Re: [RFC/PATCH 1/3] PM: QoS: Introduce a system-wakeup QoS limit

[Rafael J. Wysocki]

On Mon, Aug 11, 2025 at 7:16 PM Kevin Hilman <khilman@xxxxxxxxxxxx> wrote:
>
> "Rafael J. Wysocki" <rafael@xxxxxxxxxx> writes:
>
> > On Wed, Jul 16, 2025 at 2:33 PM Ulf Hansson <ulf.hansson@xxxxxxxxxx> wrote:
> >>
> >> Some platforms and devices supports multiple low-power-states than can be
> >> used for system-wide suspend. Today these states are selected on per
> >> subsystem basis and in most cases it's the deepest possible state that
> >> becomes selected.
> >>
> >> For some use-cases this is a problem as it isn't suitable or even breaks
> >> the system-wakeup latency constraint, when we decide to enter these deeper
> >> states during system-wide suspend.
> >>
> >> Therefore, let's introduce an interface for user-space, allowing us to
> >> specify the system-wakeup QoS limit. Subsequent changes will start taking
> >> into account the QoS limit.
> >
> > Well, this is not really a system-wakeup limit, but a CPU idle state
> > latency limit for states entered in the last step of suspend-to-idle.
> >
> > It looks like the problem is that the existing CPU latency QoS is not
> > taken into account by suspend-to-idle, so instead of adding an
> > entirely new interface to overcome this, would it make sense to add an
> > ioctl() to the existing one that would allow the user of it to
> > indicate that the given request should also be respected by
> > suspend-to-idle?
> >
> > There are two basic reasons why I think so:
> > (1) The requests that you want to be respected by suspend-to-idle
> > should also be respected by the regular "runtime" idle, or at least I
> > don't see a reason why it wouldn't be the case.
> > (2) The new interface introduced by this patch basically duplicates
> > the existing one.
>
> I also think that just using the existing /dev/cpu_dma_latency is the
> right approach here, and simply teaching s2idle to respect this value.
>
> I'm curious about the need for a new ioctl() though.  Under what
> conditions do you want normal/runtime CPUidle to respect this value and
> s2idle to not respect this value?

In a typical PC environment s2idle is a replacement for ACPI S3 which
does not take any QoS constraints into account, so users may want to
set QoS limits for run-time and then suspend with the expectation that
QoS will not affect it.