Skip to content

Latest commit

 

History

History
861 lines (668 loc) · 33 KB

mojo_and_services.md

File metadata and controls

861 lines (668 loc) · 33 KB

Intro to Mojo & Services

[TOC]

Overview

This document contains the minimum amount of information needed for a developer to start using Mojo effectively in Chromium, with example Mojo interface usage, service definition and hookup, and a brief overview of the Content layer's core services.

See other Mojo & Services documentation for introductory guides, API references, and more.

Mojo Terminology

A message pipe is a pair of endpoints. Each endpoint has a queue of incoming messages, and writing a message at one endpoint effectively enqueues that message on the other (peer) endpoint. Message pipes are thus bidirectional.

A mojom file describes interfaces, which are strongly-typed collections of messages. Each interface message is roughly analogous to a single proto message, for developers who are familiar with Google protobufs.

Given a mojom interface and a message pipe, one of the endpoints can be designated as an InterfacePtr and is used to send messages described by the interface. The other endpoint can be designated as a Binding and is used to receive interface messages.

*** aside NOTE: The above generalization is a bit oversimplified. Remember that the message pipe is still bidirectional, and it's possible for a mojom message to expect a reply. Replies are sent from the Binding endpoint and received by the InterfacePtr endpoint.


The Binding endpoint must be associated with (i.e. bound to) an implementation of its mojom interface in order to process received messages. A received message is dispatched as a scheduled task invoking the corresponding interface method on the implementation object.

Another way to think about all this is simply that an InterfacePtr makes calls on a remote implementation of its interface associated with a corresponding remote Binding.

Example: Defining a New Frame Interface

Let's apply this to Chrome. Suppose we want to send a "Ping" message from a render frame to its corresponding RenderFrameHostImpl instance in the browser process. We need to define a nice mojom interface for this purpose, create a pipe to use that interface, and then plumb one end of the pipe to the right place so the sent messages can be received and processed there. This section goes through that process in detail.

Defining the Interface

The first step involves creating a new .mojom file with an interface definition, like so:

// src/example/public/mojom/ping_responder.mojom
module example.mojom;

interface PingResponder {
  // Receives a "Ping" and responds with a random integer.
  Ping() => (int32 random);
};

This should have a corresponding build rule to generate C++ bindings for the definition here:

# src/example/public/mojom/BUILD.gn
import("//mojo/public/tools/bindings/mojom.gni")
mojom("mojom") {
  sources = [ "ping_responder.mojom" ]
}

Creating the Pipe

Now let's create a message pipe to use this interface.

*** aside As a general rule and as a matter of convenience when using Mojo, the client of an interface (i.e. the InterfacePtr side) is typically the party who creates a new pipe. This is convenient because the InterfacePtr may be used to start sending messages immediately without waiting for the InterfaceRequest endpoint to be transferred or bound anywhere.


This code would be placed somewhere in the renderer:

example::mojom::PingResponderPtr ping_responder;
example::mojom::PingResponderRequest request =
    mojo::MakeRequest(&ping_responder);

In this example, ping_responder is the InterfacePtr, and request is an InterfaceRequest, which is a Binding precursor that will eventually be turned into a Binding. mojo::MakeRequest is the most common way to create a message pipe: it yields both endpoints as strongly-typed objects, with the InterfacePtr as an output argument and the InterfaceRequest as the return value.

*** aside NOTE: Every mojom interface T generates corresponding C++ type aliases TPtr = InterfacePtr<T> and TRequest = InterfaceRequest<T>. Chromium code almost exclusively uses these aliases instead of writing out the more verbose templated name.

Also note that an InterfaceRequest doesn't actually do anything. It is an inert holder of a single message pipe endpoint. It exists only to make its endpoint more strongly-typed at compile-time, indicating that the endpoint expects to be bound by a Binding of the same interface type.


Sending a Message

Finally, we can call the Ping() method on our InterfacePtr to send a message:

ping_responder->Ping(base::BindOnce(&OnPong));

*** aside IMPORTANT: If we want to receive the the response, we must keep the ping_responder object alive until OnPong is invoked. After all, ping_responder owns its message pipe endpoint. If it's destroyed then so is the endpoint, and there will be nothing to receive the response message.


