Re: Kernel panic at boot with ondemand governor as default (2.6.24-rc2)

Re: Kernel panic at boot with ondemand governor as default (2.6.24-rc2)

[Thomas Renninger]

On Mon, 2007-11-12 at 15:02 +0100, Eric Piel wrote:
> 11/12/2007 10:16 AM, Thomas Renninger wrote/a écrit:
> > On Sun, 2007-11-11 at 20:10 +0100, Eric Piel wrote:
> >> Hello,
> >>
> >> I've tried kernel 2.6.24-rc2 and I have a problem with the new option 
> >> for setting up the cpufreq ondemand governor as default: a kernel panic 
> >> happens early at boot time. If I boot first with performance governor 
> >> and later change to ondemand, everything is fine (as usual).
> >>
> >> This seems related to the delayed_workqueue. The kernel panic message 
> >> finishes by:
> >> EIP: [<c012aa60>] wq_per_cpu+0x0/0x10 SS:ESP 0068:c191be58
> >> Kernel panic - not syncing: Attempted to kill init!
> >>
> >> A picture of the whole message is available here:
> >> http://pieleric.free.fr/unorder/1162-ondemand-panic.jpg
> >>
> >> My computer is a x86 using speedstep-ich, and you can find the full 
> >> .config leading to the panic attached.
> >>
> >> Does anyone has any idea was it going wrong?
> > 
> > Does it work if you compile speedstep-ich as a module?
> > If you load it, ondemand governor should still be set up automatically.
> > If this works, could it be that the kernel does not like that
> > queue_delayed_work_on is called that early?
> 
> Yes, it works if I compile speedstep-ich as module. I've put the module 
> in initramfs and as soon it is loaded, the ondemand governor is selected 
> and works.
> 
> However, do you think it's ok to prevent cpufreq drivers to be built-in 
> if ondemand governor is selected as default?
> Another way would be to 
> reorganise the initialisation code so that workqueue is initialised 
> before the cpufreq framework is started, do you think it's possible?
Making all this work with low-level drivers built in would be perfect of
course...
If nobody answers the next days, could you open a bug at
bugzilla.kernel.org, so that this does not get lost, pls.
Add myself and Venkatesh to CC...
I try to have a look at it ASAP.
If someone already finds out more, this would be great.

Thanks,

   Thomas

Re: Kernel panic at boot with ondemand governor as default (2.6.24-rc2)

[Dave Jones]

On Mon, Nov 12, 2007 at 04:23:09PM +0100, Thomas Renninger wrote:
 > On Mon, 2007-11-12 at 15:02 +0100, Eric Piel wrote:
 > > 11/12/2007 10:16 AM, Thomas Renninger wrote/a écrit:
 > > > On Sun, 2007-11-11 at 20:10 +0100, Eric Piel wrote:
 > > >> Hello,
 > > >>
 > > >> I've tried kernel 2.6.24-rc2 and I have a problem with the new option 
 > > >> for setting up the cpufreq ondemand governor as default: a kernel panic 
 > > >> happens early at boot time. If I boot first with performance governor 
 > > >> and later change to ondemand, everything is fine (as usual).
 > > >>
 > > >> This seems related to the delayed_workqueue. The kernel panic message 
 > > >> finishes by:
 > > >> EIP: [<c012aa60>] wq_per_cpu+0x0/0x10 SS:ESP 0068:c191be58
 > > >> Kernel panic - not syncing: Attempted to kill init!
 > > >>
 > > >> A picture of the whole message is available here:
 > > >> http://pieleric.free.fr/unorder/1162-ondemand-panic.jpg
 > > >>
 > > >> My computer is a x86 using speedstep-ich, and you can find the full 
 > > >> .config leading to the panic attached.
 > > >>
 > > >> Does anyone has any idea was it going wrong?
 > > > 
 > > > Does it work if you compile speedstep-ich as a module?
 > > > If you load it, ondemand governor should still be set up automatically.
 > > > If this works, could it be that the kernel does not like that
 > > > queue_delayed_work_on is called that early?
 > > 
 > > Yes, it works if I compile speedstep-ich as module. I've put the module 
 > > in initramfs and as soon it is loaded, the ondemand governor is selected 
 > > and works.
 > > 
 > > However, do you think it's ok to prevent cpufreq drivers to be built-in 
 > > if ondemand governor is selected as default?
 > > Another way would be to 
 > > reorganise the initialisation code so that workqueue is initialised 
 > > before the cpufreq framework is started, do you think it's possible?
 > Making all this work with low-level drivers built in would be perfect of
 > course...
 > If nobody answers the next days, could you open a bug at
 > bugzilla.kernel.org, so that this does not get lost, pls.
 > Add myself and Venkatesh to CC...
 > I try to have a look at it ASAP.
 > If someone already finds out more, this would be great.

