Boost C++ Libraries

To use Beast effectively, a prior understanding of Networking is required. This section reviews these concepts as a reminder and guide for further learning.

A network allows programs located anywhere to exchange information after opting-in to communications by establishing a connection. Data may be reliably transferred across a connection in both directions (full-duplex) with bytes arriving in the same order they were sent. These connections, along with the objects and types used to represent them, are collectively termed streams. The computer or device attached to the network is called a host, and the program on the other end of an established connection is called a peer.

The internet is a global network of interconnected computers that use a variety of standardized communication protocols to exchange information. The most popular protocol is TCP/IP, which this library relies on exclusively. The protocol takes care of the low level details so that applications see a stream, which is the reliable, full-duplex connection carrying the ordered set of bytes described above. A server is a powerful, always-on host at a well-known network name or network address which provides data services. A client is a transient peer which connects to a server to exchange data, and goes offline.

A vendor supplies a program called a device driver, enabling networking hardware such as an ethernet adaptor to talk to the operating system. This in turn permits running programs to interact with networking using various flavors of interfaces such as Berkeley sockets or Windows Sockets 2 ("Winsock").

Networking in C++, represented by Boost.Asio, Asio, and Networking TS, provides a layer of abstraction to interact portably with the operating system facilities for not just networking but general input/output ("I/O").

Buffers

A buffer holds a contiguous sequence of bytes used when performing I/O. The types net::const_buffer and net::mutable_buffer represent these memory regions as type-safe pointer/size pairs:

net::const_buffer cb("Hello, world!", 13);
assert(string_view(reinterpret_cast<char const*>(
    cb.data()), cb.size()) == "Hello, world!");

char storage[13];
net::mutable_buffer mb(storage, sizeof(storage));
std::memcpy(mb.data(), cb.data(), mb.size());
assert(string_view(reinterpret_cast<char const*>(
    mb.data()), mb.size()) == "Hello, world!");

The concepts ConstBufferSequence and MutableBufferSequence describe bidirectional ranges whose value type is convertible to const_buffer and mutable_buffer respectively. These sequences allow transacting with multiple buffers in a single function call, a technique called scatter/gather I/O. Buffers and buffer sequences are non-owning; copies produce shallow references and not duplicates of the underlying memory. Each of these statements declares a buffer sequence:

net::const_buffer b1;                   // a ConstBufferSequence by definition
net::mutable_buffer b2;                 // a MutableBufferSequence by definition
std::array<net::const_buffer, 3> b3;    // A ConstBufferSequence by named requirements

The functions net::buffer_size and net::buffer_copy determine the total number of bytes in a buffer sequence, and transfer some or all of bytes from one buffer sequence to another respectively. The function buffer_size is a customization point: user defined overloads in foreign namespaces are possible, and callers should invoke buffer_size without namespace qualification. The functions net::buffer_sequence_begin and net::buffer_sequence_end are used to obtain a pair of iterators for traversing the sequence. Beast provides a set of buffer sequence types and algorithms such as buffers_cat, buffers_front, buffers_prefix, buffers_range, and buffers_suffix. This example returns the bytes in a buffer sequence as a string:

template <class ConstBufferSequence>
std::string string_from_buffers (ConstBufferSequence const& buffers)
{
    // check that the type meets the requirements using the provided type traits
    static_assert(
        net::is_const_buffer_sequence<ConstBufferSequence>::value,
        "ConstBufferSequence type requirements not met");

    // optimization: reserve all the space for the string first
    std::string result;
    result.reserve(beast::buffer_bytes(buffers));        // beast version of net::buffer_size

    // iterate over each buffer in the sequence and append it to the string
    for(auto it = net::buffer_sequence_begin(buffers);  // returns an iterator to beginning of the sequence
        it != net::buffer_sequence_end(buffers);)       // returns a past-the-end iterator to the sequence
    {
        // A buffer sequence iterator's value_type is always convertible to net::const_buffer
        net::const_buffer buffer = *it++;

        // A cast is always required to out-out of type-safety
        result.append(static_cast<char const*>(buffer.data()), buffer.size());
    }
    return result;
}

The DynamicBuffer concept defines a resizable buffer sequence interface. Algorithms may be expressed in terms of dynamic buffers when the memory requirements are not known ahead of time, for example when reading an HTTP message from a stream. Beast provides a well-rounded collection of dynamic buffer types such as buffers_adaptor, flat_buffer, multi_buffer, and static_buffer. The following function reads data from a tcp_stream into a dynamic buffer until it encountering a newline character, using net::buffers_iterator to treat the contents of the buffer as a range of characters:

// Read a line ending in '\n' from a socket, returning
// the number of characters up to but not including the newline
template <class DynamicBuffer>
std::size_t read_line(net::ip::tcp::socket& sock, DynamicBuffer& buffer)
{
    // this alias keeps things readable
    using range = net::buffers_iterator<
        typename DynamicBuffer::const_buffers_type>;

    for(;;)
    {
        // get iterators representing the range of characters in the buffer
        auto begin = range::begin(buffer.data());
        auto end = range::end(buffer.data());

