Re: [PATCH RFC 01/10] rcu: Make CONFIG_SRCU unconditionally enabled

[Josh Triplett]

On Wed, Aug 08, 2018 at 04:02:29PM -0700, Shakeel Butt wrote:
> On Wed, Aug 8, 2018 at 11:02 AM Josh Triplett <josh@xxxxxxxxxxxxxxxx> wrote:
> >
> > On Wed, Aug 08, 2018 at 07:30:13PM +0300, Kirill Tkhai wrote:
> > > On 08.08.2018 19:23, Kirill Tkhai wrote:
> > > > On 08.08.2018 19:13, Josh Triplett wrote:
> > > >> On Wed, Aug 08, 2018 at 01:17:44PM +0300, Kirill Tkhai wrote:
> > > >>> On 08.08.2018 10:20, Michal Hocko wrote:
> > > >>>> On Tue 07-08-18 18:37:36, Kirill Tkhai wrote:
> > > >>>>> This patch kills all CONFIG_SRCU defines and
> > > >>>>> the code under !CONFIG_SRCU.
> > > >>>>
> > > >>>> The last time somebody tried to do this there was a pushback due to
> > > >>>> kernel tinyfication. So this should really give some numbers about the
> > > >>>> code size increase. Also why can't we make this depend on MMU. Is
> > > >>>> anybody else than the reclaim asking for unconditional SRCU usage?
> > > >>>
> > > >>> I don't know one. The size numbers (sparc64) are:
> > > >>>
> > > >>> $ size image.srcu.disabled
> > > >>>    text      data     bss     dec     hex filename
> > > >>> 5117546   8030506 1968104 15116156         e6a77c image.srcu.disabled
> > > >>> $ size image.srcu.enabled
> > > >>>    text      data     bss     dec     hex filename
> > > >>> 5126175   8064346 1968104 15158625         e74d61 image.srcu.enabled
> > > >>> The difference is: 15158625-15116156 = 42469 ~41Kb
> > > >>
> > > >> 41k is a *substantial* size increase. However, can you compare
> > > >> tinyconfig with and without this patch? That may have a smaller change.
> > > >
> > > > $ size image.srcu.disabled
> > > >    text        data     bss     dec     hex filename
> > > > 1105900      195456   63232 1364588  14d26c image.srcu.disabled
> > > >
> > > > $ size image.srcu.enabled
> > > >    text        data     bss     dec     hex filename
> > > > 1106960      195528   63232 1365720  14d6d8 image.srcu.enabled
> > > >
> > > > 1365720-1364588 = 1132 ~ 1Kb
> > >
> > > 1Kb is not huge size. It looks as not a big price for writing generic code
> > > for only case (now some places have CONFIG_SRCU and !CONFIG_SRCU variants,
> > > e.g. drivers/base/core.c). What do you think?
> >
> > That's a little more reasonable than 41k, likely because of
> > CONFIG_TINY_SRCU. That's still not ideal, though. And as far as I can
> > tell, the *only* two pieces of core code that use SRCU are
> > drivers/base/core.c and kernel/notifier.c, and the latter is exclusively
> > code to use notifiers with SRCU, not notifiers wanting to use SRCU
> > themselves. So, as far as I can tell, this would really just save a
> > couple of small #ifdef sections in drivers/base/core.c, and I think
> > those #ifdef sections could be simplified even further. That doesn't
> > seem worth it at all.
> 
> Hi Josh, the motivation behind enabling SRCU is not to simplify the
> code in drivers/base/core.c but rather not to introduce similar ifdefs
> in mm/vmscan.c for shrinker traversals.

Leaving aside the comment someone made about sticking with rwsem for
this, I honestly hope that someday the shrinker is optional too. :)