On Fri, 10 Jun 2016 17:57:17 +0200
Sebastian Andrzej Siewior [off-list ref] wrote:

* Steven Rostedt | 2016-06-04 07:11:31 [-0400]:

quoted

From: Steven Rostedt <redacted>
Date: Tue, 5 Jan 2016 14:53:09 -0500
Subject: [PATCH] softirq: Perform softirqs in local_bh_enable() for a limited
amount of time

To prevent starvation of tasks like ksoftirqd, if the task that is
processing its softirqs takes more than 2 jiffies to do so, and the
softirqs are constantly being re-added, then defer the processing to
ksoftirqd.  

I'm not sure about this. Alison didn't scream "yes it solves my problem"
and I am not sure what it is.
It is true that in RT we don't have such a limit like in !RT. You would
need to use __raise_softirq_irqoff_ksoft() instead the normal or +
wakeup() since you may have timers pending and those need to go to the
"other" ksoftirqd.
But then I don't see much change. ksoftirqd runs now at SCHED_OTHER so
it will end up on the CPU right away unless there other tasks that need
the CPU. So the scheduler will balance it the same way. The only change
will be that softirqs which are processed in context of any application
for more than two jiffies will be moved to ksoftirqd. This could be a
win.

We actually triggered a starvation due to this. I was just seeing if
Alison hit the same issue we did in our tests.

-- Steve