Re: [PATCH] Prevent immediate process rescheduling

[Avi Kivity <avi@redhat.com>]

On 09/18/2009 11:03 PM, Peter Zijlstra wrote:
> On Fri, 2009-09-18 at 21:54 +0200, Ingo Molnar wrote:
>
>    
>>> diff --git a/kernel/sched_fair.c b/kernel/sched_fair.c
>>> index 652e8bd..4fad08f 100644
>>> --- a/kernel/sched_fair.c
>>> +++ b/kernel/sched_fair.c
>>> @@ -353,11 +353,25 @@ static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
>>>   static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
>>>   {
>>>   	struct rb_node *left = cfs_rq->rb_leftmost;
>>> +	struct sched_entity *se, *curr;
>>>
>>>   	if (!left)
>>>   		return NULL;
>>>
>>> -	return rb_entry(left, struct sched_entity, run_node);
>>> +	se = rb_entry(left, struct sched_entity, run_node);
>>> +	curr =&current->se;
>>> +
>>> +	/*
>>> +	 * Don't select the entity who just tried to schedule away
>>> +	 * if there's another entity available.
>>> +	 */
>>> +	if (unlikely(se == curr&&  cfs_rq->nr_running>  1)) {
>>> +		struct rb_node *next_node = rb_next(&curr->run_node);
>>> +		if (next_node)
>>> +			se = rb_entry(next_node, struct sched_entity, run_node);
>>> +	}
>>> +
>>> +	return se;
>>>   }
>>>        
> Really hate this change though,. doesn't seem right to not pick the same
> task again if its runnable. Bad for cache footprint.
>
> The scenario is quite common for stuff like:
>
> CPU0                                      CPU1
>
> set_task_state(TASK_INTERRUPTIBLE)
>
> if (cond)
>    goto out;
>                                       <--- ttwu()
> schedule();
>
>    

I agree, yielding should be explicitly requested.

Also, on a heavily overcommitted box an undirected yield might take 
quite a long time to find the thread that's holding the lock.  I think a 
yield_to() will be a lot better:

- we can pick one of the vcpus belonging to the same guest to improve 
the probability that the lock actually get released
- we avoid an issue when the other vcpus are on different runqueues (in 
which case the current patch does nothing)
- we can fix the accounting by donating vruntime from current to the 
yielded-to vcpu


-- 
Do not meddle in the internals of kernels, for they are subtle and quick to panic.