Re: [PATCH v2 2/2] x86/fpu: Split fpu_setup_fpu() in two

[Jan Beulich <jbeulich@xxxxxxxx>]

On 13.08.2024 14:40, Alejandro Vallejo wrote:
> On Mon Aug 12, 2024 at 4:23 PM BST, Jan Beulich wrote:
>> On 08.08.2024 15:41, Alejandro Vallejo wrote:
>>> --- a/xen/arch/x86/hvm/hvm.c
>>> +++ b/xen/arch/x86/hvm/hvm.c
>>> @@ -1164,10 +1164,25 @@ static int cf_check hvm_load_cpu_ctxt(struct domain 
>>> *d, hvm_domain_context_t *h)
>>>      seg.attr = ctxt.ldtr_arbytes;
>>>      hvm_set_segment_register(v, x86_seg_ldtr, &seg);
>>>  
>>> -    /* Cover xsave-absent save file restoration on xsave-capable host. */
>>> -    vcpu_setup_fpu(v, xsave_enabled(v) ? NULL : v->arch.xsave_area,
>>> -                   ctxt.flags & XEN_X86_FPU_INITIALISED ? ctxt.fpu_regs : 
>>> NULL,
>>> -                   FCW_RESET);
>>> +    /*
>>> +     * On Xen 4.1 and later the FPU state is restored on later HVM context 
>>> in
>>> +     * the migrate stream, so what we're doing here is initialising the FPU
>>> +     * state for guests from even older versions of Xen.
>>> +     *
>>> +     * In particular:
>>> +     *   1. If there's an XSAVE context later in the stream what we do 
>>> here for
>>> +     *      the FPU doesn't matter because it'll be overriden later.
>>> +     *   2. If there isn't and the guest didn't use extended states it's 
>>> still
>>> +     *      fine because we have all the information we need here.
>>> +     *   3. If there isn't and the guest DID use extended states (could've
>>> +     *      happened prior to Xen 4.1) then we're in a pickle because we 
>>> have
>>> +     *      to make up non-existing state. For this case we initialise the 
>>> FPU
>>> +     *      as using x87/SSE only because the rest of the state is gone.
>>
>> Was this really possible to happen? Guests wouldn't have been able to
>> turn on CR4.OSXSAVE, would they?
> 
> You may be right, but my reading of the comment and the code was that
> xsave_enabled(v) might be set and the XSAVE hvm context might be missing in 
> the
> stream. The archives didn't shed a lot more light than what the code already
> gives away.
> 
> Otherwise it would've been far simpler to unconditionally pass
> v->arch.xsave_area to the second parameter and let the xsave area to be
> overriden by the follow-up HVM context with its actual state.
> 
> If my understanding is wrong, I'm happy to remove (3), as I don't think it
> affects the code anyway. I thought however that it was a relevant data point
> to leave paper trail for.

I would certainly agree - as long as it describes (past) reality. If it
doesn't, I consider it misleading.

Jan