Re: [Intel-gfx] [PATCH v2 1/1] vfio: remove VFIO_GROUP_NOTIFY_SET_KVM

Re: [Intel-gfx] [PATCH v2 1/1] vfio: remove VFIO_GROUP_NOTIFY_SET_KVM

[Tian, Kevin]

> From: Matthew Rosato
> Sent: Thursday, May 19, 2022 5:26 AM
> 
> Rather than relying on a notifier for associating the KVM with
> the group, let's assume that the association has already been
> made prior to device_open.  The first time a device is opened
> associate the group KVM with the device.
> 
> This fixes a user-triggerable oops in GVT.

No changelog.

> 
> Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
> Signed-off-by: Matthew Rosato <mjrosato@linux.ibm.com>

Not sure whether Christoph wants a s-o-b here when he wrote
the snippet to remove the release work of gvt...

> @@ -1083,11 +1083,22 @@ static struct file *vfio_device_open(struct
> vfio_device *device)
> 
>  	mutex_lock(&device->dev_set->lock);
>  	device->open_count++;
> +	down_read(&device->group->group_rwsem);
> +	if (device->open_count == 1 && device->group->kvm) {
> +		/*
> +		 * Here we pass the KVM pointer with the group under the
> read
> +		 * lock.  If the device driver will use it, it must obtain a
> +		 * reference and release it during close_device.
> +		 */
> +		device->kvm = device->group->kvm;
> +	}
> +
>  	if (device->open_count == 1 && device->ops->open_device) {

Merge the two branches so both are under if (device->open_count == 1) {}
(and group_rwsem can be also moved inside)

> @@ -1315,9 +1330,13 @@ static int vfio_device_fops_release(struct inode
> *inode, struct file *filep)
> 
>  	mutex_lock(&device->dev_set->lock);
>  	vfio_assert_device_open(device);
> +	down_read(&device->group->group_rwsem);
>  	if (device->open_count == 1 && device->ops->close_device)
>  		device->ops->close_device(device);
>  	device->open_count--;
> +	if (device->open_count == 0 && device->kvm)
> +		device->kvm = NULL;

This can be moved out of group_rwsem as there is no reference to
vfio_group.

> +	up_read(&device->group->group_rwsem);

otherwise,

Reviewed-by: Kevin Tian <kevin.tian@intel.com>

Thanks
Kevin

Re: [Intel-gfx] [PATCH v2 1/1] vfio: remove VFIO_GROUP_NOTIFY_SET_KVM

[hch]

On Thu, May 19, 2022 at 06:43:06AM +0000, Tian, Kevin wrote:
> > This fixes a user-triggerable oops in GVT.
> 
> No changelog.

??

the cover latter clearly states what has changed since v1, and this
patch has a good commit log.  This is exactly how it is supposed to
be done.

> 
> > 
> > Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
> > Signed-off-by: Matthew Rosato <mjrosato@linux.ibm.com>
> 
> Not sure whether Christoph wants a s-o-b here when he wrote
> the snippet to remove the release work of gvt...

That's just tivial code removal, so no.

> 
> > @@ -1083,11 +1083,22 @@ static struct file *vfio_device_open(struct
> > vfio_device *device)
> > 
> >  	mutex_lock(&device->dev_set->lock);
> >  	device->open_count++;
> > +	down_read(&device->group->group_rwsem);
> > +	if (device->open_count == 1 && device->group->kvm) {
> > +		/*
> > +		 * Here we pass the KVM pointer with the group under the
> > read
> > +		 * lock.  If the device driver will use it, it must obtain a
> > +		 * reference and release it during close_device.
> > +		 */
> > +		device->kvm = device->group->kvm;
> > +	}
> > +
> >  	if (device->open_count == 1 && device->ops->open_device) {
> 
> Merge the two branches so both are under if (device->open_count == 1) {}
> (and group_rwsem can be also moved inside)

Yeah.  And we don't really need the device->group->kvm check, as
it would otherwise assign NULL which is perfectly fine.

