Re: [Intel-gfx] [PATCH 1/2] KVM: async kvm_destroy_vm for vfio devices

[Jason Gunthorpe <jgg@nvidia.com>]

On Wed, Jan 11, 2023 at 07:54:51PM +0000, Sean Christopherson wrote:

> Something feels off.  If KVM's refcount is 0, then accessing device->group->kvm
> in vfio_device_open() can't happen unless there's a refcounting bug somewhere.

The problem is in close, not open.

Specifically it would be very hard to avoid holding the group_lock
during close which is when the put is done.

> Rather than force devices to get KVM references, why not handle that in common
> VFIO code and drop KVM refcountin from devices?  Worst case scenario KVM is pinned
> by a device that doesn't need KVM but is in a group associated with KVM.  If that's
> a concern, it seems easy enough to add a flag to vfio_device_ops to enumerate
> whether or not the device depends on KVM.

We can't make cross-dependencies between kvm and core VFIO - it is why
so much of this is soo ugly.

The few device drivers that unavoidably have KVM involvment already
have a KVM module dependency, so they can safely do the get/put

Jason