Re: [PATCH net-next] net: rename low latency sockets functions to busy poll

[Stephen Hemminger <stephen@networkplumber.org>]

On Mon, 08 Jul 2013 22:46:04 +0300
Eliezer Tamir <eliezer.tamir@linux.intel.com> wrote:

> On 08/07/2013 22:37, Linus Torvalds wrote:
> > On Mon, Jul 8, 2013 at 10:14 AM, Eliezer Tamir
> > <eliezer.tamir@linux.intel.com> wrote:
> >>
> >> I think there is no way for the compiler to know the value of
> >> can_busy_loop at compile time. It depends on the replies we get
> >> from polling the sockets. ll_flag was there to make sure the compiler
> >> will know when things are defined out.
> > 
> > No, my point was that we want to handle the easily seen register test
> > first, and not even have to load current().
> > 
> > The compiler may end up scheduling the code to load current anyway,
> > but the way you wrote it it's pretty much guaranteed that it will do
> > it.
> 
> I see. OK.
> 
> > In fact, I'd argue for initializing start_time to zero, and have the
> > "have we timed out" logic load it only if necessary, rather than
> > initializing it based on whether POLL_BUSY_WAIT was set or not.
> > Because one common case - even with POLL_BUSY_WAIT - is that we go
> > through the loop exactly once, and the data exists on the very first
> > try. And that is in fact the case we want to optimize and not do any
> > extra work for at all.
> > 
> > So I would actually argue that the whole timeout code might as well be
> > something like
> > 
> >     unsigned long start_time = 0;
> >     ...
> >     if (want_busy_poll && !need_resched()) {
> >         unsigned long now = busy_poll_sched_clock();
> >         if (!start_time) {
> >             start_time = now + sysctl.busypoll;
> >             continue;
> >         }
> >         if (time_before(start_time, now))
> >             continue;
> >     }
> > 
> 
> OK.

Since this is special case in the hot path, it looks like a good case
for static branch logic.