beast/doc/qbk/08_design/1_http_message.qbk

[/
    Copyright (c) 2016-2019 Vinnie Falco (vinnie dot falco at gmail dot com)

    Distributed under the Boost Software License, Version 1.0. (See accompanying
    file LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)

    Official repository: https://github.com/boostorg/beast

<iframe width="560" height="315" src="https://www.youtube.com/embed/WsUnnYEKPnI?rel=0" frameborder="0" allowfullscreen></iframe>
]

[section:http_message_container HTTP Message Container __video__]

The following video presentation was delivered at
[@https://cppcon.org/ CppCon]
in 2017. The presentation provides a simplified explanation of the design
process for the HTTP message container used in Beast. The slides and code
used are available in the
[@https://github.com/vinniefalco/CppCon2017 GitHub repository].

[/ "Make Classes Great Again! (Using Concepts for Customization Points)"]
[block'''
<mediaobject>
  <videoobject>
    <videodata fileref="https://www.youtube.com/embed/WsUnnYEKPnI?rel=0"
        align="center" contentwidth="560" contentdepth="315"/>
  </videoobject>
</mediaobject>
''']

In this section we describe the problem of modeling HTTP messages and explain
how the library arrived at its solution, with a discussion of the benefits
and drawbacks of the design choices. The goal for creating a message model
is to create a container with value semantics, possibly movable and/or
copyable, that contains all the information needed to serialize, or all
of the information captured during parsing. More formally, given:

* `m` is an instance of an HTTP message container

* `x` is a series of octets describing a valid HTTP message in
  the serialized format described in __rfc7230__.

* `S(m)` is a serialization function which produces a series of octets
  from a message container.

* `P(x)` is a parsing function which produces a message container from
  a series of octets.

These relations are true:

* `S(m) == x`

* `P(S(m)) == m`

We would also like our message container to have customization points
permitting the following: allocator awareness, user-defined containers
to represent header fields, and user-defined types and algorithms to
represent the body. And finally, because requests and responses have
different fields in the ['start-line], we would like the containers for
requests and responses to be represented by different types for function
overloading.

Here is our first attempt at declaring some message containers:

[table
[[
```
/// An HTTP request
template<class Fields, class Body>
struct request
{
    int         version;
    std::string method;
    std::string target;
    Fields      fields;
    
    typename Body::value_type body;
};
```
][
```
/// An HTTP response
template<class Fields, class Body>
struct response
{
    int         version;
    int         status;
    std::string reason;
    Fields      fields;

    typename Body::value_type body;
};
```
]]
]

These containers are capable of representing everything in the model
of HTTP requests and responses described in __rfc7230__. Request and
response objects are different types. The user can choose the container
used to represent the fields. And the user can choose the [*Body] type,
which is a concept defining not only the type of `body` member but also
the algorithms used to transfer information in and out of that member
when performing serialization and parsing.

However, a problem arises. How do we write a function which can accept
an object that is either a request or a response? As written, the only
obvious solution is to make the message a template type. Additional traits
classes would then be needed to make sure that the passed object has a
valid type which meets the requirements. These unnecessary complexities
are bypassed by making each container a partial specialization:
```
/// An HTTP message
template<bool isRequest, class Fields, class Body>
struct message;

/// An HTTP request
template<class Fields, class Body>
struct message<true, Fields, Body>
{
    int         version;
    std::string method;
    std::string target;
    Fields      fields;

    typename Body::value_type body;
};

/// An HTTP response
template<bool isRequest, class Fields, class Body>
struct message<false, Fields, Body>
{
    int         version;
    int         status;
    std::string reason;
    Fields      fields;

    typename Body::value_type body;
};
```

Now we can declare a function which takes any message as a parameter:
```
template<bool isRequest, class Fields, class Body>
void f(message<isRequest, Fields, Body>& msg);
```

This function can manipulate the fields common to requests and responses.
If it needs to access the other fields, it can use overloads with
partial specialization, or in C++17 a `constexpr` expression:
```
template<bool isRequest, class Fields, class Body>
void f(message<isRequest, Fields, Body>& msg)
{
    if constexpr(isRequest)
    {
        // call msg.method(), msg.target()
    }
    else
    {
        // call msg.result(), msg.reason()
    }
}
```

Often, in non-trivial HTTP applications, we want to read the HTTP header
and examine its contents before choosing a type for [*Body]. To accomplish
this, there needs to be a way to model the header portion of a message.
And we'd like to do this in a way that allows functions which take the
header as a parameter, to also accept a type representing the whole
message (the function will see just the header part). This suggests
inheritance, by splitting a new base class off of the message:
```
/// An HTTP message header
template<bool isRequest, class Fields>
struct header;
```

