IPDL: Inter-Thread and Inter-Process Message Passing

The Idea

IPDL, the “Inter-[thread|process] Protocol Definition Language”, is the Mozilla-specific language that allows code to communicate between system threads or processes in a standardized, efficient, safe, secure and platform-agnostic way. IPDL communications take place between parent and child objects called actors. The architecture is inspired by the actor model.

Note

IPDL actors differ from the actor model in one significant way – all IPDL communications are only between a parent and its only child.

The actors that constitute a parent/child pair are called peers. Peer actors communicate through an endpoint, which is an end of a message pipe. An actor is explicitly bound to its endpoint, which in turn is bound to a particular thread soon after it is constructed. An actor never changes its endpoint and may only send and receive predeclared messages from/to that endpoint, on that thread. Violations result in runtime errors. A thread may be bound to many otherwise unrelated actors but an endpoint supports top-level actors and any actors they manage (see below).

Note

More precisely, endpoints can be bound to any nsISerialEventTarget, which are themselves associated with a specific thread. By default, IPDL will bind to the current thread’s “main” serial event target, which, if it exists, is retrieved with GetCurrentSerialEventTarget. For the sake of clarity, this document will frequently refer to actors as bound to threads, although the more precise interpretation of serial event targets is also always valid.

Note

Internally, we use the “Ports” component of the Chromium Mojo library to multiplex multiple endpoints (and, therefore, multiple top-level actors). This means that the endpoints communicate over the same native pipe, which conserves limited OS resources. The implications of this are discussed in IPDL Best Practices.

Parent and child actors may be bound to threads in different processes, in different threads in the same process, or even in the same thread in the same process. That last option may seem unreasonable but actors are versatile and their layout can be established at run-time so this could theoretically arise as the result of run-time choices. One large example of this versatility is PCompositorBridge actors, which in different cases connect endpoints in the main process and the GPU process (for UI rendering on Windows), in a content process and the GPU process (for content rendering on Windows), in the main process and the content process (for content rendering on Mac, where there is no GPU process), or between threads on the main process (UI rendering on Mac). For the most part, this does not require elaborate or redundant coding; it just needs endpoints to be bound judiciously at runtime. The example in Connecting With Other Processes shows one way this can be done. It also shows that, without proper plain-language documentation of all of the ways endpoints are configured, this can quickly lead to unmaintainable code. Be sure to document your endpoint bindings throroughly!!!

The Approach

The actor framework will schedule tasks to run on its associated event target, in response to messages it receives. Messages are specified in an IPDL protocol file and the response handler tasks are defined per-message by C++ methods. As actors only communicate in pairs, and each is bound to one thread, sending is always done sequentially, never concurrently (same for receiving). This means that it can, and does, guarantee that an actor will always receive messages in the same order they were sent by its related actor – and that this order is well defined since the related actor can only send from one thread.

Warning

There are a few (rare) exceptions to the message order guarantee. They include synchronous nested messages, interrupt messages, and messages with a [Priority] or [Compress] annotation.

An IPDL protocol file specifies the messages that may be sent between parent and child actors, as well as the direction and payload of those messages. Messages look like function calls but, from the standpoint of their caller, they may start and end at any time in the future – they are asynchronous, so they won’t block their sending actors or any other components that may be running in the actor’s thread’s MessageLoop.

Note

Not all IPDL messages are asynchronous. Again, we run into exceptions for messages that are synchronous, synchronous nested or interrupt. Use of synchronous and nested messages is strongly discouraged but may not always be avoidable. They will be defined later, along with superior alternatives to both that should work in nearly all cases. Interrupt messages were prone to misuse and are deprecated, with removal expected in the near future (Bug 1729044).

Protocol files are compiled by the IPDL compiler in an early stage of the build process. The compiler generates C++ code that reflects the protocol. Specifically, it creates one C++ class that represents the parent actor and one that represents the child. The generated files are then automatically included in the C++ build process. The generated classes contain public methods for sending the protocol messages, which client code will use as the entry-point to IPC communication. The generated methods are built atop our IPC framework, defined in /ipc, that standardizes the safe and secure use of sockets, pipes, shared memory, etc on all supported platforms. See Using The IPDL compiler for more on integration with the build process.

Client code must be written that subclasses these generated classes, in order to add handlers for the tasks generated to respond to each message. It must also add routines (ParamTraits) that define serialization and deserialization for any types used in the payload of a message that aren’t already known to the IPDL system. Primitive types, and a bunch of Mozilla types, have predefined ParamTraits (here and here).

Note

Among other things, client code that uses the generated code must include chromium-config.mozbuild in its moz.build file. See Using The IPDL compiler for a complete list of required build changes.

The Steps To Making A New Actor

  1. Decide what folder you will work in and create:

    1. An IPDL protocol file, named for your actor (e.g. PMyActor.ipdl – actor protocols must begin with a P). See The Protocol Language.

    2. Properly-named source files for your actor’s parent and child implementations (e.g. MyActorParent.h, MyActorChild.h and, optionally, adjacent .cpp files). See The C++ Interface.

    3. IPDL-specific updates to the moz.build file. See Using The IPDL compiler.

  2. Write your actor protocol (.ipdl) file:

    1. Decide whether you need a top-level actor or a managed actor. See Top Level Actors.

    2. Find/write the IPDL and C++ data types you will use in communication. Write ParamTraits for C++ data types that don’t have them. See Generating IPDL-Aware C++ Data Types: IPDL Structs and Unions for IPDL structures. See Referencing Externally Defined Data Types: IPDL Includes and ParamTraits for C++ data types.

    3. Write your actor and its messages. See Defining Actors.

  3. Write C++ code to create and destroy instances of your actor at runtime.

  4. Write handlers for your actor’s messages. See Actors and Messages in C++.

  5. Start sending messages through your actors! Again, see Actors and Messages in C++.

The Protocol Language

This document will follow the integration of two actors into Firefox – PMyManager and PMyManaged. PMyManager will manage PMyManaged. A good place to start is with the IPDL actor definitions. These are files that are named for the actor (e.g. PMyManager.ipdl) and that declare the messages that a protocol understands. These actors are for demonstration purposes and involve quite a bit of functionality. Most actors will use a very small fraction of these features.

These files reference three additional files. MyTypes.ipdlh is an “IPDL header” that can be included into .ipdl files as if it were inline, except that it also needs to include any external actors and data types it uses:

MyActorUtils.h and MyDataTypes.h are normal C++ header files that contain definitions for types passed by these messages, as well as instructions for serializing them. They will be covered in The C++ Interface.

Using The IPDL compiler

To build IPDL files, list them (alphabetically sorted) in a moz.build file. In this example, the .ipdl and .ipdlh files would be alongside a moz.build containing:

IPDL_SOURCES += [
    "MyTypes.ipdlh",
    "PMyManaged.ipdl",
    "PMyManager.ipdl",
]

UNIFIED_SOURCES += [
    "MyManagedChild.cpp",
    "MyManagedParent.cpp",
    "MyManagerChild.cpp",
    "MyManagerParent.cpp",
]

include("/ipc/chromium/chromium-config.mozbuild")

chromium-config.mozbuild sets up paths so that generated IPDL header files are in the proper scope. If it isn’t included, the build will fail with #include errors in both your actor code and some internal ipc headers. For example:

c:/mozilla-src/mozilla-unified/obj-64/dist/include\ipc/IPCMessageUtils.h(13,10): fatal error: 'build/build_config.h' file not found

.ipdl files are compiled to C++ files as one of the earliest post-configure build steps. Those files are, in turn, referenced throughout the source code and build process. From PMyManager.ipdl the compiler generates two header files added to the build context and exported globally: mozilla/myns/PMyManagerParent.h and mozilla/myns/PMyManagerChild.h, as discussed in Namespaces below. These files contain the base classes for the actors. It also makes several other files, including C++ source files and another header, that are automatically included into the build and should not require attention.

