On Sun, 12 Nov 2017, Arnd Bergmann wrote:
> On Sun, Nov 12, 2017 at 10:16 PM, Thomas Gleixner <t...@linutronix.de> wrote:
> > On Mon, 6 Nov 2017, Arnd Bergmann wrote:
> >> The newly added function triggers a harmless Kbuild warning because
> >> of a missing annotation:
> >>
> >> WARNING: vmlinux.o(.text+0x448098): Section mismatch in reference from the
> >> function timer_of_exit() to the function .init.text:timer_clk_exit()
> >> The function timer_of_exit() references
> >> the function __init timer_clk_exit().
> >> This is often because timer_of_exit lacks a __init
> >> annotation or the annotation of timer_clk_exit is wrong.
> >>
> >> The function is only called from other __init functions, so it
> >> can safely be marked as __init as well.
> >
> > Hmm. I don't see any caller at all. From the intention of the patch I
> > assume this isn't designed for using from init functions, so we rather have
> > to remove the __init annotations from the called functions.
> >
> > Sudeep posted a patch which does that:
> >
> >
> > https://lkml.kernel.org/r/1509979716-10646-1-git-send-email-sudeep.ho...@arm.com
> >
> > Though I rather would know whether this function is going to be used at
> > all and what the intention of this patch was.
> >
> > Benjamin????
>
> My interpretation was that timer drivers are still supposed to be unregistered
> at module unload time, but that you might use the new timer_of_exit()
> in the failure path of whatever function calls timer_of_init() successfully
> when something fails in the next step.
>
> Sudeep's interpretation also makes sense, I had not thought of that, but
> I now found the patch that adds a user in an init function:
> https://www.mail-archive.com/linux-kernel@vger.kernel.org/msg1519644.html
>
> It seems I guessed right and Sudeep guessed wrong (both by pure chance
> I admit). Both patches solve the problem, Sudeep's version is a little
> more robust in case we ever add a caller in an __exit function (which I
> think is currently not allowed), while mine saves a little bit of memory
> and matches the current usage better.
Right, but if the only use case is the cleanup in an error path, then the
function name is a misnomer.
Thanks,
tglx
