On Thu, 6 Jul 2023 14:10:12 +0900 Masami Hiramatsu (Google) <mhiramat@xxxxxxxxxx> wrote: > With only Jiri's patch, following flow can happen; > > ------ > CPU1 CPU2 > call unregister_fprobe() > ... > __fprobe_handler() > rethook_hook() on probed function > unregister_ftrace_function() > return from probed function > rethook hooks > find rh->handler == fprobe_exit_handler > call fprobe_exit_handler() > rethook_free(): > set rh->handler = NULL; > return from unreigster_fprobe; > call fp->exit_handler() <- (*) > > (*) In this point, the exit handler is called after returning from > unregister_fprobe(). > ------ > > So, this patch changes it as following; > ------ > CPU1 CPU2 > call unregister_fprobe() > ... > rethook_stop(): > set rh->handler = NULL; > __fprobe_handler() > rethook_hook() on probed function > unregister_ftrace_function() > return from probed function > rethook hooks > find rh->handler == NULL > return from rethook > rethook_free() > return from unreigster_fprobe; > ------ > > I can also just put a synchronize_sched_rcu() right after rethook_free() > to wait for all running fprobe_exit_handler() too. > This makes more sense. Can you please add the above to the change log. Thanks, -- Steve
