Re: [PATCH 2/2] kmemtrace: Use tracepoints instead of markers.

[Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>]

* Eduard - Gabriel Munteanu (eduard.munteanu@linux360.ro) wrote:
> On Tue, Dec 30, 2008 at 10:40:17AM -0200, Arnaldo Carvalho de Melo wrote:
> > Em Tue, Dec 30, 2008 at 09:41:28AM +0200, Eduard - Gabriel Munteanu escreveu:
> > > kmemtrace now uses tracepoints instead of markers. We no longer need to
> > > use format specifiers. Also dropped kmemtrace_mark_alloc_node(), since
> > > it's easy to pass -1 as the NUMA node without providing a needless and
> > > long-named variant of the same function.
> > > 
> > > Signed-off-by: Eduard - Gabriel Munteanu <eduard.munteanu@linux360.ro>
> > > diff --git a/include/linux/slab_def.h b/include/linux/slab_def.h
> > > index 7555ce9..fe3cea2 100644
> > > --- a/include/linux/slab_def.h
> > > +++ b/include/linux/slab_def.h
> > > @@ -76,8 +76,8 @@ found:
> > >  
> > >  		ret = kmem_cache_alloc_notrace(cachep, flags);
> > >  
> > > -		kmemtrace_mark_alloc(KMEMTRACE_TYPE_KMALLOC, _THIS_IP_, ret,
> > > -				     size, slab_buffer_size(cachep), flags);
> > > +		trace_kmalloc(_THIS_IP_, ret,
> > > +			      size, slab_buffer_size(cachep), flags, -1);
> > 
> > Is there a way for a tracepoint to get the caller without having to pass
> > it explicitely?
> > 
> > - Arnaldo
> 
> It might be possible, but for correctness DEFINE_TRACE() should use
> __always_inline instead of inline.
> 

Yes, we could (and maybe should) use __always_inline there. Hrm, which
which tree to you work ? You probably mean DECLARE_TRACE() rather than
DEFINE_TRACE() ?

I just went over your patch again. it uses the old DEFINE_TRACE() API.
You should get the latest tracepoints which have DECLARE_TRACE() (for
trace/kmemtrace.h) and then a DEFINE_TRACE() in a .c.  Ah I see, you
work on 2.6.28. You should work on top of -tip, which has this new API.
Using the tracepoints present in 2.6.28 will not let you do only a
single definition of the tracepoint structure and it will lead to waste
of kernel memory by defining multiple instances of tracepoint structures
(one for each trace_*() use, so one per kmalloc()). The
Documentation/tracepoints.txt file is updated accordingly.


> But it is quite pointless. Sometimes we need _RET_IP_, sometimes
> _THIS_IP_ and sometimes a parameter we've been passed. That's because we
> want the IP of the caller, so it depends on whether this slab function
> is __always_inline, non-inlined or deeply nested within other functions
> (which can be as well __always_inline or non-inlined).
> 

Hrm ? In the case we just want

  trace_kmalloc(_THIS_IP, ......);

If we have __always_inline for the trace_*() declaration, isn't it the
same to just do this in the probe ?

void probe_kmalloc(......)
{
  ... _RET_IP_ ...;

}

This would remove a parameter from the stack created from the
instrumentation site, which is always good.

Mathieu

> 
> 	Eduard
> 

-- 
Mathieu Desnoyers
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68