On 2/1/2021 2:59 PM, Dave Hansen wrote:

On 2/1/21 2:43 PM, Yu, Yu-cheng wrote:

quoted

On 1/29/2021 2:53 PM, Dave Hansen wrote:

quoted

On 1/29/21 2:35 PM, Yu, Yu-cheng wrote:

quoted

quoted

Andy Cooper just mentioned on IRC about this nugget in the spec:

      XRSTORS on CET state will do reserved bit and canonicality
      checks on the state in similar manner as done by the WRMSR to
      these state elements.

We're using copy_kernel_to_xregs_err(), so the #GP *should* be OK.
Could we prove this out in practice, please?

quoted

Do we want to verify that setting reserved bits in CET XSAVES states
triggers GP?  Then, yes, I just verified it again.  Thanks for
reminding.  Do we have any particular case relating to this?

I want to confirm that it triggers #GP and kills userspace without the
kernel WARN'ing or otherwise being visibly unhappy.

For sigreturn, shadow stack pointer is checked against its restore token
and must be smaller than TASK_SIZE_MAX.  Sigreturn cannot set any
MSR_IA32_U_CET reserved bits.

That would be nice to at least allude to in the changelog or comments.

Ok.

quoted

quoted

What about the return-to-userspace path after a ptracer writes content
to the CET fields?   I don't see the same tolerance for errors in
__fpregs_load_activate(), for instance.

Good thought.  I have not sent out my revised PTRACE patch, but values
from user will be checked for valid address and reserved bits.

Wait a sec...  What about *THIS* series?  Will *THIS* series give us
oopses when userspace blasts a new XSAVE buffer in with NT_X86_XSTATE?

Fortunately, CET states are supervisor states.  NT_x86_XSTATE has only 
user states.