We're almost done! Of course, if everything were this easy, this document wouldn't need to exist. We've taken the hard problem of sending a message from a renderer process to the browser process, and transformed it into a problem where we just need to take the request object from above and pass it to the browser process somehow where it can be turned into a Binding that dispatches its received messages.

Sending an InterfaceRequest to the Browser

It's worth noting that InterfaceRequests (and message pipe endpoints in general) are just another type of object that can be freely sent over mojom messages. The most common way to get an InterfaceRequest somewhere is to pass it as a method argument on some other already-connected interface.

One such interface which we always have connected between a renderer's RenderFrameImpl and its corresponding RenderFrameHostImpl in the browser is DocumentInterfaceBroker. We can update this definition to add support for our new PingResponder interface:

interface DocumentInterfaceBroker {
  ...

  GetPingResponder(PingResponder& responder);
}

The & syntax is not a reference! In mojom it denotes an InterfaceRequest. Specifically in this case, the GetPingResponder takes a single PingResponderRequest argument. If the & were omitted, this would instead take a PingResponderPtr.

Now the renderer can call this method with the request object it created earlier via mojo::MakeRequest:

RenderFrame* my_frame = GetMyFrame();
my_frame->GetDocumentInterfaceBroker()->GetPingResponder(std::move(request));

This will transfer the PingResponderRequest endpoint to the browser process where it will be received by the corresponding DocumentInterfaceBroker implementation. More on that below.

Implementing the Interface

Finally, we need a browser-side implementation of our PingResponder interface as well as an implementation of the new DocumentInterfaceBroker.GetPingResponder message. Let's implement PingResponder first:

#include "example/public/mojom/ping_responder.mojom.h"

class PingResponderImpl : example::mojom::PingResponder {
 public:
  explicit PingResponderImpl(example::mojom::PingResponderRequest request)
      : binding_(this, std::move(request)) {}

  // example::mojom::PingResponder:
  void Ping(PingCallback callback) override {
    // Respond with a random 4, chosen by fair dice roll.
    std::move(callback).Run(4);
  }

 private:
  mojo::Binding<example::mojom::PingResponder> binding_;

  DISALLOW_COPY_AND_ASSIGN(PingResponderImpl);
};

And conveniently RenderFrameHostImpl implements DocumentInterfaceBroker, and any calls made on the object returned by RenderFrameImpl::GetDocumentInterfaceBroker() will be routed directly to the RenderFrameHostImpl. So the only thing left to do is update RenderFrameHostImpl to implement GetPingResponder. If you forget to do this the compiler will complain anyway, because generated mojom interface methods are pure virtual methods in C++.

// render_frame_host_impl.h
class RenderFrameHostImpl
  ...
  void GetPingResponder(example::mojom::PingResponderRequest request) override;
  ...
 private:
  ...
  std::unique_ptr<PingResponderImpl> ping_responder_;
  ...
};

// render_frame_host_impl.cc
void RenderFrameHostImpl::GetPingResponder(
    example::mojom::PingResponderRequest request) {
  ping_responder_ = std::make_unique<PingResponderImpl>(std::move(request));
}

And we're done. This setup is sufficient to plumb a new interface connection between a renderer frame and its browser-side host object!

Assuming we kept our ping_responder object alive in the renderer long enough, we would eventually see its OnPong callback invoked with the totally random value of 4, as defined by the browser-side implementation above.

Services Overview & Terminology

The previous section only scratches the surface of how Mojo IPC is used in Chromium. While renderer-to-browser messaging is simple and possibly the most prevalent usage by sheer code volume, we are incrementally decomposing the codebase into a set of services with a bit more granularity than the traditional Content browser/renderer/gpu/utility process split.

A service is a self-contained library of code which implements one or more related features or behaviors and whose interaction with outside code is done exclusively through Mojo interface connections facilitated by the Service Manager.

The Service Manager is a component which can run in a dedicated process or embedded within another process. Only one Service Manager exists globally across the system, and in Chromium the browser process runs an embedded Service Manager instance immediately on startup. The Service Manager spawns service instances on-demand, and it routes each interface request from a service instance to some destination instance of the Service Manager's choosing.

Each service instance implements the Service interface to receive incoming interface requests brokered by the Service Manager, and each service instance has a Connector it can use to issue interface requests to other services via the Service Manager.

