Missing regression tests for SOCK_RAW support in standard library

[ncoghlan]

As identified in #114887 , asyncio's support for SocketCAN regressed for a long time when datagram endpoint creation was limited specifically to SOCK_DGRAM, thus incorrectly excluding SOCK_RAW.

The other untested-but-not-deprecated-at-the-OS-level socket types (SOCK_RDM, SOCK_SEQPACKET) aren't part of any standard internet protocol, but SOCK_RAW is actively used for hardware CAN bus access, so it seems reasonable to figure out a way to cover it in the regression test suite.

The part that makes this tricky is that raw sockets are also used to craft malicious attack packets in malware, so modern operating systems restrict access to them by default. Simply attempting to create a couple of SOCK_RAW endpoints in the test suite and run some of the existing SOCK_DGRAM tests over them would routinely fail with OS permission errors.

I don't have a clear solution in mind for that problem, but the ideas I've had so far are:

• define a new test resource for netraw (named after CAP_NET_RAW in Linux), gate SOCK_RAW tests for socket and asyncio behind that
• for asyncio, test with an instance of a socket subclass that is SOCK_DGRAM under the hood (so no elevated permissions needed), but claims SOCK_RAW as its socket type (so it would still flush out incorrect asyncio level assumptions like create_datagram_endpoint() no longer compatible with CAN bus raw sockets  #114887 )
• combine the two, with the latter used as a fallback testing option when the netraw resource isn't available

I don't know a clean cross-platform way to test for raw socket permissions that doesn't risk getting the test suite flagged as potential malware by behaviour monitoring systems, so the test resource seems like the best bet as a way of determining whether or not to even try creating any raw sockets.

Whether to try to test SOCK_RAW on GitHub actions and the buildbot fleet would be separate questions, but they'd need to be running in at least some locations to be useful.

Linked PRs

• gh-115159: Add SOCK_RAW regression tests #154939

[encukou]

netraw seems like the way to go.
Perhaps we'll need a dedicated buildbot worker? Anyway, making it possible to test this seems like a prerequisite to ensuring the tests are run.

[ncoghlan]

Updated the initial post to clarify how a fallback testing option when actual raw sockets aren't available could work (while socket.type is read-only from Python, a test-only socket subclass can override the type property to always return SOCK_RAW regardless of the underlying socket type).

Given that, I'm leaning towards option 3:

• socket layer tests that only run when the new netraw resource is available, skipped otherwise
• asyncio layer tests that always run, using an actual raw socket if netraw is available, and a custom FakeRawSocket subclass instance otherwise

Exactly what those tests cover would then be based on what's easy to refactor the existing tests to execute rather than attempting to replicate the fulll level of test coverage that SOCK_DGRAM gets.