Re: perf: PERF_FLAG_FD_OUTPUT has been broken since 2.6.35

[Peter Zijlstra <peterz@infradead.org>]

On Thu, Jan 08, 2015 at 10:22:37AM -0500, Vince Weaver wrote:
> Hello,
> 
> I was working on improving the manpage by working out how the 
> perf_event_open() PERF_FLAG_FD_OUTPUT flag works.
> 
> It turns out it doesn't.
> 
> In kernel/events/core.c when opening a file this is done:
> 
>         if (output_event) {
>            err = perf_event_set_output(event, output_event);
>            if (err)
>               goto err_context;
>         }
> 
> 	and perf_event_set_output() does:
> 
>                 if (output_event->cpu == -1 && output_event->ctx != event->ctx) {
>                    goto out;
>                 }
> 
> 	But event->ctx is always NULL; it does not get set to ctx until
> 	after the call to perf_event_set_output().
> 
> 	It looks like this was broken back in 2.6.35 days with the change:
> 
> 		commit ac9721f3f54b27a16c7e1afb2481e7ee95a70318
> 		Author: Peter Zijlstra <a.p.zijlstra@chello.nl>
> 		Date:   Thu May 27 12:54:41 2010 +0200
> 
> 		perf_events: Fix races and clean up perf_event and perf_mmap_data interaction
> 
> So is this worth fixing seeing as apparently no one uses this feature?

I think there's a fair argument for removing it, Ingo, Acme?

> As an aside, added error reporting is *really* needed as I had to sprinkle
> all of events/core.c with printk() calls to even begin to understand why
> I was getting the EINVAL result.

Yes, let me see if I can find someone to work on that :/