        // search for "\n" and return if found
        auto pos = std::find(begin, end, '\n');
        if(pos != range::end(buffer.data()))
            return std::distance(begin, end);

        // Determine the number of bytes to read,
        // using available capacity in the buffer first.
        std::size_t bytes_to_read = std::min<std::size_t>(
              std::max<std::size_t>(512,                // under 512 is too little,
                  buffer.capacity() - buffer.size()),
              std::min<std::size_t>(65536,              // and over 65536 is too much.
                  buffer.max_size() - buffer.size()));

        // Read up to bytes_to_read bytes into the dynamic buffer
        buffer.commit(sock.read_some(buffer.prepare(bytes_to_read)));
    }
}

Synchronous I/O

Synchronous input and output is accomplished through blocking function calls that return with the result of the operation. Such operations typically cannot be canceled and do not have a method for setting a timeout. The SyncReadStream and SyncWriteStream concepts define requirements for synchronous streams: a portable I/O abstraction that transfers data using buffer sequences to represent bytes and either error_code or an exception to report any failures. net::basic_stream_socket is a synchronous stream commonly used to form TCP/IP connections. User-defined types which meet the requirements are possible:

// Meets the requirements of SyncReadStream
struct sync_read_stream
{
    // Returns the number of bytes read upon success, otherwise throws an exception
    template <class MutableBufferSequence>
    std::size_t read_some(MutableBufferSequence const& buffers);

    // Returns the number of bytes read successfully, sets the error code if a failure occurs
    template <class MutableBufferSequence>
    std::size_t read_some(MutableBufferSequence const& buffers, error_code& ec);
};

// Meets the requirements of SyncWriteStream
struct sync_write_stream
{
    // Returns the number of bytes written upon success, otherwise throws an exception
    template <class ConstBufferSequence>
    std::size_t write_some(ConstBufferSequence const& buffers);

    // Returns the number of bytes written successfully, sets the error code if a failure occurs
    template <class ConstBufferSequence>
    std::size_t write_some(ConstBufferSequence const& buffers, error_code& ec);
};

A synchronous stream algorithm is written as a function template accepting a stream object meeting the named requirements for synchronous reading, writing, or both. This example shows an algorithm which writes text and uses exceptions to indicate errors:

template <class SyncWriteStream>
void hello (SyncWriteStream& stream)
{
    net::const_buffer cb("Hello, world!", 13);
    do
    {
        auto bytes_transferred = stream.write_some(cb); // may throw
        cb += bytes_transferred; // adjust the pointer and size
    }
    while (cb.size() > 0);
}

The same algorithm may be expressed using error codes instead of exceptions:

template <class SyncWriteStream>
void hello (SyncWriteStream& stream, error_code& ec)
{
    net::const_buffer cb("Hello, world!", 13);
    do
    {
        auto bytes_transferred = stream.write_some(cb, ec);
        cb += bytes_transferred; // adjust the pointer and size
    }
    while (cb.size() > 0 && ! ec);
}

Asynchronous I/O

An asynchronous operation begins with a call to an initiating function, which starts the operation and returns to the caller immediately. This outstanding asynchronous operation proceeds concurrently without blocking the caller. When the externally observable side effects are fully established, a movable function object known as a completion handler provided in the initiating function call is queued for execution with the results, which may include the error code and other specific information. An asynchronous operation is said to be completed after the completion handler is queued. The code that follows shows how some text may be written to a socket asynchronously, invoking a lambda when the operation is complete:

// initiate an asynchronous write operation
net::async_write(sock, net::const_buffer("Hello, world!", 13),
    [](error_code ec, std::size_t bytes_transferred)
    {
        // this lambda is invoked when the write operation completes
        if(! ec)
            assert(bytes_transferred == 13);
        else
            std::cerr << "Error: " << ec.message() << "\n";
    });
// meanwhile, the operation is outstanding and execution continues from here

Every completion handler (also referred to as a continuation) has both an associated allocator returned by net::get_associated_allocator, and an associated executor returned by net::get_associated_executor. These associations may be specified intrusively:

// The following is a completion handler expressed
// as a function object, with a nested associated
// allocator and a nested associated executor.
struct handler
{
    using allocator_type = std::allocator<char>;
    allocator_type get_allocator() const noexcept;

    using executor_type = boost::asio::io_context::executor_type;
    executor_type get_executor() const noexcept;

    void operator()(boost::beast::error_code, std::size_t);
};

Or these associations may be specified non-intrusively, by specializing the class templates net::associated_allocator and net::associated_executor:

namespace boost {
namespace asio {

template<class Allocator>
struct associated_allocator<handler, Allocator>
{
    using type = std::allocator<void>;

    static
    type
    get(handler const& h,
        Allocator const& alloc = Allocator{}) noexcept;
};

template<class Executor>
struct associated_executor<handler, Executor>
{
    using type = boost::asio::executor;

    static
    type
    get(handler const& h,
        Executor const& ex = Executor{}) noexcept;
};

} // boost
} // asio

The function net::bind_executor may be used when the caller wants to change the executor of a completion handler.

