On Sat, Oct 5, 2019 at 3:25 AM Lucas De Marchi
<lucas.de.mar...@gmail.com> wrote:
>
> On Fri, Oct 4, 2019 at 2:57 AM Matthias Maennich <maenn...@google.com> wrote:
> >
> > depmod in its current version is not aware of symbol namespace in
> > ksymtab entries introduced with 8651ec01daed ("module: add support for
> > symbol namespaces."). They have the form
> >
> >   __ksymtab_NAMESPACE.symbol_name
> >
> > A fix for kmod's depmod has been proposed [1]. In order to support older
> > versions of depmod as well, create a System.map.no_namespaces during
> > scripts/depmod.sh that has the pre-namespaces format. That way users do
> > not immediately upgrade the userspace tool.
> >
> > [1] 
> > https://lore.kernel.org/linux-modules/20191004094136.166621-1-maenn...@google.com/
> >
> > Reported-by: Stefan Wahren <stefan.wah...@i2se.com>
> > Fixes: 8651ec01daed ("module: add support for symbol namespaces.")
> > Cc: Masahiro Yamada <yamada.masah...@socionext.com>
> > Cc: Lucas De Marchi <lucas.de.mar...@gmail.com>
> > Cc: Jessica Yu <j...@kernel.org>
> > Cc: Martijn Coenen <m...@android.com>
> > Cc: Greg Kroah-Hartman <gre...@linuxfoundation.org>
> > Cc: linux-modu...@vger.kernel.org
> > Signed-off-by: Matthias Maennich <maenn...@google.com>
> > ---
> >
> > Please note this depends on the new ksymtab entry format proposed in
> > https://lore.kernel.org/lkml/20191003075826.7478-2-yamada.masah...@socionext.com/
>
> I don't really agree with that thought, more below.
>
> >
> > That is likely to be merged soon as well as it fixes problems in 5.4-rc*, 
> > hence
> > this patch depends on it.
> >
> > Cheers,
> > Matthias
> >
> >  .gitignore        | 1 +
> >  scripts/depmod.sh | 8 +++++++-
> >  2 files changed, 8 insertions(+), 1 deletion(-)
> >
> > diff --git a/.gitignore b/.gitignore
> > index 70580bdd352c..5ed58a7cb433 100644
> > --- a/.gitignore
> > +++ b/.gitignore
> > @@ -59,6 +59,7 @@ modules.order
> >  /vmlinux-gdb.py
> >  /vmlinuz
> >  /System.map
> > +/System.map.no_namespaces
> >  /Module.markers
> >  /modules.builtin.modinfo
> >
> > diff --git a/scripts/depmod.sh b/scripts/depmod.sh
> > index e083bcae343f..602e1af072c7 100755
> > --- a/scripts/depmod.sh
> > +++ b/scripts/depmod.sh
> > @@ -39,7 +39,13 @@ if $depmod_hack_needed; then
> >         KERNELRELEASE=99.98.$KERNELRELEASE
> >  fi
> >
> > -set -- -ae -F System.map
> > +# Older versions of depmod do not support symbol namespaces in ksymtab 
> > entries,
> > +# hence create an alternative System.map with namespace patched out to use 
> > for
> > +# depmod. I.e. transform entries as follows:
> > +#    __ksymtab_NAMESPACE.symbol_name -> __ksymtab_symbol_name
> > +sed 's/__ksymtab_.*\./__ksymtab_/' System.map > System.map.no_namespaces
>
> So people with old kmod will have to know they need to pass
> System.map.no_namespaces rather than the usual
> System.map. Also, distros will need to be update to also copy the new
> file to the kernel package (or upgrade/patch kmod).
>
> I'd rather maintain the current format and fix the bug that patch is
> fixing. The namespace
> in the end IMO is just a small annoyance with a reason to  exist.

I agree, this fix is bad.
We should not bother kmod or any tools.
And System.map.no_namespaces is a cheesy workaround.


BTW, I expressed my negative opinion in the review process
for the patch set. I am still not convinced with the
namespace feature, but anyway it was merged
(with poor review and test).



Get back on track, probably the right fix would be to
stop using __ksymtab_<namespace>.<symbol>.

It is not used for any purposes but passing
<namespace> / <symbol> pairs to modpost.


For example, __kstrtabns_##sym points to
the namespace string, so it would be possible
to parse it from modpost?

Then, asm("__ksymtab_" #ns NS_SEPARATOR #sym)
will go away.



Masahiro




> Lucas De Marchi
>
> > +
> > +set -- -ae -F System.map.no_namespaces
> >  if test -n "$INSTALL_MOD_PATH"; then
> >         set -- "$@" -b "$INSTALL_MOD_PATH"
> >  fi
> > --
> > 2.23.0.581.g78d2f28ef7-goog
> >
>
>
> --
> Lucas De Marchi



-- 
Best Regards
Masahiro Yamada

Reply via email to