On 10/16/25 16:54, Christoph Berg wrote:
> Re: Tomas Vondra
>> It's probably more about the kernel version. What kernels are used by
>> these systems?
>
> It's the very same kernel, just different docker containers on the
> same system. I did not investigate yet where the problem is coming
> from, different libnuma versions seemed like the best bet.
>
> Same (differing) results on both these systems:
> Linux turing 6.16.7+deb14-amd64 #1 SMP PREEMPT_DYNAMIC Debian 6.16.7-1 (2025-09-11) x86_64 GNU/Linux
> Linux jenkins 6.1.0-39-amd64 #1 SMP PREEMPT_DYNAMIC Debian 6.1.148-1 (2025-08-26) x86_64 GNU/Linux
>
Hmmm. Those seem like relatively recent kernels.
>> Not sure how would that work. It seems this is some sort of permission
>> check in numa_move_pages, that's not what pg_numa_available does. Also,
>> it may depending on the page queried (e.g. whether it's exclusive or
>> shared by multiple processes).
>
> It's probably the lack of some process capability in that environment.
> Maybe there is a way to query that, but I don't know much about that
> yet.
>
move_page() manpage mentions PTRACE_MODE_READ_REALCREDS (man ptrace) so
maybe that's it.
--
Tomas Vondra
