power management regression in linux 3.10?

power management regression in linux 3.10?

[Michael Karcher]

Hello,

I am using an IBM ThinkPad T60 (i915, Core 2 Duo T7200) and recently
upgraded from a custom 3.7 kernel to a custom 3.10-rc4 kernel
(v3.10-rc4-1-gc026d8e).
After upgrading, I get around 17-20W power usage instead of 12-15W power
usage as I had with 3.7. powertop show that CPU0 is in state POLL all of
the idle time, while CPU1 correctly enters C3.

As this may be caused by my kernel configuration, I pastebinned
my .config to http://pastebin.ca/2388457. Note that I am using
CONFIG_NO_HZ_FULL, which might be nonstandard and actually not sensible
for a standard single-user laptop system with GUI.

Do you want me to try something to debug the problem?

Regards,
  Michael Karcher

Re: power management regression in linux 3.10?

[Rafael J. Wysocki]

On Monday, June 03, 2013 07:56:15 PM Michael Karcher wrote:
> Hello,
> 
> I am using an IBM ThinkPad T60 (i915, Core 2 Duo T7200) and recently
> upgraded from a custom 3.7 kernel to a custom 3.10-rc4 kernel
> (v3.10-rc4-1-gc026d8e).
> After upgrading, I get around 17-20W power usage instead of 12-15W power
> usage as I had with 3.7. powertop show that CPU0 is in state POLL all of
> the idle time, while CPU1 correctly enters C3.
> 
> As this may be caused by my kernel configuration, I pastebinned
> my .config to http://pastebin.ca/2388457. Note that I am using
> CONFIG_NO_HZ_FULL, which might be nonstandard and actually not sensible
> for a standard single-user laptop system with GUI.
> 
> Do you want me to try something to debug the problem?

Yes, please try with CONFIG_NO_HZ_IDLE and see if you can reproduce the
problem.

Does your system use intel_idle?  You've set that in .config, but can you check
what dmesg says, please?

Rafael


-- 
I speak only for myself.
Rafael J. Wysocki, Intel Open Source Technology Center.

Re: power management regression in linux 3.10?

[Dirk Brandewie]

On 06/03/2013 12:34 PM, Rafael J. Wysocki wrote:
> On Monday, June 03, 2013 07:56:15 PM Michael Karcher wrote:
>> Hello,
>>
>> I am using an IBM ThinkPad T60 (i915, Core 2 Duo T7200) and recently
>> upgraded from a custom 3.7 kernel to a custom 3.10-rc4 kernel
>> (v3.10-rc4-1-gc026d8e).
>> After upgrading, I get around 17-20W power usage instead of 12-15W power
>> usage as I had with 3.7. powertop show that CPU0 is in state POLL all of
>> the idle time, while CPU1 correctly enters C3.
>>
>> As this may be caused by my kernel configuration, I pastebinned
>> my .config to http://pastebin.ca/2388457. Note that I am using
>> CONFIG_NO_HZ_FULL, which might be nonstandard and actually not sensible
>> for a standard single-user laptop system with GUI.
>>
>> Do you want me to try something to debug the problem?
>
> Yes, please try with CONFIG_NO_HZ_IDLE and see if you can reproduce the
> problem.
>
> Does your system use intel_idle?  You've set that in .config, but can you check
> what dmesg says, please?
>
> Rafael
>
>

Unset CONFIG_NO_HZ it is setting CONFIG_CPU_IDLE_GOV_MENU which with 
CONFIG_NO_HZ_FULL set is confusing intel_pstate.  I am still trying to
figure out what the root cause is commit 5811d996 seems to be implicated.

--Dirk

Re: power management regression in linux 3.10?

[Rafael J. Wysocki]

On Monday, June 03, 2013 12:37:21 PM Dirk Brandewie wrote:
> On 06/03/2013 12:34 PM, Rafael J. Wysocki wrote:
> > On Monday, June 03, 2013 07:56:15 PM Michael Karcher wrote:
> >> Hello,
> >>
> >> I am using an IBM ThinkPad T60 (i915, Core 2 Duo T7200) and recently
> >> upgraded from a custom 3.7 kernel to a custom 3.10-rc4 kernel
> >> (v3.10-rc4-1-gc026d8e).
> >> After upgrading, I get around 17-20W power usage instead of 12-15W power
> >> usage as I had with 3.7. powertop show that CPU0 is in state POLL all of
> >> the idle time, while CPU1 correctly enters C3.
> >>
> >> As this may be caused by my kernel configuration, I pastebinned
> >> my .config to http://pastebin.ca/2388457. Note that I am using
> >> CONFIG_NO_HZ_FULL, which might be nonstandard and actually not sensible
> >> for a standard single-user laptop system with GUI.
> >>
> >> Do you want me to try something to debug the problem?
> >
> > Yes, please try with CONFIG_NO_HZ_IDLE and see if you can reproduce the
> > problem.
> >
> > Does your system use intel_idle?  You've set that in .config, but can you check
> > what dmesg says, please?
> >
> > Rafael
> >
> >
> 
> Unset CONFIG_NO_HZ it is setting CONFIG_CPU_IDLE_GOV_MENU which with 
> CONFIG_NO_HZ_FULL set is confusing intel_pstate.  I am still trying to
> figure out what the root cause is commit 5811d996 seems to be implicated.

