On 20.12.21 10:58, Pavel Machek wrote:

Hi!

I believe we should not change build requirements in the middle of
stable series.

To our testing team: 5.10.85 introduced new requirements for the
build. gmp.h is now required in our configs, and maybe something else.

Hi Pavel, sorry for missing this email before now.
I can look into supporting this, depending on the answers to the comments below...

Thank you.

Easiest fix might be to add

# CONFIG_GCC_PLUGINS is not set

to our configs. Alternatively I know which patch to revert.

But I believe -stable should be the one doing the revert, as the patch
does not fix serious bug and introduces problem. Faster compile is
nice but let mainline have those kind of changes.

But that commit is needed to get gcc11 plugins to work with the 5.10.y
kernel tree. So either we "break" it for old and obsolete gcc versions
(i.e. gcc7), or newer supported versions break.

Well this leads us to an interesting point.
At the moment we use GCC v8.1.0 for building all of our kernel trees (cip & stable).
What does CIP want to do mid/long term? Keep upgrading the version we use? Or try and support a specific version of GCC for 10 years?
If the former, when do we want to upgrade?
If the latter, which version?

We should do what our users are likely to do... they want stable
kernel, and will not update toolchain in middle of product
maintainance. [Updating toolchain when starting new product with given
-cip kernel is more likely].

I believe that means we should stick to specific version, but I'm not
sure what version it is. We support Debian distro, likely gcc version
from that distro would be a good option? Perhaps we should ask on TSC
meeting tommorow?

5.10 kernel was released in Dec 2020. At that time, gcc 8 and 9 were
maintained, and gcc 10 was new (https://gcc.gnu.org/releases.html).

To get some results for -stable testing, easiest options might be to
disable gcc plugin support in Kconfig.

Best regards,
Pavel

The natural pairing would be "buster/kernel 4.19/gcc-8" and
"bullseye/kernel 5.10/gcc-10", indeed.

I'm definitely not able to attend the TSC call tomorrow. If you want to
discuss this topic, someone would have to pick up the kernel WG
representation.

Jan