C++ definions of the actors are required for IPDL. They define the actions that are taken in response to messages – without this, they would have no value. There will be much more on this when we discuss Actors and Messages in C++ but note here that C++ header files named for the actor are required by the IPDL compiler. The example would expect mozilla/myns/MyManagedChild.h, mozilla/myns/MyManagedParent.h, mozilla/myns/MyManagerChild.h and mozilla/myns/MyManagerParent.h and will not build without them.

Referencing Externally Defined Data Types: IPDL Includes

Let’s begin with PMyManager.ipdl. It starts by including types that it will need from other places:

include protocol PMyManaged;
include MyTypes;                          // for MyActorPair

using MyActorEnum from "mozilla/myns/MyActorUtils.h";
using struct mozilla::myns::MyData from "mozilla/MyDataTypes.h";
[MoveOnly] using mozilla::myns::MyOtherData from "mozilla/MyDataTypes.h";
[RefCounted] using class mozilla::myns::MyThirdData from "mozilla/MyDataTypes.h";

The first line includes a protocol that PMyManager will manage. That protocol is defined in its own .ipdl file. Cyclic references are expected and pose no concern.

The second line includes the file MyTypes.ipdlh, which defines types like structs and unions, but in IPDL, which means they have behavior that goes beyond the similar C++ concepts. Details can be found in Generating IPDL-Aware C++ Data Types: IPDL Structs and Unions.

The final lines include types from C++ headers. Additionally, the [RefCounted] and [MoveOnly] attributes tell IPDL that the types have special functionality that is important to operations. These are the data type attributes currently understood by IPDL:

[RefCounted]

Type T is reference counted (by AddRef/Release). As a parameter to a message or as a type in IPDL structs/unions, it is referenced as a RefPtr<T>.

[MoveOnly]

The type T is treated as uncopyable. When used as a parameter in a message or an IPDL struct/union, it is as an r-value T&&.

Finally, note that using, using class and using struct are all valid syntax. The class and struct keywords are optional.

Namespaces

From the IPDL file:

namespace mozilla {
namespace myns {

    // ... data type and actor definitions ...

}    // namespace myns
}    // namespace mozilla

Namespaces work similar to the way they do in C++. They also mimic the notation, in an attempt to make them comfortable to use. When IPDL actors are compiled into C++ actors, the namespace scoping is carried over. As previously noted, when C++ types are included into IPDL files, the same is true. The most important way in which they differ is that IPDL also uses the namespace to establish the path to the generated files. So, the example defines the IPDL data type mozilla::myns::MyUnion and the actors mozilla::myns::PMyManagerParent and mozilla::myns::PMyManagerChild, which can be included from mozilla/myns/PMyManagerParent.h, mozilla/myns/PMyManagerParent.h and mozilla/myns/PMyManagerChild.h, respectively. The namespace becomes part of the path.

Generating IPDL-Aware C++ Data Types: IPDL Structs and Unions

PMyManager.ipdl and MyTypes.ipdlh define:

[Comparable] union MyUnion {
    float;
    MyOtherData;
};

struct MyActorPair {
    PMyManaged actor1;
    nullable PMyManaged actor2;
};

From these descriptions, IPDL generates C++ classes that approximate the behavior of C++ structs and unions but that come with pre-defined ParamTraits implementations. These objects can also be used as usual outside of IPDL, although the lack of control over the generated code means they are sometimes poorly suited to use as plain data. See ParamTraits for details.

The [Comparable] attribute tells IPDL to generate operator== and operator!= for the new type. In order for it to do that, the fields inside the new type need to define both of those operators.

Finally, the nullable keyword indicates that, when serialized, the actor may be null. It is intended to help users avoid null-object dereference errors. It only applies to actor types and may also be attached to parameters in message declarations.

Defining Actors

The real point of any .ipdl file is that each defines exactly one actor protocol. The definition always matches the .ipdl filename. Repeating the one in PMyManager.ipdl:

[RefCounted] sync protocol PMyManager {
    manages PMyManaged;

    async PMyManaged();
    // ... more message declarations ...
};

The [RefCounted] annotation causes IPDL to generate reference counted C++ classes for these actors. This should be considered the default (Bug 1736371).

Important

A form of reference counting is always used internally by IPDL to make sure that it and its clients never address an actor the other component deleted but this becomes fragile, and sometimes fails, when the client code does not respect the reference count. For example, when IPDL detects that a connection died due to a crashed remote process, deleting the actor could leave dangling pointers, so IPDL cannot delete it. On the other hand, there are many cases where IPDL is the only entity to have references to some actors (this is very common for one side of a managed actor) so IPDL must delete it. If all of those objects were reference counted then there would be no complexity here. Indeed, new actors that are not reference counted should not be approved without a very compelling reason. New non-reference-counted actors may soon be forbidden.

The sync keyword tells IPDL that the actor contains messages that block the sender using sync blocking, so the sending thread waits for a response to the message. There is more on what it and the other blocking modes mean in IPDL messages. For now, just know that this is redundant information whose value is primarily in making it easy for other developers to know that there are sync messages defined here. This list gives preliminary definitions of the options for the actor-blocking policy of messages:

async

Actor may contain only asynchronous messages.

sync

Actor has async capabilities and adds sync messages. sync messages can only be sent from the child actor to the parent.

intr (deprecated)

Actor has sync capabilities and adds intr messages. Some messages can be received while an actor waits for an intr response. This type will be removed soon.

Beyond these protocol blocking strategies, IPDL supports annotations that indicate the actor has messages that may be received in an order other than the one they were sent in. These orderings attempt to handle messages in “message thread” order (as in e.g. mailing lists). These behaviors can be difficult to design for. Their use is discouraged but is sometimes warranted. They will be discussed further in Nested messages.

[NestedUpTo=inside_sync]

Actor has high priority messages that can be handled while waiting for a sync response.

[NestedUpTo=inside_cpow]

Actor has the highest priority messages that can be handled while waiting for a sync response.

The manages clause tells IPDL that PMyManager manages the PMyManaged actor that was previously include d. As with any managed protocol, it must also be the case that PMyManaged.ipdl includes PMyManager and declares that PMyManaged is managed by PMyManager. Recalling the code:

// PMyManaged.ipdl
include protocol PMyManager;
// ...

[RefCounted] protocol PMyManaged {
  manager PMyManager;
  // ...
};

An actor has a manager (e.g. PMyManaged) or else it is a top-level actor (e.g. PMyManager). An actor protocol may be managed by more than one actor type. For example, PMyManaged could have also been managed by some PMyOtherManager not shown here. In that case, manager s are presented in a list, separated by or – e.g. manager PMyManager or PMyOtherManager. Of course, an instance of a managed actor type has only one manager actor (and is therefore managed by only one of the types of manager). The manager of an instance of a managee is always the actor that constructed that managee.

Finally, there is the message declaration async PMyManaged(). This message is a constructor for MyManaged actors; unlike C++ classes, it is found in MyManager. Every manager will need to expose constructors to create its managed types. These constructors are the only way to create an actor that is managed. They can take parameters and return results, like normal messages. The implementation of IPDL constructors are discussed in Actor Lifetimes in C++.

We haven’t discussed a way to construct new top level actors. This is a more advanced topic and is covered separately in Top Level Actors.

