Re: [PATCH] ipmi: avoid atomic_inc in exit function

[Arnd Bergmann]

On Mon, Apr 15, 2019 at 7:39 PM Corey Minyard <minyard@xxxxxxx> wrote:
>
> On Mon, Apr 15, 2019 at 09:40:22AM -0700, Christoph Hellwig wrote:
> > On Mon, Apr 15, 2019 at 05:55:00PM +0200, Arnd Bergmann wrote:
> > > This causes a link failure on ARM in certain configurations,
> > > when we reference each atomic operation from .alt.smp.init in
> > > order to patch out atomics on non-SMP systems:
> > >
> > > `.exit.text' referenced in section `.alt.smp.init' of drivers/char/ipmi/ipmi_msghandler.o: defined in discarded section `.exit.text' of drivers/char/ipmi/ipmi_msghandler.o
> > >
> > > In this case, we can trivially replace the atomic_inc() with
> > > an atomic_set() that has the same effect and does not require
> > > a fixup.
> >
> > I'd rather fÑx the arm section management.  Using atomic in exit
> > routines is perfectly valid, and it would seem odd to forbid it.
>
> That was my first thought, too.  It's kind of hard to believe that
> the IPMI driver is the only thing that does an atomic_inc() in the
> exit code.

That's what I had thought as well at first, and I carried a patch
to work around this by not dropping the .text.exit section on ARM
when SMP patching is enabled for a few years. I never sent this
because that can waste a significant amount of kernel memory,
and I knew the warning is harmless.

When revisiting it now, I found that this one was the only instance
I ever hit. It seems to be that using atomics in module_exit() is
indeed odd, because the function is rarely concurrent with anything
else.

        Arnd