Skip to contentSkip to conversation

Too many poll_idle cpu calls in kernel 3.2 rc-series

DISCUSSION Nicolas Kalkhof 4 participants 4 messages Nov 28, 2011 – Dec 10, 2011

lore
Share & tools
lore
#0 lore raw

Andi,

correct. The issue was introduced somewhere between 3.1 final and 3.2-rc1. I've tried to track down the bad commit but so far I haven't been successfull. The only workaround I found is to set the idle boot parameter to mwait and to disable CONFIG_INTEL_IDLE. Somehow the kernel seems to set idle to poll by default - no idea why :-(. I wonder if other mobile sandy bridge cpus are affected as well.

Regards,
Nic

-----Ursprüngliche Nachricht-----
Von: "Andi Kleen" <[email protected]>
Gesendet: Nov 28, 2011 8:33:28 PM
An: "Nicolas Kalkhof" <[email protected]>
Betreff: Re: Too many poll_idle cpu calls in kernel 3.2 rc-series

Quoted 13 lines
Quoted 5 lines · Nicolas Kalkhof

the kernel v3.2 series shows excessive poll_idle calls during system idle phases causing the cpu (mobile i7 2620M) to heat up and draw additional 10% more power.  All kernel versions between 3.2-rc1 and 3.2-rc3 are affected. Perf Top and Powertop both show the same issue. I've posted a similar bug report before: http://marc.info/?l=linux-kernel&m=132164909313594&w=2
and used perf top -U to verify the poll_idle calls as requested.

So it didn't happen with earlier kernels?
I would suggest a git bisect in this case.

-Andi

Quoted 7 lines

--
[email protected] -- Speaking for myself only
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

___________________________________________________________
SMS schreiben mit WEB.DE FreeMail - einfach, schnell und kostenguenstig. Jetzt gleich testen! http://f.web.de/?mc=021192

#1 lore raw
Quoted 3 lines · Nicolas Kalkhof

Andi,

correct. The issue was introduced somewhere between 3.1 final and 3.2-rc1. I've tried to track down the bad commit but so far I haven't been successfull. The only workaround I found is to set the idle boot parameter to mwait and to disable CONFIG_INTEL_IDLE. Somehow the kernel seems to set idle to poll by default - no idea why :-(. I wonder if other mobile sandy bridge cpus are affected as well.

There are only a few changesets affecting drivers/idle

Can you try to git revert

46bcfad7a819bd17ac4e831b04405152d59784ab
4202735e8ab6ecfb0381631a0d0b58fefe0bd4e2
e978aa7d7d57d04eb5f88a7507c4fb98577def77

?

-Andi

#2 lore raw
Quoted 9 lines · Andi Kleen
Quoted 3 lines · Nicolas Kalkhof

Andi,

correct. The issue was introduced somewhere between 3.1 final and 3.2-rc1. I've tried to track down the bad commit but so far I haven't been successfull. The only workaround I found is to set the idle boot parameter to mwait and to disable CONFIG_INTEL_IDLE. Somehow the kernel seems to set idle to poll by default - no idea why :-(. I wonder if other mobile sandy bridge cpus are affected as well.

There are only a few changesets affecting drivers/idle

Can you try to git revert

Quoted 7 lines

46bcfad7a819bd17ac4e831b04405152d59784ab
4202735e8ab6ecfb0381631a0d0b58fefe0bd4e2
e978aa7d7d57d04eb5f88a7507c4fb98577def77

?

-Andi

Hi,

I am guessing that for workaround just setting the idle boot parameter to mwait should work. Have u tested this without disabling CONFIG_INTEL_IDLE ?

Also one thing to note from your logs is, cpu is not entering any of the other C states except for POLL. Can you also check if your are enabling only POLL for idle. I am able to enter deeper C-states on i5-2540M with Intel_Idle enabled.

Power top o/p:
-------------

            Core    |             CPU 0               CPU 1
POLL        0.0%    | POLL        0.0%    0.0 ms      0.0%    0.0 ms
C1-SNB      1.4%    | C1-SNB      2.2%    0.7 ms      0.5%    0.9 ms
C3-SNB      0.9%    | C3-SNB      1.6%    1.2 ms      0.1%    1.2 ms
C6-SNB      0.2%    | C6-SNB      0.3%    1.6 ms      0.0%    0.9 ms
C7-SNB     95.8%    | C7-SNB     93.7%    3.7 ms     97.8%   15.1 ms

Regards,
Deepthi

#3 lore raw
Quoted 22 lines · Deepthi Dharwar
Quoted 9 lines · Andi Kleen
Quoted 3 lines · Nicolas Kalkhof

Andi,

correct. The issue was introduced somewhere between 3.1 final and 3.2-rc1. I've tried to track down the bad commit but so far I haven't been successfull. The only workaround I found is to set the idle boot parameter to mwait and to disable CONFIG_INTEL_IDLE. Somehow the kernel seems to set idle to poll by default - no idea why :-(. I wonder if other mobile sandy bridge cpus are affected as well.

There are only a few changesets affecting drivers/idle

Can you try to git revert

Quoted 7 lines

46bcfad7a819bd17ac4e831b04405152d59784ab
4202735e8ab6ecfb0381631a0d0b58fefe0bd4e2
e978aa7d7d57d04eb5f88a7507c4fb98577def77

?

-Andi

Hi,

Quoted 8 lines

I am guessing that for workaround just setting the idle boot parameter to mwait should work. Have u tested this without disabling CONFIG_INTEL_IDLE ?

Also one thing to note from your logs is, cpu is not entering any of the other C states except for POLL. Can you also check if your are enabling only POLL for idle. I am able to enter deeper C-states on i5-2540M with Intel_Idle enabled.

Seems I accidentally found a way to reproduce, boot with nohz=off. Q6600 box can be booted nohz=off without ending up as if I'd booted idle=poll, but boxen using intel_idle() end up using poll_idle().

-Mike