Re: [RFC PATCH 1/1] smi_detector: A System Management Interrupt detector

[Jon Masters <jonathan@jonmasters.org>]

On Tue, 2009-06-02 at 11:32 -0700, Paul E. McKenney wrote:
> On Sun, May 31, 2009 at 12:31:18PM -0400, Jon Masters wrote:
> > This patch introduces a new SMI (System Management Interrupt) detector module
> > that can be used to detect high latencies within the system. It was originally
> > written for use in the RT kernel, but has wider applications.
> 
> This would have been extremely handy a few years back when we were
> chasing some latency issues!!!  ;-)

Thanks. One guy on IRC already asked why it didn't find all his SMIs.
I'll add to the docs that this can obviously only detect when it's
sampling - since we can't constantly sample without "locking up" the
machine, it's a tradeoff. The defaults are intended so you could put
them on a laptop and not notice - in practice, I've been running with 1
second window size and .5 second sample width - i.e. very meaty and
highly disruptive, but great at catching the little so and sos :)

> I don't see how this handles CPU hotplug operations (see interspersed),
> but I am OK with "don't do CPU-hotplug operations while running this
> test."

It doesn't, good catch. Though in fairness, I don't think that's big
issue - but having said that, the ring_buffer handles cpu hotplug nicely
and I just need to think a bit harder, so I'll possibly do that in my
next v3 posting...gonna go get that one ready now.

Thanks guys,

Jon.