ule+smp: small optimization for turnstile priority lending

[Attilio Rao]

On Mon, Oct 29, 2012 at 6:59 PM, Attilio Rao <attilio at freebsd.org> wrote:

[ trimm ]

> This also let me wonder about the tdq_lowpri check in the self case,
> in general. Basically it forces sched_pickcpu() to select self if and
> only if the new thread to schedule has an higher priority than the
> lowest one on curcpu. Why is that like this? Exactly this check is
> used to enforce some sort of fairness?
> It would be good if Jeff spends a word or two on this check specifically.

Also, I've read the code of tdq_setlowpri() more closely and I think
the name tdq_lowpri is highly misleading, because that seems to me the
*highest* priority thread that gets returned.
Said that, this means that self will be selected in sched_pickcpu() if
and only if the new thread has an higher priority than all the ones on
the self runqueue. Isn't all this a bit too extreme? Assuming that one
cpu has only a single high-priority thread and others are very loaded
it would likely make sense to not keep loading them but switch to the
self one, maybe.

> Anyway the patch that implements what suggested, let me know your
> thinking about it:
> http://www.freebsd.org/~attilio/avg_ule2.patch

I was thinking that however, maybe we could do the tdq_choose()
calculation if self == target, to have a little more chances to get
the optimization in place, eventually.

Thanks,
Attilio


-- 
Peace can only be achieved by understanding - A. Einstein