Well, what about CCing the author (done now)?

Rafael

Re: power management regression in linux 3.10?

[Frederic Weisbecker]

On Mon, Jun 03, 2013 at 10:04:30PM +0200, Rafael J. Wysocki wrote:
> On Monday, June 03, 2013 12:37:21 PM Dirk Brandewie wrote:
> > On 06/03/2013 12:34 PM, Rafael J. Wysocki wrote:
> > > On Monday, June 03, 2013 07:56:15 PM Michael Karcher wrote:
> > >> Hello,
> > >>
> > >> I am using an IBM ThinkPad T60 (i915, Core 2 Duo T7200) and recently
> > >> upgraded from a custom 3.7 kernel to a custom 3.10-rc4 kernel
> > >> (v3.10-rc4-1-gc026d8e).
> > >> After upgrading, I get around 17-20W power usage instead of 12-15W power
> > >> usage as I had with 3.7. powertop show that CPU0 is in state POLL all of
> > >> the idle time, while CPU1 correctly enters C3.

With CONFIG_NO_HZ_FULL it is expected that CPU 0 consumes more power because
it handles the timekeeping for every other CPUs and so it never stop its tick.
Paul McKenney is working on a solution to solve that.

But if you see something similar on the other CPUs it's a bug.

Thanks.

Re: power management regression in linux 3.10?

[Michael Karcher]

Am Montag, den 03.06.2013, 12:37 -0700 schrieb Dirk Brandewie:
> On 06/03/2013 12:34 PM, Rafael J. Wysocki wrote:
> > On Monday, June 03, 2013 07:56:15 PM Michael Karcher wrote:
> >> Hello,
> > Yes, please try with CONFIG_NO_HZ_IDLE and see if you can reproduce the
> > problem.
Instead of doing that, I successfully applied Dirk's suggestion. Is
CONFIG_NO_HZ + CONFIG_NO_HZ_IDLE worth testing?

> > Does your system use intel_idle?  You've set that in .config, but can you check
> > what dmesg says, please?
No, intel_idle complains it doesn't support my CPU family.

> Unset CONFIG_NO_HZ it is setting CONFIG_CPU_IDLE_GOV_MENU which with 
> CONFIG_NO_HZ_FULL set is confusing intel_pstate.  I am still trying to
> figure out what the root cause is commit 5811d996 seems to be implicated.
Unsetting CONFIG_NO_HZ helped. Now CPU0 is using C states again.

> --Dirk
Regards,
  Michael Karcher

Re: power management regression in linux 3.10? (related to CONFIG_NO_HZ_FULL)

[Rafael J. Wysocki]

On Tuesday, June 04, 2013 03:08:48 AM Michael Karcher wrote:
> Am Montag, den 03.06.2013, 12:37 -0700 schrieb Dirk Brandewie:
> > On 06/03/2013 12:34 PM, Rafael J. Wysocki wrote:
> > > On Monday, June 03, 2013 07:56:15 PM Michael Karcher wrote:
> > >> Hello,
> > > Yes, please try with CONFIG_NO_HZ_IDLE and see if you can reproduce the
> > > problem.
> Instead of doing that, I successfully applied Dirk's suggestion. Is
> CONFIG_NO_HZ + CONFIG_NO_HZ_IDLE worth testing?

Yes, it is.

> > > Does your system use intel_idle?  You've set that in .config, but can you check
> > > what dmesg says, please?
> No, intel_idle complains it doesn't support my CPU family.

I thought so.  It is using the ACPI idle driver, then, so I think cpuidle in
general is affected.

That seems to be a time accounting problem of some sort (adding CCs to Ingo and
Thomas).

Thanks,
Rafael


-- 
I speak only for myself.
Rafael J. Wysocki, Intel Open Source Technology Center.