On Fri, Jun 18, 2021 at 10:35:21AM -0400, Jason J. Herne wrote:

On 6/18/21 10:10 AM, Jason Gunthorpe wrote:

quoted

On Fri, Jun 18, 2021 at 09:35:24AM -0400, Jason J. Herne wrote:

quoted

vfio_ap_matrix_dev_release is shadowing the global matrix_dev with driver
data that never gets set. So when release is called we end up not freeing
matrix_dev. The fix is to remove the shadow variable and just free the
global.

I would clarify this commit message to say that the drv_data of the
matrix_dev is never set and so dev_get_drvdata() always returns NULL

And I would suggest to use

   container_of(dev, struct ap_matrix_dev, dev)

instead of the global variable, and probably NULL the global
too..

The use of driver_data seems to have been completely erroneous here. In this
case the global matrix_dev is the top level struct. It is not contained in
anything. matrix_dev is created upon module load, and it is freed when the
module exits.

So I don't think using container_of makes sense. Unless I've
misunderstood your suggestion?

	matrix_dev = kzalloc(sizeof(*matrix_dev), GFP_KERNEL);
	matrix_dev->device.release = vfio_ap_matrix_dev_release;
	ret = device_register(&matrix_dev->device);

"dev" is contained inside matrix_dev which is why we should use
container of to go from a struct device pointer back to the containing
matrix_dev pointer

Jason