4 // Fraunhofer Institute for Open Communication Systems (FOKUS)
6 // The contents of this file are subject to the Fraunhofer FOKUS Public License
7 // Version 1.0 (the "License"); you may not use this file except in compliance
8 // with the License. You may obtain a copy of the License at
9 // http://senf.berlios.de/license.html
11 // The Fraunhofer FOKUS Public License Version 1.0 is based on,
12 // but modifies the Mozilla Public License Version 1.1.
13 // See the full license text for the amendments.
15 // Software distributed under the License is distributed on an "AS IS" basis,
16 // WITHOUT WARRANTY OF ANY KIND, either express or implied. See the License
17 // for the specific language governing rights and limitations under the License.
19 // The Original Code is Fraunhofer FOKUS code.
21 // The Initial Developer of the Original Code is Fraunhofer-Gesellschaft e.V.
22 // (registered association), Hansastraße 27 c, 80686 Munich, Germany.
25 // Stefan Bund <g0dil@berlios.de>
30 /** \mainpage The SENF Socket Library
32 The Socket library provides a high level and object oriented abstraction based on the BSD socket
33 API (but not limited to it).
37 \section socket_intro Introduction
38 \seechapter \ref structure \n
39 \seechapter \ref usage
41 The socket library abstraction is based on several concepts:
43 \li The basic visible interface is a \link handle_group handle object\endlink
44 \li The socket interface relies on a \link policy_group policy framework \endlink to configure
46 \li The rest of the socket API is accessible using a classic inheritance hierarchy of \link
47 protocol_group protocol classes \endlink
48 \li There is a family of auxiliary \ref addr_group to supplement the socket library
51 \section socket_handle Socket Handles
52 \seechapter \ref handle_group \n
53 \seechapter \ref concrete_protocol_group
55 The handle/body architecture provides automatic reference counted management of socket
56 instances. This is the visible interface to the socket library.
58 Each specific protocol is used primarily via a protocol specific handle (a typedef
59 symbol). However, more generic kinds of handles can be defined for more generic functionality.
63 \section socket_policy The Policy interface
64 \seechapter \ref policy_group
66 The policy framework configures the exact features, a specific type of socket handle
67 provides. This offers highly efficient access to the most important socket functions (like
68 reading and writing). The policy interface however is a \e static, non-polymorphic interface.
71 \section socket_protocol The Protocol interface
72 \seechapter \ref protocol_group
75 The protocol interface provides further protocol dependent and (possibly) polymorphic access to
76 further socket functionality. On the other hand, this type of interface is not as flexible,
77 generic and fast as the policy interface.
79 \section socket_addr Auxiliary Addressing classes
80 \seechapter \ref addr_group
82 To supplement the socket library, there are a multitude of addressing classes. These come in two
84 \li Protocol specific addresses (e.g. INet4Address, MACAddress)
85 \li Socket addresses (\c sockaddr) (e.g. INet4SocketAddress, LLSocketAddress)
87 Whereas the protocol specific addresses are custom value types which represent their
88 corresponding low-level address, the socket addresses are based on the corresponding \c sockaddr
91 \section socket_further Going further
92 \seechapter \ref extend \n
93 \seechapter \ref implementation
95 The socket library is highly flexible and extensible. The implementation is not restricted to
96 plain BSD sockets: Any type of read/write communication can be wrapped into the socket library
97 (one Example is the TapSocketHandle which provides access to a Linux \c tap device).
101 /** \page structure Overview of the Socket Library Structure
103 \image html Handle.png
105 This diagram tries to give a structural overview of the Socket Library, it does \e not directly
106 show, how the library is implemented. This will be explained later.
108 The outside interface to the library is a Handle object. This is the only object, the library
109 user directly interacts with. Every handle references some socket. This is like the ordinary
110 POSIX API: the file descriptor (also called file handle, an integer number) references a socket
111 structure which lives in kernel space. In this library, the Handle object (which is not a simple
112 integer any more but an object) references the Socket (which is part of the
113 implementation). Several handles may reference the same Socket. In contrast to the kernel API,
114 the library employs reference counting to release a socket when the last Handle to it goes out
117 The behavior of a Socket is defined by it's Protocol. It is divided into two parts: the
118 <em>policy interface</em> and the <em>protocol interface</em>. Together they provide the
119 complete API for a specific type of Socket as defined by the Protocol. The <em>policy
120 interface</em> provides highly efficient access to the most frequently used operations whereas
121 the <em>protocol interface</em> completes the interface by providing a complete set of all
122 protocol specific operations not found in the policy interface. This structure allows us to
123 combine the benefits of two design methodologies: The policy interface utilizes a policy based
124 design technique and is highly efficient albeit more complex to implement, whereas the protocol
125 interface is based on a more common inheritance architecture which is not as optimized for
126 performance but much simpler to implement. We reduce the complexity of the implementation by
127 reducing the policy interface to a minimal sensible subset of the complete API.
