Re: [PATCH net-next v4] rps: selective flow shedding during softnet overflow

[Eric Dumazet <eric.dumazet@gmail.com>]

On Tue, 2013-04-23 at 21:25 -0400, Jamal Hadi Salim wrote:
> On 13-04-23 08:09 PM, Eric Dumazet wrote:
> 
> > qdisc code has a hardcoded dev_hard_start_xmit() call, thats why ifb
> > hack is used. Not mentioning device flow control.
> >
> > It might be possible to use a q->xmit() method instead, so that it can
> > be used on ingress without ifb.
> >
> 
> If i understood correctly what you are trying to achieve:
> I dont think one qdisc per rx queue/ring will work well in
> presence of qdisc since the qdisc is attached per netdev.

MQ permits to have one qdisc per TX queue.

It would be the same concept in ingress.

> i.e when packets are fanned out across cpu backlogs, as long
> as they came in via same netdev queue, they are going to share
> the same lock with all other cpus such packets have been fanned out to
> the moment you attach an ingress qdisc to that netdev ring/queue.
> 

Not sure what you mean. The qdisc stuff would replace the 'cpu backlog',
not be added to it. Think of having possibility to control backlog using
standard qdiscs, like fq_codel ;)

> One unorthodox approach is to have a qdisc per backlog queue
> since the backlog is per cpu; given it is abstracted as a netdev,
> it becomes a natural fit (sans the fact backlog queue is
> unidirectional).

Yes, but the per cpu backlog is shared for all devices. We probably want
different qdisc for gre tunnel, eth0, ...