Re: XSAVE flavors

["Jan Beulich" <JBeulich@suse.com>]

>>> On 02.02.16 at 07:31, <shuai.ruan@linux.intel.com> wrote:
> On Tue, Jan 26, 2016 at 08:12:20AM -0700, Jan Beulich wrote:
>> >>> On 26.01.16 at 15:33, <JBeulich@suse.com> wrote:
>> > originally I only meant to inquire about the state of the promised
>> > alternatives improvement to the XSAVE code. However, while
>> > looking over the code in question again I stumbled across a
>> > separate issue: XSAVES, just like XSAVEOPT, may use the
>> > "modified" optimization. However, the fcs and fds handling code
>> > that has been present around the use of XSAVEOPT did not also
>> > get applied to the XSAVES path. I suppose this was just an
>> > oversight?
> Really sorry for late response. The alternatives on xsave code is ok a 
> couples 
> weeks ago, the patch solve xsaves use modified optimization problem.
> I will send it now.

Thanks, But this response of yours covers only one half of what
I've pointed out.

>> > With this another question then is whether, when both XSAVEC
>> > and XSAVEOPT are available, it is indeed always better to use
>> > XSAVEC (as the code is doing after your enabling).
> Yes.
> But current no machine only support xsavec not support xsaves.  
> I enable xsavec for "xsavec is a feature".

But this shouldn't preclude the code being in reasonable shape
also for the case where a CPU has XSAVEC but no XSAVES. The
more that right now we don't really need XSAVES (since we don't
yet allow any bit to get set in XSS).

>> And I'm afraid there's yet one more issue: If my reading of the
>> SDM is right, then the offsets at which components get saved
>> by XSAVEC / XSAVES aren't fixed, but depend on RFBM (as that's
>> what gets stored into xcomp_bv[62:0]). xstate_comp_offsets[],
>> otoh, gets computed based on all available features, irrespective
>> of vcpu_xsave_mask() returning four different values depending
>> on current guest state. I can't see how get_xsave_addr() can
>> work correctly without honoring xcomp_bv. Nor can I convince
>> myself that state can't get corrupted / lost, e.g. when a save
>> with v->fpu_dirtied set is followed by one with v->fpu_dirtied
>> clear.
>> 
>> Am I misunderstanding what the SDM writes?
>> 
> Yes. you are right. This is a issue. I will find a way to solve
> this.

Thanks.

Jan