Re: [PATCH] softirq: Do not loop if running under a real-time task

[Sebastian Andrzej Siewior]

On 2023-03-06 16:45:48 [+0100], Petr Malat wrote:
> Softirq processing can be a source of a scheduling jitter if it executes
> in a real-time task as in that case need_resched() is false unless there
> is another runnable task with a higher priority. This is especially bad
> if the softirq processing runs in a migration thread, which has priority
> 99 and usually runs for a short time.
> 
> One option would be to not restart the softirq processing if there is
> another runnable task to allow the high prio task to finish and yield the
> CPU, the second one is to not restart if softirq executes in a real-time
> task. Usually, real-time tasks don't want to be interrupted, so implement
> the second option.

This affects only PEEMPT_RT, right?

I have plans to redo parts of it. You shouldn't enter ksoftirqd to be
begin with. There is this ktimerd in v6.1 which mitigates this to some
point and I plan to extend it to also cover the sched-softirq.
Other than that, you are right in saying that the softirq must not
continue with a RT prio and that need_resched() is not visible here.
However ksoftirqd itself must be able to do loops unless the
need-resched flag is seen.

Since you mentioned migration thread, how ofter to you see this or how
does this trigger?

> Signed-off-by: Petr Malat <oss@xxxxxxxxx>

Sebastian