[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <20190325175710.GA25444@openwall.com>
Date: Mon, 25 Mar 2019 18:57:11 +0100
From: Solar Designer <solar@...nwall.com>
To: john-users@...ts.openwall.com
Subject: Re: can other user access secrets in GPU's memory?
On Mon, Mar 25, 2019 at 08:10:58AM +0100, magnum wrote:
> GPUs doesn't have any protection
> against users reading your memory even *while* you're using it. In my
> view is this is primarily a driver/runtime issue and ultimately even a
> hardware issue (no memory protection in the design).
Modern GPUs have MMUs, so the hardware is probably capable. I agree
it's "primarily a driver/runtime issue". Also, I'm under impression
NVIDIA cares about this issue already, but AMD doesn't care yet. We can
try reporting this as a bug against the AMDGPU-PRO driver and see what
happens. Aleksey, care to try? Also:
On Mon, Mar 25, 2019 at 01:42:07AM +0300, Aleksey Cherepanov wrote:
> I was playing around with OpenCL and it looks like memory is not zeroed
> by default on GPUs, i.e. I can read old values from gpu on another run
> of host program. So I guess it means that other users can read these
> values too, right?
Don't assume - actually test it as two different uses. While your guess
is probably correct this time, it's conceivable for a driver to be smart
and only zeroize memory before giving it to a different user, not to
another OpenCL kernel invocation by the same user. So you need to test
and make sure before reporting this as a bug.
Thanks,
Alexander
Powered by blists - more mailing lists
Confused about mailing lists and their use?
Read about mailing lists on Wikipedia
and check out these
guidelines on proper formatting of your messages.
