Skip to content

Latest commit

 

History

History
519 lines (417 loc) · 22.3 KB

mojo_ipc_conversion.md

File metadata and controls

519 lines (417 loc) · 22.3 KB

Converting Legacy IPC to Mojo

[TOC]

Overview

A number of IPC messages sent (primarily between the browser and renderer processes) are still defined using Chrome's old IPC system in //ipc. This system uses base::Pickle as the basis for message serialization and is supported by a number if IPC_* preprocessor macros defined in //ipc and used around the source tree.

There is an ongoing, distributed effort to get these messages converted to Mojo interface messages. Messages that still need to be converted are tracked in two spreadsheets:

This document is concerned primarily with rote conversion of legacy IPC messages to Mojo interface messages. If you are considering more holistic refactoring and better isolation of an entire subsystem of the browser, you may consider servicifying the feature instead of merely converting its IPCs.

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

Legacy IPC Concepts

Each Content child process has a single IPC::Channel implementation going between it and the browser process, and this is used as the sole two-way FIFO to send legacy IPC messages between the processes.

There are two fundamental types of legacy IPC messages: control messages, defined via IPC_MESSAGE_CONTROLn macros (where n is some small integer) and routed messages defined via IPC_MESSAGE_ROUTEDn macros.

Control messages generally go between a browser-side process host (e.g., RenderProcessHost or GpuProcessHost) and the child-side ChildThreadImpl subclass. All of these classes implement IPC::Sender and thus have a Send method for sending a control message to their remote counterpart, and they implement IPC::Listener to receive incoming control messages via OnMessageReceived.

Routed messages are relegated to routes which have arbitrary meaning determined by their use within a given process. For example, renderers use routes to isolate messages scoped to individual render frames, and so such routed messages will travel between a RenderFrameHostImpl and its corresponding RenderFrameImpl, both of which also implement IPC::Sender and IPC::Listener.

Mojo Interfaces as Routes

Routed messages in the old IPC system always carry a routing ID to identify to the receiving endpoint which routed object (e.g. which RenderFrameImpl or RenderViewImpl or whatever) the message is targeting. Each endpoint is thus required to do some additional book-keeping to track what each routing ID means.

Mojo interfaces obviate the need for routing IDs, as new "routes" can be established by simply creating a new interface pipe and passing one endpoint to something which knows how to bind it.

When thinking about an IPC message conversion to Mojo, it's important to consider whether the message is a control message or a routed message, as this determines where you might find an existing Mojo interface to carry your message, or where you will want to add a new end-to-end Mojo interface for that purpose. This can mean the difference between a single per-process interface going between each RenderProcessHostImpl and its corresponding RenderThreadImpl, vs a per-frame interface going between each RenderFrameHostImpl and its corresponding RenderFrameImpl.

Ordering Considerations

One very important consideration when doing IPC conversions is the relative ordering of IPC-driven operations. With the old IPC system, because every message between two processes is globally ordered, it is quite easy for parts of the system to (intentionally or often unintentionally) rely on strict ordering guarantees.

For example, imagine a WebContentsObserver in the browser processes observes a frame navigation and immediately sends an IPC message to the frame to configure some new behavior. The implementation may be inadvertently relying on this message arriving before some other tangentially related message sent to the same frame shortly after the same navigation event.

Mojo does not (and in fact cannot) make any strict ordering guarantees between separate message pipes, as message pipes may be freely moved across process boundaries and thus cannot necessarily share a common FIFO at all times.

If the two messages described above were moved to separate Mojo interfaces on separate message pipes, renderer behavior could break as the first message may arrive after the second message.

The best solution to this problem is to rethink the IPC surface and/or implementation on either side to eliminate ordering dependencies between two interfaces that otherwise seem to be logically distinct. Failing that, Mojo's solution to this problem is to support associated interfaces. In a nutshell, these allow multiple distinct interfaces to be multiplexed over a shared message pipe.

Channel-Associated Interfaces

The previous section mentions associated interfaces as a general-purpose solution for establishing a mutual FIFO between multiple logical Mojo interfaces by having them share a single message pipe.

In Chrome, the IPC::Channel which carries all legacy IPC messages between two processes is itself a Mojo message pipe. We provide a mechanism for associating arbitrary Mojo interfaces with this pipe, which means messages can be converted to Mojo while preserving strict FIFO with respect to other legacy IPC messages. Such interfaces are designated in Chrome parlance as Channel-associated interfaces.

