[linux-audio-user] -rt IRQ handler priorities (was: Re: molnar patch)

[Wolfgang Hoffmann]

On Tuesday 09 May 2006 01:08, Lee Revell wrote:
> On Tue, 2006-05-09 at 00:47 +0200, Florian Paul Schmidt wrote:
> > On Mon, 8 May 2006 21:54:53 +0200
> >
> > Wolfgang Hoffmann <woho at woho.de> wrote:
> > > Apropos: on your page on -rt setup (excellent page, btw., many thanks!
> > > :-), you suggest raising "softirq-timer/0" to prio 99, to make sleep()
> > > function right (http://tapas.affenbande.org/?page_id=40 sleep()
> > > based/system timer).
> > >
> > > I did so, and got strange latencies (> 40 ms) exactly once every 10
> > > minutes, caused by some routing-related action (rt_secret_rebuild)
> > > being run by the softirq-timer/0 thread. Don't you get bit by that,
> > > too? Kernel is 2.6.16-rt16
> >
> > I haven't been bitten by that. Do you also get xruns with [i suppose so,
> > just asking to make sure]? I haven't had as much time as before to play
> > around and test things, so maybe it has crept into the kernel recently
> > or maybe i just always had high-res timers enabled.

To be more precise, I get xruns, not scheduling latencies. Scheduling latency 
is low (~50us), but softirq-timer/0 preempts all other threads when boosted 
to prio 99, leading to xruns in jackd.

It's not a scheduler or -rt latency problem, it's a priority problem, since 
softirq-timer/0 runs different jobs that would require different priorities: 
wakeup sleep() at highest prio, rt_secret_rebuild et. al. at low prio.

With CONFIG_HIGH_RES_TIMERS=y it seems thatt softirq-timer/0 is no longer 
needed for waking up sleep()ing threads, so it can be left at it's default 
low prio (or even demoted to SCHED_OTHER?).

> > > My solution is to configure with CONFIG_HIGH_RES_TIMERS=y. Then,
> > > sleep() wakes up correctly even with softirq-timer/0 being low-priority
> > > (SCHED_FIFO 1 or even SCHED_OTHER).
> > >
> > > In general I find adjusting priorities of the various softirq kernel
> > > threads a bit of secret art. I can't find much documentation about
> > > "what kernel thread runs which job" that would help making some proper
> > > decisions here. I found my desktop "feels" most responsive when
> > > demoting all softirq thread to SCHED_OTHER. I did so after seeing that
> > > with a non-rt kernel, bottom-half handler don't run SCHED_FIFO/_RR at
> > > all. So -rt now gives me robust low latencies for jackd, and still
> > > proper desktop feeling.
> > >
> > > Well, maybe this is getting off-topic for this list. But it seems to me
> > > trimming priorities between kernel and userland threads is a bit like
> > > no man's land.
> >
> > I agree. Maybe Lee Revell knows more [CC'ing him]. Lee, you know
> > something about all these softirq threads? What exactly do they do?
>
> Not really.  The problem with making the softirq timer thread high
> priority is that lots of random code gets run from softirq timer
> context, including rt_secret_rebuild() which is a well known latency
> killer.

Ah, ok. Now well known to me, too ;-)

> Lately I've been focused on getting mainline usable for audio - I have
> not run -rt in a while.
>
> It is a bit of a black art because you need lots of knowledge of the
> kernel to know how to set the priorities.  But that's the nature of hard
> realtime - you really need to know a lot about the system to get the
> priorities right.
>
> I've never had to alter the priority of the softirq timer thread myself
> - does it work OK if you leave the softirq thread alone?

With CONFIG_HIGH_RES_TIMERS=y, yes.

With CONFIG_HIGH_RES_TIMERS=n, no. Then, high-prio theads don't wake up from 
sleep() if a mid-prio CPU hog is running. Typical scenario is a high-prio 
watchdog like Florians rt_watchdog (http://tapas.affenbande.org/?page_id=38), 
which fails to work unless softirq-timer/0 runs with higher prio than the CPU 
hog. It seems that softirq-timer/0 has to be scheduled for any thread to wake 
up. It's a bit like priority inversion.

> If not try setting the softirq thread to a lower RT priority than your
> soundcard, JACK, and all other interrupt threads.

Yes, that should cure the xruns. But I wonder if jackd's watchdog can do it's 
job when a jack client is running wild? Probably not. softirq-timer/0 would 
not get to run then, so the watchdog wouldn't wake up.

> I've added Ingo to the cc: list, maybe he has some input.

That would be helpful. Also adding Steve to cc: as he mentioned writing a 
section in an O'Reilly book about how to use the -rt patch in 
http://www.ussg.iu.edu/hypermail/linux/kernel/0604.2/0340.html

Wolfgang