Re: [PATCH v5 3/7] x86/fpu/xstate: Differentiate default features for host and guest FPUs

[Sean Christopherson <seanjc@google.com>]

On Fri, Apr 25, 2025, Rick P Edgecombe wrote:
> On Fri, 2025-04-25 at 16:24 +0800, Chao Gao wrote:
> > > 
> > > In the later patches, it doesn't seem to change the "user" parts. These
> > > configurations end up controlling the default size and features that gets
> > > copied
> > > to userspace in KVM_SET_XSAVE. I guess today there is only one default size
> > > and
> > > feature set for xstate copied to userspace. The suggestion from Chang was
> > > that
> > > it makes the code more readable, but it seems like it also breaks apart a
> > > unified concept for no functional benefit.
> > 
> > In the future, the feature and size of the uABI buffer for guest FPUs may
> > differ from those of non-guest FPUs. Sean rejected the idea of
> > saving/restoring
> > CET_S xstate in KVM partly because:
> > 
> >  :Especially because another big negative is that not utilizing XSTATE bleeds
> > into
> >  :KVM's ABI.  Userspace has to be told to manually save+restore MSRs instead
> > of just
> >  :letting KVM_{G,S}ET_XSAVE handle the state.
> 
> Hmm, interesting. I guess there are two things.
> 1. Should CET_S be part of KVM_GET_XSAVE instead of via MSRs ioctls? It never
> was in the KVM CET patches.
> 2. A feature mask far away in the FPU code controls KVM's xsave ABI.
> 
> For (1), does any userspace depend on their not being supervisor features? (i.e.
> tries to restore the guest FPU for emulation or something). There probably are
> some advantages to keeping supervisor features out of it, or at least a separate
> ioctl.

CET_S probably shouldn't be in XSAVE ABI, because that would technically leak
kernel state to userspace for the non-KVM use case.  I assume the kernel has
bigger problems if CET_S is somehow tied to a userspace task.

For KVM, it's just the one MSR, and KVM needs to support save/restore of that MSR
no matter what, so supporting it via XSAVE would be more work, a bit sketchy, and
create yet another way for userspace to do weird things when saving/restoring vCPU
state.

> (2) is an existing problem. But if we think KVM should have its own
> feature set of bits for ABI purposes, it seems like maybe it should have some
> dedicated consideration. 

Nah, don't bother.  The kernel needs to solve the exact same problems for the
signal ABI, I don't see any reason to generate more work.  From a validation
coverage perspective, I see a lot of value in shared code.