On Tue, Jul 24, 2018 at 6:44 AM Michael S. Tsirkin <m...@redhat.com> wrote:
>
> On Mon, Jul 23, 2018 at 04:11:19PM -0700, Caleb Raitto wrote:
> > From: Caleb Raitto <carai...@google.com>
> >
> > The driver disables tx napi if it's not certain that completions will
> > be processed affine with tx service.
> >
> > Its heuristic doesn't account for some scenarios where it is, such as
> > when the queue pair count matches the core but not hyperthread count.
> >
> > Allow userspace to override the heuristic. This is an alternative
> > solution to that in the linked patch. That added more logic in the
> > kernel for these cases, but the agreement was that this was better left
> > to user control.
> >
> > Do not expand the existing napi_tx variable to a ternary value,
> > because doing so can break user applications that expect
> > boolean ('Y'/'N') instead of integer output. Add a new param instead.
> >
> > Link: https://patchwork.ozlabs.org/patch/725249/
> > Acked-by: Willem de Bruijn <will...@google.com>
> > Acked-by: Jon Olson <jonol...@google.com>
> > Signed-off-by: Caleb Raitto <carai...@google.com>
>
> Is there a reason the same rule should apply to all devices?
> If not shouldn't this be an ethtool option?It not very likely that a guest will have multiple virtio_net devices, some of which have affinity_hint_set and some do not? I'd really rather not add the extra option at all, but remove the affinity_hint_set requirement for now. Without more data, I understand the concern about cacheline bouncing if napi-tx would becomes the default at some point and we don't have data on this by then. But while it isn't default and a user has to opt in to napi-tx to try it that seems enough guardrail to me. The original reason was lack of data on whether napi-tx may suffer from cache invalidations when tx and rx softirq are on different cpus and we enable tx descriptor cleaning from the rx handler (i.e., on ACK). >From those initial numbers it seemed to be a win even with those invalidations. https://patchwork.ozlabs.org/patch/746232/ In lieu of removing the affinity_hint_set, this boolean is the least amount of both new interface and implementation to allow experimentation. We can easily leave it as a noop eventually when we are confident that napi-tx can be enabled even without affinity. By comparison, an ethtool param would be quite a bit of new logic.
