On Wed, Oct 16, 2019 at 12:40:32AM +0800, Xiaoyao Li wrote:
> Current x86 arch vcpu creation flow is a little bit messed.
> Specifically, vcpu's data structure allocation and vcpu initialization
> are mixed up, which is unfriendly to read.
>
> Seperating the vcpu_create and vcpu_init just like what ARM does, that
> it first calls vcpu_create related functions for vcpu's data structure
> allocation and then calls vcpu_init related functions to initialize the
> vcpu.
My vote is to take advantage of the requirement that @vcpu must reside at
offset 0 in vmx_vcpu and svm_vcpu, and allocate the vcpu in x86 code.
That would allow kvm_arch_vcpu_create() to invoke kvm_vcpu_init() directly
instead of bouncing through the vendor code.
And if we're extra lucky and the other architectures can use a similar
pattern, kvm_vm_ioctl_create_vcpu() could be refactored to something like:
vcpu = kvm_arch_vcpu_alloc(kvm, id);
if (IS_ERR(vcpu)) {
r = PTR_ERR(vcpu);
goto vcpu_decrement;
}
r = kvm_arch_vcpu_init(vcpu);
if (r)
goto vcpu_destroy;
