Re: [PATCH v3 3/4] xen/hvm: introduce a fpu_uninitialised field to the CPU save record

["Roger Pau Monné" <roger.pau@citrix.com>]

El 24/11/15 a les 14.34, Jan Beulich ha escrit:
>>>> On 24.11.15 at 14:10, <roger.pau@citrix.com> wrote:
>> El 20/11/15 a les 16.49, Jan Beulich ha escrit:
>>>>>> On 18.11.15 at 17:37, <roger.pau@citrix.com> wrote:
>>>> @@ -2091,7 +2092,8 @@ static int hvm_load_cpu_ctxt(struct domain *d, hvm_domain_context_t *h)
>>>>          struct xsave_struct *xsave_area = v->arch.xsave_area;
>>>>  
>>>>          memcpy(v->arch.xsave_area, ctxt.fpu_regs, sizeof(ctxt.fpu_regs));
>>>> -        xsave_area->xsave_hdr.xstate_bv = XSTATE_FP_SSE;
>>>> +        xsave_area->xsave_hdr.xstate_bv = ctxt.fpu_initialised ?
>>>> +                                                    XSTATE_FP_SSE : 0;
>>>>      }
>>>>      else
>>>>          memcpy(v->arch.fpu_ctxt, ctxt.fpu_regs, sizeof(ctxt.fpu_regs));
>>>
>>> Question is - are the memcpy()s here really meaningful/valid
>>> when !ctxt.fpu_initialized? I.e. shouldn't this code rather be
>>> skipped instead of getting modified?
>>
>> If !fpu_initialized the fpu context save record is all zeroed out. I
>> don't think it matters much (apart from saving a few CPU cycles). I can
>> send a new version that doesn't save/restore the fpu context at all if
>> !fpu_initialised.
> 
> I'd appreciate that (ideally with if(!fpu_initialised) memset(); else if ...).
> 
>>>> @@ -157,6 +159,8 @@ struct hvm_hw_cpu {
>>>>      };
>>>>      /* error code for pending event */
>>>>      uint32_t error_code;
>>>> +    /* is fpu initialised? */
>>>> +    uint32_t fpu_initialised;
>>>
>>> A whole uint32_t for just one bit? Didn't we talk about making this
>>> new field a flags one, consuming just one bit from it?
>>
>> AFAIK we agreed on adding this field to the tail and making it a
>> uint32_t so that when new fields are added they can be detected by
>> looking at the size of the structure:
>>
>> http://marc.info/?l=xen-devel&m=144490321208291 
> 
> Admittedly it's a little implicit, but that mail has, in its quoting parts,
> 
> "... (and validate unused tail bits are zero, so they can be used for
> something later on)"
> 
> going back to that intention of using just a single bit here afaict.

Ack. I have to admit I've misunderstood that part. Then I guess the
field should also have a more generic name, like "flags", and
fpu_initialised should be defined as (1U << 0). Or do you want me to use
the MSB in order to store the fpu_initialised bit?

Roger.