Hi! 👋 You've found the C++ code backing Node.js. This README aims to help you get started working on it and document some idioms you may encounter while doing so.
Node.js has a document detailing its C++ coding style that can be helpful as a reference for stylistic issues.
A lot of the Node.js codebase is around what the underlying JavaScript engine, V8, provides through its API for embedders. Knowledge of this API can also be useful when working with native addons for Node.js written in C++, although for new projects N-API is typically the better alternative.
V8 does not provide much public API documentation beyond what is
available in its C++ header files, most importantly v8.h, which can be
accessed online in the following locations:
- On GitHub:
v8.hin Node.js - On GitHub:
v8.hin V8 - On the Chromium project's Code Search application:
v8.hin Code Search
V8 also provides an introduction for V8 embedders, which can be useful for understanding some of the concepts it uses in its embedder API.
Important concepts when using V8 are the ones of Isolates and
JavaScript value handles.
V8 supports fast API calls, which can be useful for improving the performance in certain cases.
The other major dependency of Node.js is libuv, providing the event loop and other operating system abstractions to Node.js.
There is a reference documentation for the libuv API.
The Node.js C++ files follow this structure:
The .h header files contain declarations, and sometimes definitions that don't
require including other headers (e.g. getters, setters, etc.). They should only
include other .h header files and nothing else.
The -inl.h header files contain definitions of inline functions from the
corresponding .h header file (e.g. functions marked inline in the
declaration or template functions). They always include the corresponding
.h header file, and can include other .h and -inl.h header files as
needed. It is not mandatory to split out the definitions from the .h file
into an -inl.h file, but it becomes necessary when there are multiple
definitions and contents of other -inl.h files start being used. Therefore, it
is recommended to split a -inl.h file when inline functions become longer than
a few lines to keep the corresponding .h file readable and clean. All visible
definitions from the -inl.h file should be declared in the corresponding .h
header file.
The .cc files contain definitions of non-inline functions from the
corresponding .h header file. They always include the corresponding .h
header file, and can include other .h and -inl.h header files as needed.
A number of concepts are involved in putting together Node.js on top of V8 and libuv. This section aims to explain some of them and how they work together.
The v8::Isolate class represents a single JavaScript engine instance, in
particular a set of JavaScript objects that can refer to each other
(the “heap”).
The v8::Isolate is often passed to other V8 API functions, and provides some
APIs for managing the behaviour of the JavaScript engine or querying about its
current state or statistics such as memory usage.
V8 APIs are not thread-safe unless explicitly specified. In a typical Node.js
application, the main thread and any Worker threads each have one Isolate,
and JavaScript objects from one Isolate cannot refer to objects from
another Isolate.
Garbage collection, as well as other operations that affect the entire heap,
happen on a per-Isolate basis.
Typical ways of accessing the current Isolate in the Node.js code are:
- Given a
FunctionCallbackInfofor a binding function, usingargs.GetIsolate(). - Given a
Context, usingcontext->GetIsolate(). - Given a
Environment, usingenv->isolate(). - Given a
Realm, usingrealm->isolate().
V8 provides classes that mostly correspond to JavaScript types; for example,
v8::Value is a class representing any kind of JavaScript type, with
subclasses such as v8::Number (which in turn has subclasses like v8::Int32),
v8::Boolean or v8::Object. Most types are represented by subclasses
of v8::Object, e.g. v8::Uint8Array or v8::Date.
V8 provides the ability to store data in so-called “internal fields” inside
v8::Objects that were created as instances of C++-backed classes. The number
of fields needs to be defined when creating that class.
Both JavaScript values and void* pointers may be stored in such fields.
In most native Node.js objects, the first internal field is used to store a
pointer to a BaseObject subclass, which then contains all relevant
information associated with the JavaScript object.
Typical ways of working with internal fields are:
obj->InternalFieldCount()to look up the number of internal fields for an object (0for regular JavaScript objects).obj->GetInternalField(i)to get a JavaScript value from an internal field.obj->SetInternalField(i, v)to store a JavaScript value in an internal field.obj->GetAlignedPointerFromInternalField(i)to get avoid*pointer from an internal field.obj->SetAlignedPointerInInternalField(i, p)to store avoid*pointer in an internal field.
Contexts provide the same feature under the name “embedder data”.
All JavaScript values are accessed through the V8 API through so-called handles,
of which there are two types: Locals and Globals.
A v8::Local handle is a temporary pointer to a JavaScript object, where
“temporary” usually means that is no longer needed after the current function
is done executing. Local handles can only be allocated on the C++ stack.
Most of the V8 API uses Local handles to work with JavaScript values or return
them from functions.
Whenever a Local handle is created, a v8::HandleScope or
v8::EscapableHandleScope object must exist on the stack. The Local is then
added to that scope and deleted along with it.
When inside a binding function, a HandleScope already exists outside of
it, so there is no need to explicitly create one.
EscapableHandleScopes can be used to allow a single Local handle to be
passed to the outer scope. This is useful when a function returns a Local.
The following JavaScript and C++ functions are mostly equivalent:
function getFoo(obj) {
return obj.foo;
}v8::Local<v8::Value> GetFoo(v8::Local<v8::Context> context,
v8::Local<v8::Object> obj) {
v8::Isolate* isolate = context->GetIsolate();
v8::EscapableHandleScope handle_scope(isolate);
// The 'foo_string' handle cannot be returned from this function because
// it is not “escaped” with `.Escape()`.
v8::Local<v8::String> foo_string =
v8::String::NewFromUtf8(isolate, "foo").ToLocalChecked();
v8::Local<v8::Value> return_value;
if (obj->Get(context, foo_string).ToLocal(&return_value)) {
return handle_scope.Escape(return_value);
} else {
// There was a JS exception! Handle it somehow.
return v8::Local<v8::Value>();
}
}See exception handling for more information about the usage of .To(),
.ToLocalChecked(), v8::Maybe and v8::MaybeLocal usage.
