Re: [tip:perf/core] events: Rename TRACE_EVENT_TEMPLATE() to DECLARE_EVENT_CLASS()

[Frederic Weisbecker <fweisbec@gmail.com>]

On Thu, Nov 26, 2009 at 09:45:30AM -0500, Steven Rostedt wrote:
> On Thu, 2009-11-26 at 09:40 +0100, Ingo Molnar wrote:
> > * Steven Rostedt <rostedt@goodmis.org> wrote:
> > 
> 
> > > I would like to hear what others think about this change before we go 
> > > ahead and implement it.
> > 
> > You mean TRACE_EVENT() -> DEFINE_SINGLE_EVENT()? Sure, we want todo it 
> > in a more quiet moment of the kernel cycle, not now.
> > 
> > (TRACE_EVENT_TEMPLATE OTOH has existed for just a few days so it's not a 
> > problem.)
> 
> Yes the template name is new, I'm not talking about that on
> particularly.


I personally don't mind much about the name, especially
between class and template. Both make equally sense to me.

But DECLARE sounds like a misnomer here (like DEFINE_EVENT
somehow) as TRACE_EVENT, DEFINE_EVENT and TRACE_EVENT_TEMPLATE
all behave either as a declaration or a definition, depending
on the CREATE_TRACE_POINT macro.

Also, considering the arising question of notifiers and TRACE_EVENT
that are starting to collide in that we have two event callbacks
subsystems that could be gathered in one, I guess TRACE_EVENT will
become too general in the future.

If we consider improving the TRACE_EVENT to support tracing (like
it does already) but also blocking notifiers, atomic notifiers, etc...
by migrating the notifier code to TRACE_EVENT,
then the name should probably be reconsidered as a more general thing.

KERNEL_EVENT ? NOTIFY_EVENT ?

And then the CPP callbacks we are currently using for tracing should
probably be renamed as they won't concern the notifier callbacks.

TP_printk could be renamed as TRACE_print, TP_fast_assing could be
TRACE_fast_assign, etc...

 
> > 
> > > A lot of developers have just learned about TRACE_EVENT and now it 
> > > just disappeared. Well, not really, but in the sense of ' find 
> > > linux.git -name '*.[ch]' | xargs grep TRACE_EVENT' it no longer 
> > > exists.
> > 
> > A second problem with the TRACE_EVENT name is that it's not just for 
> > tracing - we dont necessarily 'trace' events here. We can use the event 
> > callbacks to collect pure counts:
> 
> Then we might as well rename the "trace_*" all over the kernel.


I think this should be kept. Although if notifier goes migrated in
TRACE_EVENT, having notify_event() would gather the two meanings of
trace_* and notify_*

 
> > DEFINE_SINGLE_EVENT() solves all these problems:
> > 
> >  - It's obvious what it does
> > 
> >  - It suggests users of it that there's another non-single-event 
> >    facility, gently nudging them towards the use of the more efficient
> >    DEFINE_EVENT_CLASS() + DEFINE_EVENT() method.
> > 
> >  - It fits nicely into the rest of the naming scheme.
> 
> Like I said earlier, I'm not really attached to the name. Except that
> there's already a lot of documentation (I've given tutorials about it)
> using the TRACE_EVENT name. But who am I to decide?


Not that I like much DEFINE_SINGLE_EVENT(), because DEFINE is ambiguous
and SINGLE too (single can indeed be interpreted as something that doesn't
need a class, but is also confusing as it suggests that DEFINE_EVENT defines
several events in once), but I think a tutorial shouldn't paralyze a
subsystem progression.

Why not having BUILD_EVENT_CLASS(), BUILD_EVENT_FROM_CLASS(), and BUILD_EVENT() ?

That said, TRACE_EVENT() can still remain as an alias.