On Mon, Nov 06, 2017 at 05:12:03PM +0000, Arnd Bergmann wrote:
> On Mon, Nov 6, 2017 at 5:50 PM, Dave Martin <Dave.Martin@xxxxxxx> wrote:
> > On Mon, Nov 06, 2017 at 01:36:40PM +0000, Arnd Bergmann wrote:
> >> The register_sysctl() function has been around for five years with commit
> >> fea478d4101a ("sysctl: Add register_sysctl for normal sysctl users")
> >> but now that arm64 started using it, I ran into a compile error:
> >>
> >> arch/arm64/kernel/armv8_deprecated.c: In function 'register_insn_emulation_sysctl':
> >> arch/arm64/kernel/armv8_deprecated.c:257:2: error: implicit declaration of function 'register_sysctl'
> >
> > Hmmm, looks like I missed this combination in testing.
> >
> > I wonder whether ARMV8_DEPRECATED without SYSCTL is really a good idea
> > though: in that config, we build a lot of dead code and leak some
> > memory today. The default emulation is still potentially useful, but
> > all the support for runtime twiddling of the emulation modes becomes
> > useless.
> >
> > For parallel reasons, the SVE sysctl stuff is protected by #ifdef
> > CONFIG_SYSCTL, which is why I didn't get a similar splat there.
> >
> >
> > So, although this patch is superficially sensible, it may tend to hide
> > bugs: code that calls register_sysctl() when CONFIG_SYSCTL=n is
> > suspicious and probably needs review... no?
>
> I think your analysis for this code is correct, we waste a lot of memory
> if we do it like this. However in the general case where we register a
> statically allocated 'struct ctl_table', it wouldn't be an issue, because
> gcc could then eliminate all the dead code.
>
> Adding a CONFIG_SYSCTL #ifdef or Kconfig dependency would
> probably be reasonable for armv8_deprecated.c, for the rest of
> the kernel, having that wrapper is probably better.
>
> I don't really care how this gets fixed, as long as some solution gets
> merged.
Well, since your patch makes things more consistent, I'm happy to
Reviewed-by: Dave Martin <Dave.Martin@xxxxxxx>
The armv8_deprecated code looks suspicious in any case, so I'll propose
a fix for that separately -- adding a Kconfig dependency seems simplest.
Cheers
---Dave
