Re: [PATCH v2 03/13] stackleak: remove redundant check

[Alexander Popov]

On 12.05.2022 12:14, Mark Rutland wrote:
> On Wed, May 11, 2022 at 07:44:41AM -0700, Kees Cook wrote:
> > On May 11, 2022 1:02:45 AM PDT, Mark Rutland <mark.rutland@xxxxxxx> wrote:
> > > On Tue, May 10, 2022 at 08:00:38PM -0700, Kees Cook wrote:
> > > > On Tue, May 10, 2022 at 12:46:48PM +0100, Mark Rutland wrote:
> > > > > On Sun, May 08, 2022 at 09:17:01PM +0300, Alexander Popov wrote:
> > > > > > On 27.04.2022 20:31, Mark Rutland wrote:
> > > > > > > In __stackleak_erase() we check that the `erase_low` value derived from
> > > > > > > `current->lowest_stack` is above the lowest legitimate stack pointer
> > > > > > > value, but this is already enforced by stackleak_track_stack() when
> > > > > > > recording the lowest stack value.
> > > > > > >
> > > > > > > Remove the redundant check.
> > > > > > >
> > > > > > > There should be no functional change as a result of this patch.
> > > > > >
> > > > > > Mark, I can't agree here. I think this check is important.
> > > > > > The performance profit from dropping it is less than the confidence decrease :)
> > > > > >
> > > > > > With this check, if the 'lowest_stack' value is corrupted, stackleak doesn't
> > > > > > overwrite some wrong kernel memory, but simply clears the whole thread
> > > > > > stack, which is safe behavior.
> > > > >
> > > > > If you feel strongly about it, I can restore the check, but I struggle to
> > > > > believe that it's worthwhile. The `lowest_stack` value lives in the
> > > > > task_struct, and if you have the power to corrupt that you have the power to do
> > > > > much more interesting things.
> > > > >
> > > > > If we do restore it, I'd like to add a big fat comment explaining the
> > > > > rationale (i.e. that it only matter if someone could corrupt
> > > > > `current->lowest_stack`, as otherwise that's guarnateed to be within bounds).
> > > >
> > > > Yeah, let's restore it and add the comment. While I do agree it's likely
> > > > that such an corruption would likely mean an attacker had significant
> > > > control over kernel memory already, it is not uncommon that an attack
> > > > only has a limited index from a given address, etc. Or some manipulation
> > > > is possible via weird gadgets, etc. It's unlikely, but not impossible,
> > > > and a bounds-check for that value is cheap compared to the rest of the
> > > > work happening. :)
> > >
> > > Fair enough; I can go spin a patch restoring this. I'm somewhat unhappy with
> > > silently fixing that up, though -- IMO it'd be better to BUG() or similar in
> > > that case.
> >
> > I share your desires, and this was exactly what Alexander originally proposed, but Linus rejected it violently. :(
> > https://lore.kernel.org/lkml/CA+55aFy6jNLsywVYdGp83AMrXBo_P-pkjkphPGrO=82SPKCpLQ@xxxxxxxxxxxxxx/
>
> I see. :/
>
> Thinking about this some more, if we assume someone can corrupt *some* word of
> memory, then we need to consider that instead of corrupting
> task_struct::lowest_stack, they could corrupt task_struct::stack (or x86's
> cpu_current_top_of_stack prior to this series).
>
> With that in mind, if we detect that task_struct::lowest_stack is
> out-of-bounds, we have no idea whether it has been corrupted or the other bound
> values have been corrupted, and so we can't do the erase safely anyway.

:)

IMO, even if a kernel thread stack is moved somewhere for any weird reason, 
stackleak must erase it at the end of syscall and do its job.

> So AFAICT we must *avoid* erasing when that goes wrong. Maybe we could WARN()
> instead of BUG()?

Mark, I think security features must not go out of service.

The 'lowest_stack' value is for making stackleak faster. I believe if the 
'lowest_stack' value is invalid, stackleak must not skip its main job and should 
erase the whole kernel thread stack.

When I developed 'stackleak_erase()' I tried adding 'WARN_ON()', but it didn't 
work properly there, as I remember. Warning handling code is very complex. So I 
dropped that fragile part.

Best regards,
Alexander