Re: pcie_bandwidth_available and USB4/TBT3

[Mika Westerberg <mika.westerberg@linux.intel.com>]

Hi Mario,

On Mon, Oct 30, 2023 at 11:46:20AM -0500, Mario Limonciello wrote:
> Hi,
> 
> Recently we’ve been looking at some issues with AMD dGPUs being put into a
> TBT3 eGPU enclosure and various issues that come up.  Several of them are
> root caused to bugs in the amdgpu driver that we’ll fix there.
> 
> However one thing stands out is a performance problem where the cards are
> artificially limited to a lower speed than necessary.
> 
> The amdgpu driver uses pcie_bandwidth_available() to decide what values to
> use for the platform speed cap and bandwidth cap.
> The value returned for the platform speed cap is always hardcoded to 2.5
> GT/s.
> 
> This happens because the USB4 spec explicitly states[1]
> 
> ---
> 11.2.1 PCIe Physical Layer Logical Sub-block
> The Logical sub-block shall update the PCIe configuration registers with the
> following
> characteristics:
> • PCIe Gen 1 protocol behavior.
> • Max Link Speed field in the Link Capabilities Register set to 0001b (data
> rate of 2.5 GT/s
> only).
> Note: These settings do not represent actual throughput. Throughput is
> implementation specific
> and based on the USB4 Fabric performance.
> ---
> 
> So I wanted to ask – is it better to:
> 1. Catch this case in pcie_bandwidth_available() to skip PCIe root ports
> associated with a USB4 controller.
> 
> 2. Special case the usage of pcie_bandwidth_available() to ignore any
> limiting devices when dev_is_removable() for the dGPU.
> 
> I'm personally tending to think it's better to fix in
> pcie_bandwidth_available() because papering over it in amdgpu means that the
> discovering the upper bound isn't possible if you must ignore the return
> value for pcie_bandwidth_available().

I agree, handling this in pcie_bandwidth_available() makes sense but I
suggest also to document this the kernel-doc so that the GPU driver
writers (probably the only ones using this function) can find the
information easily. Maybe the simplest is to skip any "tunneled" links
and stick to the real PCIe links when calculating the bandwidth.