late code injection

[gireeshpunathil]

Is your feature request related to a problem? Please describe.

Currently, a node process cannot receive and execute custom code after it has started. For monitoring tools, a common use case is to be able to attach to an arbitrary process, inject a (tracing) code, instruct the runtime to execute that, and optionally detach from the process.

Describe the solution you'd like

• define a mechanism to interrupt the runtime (signals in platforms where supported)
• specify a Javascript interface for the code injection:
  • the signal is delegated to a Javascript handler function
  • a module with a predefined name (such as agent) is loaded
  • a function with a predefined name (such as execute) is executed
  • a user can define this module and place it in the search path, prior to signalling
  • upon completion, the module is unloaded
• define security restrictions (only requests from matching UID / GID allowed)

Describe alternatives you've considered

• modify the application and integrate with the monitoring tools
• define a specific function in the application that receives code at runtime

[vdeturckheim]

@gireeshpunathil, I have been working on a somehow linked use case for that through using the debugger: https://blog.sqreen.com/remote-debugging-nodejs-runtime-code-injection/ would that match your user story ?
This method require local access which actually provides pretty a good security scope.

[gireeshpunathil]

thanks @vdeturckheim , this is a great article!

technically this matches my use case in that:

• it allows the process to start without any special configuration
• it allows the process be interrupted
• it allows the process to run arbitrary code

here are potential divergence:

• my deployment is container based, so no chrome dev tools
• my target environment is production, so inspector and usage of chrome-remote-interface may not be great
• we don't want the inspector or any open port to linger, after the injected work is complete.

[RaisinTen]

PR: #37503

[addaleax]

my target environment is production, so inspector and usage of chrome-remote-interface may not be great

Can you explain why that would not be good?

we don't want the inspector or any open port to linger, after the injected work is complete.

Doesn't https://nodejs.org/api/inspector.html#inspector_inspector_close solve this for you?

[devsnek]

I can agree that using the inspector is not always ideal. Certain commands may randomly deoptimize code, for example.

I also agree that NODE_OPTIONS=--require seems like an appropriate solution to this.

[gireeshpunathil]

Can you explain why that would not be good?

@addaleax - it is not about functional correctness, or my conviction of its greatness. Containerized production deployments run with a lot of audit / restrictions / compliance, and I don't think ability to opening a debug port will pass such an audit.

Doesn't https://nodejs.org/api/inspector.html#inspector_inspector_close solve this for you?

ok, I wasn't sure how can I get a reference to the running instance of the inspector, from my handler code.

[addaleax]

Can you explain why that would not be good?

@addaleax - it is not about functional correctness, or my conviction of its greatness. Containerized production deployments run with a lot of audit / restrictions / compliance, and I don't think ability to opening a debug port will pass such an audit.

But executing arbitrary code at runtime will pass such an audit...? That sounds very odd to me.

Ultimately, if you're trying to do that, then you'll need to communicate with that process, and there's a limited number of ways to do so. If raising a signal doesn't work because passing parameters that way is hard and you don't want to pre-set paramters in advance (which is fair), then you will need another communication channel, and most likely that's going to look like a debug port or a unix domain socket or something similar, I think. It might make sense to be more specific about the exact requirements you have here.

[gireeshpunathil]

@addaleax - the inspector is experimental, and according to our own documentation, use of such feature is not recommended in production environments. @vdeturckheim's blog also specifically warns against using this method in production.

Exact requirement:

• run an arbitrary code in a running node process
• the target process needs zero preparation (by the user)
• use case: monitoring tools, that can attach to and monitor any (legacy) apps
• how the communication occurs and how the code is passed is left to implementation

Reference: Java has a feature called late attach. With only the knowledge of the target process's PID, a user is able to attach to a JVM and inject a monitoring code.

https://docs.oracle.com/en/java/javase/15/docs/api/jdk.attach/com/sun/tools/attach/VirtualMachine.html

VirtualMachine vm = VirtualMachine.attach("2177");

They use a second Java process as an attacher and complex hand-shaking process before the code is attached I guess, but I don't know the internals fully to comment further.

[devsnek]

aside from googling "java late attach" showing a bunch of articles recommending that you disable it for security reasons, doesn't that go against your own point about about auditing? it might help if you were more concrete in your requirements.

[gireeshpunathil]

@devsnek - the provider does not warn against any security issues.

In terms of concreteness, didn't my previous comment cover that? I specified the requirement, the use case, and the choice that can be made in the implementation. What else would you like to know?

here is a user scenario:

• I have a node.js application deployed (say in container)
• It can be running as a single or a thousand replicas.
• I have no ability to change its code or config.
• I have a monitoring module. I need to monitor those deployments.
• I am able to install the module, alert the target process to launch the module & run some code.

[devsnek]

For example, what would or would not pass the audits you say need to be passed? And assuming that between now and when you want to use this feature, there will have to be another deploy, at least to update node, why can't you use that opportunity to deploy your own system as was suggested in the other thread?

[gireeshpunathil]

For example, what would or would not pass the audits you say need to be passed?

standard practices: usage of stable features.

why can't you use that opportunity to deploy your own system as was suggested in the other thread?

this is not just about fixing any currently running deployments. In future, if a deployment needs to be monitored on demand for a fixed period of time (which is otherwise not configured to be monitored) and detached from the target, it does not make sense to wait for the next available maintenance window?