Re: [PATCH v5 4/4] platform/x86/intel/pmc: core: Report duration of time in HW sleep state

[Rafael J. Wysocki]

On Mon, Apr 3, 2023 at 8:07 PM Limonciello, Mario
<mario.limonciello@xxxxxxx> wrote:
>
> On 4/3/2023 13:00, Box, David E wrote:
> > On Fri, 2023-03-31 at 20:05 +0200, Rafael J. Wysocki wrote:
> >> On Thu, Mar 30, 2023 at 9:45 PM Mario Limonciello
> >> <mario.limonciello@xxxxxxx> wrote:
> >>>
> >>> intel_pmc_core displays a warning when the module parameter
> >>> `warn_on_s0ix_failures` is set and a suspend didn't get to a HW sleep
> >>> state.
> >>>
> >>> Report this to the standard kernel reporting infrastructure so that
> >>> userspace software can query after the suspend cycle is done.
> >>>
> >>> Signed-off-by: Mario Limonciello <mario.limonciello@xxxxxxx>
> >>> ---
> >>> v4->v5:
> >>>   * Reword commit message
> >>> ---
> >>>   drivers/platform/x86/intel/pmc/core.c | 2 ++
> >>>   1 file changed, 2 insertions(+)
> >>>
> >>> diff --git a/drivers/platform/x86/intel/pmc/core.c
> >>> b/drivers/platform/x86/intel/pmc/core.c
> >>> index e2f171fac094..980af32dd48a 100644
> >>> --- a/drivers/platform/x86/intel/pmc/core.c
> >>> +++ b/drivers/platform/x86/intel/pmc/core.c
> >>> @@ -1203,6 +1203,8 @@ static inline bool pmc_core_is_s0ix_failed(struct
> >>> pmc_dev *pmcdev)
> >>>          if (pmc_core_dev_state_get(pmcdev, &s0ix_counter))
> >>>                  return false;
> >>>
> >>> +       pm_set_hw_sleep_time(s0ix_counter - pmcdev->s0ix_counter);
> >>> +
> >>
> >> Maybe check if this is really accumulating?  In case of a counter
> >> overflow, for instance?
> >
> > Overflow is likely on some systems. The counter is only 32-bit and at our
> > smallest granularity of 30.5us per tick it could overflow after a day and a half
> > of s0ix time, though most of our systems have a higher granularity that puts
> > them around 6 days.
> >
> > This brings up an issue that the attribute cannot be trusted if the system is
> > suspended for longer than the maximum hardware counter time. Should be noted in
> > the Documentation.
>
> I think it would be rather confusing for userspace having to account for
> this and it's better to abstract it in the kernel.
>
> How can you discover the granularity a system can support?
> How would you know overflow actually happened?  Is there a bit somewhere
> else that could tell you?

I'm not really sure if there is a generally usable overflow detection for this.

> In terms of ABI how about when we know overflow occurred and userspace
> reads the sysfs file we return -EOVERFLOW instead of a potentially bad
> value?

So if the new value is greater than the old one, you don't really know
whether or not an overflow has taken place.

And so I would just document the fact that the underlying HW/firmware
counter overflows as suggested by Dave.