Re: [BUG] cpuset: empty sched domain with isolcpus=domain, panic at CPU offline

From: Waiman Long

Date: Sun Nov 08 2026 - 14:07:08 EST


On 11/8/26 1:46 PM, Mathieu Desnoyers wrote:
On 2026-11-08 13:35, Waiman Long wrote:
[...]
Yes, we are going to add the check to make sure that there will be at
least one housekeeping CPU left for the top_cpuset sched domain.

See
https://lore.kernel.org/lkml/20261105130344.664336-1-guopeng.zhang@xxxxxxxxx/
Thanks, I had missed that series. Refusing such a partition is a better
fix than giving the top cpuset no sched domain.

This patch is correct AFAICS. I am not against it but it may not be
necessary after we make sure that there is at least one HK online CPU in
the top sched domain.
Then I will drop my patch and rely on Guopeng's series. I can test its
v3 on the setup that hit the panic: 5 vCPUs with isolcpus=domain,0,1.

By my reading, your _cpu_down() patch covers a case that cpuset cannot
fix. With no partition, the last online housekeeping CPU could go
offline. No partition could then give one back.

Yes, that is the point.

Cheers,
Longman