Re: [RFC][PATCH] x86/ftrace: Have ftrace trampolines turn read-only at the end of system boot up

[Josh Poimboeuf]

On Fri, May 01, 2020 at 09:24:04AM -0400, Steven Rostedt wrote:
> On Fri, 1 May 2020 00:17:06 -0500
> Josh Poimboeuf <jpoimboe@xxxxxxxxxx> wrote:
> 
> > > Would it be easier to just call a new __text_poke_bp() which skips the
> > > SYSTEM_BOOTING check, since you know the trampoline will always be
> > > read-only?
> > > 
> > > Like:  
> > 
> > early_trace_init() is called after mm_init(), so I thought it might
> > work, but I guess not:
> 
> Yeah, I was about to say that this happens before mm_init() ;-)

It happens *after* mm_init().  But now text_poke() has a dependency on
poking_init(), has a dependency on proc_caches_init(), which has a
dependency on kmem_cache_init_late(), etc.

So how early do you need early_trace_init()?  I'm assuming moving it to
after kmem_cache_init_late() would be too late.

> It's why we already have magic for enabling function tracing the first time.
> 
> Do you see anything wrong with this current solution? It probably needs
> more comments, but I wanted to get acceptance on the logic before I go and
> pretty it up and send a non RFC patch.

Assuming we can't get text_poke() working earlier, it seems reasonable
to me.

-- 
Josh