Date:	Wed, 10 Feb 2016 16:12:31 +0100
From:	Karsten Malcher <debian@...e.decotrain.de>
To:	Ken Moffat <zarniwhoop@...world.com>
CC:	linux-kernel@...r.kernel.org
Subject: Re: Freezing system after kernel 3.2

Hello Ken,

Am 09.02.2016 um 20:42 schrieb Ken Moffat:
> On Tue, Feb 09, 2016 at 05:03:57PM +0100, Karsten Malcher wrote:
>
>> I have to found out that freezing can occur under kernel 3.2 too, but far less common.
>> So the interesting question is why in newer kernels this will occur very often?
>>
> An interesting question, but almost impossible to track down (I've
> had a problem myself where the "is it ok?" question could not be
> reliably answered - al you can do is find that some kernel versions
> (or some stable revisions) seem worse / less bad.

In my case it seems clear that something is wrong with the frequency scaling.
I can see with cpufreq-info:
driver: acpi-cpufreq
profil: "ondemand"
actual: 800 MHz

You find some good explanations here:
https://wiki.archlinux.org/index.php/CPU_frequency_scaling#CPU_frequency_driver

acpi-cpufreq: "CPUFreq driver which utilizes the ACPI Processor Performance States. "
So the cpu is not touched direct through the driver - only over the ACPI.
ACPI is a BIOS function - so it is clear that bugs in the BIOS will cause the problem there!


>> I could found a solution for the problem in the linked Blog.
>> When you disable Cool' n' Quiet the system is running stable with newer kernel too!
>>
> Ouch.  Full heat, higher power bills.

Yes - that's not nice!

>
>> So it seems to be a buggy BIOS that is causing 2 big problems:
>> 1. RAM can only be used with 3264 MB of 4096 MB.
>> 2. When Cool' n' Quiet is enabled the system is freezing within an minute.
>>
>> I have reported this to Asrock, but i don't think they will do anything for an older mainboard ...
>>
>> You can found several problems regarding freezing with the RS480 chip,
>> but not together with the frequency scaling (Cool'n'Quiet).
>> This problems occur with other chipsets too.
>>
>> How the frequency scaling is working?
>> I have found some hints about disabling C1 and C6 states will help to solve the problem:
>>
>> https://forum.teksyndicate.com/t/amd-system-keeps-freezing/78380/17
>> https://rog.asus.com/forum/showthread.php?60224-amd-system-keeps-freezing&s=03842b8ef2cb1a25bc458b6fe56c9213&p=492388&viewfull=1#post492388
>> https://community.amd.com/message/2645600#2645600
>>
>> But i have no settings about this states in my BIOS.
> Google found a post (in chinese, I think) which mentioned
>
> echo 1 > /sys/devices/system/cpu/cpuX/cpuidle/stateY/disable
>
> where X and Y should be replaced by the cpu number and the state
> number.  I would be very reluctant to try that, particularly for C1
> because that is the standard halt state and power usage might go up
> significantly.

I will have a closer look at this.
But when all functionality is going though the BIOS i am not hopefully.


>
>> Has the behaviour of the frequency scaling drivers changed somehow?
> The behaviour of the cpufreq drivers has definitely changed over the
> years.  When I bought my phenom x4 I also bought an i3 SandyBridge
> intel.  At that time (2012) both were running the ondemand governor
> and the phenom was faster in all my compiles (I build
> linuxfromscratch and work on beyond linuxfromscratch, as well as
> trying to keep an eye on test kernels in case there are problems for
> my hardware).

For interest i have a look at the sources of the acpi-driver.
But it is not understandable without background knowledge of the details. :-)

As i know there is an individual table stored in the cpu that keeps the possible settings
for voltage, multiplier and states.
This should be read and managed through the BIOS and give standard ACPI functionality to the OS.
Really complex.

> At some point, the intel governor moved to 'performance' - I changed
> to the altered kernel driver when it was first available (on intels
> 'performance' works well - generally low power except when busy).

But you can switch between the different scaling governors.
https://wiki.archlinux.org/index.php/CPU_frequency_scaling#Scaling_governors

Normally this should fit to the needs.

> A bit later, there were changes affecting AMD - I forget the details,
> I think early K8s were not altered, only K10 and later.  Since then,
> my phenom has been slower than the i3 when doing single-threaded
> compiles.  Some of that might be because jobs deliberately bounce
> around the cores to even out wear and heat, possibly on my AMD
> machines (I've also got a recent A10) there is a cache penalty when
> moving between the processor 'units'.

I read a german article about possible overclocking of Intel CPU's some days ago.
http://www.heise.de/newsticker/meldung/Skylake-Overclocking-Katz-und-Mausspiel-zwischen-Intel-und-Mainboard-Herstellern-3097870.html
Maybe you can translate with google.
There is a possibility with some BIOS and CPU's to make overclocking that is only available with much more expensive
type of CPU's.

> At the risk of pissing you off with tales of new hardware, I suspect
> there are still changes in the kernel cpufreq area - I bought a
> haswell i7 last week, nominally 3.6 GHz but the specs say it can
> boost a thread to 3.9GHz.  I started with SystemRescueCD and saw the
> frequencies at 3.6 GHz maximum, and often at lower frequencies when
> idle.  Then I installed Mint, I think the frequencies were similar.
> After that I put Fedora23 on it and was shocked to see that all
> cores were often around 3.9GHz - what was odd was that the system
> power draw (PC, monitor, net switch, kvm switch) remained at about
> 90 Watts when idle, up to about 166W when running make -j 8.

I could not find any information what has changed over the time,
but here some interesting Doku:

http://kernel.org/doc/Documentation/cpu-freq/governors.txt
http://kernel.org/doc/Documentation/cpu-freq/cpu-drivers.txt
http://www.pantz.org/software/cpufreq/usingcpufreqonlinux.html
http://unix.stackexchange.com/questions/121410/setting-cpu-governor-to-on-demand-or-conservative

>
> Now that I have installed linuxfromscratch (a binary copy from the
> SandyBridge, then I used that to compile a current version with a
> 4.4.1 kernel) I never see frequencies less than 3.7 GHz reported,
> but the power consumption is nice and low when idle.

For me i was always satisfied with the functionality of the linux frequency drivers.
But it is not fine when i have to miss it.

>
> Summary: over time, everything changes.

And get's more complicated.
Then the time will come that nobody can solve problems when they will occur.

>
> Good luck with taming that machine to a mostly-usable state.  The
> only thing I ever had to do to get a box stable was on an early K8,
> and with two pairs of memory sticks I had to back off the memory
> frequency from the default.  But that was years ago, and the
> then-current version of memtest86 showed the problem in an
> overnight run.  Note that my phenom continues as "from time to time,
> unstable" (some odd crashes, many internal compiler errors).

It's better to have a stable PC that runs at full speed then a PC that is freezing and not usable.
But it is not amazing to search and fix such problems.

Karsten

Powered by blists - more mailing lists