fix a race where the timer signal could remain turned off, leading to deadlock

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

diff --git a/rts/Schedule.c b/rts/Schedule.c
index 33715b1..9baf755 100644 (file)
--- a/rts/Schedule.c
+++ b/rts/Schedule.c
@@ -777,9 +777,11 @@ schedulePushWork(Capability *cap USED_IF_THREADS,
 
     // Check whether we have more threads on our run queue, or sparks
     // in our pool, that we could hand to another Capability.
-    if ((emptyRunQueue(cap) || cap->run_queue_hd->_link == END_TSO_QUEUE)
-       && sparkPoolSizeCap(cap) < 2) {
-       return;
+    if (cap->run_queue_hd == END_TSO_QUEUE) {
+        if (sparkPoolSizeCap(cap) < 2) return;
+    } else {
+        if (cap->run_queue_hd->_link == END_TSO_QUEUE &&
+            sparkPoolSizeCap(cap) < 1) return;
     }
 
     // First grab as many free Capabilities as we can.
@@ -1400,6 +1402,13 @@ scheduleHandleThreadFinished (Capability *cap STG_UNUSED, Task *task, StgTSO *t)
     debugTrace(DEBUG_sched, "--++ thread %lu (%s) finished", 
               (unsigned long)t->id, whatNext_strs[t->what_next]);
 
+    // blocked exceptions can now complete, even if the thread was in
+    // blocked mode (see #2910).  The thread is already marked
+    // ThreadComplete, so any further throwTos will complete
+    // immediately and we don't need to worry about synchronising with
+    // those.
+    awakenBlockedExceptionQueue (cap, t);
+
       //
       // Check whether the thread that just completed was a bound
       // thread, and if so return with the result.  
@@ -1590,6 +1599,21 @@ delete_threads_and_gc:
     
     heap_census = scheduleNeedHeapProfile(rtsTrue);
 
+    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();
+    }
+
 #if defined(THREADED_RTS)
     debugTrace(DEBUG_sched, "doing GC");
     // reset waiting_for_gc *before* GC, so that when the GC threads
@@ -1631,16 +1655,6 @@ delete_threads_and_gc:
     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)
     if (gc_type == PENDING_GC_SEQ) {
         // release our stash of capabilities.