Re: [PATCH] netdevsim: remove incorrect __net_initdata annotations

From: Arnd Bergmann
Date: Wed Apr 04 2018 - 12:03:49 EST


On Wed, Apr 4, 2018 at 5:52 PM, David Miller <davem@xxxxxxxxxxxxx> wrote:
> From: Arnd Bergmann <arnd@xxxxxxxx>
> Date: Wed, 4 Apr 2018 14:12:39 +0200
>
>> The __net_initdata section cannot currently be used for structures that
>> get cleaned up in an exitcall using unregister_pernet_operations:
>>
>> WARNING: vmlinux.o(.text+0x868c34): Section mismatch in reference from the function nsim_devlink_exit() to the (unknown reference) .init.data:(unknown)
>> The function nsim_devlink_exit() references
>> the (unknown reference) __initdata (unknown).
>> This is often because nsim_devlink_exit lacks a __initdata
>> annotation or the annotation of (unknown) is wrong.
>> WARNING: vmlinux.o(.text+0x868c64): Section mismatch in reference from the function nsim_devlink_init() to the (unknown reference) .init.data:(unknown)
>> WARNING: vmlinux.o(.text+0x8692bc): Section mismatch in reference from the function nsim_fib_exit() to the (unknown reference) .init.data:(unknown)
>> WARNING: vmlinux.o(.text+0x869300): Section mismatch in reference from the function nsim_fib_init() to the (unknown reference) .init.data:(unknown)
>>
>> As that warning tells us, discarding the structure after a module is
>> loaded would lead to a undefined behavior when that module is removed.
>>
>> It might be possible to change that annotation so it has no effect for
>> loadable modules, but I have not figured out exactly how to do that, and
>> we want this to be fixed in -rc1.
>>
>> This just removes the annotations, just like we do for all other such
>> modules.
>>
>> Fixes: 37923ed6b8ce ("netdevsim: Add simple FIB resource controller via devlink")
>> Signed-off-by: Arnd Bergmann <arnd@xxxxxxxx>
>
> Hmmm...
>
> __net_initdata defines to nothing if network namespaces are enabled.
>
> That's the whole point.
>
> Therefore, if NETNS is disabled, "unregister_pernet_operations" cannot
> happen and this is the only time when __net_initdata actually does
> something.

The problem happens with CONFIG_NETNS=n: __net_initdata turns into
__initdata, and nsim_fib_net_ops is referenced from a function that is
not marked __init (i.e. nsim_fib_exit).

The logic to turn __net_initdata into __init only makes sense for
subsystems that have no way to be unregistered, i.e. code which is
always built-in, or non-unloadable modules.

Arnd