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

Jerin Jacob jerinjacobk at gmail.com
Mon May 4 19:54:40 CEST 2020


On Mon, May 4, 2020 at 11:10 PM David Marchand
<david.marchand at redhat.com> wrote:
>
> On Mon, May 4, 2020 at 7:19 PM Jerin Jacob <jerinjacobk at gmail.com> wrote:
> >
> > On Mon, May 4, 2020 at 10:38 PM David Marchand
> > <david.marchand at redhat.com> wrote:
> > >
> > > On Mon, May 4, 2020 at 4:39 PM Jerin Jacob <jerinjacobk at gmail.com> wrote:
> > > >
> > > > On Mon, May 4, 2020 at 7:34 PM David Marchand <david.marchand at redhat.com> wrote:
> > > > >
> > > > > On Mon, May 4, 2020 at 4:47 AM Jerin Jacob <jerinjacobk at gmail.com> wrote:
> > > > > >
> > > > > > On Mon, May 4, 2020 at 2:02 AM David Marchand <david.marchand at redhat.com> wrote:
> > > > > > >
> > > > > > > RTE_TRACE_POINT_DEFINE and RTE_TRACE_POINT_REGISTER must come in pairs.
> > > > > > > Merge them and let RTE_TRACE_POINT_REGISTER handle the constructor part.
> > > > > >
> > > > > >
> > > > > > Initially, I thought of doing the same. But, later I realized that
> > > > > > this largely grows the number of constructors been called.
> > > > > > I had concerns about the boot time of the application and/or loading
> > > > > > the shared library, that the reason why spitting
> > > > > > as two so that constructor registers a burst of traces like rte_log.
> > > > >
> > > > > I am a bit skeptical.
> > > > > In terms of cycles and looking at __rte_trace_point_register() (which
> > > > > calls malloc), the cost of calling multiple constructors instead of
> > > > > one is negligible.
> > > >
> > > > We will have a lot tracepoints latter, each one translates to the
> > > > constructor may not be a good
> > > > improvement. The scope is limited only to register function so IMO it
> > > > is okay to have split
> > > > just like rte_log. I don't see any reason why it has to be a different
> > > > than rte_log.
> > >
> > > What is similar to rte_log?
> > > There is neither RTE_LOG_REGISTER macro, nor two-steps declaration of
> > > dynamic logtypes.
> >
> >
> > Here is an example of rte_log registration. Which has _one_
> > constructor and N number of
> > rte_log_register() underneath.
>
> rte_log is one thing, rte_trace is already different.
>
> There is _no macro_ in rte_log for registration.
> The reason being in that a rte_log logtype is a simple integer without
> any special declaration requiring a macro.

I just wrapped in macro for convincing, but it has the same semantics.
global variable and API/macro to register.


>
> For tracepoints, we have a special two steps thing: the tracepoint
> handle must be derived from the tracepoint name.
> Then this handle must be registered.
> What I proposed is to make life easier for developers that want to add
> tracepoints and I suppose you noticed patch 1 of this series.

To reduce the constructors. I don't want trace libraries to add lot of
constructors.
I don't think it simplifies a lot as the scope of only for registration.


>
>
> > > > One of the thought process is, we probably remove the constructor
> > > > scheme to all other with DPDK
> > > > and replace it with a more register scheme. In such a case, we can
> > > > skip calling the constructor all tother
> > > > when trace is disabled.
> > >
> > > Sorry, but I have a hard time understanding your point.
> > > Are you talking about application boot time?
> >
> > Yes. The optimization of application boottime time in case of static
> > binary and/or shared library(.so) load time.
>
> As Thomas mentioned, do you have numbers?

No. But I know, it is obvious that current code is better in terms of
boot time than the proposed one.
I would like to not add a lot of constructor for trace as the FIRST
module in DPDK.

>
>
> --
> David Marchand
>


More information about the dev mailing list