Comment 32 for bug 1874056

------- Comment From <email address hidden> 2020-05-20 06:22 EDT-------
(In reply to comment #37)
> @IBM Please can you test the patched kernel from the PPA and leave a quick
> feedback here if it works like expected?
> This would lower the risk a bit and increase the confidence and would at the
> end help regarding the acceptance by the kernel team.

Thanks for the fast turn around, I can confirm that this package from your PPA
works as expected enabling the correct multi-function enumeration.

But let me take this opportunity to talk about a related issue because

a) It is another reason these patches fix real problems and
b) It is needed to make full use of these patches when using SR-IOV
c) I want to give you this as a heads-up to consider during the discussion
of whether you take the patches discussed here.

echo 0 > /sys/bus/pci/devices/<mlx5_pf>/sriov_numvfs

With multiple PCI Virtual Functions (VFs) active, the drivers for the VFs just see the devices gone.
This of course only affects systems with PCI Physical Functions (none
at customers currently) and indeed the same happens without these patches
(just confirmed that again) so this isn't a regression.

The thing is, I have 2 patches queued for upstream that make this VF
shutdown as graceful as on x86_64.

These patches need to go on top of the patches discussed here as they
make use of the RID introduced with these patches. However
they also need a (very small) common code change. So I'm not sure
how we will handle those.