Hi,
On Mon, 2009-11-02 at 22:43 +0100, Frederic Weisbecker wrote:
[...]
> > >
> > > But I would rather see a __mips_notrace on these two core functions.
> >
> > What about this: __arch_notrace? If the arch need this, define it,
> > otherwise, ignore it! if only graph tracer need it, define it in "#ifdef
> > CONFIG_FUNCTION_GRAPH_TRACER ... #endif".
>
>
>
> The problem is that archs may want to disable tracing on different
> places.
> For example mips wants to disable tracing in timecounter_read_delta,
> but another arch may want to disable tracing somewhere else.
>
> We'll then have several unrelated __arch_notrace. One that is relevant
> for mips, another that is relevant for arch_foo, but all of them will
> apply for all arch that have defined a __arch_notrace.
>
> It's true that __mips_notrace is not very elegant as it looks like
> a specific arch annotation intruder.
>
>
> But at least that gives us a per arch filter granularity.
>
> If only static ftrace could disappear, we could keep only dynamic
> ftrace and we would then be able to filter dynamically.
> But I'm not sure it's a good idea for archs integration.
>
I think if we use something like __mips_notrace here, we may get lots of
__ARCH_notraces here too, 'Cause some other platforms(at least, as I
know, Microblaze will do it too) may also need to add one here, it will
become:
__mips_notrace __ARCH1_notrace __ARCH2_notrace .... foo() {...}
A little ugly ;)
and If a new platform need it's __ARCH_notrace, they need to touch the
common part of ftrace, more side-effects!
but with __arch_notrace, the archs only need to touch it's own part,
Although there is a side-effect as you mentioned above ;)
So, what should we do?
Regards,
Wu Zhangjin
