test: improve test-child-process-fork-dgram

[mhdawson]

Checklist

• make -j4 test (UNIX), or vcbuild test nosign (Windows) passes
• commit message follows commit guidelines

Affected core subsystem(s)

test

Description of change

Previous implementation was subject to timing out as there
was no guarantee that the parent or child would ever
receive a message. Modified test so that once parent or
child receives a message it stops receiving them so that
we are guaranteed the next message will be received by the
other end.

The test also left lingering processes when it timed out
so pulled the timeout into the test itself so that it
could properly cleanup in case of failure.

[imyller]

LGTM

[gibfahn]

Commit message nit:

Modified test so that once parent or child receives a message it stops receiving them so that we are guaranteed the next message will be received by the other end.

Maybe: Modified test so that once parent or child receives a message it closes its connection to the socket, so we can guarantee the next message will be received by the other node.

[gibfahn]

s/guarranteed/guaranteed/

[mhdawson]

Ok will update commit message.

[gibfahn]

@mhdawson The typo is in the test comment on line 12.

[gibfahn]

Is there a reason to add the serverPort variable? It only seems to be used here, and makes things a bit more complicated.

[mhdawson]

Its needed because the server is closed before we have stopped sending. Therefore, you can't call server.address().port. I found this the hard way while building up the changeset.

[gibfahn]

So if we're now closing the child/parent server once they receive a message, the test should finish after two messages unless some are disappearing (this is UDP so I guess messages can fail). I haven't seen any dropped messages at all in any of the test runs, so having a 45s internal timeout (i.e. 45,000 messages) may be overkill?

Changes LGTM if CI passes (probably also worth running a stress test on AIX as well to make sure).

[gibfahn]

Also just to document, this fixes #8271

[mhdawson]

Already ran the stress run of 500 on AIX and all passed versus about at 1/10 failure before.

CI run here: https://ci.nodejs.org/job/node-test-pull-request/4223/

[mhdawson]

In terms of the timeout as long as it is shorter than the 60s from test harness I don't think it really matters as we will normally finish fast anyway.

[mhdawson]

Commit message updated.

[mhdawson]

Lots of failures in the CI run, but none related to the test being changed. Several test tick processor related timeouts as well as checkouts failing across a number of platforms.
I'll try again later when things seem quieter.

[mhdawson]

Another CI run: https://ci.nodejs.org/job/node-test-pull-request/4225/

[mhdawson]

Ok second CI run only had failures (windows and smartos) related to test-tick-processor which were pre-existing and not related to the changes to this test.

[gibfahn]

@mhdawson You missed a nit on L20, otherwise LGTM.

@misterdjules LGTY? (also @AndreasMadsen , but feel free to ignore if you don't think you could usefully review).

[AndreasMadsen]

@gibfahn Could you explain why this comment isn't a concern here:

Not too sure. If the intention of the test was that messages reach both the parent and the child while both are listening, then probably the fix is not correct. If not, probably it is.
by @santigimeno #8549 (comment)

[gibfahn]

@AndreasMadsen I think the point is that due to the unreliable nature of the protocol, we have no grounds to expect that messages will reach both while both are listening. The system may route all messages to either child or parent.

[mhdawson]

@gibfahn thats for the catch, think I got the nit this time :)

[gibfahn]

LGTM

[imyller]

New CI: https://ci.nodejs.org/job/node-test-pull-request/4247/

[mhdawson]

There are 2 hung processes on test-osuosl-aix61-ppc64_be-1, one from Oct 3 and one from Oct 7 so I think the problem still exists unless it was fixed by something after Oct 7.

[mhdawson]

@gibfahn I think the two get separate copies of the handle so one closing may not affect the other. It should however be possible to write a test that validates it one way or the other by making sure it either passes/fails after one side has done the close.

[mhdawson]

Lastest process on test-osuosl-aix61-ppc64_be-1 is from Oct 8th.

[Trott]

Annnnnddddd....now it's failing consistently on freebsd in CI.

So, this is probably a thing again.

I guess someone should write some code to confirm the speculation in #8697 (comment) that the two copies get separate handles and then, if so, this can get a rebase and land? Or am I mistaken? /cc @mhdawson @gibfahn

[gibfahn]

@Trott Well we need to work out why it's failing on macOS 10.12 first I think

[Trott]

Ooof. OK, for fun, or at least "fun", CI stress test across all available platforms: https://ci.nodejs.org/job/node-stress-single-test/999/

[mhdawson]

@gibfahn and I did discuss writing the test to confirm close behavior across processes. I think he'll be looking at that and the failures on macOS 10.12 soon.

[santigimeno]

I think the OS X could be similar to the one described in #7512: if for some reason, the server on the parent side is closed before the handle has been received by the child process, the server remains closed in the child, thus never receiving any message. This patch, fixes the issue for me:

diff --git a/test/parallel/test-child-process-fork-dgram.js b/test/parallel/test-child-process-fork-dgram.js
index 57d8fb0..bd72631 100644
--- a/test/parallel/test-child-process-fork-dgram.js
+++ b/test/parallel/test-child-process-fork-dgram.js
@@ -39,6 +39,7 @@ if (process.argv[2] === 'child') {
         server.close();
       });

+      process.send('handleReceived');
     } else if (msg === 'stop') {
       process.removeListener('message', removeMe);
       if (!serverClosed) {
@@ -69,13 +70,13 @@ if (process.argv[2] === 'child') {
     serverPort = server.address().port;
     child.send('server', server);

-    child.once('message', function(msg) {
+    child.on('message', function(msg) {
       if (msg === 'gotMessage') {
         childGotMessage = true;
+      } else if (msg === 'handleReceived') {
+        sendMessages();
       }
     });
-
-    sendMessages();
   });

   var iterations = 0;

[gibfahn]

@santigimeno This might well be the reason that the test was failing on AIX as well. If it fixes the test for all platforms, then I think it is preferable over this PR.

[Trott]

@santigimeno @gibfahn That unfortunately does not fix it on FreeBSD on our CI, but it is a good thing to add to the test anyway, IMO. And if it clears it up on AIX and/or macos, awesome.

[Trott]

Perhaps this is a libuv issue ultimately? Like, ignorant guess here, but maybe this bit of code explains why the test is reliable on Linux, but not so on other Unices? Or something else along those lines?

EDIT: Eh, or maybe not...

[Trott]

Alternative fix: #9098

Haven't tested it yet on AIX. About to run CI. But if someone is doing these tests locally and whatnot, maybe give it a shot?

That alternative definitely works on macos and works on FreeBSD.

[mhdawson]

Closing since alternative has landed