win, cluster: no callback called after write

[bzoz]

• Version: 9.6.0
• Platform: Windows
• Subsystem: cluster

On Windows, after calling process.disconnect callbacks for previous process.send calls are not called. Example:

'use strict';
const cluster = require('cluster');

if (cluster.isMaster) {
  cluster.fork().on('message', (msg) => {
    console.log(msg.msg);
  });
} else {
  process.send({msg: 'hello'}, () => {
    console.log('cb!');
  });
  process.disconnect();
}

On Windows this will print:

hello

whereas on Linux it will print

hello
cb!

This was reported in https://github.com/libuv/libuv/issues/1729, but it looks like some issue with cluster module. Callbacks are called on both platforms in the following code, which uses bare IPC pipe:

const cp = require('child_process');

if (process.argv[2] === 'child') {
  process.send({msg: 'hello'}, () => {
    console.log('cb!');
  });
  process.disconnect();
} else {
  const child = cp.spawn(process.argv0, [__filename, 'child'], {
    stdio: ['inherit', 'inherit', 'inherit', 'ipc']
  });
  child.on('message', (msg) => {
    console.log(msg.msg);
  })
}

On Windows this will print:

hello
cb!

and on Linux:

cb!
hello

/cc @nodejs/platform-windows @nodejs/cluster

[oyyd]

Try to investigate why this happens these days:

• When calling process.disconnect(), a child process will call process.exit() to exit automatically.

https://github.com/nodejs/node/blob/af7164ebccd21d9fc5b0782e0427257f7637a4db/lib/internal/cluster/child.js#L33-L37

• According to the code here, IO operations like process.send() will firstly call uv_try_write and expect it's done immediately. If it's done asynchronously, the process will exit before callbacks are called.

https://github.com/nodejs/node/blob/af7164ebccd21d9fc5b0782e0427257f7637a4db/src/stream_wrap.cc#L303

• However, the uv_try_write for pipes is not implemented on Windows and it will always return EAGAIN. So that on Windows the callback won't be call if it exits immediately. (libuv code here)

I also found that @bnoordhuis had a PR before to implement uv_try_write for pipes but it didn't land finally.

I think implementing uv_try_write for pipes on Windows is the most elegant way to solve this issue if possible. Otherwise, we may need to prevent child process exit by calling procecss.exit() directly.

net.js and http2.js might have the same problem on Windows as they also call uv_try_write.