[Cbe-oss-dev] [PATCH 7/8] spufs: Add more markers for sputrace

[Julio M. Merino Vidal]

On Fri, Apr 25, 2008 at 02:12:44PM +0200, Christoph Hellwig wrote:
> On Fri, Apr 25, 2008 at 12:07:28PM +0200, jmerino at ac.upc.edu wrote:
> > Define the following new markers for spufs to allow more finer-grained
> > tracing of scheduling decisions:
> > 
> > * __spu_add_to_rq__enter
> > * __spu_del_from_rq__enter
> > * alloc_spu_context__return
> > * spu_activate__enter
> > * spu_bind_context__return
> > * spu_find_victim__return
> > * spu_process_callback__syscall
> > * spu_run_init__enter
> > * spu_run_init__write
> > * spu_unbind_context__return
> > * spufs_run_spu__cont
> > * spufs_run_spu__enter
> > * spufs_run_spu__return
> 
> Adding these is fine with me, but please also add handlers because
> they're completely useless otherwise.

You mean handlers in sputrace similar to spu_context_event?  I was thinking
on adapting sputrace to make it possible to get info from markers that do
not follow the current conventions (ctx alone or ctx/spu pairs).  I.e. add
additional functions to capture unknown parameters (npc, flags and such).

However, the current patch uses spu_context_event and spu_context_nospu_event
for all these markers, which means that they do work but the user will only
get information of the known parameters (which are, again, the ctx alone or
the ctx/spu pairs).

Anyway, these are all useful with SystemTap.  Whenever this patch set is
integrated mainstream, I'll submit a cbe.stp tapset to the SystemTap
developers to support these markers "out of the box".

Or where you referring to something else?

-- 
Julio M. Merino Vidal <jmerino at ac.upc.edu>