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

[Thomas Monjalon <thomas@monjalon.net>]

05/05/2020 09:17, Jerin Jacob:
> On Tue, May 5, 2020 at 12:31 PM Thomas Monjalon <thomas@monjalon.net> wrote:
> > 05/05/2020 05:43, Jerin Jacob:
> > > On Tue, May 5, 2020 at 3:01 AM Thomas Monjalon <thomas@monjalon.net> wrote:
> > > > 04/05/2020 19:54, Jerin Jacob:
> > > > > On Mon, May 4, 2020 at 11:10 PM David Marchand
> > > > > > On Mon, May 4, 2020 at 7:19 PM Jerin Jacob <jerinjacobk@gmail.com> wrote:
> > > > > > > On Mon, May 4, 2020 at 10:38 PM David Marchand
> > > > > > > > On Mon, May 4, 2020 at 4:39 PM Jerin Jacob <jerinjacobk@gmail.com> wrote:
> > > > > > > > > On Mon, May 4, 2020 at 7:34 PM David Marchand <david.marchand@redhat.com> wrote:
> > > > > > > > > > On Mon, May 4, 2020 at 4:47 AM Jerin Jacob <jerinjacobk@gmail.com> wrote:
> > > > > > > > > > > On Mon, May 4, 2020 at 2:02 AM David Marchand <david.marchand@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.
> > > >
> > > > No, it is not obvious.
> > > > The version from David looks simpler to use and to understand.
> > > > Without any number, I consider usability (and maintenance) wins.
> > >
> > > Thanks for the feedback.
> > >
> > > As the trace maintainer, I would like not to explode constructor usage
> > > for trace library.
> > > My reasoning, We could do trace registration without this constructor scheme.
> > ???
> 
> We don't need this patch to make trace to work.
> 
> >
> >
> > > If you think, it is better usability, lets add an option for rte_log
> > > for the constructor scheme.
> >
> > It makes non-sense.
> > rte_log requires only one function call per log type.
> 
> Here is the example of the log registration:
> 
> global variable:
> int otx2_logtype_base;
> int otx2_logtype_mbox;
> int otx2_logtype_npa;
> 
> RTE_INIT(otx2_log_init);
> static void
> otx2_log_init(void)
> {
>         otx2_logtype_base = rte_log_register("pmd.octeontx2.base");
>         otx2_logtype_mbox = rte_log_register("pmd.octeontx2.mbox");
>         otx2_logtype_npa = rte_log_register("pmd.mempool.octeontx2");
> }
> 
> What the proposed patch here.
> # Making N constructors from one
> # Grouping global variable and register function under a single Marco
> and making it as N constructors.
> Why can we do the same logic for rte_log?

rte_log is simple, there is nothing to simplify.

This comparison makes no sense.


> > rte_trace requires 3 macros calls per trace type:
> > RTE_TRACE_POINT_REGISTER, RTE_TRACE_POINT_DEFINE, RTE_TRACE_POINT_ARGS
> > This patch is unifying the first 2 macro calls to make usage simpler,
> > and ease rte_trace adoption.
> 
> RTE_TRACE_POINT_ARGS is NOP and for the syntax.
> It is similar to rte_log. rte_log don't have RTE_TRACE_POINT_REGISTER instead
> it is creating global variable  see, "int otx2_logtype_base;
> 
> >
> > Note: the other usability weirdness is mandating declaring each trace
> > function with a magic double underscore prefix which appears nowhere else.
> >
> >
> > > Analyze the impact wrt boot time and cross-platform pov as the log
> > > has a lot of entries to test. If the usage makes sense then it should make sense
> > > for rte_log too. I would like to NOT have trace to be the first
> > > library to explode
> > > with the constructor scheme. I suggest removing this specific patch from RC2 and
> > > revisit later.
> >
> > You still did not give any argument against increasing the number
> > of constructors.
> 
> If you are proposing the new scheme, you have to prove the overhead
> with a significant number of constructors
> and why it has differed from existing scheme of things. That's is the
> norm in opensource.

I say there is no overhead.
The target is to simplify the usage and I prove it:
	1 call replacing 2 calls.