I had a quick browse through the Harmony SVN and spotted what appears to
be a vulnerability in the java.lang.ThreadLocal implementation. I have
briefly discussed this with Tim Ellison and Geir Magnusson Jr., off list
before posting here.
Harmony uses a per Thread HashMap (WeakHashMap in classlibadapter) to
map ThreadLocals onto values. HashMaps (should) check for equality with
Object.equals and Object.hashCode instead of == and System.identityHashCode.
Malicious subclasses of ThreadLocal can override hashCode to run through
all possible hash codes, extracting all the ThreadLocals present in the
current thread through an overridden equals. Some of these ThreadLocals
may contain sensitive values. Even if Harmony generates identity hash
codes entirely at random, the process should be completable in the order
of a few minutes of CPU time.
Tim Ellison suggests replacing the HashMap with an IdentityHashMap. I
agree that this would fix the security vulnerability. Some modern code,
such as I believe Spring, creates many ThreadLocal instances, so you may
wish to look further at quality of implementation issues.
FWIW, I believe early versions of Sun's 1.3 J2SE suffered a similar problem.
Tom Hawtin
