RE: [RFC 02/20] vfio: Add device class for /dev/vfio/devices

[Liu, Yi L]

> From: Jason Gunthorpe <jgg@xxxxxxxxxx>
> Sent: Monday, November 1, 2021 8:50 PM
> 
> On Fri, Oct 29, 2021 at 09:47:27AM +0000, Liu, Yi L wrote:
> > Hi Jason,
> >
> > > From: Jason Gunthorpe <jgg@xxxxxxxxxx>
> > > Sent: Monday, October 25, 2021 8:53 PM
> > >
> > > On Mon, Oct 25, 2021 at 06:28:09AM +0000, Liu, Yi L wrote:
> > > >    thanks for the guiding. will also refer to your vfio_group_cdev series.
> > > >
> > > >    Need to double confirm here. Not quite following on the kfree. Is
> > > >    this kfree to free the vfio_device structure? But now the
> > > >    vfio_device pointer is provided by callers (e.g. vfio-pci). Do
> > > >    you want to let vfio core allocate the vfio_device struct and
> > > >    return the pointer to callers?
> > >
> > > There are several common patterns for this problem, two that would be
> > > suitable:
> > >
> > > - Require each driver to provide a release op inside vfio_device_ops
> > >   that does the kfree. Have the core provide a struct device release
> > >   op that calls this one. Keep the kalloc/kfree in the drivers
> >
> > this way sees to suit the existing vfio registration manner listed
> > below. right?
> 
> Not really, most drivers are just doing kfree. The need for release
> comes if the drivers are doing more stuff.
> 
> > But device drivers needs to do the kfree in the
> > newly added release op instead of doing it on their own (e.g.
> > doing kfree in remove).
> 
> Yes
> 
> > > struct ib_device *_ib_alloc_device(size_t size);
> > > #define ib_alloc_device(drv_struct, member)                                    \
> > >         container_of(_ib_alloc_device(sizeof(struct drv_struct) +              \
> > >                                       BUILD_BUG_ON_ZERO(offsetof(              \
> > >                                               struct drv_struct, member))),    \
> > >                      struct drv_struct, member)
> > >
> >
> > thanks for the example. If this way, still requires driver to provide
> > a release op inside vfio_device_ops. right?
> 
> No, it would optional. It would contain the stuff the driver is doing
> before kfree()
> 
> For instance mdev looks like the only driver that cares:
> 
> 	vfio_uninit_group_dev(&mdev_state->vdev);
> 	kfree(mdev_state->pages);
> 	kfree(mdev_state->vconfig);
> 	kfree(mdev_state);
> 
> pages/vconfig would logically be in a release function

I see. So the criteria is: the pointer fields pointing to a memory buffer
allocated by the device driver should be logically be free in a release
function. right? I can see there are such fields in struct vfio_pci_core_device
and mdev_state (both mbochs and mdpy). So we may go with your option #2.
Is it? otherwise, needs to add release callback for all the related drivers.

struct vfio_pci_core_device {
	struct vifo_device vdev;
...
	u8 *pci_config_map;
	u8 *vconfig;
...
};

struct mdev_state {
	struct vifo_device vdev;
...
	u8 *vconfig;
	struct page **pages;
...
};

> On the other hand ccw needs to rcu free the vfio_device, so that would
> have to be global overhead with this api design.

not quite get. why ccw is special here? could you elaborate?

Thanks,
Yi Liu