On 17/08/2017 11:55, David Hildenbrand wrote: > On 17.08.2017 11:44, Paolo Bonzini wrote: >> On 17/08/2017 11:28, Cornelia Huck wrote: >>> On Thu, 17 Aug 2017 11:16:59 +0200 >>> Paolo Bonzini <pbonzini@xxxxxxxxxx> wrote: >>> >>>> On 17/08/2017 09:36, Cornelia Huck wrote: >>>>>> What if we just sent a "vcpu move" request to all vcpus with the new >>>>>> pointer after it moved? That way the vcpu thread itself would be >>>>>> responsible for the migration to the new memory region. Only if all >>>>>> vcpus successfully moved, keep rolling (and allow foreign get_vcpu again). >>>>>> >>>>>> That way we should be basically lock-less and scale well. For additional >>>>>> icing, feel free to increase the vcpu array x2 every time it grows to >>>>>> not run into the slow path too often. >>>>> >>>>> I'd prefer the rcu approach: This is a mechanism already understood >>>>> well, no need to come up with a new one that will likely have its own >>>>> share of problems. >>>> >>>> What Alex is proposing _is_ RCU, except with a homegrown >>>> synchronize_rcu. Using kvm->srcu seems to be the best of both worlds. >>> >>> I'm worried a bit about the 'homegrown' part, though. >> >> I agree, that's why I'm suggesting SRCU instead. But it's a trick that >> has its uses. For example, if you were only doing reads from a work >> queue, flush_work_queue could be used as the "homegrown >> synchronize_rcu". In KVM you might use kvm_make_all_cpus_request, I guess. >> >>> I also may be misunderstanding what Alex means with "vcpu move"... >> >> My interpretation was "resizing the array" (so it moves in memory). > > Unpopular opinion: Let's keep it simple first (straight rcu) and > optimize later on. RCU vs. SRCU is about correctness, not optimization... Paolo