Code which accesses the fields has to laboriously mention the `fields`
member, so we'll not only make `header` a base class but we'll make
a quality of life improvement and derive the header from the fields
for notational convenience. In order to properly support all forms
of construction of [*Fields] there will need to be a set of suitable
constructor overloads (not shown):
```
/// An HTTP request header
template<class Fields>
struct header<true, Fields> : Fields
{
    int         version;
    std::string method;
    std::string target;
};

/// An HTTP response header
template<class Fields>
struct header<false, Fields> : Fields
{
    int         version;
    int         status;
    std::string reason;
};

/// An HTTP message
template<bool isRequest, class Fields, class Body>
struct message : header<isRequest, Fields>
{
    typename Body::value_type body;

    /// Construct from a `header`
    message(header<isRequest, Fields>&& h);
};

```

Note that the `message` class now has a constructor allowing messages
to be constructed from a similarly typed `header`. This handles the case
where the user already has the header and wants to make a commitment to the
type for [*Body]. A function can be declared which accepts any header:
```
template<bool isRequest, class Fields>
void f(header<isRequest, Fields>& msg);
```

Until now we have not given significant consideration to the constructors
of the `message` class. But to achieve all our goals we will need to make
sure that there are enough constructor overloads to not only provide for
the special copy and move members if the instantiated types support it,
but also allow the fields container and body container to be constructed
with arbitrary variadic lists of parameters. This allows the container
to fully support allocators.

The solution used in the library is to treat the message like a `std::pair`
for the purposes of construction, except that instead of `first` and `second`
we have the `Fields` base class and `message::body` member. This means that
single-argument constructors for those fields should be accessible as they
are with `std::pair`, and that a mechanism identical to the pair's use of
`std::piecewise_construct` should be provided. Those constructors are too
complex to repeat here, but interested readers can view the declarations
in the corresponding header file.

There is now significant progress with our message container but a stumbling
block remains. There is no way to control the allocator for the `std::string`
members. We could add an allocator to the template parameter list of the
header and message classes, use it for those strings. This is unsatisfying
because of the combinatorial explosion of constructor variations needed to
support the scheme. It also means that request messages could have [*four]
different allocators: two for the fields and body, and two for the method
and target strings. A better solution is needed.

To get around this we make an interface modification and then add
a requirement to the [*Fields] type. First, the interface change:
```
/// An HTTP request header
template<class Fields>
struct header<true, Fields> : Fields
{
    int         version;

    verb        method() const;
    string_view method_string() const;
    void        method(verb);
    void        method(string_view);

    string_view target(); const;
    void        target(string_view);

private:
    verb method_;
};

/// An HTTP response header
template<class Fields>
struct header<false, Fields> : Fields
{
    int         version;
    int         result;
    string_view reason() const;
    void        reason(string_view);
};
```

The start-line data members are replaced by traditional accessors
using non-owning references to string buffers. The method is stored
using a simple integer instead of the entire string, for the case
where the method is recognized from the set of known verb strings.

Now we add a requirement to the fields type: management of the
corresponding string is delegated to the [*Fields] container, which can
already be allocator aware and constructed with the necessary allocator
parameter via the provided constructor overloads for `message`. The
delegation implementation looks like this (only the response header
specialization is shown):
```
/// An HTTP response header
template<class Fields>
struct header<false, Fields> : Fields
{
    int     version;
    int     status;

    string_view
    reason() const
    {
        return this->reason_impl(); // protected member of Fields
    }

    void
    reason(string_view s)
    {
        this->reason_impl(s);       // protected member of Fields
    }
};
```

Now that we've accomplished our initial goals and more, there are a few 
more quality of life improvements to make. Users will choose different
types for `Body` far more often than they will for `Fields`. Thus, we
swap the order of these types and provide a default. Then, we provide
type aliases for requests and responses to soften the impact of using
`bool` to choose the specialization:

```
/// An HTTP header
template<bool isRequest, class Body, class Fields = fields>
struct header;

/// An HTTP message
template<bool isRequest, class Body, class Fields = fields>
struct message;

/// An HTTP request
template<class Body, class Fields = fields>
using request = message<true, Body, Fields>;

/// An HTTP response
template<class Body, class Fields = fields>
using response = message<false, Body, Fields>;
```

This allows concise specification for the common cases, while
allowing for maximum customization for edge cases:
```
request<string_body> req;

response<file_body> res;
```

This container is also capable of representing complete HTTP/2 messages.
Not because it was explicitly designed for, but because the IETF wanted to
preserve message compatibility with HTTP/1. Aside from version specific
fields such as Connection, the contents of HTTP/1 and HTTP/2 messages are
identical even though their serialized representation is considerably
different. The message model presented in this library is ready for HTTP/2.

In conclusion, this representation for the message container is well thought
out, provides comprehensive flexibility, and avoids the necessity of defining
additional traits classes. User declarations of functions that accept headers
or messages as parameters are easy to write in a variety of ways to accomplish
different results, without forcing cumbersome SFINAE declarations everywhere.

[endsect]