de-HZ-ification of the granularity defaults unearthed a pre-existing
property of CFS: while it correctly converges to the granularity goal,
it does not prevent run-time fluctuations in the range of
[-gran ... 0 ... +gran].

With the increase of the granularity due to the removal of HZ
dependencies, this becomes visible in chew-max output (with 5 tasks
running):

 out:  28 . 27. 32 | flu:  0 .  0 | ran:    9 .   13 | per:   37 .   40
 out:  27 . 27. 32 | flu:  0 .  0 | ran:   17 .   13 | per:   44 .   40
 out:  27 . 27. 32 | flu:  0 .  0 | ran:    9 .   13 | per:   36 .   40
 out:  29 . 27. 32 | flu:  2 .  0 | ran:   17 .   13 | per:   46 .   40
 out:  28 . 27. 32 | flu:  0 .  0 | ran:    9 .   13 | per:   37 .   40
 out:  29 . 27. 32 | flu:  0 .  0 | ran:   18 .   13 | per:   47 .   40
 out:  28 . 27. 32 | flu:  0 .  0 | ran:    9 .   13 | per:   37 .   40

average slice is the ideal 13 msecs and the period is picture-perfect 40
msecs. But the 'ran' field fluctuates around 13.33 msecs and there's no
mechanism in CFS to keep that from happening: it's a perfectly valid
solution that CFS finds.

to fix this we add a granularity/preemption rule that knows about
the "target latency", which makes tasks that run longer than the ideal
latency run a bit less. The simplest approach is to simply decrease the
preemption granularity when a task overruns its ideal latency. For this
we have to track how much the task executed since its last preemption.

( this adds a new field to task_struct, but we can eliminate that
  overhead in 2.6.24 by putting all the scheduler timestamps into an
  anonymous union. )

with this change in place, chew-max output is fluctuation-less all
around:

 out:  28 . 27. 39 | flu:  0 .  2 | ran:   13 .   13 | per:   41 .   40
 out:  28 . 27. 39 | flu:  0 .  2 | ran:   13 .   13 | per:   41 .   40
 out:  28 . 27. 39 | flu:  0 .  2 | ran:   13 .   13 | per:   41 .   40
 out:  28 . 27. 39 | flu:  0 .  2 | ran:   13 .   13 | per:   41 .   40
 out:  28 . 27. 39 | flu:  0 .  1 | ran:   13 .   13 | per:   41 .   40
 out:  28 . 27. 39 | flu:  0 .  1 | ran:   13 .   13 | per:   41 .   40

this patch has no impact on any fastpath or on any globally observable
scheduling property. (unless you have sharp enough eyes to see
millisecond-level ruckles in glxgears smoothness :-)
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Mike Galbraith <efault@gmx.de>