*** aside NOTE: Channel-associated interface acquisition is not constrained by the Service Manager in any way, so security reviewers need to be careful to inspect new additions and uses of such interfaces.


Usage of Channel-associated interfaces should be rare but is considered a reasonable intermediate solution for incremental IPC conversions where it would be too risky or noisy to convert a large IPC surface all at once, but it would also be impossible to split the IPC surface between legacy IPC and a dedicated Mojo interface pipe without introducing timing bugs.

At this point in Chrome's development, practical usage of Channel-associated interfaces is restricted to the IPC::Channel between the browser process and a renderer process, as this is the most complex IPC surface with the most implicit ordering dependencies. A few simple APIs exist to support this.

RenderProcessHostImpl owns an IPC::Channel to its corresponding RenderThreadImpl in the render process. This object has a GetRemoteAssociatedInterfaces method which can be used to pass arbitrary associated interface requests:

mojo::PendingAssociatedRemote<magic::mojom::GoatTeleporter> teleporter;
channel_->GetRemoteAssociatedInterfaces()->GetInterface(teleporter.BindNewEndpointAndPassReceiver());

// These messages are all guaranteed to arrive in the same order they were sent.
channel_->Send(new FooMsg_SomeLegacyIPC);
teleporter->TeleportAllGoats();
channel_->Send(new FooMsg_AnotherLegacyIPC);

Likewise, ChildThreadImpl has an IPC::Channel that can be used in the same way to send such messages back to the browser.

To receive and bind incoming Channel-associated interface requests, the above objects also implement IPC::Listener::OnAssociatedInterfaceRequest.

For supplementation of routed messages, both RenderFrameHostImpl and RenderFrameImpl define a GetRemoteAssociatedInterfaces method which works like the one on IPC::Channel, and both objects also implement IPC::Listener::OnAssociatedInterfaceRequest for processing incoming associated interface requests specific to their own frame.

There are some example conversion CLs which use Channel-associated interfaces here and here.

Deciding How to Approach a Conversion

There are a few questions you should ask before embarking upon any IPC message conversion journey, and there are many potential approaches to consider. The right one depends on context.

Note that this section assumes the message is traveling between the browser process and a renderer process. Other cases are rare and developers may wish to consult chromium-mojo@chromium.org before proceeding with them. Otherwise, apply the following basic algorithm to decide how to proceed:

  • General note: If the message is a reply to some other message (typically these take a "request ID" argument), see the note about message replies at the bottom of this section.
  • Consider whether or not the message makes sense as part of the IPC surface of a new or existing service somewhere in //services or //chrome/services, etc. This is less and less likely to be the case as time goes on, as many remaining IPC conversions are quite narrowly dealing with specific browser/renderer details rather than the browser's supporting subsystems. If defining a new service, you may wish to consult some of the other Mojo & Services documentation first.
  • If the message is an IPC_MESSAGE_CONTROL message:
    • If there are likely to be strict ordering requirements between this message and other legacy IPC or Channel-associated interface messages, consider using a new or existing Channel-associated interface between RenderProcessHostImpl and RenderThreadImpl.
    • If the message is sent from a renderer to the browser:
      • If an existing interface is bound by RenderProcessHostImpl and requested through RenderThread's Connector and seems to be a good fit for the message, add the equivalent Mojo message to that interface.
      • If no such interface exists, consider adding one for this message and any related messages.
    • If the message is sent from the browser to a renderer:
      • If an existing interface is bound by RenderThreadImpl and requested through a BrowserContext Connector referencing a specific RenderProcessHost identity, and the interface seems to be a good fit for the message, add the equivalent Mojo message to that interface.
      • If no such interface exists, consider adding one for this message and any related messages.
  • If the message is an IPC_MESSAGE_ROUTED message:
    • Determine what the routing endpoints are. If they are RenderFrameHostImpl and RenderFrameImpl:
      • If there are likely to be strict ordering requirements between this message and other legacy IPC or Channel-associated interface messages, consider using a new or existing Channel-associated interface between RenderFrameHostImpl and RenderFrameImpl.
      • If the message is sent from a renderer to the browser:
        • If an existing interface is bound by RenderFrameHostImpl and acquired either via RenderFrame::GetRemoteInterfaces or RenderFrame::GetBrowserInterfaceBroker and the interface seems to be a good fit for this message, add the equivalent Mojo message to that interface.
        • If no such interface exists, consider adding one and registering it with RenderFrameHostImpl's BrowserInterfaceBroker. See the simple example in the "Intro to Mojo & Services" document.
      • If the message is sent from the browser to a renderer, consider adding a Mojo equivalent to the content.mojom.Frame interface defined here.
    • If the routing endpoints are not frame objects (for example, they may be RenderView/RenderViewHost objects), this is a special case which does not yet have an easy conversion approach readily available. Contact chromium-mojo@chromium.org to propose or discuss options.