If it is known that a Local<Value> refers to a more specific type, it can
be cast to that type using .As<...>():
v8::Local<v8::Value> some_value;
// CHECK() is a Node.js utilitity that works similar to assert().
CHECK(some_value->IsUint8Array());
v8::Local<v8::Uint8Array> as_uint8 = some_value.As<v8::Uint8Array>();Generally, using val.As<v8::X>() is only valid if val->IsX() is true, and
failing to follow that rule may lead to crashes.
If it is expected that no Local handles should be created within a given
scope unless explicitly within a HandleScope, a SealHandleScope can be used.
For example, there is a SealHandleScope around the event loop, forcing
any functions that are called from the event loop and want to run or access
JavaScript code to create HandleScopes.
A v8::Global handle (sometimes also referred to by the name of its parent
class Persistent, although use of that is discouraged in Node.js) is a
reference to a JavaScript object that can remain active as long as the engine
instance is active.
Global handles can be either strong or weak. Strong global handles are so-called “GC roots”, meaning that they will keep the JavaScript object they refer to alive even if no other objects refer to them. Weak global handles do not do that, and instead optionally call a callback when the object they refer to is garbage-collected.
v8::Global<v8::Object> reference;
void StoreReference(v8::Isolate* isolate, v8::Local<v8::Object> obj) {
// Create a strong reference to `obj`.
reference.Reset(isolate, obj);
}
// Must be called with a HandleScope around it.
v8::Local<v8::Object> LoadReference(v8::Isolate* isolate) {
return reference.Get(isolate);
}v8::Eternal handles are a special kind of handles similar to v8::Global
handles, with the exception that the values they point to are never
garbage-collected while the JavaScript Engine instance is alive, even if
the v8::Eternal itself is destroyed at some point. This type of handle
is rarely used.
JavaScript allows multiple global objects and sets of built-in JavaScript
objects (like the Object or Array functions) to coexist inside the same
heap. Node.js exposes this ability through the vm module.
V8 refers to each of these global objects and their associated builtins as a
Context.
Currently, in Node.js there is one main Context associated with the
principal Realm of an Environment instance, and a number of
subsidiary Contexts that are created with vm.Context or associated with
ShadowRealm.
Most Node.js features will only work inside a context associated with a
Realm. The only exception at the time of writing are MessagePort
objects. This restriction is not inherent to the design of Node.js, and a
sufficiently committed person could restructure Node.js to provide built-in
modules inside of vm.Contexts.
Often, the Context is passed around for exception handling.
Typical ways of accessing the current Context in the Node.js code are:
- Given an
Isolate, usingisolate->GetCurrentContext(). - Given an
Environment, usingenv->context()to get theEnvironment's principalRealm's context. - Given a
Realm, usingrealm->context()to get theRealm's context.
The main abstraction for an event loop inside Node.js is the uv_loop_t struct.
Typically, there is one event loop per thread. This includes not only the main
thread and Workers, but also helper threads that may occasionally be spawned
in the course of running a Node.js program.
The current event loop can be accessed using env->event_loop() given an
Environment instance. The restriction of using a single event loop
is not inherent to the design of Node.js, and a sufficiently committed person
could restructure Node.js to provide e.g. the ability to run parts of Node.js
inside an event loop separate from the active thread's event loop.
Node.js instances are represented by the Environment class.
Currently, every Environment class is associated with:
- One event loop
- One
Isolate - One principal
Realm
The Environment class contains a large number of different fields for
different built-in modules that can be shared across different Realm
instances, for example, the inspector agent, async hooks info.
Typical ways of accessing the current Environment in the Node.js code are:
- Given a
FunctionCallbackInfofor a binding function, usingEnvironment::GetCurrent(args). - Given a
BaseObject, usingenv()orself->env(). - Given a
Context, usingEnvironment::GetCurrent(context). This requires thatcontexthas been associated with theEnvironmentinstance, e.g. is the mainContextfor theEnvironmentor one of itsvm.Contexts. - Given an
Isolate, usingEnvironment::GetCurrent(isolate). This looks up the currentContextand then uses that.
The Realm class is a container for a set of JavaScript objects and functions
that are associated with a particular ECMAScript realm.
Each ECMAScript realm comes with a global object and a set of intrinsic
objects. An ECMAScript realm has a [[HostDefined]] field, which represents
the Node.js Realm object.
Every Realm instance is created for a particular Context. A Realm
can be a principal realm or a synthetic realm. A principal realm is created
for each Environment's main Context. A synthetic realm is created
for the Context of each ShadowRealm constructed from the JS API. No
Realm is created for the Context of a vm.Context.
Native bindings and built-in modules can be evaluated in either a principal realm or a synthetic realm.
The Realm class contains a large number of different fields for
different built-in modules, for example the memory for a Uint32Array that
the url module uses for storing data returned from a
urlBinding.update() call.
It also provides cleanup hooks and maintains a list of BaseObject
instances.
Typical ways of accessing the current Realm in the Node.js code are:
- Given a
FunctionCallbackInfofor a binding function, usingRealm::GetCurrent(args). - Given a
BaseObject, usingrealm()orself->realm(). - Given a
Context, usingRealm::GetCurrent(context). This requires thatcontexthas been associated with theRealminstance, e.g. is the principalRealmfor theEnvironment. - Given an
Isolate, usingRealm::GetCurrent(isolate). This looks up the currentContextand then uses itsRealm.
