Re: [RFC PATCH] x86/entry/64: randomize kernel stack offset upon system call

[Andy Lutomirski <luto@kernel.org>]

> On Feb 20, 2019, at 2:20 PM, Kees Cook <keescook@chromium.org> wrote:
>
> On Wed, Feb 13, 2019 at 11:52 PM Reshetova, Elena
> <elena.reshetova@intel.com> wrote:
>> Now back to our proposed countermeasures given that attacker has found a way to do
>> a crafted overflow and overwrite:
>>
>>  1) pt_regs is not predictable, but can be discovered in ptrace-style scenario or cache-probing.
>>     If discovered, then attack succeeds as of now.
>>  2) relative stack offset is not predictable and randomized, cannot be probed very easily via
>>      cache or ptrace. So, this is an additional hurdle on the attacker's way since stack is non-
>>      deterministic now.
>>  3) nothing changed for this type of attack, given that attacker's goal is not to overwrite CS
>>      in adjusted pt_regs. If it is his goal, then it helps with that.
>>
>>
>> Now summary:
>>
>> It would seem to me that:
>>
>> - regs->cs |= 3 on exit is a thing worth doing anyway, just because it is cheap, as Andy said, and it
>> might make a positive difference in two out of three attack scenarios. Objections?
>
> I would agree, let's just do this.

Thinking slightly more about this, it’s an incomplete protection.  It
keeps an attacker from returning to kernel mode, but it does not
protect the privileged flag bits.  I think that IOPL is the only thing
we really care about, and doing anything useful about IOPL would be
rather more complex, unfortunately.  I suppose we could just zero it
and guard that with a static branch that is switched off the first
time anyone uses iopl(3).

I suppose we could also add a config option to straight-up disable
IOPL.  I sincerely hope that no one uses it any more. Even the small
number of semi-legit users really ought to be using ioperm() instead.

>
>> - randomization of stack top is only worth doing in ptrace-blocked scenario.
>> Do we have such scenarios left that people care about?
>> Because if we do, then we know that there is a real attack vector that we close this way, otherwise not.
>> This is actually interesting, because we need to remember to take ptrace into our overall
>> kernel hardening threat model (smth that at least I haven't quite realized before) and evaluate every new
>> feature (especially randomization ones) being robust against ptrace probing.
>>
>> - randomization after pt_regs only would make a difference in attack scenario "c", for which
>>  we don't yet have a proof of concept exploit or technique that would work (does not guarantee that
>> attackers don't have the exploits ready through :( ).
>> So, if we implement this, the "justification part" for the feature would be smth like "to make it
>> harder for future possible stack-based exploits that utilize overflows", if/when someone find a new
>> 'ala VLA' way of doing the controlled overflow.
>> How do people feel about it? Is it worth having? I can work on the POC for this in direction that Andy
>> outlined and can provide performance impact/etc., but it is good that we understand that we cannot
>> provide a better justification for this feature at the moment unless someone is ready to share some
>> new exploit technique with us.
>
> I think this make sense. I do think, however, the work should be done
> at syscall entry, though. Thoughts?
>
>

Seems reasonable to me.