Re: [PATCH v1 1/1] PCI/AER: Use _OSC negotiation to determine AER ownership

[Bjorn Helgaas]

On Wed, Apr 29, 2020 at 03:24:41PM +0000, Austin.Bolen@xxxxxxxx wrote:
> On 4/28/2020 3:37 PM, Bjorn Helgaas wrote:
> > [EXTERNAL EMAIL] 
> >
> > [+to Mario, Austin, Rafael; Dell folks, I suspect this commit will
> > break Dell servers but I'd like your opinion]
> >
> > <snip>
> Thanks Bjorn, for the heads up. I checked with our server BIOS team and
> they say that only checking _OSC for AER should work on our servers.  We
> always configure_OSC and the HEST FIRMWARE_FIRST flag to retain firmware
> control of AER so either could be checked.
> 
> > I *really* want the patch because the current mix of using both _OSC
> > and FIRMWARE_FIRST to determine AER capability ownership is a mess and
> > getting worse, but I'm more and more doubtful.
> >
> > My contention is that if firmware doesn't want the OS to use the AER
> > capability it should simply decline to grant control via _OSC.
>
> I agree per spec that _OSC should be used and this was confirmed by the
> ACPI working group. Alex had submitted a patch for us [2] to switch to
> using _OSC to determine AER ownership following the decision in the ACPI
> working group.

Perfect, thank you!  I had forgotten that Alex posted that.  We should
add credit to him and a link to that discussion.  Thanks again!

> [2] https://lkml.org/lkml/2018/11/16/202
> 
> > But things like 0584396157ad ("PCI: PCIe AER: honor ACPI HEST FIRMWARE
> > FIRST mode") [1] suggest that some machines grant AER control to the
> > OS via _OSC, but still expect the OS to leave AER alone for certain
> > devices.
>
> AFAIK, no Dell server, including the 11G servers mentioned in that
> patch, have granted control of AER via _OSC and set HEST FIRMWARE_FIRST
> for some devices. I don't think this model is even support by the
> ACPI/PCIe standards.  Yes, you can set the bits that way, but there is
> no text I've found that says how the OS/firmware should behave in that
> scenario. In order to be interoperable, I think someone would need to
> standardized how the OS/firmware would could co-ordinate in such a model.

I agree and I want to get Linux out of the current muddle where we
try to make sense out of it.

> > I think the FIRMWARE_FIRST language in the ACPI spec is really too
> > vague to tell the OS not to use the AER Capability, but it seems like
> > that's what commits like [1] rely on.
> >
> > The current _OSC definition (PCI Firmware r3.2) applies only to
> > PNP0A03/PNP0A08 devices, but it's conceivable that it could be
> > extended to other devices if we need per-device AER Capability
> > ownership.
> >
> > [1] https://git.kernel.org/linus/0584396157ad
<snip>