Re: [RFC][PATCH 7/7] livepatch,context_tracking: Avoid disturbing NOHZ_FULL tasks

[Petr Mladek]

On Thu 2021-09-23 15:28:56, Peter Zijlstra wrote:
> On Thu, Sep 23, 2021 at 03:14:48PM +0200, Petr Mladek wrote:
> 
> > IMHO, this is not safe:
> > 
> > CPU0				CPU1
> > 
> > klp_check_task(A)
> >   if (context_tracking_state_cpu(task_cpu(task)) == CONTEXT_USER)
> >      goto complete;
> > 
> >   clear_tsk_thread_flag(task, TIF_PATCH_PENDING);
> > 
> > 				# task switching to kernel space
> > 				klp_update_patch_state(A)
> > 				       if (test_and_clear_tsk_thread_flag(task,	TIF_PATCH_PENDING))
> > 				       //false
> > 
> > 				# calling kernel code with old task->patch_state
> > 
> > 	task->patch_state = klp_target_state;
> > 
> > BANG: CPU0 sets task->patch_state when task A is already running
> > 	kernel code on CPU1.
> 
> Why is that a problem? That is, who actually cares about
> task->patch_state ? I was under the impression that state was purely for
> klp itself, to track which task has observed the new state.

It is the other way. The patch_state is used in klp_ftrace_handler()
to decide which code must be used (old or new).

The state must change only when the given task is _not_ using
any patched function. Hence we do it when:

     + no patched function is on the stack  (needed primary for kthreads)
     + entering/leaving kernel              (reliable way for user space)

See "Consistency model" in Documentation/livepatch/livepatch.rst
for more details.

Best Regards,
Petr