[Security] [PATCH 00/20] world-writable files in sysfs and debugfs

[James Bottomley]

On Tue, 2011-03-15 at 19:08 +0300, Vasiliy Kulikov wrote:
> On Tue, Mar 15, 2011 at 07:50 -0400, James Bottomley wrote:
> >      1. Did anyone actually check for capabilities before assuming world
> >         writeable files were wrong?
> 
> I didn't check all these files as I haven't got these hardware :-)

You don't need the hardware to check ... the question becomes is a
capabilities test sitting in the implementation or not.

>   But
> as I can "chmod a+w" all sysfs files on my machine and they all become
> sensible to nonroot writes, I suppose there is nothing preventing
> nonroot users from writing to these buggy sysfs files.  As you can see,
> there are no capable() checks in these drivers in open() or write().
> 
> >      2. Even if there aren't any capabilities checks in the implementing
> >         routines, should there be (are we going the separated
> >         capabilities route vs the monolithic root route)?
> 
> IMO, In any case old good DAC security model must not be obsoleted just
> because someone thinks that MAC or anything else is more convenient for
> him.  If sysfs is implemented via filesystem then it must support POSIX
> permissions semantic.  MAC is very good in _some_ cases, but not instead
> of DAC.

Um, I'm not sure that's even an issue.  capabilities have CAP_ADMIN
which is precisely the same check as owner == root.  We use this a lot
because ioctls ignore the standard unix DAC model.

James