Re: [PATCH] kthread: fall back when no preferred housekeeping CPU is online

[Frederic Weisbecker]

Le Tue, Oct 20, 2026 at 03:55:18PM +0300, Avi Weiss a écrit :
> kthread_fetch_affinity() falls back to the housekeeping mask
> only when the intersection with the preferred mask is empty.
> 
> If the intersection contains only offline CPUs,
> set_cpus_allowed_ptr() returns -EINVAL and the kthread keeps
> its previous affinity, which may still include an isolated CPU.
> 
> Check whether the intersection contains an online CPU instead.
> Fall back to the full housekeeping mask when none are online.
> 
> Fixes: e894f6339808 ("kthread: Honour kthreads preferred affinity after cpuset changes")
> Signed-off-by: Avi Weiss <thnkslprpt@xxxxxxxxx>
> ---
> 
> Tested on x86_64 Linux 7.3.0 with vng/qemu and a kthread with 
> preferred CPUs 1-2.
> 
> With CPU 2 offline and CPU 1 isolated, the unpatched kernel kept
> affinity 1-2 and set_cpus_allowed_ptr() returned -EINVAL. With the
> fix, affinity changed to 0,2. Bringing CPU 2 online restored 
> affinity to 2, and removing isolation restored 1-2.
> 
>  kernel/kthread.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/kernel/kthread.c b/kernel/kthread.c
> index cc8cb5d3eab7..fdb8fb07f40b 100644
> --- a/kernel/kthread.c
> +++ b/kernel/kthread.c
> @@ -343,7 +343,7 @@ static void kthread_fetch_affinity(struct kthread *kthread, struct cpumask *cpum
>  	}
>  
>  	cpumask_and(cpumask, pref, housekeeping_cpumask(HK_TYPE_DOMAIN));
> -	if (cpumask_empty(cpumask))
> +	if (!cpumask_intersects(cpumask, cpu_online_mask))
>  		cpumask_copy(cpumask, housekeeping_cpumask(HK_TYPE_DOMAIN));
>  }

Good catch thank you! Queueing it.



>  
> 
> base-commit: f47d223627f80cb7a1eb14c585fe6cb35de5e984
> -- 
> 2.43.0
> 

-- 
Frederic Weisbecker
SUSE Labs