Re: [PATCH v2 02/13] vfio/mdev: Allow the mdev_parent_ops to specify the device driver to bind

[Leon Romanovsky <leon@kernel.org>]

On Wed, Apr 28, 2021 at 11:14:46AM -0300, Jason Gunthorpe wrote:
> On Wed, Apr 28, 2021 at 09:44:07AM +0300, Leon Romanovsky wrote:
> > On Mon, Apr 26, 2021 at 05:00:04PM -0300, Jason Gunthorpe wrote:
> > > This allows a mdev driver to opt out of using vfio_mdev.c, instead the
> > > driver will provide a 'struct mdev_driver' and register directly with the
> > > driver core.
> > > 
> > > Much of mdev_parent_ops becomes unused in this mode:
> > > - create()/remove() are done via the mdev_driver probe()/remove()
> > > - mdev_attr_groups becomes mdev_driver driver.dev_groups
> > > - Wrapper function callbacks are replaced with the same ones from
> > >   struct vfio_device_ops
> > > 
> > > Following patches convert all the drivers.
> > > 
> > > Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
> > >  drivers/vfio/mdev/mdev_core.c   | 64 ++++++++++++++++++++++++++++-----
> > >  drivers/vfio/mdev/mdev_driver.c | 17 ++++++++-
> > >  include/linux/mdev.h            |  3 ++
> > >  3 files changed, 75 insertions(+), 9 deletions(-)
> > 
> > <...>
> > 
> > > +/*
> > > + * mdev drivers can refuse to bind during probe(), in this case we want to fail
> > > + * the creation of the mdev all the way back to sysfs. This is a weird model
> > > + * that doesn't fit in the driver core well, nor does it seem to appear any
> > > + * place else in the kernel, so use a simple hack.
> > > + */
> > > +static int mdev_bind_driver(struct mdev_device *mdev)
> > > +{
> > > +	struct mdev_driver *drv = mdev->type->parent->ops->device_driver;
> > > +	int ret;
> > > +
> > > +	if (!drv)
> > > +		drv = &vfio_mdev_driver;
> > > +
> > > +	while (1) {
> > > +		device_lock(&mdev->dev);
> > > +		if (mdev->dev.driver == &drv->driver) {
> > > +			ret = 0;
> > > +			goto out_unlock;
> > > +		}
> > > +		if (mdev->probe_err) {
> > > +			ret = mdev->probe_err;
> > > +			goto out_unlock;
> > > +		}
> > > +		device_unlock(&mdev->dev);
> > > +		ret = device_attach(&mdev->dev);
> > 
> > The sequence above looks sketchy:
> > 1. lock
> > 2. check for driver
> > 3. unlock
> > 4. device_attach - it takes internally same lock as in step 1.
> > 
> > Why don't you rely on internal to device_attach() driver check?
> 
> This is locking both probe_err and the check that the right driver is
> bound. device_attach() doesn't tell you the same information

device_attach() returns you the information that driver is already
bound, which is the same as you are doing here, because you don't
unbind "the wrong driver".

Thanks

> 
> Jason