On 4/22/2023 12:20 AM, Joshua Washington wrote:
> Google cloud routes traffic using IP addresses without the support of MAC
> addresses, so changing source IP address for txonly-multi-flow can have
> negative performance implications for net/gve when using testpmd. This
> patch updates txonly multiflow mode to modify source ports instead of
> source IP addresses.
>
> The change can be tested with the following command:
> dpdk-testpmd -- --forward-mode=txonly --txonly-multi-flow \
> --tx-ip=<SRC>,<DST>
>
> Signed-off-by: Joshua Washington <joshw...@google.com>
> Reviewed-by: Rushil Gupta <rush...@google.com>
> ---
> app/test-pmd/txonly.c | 39 +++++++++++++++++++++++----------------
> 1 file changed, 23 insertions(+), 16 deletions(-)
>
> diff --git a/app/test-pmd/txonly.c b/app/test-pmd/txonly.c
> index b3d6873104..f79e0e5d0b 100644
> --- a/app/test-pmd/txonly.c
> +++ b/app/test-pmd/txonly.c
> @@ -56,7 +56,7 @@ uint32_t tx_ip_dst_addr = (198U << 24) | (18 << 16) | (0 <<
> 8) | 2;
> #define IP_DEFTTL 64 /* from RFC 1340. */
>
> static struct rte_ipv4_hdr pkt_ip_hdr; /**< IP header of transmitted
> packets. */
> -RTE_DEFINE_PER_LCORE(uint8_t, _ip_var); /**< IP address variation */
> +RTE_DEFINE_PER_LCORE(uint8_t, _src_var); /**< Source port variation */
What about '_src_port_var' as variable name?
> static struct rte_udp_hdr pkt_udp_hdr; /**< UDP header of tx packets. */
>
> static uint64_t timestamp_mask; /**< Timestamp dynamic flag mask */
> @@ -230,28 +230,35 @@ pkt_burst_prepare(struct rte_mbuf *pkt, struct
> rte_mempool *mbp,
> 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 rte_ether_hdr));
> + copy_buf_to_pkt(&pkt_udp_hdr, sizeof(pkt_udp_hdr), pkt,
> + sizeof(struct rte_ether_hdr) +
> + sizeof(struct rte_ipv4_hdr));
> if (txonly_multi_flow) {
> - uint8_t ip_var = RTE_PER_LCORE(_ip_var);
> - struct rte_ipv4_hdr *ip_hdr;
> - uint32_t addr;
> + uint16_t src_var = RTE_PER_LCORE(_src_var);
> + struct rte_udp_hdr *udp_hdr;
> + uint16_t port;
'port' is used in testpmd for port_id in many places, what do you think
to rename variable as 'src_port'?
>
> - ip_hdr = rte_pktmbuf_mtod_offset(pkt,
> - struct rte_ipv4_hdr *,
> - sizeof(struct rte_ether_hdr));
> + udp_hdr = rte_pktmbuf_mtod_offset(pkt,
> + struct rte_udp_hdr *,
> + sizeof(struct rte_ether_hdr) +
> + sizeof(struct rte_ipv4_hdr));
> /*
> - * Generate multiple flows by varying IP src addr. This
> - * enables packets are well distributed by RSS in
> + * Generate multiple flows by varying UDP source port.
> + * This enables packets are well distributed by RSS in
> * receiver side if any and txonly mode can be a decent
> * packet generator for developer's quick performance
> * regression test.
> + *
> + * Only ports in the range 49152 (0xC000) and 65535 (0xFFFF)
> + * will be used, with the least significant byte representing
> + * the lcore ID. As such, the most significant byte will cycle
> + * through 0xC0 and 0xFF.
> */
> - addr = (tx_ip_dst_addr | (ip_var++ << 8)) + rte_lcore_id();
> - ip_hdr->src_addr = rte_cpu_to_be_32(addr);
> - RTE_PER_LCORE(_ip_var) = ip_var;
> + port = ((((src_var++) % (0xFF - 0xC0) + 0xC0) & 0xFF) << 8)
> + + rte_lcore_id();
To prevent '%' in the datapath, I think following does the same thing:
port = ((src_var++ | 0xC0) << 8) + rte_lcore_id();
Although it is not big deal, in original version you never got 0xFFXX
values:
3E % 3F + C0 = FE
3F % 3F + C0 = C0
