Re: [PATCH] r8169: don't use MSI-X on RTL8106e

On 22.08.2018 13:44, Thomas Gleixner wrote:

On Tue, 21 Aug 2018, Heiner Kallweit wrote:

quoted

On 21.08.2018 21:31, David Miller wrote:

quoted

From: Heiner Kallweit <hkallweit1@gmail.com>
Date: Mon, 20 Aug 2018 22:46:48 +0200

quoted

I'm in contact with Realtek and according to them few chip versions
seem to clear MSI-X table entries on resume from suspend. Checking
with them how this could be fixed / worked around.
Worst case we may have to disable MSI-X in general.

I worry that if the chip does this, and somehow MSI-X is enabled and
an interrupt is generated, the chip will write to the cleared out
MSI-X address.  This will either write garbage into memory or cause
a bus error and require PCI error recovery.

It also looks like your test patch doesn't fix things for people who
have tested it.

The test patch was based on the first info from Realtek which made me
think that the base address of the MSI-X table is cleared, what
obviously is not the case.

After some further tests it seems that the solution isn't as simple
as storing the MSI-X table entries on suspend and restore them on
resume. On my system (where MSI-X works fine) MSI-X table entries
on resume are partially different from the ones on suspend.

Which is not a surprise. Please don't try to fiddle with that at the driver
level. The irq and PCI core code are the ones in charge and if you'd
restore at the wrong point then hell breaks lose.

Instead of spending a lot of effort on a workaround which may not be
acceptable, it may be better to fall back to MSI on all affected chip
versions. For two chip versions which were reported to have this issues
we're doing this already. I asked Realtek whether they have an overview
which chip versions are affected, let's see ..

The Realtek chips provide an alternative, register-based way to access
the MSI-X table, and their Windows driver seems to use it. See here:
https://patchwork.kernel.org/patch/4149171/

But as we handle all MSI-X basics in the PCI core, this isn't an option.

Can you please do the following:

 1) Store the PCI config space at suspend time
 2) Compare the PCI config space at resume time and print the difference

Do that on a working and a non-working version of Realtek NICs.

Thanks,

	tglx