On Mon, Jan 11, 2016 at 04:29:03PM +0000, Peter Maydell wrote:

On 11 January 2016 at 16:21, Andrew Jones [off-list ref] wrote:

quoted

On Mon, Jan 11, 2016 at 05:09:27PM +0100, Andrew Jones wrote:

quoted

On Mon, Jan 11, 2016 at 04:09:29PM +0100, Christoffer Dall wrote:

quoted

Are vcpu ids already exposed to userspace (beyond the stupid
KVM_IRQ_LINE) ioctl and as such we're bound to whatever upper limit and
format they have?

The only other place I found is KVM_CREATE_VCPU. I suppose we could move
to MPIDR for that, and it would be a nice way to handle the "userspace
determines MPIDR" work that I plan to do. Both KVM and its userspaces
would still use some counter-based vcpu identifiers internally, to avoid
large, sparse structures, but I guess the advantage is that they don't
have to agree on how they do that. The 'vcpu id' used by KVM_CREATE_VCPU
is already 32-bits, and is supposed to be an arbitrary identifier. That
all looks good for converting to MPIDR.

Correction. I understand that vcpu-id is "supposed" to be an arbitrary
identifier now, but it doesn't appear that all the assumptions that it's
a counter are gone yet... virt/kvm/kvm_main.c has

static int kvm_vm_ioctl_create_vcpu(struct kvm *kvm, u32 id)
...
        if (id >= KVM_MAX_VCPUS)
                return -EINVAL;

I think the last time we talked about supporting "userspace
determines MPIDR" the idea was to do it by allowing userspace to
write to the MPIDR register with KVM_SET_ONE_REG. So you'd
create a bunch of CPUs with vcpu-ids as usual, and then the
MPIDRs would be set for them later as appropriate (or not
at all, if userspace was an older qemu).

Yup, I recall that. I'm just expanding this discussion into that one.
If we wanted to single vcpu identifier type, and we wanted it to be
MPIDR, then I guess we'd want to pass it in to KVM_CREATE_VCPU too, at
which point we no longer need to set it later with KVM_SET_ONE_REG.

Thanks,
drew