Vscode: Support auto attach for node.js subprocesses (aka cluster support)

[weinand]

While working on the Process viewer, I noticed that we could easily support node.js "Clusters" (or any other spawned processes) with a similar approach.

Here is a sketch:

• introduce a new launch config attribute "autoAttachChildren"
• if a launch config has "autoAttachChildren" set to true, node-debug starts to track creation of subprocesses of the debuggee.
• if a subprocess is recognised as being in debug mode (e.g. started with one of more of the flags --inspect, --inspect-brk, --inspect-port, --debug, --debug-brk, --debug-port), node-debug starts an "attach" debug session for the given port.

The result looks like this (and a first cut will be available in the next Insiders).

I've verified the feature (for macOS only) with the example from the node.js documentation for 'cluster':

const cluster = require('cluster');
const http = require('http');
const numCPUs = require('os').cpus().length;

if (cluster.isMaster) {
    console.log(`Master ${process.pid} is running`);

    // Fork workers.
    for (let i = 0; i < numCPUs; i++) {
        cluster.fork();
    }

    cluster.on('exit', (worker, code, signal) => {
        console.log(`worker ${worker.process.pid} died`);
    });
} else {
    // Workers can share any TCP connection
    // In this case it is an HTTP server
    http.createServer((req, res) => {
        res.writeHead(200);
        res.end('hello world\n');
    }).listen(8000);

    console.log(`Worker ${process.pid} started`);
}

Here is the launch config:

{
    "type": "node",
    "request": "launch",
    "name": "Cluster",
    "program": "${workspaceFolder}/test.js",
    "autoAttachChildren": true
},

This already works fine for node.js < v8 (legacy).
For node.js >= v8 the problem is that all workers stop on entry, which requires some manual "continuing". The problem is that node-debug2 does not automatically send a "continue" request because it guesses that a "stop on entry" is desired. Node-debug1 had the same problem but I made it "guess" better in this case...

The work around for node.js >= v8 uses an explicit "--inspect" instead of "--inspect-brk":

{
    "type": "node",
    "request": "launch",
    "name": "Cluster (--inspect)",
    "runtimeArgs": [
        "--inspect=50000"
    ],
    "port": 50000,
    "program": "${workspaceFolder}/test.js",
    "autoAttachChildren": true
},

The fix for node-debug was to only set this._stopOnEntry if it hasn't been set in the launch config:
see https://github.com/Microsoft/vscode-node-debug/blob/9a9ccb1703741fbad10e29d1769c0b8fc5cd3228/src/node/nodeDebug.ts#L1517

/cc @roblourens @auchenberg @isidorn

[roblourens]

This is cool. Here's an issue from someone debugging with the cluster module: https://github.com/Microsoft/vscode/issues/39033

How does cluster.fork select a debug port? Does it automatically pick incrementing port numbers?

[weinand]

@roblourens yes, cluster increments port numbers like this (from the Process Viewer):

"legacy" as expected:

"inspect" by using an interesting "inspect-port" flag:

If a port is specified, "inspect-port" overrides the "inspect" port:

[weeco]

Finally! Looks like a nice feature. Does it have any downsides or would this feature be enabled by default?

[weinand]

"autoAttachChildren" is an idiotic attribute name.
Any suggestions for a better name?

[roblourens]

I don't think it's that bad. That's exactly what it does. I can only think of some variations on it like attachToChildProcs. You probably don't want to tie it to cluster because it would work for any type of spawned process.

[roblourens]

The generated config maybe should copy outFiles and other props from the original config.

[weeco]

I agree with rob, for me the name of this setting tells what it is supposed to do. I prefer childProcess over children though.

[weinand]

Then let's go with "autoAttachChildProcesses" (yes, it's long but we have IntelliSense...)

[jez9999]

This is great - it should be built into VS Code. Anyone who wants to debug a parent and its child processes in Node wants this.

[weinand]

A first cut is already in VS Code Insiders.

[weinand]

@roblourens I've verified (by patching Insiders) that your fix would work fine.
But I couldn't build/package node-debug2 for the marketplace (how do you build node-debug2?). It would be great if you could publish a 1.20.0 version.

[roblourens]

I just press cmd+shift+b :) What problem did you have? I just published 1.20.0.

[weinand]

gulp package fails for me (I wanted to produce a VSIX for installing it in Insiders).

[roblourens]

It works for me

[weinand]

Strange.
I did:

git clean -xfd .
npm install
gulp package

and got:

Are you using an 'npm link' for vscode-chrome-debug-core?

[roblourens]

No, but I just cleaned everything and I see the same thing. Something wrong with my package-lock.json?

[roblourens]

If I delete and regenerate the package-lock, there are a million changes and npm list doesn't fail...?

[weinand]

Interesting: after regenerate the package-lock not only all the error are gone, but the strange problem with the version attribute being a url is gone too:

[roblourens]

The last few times I've done something that changes the package-lock.json, I've seen it randomly add and remove those

[weinand]

This feature works now in Insiders.

Try this sample:

const cluster = require('cluster');
const http = require('http');
const numCPUs = require('os').cpus().length;

if (cluster.isMaster) {
    console.log(`Master ${process.pid} is running`);

    // Fork workers.
    for (let i = 0; i < numCPUs; i++) {
        cluster.fork();
    }

    cluster.on('exit', (worker, code, signal) => {
        console.log(`worker ${worker.process.pid} died`);
    });
} else {
    // Workers can share any TCP connection
    // In this case it is an HTTP server
    http.createServer((req, res) => {
        res.writeHead(200);
        res.end('hello world\n');
    }).listen(8000);

    console.log(`Worker ${process.pid} started`);
}

with this launch config:

{
    "type": "node",
    "request": "launch",
    "name": "Cluster",
    "program": "${workspaceFolder}/test.js",
    "autoAttachChildProcesses": true
},

[ddorkin-issart]

Hello @weinand I can't set up http server on each cluster according with your code.
My environments:

1. windows 10 x64
2. vs code version 1.20.0-insider.
3. node v6.3.1

Output data:

Debugger listening on [::]:11036
Master 16144 is running
Debugger listening on [::]:11037
Debugger listening on [::]:11038
Debugger listening on [::]:11039
Debugger listening on [::]:11040

May be you can clarify why cluster.fork() command doesn't achieve next parts of script.

[weinand]

@ddorkin-issart How does your launch config look like?

[mdhornet90]

@weinand This is excellent when using npm, but is there anything special that needs to be done to incorporate nodemon on a cluster?

[weinand]

@mdhornet90 this feature has nothing to do with npm.
Here are the requirements:

• in order to be able to track the subprocesses, we need the process id. For this we require that the main debuggee launched from the launch config is a node.js process and we use an "evaluate" to find its process id.
• in order to be able to attach to the subprocesses they must be in debug mode. The feature determines whether a process is in debug mode by analysing the program arguments. Currently we detect the patterns --inspect, --inspect-brk, --inspect-port, --debug, --debug-brk, --debug-port (all optionally followed by a '=' and a port number).

[mdhornet90]

Alright, so this feature is independent of _how_ node is started, thanks for the insight!