On Thu, 2015-11-19 at 22:44 +0100, Francois Romieu wrote:

Eric Dumazet [off-list ref] :
[...]

quoted

This looks like a race when/if RX happens very shortly after one
transmit.

Francois, are we really saving lot of cpu cycles testing status ?

I'm not sure. An extra smp_rmb and three cache lines at most when
there really isn't any work ? Probably not worth it.

Can you clarify which kind of race you're thinking about ?

I was mostly thinking of some hardware bug.

Or maybe a software barrier is missing, but I could not see it.

(Assuming rtl_get_events() is accurate, ie really returns a mask of
pending events in all circumstances, including if run in // from
rtl_slow_event_work() and rtl8169_poll())