Add new socket type for raw ethernet protocols

[ususdei]

Hi everyone!

I'm currently working on a rust library for a protocol which runs on plain ethernet using a custom ethertype (TECMP on 0x99FE).
It also uses an additional IP-based protocol for discovery and configuration.

I would also like to use this library on some STM32, so I tried to port it from native linux sockets to the smoltcp stack.

Thus I started to implement some basic support for raw ethernet sockets.
This is still a work in progress but the basic functionality already works on an STM32F7.
It is still missing support for Metadata and the send API could be more convenient
but I would like to get this out there and gather some early feedback as to whether you would even consider merging something like this.

So just some initial questions:

• Is this feature even in scope for the smoltcp project?
• Do you agree with my overall design?
• Do you already have any other initial feedback?

Thank you all for your feedback and your great contribution to the embedded rust ecosystem!

[codecov]

Codecov Report

❌ Patch coverage is 82.12435% with 69 lines in your changes missing coverage. Please review.
✅ Project coverage is 80.66%. Comparing base (39cd44e) to head (4114e16).

Files with missing lines	Patch %	Lines
src/socket/eth.rs	81.54%	62 Missing ⚠️
src/iface/interface/mod.rs	85.71%	6 Missing ⚠️
src/socket/mod.rs	0.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1013      +/-   ##
==========================================
+ Coverage   80.45%   80.66%   +0.20%     
==========================================
  Files          81       82       +1     
  Lines       24461    24829     +368     
==========================================
+ Hits        19681    20028     +347     
- Misses       4780     4801      +21     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[nikvoid]

I missed this PR and made similar thing, so... +1 for raw ethernet sockets

[i404788]

We nearly did the same as @nikvoid (hey btw), +1. Also should link #942 to help the next person

[not-Ryan]

Something you might want to add the socket-eth feature to is the compile error on lib.rs:122

Right now I'm getting the compile error:

compile_error!("If you enable the socket feature, you must enable at least one of the following features: socket-raw, socket-udp, socket-tcp, socket-icmp, socket-dhcpv4, socket-dns");

[ususdei]

It's cool to hear that others would like to use this lib in similar ways as we do.

But I'm not sure if it's really worth investing more time into polishing this PR.
There hasn't been any feedback to this in almost a year now.

[thvdveld]

@ususdei sorry it took so long before the first reply. I think this PR looks good. Could you, as @not-Ryan suggested, edit the compile error as well? I don't have any other remarks.

@Dirbaio @whitequark any remarks? I think this feature was requested already in the issues (#942, #902 could be implemented as a Eth socket).

[whitequark]

@ususdei Apologies for the lack of feedback. I don't see any reason not to merge this; it fits the overall theme of the library and doesn't add any overhead when disabled.

The implementation is basically copied from raw sockets, right? That's also an easy thing to approve.

[not-Ryan]

Since your force-push of 833d319 it seems to be broken at:

ususdei@833d319#diff-57320a1d1f958a93a7d782cab2e9e6409dd32f50a98a86e854e88b28559de4b0R62-R65

This is only when you disable the default features:

error[E0308]: mismatched types
  --> .../smoltcp-426d9b36010b4974/833d319/src/socket/eth.rs:62:56
   |
62 |     fn fmt(&self, f: &mut core::fmt::Formatter<'_>) -> core::fmt::Result {
   |        ---                                             ^^^^^^^^^^^^^^^^^ expected `Result<(), Error>`, found `()`
   |        |
   |        implicitly returns `()` as its body has no tail or `return` expression
   |
   = note:   expected enum `core::result::Result<(), core::fmt::Error>`
           found unit type `()`

[ususdei]

Thanks @not-Ryan !
I found some more feature combinations which don't work yet.
I am working on them while also improving the test coverage.

[ususdei]

I just finished some remaining TODOs from my list and reran the code on my test-device.

So from my side, this MR is now ready for a thorough review.
Let me know if you are unhappy with anything.

[ususdei]

The implementation is basically copied from raw sockets, right? That's also an easy thing to approve.

Exactly. I didn't see a way to extend raw sockets to also handle plain ethernet frames without breaking the
API, type safety and performance. So eth-sockets are now almost a copy of raw sockets, just on a different layer.

And then copied over metadata support from UDP sockets (which is still missing from raw sockets but I need the metadata-id for accurate timestamps).

[whitequark]

Can you explain the changes to ARP logic you've made?

[ususdei]

Can you explain the changes to ARP logic you've made?

None, I hope...

Since the ethernet packet handling cannot generate any automatic reply, we can still process ARP packages using process_arp() as we did before (same for all IP packets).
For any other ethertype, we can ignore ARP completely.

Or did I miss something here?