Re: On trace_*_rcuidle functions in modules

[Steven Rostedt <rostedt@goodmis.org>]

On Wed, 15 Apr 2020 18:02:58 -0700
Bjorn Andersson <bjorn.andersson@linaro.org> wrote:

> Forgive me, but how is this problem related to the fact that the code is
> dynamically loaded, i.e. encapsulated in a module?

It's not.

> 
> Per the example earlier in this thread, the thing we're worried about is
> a use after free in the following scenario, right?
> 
>         cpu_pm_unregister_notifier(my_notifier);
> 	kfree(my_data);
> 
> But a driver implementing this snippet might do this regardless of being
> builtin or module and afaict exiting probe() unsuccessfully or unbinding
> the device would risk triggering this issue?

I know my email was confusing. I was talking about a bug that does not
exist. (There is no bug!)

The reason is that rcu is enabled during the call to the notifiers. The
above assumes that the my_data usage in the notifier callback is
surrounded by rcu_read_lock() (otherwise it's broken regardless of this
code or not). The above unregister will call synchronize_rcu() after it
removes the notifier which will guarantee that the rcu_read_lock()
critical sections would be completed. Then the kfree(my_data) can free
my_data with no possible users.

-- Steve