// Copyright (C) 2007
// Fraunhofer Institut fuer offene Kommunikationssysteme (FOKUS)
// Kompetenzzentrum fuer Satelitenkommunikation (SatCom)
// Stefan Bund <g0dil@berlios.de>
//
// This program is free software; you can redistribute it and/or modify
// it under the terms of the GNU General Public License as published by
// the Free Software Foundation; either version 2 of the License, or
// (at your option) any later version.
//
// This program is distributed in the hope that it will be useful,
// but WITHOUT ANY WARRANTY; without even the implied warranty of
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
// GNU General Public License for more details.
//
// You should have received a copy of the GNU General Public License
// along with this program; if not, write to the
// Free Software Foundation, Inc.,
// 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
/** \mainpage libPPI : The Packet Processing Infrastructure
The PPI provides an infrastructure to create packet oriented network processing
applications. A PPI application is built by combining processing modules in a very flexible
manner.
\image html scenario.png Target Scenario
The PPI concept is built around some key concepts
\li The PPI is based on processing \ref packets. It does not handle stream oriented channels.
\li The PPI is built around reusable \ref modules. Each module is completely independent.
\li Each module has an arbitrary number of \ref connectors, inputs and outputs.
\li The modules are connected to each other using flexible \ref connections.
\li Data flow throughout the network is governed via flexible automatic or manual \ref
throttling.
\li Modules may register additional external \ref events (file descriptor events or timers).
The PPI thereby builds on the facilities provided by the other components of the SENF
framework. The target scenario above depicts a diffserv capable UDLR/ULE router including
performance optimizations for TCP traffic (PEP). This router is built by combining several
modules.
\section design Design considerations
The PPI interface is designed to be as simple as possible. It provides sane defaults for all
configurable parameters to simplify getting started. It also automates all resource
management. Especially to simplify resource management, the PPI will take many configuration
objects by value. Even though this is not as efficient, it frees the user from most resource
management chores. This decision does not affect the runtime performance since it only affects
the configuration step.
\section packets Packets
The PPI processes packets and uses the <a href="@TOPDIR@/Packets/doc/html/index.html">Packet
library</a> to handle them. All packets are passed around as generic Packet::ptr's, the PPI
does not enforce any packet type restrictions.
\section modules Modules
A module is represented by a class type. Each module has several components:
\li It may have any number of connectors (inputs and outputs)
\li Each module declares flow information which details the route packets take within the
module. This information does not define how the information is processed, it only tells,
where data arriving on some input will be directed at.
\li The module might take additional parameters.
\li The module might also register additional events.
Modules are divided roughly in to two categories: I/O modules provide packet sources and sinks
(network connection, writing packets to disk, generating new packets) whereas processing modules
process packets internally. In the target scenario, <em>TAP</em>, <em>ASI Out</em>, <em>Raw
Socket</em> and in a limited way <em>Generator</em> are I/O modules whereas <em>PEP</em>,
<em>DiffServ</em>, <em>DVB Enc</em>, <em>GRE/UDLR</em>, <em>TCP Filter</em> and <em>Stuffer</em>
are processing modules. <em>ASI/MPEG</em> and <em>Net</em> are external I/O ports which are
integrated via the <em>TAP</em>, <em>ASI Out</em> and <em>Raw Sock</em> modules using external
events.
The following example module declares three I/O connectors (see below): <tt>payload</tt>,
<tt>stuffing</tt> and <tt>output</tt>. These connectors are defined as <em>public</em> data
members so they can be accessed from the outside. This is important as we will see below.
\code
class RateStuffer
: public senf::ppi::module::Module
{
senf::ppi::IntervalTimer timer_;
public:
senf::ppi::connector::ActiveInput payload;
senf::ppi::connector::ActiveInput stuffing;
senf::ppi::connector::ActiveOutput output;
RateStuffer(unsigned packetsPerSecond)
: timer_(1000u, packetsPerSecond)
{
route(payload, output);
route(stuffing, output);
registerEvent(&RateStuffer::tick, timer_);
}
private:
void tick()
{
if (payload)
output(payload());
else
output(stuffing());
}
};
\endcode
On module instantiation, it will declare it's flow information with <tt>route</tt> (which is
inherited from <tt>senf::ppi::module::Module</tt>). Then the module registers an interval timer
which will fire <tt>packetsPerSecond</tt> times every <tt>1000</tt> milliseconds.
The processing of the module is very simple: Whenever a timer tick arrives a packet is sent. If
the <tt>payload</tt> input is ready (see throttling below), a payload packet is sent, otherwise
a stuffing packet is sent. The module will therefore provide a constant stream of packets at a
fixed rate on <tt>output</tt>
An example module to generate the stuffing packets could be
\code
class CopyPacketGenerator
: public senf::ppi::module::Module
{
public:
senf::ppi::connector::PassiveOutput output;
CopyPacketGenerator(Packet::ptr template)
: template_ (template)
{
noroute(output);
output.onRequest(&CopyPacketGenerator::makePacket);
}
private:
Packet::ptr template_;
void makePacket()
{
output(template_.clone());
}
};
\endcode
This module just produces a copy of a given packet whenever output is requested.
