[Flaky Test]: <TestConnInfoConnCloseThenAnotherConn> – failed to start connection credentials listener: listen unix ...: bind: invalid argument

[belimawr]

Failing test case

TestConnInfoConnCloseThenAnotherConn

Error message

conn_info_server_test.go:190: failed to start connection credentials listener: listen unix /var/folders/1w/pb98dgl15sd6jdcx2yy6j2500000gn/T/TestConnInfoConnCloseThenAnotherConn3117723250/001/.teaci.sock: bind: invalid argument

Build

Local run of the tests

OS

Linux, Mac

Stacktrace and notes

This test is failing because the test creates a unix socket on a path that can exceed the maximum size for the OS.

man unix on darwin says:

UNIX-domain addresses are variable-length filesystem pathnames of at most 104 characters.

On the listed example the path is 111 characters long.

The issue can also affect Linux:

$ cat /usr/include/linux/un.h | grep "define UNIX_PATH_MAX"
#define UNIX_PATH_MAX   108

The test usually passes on Linux because the generated path is < 100 characters. E.g: /tmp/TestConnInfoConnCloseThenAnotherConn233963498/001/.teaci.sock

tiago@Not-A-Linux~/devel/elastic-agent/pkg/component/runtime % go test -run=TestConnInfoConnCloseThenAnotherConn -v -count=1
=== RUN   TestConnInfoConnCloseThenAnotherConn
=== RUN   TestConnInfoConnCloseThenAnotherConn/port
=== RUN   TestConnInfoConnCloseThenAnotherConn/local
    conn_info_server_test.go:190: failed to start connection credentials listener: listen unix /var/folders/1w/pb98dgl15sd6jdcx2yy6j2500000gn/T/TestConnInfoConnCloseThenAnotherConn3117723250/001/.teaci.sock: bind: invalid argument
--- FAIL: TestConnInfoConnCloseThenAnotherConn (0.04s)
    --- PASS: TestConnInfoConnCloseThenAnotherConn/port (0.00s)
    --- FAIL: TestConnInfoConnCloseThenAnotherConn/local (0.04s)
FAIL
exit status 1
FAIL    github.com/elastic/elastic-agent/pkg/component/runtime  0.456s
1:WARN tiago@Not-A-Linux~/devel/elastic-agent/pkg/component/runtime %

The problem comes from the getAddress called by runTests function that generates the address for the sockets without doing any length validation.

elastic-agent/pkg/component/runtime/conn_info_server_test.go

Lines 77 to 91 in c1d42ae

	func getAddress(dir string, isLocal bool) string {
	if isLocal {
	u := url.URL{}
	u.Path = "/"
	if runtime.GOOS == "windows" {
	u.Scheme = "npipe"
	return u.JoinPath("/", testSock).String()
	}
	u.Scheme = "unix"
	return u.JoinPath(dir, testSock).String()
	}
	return fmt.Sprintf("127.0.0.1:%d", testPort)
	}

elastic-agent/pkg/component/runtime/conn_info_server_test.go

Lines 93 to 115 in c1d42ae

	func runTests(t *testing.T, fn func(*testing.T, string)) {
	sockdir := t.TempDir()
	tests := []struct {
	name string
	address string
	}{
	{
	name: "port",
	address: getAddress("", false),
	},
	{
	name: "local",
	address: getAddress(sockdir, true),
	},
	}
	for _, tc := range tests {
	t.Run(tc.name, func(t *testing.T) {
	fn(t, tc.address)
	})
	}
	}

[elasticmachine]

Pinging @elastic/elastic-agent (Team:Elastic-Agent)

[belimawr]

We already have some code to handle socket paths that are too long:

• https://github.com/elastic/elastic-agent/blob/main/pkg/utils/socket_unix.go
• https://github.com/elastic/elastic-agent/blob/main/pkg/utils/socket_windows.go

Using it in our tests should fix the problem