Messages in this thread | | | From | David Sharp <> | Date | Tue, 10 May 2011 12:13:48 -0700 | Subject | Re: Fix powerTOP regression with 2.6.39-rc5 |
| |
On Tue, May 10, 2011 at 3:33 AM, Steven Rostedt <rostedt@goodmis.org> wrote: > On Tue, 2011-05-10 at 10:47 +0200, Ingo Molnar wrote: >> * Steven Rostedt <rostedt@goodmis.org> wrote: >> >> > [...] Thus a library is a perfect solution. [...] >> >> Btw., just to make things clear, if we indeed have a library to parse things >> and if all apps use that then the ABI moves to another (library) level. >> >> The requirement from my maintenance POV is very, very simple: apps should not >> break on new kernels. If this is achieved by making apps smarter then that's a >> valid solution. >> > > Great! Because this is what I want. I would also want a way to designate > events as stable. I'll add a TRACE_EVENT_STABLE() that can only have the > events that maintainers agree to maintain. And give the apps an ability > to only see these.
A TRACE_EVENT_STABLE() would mark the entire event as stable. I was wondering if we should instead mark fields within events as stable. Even within a "stable" event, certain fields we might not want to guarantee to be in the next release. This might also make it clearer that the position of a field (stable or not) in an event can change, and tools really should parse the event format.
> Have the other events need either a separate library, > or perhaps just separate calls from within the same library, so the XFS > developers can feel safe that their tracepoints will not be depended on. > And perhaps have two tracepoints for sched_switch such that Peter > Zijlsta is happy that he's not bound by an tracepoint that keeps him > from getting rid of FIFO ;) > > I'm happy to write a libperf.so and I can discuss with Arnaldo, Arjan > and yourself what is the best way of doing this. > > Thanks, > > -- Steve > > >
| |