[dpdk-dev] [PATCH] net/bonding: improve non-ip packets RSS

Declan Doherty declan.doherty at intel.com
Fri Feb 10 17:30:32 CET 2017


On 18/11/16 09:08, haifeng.lin at huawei.com (Haifeng Lin) wrote:
> Most ethernet not support non-ip packets RSS and only first
> queue can used to receive. In this scenario lacp bond can
> only use one queue even if multi queue configured.
>
> We use below formula to change the map between bond_qid and
> slave_qid to let at least slave_num queues to receive packets:
>
> 	slave_qid = (bond_qid + slave_id) % queue_num
>
> Signed-off-by: Haifeng Lin <haifeng.lin at huawei.com>
> ---
>  drivers/net/bonding/rte_eth_bond_pmd.c | 6 +++++-
>  1 file changed, 5 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/net/bonding/rte_eth_bond_pmd.c b/drivers/net/bonding/rte_eth_bond_pmd.c
> index 09ce7bf..8ad843a 100644
> --- a/drivers/net/bonding/rte_eth_bond_pmd.c
> +++ b/drivers/net/bonding/rte_eth_bond_pmd.c
> @@ -141,6 +141,8 @@ bond_ethdev_rx_burst_8023ad(void *queue, struct rte_mbuf **bufs,
>  	uint8_t collecting;  /* current slave collecting status */
>  	const uint8_t promisc = internals->promiscuous_en;
>  	uint8_t i, j, k;
> +	int slave_qid, bond_qid = bd_rx_q->queue_id;
> +	int queue_num = internals->nb_rx_queues;
>
>  	rte_eth_macaddr_get(internals->port_id, &bond_mac);
>  	/* Copy slave list to protect against slave up/down changes during tx
> @@ -154,7 +156,9 @@ bond_ethdev_rx_burst_8023ad(void *queue, struct rte_mbuf **bufs,
>  		collecting = ACTOR_STATE(&mode_8023ad_ports[slaves[i]], COLLECTING);
>
>  		/* Read packets from this slave */
> -		num_rx_total += rte_eth_rx_burst(slaves[i], bd_rx_q->queue_id,
> +		slave_qid = queue_num ? (bond_qid + slaves[i]) % queue_num :
> +				bond_qid;
> +		num_rx_total += rte_eth_rx_burst(slaves[i], slave_qid,
>  				&bufs[num_rx_total], nb_pkts - num_rx_total);
>
>  		for (k = j; k < 2 && k < num_rx_total; k++)
>

Nack, I think this could introduce unexpected behaviour as could then be 
read from a different of a slave queue that the queue id specified by 
the calling function, where the expected behaviour is that there is a 
1:1 queue mapping from bond to slave queues. If RSS is needed for 
ethdevs which don't support it natively I think the appropriate solution 
is to create a software RSS solution which can be enabled at the slave 
ethdev level itself. I don't think the bonding layer should be 
implementing this functionality.


More information about the dev mailing list