Re: LTP tracing crashed on arm64 rk3399-rock-pi-4 - pc : ftrace_ops_test

[Steven Rostedt]

On Thu, 13 Jun 2024 13:29:58 +0530
Naresh Kamboju <naresh.kamboju@xxxxxxxxxx> wrote:

> > --- a/kernel/trace/fgraph.c
> > +++ b/kernel/trace/fgraph.c
> > @@ -641,7 +641,7 @@ int function_graph_enter(unsigned long ret, unsigned long func,
> >         {
> >                 for_each_set_bit(i, &fgraph_array_bitmask,
> >                                          sizeof(fgraph_array_bitmask) * BITS_PER_BYTE) {
> > -                       struct fgraph_ops *gops = fgraph_array[i];
> > +                       struct fgraph_ops *gops = READ_ONCE(fgraph_array[i]);
> >                         int save_curr_ret_stack;
> >
> >                         if (gops == &fgraph_stub)
> >
> >
> > Because if the compiler decides to re-read gops from fgraph_array[i] after the
> > above check for the following line that does:
> >
> >                         save_curr_ret_stack = current->curr_ret_stack;
> >                         if (ftrace_ops_test(&gops->ops, func, NULL) &&
> >                             gops->entryfunc(&trace, gops))
> >                                 bitmap |= BIT(i);
> >
> >
> > and gops now points to fgraph_stub, it will trigger this bug.
> >
> > Can you apply the above change and see if the bug goes away?  
> 
> I will apply this patch and run the test in a loop.
> Since it is only seen once. Not sure I could validate this and confirm.

We could just look at the code that clang produced and see if it
accesses the fgraph_array[] again. If this was the cause, it would show
up it in the code.

But regardless, I think I'm going to add that READ_ONCE() anyway,
because it is legitimate for the compiler to do the above without it.

Thanks,

-- Steve