Re: [PATCH] capability: WARN when invalid capability is requested rather than BUG/panic

[Eric Paris <eparis@redhat.com>]

On Tue, 2008-09-30 at 10:38 -0500, Serge E. Hallyn wrote:
> Quoting Eric Paris (eparis@redhat.com):
> > On Wed, 2008-10-01 at 00:23 +1000, James Morris wrote:
> > > On Tue, 30 Sep 2008, Eric Paris wrote:
> > > 
> > > > This patch adds a WARN_ONCE() to cap_capable() so we will stop
> > > > dereferencing random spots of memory and will cleanly tell the obviously
> > > > broken driver that it doesn't have that ridiculous permissions.  No idea
> > > > if the driver is going to handle EPERM but anything that calls capable
> > > > and doesn't expect a denial has got to be the worst piece of code ever
> > > > written.....  I could return EINVAL, but I think its clear that noone
> > > > has capabilities over 64 so clearly they don't have that permission.
> > > > 
> > > > This 'could' be considered a regression since 2.6.24.  Neither SELinux
> > > > nor the capabilities system had a problem with ginormous request values
> > > > until we got 64 bit support, although this is OBVIOUSLY a bug with the
> > > > out of tree closed source driver....
> > > 
> > > An issue here is whether we should be adding workarounds in the mainline 
> > > kernel for buggy closed drivers.  Papering over problems rather than 
> > > getting them fixed does not seem like a winning approach.  Especially 
> > > problems which are unexpectedly messing with kernel security APIs.
> > 
> > I don't know, looking at the feelings on "Can userspace bugs be kernel
> > regressions" leads me to believe that when we break something that once
> > worked we are supposed to fix it.
> > 
> > http://lwn.net/Articles/292143/
> > 
> > I don't think the proprietary closed source nature of the driver makes
> > it any less our problem
> 
> The kernel-space nature of the driver is the distinction here.
> 
> > to not make changes which cause the kernel to
> > esplode.
> > 
> > > Also, won't this encourage vendors of such drivers to continue with this 
> > > behavior, while discouraging those vendors who are doing the right thing?
> > 
> > Discouraging people who open source their drivers and put them in the
> > kernel?  obviously not.  encouraging crap?  well, I hope we fix
> > regressions no matter how they are found...
> > 
> > > Do we know if this even really helps the user?  For all we know, the 
> > > driver may simply crash differently with an -EPERM.
> > 
> > Well, before the 64 bit capabilities change we did:
> > 
> > (cap_t(c) & CAP_TO_MASK(flag))
> > 
> > so a huge value for "flag" got masked off.
> > 
> > After 64 bit capabilities we do:
> > 
> > ((c).cap[CAP_TO_INDEX(flag)] & CAP_TO_MASK(flag))
> 
> Perhaps we should have CAP_TO_INDEX mask itself?
> 
> #define CAP_TO_INDEX(x)		(((x) >> 5) & _KERNEL_CAPABILITY_U32S)

Well, you save a branch and won't get the pagefault so it does 'fix' the
pagefault/panic from cap code.  It doesn't tell us when others screw up
and SELinux is still possibly going to BUG().  We are also going to
actually be returning a permission decision not on what was requested
but on something wholely different.

I like mine better, but I'm ok with yours and can just do my changes in
SELinux if this is how cap wants to handle it.  I don't really like the
idea of mutating the inputs and then making the security decision based
on that mutation rather than on the original inputs (and yes, I realize
that exactly what 2.6.24 was doing)

> Though I still think it's not unreasonable to simply ask for the driver
> to be fixed.

I'm not going to argue that the driver needs fixed and that is the real
problem.  I know its been filed with them and the response was that
there is no support for linux.  I have today tried to poke the path I
know of between Red Hat and them to ask them to take a look.