service name not available for the specified socket type

[mtelka]

Things to check first

• I have searched the existing issues and didn't find my bug already reported there

• I have checked that my bug is still present in the latest release

AnyIO version

4.3.0

Python version

3.9.18

What happened?

Several anyio tests fails with this error:

_________________ TestTLSListener.test_handshake_fail[asyncio] _________________
tests/streams/test_tls.py:417: in test_handshake_fail
    listener = await create_tcp_listener(local_host="127.0.0.1")
../prototype/i386/usr/lib/python3.9/vendor-packages/anyio/_core/_sockets.py:290: in create_tcp_listener
    gai_res = await getaddrinfo(
../prototype/i386/usr/lib/python3.9/vendor-packages/anyio/_core/_sockets.py:573: in getaddrinfo
    gai_res = await get_async_backend().getaddrinfo(
../prototype/i386/usr/lib/python3.9/vendor-packages/anyio/_backends/_asyncio.py:2370: in getaddrinfo
    return await get_running_loop().getaddrinfo(
/usr/lib/python3.9/asyncio/base_events.py:861: in getaddrinfo
    return await self.run_in_executor(
/usr/lib/python3.9/concurrent/futures/thread.py:58: in run
    result = self.fn(*self.args, **self.kwargs)
/usr/lib/python3.9/asyncio/base_events.py:844: in _getaddrinfo_debug
    addrinfo = socket.getaddrinfo(host, port, family, type, proto, flags)
/usr/lib/python3.9/socket.py:954: in getaddrinfo
    for res in _socket.getaddrinfo(host, port, family, type, proto, flags):
E   socket.gaierror: [Errno 9] service name not available for the specified socket type

Here is the complete list of tests with this failure:

TestTLSListener.test_handshake_fail[asyncio]
TestTLSListener.test_extra_attributes[asyncio]
TestTCPListener.test_extra_attributes[asyncio-ipv4]
TestTCPListener.test_extra_attributes[asyncio-ipv6]
TestTCPListener.test_accept[asyncio-ipv4]
TestTCPListener.test_accept[asyncio-ipv6]
TestTCPListener.test_accept[asyncio-both]
TestTCPListener.test_accept_after_close[asyncio-ipv4]
TestTCPListener.test_accept_after_close[asyncio-ipv6]
TestTCPListener.test_socket_options[asyncio-ipv4]
TestTCPListener.test_socket_options[asyncio-ipv6]
TestTCPListener.test_close_from_other_task[asyncio-ipv4]
TestTCPListener.test_close_from_other_task[asyncio-ipv6]
TestTCPListener.test_send_after_eof[asyncio-ipv4]
TestTCPListener.test_send_after_eof[asyncio-ipv6]
TestTCPListener.test_bind_link_local[asyncio]
test_multi_listener[asyncio]
test_getaddrinfo[asyncio]

The problem seems to be that getaddrinfo is called with port=0. The similar issue was already fixed in eventlet (please read the link for more details about the issue).

How can we reproduce the bug?

Run anyio tests on OpenIndiana.

[agronholm]

I'm just wondering why this fails only on this specific distro and not any other.

[mtelka]

I'm just wondering why this fails only on this specific distro and not any other.

Because other distros are usually Linux?

[agronholm]

I'm just wondering why this fails only on this specific distro and not any other.

Because other distros are usually Linux?

Ahh...I didn't realize.

[mtelka]

BTW, I tried this:

--- anyio-4.3.0/src/anyio/_core/_sockets.py.orig
+++ anyio-4.3.0/src/anyio/_core/_sockets.py
@@ -264,7 +264,7 @@
 async def create_tcp_listener(
     *,
     local_host: IPAddressType | None = None,
-    local_port: int = 0,
+    local_port: int | None = None,
     family: AnyIPAddressFamily = socket.AddressFamily.AF_UNSPEC,
     backlog: int = 65536,
     reuse_port: bool = False,

This allowed the create_tcp_listener() to succeed, but tests/streams/test_tls.py still failed later with:

_________________ TestTLSListener.test_handshake_fail[asyncio] _________________
tests/streams/test_tls.py:422: in test_handshake_fail
    sock.connect(listener.extra(SocketAttribute.local_address))
../prototype/i386/usr/lib/python3.9/vendor-packages/anyio/_core/_typedattr.py:79: in extra
    raise TypedAttributeLookupError("Attribute not found") from None
E   anyio.TypedAttributeLookupError: Attribute not found

So this apparently needs more work :-(.

[agronholm]

I think it'd be best if you submitted a PR. I don't think I can reasonably verify this, as I'm not familiar with this OS.

[mtelka]

Is what I propose above for _socket.py the right way to go? If so, then you can start with that and IMHO you'll see the same failure at tests/streams/test_tls.py:422 as I do. I believe this is not platform specific.

