Re: s390 depending on cc-options makes it difficult to configure

[Masahiro Yamada]

On Sat, Feb 15, 2020 at 5:35 AM Jeremy Cline <jcline@xxxxxxxxxx> wrote:
>
> On Fri, Feb 14, 2020 at 12:31:05PM +0900, Masahiro Yamada wrote:
> > Hi.
> >
> > On Tue, Feb 11, 2020 at 3:49 AM Philipp Rudo <prudo@xxxxxxxxxxxxx> wrote:
> > >
> > > Hey Jeremy,
> > > Hey Michal,
> > >
> > > sorry for the late response. The mail got lost in the pre-xmas rush...
> > >
> > > In my opinion the problem goes beyond s390 and the commit you mentioned. So I'm
> > > also adding Masahiro as Kconfig maintainer and author of cc-option.
> >
> >
> > I did not notice the former discussion.
> > Thanks for CC'ing me.
> >
> >
> >
> >
> > > On Wed, 11 Dec 2019 12:18:22 -0500
> > > Jeremy Cline <jcline@xxxxxxxxxx> wrote:
> > >
> > > > On Tue, Dec 10, 2019 at 10:01:08AM +0100, Michal Kubecek wrote:
> > > > > On Mon, Dec 09, 2019 at 11:41:55AM -0500, Jeremy Cline wrote:
> > > > > > Hi folks,
> > > > > >
> > > > > > Commit 5474080a3a0a ("s390/Kconfig: make use of 'depends on cc-option'")
> > > > > > makes it difficult to produce an s390 configuration for Fedora and Red
> > > > > > Hat kernels.
> > > > > >
> > > > > > The issue is I have the following configurations:
> > > > > >
> > > > > > CONFIG_MARCH_Z13=y
> > > > > > CONFIG_TUNE_Z14=y
> > > > > > # CONFIG_TUNE_DEFAULT is not set
> > > > > >
> > > > > > When the configuration is prepared on a non-s390x host without a
> > > > > > compiler with -march=z* it changes CONFIG_TUNE_DEFAULT to y which, as
> > > > > > far as I can tell, leads to a kernel tuned for z13 instead of z14.
> > > > > > Fedora and Red Hat build processes produce complete configurations from
> > > > > > snippets on any available host in the build infrastructure which very
> > > > > > frequently is *not* s390.
> > > > >
> > > > > We have exactly the same problem. Our developers need to update config
> > > > > files for different architectures and different kernel versions on their
> > > > > machines which are usually x86_64 but that often produces different
> > > > > configs than the real build environment.
> > > > >
> > > > > This is not an issue for upstream development as one usually updates
> > > > > configs on the same system where the build takes place but it's a big
> > > > > problem for distribution maintainers.
> > >
> > > If I recall correct the goal was to avoid trouble with clang, as it does not
> > > support all processor types with -march. But yeah, in the original
> > > consideration we only thought about upstream development and forgot the
> > > distros.
> > > > > > I did a quick search and couldn't find any other examples of Kconfigs
> > > > > > depending on march or mtune compiler flags and it seems like it'd
> > > > > > generally problematic for people preparing configurations.
> > >
> > > True, but not the whole story. Power and Arm64 use cc-option to check for
> > > -mstack-protector*, which do not exist on s390. So you have the same problem
> > > when you prepare a config for any of them on s390. Thus simply reverting the
> > > commit you mentioned above does not solve the problem but merely hides one
> > > symptom. Which also means that the original problem will return over and over
> > > again in the future.
> > >
> > > An other reason why I don't think it makes sens to revert the commit is that it
> > > would make cc-option as a whole useless. What's the benefit in having cc-option
> > > when you are not allowed to use it? Or less provocative, in which use cases is
> > > allowed to use cc-option?
> >
> >
> > You are right.
> > Reverting the particular s390 commit is not the solution.
> >
> >
> > > > > There are more issues like this. In general, since 4.17 or 4.18, the
> > > > > resulting config depends on both architecture and compiler version.
> > > > > Earlier, you could simply run "ARCH=... make oldconfig" (or menuconfig)
> > > > > to update configs for all architectures and distribution versions.
> > > > > Today, you need to use the right compiler version (results with e.g.
> > > > > 4.8, 7.4 and 9.2 differ) and architecture.
> > > > >
> > > >
> > > > Yeah, that's also troublesome. This is by no means the first problem
> > > > related to the environment at configuration time, but it the most
> > > > bothersome to work around (at least for Fedora kernel configuration).
> > > >
> > > > > At the moment, I'm working around the issue by using chroot environments
> > > > > with target distributions (e.g. openSUSE Tumbleweed) and set of cross
> > > > > compilers for supported architectures but it's far from perfect and even
> > > > > this way, there are problemantic points, e.g. BPFILTER_UMH which depends
> > > > > on gcc being able to not only compile but also link.
> > > > >
> > > > > IMHO the key problem is that .config mixes configuration with
> > > > > description of build environment. I have an idea of a solution which
> > > > > would consist of
> > > > >
> > > > >   - an option to extract "config" options which describe build
> > > > >     environment (i.e. their values are determined by running some
> > > > >     command, rather than reading from a file or asking user) into
> > > > >     a cache file
> > > > >   - an option telling "make *config" to use such cache file for these
> > > > >     environment "config" options instead of running the test scripts
> > > > >     (and probably issue an error if an environment option is missing)
> > > > >
> > > >
> > > > I agree that the issue is mixing kernel configuration with build
> > > > environment. I suppose a cache file would work, but it still sounds like
> > > > a difficult process that is working around that fact that folks are
> > > > coupling the configuration step with the build step.
> > >
> > > An other solution would be a "I know better" switch which simply disables
> > > cc-option for that run. That would allow the use of cc-option for upstream
> > > development and provide a simple way for distros to turn it off.
> > >
> > > > I would advocate that this patch be reverted and an effort made to not
> > > > mix build environment checks into the configuration. I'm much happier
> > > > for the build to fail because the configuration can't be satisfied by
> > > > the environment than I am for the configuration to quietly change or for
> > > > the tools to not allow me to make the configuration in the first place.
> > > > Ideally the tools would warn the user if their environment won't build
> > > > the configuration, but that's a nice-to-have.
> > >
> > > I too would prefer to have a warning instead of the config being silently
> > > changed. But again, the problem goes beyond what was reported.
> > >
> > > @Masahiro: What do you think about it?
> > >
> > > Thanks
> > > Philipp
> > >
> >
> >
> > The problem for Jeremy and Michal is,
> > it is difficult to get a full-feature cross-compiler
> > for every arch.
> >
>
> Indeed.
>
> > One idea to workaround this is
> > to use a fake script that accepts any flag,
> > and use it as $(CC) in Kconfig.
> >
> > RFC patch is attached.
> >
> > This is not a perfect solution, of course.
> >
>
> The attached patch doesn't looks like it'd work for what we need,

