Re: ~90s reboot delay with v6.19 and PREEMPT_RT

[Steven Rostedt]

On Thu, 26 Feb 2026 14:24:34 +0100
Bert Karwatzki <spasswolf@xxxxxx> wrote:

> I think I've found the reason for the stall. I was looking at commit 9311e6c29b34
> ("cgroup: Fix sleeping from invalid context warning on PREEMPT_RT"),
> and noticed this:
> 
> +static void __init cgroup_rt_init(void)
> +{
> +       int cpu;
> +
> +       for_each_possible_cpu(cpu) {
> +               init_llist_head(per_cpu_ptr(&cgrp_dead_tasks, cpu));
> +               per_cpu(cgrp_dead_tasks_iwork, cpu) =
> +                       IRQ_WORK_INIT_LAZY(cgrp_dead_tasks_iwork_fn);
> +       }
> 
> IRQ_WORK_INIT_LAZY() expands to __IRQ_WORK_INIT(_func, IRQ_WORK_LAZY):
> (in include/linux/irq_work.h)
> IRQ_WORK_LAZY is declare in include/linux/smp-types.h:
> 	IRQ_WORK_LAZY		= 0x04, /* No IPI, wait for tick */
> 
> The "wait for tick" gave me an idea as I'm also using
> CONFIG_NO_HZ_FULL=y

If you disable NO_HZ_FULL, does the problem also go away?

I wonder if you add irq events and trace a good and bad boot to see if it
definitely shows the delayed tick being an issue.

  echo 1 > /sys/kernel/tracing/instances/boot_map/events/irq/enable

-- Steve