Re: [PATCH 2/2] intel_idle: Introduce 'states_off' module parameter

[Rafael J. Wysocki]

On Fri, Jan 31, 2020 at 12:54 PM David Laight <David.Laight@xxxxxxxxxx> wrote:
>
> From: Artem Bityutskiy >
> > Sent: 31 January 2020 11:24
> > On Fri, 2020-01-31 at 11:07 +0000, David Laight wrote:
> > > Unless you know exactly which cpu table is being used the
> > > only constraint a user can request is the latency.
> >
> > Hi David,
> >
> > in all my use-cases I always know what is the CPU I am dealing with and
> > what are the C-states. Simply because in my view they are always CPU-
> > dependent in terms of what they do and how are they named.
> >
> > What you say sounds to me like you would want to disable some C-states
> > without knowing anything (or much) about the CPU you are dealing with
> > and the C-state names.
> >
> > If so, could you please share examples of such use-cases?
>
> Dunno, but clearly you want to disable (say) C3 while leaving C6
> enabled.
>
> I was trying to find why it was taking 600+us for a RT process
> to get rescheduled when it had only been sleeping for a few us.
>
> I found where it was sleeping, but that didn't help at all.
> Someone pointed me at a 'random' pdf that referred to /dev/cpu_dma_latency.
> Setting that to a small value (eg 20) helps no end.
> But there are no references in the code or man pages to that.

There is a piece of kernel documentation regarding it, however:

https://www.kernel.org/doc/html/latest/admin-guide/pm/cpuidle.html#cpu-pm-qos