Fix a race in the deadlock-detection code

After a deadlock it was possible for the timer signal to remain off,
which meant that the next deadlock would not be detected, and the
system would hang.  Spotted by conc047(threaded2).

diff --git a/rts/Schedule.c b/rts/Schedule.c
index ed4d4b0..7dd0634 100644 (file)
--- a/rts/Schedule.c
+++ b/rts/Schedule.c
@@ -1001,12 +1001,11 @@ scheduleDetectDeadlock (Capability *cap, Task *task)
        // they are unreachable and will therefore be sent an
        // exception.  Any threads thus released will be immediately
        // runnable.
-       cap = scheduleDoGC (cap, task, rtsTrue/*force  major GC*/);
+       cap = scheduleDoGC (cap, task, rtsTrue/*force major GC*/);
+        // when force_major == rtsTrue. scheduleDoGC sets
+        // recent_activity to ACTIVITY_DONE_GC and turns off the timer
+        // signal.
 
-       recent_activity = ACTIVITY_DONE_GC;
-        // disable timer signals (see #1623)
-        stopTimer();
-       
        if ( !emptyRunQueue(cap) ) return;
 
 #if defined(RTS_USER_SIGNALS) && !defined(THREADED_RTS)
@@ -1577,6 +1576,16 @@ scheduleDoGC (Capability *cap, Task *task USED_IF_THREADS, rtsBool force_major)
     balanceSparkPoolsCaps(n_capabilities, capabilities);
 #endif
 
+    if (force_major)
+    {
+        // We've just done a major GC and we don't need the timer
+        // signal turned on any more (#1623).
+        // NB. do this *before* releasing the Capabilities, to avoid
+        // deadlocks!
+        recent_activity = ACTIVITY_DONE_GC;
+        stopTimer();
+    }
+
 #if defined(THREADED_RTS)
     // release our stash of capabilities.
     for (i = 0; i < n_capabilities; i++) {