Re: [PATCH v2] eal: add build-time option to omit trace

[Jerin Jacob <jerinjacobk@gmail.com>]

On Thu, Sep 19, 2024 at 9:38 PM Morten Brørup <mb@smartsharesystems.com> wrote:
>
> > From: Jerin Jacob [mailto:jerinjacobk@gmail.com]
> > Sent: Thursday, 19 September 2024 17.49
> >
> > On Thu, Sep 19, 2024 at 9:05 PM Morten Brørup <mb@smartsharesystems.com>
> > wrote:
> > >
> > > > Could you share data with a real-world application of the elf size?
> > > > 1)Without any change
> > >
> > > Size of the statically linked executable on the target file system:
> > > 3,800,528 byte
> > >
> > > > 2)Only disabling via __rte_trace_point_emit_header_generic() .. aka below
> > > > patch.
> > >
> > > 3,572,032 byte
> > >
> > > > 3)Full disable.
> > >
> > > 3,572,032 byte
> > >
> > > >
> > > > I think, size command will spit out with different section's size.
> > > > This data can be used to take decision and to know how much % it
> > > > adding up?
> > >
> > > I.e. trace adds 228,496 byte = 6.4 % to the size of this specific
> > executable.
> > >
> > > ~200 KB might not seem as much, but this is just omitting one module.
> > > If other unused modules also add ~200 KB, it adds up.
> > >
> > > And I haven't done any significant additional memory footprint tuning.
> > >
> > >
> > > > > > #define __rte_trace_point_emit_header_generic(t) \
> > > > > > void *mem; \
> > > > > > do { \
> > > > > >   +      if (RTE_TRACE == 0) \
> > > > > >   +             return \
> > >
> > > Tested without the RTE_TRACE==0 check, simply:
> > >   void *mem; \
> > >   do { \
> > > +         return; \
> > >           const uint64_t val = rte_atomic_load_explicit(t,
> > rte_memory_order_acquire); \
> > >
> > > > > >         const uint64_t val = rte_atomic_load_explicit(t,
> > > > > > rte_memory_order_acquire); \
> > > > > >         if (likely(!(val & __RTE_TRACE_FIELD_ENABLE_MASK))) \
> > > > > >                 return; \
> > > > > >         mem = __rte_trace_mem_get(val); \
> > > > > >         if (unlikely(mem == NULL)) \
> > > > > >                 return; \
> > > > > >         mem = __rte_trace_point_emit_ev_header(mem, val); \
> > > > > > } while (0)
> > >
> > > I don't understand why your method uses as little memory as mine...
> >
> > Compiler is start enough to understand those functions are NOP.
>
> Yes, I understand why the compiler understands that this code is NOP.
>
> But I don't understand why the tracepoints' values and names are not registered by the unmodified RTE_TRACE_POINT_REGISTER().

My best _guess_ is __rte_trace_point_register()'s register_fn()
execution comes as NOP.




>
> >
> >
> > > My method should also omit the trace points themselves, with their names and
> > numbers, and their initialization.
> > >
> > > I haven't looked deeper into it.
> > >
> > > If your method of omitting trace is as efficient as all my ifdefs, I also
> > prefer your method.
> > > Simpler is better.
> > >