Re: [PATCH v3 6/9] vfio: ap: register IOMMU VFIO notifier

[Tony Krowiak <akrowiak@linux.ibm.com>]

On 2/14/19 8:51 AM, Pierre Morel wrote:
> To be able to use the VFIO interface to facilitate the
> mediated device memory pining/unpining we need to register
> a notifier for IOMMU.
> 
> Signed-off-by: Pierre Morel <pmorel@linux.ibm.com>
> ---
>   drivers/s390/crypto/vfio_ap_ops.c     | 64 +++++++++++++++++++++++++++++++----
>   drivers/s390/crypto/vfio_ap_private.h |  2 ++
>   2 files changed, 60 insertions(+), 6 deletions(-)
> 
> diff --git a/drivers/s390/crypto/vfio_ap_ops.c b/drivers/s390/crypto/vfio_ap_ops.c
> index 1851b24..6eddc2c 100644
> --- a/drivers/s390/crypto/vfio_ap_ops.c
> +++ b/drivers/s390/crypto/vfio_ap_ops.c
> @@ -781,6 +781,36 @@ static const struct attribute_group *vfio_ap_mdev_attr_groups[] = {
>   };
>   
>   /**
> + * vfio_ap_mdev_iommu_notifier: IOMMU notifier callback
> + *
> + * @nb: The notifier block
> + * @action: Action to be taken (VFIO_IOMMU_NOTIFY_DMA_UNMAP)
> + * @data: the specific unmap structure for vfio_iommu_type1
> + *
> + * Unpins the guest IOVA. (The NIB guest address we pinned before).
> + * Return NOTIFY_OK after unpining on a UNMAP request.
> + * otherwise, returns NOTIFY_DONE .
> + */
> +static int vfio_ap_mdev_iommu_notifier(struct notifier_block *nb,
> +				       unsigned long action, void *data)
> +{
> +	struct ap_matrix_mdev *matrix_mdev;
> +
> +	matrix_mdev = container_of(nb, struct ap_matrix_mdev, iommu_notifier);
> +
> +	if (action == VFIO_IOMMU_NOTIFY_DMA_UNMAP) {
> +		struct vfio_iommu_type1_dma_unmap *unmap = data;
> +		unsigned long g_pfn = unmap->iova >> PAGE_SHIFT;
> +
> +		vfio_unpin_pages(mdev_dev(matrix_mdev->mdev), &g_pfn, 1);
> +		return NOTIFY_OK;
> +	}
> +
> +	return NOTIFY_DONE;
> +}
> +
> +
> +/**
>    * vfio_ap_mdev_set_kvm
>    *
>    * @matrix_mdev: a mediated matrix device
> @@ -904,8 +934,7 @@ static void vfio_ap_dissociate_queues(struct ap_matrix_mdev *matrix_mdev)
>    * In the case a queue could not be found return -ENODEV.
>    * Otherwise return 0.
>    */
> -static __attribute__((unused))
> -	int vfio_ap_associate_queues(struct ap_matrix_mdev *matrix_mdev)
> +static int vfio_ap_associate_queues(struct ap_matrix_mdev *matrix_mdev)

Maybe this function should be introduced in this patch instead?

>   {
>   	unsigned long apid, apqi;
>   	struct vfio_ap_queue *q;
> @@ -967,12 +996,32 @@ static int vfio_ap_mdev_open(struct mdev_device *mdev)
>   
>   	ret = vfio_register_notifier(mdev_dev(mdev), VFIO_GROUP_NOTIFY,
>   				     &events, &matrix_mdev->group_notifier);
> -	if (ret) {
> -		module_put(THIS_MODULE);
> -		return ret;
> -	}
> +	if (ret)
> +		goto err_group;
> +
> +	matrix_mdev->iommu_notifier.notifier_call = vfio_ap_mdev_iommu_notifier;
> +	events = VFIO_IOMMU_NOTIFY_DMA_UNMAP;
> +
> +	ret = vfio_register_notifier(mdev_dev(mdev), VFIO_IOMMU_NOTIFY,
> +				     &events, &matrix_mdev->iommu_notifier);
> +	if (ret)
> +		goto err_iommu;
> +
> +	ret = vfio_ap_associate_queues(matrix_mdev);
> +	if (ret)
> +		goto err_associate;

I think the matrix_mdev should be associated with queues when an 
assignment of an adapter or domain is made to the mdev device via its
sysfs interfaces. I say this because assigning an adapter or domain to
an mdev device effectively grants ownership of any additional AP queues 
added to the mdev device's AP matrix as a result of the assignment. It
only makes sense to assign ownership to the vfio_ap_queue objects
representing the queues at that time. If an adapter or domain is
dynamically assigned while a guest is using the affected queues, then
the associations will have to be made at that time and this code will
likely go bye bye.

>   
>   	return 0;
> +
> +err_associate:
> +	vfio_unregister_notifier(mdev_dev(mdev), VFIO_IOMMU_NOTIFY,
> +				 &matrix_mdev->iommu_notifier);
> +err_iommu:
> +	vfio_unregister_notifier(mdev_dev(mdev), VFIO_GROUP_NOTIFY,
> +				 &matrix_mdev->group_notifier);
> +err_group:
> +	module_put(THIS_MODULE);
> +	return ret;
>   }
>   
>   static void vfio_ap_mdev_release(struct mdev_device *mdev)
> @@ -985,6 +1034,9 @@ static void vfio_ap_mdev_release(struct mdev_device *mdev)
>   	vfio_ap_mdev_reset_queues(mdev);
>   	vfio_unregister_notifier(mdev_dev(mdev), VFIO_GROUP_NOTIFY,
>   				 &matrix_mdev->group_notifier);
> +	vfio_unregister_notifier(mdev_dev(mdev), VFIO_IOMMU_NOTIFY,
> +				 &matrix_mdev->iommu_notifier);
> +	vfio_ap_dissociate_queues(matrix_mdev);

I think the matrix_mdev should be dissociated from queues when an
adapter or domain is unassigned from the mdev device via its
sysfs interfaces. I say this because unassigning an adapter or domain 
from an mdev device effectively takes away ownership of any AP queues
removed from the mdev device's AP matrix as a result of the
unassignment. It only makes sense to remove ownership from the
vfio_ap_queue objects representing the queues at that time. This will
become necessary for the forthcoming dynamic configuration patches.
If an adapter or domain is dynamically unassigned while a guest is
using the affected queues, then the dissociation will have to be made
at that time and this code will likely go bye bye.

>   	matrix_mdev->kvm = NULL;
>   	module_put(THIS_MODULE);
>   }
> diff --git a/drivers/s390/crypto/vfio_ap_private.h b/drivers/s390/crypto/vfio_ap_private.h
> index 10bc8b5..2781720 100644
> --- a/drivers/s390/crypto/vfio_ap_private.h
> +++ b/drivers/s390/crypto/vfio_ap_private.h
> @@ -80,7 +80,9 @@ struct ap_matrix_mdev {
>   	struct list_head node;
>   	struct ap_matrix matrix;
>   	struct notifier_block group_notifier;
> +	struct notifier_block iommu_notifier;
>   	struct kvm *kvm;
> +	struct mdev_device *mdev;
>   };
>   
>   extern int vfio_ap_mdev_register(void);
>