*** aside NOTE: If you are converting a sync IPC, see the section on Synchronous Calls in the Mojo documentation.


Dealing With Replies

If the message is a reply, meaning it has a "request ID" which correlates it to a prior message in the opposite direction, consider converting the request message following the algorithm above. Unlike with legacy IPC, Mojo messages support replies as a first-class concept. So for example if you have:

IPC_CONTROL_MESSAGE2(FooHostMsg_DoTheThing,
                     int /* request_id */,
                     std::string /* name */);
IPC_CONTROL_MESSAGE2(FooMsg_DidTheThing,
                     int /* request_id */,
                     bool /* success */);

You should consider defining an interface Foo which is bound in RenderProcessHostImpl and acquired from RenderThreadImpl, with the following mojom definition:

interface Foo {
  DoTheThing(string name) => (bool success);
};

See Receiving responses for more information.

Repurposing IPC::ParamTraits and IPC_STRUCT* Invocations

Occasionally it is useful to do partial IPC conversions, where you want to convert a message to a Mojo interface method but you don't want to necessarily convert every structure passed by the message. In this case, you can leverage Mojo's type-mapping system to repurpose existing IPC::ParamTraits.

*** aside NOTE: Although in some cases IPC::ParamTraits<T> specializations are defined manually in library code, the IPC_STRUCT* macro helpers also define IPC::ParamTraits<T> specializations under the hood. All advice in this section pertains to both kinds of definitions.


If a mojom struct is declared without a struct body and is tagged with [Native], and a corresponding typemap is provided for the struct, the emitted C++ bindings will -- as if by magic -- replace the mojom type with the typemapped C++ type and will internally use the existing IPC::ParamTraits<T> specialization for that type in order to serialize and deserialize the struct.

For example, given the resource_messages.h header which defines an IPC mapping for content::ResourceRequest:

IPC_STRUCT_TRAITS_BEGIN(content::ResourceRequest)
  IPC_STRUCT_TRAITS_MEMBER(method)
  IPC_STRUCT_TRAITS_MEMBER(url)
  // ...
IPC_STRUCT_TRAITS_END()

and the resource_request.h header which actually defines the content::ResourceRequest type:

namespace content {

struct CONTENT_EXPORT ResourceRequest {
  // ...
};

}  // namespace content

we can declare a corresponding "native" mojom struct:

module content.mojom;

[Native]
struct URLRequest;

and add a typemap like url_request.typemap to define how to map between them:

mojom = "//content/public/common/url_loader.mojom"
public_headers = [ "//content/common/resource_request.h" ]
traits_headers = [ "//content/common/resource_messages.h" ]
...
type_mappings = [ "content.mojom.URLRequest=content::ResourceRequest" ]

Note specifically that public_headers includes the definition of the native C++ type, and traits_headers includes the definition of the legacy IPC traits.

As a result of all this, other mojom files can now reference content.mojom.URLRequest as a type for method parameters and other struct fields, and the generated C++ bindings will represent those values exclusively as content::ResourceRequest objects.

This same basic approach can be used to leverage existing IPC_ENUM_TRAITS for invocations for [Native] mojom enum aliases.

*** aside NOTE: Use of [Native] mojom definitions is strictly limited to C++ bindings. If a mojom message depends on such definitions, it cannot be sent or received by other language bindings. This feature also depends on continued support for legacy IPC serialization and all uses of it should therefore be treated as technical debt.