Declaring IPDL Messages

The final part of the actor definition is the declaration of messages:

[RefCounted] sync protocol PMyManager {
  // ...
  parent:
    async __delete__(nsString aNote);
    sync SomeMsg(MyActorPair? aActors, MyData[] aMyData)
        returns (int32_t x, int32_t y, MyUnion aUnion);
    async PMyManaged();
  both:
    [Tainted] async AnotherMsg(MyActorEnum aEnum, int32_t aNumber)
        returns (MyOtherData aOtherData);
};

The messages are grouped into blocks by parent:, child: and both:. These labels work the way public: and private: work in C++ – messages after these descriptors are sent/received (only) in the direction specified.

Note

As a mnemonic to remember which direction they indicate, remember to put the word “to” in front of them. So, for example, parent: preceeds __delete__, meaning __delete__ is sent from the child to the parent, and both: states that AnotherMsg can be sent to either endpoint.

IPDL messages support the following annotations:

[Compress]

Indicates repeated messages of this type will consolidate.

[Tainted]

Parameters are required to be validated before using them.

[Priority=Foo]

Priority of MessageTask that runs the C++ message handler. Foo is one of: normal, input, vsync, mediumhigh, or control. See the IPC::Message::PriorityValue enum.

[Nested=inside_sync]

Indicates that the message can sometimes be handled while a sync message waits for a response.

[Nested=inside_cpow]

Indicates that the message can sometimes be handled while a sync message waits for a response.

[Compress] provides crude protection against spamming with a flood of messages. When messages of type M are compressed, the queue of unprocessed messages between actors will never contain an M beside another one; they will always be separated by a message of a different type. This is achieved by throwing out the older of the two messages if sending the new one would break the rule. This has been used to throttle pointer events between the main and content processes.

[Compress=all] is similar but applies whether or not the messages are adjacent in the message queue.

[Tainted] is a C++ mechanism designed to encourage paying attentiton to parameter security. The values of tainted parameters cannot be used until you vouch for their safety. They are discussed in Actors and Messages in C++.

The Nested annotations are deeply related to the message’s blocking policy that follows it and which was briefly discussed in Defining Actors. See Nested messages for details.

The following is a complete list of the available blocking policies. It resembles the list in Defining Actors:

async

Actor may contain only asynchronous messages.

sync

Actor has async capabilities and adds sync messages. sync messages can only be sent from the child actor to the parent.

intr (deprecated)

Actor has sync capabilities and adds intr messages. This type will be removed soon.

The policy defines whether an actor will wait for a response when it sends a certain type of message. A sync actor will wait immediately after sending a sync message, stalling its thread, until a response is received. This is an easy source of browser stalls. It is rarely required that a message be synchronous. New sync messages are therefore required to get approval from an IPC peer. The IPDL compiler will require such messages to be listed in the file sync-messages.ini.

The notion that only child actors can send sync messages was introduced to avoid potential deadlocks. It relies on the belief that a cycle (deadlock) of sync messages is impossible because they all point in one direction. This is no longer the case because any endpoint can be a child or parent and some, like the main process, sometimes serve as both. This means that sync messages should be used with extreme care.

Note

The notion of sync messages flowing in one direction is still the main mechanism IPDL uses to avoid deadlock. New actors should avoid violating this rule as the consequences are severe (and complex). Actors that break these rules should not be approved without extreme extenuating circumstances. If you think you need this, check with the IPC team on Element first (#ipc).

An async actor will not wait. An async response is essentially identical to sending another async message back. It may be handled whenever received messages are handled. The value over an async response message comes in the ergonomics – async responses are usually handled by C++ lambda functions that are more like continuations than methods. This makes them easier to write and to read. Additionally, they allow a response to return message failure, while there would be no such response if we were expecting to send a new async message back, and it failed.

Following synchronization is the name of the message and its parameter list. The message __delete__ stands out as strange – indeed, it terminates the actor’s connection. It does not delete any actor objects itself! It severs the connections of the actor and any actors it manages at both endpoints. An actor will never send or receive any messages after it sends or receives a __delete__. Note that all sends and receives have to happen on a specific worker thread for any actor tree so the send/receive order is well defined. Anything sent after the actor processes __delete__ is ignored (send returns an error, messages yet to be received fail their delivery). In other words, some future operations may fail but no unexpected behavior is possible.

In our example, the child can break the connection by sending __delete__ to the parent. The only thing the parent can do to sever the connection is to fail, such as by crashing. This sort of unidirectional control is both common and desirable.

PMyManaged() is a managed actor constructor. Note the asymmetry – an actor contains its managed actor’s constructors but its own destructor.

The list of parameters to a message is fairly straight-forward. Parameters can be any type that has a C++ ParamTraits specialization and is imported by a directive. That said, there are some surprises in the list of messages:

int32_t,…

The standard primitive types are included. See builtin.py for a list. Pointer types are, unsurprisingly, forbidden.

?

When following a type T, the parameter is translated into Maybe<T> in C++.

[]

When following a type T, the parameter is translated into nsTArray<T> in C++.

Finally, the returns list declares the information sent in response, also as a tuple of typed parameters. As previously mentioned, even async messages can receive responses. A sync message will always wait for a response but an async message will not get one unless it has a returns clause.

This concludes our tour of the IPDL example file. The connection to C++ is discussed in the next chapter; messages in particular are covered in Actors and Messages in C++. For suggestions on best practices when designing your IPDL actor approach, see IPDL Best Practices.

IPDL Syntax Quick Reference

The following is a list of the keywords and operators that have been introduced for use in IPDL files:

include

Include a C++ header (quoted file name) or .ipdlh file (unquoted with no file suffix).

using (class|struct) from

Similar to include but imports only a specific data type.

include protocol

Include another actor for use in management statements, IPDL data types or as parameters to messages.

[RefCounted]

Indicates that the actor’s classes or the imported C++ data types are reference counted. IPDL will respect the ref count. Refcounted imported types require a different ParamTraits interface than non-reference-counted types.

[MoveOnly]

Indicates that an imported C++ data type should not be copied. IPDL code will move it instead.

namespace

Specifies the namespace for IPDL generated code.

union

An IPDL union definition.

struct

An IPDL struct definition.

[Comparable]

Indicates that IPDL should generate operator== and operator!= for the given IPDL struct/union.

nullable

Indicates that an actor reference in an IPDL type may be null when sent over IPC.

protocol

An IPDL protocol (actor) definition.

sync/async

These are used in two cases: (1) to indicate whether a message blocks as it waits for a result and (2) because an actor that contains sync messages must itself be labeled sync or intr.

[NestedUpTo=inside_sync]

Indicates that an actor contains [Nested=inside_sync] messages, in addition to normal messages.

[NestedUpTo=inside_cpow]

Indicates that an actor contains [Nested=inside_cpow] messages, in addition to normal messages.

intr

Used to indicate either that (1) an actor contains sync, async and (deprecated) intr messages, or (2) a message is intr type.

[Nested=inside_sync]

Indicates that the message can be handled while waiting for lower-priority, or in-message-thread, sync responses.

[Nested=inside_cpow]

Indicates that the message can be handled while waiting for lower-priority, or in-message-thread, sync responses. Cannot be sent by the parent actor.

manager

Used in a protocol definition to indicate that this actor manages another one.

manages

Used in a protocol definition to indicate that this actor is managed by another one.

or

Used in a manager clause for actors that have multiple potential managers.

parent: / child: / both:

Indicates direction of subsequent actor messages. As a mnemonic to remember which direction they indicate, put the word “to” in front of them.

returns

Defines return values for messages. All types of message, including async, support returning values.

__delete__

A special message that destroys the related actors at both endpoints when sent. Recv__delete__ and ActorDestroy are called before destroying the actor at the other endpoint, to allow for cleanup.

int32_t,…

The standard primitive types are included.

String

Translated into nsString in C++.

?

