Diving into C++ internals of node

Fri May 15 2015 20:00:00 GMT-0400 (EDT)

Intro

There is nothing to be scared about in the C++ internals of the project, especially in internals of io.js and node.js.

If you ever tried to optimize JavaScript code to squeeze out every possible performance or memory usage improvement out of it - you already wrote some C++ code.

Many blogs, workshops mention JavaScript optimizations, and some of the popular suggestions are:

Hidden Classes

Declare all properties in the constructor to avoid creating extra "hidden classes". This makes them pretty much the same as a C structures, or C++ classes, where properties are declared ahead of time to help the compiler optimize access to them.

Example:

function Point(x, y, z) {
  this.x = x
  this.y = y
  this.z = z
}

Similar code in C++:

class Point {
 public:
  double x;
  double y;
  double z;
};

Avoid Polymorphism

Avoid storing different types of values in a variables, and avoid passing different types of values as an arguments to the function. This principle could also be called "Make your code monomorphic", or "don't mess with Compiler". This makes code look like as it has static typing, which is what we do in C++.

function add(x, y) {
  return x + y;
}

add(0, 1); // <- good
add('foo', 'bar'); // <- polymorphism!

Compare to:

int add(int x, int y) {
  return x + y;
}

Cache and Reuse

Cache and reuse instances of objects that are expensive to create and are allocated often. This is one is similar to manual memory allocation in C++.

function Parser() {
}

Parser.freelist = [];

Parser.get = function() {
  if (this.freelist.length)
    return this.freelist.pop();
  return new Parser();
};

Parser.prototype.release = ...;

In C++:

Parser* p = new Parser();
delete p;

To conclude, even if you never wrote C++ code, you actually very likely did it in JS.

It is no surprise we use C++ in io.js/node.js. After all, V8 is written in C++ and it provides only a limited set of ECMAScript JavaScript APIs. They are definitely cool, but if you got used to setTimeout() / clearTimeout() - you'll be pretty disappointed to use just plain ECMA.

Our C++ layer lives on top of the event-loop and provides all sorts of APIs: from net sockets to dns queries, from file system events to the zlib bindings. Which is the main reason why node.js was created in the first place!

Short History of C++ layer

History of git blame

To better understand all of these, and to ease the contribution process - it might be a good idea to start with the history of the subject. Luckily, from its inception, node.js is using VCS, in particular git, so the history of the development might be revealed by running git log and git blame on it.

Briefly, git log deps/v8 - has the history of v8 fighting us, and git log src/ - has the history of us fighting v8.

Very first version

Jokes aside, everything started from 61890720 commit. The commit log just says:

add readme and initial code

Unfortunately, we can't elaborate much from it, and need to figure out the details ourselves. What do we see there?

  • libebb - which was used as an HTTP parser. Ryan used the code from the Ebb server that he has previously written for Ruby
  • liboi- which was as a TCP server framework on top of the libev. liboi stands for Library for Output Input

So the first code (that actually started compiling only at 7b7ceea) only had one HTTP server and supplied JavaScript source code was just a handler for it.

function Process(request) {
  if (options.verbose) {
    log("Processing " + request.host +
        request.path +
        " from " + request.referrer +
        "@" + request.userAgent);
  }
  if (!output[request.host])
    output[request.host] = 1;
  else
    output[request.host]++
}

How was it organized internally?

There was a server.cc file which was reading the command line options, loading the JavaScript source file, feeding all of these into V8, and starting the HTTP server.

Second C++ file was js_http_request_processor.cc and it was responsible for invoking the JavaScript http request handler. Not that much for a separate C++ file, right?

It wasn't working that much at that point, and didn't have any of functionality that is provided today. So let's conclude and move on from it quickly.

This version is characterized by following:

  • One file to setup V8 and let JavaScript know about command-line arguments
  • HTTP server fully implemented in C/C++, not invoking the JavaScript for any networking activities
  • One C++ instance per every incoming request, this instance maps some of the HTTP fields (like host, url, method) to the JavaScript object.

The last bullet point is very important to note: the C++ instance <-> JS object mapping is a building brick of all future releases of node.js (including the present one).

064c8f02

Now we quickly jump to 064c8f02. The commit log says:

Use ObjectWrap base class for File, Socket, Server.

