MERGE: Fix bug exposed by conc052.

A thread that was blocked on a blackhole but can now be woken up could
possibly be treated as unreachable by the GC, and sent the
NonTermination exception.

This can give rise to spurious <<loop>>s in concurrent programs, so
it's a good one to fix.

diff --git a/rts/sm/MarkWeak.c b/rts/sm/MarkWeak.c
index 49134da..455b586 100644 (file)
--- a/rts/sm/MarkWeak.c
+++ b/rts/sm/MarkWeak.c
@@ -284,17 +284,26 @@ traverseBlackholeQueue (void)
 {
     StgTSO *prev, *t, *tmp;
     rtsBool flag;
+    nat type;
 
     flag = rtsFalse;
     prev = NULL;
 
     for (t = blackhole_queue; t != END_TSO_QUEUE; prev=t, t = t->link) {
+        // if the thread is not yet alive...
        if (! (tmp = (StgTSO *)isAlive((StgClosure*)t))) {
-           if (isAlive(t->block_info.closure)) {
-               t = (StgTSO *)evacuate((StgClosure *)t);
-               if (prev) prev->link = t;
-               flag = rtsTrue;
-           }
+            // if the closure it is blocked on is either (a) a
+            // reachable BLAKCHOLE or (b) not a BLACKHOLE, then we
+            // make the thread alive.
+           if (!isAlive(t->block_info.closure)) {
+                type = get_itbl(t->block_info.closure)->type;
+                if (type == BLACKHOLE || type == CAF_BLACKHOLE) {
+                    continue;
+                }
+            }
+            t = (StgTSO *)evacuate((StgClosure *)t);
+            if (prev) prev->link = t;
+            flag = rtsTrue;
        }
     }
     return flag;