Re: [PATCH v2 8/8] x86/fpu/xstate: Restore supervisor xstates for __fpu__restore_sig()

[Yu-cheng Yu]

On Fri, 2020-02-28 at 18:22 +0100, Borislav Petkov wrote:
> On Fri, Feb 28, 2020 at 08:20:27AM -0800, Yu-cheng Yu wrote:
> > When XSAVES writes to an xsave buffer, xsave->header.xcomp_bv is set to
> > include only saved components, effectively changing the buffer's format.
> 
> So you want to *save* the supervisor states and xcomp_bv will be set to
> supervisor states only and since we don't care about the user states
> there - they will be loaded later - we're good.

No!

> Or do you have to set xcomp_bv later in order to save the user
> components too and also rearrange the buffer to undo the format change
> above?

That is the case.  If we save only supervisor states, the buffer becomes
smaller and has only supervisor states.

> We have using_compacted_format() and we do conversion from compacted to
> standard buffers - I'm looking at copy_xstate_to_kernel() et al - so it
> shouldn't be impossible. So to repeat Sebastian's question which you
> ignored:
> 
> "How large is this supervisor state at most? I guess saving the AVX512
> state just to get the 2 bytes of the supervisor state at the right spot
> is not really optimal."

I thought Sebastian was saying XSAVES is not optimal.

CET has 16 bytes for ring-3 setting, 24 bytes for ring-0.
Saving supervisor states somewhere else and copying back is not better
either.

> In any case, this performance penalty better be paid only by those
> who are actually using some supervisor states. I haven't looked at
> the CET patchset but I'm assuming you're setting the CET bit in
> xfeatures_mask_all only when the feature is being actually used?

We save supervisor states only when xfeatures_mask_supervisor() is not
zero.

Yu-cheng