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

[Vasiliy Kulikov <segoon@openwall.com>]

On Fri, Feb 17, 2012 at 09:54 -0800, Greg KH wrote:
> I'm referring to the fact that the use of kref in this type of error or
> problem is rare.
> 
> Yes, we have these types of problems at times, but a kref doesn't seem
> to be involved in them that I know of, so changing the kref code
> wouldn't help here from what I can tell.

Ehr, what's the difference between kref and "raw" atomic_t in a refcounting case?
There is _no_ difference in sense of overflows as a kref uses the same atomic_t.

I second David that we should use kref for overflow protection: we want to
hook an overflow case somehow in cases atomic_t is used as a refcounter.  It is
_ideally_ handled by introducing atomic_t's subtype.  And this subtype already
exists - it is called kref.


I expect all atomic_t refcounters users have

	if (atomic_dec_and_test()) smth_put() 

pattern, otherwise it is not a true refcounter :)  It should be straightforward to
move to kref.


Moving to atomic64_t is attractive, but:

1) we still should find all atomic_t refcounters.  Why not move to kref then?

2) what to do with architectures-loosers?


Thanks,

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