Re: [PATCH v7 23/36] function_graph: Add a new exit handler with parent_ip and ftrace_regs

[Masami Hiramatsu (Google) <mhiramat@kernel.org>]

Hi Steve,

On Sun, 18 Feb 2024 11:53:28 +0900
Masami Hiramatsu (Google) <mhiramat@kernel.org> wrote:

> On Fri, 16 Feb 2024 17:51:08 +0900
> Masami Hiramatsu (Google) <mhiramat@kernel.org> wrote:
> 
> > > > @@ -798,10 +798,6 @@ ftrace_pop_return_trace(struct ftrace_graph_ret *trace, unsigned long *ret,
> > > >  
> > > >  	*index += FGRAPH_RET_INDEX;
> > > >  	*ret = ret_stack->ret;
> > > > -	trace->func = ret_stack->func;
> > > > -	trace->calltime = ret_stack->calltime;
> > > > -	trace->overrun = atomic_read(&current->trace_overrun);
> > > > -	trace->depth = current->curr_ret_depth;
> > > 
> > > There's a lot of information stored in the trace structure. Why not pass
> > > that to the new retregfunc?
> > > 
> > > Then you don't need to separate this code out.
> > 
> > Sorry, I couldn't catch what you meant, Would you mean to call
> > ftrace_pop_return_trace() before calling retregfunc()?? because some of the
> > information are found from ret_stack, which is poped from shadow stack.
> 
> Ah, sorry I got what you said. I think this `trace` is not usable for the new
> interface. Most of the information is only used for the function-graph tracer.
> For example, trace->calltime and trace->overrun, trace->depth are used only
> for the function-graph tracer, but not for the other tracers.
> 
> But yeah, this idea is considerable. It also allows us to just update
> entryfunc() and retfunc() to pass fgraph_regs and return address.

The reason why I didn't use the those for *regfunc() is not only those
have unused information, but those does not have some params.

 - ftrace_graph_ent only have current `func`, but entryregfunc()
    needs `parent_ip` (== return address)
 - ftrace_graph_ret only have current `func`, but retregfunc()
    needs `ret` (== return address) too.

If I update the ftrace_graph_ent/ret to add 'retaddr', we can just pass
ftrace_graph_ent/ret, ftrace_regs, and fgraph_ops to *regfunc().
Moreover, maybe we don't need *regfunc, but just update entryfunc/retfunc
to pass ftrace_regs *, which will be NULL if it is not supported.

Thank you,

-- 
Masami Hiramatsu (Google) <mhiramat@kernel.org>