diff mbox series

[v5,1/2] app/testpmd: optimize testpmd txonly mode

Message ID 20190331131341.12924-1-pbhagavatula@marvell.com (mailing list archive)
State Superseded, archived
Delegated to: Ferruh Yigit
Headers
Series [v5,1/2] app/testpmd: optimize testpmd txonly mode |

Checks

Context Check Description
ci/checkpatch success coding style OK
ci/intel-Performance-Testing success Performance Testing PASS
ci/mellanox-Performance-Testing fail Performance Testing issues
ci/Intel-compilation fail Compilation issues

Commit Message

Pavan Nikhilesh Bhagavatula March 31, 2019, 1:14 p.m. UTC 
  From: Pavan Nikhilesh <pbhagavatula@marvell.com>

Optimize testpmd txonly mode by
1. Moving per packet ethernet header copy above the loop.
2. Use bulk ops for allocating segments instead of having a inner loop
for every segment.

Also, move the packet prepare logic into a separate function so that it
can be reused later.

Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
---
 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 | 139 +++++++++++++++++++++++-------------------
 1 file changed, 76 insertions(+), 63 deletions(-)

--
2.21.0

Comments

Ferruh Yigit April 1, 2019, 8:25 p.m. UTC | #1 
On 3/31/2019 2:14 PM, Pavan Nikhilesh Bhagavatula wrote:
> From: Pavan Nikhilesh <pbhagavatula@marvell.com>
> 
> Optimize testpmd txonly mode by
> 1. Moving per packet ethernet header copy above the loop.
> 2. Use bulk ops for allocating segments instead of having a inner loop
> for every segment.
> 
> Also, move the packet prepare logic into a separate function so that it
> can be reused later.
> 
> Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
> ---
>  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)

Hi Thomas, Shahafs,

I guess there was a performance issue on Mellanox with this patch, I assume it
is still valid, since this version only has some cosmetic change, but can you
please confirm?

And what is the next step, can you guys provide some info to Pavan to solve the
issue, or perhaps even better a fix?
Thomas Monjalon April 1, 2019, 8:53 p.m. UTC | #2 
01/04/2019 22:25, Ferruh Yigit:
> On 3/31/2019 2:14 PM, Pavan Nikhilesh Bhagavatula wrote:
> > From: Pavan Nikhilesh <pbhagavatula@marvell.com>
> > 
> > Optimize testpmd txonly mode by
> > 1. Moving per packet ethernet header copy above the loop.
> > 2. Use bulk ops for allocating segments instead of having a inner loop
> > for every segment.
> > 
> > Also, move the packet prepare logic into a separate function so that it
> > can be reused later.
> > 
> > Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
> > ---
> >  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)
> 
> Hi Thomas, Shahafs,
> 
> I guess there was a performance issue on Mellanox with this patch, I assume it
> is still valid, since this version only has some cosmetic change, but can you
> please confirm?

We will check it.

> And what is the next step, can you guys provide some info to Pavan to solve the
> issue, or perhaps even better a fix?

Looking at the first patch, there are still 3 changes merged together.
Why not splitting even more?
Jerin Jacob Kollanukkaran April 2, 2019, 1:03 a.m. UTC | #3 
On Mon, 2019-04-01 at 22:53 +0200, Thomas Monjalon wrote:
> 01/04/2019 22:25, Ferruh Yigit:
> > On 3/31/2019 2:14 PM, Pavan Nikhilesh Bhagavatula wrote:
> > > From: Pavan Nikhilesh <pbhagavatula@marvell.com>
> > > 
> > > Optimize testpmd txonly mode by
> > > 1. Moving per packet ethernet header copy above the loop.
> > > 2. Use bulk ops for allocating segments instead of having a inner
> > > loop
> > > for every segment.
> > > 
> > > Also, move the packet prepare logic into a separate function so
> > > that it
> > > can be reused later.
> > > 
> > > Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
> > > ---
> > >  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)
> > 
> > Hi Thomas, Shahafs,
> > 
> > I guess there was a performance issue on Mellanox with this patch,
> > I assume it
> > is still valid, since this version only has some cosmetic change,
> > but can you
> > please confirm?
> 
> We will check it.
> 
> > And what is the next step, can you guys provide some info to Pavan
> > to solve the
> > issue, or perhaps even better a fix?
> 
> Looking at the first patch, there are still 3 changes merged
> together.
> Why not splitting even more?

