Re: [RFC PATCH v1 06/38] iommufd: Add and option to request for bar mapping with IORESOURCE_EXCLUSIVE

[Xu Yilun]

> diff --git a/drivers/vfio/pci/vfio_pci_core.c b/drivers/vfio/pci/vfio_pci_core.c
> index 6328c3a05bcd..bee3cf3226e9 100644
> --- a/drivers/vfio/pci/vfio_pci_core.c
> +++ b/drivers/vfio/pci/vfio_pci_core.c
> @@ -1753,8 +1753,15 @@ int vfio_pci_core_mmap(struct vfio_device *core_vdev, struct vm_area_struct *vma
>  	 * we need to request the region and the barmap tracks that.
>  	 */
>  	if (!vdev->barmap[index]) {
> -		ret = pci_request_selected_regions(pdev,
> +
> +		if (core_vdev->iommufd_device &&
> +		    iommufd_device_need_exclusive_range(core_vdev->iommufd_device))
> +			ret = pci_request_selected_regions_exclusive(pdev,
> +							1 << index, "vfio-pci");
> +		else
> +			ret = pci_request_selected_regions(pdev,
>  						   1 << index, "vfio-pci");
> +
>  		if (ret)
>  			return ret;

I did't get the idea.

The purpose of my original patch [1] is not to make VFIO choose between
pci_request_regions_exclusive() or pci_request_regions(). It is mainly
to prevent userspace mmap/read/write against a vfio_cdev FD. For
example:

  If pci_request_selected_regions() is succesfully executed on mmap(),
  later TSM Bind would fail on its pci_request_regions_exclusive(). It
  means userspace should not mmap otherwise you can't do private
  assignment. Vice versa, if you've done TSM Bind, you cannot mmap
  anymore.

The _exclusive is just a bonus that further prevents "/dev/mem and the
sysfs MMIO access"

[1]: https://lore.kernel.org/all/20250529053513.1592088-20-yilun.xu@xxxxxxxxxxxxxxx/

Thanks,
Yilun