Re: [PATCH 2/2] pci: Don't set RCB bit in LNKCTL if the upstream bridge hasn't
From: Don Dutile
Date: Mon Nov 14 2016 - 11:26:16 EST
On 11/14/2016 06:56 AM, Johannes Thumshirn wrote:
On Wed, Nov 09, 2016 at 11:11:40AM -0600, Bjorn Helgaas wrote:
Hi Johannes,
On Wed, Nov 02, 2016 at 04:35:52PM -0600, Johannes Thumshirn wrote:
The Read Completion Boundary (RCB) bit must only be set on a device or
endpoint if it is set on the root complex.
Certain BIOSes erroneously set the RCB Bit in their ACPI _HPX Tables
even if it is not set on the root port. This is a violation to the PCIe
Specification and is known to bring some Mellanox Connect-X 3 HCAs into
a state where they can't map their firmware and go into error recovery.
BIOS Information
Vendor: IBM
Version: -[A8E120CUS-1.30]-
Release Date: 08/22/2016
This seems like a pretty serious problem (sounds like maybe the HCA is
completely useless?)
Correct.
Can you point us at a bugzilla or other problem report? It's nice to
have details of what this looks like to a user, so people who trip
over this problem have a little more chance of finding the solution.
As we already said, our bugzilla entry for this is not accessible from the
outside, but I know Red Hat does have a bugzilla entry for the same issue as
well. Maybe this is reachable from the outside (adding Don for this, as I know
he has worked on this problem as well).
RHEL bz's are not accessible from the outside.
I suggest capturing the content of the RH bz issue and creating a k.o. bz
with the information.
7a1562d4f2d0 ("PCI: Apply _HPX Link Control settings to all devices
with a link") appeared in v3.18, so it's probably not a *new* problem,
so my guess is that this is v4.10 material.
Yes 4.10 sounds good to me. I personally think, this problem hasn't
materialized yet, as this is the kind of hardware you run on a rather /stable/
kernel either you built on your own or get from an enterprise distribution and
until recently these kernels haven't been updated to something newer than
3.18.
Thanks,
Johannes