Every Node.js instance (Environment) is associated with one IsolateData
instance that contains information about or associated with a given
Isolate.
IsolateData contains a list of strings that can be quickly accessed
inside Node.js code, e.g. given an Environment instance env the JavaScript
string “name” can be accessed through env->name_string() without actually
creating a new JavaScript string.
Every process that uses V8 has a v8::Platform instance that provides some
functionalities to V8, most importantly the ability to schedule work on
background threads.
Node.js provides a NodePlatform class that implements the v8::Platform
interface and uses libuv for providing background threading abilities.
The platform can be accessed through isolate_data->platform() given an
IsolateData instance, although that only works when:
- The current Node.js instance was not started by an embedder; or
- The current Node.js instance was started by an embedder whose
v8::Platformimplementation also implement's thenode::MultiIsolatePlatforminterface and who passed this to Node.js.
C++ functions exposed to JS follow a specific signature. The following example
is from node_util.cc:
void ArrayBufferViewHasBuffer(const FunctionCallbackInfo<Value>& args) {
CHECK(args[0]->IsArrayBufferView());
args.GetReturnValue().Set(args[0].As<ArrayBufferView>()->HasBuffer());
}(Namespaces are usually omitted through the use of using statements in the
Node.js source code.)
args[n] is a Local<Value> that represents the n-th argument passed to the
function. args.This() is the this value inside this function call.
args.GetReturnValue() is a placeholder for the return value of the function,
and provides a .Set() method that can be called with a boolean, integer,
floating-point number or a Local<Value> to set the return value.
Node.js provides various helpers for building JS classes in C++ and/or attaching C++ functions to the exports of a built-in module:
void Initialize(Local<Object> target,
Local<Value> unused,
Local<Context> context,
void* priv) {
Environment* env = Environment::GetCurrent(context);
SetMethod(context, target, "getaddrinfo", GetAddrInfo);
SetMethod(context, target, "getnameinfo", GetNameInfo);
// 'SetMethodNoSideEffect' means that debuggers can safely execute this
// function for e.g. previews.
SetMethodNoSideEffect(context, target, "canonicalizeIP", CanonicalizeIP);
// ... more code ...
Isolate* isolate = env->isolate();
// Building the `ChannelWrap` class for JS:
Local<FunctionTemplate> channel_wrap =
NewFunctionTemplate(isolate, ChannelWrap::New);
// Allow for 1 internal field, see `BaseObject` for details on this:
channel_wrap->InstanceTemplate()->SetInternalFieldCount(1);
channel_wrap->Inherit(AsyncWrap::GetConstructorTemplate(env));
// Set various methods on the class (i.e. on the prototype):
SetProtoMethod(isolate, channel_wrap, "queryAny", Query<QueryAnyWrap>);
SetProtoMethod(isolate, channel_wrap, "queryA", Query<QueryAWrap>);
// ...
SetProtoMethod(isolate, channel_wrap, "querySoa", Query<QuerySoaWrap>);
SetProtoMethod(isolate, channel_wrap, "getHostByAddr", Query<GetHostByAddrWrap>);
SetProtoMethodNoSideEffect(isolate, channel_wrap, "getServers", GetServers);
SetConstructorFunction(context, target, "ChannelWrap", channel_wrap);
}
// Run the `Initialize` function when loading this binding through
// `internalBinding('cares_wrap')` in Node.js's built-in JavaScript code:
NODE_BINDING_CONTEXT_AWARE_INTERNAL(cares_wrap, Initialize)If the C++ binding is loaded during bootstrap, it needs to be registered
with the utilities in node_external_reference.h, like this:
namespace node {
namespace util {
void RegisterExternalReferences(ExternalReferenceRegistry* registry) {
registry->Register(GetHiddenValue);
registry->Register(SetHiddenValue);
// ... register all C++ functions used to create FunctionTemplates.
}
} // namespace util
} // namespace node
// The first argument passed to `NODE_BINDING_EXTERNAL_REFERENCE`,
// which is `util` here, needs to be added to the
// `EXTERNAL_REFERENCE_BINDING_LIST_BASE` list in node_external_reference.h
NODE_BINDING_EXTERNAL_REFERENCE(util, node::util::RegisterExternalReferences)Otherwise, you might see an error message like this when building the executables:
FAILED: gen/node_snapshot.cc
cd ../../; out/Release/node_mksnapshot out/Release/gen/node_snapshot.cc
Unknown external reference 0x107769200.
<unresolved>
/bin/sh: line 1: 6963 Illegal instruction: 4 out/Release/node_mksnapshot out/Release/gen/node_snapshot.ccYou can try using a debugger to symbolicate the external reference. For example,
with lldb's image lookup --address command (with gdb it's info symbol):
$ lldb -- out/Release/node_mksnapshot out/Release/gen/node_snapshot.cc
(lldb) run
Process 7012 launched: '/Users/joyee/projects/node/out/Release/node_mksnapshot' (x86_64)
Unknown external reference 0x1004c8200.
<unresolved>
Process 7012 stopped
(lldb) image lookup --address 0x1004c8200
Address: node_mksnapshot[0x00000001004c8200] (node_mksnapshot.__TEXT.__text + 5009920)
Summary: node_mksnapshot`node::util::GetHiddenValue(v8::FunctionCallbackInfo<v8::Value> const&) at node_util.cc:159Which explains that the unregistered external reference is
node::util::GetHiddenValue defined in node_util.cc.
