This package implements faces using DPDK ethdev as transport.
Face type represents an Ethernet-based face or a memif face. See package ethface and package memifface for more information.
Port type organizes faces on the same DPDK ethdev. It manages ethdev resources and prevents conflicts among the faces.
There are three receive path implementations. One of them is chosen during port creation; the choice cannot be changed afterwards.
RxFlow is a hardware-accelerated receive path. It uses one or more RX queues per face, and creates a flow via rte_flow API to steer incoming frames to those queues. The hardware performs header matching; there is minimal checking on software side.
RxTable is a software receive path.
It continuously polls ethdev RX queue 0 for incoming frames.
For each incoming frame, the software performs header matching (implemented in EthRxMatch struct), and then labels each matched frame with the face ID.
Matchings are attempted iteratively for each face that are arranged in an RCU-protected linked list; if the port has a pass-through face, it is arranged last and would always match.
If no match is found for an incoming frame, the Ethernet frame is sent to packet dumper if enabled, otherwise it is dropped.
RxMemif is a memif-specific receive path, where each port has only one face. It continuously polls ethdev RX queue 0 for incoming frames, and then labels each frame with the only face ID.
EthFace_TxBurst function implements the send path.
Currently, the send path only uses ethdev TX queue 0.
It prepends Ethernet/UDP/VXLAN headers to each frame (implemented in EthTxHdr struct), and requires every outgoing packet to have sufficient headroom for the headers.
The send path is thread-safe only if the underlying DPDK PMD is thread safe, which generally is not the case.
Therefore, iface.TxLoop calls EthFace_TxBurst from the same thread for all faces on the same port.