Every service has a manifest which declares some static metadata about the service. This metadata is used by the Service Manager for various purposes, including as a declaration of what interfaces are exposed to other services in the system. This eases the security review process.

Inside its manifest every service declares its service name, used to identify instances of the service in the most general sense. Names are free-form and usually short strings which must be globally unique. Some services defined in Chromium today include "device", "identity", and "network" services.

For more complete and in-depth coverage of the concepts covered here and other related APIs, see the Service Manager documentation.

Example: Building a Simple Out-of-Process Service

There are multiple steps required to get a new service up and running in Chromium. You must:

  • Define the Service implementation
  • Define the service's manifest
  • Tell Chromium's Service Manager about the manifest
  • Tell Chromium how to instantiate the Service implementation when it's needed

This section walks through these steps with some brief explanations. For more thorough documentation of the concepts and APIs used herein, see the Service Manager and Mojo documentation.

Defining the Service

Typically service definitions are placed in a services directory, either at the top level of the tree or within some subdirectory. In this example, we'll define a new service for use by Chrome specifically, so we'll define it within //chrome/services.

We can create the following files. First some mojoms:

// src/chrome/services/math/public/mojom/constants.mojom
module math.mojom;

// These are not used by the implementation directly, but will be used in
// following sections.
const string kServiceName = "math";
const string kArithmeticCapability = "arithmetic";
// src/chrome/services/math/public/mojom/divider.mojom
module math.mojom;

interface Divider {
  Divide(int32 dividend, int32 divisor) => (int32 quotient);
};
# src/chrome/services/math/public/mojom/BUILD.gn
import("//mojo/public/tools/bindings/mojom.gni")

mojom("mojom") {
  sources = [
    "constants.mojom",
    "divider.mojom",
  ]
}

Then the actual Service implementation:

// src/chrome/services/math/math_service.h
#include "services/service_manager/public/cpp/service.h"

#include "base/macros.h"
#include "chrome/services/math/public/mojom/divider.mojom.h"

namespace math {

class MathService : public service_manager::Service,
                    public mojom::Divider {
 public:
  explicit MathService(service_manager::mojom::ServiceRequest request);
  ~MathService() override;

 private:
  // service_manager::Service:
  void OnBindInterface(const service_manager::BindSourceInfo& source,
                       const std::string& interface_name,
                       mojo::ScopedMessagePipeHandle interface_pipe) override;

  // mojom::Divider:
  void Divide(int32_t dividend,
              int32_t divisor,
              DivideCallback callback) override;

  service_manager::ServiceBinding service_binding_;

  // You could also use a Binding. We use BindingSet to conveniently allow
  // multiple clients to bind to the same instance of this class. See Mojo
  // C++ Bindings documentation for more information.
  mojo::BindingSet<mojom::Divider> divider_bindings_;

  DISALLOW_COPY_AND_ASSIGN(MathService);
};

}  // namespace math
// src/chrome/services/math/math_service.cc
#include "chrome/services/math/math_service.h"

namespace math {

MathService::MathService(service_manager::ServiceRequest request)
    : service_binding_(this, std::move(request)) {}

MathService::~MathService() = default;

void MathService::OnBindInterface(
    const service_manager::BindSourceInfo& source,
    const std::string& interface_name,
    mojo::ScopedMessagePipeHandle interface_pipe) {
  // Note that services typically use a service_manager::BinderRegistry if they
  // plan on handling many different interface request types.
  if (interface_name == mojom::Divider::Name_) {
    divider_bindings_.AddBinding(
        this, mojom::DividerRequest(std::move(interface_pipe)));
  }
}

void MathService::Divide(int32_t dividend,
                         int32_t divisor,
                         DivideCallback callback) {
  // Respond with the quotient!
  std::move(callback).Run(dividend / divisor);
}

}  // namespace math
# src/chrome/services/math/BUILD.gn

source_set("math") {
  sources = [
    "math.cc",
    "math.h",
  ]

  deps = [
    "//base",
    "//chrome/services/math/public/mojom",
    "//services/service_manager/public/cpp",
  ]
}

Now we have a fully defined math service implementation, including a nice little Divider interface for clients to play with. Next we need to define the service's manifest to declare how the service can be used.

