RE: Feature request: MLX5 DPDK flow item type RAW support

[Asaf Penso <asafp@nvidia.com>]

[-- Attachment #1: Type: text/plain, Size: 3204 bytes --]

Thanks for the explanation!
I understand your use case, and it cannot be supported with ConnectX-5.

Regards,
Asaf Penso

From: Vladimir Yesin <v.yesin@gmail.com>
Sent: Tuesday, March 1, 2022 10:02 AM
To: Asaf Penso <asafp@nvidia.com>
Cc: users@dpdk.org
Subject: Re: Feature request: MLX5 DPDK flow item type RAW support

Hello Asaf,
I am currently working on forwarding IQ samples from EttusResearch USRP n320 SDR receiver to Nvidia GPU (via GPUDirect RDMA) with DPDK.
We are using Connectx-5 NICs.

USRP uses CHDR network protocol (https://files.ettus.com/manual_archive/release_003_009_000/html/page_rtp.html<https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Ffiles.ettus.com%2Fmanual_archive%2Frelease_003_009_000%2Fhtml%2Fpage_rtp.html&data=04%7C01%7Casafp%40nvidia.com%7Cf1ac28854ffa4ae872b008d9fb59bc44%7C43083d15727340c1b7db39efd9ccc17a%7C0%7C0%7C637817185558439757%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&sdata=%2BbSrEtZBIEfQkPU52Ed0A%2FGcwU8xC0CXrlAu4SSRL%2BY%3D&reserved=0>) over UDP for CONTROL and DATA planes. CHDR is very similar to eCPRI, supported via DPDK flex item.
I am currently working on UHD driver (https://github.com/EttusResearch/uhd<https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2FEttusResearch%2Fuhd&data=04%7C01%7Casafp%40nvidia.com%7Cf1ac28854ffa4ae872b008d9fb59bc44%7C43083d15727340c1b7db39efd9ccc17a%7C0%7C0%7C637817185558439757%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&sdata=%2BCiDDWu6XTcbjQiTF3UsOPCWx75TQ7QXyEqj%2Bhtpz%2FY%3D&reserved=0>) modification of to split CONTROL and DATA plane CHDR flows to CPU and GPU respectively with DPDK flow API via distinct HW queues and mempools.
CHDR header contains packet type field to separate these streams using RAW or FLEX items...

Currently, DPDK mlx5 PMD for Connectx-5 NICs does not implement flex and raw items.
The absence of these functions does not allow us to solve our problem.




сб, 26 февр. 2022 г. в 21:08, Asaf Penso <asafp@nvidia.com<mailto:asafp@nvidia.com>>:
Hello Vladimir,

Would you please elaborate more on your use case?
What exactly do you wish to match on?

Currently, there is no plan to implement the RAW item.
However, with Bluefield 2 DPU, we implemented the flex item, that can be useful for you.

Regards,
Asaf Penso
________________________________
From: Vladimir Yesin <v.yesin@gmail.com<mailto:v.yesin@gmail.com>>
Sent: Friday, February 25, 2022 4:28:57 PM
To: users@dpdk.org<mailto:users@dpdk.org> <users@dpdk.org<mailto:users@dpdk.org>>
Subject: Feature request: MLX5 DPDK flow item type RAW support


Current DPDK 21.11 flow API does not support RTE_FLOW_ITEM_TYPE_RAW for MLX5.

I need support of RTE_FLOW_ITEM_TYPE_RAW in DPDK flow API to enqueue some ingress packets by content to GPU with support of GPUDirect RDMA and other to CPU via distinct HW queues (RTE_FLOW_ACTION_TYPE_QUEUE).

For now RTE_FLOW_ITEM_TYPE_UDP and RTE_FLOW_ITEM_TYPE_IPV4 filtering and enqueueing with address and ports are supported.

Are there any plans to support RTE_FLOW_ITEM_TYPE_RAW for MLX5?

[-- Attachment #2: Type: text/html, Size: 8986 bytes --]