Re: [PATCH net-next 2/2] net: exit busy loop when another process is runnable

[Ingo Molnar <mingo@kernel.org>]

* Michael S. Tsirkin <mst@redhat.com> wrote:

> On Fri, Aug 22, 2014 at 09:36:53AM +0200, Ingo Molnar wrote:
> > 
> > > > diff --git a/include/net/busy_poll.h b/include/net/busy_poll.h
> > > > index 1d67fb6..8a33fb2 100644
> > > > --- a/include/net/busy_poll.h
> > > > +++ b/include/net/busy_poll.h
> > > > @@ -109,7 +109,8 @@ static inline bool sk_busy_loop(struct sock *sk, int nonblock)
> > > >  		cpu_relax();
> > > >  
> > > >  	} while (!nonblock && skb_queue_empty(&sk->sk_receive_queue) &&
> > > > -		 !need_resched() && !busy_loop_timeout(end_time));
> > > > +		 !need_resched() && !busy_loop_timeout(end_time) &&
> > > > +		 nr_running_this_cpu() < 2);
> > 
> > So it's generally a bad idea to couple to the scheduler through 
> > such a low level, implementation dependent value like 
> > 'nr_running', causing various problems:
> > 
> >  - It misses important work that might be pending on this CPU,
> >    like RCU callbacks.
> > 
> >  - It will also over-credit task contexts that might be
> >    runnable, but which are less important than the currently
> >    running one: such as a SCHED_IDLE task
> > 
> >  - It will also over-credit even regular SCHED_NORMAL tasks, if
> >    this current task is more important than them: say
> >    SCHED_FIFO. A SCHED_FIFO workload should run just as fast 
> >    with SCHED_NORMAL tasks around, as a SCHED_NORMAL workload 
> >    on an otherwise idle system.
> > 
> > So what you want is a more sophisticated query to the 
> > scheduler, a sched_expected_runtime() method that returns the 
> > number of nsecs this task is expected to run in the future, 
> > which returns 0 if you will be scheduled away on the next 
> > schedule(), and returns infinity for a high prio SCHED_FIFO 
> > task, or if this SCHED_NORMAL task is on an otherwise idle CPU.
> > 
> > It will return a regular time slice value in other cases, when 
> > there's some load on the CPU.
> > 
> > The polling logic can then do its decision based on that time 
> > value.
> > 
> > All this can be done reasonably fast and lockless in most 
> > cases, so that it can be called from busy-polling code.
> >
> > An added advantage would be that this approach consolidates the 
> > somewhat random need_resched() checks into this method as well.
> > 
> > In any case I don't agree with the nr_running_this_cpu() 
> > method.
> > 
> > (Please Cc: me and lkml to future iterations of this patchset.)
> > 
> > Thanks,
> > 
> > 	Ingo
> 
> I tried to look into this: it might be even nicer to add
> sched_expected_to_run(time) which tells us whether we expect the current
> task to keep running for the next XX nsecs.
> 
> For the fair scheduler, it seems that it could be as simple as
> 
> +static bool expected_to_run_fair(struct cfs_rq *cfs_rq, s64 t)
> +{
> +       struct sched_entity *left;
> +       struct sched_entity *curr = cfs_rq->curr;
> +
> +       if (!curr || !curr->on_rq)
> +               return false;
> +
> +       left = __pick_first_entity(cfs_rq);
> +       if (!left)
> +               return true;
> +
> +       return (s64)(curr->vruntime + calc_delta_fair(t, curr) -
> +                    left->vruntime) < 0;
> +}
> 
> The reason it seems easier is because that way we can reuse
> calc_delta_fair and don't have to do the reverse translation
> from vruntime to nsec.
> 
> And I guess if we do this with interrupts disabled, and only poke
> at the current CPU's rq, we know first entity
> won't go away so we don't need locks? 
> 
> Is this close to what you had in mind?

Yeah, fair enough ;-)

I'm not 100% convinced about the interface, but the model looks good to me.
Let's try it - I don't have fundamental objections anymore.

I also agree that it could be done lockless - although I'd suggest two steps: 
first do the dumb thing with the proper scheduler lock(s) held, then another patch 
which removes the locks for a bit more performance. That will make any subtle 
crashes/races bisectable.

Thanks,

	Ingo