\section connectors Connectors
Inputs and Outputs can be active and passive. An \e active I/O is <em>activated by the
module</em> to send data or to poll for available packets. A \e passive I/O is <em>signaled by
the framework</em> to fetch data from the module or to pass data into the module.
To send or receive a packet (either actively or after packet reception has been signaled), the
module just calls the connector. This allows to generate or process multiple packets in one
iteration. However, reading will only succeed, as long as packets are available from the
connection.
Since a module is free to generate more than a single packet on incoming packet requests, all
input connectors incorporate a packet queue. This queue is exposed to the module and allows the
module to process packets in batches.
\section connections Connections
To make use of the modules, they have to be instantiated and connections have to be created
between the I/O connectors. It is possible to connect any pair of input/output connectors as
long as one of them is active and the other is passive.
It is possible to connect two active connectors with each other using a special adaptor
module. This Module has a passive input and a passive output. It will queue any incoming packets
and automatically handle throttling events (see below). This adaptor is automatically added by
the connect method if needed.
To complete our simplified example: Lets say we have an <tt>ActiveSocketInput</tt> and a
<tt>PassiveUdpOutput</tt> module. We can then use our <tt>RateStuffer</tt> module to build an
application which will create a fixed-rate UDP stream:
\code
RateStuffer rateStuffer (10);
senf::Packet::ptr stuffingPacket = senf::Packet::create<...>(...);
CopyPacketGenerator generator (stuffingPacket);
senf::UDPv4ClientSocketHandle inputSocket (1111);
senf::ppi::module::ActiveSocketReader udpInput (inputSocket);
senf::UDPv4ClientSocketHandle outputSocket ("2.3.4.5:2222");
senf::ppi::module::PassiveSocketWriter udpOutput (outputSocket);
senf::ppi::module::PassiveQueue adaptor;
senf::ppi::connect(udpInput.output, adaptor.input);
senf::ppi::connect(adaptor.output, rateStuffer.payload);
adaptor.qdisc(ThresholdQueueing(10,5));
senf::ppi::connect(generator.output, rateStuffer.stuffing);
senf::ppi::connect(rateStuffer.output, udpOutput.input);
senf::ppi::run();
\endcode
First all necessary modules are created. Then the connections between these modules are set
up. The buffering on the udpInput <-> rateStuffer adaptor is changed so the queue will begin to
throttle only if more than 10 packets are in the queue. The connection will be unthrottled as
soon as there are no more than 5 packets left in the queue. This application will read
udp-packets coming in on port 1111 and will forward them to port 2222 on host 2.3.4.5 with a
fixed rate of 10 packets / second.
\section throttling Throttling
If a passive connector cannot handle incoming requests, this connector may be \e
throttled. Throttling a request will forward a throttle notification to the module connected
to that connector. The module then must handle this throttle notification. If automatic
throttling is enabled for the module (which is the default), the notification will automatically
be forwarded to all dependent connectors as taken from the flow information. For there it will
be forwarded to further modules and so on.
A throttle notification reaching an I/O module will normally disable the input/output by
disabling any external I/O events registered by the module. When the passive connector which
originated the notification becomes active again, it creates an unthrottle notification which
will be forwarded in the same way. This notification will re-enable any registered I/O events.
The above discussion shows, that throttle events are always generated on passive connectors and
received on active connectors. To differentiate further, the throttling originating from a
passive input is called <em>backward throttling</em> since it is forwarded in the direction \e
opposite to the data flow. Backward throttling notifications are sent towards the input
modules. On the other hand, the throttling originating from a passive output is called
<em>forward throttling</em> since it is forwarded along the \e same direction the data
is. Forward throttling notifications are therefore sent towards the output modules.
Since throttling a passive input may not disable all further packet delivery immediately, all
inputs contains an input queue. In it's default configuration, the queue will send out throttle
notifications when it becomes non-empty and unthrottle notifications when it becomes empty
again. This automatic behavior may however be disabled. This allows a module to collect incoming
packets in it's input queue before processing a bunch of them in one go.
\section events Events
Modules may register additional events. These external events are very important since they
drive the PPI framework. Possible event sources are
\li time based events
\li file descriptors.
\li internal events (e.g. IdleEvent)
Here some example code implementing the ActiveSocketInput Module:
\code
class ActiveSocketReader
: public senf::ppi::module::Module
{
typedef senf::ClientSocketHandle<
senf::MakeSocketPolicy< senf::ReadablePolicy,
senf::DatagramFramingPolicy > > SocketHandle;
SocketHandle socket_;
DataParser const & parser_;
senf::ppi:IOSignaler event_;
static PacketParser<senf::DataPacket> defaultParser_;
public:
senf::ppi::connector::ActiveOutput output;
// I hestitate taking parser by const & since a const & can be bound to
// a temporary even though a const & is all we need. The real implementation
// will probably make this a template arg. This simplifies the memory management
// from the users pov.
