Re: [RFC PATCH v5 16/16] dcache: Add CONFIG_DCACHE_SMO

[Roman Gushchin]

On Tue, May 21, 2019 at 11:31:18AM +1000, Tobin C. Harding wrote:
> On Tue, May 21, 2019 at 12:57:47AM +0000, Roman Gushchin wrote:
> > On Mon, May 20, 2019 at 03:40:17PM +1000, Tobin C. Harding wrote:
> > > In an attempt to make the SMO patchset as non-invasive as possible add a
> > > config option CONFIG_DCACHE_SMO (under "Memory Management options") for
> > > enabling SMO for the DCACHE.  Whithout this option dcache constructor is
> > > used but no other code is built in, with this option enabled slab
> > > mobility is enabled and the isolate/migrate functions are built in.
> > > 
> > > Add CONFIG_DCACHE_SMO to guard the partial shrinking of the dcache via
> > > Slab Movable Objects infrastructure.
> > 
> > Hm, isn't it better to make it a static branch? Or basically anything
> > that allows switching on the fly?
> 
> If that is wanted, turning SMO on and off per cache, we can probably do
> this in the SMO code in SLUB.

Not necessarily per cache, but without recompiling the kernel.
> 
> > It seems that the cost of just building it in shouldn't be that high.
> > And the question if the defragmentation worth the trouble is so much
> > easier to answer if it's possible to turn it on and off without rebooting.
> 
> If the question is 'is defragmentation worth the trouble for the
> dcache', I'm not sure having SMO turned off helps answer that question.
> If one doesn't shrink the dentry cache there should be very little
> overhead in having SMO enabled.  So if one wants to explore this
> question then they can turn on the config option.  Please correct me if
> I'm wrong.

The problem with a config option is that it's hard to switch over.

So just to test your changes in production a new kernel should be built,
tested and rolled out to a representative set of machines (which can be
measured in thousands of machines). Then if results are questionable,
it should be rolled back.

What you're actually guarding is the kmem_cache_setup_mobility() call,
which can be perfectly avoided using a boot option, for example. Turning
it on and off completely dynamic isn't that hard too.

Of course, it's up to you, it's just probably easier to find new users
of a new feature, when it's easy to test it.

Thanks!