And this is the point where node.js has introduced one API to wrap all objects.

net.Server, net.Socket, and File C++ classes are children of this ObjectWrap class. Which means that for every instance of them - there will be one instance of a JS object. Invoking methods on this JS object will invoke C++ methods on the corresponding C++ class, and the constructor itself is a C++ class constructor.

There are now different files for different parts of the provided API:

  • src/node.cc to set up C++ libraries and invoke src/main.js which loads the script file and does some JavaScript initialization. (At this commit we started to write as much code as possible in JavaScript, and leave the rest in the C++ land. This pattern is used in io.js and node.js now too)
  • src/http.cc - http server API, Connection, HttpRequest objects
  • src/file.cc, src/file.js - future fs module. src/file.js consists of the API abstractions for the C++ layer, basically the same thing as with src/node.cc and src/main.js
  • src/process.cc has only exit() method so far, will evolve into the process object
  • src/timers.cc is about setTimeout/setInterval

Just a side note: HTTP server is still provided by liboi, and node.js is using libev.

v0.2

There was lots of growing and maturing from that commit to the v0.2, and most notable of them were about separating the JS parts from the C++ ones, adding CommonJS support, and tons of new modules! The file structure is beginning to look like what we have now:

  • lib/ folder for all JavaScript CommonJS modules
  • src/ for their C++ counterparts
  • deps/ for all dependencies: v8, http-parser, c-ares (for async DNS), libeio (for async FS), and libev (for async networking and auxiliary stuff)

Previously barely used through the src/, ObjectWrap now became a public API, which helped polish it out a lot and improved our core use case as well.

Very importantly, in 064c8f02 all C++ interfaces were global objects. In v0.2 they are provided by process.binding and are thus not directly visible to the user's code.

For example, process.binding('fs'):

> process.binding('fs');
{ access: [Function: access],
  close: [Function: close],
  open: [Function: open],
  ...lots of stuff...

returns lots of C++ methods and classes that are heavily used for interoperation between C++ and JS in lib/fs.js. Similar stuff is done for the rest of the lib/ modules.

v0.6

Just a short note: libev was removed and replaced by libuv. A product of lots of work by Ben Noordhuis, Bert Belder, Ryan Dahl, and others!

The v0.6 version is a major milestone in evolution of node.js. Partly because Windows is now in the list of the officially supported platforms, partly because we have our own single event-loop platform that supports both async File System and async networking operations.

v0.10

Good, stable, but boring...

v0.12 and io.js

Lots of new stuff! :)

Mainly, we have outgrown the ObjectWrap to accommodate the tracing API (which is still needs lots of rework, AFAIK). The hip thing now is AsyncWrap which is in many ways the same thing, but now is attached to some particular domain of operation (i.e. http, dns, tls) and which might have the another AsyncWrap as a parent. Note that ObjectWrap lives in src/node_object_wrap.h, and AsyncWrap in src/async-wrap.h.

This is now the present point of the node.js evolution, and I would like to stop with the Software Archeology at this point.

Interoperation, handles, wraps, and unicorns!

We are finally ready to dive into the C++ internals, and explore them in a greater detail.

As we already figured out - whole APIs provided by the node.js/io.js live in two folders: lib and src. lib holds the core modules, src holds their C++ counterparts.

When you call require('fs') - it does nothing but just executes the contents of the lib/fs.js file. No magic here.

Now comes the interesting part, JavaScript is not capable of file system operations, nor it is capable of networking. This is actually for the best! (You don't want your browser to mess up whole file system, right?) So when you do fs.writeFileSync, or when you are calling http.request() there is a lot of low-level C++ stuff happening outside of the JS-land.

While the fs module is quite simple to explain, it is quite boring too. After all, in most of the cases it is just using number to represent the opened file (so called file descriptor), and it is passing this number around: from C++ to JS, and from JS to C++. Nothing interesting, let's move on!

Certainly much more attractive is the net module. We create sockets, get the connect events, and expect the .write() callbacks to be eventually invoked. All of these should be powered by the C++ machinery!

Here is where most of the interoperation is happening. The tcp_wrap and stream_wrap bindings (remember, process.binding(), right?) provide very useful classes for JS-land: TCP, TCPConnectWrap, WriteWrap, ShutdownWrap.

