ULE/sched issues on stable/9 - why isn't preemption occuring?

[Attilio Rao]

2012/5/29 Adrian Chadd <adrian at freebsd.org>:
> Hi Alexander and others,
>
> I've been tinkering with ath(4) IO scheduling and taskqueues. In order
> to get proper "in order" TX IO occuring, I've placed ath_start() into
> a taskqueue so now whenever ath_start() is called, it just schedules a
> taskqueue entry to run.
>
> However, performance is worse. :-)
>
> Here's a schedgraph trace.
>
> http://people.freebsd.org/~adrian/ath/ktr.4-ath-iperf-using-taskqueue-for-tx.ktr.gz
>
> I've thrown this through schedgraph.py on stable/9 and I've found some
> rather annoying behaviour. It seems that the ath0 taskqueue stays in
> the "runq add" state for quite a long time (1.5ms and longer) because
> something else is going on on CPU #0.
>
> I'm very confused about what's going on. I'd like a hand trying to
> figure out why the schedgraph output is the way it is.

What I would usually do for this cases, is to patch your kernel with
more KTR traces (handmade class, add a fictious KTR class after
KTR_BUF) in the interested code paths to log why your task is not
really scheduled.

Attilio


-- 
Peace can only be achieved by understanding - A. Einstein