Re: [PATCH] x86: fix improper .init-type section references

["Jan Beulich" <jbeulich@novell.com>]

>> The cacheinfo_cpu_notifier itself is called out from _cpu_down() which
>> is ... yep, *not* __cpuinit (and obviously *cannot* be so). _cpu_down()
>> is called from cpu_down() which is (thankfully!) protected inside
>> #ifdef CONFIG_HOTPLUG_CPU, so we've /just about/ escaped trouble
>> so far, but another of it's callers is disable_nonboot_cpus() which does
>> *not* depend on HOTPLUG_CPU, but is #ifdef'ed inside SUSPEND_SMP
>> (gargh!) instead, and ... is _not_ __cpuinit either (obviously, again).
>
>Wait, SUSPEND_SMP again depends on HOTPLUG_CPU, so there
>are no issues here in marking cache_remove_dev() __cpuexit
>(which effectively simply becomes #ifdef HOTPLUG_CPU) as all
>callsites that call out the notifier with CPU_DEAD/FROZEN are also
>going to be disabled. [ Ok, looks like using __cpuexit as this kind of
>pseudo-#ifdef HOTPLUG_CPU is a standard practice? ]
>
>But modpost will still complain (bogus warning) about calling __exit
>from __init (when HOTPLUG_CPU=n) for cacheinfo_cpu_callback()
>calling cache_remove_dev(), no?

No, it doesn't, at least not for me. And from a purely theoretical
perspective I don't think such references should be considered bad -
.exit.* should be discarded together with .init.* if unloading is
impossible (built-in or configured off), not before module/kernel
initialization. This is specifically because init code may want to utilize
exit code in case of initialization failure.

Jan