Xen devel - Apr 2013 - Is it safe to hvm_emulate_one() an MMX instruction?

Hello,

assuming that RIP points to just before a MMX instruction (such as
''movdqu [some_address],xmm0''), is it safe to call
hvm_emulate_one() in
the hypervisor? Are MMX instructions handled in a special manner?


Thanks,
Razvan Cojocaru

>>> On 17.04.13 at 11:30, Razvan Cojocaru <rzvncj@gmail.com>
wrote:
> assuming that RIP points to just before a MMX instruction (such as
> ''movdqu [some_address],xmm0''), is it safe to call
hvm_emulate_one() in
> the hypervisor? Are MMX instructions handled in a special manner?
So do you mean an MMX instruction, or MOVDQU (which is an XMM
one)? In any event, 4.2 and later should be fine with this particular
instruction (and a few more move variants), but 4.1.x and earlier
won''t be, nor does the emulator support the whole spectrum of
MMX, XMM, and AVX instructions.

Jan

> So do you mean an MMX instruction, or MOVDQU (which is an XMM
> one)? In any event, 4.2 and later should be fine with this particular
> instruction (and a few more move variants), but 4.1.x and earlier
> won''t be, nor does the emulator support the whole spectrum of
> MMX, XMM, and AVX instructions.
In this particular case, MOVDQ, but I''m interested in all the mentioned
types of instructions.

What happens when a HVM guest tries to execute such an unsupported
instruction, via the hypervisor calling hvm_emulate_one()? Does
hvm_emulate_one() not return X86EMUL_OKAY - is testing for that enough?


Thanks,
Razvan Cojocaru

>>> On 17.04.13 at 12:40, Razvan Cojocaru <rzvncj@gmail.com>
wrote:
>>  So do you mean an MMX instruction, or MOVDQU (which is an XMM
>> one)? In any event, 4.2 and later should be fine with this particular
>> instruction (and a few more move variants), but 4.1.x and earlier
>> won''t be, nor does the emulator support the whole spectrum of
>> MMX, XMM, and AVX instructions.
> 
> In this particular case, MOVDQ, but I''m interested in all the
mentioned
> types of instructions.
> 
> What happens when a HVM guest tries to execute such an unsupported
> instruction, via the hypervisor calling hvm_emulate_one()? Does
> hvm_emulate_one() not return X86EMUL_OKAY - is testing for that enough?
Supposedly yes.

Jan