Some internal bindings, such as the HTTP parser, maintain internal state that
only affects that particular binding. In that case, one common way to store
that state is through the use of Realm::AddBindingData, which gives
binding functions access to an object for storing such state.
That object is always a BaseObject.
In the binding, call SET_BINDING_ID() with an identifier for the binding
type. For example, for http_parser::BindingData, the identifier can be
http_parser_binding_data.
If the binding should be supported in a snapshot, the id and the
fully-specified class name should be added to the SERIALIZABLE_BINDING_TYPES
list in base_object_types.h, and the class should implement the serialization
and deserialization methods. See the comments of SnapshotableObject on how to
implement them. Otherwise, add the id and the class name to the
UNSERIALIZABLE_BINDING_TYPES list instead.
// In base_object_types.h, add the binding to either
// UNSERIALIZABLE_BINDING_TYPES or SERIALIZABLE_BINDING_TYPES.
// The second parameter is a descriptive name of the class, which is
// usually the fully-specified class name.
#define UNSERIALIZABLE_BINDING_TYPES(V) \
V(http_parser_binding_data, http_parser::BindingData)
// In the HTTP parser source code file:
class BindingData : public BaseObject {
public:
BindingData(Realm* realm, Local<Object> obj) : BaseObject(realm, obj) {}
SET_BINDING_ID(http_parser_binding_data)
std::vector<char> parser_buffer;
bool parser_buffer_in_use = false;
// ...
};
// Available for binding functions, e.g. the HTTP Parser constructor:
static void New(const FunctionCallbackInfo<Value>& args) {
BindingData* binding_data = Realm::GetBindingData<BindingData>(args);
new Parser(binding_data, args.This());
}
// ... because the initialization function told the Realm to store the
// BindingData object:
void InitializeHttpParser(Local<Object> target,
Local<Value> unused,
Local<Context> context,
void* priv) {
Realm* realm = Realm::GetCurrent(context);
BindingData* const binding_data = realm->AddBindingData<BindingData>(target);
if (binding_data == nullptr) return;
Local<FunctionTemplate> t = NewFunctionTemplate(realm->isolate(), Parser::New);
...
}The V8 engine provides multiple features to work with JavaScript exceptions, as C++ exceptions are disabled inside of Node.js:
V8 provides the v8::Maybe<T> and v8::MaybeLocal<T> types, typically used
as return values from API functions that can run JavaScript code and therefore
can throw exceptions.
Conceptually, the idea is that every v8::Maybe<T> is either empty (checked
through .IsNothing()) or holds a value of type T (checked through
.IsJust()). If the Maybe is empty, then a JavaScript exception is pending.
A typical way of accessing the value is using the .To() function, which
returns a boolean indicating success of the operation (i.e. the Maybe not
being empty) and taking a pointer to a T to store the value if there is one.
maybe.Check() can be used to assert that the maybe is not empty, i.e. crash
the process otherwise. maybe.FromJust() (aka maybe.ToChecked()) can be used
to access the value and crash the process if it is not set.
This should only be performed if it is actually sure that the operation has not failed. A lot of the Node.js source code does not follow this rule, and can be brought to crash through this.
In particular, it is often not safe to assume that an operation does not throw an exception, even if it seems like it would not do that. The most common reasons for this are:
- Calls to functions like
object->Get(...)orobject->Set(...)may fail on most objects, if theObject.prototypeobject has been modified from userland code that added getters or setters. - Calls that invoke any JavaScript code, including JavaScript code that is
provided from Node.js internals or V8 internals, will fail when JavaScript
execution is being terminated. This typically happens inside Workers when
worker.terminate()is called, but it can also affect the main thread when e.g. Node.js is used as an embedded library. These exceptions can happen at any point. It is not always obvious whether a V8 call will enter JavaScript. In addition to unexpected getters and setters, accessing some types of built-in objects likeMaps andSets can also run V8-internal JavaScript code.
v8::MaybeLocal<T> is a variant of v8::Maybe<T> that is either empty or
holds a value of type Local<T>. It has methods that perform the same
operations as the methods of v8::Maybe, but with different names:
Maybe |
MaybeLocal |
|---|---|
maybe.IsNothing() |
maybe_local.IsEmpty() |
maybe.IsJust() |
!maybe_local.IsEmpty() |
maybe.To(&value) |
maybe_local.ToLocal(&local) |
maybe.ToChecked() |
maybe_local.ToLocalChecked() |
maybe.FromJust() |
maybe_local.ToLocalChecked() |
maybe.Check() |
– |
v8::Nothing<T>() |
v8::MaybeLocal<T>() |
v8::Just<T>(value) |
v8::MaybeLocal<T>(value) |
Usually, the best approach to encountering an empty Maybe is to just return
from the current function as soon as possible, and let execution in JavaScript
land resume. If the empty Maybe is encountered inside a nested function,
is may be a good idea to use a Maybe or MaybeLocal for the return type
of that function and pass information about pending JavaScript exceptions along
that way.
Generally, when an empty Maybe is encountered, it is not valid to attempt
to perform further calls to APIs that return Maybes.
A typical pattern for dealing with APIs that return Maybe and MaybeLocal is
using .ToLocal() and .To() and returning early in case there is an error:
// This could also return a v8::MaybeLocal<v8::Number>, for example.
v8::Maybe<double> SumNumbers(v8::Local<v8::Context> context,
v8::Local<v8::Array> array_of_integers) {
v8::Isolate* isolate = context->GetIsolate();
v8::HandleScope handle_scope(isolate);
double sum = 0;
for (uint32_t i = 0; i < array_of_integers->Length(); i++) {
v8::Local<v8::Value> entry;
if (!array_of_integers->Get(context, i).ToLocal(&entry)) {
// Oops, we might have hit a getter that throws an exception!
