regression in 2.6.35.4 'load is to heavy (video subsystem?)'

regression in 2.6.35.4 'load is to heavy (video subsystem?)'

[Karsten Mehrhoff]

Using the same .config from 2.6.35.3 to compile 2.5.36.4 results in a  
heavy load with 2.6.35.4.

Example:

Difference between 2.6.35.1/2/3 and 2.6.35.4 while watching some videos:
2.6.35.4 switches the cpu for flash videos in the browser (opera or  
iceweasel) or other video outputs to 2200/2400/2600 MHz meanwhile 2.6.35.3  
(or older) stays at 1000 Mhz. That results in a higher cpu temperature,  
more power consumption and so one.

Using other GUI program results in nearly the same problems with 2.6.35.4,  
so this kernel is unusable for me.

Results to see the difference for the same action
2.6.35.4
Core0 Temp:  +45.0°C
Core1 Temp:  +43.0°C
cpu MHz:	2200.000 or higher

2.6.35.3
Core0 Temp:  +32.0°C
Core1 Temp:  +31.0°C
cpu MHz:	1000.000 (max. 1800, but falling back to 1000)

kernel compiled with 'CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND=y'
results for me in 1000, 1800, 2000, 2200, 2400, 2600 MHz.

I'm not the only one with this problem, other users experienced the same  
behavior on other systems on 386 systems, i.e. a regression for glxgears  
about 30% on slower systems. We all uses differnet AMD cpus and nNida  
graphic controllers. Same results for the nvidia-kernel from the repos or  
the nVidia driver from nvidia.com.

There must something be wrong in the video subsystem, which is causing  
this regression.

My system (overview using 2.5.35.3):
=====================
Processor:	2x AMD Athlon(tm) 64 X2 Dual Core Processor 5000+
Memory:		 4060MB

Display
Resolution:       1920x1080 pixels
OpenGL Renderer:  GeForce 9500 GT/PCI/SSE2
X11 Vendor:	      The X.Org Foundation
Version:          1.7.7

Version
Kernel:			Linux 2.6.35.3-kmt (x86_64)
Compiled:	SMP Mon Aug 23 00:58:37 CEST 2010
C Library:      GNU C Library version 2.11.2 (stable)
Default C Compiler: GNU C Compiler version 4.4.5 20100824 (prerelease)  
(Debian 4.4.4-11)
Distribution:	Debian GNU/Linux squeeze/sid

OpenGL
Vendor:        NVIDIA Corporation
Renderer:      GeForce 9500 GT/PCI/SSE2
Version:       3.3.0 NVIDIA 256.44
Direct Rendering: Yes


Regards
Karsten
--

Re: regression in 2.6.35.4 'load is to heavy (video subsystem?)'

[Andrew Morton]

On Mon, 30 Aug 2010 10:02:36 +0200
Karsten Mehrhoff <kawime@gmx.de> wrote:

> Using the same .config from 2.6.35.3 to compile 2.5.36.4 results in a  
> heavy load with 2.6.35.4.

A regression within -stable is rather bad.

