2017-04-24 18:22+0200, Radim Krčmář: > 2017-04-18 16:30+0200, Paolo Bonzini: >> On 18/04/2017 16:16, Paolo Bonzini wrote: >>>> This patch allows userspace to tell how many VCPUs it is going to use, >>>> which can save memory when allocating the kvm->vcpus array. This will >>>> be done with a new KVM_CREATE_VM2 IOCTL. >>>> >>>> An alternative would be to redo kvm->vcpus as a list or protect the >>>> array with RCU. RCU is slower and a list is not even practical as >>>> kvm->vcpus are being used for index-based accesses. >>>> >>>> We could have an IOCTL that is called in between KVM_CREATE_VM and first >>>> KVM_CREATE_VCPU and sets the size of the vcpus array, but we'd be making >>>> one useless allocation. Knowing the desired number of VCPUs from the >>>> beginning is seems best for now. >>>> >>>> This patch also prepares generic code for architectures that will set >>>> KVM_CONFIGURABLE_MAX_VCPUS to a non-zero value. >>> Why is KVM_MAX_VCPU_ID or KVM_MAX_VCPUS not enough? >> >> Ok, for KVM_MAX_VCPUS I should have read the cover letter more carefully. :) > > KVM_MAX_VCPU_ID makes sense as the upper bound, I just didn't want to > mingle the concepts, because the kvm->vcpus array is not indexed by > VCPU_ID ... > > In hindsight, it would be best to change that and get rid of the search. > I'll see how that looks in v2. I realized why not: * the major user of kvm->vcpu is kvm_for_each_vcpu and it works best with a packed array * at least arm KVM_IRQ_LINE uses the order in which cpus were created to communicate with userspace Putting this work into a drawer with the "do not share data structure between kvm_for_each_vcpu and kvm_get_vcpu" idea. :)