When following a type T in an IPDL data structure or message parameter, the parameter is translated into Maybe<T> in C++.

[]

When following a type T in an IPDL data structure or message parameter, the parameter is translated into nsTArray<T> in C++.

[Tainted]

Used to indicate that a message’s handler should receive parameters that it is required to manually validate. Parameters of type T become Tainted<T> in C++.

[Compress]

Indicates repeated messages of this type will consolidate. When two messages of this type are sent and end up side-by-side in the message queue then the older message is discarded (not sent).

[Compress=all]

Like [Compress] but discards the older message regardless of whether they are adjacent in the message queue.

[Priority=Foo]

Priority of MessageTask that runs the C++ message handler. Foo is one of: normal, input, vsync, mediumhigh, or control.

The C++ Interface

ParamTraits

Before discussing how C++ represents actors and messages, we look at how IPDL connects to the imported C++ data types. In order for any C++ type to be (de)serialized, it needs an implementation of the ParamTraits C++ type class. ParamTraits is how your code tells IPDL what bytes to write to serialize your objects for sending, and how to convert those bytes back to objects at the other endpoint. Since ParamTraits need to be reachable by IPDL code, they need to be declared in a C++ header and imported by your protocol file. Failure to do so will result in a build error.

Most basic types and many essential Mozilla types are always available for use without inclusion. An incomplete list includes: C++ primitives, strings (std and mozilla), vectors (std and mozilla), RefPtr<T> (for serializable T), UniquePtr<T>, nsCOMPtr<T>, nsTArray<T>, std::unordered_map<T>, nsresult, etc. See builtin.py, ipc_message_utils.h and IPCMessageUtilsSpecializations.h.

ParamTraits typically bootstrap with the ParamTraits of more basic types, until they hit bedrock (e.g. one of the basic types above). In the most extreme cases, a ParamTraits author may have to resort to designing a binary data format for a type. Both options are available.

We haven’t seen any of this C++ yet. Let’s look at the data types included from MyDataTypes.h:

// MyDataTypes.h
namespace mozilla::myns {
    struct MyData {
        nsCString s;
        uint8_t bytes[17];
        MyData();      // IPDL requires the default constructor to be public
    };

    struct MoveonlyData {
        MoveonlyData();
        MoveonlyData& operator=(const MoveonlyData&) = delete;

        MoveonlyData(MoveonlyData&& m);
        MoveonlyData& operator=(MoveonlyData&& m);
    };

    typedef MoveonlyData MyOtherData;

    class MyUnusedData {
    public:
        NS_INLINE_DECL_REFCOUNTING(MyUnusedData)
        int x;
    };
};

namespace IPC {
    // Basic type
    template<>
    struct ParamTraits<mozilla::myns::MyData> {
        typedef mozilla::myns::MyData paramType;
        static void Write(Message* m, const paramType& in);
        static bool Read(const Message* m, PickleIterator* iter, paramType* out);
    };

    // [MoveOnly] type
    template<>
    struct ParamTraits<mozilla::myns::MyOtherData> {
        typedef mozilla::myns::MyOtherData paramType;
        static void Write(Message* m, const paramType& in);
        static bool Read(const Message* m, PickleIterator* iter, paramType* out);
    };

    // [RefCounted] type
    template<>
    struct ParamTraits<mozilla::myns::MyUnusedData> {
        typedef mozilla::myns::MyUnusedData paramType;
        static void Write(Message* m, paramType* in);
        static bool Read(const Message* m, PickleIterator* iter, RefPtr<paramType>* out);
    };
}

MyData is a struct and MyOtherData is a typedef. IPDL is fine with both. Additionally, MyOtherData is not copyable, matching its IPDL [MoveOnly] annotation.

ParamTraits are required to be defined in the IPC namespace. They must contain a Write method with the proper signature that is used for serialization and a Read method, again with the correct signature, for deserialization.

Here we have three examples of declarations: one for an unannotated type, one for [MoveOnly] and a [RefCounted] one. Notice the difference in the [RefCounted] type’s method signatures. The only difference that may not be clear from the function types is that, in the non-reference-counted case, a default-constructed object is supplied to Read but, in the reference-counted case, Read is given an empty RefPtr<MyUnusedData> and should only allocate a MyUnusedData to return if it so desires.

These are straight-forward implementations of the ParamTraits methods for MyData:

/* static */ void IPC::ParamTraits<MyData>::Write(Message* m, const paramType& in) {
    WriteParam(m, in.s);
    m->WriteBytes(in.bytes, sizeof(in.bytes));
}
/* static */ bool IPC::ParamTraits<MyData>::Read(const Message* m, PickleIterator* iter, paramType* out) {
    return ReadParam(m, iter, &out->s) &&
           m->ReadBytesInto(iter, out->bytes, sizeof(out->bytes));
}

WriteParam and ReadParam call the ParamTraits for the data you pass them, determined using the type of the object as supplied. WriteBytes and ReadBytesInto work on raw, contiguous bytes as expected. Message and PickleIterator are IPDL internal objects; Message hold the incoming/outgoing message as a stream of bytes and the PickleIterator records our current spot in the stream. It is very rare for client code to use them in a manner different than this. Their advanced use is beyond the scope of this document.

Important

Potential failures in Read include everyday C++ failures like out-of-memory conditions, which can be handled as usual. But Read can also fail due to things like data validation errors. ParamTraits read data that is considered insecure. It is important that they catch corruption and properly handle it. Returning false from Read will usually result in crashing the process (everywhere except in the main process). This is the right behavior as the browser would be in an unexpected state, even if the serialization failure was not malicious (since it cannot process the message). Other responses, such as failing with a crashing assertion, are inferior. IPDL fuzzing relies on ParamTraits not crashing due to corruption failures. Occasionally, validation will require access to state that ParamTraits can’t easily reach. (Only) in those cases, validation can be reasonably done in the message handler. Such cases are a good use of the Tainted annotation. See Actors and Messages in C++ for more.

IPC::ParamTraits<T> is a convenience class used by IPDLParamTraits<T> here. You can choose to specialize mozilla::ipc::IPDLParamTraits<T> instead of IPC::ParamTraits<T> if you need the actor object itself during serialization or deserialization. The only difference between the two is that IPDLParamTraits<T> methods are given the actor. So we could have written serialization for MyData this way:

namespace mozilla::ipc {
    // Basic type
    template<>
    struct IPDLParamTraits<mozilla::myns::MyData> {
        typedef mozilla::myns::MyData paramType;
        static void Write(Message* m, IProtocol* aActor, const paramType& in);
        static bool Read(const Message* m, PickleIterator* iter, IProtocol* aActor, paramType* out);
    };

    /* static */ void IPDLParamTraits<mozilla::myns::MyData>::Write(
            Message* m, IProtocol* aActor, const paramType& in) {
        WriteParam(m, in.s);
        m->WriteBytes(in.bytes, sizeof(in.bytes));
    }
    /* static */ bool IPDLParamTraits<mozilla::myns::MyData>::Read(
            const Message* m, PickleIterator* iter, IProtocol* aActor, paramType* out) {
        return ReadParam(m, iter, &out->s) &&
               m->ReadBytesInto(iter, out->bytes, sizeof(out->bytes));
    }
}

We didn’t do this because we didn’t need the IProtocol for our case. It is rare that the actor is useful in serialization.