Defining the Manifest

Manifests are defined as Manifest objects, typically built using a ManifestBuilder. As a general rule, services should define their manifest in a dedicated source_set or component target under their public/cpp subdirectory (typically referred to as the service's C++ client library).

We can create the following files for this purpose:

// src/chrome/services/math/public/cpp/manifest.h
#include "services/service_manager/public/cpp/manifest.h"

namespace math {

const service_manager::Manifest& GetManifest();

}  // namespace math
// src/chrome/services/math/public/cpp/manifest.cc
#include "chrome/services/math/public/cpp/manifest.h"

#include "base/no_destructor.h"
#include "chrome/services/math/public/mojom/constants.mojom.h"
#include "chrome/services/math/public/mojom/divider.mojom.h"
#include "services/service_manager/public/cpp/manifest_builder.h"

namespace math {

const service_manager::Manifest& GetManifest() {
  static base::NoDestructor<service_manager::Manifest> manifest{
      service_manager::ManifestBuilder()
          .WithServiceName(mojom::kServiceName)
          .ExposeCapability(
              mojom::kArithmeticCapability,
              service_manager::Manifest::InterfaceList<mojom::Divider>())
          .Build()};
  return *manifest
}

}  // namespace math

We also need to define a build target for our manifest sources:

# src/chrome/services/math/public/cpp/BUILD.gn

source_set("manifest") {
  sources = [
    "manifest.cc",
    "manifest.h",
  ]

  deps = [
    "//base",
    "//chrome/services/math/public/mojom",
    "//services/service_manager/public/cpp",
  ]
}

The above Manifest definition declares that the service is named math and that it exposes a single capability named arithmetic which allows access to the Divider interface.

Another service may require this capability from its own manifest in order for the Service Manager to grant it access to a Divider. We'll see this a few sections below. First, let's get the manifest and service implementation registered with Chromium's Service Manager.

Registering the Manifest

For the most common out-of-process service cases, we register service manifests by packaging them in Chrome. This can be done by augmenting the value returned by GetChromePackagedServiceManifests.

We can add our manifest there:

// Deep within src/chrome/app/chrome_packaged_service_manifests.cc...
const std::vector<service_manager::Manifest>
GetChromePackagedServiceManifests() {
      ...
      math::GetManifest(),
      ...

And don't forget to add a GN dependency from //chrome/app:chrome_packaged_service_manifests onto //chrome/services/math/public/cpp:manifest!

We're almost done with service setup. The last step is to teach Chromium (and thus the Service Manager) how to launch an instance of our beautiful math service.

Hooking Up the Service Implementation

There are two parts to this for an out-of-process Chrome service.

First, we need to inform the embedded Service Manager that this service is an out-of-process service. The goofiness of this part is a product of some legacy issues and it should be eliminated soon, but for now it just means teaching the Service Manager how to label the process it creates for this service (e.g. how the process will appear in the system task manager). We modify ChromeContentBrowserClient::RegisterOutOfProcessServices for this:

void ChromeContentBrowserClient::RegisterOutOfProcessServices(
    OutOfProcessServicesMap* services) {
  ...

  (*services)[math::mojom::kServiceName] =
      base::BindRepeating([]() -> base::string16 {
        return "Math Service";
      });

  ...
}

And finally, since nearly all out-of-process services run in a "utility" process today, we need to add a dependency on our actual Service implementation to Chrome's service spawning code within the utility process.

For this step we just modify ChromeContentUtilityClient::MaybeCreateMainThreadService by adding a block of code as follows:

std::unique_ptr<service_manager::Service> ChromeContentUtilityClient::MaybeCreateMainThreadService(
    const std::string& service_name,
    service_manager::mojom::ServiceRequest request) {
  ...

  if (service_name == math::mojom::kServiceName)
    return std::make_unique<math::MathService>(std::move(request));

  ...
}

And we're done!

As one nice follow-up step, let's use our math service from the browser.

Using the Service

We can grant the browser process access to our Divider interface by requiring the math service's arithmetic capability within the content_browser service manifest.

*** aside NOTE: See the following section for an elaboration on what content_browser is. For the sake of this example, it's magic.


