Re: [dpdk-dev] [PATCH 2/8] trace: simplify trace point registration

[Thomas Monjalon <thomas@monjalon.net>]

05/05/2020 19:28, Jerin Jacob:
> On Tue, May 5, 2020 at 10:50 PM Thomas Monjalon <thomas@monjalon.net> wrote:
> > 05/05/2020 19:09, Jerin Jacob:
> > > On Tue, May 5, 2020 at 10:38 PM Jerin Jacob <jerinjacobk@gmail.com> wrote:
> > > > On Tue, May 5, 2020 at 10:28 PM Thomas Monjalon <thomas@monjalon.net> wrote:
> > > > > 05/05/2020 18:46, Jerin Jacob:
> > > > > > On Tue, May 5, 2020 at 9:58 PM David Marchand <david.marchand@redhat.com> wrote:
> > > > > > > On Tue, May 5, 2020 at 5:25 PM Jerin Jacob <jerinjacobk@gmail.com> wrote:
> > > > > > > > On Tue, May 5, 2020 at 5:56 PM Jerin Jacob <jerinjacobk@gmail.com> wrote:
> > > > > > > > > On Tue, May 5, 2020 at 5:06 PM David Marchand <david.marchand@redhat.com> wrote:
> > > > > > > > > > On Tue, May 5, 2020 at 12:13 PM Jerin Jacob <jerinjacobk@gmail.com> wrote:
> > > > > > > > > > > > > Please share the data.
> > > > > > > > > > > >
> > > > > > > > > > > > Measured time between first rte_trace_point_register and last one with
> > > > > > > > > > > > a simple patch:
> > > > > > > > > > >
> > > > > > > > > > > I will try to reproduce this, once we finalize on the above synergy
> > > > > > > > > > > with rte_log.
> > > > > > > > > >
> > > > > > > > > > I took the time to provide measure but you won't take the time to look at this.
> > > > > > > > >
> > > > > > > > > I will spend time on this. I would like to test with a shared library
> > > > > > > > > also and more tracepoints.
> > > > > > > > > I was looking for an agreement on using the constructor for rte_log as
> > > > > > > > > well(Just make sure the direction is correct).
> > > > > > > > >
> > > > > > > > > Next steps:
> > > > > > > > > - I will analyze the come back on this overhead on this thread.
> > > > > > > >
> > > > > > > > I have added 500 constructors for testing the overhead with the shared
> > > > > > > > build and static build.
> > > > > > > > My results inline with your results aka negligible overhead.
> > > > > > > >
> > > > > > > > David,
> > > > > > > > Do you have plan for similar RTE_LOG_REGISTER as mentioned earlier?
> > > > > > > > I would like to have rte_log and rte_trace semantics similar to registration.
> > > > > > > > If you are not planning to submit the rte_log patch then I can send
> > > > > > > > one for RC2 cleanup.
> > > > > > >
> > > > > > > It won't be possible for me.
> > > > > >
> > > > > > I can do that if we agree on the specifics.
> > > > > >
> > > > > >
> > > > > > >
> > > > > > > Relying on the current rte_log_register is buggy with shared builds,
> > > > > > > as drivers are calling rte_log_register, then impose a default level
> > > > > > > without caring about what the user passed.
> > > > > > > So if we introduce a RTE_LOG_REGISTER macro now at least this must be fixed too.
> > > > > > >
> > > > > > > What I wanted to do:
> > > > > > > - merge rte_log_register_and_pick_level() (experimental) into
> > > > > > > rte_log_register, doing this should be fine from my pov,
> > > > > > > - reconsider the relevance of a fallback logtype when registration fails,
> > > > > > > - shoot the default level per component thing: levels meaning is
> > > > > > > fragmented across the drivers/libraries because of it, but this will
> > > > > > > open a big box of stuff,
> > > > > >
> > > > > > This you are referring to internal implementation improvement. Right?
> > > > > > I was referring to remove the current clutter[1]
> > > > > > If we stick the following as the interface. Then you can do other
> > > > > > improvements when you get time
> > > > > > that won't change the consumer code or interference part.
> > > > > >
> > > > > > #define RTE_LOG_REGISTER(type, name, level)
> > > > >
> > > > > This discussion is interesting but out of scope for rte_trace.
> > > > > I am also interested in rte_log registration cleanup,
> > > > > but I know it is too much work for the last weeks of 20.05.
> > > > >
> > > > > As Olivier said about rte_trace,
> > > > > "Since it's a new API, it makes sense to make
> > > > > it as good as possible for the first version."
> > > > >
> > > > > So please let's conclude on this rte_trace patch for 20.05-rc2,
> > > > > and commit to fix rte_log registration in the first days of 20.08.
> > > >
> > > > Why not hold the trace registration patch 2/8 and apply rest for RC2.
> > > > Once we have synergy between the registration scheme between rte_log
> > > > and rte_trace
> > > > apply the patch for RC2.
> > >
> > > I meant, Once we have synergy between the registration scheme between
> > > rte_log and rte_trace
> > > apply the patch for _20.08_?
> >
> > Because of what I wrote above:
> > As Olivier said about rte_trace,
> > "Since it's a new API, it makes sense to make
> > it as good as possible for the first version."
> >
> > The intent is to show an API as simple as possible
> > in order to have a maximum of developers integrating it,
> > and getting more interesting feedbacks.
> >
> > In other words, we want to make your work shine for prime time.
> 
> I like that, If it is not shining just because of 2/8 not applying now
> then I fine with that.
> Anyway, it is an experimental API, There is still room to change and
> nothing is set and stone.
> For me, the synergy between log/trace interface important as trace
> needs to replace rte_log.

Now that I better understand what rte_trace (and tracing in general) is,
I believe rte_log cannot be replaced.
I think we can write logs in traces, as a log option, but it should be
just one possible output among others.

I think everybody agree to use one constructor per log type and
per trace type. We are ready to do this change for rte_trace first.
This is your call to accept it or not, even if don't understand
why you would like both to be done at the exact same time.