Am 10.01.2011 22:06, Jan Kiszka wrote:
>>  kvmclock should be created with
>> kvm_state as a parameter and kvm_vm_ioctl() is passed the stored
>> reference.   Taking a global reference to kvm_state in machine_init is
>> not a bad thing, obviously the machine initialization function needs
>> access to the kvm_state.
> 
> This would also require changing sysbus interfaces for the sake of KVM's
> "abstraction". If this is the only way forward, I could look into this.

Actually, there is already a channel to pass pointers to qdev devices:
the pointer property hack. I'm not sure we should contribute to its user
base or take the chance for a cleanup, but we are not alone with this
requirement. Point below remains valid, though.

> 
> Still, I do not see any benefit for the affected code. You then either
> need to "steal" a kvm_state reference from the first cpu or introduce a
> marvelous interface like kvm_get_state() to make this work from outside
> of the KVM core.
> 

Jan