[lttng-dev] [PATCH lttng-ust] Improve tracelog handling, reduce exported functions

[Norbert Lange]

Am Do., 20. Mai 2021 um 17:21 Uhr schrieb Mathieu Desnoyers
<mathieu.desnoyers at efficios.com>:
>
> ----- On May 20, 2021, at 10:57 AM, Norbert Lange nolange79 at gmail.com wrote:
>
> > Am Do., 20. Mai 2021 um 16:19 Uhr schrieb Mathieu Desnoyers
> > <mathieu.desnoyers at efficios.com>:
> >>
> >> ----- On May 20, 2021, at 8:18 AM, lttng-dev lttng-dev at lists.lttng.org wrote:
> >>
> >> > Instead of creating functions for each loglevel, simply pass the
> >> > callback as argument.
> >> >
> >> > Further pack all preprocessor information into a struct that
> >> > the compiler already can prepare.
> >>
> >> This introduces an ABI break too late in the cycle.
> >
> > So 2.14 would be the next chance I guess
>
> No. The original ABI was introduced about 10 years ago with lttng-ust 2.0,
> and lttng-ust 2.13 introduces the first ABI break since. I don't
> plan on doing any ABI break in lttng-ust in the foreseeable future.
>
> ABI breaks require that our users recompile all their instrumented
> applications, which is really cumbersome for large software deployments.
> We don't break ABI lightly.

Yeah, I understand.


> >> Also, I'm not so keen on adding an indirect call on the fast-path
> >> when it's not absolutely needed.
> >
> > Code seems pretty similar: https://godbolt.org/z/oK1WhWqGT
>
> By fast-path, I also mean:
>
> +        (*callback)(source->file, source->line, source->func, msg, len,
> +                LTTNG_UST_CALLER_IP());
>
> Which introduces an indirect call which needs to be taken when tracing
> is active.

The worst thing is that it would tax branch-predictors. Indirect jumps aren't
that horrible, and if you have public interpose-able ELF symbols you
have more of them than you might know...

And that's a function that calls a printf variant, and did alloc memory.

> >> What is wrong with having one symbol per loglevel ?
> >
> > Macro-magic is cumbersome to edit, more code, more relocations.
>
> If it was still time for ABI breaks, I would be tempted to consider it
> especially given that tracelog and tracef are not expected to be "high-speed",
> but now is too late for breaking ABI.
>
> >
> > Easier to adapt aswell, could roll my own tracelog functions while
> > using lttng_ust__tracelog_printf (started soind that as I don't want
> > to link to lttng-ust.so)
>
> What prevents you from linking against lttng-ust.so again ?

I did not poke around enough with Lttng to be confident it wont have
side effects,
I really don't want it active in production. It doesn't seem there is
much public knowledge with Xenomai either.
lttng-ust.so will spawn threads, lttng-ust-tracepoint.so is mostly passive,
So Id want a dynamic tracepoint-provider than i can dlopen (so that
the signal masks are inherited,
I hope you dont touch them).

Of course I could just remove all lttng libraries on the production
system aswell. Still doesnt change that
tracelog and tracef doesnt work that way.

I implemented my own tracelog/tracef using the normal lttng
tracepoints for now, they totally break on source level with 2.13
aswell ;)
is it ok if I do this to access them:

#define TRACEPOINT_DEFINE
#define TRACEPOINT_PROBE_DYNAMIC_LINKAGE
// 2.12
// #include <lttng/lttng-ust-tracelog.h>
// #include <lttng/lttng-ust-tracef.h>
// 2.13
#include <lttng/tp/lttng-ust-tracelog.h>
#include <lttng/tp/lttng-ust-tracef.h>

ie. I would load lttng-ust.so later and can then use those tracepoints.

Norbert