> Example:
> 
> Difference between 2.6.35.1/2/3 and 2.6.35.4 while watching some videos:
> 2.6.35.4 switches the cpu for flash videos in the browser (opera or  
> iceweasel) or other video outputs to 2200/2400/2600 MHz meanwhile 2.6.35.3  
> (or older) stays at 1000 Mhz. That results in a higher cpu temperature,  
> more power consumption and so one.
> 
> Using other GUI program results in nearly the same problems with 2.6.35.4,  
> so this kernel is unusable for me.
> 
> Results to see the difference for the same action
> 2.6.35.4
> Core0 Temp:  +45.0__C
> Core1 Temp:  +43.0__C
> cpu MHz:	2200.000 or higher
> 
> 2.6.35.3
> Core0 Temp:  +32.0__C
> Core1 Temp:  +31.0__C
> cpu MHz:	1000.000 (max. 1800, but falling back to 1000)
> 
> kernel compiled with 'CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND=y'
> results for me in 1000, 1800, 2000, 2200, 2400, 2600 MHz.
> 
> I'm not the only one with this problem, other users experienced the same  
> behavior on other systems on 386 systems, i.e. a regression for glxgears  
> about 30% on slower systems. We all uses differnet AMD cpus and nNida  
> graphic controllers. Same results for the nvidia-kernel from the repos or  
> the nVidia driver from nvidia.com.
> 
> There must something be wrong in the video subsystem, which is causing  
> this regression.
> 
> My system (overview using 2.5.35.3):
> =====================
> Processor:	2x AMD Athlon(tm) 64 X2 Dual Core Processor 5000+
> Memory:		 4060MB
> 
> Display
> Resolution:       1920x1080 pixels
> OpenGL Renderer:  GeForce 9500 GT/PCI/SSE2
> X11 Vendor:	      The X.Org Foundation
> Version:          1.7.7
> 
> Version
> Kernel:			Linux 2.6.35.3-kmt (x86_64)
> Compiled:	SMP Mon Aug 23 00:58:37 CEST 2010
> C Library:      GNU C Library version 2.11.2 (stable)
> Default C Compiler: GNU C Compiler version 4.4.5 20100824 (prerelease)  
> (Debian 4.4.4-11)
> Distribution:	Debian GNU/Linux squeeze/sid
> 
> OpenGL
> Vendor:        NVIDIA Corporation
> Renderer:      GeForce 9500 GT/PCI/SSE2
> Version:       3.3.0 NVIDIA 256.44
> Direct Rendering: Yes

I'm not seeing any relevant cpufreq changes in 2.6.35.3 -> 2.6.35.4
(ftp://ftp.kernel.org/pub/linux/kernel/v2.6/ChangeLog-2.6.35.4).  There
were a lot of DRM changes.

Re: regression in 2.6.35.4 'load is to heavy (video subsystem?)'

[Andrew Morton]

On Mon, 30 Aug 2010 10:02:36 +0200
Karsten Mehrhoff <kawime@gmx.de> wrote:

> Using the same .config from 2.6.35.3 to compile 2.5.36.4 results in a  
> heavy load with 2.6.35.4.

A regression within -stable is rather bad.

> Example:
> 
> Difference between 2.6.35.1/2/3 and 2.6.35.4 while watching some videos:
> 2.6.35.4 switches the cpu for flash videos in the browser (opera or  
> iceweasel) or other video outputs to 2200/2400/2600 MHz meanwhile 2.6.35.3  
> (or older) stays at 1000 Mhz. That results in a higher cpu temperature,  
> more power consumption and so one.
> 
> Using other GUI program results in nearly the same problems with 2.6.35.4,  
> so this kernel is unusable for me.
> 
> Results to see the difference for the same action
> 2.6.35.4
> Core0 Temp:  +45.0__C
> Core1 Temp:  +43.0__C
> cpu MHz:	2200.000 or higher
> 
> 2.6.35.3
> Core0 Temp:  +32.0__C
> Core1 Temp:  +31.0__C
> cpu MHz:	1000.000 (max. 1800, but falling back to 1000)
> 
> kernel compiled with 'CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND=y'
> results for me in 1000, 1800, 2000, 2200, 2400, 2600 MHz.
> 
> I'm not the only one with this problem, other users experienced the same  
> behavior on other systems on 386 systems, i.e. a regression for glxgears  
> about 30% on slower systems. We all uses differnet AMD cpus and nNida  
> graphic controllers. Same results for the nvidia-kernel from the repos or  
> the nVidia driver from nvidia.com.
> 
> There must something be wrong in the video subsystem, which is causing  
> this regression.
> 
> My system (overview using 2.5.35.3):
> =====================
> Processor:	2x AMD Athlon(tm) 64 X2 Dual Core Processor 5000+
> Memory:		 4060MB
> 
> Display
> Resolution:       1920x1080 pixels
> OpenGL Renderer:  GeForce 9500 GT/PCI/SSE2
> X11 Vendor:	      The X.Org Foundation
> Version:          1.7.7
> 
> Version
> Kernel:			Linux 2.6.35.3-kmt (x86_64)
> Compiled:	SMP Mon Aug 23 00:58:37 CEST 2010
> C Library:      GNU C Library version 2.11.2 (stable)
> Default C Compiler: GNU C Compiler version 4.4.5 20100824 (prerelease)  
> (Debian 4.4.4-11)
> Distribution:	Debian GNU/Linux squeeze/sid
> 
> OpenGL
> Vendor:        NVIDIA Corporation
> Renderer:      GeForce 9500 GT/PCI/SSE2
> Version:       3.3.0 NVIDIA 256.44
> Direct Rendering: Yes

