Re: [PATCH 1/4] sched: move IO scheduling accounting from io_schedule_timeout() to __schedule()

[Tejun Heo]

Hello,

On Sat, Oct 29, 2016 at 05:21:26AM +0200, Peter Zijlstra wrote:
> On Fri, Oct 28, 2016 at 03:12:32PM -0400, Tejun Heo wrote:
> > Hello, Peter.
> > 
> > On Fri, Oct 28, 2016 at 09:07:02PM +0200, Peter Zijlstra wrote:
> > > One alternative is to inherit the iowait state of the task we block on.
> > > That'll not get rid of the branches much, but it will remove the new
> > > mutex APIs.
> > 
> > Yeah, thought about that briefly but we don't necessarily track mutex
> 
> This one I actually fixed and should be in -next. And it would be
> sufficient to cover the use case here.

Tracking the owners of mutexes and rwsems does help quite a bit.  I
don't think it's as simple as inheriting io sleep state from the
current owner tho.  The owner might be running or in a non-IO sleep
when others try to grab the mutex.  It is an option to ignore those
cases but this would have a real possibility to lead to surprising
results in some corner cases.  If we choose to propagate dynamically,
it becomes an a lot more complex problem and I don't think it'd be
justfiable.

Unless there can be a simple enough and reliable solution, I think
it'd be better to stick with explicit marking.

Thanks.

-- 
tejun