Splitting further more is not a issue. But we should not start the
thread for squashing it patch latter. What would be interesting to know
if there is any performance degradation with Mellanox NIC? If so, Why?
Based on that, We can craft the patch as you need.


> 
> 
>
Thomas Monjalon April 2, 2019, 7:06 a.m. UTC | #4 
02/04/2019 03:03, Jerin Jacob Kollanukkaran:
> On Mon, 2019-04-01 at 22:53 +0200, Thomas Monjalon wrote:
> > 01/04/2019 22:25, Ferruh Yigit:
> > > On 3/31/2019 2:14 PM, Pavan Nikhilesh Bhagavatula wrote:
> > > > From: Pavan Nikhilesh <pbhagavatula@marvell.com>
> > > > 
> > > > Optimize testpmd txonly mode by
> > > > 1. Moving per packet ethernet header copy above the loop.
> > > > 2. Use bulk ops for allocating segments instead of having a inner
> > > > loop
> > > > for every segment.
> > > > 
> > > > Also, move the packet prepare logic into a separate function so
> > > > that it
> > > > can be reused later.
> > > > 
> > > > Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
> > > > ---
> > > >  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)
> > > 
> > > Hi Thomas, Shahafs,
> > > 
> > > I guess there was a performance issue on Mellanox with this patch,
> > > I assume it
> > > is still valid, since this version only has some cosmetic change,
> > > but can you
> > > please confirm?
> > 
> > We will check it.
> > 
> > > And what is the next step, can you guys provide some info to Pavan
> > > to solve the
> > > issue, or perhaps even better a fix?
> > 
> > Looking at the first patch, there are still 3 changes merged
> > together.
> > Why not splitting even more?
> 
> Splitting further more is not a issue. But we should not start the
> thread for squashing it patch latter. What would be interesting to know
> if there is any performance degradation with Mellanox NIC? If so, Why?
> Based on that, We can craft the patch as you need.

Regarding Mellanox degradation, we need to check if this is in this patch.

Not related to performance degradation, it is a good practice
to split logical changes of a rework.
Jerin Jacob Kollanukkaran April 2, 2019, 8:31 a.m. UTC | #5 
On Tue, 2019-04-02 at 09:06 +0200, Thomas Monjalon wrote:
> 02/04/2019 03:03, Jerin Jacob Kollanukkaran:
> > On Mon, 2019-04-01 at 22:53 +0200, Thomas Monjalon wrote:
> > > 01/04/2019 22:25, Ferruh Yigit:
> > > > On 3/31/2019 2:14 PM, Pavan Nikhilesh Bhagavatula wrote:
> > > > > From: Pavan Nikhilesh <pbhagavatula@marvell.com>
> > > > > 
> > > > > Optimize testpmd txonly mode by
> > > > > 1. Moving per packet ethernet header copy above the loop.
> > > > > 2. Use bulk ops for allocating segments instead of having a
> > > > > inner
> > > > > loop
> > > > > for every segment.
> > > > > 
> > > > > Also, move the packet prepare logic into a separate function
> > > > > so
> > > > > that it
> > > > > can be reused later.
> > > > > 
> > > > > Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
> > > > > ---
> > > > >  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)
> > > > 
> > > > Hi Thomas, Shahafs,
> > > > 
> > > > I guess there was a performance issue on Mellanox with this
> > > > patch,
> > > > I assume it
> > > > is still valid, since this version only has some cosmetic
> > > > change,
> > > > but can you
> > > > please confirm?
> > > 
> > > We will check it.
> > > 
> > > > And what is the next step, can you guys provide some info to
> > > > Pavan
> > > > to solve the
> > > > issue, or perhaps even better a fix?
> > > 
> > > Looking at the first patch, there are still 3 changes merged
> > > together.
> > > Why not splitting even more?
> > 
> > Splitting further more is not a issue. But we should not start the
> > thread for squashing it patch latter. What would be interesting to
> > know
> > if there is any performance degradation with Mellanox NIC? If so,
> > Why?
> > Based on that, We can craft the patch as you need.
> 
> Regarding Mellanox degradation, we need to check if this is in this
> patch.

