On Fri, Nov 01, 2024 at 04:18:55AM +0200, Wei Fang wrote:

quoted

On Sat, Oct 26, 2024 at 06:01:37AM +0300, Wei Fang wrote:

quoted

system-controller not only configure the endpoints of the NETC, but also
can configure the ECAM space, such as the vendor ID, device ID, the RID
of endpoint, VF stride and so on. For this perspective, I don't think the
ECAM space should placed at the same hierarchical level with system-controller.

If they are placed at the same level, then before pci_host_common_probe() is
called, we need to ensure that IERB completes probe(), which means we need
to modify the PCI host common driver, component API or add a callback function
or something else, which I don't think is a good idea.

Ok, that does sound important. If the NETCMIX block were to actually
modify the ECAM space, what would be the primary source of information
for how the ECAM device descriptions should look like?

I think the related info should be provided by DTS, but currently, we do not
have such requirement that needs Linux to change the ECAM space, this may
be supported in the future if we have the requirement.

quoted

I remember a use case being discussed internally a while ago was that
where the Cortex-A cores are only guests which only have ownership of
some Ethernet ports discovered through the ECAM, but not of the entire
NETCMIX block and not of physical Ethernet ports. How would that be
described in the device tree? The ECAM node would no longer be placed
under system-controller?

Yes, we indeed have this use case on i.MX95, only the VFs of 10G ENETC
are owned by Cortex-A, the entire ECAM space and other NETC devices
are all owned by Cortex-M. In this case, the system-controller is no needed
in DTS, because Linux have no permission to access these resources.

quoted

At what point does it simply just make more sense to have a different
PCIe ECAM driver than pcie-host-ecam-generic, which just handles
internally the entire NETCMIX?

Currently, I have not idea in what use case we need a different ECAM driver
to handle internally the entire system-controller.

For the use case I mentioned above, we use a different ECAM driver, which
is implemented by RPMSG, because the entire ECAM space is owned by
Cortex-M. So we use the ECAM driver to notify the Cortex-M to enable/disable
VFs or do FLR for VFs and so on. But this ECAM driver does not need to
configure the system-controller.

Ok, I was actually wondering if it makes sense for the the parent bus of
the NETC PCIe functions to be described through a unified binding that
covers all of the above use cases, so that major device tree modifications
aren't necessary to adapt between the 'Linux as host' and 'Linux as guest'
use cases. But you're saying it doesn't make much sense, because the
device tree in the guest case would contain descriptions of inaccessible
resources (the NETCMIX block). Oh well, this is just another case where
"device tree should describe hardware" actually means "device tree describes
what software wants to know about the hardware".

Anyway, I am now convinced by your design choices at least to the extent
that they appear self-consistent to me (I still don't really have an
independent opinion). If somebody has a different idea on how the PCIe
bus should be described, feel free to chime in.