Re: [PATCH 1/9] exec: add a global execve counter

[Solar Designer <solar@openwall.com>]

On Sat, Mar 10, 2012 at 04:12:20PM -0800, Linus Torvalds wrote:
> On Sat, Mar 10, 2012 at 3:25 PM, Djalal Harouni <tixxdz@opendz.org> wrote:
> >
> > Given that consideration this patch introduces two counters:
> > A global atomic execve counter that will be incremented on every
> > do_execve_common() call, and an atomic exec_id member for the task_struct.
> 
> This seems horribly expensive on most 32-bit architectures, including
> very much x86-32. That atomic64_inc_return() is not cheap.  It's
> possible that it's basically an impossible operation to do atomically
> on certain platforms, causing it to use some random spinlock instead.

I think these are _relatively_ cheap considering that it's execve().
If we can do e.g. 10 million of atomic64_inc_return()'s per second (and
I think we can do a lot more, although I did not benchmark this), but
only 10000 of execve()'s per second, then the performance impact is 0.1%.
I admit that 0.1% is significant, but I used worst-case guesstimates
here; the actual number might be more like 0.01% (50 million vs. 5 thousand)
or even 0.002% (200 million vs. 4 thousand).  (200 million is 15 CPU
cycles at 3 GHz, which may be a reasonable optimistic estimate.
4 thousand is a realistic execve() speed for dynamically-linked programs.)

> I wonder if we couldn't make something much cheaper, since we don't
> actually care about it being globally incrementing, we just care about
> it being globally unique. IOW, it could easily be a 56-bit per-cpu
> counter along with the CPU number in the high bits or something like
> that. Avoiding the whole atomicity issue, and thus avoiding the main
> reason those things are really expensive.
> 
> IOW, something like
> 
>    cpu = get_cpu();
>    .. increment percpu 64-bit counter ..
>    id = counter * MAX_CPUS + cpu;
>    put_cpu(cpu);
> 
> or equivalent would seem to be a potentially much cheaper approach.

This makes sense to me.  We just need to ensure that the per-CPU counter
is still large enough that it won't wrap.  56 bits is enough
(considering that the attack has to run on a single CPU of the target
system, unlike e.g. an attack on a cipher with a 56-bit key could be),
but there's little room for reducing this further (such as to support
many more than 256 CPUs in a system).  So if we use this approach, maybe
we should simply keep a 64-bit per-CPU counter and a 32-bit CPU number,
for a 96-bit id.  This may even be faster on execve() (no need to
combine the two values into one).

I don't know which one of these approaches has lower overhead on current
systems.  My _guess_ is that atomic64_inc_return() may well be faster in
many cases.

Alexander