Re: [pull request][for-next] Mellanox mlx5 Reorganize core driver directory layout

[Tom Herbert <tom@herbertland.com>]

On Fri, Jan 13, 2017 at 12:29 PM, Leon Romanovsky <leon@kernel.org> wrote:
> On Fri, Jan 13, 2017 at 12:14:07PM -0500, David Miller wrote:
>> From: Saeed Mahameed <saeedm@mellanox.com>
>> Date: Thu, 12 Jan 2017 19:22:34 +0200
>>
>> > This pull request includes one patch from Leon, this patch as described
>> > below will change the driver directory structure and layout for better,
>> > logical and modular driver files separation.
>> >
>> > This change is important to both rdma and net maintainers in order to
>> > have smoother management of driver patches for different mlx5 sub modules
>> > and smoother rdma-next vs. net-next features submissions.
>> >
>> > Please find more info below -in the tag commit message-,
>> > review and let us know if there's any problem.
>> >
>> > This change doesn't introduce any conflicts with the current mlx5
>> > fixes and cleanups posted on 2017-01-10 to net branch, and merge tests
>> > worked flawlessly with no issues.
>> >
>> > This is the last pull request meant for both rdma-next and net-next.
>> > Once pulled, this will be the base shared code for both trees.
>>
>> This is pretty crazy, it will make all bug fix backporting to -stable
>> a complete nightmare for myself, Doug, various distribution maintainers
>> and many other people who quietly have to maintain their own trees and
>> do backporting.
>
> Hi Dave,
>
> I understand your worries, but our case is similar to various other
> drivers, for example hfi1 which was in staging for years while
> supported in RedHat and moved from there to IB. The Chelsio drivers did
> similar reorg in 2016 (drivers/net/ethernet/chelsio/libcxgb) while their
> drivers were in the tree for long time before.
>
> Additionally, Doug doesn't need to maintain -stable queue and it is done
> by relevant submaintainers who are adding stable tags by themselves. In
> the IB case, the burden will continue to be on me and not on Doug.
>
Recently I had to backport the mlx5 driver from 4.9 to 4.6 in order to
get support for XDP. The biggest issue I faced was the lack of
modularity in the many driver features that are now supported. The
problem with backporting these new features is the spider web of
dependencies that they bring in from the rest of the kernel. I ended
up taking out en_rep, vxlan, en_tc, eswitch, and dcbnl. The result was
~340 patches which is still a lot but at least this was constrained to
patches in the mlx5 directories and are relevant to what we want to
do.

In lieu of restructuring the directories, I would much rather see more
config options so that we can build drivers that don't unnecessarily
complicate our lives with features we don't use. This is not just true
for Mellanox, but I would say it would be true of any driver that
someone is trying to deploy and maintain at large scale.

Btw, we did hit one issue in the backport. We started to get rx csum
faults (checksum complete value indicates TCP checksum is bad, but
host computation says checksum is good). I ran against 4.9 upstream
kernel and do see these, however don't see them in 4.10. I haven't
bisected yet. Is this a known issue?

Thanks,
Tom

>>
>> I really don't think you can justify this rearrangement based upon the
>> consequences and how much activity happens in this driver.
>>
>> You should have thought long and hard about the layout a long time ago
>> rather than after the driver has been in the tree for many years.
>>
>> Sorry.