Hi,
On Mon, Jan 6, 2025 at 10:00 AM Elliot Berman
<elliot.berman@xxxxxxxxxxxxxxxx> wrote:
>
> On Thu, Jan 02, 2025 at 04:24:22PM -0800, Sami Tolvanen wrote:
> > Hi Elliot,
> >
> > On Mon, Dec 30, 2024 at 7:33 PM Elliot Berman
> > <elliot.berman@xxxxxxxxxxxxxxxx> wrote:
> > >
> > > On Fri, Dec 20, 2024 at 06:41:42PM +0100, Sebastian Andrzej Siewior wrote:
> > > > __module_address() can be invoked within a RCU section, there is no
> > > > requirement to have preemption disabled.
> > > >
> > > > I'm not sure if using rcu_read_lock() will introduce the regression that
> > > > has been fixed in commit 14c4c8e41511a ("cfi: Use
> > > > rcu_read_{un}lock_sched_notrace").
> > > >
> > >
> > > You can replace the rcu_read_lock_sched_notrace() with guard(rcu)().
> > > Regular rcu lock doesn't generate function traces, so the recursive loop
> > > isn't possible.
> > >
> > > I've tested:
> > > - the current kernel (no recursive loop)
> > > - Revert back to rcu_read_lock_sched() (fails)
> >
> > Which kernel version did you test? I assume something pre-KCFI as
> > arm64 doesn't use this code since v6.1.
> >
>
> Ah, thanks for calling me out. I dug a bit more, I thought I was looking
> at a recursive loop in the ftrace buffers, but was actually the expected
> behavior. When I tested on the other configurations, the stm dummy
> driver hadn't kicked in yet by the time I looked at the ftrace. Indeed,
> this function code is not used on arm64.
>
> I experimented with an x86 build as well and I was able to get the hang
> I remember seeing after some tweaks to force a CFI failure. Still,
> guard(rcu)() is okay by me :)
OK, great. That makes sense. Thanks for taking the time to test this!
Sami