A special case worth mentioning is that of enums. Enums are a common source of security holes since code is rarely safe with enum values that are not valid. Since data obtained through IPDL messages should be considered tainted, enums are of principal concern. ContiguousEnumSerializer and ContiguousEnumSerializerInclusive safely implement ParamTraits for enums that are only valid for a contiguous set of values, which is most of them. The generated ParamTraits confirm that the enum is in valid range; Read will return false otherwise. As an example, here is the MyActorEnum included from MyActorUtils.h:

enum MyActorEnum { e1, e2, e3, e4, e5 };

template<>
struct ParamTraits<MyActorEnum>
  : public ContiguousEnumSerializerInclusive<MyActorEnum, MyActorEnum::e1, MyActorEnum::e5> {};

IPDL Structs and Unions in C++

IPDL structs and unions become C++ classes that provide interfaces that are fairly self-explanatory. Recalling MyUnion and MyActorPair from IPDL Structs and Unions :

union MyUnion {
    float;
    MyOtherData;
};

struct MyActorPair {
    PMyManaged actor1;
    nullable PMyManaged actor2;
};

These compile to:

class MyUnion {
    enum Type { Tfloat, TMyOtherData };
    Type type();
    MyUnion(float f);
    MyUnion(MyOtherData&& aOD);
    MyUnion& operator=(float f);
    MyUnion& operator=(MyOtherData&& aOD);
    operator float&();
    operator MyOtherData&();
};

class MyActorPair {
    MyActorPair(PMyManagedParent* actor1Parent, PMyManagedChild* actor1Child,
                PMyManagedParent* actor2Parent, PMyManagedChild* actor2Child);
    // Exactly one of { actor1Parent(), actor1Child() } must be non-null.
    PMyManagedParent*& actor1Parent();
    PMyManagedChild*& actor1Child();
    // As nullable, zero or one of { actor2Parent(), actor2Child() } will be non-null.
    PMyManagedParent*& actor2Parent();
    PMyManagedChild*& actor2Child();
}

The generated ParamTraits use the ParamTraits for the types referenced by the IPDL struct or union. Fields respect any annotations for their type (see IPDL Includes). For example, a [RefCounted] type T generates RefPtr<T> fields.

Note that actor members result in members of both the parent and child actor types, as seen in MyActorPair. When actors are used to bridge processes, only one of those could ever be used at a given endpoint. IPDL makes sure that, when you send one type (say, PMyManagedChild), the adjacent actor of the other type (PMyManagedParent) is received. This is not only true for message parameters and IPDL structs/unions but also for custom ParamTraits implementations. If you Write a PFooParent* then you must Read a PFooChild*. This is hard to confuse in message handlers since they are members of a class named for the side they operate on, but this cannot be enforced by the compiler. If you are writing MyManagerParent::RecvSomeMsg(Maybe<MyActorPair>&& aActors, nsTArray<MyData>&& aMyData) then the actor1Child and actor2Child fields cannot be valid since the child (usually) exists in another process.

Actors and Messages in C++

As mentioned in Using The IPDL compiler, the IPDL compiler generates two header files for the protocol PMyManager: PMyManagerParent.h and PMyManagerChild.h, which declare the actor’s base classes. There, we discussed how the headers are visible to C++ components that include chromium-config.mozbuild. We, in turn, always need to define two files that declare our actor implementation subclasses (MyManagerParent.h and MyManagerChild.h). The IPDL file looked like this:

So MyManagerParent.h looks like this:

#include "PMyManagerParent.h"

namespace mozilla {
namespace myns {

class MyManagerParent : public PMyManagerParent {
    NS_INLINE_DECL_REFCOUNTING(MyManagerParent, override)
protected:
    IPCResult Recv__delete__(const nsString& aNote);
    IPCResult RecvSomeMsg(const Maybe<MyActorPair>& aActors, const nsTArray<MyData>& aMyData,
                          int32_t* x, int32_t* y, MyUnion* aUnion);
    IPCResult RecvAnotherMsg(const Tainted<MyActorEnum>& aEnum, const Tainted<int32_t>& aNumber,
                             AnotherMsgResolver&& aResolver);

    already_AddRefed<PMyManagerParent> AllocPMyManagedParent();
    IPCResult RecvPMyManagedConstructor(PMyManagedConstructor* aActor);

    // ... etc ...
};

} // namespace myns
} // namespace mozilla

All messages that can be sent to the actor must be handled by Recv methods in the proper actor subclass. They should return IPC_OK() on success and IPC_FAIL(actor, reason) if an error occurred (where actor is this and reason is a human text explanation) that should be considered a failure to process the message. The handling of such a failure is specific to the process type.

Recv methods are called by IPDL by enqueueing a task to run them on the MessageLoop for the thread on which they are bound. This thread is the actor’s worker thread. All actors in a managed actor tree have the same worker thread – in other words, actors inherit the worker thread from their managers. Top level actors establish their worker thread when they are bound. More information on threads can be found in Top Level Actors. For the most part, client code will never engage with an IPDL actor outside of its worker thread.

Received parameters become stack variables that are std::move-d into the Recv method. They can be received as a const l-value reference, rvalue-reference, or by value (type-permitting). [MoveOnly] types should not be received as const l-values. Return values for sync messages are assigned by writing to non-const (pointer) parameters. Return values for async messages are handled differently – they are passed to a resolver function. In our example, AnotherMsgResolver would be a std::function<> and aResolver would be given the value to return by passing it a reference to a MyOtherData object.

MyManagerParent is also capable of sending an async message that returns a value: AnotherMsg. This is done with SendAnotherMsg, which is defined automatically by IPDL in the base class PMyManagerParent. There are two signatures for Send and they look like this:

// Return a Promise that IPDL will resolve with the response or reject.
RefPtr<MozPromise<MyOtherData, ResponseRejectReason, true>>
SendAnotherMsg(const MyActorEnum& aEnum, int32_t aNumber);

// Provide callbacks to process response / reject.  The callbacks are just
// std::functions.
void SendAnotherMsg(const MyActorEnum& aEnum, int32_t aNumber,
    ResolveCallback<MyOtherData>&& aResolve, RejectCallback&& aReject);

The response is usually handled by lambda functions defined at the site of the Send call, either by attaching them to the returned promise with e.g. MozPromise::Then, or by passing them as callback parameters. See docs on MozPromise for more on its use. The promise itself is either resolved or rejected by IPDL when a valid reply is received or when the endpoint determines that the communication failed. ResponseRejectReason is an enum IPDL provides to explain failures.

Additionally, the AnotherMsg handler has Tainted parameters, as a result of the [Tainted] annotation in the protocol file. Recall that Tainted is used to force explicit validation of parameters in the message handler before their values can be used (as opposed to validation in ParamTraits). They therefore have access to any state that the message handler does. Their APIs, along with a list of macros that are used to validate them, are detailed here.

Send methods that are not for async messages with return values follow a simpler form; they return a bool indicating success or failure and return response values in non-const parameters, as the Recv methods do. For example, PMyManagerChild defines this to send the sync message SomeMsg:

// generated in PMyManagerChild
bool SendSomeMsg(const Maybe<MyActorPair>& aActors, const nsTArray<MyData>& aMyData,
                 int32_t& x, int32_t& y, MyUnion& aUnion);

Since it is sync, this method will not return to its caller until the response is received or an error is detected.

All calls to Send methods, like all messages handler Recv methods, must only be called on the worker thread for the actor.

Constructors, like the one for MyManaged, are clearly an exception to these rules. They are discussed in the next section.

