Re: ksoftirqd/n permanently eating 60% of a CPU

[Louis-David Mitterrand <vindex+lists-linux-kernel@apartia.org>]

On Sun, Sep 12, 2010 at 12:31:30PM +0200, Andreas Mohr wrote:
> Hi,
> 
> Since there hasn't been a reply in a week, let me try to provide at
> least a semi-useless one ;)

Hi and thanks for offering your help.

> Such high use probably means that something in the kernel keeps
> retriggering some
> tasklet(?) since it thinks that something hasn't been done yet
> (i.e., the tasklet that got triggered didn't manage to satisfy the success
> criteria of its owner),
> or it means that something "legitimately" uses obscene amounts of
> softirqd activity (perhaps some debug infrastructure such as memory tracing
> or logging or some such).
> 
> Thus:
> 1. boot with "single" or "init=/bin/bash" kernel commandline and check
>    whether the activity still happens

Booting 2.6.35.4 with init=/bin/bash doesn't fix the problem:
ksoftirqd/n keeps firing at 50/60% CPU. No daemons are running of
course.

Examining /proc/interrupts shows that "Local timer interrupts" are
mainly firing.

Strangely enough, the interrupt storm starts only 5 minutes or so after
bootup.

I'd like to get to the bottom of this. As it happens on several, very
different machines.