Re: [PATCH 3/5] perf tools: Reference count map_groups objects

[Jiri Olsa <jolsa@redhat.com>]

On Tue, Apr 15, 2014 at 12:17:22PM -0600, David Ahern wrote:
> On 4/15/14, 3:56 AM, Jiri Olsa wrote:
> >
> >hum.. I think using atomic type is not enough, we'd need
> >to make map_groups__put/get atomic as well
> >
> >not sure what's the support in user space for that.. will check
> >
> >also not to be negative, but libperf is not thread safe anyway, right? ;-)
> 
> should be. pretty certain there have been bug reports (e.g,
> perf-top) that stumble onto problems and they get fixed.

I checked and there's no problem with this in perf top, because
all thread allocations and frees (included map_groups) are done
within the main (one) thread

You could get in trouble if you'd call map_groups__get and
map_groups__put over same map_groups object from different
threads, but this never happens in perf

Thats the same sort of issue you would get if you called
machine__findnew_thread over same machine from 2 threads

I think that potential libperf users should be aware of the
proper usage or they should ask for change

thoughts? ;-) thanks,
jirka