#networking #ethernet #tcp

ethox-io-uring

An io-uring device implementation for ethox

1 unstable release

0.0.2 Feb 8, 2020

#42 in #ethernet

AGPL-3.0-only

1MB
20K SLoC

Ethox

Ethox is a standalone library for user-space networking and unikernel systems with a focus on speeed and stability. There exist basic structures for compiling its features into a complete network stack. It is also possibly useful for bare-metal microcontrollers but it is not engineered towards that primary goal.

As part of the consistent performance promises, Ethox restricts itself to never perform internal allocation. However, the user may choose to do so where deemed nessary. It targets stable and nightly Rust only, and does not aim to be compatible to previous stable releases.

Overview and Features

The initial feature set will be eth+(ipv4|ipv6)+(udp+tcp). Also arp and icmp are supported. See the more complete feature list below.

CI Status License Scc lines of code Scc comments

Usage

The main interface of Ethox is built around zero-copy, for as long as possible. It is not socket oriented but based on a dynamically built tree of trait implementors which offer callbacks for layer specific functionality and packet receiving and transmission. The packet buffer is never owned within these callbacks but a mutable, unique reference to a network device specific structure that can be resized and reused at will. Batching of packets for both ingress and egress is enabled at the NIC level. Batching is not provided in the layers above that.

License

Ethox is distributed under the AGPLv3 (Why). Code contributions are only accepted under waiver of copyright, at the moment, to allow freely choosing other licensing options further down the road. These conditions may be opened up a bit in the future.

A significant but shrinking portion of the original network code comes from smoltcp, copyright whitequark@whitequark.org, and reproduced and modified here under the terms allowed by its 0-clause BSD license. It may have changed a lot by the time you read this.

Details

More details about each layer, with supported, unsupported and work-in-progress features. Keep in mind that this may still evolve rapidly and is maybe not up-to-date. This also documents possible future additions.

Ethernet

Ethernet is the only supported medium/link layer.

  • Regular Ethernet II frames are supported.
  • Unicast and broadcast packets (not multicast) are supported.
  • 802.3 frames and 802.1Q are not supported.
  • Jumbo frames are not supported.

IPv4

  • IPv4 header checksum is generated and validated. May be ignored.
  • CIDR tables are supported.
  • QoS and TTL per route are not supported.
  • Link local routing is supported.
  • Broadcast and Network addressing is supported.
  • Prefix 31 and 32 networks are supported.
  • IGMP is not supported.
  • IPv4 fragmentation is not supported.
  • IPv4 options are not supported and silently discarded.

IPv4 — Icmpv4

  • Icmpv4 echo replies are generated.
  • Icmpv4 header checksums are supported.
  • Messages (including unreachable errors) may be passed to custom receiver logic.
  • Icmpv4 errors are not generated.

IPv4 — Arp

  • ARP packets (requests and queries) are automatically performed supported.
  • ARP entries are revalidated periodically (1 minute).
  • Gratuitous requests and replies are not supported and ignored.
  • Auto-configuration and collision detection is not supported.

IPv6

  • IPv6 options are not supported.
  • IPv6 fragmentation is not supported.

IPv6 — Icmpv6

Is not supported yet.

IPv6 — NDISC

Neighbor discovery is not supported.

Tcp

  • Header checksums are generated and validated. May be ignored.
  • Zero-copy receiving and sending of messages, all buffers under user control.
  • Adheres to maximum segment size.
  • Windows scaling is negotiated and utilized. May be configured.
  • Predefined structures for arbitrary length reassembly are available.
  • Bytes-in-flight are not limited by segment sizes.
  • Initial sequence number is generated according to rfc6528 (keyed siphash-2-4).
  • Exponential backoff is not supported.
  • Selective acknowledgment is ignored.
  • Congestion control is not implemented.
  • Round-trip-time estimation is not implemented.
  • MTU discovery is not implemented.
  • Delayed acknowledgments are not implemented.
  • Silly window syndrome avoidance is not implemented.
  • Nagle's algorithm is not implemented.
  • Timestamping is not supported.
  • Urgent pointer is ignored.
  • Probing Zero Windows is not implemented.

Udp

  • Header checksum is generated, validated, can be elided, may be ignored.
  • Zero-copy receiving and sending of messages

Why AGPL

I will likely at some point want to relicense the whole lot under the most permissive license possible in the spirit of the original library. However, I can not commit to actively developing this library at the moment. It is my hope that the license encourages users, and especially those with finished products, to contribute code back and expand the scope. If you disagree with this stance, ethically or legally or otherwise, then a friendly e-mail detailing your point of view is most welcome.

Dependencies