Re: [PATCH 5/7] percpu-rwsem: change it to rely on rss_sync infrastructure

[Oleg Nesterov <oleg@redhat.com>]

On 07/15, Paul E. McKenney wrote:
>
> On Sun, Jul 12, 2015 at 01:36:01AM +0200, Oleg Nesterov wrote:
> > Currently down_write/up_write calls synchronize_sched_expedited()
> > twice which is evil. Change this code to rely on rcu-sync primitives.
> > This avoids the _expedited "big hammer", and this can be faster in
> > the contended case or even in the case when a single thread does
> > down_write/up_write in a loop.
>
> But "evil" is such a strong word!  ;-)

I tried to be convincive ;)

But yes, to me this synchronize_sched_expedited() looks really annoying
because it is not friendly to other workloads.

> More seriously, introducing a read-side smp_mb()

Sure, we can penalize the readers and simplify this code. But the
main purpose of this primitive was "make the readers fast".

I'll write another email about this...

> My main concern would be the introduction of an exclusive lock for the
> writer, but if you are hitting the write-side acquisition that hard,
> perhaps you are having other bigger problems.

Sorry, I don't really understand... could you explain?

> Another concern would be performance in a workload where there are way
> more readers than writers, but enough writers so that there is not much
> more than one grace period's worth of spacing between them.  In that case,
> the readers are having to hit the global lock more frequently with this
> change than in the original.

Of course, this is true.

Oleg.