Re: [PATCH v2 net-next] mlx4: optimize xmit path

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

On Mon, 2014-09-29 at 10:46 -0700, Alexei Starovoitov wrote:
> On Sun, Sep 28, 2014 at 9:19 PM, Eric Dumazet <eric.dumazet@gmail.com> wrote:
> >         send_doorbell = !skb->xmit_more || netif_xmit_stopped(ring->tx_queue);
> >
> >         if (ring->bf_enabled && desc_size <= MAX_BF && !bounce &&
> >             !vlan_tx_tag_present(skb) && send_doorbell) {
> 
> This patch is good,
> 
> but I've been thinking more about bf+xmit_more and want
> to double check my understanding in that scenario:
> xmit_more=true will queue descriptors normally and
> the last xmit_more=false packet will write into BF.
> I guess BF suppose to pick up the earlier ones from
> the queue, otherwise the whole thing is broken.
> So if indeed BF can pick up the whole chain, then
> it should be the faster way than doing iowrite32(), right?
> --


Right, this is what my patch is doing.

Say we send a burst of 8 packets.

7 first packets are queued, but no iowrite32() is performed.

final packet is queued, and bf is used to send the doorbell.

With our current stack (net-next), its very easy to check :

Disable tso (ethtool -K eth0 tso off)

Then run netperf -t TCP_RR -H ... -- -r 6000,6000

GSO makes great use of skb->xmit_more right now ;)

(Assuming we do not have GSO preparing hundred of segments out of a
single TSO packet)