Re: Xen on Zen 3

[Jan Beulich <jbeulich@xxxxxxxx>]

On 07.04.2026 10:08, Matyáš Kroupa wrote:
> Hi,
> I am trying to run Xen on a bare metal with Zen 3 microarchitecture (to be 
> specific Ryzen 9 5950X). I am using Fedora 43 as Dom0 in PVH mode. However, 
> the 
> Linux kernel crashes early in the boot causing Xen to reboot. I am not 
> exactly 
> sure if Linux or Xen is the problematic part. The kernel boots fine without 
> Xen. Before Ryzen 9 I had Ryzen 5 3400G (Zen+) which booted even with Xen.
> 
> The Xen is a backported from Fedora 44 (not yet released) because I need Xen 
> 4.21 features, mainly ReBAR support for the Arc B50 GPU. Package 
> specification 
> with build commands can be found on https://src.fedoraproject.org/rpms/xen/
> blob/f44/f/xen.spec. There are also downstream patches and configuration.
> 
> I have set up a serial link, so I have some logs. They are in the attachment. 
> I have used gdb to translate the binary offset to source code line. The 
> problematic line is https://elixir.bootlin.com/linux/v6.19.10/source/arch/x86/
> kernel/amd_node.c#L296. It seems that linux is detecting zero PCI roots on 
> non-zero number of nodes causing division by zero.

Not exactly, there is an earlier exit from the function when num_roots is 0.
If that line is the problem one, then presumably num_roots < num_nodes, thus
yielding roots_per_node as 0. Sadly you didn't enable enough verbosity for

        pr_debug("Found %d AMD root devices\n", num_roots);

to actually leave a trace in the log. I'd guess the value to be 1, but there
being multiple nodes at the same time. You may want to instrument the
function a little more to be certain.

Seeing

static inline u16 amd_num_nodes(void)
{
        return topology_amd_nodes_per_pkg() * topology_max_packages();
}

I'm not overly surprised of problems, though. Our topology representation
still has a lot of work to do before the data can be reliably used.

For your immediate purpose you may want to change the "!num_roots" check
into a "num_roots < num_nodes" one (on the assumption that num_nodes
can't be 0). Whether that's acceptable upstream I don't know, of course.

Jan