Multiple failures in test_vsock

[asomers]

When we first added the vsock code in #1091 the test contained an assertion that connecting to the socket would fail, because "The current implementation does not support loopback devices". But it did not check the particular failure mode. A recent change (#1915) has caused connect to start succeeding, in at least some cases. It seems that at least some of the original failures were due to a race condition in the test: it closed a socket too early. Now that that is fixed, we see three behaviors:

• Fails with EINVAL on Linux x86_64, presumably due to the "does not support loopback devices" part.
• Passes on most qemu platforms. Perhaps it's connecting to the qemu host?
• Fails with ENODEV on mips and mips64

See #1917 to view the behavior on each platform.

@stefano-garzarella could you please clarify what we should expect from connect, and fix the test? For now I'm going to disable it so it doesn't hold up other PRs.

[stefano-garzarella]

@asomers yep, I'll fix the test, making it more reliable.

About connect, it can succeed if it runs in the host that has also the new vsock_loopback module available.
This module will be automatically loaded if there aren't other vsock transports loaded (e.g. in a host or a guest without any vsock devices).

This will make the test to fail, since now it is able to handle loopback.
The different errno depends on the availability of the device.

I think we should avoid trying the connection, since it can change depending on where the test is running.
So, we should always test a failing scenario, maybe using another port. I don't think our goal is to test that vsock is working, but just the interface.

Something like this:

diff --git a/test/sys/test_socket.rs b/test/sys/test_socket.rs
index 2ab6c54..d3fbbaa 100644
--- a/test/sys/test_socket.rs
+++ b/test/sys/test_socket.rs
@@ -2080,10 +2080,9 @@ pub fn test_vsock() {
         )
         .expect("socket failed");
 
-        let sockaddr_host = VsockAddr::new(cid, port);
+        let sockaddr_host = VsockAddr::new(cid, port + 1);
 
-        // The current implementation does not support loopback devices, so,
-        // for now, we expect a failure on the connect.
+        // We expect a failure on the connect since we are using a wrong port.
         assert_ne!(connect(s2, &sockaddr_host), Ok(()));
 
         close(s2).unwrap();

I'll send a PR ASAP.

[stefano-garzarella]

I'll send a PR ASAP.

@asomers or do you prefer to include these changed in #1915 ?

[asomers]

I don't think our goal is to test that vsock is working, but just the interface.

Yes, definitely. In fact, we don't even need to try connect. We just need to be sure that the address type works.

@asomers or do you prefer to include these changed in #1915 ?

No, let's do it standalone.

[stefano-garzarella]

I don't think our goal is to test that vsock is working, but just the interface.

Yes, definitely. In fact, we don't even need to try connect. We just need to be sure that the address type works.

I agree, I'll remove that part.

@asomers or do you prefer to include these changed in #1915 ?

No, let's do it standalone.

Okay, I'll send it!