Re: [dpdk-dev] [PATCH v6 1/4] app/testpmd: move eth header generation outside the loop

[Raslan Darawsheh <rasland@mellanox.com>]

Hi,

The performance issue that we saw in Mellanox is now fixed with latest version:

The issue was with the packet len calculation that was changed in the latest versions > 4:
diff --git a/app/test-pmd/txonly.c b/app/test-pmd/txonly.c
index af0be89..2f40949 100644
--- a/app/test-pmd/txonly.c
+++ b/app/test-pmd/txonly.c
@@ -176,6 +176,7 @@ pkt_burst_prepare(struct rte_mbuf *pkt, struct rte_mempool *mbp,
        pkt->l3_len = sizeof(struct ipv4_hdr);

        pkt_seg = pkt;
+       pkt_len = pkt->data_len;
        for (i = 1; i < nb_segs; i++) {
                pkt_seg->next = pkt_segs[i - 1];
                pkt_seg = pkt_seg->next;
@@ -198,7 +199,7 @@ pkt_burst_prepare(struct rte_mbuf *pkt, struct rte_mempool *mbp,
         * burst of packets to be transmitted.
         */
        pkt->nb_segs = nb_segs;
-       pkt->pkt_len += pkt_len;
+       pkt->pkt_len = pkt_len;

        return true;

and now we see ~20% improvement in performance.

Kindest regards,
Raslan Darawsheh

> -----Original Message-----
> From: dev <dev-bounces@dpdk.org> On Behalf Of Pavan Nikhilesh
> Bhagavatula
> Sent: Tuesday, April 2, 2019 12:53 PM
> To: Jerin Jacob Kollanukkaran <jerinj@marvell.com>; Thomas Monjalon
> <thomas@monjalon.net>; arybchenko@solarflare.com;
> ferruh.yigit@intel.com; bernard.iremonger@intel.com; Ali Alnubani
> <alialnu@mellanox.com>
> Cc: dev@dpdk.org; Pavan Nikhilesh Bhagavatula
> <pbhagavatula@marvell.com>
> Subject: [dpdk-dev] [PATCH v6 1/4] app/testpmd: move eth header
> generation outside the loop
> 
> From: Pavan Nikhilesh <pbhagavatula@marvell.com>
> 
> Testpmd txonly copies the src/dst mac address of the port being
> processed to ethernet header structure on the stack for every packet.
> Move it outside the loop and reuse it.
> 
> Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
> ---
>  v6 Changes
>  - Rebase onto ToT.
>  - Split the changes further
> 
>  v5 Changes
>  - Remove unnecessary change to struct rte_port *txp (movement).
> (Bernard)
> 
>  v4 Changes:
>  - Fix packet len calculation.
> 
>  v3 Changes:
>  - Split the patches for easier review. (Thomas)
>  - Remove unnecessary assignments to 0. (Bernard)
> 
>  v2 Changes:
>  - Use bulk ops for fetching segments. (Andrew Rybchenko)
>  - Fallback to rte_mbuf_raw_alloc if bulk get fails. (Andrew Rybchenko)
>  - Fix mbufs not being freed when there is no more mbufs available for
>  segments. (Andrew Rybchenko)
> 
>  app/test-pmd/txonly.c | 15 ++++++++-------
>  1 file changed, 8 insertions(+), 7 deletions(-)
> 
> diff --git a/app/test-pmd/txonly.c b/app/test-pmd/txonly.c
> index def52a048..0d411dbf4 100644
> --- a/app/test-pmd/txonly.c
> +++ b/app/test-pmd/txonly.c
> @@ -190,6 +190,14 @@ pkt_burst_transmit(struct fwd_stream *fs)
>  		ol_flags |= PKT_TX_QINQ_PKT;
>  	if (tx_offloads & DEV_TX_OFFLOAD_MACSEC_INSERT)
>  		ol_flags |= PKT_TX_MACSEC;
> +
> +	/*
> +	 * Initialize Ethernet header.
> +	 */
> +	ether_addr_copy(&peer_eth_addrs[fs->peer_addr],
> &eth_hdr.d_addr);
> +	ether_addr_copy(&ports[fs->tx_port].eth_addr, &eth_hdr.s_addr);
> +	eth_hdr.ether_type = rte_cpu_to_be_16(ETHER_TYPE_IPv4);
> +
>  	for (nb_pkt = 0; nb_pkt < nb_pkt_per_burst; nb_pkt++) {
>  		pkt = rte_mbuf_raw_alloc(mbp);
>  		if (pkt == NULL) {
> @@ -226,13 +234,6 @@ pkt_burst_transmit(struct fwd_stream *fs)
>  		}
>  		pkt_seg->next = NULL; /* Last segment of packet. */
> 
> -		/*
> -		 * Initialize Ethernet header.
> -		 */
> -		ether_addr_copy(&peer_eth_addrs[fs-
> >peer_addr],&eth_hdr.d_addr);
> -		ether_addr_copy(&ports[fs->tx_port].eth_addr,
> &eth_hdr.s_addr);
> -		eth_hdr.ether_type =
> rte_cpu_to_be_16(ETHER_TYPE_IPv4);
> -
>  		/*
>  		 * Copy headers in first packet segment(s).
>  		 */
> --
> 2.21.0