On 9/1/21 6:00 AM, Borislav Petkov wrote:

quoted

So, this code takes the coward's way out: it *forces* TIF_NEED_FPU_LOAD
to be clear by making the registers live with fpregs_restore_userregs().
 That lets it just use WRMSR instead of dealing with the XSAVE buffer
directly.  If it didn't do this with the *WHOLE* set of user FPU state,
we'd need more fine-granted "NEED_*_LOAD" tracking than our one FPU bit.

This is also *only* safe because the task is newly-exec()'d and the FPU
state was just reset.  Otherwise, we might have had to worry that the
non-PL3 SSPs have garbage or that non-SHSTK bits are set in MSR_IA32_U_CET.

That said, after staring at it, I *think* this code is functionally
correct and OK performance-wise.

Right, except that that is being done in
setup_signal_shadow_stack()/restore_signal_shadow_stack() too, for the
restore token.

Which means, a potential XRSTOR each time just for a single MSR. That
means, twice per signal in the worst case.

Which means, shadow stack should be pretty noticeable in signal-heavy
benchmarks...

Ahh, good point.  I totally missed the signal side.

Yu-cheng, it would probably be wise to take a look at where
TIF_NEED_FPU_LOAD is set in the signal paths.  I suspect the new shadow
stack code is clearing it pretty quickly.  That's not *necessarily* a
waste because it has to be XRSTOR'd somewhere before returning to
userspace.  But, it does impose an extra XSAVE/XRSTOR burden if the code
is preempted somewhere between
setup_signal_shadow_stack()/restore_signal_shadow_stack() and the return
to usersspace.

Some performance numbers would be nice.

Even nicer would be to make your code work without requiring
TIF_NEED_FPU_LOAD to be clear, or actively clearing it.