Actor Lifetimes in C++

The constructor message for MyManaged becomes two methods at the receiving end. AllocPMyManagedParent constructs the managed actor, then RecvPMyManagedConstructor is called to update the new actor. The following diagram shows the construction of the MyManaged actor pair:

%%{init: {'sequence': {'boxMargin': 4, 'actorMargin': 10} }}%% sequenceDiagram participant d as Driver participant mgdc as MyManagedChild participant mgrc as MyManagerChild participant ipc as IPC Child/Parent participant mgrp as MyManagerParent participant mgdp as MyManagedParent d->>mgdc: new mgdc->>d: [mgd_child] d->>mgrc: SendPMyManagedConstructor<br/>[mgd_child, params] mgrc->>ipc: Form actor pair<br/>[mgd_child, params] par mgdc->>ipc: early PMyManaged messages and ipc->>mgrp: AllocPMyManagedParent<br/>[params] mgrp->>mgdp: new mgdp->>mgrp: [mgd_parent] ipc->>mgrp: RecvPMyManagedConstructor<br/>[mgd_parent, params] mgrp->>mgdp: initialization ipc->>mgdp: early PMyManaged messages end Note over mgdc,mgdp: Bi-directional sending and receiving will now happen concurrently.

A MyManaged actor pair being created by some Driver object. Internal IPC objects in the parent and child processes are combined for compactness. Connected par blocks run concurrently. This shows that messages can be safely sent while the parent is still being constructed.

The next diagram shows the destruction of the MyManaged actor pair, as initiated by a call to Send__delete__. __delete__ is sent from the child process because that is the only side that can call it, as declared in the IPDL protocol file.

%%{init: {'sequence': {'boxMargin': 4, 'actorMargin': 10} }}%% sequenceDiagram participant d as Driver participant mgdc as MyManagedChild participant ipc as IPC Child/Parent participant mgdp as MyManagedParent d->>mgdc: Send__delete__ mgdc->>ipc: Disconnect<br/>actor pair par ipc->>mgdc: ActorDestroy ipc->>mgdc: Release and ipc->>mgdp: Recv__delete__ ipc->>mgdp: ActorDestroy ipc->>mgdp: Release end

A MyManaged actor pair being disconnected due to some Driver object in the child process sending __delete__.

Finally, let’s take a look at the behavior of an actor whose peer has been lost (e.g. due to a crashed process).

%%{init: {'sequence': {'boxMargin': 4, 'actorMargin': 10} }}%% sequenceDiagram participant mgdc as MyManagedChild participant ipc as IPC Child/Parent participant mgdp as MyManagedParent Note over mgdc: CRASH!!! ipc->>ipc: Notice fatal error. ipc->>mgdp: ActorDestroy ipc->>mgdp: Release

A MyManaged actor pair being disconnected when its peer is lost due to a fatal error. Note that Recv__delete__ is not called.

The Alloc and Recv...Constructor methods are somewhat mirrored by Recv__delete__ and ActorDestroy but there are a few differences. First, the Alloc method really does create the actor but the ActorDestroy method does not delete it. Additionally, ActorDestroy is run at both endpoints, during Send__delete__ or after Recv__delete__. Finally and most importantly, Recv__delete__ is only called if the __delete__ message is received but it may not be if, for example, the remote process crashes. ActorDestroy, on the other hand, is guaranteed to run for every actor unless the process terminates uncleanly. For this reason, ActorDestroy is the right place for most actor shutdown code. Recv__delete__ is rarely useful, although it is occasionally beneficial to have it receive some final data.

The relevant part of the parent class looks like this:

class MyManagerParent : public PMyManagerParent {
    already_AddRefed<PMyManagerParent> AllocPMyManagedParent();
    IPCResult RecvPMyManagedConstructor(PMyManagedConstructor* aActor);

    IPCResult Recv__delete__(const nsString& aNote);
    void ActorDestroy(ActorDestroyReason why);

    // ... etc ...
};

The Alloc method is required for managed actors that are constructed by IPDL receiving a Send message. It is not required for the actor at the endpoint that calls Send. The Recv...Constructor message is not required – it has a base implementation that does nothing.

If the constructor message has parameters, they are sent to both methods. Parameters are given to the Alloc method by const reference but are moved into the Recv method. They differ in that messages can be sent from the Recv method but, in Alloc, the newly created actor is not yet operational.

The Send method for a constructor is similarly different from other Send methods. In the child actor, ours looks like this:

IPCResult SendPMyManagedConstructor(PMyManagedChild* aActor);

The method expects a PMyManagedChild that the caller will have constructed, presumably using new (this is why it does not require an Alloc method). Once Send...Constructor is called, the actor can be used to send and receive messages. It does not matter that the remote actor may not have been created yet due to asynchronicity.

The destruction of actors is as unusual as their construction. Unlike construction, it is the same for managed and top-level actors. The use of [RefCounted] actors removes a lot of the complexity but there is still a process to understand. Actor destruction begins when an __delete__ message is sent. In PMyManager, this message is declared from child to parent. The actor calling Send__delete__ is no longer connected to anything when the method returns. Future calls to Send return an error and no future messages will be received. This is also the case for an actor that has run Recv__delete__; it is no longer connected to the other endpoint.

Note

Since Send__delete__ may release the final reference to itself, it cannot safely be a class instance method. Instead, unlike other Send methods, it’s a static class method and takes the actor as a parameter:

static IPCResult Send__delete__(PMyManagerChild* aToDelete);

Additionally, the __delete__ message tells IPDL to disconnect both the given actor and all of its managed actors. So it is really deleting the actor subtree, although Recv__delete__ is only called for the actor it was sent to.

During the call to Send__delete__, or after the call to Recv__delete__, the actor’s ActorDestroy method is called. This method gives client code a chance to do any teardown that must happen in all circumstances were it is possible – both expected and unexpected. This means that ActorDestroy will also be called when, for example, IPDL detects that the other endpoint has terminated unexpectedly, so it is releasing its reference to the actor, or because an ancestral manager (manager or manager’s manager…) received a __delete__. The only way for an actor to avoid ActorDestroy is for its process to crash first. ActorDestroy is always run after its actor is disconnected so it is pointless to try to send messages from it.

Why use ActorDestroy instead of the actor’s destructor? ActorDestroy gives a chance to clean up things that are only used for communication and therefore don’t need to live for however long the actor’s (reference-counted) object does. For example, you might have references to shared memory (Shmems) that are no longer valid. Or perhaps the actor can now release a cache of data that was only needed for processing messages. It is cleaner to deal with communication-related objects in ActorDestroy, where they become invalid, than to leave them in limbo until the destructor is run.

Consider actors to be like normal reference-counted objects, but where IPDL holds a reference while the connection will or does exist. One common architecture has IPDL holding the only reference to an actor. This is common with actors created by sending construtor messages but the idea is available to any actor. That only reference is then released when the __delete__ message is sent or received.

The dual of IPDL holding the only reference is to have client code hold the only reference. A common pattern to achieve this has been to override the actor’s AddRef to have it send __delete__ only when it’s count is down to one reference (which must be IPDL if actor.CanSend() is true). A better approach would be to create a reference-counted delegate for your actor that can send __delete__ from its destructor. IPDL does not guarantee that it will not hold more than one reference to your actor.

Top Level Actors

Recall that top level actors are actors that have no manager. They are at the root of every actor tree. There are two settings in which we use top-level actors that differ pretty dramatically. The first type are top-level actors that are created and maintained in a way that resembles managed actors, but with some important differences we will cover in this section. The second type of top-level actors are the very first actors in a new process – these actors are created through different means and closing them (usually) terminates the process. The new process example demonstrates both of these. It is discussed in detail in Adding a New Type of Process.

Value of Top Level Actors

Top-level actors are harder to create and destroy than normal actors. They used to be more heavyweight than managed actors but this has recently been dramatically reduced.

Note

Top-level actors previously required a dedicated message channel, which are limited OS resources. This is no longer the case – message channels are now shared by actors that connect the same two processes. This message interleaving can affect message delivery latency but profiling suggests that the change was basically inconsequential.

So why use a new top level actor?