ActiveSocketReader(SocketHandle socket,
DataParser & parser = ActiveSocketReader::defaultParser_)
: socket_ (socket),
parser_ (parser)
event_ (socket, senf::ppi::IOSignaler::Read)
{
registerEvent( &ActiveSocketReader::data, event_ );
route(event_, output);
}
private:
void data()
{
std::string data;
socket_.read(data);
output(parser_(data));
}
};
\endcode
First we declare our own socket handle type which allows us to read packets. The constructor
then takes two arguments: A compatible socket and a parser object. This parser object gets
passed the packet data as read from the socket (an \c std::string) and returns a
senf::Packet::ptr. The \c PacketParser is a simple parser which interprets the data as specified
by the template argument.
We register an IOSignaler event. This event will be signaled whenever the socket is
readable. This event is routed to the output. This routing automates throttling for the socket:
Whenever the output receives a throttle notifications, the event will be temporarily disabled.
Processing arriving packets happens in the \c data() member: This member simple reads a packet
from the socket. It passes this packet to the \c parser_ and sends the generated packet out.
\section flows Information Flow
The above description conceptually introduces three different flow levels:
\li The <em>data flow</em> is, where the packets are flowing. This flow always goes from output
to input connector.
\li The <em>execution flow</em> describes the flow of execution from one module to another. This
flow always proceeds from active to passive connector.
\li The <em>control flow</em> is the flow of throttling notifications. This flow always proceeds
\e opposite to the execution flow, from passive to active connector.
This is the outside view, from without any module. These flows are set up using
senf::ppi::connect() statements.
Within a module, the different flow levels are defined differently depending on the type of
flow:
\li The <em>data flow</em> is defined by how data is processed. The different event and
connector callbacks will pass packets around and thereby define the data flow
\li Likewise, the <em>execution flow</em> is defined parallel to the data flow (however possible
in opposite direction) by how the handler of one connector calls other connectors.
\li The <em>control flow</em> is set up using senf::ppi::Module::route statements (as long as
automatic throttling is used. Manual throttling defines the control flow within the
respective callbacks).
In nearly all cases, these flows will be parallel. Therefore it makes sense to define the \c
route statement as defining the 'conceptual data flow' since this is also how control messages
should flow (sans the direction, which is defined by the connectors active/passive property).
\see \ref ppi_implementation \n
<a href="http://openfacts.berlios.de/index-en.phtml?title=SENF:_Packet_Processing_Infrastructure">Implementation plan</a>
*/
/** \page ppi_implementation Implementation Overview
\section processing Data Processing
The processing in the PPI is driven by events. Without events <em>nothing will happen</em>. When
an event is generated, the called module will probably call one of it's active connectors.
Calling an active connector will directly call the handler registered at the connected passive
connector. This way the call and data are handed across the connections until an I/O module will
finally handle the request (by not calling any other connectors).
Throttling is handled in the same way: Throttling a passive connector will call a corresponding
(internal) method of the connected active connector. This method will call registered handlers
and will analyze the routing information of the module for other (passive) connectors to call
and throttle. This will again create a call chain which terminates at the I/O modules. An event
which is called to be throttled will disable the event temporarily. Unthrottling works in the
same way.
This simple structure is complicated by the existence of the input queues. This affects both
data forwarding and throttling:
\li A data request will only be forwarded, if no data is available in the queue
\li The connection will only be throttled when the queue is empty
\li Handlers of passive input connectors must be called repeatedly until either the queue is
empty or the handler does not take any packets from the queue
\section logistics Managing the Data Structures
The PPI itself is a singleton. This simplifies many of the interfaces (We do not need to pass
the PPI instance). Should it be necessary to have several PPI systems working in parallel
(either by registering all events with the same event handler or by utilizing multiple threads),
we can still extend the API by adding an optional PPI instance argument.
Every module manages a collection of all it's connectors and every connector has a reference to
it's containing module. In addition, every connector maintains a collection of all it's routing
targets.
All this data is initialized via the routing statements. This is, why \e every connector must
appear in at least one routing statement: These statements will as a side effect initialize the
connector with it's containing module.
Since all access to the PPI via the module is via it's base class, unbound member function
pointers can be provided as handler arguments: They will automatically be bound to the current
instance. This simplifies the PPI usage considerably. The same is true for the connectors: Since
they know the containing module, they can explicitly bind unbound member function pointers to
the instance.
\section random_notes Random implementation notes
Generation of throttle notifications: Backward throttling notifications are automatically
generated (if this is not disabled) whenever the input queue is non-empty \e after the event
handler has finished processing. Forward throttling notifications are not generated
automatically within the connector. However, the Passive-Passive adaptor will generate
Forward-throttling notifications whenever the input queue is empty.
*/
// Local Variables:
// mode: c++
// fill-column: 100
// c-file-style: "senf"
// indent-tabs-mode: nil
// ispell-local-dictionary: "american"
// mode: flyspell
// mode: auto-fill
// End: