Kernel 3.6.2 Power Regressions (Sandy Bridge)

cris9288 · 2013-05-13 20:10:26

hrkfdn wrote:

I just built 3.9.2 and I'm still having the problem that after wakeup from suspend the CPU core temperature goes up to 70°C (in idle 40-50°C) with no CPU usage.
Anyone else encountering this? It's so incredibly annoying.. I need to suspend/wakeup again and then it's gone.

I just had it happen right now. I was hoping it would go away after the upgrade to 3.9.

Munksgaard · 2013-05-14 08:17:54

cris9288 wrote:

hrkfdn wrote:
I just built 3.9.2 and I'm still having the problem that after wakeup from suspend the CPU core temperature goes up to 70°C (in idle 40-50°C) with no CPU usage.
Anyone else encountering this? It's so incredibly annoying.. I need to suspend/wakeup again and then it's gone.

I just had it happen right now. I was hoping it would go away after the upgrade to 3.9.

I'm still experiencing the problem as well, running a Thinkpad X220

Right after unsuspending:

$ uname -a
Linux hertz 3.9.2-1-ARCH #1 SMP PREEMPT Sat May 11 20:31:08 CEST 2013 x86_64 GNU/Linux

$ sensors
thinkpad-isa-0000
Adapter: ISA adapter
fan1:        3886 RPM

acpitz-virtual-0
Adapter: Virtual device
temp1:        +77.0°C  (crit = +99.0°C)

coretemp-isa-0000
Adapter: ISA adapter
Physical id 0:  +70.0°C  (high = +86.0°C, crit = +100.0°C)
Core 0:         +70.0°C  (high = +86.0°C, crit = +100.0°C)
Core 1:         +66.0°C  (high = +86.0°C, crit = +100.0°C)

stupidus · 2013-05-14 09:21:24

I have to confirm the previous reports. I just did a suspend/resume cycle and the temperature started increasing right afterwards. i7z reported that the CPU was running will 3GHz (I have a i5 2520M, so it was running with TurboBoost) while all cores where idle. After another suspend/resume it went back to 800 MHz.

EDIT: This is with linux-3.9.2 from [core].

Last edited by stupidus (2013-05-14 09:22:04)

Druedain · 2013-05-14 10:58:24

Hi!

As I can see another missing thing with intel_pstate is „memory” o_O . I have to manually put this

cpupower frequency-set -g performance

after every reboot…

I tried to find any discussion about this in the internet, but I find only discussions about other missing features (like checking actual cpu frequency). I'm curious why is it so hard to find info about such basic thing when there is so much talk about more technical aspects. Anyone know which kernel/acpi mailing list is the one from which I should start with questions to devs?

For now the only workaround I can imagine is to create script, and load it at boot…

Chais · 2013-05-14 12:42:42

So far the update is working out for me. Predicted runtime 4-5 hrs.
@Druedain: If you're using cpupower only you need to set performance the default governor in

/etc/cpufreq-bench.conf

I reactivated laptop-mode-tools and have it switch the governor according AC/battery mode, which works fine.

Last edited by Chais (2013-05-14 12:43:18)

Druedain · 2013-05-14 12:51:05

@Chais
I'm afraid it's not working…

cris9288 · 2013-05-14 13:43:15

Unless you are dead set on using Intel_pstate Just add "intel_pstate=disabled" to your kernel parameters. This should load acpi_cpufreq As before.

Druedain · 2013-05-14 13:48:23

Hmmm… This won't solve the problem . I don't want to treat this as some people treat pulseaudio or systemd: I don't understand it so I switch it off. I really want know how to use intel_pstate, how to set performance governor permanently, or just where can I ask devs such question and don't be treated as noob that should google it…

Chais · 2013-05-14 14:48:58

It's all your fault. Now I dug deeper and found loads of weird behaviour. I was so happy yesterday that it finally worked...
Hardware is a i5-2450M.
After the wakeup the temp started at a good 50°C but would rise steadily. Asking cpupower for the status I got:

Treiber: intel_pstate
  Folgende CPUs laufen mit der gleichen Hardware-Taktfrequenz: 0
  Die Taktfrequenz folgender CPUs werden per Software koordiniert: 0
  Maximale Dauer eines Taktfrequenzwechsels: 0.97 ms.
  Hardwarebedingte Grenzen der Taktfrequenz: 800 MHz - 3.10 GHz
  mögliche Regler: performance, powersave
  momentane Taktik: die Frequenz soll innerhalb 800 MHz und 3.10 GHz.
                    liegen. Der Regler "performance" kann frei entscheiden,
                    welche Taktfrequenz innerhalb dieser Grenze verwendet wird.
  boost state support:
    Supported: yes
    Active: yes
    25500 MHz max turbo 4 active cores
    25500 MHz max turbo 3 active cores
    25500 MHz max turbo 2 active cores
    25500 MHz max turbo 1 active cores

You'll notice that it lacks an actual current freq.

cpupower -c all frequency-info -f

also wouldn't get me any info.
Since the machine was getting warmer I switched to the powersave governor. Now shit got weirder. Temperature kept rising and is currently at 76°C while "powersaving". Taking a look into

/sys/devices/system/cpu/cpu0/cpufreq/

I found the file scaling_cur_freq to be missing. That's probably why cpupower is unable to determine the frequency. Looking through the files I found something weird:

sudo cat /sys/devices/system/cpu/cpu{0..3}/cpufreq/scaling_min_freq 
800000
3100000
3100000
3100000

So only the first thread of the first core is clocked to 800MHz, the others run at full throttle. No wonder this thing is boiling. So how to set cpupower the correct values? I couldn't find anything in the man page or wiki.
I'll report back after another sleep/wake-cycle.

EDIT:
After another cycle all threads run at 800MHz, cpupower is able to read the current frequency and the box is cooling down.

sudo cat /sys/devices/system/cpu/cpu{0..3}/cpufreq/cpuinfo_cur_freq
800000
800000
800000
800000

Weird...

Last edited by Chais (2013-05-14 14:57:07)

Druedain · 2013-05-14 14:55:40

@Chais
The only thing I can tell you now after digging into mailing lists is that, we are not the only who suffer from such bugs, and devs now tries to patch intel_pstate. Hopefully bugs will be fixed shortly…

stupidus · 2013-05-14 14:58:05

Are you running cpupower as root? It does not give me any information either, if I'm running it as a normal user. I need to run it as root. Anyway, you could try i7z, it's supposed to give more accurate information about your CPU frequency for Sandy Bridge hardware (need to run it as root aswell).

And what you observed is actually similar to what I've seen, that the CPU goes into full Turbo Boost mode.

Druedain · 2013-05-14 15:41:09

This is quite strange ._.

From what I've observed:
– if after boot / suspend is set powersafe governor then switch to performance works fine,
– if after first suspend is set performance governor then cpu overheats / at max frequ,
– when after first suspend governor wasn't change, then after third suspend again it works good.
– when it finally works after third suspend it again break ._.

P state is definitely not ready for use. I though that it only need more love and understanding, but now it seems that it's just broken…

_____

I send an e-mail to one guy that wrote about using P state. We will see if it was the right person to ask about it. I'll tell you about it when / if I get useful reply.

Last edited by Druedain (2013-05-14 16:58:48)

cfr · 2013-05-14 23:23:07

It seems to work reasonably well for me except that I need sudo to get the current frequency (so cpupower doesn't get it at least for me as a normal user).

I'm wondering if it is relevant that turbo boost is reported as unsupported on my machine.

Or if I'm just yet to be bit.

I *did* see my temps rocket on one occasion but that was before intel_pstate and it only happened once. (I rebooted and all was well.)

Druedain · 2013-05-14 23:48:45

But if you
are building your own kernel, under the CPU Frequency scaling menu
in menuconfig, just above where you choose which drivers to build is
a setting for
Default CPUFreq governor (currentname) --->
Key 'enter' on that and the performance governor should be top of
the list.

If this is the only way to set the default cpu governor, then… Intel, WTF have you done??? o_O

Thaodan · 2013-05-15 21:42:07

Its not instells fault, its the fault of the build settting. An option would be to supply the governess as module or at least shift the intelp state governor as module.

graysky · 2013-05-17 19:39:37

Earlier in this thread, there was some confusion over the pstate driver working with just sandy bridge CPUs. This thread shows a link to a simple patch to allow ivybridge chips to use the pstate driver. Running it here on my i7-3770k.

