On Fri, Oct 1, 2021 at 5:31 AM Lee Jones [off-list ref] wrote:

On Fri, 01 Oct 2021, Arnd Bergmann wrote:

quoted

The vmlinux file is clearly too big and includes too much stuff that should
be in loadable modules

This for me is the crux of the matter.

The ability to replace modules was only brought to light as an "and
also, this is possible".  However in retrospect, given the attention
this has received, it probably shouldn't have even mentioned, as it's
not that important.

Too late, unfortunately.

I would actually argue that given the benefit of needing more vendor
engagement upstream for GKI to be smooth, it's in our interest to
welcome those engagements and make the most of it and help vendors get
there, and it's against those interests to make it easier to be
out-of-tree if it comes at the expense of our in-tree users and
maintainers which this does.

We should focus on the benefits of making parts of the kernel modular
if technically possible.  The most prominent of those is core binary
size, since this has a direct impact on boot-time and RAM usage.

The way forward here should be to focus on the problem that needs to
be solved (vmlinux size) and not overly fixate on whether this
patchset is what needs to be merged to reach there, given the
downsides observed. I'm not saying let's not improve vmlinux binary
size, but this particular approach isn't appealing.

Reclaiming dead code after boot is certainly one way to tackle part of
the problem.  Ensuring that it's not even loaded into RAM in the first
place is a better more encompassing solution to both issues IMHO.

See my reply to Arnd; the reason some of these drivers aren't modules
today is because they are needed during early boot to bring the
platform to a stable operating point. Work on fixing the binary size
is terrific, but this approach seems to be shortsighted and it's been
done in a way that rubs the maintainers the wrong way.


-Olof