On Tue, Aug 30, 2016 at 18:06:36 -0400, John Ferlan wrote:
> 
> 
> On 08/12/2016 09:33 AM, Jiri Denemark wrote:
> > It really doesn't belong to the generic CPU driver.
> > 
> > Signed-off-by: Jiri Denemark <jdene...@redhat.com>
> > ---
> >  src/cpu/cpu_x86.c            | 16 ++--------------
> >  src/qemu/qemu_capabilities.c | 16 +++++++++++++++-
> >  2 files changed, 17 insertions(+), 15 deletions(-)
> > 
> 
> Hmm... interesting is this something that the online perf add more stats
> will need to also adjust, see (8/8):
> 
> http://www.redhat.com/archives/libvir-list/2016-August/msg00209.html
> 
> It doesn't seem so, but since I recognized the acronyms I figured I'd
> check...

The support for perf events is advertised by CMT CPU feature (and a few
others), but otherwise the parts of our code dealing with CPU features
and perf events are unrelated. There's no need to adjust anything in the
perf events code.

> So here we are again at a summary - if I didn't comment on something
> consider it an implicit ACK.
> 
> There's a couple of reviews that are simple and ACK'able - I think
> they're obvious.
> 
> However, there's also a couple where I'm just looking for information. I
> have no reason to not ACK, just wanted some clarity. I don't necessarily
> need to see a whole new series.  I think it just the interaction noted
> in patch 40, 35, and 26 (update and compare callbacks).

So after my replies to your comments, do you want me to resend any
patches from the series? I think the best option is to resend them all
and mark all unchanged patches so that reviewers do not need to look at
them again, but it's going to be a series of 45 patches.

Anyway, thanks for the review, I can imagine going through so many
patches to the ugly CPU code was not easy. But I think the code is
going to get better thanks to the patches :-)

Jirka

--
libvir-list mailing list
libvir-list@redhat.com
https://www.redhat.com/mailman/listinfo/libvir-list

Reply via email to