On 10/16/2019 3:35 PM, Paolo Bonzini wrote:
On 16/10/19 03:52, Xiaoyao Li wrote:
user_fpu could be made percpu too... That would save a bit of memory
for each vCPU. I'm holding on Xiaoyao's patch because a lot of the code
he's touching would go away then.
Sorry, I don't get clear your attitude.
Do you mean the generic common function is not so better that I'd better
to implement the percpu solution?
I wanted some time to give further thought to the percpu user_fpu idea.
But kvm_load_guest_fpu and kvm_put_guest_fpu are not part of vcpu_load,
so it would not be so easy. I'll just apply your patch now.
Got it, thanks.
BTW, could you have a look at the series I sent yesterday to refactor
the vcpu creation flow, which is inspired partly by this issue. Any
comment and suggestion is welcomed since I don't want to waste time on
wrong direction.
