Re: [PATCH v4 RESEND 3/5] KVM: x86: hyperv: use get_vcpu_by_vpidx() in kvm_hv_flush_tlb()

From: Paolo Bonzini
Date: Tue Sep 25 2018 - 05:30:05 EST


On 25/09/2018 10:57, Roman Kagan wrote:
> Speaking of the options we have, the choice depends on the assumptions
> we take. (And I guess when you spoke of quadratic complexity you
> referred to the algorithm to convert the vp_index mask into the KVM cpu
> mask.)

Right; with Vitaly's patch, if you have a random mapping between
vp_index and cpu index, each loop requires a list walk, and so you have
O(#VMcpus * #IPIcpus) worst case for sending an IPI to #IPIcpus CPUs in
a guest with #VMcpus.

> If we can assume that in all relevant cases vp_index coincides with the
> cpu index (which I think we can) then Vitaly's approach is the most
> efficient.
>
> If, on the opposite, we want to optimize for random mapping between
> vp_index and cpu index, then it's probably better instead to iterate
> over vcpus and test if their vp_index belongs to the requested mask.

Yes, that would work too. Perhaps we can do both? You can have a
kvm->num_mismatched_vp_indexes count to choose between the two.

Paolo

> Neither of the above is quadratic.

> Dunno if we need to specifically consider intermediate situations.