Re: [PATCH] PCI: Mark Nvidia GB10 to avoid bus reset

[Bjorn Helgaas <helgaas@kernel.org>]

On Thu, Jan 15, 2026 at 12:11:09PM -0800, Terje Bergstrom wrote:
> On 1/14/26 09:28, Bjorn Helgaas wrote:
> 
> > What sort of crash happens?  It's useful if we can include a bread
> > crumb that will help people identify the crash and find a fix.
>
> We observed retraining to lower PCIe lane count and config read
> timeout.  So yes crash is not the best way to describe it.
> 
> > I'm confused about what the topology is.  I first assumed GB10 was
> > a PCIe Endpoint, since Secondary Bus Reset only applies to devices
> > below a bridge, so SBR would be applied to a device by a config
> > write to that bridge.
>
> gb10 is an SoC designed by NVIDIA and Mediatek in collaboration.
> It's not an endpoint, but has its own PCIe controller for connecting
> PCIe peripherals like NVMe drives, NIC, etc.

OK, so you do SBR to some endpoint below a GB10 Root Port, and after
the SBR, the link to the endpoint retrains with a lower lane count and
config reads to the endpoint time out?

I see you're from NVIDIA, so if you're confirming that this is a
hardware erratum (not an issue with the GB10 PCI controller driver),
we should definitely apply this, and I'll wordsmith the commit log and
comment something like this:

  When asserting Secondary Bus Reset to downstream devices via a GB10
  Root Port, the link doesn't retrain correctly.  The link may retrain
  with a lower lane count, and config accesses to downstream devices
  may fail.

Bjorn