Re: [patch] voluntary-preempt-2.6.9-rc1-bk12-R5

[Rui Nuno Capela]

Hi,

Lee Revell wrote:
>
>Rui Nuno Capela wrote:
>>
>> 8) Indeed, only by disabling both softirq and hardirq preeemption I get
>> an usable VP+SMP kernel. But that's no surprise either, it has been
>> always like that until Q3, which was the latest VP+SMP combination that
>> didn't suffer with the Wacom tablet presence at boot/init time. I only
>> hoped the (soft|hard)irq trouble would be solved by R9 time.
>
> Rui, did you ever get this working?  Other testers are not reporting
> problems, it would be good to know if there are still bugs lurking.
>
> Have you tried booting with hard and softirq preemption disabled and
> enabling them one at a time?
>

I've beeing doing a lot experiments, the trial-and-error way, by tweaking
kernel config options on and off, and (re)building the
linux-2.6.9-rc1-bk12-S0 SMP kernel.

I have some news indeed. As you may recall, I'm trying to run jackd
realtime low-latency audio on a P4 2.80C HT (Hyperthreading), and I keep
CONFIG_SMP=y always set.

I found, almost by mistake, that whether CONFIG_SCHED_SMT is set makes a
lot of difference.

a) With CONFIG_SCHED_SMT=y, which I've being doing since ever, the system
behavior is that same one I've been complaining about: having
softirq-preempt=0 and hardirq-preempt=0 is the minimal setting to run
jackd in realtime mode without hard-locking the whole system. Even then, I
get the system completely frozen more times than I like, almost twice a
day! Can't figure out who's or what's the culprit here. It's quite random.

b) When CONFIG_SCHED_SMT is not set, I can run all along with
softirq-preempt=1, hardirq-preempt=1 et al. While running jackd in
realtime mode, I get NO hard-locks, but unfortunately XRUNs are plenty. A
real storm. However I've noticed that the whole seems pretty much stable,
as I didn't experience one single system hang. Regression to
softirq-preempt=0 and hardirq-preempt=0 dissolves the xrun storm to
nothing again.

All my experiments were done based on starting jackd -R -p 128 -n 2 ...
using an onboard Intel ICH5 soundcrap driven by snd-intel8x0 (alsa). Oh, I
forgot to say that it's been always with kernel-preempt=1 and
voluntary-preempt=1.

I'm preparing to take some latency traces later on, while regarding the
SMP=1 SMT=0 configuration and softirq=1 hardirq=1 setting, in a effort to
let that horrible XRUN flux getting out of the way somehow, someday ;)

So, bottomline is that the SMT-aware scheduler is not ready for VP, isn't
it? Does anyone care to confirm this out?

Cheers.
-- 
rncbc aka Rui Nuno Capela
rncbc@xxxxxxxxx

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/