Re: [PATCH 1/3] AMD Family10h+ IBS support for oProfile driver: Setup routines

[Ingo Molnar <mingo@elte.hu>]

* Andrew Morton <akpm@linux-foundation.org> wrote:

> On Thu, 3 Jul 2008 11:20:52 +0200 Ingo Molnar <mingo@elte.hu> wrote:
> 
> > 1) please rebase your patches ontop of -mm, which carries the 
> > oprofile multiplexing cleanups and enhancements already which 
> > heavily interact with your patchset.
> 
> I dropped those.  Because it's obvious that significant changes are to 
> be made.  So I guess working against linux-next is appropriate for 
> this change.

hm, okay. All the pending oprofile enhancements are nice in principle 
and there's like 3-4 oprofile topics that we should really start 
reviewing, integrating and testing:

- the gist of the perfmon2 changes which make oprofile work across
  context-switches. Oprofile should have done this from the get go.

- the oprofile syscall enhancements from perfmon2 for lightning-fast
  profiling feedback.

- oprofile multiplexing: makes hw counter constraints largely invisible.
  Should have been done years ago.

- new hw support like this IBS patch but it would also be nice to see
  BTS/PEBS support perhaps.

- [ personally i'd love to see /debug/oprofile/prof.txt that would do
    user-space-less parsing of oprofile data and would in essence 
    replace readprofile for all practical purposes. (while still being 
    as simple and self-contained as readprofile) Full in-kernel 
    generation of human-readable text output. This would be the thing
    that would bring oprofile a lot closer to the average kernel 
    developer IMO. (Also, various knobs under /debug/oprofile/* to
    configure oprofile details on the fly, without any userspace 
    dependencies.) ]

- [ oprofile .config driven in-kernel self-tests. Right now we notice it 
    quite late when it breaks. If we do /debug/oprofile/ that would be a 
    great step towards that. ]

tip/x86/oprofile (which already has some smaller changes) could host 
them, if there's interest. It would be renamed to tip/oprofile or so, 
kept separate and exported separately for import into -mm and/or 
linux-next - not mixed into x86 and other -tip bits.

v2.6.28 stuff at the earliest, obviously.

	Ingo