On 03/29/2014 04:02 PM, Eric Dumazet wrote:

On Fri, 2014-03-28 at 22:13 +0100, Daniel Borkmann wrote:

quoted

On 03/28/2014 09:50 PM, David Miller wrote:

quoted

quoted

The only remaining issue is that Eric seems to suggest that a tx
dropped counter bump should be added.  Please submit that as a
follow-on if my reading of his feedback is accurate.

Yes, will do next week. I think at time of submission we didn't
have that yet. :)

BTW, how direct xmit performs in case of BQL enabled driver ?

dummy is nice, but not really representative of real device ;)

netif_xmit_frozen_or_stopped() tests (dev_queue->state &
QUEUE_STATE_ANY_XOFF_OR_FROZEN), so will not fill TX ring.

Since you have no queue, I would rather use a test without
BQL bit (1 << __QUEUE_STATE_STACK_XOFF), so that you can absorb bursts
into TX ring ...

Thanks Eric, I'll look into it on next Tuesday [I was on travel
whole weekend and will be on Monday].