Re: v5.10: sched_cpu_dying() hits BUG_ON during hibernation: kernel BUG at kernel/sched/core.c:7596!

[Valentin Schneider <valentin.schneider@arm.com>]

Hi,

On 22/12/20 09:13, Dexuan Cui wrote:
> Hi,
> I'm running a Linux VM with the recent mainline (48342fc07272, 12/20/2020) on Hyper-V.
> When I test hibernation, the VM can easily hit the below BUG_ON during the resume
> procedure (I estimate this can repro about 1/5 of the time). BTW, my VM has 40 vCPUs.
>
> I can't repro the BUG_ON with v5.9.0, so I suspect something in v5.10.0 may be broken?
>
> In v5.10.0, when the BUG_ON happens, rq->nr_running==2, and rq->nr_pinned==0:
>
> 7587 int sched_cpu_dying(unsigned int cpu)
> 7588 {
> 7589         struct rq *rq = cpu_rq(cpu);
> 7590         struct rq_flags rf;
> 7591
> 7592         /* Handle pending wakeups and then migrate everything off */
> 7593         sched_tick_stop(cpu);
> 7594
> 7595         rq_lock_irqsave(rq, &rf);
> 7596         BUG_ON(rq->nr_running != 1 || rq_has_pinned_tasks(rq));
> 7597         rq_unlock_irqrestore(rq, &rf);
> 7598
> 7599         calc_load_migrate(rq);
> 7600         update_max_interval();
> 7601         nohz_balance_exit_idle(rq);
> 7602         hrtick_clear(rq);
> 7603         return 0;
> 7604 }
>
> The last commit that touches the BUG_ON line is the commit
> 3015ef4b98f5 ("sched/core: Make migrate disable and CPU hotplug cooperative")
> but the commit looks good to me.
>
> Any idea?
>

I'd wager this extra task is a kworker; could you give this series a try?

  https://lore.kernel.org/r/20201218170919.2950-1-jiangshanlai@gmail.com

> Thanks,
> -- Dexuan