Re: Assumably a BUG in Linux Kernel (scheduler part)

[Ingo Molnar <mingo@elte.hu>]

* Andrey Gelman <agelman@012.net.il> wrote:

> Hello there !
> Assumably, I've discovered a bug in Linux kernel (version 2.6.16), at:
> kernel\sched.c   function set_user_nice()
[...]

> //-------------------------------------------------
> /*
> 	//BUGGED FORMULA : 5 lines
>         old_prio = p->prio;
>         new_prio = NICE_TO_PRIO(nice);
>         delta = new_prio - old_prio;
>         p->static_prio = NICE_TO_PRIO(nice);
>         p->prio += delta;
> */
>     //BUG FIX : 5 lines
>     old_prio = p->static_prio;
>     new_prio = NICE_TO_PRIO(nice);
>     delta = new_prio - old_prio;
>     p->static_prio = new_prio;
>     p->prio += delta;
> //-------------------------------------------------

you are right, this is a bug in the scheduler - good find.

I did accidentally fix it months ago via one of the PI-scheduling 
cleanups, and thus the fix is in the current -mm tree and is scheduled 
for 2.6.18 inclusion:

 http://kernel.org/pub/linux/kernel/people/akpm/patches/2.6/2.6.17-rc6/2.6.17-rc6-mm2/broken-out/pi-futex-scheduler-support-for-pi.patch

(see the adding of effective_prio() to set_user_nice(), instead of an 
open-coded calculation of the priority)

But i did not realize that this also fixed a bug. The effects of the bug 
are minor: a user can renice only 19 times (to go from 0 to +19), so 
there's a finite amount of extra timeslices a CPU hog might win due to 
this. I'd not include the effective_prio() change in 2.6.17 - it touches 
quite some code and we are close to the release of 2.6.17. Nevertheless 
kudos for finding and pointing out this bug!

Andrew, please add this to the changelog of
pi-futex-scheduler-support-for-pi.patch:

the effective_prio() cleanups also fix a priority-calculation bug 
pointed out by Andrey Gelman, in set_user_nice().

	Ingo