129 \section over_policy The Policy Interface
131 The policy of a Socket consists of several parts, called <em>policy axis</em>. Each axis
132 corresponds to one specific interface aspect of the Socket. The exact meaning of the policy axis
133 are defined elsewhere (see \ref policy_group). The Protocol will always provide a complete set
134 of <em>policy classes</em>, one for each axis.
136 This <em>complete socket policy</em> defines the policy interface of the protocol. This
137 interface is carried over into the Handle. The socket policy as defined in the Handle however
138 may be <em>incomplete</em>. This mans, that the \e accessible interface of the Socket depends on
139 the type of Handle used. The inherent interface does not change but the view of this interface
140 does if the Handle does not provide the \e complete policy interface. This feature is very
141 important. It allows to define generic Handle types. A generic Handle with an incompletely
142 defined policy can point to an arbitrary Socket as long as all those policy axis which \e are
143 defined match those defined in that Socket's protocol. Using such a generic handle decouples the
144 implementation parts using this handle from the other socket aspects (e.g. you may define a
145 generic socket handle for TCP based communication leaving the addressingPolicy undefined which
146 makes your code independent of the type of addressing, IPv4 or IPv6).
148 This can be described as generalized compile-time polymorphism: A base class reference to some
149 derived class will only give access to a reduced interface (the base class interface) of a
150 class. The class still is of it's derived type (and inherently has the complete interface) but
151 only part of it is accessible via the base class reference. Likewise a generic handle (aka base
152 class reference) will only provide a reduced interface (aka base class interface) to the derived
153 class instance (aka socket).
155 \section over_protocol The Protocol Interface
157 The protocol interface is provided by a set of <em>protocol facets</em>. Each facet provides a
158 part of the interface. Whereas the policy interface is strictly defined (the number and type of
159 policy axis is fixed and also the possible members provided by the policy interface are fixed),
160 the protocol interface is much more flexible. Any member needed to provide a complete API for
161 the specific protocol may be defined, the number and type of facets combined to provide the
162 complete interface is up to the Protocol implementor. This flexibility is necessary to provide a
163 complete API for every possible protocol.
165 However this flexibility comes at a cost: To access the protocol interface the user must know
166 the exact protocol of the socket. With other words, the protocol is only accessible if the
167 handle you use is a <em>protocol specific</em> handle. A protocol specific Handle differs from a
168 generic Handle in two ways: It always has a complete policy and it knows the exact protocol type
169 of the socket (which generic handles don't). This allows to access to the complete protocol
172 \section over_impl Implementation of the Socket Library Structure
174 In the Implementation, the socket policy is identified by an instance of the senf::SocketPolicy
175 template. The Socket representation is internally represented in a senf::SocketBody which is not
176 outside visible. The Handle is provided by a hierarchy of handle templates. Each Handle template
177 uses template arguments for the policy and/or protocol as needed (see \ref handle_group).
179 The Handle hierarchy divides the interface into two separate strains: the client interface
180 (senf::ClientSocketHandle and senf::ProtocolClientSocketHandle) provides the interface of a
181 client socket whereas the server interface (senf::ServerSocketHandle and
182 senf::ProtocolServerSocketHandle) provides the interface as used by server sockets.
184 The protocol interface is implemented using inheritance: The Protocol class inherits from each
185 protocol facet using multiple (virtual public) inheritance. The Protocol class therefore
186 provides the complete protocol API in a unified (see \ref protocol_group).
189 /** \page usage Using the Socket Library
191 Whenever you use the socket library, what you will be dealing with are FileHandle derived
192 instances. The socket library relies on reference counting to automatically manage the
193 underlying socket representation. This frees you of having to manage the socket lifetime
196 \section usage_create Creating a Socket Handle
198 To create a new socket handle (opening a socket), you will need to use
199 ProtocolClientSocketHandle or ProtocolServerSocketHandle. You will probably not use these
200 templates as is but use proper typedefs (for example TCPv4ClientSocketHandle or
201 PacketSocketHandle). The documentation for these socket handles are found in the protocol class
202 (for example TCPv4SocketProtocol or PacketSocketProtocol).
