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>
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++)