The allocator is used by the implementation to obtain any temporary storage necessary to perform the operation. Temporary allocations are always freed before the completion handler is invoked. The executor is a cheaply copyable object providing the algorithm used to invoke the completion handler. Unless customized by the caller, a completion handler defaults to using std::allocator<void> and the executor of the corresponding I/O object.

Networking prescribes facilities to determine the context in which handlers run. Every I/O object refers to an ExecutionContext for obtaining the Executor instance used to invoke completion handlers. An executor determines where and how completion handlers are invoked. Executors obtained from an instance of net::io_context offer a basic guarantee: handlers will only be invoked from threads which are currently calling net::io_context::run.

The AsyncReadStream and AsyncWriteStream concepts define requirements for asynchronous streams: a portable I/O abstraction that exchanges data asynchronously using buffer sequences to represent bytes and error_code to report any failures. An asynchronous stream algorithm is written as a templated initiating function template accepting a stream object meeting the named requirements for asynchronous reading, writing, or both. This example shows an algorithm which writes some text to an asynchronous stream:

template <class AsyncWriteStream, class WriteHandler>
void async_hello (AsyncWriteStream& stream, WriteHandler&& handler)
{
    net::async_write (stream,
        net::buffer("Hello, world!", 13),
        std::forward<WriteHandler>(handler));
}

Concurrency

I/O objects such as sockets and streams are not thread-safe. Although it is possible to have more than one operation outstanding (for example, a simultaneous asynchronous read and asynchronous write) the stream object itself may only be accessed from one thread at a time. This means that member functions such as move constructors, destructors, or initiating functions must not be called concurrently. Usually this is accomplished with synchronization primitives such as a mutex, but concurrent network programs need a better way to access shared resources, since acquiring ownership of a mutex could block threads from performing uncontended work. For efficiency, networking adopts a model of using threads without explicit locking by requiring all access to I/O objects to be performed within a strand.

Universal Model

Because completion handlers cause an inversion of the flow of control, sometimes other methods of attaching a continuation are desired. Networking provides the Universal Model for Asynchronous Operations, providing a customizable means for transforming the signature of the initiating function to use other types of objects and methods in place of a completion handler callback. For example to call to write a string to a socket asynchronously, using a std::future to receive the number of bytes transferred thusly looks like this:

std::future<std::size_t> f = net::async_write(sock,
    net::const_buffer("Hello, world!", 13), net::use_future);

This functionality is enabled by passing the variable net::use_future (of type net::use_future_t<>) in place of the completion handler. The same async_write function overload can work with a fiber launched with asio::spawn:

asio::spawn(
    [&sock](net::yield_context yield)
    {
        std::size_t bytes_transferred = net::async_write(sock,
            net::const_buffer("Hello, world!", 13), yield);
        (void)bytes_transferred;
    });

In both of these cases, an object with a specific type is used in place of the completion handler, and the return value of the initiating function is transformed from void to std::future<std::size_t> or std::size_t. The handler is sometimes called a CompletionToken when used in this context. The return type transformation is supported by customization points in the initiating function signature. Here is the signature for net::async_write:

template<
    class AsyncWriteStream,
    class ConstBufferSequence,
    class CompletionToken>
auto
async_write(
    AsyncWriteStream* stream,                       // references are passed as pointers
    ConstBufferSequence const& buffers,
    CompletionToken&& token)                        // a handler, or a special object.
    ->
    typename net::async_result<                     // return-type customization point.
        typename std::decay<CompletionToken>::type, // type used to specialize async_result.
        void(error_code, std::size_t)               // underlying completion handler signature.
            >::return_type;

The type of the function's return value is determined by the net::async_result customization point, which comes with specializations for common library types such as std::future and may also be specialized for user-defined types. The body of the initiating function calls the net::async_initiate helper to capture the arguments and forward them to the specialization of async_result. An additional "initiation function" object is provided which async_result may use to immediately launch the operation, or defer the launch of the operation until some point in the future (this is called "lazy execution"). The initiation function object receives the internal completion handler which matches the signature expected by the initiating function:

return net::async_initiate<
    CompletionToken,
    void(error_code, std::size_t)>(
        run_async_write{},              // The "initiation" object.
        token,                          // Token must come before other arguments.
        &stream,                        // Additional captured arguments are
        buffers);                       //   forwarded to the initiation object.

This transformed, internal handler is responsible for the finalizing step that delivers the result of the operation to the caller. For example, when using net::use_future the internal handler will deliver the result by calling std::promise::set_value on the promise object returned by the initiating function.

Using Networking

Most library stream algorithms require a tcp::socket, net::ssl::stream, or other Stream object that has already established communication with a remote peer. This example is provided as a reminder of how to work with sockets:

// The resolver is used to look up IP addresses and port numbers from a domain and service name pair
tcp::resolver r{ioc};

// A socket represents the local end of a connection between two peers
tcp::socket stream{ioc};

// Establish a connection before sending and receiving data
net::connect(stream, r.resolve("www.example.com", "http"));

// At this point `stream` is a connected to a remote
// host and may be used to perform stream operations.

Throughout this documentation identifiers with the following names have special meaning: