DPDK: Create multiple RSS flows with unique set of queues for different patterns

DPDK: Create multiple RSS flows with unique set of queues for different patterns

[Raghavan V]

[-- Attachment #1: Type: text/plain, Size: 2682 bytes --]

Hi Team,

Is there any possible way to create multiple RSS flows to distribute packets to set of queues when a particular Ip pattern matches.
For e.g.,
I have 40 RX queues setup,
Flow 1: Match a particular IP pattern and distribute it to specified RX queues set (0 - 29)
Flow 2: Match a particular IP pattern and distribute it to specified RX queues set (30 - 39)

Is the below code snippet possible? Also is it possible to use symmetric toeplitz  hash and ETH_RSS_L3_SRC_ONLY | ETH_RSS_L3_DST_ONLY and let me know any other config should be included.

uint16_t queue_indices[] = {0,...29};
uint16_t queue_indices2[] = {30,..39};

action_rss1 = (struct rte_flow_action_rss){
                    .types = rss_conf.rss_hf,
                    .key_len = rss_conf.rss_key_len,
                    .queue_num = 30,
                    .key = rss_key,
                    .queue = queue_indices,
            };

action_rss2 = (struct rte_flow_action_rss){
                    .types = rss_conf.rss_hf,
                    .key_len = rss_conf.rss_key_len,
                    .queue_num = 10,
                    .key = rss_key,
                    .queue = queue_indices2,
            };

struct rte_flow_item pattern[] = {
                [0]={
                        .type = RTE_FLOW_ITEM_TYPE_ETH,

                },
                [1]={
                        .type = RTE_FLOW_ITEM_TYPE_IPV4,
                        .spec = &ipv4_spec,
                        .mask = &ipv4_mask,


                },
                [2]={
                        .type = RTE_FLOW_ITEM_TYPE_UDP,
                },

                [3]={
                        .type = RTE_FLOW_ITEM_TYPE_END,
                }
        };
        struct rte_flow_item pattern2[] = {
                [0]={
                        .type = RTE_FLOW_ITEM_TYPE_ETH,

                },

                [1]={
                        .type = RTE_FLOW_ITEM_TYPE_IPV4,
                        .spec = &ipv4_spec2,
                        .mask = &ipv4_mask2,


                },
                [2]={
                        .type = RTE_FLOW_ITEM_TYPE_UDP,
                },

                [3]={
                        .type = RTE_FLOW_ITEM_TYPE_END,
                }
        };
rte_flow_validate(bond_port, &attr, pattern, action_rss1, &error);
rte_flow_create(bond_port, &attr, pattern, action_rss1, &error);

rte_flow_validate(bond_port, &attr, pattern2, action_rss2, &error);
rte_flow_create(bond_port, &attr, pattern2, action_rss2, &error);


Pls let me know any other approach will match my requirement.

Thanks,
Raghavan V


[-- Attachment #2: Type: text/html, Size: 13534 bytes --]

Re: DPDK: Create multiple RSS flows with unique set of queues for different patterns

[Stephen Hemminger]

On Thu, 8 Aug 2024 11:19:13 +0000
Raghavan V <Raghavan.V2@tatacommunications.com> wrote:

> Hi Team,
> 
> Is there any possible way to create multiple RSS flows to distribute packets to set of queues when a particular Ip pattern matches.
> For e.g.,
> I have 40 RX queues setup,
> Flow 1: Match a particular IP pattern and distribute it to specified RX queues set (0 - 29)
> Flow 2: Match a particular IP pattern and distribute it to specified RX queues set (30 - 39)

You need to address what to with packets that do match either of the patterns.
Usually best to reserve one queue (0) for those.

> 
> Is the below code snippet possible? Also is it possible to use symmetric toeplitz  hash and ETH_RSS_L3_SRC_ONLY | ETH_RSS_L3_DST_ONLY and let me know any other config should be included.

That is HW dependent I suspect most hardware doesn't support it. Probably only mlx5.
You are best to read the driver source to find out if you want to do special case like this.

> 
> uint16_t queue_indices[] = {0,...29};
> uint16_t queue_indices2[] = {30,..39};
> 
> action_rss1 = (struct rte_flow_action_rss){
>                     .types = rss_conf.rss_hf,
>                     .key_len = rss_conf.rss_key_len,

Usually best to go with default rss_key and leave key_len and key as default of 0.

>                     .queue_num = 30,
>                     .key = rss_key,
>                     .queue = queue_indices,
>             };
> 
> action_rss2 = (struct rte_flow_action_rss){
>                     .types = rss_conf.rss_hf,
>                     .key_len = rss_conf.rss_key_len,
>                     .queue_num = 10,
>                     .key = rss_key,
>                     .queue = queue_indices2,
>             };
> 
> struct rte_flow_item pattern[] = {
>                 [0]={
>                         .type = RTE_FLOW_ITEM_TYPE_ETH,
> 
>                 },
>                 [1]={
>                         .type = RTE_FLOW_ITEM_TYPE_IPV4,
>                         .spec = &ipv4_spec,
>                         .mask = &ipv4_mask,
> 
> 
>                 },
>                 [2]={
>                         .type = RTE_FLOW_ITEM_TYPE_UDP,
>                 },
> 
>                 [3]={
>                         .type = RTE_FLOW_ITEM_TYPE_END,
>                 }
>         };
>         struct rte_flow_item pattern2[] = {
>                 [0]={
>                         .type = RTE_FLOW_ITEM_TYPE_ETH,
> 
>                 },
> 
>                 [1]={
>                         .type = RTE_FLOW_ITEM_TYPE_IPV4,
>                         .spec = &ipv4_spec2,
>                         .mask = &ipv4_mask2,
> 
> 
>                 },
>                 [2]={
>                         .type = RTE_FLOW_ITEM_TYPE_UDP,
>                 },
> 
>                 [3]={
>                         .type = RTE_FLOW_ITEM_TYPE_END,
>                 }
>         };
> rte_flow_validate(bond_port, &attr, pattern, action_rss1, &error);
> rte_flow_create(bond_port, &attr, pattern, action_rss1, &error);
> 
> rte_flow_validate(bond_port, &attr, pattern2, action_rss2, &error);
> rte_flow_create(bond_port, &attr, pattern2, action_rss2, &error);
> 
> 
> Pls let me know any other approach will match my requirement.
> 
> Thanks,
> Raghavan V
>