Re: Using perf_event_open() to sample multiple events of a process

[Peter Zijlstra]

On Sat, Nov 06, 2021 at 01:45:57AM +0100, Peter Zijlstra wrote:
> On Thu, Nov 04, 2021 at 10:57:50PM -0700, Nadav Amit wrote:
> > Hello Ken, Peter,
> > 
> > I would appreciate some help regarding the use of perf_event_open()
> > to have multiple samples getting into the same mmap’d memory when they
> > are both attached to the same process.
> > 
> > I am doing so (using both PERF_FLAG_FD_NO_GROUP and PERF_FLAG_FD_OUTPUT),
> > but it results in -EINVAL. Debugging the code shows that
> > perf_event_set_output() fails due to the following check:
> > 
> >         /*
> >          * If its not a per-cpu rb, it must be the same task.
> >          */
> >         if (output_event->cpu == -1 && output_event->ctx != event->ctx)
> >                 goto out;
> > 
> > However, it appears that at this point, event->ctx is not initialized
> > (it is null) so the test fails and the whole perf_event_open() syscall
> > fails.
> > 
> > Am I missing something? If not, I am unsure, unfortunately, what the
> > proper way to fix it is…
> > 
> > I include a small test that fails on my system. The second
> > perf_event_open fails due to the check in perf_event_set_output():
> > 
> 
> Works when you use the SET_OUTPUT ioctl()...
> 
> I think something went sideways in the syscall path and things went out
> of order :/ I'll try and have a look.

The problem seems to be that we call perf_event_set_output() before we
set event->ctx, which is a bit of a problem.

Now, afaict it's been broken since c3f00c70276d ("perf: Separate
find_get_context() from event initialization"), which is ages ago :/

It's waaay too late to try and fix it; I'll be likely to make an even
bigger mess if I tried. Perhaps tomorrow.

Clearly FD_OUTPUT isn't much used :-(