Re: [PATCH V4 08/13] ftrace: Add perf text poke events for ftrace trampolines

[Peter Zijlstra <peterz@infradead.org>]

On Wed, Apr 01, 2020 at 01:42:50PM +0300, Adrian Hunter wrote:
> On 1/04/20 1:09 pm, Peter Zijlstra wrote:
> > On Wed, Mar 04, 2020 at 11:06:28AM +0200, Adrian Hunter wrote:
> >> Add perf text poke events for ftrace trampolines when created and when
> >> freed.
> > 
> > If I'm not mistaken that ends up like so:
> > 
> > static void ftrace_update_trampoline(struct ftrace_ops *ops)
> > {
> > +       unsigned long trampoline = ops->trampoline;
> > +
> > 	arch_ftrace_update_trampoline(ops);
> > +       if (ops->trampoline && ops->trampoline != trampoline &&
> >> +	    (ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP)) {
> >> +		/* Add to kallsyms before the perf events */
> > +               ftrace_add_trampoline_to_kallsyms(ops);
> >> +		perf_event_ksymbol(PERF_RECORD_KSYMBOL_TYPE_OOL,
> >> +				   ops->trampoline, ops->trampoline_size, false,
> >> +				   FTRACE_TRAMPOLINE_SYM);
> >> +		/*
> >> +		 * Record the perf text poke event after the ksymbol register
> >> +		 * event.
> >> +		 */
> >> +		perf_event_text_poke((void *)ops->trampoline, NULL, 0,
> >> +				     (void *)ops->trampoline,
> >> +				     ops->trampoline_size);
> > 	}
> > }
> > 
> > And afaict, that is wrong.
> > 
> > The thing is; arch_ftrace_update_trampoline() can actually *update* an
> > existing trampoline, as per the name. Yes it also creates a trampoline
> > if there isn't one already, but if there already is one, it will modify
> > it in-place.
> > 
> > I see the appeal of having this event in generic code; but I'm thinking
> > you'll need the update even in arch code anyway, at which point it'd
> > probably be easier to do all of this in arch code.
> 
> For x86, we use text_poke_bp for updates which already does text_poke events
> via text_poke_bp_batch.

Ah, indeed! Damn, I even wrote that code :/

> It might be reasonable to assume other architectures will also need to put
> updates through a common text poker which will take care of text_poke events.

You'd better look, I recently rewrote the x86/ftrace code to be 'sane'
and not re-implement all the text poking stuff itself.

But I suppose that any arch adding support for this can fix that up if
needed.