  • TCP holds the TCP socket and provides methods for writing and reading stuff
  • *Wrap objects are what you pass to the TCP methods when you expect some async action to happen, and need to receive notification (callback) on their completion.

For example, the normal workflow for net.connect() follows:

  • Create TCP instance in lib/net.js, store it in the _handle property of the net.Socket object
  • Parse all arguments to net.connect()
  • Create TCPConnectWrap instance (usually named req)
  • Invoke .connect() method with req, port, host
  • Get req.oncomplete function invoked eventually, once the connection was established, or once the kernel reported an error

In conclusion: most of the C++ classes are either handles, or requests. Requests are very temporary and never outlive the handle that they are bound to, while the handles are something that live much longer (i.e. for the entire life time of the TCP connection).

Speaking of the file structure: TCP is represented by the TCPWrap class in src/tcp_wrap.cc, TCPConnectWrap lives in the same place, and WriteWrap is in the stream_base.cc file (in io.js).

Structure of C++ files

But how does the C++ provide this classes to JavaScript?

Each binding has a NODE_MODULE_CONTEXT_AWARE_BUILTIN macro that registers it in the node.cc. This has the same effect as following JavaScript snippet:

modules[moduleName] = {
  initialized: false,
  initFn: moduleInitFn
};

When process.binding('moduleName') is invoked, node.cc looks up the proper internal binding in this hashmap and initializes it (if it wasn't previously initialized) by calling the supplied function.

process.binding = function(moduleName) {
  var module = modules[moduleName];
  if (module.initialized)
    return module.exports;

  module.exports = {};
  module.initFn(module.exports);
  return module.exports;
};

This initialization function receives exports object as an input, and exports the methods and classes to it in pretty much the same way as you normally do in CommonJS modules.

Each of the exported classes are bound to some C++ classes, and most of them are actually derived from the AsyncWrap C++ class.

The Handle instances are destroyed automatically by V8's GC (once they are closed in JS), and the Wraps are manually destroyed by the Handle, once they are not used anymore.

Side-note:

there are two types of references to the JS objects from C++ land: normal and weak. By default AsyncWraps are referencing their objects in a normal way, which means that the JS objects representing the C++ classes won't be garbage collected until C++ class will dispose the reference. The weak mode is turned on only when the MakeWeak is called somewhere in C++. This might be very useful when debugging memory leaks.

Small exam

Situation

You debug some io.js/node.js issue, and find that it is crashing when instantiating a class provided by process.binding('broken'). Where will you attempt to search for the C++ source code of that class?

Answer

Somewhere in src/. Find NODE_MODULE_CONTEXT_AWARE_BUILTIN(broken, ...) and it is most like going to be in src/broken_something.cc.

C++ Streams

Now comes one of my recent obsessions. The C++ Stream API.

It is a established fact for me that exposing the building blocks of APIs helps to renovate, reshape and make them better a lot better. One of such thing that I was always keen to re-do was a StreamWrap instance.

It was ok-ish in v0.10, but when we moved TLS (SSL) implementation into C++ land it changed dramatically... and, honestly saying, not in a good way.

The previously singular StreamWrap instance, now became a monster that was capable of passing the incoming data elsewhere, skipping the JavaScript callbacks completely and doing some dark-magic OpenSSL machinery on top of it. The implementation worked like a charm, providing much better TLS performance, but the source code became cluttered and rigid.

This "move-parsing-to-elsewhere" thing reminded me a lot about the stream.pipe that we had for JavaScript streams for ages. The natural thing to do about it was to introduce something similar in the C++ land too. This is exactly what was done in io.js, and the results of this live in src/stream_base.cc.

Next step with the C++ Stream APIs

Now we have a very general implementation of this thing that could be reused in many places. The first thing that I expect will be using this might be an HTTP2 stream. To do it in core, we should do it in user-land first, and it could be accomplished only by exposing the C++ Stream API, in the same way as we did it with ObjectWrap.

Epilogue

I'm going to ask you to:

  • Clone the io.js repo
  • Open the src/
  • Go through files in it, and check what you read about it
  • Open src/stream_base.h, src/stream_base.cc and friends and figure out what seems to be wrong to you
  • Send a PR
  • Have fun!