Re: [REGRESSION] [Linux 3.2] top/htop and all other CPU usage metering applications has gone crackers

[Michal Hocko <mhocko@suse.cz>]

[-- Attachment #1: Type: text/plain, Size: 3182 bytes --]

On Mon 28-11-11 22:41:25, Michal Hocko wrote:
> Hi,
> 
> On Mon 28-11-11 21:19:26, Rafael J. Wysocki wrote:
> > On Monday, November 28, 2011, Tino Keitel wrote:
> > > On Sun, Nov 27, 2011 at 12:45:57 +0100, Rafael J. Wysocki wrote:
> > > > On Sunday, November 27, 2011, Tino Keitel wrote:
> > > > > On Thu, Nov 24, 2011 at 21:05:53 +0100, Tino Keitel wrote:
> > > > > > On Thu, Nov 24, 2011 at 10:30:15 +0000, Artem S. Tashkinov wrote:
> > > > > > > Hello,
> > > > > > > 
> > > > > > > I'd like to report a weird regression in Linux 3.2 (running rc3 now) - all CPU metering applications have gone terribly mad
> > > > > > > under this kernel:
> > > > > > 
> > > > > > I get the same using top, htop and the gnome system monitor with kernel
> > > > > > 3.2 on a Sandy Bridge quad core box, running Debian unstable.
> > > > > 
> > > > > I just tested 3.2-rc2, and see the same bug.
> > > > 
> > > > I'm seeing that too on one of my test boxes, but not all the time
> > > > (i.e. there are periods in which the readings are correct).  The other boxes
> > > > I've tested with 3.2-rc are fine in that respect.
> > > > 
> > > > Also, it seems that it shows 100%-(real load) when it is wrong.  So, it looks
> > > > like there's an overflow somewhere in the CPU load measuring code, at least
> > > > on some CPUs.
> > > 
> > > Hi,
> > > 
> > > I reverted this commit and so far it looks good:
> > > 
> > > commit a25cac5198d4ff2842ccca63b423962848ad24b2
> > > Author: Michal Hocko <mhocko@suse.cz>
> > > Date:   Wed Aug 24 09:40:25 2011 +0200
> > > 
> > >     proc: Consider NO_HZ when printing idle and iowait times
> > > 
> > > I'll report back tomorrow how the kernel behaves.
> > 
> > Hmm.  Michal, can you have a look at that, please?
> 
> Hmm, my testing didn't show anything like that. Could you post
> cat /proc/stat collected every second during 30s or so?
> 
> Here is the output of my run with 3.2.0-rc3-00004-gdd38d29 and the attached config:
> for i in `seq 30`; 
> do 
> 	cat /proc/stat > `date +'%s'`
> 	sleep 1
> done
> export old_user=0 old_nice=0 old_sys=0 old_idle=0 old_iowait=0; 
> grep cpu0 * | while read cpu user nice sys idle iowait rest; 
> do 
> 	echo $cpu $(($user-$old_user)) $(($nice-$old_nice)) $(($sys-$old_sys)) $(($idle-$old_idle)) $(($iowait-$old_iowait))
> 	old_user=$user old_nice=$nice old_sys=$sys old_idle=$idle old_iowait=$iowait
> done
> 

Same results (attached) x86_64 with AMD 16CPUs in my lab with a
different cpuidle driver:
grep . -r /sys/devices/system/cpu/cpuidle/
$ /sys/devices/system/cpu/cpuidle/current_driver:none
/sys/devices/system/cpu/cpuidle/current_governor_ro:menu
$ grep . -r /sys/devices/system/cpu/cpufreq/
/sys/devices/system/cpu/cpufreq/ondemand/sampling_rate_min:10000
/sys/devices/system/cpu/cpufreq/ondemand/sampling_rate:38000
/sys/devices/system/cpu/cpufreq/ondemand/up_threshold:40
/sys/devices/system/cpu/cpufreq/ondemand/sampling_down_factor:1
/sys/devices/system/cpu/cpufreq/ondemand/ignore_nice_load:0
/sys/devices/system/cpu/cpufreq/ondemand/powersave_bias:0
/sys/devices/system/cpu/cpufreq/ondemand/io_is_busy:0

-- 
Michal Hocko
SUSE Labs
SUSE LINUX s.r.o.
Lihovarska 1060/12
190 00 Praha 9    
Czech Republic

[-- Attachment #2: config.gz --]
[-- Type: application/octet-stream, Size: 33394 bytes --]

[-- Attachment #3: amd_16cpus.tar.bz2 --]
[-- Type: application/octet-stream, Size: 3958 bytes --]