lkml.org 
[lkml]   [2009]   [May]   [23]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [patch 0/5] Support for sanitization flag in low-level page allocator

* Larry H. <research@subreption.com> wrote:

> NOTE: Let's keep the PaX Team on CC from now on, they might have further
> input to this discussion. (pageexec at freemail dot hu)
>
> On 09:34 Fri 22 May , Ingo Molnar wrote:
> > The whole kernel contains data that 'should not be leaked'.
> > _If_ any of this is done, i'd _very_ strongly suggest to describe it
> > by what it does, not by what its subjective security attribute is.
> >
> > 'PG_eyes_only' or 'PG_eagle_azf_compartmented' is silly naming. It
> > is silly because it hardcodes one particular expectation/model of
> > 'security'.
> >
> > GFP_NON_PERSISTENT & PG_non_persistent is a _lot_ better, because it
> > is a technical description of how information spreads. (which is the
> > underlying principle of every security model)
> >
> > That name alone tells us everyting what this does: it does not
> > allow this data to reach or touch persistent storage. It wont be
> > swapped and it wont by saved by hibernation. It will also be
> > cleared when freed, to achieve its goal of never touching
> > persistent storage.
>
> The problem is that these patches have a more broad purpose and I
> never mentioned persistent storage as one of them (initially).
> Check earlier messages to see what has been discussed so far.

You need to address my specific concerns instead of referring back
to an earlier discussion. The patches touch code i maintain and i
find them (and your latest resend) unacceptable.

> Regarding the naming changes, those have been done as of Rik's
> comments and I would rather focus on the technical and
> implementation side now.

Naming _is_ a technical issue. Especially here.

> > In-kernel crypto key storage using GFP_NON_PERSISTENT makes some
> > sense - as long as the kernel stack itself is mared
> > GFP_NON_PERSISTENT as well ... which is quite hairy from a
> > performance point of view: we _dont_ want to clear the full
> > stack page for every kernel thread exiting.
>
> Burning the stack there is beyond overkill.

What you are missing is that your patch makes _no technical sense_
if you allow the same information to leak over the kernel stack.
Kernel stacks can be freed and reused, swapped out and thus
'exposed'.

Ingo


\
 
 \ /
  Last update: 2009-05-23 14:53    [from the cache]
©2003-2011 Jasper Spaans