On 29.09.14 14:21, Paolo Bonzini wrote:
> Il 29/09/2014 13:53, Alexander Graf ha scritto:
>>
>> cpu_handle_ioreq()
>> {
>> ...
>>
>> if (req->type == IOPORT_REGISTERS_SYNCHRONIZED) {
>> cpu->xen_vcpu_dirty = true;
>> synchronize_xen_to_env(xenptr, cpu);
>> }
>>
>> handle_ioreq();
>>
>> if (req->type == IOPORT_REGISTERS_SYNCHRONIZED) {
>> cpu->xen_vcpu_dirty = false;
>> synchronize_env_to_xen(xenptr, cpu);
>> }
>>
>> ...
>> }
>>
>> void xen_cpu_synchronize_state(CPUState *cpu)
>> {
>> assert(cpu->xen_vcpu_dirty);
>> }
>>
>> Then no changes to the vmport code would be necessary and this problems
>> where some code path wants to do direct access to registers
>> automatically tells us that things are broken.
>
> Yeah, that would be possible. You do not even need synchronize_state,
> it seems to me that it introduces more complication for little gain.
Well, it makes all accels behave the same and keep information always at
a single entity (the env struct). I don't think the vmport code should
have knowledge of a xen env struct.
Alex
