Re: [PATCH v2] vfio/pci: Fix vf_token mechanism when device-specific VF drivers are used

[Alex Williamson <alex.williamson@redhat.com>]

On Mon, 11 Apr 2022 10:56:31 -0300
Jason Gunthorpe <jgg@nvidia.com> wrote:

> @@ -1732,10 +1705,28 @@ static int vfio_pci_bus_notifier(struct notifier_block *nb,
>  static int vfio_pci_vf_init(struct vfio_pci_core_device *vdev)
>  {
>  	struct pci_dev *pdev = vdev->pdev;
> +	struct vfio_pci_core_device *cur;
> +	struct pci_dev *physfn;
>  	int ret;
>  
> -	if (!pdev->is_physfn)
> +	if (!pdev->is_physfn) {
> +		/*
> +		 * If this VF was created by our vfio_pci_core_sriov_configure()
> +		 * then we can find the PF vfio_pci_core_device now, and due to
> +		 * the locking in pci_disable_sriov() it cannot change until
> +		 * this VF device driver is removed.
> +		 */
> +		physfn = pci_physfn(vdev->pdev);
> +		mutex_lock(&vfio_pci_sriov_pfs_mutex);
> +		list_for_each_entry (cur, &vfio_pci_sriov_pfs, sriov_pfs_item) {
> +			if (cur->pdev == physfn) {
> +				vdev->sriov_pf_core_dev = cur;
> +				break;
> +			}
> +		}
> +		mutex_unlock(&vfio_pci_sriov_pfs_mutex);
>  		return 0;
> +	}
>  
>  	vdev->vf_token = kzalloc(sizeof(*vdev->vf_token), GFP_KERNEL);
>  	if (!vdev->vf_token)

One more comment on final review; are we equating !is_physfn to
is_virtfn above?  This branch was originally meant to kick out both VFs
and non-SRIOV PFs.  Calling pci_physfn() on a !is_virtfn device will
return itself, so we should never find a list match, but we also don't
need to look for a match for !is_virtfn, so it's a bit confusing and
slightly inefficient.  Should the new code be added in a separate
is_virtfn branch above the existing !is_physfn test?  Thanks,

Alex