Please check.

> 
> Not related to performance degradation, it is a good practice
> to split logical changes of a rework.

Yes. No disagreement here. What I would like to emphasis here is that,
It is critical to know, Is there any degradation as this patch has been
blocked by saying there is a degradation with Mellanox NIC.

It is trivial to split the patches N more logical one. The former one
is complex and it has dependency.




> 
>
Ali Alnubani April 2, 2019, 9:03 a.m. UTC | #6 
Hi Pavan,

> -----Original Message-----
> From: dev <dev-bounces@dpdk.org> On Behalf Of Pavan Nikhilesh
> Bhagavatula
> Sent: Sunday, March 31, 2019 4:14 PM
> To: Jerin Jacob Kollanukkaran <jerinj@marvell.com>; Thomas Monjalon
> <thomas@monjalon.net>; arybchenko@solarflare.com;
> ferruh.yigit@intel.com; bernard.iremonger@intel.com
> Cc: dev@dpdk.org; Pavan Nikhilesh Bhagavatula
> <pbhagavatula@marvell.com>
> Subject: [dpdk-dev] [PATCH v5 1/2] app/testpmd: optimize testpmd txonly
> mode
> 
> From: Pavan Nikhilesh <pbhagavatula@marvell.com>
> 
> Optimize testpmd txonly mode by
> 1. Moving per packet ethernet header copy above the loop.
> 2. Use bulk ops for allocating segments instead of having a inner loop for
> every segment.
> 
> Also, move the packet prepare logic into a separate function so that it can be
> reused later.
> 
> Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
> ---
>  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 | 139 +++++++++++++++++++++++-------------------
>  1 file changed, 76 insertions(+), 63 deletions(-)
> 

The patch doesn't apply. It conflicts with 82010ef55 app/testpmd: make txonly mode generate multiple flows (http://patches.dpdk.org/patch/51869/).
Can you please rebase?

Thanks,
Ali
Pavan Nikhilesh Bhagavatula April 2, 2019, 9:06 a.m. UTC | #7 
Hi Ali,

> -----Original Message-----
> From: Ali Alnubani <alialnu@mellanox.com>
> Sent: Tuesday, April 2, 2019 2:33 PM
> To: Pavan Nikhilesh Bhagavatula <pbhagavatula@marvell.com>
> Cc: dev@dpdk.org; Jerin Jacob Kollanukkaran <jerinj@marvell.com>; Thomas
> Monjalon <thomas@monjalon.net>; arybchenko@solarflare.com;
> ferruh.yigit@intel.com; bernard.iremonger@intel.com
> Subject: [EXT] RE: [dpdk-dev] [PATCH v5 1/2] app/testpmd: optimize testpmd
> txonly mode
> 
> External Email
> 
> ----------------------------------------------------------------------
> Hi Pavan,
> 
> > -----Original Message-----
> > From: dev <dev-bounces@dpdk.org> On Behalf Of Pavan Nikhilesh
> > Bhagavatula
> > Sent: Sunday, March 31, 2019 4:14 PM
> > To: Jerin Jacob Kollanukkaran <jerinj@marvell.com>; Thomas Monjalon
> > <thomas@monjalon.net>; arybchenko@solarflare.com;
> > ferruh.yigit@intel.com; bernard.iremonger@intel.com
> > Cc: dev@dpdk.org; Pavan Nikhilesh Bhagavatula
> > <pbhagavatula@marvell.com>
> > Subject: [dpdk-dev] [PATCH v5 1/2] app/testpmd: optimize testpmd
> > txonly mode
> >
> > From: Pavan Nikhilesh <pbhagavatula@marvell.com>
> >
> > Optimize testpmd txonly mode by
> > 1. Moving per packet ethernet header copy above the loop.
> > 2. Use bulk ops for allocating segments instead of having a inner loop
> > for every segment.
> >
> > Also, move the packet prepare logic into a separate function so that
> > it can be reused later.
> >
> > Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
> > ---
> >  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 | 139
> > +++++++++++++++++++++++-------------------
> >  1 file changed, 76 insertions(+), 63 deletions(-)
> >
> 
> The patch doesn't apply. It conflicts with 82010ef55 app/testpmd: make
> txonly mode generate multiple flows
> (http://patches.dpdk.org/patch/51869/).
> Can you please rebase?

