On Fri, May 15, 2020 at 2:56 PM Bartosz Golaszewski <brgl@xxxxxxxx> wrote:
> pt., 15 maj 2020 o 14:04 Arnd Bergmann <arnd@xxxxxxxx> napisał(a):
> > On Fri, May 15, 2020 at 9:11 AM Bartosz Golaszewski <brgl@xxxxxxxx> wrote:
> > > >
> > > > It looks like most of the stuff inside of the loop can be pulled out
> > > > and only done once here.
> > > >
> > >
> > > I did that in one of the previous submissions but it was pointed out
> > > to me that a parallel TX path may fill up the queue before I wake it.
> >
> > Right, I see you plugged that hole, however the way you hold the
> > spinlock across the expensive DMA management but then give it
> > up in each loop iteration feels like this is not the most efficient
> > way.
> >
>
> Maybe my thinking is wrong here, but I assumed that with a spinlock
> it's better to give other threads the chance to run in between each
> iteration. I didn't benchmark it though.
It depends. You want to avoid lock contention (two threads trying to
get the lock at the same time) but you also want to avoid bouncing
around the spinlock between the caches.
In the contention case, what I think would happen here is that the
cleanup thread gives up the lock and the xmit function gets it, but
then the cleanup thread is spinning again, so you are still blocked
on one of the two CPUs but also pay the overhead of synchronizing
between the two.
Holding the lock the whole time would speed up both the good case
(no contention) and the bad case (bouncing the lock) a little bit
because it saves some overhead. Holding the lock for shorter
times (i.e. not during the cache operations) would reduce the
amount of lock-contention but not help in the good case.
Not needing a lock at all is generally best, but getting it right
is tricky ;-)
Arnd
