Date:	Mon, 1 Dec 2014 18:27:38 +0100
From:	Martin Schwidefsky <schwidefsky@...ibm.com>
To:	Thomas Gleixner <tglx@...utronix.de>
Cc:	Frederic Weisbecker <fweisbec@...il.com>,
	LKML <linux-kernel@...r.kernel.org>,
	Tony Luck <tony.luck@...el.com>,
	Peter Zijlstra <peterz@...radead.org>,
	Heiko Carstens <heiko.carstens@...ibm.com>,
	Benjamin Herrenschmidt <benh@...nel.crashing.org>,
	Oleg Nesterov <oleg@...hat.com>,
	Paul Mackerras <paulus@...ba.org>,
	Wu Fengguang <fengguang.wu@...el.com>,
	Ingo Molnar <mingo@...nel.org>, Rik van Riel <riel@...hat.com>
Subject: Re: [RFC PATCH 07/30] cputime: Convert kcpustat to nsecs

On Mon, 1 Dec 2014 18:15:36 +0100 (CET)
Thomas Gleixner <tglx@...utronix.de> wrote:

> On Mon, 1 Dec 2014, Martin Schwidefsky wrote:
> > On Mon, 1 Dec 2014 17:10:34 +0100
> > Frederic Weisbecker <fweisbec@...il.com> wrote:
> > 
> > > Speaking about the degradation in s390:
> > > 
> > > s390 is really a special case. And it would be a shame if we prevent from a
> > > real core cleanup just for this special case especially as it's fairly possible
> > > to keep a specific treatment for s390 in order not to impact its performances
> > > and time precision. We could simply accumulate the cputime in per-cpu values:
> > > 
> > > struct s390_cputime {
> > >        cputime_t user, sys, softirq, hardirq, steal;
> > > }
> > > 
> > > DEFINE_PER_CPU(struct s390_cputime, s390_cputime);
> > > 
> > > Then on irq entry/exit, just add the accumulated time to the relevant buffer
> > > and account for real (through any account_...time() functions) only on tick
> > > and task switch. There the costly operations (unit conversion and call to
> > > account_...._time() functions) are deferred to a rarer yet periodic enough
> > > event. This is what s390 does already for user/system time and kernel
> > > boundaries.
> > > 
> > > This way we should even improve the situation compared to what we have
> > > upstream. It's going to be faster because calling the accounting functions
> > > can be costlier than simple per-cpu ops. And also we keep the cputime_t
> > > granularity. For archs like s390 which have a granularity higher than nsecs,
> > > we can have:
> > > 
> > >    u64 cputime_to_nsecs(cputime_t time, u64 *rem);
> > > 
> > > And to avoid remainder losses, we can do that from the tick:
> > > 
> > >     delta_cputime = this_cpu_read(s390_cputime.hardirq);
> > >     delta_nsec = cputime_to_nsecs(delta_cputime, &rem);
> > >     account_system_time(delta_nsec, HARDIRQ_OFFSET);
> > >     this_cpu_write(s390_cputime.hardirq, rem);
> > > 
> > > Although I doubt that remainders below one nsec lost each tick matter that much.
> > > But if it does, it's fairly possible to handle like above.
> >  
> > To make that work we would have to move some of the logic from account_system_time
> > to the architecture code. The decision if a system time delta is guest time,
> > irq time, softirq time or simply system time is currently done in 
> > kernel/sched/cputime.c.
> > 
> > As the conversion + the accounting is delayed to a regular tick we would have
> > to split the accounting code into decision functions which bucket a system time
> > delta should go to and introduce new function to account to the different buckets.
> > 
> > Instead of a single account_system_time we would have account_guest_time,
> > account_system_time, account_system_time_irq and account_system_time_softirq.
> > 
> > In principle not a bad idea, that would make the interrupt path for s390 faster
> > as we would not have to call account_system_time, only the decision function
> > which could be an inline function.
> 
> Why make this s390 specific?
> 
> We can decouple the accounting from the time accumulation for all
> architectures.
> 
> struct cputime_record {
>        u64 user, sys, softirq, hardirq, steal;
> };
> 
> DEFINE_PER_CPU(struct cputime_record, cputime_record);
> 
> Now let account_xxx_time() just work on that per cpu data
> structures. That would just accumulate the deltas based on whatever
> the architecture uses as a cputime source with whatever resolution it
> provides.
> 
> Then we collect that accumulated results for the various buckets on a
> regular base and convert them to nano seconds. This is not even
> required to be at the tick, it could be done by some async worker and
> on idle enter/exit.

And leave the decision making in kernel/sched/cputime.c. Yes, that is good.
This would make the arch and the account_xxx_time() function care about
cputime_t and all other common code would use nano-seconds. With the added
benefit that I do not have to change the low level code too much ;-)

-- 
blue skies,
   Martin.

"Reality continues to ruin my life." - Calvin.

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

Powered by blists - more mailing lists