Re: [RFC PATCH 1/3] sched/fair: Call newidle_balance() from finish_task_switch()

[Scott Wood <swood@redhat.com>]

On Wed, 2020-04-29 at 11:05 +0200, Peter Zijlstra wrote:
> On Tue, Apr 28, 2020 at 06:20:32PM -0500, Scott Wood wrote:
> > On Wed, 2020-04-29 at 01:02 +0200, Peter Zijlstra wrote:
> > > On Tue, Apr 28, 2020 at 05:55:03PM -0500, Scott Wood wrote:
> > > > On Wed, 2020-04-29 at 00:09 +0200, Peter Zijlstra wrote:
> > > > > Also, if you move it this late, this is entirely the wrong
> > > > > place.  If you do it after the context switch either use the
> > > > > balance_callback or put it in the idle path.
> > > > > 
> > > > > But what Valentin said; this needs a fair bit of support, the
> > > > > whole reason we've never done this is to avoid that double
> > > > > context switch...
> > > > > 
> > > > 
> > > > balance_callback() enters with the rq lock held but BH not
> > > > separately
> > > 
> > > BH? softirqs you mean? Pray tell more.
> > 
> > In https://lore.kernel.org/lkml/5122CD9C.9070702@oracle.com/ the need to
> > keep softirqs disabled during rebalance was brought up, but simply
> > wrapping
> > the lock dropping in local_bh_enable()/local_bh_disable() meant that
> > local_bh_enable() would be called with interrupts disabled, which isn't
> > allowed.
> 
> That thread, nor your explanation make any sense -- why do we care about
> softirqs?, 

I was trusting Steve's claim that that was the issue (it seemed plausible
given that system-wide rebalancing is done from a softirq).  If things have
changed since then, great.  If that was never the issue, then there's the
question of what caused the bug Sasha saw.

> nor do I see how placing it in finish_task_switch() helps
> with any of this.

It lets us do the local_bh_enable() after IRQs are enabled, since we don't
enter with any existing atomic context.  Though I suppose we could instead
do another lock drop at the end of newidle_balance() just to enable
softirqs.

> > > > disabled, which interferes with the ability to enable interrupts
> > > > but not BH.  It also gets called from rt_mutex_setprio() and
> > > > __sched_setscheduler(), and I didn't want the caller of those to
> > > > be stuck with the latency.
> > > 
> > > You're not reading it right.
> > 
> > Could you elaborate?
> 
> If you were to do a queue_balance_callback() from somewhere in the
> pick_next_task() machinery, then the balance_callback() at the end of
> __schedule() would run it, and it'd be gone. How would
> rt_mutex_setprio() / __sched_setscheduler() be affected?

The rq lock is dropped between queue_balance_callback() and the
balance_callback() at the end of __schedule().  What stops
setprio/setscheduler on another cpu from doing the callback at that
point?

-Scott