[dpdk-dev] DPDK19.11 Mellanox Roadmap

[Asaf Penso]

This is Mellanox's roadmap for DPDK19.11, which we are working on currently.



Introducing rte_flow APIs:

* Setting Metadata in RX and receiving it as part of mbuf, to have more info from packet processing in the NIC.

   In case of multiple tables offloads, like Connection Tracking with OVS-DPDK, it can be used to understand which flow was missed.

Introduce rte_eth_dev APIs for:

* Hairpin, which acts as a "bump on the wire", saving the host’s cores processing.

   This can be used, for example, in a gateway application, where decap, header rewrite and encap offload can be done to the packet and sent it back to the wire, without any SW involvement.

* Use generic functions for VF configuration, like MAC address, by using representors ports.



Those APIs will be supported via mlx5 PMD for ConnectX-5 SmartNIC, BlueField IPU and above.

Introduce hash_list data structure, for optimized insertion and query rate, that has a re-sizeable option.
    Using hash_list data structure improves lookup performance of mark ids and table ids; especially important when using millions of elements.
    The existing data structure requires pre-allocation of memory for the max amount of elements, which is not efficient.

    The new implementation starts with a small number of elements and allocates more dynamically during the runtime.


Additions to mlx5 PMD (ConnectX-5 SmartNIC, BlueField IPU and above):

* Support for reading the plugin module EEPROM module in specific, the below functions will be implemented:

    * rte_eth_dev_get_module_info

    * rte_eth_dev_get_module_eeprom


* VF LAG (LACP) to bond 2 VFs of the same device into a single interface.
   With this, no need of DPDK bonding for mlx5 VFs, and the LAG functionality is offloaded to the HW with bifurcated driver.

* Flow metering offload to HW, to provide better bandwidth QoS per flow.
   The supported bandwidth metering type is srTCM (IETF RFC 2697) color blind mode, and the support is for 4K meters.
   Seamless support, without any API changes, for millions of meters which will be supported in ConnectX-6 Dx.

* HW offload for VLAN push and pop

* HW offload for header modify of VLAN ID field

* Improved rte_flow_create error messages. Erroneous calls will be returned with meaningful information, which provides clearer reasons about the error root cause.