[PATCH v2 06/22] vfio/pci: Retrieve preserved device files after Live Update

[David Matlack]

On 2026-02-26 03:52 PM, Alex Williamson wrote:
> On Thu, 29 Jan 2026 21:24:53 +0000 David Matlack <dmatlack at google.com> wrote:

> > diff --git a/drivers/vfio/device_cdev.c b/drivers/vfio/device_cdev.c
> > index 8ceca24ac136..935f84a35875 100644
> > --- a/drivers/vfio/device_cdev.c
> > +++ b/drivers/vfio/device_cdev.c
> > @@ -52,6 +46,19 @@ int vfio_device_fops_cdev_open(struct inode *inode, struct file *filep)
> >  	vfio_device_put_registration(device);
> >  	return ret;
> >  }
> > +EXPORT_SYMBOL_GPL(__vfio_device_fops_cdev_open);
> 
> I really dislike that we're exporting the underscore variant, which
> implies it's an internal function that the caller should understand the
> constraints, without outlining any constraints.
> 
> I'm not sure what a good alternative is.  We can drop fops since this
> isn't called from file_operations.  Maybe vfio_device_cdev_open_file().

Ack. Due to the bug you pointed out below, I think the changes in this
file will look fairly different in the next version. But no matter what
I'll avoid exporting a underscore variant without outlining the
constraints.

> > +	/*
> > +	 * Simulate opening the character device using an anonymous inode. The
> > +	 * returned file has the same properties as a cdev file (e.g. operations
> > +	 * are blocked until BIND_IOMMUFD is called).
> > +	 */
> > +	file = anon_inode_getfile_fmode("[vfio-device-liveupdate]",
> > +					&vfio_device_fops, NULL,
> > +					O_RDWR, FMODE_PREAD | FMODE_PWRITE);
> > +	if (IS_ERR(file)) {
> > +		ret = PTR_ERR(file);
> > +		goto out;
> > +	}
> > +
> > +	ret = __vfio_device_fops_cdev_open(device, file);
> > +	if (ret) {
> > +		fput(file);
> 
> Don't we end up calling vfio_device_fops.release with NULL
> file->private_data here with inevitable segfaults?  Thanks,

Yes indeed. In that case I think we need to call
vfio_device_try_get_registration() and vfio_allocate_device_file()
before anon_inode_getfile_fmode(). We could play games with file->fops
to avoid it calling vfio_device_fops.release here instead, but that
feels hacky.