Re: [PATCH v2 3/3] x86/fpu/xstate: Invalidate fpregs when __fpu_restore_sig() fails

[Yu-cheng Yu <yu-cheng.yu@intel.com>]

On Wed, 2019-12-18 at 16:54 +0100, Sebastian Andrzej Siewior wrote:
> On 2019-12-12 13:08:55 [-0800], Yu-cheng Yu wrote:
> > In __fpu_restore_sig(),'init_fpstate.xsave' and part of 'fpu->state.xsave'
> > are restored separately to xregs.  However, as stated in __cpu_invalidate_
> > fpregs_state(),
> > 
> >   Any code that clobbers the FPU registers or updates the in-memory
> >   FPU state for a task MUST let the rest of the kernel know that the
> >   FPU registers are no longer valid for this task.
> > 
> > and this code violates that rule.  Should the restoration fail, the other
> > task's context is corrupted.
> > 
> > This problem does not occur very often because copy_*_to_xregs() succeeds
> > most of the time.  
> 
> why "most of the time"? It should always succeed. We talk here about
> __fpu__restore_sig() correct? Using init_fpstate as part of restore
> process isn't the "default" case. If the restore _here_ fails then it
> fails.
> 
> >                    It occurs, for instance, in copy_user_to_fpregs_
> > zeroing() when the first half of the restoration succeeds and the other
> > half fails.  This can be triggered by running glibc tests, where a non-
> > present user stack page causes the XRSTOR to fail.
> 
> So if copy_user_to_fpregs_zeroing() fails then we go to the slowpath.
> Then we load the FPU register with copy_kernel_to_xregs_err().
> In the end they are either enabled (fpregs_mark_activate()) or cleared
> if it failed (fpu__clear()). Don't see here a problem.

I could have explained this better, sorry!  I will explain the first
case below; other cases are similar.

In copy_user_to_fpregs_zeroing(), we have:

    if (user_xsave()) {
        ...
        if (unlikely(init_bv))
            copy_kernel_to_xregs(&init_fpstate.xsave, init_bv);
        return copy_user_to_xregs(buf, xbv);
        ...
    }

The copy_user_to_xregs() may fail, and when that happens, before going to
the slow path, there is fpregs_unlock() and context switches may happen.
However, at this point, fpu_fpregs_owner_ctx has not been changed; it could
still be another task's FPU.  For this to happen and to be detected, the user
stack page needs to be non-present, fpu_fpregs_owner_ctx need to be another task,
and that other task needs to be able to detect its registers are modified.
The last factor is not easy to reproduce, and a CET control-protection fault
helps.

> 
> Can you tell me which glibc test? I would like to reproduce this.
> 
> > The introduction of supervisor xstates and CET, while not contributing to
> > the problem, makes it more detectable.  After init_fpstate and the Shadow
> > Stack pointer have been restored to xregs, the XRSTOR from user stack
> > fails and fpu_fpregs_owner_ctx is not updated.  The task currently owning
> > fpregs then uses the corrupted Shadow Stack pointer and triggers a control-
> > protection fault.
> 
> So I don't need new HW with supervisor and CET? A plain KVM box with
> SSE2 and so should be enough?

What I do is, clone the whole glibc source, and run mutiple copies of
"make check".  In about 40 minutes or so, there are unexplained seg faults,
or a few control-protection faults (if you enable CET).  Please let me
know if more clarification is needed.

Thanks,
Yu-cheng