feat(phy): add TCP segmentation support

[cagatay-y]

The Windows and Linux documentations for the same feature were used as references. It was more difficult to find information on how devices support the feature, so the interface is mostly based on how virtio-net devices expect to be driven. The code was tested on the Hermit kernel with a virtio-net device, but only with IPv4, as Hermit does not support IPv6.

Design rationale

• The segmentation capabilities are specified separately for TCP on IPv4 and IPv6. This follows the example of Linux and Windows. This should provide flexibility for devices that support only one of the IP versions and should not be a problem for devices that support both as a single feature.
• As it is necessary to communicate to the device what the target segment size is, a field is added to PacketMeta. To keep the compiler able to optimise out the structure if the features associated with the fields, the segmentation offload feature is optional, following the example of packetmeta-id.
• For pre-segmentation IP packets whose size exceeds the maximum that is allowed by the length field of the IP header, we fall back to zero. An alternative would be to set the field to zero unconditionally, and this seems to be a "should" level expectation for at least for some Intel cards. The fallback approach has the advantage of keeping code that uses the IP header field (for example for constructing TcpPackets to set the partial checksum in our case) functional for sufficiently small unsegmented packets. On the other hand, it can also be argued that causing a failure even for smaller packets can help uncover the error case with the larger packets earlier in the development of the device drivers.

Comparison to smoltcp-rs#830

• The segmentation capabilities struct in the older PR follows the structure of ChecksumCapabilities. This PR does not do so for the following reasons:
  • The segmentation capability for a given protocol is not binary. Different devices may support large (i.e. unsegmented) packets of different maximum sizes (references for various devices)
  • On the receive side the related feature seems to be called "coalescing." It may be somewhat confusing to represent information for the receive side in a structure whose name refers to "segmentation."
• The capability structure is named SegmentationCapabilities rather than the TCP specific name TsoCapabilities. I believe the current structure can be expanded to support UDP, though I currently do not have a prototype for that.
• The current PR has support for communicating the target segment size to the device.

[mkroening]

The code looks good to me, though I am not too familiar with smoltcp's internals.

The PR could use a description discussing the design and differentiating itself from smoltcp-rs#830, though.

Furthermore, CI does not pass at the moment.

I have updated our fork, so if you force-push, the commits should show up more nicely.

[codecov-commenter]

Welcome to Codecov 🎉

Once you merge this PR into your default branch, you're all set! Codecov will compare coverage reports and display results in all future pull requests.

Thanks for integrating Codecov - We've got you covered ☂️