On Mon, May 29, 2017 at 11:20:24AM +0200, Dietmar Eggemann wrote:

Hi Greg,

On 05/26/2017 08:36 PM, Greg KH wrote:

quoted

On Fri, May 26, 2017 at 11:10:32AM +0100, Juri Lelli wrote:

quoted

Hi,

On 25/05/17 15:18, Greg KH wrote:

quoted

On Thu, Apr 20, 2017 at 03:43:16PM +0100, Juri Lelli wrote:

[...]

quoted

quoted

quoted

But this is all really topology stuff, right?  Why use "capacity" at
all:
	topology_normalize_cpu()
	topology_parse_cpu()
	topology_scale_cpu()
	topology_set_scale()
?

It's always best to put the "subsystem" name first, we have a bad
history of getting this wrong in the past by putting the verb first, not
the noun.

topology_ works for me. However, I'd keep "capacity" in the names, as we
might need to topology_normalize_cpu_somethingelse() (etc.) in the
future?

Worry about the future, in the future.  Change the names then, _IF_ it
becomes an issue.  Try to be short and simple please.

quoted

Updated patch follows. I kept Catalin and Russell's acks as I only
renamed the functions, please shout if that's not OK.

Greg, if you are fine with this approach, do you still want a complete
v5 of the set or can you pick this up?

Am I the one who is supposed to take all of these arm-specific patches?
If so, that's fine, but I need to have acks from the arm maintainers...

Oh, and drop "capacity" please :)

Once we have driver/base/arch_topology.c in, we want to enable (cpu
micro-architectural + max frequency (OPPmax)) invariant and frequency
(OPPmin..OPPmax) invariant load-tracking/accounting in the task scheduler
for arm and arm64.

The way to do this is to define the task scheduler interfaces
arch_scale_cpu_capacity() and arch_scale_freq_capacity() in arch specific
code:

#define arch_scale_cpu_capacity topology_scale_cpu_capacity
#define arch_scale_freq_capacity topology_scale_freq_capacity

In case an arch is not defining them, the default definitions in
kernel/sched/sched.h are used.

So topology_scale_cpu() wouldn't be correct since we scale the _capacity_ by
the micro-architectural differences (hence cpu) and not the cpu.

Likewise we will have a function topology_scale_freq_capacity indicating
that we scale the capacity by the frequency.

Or would you prefer something like topology_scale_capacity_by_cpu() and
topology_scale_capacity_by_freq()?

I think that if you are creating an api that the scheduler will use, you
need to ask the scheduler maintainers/developers what they want to see
here, as that would be up to them, not me...

thanks,

greg k-h