On Fri, 13 Nov 2015 15:46:06 +0100
Cornelia Huck <cornelia.h...@de.ibm.com> wrote:
> On Fri, 13 Nov 2015 09:26:26 +0100
> Greg Kurz <gk...@linux.vnet.ibm.com> wrote:
>
> > On Thu, 12 Nov 2015 18:52:55 +0100
> > Cornelia Huck <cornelia.h...@de.ibm.com> wrote:
> >
> > > On Mon, 09 Nov 2015 18:41:33 +0100
> > > Greg Kurz <gk...@linux.vnet.ibm.com> wrote:
>
> > > > +static bool virtio_net_needs_hdr_swap(VirtIONet *n)
> > > > +{
> > > > + /* virtio_needs_swap() is constant for fixed endian targets: call
> > > > it
> > > > + * first to filter them out without penalty.
> > >
> > > What do you mean with 'constant' here? It still needs to retrieve the
> > > feature bit from the device, no?
> > >
> >
> > Yes the comment is inaccurate... With the "virtio: cross-endian helpers
> > fixes"
> > series, virtio_needs_swap() indeed resolves to { return 0; } for fixed
> > little
> > endian targets but we still need to check the feature bit when the target
> > is big
> > endian.
> >
> > If I drop the call to virtio_needs_swap(), all targets will have the same
> > penalty. If I keep it, fixed little endian targets have no penalty but fixed
> > big endian targets get an extra check and bi-endian targets get two extra
> > checks...
> >
> > Not sure what to decide...
>
> The impact probably isn't too large, but improving one configuration
> via penalizing others feels wrong.
>
> Just check the bool value to make the decision?
Sounds fair. I'll drop the call to virtio_needs_swap() then.
Thanks.
--
Greg
