Re: [PATCH v2] vmpressure: implement strict mode

[Andrew Morton <akpm@linux-foundation.org>]

On Fri, 28 Jun 2013 09:02:01 +0900 Minchan Kim <minchan@kernel.org> wrote:

> Hi Andrew,
> 
> On Thu, Jun 27, 2013 at 03:02:31PM -0700, Andrew Morton wrote:
> > On Wed, 26 Jun 2013 23:17:12 -0400 Luiz Capitulino <lcapitulino@redhat.com> wrote:
> > 
> > > Currently, an eventfd is notified for the level it's registered for
> > > _plus_ higher levels.
> > > 
> > > This is a problem if an application wants to implement different
> > > actions for different levels. For example, an application might want
> > > to release 10% of its cache on level low, 50% on medium and 100% on
> > > critical. To do this, an application has to register a different
> > > eventfd for each pressure level. However, fd low is always going to
> > > be notified and and all fds are going to be notified on level critical.
> > > 
> > > Strict mode solves this problem by strictly notifiying an eventfd
> > > for the pressure level it registered for. This new mode is optional,
> > > by default we still notify eventfds on higher levels too.
> > > 
> > 
> > It didn't take long for this simple interface to start getting ugly :(
> > And having the fd operate in different modes is ugly.
> > 
> > Can we instead pass the level in the event payload?
> 
> You mean userland have to look the result of read(2) to confirm what
> current level is and if it's no interest for us, we don't do any reaction.

Something like that.  It's flexible, simple, keeps policy in userspace.

> If so, userland daemon would receive lots of events which are no interest.

"lots"?  If vmpressure is generating events at such a high frequency that
this matters then it's already busted?