// It's better to not continue return an empty (“nothing”) Maybe.
return v8::Nothing<double>();
}
if (!entry->IsNumber()) {
// Let's just skip any non-numbers. It would also be reasonable to throw
// an exception here, e.g. using the error system in src/node_errors.h,
// and then to return an empty Maybe again.
continue;
}
// This cast is valid, because we've made sure it's really a number.
v8::Local<v8::Number> entry_as_number = entry.As<v8::Number>();
sum += entry_as_number->Value();
}
return v8::Just(sum);
}
// Function that is exposed to JS:
void SumNumbers(const v8::FunctionCallbackInfo<v8::Value>& args) {
// This will crash if the first argument is not an array. Let's assume we
// have performed type checking in a JavaScript wrapper function.
CHECK(args[0]->IsArray());
double sum;
if (!SumNumbers(args.GetIsolate()->GetCurrentContext(),
args[0].As<v8::Array>()).To(&sum)) {
// Nothing to do, we can just return directly to JavaScript.
return;
}
args.GetReturnValue().Set(sum);
}If there is a need to catch JavaScript exceptions in C++, V8 provides the
v8::TryCatch type for doing so, which we wrap into our own
node::errors::TryCatchScope in Node.js. The latter has the additional feature
of providing the ability to shut down the program in the typical Node.js way
(printing the exception + stack trace) if an exception is caught.
A TryCatch will catch regular JavaScript exceptions, as well as termination
exceptions such as the ones thrown by worker.terminate() calls.
In the latter case, the try_catch.HasTerminated() function will return true,
and the exception object will not be a meaningful JavaScript value.
try_catch.ReThrow() should not be used in this case.
Two central concepts when working with libuv are handles and requests.
Handles are subclasses of the uv_handle_t “class”, and generally refer to
long-lived objects that can emit events multiple times, such as network sockets
or file system watchers.
In Node.js, handles are often managed through a HandleWrap subclass.
Requests are one-time asynchronous function calls on the event loop, such as file system requests or network write operations, that either succeed or fail.
In Node.js, requests are often managed through a ReqWrap subclass.
When a Node.js Environment is destroyed, it generally needs to clean up
any resources owned by it, e.g. memory or libuv requests/handles.
Cleanup hooks are provided that run before the Environment or the
Realm is destroyed. They can be added and removed by using
env->AddCleanupHook(callback, hint); and
env->RemoveCleanupHook(callback, hint);, or
realm->AddCleanupHook(callback, hint); and
realm->RemoveCleanupHook(callback, hint); respectively, where callback takes
a void* hint argument.
Inside these cleanup hooks, new asynchronous operations may be started on the event loop, although ideally that is avoided as much as possible.
For every ReqWrap and HandleWrap instance, the cleanup of the
associated libuv objects is performed automatically, i.e. handles are closed
and requests are cancelled if possible.
Realm cleanup depends on the realm types. All realms are destroyed when the
Environment is destroyed with the cleanup hook. A ShadowRealm can
also be destroyed by the garbage collection when there is no strong reference
to it.
Every BaseObject is tracked with its creation realm and will be destroyed
when the realm is tearing down.
If a libuv handle is not managed through a HandleWrap instance,
it needs to be closed explicitly. Do not use uv_close() for that, but rather
env->CloseHandle(), which works the same way but keeps track of the number
of handles that are still closing.
There is no way to abort libuv requests in general. If a libuv request is not
managed through a ReqWrap instance, the
env->IncreaseWaitingRequestCounter() and
env->DecreaseWaitingRequestCounter() functions need to be used to keep track
of the number of active libuv requests.
Calling into JavaScript is not allowed during cleanup. Worker threads explicitly forbid this during their shutdown sequence, but the main thread does not for backwards compatibility reasons.
When calling into JavaScript without using MakeCallback(), check the
env->can_call_into_js() flag and do not proceed if it is set to false.
A large number of classes in the Node.js C++ codebase refer to other objects.
The MemoryRetainer class is a helper for annotating C++ classes with
information that can be used by the heap snapshot builder in V8, so that
memory retained by C++ can be tracked in V8 heap snapshots captured in
Node.js applications.
Inheriting from the MemoryRetainer class enables objects (both from JavaScript
and C++) to refer to instances of that class, and in turn enables that class
to point to other objects as well, including native C++ types
such as std::string and track their memory usage.
This can be useful for debugging memory leaks.
The memory_tracker.h header file explains how to use this class.
A frequently recurring situation is that a JavaScript object and a C++ object
need to be tied together. BaseObject is the main abstraction for that in
Node.js, and most classes that are associated with JavaScript objects are
subclasses of it. It is defined in base_object.h.
Every BaseObject is associated with one Realm and one
v8::Object. The v8::Object needs to have at least one internal field
that is used for storing the pointer to the C++ object. In order to ensure this,
the V8 SetInternalFieldCount() function is usually used when setting up the
class from C++.
The JavaScript object can be accessed as a v8::Local<v8::Object> by using
self->object(), given a BaseObject named self.
Accessing a BaseObject from a v8::Local<v8::Object> (frequently that is
args.This() in a binding function) can be done using
the Unwrap<T>(obj) function, where T is a subclass of BaseObject.
A helper for this is the ASSIGN_OR_RETURN_UNWRAP macro that returns from the
current function if unwrapping fails (typically that means that the BaseObject
has been deleted earlier).
void Http2Session::Request(const FunctionCallbackInfo<Value>& args) {
Http2Session* session;
ASSIGN_OR_RETURN_UNWRAP(&session, args.This());
Environment* env = session->env();
Local<Context> context = env->context();
Isolate* isolate = env->isolate();
// ...
