[PATCH v2 2/3] net/zxdh: optimize Rx recv pkts performance

[Stephen Hemminger]

On Thu, 23 Apr 2026 09:18:17 +0800
Junlong Wang <wang.junlong1 at zte.com.cn> wrote:

> +
> +	PMD_DRV_LOG(DEBUG, "port %d min_rx_buf_size %d",
> +		eth_dev->data->port_id, eth_dev->data->min_rx_buf_size);

Don't use %d when printing unsigned values.

+	/* If device is started, refuse mtu that requires the support of
+	 * scattered packets when this feature has not been enabled before.
+	 */
+	if (dev->data->dev_started &&
+		((!dev->data->scattered_rx &&
+		 ((uint32_t)ZXDH_MTU_TO_PKTLEN(new_mtu) >
+		 (dev->data->min_rx_buf_size - RTE_PKTMBUF_HEADROOM))) ||
+		 (dev->data->scattered_rx &&
+		 ((uint32_t)ZXDH_MTU_TO_PKTLEN(new_mtu) <=
+		 (dev->data->min_rx_buf_size - RTE_PKTMBUF_HEADROOM))))) {
+		PMD_DRV_LOG(ERR, "Stop port first.");
+		return -EINVAL;
+	}

You can use lines up to 100 characters, and break up this into multiple
if statements to avoid such a complex expression. Looks like multiple
parts are the same?

>  
> +#define ZXDH_VLAN_TAG_LEN   4
Why not use RTE_VLAN_HLEN?

> +#define ZXDH_ETH_OVERHEAD  (RTE_ETHER_HDR_LEN + RTE_ETHER_CRC_LEN + ZXDH_VLAN_TAG_LEN * 2)
> +#define ZXDH_MTU_TO_PKTLEN(mtu) ((mtu) + ZXDH_ETH_OVERHEAD)

> +static inline int zxdh_init_mbuf(struct rte_mbuf *rxm, uint16_t len,
> +		struct zxdh_hw *hw, struct zxdh_virtnet_rx *rxvq)
> +{
> +	uint16_t hdr_size = 0;
> +	struct zxdh_net_hdr_ul *header;
> +
> +	header = (struct zxdh_net_hdr_ul *)((char *)
> +					rxm->buf_addr + RTE_PKTMBUF_HEADROOM);

Please use rte_pktmbuf_mtod instead for this.

> +uint16_t zxdh_recv_single_pkts(void *rx_queue, struct rte_mbuf **rcv_pkts, uint16_t nb_pkts)
> +{
> +	struct zxdh_virtnet_rx *rxvq = rx_queue;
> +	struct zxdh_virtqueue *vq = rxq_get_vq(rxvq);
> +	struct zxdh_hw *hw = vq->hw;
> +	struct rte_mbuf *rxm;
> +	uint32_t lens[ZXDH_MBUF_BURST_SZ];
> +	uint16_t len = 0;
> +	uint16_t nb_rx = 0;
> +	uint16_t num;
> +	uint16_t i = 0;

Useless initialization of i.

>  
> -			dev->data->rx_mbuf_alloc_failed += free_cnt;
> +	num = nb_pkts;
> +	if (unlikely(num > ZXDH_MBUF_BURST_SZ))
> +		num = ZXDH_MBUF_BURST_SZ;



> +	num = zxdh_dequeue_burst_rx_packed(vq, rcv_pkts, lens, num);
> +	if (num == 0) {
> +		rxvq->stats.idle++;
> +		goto refill;

Since this is normal path on idle network, the counter will grow
rapidly. Do you need it?

> +	}
> +
> +	for (i = 0; i < num; i++) {
> +		rxm = rcv_pkts[i];
> +		len = lens[i];
> +		if (unlikely(zxdh_init_mbuf(rxm, len, hw, &vq->rxq) < 0)) {
> +			rte_pktmbuf_free(rxm);
> +			continue;
>  		}

Better practice to make rxm and len variables scoped to the loop.

AI review noticed that there is now a double free in the error path.
Both error paths inside zxdh_init_mbuf() already call rte_pktmbuf_free(rxm) before returning -1. The caller's rte_pktmbuf_free(rxm) then frees it a second time. Remove the caller's free, or stop freeing inside zxdh_init_mbuf().

(zxdh_set_rxtx_funcs) — dropped mergeable-rxbuf feature check: The old code returned -1 with an error log when the peer did not negotiate ZXDH_NET_F_MRG_RXBUF. The new code silently drops that check. If the negotiated feature set doesn't include MRG_RXBUF, the multi-segment rx path may now be selected against a peer that doesn't support it.