For Chrome-specific features such as our glorious new math service, we can amend the content_browser manifest by modifying GetChromeContentBrowserOverlayManifest as follows:

// src/chrome/app/chrome_content_browser_overlay_manifest.cc

...
const service_manager::Manifest& GetChromeContentBrowserOverlayManifest() {
  ...
        .RequireCapability(math::mojom::kServiceName,
                           math::mojom::kArithmeticCapability)
  ...
}

Finally, we can use the global content_browser instance's Connector to send an interface request to our service. This is accessible from the main thread of the browser process. Somewhere in src/chrome/browser, we can write:

// This gives us the global content_browser's Connector
service_manager::Connector* connector =
    content::ServiceManagerConnection::GetForProcess()->GetConnector();

// Recall from the earlier Mojo section that mojo::MakeRequest creates a new
// message pipe for our interface. Connector passes the request endpoint to
// the Service Manager along with the name of our target service, "math".
math::mojom::DividerPtr divider;
connector->BindInterface(math::mojom::kServiceName,
                         mojo::MakeRequest(&divider));

// As a client, we do not have to wait for any acknowledgement or confirmation
// of a connection. We can start queueing messages immediately and they will be
// delivered as soon as the service is up and running.
divider->Divide(
    42, 6, base::BindOnce([](int32_t quotient) { LOG(INFO) << quotient; }));

*** aside NOTE: To ensure the execution of the response callback, the DividerPtr object must be kept alive (see this section and this note from an earlier section).


This should successfully spawn a new process to run the math service if it's not already running, then ask it to do a division, and ultimately log the result after it's sent back to the browser process.

Finally it's worth reiterating that every service instance in the system has its own Connector and there's no reason we have to limit ourselves to content_browser as the client, as long as the appropriate manifest declares that it requires our arithmetic capability.

If we did not update the content_browser manifest overlay as we did in this example, the Divide call would never reach the math service (in fact the service wouldn't even be started) and instead we'd get an error message (or in developer builds, an assertion failure) informing us that the Service Manager blocked the BindInterface call.

Content-Layer Services Overview

Apart from very early initialization steps in the browser process, every bit of logic in Chromium today is effectively running as part of one service instance or another.

Although we continue to migrate parts of the browser's privileged functionality to more granular services defined below the Content layer, the main services defined in Chromium today continue to model the Content layer's classical multiprocess architecture which defines a handful of process types: browser, renderer, gpu, utility, and plugin processes. For each of these process types, we now define corresponding services.

Manifest definitions for all of the following services can be found in //content/public/app.

The Browser Service

content_browser is defined to encapsulate general-purpose browser process code. There are multiple instances of this service, all running within the singular browser process. There is one shared global instance as well an additional instance for each BrowserContext (i.e. per Chrome profile).

The global instance exists primarily so that arbitrary browser process code can reach various system services conveniently via a global Connector instance on the main thread.

Each instance associated with a BrowserContext is placed in an isolated instance group specific to that BrowserContext. This limits the service instances with which its Connector can make contact. These instances are used primarily to facilitate the spawning of other isolated per-profile service instances, such as renderers and plugins.

The Renderer Service

A content_renderer instance is spawned in its own sandboxed process for every site-isolated instance of Blink we require. Instances are placed in the same instance group as the renderer's corresponding BrowserContext, i.e. the profile which navigated to the site being rendered.

Most interfaces used by content_renderer are not brokered through the Service Manager but instead are brokered through dedicated interfaces implemented by content_browser, with which each renderer maintains persistent connections.

The GPU Service

Only a single instance of content_gpu exists at a time and it always runs in its own isolated, sandboxed process. This service hosts the code in content/gpu and whatever else Content's embedder adds to that for GPU support.

The Plugin Service

content_plugin hosts a plugin in an isolated process. Similarly to content_renderer instances, each instance of content_plugin belongs to an instance group associated with a specific BrowserContext, and in general plugins get most of their functionality by talking directly to content_browser rather than brokering interface requests through the Service Manager.

The Utility Service

content_utility exists only nominally today, as there is no remaining API surface within Content which would allow a caller to explicitly create an instance of it. Instead, this service is used exclusively to bootstrap new isolated processes in which other services will run.

Exposing Interfaces Between Content Processes

Apart from the standard Service Manager APIs, the Content layer defines a number of additional concepts for Content and its embedder to expose interfaces specifically between Content processes in various contexts.

