Re: [PATCH] PM/QoS: resume latency 0 should be not latency allowed

["Rafael J. Wysocki" <rafael@kernel.org>]

On Mon, Nov 1, 2021 at 1:50 PM Youquan Song <youquan.song@intel.com> wrote:
>
> ICX and more recent Intel CPU has C1 exit_latency equal 1 in intel_idle

Please expand "ICX".

> driver, There are PM QoS interfaces: global /dev/cpu_dma_latency and
> per-device latency constraints. per-CPU set pm_qos_resume_latency_us
> to limit single CPU entery into deepest C-state.
> When it is required to disable entry into even C1, we need to set "0"
>  to /sys pm_qos_resume_latency_us, but "0" of pm_qos_resume_latency_us
> is redefined to no constraint now.
>
> so this patch set "0" back to original meaning "no latency at all".

It is not "back", it simply changes the meaning of 0.  And that's why
Documentation/ABI/testing/sysfs-devices-power needs to be updated.

> Signed-off-by: Youquan Song <youquan.song@intel.com>
> ---
>  drivers/base/power/sysfs.c | 8 ++------
>  1 file changed, 2 insertions(+), 6 deletions(-)
>
> diff --git a/drivers/base/power/sysfs.c b/drivers/base/power/sysfs.c
> index a1474fb67db9..b4bd67e6f440 100644
> --- a/drivers/base/power/sysfs.c
> +++ b/drivers/base/power/sysfs.c
> @@ -214,10 +214,8 @@ static ssize_t pm_qos_resume_latency_us_show(struct device *dev,
>  {
>         s32 value = dev_pm_qos_requested_resume_latency(dev);
>
> -       if (value == 0)
> -               return sysfs_emit(buf, "n/a\n");
>         if (value == PM_QOS_RESUME_LATENCY_NO_CONSTRAINT)
> -               value = 0;
> +               return sysfs_emit(buf, "no constraint\n");
>
>         return sysfs_emit(buf, "%d\n", value);
>  }
> @@ -237,10 +235,8 @@ static ssize_t pm_qos_resume_latency_us_store(struct device *dev,
>                 if (value < 0 || value == PM_QOS_RESUME_LATENCY_NO_CONSTRAINT)
>                         return -EINVAL;
>
> -               if (value == 0)
> -                       value = PM_QOS_RESUME_LATENCY_NO_CONSTRAINT;
>         } else if (sysfs_streq(buf, "n/a")) {
> -               value = 0;
> +               value = PM_QOS_RESUME_LATENCY_NO_CONSTRAINT;
>         } else {
>                 return -EINVAL;
>         }
> --

But do you really need to make these changes?  What problem is there
with using the interface as currently documented, ie. writing "n/a" to
it to get "no latency at all"?