Re: [PATCH 01/15] KVM: PPC: Make register read/write wrappers always work

[Alexander Graf <agraf@suse.de>]

Avi Kivity wrote:
> On 03/08/2010 03:44 PM, Alexander Graf wrote:
>> Avi Kivity wrote:
>>   
>>> On 03/05/2010 06:50 PM, Alexander Graf wrote:
>>>     
>>>> We have wrappers to do for example gpr read/write accesses with,
>>>> because the contents of registers could be either in the PACA
>>>> or in the VCPU struct.
>>>>
>>>> There's nothing that says we have to have the guest vcpu loaded
>>>> when using these wrappers though, so let's introduce a flag that
>>>> tells us whether we're inside a vcpu_load context.
>>>>
>>>>
>>>>        
>>> On x86 we always access registers within vcpu_load() context.  That
>>> simplifies things.  Does this not apply here?
>>>
>>> Even so, sometimes guest registers are present on the cpu, and
>>> sometimes in shadow variables (for example, msrs might be loaded or
>>> not).  The approach here is to always unload and access the variable
>>> data.  See for example vmx_set_msr() calling vmx_load_host_state()
>>> before accessing msrs.
>>>
>>> Seems like this could reduce the if () tree?
>>>      
>> Well - it would probably render this particular patch void. In fact, I
>> think it is already useless thanks to the other "always do vcpu_load"
>> patch.
>>
>> As far as the already existing if goes, we can't really get rid of that.
>> I want to be fast in the instruction emulation. Copying around the
>> registers won't help there.
>>    
>
> So do it the other way around.  Always load the registers (of course,
> do nothing if already loaded) and then access them in just one way.  I
> assume during emulation the registers will always be loaded?

During emulation we're always in VCPU_RUN, so the vcpu is loaded.

Do you mean something like:

read_register(num) {
  vcpu_load();
  read register from PACA(num);
  vcpu_put();
}

? Does vcpu_load incur overhead when it doesnt' need to do anything?


Alex