On Wed, 26 Jun 2019, Petr Mladek wrote:

On Wed 2019-06-26 10:22:45, Miroslav Benes wrote:
It is similar problem that has been solved by 2d1e38f56622b9bb5af8
("kprobes: Cure hotplug lock ordering issues"). This commit solved
it by always taking cpu_hotplug_lock.rw_sem before text_mutex inside.

If we follow the lock ordering then ftrace has to take text_mutex
only when stop_machine() is not called or from code called via
stop_machine() parameter.

This is not easy with the current design. For example, arm calls
set_all_modules_text_rw() already in ftrace_arch_code_modify_prepare(),
see arch/arm/kernel/ftrace.c. And it is called:

  + outside stop_machine() from ftrace_run_update_code()
  + without stop_machine() from ftrace_module_enable()

A conservative solution for 5.2 release would be to move text_mutex
locking from the generic kernel/trace/ftrace.c into
arch/x86/kernel/ftrace.c:

   ftrace_arch_code_modify_prepare()
   ftrace_arch_code_modify_post_process()

It should be enough to fix the original problem because
x86 is the only architecture that calls set_all_modules_text_rw()
in ftrace path and supports livepatching at the same time.

Looks correct, but I've paged out all the gory details vs. lock ordering in
that area.

Thanks,

	tglx