// The actual function body, which can now use the `session` object.
// ...
}The BaseObject class comes with a set of features that allow managing the
lifetime of its instances, either associating it with the lifetime of the
corresponding JavaScript object or untying the two.
The BaseObject::MakeWeak() method turns the underlying Global handle
into a weak one, and makes it so that the BaseObject::OnGCCollect() virtual
method is called when the JavaScript object is garbage collected. By default,
that methods deletes the BaseObject instance.
BaseObject::ClearWeak() undoes this effect.
It generally makes sense to call MakeWeak() in the constructor of a
BaseObject subclass, unless that subclass is referred to by e.g. the event
loop, as is the case for the HandleWrap and ReqWrap classes.
In addition, there are two kinds of smart pointers that can be used to refer
to BaseObjects.
BaseObjectWeakPtr<T> is similar to std::weak_ptr<T>, but holds on to
an object of a BaseObject subclass T and integrates with the lifetime
management of the former. When the BaseObject no longer exists, e.g. when
it was garbage collected, accessing it through weak_ptr.get() will return
nullptr.
BaseObjectPtr<T> is similar to std::shared_ptr<T>, but also holds on to
objects of a BaseObject subclass T. While there are BaseObjectPtrs
pointing to a given object, the BaseObject will always maintain a strong
reference to its associated JavaScript object. This can be useful when one
BaseObject refers to another BaseObject and wants to make sure it stays
alive during the lifetime of that reference.
A BaseObject can be “detached” through the BaseObject::Detach() method.
In this case, it will be deleted once the last BaseObjectPtr referring to
it is destroyed. There must be at least one such pointer when Detach() is
called. This can be useful when one BaseObject fully owns another
BaseObject.
AsyncWrap is a subclass of BaseObject that additionally provides tracking
functions for asynchronous calls. It is commonly used for classes whose methods
make calls into JavaScript without any JavaScript stack below, i.e. more or less
directly from the event loop. It is defined in async_wrap.h.
Every AsyncWrap subclass has a “provider type”. A list of provider types is
maintained in src/async_wrap.h.
Every AsyncWrap instance is associated with two numbers, the “async id”
and the “async trigger id”. The “async id” is generally unique per AsyncWrap
instance, and only changes when the object is re-used in some way.
See the async_hooks module documentation for more information about how
this information is provided to async tracking tools.
The AsyncWrap class has a set of methods called MakeCallback(), with the
intention of the naming being that it is used to “make calls back into
JavaScript” from the event loop, rather than making callbacks in some way.
(As the naming has made its way into the Node.js public API, it's not worth
the breakage of fixing it).
MakeCallback() generally calls a method on the JavaScript object associated
with the current AsyncWrap, and informs async tracking code about these calls
as well as takes care of running the process.nextTick() and Promise task
queues once it returns.
Before calling MakeCallback(), it is typically necessary to enter both a
HandleScope and a Context::Scope.
void StatWatcher::Callback(uv_fs_poll_t* handle,
int status,
const uv_stat_t* prev,
const uv_stat_t* curr) {
// Get the StatWatcher instance associated with this call from libuv,
// StatWatcher is a subclass of AsyncWrap.
StatWatcher* wrap = ContainerOf(&StatWatcher::watcher_, handle);
Environment* env = wrap->env();
HandleScope handle_scope(env->isolate());
Context::Scope context_scope(env->context());
// Transform 'prev' and 'curr' into an array:
Local<Value> arr = ...;
Local<Value> argv[] = { Integer::New(env->isolate(), status), arr };
wrap->MakeCallback(env->onchange_string(), arraysize(argv), argv);
}See Callback scopes for more information.
HandleWrap is a subclass of AsyncWrap specifically designed to make working
with libuv handles easier. It provides the .ref(), .unref() and
.hasRef() methods as well as .close() to enable easier lifetime management
from JavaScript. It is defined in handle_wrap.h.
HandleWrap instances are cleaned up automatically when the
current Node.js Environment is destroyed, e.g. when a Worker thread stops.
HandleWrap also provides facilities for diagnostic tooling to get an
overview over libuv handles managed by Node.js.
ReqWrap is a subclass of AsyncWrap specifically designed to make working
with libuv requests easier. It is defined in req_wrap.h.
In particular, its Dispatch() method is designed to avoid the need to keep
track of the current count of active libuv requests.
ReqWrap also provides facilities for diagnostic tooling to get an
overview over libuv handles managed by Node.js.
V8 comes with a trace-based C++ garbage collection library called
Oilpan, whose API is in headers underdeps/v8/include/cppgc.
In this document we refer to it as cppgc since that's the namespace
of the library.
C++ objects managed using cppgc are allocated in the V8 heap
and traced by V8's garbage collector. The cppgc library provides
APIs for embedders to create references between cppgc-managed objects
and other objects in the V8 heap (such as JavaScript objects or other
objects in the V8 C++ API that can be passed around with V8 handles)
in a way that's understood by V8's garbage collector.
This helps avoiding accidental memory leaks and use-after-frees coming
from incorrect cross-heap reference tracking, especially when there are
cyclic references. This is what powers the
unified heap design in Chromium to avoid cross-heap memory issues,
and it's being rolled out in Node.js to reap similar benefits.
