Re: CPU scheduler weirdness?

[Marton Balint]

On Thu, 20 Aug 2009, Marton Balint wrote:
> On Thu, 20 Aug 2009, Ingo Molnar wrote:
>
> > * Marton Balint <cus@xxxxxxxxxx> wrote:
> >
> > > On Wed, 19 Aug 2009, Peter Zijlstra wrote:
> > >
> > > > On Wed, 2009-08-19 at 14:34 +0200, Marton Balint wrote:
> > > > > On Wed, 19 Aug 2009, Peter Zijlstra wrote:
> > > > >
> > > > > > On Wed, 2009-08-19 at 14:01 +0200, Marton Balint wrote:
> > > > > > > On Wed, 19 Aug 2009, Peter Zijlstra wrote:
> > > > > > > > On Tue, 2009-08-18 at 21:49 +0200, Marton Balint wrote:
> > > > > > > >
> > > > > > > > > In the meantime, I was able to create a tiny C program which always
> > > > > > > > > succesfully reproduces the bug. It's basically an endless loop which 
> > > > > > > > > does
> > > > > > > > > not stop while the process is running on the last CPU core. The 
> > > > > > > > > program
> > > > > > > > > creates multiple instances of itself, to be able to keep all of the 
> > > > > > > > > CPU
> > > > > > > > > cores busy. After 1 second, the processes running on other than the 
> > > > > > > > > last
> > > > > > > > > CPU core die, the processes running on the last CPU core remain 
> > > > > > > > > stuck
> > > > > > > > > there...
> > > > > > > > >
> > > > > > > > > I tested it on my dual core system, if someone could test it on a 
> > > > > > > > > quad
> > > > > > > > > core and report back that would probably be useful.
> > > > > > > > >
> > > > > > > > > Usage: ./schedtest <number of CPU cores>
> > > > > > > > >
> > > > > > > > > And don't forget to kill the stuck processes after using the 
> > > > > > > > > program! :)
> > > > > > > >
> > > > > > > > So what's the bug? Sure one task will stay on the cpu, and because 
> > > > > > > > there
> > > > > > > > is no contention it doesn't get migrated, and therefore won't quit,
> > > > > > > > how's that a problem?
> > > > > > >
> > > > > > > Problem is that more than one processes remain on that CPU core, and 
> > > > > > > none
> > > > > > > of them get migrated to other (idle) cores. I tested it with my E8400
> > > > > > > processor and 2.6.31-rc5-git3 kernel.
> > > > > >
> > > > > > Only one remains here.. on a c2q running 2.6.31-rc6-tip
> > > > > >
> > > > > > Do you have a .config handy?
> > > > >
> > > > > Yes it's in my original post:
> > > > >
> > > > > http://marc.info/?l=linux-kernel&m=125012584709800&w=2
> > > >
> > > > Right you are,.. so I build a kernel with the cgroup scheduler in and
> > > > tested it on a dual-core opteron machine, but I can't seem to reproduce
> > > > this.
> > > >
> > > > Are you using cgroups in any way, or do you simply have it enabled in
> > > > your config?
> > >
> > > No, it's just enabled. Actually the kernel is from the
> > > openSUSE build service:
> > >
> > > http://download.opensuse.org/repositories/Kernel:/HEAD/openSUSE_11.1/x86_64/
> > >
> > > But the problem is present for both the kernel-default
> > > kernel and the kernel-vanilla kernel which does not
> > > contain any suse-specific patches.
> > >
> > > This evening I had a bit more time to test, and I've
> > > made a surprising discovery: I can only reproduce the
> > > bug if the kernel module of my TV tuner card is loaded.
> > > I have a Leadtek Winfast 2000 XP Expert TV card, it
> > > uses the cx8800 kernel module. It seems that the
> > > problem is somehow related to the infrared sensor of
> > > the TV card, because I recompiled the module with the
> > > 'case CX88_BOARD_WINFAST2000XP_EXPERT:' line removed
> > > from cx88-input.c and I couldn't reproduce the bug with
> > > the new kernel module.
> >
> > Extremely weird. Are timers somehow busted?
>
> How can I check that?
>
> In the meantime, I updated my original C program and also created a kernel 
> module (schedtest_mod.c) which causes the same scheduling problems as the 
> kernel module of my TV card. The kernel module is a skeleton of the infrared 
> sensor polling code in cx88-input.c. It uses schedule_delayed_work, this 
> seems to cause the problem. The C program (schedtest.c) is also updated, it 
> now detects the number of CPU cores, from now, what you can set as a command 
> line parameter is the CPU core number, on which the schedtest processes will 
> not quit. (previously this was always the last core).
>
> So to reproduce the bug on a dual core system, compile and insert the kernel 
> module (schedtest_mod.c). Then check dmesg, it should contain on which CPU 
> core is the delayed_work running. You should use the CPU core id of the 
> _other_ CPU core as a command line parameter to the updated schedtest 
> program.
>
> And by the way, thank you guys for the help so far, hopefully we'll get to 
> the bottom of this :)

I reproduced the bug with the previously provided kernel module and 
C program on a different computer (it's a laptop with a core2 duo P8400 
CPU), and also bisected the bug to this commit:

sched: fine-tune SD_MC_INIT:
14800984706bf6936bbec5187f736e928be5c218

If I add again the removed SD_BALANCE_NEWIDLE to flags, then everything 
works as expected. So what would be the correct fix for this bug? 
Revert the patch? Or just add SD_BALANCE_NEWIDLE to flags?

Regards,
  Marton
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/