On Fri, Mar 5, 2021 at 12:08 AM Josh Poimboeuf <jpoimboe@xxxxxxxxxx> wrote:
>
> On Thu, Mar 04, 2021 at 09:27:28PM +0900, Masahiro Yamada wrote:
> > I agree with rebuilding GCC plugins when the compiler is upgraded
> > for *in-tree* building.
> > Linus had reported it a couple of months before,
> > and I just submitted a very easy fix.
>
> Hm? So does that mean that a GCC version change won't trigger a
> tree-wide rebuild? So you're asserting that a GCC mismatch is ok for
> in-tree code, but not for external modules??? That seems backwards.
>
> For in-tree, why not just rebuild the entire tree? Some kernel features
> are dependent on compiler version or capability, so not rebuilding the
> tree could introduce silent breakage.
All the kernel-space objects are rebuilt
when the compiler is upgraded.
(See commit 8b59cd81dc5e724eaea283fa6006985891c7bff4)
Linus complaint about GCC plugins not being rebuilt.
^^^^^^^^^^^
https://lore.kernel.org/lkml/CAHk-=wieoN5ttOy7SnsGwZv+Fni3R6m-Ut=oxih6bbZ28G+4dw@xxxxxxxxxxxxxx/
That is easy to fix. I submitted a patch:
https://patchwork.kernel.org/project/linux-kbuild/patch/20210304113708.215121-1-masahiroy@xxxxxxxxxx/
> For external modules, a tree-wide rebuild isn't an option so the risk is
> assumed by the user. I posted a patch earlier [1] which prints a
> warning if the compiler major/minor version changes with an external
> module build.
>
> [1] https://lkml.kernel.org/r/20210201211322.t2rxmvnrystc2ky7@treble
>
> > Rebuilding plugins for external modules is not easy;
> > plugins are placed in the read-only directory,
> > /usr/src/linux-headers-$(uname -r)/scripts/gcc-plugins/.
> >
> > The external modules must not (cannot) update in-tree
> > build artifacts. "Rebuild" means creating copies in a different
> > writable directory.
> > Doing that requires a lot of design changes.
>
> Ok. So it sounds like the best/easiest option is the original patch in
> this thread: when building an external module with a GCC mismatch, just
> disable the GCC plugin, with a warning (or an error for randstruct).
It was rejected.
If a distribution wants to enable CONFIG_GCC_PLUGINS,
it must provide GCC whose version is the same as
used for building the kernel.
If a distribution cannot manage release in that way,
do not enable CONFIG_GCC_PLUGINS in the first place.
--
Best Regards
Masahiro Yamada
