Re: [patch] O(1) scheduler, -H5

[Ingo Molnar]

On Fri, 11 Jan 2002, Mike Kravetz wrote:

>         reacquire_kernel_lock(current);
>         if (unlikely(current->need_resched))
>                 goto need_resched_back;
>         return;
>
> The question is why do we reacquire the kernel lock before checking
> for need_resched?.  If it is not needed, we can save a lock/unlock
> cycle in the case of need_resched.

that code is something i discovered when trying to reduce scheduling
latencies for the very first lowlatency patchset i released. Back then,
1-2-3 years ago, kernel_lock usage was still common in the kernel. So it
often happened that tasks were spending more than 1 msec spinning for the
kernel lock, and often they did that in the reacquire code. So to reduce
latencies, i've added ->need_resched checking to kernel_lock() and
reacquire_kernel_lock() as well.

these days kernel_lock contention doesnt happen all that often, so i think
we should remove it from the 2.5 tree. I consider the preemptible kernel
patch to be the most advanced method to get low scheduling-latencies
anyway.

> This code isn't new with the O(1) scheduler, so I'm guessing there is
> a need to hold the kernel_lock when checking need_resched.  I just
> don't know what it is.

there is no need for it to be under the kernel_lock - i simply found it to
be an easy and common preemption point.

	Ingo