[agronholm]

If the problem is just with getaddrinfo(), then how about a simpler change?

    gai_res = await getaddrinfo(
        local_host,
        local_port or None,
        family=family,
        type=socket.SocketKind.SOCK_STREAM if sys.platform == "win32" else 0,
        flags=socket.AI_PASSIVE | socket.AI_ADDRCONFIG,
    )

This at least passes my local tests.

[mtelka]

I tried:

--- anyio-4.3.0/src/anyio/_core/_sockets.py.orig
+++ anyio-4.3.0/src/anyio/_core/_sockets.py
@@ -289,7 +289,7 @@
     local_host = str(local_host) if local_host is not None else None
     gai_res = await getaddrinfo(
         local_host,
-        local_port,
+        local_port or None,
         family=family,
         type=socket.SocketKind.SOCK_STREAM if sys.platform == "win32" else 0,
         flags=socket.AI_PASSIVE | socket.AI_ADDRCONFIG,

And I see the same failure here:

_________________ TestTLSListener.test_handshake_fail[asyncio] _________________
tests/streams/test_tls.py:422: in test_handshake_fail
    sock.connect(listener.extra(SocketAttribute.local_address))
../prototype/i386/usr/lib/python3.9/vendor-packages/anyio/_core/_typedattr.py:79: in extra
    raise TypedAttributeLookupError("Attribute not found") from None
E   anyio.TypedAttributeLookupError: Attribute not found

So it looks like this failure is not platform agnostic as I hoped :-(.

[agronholm]

It's just one failure, or are there more?

[mtelka]

This is the complete list of failed test with anyio.TypedAttributeLookupError: Attribute not found error I see with the above change:

TestTLSListener.test_handshake_fail[asyncio]
TestTLSListener.test_extra_attributes[asyncio]
TestTCPListener.test_extra_attributes[asyncio-ipv4]
TestTCPListener.test_extra_attributes[asyncio-ipv6]
TestTCPListener.test_send_after_eof[asyncio-ipv4]
TestTCPListener.test_send_after_eof[asyncio-ipv6]

The test_getaddrinfo[asyncio] test is still failing with service name not available for the specified socket type (apparently there are more places where 0 should be replaced by None), two are failing with different error now, namely these:

___________ TestTCPListener.test_close_from_other_task[asyncio-ipv4] ___________
tests/test_sockets.py:650: in test_close_from_other_task
    tg.cancel_scope.cancel()
E   Failed: DID NOT RAISE <class 'exceptiongroup.ExceptionGroup'>
___________ TestTCPListener.test_close_from_other_task[asyncio-ipv6] ___________
tests/test_sockets.py:650: in test_close_from_other_task
    tg.cancel_scope.cancel()
E   Failed: DID NOT RAISE <class 'exceptiongroup.ExceptionGroup'>

and all other tests from the initial list are passing now.

[mtelka]

... and these failures are not related to port=0 vs. port=None. Here is another example:

_____________ TestTCPListener.test_extra_attributes[asyncio-ipv4] ______________
tests/test_sockets.py:521: in test_extra_attributes
    assert multi.extra(SocketAttribute.family) == family
../prototype/i386/usr/lib/python3.9/vendor-packages/anyio/_core/_typedattr.py:79: in extra
    raise TypedAttributeLookupError("Attribute not found") from None
E   anyio.TypedAttributeLookupError: Attribute not found

[mtelka]

... and these failures are not related to port=0 vs. port=None. Here is another example:

In other words, your proposed change (or mine, they are effectively very similar) is clearly improving things. But it looks like the similar issue is in create_udp_socket(), and maybe at some other places too.

[agronholm]

I don't think I can effectively work on these if I can't reproduce them.

[mtelka]

I don't think I can effectively work on these if I can't reproduce them.

Okay, np. Will you accept a PR if I create it based on my local_port: int | None = None proposal?

[agronholm]

I'd much prefer a PR that fixes all of the problems you're seeing. I can help you figure out the rest.

[mtelka]

Hmmm, I see those as different issues, so I'd prefer to handle them separately...

Anyway, what do you suggest for the assert multi.extra(SocketAttribute.family) == family issue?

[agronholm]

Hmmm, I see those as different issues, so I'd prefer to handle them separately...

From my PoV, they all fall under the umbrella of "fails on Illumos, works everywhere else".

Anyway, what do you suggest for the assert multi.extra(SocketAttribute.family) == family issue?

That's something you need to dive into – check why the attribute isn't available. The _SocketProvider class might be a good place to start.

[agronholm]

I think that what's happening here is that the attribute is actually found, but its getter raises KeyError, thus leading into this error. This in itself is a subtle bug.

[agronholm]

That said, the accessor doesn't do any dict lookups so I don't understand where the KeyError is coming from. Setting a breakpoint after line 77 in _typedattr.py should help figure it out.

[agronholm]

This should be easier to debug on latest AnyIO as the aforementioned bug with extra() has been fixed. Also, let me know if you still want me to pursure a fix for this.