Theoretical data race on java.util.logging.Handler.sealed
peter.levart at gmail.com
Tue Dec 3 09:44:40 UTC 2013
On 12/03/2013 09:51 AM, Peter Levart wrote:
> While browsing the code of java.util.logging.Handler, I noticed a
> theoretical possibility that a security check in a j.u.l.StreamHandler
> be circumvented using a data race.
> There is a plain boolean instance field 'sealed' in j.u.l.Handler that
> is pre-initialized to 'true' in field initializer. StreamHandler
> sublcass' constructors overwrite this value with 'false' at the
> beginning, then issue some operations which circumvent security
> checks, and finally they reset the 'sealed' value back to 'true' at
> the end.
> If a reference to an instance of StreamHandler or subclass is passed
> to some thread without synchronization via data-race, this thread can
> see 'true' or 'false' as the possible values of 'sealed' variable,
> thus it is possible to circumvent security checks.
> One possibility to fix this is moving the field to StreamHandler and
> making it final:
> Just making the field volatile might not work. There is an ongoing
> debate on concurrency-interest which suggests that volatile fields are
> not exceptional in constructors like final fields are...
> Regards, Peter
The proposed patch is not complete. There are several subclasses of
StreamHandler in the java.util.logging package that also need a way to
bypass security checks for some operations in their constructors. So
here's the updated webrev which updates them with the same code as
StreamHandler. This means that there are two copies of 'sealed' flag in
object of type ConsoleHandler, for example, but only the one declared in
ConsoleHandler is relevant for governing access checks:
Before filing the bug, I'm asking the list whether this can be
considered a bug...
More information about the core-libs-dev