, cap);
-#else
-    GarbageCollect(force_major || heap_census, 0, cap);
+    if (gc_type == PENDING_GC_PAR)
+    {
+        releaseGCThreads(cap);
+    }
 #endif
 
     if (heap_census) {
@@ -2203,22 +2212,16 @@ exitScheduler(
 {
     Task *task = NULL;
 
-#if defined(THREADED_RTS)
     ACQUIRE_LOCK(&sched_mutex);
     task = newBoundTask();
     RELEASE_LOCK(&sched_mutex);
-#endif
 
     // If we haven't killed all the threads yet, do it now.
     if (sched_state < SCHED_SHUTTING_DOWN) {
        sched_state = SCHED_INTERRUPTING;
-#if defined(THREADED_RTS)
         waitForReturnCapability(&task->cap,task);
        scheduleDoGC(task->cap,task,rtsFalse);    
         releaseCapability(task->cap);
-#else
-       scheduleDoGC(&MainCapability,task,rtsFalse);    
-#endif
     }
     sched_state = SCHED_SHUTTING_DOWN;