204 \section usage_reusable Writing Reusable Components
206 To make your code more flexible, you should not pass around your socket in this form. Most of
207 your code will be using only a small subset of the ProtocolClientSocketHandle or
208 ProtocolServerSocketHandle API.
210 If instead of using the fully specified handle type you use a more incomplete type, you allow
211 your code to be used with all sockets which fulfill the minimal requirements of your code. These
212 types are based on the ClientSocketHandle and ServerSocketHandle templates which implement the
213 policy interface without providing the concrete protocol interface. To use those templates you
214 may define a special reduced policy or handle for your code. By giving only an incomplete policy
215 you thereby reduce the interface to that required by your module:
218 typedef ClientSocketHandle<
222 ConnectedCommunicationPolicy > > MyReadableHandle;
226 This defines \c MyReadableHandle as a ClientSocketHandle which will have only read
227 functionality. Your code expects a stream interface (in contrast to a packet or datagram based
228 interface). You will not have \c write or \c readfrom members. \c write will be disabled since
229 the WritePolicy is unknown, \c readfrom will be disabled since a socket with the
230 ConnectedCommunicationPolicy does not have a \c readfrom member.
238 /** \page extend Extending the Library
240 There are two layers, on which the socket library can be extended: On the protocol layer and on
241 the policy layer. Extending the protocol layer is quite simple and works as long as the desired
242 protocol does use the same BSD API used by the standard internet protocols as implemented in the
243 standard policies (i.e. it uses ordinary read() and write() or rcvfrom() or sendto() calls and
246 If however the implementation of a policy feature needs to be changed, a new policy class has to
247 be written. This also is not very complicated however the integration is more complex.
249 \section extend_protocol Writing a new protocol class
251 Most protocols can be implemented by just implementing a new protocol class. The protocol class
252 must be derived from ConcreteSocketProtocol and takes the socket policy (as created by
253 MakeSocketPolicy) as a template argument. See the documentation of this class for the interface.
255 \attention You may want to use multiple inheritance as it is used in the implementation of the
256 standard protocols (See \ref protocol_group). You must however be extra careful to ensure, that
257 every class ultimately has SocketPolicy as a public \e virtual base.
259 After the protocol class has been defined, you will probably want to provide typedefs for the
260 new protocol sockets. If the new protocol is connection oriented, this will be like
262 typedef ProtocolClientSocketHandle<MySocketProtocolClass> MySocketProtocolClientSocketHandle;
263 typedef ProtocolServerSocketHandle<MySocketProtocolClass> MySocketProtocolServerSocketHandle;
266 \section extend_policy Extending the policy framework
268 If you have to extend the policy framework, you will need to be aware of some important
269 limitations of the socket library:
271 \li When you define a new policy for some axis, this new policy <em>must not</em> be derived
272 from one of the existing concrete policy classes (except of course the respective policy
273 axis base class). This is important since the policy type is \e not polymorphic. The policy
274 to be used is selected by the compiler using the \e static type, which is exactly what is
275 desired, since this allows calls to be efficiently inlined.
277 \li Therefore, extending the policy framework will make the new socket probably \e incompatible
278 with generic code which relies on the policy axis which is extended. Example: If you write a
279 new write policy because your protocol does not use ordinary write() system calls but some
280 protocol specific API, Then any generic function relying on WritablePolicy will \e not work
281 with the new socket, since the socket does \e not have this policy, it has some other kind
284 Therefore you need to be careful of what you are doing. The first step is to find out, which
285 policy you will have to implement. For this, find the ClientSocketHandle and/or
286 ServerSocketHandle members you want to change (see \ref ClientSocketHandle and \ref
287 ServerSocketHandle). Not all policy axis directly contribute to the SocketHandle
288 interface. However, some policy members additionally depend on other policy axis (example:
289 AddressingPolicy::connect is only defined if the communication policy is
290 ConnectedCommunication).
295 /** \page implementation Implementation notes
297 \section class_diagram Class Diagram
299 \diaimage SocketLibrary-classes.dia
301 \section impl_notes Arbitrary Implementation Notes
303 \li The implementation tries to isolate the library user as much as possible from the system
304 header files since those headers define a lot of define symbols and introduce a host of
305 symbols into the global namespace. This is, why some classes define their own \c enum types
306 to replace system defined define constants. This also precludes inlining some functionality.
308 \li To reduce overhead, template functions/members which are more than one-liners are often
309 implemented in terms of a non-template function/member. This is also used to further the
310 isolation from system headers as defined above (template code must always be included into
311 every compilation unit together with all headers need for the implementation).
320 // c-file-style: "senf"
321 // indent-tabs-mode: nil
322 // ispell-local-dictionary: "american"
325 // compile-command: "scons -u doc"