On Fri, Jun 29, 2018 at 05:21:47PM +0200, Vitaly Kuznetsov wrote: > Roman Kagan <rkagan@xxxxxxxxxxxxx> writes: > > > On Fri, Jun 29, 2018 at 04:14:53PM +0200, Vitaly Kuznetsov wrote: > >> VP_INDEX almost always matches VCPU id and get_vcpu_by_vpidx() is fast, > >> use it instead of traversing full vCPU list every time. > >> > >> To support the change switch kvm_make_vcpus_request_mask() to checking > >> vcpu_id instead of vcpu index, > > > > I'm afraid you can't do this: vcpu_id (== apic id) can be sparse, i.e. > > it's not very well suited for bitmaps and can exceed the max number of > > vcpus. > > True. The bitmap should be of KVM_MAX_VCPU_ID size, not > KVM_MAX_VCPUS. > > Unfortunately there's no convenient way to get VCPU idx from VCPU > id, kvm_vcpu_get_idx() just walks the whole list :-( I see two possible > options: > 1) Add vcpu_idx fields to struct kvm_vcpu > 2) Keep the change expecting masks of KVM_MAX_VCPU_ID in > kvm_make_vcpus_request_mask(). KVM_MAX_VCPU_ID is currently 1023 so our > bitmaps will be 16 longs long. Not sure if it's too much. 3) rework get_vcpu_by_vpidx into get_vcpu_idx_by_vpidx followed by get_cpu, and use the former for your purposes 4) duplicate get_vcpu_by_vpidx logic in get_vcpu_idx_by_vpidx Roman. P.S. I'm starting to wonder how safe this get_vcpu_* thing is WRT vcpu removal, but that's a different story anyway.
