test: fix localhost error in test-dns-ipv6

[gibfahn]

Checklist

• make -j4 test (UNIX), or vcbuild test nosign (Windows) passes
• tests and/or benchmarks are included
• commit message follows commit guidelines

Affected core subsystem(s)

test

Description of change

If localhost doesn't resolve to ::1 in IPv6, we should skip the test.

This is the same fix as 814b8c3 (#7766)

cc: @Trott

[jasnell]

This should assert instead of throw

[Trott]

This should assert instead of throw

Why is that? (Not trolling; honest question. If there's an err object, then that's basically an exception, not a test failure per se. Might be system DNS misconfiguration, DNS server unavailability, network problems...)

[cjihrig]

I think this should be assert.ifError().

[gibfahn]

I'm happy to change it, but if assert.ifError() is just this:

assert.ifError = function(err) { if (err) throw err; };

what is the reason for having the extra layer of indirection?

[Trott]

This looks incorrect to me. It will skip these tests if localhost does not resolve to ::1 but none of these tests depend on that and they should run regardless. The exception is the test that's commented out and that test I suppose should be skipped if this lookup fails, perhaps. But that test is already being skipped everywhere. :-(

Is there an issue this is trying to fix? Can you link to that?

[gibfahn]

@Trott Yep you're right! I got confused. The problem we were having is that some machines didn't have the ::1 localhost line at all, so the test was failing on L147 with:

# /home/gib/code/node/labels/bvt/node/test/internet/test-dns-ipv6.js:182
#     if (err) throw err;
#              ˆ
#
# Error: getnameinfo ENOTFOUND ::1
#     at errnoException (dns.js:28:10)
#     at GetNameInfoReqWrap.onlookupservice [as oncomplete] (dns.js:180:26)

common.hasIPv6 returns true on the machine.

Obviously you could say that this is a machine issue, but I'm pretty sure this is the default for this machine, so it'd be nice to handle the failure in this test.

Also I agree it's not ideal to skip the entire test if one of the subtests has an issue, but I'm not aware of any mechanism for skipping subtests (in fact I don't think we've implemented subtests in node, although I know npm has them).

[Trott]

Would the solution be to change line 182 from this:

    if (err) throw err;

To this?:

    if (err) {
      if (err.code === 'ENOTFOUND') {
        return done();
      }
      throw err;
    }

[gibfahn]

That does indeed stop the test from failing, but it also basically hides the fact that the test didn't actually pass due to ::1 not being resolved. Presumably this partially defeats the purpose of the having this particular test?

I understand there may not be any better solution.

[Trott]

I think the problem is that the test is faulty. The test is making an assumption about a system that isn't actually true on many default Linux variants as we've seen in other tests and now we're seeing in this one with AIX.

The test amounts to "If you look up the IPv6 resolution for localhost, then you must get ::1." But as we've seen, there are multiple systems where IPv6 can be configured but localhost is not set to ::1 in /etc/hosts or whatever. As much as possible, we want to isolate out system configuration quirks from the tests. That's not always possible, but as much as we can, we should do it.

To that end, I'd propose that although still faulty, this is an improvement: "If you look up the IPv6 resolution for localhost, then you must get either ::1 or ENOTFOUND."

To that end, I'd change the code I wrote above to this:

  if (err) {
    assert.strictEqual(err.code, 'ENOTFOUND');
    return done();
  }

That makes it a little more clear that we're not skipping the test so much as checking an alternative acceptable result.

[Trott]

By the way, you may be wondering why you are seeing this on AIX while it never caused problems with the Linux variants that also do not have the localhost/::1 mapping set up by default.

The answer is because the test/internet tests are not run in CI and therefore are mostly not run at all! So thanks for running them and thanks for chasing down this issue!

[gibfahn]

@Trott That makes a lot of sense. I'll make that change and add a comment to note down why ENOTFOUND is an acceptable answer. Thanks for the help!

[gibfahn]

@Trott Updated, PTAL

[jasnell]

It may be worthwhile to expand the code comment in the test to explain a bit more. Specifically, calling out something like:

// We're not skipping the test so much as checking an alternative acceptable result.

Or some such...

Otherwise, LGTM

[gibfahn]

Sounds good to me, updated.

[jasnell]

LGTM

[Trott]

LGTM

CI: https://ci.nodejs.org/job/node-test-pull-request/3914/

[gibfahn]

Looks like CI passed except for a Git checkout issue on linuxone. As the internet tests aren't actually run, I guess the only thing the CI really tells us is that the linter passed.

[Trott]

Landed in 3504a98