On Tue, Aug 06, 2024 at 05:04:06PM -0500, Bjorn Helgaas wrote:
> > +static bool pcie_is_tunneled(struct pci_dev *pdev)
> > +{
> > + struct pci_dev *parent, *root;
> > +
> > + parent = pci_upstream_bridge(pdev);
> > + /* If pdev doesn't have a parent, then there's no way it is tunneled.*/
> > + if (!parent)
> > + return false;
> > +
> > + root = pcie_find_root_port(pdev);
> > + /* If pdev doesn't have a root, then there's no way it is tunneled.*/
> > + if (!root)
> > + return false;
> > +
> > + /* Internal PCIe devices are not tunneled. */
> > + if (!root->external_facing)
> > + return false;
> > +
> > + /* Anything directly behind a "usb4-host-interface" is tunneled. */
> > + if (pcie_has_usb4_host_interface(parent))
> > + return true;
> > +
> > + /*
> > + * Check if this is a discrete Thunderbolt/USB4 controller that is
> > + * directly behind the non-USB4 PCIe Root Port marked as
> > + * "ExternalFacingPort". These PCIe devices are used to add Thunderbolt
> > + * capabilities to CPUs that lack integrated Thunderbolt.
> > + * These are not behind a PCIe tunnel.
>
> I need more context to be convinced that this is a reliable heuristic.
> What keeps somebody from plugging a discrete Thunderbolt/USB4
> controller into an external port?
In case of integrated host controller the above function returns true for
anything connected behind its tunneled PCIe root ports so if you plug in
a discrete controller into that it will be treated as tunneled and gets
full IOMMU mappings.
In case of discrete host controller there are two ways. With USB4 the above
function finds "usb4-host-interface" firmware description and returns
true the discrete controller you plug in will be treated as tunneled and
gets full IOMMU mappings.
With non-USB4 (that's Thunderbolt 3 and below) discrete host controller we
have root->external_facing true (it is marked with "ExternalFacingPort"
firmware property) the check below fails because the discrete controller
you plug in is not directly under that so it will be treated as tunneled
and gets full IOMMU mappings.
> Maybe this just needs a sentence or
> two from Lukas's (?) helpful intro to tunneling?
>
> > + if (pcie_switch_directly_under(root, pdev))
> > + return false;
> > +
> > + /* PCIe devices after the discrete chip are tunneled. */
> > + return true;
> > +}
