PCIe errata and GPU patches

There have been patches floating around for a while now to ensure that GPUs work; the current approach, AFAIK, is not upstreamable and was rejected. When rebasing them onto Linux 7.1.9, I asked Claude to refactor to make something upstreamable. I promptly decided to update my Ampere box to Fedora 45, which has a 7.2 kernel. So I fixed a couple of minor issues to get them to apply there. I have tested on my ADLink box, and they seem okay with my minimal use. I would appreciate some wider testing before trying to take them upstream and possibly someone smarter than me to fix any issues with them. linux-ampere-altra-erratum-pcie-65/v7.2 at main · ausil/linux-ampere-altra-erratum-pcie-65 · GitHub . The two patches live in that directory of my fork

1 Like

Why do you think this will be accepted upstream where the old patches wouldn’t? This looks like the same approach to me.

1 Like

GitHub - ausil/linux-ampere-altra-erratum-pcie-65: Kernel patches with workaround for Altra erratum PCIE_65 · GitHub is a fork,

GitHub - AmpereComputing/linux-ampere-altra-erratum-pcie-65: Kernel patches with workaround for Altra erratum PCIE_65 · GitHub is the official upstream repo by Ampere Computing.

Is Ampere’s repo officially dead and is ausil’s repo its (de facto) official successor?

1 Like

Chimera Linux has been maintaining the patch with up to date linux kernels for a long time. You could just pull the patch from their git. I don’t see what passing the patch through AI is going to do to make it upstreamable. It’s the same workaround approach as far as I can tell; just because you told the AI to make it upstreamable doesn’t make it so.

Please offer correction is I’m missing something.

1 Like

I am not sure they will be accepted. the biggest difference is that Affected apertures are learned per root complex from the matching PCI host bridge’s own resource windows (drivers/pci/quirks.c’s early fixup on the RC’s PCI ID calls into arch/arm64/kernel/altra_82288.c), rather than assumed from fixed physical-address ranges, since aperture placement is board/firmware dependent and Altra platforms may expose several independently-windowed root complexes.

Most of that came from the commit message. and it makes sense to me why it may be accepted, but this is not my strongest suite. I wanted to kick off a discussion because carrying patches and maintaining a kernel fork for it is really not ideal, but it is the only way I can use the system as I want to. This bug is a major hindrance to using the hardware.

1 Like

My fork is definelty not a sucessor, it was something I threw together to facilitate a discussion and hopefully get people smarter than me to talk and actually come up with a solution

1 Like