Exposing Browser Interfaces to Renderer Documents and Workers

Documents and workers are somewhat of a special case since interface access decisions often require browser-centric state that the Service Manager cannot know about, such as details of the current BrowserContext, the origin of the renderered content, installed extensions in the renderer, etc. For this reason, interface brokering decisions are increasingly being made by the browser.

There are two ways this is done: the Deprecated way and the New way.

The Deprecated Way: InterfaceProvider

This is built on the concept of interface filters and the InterfaceProvider interface. It is deprecated and new features should use The New Way instead. This section only briefly covers practical usage in Chromium.

The content_browser manifest exposes capabilities on a few named interface filters, the main one being "navigation:frame". There are others scoped to different worker contexts, e.g. "navigation:service_worker". RenderProcessHostImpl or RenderFrameHostImpl sets up an InterfaceProvider for each known execution context in the corresponding renderer, filtered through the Service Manager according to one of the named filters.

The practical result of all this means the interface must be listed in the content_browser manifest under the ExposeInterfaceFilterCapability_Deprecated("navigation:frame", "renderer", ...) entry, and a corresponding interface request handler must be registered with the host's registry_ in RenderFrameHostImpl::RegisterMojoInterfaces

Similarly for worker contexts, an interface must be exposed by the "renderer" capability on the corresponding interface filter (e.g., "navigation:shared_worker") and a request handler must be registered within RendererInterfaceBinders::InitializeParameterizedBinderRegistry.

The best way to understand all of this after reading this section is to look at the linked code above and examine a few examples. They are fairly repetitive. For additional convenience, here is also a link to the content_browser manifest.

The New Way: Interface Brokers

*** aside In classic Google tradition, the New Way is not entirely ready yet. As of this writing, worker-scoped interfaces must still use the Old Way described above.


Rather than the confusing spaghetti of interface filter logic, we now define an explicit mojom interface with a persistent connection between a renderer's frame object and the corresponding RenderFrameHostImpl in the browser process. This interface is called DocumentInterfaceBroker and is fairly easy to work with: you simply add a new factory method to the interface definition:

interface DocumentInterfaceBroker {
  ...

  GetGoatTeleporter(magic.mojom.GoatTeleporter& request);
};

and implement this new method on RenderFrameHostImpl, which is an implementation (the production implementation) of DocumentInterfaceBroker:

void RenderFrameHostImpl::GetGoatTeleporter(
    magic::mojom::GoatTeleporterRequest request) {
  goat_teleporter_binding_.Bind(std::move(request));
}

Exposing Browser Interfaces to Render Processes

Sometimes (albeit rarely) it's useful to expose a browser interface directly to a renderer process. This can be done as for any other interface exposed between two services. In this specific instance, the content_browser manifest exposes a capability named "renderer" which content_renderer requires. Any interface listed as part of that capability can be accessed by a content_renderer instance by using its own Connector. See below.

Exposing Browser Interfaces to Content Child Processes

All Content child process types (renderer, GPU, and plugin) share a common API to interface with the Service Manager. Their Service Manager connection is initialized and maintained by ChildThreadImpl on process startup, and from the main thread, you can access the process's Connector as follows:

auto* connector = content::ChildThread::Get()->GetConnector();

// For example...
connector->BindInterface(content::mojom::kBrowserServiceName,
                         std::move(some_request));

Exposing Content Child Process Interfaces to the Browser

Content child processes may also expose interfaces to the browser, though this is much less common and requires a fair bit of caution since the browser must be careful to only call Connector.BindInterface in these cases with an exact service_manager::Identity to avoid unexpected behavior.

Every child process provides a subclass of ChildThreadImpl, and this can be used to install a new ConnectionFilter on the process's Service Manager connection before starting to accept requests.

This behavior should really be considered deprecated, but for posterity, here is how the GPU process does it:

  1. Disable Service Manager connection auto-start
  2. Register a new ConnectionFilter impl to handle certain interface requests
  3. Start the Service Manager connection manually

It's much more common instead for there to be some primordial interface connection established by the child process which can then be used to facilitate push communications from the browser, so please consider not duplicating this behavior.

Additional Support

If this document was not helpful in some way, please post a message to your friendly chromium-mojo@chromium.org or services-dev@chromium.org mailing list.