Re: sys_sched_yield fast path

[Anton Blanchard <anton@linuxcare.com.au>]

 
Hi,

> 2.4.x has changed the scheduler behaviour so that the task that call
> sched_yield() is not rescheduled by the incoming schedule().  A flag is
> set ( under certain conditions in SMP ) and the goodness() calculation
> assign the lower value to the exiting task ( this flag is cleared in
> schedule_tail() ).

The behaviour I am talking about is when there is a heavily contended
spinlock, and more than one task is trying to obtain it. Since SCHED_YIELD
only changes the goodness when we are trying to reschedule the task we
can bounce between two or more tasks doing sched_yield() for a while before
finally running the task that has the spinlock.

Of course with short lived spinlocks you should rarely get the case where
a task grabs a spinlock just before its timeslice is up, so maybe the answer
is just to spin a few times on sched_yield() then back off to nanosleep()
like pthreads does.

Anton