Am Donnerstag, 15. Oktober 2015, 11:48:12 schrieb Dan Carpenter:
> On Thu, Oct 15, 2015 at 10:39:02AM +0200, Marc Dietrich wrote:
> > > > - if (unlikely(nvec->rx == NULL)) {
> > > > + if (!unlikely(nvec->rx)) {
> > >
> > > This isn't right. You intented to say:
> > > if (unlikely(!nvec->rx)) {
> > >
> > > But even better to just remove the unlikely entirely.
> > >
> > > if (!nvec->rx) {
> >
> > why? the "unlikely" is there to optimize a critical interrupt path.
>
> The rule is that drivers should not use likely/unlikely() unless there
> is a difference in benchmark numbers.
well, we know that additional cpu cycles in this path break transfer for
unknown reasons. However, the unlikely here may be overkill. On the other
hand, I prefer not to change something here until these timing effects are
better understood.
> How critical can it be when it's
> always followed by a udelay(100)???
yes, this delay shouldn't be there at all. This is one of the timing mysteries
we still have to resolve.
> There are more important optimizations needed here.
sure. We are currently trying to move all this out into the tegra-i2c driver,
so this code block will get a major review/rewrite in the near future anyway.
Marc
Attachment:
signature.asc
Description: This is a digitally signed message part.