I'm actually surprised that speedstep-ich works with a low enough
latency for ondemand.  On the kernel that works, can you boot
with cpufreq.debug=7 and send the cpufreq messages from dmesg?

	Dave

-- 
http://www.codemonkey.org.uk

Re: Kernel panic at boot with ondemand governor as default (2.6.24-rc2)

[Eric Piel]

2007年11月12日 21:26, Dave Jones wrote/a écrit:
> On Mon, Nov 12, 2007 at 04:23:09PM +0100, Thomas Renninger wrote:
>  > On Mon, 2007-11-12 at 15:02 +0100, Eric Piel wrote:
:
>  > > Yes, it works if I compile speedstep-ich as module. I've put the module 
>  > > in initramfs and as soon it is loaded, the ondemand governor is selected 
>  > > and works.
>  > > 
>  > > However, do you think it's ok to prevent cpufreq drivers to be built-in 
>  > > if ondemand governor is selected as default?
>  > > Another way would be to 
>  > > reorganise the initialisation code so that workqueue is initialised 
>  > > before the cpufreq framework is started, do you think it's possible?
>  > Making all this work with low-level drivers built in would be perfect of
>  > course...
>  > If nobody answers the next days, could you open a bug at
>  > bugzilla.kernel.org, so that this does not get lost, pls.
>  > Add myself and Venkatesh to CC...
>  > I try to have a look at it ASAP.
>  > If someone already finds out more, this would be great.
> 
> I'm actually surprised that speedstep-ich works with a low enough
> latency for ondemand.  On the kernel that works, can you boot
> with cpufreq.debug=7 and send the cpufreq messages from dmesg?
I'm not surprised, contrarly to speedstep-smi, speedstep-ich has been 
working with ondemand for quite a long time, since commit 
1a10760c91c394dfe4adfefeeaf85cd8098c4894 precisely :-) As stated in the 
commit message, the latency transition is about 200µs, much smaller than 
the 10ms required by ondemand. Actually, on my laptop the transition 
time is set to even higher due to problems of measurement, but it's 
still fine for ondemand. I get this at boot:
speedstep: frequency transition measured seems out of range (4800 nSec), 
falling back to a safe one of 500000 nSec.


See you,
Eric

Re: Kernel panic at boot with ondemand governor as default (2.6.24-rc2)

[Eric Piel]

2007年11月12日 16:23, Thomas Renninger wrote/a écrit:
> On Mon, 2007-11-12 at 15:02 +0100, Eric Piel wrote:
:
>> Another way would be to 
>> reorganise the initialisation code so that workqueue is initialised 
>> before the cpufreq framework is started, do you think it's possible?
> Making all this work with low-level drivers built in would be perfect of
> course...
Hi, I've just checked and it seems a bit weird, at least not as I 
expected: the workqueue is already initialized _before_ cpufreq! At 
least, from what I read in init/main.c, in do_basic_setup(), first there 
is a call to init_workqueues(), then there is a call to do_initcalls() 
(which indirectly calls cpufreq_core_init()). So maybe workqueues need 
something more than being initialized to work? What could it be? (My 
kernel is compiled for monoprocessor,  I can't see what goes wrong in 
wq_per_cpu()).

See you,
Eric