But otherwise this also looks good to me:

Reviewed-by: Christoph Hellwig <hch@lst.de>

Re: [Intel-gfx] [PATCH v2 1/1] vfio: remove VFIO_GROUP_NOTIFY_SET_KVM

[Tian, Kevin]

> From: hch@infradead.org <hch@infradead.org>
> Sent: Thursday, May 19, 2022 2:48 PM
> 
> On Thu, May 19, 2022 at 06:43:06AM +0000, Tian, Kevin wrote:
> > > This fixes a user-triggerable oops in GVT.
> >
> > No changelog.
> 
> ??
> 
> the cover latter clearly states what has changed since v1, and this
> patch has a good commit log.  This is exactly how it is supposed to
> be done.

sigh... don't know why I missed the coverletter.

Re: [Intel-gfx] [PATCH v2 1/1] vfio: remove VFIO_GROUP_NOTIFY_SET_KVM

[Jason Gunthorpe]

On Wed, May 18, 2022 at 05:26:07PM -0400, Matthew Rosato wrote:
> Rather than relying on a notifier for associating the KVM with
> the group, let's assume that the association has already been
> made prior to device_open.  The first time a device is opened
> associate the group KVM with the device.
> 
> This fixes a user-triggerable oops in GVT.
> 
> Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
> Signed-off-by: Matthew Rosato <mjrosato@linux.ibm.com>
> ---
>  drivers/gpu/drm/i915/gvt/gtt.c        |  4 +-
>  drivers/gpu/drm/i915/gvt/gvt.h        |  3 -
>  drivers/gpu/drm/i915/gvt/kvmgt.c      | 82 ++++++---------------------
>  drivers/s390/crypto/vfio_ap_ops.c     | 38 ++++---------
>  drivers/s390/crypto/vfio_ap_private.h |  3 -
>  drivers/vfio/vfio.c                   | 75 ++++++++----------------
>  include/linux/vfio.h                  |  5 +-
>  7 files changed, 56 insertions(+), 154 deletions(-)

I'm fine with this, thanks for finishing it

> diff --git a/drivers/s390/crypto/vfio_ap_ops.c b/drivers/s390/crypto/vfio_ap_ops.c
> index e8914024f5b1..17a56bb4cf25 100644
> --- a/drivers/s390/crypto/vfio_ap_ops.c
> +++ b/drivers/s390/crypto/vfio_ap_ops.c
> @@ -1196,6 +1196,8 @@ static int vfio_ap_mdev_set_kvm(struct ap_matrix_mdev *matrix_mdev,
>  {
>         struct ap_matrix_mdev *m;
>
> +       kvm_get_kvm(kvm);
> +

[..]

> +       if (kvm)
> +               kvm_put_kvm(kvm);
>  }

This extra ref traffic is not necessary, the kvm will have a valid ref
because we are under the group_rwsem here and then the function will
obtain a ref when it stores it in matrix_mdev->kvm.

> diff --git a/drivers/vfio/vfio.c b/drivers/vfio/vfio.c
> index cfcff7764403..272acb62facd 100644
> --- a/drivers/vfio/vfio.c
> +++ b/drivers/vfio/vfio.c
> @@ -1083,11 +1083,22 @@ static struct file *vfio_device_open(struct vfio_device *device)
>  
>  	mutex_lock(&device->dev_set->lock);
>  	device->open_count++;
> +	down_read(&device->group->group_rwsem);
> +	if (device->open_count == 1 && device->group->kvm) {
> +		/*
> +		 * Here we pass the KVM pointer with the group under the read
> +		 * lock.  If the device driver will use it, it must obtain a
> +		 * reference and release it during close_device.
> +		 */
> +		device->kvm = device->group->kvm;
> +	}

But it is a bit ugly to keep the device->kvm in the struct device
without holding a reference count.

It is probably worth adding a comment to the definition in the struct
as well that the device driver must reference it or never touch it.

Jason