I thought turning all cc-options to y would work
for what you need.

With this, you can enable
CONFIG_MARCH_Z13=y and CONFIG_TUNE_Z14=y
instead of CONFIG_TUNE_DEFAULT.

If this approach does not work for you,
what is your requirement?



> although I wonder if it's easier to just check when cc-options is
> defined for an environment variable or something and always return y
> instead of calling out to $(CC) at all. Comes to the same thing, I
> suppose.


The macro definition in scripts/Kconfig.include
takes precedence over any environment variable.

So, if you want to hack it from the environment,
you need to change the code somehow.

The scripts/dummy-tools/ approach does not change
anything for the use-case in upstream.

The result is the same, of course.


Masahiro Yamada




> >
> > Evaluating the compiler in the Kconfig stage
> > conceptually has a conflict with the workflow
> > of distro maintainers.
> >
> > I think the only way to solve it completely is,
> > ultimately, go back to pre 4.18 situation.
> > But, I am not sure if upstream people want to do it.
> > At least, Linus was happy to do compiler-tests
> > in Kconfig.
> >
> > I already got several criticism about the
> > new feature in Kconfig because it broke the
> > workflow of distro maintainers. Sorry about that.
> >
>
> No worries, it's a tough balancing act between upstream users and
> distros. It's not caused me *that* much bother.
>
> >
> > The idea from Michal, separation of the build environment
> > description, would work too.
> > IIRC, the crosstool-ng project generates some
> > Kconfig files based on the environment.
> > In hindsight, Kconfig did not need to have cc-option
> > but it was how I implemented. I just thought it would be cleaner to
> > put cc-option and the CONFIG option depending on it very close.
> >
> > Anyway, comments to the attachment are appreciated.
> >
>
> I believe it would solve our problem so from that perspective, it looks
> good to me.
>
> Thanks,
> Jeremy
>


-- 
Best Regards
Masahiro Yamada