Re: [PATCH 0/18] Allow different tracers to be compiled independently

[Frederic Weisbecker <fweisbec@gmail.com>]

On Fri, Mar 26, 2010 at 01:20:29AM +0100, Jan Kara wrote:
> On Tue 23-03-10 02:04:21, Frederic Weisbecker wrote:
> > On Tue, Mar 23, 2010 at 01:32:02AM +0100, Jan Kara wrote:
> > > 
> > >   Hi,
> > > 
> > >   currently, when one tracer is selected, most of tracepoints for other
> > > tracers also gets pulled into the kernel. So for example it's not possible
> > > to enable BLK_DEV_IO_TRACE without polluting slab allocation paths with
> > > tracepoint checks (see changelog of patch 01). This patch set adds a
> > > possibility for each set of trace points to be compile-enabled separately.
> > >   The first patch contains the necessary magic in linux/tracepoint.h. Other
> > > patches just tell tracing framework about correspoding config options
> > > and possibly introduce them if they did not exist before.
> > >   The patches in this patch set are actually completely independent so 
> > > they can be merged via respective subsystem trees. But changes are rather
> > > tiny so I don't expect much conflicts...
> > > 
> > > 								Honza
> > 
> > (Adding more people in Cc)
> > 
> > I don't know. Yeah this first looks like a good idea but once
> > CONFIG_EVENT_TRACING is enabled, each tracepoint is a lightweight
> > thing and induce a tiny overhead, probably hard to notice, and
> > this is going to be even more the case after the jmp label
> > optimization patches.
>   Sorry for replying late, I was on vacation. My motivation was that we
> wanted BLK_DEV_IO_TRACE enabled in all our distro kernels but there is
> a concern that this could have some impact on performance especially
> in SLAB allocator due to more icache pressure or so. This is not completely
> bogus concern if you look at bloat-o-meter output. For example:
> function                                     old     new   delta
> kmem_cache_alloc                             542     768    +226
> 
>   But looking at the disassembly now, I can see that the difference in
> inline code is actually only ~40 bytes (on x86_64) - so that's about 7%.
> Not a huge deal but still noticeable. 
> 
> > I liked the fact we had a general tracing kernel once the above
> > config is selected. And we don't bother telling people that to
> > use tool X you need CONFIG_EVENT_Y, and you need to rebuild your
> > kernel, etc...
>   OK, I understand this. I guess I should go and measure whether disabling
> tracepoints really makes some performance difference or not.



If you do so, please try with with the jmp label patches:
http://lwn.net/Articles/379776/