Re: [PATCH 2/2] scripts: ftrace - move the nop-processing in ftrace_init to compile time

[Yinan Liu]

This is my GCC version: GCC version 4.8.5 20150623 (red hat 4.8.5-44) 
(GCC) .

In fact, I see the make_nop processing in recordmcount. I'm still 
confused why this part can be directly replaced?


在 2021/9/11 下午10:12, Steven Rostedt 写道:
> On Sat, 11 Sep 2021 21:50:43 +0800
> Yinan Liu <yinan@xxxxxxxxxxxxxxxxx> wrote:
>
> > When ftrace is enabled, ftrace_init will consume a period of
> > time, usually around 15~20ms. Approximately 60% of the time is
> > consumed by nop-processing. Moving the nop-processing to the
> > compile time can speed up the kernel boot process.
> >
> > performance test:
> >          env:    Intel(R) Xeon(R) CPU E5-2682 v4 @ 2.50GHz
> >          method: before and after patching, compare the
> >                  total time of ftrace_init(), and verify
> >                  the functionality of ftrace.
> >
> >          avg_time of ftrace_init:
> >                  with patch: 7.114ms
> >                  without patch: 15.763ms
> What compiler are you using? Because by default, gcc should already do
> this for you. In fact, recordmcount isn't even called with the latest
> gcc, as gcc creates mcount_loc and inserts nops.
>
> This was implemented before, but because we use to have "ideal nops"
> that was determined at run time, because the different CPUs had
> different efficiency on what nop was used, we had to do it at run time.
> But that is no longer the case today, so we can revisit this.
>
> > Signed-off-by: Yinan Liu <yinan@xxxxxxxxxxxxxxxxx>
> > ---
> >   kernel/trace/ftrace.c  |  4 ++++
> >   scripts/recordmcount.h | 14 ++++++++++++++
> >   2 files changed, 18 insertions(+)
> >
> > diff --git a/kernel/trace/ftrace.c b/kernel/trace/ftrace.c
> > index c236da868990..ae3fba331179 100644
> > --- a/kernel/trace/ftrace.c
> > +++ b/kernel/trace/ftrace.c
> > @@ -6261,6 +6261,10 @@ static int ftrace_process_locs(struct module *mod,
> >   	 * until we are finished with it, and there's no
> >   	 * reason to cause large interrupt latencies while we do it.
> >   	 */
> > +#if defined CONFIG_X86 || defined CONFIG_X86_64 || defined CONFIG_ARM || defined CONFIG_ARM64
> We don't list archs in generic files. The above needs to be something like:
>
> #ifdef ARCH_HAS_MCOUNT_NOP
>
> or some name like that, and then that macro gets defined by the arch
> header (include/asm/ftrace.h)
>
>
>
> > +	ret = 0;
> > +	goto out;
> > +#endif
> space should be here.
>
> >   	if (!mod)
> >   		local_irq_save(flags);
> >   	ftrace_update_code(mod, start_pg);
> -- Steve