I'm not seeing any relevant cpufreq changes in 2.6.35.3 -> 2.6.35.4
(ftp://ftp.kernel.org/pub/linux/kernel/v2.6/ChangeLog-2.6.35.4).  There
were a lot of DRM changes.

Re: regression in 2.6.35.4 'load is to heavy (video subsystem?)'

[Dave Airlie]

On Tue, 2010-09-21 at 16:50 -0700, Andrew Morton wrote:
> On Mon, 30 Aug 2010 10:02:36 +0200
> Karsten Mehrhoff <kawime@gmx.de> wrote:
> 
> > Using the same .config from 2.6.35.3 to compile 2.5.36.4 results in a  
> > heavy load with 2.6.35.4.
> 
> A regression within -stable is rather bad.
> 
> > Example:
> > 
> > Difference between 2.6.35.1/2/3 and 2.6.35.4 while watching some videos:
> > 2.6.35.4 switches the cpu for flash videos in the browser (opera or  
> > iceweasel) or other video outputs to 2200/2400/2600 MHz meanwhile 2.6.35.3  
> > (or older) stays at 1000 Mhz. That results in a higher cpu temperature,  
> > more power consumption and so one.
> > 
> > Using other GUI program results in nearly the same problems with 2.6.35.4,  
> > so this kernel is unusable for me.
> > 
> > Results to see the difference for the same action
> > 2.6.35.4
> > Core0 Temp:  +45.0__C
> > Core1 Temp:  +43.0__C
> > cpu MHz:	2200.000 or higher
> > 
> > 2.6.35.3
> > Core0 Temp:  +32.0__C
> > Core1 Temp:  +31.0__C
> > cpu MHz:	1000.000 (max. 1800, but falling back to 1000)
> > 
> > kernel compiled with 'CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND=y'
> > results for me in 1000, 1800, 2000, 2200, 2400, 2600 MHz.
> > 
> > I'm not the only one with this problem, other users experienced the same  
> > behavior on other systems on 386 systems, i.e. a regression for glxgears  
> > about 30% on slower systems. We all uses differnet AMD cpus and nNida  
> > graphic controllers. Same results for the nvidia-kernel from the repos or  
> > the nVidia driver from nvidia.com.
> > 
> > There must something be wrong in the video subsystem, which is causing  
> > this regression.
> > 
> > My system (overview using 2.5.35.3):
> > =====================
> > Processor:	2x AMD Athlon(tm) 64 X2 Dual Core Processor 5000+
> > Memory:		 4060MB
> > 
> > Display
> > Resolution:       1920x1080 pixels
> > OpenGL Renderer:  GeForce 9500 GT/PCI/SSE2
> > X11 Vendor:	      The X.Org Foundation
> > Version:          1.7.7
> > 
> > Version
> > Kernel:			Linux 2.6.35.3-kmt (x86_64)
> > Compiled:	SMP Mon Aug 23 00:58:37 CEST 2010
> > C Library:      GNU C Library version 2.11.2 (stable)
> > Default C Compiler: GNU C Compiler version 4.4.5 20100824 (prerelease)  
> > (Debian 4.4.4-11)
> > Distribution:	Debian GNU/Linux squeeze/sid
> > 
> > OpenGL
> > Vendor:        NVIDIA Corporation
> > Renderer:      GeForce 9500 GT/PCI/SSE2
> > Version:       3.3.0 NVIDIA 256.44
> > Direct Rendering: Yes
> 
> I'm not seeing any relevant cpufreq changes in 2.6.35.3 -> 2.6.35.4
> (ftp://ftp.kernel.org/pub/linux/kernel/v2.6/ChangeLog-2.6.35.4).  There
> were a lot of DRM changes.

No us, he's using a binary driver.

Dave.

Re: [stable] regression in 2.6.35.4 'load is to heavy (video subsystem?)'

[Greg KH]

On Tue, Sep 21, 2010 at 04:50:07PM -0700, Andrew Morton wrote:
> On Mon, 30 Aug 2010 10:02:36 +0200
> Karsten Mehrhoff <kawime@gmx.de> wrote:
> 
> > Using the same .config from 2.6.35.3 to compile 2.5.36.4 results in a  
> > heavy load with 2.6.35.4.
> 
> A regression within -stable is rather bad.

Agreed.

> > Example:
> > 
> > Difference between 2.6.35.1/2/3 and 2.6.35.4 while watching some videos:
> > 2.6.35.4 switches the cpu for flash videos in the browser (opera or  
> > iceweasel) or other video outputs to 2200/2400/2600 MHz meanwhile 2.6.35.3  
> > (or older) stays at 1000 Mhz. That results in a higher cpu temperature,  
> > more power consumption and so one.
> > 
> > Using other GUI program results in nearly the same problems with 2.6.35.4,  
> > so this kernel is unusable for me.
> > 
> > Results to see the difference for the same action
> > 2.6.35.4
> > Core0 Temp:  +45.0__C
> > Core1 Temp:  +43.0__C
> > cpu MHz:	2200.000 or higher
> > 
> > 2.6.35.3
> > Core0 Temp:  +32.0__C
> > Core1 Temp:  +31.0__C
> > cpu MHz:	1000.000 (max. 1800, but falling back to 1000)

Can you run 'git bisect' between 2.6.35.3 and 2.6.35.4 to try to find
out the offending patch that caused this issue?

thanks,

greg k-h

Re: [stable] regression in 2.6.35.4 'load is to heavy (video subsystem?)'

[Karsten Mehrhoff]

[Am 22.09.2010, 02:01 Uhr, schrieb Greg KH <greg@kroah.com>]

> On Tue, Sep 21, 2010 at 04:50:07PM -0700, Andrew Morton wrote:
>> On Mon, 30 Aug 2010 10:02:36 +0200
>> Karsten Mehrhoff <kawime@gmx.de> wrote:
>>
>>> Using the same .config from 2.6.35.3 to compile 2.5.36.4 results in a
>>> heavy load with 2.6.35.4.
>>
>> A regression within -stable is rather bad.
>
> Agreed.
>
>>> Example:
>>>
>>> Difference between 2.6.35.1/2/3 and 2.6.35.4 while watching some  
>> videos:
>>> 2.6.35.4 switches the cpu for flash videos in the browser (opera or
>>> iceweasel) or other video outputs to 2200/2400/2600 MHz meanwhile  
>> 2.6.35.3
>>> (or older) stays at 1000 Mhz. That results in a higher cpu  
>> temperature,
>>> more power consumption and so one.
>>>
>>> Using other GUI program results in nearly the same problems with  
>> 2.6.35.4,
>>> so this kernel is unusable for me.
>>>
>>> Results to see the difference for the same action
>>> 2.6.35.4
>>> Core0 Temp:  +45.0__C
>>> Core1 Temp:  +43.0__C
>>> cpu MHz:	2200.000 or higher
>>>
>>> 2.6.35.3
>>> Core0 Temp:  +32.0__C
>>> Core1 Temp:  +31.0__C
>>> cpu MHz:	1000.000 (max. 1800, but falling back to 1000)
>
> Can you run 'git bisect' between 2.6.35.3 and 2.6.35.4 to try to find
> out the offending patch that caused this issue?
>
> thanks,
>
> greg k-h

