Re: [RFC PATCH 0/2] sched: move content out of core files for load average

[Ingo Molnar <mingo@kernel.org>]

* Paul Gortmaker <paul.gortmaker@windriver.com> wrote:

> On 13-04-18 07:14 AM, Peter Zijlstra wrote:
> > On Mon, 2013-04-15 at 11:33 +0200, Ingo Molnar wrote:
> >> * Paul Gortmaker <paul.gortmaker@windriver.com> wrote:
> >>
> >>> Recent activity has had a focus on moving functionally related blocks of stuff 
> >>> out of sched/core.c into stand-alone files.  The code relating to load average 
> >>> calculations has grown significantly enough recently to warrant placing it in a 
> >>> separate file.
> >>>
> >>> Here we do that, and in doing so, we shed ~20k of code from sched/core.c (~10%).
> >>>
> >>> A couple small static functions in the core sched.h header were also localized 
> >>> to their singular user in sched/fair.c at the same time, with the goal to also 
> >>> reduce the amount of "broadcast" content in that sched.h file.
> >>
> >> Nice!
> >>
> >> Peter, is this (and the naming of the new file) fine with you too?
> > 
> > Yes and no.. that is I do like the change, but I don't like the
> > filename. We have _waaaay_ too many different things we call load_avg.
> > 
> > That said, I'm having a somewhat hard time coming up with a coherent
> > alternative :/
> 
> Several of the relocated functions start their name with "calc_load..."
> Does "calc_load.c" sound any better?

Peter has a point about load_avg being somewhat of a misnomer: that's not your 
fault in any way, we created overlapping naming within the scheduler and are now 
hurting from it.

Here are the main scheduler 'load' concepts we have right now:

 - The externally visible 'average load' value extracted by tools like 'top' via 
   /proc/loadavg and handled by fs/proc/loadavg.c. Internally the naming is all 
   over the map: the fields that are updated are named 'avenrun[]', most other 
   variables and methods are named calc_load_*(), and a few callbacks are named 
   *_cpu_load_*().

 - rq->cpu_load, a weighted, vectored scheduler-internal notion of task load 
   average with multiple run length averages. Only exposed by debug interfaces but 
   otherwise relied on by the scheduler for SMP load balancing.

 - se->avg - per entity (per task) load average. This is integrated differently 
   from the cpu_load - but work is ongoing to possibly integrate it with the
   rq->cpu_load metric. This metric is used for CPU internal execution time
   allocation and timeslicing, based on nice value priorities and cgroup 
   weights and constraints.

Work is ongoing to integrate rq->cpu_load and se->avg - eventually they will 
become one metric.

It might eventually make sense to integrate the 'average load' calculation as well 
with all this - as they really have a similar purpose, the avenload[] vector of 
averages is conceptually similar to the rq->cpu_load[] vector of averages.

So I'd suggest to side-step all that existing confusion and simply name the new 
file kernel/sched/proc.c - our external /proc scheduler ABI towards userspace. 
This is similar to the already existing kernel/irq/proc.c pattern.

A technical request: mind doing your patch against the tip:master tree? It's at:

      git git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip.git master

We have changes pending both in the sched/core, timers/nohz, core/locking and 
smp/hotplug trees, and your split-up interacts with all that pending work, 
creating conflicts.

Thanks,

	Ingo