Declan, Mohammad, The submission [1] of steering action between switch ports clearly requires a switch model in DPDK. The Port Representor based on a virtual PMD broker on NIC ops (rte_dev_ops) does not provide the required functionality. Using NIC terminology and not Switch API's will lead to a dead-end. Moreover, it does not fit the Kernel design. We need to be careful from this ending up as two different deployment models for users, which is very bad. There was a long discussion about this in netdev ML [2], including the VEPA mode support.
As described in the links Alejandro referenced earlier, each of the switch ports should be a real PMD, and switch operations should be applied on these PMD ports. This includes the steering redirection of traffic between switch ports [1], port ACL's to block/allow traffic, VST/VGT modes and anti spoofing, link trust mode [3] for promiscuous configuration, mirroring of switch port traffic, and Tx and Rx of switch port traffic to/from VF's port. More over, building this as real PMD ports of a switch device removes the need to add a new broker framework all together. Each vendor just needs to map additional PMD ports during the probing stage. By adding a switchdev_id we can define these are ports associated to the same switching device, and can allow new port and inter-port actions. [1] http://dpdk.org/dev/patchwork/patch/32550/ [2] https://www.spinics.net/lists/netdev/msg467375.html [2] https://www.systutorials.com/docs/linux/man/8-ip-link/ Alex
