Re: [REGRESSION]: mlx5_core does not autoload in Linux 6.18

Re: [REGRESSION]: mlx5_core does not autoload in Linux 6.18

[Thorsten Leemhuis]

Lo! Thx for the report.

Let me CC the maintainers of said driver, maybe they have an idea or
even head about the problem already. If not, you might need to bisect this.

On 12/4/25 08:54, Demi Marie Obenour wrote:
> In Linux 6.18, mlx5_core does not load automatically.  This causes
> my server to not be accessible from the network.  Manually loading
> the module in the initramfs fixes the problem.  Everything worked
> with Linux 6.17.x.
> 
> I'm using the linuxPackages_latest package from Nixpkgs.

Mentioning it like that might scare kernel developers, as they have no
idea what this is and thus might suspect that it's a vendor kernel with
lots of patches applied. Is that the case? Or is that vanilla or at
least close to vanilla?

>  The server is a RISE-7 OVH dedicated server with an AMD Epyc 7402 CPU.

Ciao, Thorsten

#regzbot ^introduced: v6.17..v6.18
#regzbot title: net: mlx5_core: module does not autoload anymore Linux 6.18

Re: [REGRESSION]: mlx5_core does not autoload in Linux 6.18

[Demi Marie Obenour]

[-- Attachment #1.1.1: Type: text/plain, Size: 2232 bytes --]

On 12/4/25 06:53, Thorsten Leemhuis wrote:
> Lo! Thx for the report.
> 
> Let me CC the maintainers of said driver, maybe they have an idea or
> even head about the problem already. If not, you might need to bisect this.

Interestingly, the problem does not reproduce after a reboot and
a further Nixpkgs upgrade.  This either means that NixOS failed to
regenerate the initramfs, or that the problem is not reproducible
reliably.

> On 12/4/25 08:54, Demi Marie Obenour wrote:
>> In Linux 6.18, mlx5_core does not load automatically.  This causes
>> my server to not be accessible from the network.  Manually loading
>> the module in the initramfs fixes the problem.  Everything worked
>> with Linux 6.17.x.
>>
>> I'm using the linuxPackages_latest package from Nixpkgs.
> 
> Mentioning it like that might scare kernel developers, as they have no
> idea what this is and thus might suspect that it's a vendor kernel with
> lots of patches applied. Is that the case? Or is that vanilla or at
> least close to vanilla?

https://github.com/NixOS/nixpkgs/tree/7905606cfa51a1815787377b9cb04291e87ebcb4/pkgs/os-specific/linux/kernel
shows that there are very few patches.  The only ones I see are:

- Change the RANDSTRUCT seed to one provided by NixOS.
- Replace hard-coded paths with ones that are present in NixOS.
- Make Kconfig print ### after every question.
- Change two EXPORT_SYMBOL_GPL to EXPORT_SYMBOL
  so that ZFS works with PREEMPT_RT.

Of those, the first one is needed for reproducible builds with
RANDSTRUCT.  The second is needed because of NixOS's filesystem
structure, which isn't FHS-conforming.  The third is used by
NixOS-specific automation, and the last one is clearly not upstreamable
but also will not break anything.

>>  The server is a RISE-7 OVH dedicated server with an AMD Epyc 7402 CPU.
> 
> Ciao, Thorsten
> 
> #regzbot ^introduced: v6.17..v6.18
> #regzbot title: net: mlx5_core: module does not autoload anymore Linux 6.18
I wonder if this is even a kernel bug.  It could be that NixOS
generated a broken userspace that couldn't load the module.
The problem going away after nixos-rebuild supports this.
-- 
Sincerely,
Demi Marie Obenour (she/her/hers)

[-- Attachment #1.1.2: OpenPGP public key --]
[-- Type: application/pgp-keys, Size: 7253 bytes --]

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 833 bytes --]