Last edited by graysky (2013-05-17 19:40:17)

Thaodan · 2013-05-18 01:31:14

Do you added it to linux-ck?

graysky · 2013-05-18 06:34:21

Thaodan wrote:

Do you added it to linux-ck?

https://aur.archlinux.org/packages/Linux-ck/
See my comment from 2013-05-17 20:02

stupidus · 2013-05-25 09:34:12

I just installed linux-3.9.3 from [core]. The behaviour of the pstate driver changed a bit compared to 3.9.2. Back then the frequency of the CPU was usually at minimum (800 MHz), now it stays at a higher level (~1.7 GHz). But because it is still idle most of the time, the power consumption does not change.

But the suspend/resume bug is still not fixed. Sometimes, it will go into full turbo boost mode after a suspend/resume cycle. Another suspend/resume usually fixes this. This behaviour is quite annoying, as I have to check the CPU frequency after every resume. (Or just check the temperature, that is also a good indicator. Anyhow, this bug is really annoying)

donniezazen · 2013-05-26 15:14:55

@stupidus I have been noticing the same stuff. It seems my CPU not runs over 2GHz sometimes even 3.4GHz for no obvious reasons. My battery life has taken a narrow dive to let's say about 2 hours. Also I have been noticing suspend/resume bug quite too often on 3.9.3-1-ARCH.

cfr · 2013-05-26 16:25:02

Does the bug go away if you disable turbo boost? I'm not suggesting this as a solution but just to try to narrow the issue down. The reason I ask is that I have only seen the suspend/resume bug maybe once or twice and only before the new pstate driver. But turbo boost isn't supported on my hardware and I'm wondering if that is why.

firewalker · 2013-05-26 16:42:50

I use intel_pstate=disable and I have the same issues after suspend to RAM.

https://bbs.archlinux.org/viewtopic.php?id=163854

Has anyone reported this as a bug upstream?

Last edited by firewalker (2013-05-26 16:44:24)

donniezazen · 2013-05-27 08:09:05

@cfr I have no idea. I haven't messed up with turbo boost. I suppose I am not running turbo boost. My arch used to settle down at 800Hz but not it's running at 2.4 to 3.4GHz. Max speed on my system should be 2.7GHz.

Svenstaro · 2013-05-27 09:04:44

Be advised that pstate is actually usable in 3.9.4. Also, cpu frequency is not meant to stay at its lowest possible frequency at all times by design. Intel engineers determined that frequency is not actually what draws all the power and that continually waking up the processor to see whether a change in frequency is in order actually drains more power than clocking it a little higher to begin with. As far as I can see, pstate works as designed in 3.9.4.

firewalker · 2013-05-27 09:23:22

So, someone should fill a bug report?

Arch Linux

#451 2013-05-13 20:10:26

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#452 2013-05-14 08:17:54

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#453 2013-05-14 09:21:24

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#454 2013-05-14 10:58:24

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#455 2013-05-14 12:42:42

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#456 2013-05-14 12:51:05

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#457 2013-05-14 13:43:15

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#458 2013-05-14 13:48:23

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#459 2013-05-14 14:48:58

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#460 2013-05-14 14:55:40

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#461 2013-05-14 14:58:05

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#462 2013-05-14 15:41:09

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#463 2013-05-14 23:23:07

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#464 2013-05-14 23:48:45

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#465 2013-05-15 21:42:07

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#466 2013-05-17 19:39:37

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#467 2013-05-18 01:31:14

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#468 2013-05-18 06:34:21

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#469 2013-05-25 09:34:12

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#470 2013-05-26 15:14:55

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#471 2013-05-26 16:25:02

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#472 2013-05-26 16:42:50

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#473 2013-05-27 08:09:05

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#474 2013-05-27 09:04:44

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

#475 2013-05-27 09:23:22

Re: Kernel 3.6.2 Power Regressions (Sandy Bridge)

Board footer