Re: [RFC v2 2/7] perf: add AUX area to ring buffer for raw data streams

[Peter Zijlstra <peterz@infradead.org>]

On Wed, Jun 25, 2014 at 02:09:31PM +0300, Alexander Shishkin wrote:
> Peter Zijlstra <peterz@infradead.org> writes:
> 
> > On Wed, Jun 11, 2014 at 06:41:45PM +0300, Alexander Shishkin wrote:
> >> +	/*
> >> +	 * Set up pmu-private data structures for an AUX area
> >> +	 */
> >> +	void *(*setup_aux)		(int cpu, void **pages,
> >> +					 int nr_pages, bool overwrite);
> >> +					/* optional */
> >> +
> >> +	/*
> >> +	 * Free pmu-private AUX data structures
> >> +	 */
> >> +	void (*free_aux)		(void *aux); /* optional */
> >
> > I was hoping you could replace those with a PERF_CAP_AUX or something
> > and then have one generic allocation routine like you provide in the
> > subsequent patches.
> 
> I need these to allocate pmu-specific SG tables now, which I don't see
> how to generalize nicely. User-visible aux_pages are allocated in the
> generic rb_alloc_aux().
> 
> As for sg tables, the alternative would be to allocate them in the
> event::pmu::add() path, which is probably not very good for
> performance, unless I'm missing something? Like, I can allocate stuff in
> the first add() and then free it in event::destroy().
> 
> Or, we can assume a generic sg table format and allocate them in the
> generic code, but then we'll need more capabilities to indicate at least
> the size of a table entry. PT needs 64 bits per entry, ARM's TMC needs
> 32 and the non-SG PT will still need some tables. But I'd rather leave
> this part to the pmu drivers. What do you think?

Ah, I wasn't aware.. too bad. Maybe add this to the changelog.