For general guidance on how to use cppgc, see the
Oilpan documentation in Chromium. In Node.js there is a helper
mixin node::CppgcMixin from cppgc_helpers.h to help implementing
cppgc-managed wrapper objects with a BaseObject-like interface.
cppgc-manged objects in Node.js internals should extend this mixin,
while non-cppgc-managed objects typically extend BaseObject - the
latter are being migrated to be cppgc-managed wherever it's beneficial
and practical. Typically cppgc-managed objects are more efficient to
keep track of (which lowers initialization cost) and work better
with V8's GC scheduling.
A cppgc-managed native wrapper should look something like this:
#include "cppgc_helpers.h"
// CPPGC_MIXIN is a helper macro for inheriting from cppgc::GarbageCollected,
// cppgc::NameProvider and public CppgcMixin. Per cppgc rules, it must be
// placed at the left-most position in the class hierarchy.
class MyWrap final : CPPGC_MIXIN(MyWrap) {
public:
SET_CPPGC_NAME(MyWrap) // Sets the heap snapshot name to "Node / MyWrap"
// The constructor can only be called by `cppgc::MakeGarbageCollected()`.
MyWrap(Environment* env, v8::Local<v8::Object> object);
// Helper for constructing MyWrap via `cppgc::MakeGarbageCollected()`.
// Can be invoked by other C++ code outside of this class if necessary.
// In that case the raw pointer returned may need to be managed by
// cppgc::Persistent<> or cppgc::Member<> with corresponding tracing code.
static MyWrap* New(Environment* env, v8::Local<v8::Object> object);
// Binding method to help constructing MyWrap in JavaScript.
static void New(const v8::FunctionCallbackInfo<v8::Value>& args);
void Trace(cppgc::Visitor* visitor) const final;
}cppgc::GarbageCollected types are expected to implement a
void Trace(cppgc::Visitor* visitor) const method. When they are the
final class in the hierarchy, this method must be marked final. For
classes extending node::CppgcMixn, this should typically dispatch a
call to CppgcMixin::Trace() first, then trace any additional owned data
it has. See deps/v8/include/cppgc/garbage-collected.h see what types of
data can be traced.
void MyWrap::Trace(cppgc::Visitor* visitor) const {
CppgcMixin::Trace(visitor);
visitor->Trace(...); // Trace any additional data MyWrap has
}C++ objects subclassing node::CppgcMixin have a counterpart JavaScript object.
The two references each other internally - this cycle is well-understood by V8's
garbage collector and can be managed properly.
Similar to BaseObjects, cppgc-managed wrappers objects must be created from
object templates with at least node::CppgcMixin::kInternalFieldCount internal
fields. To unify handling of the wrappers, the internal fields of
node::CppgcMixin wrappers would have the same layout as BaseObject.
// To create the v8::FunctionTemplate that can be used to instantiate a
// v8::Function for that serves as the JavaScript constructor of MyWrap:
Local<FunctionTemplate> ctor_template = NewFunctionTemplate(isolate, MyWrap::New);
ctor_template->InstanceTemplate()->SetInternalFieldCount(
ContextifyScript::kInternalFieldCount);cppgc::GarbageCollected objects should not be allocated with usual C++
primitives (e.g. using new or std::make_unique is forbidden). Instead
they must be allocated using cppgc::MakeGarbageCollected - this would
allocate them in the V8 heap and allow V8's garbage collector to trace them.
It's recommended to use a New method to wrap the cppgc::MakeGarbageCollected
call so that external C++ code does not need to know about its memory management
scheme to construct it.
MyWrap* MyWrap::New(Environment* env, v8::Local<v8::Object> object) {
// Per cppgc rules, the constructor of MyWrap cannot be invoked directly.
// It's recommended to implement a New() static method that prepares
// and forwards the necessary arguments to cppgc::MakeGarbageCollected()
// and just return the raw pointer around - do not use any C++ smart
// pointer with this, as this is not managed by the native memory
// allocator but by V8.
return cppgc::MakeGarbageCollected<MyWrap>(
env->isolate()->GetCppHeap()->GetAllocationHandle(), env, object);
}
// Binding method to be invoked by JavaScript.
void MyWrap::New(const FunctionCallbackInfo<Value>& args) {
Environment* env = Environment::GetCurrent(args);
Isolate* isolate = env->isolate();
Local<Context> context = env->context();
CHECK(args.IsConstructCall());
// Get more arguments from JavaScript land if necessary.
New(env, args.This());
}In the constructor of node::CppgcMixin types, use
node::CppgcMixin::Wrap() to finish the wrapping so that
V8 can trace the C++ object from the JavaScript object.
MyWrap::MyWrap(Environment* env, v8::Local<v8::Object> object) {
// This cannot invoke the mixin constructor and has to invoke via a static
// method from it, per cppgc rules.
CppgcMixin::Wrap(this, env, object);
}When given a v8::Local<v8::Object> that is known to be the JavaScript
wrapper object for MyWrap, uses the node::CppgcMixin::Unwrap() to
get the C++ object from it:
v8::Local<v8::Object> object = ...; // Obtain the JavaScript from somewhere.
MyWrap* wrap = CppgcMixin::Unwrap<MyWrap>(object);Similar to ASSIGN_OR_RETURN_UNWRAP, there is a ASSIGN_OR_RETURN_UNWRAP_CPPGC
that can be used in binding methods to return early if the JavaScript object does
not wrap the desired type. And similar to BaseObject, node::CppgcMixin
provides env() and object() methods to quickly access the associated
node::Environment and its JavaScript wrapper object.
