Re: [PATCH RFC] intel_pstate: play well with frequency limits set by acpi

[Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com>]

On Tue, 2015-07-21 at 19:37 +0300, Konstantin Khlebnikov wrote:
> On 21.07.2015 18:37, Srinivas Pandruvada wrote:
> > On Tue, 2015-07-21 at 13:25 +0300, Konstantin Khlebnikov wrote:
> >> On 21.07.2015 00:08, Srinivas Pandruvada wrote:
> >>> On Fri, 2015-07-17 at 07:36 +0300, Konstantin Khlebnikov wrote:
> >>>> On Fri, Jul 17, 2015 at 1:08 AM, Srinivas Pandruvada
> >>>> <srinivas.pandruvada@linux.intel.com> wrote:
> >>>>> On Thu, 2015-07-16 at 21:17 +0300, Konstantin Khlebnikov wrote:
> >>>>>> IPMI can control CPU P-states remotely: configuration is reported via
> >>>>>> common ACPI interface (_PPC/_PSS/etc). This patch adds required minimal
> >>>>>> support in intel_pstate to receive and use these P-state limits.
> >>>>>>
> >>>>>> * ignore limit of top state in _PPC: it lower than turbo boost frequency
> >>>>>> * register intel_pstate in acpi-processor to get states from _PSS
> >>>>>> * link acpi_processor_get_bios_limit: this adds attribute "bios_limit"
> >>>>>>
> >>>>>> Signed-off-by: Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
> >>>>>> ---
> >>>>>>    drivers/acpi/processor_perflib.c |    3 +-
> >>>>>>    drivers/cpufreq/intel_pstate.c   |   57 ++++++++++++++++++++++++++++++++++++++
> >>>>>>    2 files changed, 59 insertions(+), 1 deletion(-)
> >>>>>>
> >>>>>> diff --git a/drivers/acpi/processor_perflib.c b/drivers/acpi/processor_perflib.c
> >>>>>> index cfc8aba72f86..781e328c9d5f 100644
> >>>>>> --- a/drivers/acpi/processor_perflib.c
> >>>>>> +++ b/drivers/acpi/processor_perflib.c
> >>>>>> @@ -98,7 +98,8 @@ static int acpi_processor_ppc_notifier(struct notifier_block *nb,
> >>>>>>
> >>>>>>         ppc = (unsigned int)pr->performance_platform_limit;
> >>>>>>
> >>>>>> -     if (ppc >= pr->performance->state_count)
> >>>>>> +     /* Ignore limit of top state: it lower than turbo boost frequency */
> >>>>>> +     if (!ppc || ppc >= pr->performance->state_count)
> >>>>> Why? Isn't the previous check enough?
> >>>>
> >>>> Zero _PPC state must be top performance state but as I see frequency in
> >>>> _PSS is lower than maximum possible turbo frequency. So, in this case
> >>>> intel_pstate cannnot get "100%"  for max bound even it there is no limit set.
> >>>>
> >>>> For example: I saw _PSS[0] = 2601 Mhz, PSS[1] = 2600 Mhz while turbo
> >>>> state is 3400 Mhz.
> >>>>
> >>> Have you tested dynamic _PPC modification with acpi cpufreq with this
> >>> change (after boot)? Suppose _PPC is changed from 3 to 0, then
> >>> cpufreq_verify_within_limits will not be called to change to new max
> >>> turbo performance state.
> >>>
> >>
> >> I haven't checked that but as I see acpi_processor_ppc_notifier()
> >> can only reduce maximum frequency. So, there should be no problem
> >> in this case.
> > No, it can also be used in both ways. Once reduced, it can increase as
> > well. _PPC can be dynamically modified by BIOS to reduce and also to
> > increase.
> 
> Well, in this case BIOS will trigger ACPI_PROCESSOR_NOTIFY_PERFORMANCE:
> kernel evaluate new _PPC and call cpufreq_update_policy()
> which set initial frequency min/max range according to user setup and
> apply all limits after that. Initial policy->user_policy.min/max stay
> unchanged. So, that dynamic modification works in both ways.
> 
Fair enough. We need to take account for _PSS. We have some changes for
this, but not gone through test cycle. I will post them as RFC, please
check. Thanks for your patience.

- Srinivas