Fix race in finalizeAllEnqueued.

The problem was as follows:
The GC finishes and enqueues all cleared references on the
ReferenceQueue.unenqueued list. Then someone calls
finalizeAllEnqueued which enqueues a finalizer reference directly on
the finalizer reference queue. However the recently cleared references
had not yet been processed by the ReferenceQueueDaemon resulting
in the sentinal being finalized before the objects which were freed
in the GC.

The fix is enqueuing the sentinel finalizer reference on the unenqueued
queue.

(cherry picked from commit 2cc3d41ff2dfd673fa1ab4ef2a46525933d523ec)

Change-Id: I0598199ef56aa8bd9e1fa40c39d50b2785dc5c10

1 files changed, 7 insertions, 1 deletions

diff --git a/luni/src/main/java/java/lang/ref/FinalizerReference.java b/luni/src/main/java/java/lang/ref/FinalizerReference.java
index 14eaae4..47dc0b4 100644
--- a/luni/src/main/java/java/lang/ref/FinalizerReference.java
+++ b/luni/src/main/java/java/lang/ref/FinalizerReference.java
@@ -98,7 +98,13 @@ public final class FinalizerReference<T> extends Reference<T> {
                     FinalizerReference<Sentinel> sentinelReference = (FinalizerReference<Sentinel>) r;
                     sentinelReference.referent = null;
                     sentinelReference.zombie = sentinel;
-                    sentinelReference.enqueueInternal();
+                    // Make a single element list, then enqueue the reference on the daemon unenqueued
+                    // list. This is required instead of enqueuing directly on the finalizer queue
+                    // since there could be recently freed objects in the unqueued list which are not
+                    // yet on the finalizer queue. This could cause the sentinel to run before the
+                    // objects are finalized. b/17381967
+                    sentinelReference.pendingNext = sentinelReference;
+                    ReferenceQueue.add(sentinelReference);
                     return;
                 }
             }