Re: [PATCH v4 16/27] tracing: Remove regular RCU context for _rcuidle tracepoints (again)

[Joel Fernandes]

On Fri, Mar 06, 2020 at 07:51:18AM -0800, Alexei Starovoitov wrote:
> On Fri, Mar 6, 2020 at 3:31 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> >
> > On Fri, Mar 06, 2020 at 11:43:35AM +0100, Peter Zijlstra wrote:
> > > On Fri, Feb 21, 2020 at 02:34:32PM +0100, Peter Zijlstra wrote:
> > > > Effectively revert commit 865e63b04e9b2 ("tracing: Add back in
> > > > rcu_irq_enter/exit_irqson() for rcuidle tracepoints") now that we've
> > > > taught perf how to deal with not having an RCU context provided.
> > > >
> > > > Signed-off-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
> > > > Reviewed-by: Steven Rostedt (VMware) <rostedt@xxxxxxxxxxx>
> > > > ---
> > > >  include/linux/tracepoint.h |    8 ++------
> > > >  1 file changed, 2 insertions(+), 6 deletions(-)
> > > >
> > > > --- a/include/linux/tracepoint.h
> > > > +++ b/include/linux/tracepoint.h
> > > > @@ -179,10 +179,8 @@ static inline struct tracepoint *tracepo
> > > >              * For rcuidle callers, use srcu since sched-rcu        \
> > > >              * doesn't work from the idle path.                     \
> > > >              */                                                     \
> > > > -           if (rcuidle) {                                          \
> > > > +           if (rcuidle)                                            \
> > > >                     __idx = srcu_read_lock_notrace(&tracepoint_srcu);\
> > > > -                   rcu_irq_enter_irqsave();                        \
> > > > -           }                                                       \
> > > >                                                                     \
> > > >             it_func_ptr = rcu_dereference_raw((tp)->funcs);         \
> > > >                                                                     \
> > > > @@ -194,10 +192,8 @@ static inline struct tracepoint *tracepo
> > > >                     } while ((++it_func_ptr)->func);                \
> > > >             }                                                       \
> > > >                                                                     \
> > > > -           if (rcuidle) {                                          \
> > > > -                   rcu_irq_exit_irqsave();                         \
> > > > +           if (rcuidle)                                            \
> > > >                     srcu_read_unlock_notrace(&tracepoint_srcu, __idx);\
> > > > -           }                                                       \
> > > >                                                                     \
> > > >             preempt_enable_notrace();                               \
> > > >     } while (0)
> > >
> > > So what happens when BPF registers for these tracepoints? BPF very much
> > > wants RCU on AFAIU.
> >
> > I suspect we needs something like this...
> >
> > diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
> > index a2f15222f205..67a39dbce0ce 100644
> > --- a/kernel/trace/bpf_trace.c
> > +++ b/kernel/trace/bpf_trace.c
> > @@ -1475,11 +1475,13 @@ void bpf_put_raw_tracepoint(struct bpf_raw_event_map *btp)
> >  static __always_inline
> >  void __bpf_trace_run(struct bpf_prog *prog, u64 *args)
> >  {
> > +       int rcu_flags = trace_rcu_enter();
> >         rcu_read_lock();
> >         preempt_disable();
> >         (void) BPF_PROG_RUN(prog, args);
> >         preempt_enable();
> >         rcu_read_unlock();
> > +       trace_rcu_exit(rcu_flags);
> 
> One big NACK.
> I will not slowdown 99% of cases because of one dumb user.
> Absolutely no way.

For the 99% usecases, they incur an additional atomic_read and a branch, with
the above.  Is that the concern? Just want to make sure we are talking about
same thing.

Speaking of slowdowns, you don't really need that rcu_read_lock/unlock()
pair in __bpf_trace_run() AFAICS. The rcu_read_unlock() can run into the
rcu_read_unlock_special() slowpath and if not, at least has branches.  Most
importantly, RCU is consolidated which means preempt_disable() implies
rcu_read_lock().

thanks,

 - Joel