Re: [PATCH] [RFC] rt: kernel/sched/core: fix kthread_park() pending too long when CPU un-plugged

[Sebastian Siewior]

On 2021-01-07 11:45:39 [+0100], Peter Zijlstra wrote:
> On Thu, Jan 07, 2021 at 05:18:41PM +0800, Ran Wang wrote:
> > +
> > +			if (IS_ENABLED(CONFIG_PREEMPT_RT) &&
> > +			    !strncmp(p->comm, "ksoftirqd/", 10))
> > +				schedule_hrtimeout(&to,
> > +					HRTIMER_MODE_REL | HRTIMER_MODE_HARD);
> > +			else
> > +				schedule_hrtimeout(&to, HRTIMER_MODE_REL);
> 
> This is horrific, why did you not self-censor and spare me the mental
> anguish of having to formulate a CoC compliant response?
> 
> It also violates coding style, but given the total lack of any sense,
> that seems like a minor detail.
> 
> Why can't we use HRTIMER_MODE_HARD unconditionally?

I had a similar patch in -RT and dropped it in v5.10-rc7-rt16.
It was added because RT couldn't boot since creating the boot-threads
didn't work before the ksoftirqd was up. This was fixed by commit
   26c7295be0c5e ("kthread: Do not preempt current task if it is going to call schedule()")

and live was good again.
tglx (also) suggested to add HRTIMER_MODE_HARD unconditionally (it
looked at SYSTEM_STATE back then) and I was only worried some abuse via
userland.
This sleep can be triggered by ptrace/strace() and with brief testing I
can trigger the sleep there but I don't get it anywhere near where I
would notice it with cyclictest.

Sebastian