On 10 April 2013 09:26, Peter Zijlstra peterz@infradead.org wrote:
On Tue, 2013-04-09 at 11:06 +0200, Vincent Guittot wrote:
+void idle_enter(struct rq *this_rq) +{
update_rq_runnable_avg(this_rq, 1);+}
+void idle_exit(struct rq *this_rq) +{
update_rq_runnable_avg(this_rq, 0);+}
These seem like fairly unfortunate names to expose to the global namespace, why not expose update_rq_runnable_avg() instead?
Just to gather in one place all cfs actions that should be done when exiting idle even if we only have update_rq_runnable_avg right now. I have distinguished that from idle_balance because this sequence can't generate extra context switch like idle_balance and they would finally not be called in the same time
OK, but could we then please give then more scheduler specific names? It just seems to me that idle_enter/idle_exit() are very obvious function names for unrelated things.
How about calling it idle_{enter,exit}_fair; so that once other classes grow hooks we can do something like:
My primary goal was to align with idle_balance name but idle_{enter,exit}_fair is better
In the same way, should we change idle_balance to idle_balance_fair ?
and since we don't have Steve's irq constraint anymore, we could move idle_balance in the beginning of the function pick_next_task_fair ? We will not have spurious switch context and we will remove fair function from __schedule function
Vincent
static void pre_schedule_idle(struct rq *rq, struct task_struct *p) { struct sched_class *class;
for_each_class(class) { if (class->idle_enter) class->idle_enter(rq); }}
or whatnot..