On Fri, Aug 23, 2024 at 10:03:05PM +0900, Masahiro Yamada wrote:
> On Thu, Aug 22, 2024 at 5:36 PM Michal Suchánek <msuchanek@xxxxxxx> wrote:
> >
> > Hello,
> >
> > On Thu, Aug 22, 2024 at 01:05:11AM -0500, Lucas De Marchi wrote:
> > > On Tue, Dec 19, 2023 at 05:37:31PM GMT, Masahiro Yamada wrote:
> > > > On Thu, Dec 7, 2023 at 3:37 AM Lucas De Marchi <lucas.demarchi@xxxxxxxxx> wrote:
> > > > >
> > > > > On Fri, Nov 10, 2023 at 01:13:53PM +0100, Michal Suchanek wrote:
> > > > > >Hello,
> > > > > >
> > > > > >This is resend of the last patch in the series that adds prefix support
> > > > > >to kernel module location together with additional patch for validating
> > > > > >the user supplied input to options that are interpreted as directories.
> > > > > >
> > > > > >Thanks
> > > > >
> > > > > applied, thanks
> > > > >
> > > > > Lucas De Marchi
> > > >
> > > >
> > > >
> > > > If I understood this correctly, MODULE_DIRECTORY is determined
> > > > by "configure --with-module-directory=...", and there is no
> > > > way to change it after that.
> > > >
> > > >
> > > > If so, how to work with cross-building?
> > > >
> > > > Cross-building is typical when building embedded Linux systems.
> > >
> > > I was thinking the `pkg-config --variable=module_directory`
> > > from the target would be sufficient, but...
> > >
> > > >
> > > >
> > > > Consider this scenario:
> > > >
> > > > - Your build machine adopts
> > > > MODULE_DIRECTORY=/usr/lib/modules
> > > > - The target embedded system adopts
> > > > MODULE_DIRECTORY=/lib/modules
> > > >
> > > > (or vice a versa)
> > > > depmod is used also for cross-building because
> > > > it is executed as a part of "make module_install".
> > > >
> > > >
> > > > The counterpart patch set for Kbuild provides
> > > > KERNEL_MODULE_DIRECTORY, which only changes
> > > > the destination directory to which *.ko are copied.
> > > >
> > > > You cannot change the directory where the
> > > > depmod searches for modules, as it is fixed
> > > > at the compile-time of kmod.
> > > >
> > > >
> > > >
> > > >
> > > > In this case, what we can do is to build another
> > > > instance of kmod configured for the target system,
> > >
> > > the target system may not even have depmod actually, so using just the
> > > host one seems more appropriate. But target should have the kmod.pc for
> > > the pkg-config call to work.
> > >
> > > > and use it for modules_install:
> > > >
> > > > 1. In the kmod source directory
> > > > ./configure --with=module-directory=/lib/modules
> > > > make
> > > >
> > > > 2. make modules_install INSTALL_MOD_PATH=<staging-dir>
> > > > KERNEL_MODULE_DIRECTORY=/lib/modules
> > > > DEPMOD=<new-depmod-you-has-just-built>
> > > >
> > > >
> > > >
> > > > If you use OpenEmbedded etc., this is what you do
> > > > because host tools are built from sources.
> > > >
> > > > But, should it be required all the time?
> > > > Even when the target embedded system uses
> > > > busybox-based modprobe instead of kmod?
> > >
> > > no, I don't think we can rely on depmod from the target.
> > >
> > > >
> > > >
> > > >
> > > > depmod provides --basedir option, which changes
> > > > the prefix part, but there is no way to override
> > > > the stem part, MODULE_DIRECTRY.
> > > >
> > > > In the review of the counter patch set,
> > > > I am suggesting an option to override MODULE_DIRECTRY
> > > > (let's say --moduledir) at least for depmod.
> > >
> > > ok
> > >
> > > >
> > > > (Perhaps modinfo too, as it also supports --basedir)
> > > >
> > > >
> > > >
> > > > Then, we can change scripts/depmod.sh so that
> > > > Kbuild can propagate KERNEL_MODULE_DIRECTORY
> > > > to depmod.
> > > >
> > > >
> > > > if <depmod supports --moduledir>; then
> > > > set -- "$@" --moduledir "${KERNEL_MODULE_DIRECTORY}"
> > > > fi
> > > >
> > > >
> > > >
> > > > Does it make sense?
> >
> > It does not make sense for the common case: building kernel for the host
> > system.
> >
> > Then overriding the directory is wrong, and using what kmod was compiled
> > with is needed to get correct module directory layout.
>
>
> NACK.
>
> scripts/Makefile.modinst and depmod must agree about
> the install destination.
>
> Both must refer to the same ${KERNEL_MODULE_DIRECTORY}.
Nack to what, exactly?
And what needs to agree here, exactly?
If the kmod was compiled with a non-default kernel module directory then
for depmod and scripts/Makefile.modinst to agree the kernel makefile
should extract the directory it was compile with from kmod kmod, and
that is the change that was not merged.
Overriding the directory with an option only for the kernel build will
make modules_install install the modules in the wrong directory.
Of course, the user is free to move them elsewhere afterwards but I
would say they should not need to do that.
Then there is the possibility that the build is for a different target
system, and the host kmod and scripts/Makefile.modinst agreeing is not
enough.
Then either a 'cross' kmod can be built, and this will again work so
long as both path and pkgconfig path point to this modified kmod.
Or kmod can grow an option to set the kernel module directory
dynamically. However, setting it to the current hardcoded value will
again break the case when building for the current system with
non-default kernel module directory location.
Unless both is done - that is the default is read from kmod, user is
provided with an option to override the default, and whatever ends up
being used by scripts/Makefile.modinst is then passed back to depmod.
Or do you envision some other solution?
Thanks
Michal
> > Or it would make sense if both was done:
> >
> > Default KERNEL_MODULE_DIRECTORY to what kmod was compiled with, and
> > then pass the actual value to depmod so that depmod uses the compiled-in
> > value by default, and the user-provided value when
> > KERNEL_MODULE_DIRECTORY was overridden by the user.
> >
> > Thanks
> >
> > Michal
>
>
>
> --
> Best Regards
> Masahiro Yamada
