Hi Kevin, Kevin Traynor, Nov 11, 2022 at 19:15: > > + reta_conf_size = (dev->reta_size / RTE_ETH_RETA_GROUP_SIZE) > > + * sizeof(struct rte_eth_rss_reta_entry64); > > In dpdk_eth_dev_init, we get reta_size from driver, > > mlx5_ethdev.c > > 333├> info->reta_size = priv->reta_idx_n ? > 334│ priv->reta_idx_n : config->ind_table_max_size; > > (gdb) p priv->reta_idx_n > $5 = 1 > (gdb) p config->ind_table_max_size > $6 = 512 > > and store: > dev->reta_size = info.reta_size; > > Now we use it, > dev->reta_size = 1 / RTE_ETH_RETA_GROUP_SIZE (64) > but it results in reta_conf_size = 0 > > > + reta_conf = xmalloc(reta_conf_size); > > xmalloc only allocates 1 byte (void *p = malloc(size ? size : 1);)
Hmm indeed this is bad :) Since reworking the RSS fallback I may not have re-tested with CX-5, only with Intel NICs. I am quite surprised of this return value. Could it be an issue with the mlx5 dpdk driver? Thanks for testing!