  • The most dramatic property distinguishing top-level actors is the ability to bind to whatever EventTarget they choose. This means that any thread that runs a MessageLoop can use the event target for that loop as the place to send incoming messages. In other words, Recv methods would be run by that message loop, on that thread. The IPDL apparatus will asynchronously dispatch messages to these event targets, meaning that multiple threads can be handling incoming messages at the same time. The PBackground approach was born of a desire to make it easier to exploit this, although it has some complications, detailed in that section, that limit its value.

  • Top level actors suggest modularity. Actor protocols are tough to debug, as is just about anything that spans process boundaries. Modularity can give other developers a clue as to what they need to know (and what they don’t) when reading an actor’s code. The alternative is proverbial dumpster classes that are as critical to operations (because they do so much) as they are difficult to learn (because they do so much).

  • Top level actors are required to connect two processes, regardless of whether the actors are the first in the process or not. As said above, the first actor is created through special means but other actors are created through messages. In Gecko, apart from the launcher and main processes, all new processes X are created with their first actor being between X and the main process. To create a connection between X and, say, a content process, the main process has to send connected Endpoints to X and to the content process, which in turn use those endpoints to create new top level actors that form an actor pair. This is discussed at length in Connecting With Other Processes.

Top-level actors are not as frictionless as desired but they are probably under-utilized relative to their power. In cases where it is supported, PBackground is sometimes a simpler alternative to achieve the same goals.

Creating Top Level Actors From Other Actors

The most common way to create new top level actors is by creating a pair of connected Endpoints and sending one to the other actor. This is done exactly the way it sounds. For example:

bool MyPreexistingActorParent::MakeMyActor() {
    Endpoint<PMyActorParent> parentEnd;
    Endpoint<PMyActorChild> childEnd;
    if (NS_WARN_IF(NS_FAILED(PMyActor::CreateEndpoints(
          base::GetCurrentProcId(), OtherPid(), &parentEnd, &childEnd)))) {
        // ... handle failure ...
        return false;
    }
    RefPtr<MyActorParent> parent = new MyActorParent;
    if (!parentEnd.Bind(parent)) {
        // ... handle failure ...
        delete parent;
        return false;
    }
    // Do this second so we skip child if parent failed to connect properly.
    if (!SendCreateMyActorChild(std::move(childEnd))) {
        // ... assume an IPDL error will destroy parent.  Handle failure beyond that ...
        return false;
    }
    return true;
}

Here MyPreexistingActorParent is used to send a child endpoint for the new top level actor to MyPreexistingActorChild, after it hooks up the parent end. In this example, we bind our new actor to the same thread we are running on – which must be the same thread MyPreexistingActorParent is bound to since we are sending CreateMyActorChild from it. We could have bound on a different thread.

At this point, messages can be sent on the parent. Eventually, it will start receiving them as well.

MyPreexistingActorChild still has to receive the create message. The code for that handler is pretty similar:

IPCResult MyPreexistingActorChild::RecvCreateMyActorChild(Endpoint<PMyActorChild>&& childEnd) {
    RefPtr<MyActorChild> child = new MyActorChild;
    if (!childEnd.Bind(child)) {
        // ... handle failure and return ok, assuming a related IPDL error will alert the other side to failure ...
        return IPC_OK();
    }
    return IPC_OK();
}

Like the parent, the child is ready to send as soon as Bind is complete. It will start receiving messages soon afterward on the event target for the thread on which it is bound.

Creating First Top Level Actors

The first actor in a process is an advanced topic that is covered in the documentation for adding a new process.

PBackground

Developed as a convenient alternative to top level actors, PBackground is an IPDL protocol whose managees choose their worker threads in the child process and share a thread dedicated solely to them in the parent process. When an actor (parent or child) should run without hogging the main thread, making that actor a managee of PBackground (aka a background actor) is an option.

Warning

Background actors can be difficult to use correctly, as spelled out in this section. It is recommended that other options – namely, top-level actors – be adopted instead.

Background actors can only be used in limited circumstances:

  • PBackground only supports the following process connections (where ordering is parent <-> child): main <-> main, main <-> content, main <-> socket and socket <-> content.

Important

Socket process PBackground actor support was added after the other options. It has some rough edges that aren’t easy to anticipate. In the future, their support may be broken out into a different actor or removed altogether. You are strongly encouraged to use new Top Level Actors instead of PBackground actor when communicating with socket process worker threads.

  • Background actor creation is always initiated by the child. Of course, a request to create one can be sent to the child by any other means.

  • All parent background actors run in the same thread. This thread is dedicated to serving as the worker for parent background actors. While it has no other functions, it should remain responsive to all connected background actors. For this reason, it is a bad idea to conduct long operations in parent background actors. For such cases, create a top level actor and an independent thread on the parent side instead.

  • Background actors are currently not reference-counted. IPDL’s ownership has to be carefully respected and the (de-)allocators for the new actors have to be defined. See The Old Ways for details.

A hypothetical layout of PBackground threads, demonstrating some of the process-type limitations, is shown in the diagram below.

flowchart LR subgraph content #1 direction TB c1tm[main] c1t1[worker #1] c1t2[worker #2] c1t3[worker #3] end subgraph content #2 direction TB c2tm[main] c2t1[worker #1] c2t2[worker #2] end subgraph socket direction TB stm[main] st1[background parent /\nworker #1] st2[worker #2] end subgraph main direction TB mtm[main] mt1[background parent] end %% PBackground connections c1tm --> mt1 c1t1 --> mt1 c1t2 --> mt1 c1t3 --> mt1 c1t3 --> st1 c2t1 --> st1 c2t1 --> mt1 c2t2 --> mt1 c2tm --> st1 stm --> mt1 st1 --> mt1 st2 --> mt1

Hypothetical PBackground thread setup. Arrow direction indicates child-to-parent PBackground-managee relationships. Parents always share a thread and may be connected to multiple processes. Child threads can be any thread, including main.

Creating background actors is done a bit differently than normal managees. The new managed type and constructor are still added to PBackground.ipdl as with normal managees but, instead of new-ing the child actor and then passing it in a SendFooConstructor call, background actors issue the send call to the BackgroundChild manager, which returns the new child:

// Bind our new PMyBackgroundActorChild to the current thread.
PBackgroundChild* bc = BackgroundChild::GetOrCreateForCurrentThread();
if (!bc) {
    return false;
}
PMyBackgroundActorChild* pmyBac = bac->SendMyBackgroundActor(constructorParameters);
if (!pmyBac) {
    return false;
}
auto myBac = static_cast<MyBackgroundActorChild*>(pmyBac);

Note

PBackgroundParent still needs a RecvMyBackgroundActorConstructor handler, as usual. This must be done in the ParentImpl class. ParentImpl is the non-standard name used for the implementation of PBackgroundParent.

To summarize, PBackground attempts to simplify a common desire in Gecko: to run tasks that communicate between the main and content processes but avoid having much to do with the main thread of either. Unfortunately, it can be complicated to use correctly and has missed on some favorable IPDL improvements, like reference counting. While top level actors are always a complete option for independent jobs that need a lot of resources, PBackground offers a compromise for some cases.

IPDL Best Practices

IPC performance is affected by a lot of factors. Many of them are out of our control, like the influence of the system thread scheduler on latency or messages whose travel internally requires multiple legs for security reasons. On the other hand, some things we can and should control for:

  • Messages incur inherent performance overhead for a number of reasons: IPDL internal thread latency (e.g. the I/O thread), parameter (de-)serialization, etc. While not usually dramatic, this cost can add up. What’s more, each message generates a fair amound of C++ code. For these reasons, it is wise to reduce the number of messages being sent as far as is reasonable. This can be as simple as consolidating two asynchronous messages that are always in succession. Or it can be more complex, like consolidating two somewhat-overlapping messages by merging their parameter lists and marking parameters that may not be needed as optional. It is easy to go too far down this path but careful message optimization can show big gains.

