Re: [PATCH] arch/x86: reset MXCSR to default in kernel_fpu_begin()

[Shuah Khan]

On 6/2/20 1:50 PM, Andy Lutomirski wrote:
> > On Jun 2, 2020, at 10:27 AM, Shuah Khan <skhan@xxxxxxxxxxxxxxxxxxx> wrote:
> >
> > ïOn 6/2/20 11:03 AM, Andy Lutomirski wrote:
> > > > On Tue, Jun 2, 2020 at 3:56 AM Borislav Petkov <bp@xxxxxxxxx> wrote:
> > > >
> > > > Hi,
> > > >
> > > > On Tue, Jun 02, 2020 at 01:29:51PM +0300, Petteri Aimonen wrote:
> > > > > The kernel module is not actually x86-specific, even though it is
> > > > > currently only enabled for x86. amdgpu driver already does kernel mode
> > > > > floating point operations on PPC64 also, and the same module could be
> > > > > used to test the same thing there.
> > > >
> > > > Then make it generic please and put the user portion in, say,
> > > > tools/testing/selftests/fpu/ and we can ask ppc people to test it too.
> > > > People might wanna add more stuff to it in the future, which would be
> > > > good.
> > > >
> > > > > To deterministically trigger the bug, the syscall has to come from the
> > > > > same thread that has modified MXCSR. Going through /usr/sbin/modprobe
> > > > > won't work, and manually doing the necessary syscalls for module loading
> > > > > seems too complicated.
> > > >
> > > > Ok, fair enough. But put that file in debugfs pls.
> > > I think I agree.  While it would be delightful to have general
> > > selftest tooling for kernel modules, we don't have that right now, and
> > > having the test just work with an appropriately configured kernel
> > > would be nice.
> >
> > Let's extend it to do what we want it to do. I will happy to take
> > patches. If you have some concrete ideas on what we can add, please
> > do a short summary of what is missing. I will find a way to get this
> > done.
> >
> > > How about putting the file you frob in
> > > /sys/kernel/debug/selftest_helpers/something_or_other.  The idea would
> > > be that /sys/kernel/debug/selftest_helpers would be a general place
> > > for kernel helpers needed to make selftests work.
> >
> > Is this a workaround for the lack of selftest tooling for kernel
> > modules? In which case, let's us focus on fix selftest tooling.
>
> The goal here is to have a selftest that runs kernel code as part of its operation. That is, the selftest is, logically, starting in userspace:
>
> setup_evil_state();

Is it correct to assume the stuff checked differs from test to test
and done in user-space.

> ret = call_kernel_helper();

> check_some_other_stuff();

Is it correct to assume the stuff checked differs from test to test
and done in user-space.

> undo_evil_state();

Is it correct to assume undoing evil differs from test to test
and done in user-space, provide it can be done from userspace.

> And the call_kernel_helper() could be moderately specific to the test.
The overall plan sounds good to me. I am all for adding support to
selftests so we can keep extending it.

thanks,
-- Shuah