Re: [PATCH v3 5/5] mips/kvm: Fix ABI by moving manipulation of CP0 registers to KVM_{G,S}ET_MSRS

[David Daney <ddaney@caviumnetworks.com>]

On 05/21/2013 09:28 AM, Gleb Natapov wrote:
> On Tue, May 21, 2013 at 09:21:06AM -0700, David Daney wrote:
>> On 05/21/2013 08:37 AM, Gleb Natapov wrote:
>>> On Mon, May 20, 2013 at 02:01:26PM -0700, David Daney wrote:
>>>> From: David Daney <david.daney@cavium.com>
>>>>
>>>> Because not all 256 CP0 registers are ever implemented, we need a
>>>> different method of manipulating them.  Use the
>>>> KVM_GET_MSRS/KVM_SET_MSRS mechanism as x86 does for its MSRs.
>>>>
>>> Have you looked at KVM_(GET|SET)_ONE_REG interface (not used by x86, but is
>>> used bu arm/ppc/s390). It looks like it is more suitable for your case.
>>> Actually you can use it instead of KVM_(GET|SET)_REGS for all registers.
>>
>> Yes, I suppose it could be used.  One problem it has is that there
>> is no way to query the set of supported registers.
> KVM_GET_REG_LIST

OK, I guess we can switch to KVM_GET_REG_LIST/KVM_(GET|SET)_ONE_REG.

I will revise this patch.

Thanks.

>
>>                                                      Also you have to
>> make multiple calls to set multiple registers, which involves
>> vcpu_{load,put} for each register.
>>
> How often this happens on the fast path on mips?

Rarely, if ever once, the VCPU is running.

Only when initializing the VCPU do we need to set a bunch of registers.


> On x86 this never
> happens on the fast path so it uses KVM_(GET|SET)_REGS mostly for
> historical reasons.
>
>> We will definitely implement it for all the FP and General Purpose
>> registers.
>>
>>>
>>>> Code related to implementing KVM_GET_MSRS/KVM_SET_MSRS is consolidated
>>
>
> --
> 			Gleb.
>