[PATCH 0/4] ARM: mvebu: minor Armada 370 RD improvements

[ijc@hellion.org.uk (Ian Campbell)]

On Thu, 2014-09-11 at 10:26 -0400, Jason Cooper wrote:
> merf.  It helps if I actually _add_ Ian to the Cc. :)
> 
> On Thu, Sep 11, 2014 at 10:26:01AM -0400, Jason Cooper wrote:
> > +Ian
> > 
> > On Thu, Sep 11, 2014 at 04:00:00PM +0200, Thomas Petazzoni wrote:
> > > On Thu, 11 Sep 2014 15:43:23 +0200, Andrew Lunn wrote:
> > > > A point for discussion, triggered by 3/4. We have repeatedly seen
> > > > issues with modular builds, which we don't see with everything built
> > > > in. Can we improve our testing? Is there is way to take for example
> > > > mvebu_v7_defconfig and turn any y tristate options into m? Build and
> > > > boot? Get this added to the boot test farms?
> > > 
> > > I indeed agree that we are not testing much the modular case, but:
> > > 
> > >  1/ In my case, a mvebu_v7_defconfig where most the stuff would become
> > >     enabled as a module would make mvebu_v7_defconfig pretty much
> > >     useless for my daily work. When I'm doing active development, I
> > >     typically have everything built-in including the initramfs itself,
> > >     so that I don't rely on any storage/network to get the rootfs and
> > >     additional kernel drivers.
> > > 
> > >  2/ I don't think it would change much in terms of runtime coverage:
> > >     the tests done on the boot farm are just boot tests, I don't think
> > >     they load any kernel module. We could ask Kevin and Olof what
> > >     rootfs they use, but most likely it's some very small rootfs that
> > >     doesn't even have udev/mdev for automatic module loading. So no
> > >     module would be loaded, which would mean that a mvebu_v7_defconfig
> > >     with everything as a module would actually test *less* thing than a
> > >     configuration that has everything built in.
> > > 
> > > So, I agree on the observation, but I'm unsure the proposed solution
> > > will really solve or even improve the situation :/
> > 
> > Since debian does modular builds on a regular basis, maybe they have
> > some recommendations.

Rather than modularising $socfamily_defconfig perhaps this is something
which can be achieved by also testing multiplatform defconfigs builds,
and perhaps making those more modular (if they aren't already)?

It seems like a multiplatform kernel is something which would be more
naturally quite modular anyway, and certainly distros are mostly
interested in moving to configurations which are both multiplatform and
modular (at least in the long run, Debian has some legacy v5 flavours
which may or may not get moved over).

That would avoid needing to mess with mvebu_v7_defconfig, since it seems
quite reasonable and natural for that to be pretty non-modular.

Ian.