On Mon, May 25, 2009 at 3:37 PM, Brendan Eich <brendan@...
> On May 25, 2009, at 2:56 PM, Tyler Close wrote:
>> On Sun, May 24, 2009 at 7:49 AM, Douglas Crockford
>> > I am considering the blocking of try/catch in ADsafe. I am
>> concerned about the
>> > potential of using exceptions to deliver capabilities between
>> isolated widgets.
>> stack overflow and out of memory errors.
> Out of memory is not catchable in SpiderMonkey.
What about stack overflow?
> What browsers did you test?
I did the testing during the caja security review and I believe I got
an exploit working in both IE 6 and Firefox 2 on Windows using the
stack overflow Error. The stack overflow Error was easier to work with
than the out of memory Error, since it's more predictable.
>> A widget could use this
>> ability to put another object, or perhaps even the browser, in an
>> inconsistent state. For example, the widget could use up all but one
>> stack frame and then make a call to a browser object which mutates
>> part of its state and then attempts a function call before making
>> additional mutations. The victim object would make the first mutation,
>> but suffer a stack overflow error before being able to complete the
>> rest of the mutations. The widget code could catch the Error, leaving
>> the victim object in the inconsistent state.
> Sounds like a bug in the victim object. Why didn't it catch and clean
Because it wasn't expecting the function to throw. For example, the
called function may have been one implemented in the same block of
code as the calling function. Seeing that the called function didn't
throw, the programmer of the victim object didn't write a try catch
block. In general, protecting yourself against this kind of attack by
defensively writing try catch blocks is too cumbersome and error prone
to be feasible.
> Really, there are lots of potential bugs where an inconsistent state
> could result from errors. Making the errors fatal to the currently
> exeucting script only increases consistency in that particular script
> or event handler's control flow. The next script or event can still
> the inconsistency.
Sounds like an excellent argument for making errors fail-stop. ADsafe
is in a good position to implement this, since it can wrap all event
handlers created by a widget.
> If you poison the whole well, meaning both fail-stop the script and
> make the entire reachable object graph inaccessible or error-tainted,
> then you can limit the leak. But there's still a termination channel.
Allowing the widget to explicitly terminate the page is fine, since it
implicitly has this authority anyways, since it could always just
enter an infinite loop, or use other DOS techniques.
"Waterken News: Capability security on the Web"