Re: [RFC 11/12] vfio: Add ioctls for device cdev iommufd

[Yi Liu <yi.l.liu@intel.com>]

On 2023/1/9 22:14, Jason Gunthorpe wrote:
> On Mon, Jan 09, 2023 at 07:47:03AM +0000, Tian, Kevin wrote:
>>> From: Liu, Yi L <yi.l.liu@intel.com>
>>> Sent: Monday, December 19, 2022 4:47 PM
>>>
>>> @@ -415,7 +416,7 @@ static int vfio_device_first_open(struct
>>> vfio_device_file *df,
>>>   	if (!try_module_get(device->dev->driver->owner))
>>>   		return -ENODEV;
>>>
>>> -	if (iommufd)
>>> +	if (iommufd && !IS_ERR(iommufd))
>>>   		ret = vfio_iommufd_bind(device, iommufd, dev_id, pt_id);
>>>   	else
>>>   		ret = vfio_device_group_use_iommu(device);
>>
>> can you elaborate how noiommu actually works in the cdev path?
> 
> I still need someone to test the noiommu iommufd patch with dpdk, I'll
> post it in a minute

I remembered I had mentioned I found one guy to help. But he mentioned
to me he has got some trouble. I think it may be due to environment. Let
me check with him tomorrow about if any update.

> For cdev conversion no-iommu should be triggered by passing in -1 for
> the iommufd file descriptor to indicate there is no translation. Then
> the module parameter and security checks should be done before
> allowing the open to succeed with an identity translation in place.
> 
> There should be a check that there is no iommu driver controlling the
> device as well..

yes. I used ERR_PTR(-EINVAL) as an indicator of noiommu in this patch.
I admit this logic is incorrect. Should be

	if (iommufd) {
		if (IS_ERR(iommufd))
			ret = 0;
   		else
			ret = vfio_iommufd_bind(device, iommufd, dev_id, pt_id);
    	} else {
   		ret = vfio_device_group_use_iommu(device);
	}

-- 
Regards,
Yi Liu