Yes, I'm in the process of rebasing/splitting the patch further.

> 
> Thanks,
> Ali

Regards,
Pavan.
diff mbox series

Patch

diff --git a/app/test-pmd/txonly.c b/app/test-pmd/txonly.c
index 1f08b6ed3..9c0147089 100644
--- a/app/test-pmd/txonly.c
+++ b/app/test-pmd/txonly.c
@@ -147,6 +147,63 @@  setup_pkt_udp_ip_headers(struct ipv4_hdr *ip_hdr,
 	ip_hdr->hdr_checksum = (uint16_t) ip_cksum;
 }

+static inline bool
+pkt_burst_prepare(struct rte_mbuf *pkt, struct rte_mempool *mbp,
+		struct ether_hdr *eth_hdr, const uint16_t vlan_tci,
+		const uint16_t vlan_tci_outer, const uint64_t ol_flags)
+{
+	struct rte_mbuf *pkt_segs[RTE_MAX_SEGS_PER_PKT];
+	struct rte_mbuf *pkt_seg;
+	uint32_t nb_segs, pkt_len;
+	uint8_t i;
+
+	if (unlikely(tx_pkt_split == TX_PKT_SPLIT_RND))
+		nb_segs = random() % tx_pkt_nb_segs + 1;
+	else
+		nb_segs = tx_pkt_nb_segs;
+
+	if (nb_segs > 1) {
+		if (rte_mempool_get_bulk(mbp, (void **)pkt_segs, nb_segs))
+			return false;
+	}
+
+	rte_pktmbuf_reset_headroom(pkt);
+	pkt->data_len = tx_pkt_seg_lengths[0];
+	pkt->ol_flags = ol_flags;
+	pkt->vlan_tci = vlan_tci;
+	pkt->vlan_tci_outer = vlan_tci_outer;
+	pkt->l2_len = sizeof(struct ether_hdr);
+	pkt->l3_len = sizeof(struct ipv4_hdr);
+
+	pkt_len = pkt->data_len;
+	pkt_seg = pkt;
+	for (i = 1; i < nb_segs; i++) {
+		pkt_seg->next = pkt_segs[i - 1];
+		pkt_seg = pkt_seg->next;
+		pkt_seg->data_len = tx_pkt_seg_lengths[i];
+		pkt_len += pkt_seg->data_len;
+	}
+	pkt_seg->next = NULL; /* Last segment of packet. */
+	/*
+	 * Copy headers in first packet segment(s).
+	 */
+	copy_buf_to_pkt(eth_hdr, sizeof(eth_hdr), pkt, 0);
+	copy_buf_to_pkt(&pkt_ip_hdr, sizeof(pkt_ip_hdr), pkt,
+			sizeof(struct ether_hdr));
+	copy_buf_to_pkt(&pkt_udp_hdr, sizeof(pkt_udp_hdr), pkt,
+			sizeof(struct ether_hdr) +
+			sizeof(struct ipv4_hdr));
+
+	/*
+	 * Complete first mbuf of packet and append it to the
+	 * burst of packets to be transmitted.
+	 */
+	pkt->nb_segs = nb_segs;
+	pkt->pkt_len = pkt_len;
+
+	return true;
+}
+
 /*
  * Transmit a burst of multi-segments packets.
  */
@@ -156,7 +213,6 @@  pkt_burst_transmit(struct fwd_stream *fs)
 	struct rte_mbuf *pkts_burst[MAX_PKT_BURST];
 	struct rte_port *txp;
 	struct rte_mbuf *pkt;
-	struct rte_mbuf *pkt_seg;
 	struct rte_mempool *mbp;
 	struct ether_hdr eth_hdr;
 	uint16_t nb_tx;
@@ -164,14 +220,12 @@  pkt_burst_transmit(struct fwd_stream *fs)
 	uint16_t vlan_tci, vlan_tci_outer;
 	uint32_t retry;
 	uint64_t ol_flags = 0;
-	uint8_t  i;
 	uint64_t tx_offloads;
 #ifdef RTE_TEST_PMD_RECORD_CORE_CYCLES
 	uint64_t start_tsc;
 	uint64_t end_tsc;
 	uint64_t core_cycles;
 #endif
-	uint32_t nb_segs, pkt_len;

