On 2024/7/10 1:36, Jason Gunthorpe wrote:

On Mon, Jul 01, 2024 at 01:55:12PM +0800, Baolu Lu wrote:

quoted

On 2024/6/29 5:17, Jason Gunthorpe wrote:

quoted

On Sun, Jun 16, 2024 at 02:11:52PM +0800, Lu Baolu wrote:

quoted

+static int iommufd_fault_iopf_enable(struct iommufd_device *idev)
+{
+	struct device *dev = idev->dev;
+	int ret;
+
+	/*
+	 * Once we turn on PCI/PRI support for VF, the response failure code
+	 * should not be forwarded to the hardware due to PRI being a shared
+	 * resource between PF and VFs. There is no coordination for this
+	 * shared capability. This waits for a vPRI reset to recover.
+	 */
+	if (dev_is_pci(dev) && to_pci_dev(dev)->is_virtfn)
+		return -EINVAL;

I don't quite get this remark, isn't not supporting PRI on VFs kind of
useless? What is the story here?

This remark is trying to explain why attaching an iopf-capable hwpt to a
VF is not supported for now. The PCI sepc (section 10.4.2.1) states that
a response failure will disable the PRI on the function. But for PF/VF
case, the PRI is a shared resource, therefore a response failure on a VF
might cause iopf on other VFs to malfunction. So, we start from simple
by not allowing it.

You are talking about IOMMU_PAGE_RESP_FAILURE ?

But this is bad already, something like SVA could trigger
IOMMU_PAGE_RESP_FAILURE on a VF without iommufd today. Due to memory
allocation failure in iommu_report_device_fault()

And then we pass in code from userspace and blindly cast it to
enum iommu_page_response_code ?

Probably we should just only support IOMMU_PAGE_RESP_SUCCESS/INVALID
from userspace and block FAILURE entirely. Probably the VMM should
emulate FAILURE by disabling PRI on by changing to a non PRI domain.

And this subtle uABI leak needs a fix:

		iopf_group_response(group, response.code);

response.code and enum iommu_page_response_code are different
enums, and there is no range check. Need a static assert at least and
a range check. Send a followup patch please

Yes, sure.

Thanks,
baolu