Hi Fabio,
Have you considered to select MQ mode 'RSS' [1] via 'dev_conf.rxmode.mq_mode'
during 'rte_eth_dev_configure' [2] stage? Just to double-check.
Also, the 'RSS_HASH' flag could possibly mean the availability of 32-bit hash
value in received mbufs rather than support for RSS distribution as a whole.
[1]
https://doc.dpdk.org/api-24.07/rte__ethdev_8h.html#a586b8e86131b4ec0ccaf464e847ccf3e
[2]
https://doc.dpdk.org/api-24.07/rte__ethdev_8h.html#a1a7d3a20b102fee222541fda50fd87bd
Thank you.
On Sat, 10 Aug 2024, Fabio Fernandes wrote:
Hi,
I'm a bit puzzled why DPDK is reporting RSS to be not available on my ENA port
(at AWS) when I'm expecting it to be.
My setup is:
c7a.4xlarge
[ENA_COM: ena_com_validate_version]ENA device version: 0.10
[ENA_COM: ena_com_validate_version]ENA controller version: 0.0.1
implementation version 1
16-cores (hyper-threading disabled),
4 ENA interfaces/network adapters, but only one (with public IP) is bound to
igb_uio
DPDK 24.07
6.5.0-1020-aws,
Ubuntu 22.04.4 LTS,
The device reports that 8 queues are available, but
rte_eth_dev_info::rx_offload_capa does not have RTE_ETH_RX_OFFLOAD_RSS_HASH set
after calling rte_eth_dev_info_get().
The flag value seems to be coming from here,
https://github.com/DPDK/dpdk/blob/b3485f4293997d35b6daecc3437bb0c183a51fb3/drivers/net/ena/ena_ethdev.c#L2402
Where the offload mask is:
(ena_com_dev_get_features_ctx*) get_feat_ctx->offload
tx = 3
rx_supported = 7
rx_enabled = 0
Isn't RSS supposed to be available on this ENA setup?
If not, how am I supposed to distribute inbound packets across those 8 receive
queues/CPUs?
Some notes, log:
at eth_ena_dev_init(), ena_use_large_llq_hdr() returned true
ena_get_metrics_entries(): 0x6 customer metrics are supported
ena_use_large_llq_hdr(): Recommended device entry size policy 4
eth_ena_dev_init(): LLQ entry size 256B
eth_ena_dev_init(): Placement policy: Low latency
Any advice?
Thanks,
Fabio