 #ifdef RTE_TEST_PMD_RECORD_CORE_CYCLES
 	start_tsc = rte_rdtsc();
@@ -188,72 +242,31 @@  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) {
-		nomore_mbuf:
-			if (nb_pkt == 0)
-				return;
+		if (pkt == NULL)
+			break;
+		if (unlikely(!pkt_burst_prepare(pkt, mbp,
+						&eth_hdr, vlan_tci,
+						vlan_tci_outer,
+						ol_flags))) {
+			rte_mempool_put(mbp, pkt);
 			break;
 		}
-
-		/*
-		 * Using raw alloc is good to improve performance,
-		 * but some consumers may use the headroom and so
-		 * decrement data_off. We need to make sure it is
-		 * reset to default value.
-		 */
-		rte_pktmbuf_reset_headroom(pkt);
-		pkt->data_len = tx_pkt_seg_lengths[0];
-		pkt_seg = pkt;
-		if (tx_pkt_split == TX_PKT_SPLIT_RND)
-			nb_segs = random() % tx_pkt_nb_segs + 1;
-		else
-			nb_segs = tx_pkt_nb_segs;
-		pkt_len = pkt->data_len;
-		for (i = 1; i < nb_segs; i++) {
-			pkt_seg->next = rte_mbuf_raw_alloc(mbp);
-			if (pkt_seg->next == NULL) {
-				pkt->nb_segs = i;
-				rte_pktmbuf_free(pkt);
-				goto nomore_mbuf;
-			}
-			pkt_seg = pkt_seg->next;
-			pkt_seg->data_len = tx_pkt_seg_lengths[i];
-			pkt_len += pkt_seg->data_len;
-		}
-		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).
-		 */
-		copy_buf_to_pkt(&eth_hdr, sizeof(eth_hdr), pkt, 0);
-		copy_buf_to_pkt(&pkt_ip_hdr, sizeof(pkt_ip_hdr), pkt,
-				sizeof(struct ether_hdr));
-		copy_buf_to_pkt(&pkt_udp_hdr, sizeof(pkt_udp_hdr), pkt,
-				sizeof(struct ether_hdr) +
-				sizeof(struct ipv4_hdr));
-
-		/*
-		 * Complete first mbuf of packet and append it to the
-		 * burst of packets to be transmitted.
-		 */
-		pkt->nb_segs = nb_segs;
-		pkt->pkt_len = pkt_len;
-		pkt->ol_flags = ol_flags;
-		pkt->vlan_tci = vlan_tci;
-		pkt->vlan_tci_outer = vlan_tci_outer;
-		pkt->l2_len = sizeof(struct ether_hdr);
-		pkt->l3_len = sizeof(struct ipv4_hdr);
 		pkts_burst[nb_pkt] = pkt;
 	}
+
+	if (nb_pkt == 0)
+		return;
+
 	nb_tx = rte_eth_tx_burst(fs->tx_port, fs->tx_queue, pkts_burst, nb_pkt);
 	/*
 	 * Retry if necessary