Date:	Tue, 13 Oct 2009 17:21:26 -0400
From:	Mathieu Desnoyers <mathieu.desnoyers@...ymtl.ca>
To:	Steven Rostedt <rostedt@...dmis.org>
Cc:	linux-kernel@...r.kernel.org, Ingo Molnar <mingo@...e.hu>,
	Andrew Morton <akpm@...ux-foundation.org>,
	Frederic Weisbecker <fweisbec@...il.com>
Subject: Re: [PATCH 1/5] [PATCH 1/5] function-graph/x86: replace unbalanced
	ret with jmp

* Steven Rostedt (rostedt@...dmis.org) wrote:
> On Tue, 2009-10-13 at 16:47 -0400, Mathieu Desnoyers wrote:
> > * Steven Rostedt (rostedt@...dmis.org) wrote:
> > > From: Steven Rostedt <srostedt@...hat.com>
> > > 
> > > The function graph tracer replaces the return address with a hook to
> > > trace the exit of the function call. This hook will finish by returning
> > > to the real location the function should return to.
> > > 
> > > But the current implementation uses a ret to jump to the real return
> > > location. This causes a imbalance between calls and ret. That is
> > > the original function does a call, the ret goes to the handler
> > > and then the handler does a ret without a matching call.
> > > 
> > > Although the function graph tracer itself still breaks the branch
> > > predictor by replacing the original ret, by using a second ret and
> > > causing an imbalance, it breaks the predictor even more.
> > > 
> > > This patch replaces the ret with a jmp to keep the calls and ret
> > > balanced. I tested this on one box and it showed a 1.7% increase in
> > > performance. Another box only showed a small 0.3% increase. But no
> > > box that I tested this on showed a decrease in performance by making this
> > > change.
> > 
> > This sounds exactly like what I proposed at LPC. I'm glad it shows
> > actual improvements.
> 
> This is what we discussed at LPC. We both were under the assumption that
> a jump would work. The question was how to make that jump without hosing
> registers.
> 
> We lucked out that this is the back end of the return sequence. Where we
> can still clobber callie registers. (just not the ones holding the
> return code).
> 
> > 
> > Just to make sure I understand, the old sequence was:
> > 
> > call fct
> >   call ftrace_entry
> >   ret to fct
> > ret to ftrace_exit
> > ret to caller
> > 
> > and you now have:
> > 
> > call fct
> >   call ftrace_entry
> >   ret to fct
> > ret to ftrace_exit
> > jmp to caller
> > 
> > Am I correct ?
> 
> Almost.
>  
> What it was:
> 
> call function
>   function:
>   call mcount
>      mcount:
>      call ftrace_entry
>        ftrace_entry:
>        mess up with return code of caller
>        ret
>      ret
> 
>    [function code]
> 
>    ret to ftrace_exit
>      ftrace_exit:
>      get real return
>      ret to original
> 
> So for the function we have 3 calls and 4 rets
> 
> Now we have:
> 
> What it was:
> 
> call function
>   function:
>   call mcount
>      mcount:
>      call ftrace_entry

Can we manage to change this call

>        ftrace_entry:
>        mess up with return code of caller
>        ret

.. and this ret for 2 jmp instructions too ?

Given that we have no choice but to kill call/ret prediction logic, I
think it might be good to try to use this logic as little as possible
(by favoring jmp jmp over call/ret when the return target is invariant).

That's just an idea, benchmarks could prove me right/wrong.

Mathieu

>      ret
> 
>    [function code]
> 
>    ret to ftrace_exit
>      ftrace_exit:
>      get real return
>      jmp to original
> 
> Now we have 3 calls and 3 rets
> 
> Note the first call still does not match the ret, but we don't do two
> rets anymore.
> 
> -- Steve
> 
> 
> 

-- 
Mathieu Desnoyers
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists