Failing node 24 on macos

[Uzlopak]

@metcoder95

Can it be that your last PR regarding http2 broke on macos and node24?

[metcoder95]

Hard to say, seems related with set-cookies but possible some connect timeout; have we tried to isolate it?

[Uzlopak]

The thing is, that before it was only node20. but now node 24 ci times out, meaning it does not close properly.

Seems related to

test at test/http2.js:567:1
✖ Dispatcher#Connect (393.716833ms)
  Error [ERR_HTTP2_STREAM_ERROR]: Stream closed with error code NGHTTP2_INTERNAL_ERROR
      at ClientHttp2Stream._destroy (node:internal/http2/core:2472:13)
      at _destroy (node:internal/streams/destroy:122:10)
      at ClientHttp2Stream.destroy (node:internal/streams/destroy:84:5)
      at Writable.destroy (node:internal/streams/writable:1122:11)
      at Http2Stream.onStreamClose (node:internal/http2/core:597:12) {
    code: 'ERR_HTTP2_STREAM_ERROR'
  }

test at test/http2.js:1:1
✖ test/http2.js (30002.941292ms)

Probably a http2.js not closing a server or so, so the test hangs.

[metcoder95]

I'll be trying to debug it but so far seems like a side effect coming from another source; cannot reproduce it in local with v24 and v22 and it seems like a re-run fix it.

Let's keep the issue open until sorting it out

[Uzlopak]

Yes, it is very hard to reproduce. I am investigating core and looking for some kind of potential race condition. No luck till now.

Do you have a mac?

Sometimes you can reproduce such a bug by calling the script in bash script over and over. Maybe even setting the performance to low, maybe radically run prime95, hahaha.

Also another thing: github runners labeled with macos-latest are using arm64 or whatever the m1/2/3/4 chips are. Github runners with macos-13 label are using intel cpus. But I guess we dont compile nodejs for amd64.

Well, idk...

[metcoder95]

Yeah, I've a Mac, I've been trying to do it for a while but if it happens until the 1000th time, it might be an inconsistency on the OS then; I will still keep an eye on it and see if I can find a way to reproduce it consistently.

But I guess we dont compile nodejs for amd64.

Do you mean in the CI or Node binaries itself?

[Uzlopak]

i mean compiling node 20 for macos on intel.

Anyhow:

I merged #4561, so now we can a little bit better investigate

Try to run test/http2-dispatcher.js with bun with bun test ./test/http2-dispatcher.js

I think this is the same issue we have on mac.

I already created #4564, to pass the bun issue.

So basically goAway is already setting the socket to null, so close can not process it properly.

But I lack knowledge regarding http2, so it feels like looking for the needle in the haystack.

I assume that if we get the problem solved for bun, we will also solve it for macos.

[metcoder95]

Let me take a look at it, thanks for that, it will definitely make things easier!

[Uzlopak]

@metcoder95

Sry, i was inprecise in my words. One of the issues in bun is solved with not setting socket to null in goaway but there is still an issue in bun.