ASSIGN_OR_RETURN_UNWRAP_CPPGC(&wrap, object);
CHECK_EQ(wrap->object(), object);Unlike BaseObject which typically uses a v8::Global (either weak or strong)
to reference an object from the V8 heap, cppgc-managed objects are expected to
use v8::TracedReference (which supports any v8::Data). For example if the
MyWrap object owns a v8::UnboundScript, in the class body the reference
should be declared as
class MyWrap : ... {
v8::TracedReference<v8::UnboundScript> script;
}V8's garbage collector traces the references from MyWrap through the
MyWrap::Trace() method, which should call cppgc::Visitor::Trace on the
v8::TracedReference.
void MyWrap::Trace(cppgc::Visitor* visitor) const {
CppgcMixin::Trace(visitor);
visitor->Trace(script); // v8::TracedReference is supported by cppgc::Visitor
}As long as a MyWrap object is alive, the v8::UnboundScript in its
v8::TracedReference will be kept alive. When the MyWrap object is no longer
reachable from the V8 heap, and there are no other references to the
v8::UnboundScript it owns, the v8::UnboundScript will be garbage collected
along with its owning MyWrap. The reference will also be automatically
captured in the heap snapshots.
To create a reference from another JavaScript object to a C++ wrapper
extending node::CppgcMixin, just create a JavaScript to JavaScript
reference using the JavaScript side of the wrapper, which can be accessed
using node::CppgcMixin::object().
MyWrap* wrap = ....; // Obtain a reference to the cppgc-managed object.
Local<Object> referrer = ...; // This is the referrer object.
// To reference the C++ wrap from the JavaScript referrer, simply creates
// a usual JavaScript property reference - the key can be a symbol or a
// number too if necessary, or it can be a private symbol property added
// using SetPrivate(). wrap->object() can also be passed to the JavaScript
// land, which can be referenced by any JavaScript objects in an invisible
// manner using a WeakMap or being inside a closure.
referrer->Set(
context, FIXED_ONE_BYTE_STRING(isolate, "ref"), wrap->object()
).ToLocalChecked();Typically, a newly created cppgc-managed wrapper object should be held alive
by the JavaScript land (for example, by being returned by a method and
staying alive in a closure). Long-lived cppgc objects can also
be held alive from C++ using persistent handles (see
deps/v8/include/cppgc/persistent.h) or as members of other living
cppgc-managed objects (see deps/v8/include/cppgc/member.h) if necessary.
Its destructor will be called when no other objects from the V8 heap reference
it, this can happen at any time after the garbage collector notices that
it's no longer reachable and before the V8 isolate is torn down.
See the Oilpan documentation in Chromium for more details.
The public CallbackScope and the internally used InternalCallbackScope
classes provide the same facilities as MakeCallback(), namely:
- Emitting the
'before'event for async tracking when entering the scope - Setting the current async IDs to the ones passed to the constructor
- Emitting the
'after'event for async tracking when leaving the scope - Running the
process.nextTick()queue - Running microtasks, in particular
Promisecallbacks and async/await functions
Usually, using AsyncWrap::MakeCallback() or using the constructor taking
an AsyncWrap* argument (i.e. used as
InternalCallbackScope callback_scope(this);) suffices inside of the Node.js
C++ codebase.
Node.js uses a few custom C++ utilities, mostly defined in util.h.
Node.js provides Malloc(), Realloc() and Calloc() functions that work
like their C stdlib counterparts, but crash if memory cannot be allocated.
(As V8 does not handle out-of-memory situations gracefully, it does not make
sense for Node.js to attempt to do so in all cases.)
The UncheckedMalloc(), UncheckedRealloc() and UncheckedCalloc() functions
return nullptr in these cases (or when size == 0).
The MaybeStackBuffer class provides a way to allocate memory on the stack
if it is smaller than a given limit, and falls back to allocating it on the
heap if it is larger. This can be useful for performantly allocating temporary
data if it is typically expected to be small (e.g. file paths).
The Utf8Value, TwoByteValue (i.e. UTF-16 value) and BufferValue
(Utf8Value but copy data from a Buffer if one is passed) helpers
inherit from this class and allow accessing the characters in a JavaScript
string this way.
static void Chdir(const FunctionCallbackInfo<Value>& args) {
Environment* env = Environment::GetCurrent(args);
// ...
CHECK(args[0]->IsString());
Utf8Value path(env->isolate(), args[0]);
int err = uv_chdir(*path);
if (err) {
// ... error handling ...
}
}Node.js provides a few macros that behave similar to assert():
CHECK(expression)aborts the process with a stack trace ifexpressionis false.CHECK_EQ(a, b)checks fora == bCHECK_GE(a, b)checks fora >= bCHECK_GT(a, b)checks fora > bCHECK_LE(a, b)checks fora <= bCHECK_LT(a, b)checks fora < bCHECK_NE(a, b)checks fora != bCHECK_NULL(val)checks fora == nullptrCHECK_NOT_NULL(val)checks fora != nullptrCHECK_IMPLIES(a, b)checks thatbis true ifais true.UNREACHABLE([message])aborts the process if it is reached.
CHECKs are always enabled. For checks that should only run in debug mode, use
DCHECK(), DCHECK_EQ(), etc.
The OnScopeLeave() function can be used to run a piece of code when leaving
the current C++ scope.
static void GetUserInfo(const FunctionCallbackInfo<Value>& args) {
Environment* env = Environment::GetCurrent(args);
uv_passwd_t pwd;
// ...
const int err = uv_os_get_passwd(&pwd);
if (err) {
// ... error handling, return early ...
}
auto free_passwd = OnScopeLeave([&]() { uv_os_free_passwd(&pwd); });
// ...
// Turn `pwd` into a JavaScript object now; whenever we return from this
// function, `uv_os_free_passwd()` will be called.
// ...
}