sunvnet: Do not ask for an ACK for every dring transmit

No need to ask for an ack with every vnet_start_xmit()- the single
ACK with DRING_STOPPED is sufficient for the protocol, and we free
the sk_buff in vnet_start_xmit itself, so we dont need an ACK back.

Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Acked-by: Raghuram Kothakota <raghuram.kothakota@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This commit is contained in:
Sowmini Varadhan 2014-08-12 10:27:24 -04:00 committed by David S. Miller
parent 8356f9d564
commit 1f6394e382
1 changed files with 9 additions and 1 deletions

View File

@ -691,7 +691,15 @@ static int vnet_start_xmit(struct sk_buff *skb, struct net_device *dev)
memset(tx_buf+VNET_PACKET_SKIP+skb->len, 0, len - skb->len);
}
d->hdr.ack = VIO_ACK_ENABLE;
/* We don't rely on the ACKs to free the skb in vnet_start_xmit(),
* thus it is safe to not set VIO_ACK_ENABLE for each transmission:
* the protocol itself does not require it as long as the peer
* sends a VIO_SUBTYPE_ACK for VIO_DRING_STOPPED.
*
* An ACK for every packet in the ring is expensive as the
* sending of LDC messages is slow and affects performance.
*/
d->hdr.ack = VIO_ACK_DISABLE;
d->size = len;
d->ncookies = port->tx_bufs[dr->prod].ncookies;
for (i = 0; i < d->ncookies; i++)