On Sun, 23 Sep 2007, Mike Galbraith wrote:

quoted text
> On Sat, 2007-09-22 at 12:01 +0200, Mike Galbraith wrote:
>> On Fri, 2007-09-21 at 20:27 -0700, Tong Li wrote:
>>> Mike,
>>>
>>> Could you try this patch to see if it solves the latency problem?
>>
>> No, but it helps some when running two un-pinned busy loops, one at nice
>> 0, and the other at nice 19.  Yesterday I hit latencies of up to 1.2
>> _seconds_ doing this, and logging sched_debug and /proc/`pidof
>> Xorg`/sched from SCHED_RR shells.
>
> Looking at a log (snippet attached) from this morning with the last hunk
> of your patch still intact, it looks like min_vruntime is being modified
> after a task is queued.  If you look at the snippet, you'll see the nice
> 19 bash busy loop on CPU1 with a vruntime of 3010385.345325, and one
> second later on CPU1 with it's vruntime at 2814952.425082, but
> min_vruntime is 3061874.838356.

I think this could be what was happening: between the two seconds, CPU 0 
becomes idle and it pulls the nice 19 task over via pull_task(), which 
calls set_task_cpu(), which changes the task's vruntime to the current 
min_vruntime of CPU 0 (in my patch). Then, after set_task_cpu(), CPU 0 
calls activate_task(), which calls enqueue_task() and in turn 
update_curr(). Now, nr_running on CPU 0 is 0, so sync_vruntime() gets 
called and CPU 0's min_vruntime gets synced to the system max. Thus, the 
nice 19 task now has a vruntime less than CPU 0's min_vruntime. I think 
this can be fixed by adding the following code in set_task_cpu() before we 
adjust p->vruntime:

if (!new_rq->cfs.nr_running)
 	sync_vruntime(new_rq);

quoted text
> static void sync_vruntime(struct cfs_rq *cfs_rq)
> {
> 	struct rq *rq;
> -	int cpu;
> +	int cpu, wrote = 0;
>
> 	for_each_online_cpu(cpu) {
> 		rq = &per_cpu(runqueues, cpu);
> +		if (spin_is_locked(&rq->lock))
> +			continue;
> +		smp_wmb();
> 		cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime,
> 				rq->cfs.min_vruntime);
> +		wrote++;
> 	}
> -	schedstat_inc(cfs_rq, nr_sync_min_vruntime);
> +	if (wrote)
> +		schedstat_inc(cfs_rq, nr_sync_min_vruntime);
> }

I think this rq->lock check works because it prevents the above scenario 
(CPU 0 is in pull_task so it must hold the rq lock). But my concern is 
that it may be too conservative, since sync_vruntime is called by 
update_curr, which mostly gets called in enqueue_task() and 
dequeue_task(), both of which are often invoked with the rq lock being 
held. Thus, if we don't allow sync_vruntime when rq lock is held, the sync 
will occur much less frequently and thus weaken cross-CPU fairness.

   tong
-
unsubscribe notice
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/