On Sat, Oct 2, 2010 at 5:34 PM, Peter Zijlstra <peterz@infradead.org> wrote:
quoted text
> On Sat, 2010-10-02 at 08:42 -0700, Venkatesh Pallipadi wrote:
>>
>> > The make in_serving_softirq() be something like:
>> >  (preempt_count() & SOFTIRQ_OFFSET) || (current->flags & PF_SOFTIRQ)
>> >
>>
>> Yes. I would also need in_softirq_at_hardirq_tail() for accounting calls :)
>
> Well, you could open-code it there or something.
>
>> >> Also, ksoftirqd adds to softirq counts, does trace softirq, etc. So,
>> >> it kind of made sense to add the time also to softirq stats as well.
>> >> If we dont account time to softirq stats, then if some user is looking
>> >> at say time per softirq using the softirq count will be misled. No?
>> >
>> > Simply add back the task accounting when you report it?
>> >
>> >> In the other thread you mentioned doing that will cause problems. Were
>> >> you thinking of scheduler issues or are there other problems charging
>> >> softirq time this way?
>> >
>> > Of course there are.. you're double accounting the time of ksoftirqd,
>> > and worse, you're adding that back into the equation as part of the !
>> > sched_fair time.
>> >
>>
>> No. There should not be any double accounting with this current
>> change. We account softirq processing both at hardirq tail and
>> ksoftirqd as CPU softirq time. It will be taken out of ksoftirqd sched
>> exec time as with any other thread. And it will be taken out of fair
>> time available on the CPU as well. Which to me seems to be the right
>> thing to do, as the this is more coupled with the CPU and ksoftirqd is
>> just giving context for softirqd to run.
>
> But ksoftirqd is a SCHED_OTHER task, so by taking it out of its runtime
> the scheduler will get all confused.
>
>> Changing only hardirq tail to have SOFTIRQ_OFFSET and changing
>> ksoftirqd to SOFTIRQ_OFFSET*2 would result in these additional
>> ksoftirqd softirqs staying as the part of sched_fair time.
>>
>> Or I am totally missing something here?
>
> Please keep ksoftirq scheduling normal, if people want it to be another
> scheduling class, let them change that, but since its a task it should
> be a normal task and get scheduled like everybody else, not have some
> magic properties.
>

OK. In that case we can leave the existing softirq accounting
(!IRQ_TIME_ACCOUNTING) users as they are. They don't have any
ksoftirqd and scheduler issues. So, this patch should be good as it
is.
I will however, change IRQ_TIME_ACCOUNTING case to not take out
scheduler time out of ksoftirqd due to softirqs. Sounds good?

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