Blink-Specific Advice

Variants

Let's assume we have a mojom file such as this:

module example.mojom;

interface Foo {
  SendData(string param1, array<int32> param2);
};

The following GN snippet will generate two concrete targets: example and example_blink:

mojom("example") {
  sources = [ "example.mojom" ]
}

The target example will generate Chromium-style C++ bindings using STL types:

// example.mojom.h
namespace example {
namespace mojom {

class Example {
  virtual void SendArray(const std::string& param1, const std::vector<int32_t>& param2) = 0;
}

} // namespace mojom
} // namespace example

The target example_blink will generate Blink-style C++ bindings using WTF types:

// example.mojom-blink.h
namespace example {
namespace mojom {
namespace blink {

class Example {
  virtual void SendArray(const WTF::String& param1, const WTF::Vector<int32_t>& param2) = 0;
}

} // namespace blink
} // namespace mojom
} // namespace example

Thanks to these separate sets of bindings no work is necessary to convert types between Blink-style code and Chromium-style code. It is handled automatically during message serialization and deserialization.

For more information about variants, see this section of the C++ bindings documentation.

Binding callbacks

Mojo methods that return a value take an instance of base::OnceCallback. Use WTF::Bind() and an appropriate wrapper function depending on the type of object and the callback.

For garbage-collected (Oilpan) classes owning the mojo::Remote, it is recommended to use WrapWeakPersistent(this) for connection error handlers since they are not guaranteed to get called in a finite time period (wrapping the object with WrapPersistent in this case would cause memory leaks).

If the response can be discarded in case the object is not alive by the time the response is received, use WrapWeakPersistent(this) for binding the response callback:

// src/third_party/blink/renderer/modules/device_orientation/device_sensor_entry.cc
sensor_.set_connection_error_handler(WTF::Bind(
    &DeviceSensorEntry::HandleSensorError, WrapWeakPersistent(this)));
sensor_->ConfigureReadingChangeNotifications(/*enabled=*/false);
sensor_->AddConfiguration(
    std::move(config), WTF::Bind(&DeviceSensorEntry::OnSensorAddConfiguration,
                                 WrapWeakPersistent(this)));

Otherwise (for example, if the response callback is used to resolve a Promise), use WrapPersistent(this) to keep the object alive:

// src/third_party/blink/renderer/modules/nfc/nfc.cc
ScriptPromiseResolver* resolver = ScriptPromiseResolver::Create(script_state);
... 
nfc_->CancelAllWatches(WTF::Bind(&NFC::OnRequestCompleted,
                                 WrapPersistent(this),
                                 WrapPersistent(resolver)));

Non-garbage-collected objects can use WTF::Unretained(this) for both response and error handler callbacks when the mojo::Remote is owned by the object bound to the callback or the object is guaranteed to outlive the Mojo connection for another reason. Otherwise a weak pointer should be used. However, it is not a common pattern since using Oilpan is recommended for all Blink code.

Implementing Mojo interfaces in Blink

Only a mojo::Receiver or mojo::ReceiverSet should be used when implementing a Mojo interface in an Oilpan-managed object. The object must then have a pre-finalizer to close any open pipes when the object is about to be swept as lazy sweeping means that it may be invalid long before the destructor is called. This requires setup in both the object header and implementation.

// MyObject.h
class MyObject : public GarbageCollected,
                 public example::mojom::blink::Example {
  USING_PRE_FINALIZER(MyObject, Dispose);

 public:
  MyObject();
  void Dispose();

  // Implementation of example::mojom::blink::Example.

 private:
  mojo::Receiver<example::mojom::blink::Example> m_receiver{this};
};

// MyObject.cpp
void MyObject::Dispose() {
  m_receiver.Close();
}

For more information about Blink's Garbage Collector, see Blink GC API Reference.

Typemaps For Content and Blink Types

Using typemapping for messages that go between Blink and content browser code can sometimes be tricky due to things like dependency cycles or confusion over the correct place for some definition to live. There are some example CLs provided here, but feel free to also contact chromium-mojo@chromium.org with specific details if you encounter trouble.

This CL introduces a Mojom definition and typemap for ui::WindowOpenDisposition as a precursor to the IPC conversion below.

The follow-up CL uses that definition along with several other new typemaps (including native typemaps as described above) to convert the relatively large ViewHostMsg_CreateWindow message to Mojo.

Additional Support

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