Add support for binding Unix sockets in Linux's abstract namespace.

[Dadeos-Menlo]

Linux offers an "abstract namespace" for Unix sockets, distinguished by the address comprising an initial null byte. Specifying a file argument describing an abstract namespace address to netutil.bind_unix_socket(…) fails as a consequence of the additional, file-system orientated, processing performed:

tornado/tornado/netutil.py

Lines 212 to 222 in bdfc017

	try:
	st = os.stat(file)
	except FileNotFoundError:
	pass
	else:
	if stat.S_ISSOCK(st.st_mode):
	os.remove(file)
	else:
	raise ValueError("File %s exists and is not a socket", file)
	sock.bind(file)
	os.chmod(file, mode)

The proposed changes avoid performing the file-system orientated processing for abstract namespace addresses.

[bdarnell]

Thanks, this looks pretty good, but I think the test could be improved.

[bdarnell]

This should have a comment to say that filenames with a leading NUL byte are used by the linux-specific abstract namespace. (Should we have a check for the platform name so that attempts to use a leading NUL are errors on other platforms? Probably not, IMHO)

[Dadeos-Menlo]

Added comment in be9e153.

[bdarnell]

This is a clever way to get a unique name for the abstract socket in most cases, although it does have some weaknesses since the abstract socket namespace may be shared more widely than the filesystem (e.g. containers with host networking, or just processes with different values of TMPDIR. So I'd rather use a uuid4 here than tie the abstract socket name to the temp directory.

[Dadeos-Menlo]

Switched to generating a unique abstract namespace identifier using uuid4() in be9e153.

[bdarnell]

This test is about introspecting socket addresses, so it feels like something that should be tested for abstract sockets too. Instead of binding both abstract and filesystem sockets in setUp (half of which is going to be wasted work in any given test), and only testing the abstract socket with certain tests that must be duplicated, I think it would be better to use subclasses: factor the socket binding logic out of setUp into an overridable method, and then make an AbstractUnixSocketTest that subclasses this one.

[Dadeos-Menlo]

Refactored unit-tests as suggested in be9e153.

[Dadeos-Menlo]

Sorry, I'm not getting on very well with the pull-request checks…

Any hints as to how you'd like to make mypy happy about the refactored test case. I didn't want to make the UnixSocketTest base class an AsyncTestCase because I don't want the test functions to be executed outside of one of the, UnixSocketTestAbstract or UnixSocketTestFile concrete implementations; but a failure to do so makes mypy sad.

[bdarnell]

Hmm, I thought I had a clever solution for this but it looks like the other places where I've done this are just hacks with a lot of typing.Any and type:ignore annotations. (Other discussions of the issue at abseil/abseil-py#166 and python/cpython#61721)

In this case I think I'd make UnixSocketTest a subclass of AsyncTestCase to make type checking happy, and then skip the base class at runtime: In UnixSocketTest.setUp(), add if type(self) == UnixSocketTest: raise unittest.SkipTest("abstract base class"). That might have the fortunate side effect of making pytest happier with this pattern too (#3390)

[Dadeos-Menlo]

Implemented as suggested in 6e14f25.

[bdarnell]

Looks good, thanks for your patience with all the CI headaches (next time it'll be a little less painful because I won't have to come click "approve CI" for every new commit)

[Dadeos-Menlo]

Thank you!

I did wonder whether I was missing something in being able to trigger the CI myself…

I generally use single-quoted strings; and don't bother to run mypy on my unit tests (life's too short).

Thanks for all of your work on Tornado; we use it extensively.