On Thu, 6 Jul 2023 09:56:24 -0400 Steven Rostedt <rostedt@xxxxxxxxxxx> wrote: > 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. OK, let me update it. Thanks! > > Thanks, > > -- Steve -- Masami Hiramatsu (Google) <mhiramat@xxxxxxxxxx>