  • Even [moveonly] parameters are “copied” in the sense that they are serialized. The pipes that transmit the data are limited in size and require allocation. So understand that the performance of your transmission will be inversely proportional to the size of your content. Filter out data you won’t need. For complex reasons related to Linux pipe write atomicity, it is highly desirable to keep message sizes below 4K (including a small amount for message metadata).

  • On the flip side, very large messages are not permitted by IPDL and will result in a runtime error. The limit is currently 256M but message failures frequently arise even with slightly smaller messages.

  • Parameters to messages are C++ types and therefore can be very complex in the sense that they generally represent a tree (or graph) of objects. If this tree has a lot of objects in it, and each of them is serialized by ParamTraits, then we will find that serialization is allocating and constructing a lot of objects, which will stress the allocator and cause memory fragmentation. Avoid this by using larger objects or by sharing this kind of data through careful use of shared memory.

  • As it is with everything, concurrency is critical to the performance of IPDL. For actors, this mostly manifests in the choice of bound thread. While adding a managed actor to an existing actor tree may be a quick implementation, this new actor will be bound to the same thread as the old one. This contention may be undesirable. Other times it may be necessary since message handlers may need to use data that isn’t thread safe or may need a guarantee that the two actors’ messages are received in order. Plan up front for your actor hierarchy and its thread model. Recognize when you are better off with a new top level actor or PBackground managee that facilitates processing messages simultaneously.

  • Remember that latency will slow your entire thread, including any other actors/messages on that thread. If you have messages that will need a long time to be processed but can run concurrently then they should use actors that run on a separate thread.

  • Top-level actors decide a lot of properties for their managees. Probably the most important are the process layout of the actor (including which process is “Parent” and which is “Child”) and the thread. Every top-level actor should clearly document this, ideally in their .ipdl file.

The Old Ways

TODO:

The FUD

TODO:

The Rest

Nested messages

The Nested message annotations indicate the nesting type of the message. They attempt to process messages in the nested order of the “conversation thread”, as found in e.g. a mailing-list client. This is an advanced concept that should be considered to be discouraged, legacy functionality. Essentially, Nested messages can make other sync messages break the policy of blocking their thread – nested messages are allowed to be received while a sync messagee is waiting for a response. The rules for when a nested message can be handled are somewhat complex but they try to safely allow a sync message M to handle and respond to some special (nested) messages that may be needed for the other endpoint to finish processing M. There is a comment in MessageChannel with info on how the decision to handle nested messages is made. For sync nested messages, note that this implies a relay between the endpoints, which could dramatically affect their throughput.

Declaring messages to nest requires an annotation on the actor and one on the message itself. The nesting annotations were listed in Defining Actors and Declaring IPDL Messages. We repeat them here. The actor annotations specify the maximum priority level of messages in the actor. It is validated by the IPDL compiler. The annotations are:

[NestedUpTo=inside_sync]

Indicates that an actor contains messages of priority [Nested=inside_sync] or lower.

[NestedUpTo=inside_cpow]

Indicates that an actor contains messages of priority [Nested=inside_cpow] or lower.

Note

The order of the nesting priorities is: (no nesting priority) < inside_sync < inside_cpow.

The message annotations are:

[Nested=inside_sync]

Indicates that the message can be handled while waiting for lower-priority, or in-message-thread, sync responses.

[Nested=inside_cpow]

Indicates that the message can be handled while waiting for lower-priority, or in-message-thread, sync responses. Cannot be sent by the parent actor.

Note

[Nested=inside_sync] messages must be sync (this is enforced by the IPDL compiler) but [Nested=inside_cpow] may be async.

Nested messages are obviously only interesting when sent to an actor that is performing a synchronous wait. Therefore, we will assume we are in such a state. Say actorX is waiting for a sync reply from actorY for message m1 when actorY sends actorX a message m2. We distinguish two cases here: (1) when m2 is sent while processing m1 (so m2 is sent by the RecvM1() method – this is what we mean when we say “nested”) and (2) when m2 is unrelated to m1. Case (2) is easy; m2 is only dispatched while m1 waits if priority(m2) > priority(m1) > (no priority) and the message is being received by the parent, or if priority(m2) >= priority(m1) > (no priority) and the message is being received by the child. Case (1) is less straightforward.

To analyze case (1), we again distinguish the two possible ways we can end up in the nested case: (A) m1 is sent by the parent to the child and m2 is sent by the child to the parent, or (B) where the directions are reversed. The following tables explain what happens in all cases:

Case (A): Child sends message to a parent that is awaiting a sync response

sync m1 type (from parent)

m2 type (from child)

m2 handled or rejected

sync (no priority)

*

IPDL compiler error: parent cannot send sync (no priority)

sync inside_sync

async (no priority)

m2 delayed until after m1 completes
Currently m2 is handled during the sync wait (bug?)

sync inside_sync

sync (no priority)

m2 send fails: lower priority than m1
Currently m2 is handled during the sync wait (bug?)

sync inside_sync

sync inside_sync

m2 handled during m1 sync wait: same message thread and same priority

sync inside_sync

async inside_cpow

m2 handled during m1 sync wait: higher priority

sync inside_sync

sync inside_cpow

m2 handled during m1 sync wait: higher priority

sync inside_cpow

*

IPDL compiler error: parent cannot use inside_cpow priority

Case (B): Parent sends message to a child that is awaiting a sync response

sync m1 type (from child)

m2 type (from parent)

m2 handled or rejected

*

async (no priority)

m2 delayed until after m1 completes

*

sync (no priority)

IPDL compiler error: parent cannot send sync (no priority)

sync (no priority)

sync inside_sync

m2 send fails: no-priority sync messages cannot handle incoming messages during wait

sync inside_sync

sync inside_sync

m2 handled during m1 sync wait: same message thread and same priority

sync inside_cpow

sync inside_sync

m2 send fails: lower priority than m1

*

async inside_cpow

IPDL compiler error: parent cannot use inside_cpow priority

*

sync inside_cpow

IPDL compiler error: parent cannot use inside_cpow priority

We haven’t seen rule #2 from the comment in MessageChannel in action but, as the comment mentions, it is needed to break deadlocks in cases where both the parent and child are initiating message-threads simultaneously. It accomplishes this by favoring the parent’s sent messages over the child’s when deciding which message-thread to pursue first (and blocks the other until the first completes). Since this distinction is entirely thread-timing based, client code needs only to be aware that IPDL internals will not deadlock because of this type of race, and that this protection is limited to a single actor tree – the parent/child messages are only well-ordered when under the same top-level actor so simultaneous sync messages across trees are still capable of deadlock.

Clearly, tight control over these types of protocols is required to predict how they will coordinate within themselves and with the rest of the application objects. Control flow, and hence state, can be very difficult to predict and are just as hard to maintain. This is one of the key reasons why we have stressed that message priorities should be avoided whenever possible.

Message Logging

The environment variable MOZ_IPC_MESSAGE_LOG controls the logging of IPC messages. It logs details about the transmission and reception of messages. This isn’t controlled by MOZ_LOG – it is a separate system. Set this variable to 1 to log information on all IPDL messages, or specify a comma-separated list of top-level protocols to log (e.g. MOZ_IPC_MESSAGE_LOG="PMyManagerChild,PMyManagedParent,PMyManagedChild"). Debugging with IPDL Logging has an example where IPDL logging is useful in tracking down a bug.

Important

The preceeding P and the Parent or Child suffix are required when listing individual protocols in MOZ_IPC_MESSAGE_LOG.