Re: [PATCH] tracing: Use linker magic instead of recasting ftrace_ops_list_func()

[Jann Horn]

On Thu, Jun 18, 2020 at 6:42 PM Steven Rostedt <rostedt@xxxxxxxxxxx> wrote:
>
> On Thu, 18 Jun 2020 01:12:37 +0200
> Jann Horn <jannh@xxxxxxxxxx> wrote:
>
> > static ftrace_func_t ftrace_ops_get_list_func(struct ftrace_ops *ops)
> > +static ftrace_asm_func_t ftrace_ops_get_list_func(struct ftrace_ops *ops)
> >  {
> > +#if FTRACE_FORCE_LIST_FUNC
> > +       return ftrace_ops_list_func;
> > +#else
> >         /*
> >          * If this is a dynamic, RCU, or per CPU ops, or we force list func,
> >          * then it needs to call the list anyway.
> >          */
> > -       if (ops->flags & (FTRACE_OPS_FL_DYNAMIC | FTRACE_OPS_FL_RCU) ||
> > -           FTRACE_FORCE_LIST_FUNC)
> > +       if (ops->flags & (FTRACE_OPS_FL_DYNAMIC | FTRACE_OPS_FL_RCU))
> >                 return ftrace_ops_list_func;
> >
> >         return ftrace_ops_get_func(ops);
>
> But ftrace_ops_get_func() returns ftrace_func_t type, wont this complain?

No, because we only compile this case under FTRACE_FORCE_LIST_FUNC==0,
which means ARCH_SUPPORTS_FTRACE_OPS, which means the preprocessor
turns all occurrences of ftrace_asm_func_t into ftrace_func_t.

Essentially my idea here is to take the high-level rule "you can only
directly call ftrace_func_t-typed functions from assembly if
ARCH_SUPPORTS_FTRACE_OPS", and encode it in the type system. And then
the compiler won't complain as long as we make sure that we never cast
between the two types under ARCH_SUPPORTS_FTRACE_OPS==0.