RE: [PATCH v4] eal: add build-time option to omit trace

["Morten Brørup" <mb@smartsharesystems.com>]

> From: Jerin Jacob [mailto:jerinjacobk@gmail.com]
> Sent: Wednesday, 2 October 2024 18.18
> 
> On Tue, Oct 1, 2024 at 9:45 PM Morten Brørup <mb@smartsharesystems.com>
> wrote:
> >
> > > From: Jerin Jacob [mailto:jerinjacobk@gmail.com]
> > > Sent: Tuesday, 1 October 2024 18.06
> > >
> > > On Tue, Oct 1, 2024 at 9:32 PM Morten Brørup
> <mb@smartsharesystems.com>
> > > wrote:
> > > >
> > > > > From: Jerin Jacob [mailto:jerinjacobk@gmail.com]
> > > > > Sent: Tuesday, 1 October 2024 17.02
> > > > >
> > > > > On Tue, Oct 1, 2024 at 7:44 PM Morten Brørup
> > > <mb@smartsharesystems.com>
> > > > > wrote:
> > > > > >
> > > > > > > From: Jerin Jacob [mailto:jerinjacobk@gmail.com]
> > > > > > > Sent: Tuesday, 1 October 2024 16.05
> > > > > > >
> > > > > > > On Tue, Oct 1, 2024 at 7:19 PM Morten Brørup
> > > > > <mb@smartsharesystems.com>
> > > > > > > wrote:
> > > > > > > >
> > > > > > > > Jerin,
> > > > > > > >
> > > > > > > > If you have no further comments, please add review/ack
> tag,
> > > to
> > > > > help
> > > > > > > Thomas see that the patch has been accepted by the
> maintainer,
> > > and
> > > > > can
> > > > > > > be merged.
> > > > > > >
> > > > > > > There was a comment to make the function as
> > > rte_trace_is_enabled()
> > > > > and
> > > > > > > remove internal. The rest looks good to me. I will Ack in
> the
> > > next
> > > > > > > version.
> > > > > >
> > > > > > Perhaps my reply to that comment was unclear... such a public
> > > > > function already exists in the previous API:
> > > > >
> > > > > I see. It was not clear.
> > > > >
> > > > > >
> > > > >
> > >
> https://elixir.bootlin.com/dpdk/v24.07/source/lib/eal/include/rte_trace
> > > > > .h#L36
> > > > > >
> > > > > > That function tells if trace enabled at both build time and
> > > runtime,
> > > > > and returns false if not.
> > > > > >
> > > > > > A separate public function to tell if trace is enabled at
> build
> > > time
> > > > > seems like overkill to me. Is that what you are asking for?
> > > > >
> > > > > No. Just use rte_trace_is_enabled() in app/test instead of
> > > > > __rte_trace_point_generic_is_enabled() as it is internal.
> > > >
> > > > Just tested it, and it didn't have the wanted effect.
> > > > I think rte_trace_is_enabled() returns false until at least one
> > > tracepoint has been enabled, which seems like a good optimization.
> > > > But it also means that we cannot use it to replace
> > > __rte_trace_point_generic_is_enabled() in test/app, because no
> > > tracepoints have been enabled at this point of execution, so it
> returns
> > > false here.
> > > >
> > > > I looked around in the code, and cannot find a method without
> looking
> > > at internals, or duplicating a test case.
> > > >
> > > > I could test if rte_trace_point_lookup("app.dpdk.test.tp")
> returns
> > > non-NULL, but that would duplicate the same test in
> > > test_trace_points_lookup().
> > > >
> > > > What do you think...
> > > > Keep using internal function
> __rte_trace_point_generic_is_enabled(),
> > > > test rte_trace_point_lookup("app.dpdk.test.tp") != NULL,
> > > > or any other idea?
> > >
> > > How about the following, it is anyway the correct thing to do
> > >
> > > bool
> > > rte_trace_is_enabled(void)
> > > {
> > >  +       if (__rte_trace_point_generic_is_enabled() == false)
> > >   +              return false;
> > >         return rte_atomic_load_explicit(&trace.status,
> > > rte_memory_order_acquire) != 0;
> > > }
> > >
> >
> > It's the opposite that's causing problems:
> > Even when built with trace, rte_trace_is_enabled() returns false,
> because no trace points have been enabled when the test application
> checks if it should run test cases or not. At this point, trace.status
> is zero, so it skips testing.
> >
> > We don't need to add the test for
> rte_trace_point_generic_is_enabled()==false in rte_trace_is_enabled(),
> because nothing can increase trace.status if no tracepoints exist. (As
> far as I understand.)
> 
> OK. I see. IMO, It is not good to expose internal symbol to app/test.

Agree. Fixed in v5.

> How about,
> Changing __rte_trace_point_generic_is_enabled() as
> rte_trace_feature_is_enabled() or similar. This variant is more like
> feature is disabled at compiled time or not? And make it as public and
> use in app/test.

In v5, I have added the public function rte_trace_feature_is_enabled(), which returns true iff the application is built with trace enabled (this is tested using inline), and the DPDK - in case it is built separately - is built with trace enabled (this is compiled with the trace lib's C file).

App/test now uses this function instead of the internal one.

__rte_trace_point_generic_is_enabled() needs to remain static inline, due to the way it is used for omitting trace at build time.

PS: Both trace_autotest and trace_perf_autotest have been tested with and without RTE_TRACE, and they behave as expected.