Re: [kernel-hardening] Re: Add overflow protection to kref

[Vasiliy Kulikov <segoon@openwall.com>]

On Sat, Feb 18, 2012 at 11:15 -0500, David Windsor wrote:
> On Fri, Feb 17, 2012 at 8:44 PM, Roland Dreier <roland@purestorage.com> wrote:
> > On Fri, Feb 17, 2012 at 3:39 PM, Djalal Harouni <tixxdz@opendz.org> wrote:
> >>> 2) what to do with architectures-loosers?
> >> There is lib/atomic64.c but with a static hashed array of raw_spinlocks.
> >
> > Even leaving aside performance impact of atomic64_t (and probably
> > in most cases the performance of kref is not important at all), it is
> > unfortunate to bloat the size from 4 bytes to 8 bytes.
> >
> > It seems much better to have some out-of-line code for overflow
> > checking rather than increasing the size of every data structure
> > that embeds a kref.
> >
> 
> kref is mostly a set of operations (init, get, sub, put) to be
> performed on an atomic_t object.
> 
> >From linux/kref.h:
> 
> struct kref {
>     atomic_t refcount;
> };
> 
> Moving overflow protection into kref amounts to placing some
> procedural code into kref_get and kref_sub, adding a rather small
> constant factor of time, not space, to users of kref.  Introducing
> overflow protection doesn't necessitate adding anything to kref for
> greater state tracking.
> 
> Did you have something else in mind when you suggested a potential
> increase in the size of kref?

4 bytes => 8 bytes of atomic_t => atomic64_t in case we increase the refcounter
range to make it impossible to overflow the refcounter

compared to

add checks into kref_get()/atomic_inc*() without changing refcounter ranges.


Thanks,

-- 
Vasiliy Kulikov
http://www.openwall.com - bringing security into open computing environments