On Fri, Feb 14, 2025 at 05:30:17PM -0500, Steven Rostedt wrote:
> On Thu, 6 Feb 2025 10:27:20 -0500
> Steven Rostedt <rost...@goodmis.org> wrote:
>
> > > BTW, do we really need to disable preempt or is it enough to call
> > > rcu_read_lock()?
> >
> > Bah, as I expected this function to be changed, I didn't spend too much
> > time on looking at its implementation. I just cut and pasted how the other
> > loops worked. But yes, it should not be disabling preemption. In fact, I
> > think the module code itself should not be disabling preemption!
> >
> > I'll have to go and look into that.
>
> It really looks like it requires preempt_disable(), as the code in
> kernel/module/main.c has in several places:
>
> preempt_disable();
>
> list_for_each_entry_rcu(mod, &modules, list) {
> [..]
> }
>
> preempt_enable();
>
> Or
>
> module_assert_mutex_or_preempt();
>
> [..]
>
> list_for_each_entry_rcu(mod, &modules, list,
> lockdep_is_held(&module_mutex)) {
>
>
> So it looks like it either requires preempt_disable or holding the
> module_mutex.
>
> As I need to call this with trace_types_lock held, and there's a place
> where trace_types_lock is within a module callback, I don't think it's safe
> to take that lock in that loop, otherwise we have the ABBA deadlock.
>
> Luis,
>
> Is this patch OK, and also is there any plan to move the module code to
> using just rcu_read_lock instead of preempt_disable?
The patch is not OK, you're looking at old code, look at
modules-next and as Petr suggested look at Sebastian's recently
merged work.
git remote add korg-modules
git://git.kernel.org/pub/scm/linux/kernel/git/modules/linux.git
Luis