Same for 2.6.35.5 using 256.53

For your info, I did run some tests today using a nVidia 9500GT

    Kernel       |   Performance with NVIDIA-Linux-x86_64-
                 |       256.53      |    260.19.04 (BETA)
----------------------------------------------------------
  2.6.35.3       |      good         |      good
----------------------------------------------------------
  2.6.35.4       |       bad         |     not tested
----------------------------------------------------------
  2.6.35.5       |       bad         |      ~ good
----------------------------------------------------------
  2.6.36-rc4     |                   |
    (git1-5)     |     failed        |      good
----------------------------------------------------------
  2.6.36-rc4     |                   |
    (git1)       |     failed        |      good
----------------------------------------------------------


Karsten

Re: [stable] regression in 2.6.35.4 'load is to heavy (video subsystem?)'

[Florian Mickler]

On Wed, 22 Sep 2010 11:42:41 +0200
Karsten Mehrhoff <kawime@gmx.de> wrote:

> [Am 22.09.2010, 02:01 Uhr, schrieb Greg KH <greg@kroah.com>]
> 

> >>> Difference between 2.6.35.1/2/3 and 2.6.35.4 while watching some  
> >> videos:
> >>> 2.6.35.4 switches the cpu for flash videos in the browser (opera or
> >>> iceweasel) or other video outputs to 2200/2400/2600 MHz meanwhile  
> >> 2.6.35.3
> >>> (or older) stays at 1000 Mhz. That results in a higher cpu  
> >> temperature,
> >>> more power consumption and so one.
> >>>
> >>> Using other GUI program results in nearly the same problems with  
> >> 2.6.35.4,
> >>> so this kernel is unusable for me.
> >>>
> >>> Results to see the difference for the same action
> >>> 2.6.35.4
> >>> Core0 Temp:  +45.0__C
> >>> Core1 Temp:  +43.0__C
> >>> cpu MHz:	2200.000 or higher
> >>>
> >>> 2.6.35.3
> >>> Core0 Temp:  +32.0__C
> >>> Core1 Temp:  +31.0__C
> >>> cpu MHz:	1000.000 (max. 1800, but falling back to 1000)
> >
> > Can you run 'git bisect' between 2.6.35.3 and 2.6.35.4 to try to find
> > out the offending patch that caused this issue?
> >
> > thanks,
> >
> > greg k-h
> 
> Same for 2.6.35.5 using 256.53
> 
> For your info, I did run some tests today using a nVidia 9500GT
> 
>     Kernel       |   Performance with NVIDIA-Linux-x86_64-
>                  |       256.53      |    260.19.04 (BETA)
> ----------------------------------------------------------
>   2.6.35.3       |      good         |      good
> ----------------------------------------------------------
>   2.6.35.4       |       bad         |     not tested
> ----------------------------------------------------------

What would be tremendously more interesting is if you can
reproduce the issue with the in-kernel nouveau driver or just without
the binary driver. As we have no sourcecode for the binary driver, we
can not tell what it does and are thus unable to debug any issues.

If it is an issue that is not reproducible without the binary driver,
please contact the vendor of that driver. 

It it is reproducible even without that driver, it would help if you
could tell exactly which patch in 2.6.35.4 makes the difference
between good and bad in your test below.

There are 114 patches between 2.6.35.3 and 2.6.35.4. If you test
between them, you can pinpoint the exact patch with about 7 tests. 

git bisect does this for you, just do 

$ git clone git://git.kernel.org/pub/scm/linux/kernel/git/stable/linux-2.6.35.y.git
$ cd linux-2.6.35.y 
$ git bisect bad v2.6.35.4
$ git bisect good v2.6.35.3

Git then checks out a testcandidate for you, which you should compile
and test. If it's good , type 
$ git bisect good
, if its bad
$ git bisect bad

If all goes well, after about 7 tests, it will tell you
what patch did introduce the regression.

Regards,
Flo
 
> 
> Karsten