Re: [patch V6 04/37] x86: Make hardware latency tracing explicit

[Andy Lutomirski]

On Sun, May 17, 2020 at 1:48 AM Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
>
> Andy Lutomirski <luto@xxxxxxxxxx> writes:
> > On Fri, May 15, 2020 at 5:10 PM Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
> >>
> >>
> >> The hardware latency tracer calls into trace_sched_clock and ends up in
> >> various instrumentable functions which is problemeatic vs. the kprobe
> >> handling especially the text poke machinery. It's invoked from
> >> nmi_enter/exit(), i.e. non-instrumentable code.
> >>
> >> Use nmi_enter/exit_notrace() instead. These variants do not invoke the
> >> hardware latency tracer which avoids chasing down complex callchains to
> >> make them non-instrumentable.
> >>
> >> The real interesting measurement is the actual NMI handler. Add an explicit
> >> invocation for the hardware latency tracer to it.
> >>
> >> #DB and #BP are uninteresting as they really should not be in use when
> >> analzying hardware induced latencies.
> >>
> >
> >> @@ -849,7 +851,7 @@ static void noinstr handle_debug(struct
> >>  static __always_inline void exc_debug_kernel(struct pt_regs *regs,
> >>                                              unsigned long dr6)
> >>  {
> >> -       nmi_enter();
> >> +       nmi_enter_notrace();
> >
> > Why can't exc_debug_kernel() handle instrumentation?  We shouldn't
> > recurse into #DB since we've already cleared DR7, right?
>
> It can later on. The point is that the trace stuff calls into the world
> and some more before the entry handling is complete.
>
> Remember this is about ensuring that all the state is properly
> established before any of this instrumentation muck can happen.
>
> DR7 handling is specific to #DB and done even before nmi_enter to
> prevent recursion.

So why is this change needed?