Introduction
This repository contains a variety of potentially useful information for those wanting to work with the P4 programming language.
Installing open source P4 development tools
There are install scripts in this repository that can assist you
in building and installing the p4c
and behavioral-model
projects
and their dependencies on Ubuntu Linux systems with these versions:
- 20.04 or 22.04
See these instructions and trouble-shooting tips for using these scripts.
There are also some newer instructions here, perhaps a little less beginner-friendly, for installing IPDK 23.01 on an Ubuntu 20.04 Linux system, and using the DPDK software switch to load and run P4 programs.
Articles on various P4 topics
See this page for a list of articles included in this repository.
Overview of public P4.org documentation and code repositories
- An overview of the p4lang organization Github repositories
- A figure with the dependencies between these repositories (last updated June 2017 so may be slightly out of date).
Small demo P4 programs
A collection of small demo P4 programs, some of them with equivalent versions written in both P4_14 and P4_16 versions of the language.
- includes a heavily commented P4_16 program, which by reading carefully one may learn some things about the P4_16 language.
- Each of the demo directories includes its own README.md file with
instructions for compiling the program and running it with the
simple_switch
software switch from thebehavioral-model
repository, including interactively adding table entries to the tables, and send packets using Scapy that will be processed by the software switch.
Related development tools and tips
-
Instructions for setting up several text editing programs for color highlighting P4 programs, and quickly finding definitions for named things like control blocks and tables.
-
'Cheat sheet' of where P4_16 language constructs are allowed to be used within a program.
- figure
- text
- NOTE: This describes where language constructs are allowed by the P4_16 language specification (version 1.1.0). P4_16 implementations may restrict your program to less than what is shown here, or enable extensions not described here. Ask the creator of the implementation you are using to learn more.
-
Another useful 'cheat sheet' with example snippets of code is in the p4lang/tutorials repository.
-
Some notes on using the Scapy library for creating, editing, and decoding packets, useful when creating test packets to send to a P4 program being tested, especially within the control of a general purpose programming language (i.e. Python).
Brief overview of P4
This is just enough to give you a 'flavor' of what P4 is like, in about 500 words. It assumes you are familiar with the programming language C or something very similar to it, e.g. C++ or Java.
- Start with C.
- Remove loops, recursive calls, pointers, malloc, and free. When
your mind recovers from reeling over these drastic limitations,
you will clearly realize that P4 is not a general purpose
programming language. It was not designed to be.
- Without loops or recursive calls, the work done per packet can be bounded at compilation time, which helps when targeting the highest performance P4-programmable devices.
- Without pointers, malloc, and free, general purpose data structures like linked lists, trees, etc. having arbitrary size is not possible. You can implement trees with fixed depth, e.g. a depth 3 tree could be implemented via a sequence of 3 P4 tables.
- Add special constructs called parsers, focused on the capabilities
most needed when parsing the contents of a received packet into a
sequence of headers.
- Parsers are defined as finite state machines, with states that you must name and define what the possible transitions are between them. It is actually allowed to have loops in the parser finite state machine, but the highest performance targets will typically restrict you to loops that can be unrolled to a compile-time known maximum number of iterations, e.g. for parsing a sequence of MPLS headers at most 5 headers long (where 5, or some other value, is a number you pick in your source code).
- P4 is focused on packet header processing. Whatever part of a packet you do not parse into some header is, for that P4 program, considered the "payload" of the packet, which is typically carried along, unmodified, with the packet when you are done processing it. You can modify header fields however you like.
- Add special constructs called tables, where for each one you
define a search key consisting of a number of packet header fields
and/or values of variables in your P4 program. Each table can
also have one or more actions defined for them.
- A P4 program represents only a small fraction of a complete working system. Control plane software that would typically be running on a general purpose CPU, written in one or more general purpose programming languages like C, C++, Java, Python, etc., is responsible for adding and removing entries to these tables, selecting for each entry the search key to be matched against, and the action to be executed if, while processing a packet, that table entry is matched.
- A P4_16 "architecture" like the Portable Switch Architecture (PSA) also defines a library of other constructs, such as packet/byte counters, meters for enforcing average packet and/or bit rates on forwarded traffic, registers for some limited kinds of stateful packet processing, and methods for recirculating a packet, multicasting it to multiple destinations, etc.
- Fields and variables can be integers of arbitrary bit width (up to some maximum size allowed by a particular implementation), with results of arithmetic operations well defined for all operations. P4 implementations need not implement floating point arithmetic, and I expect most would not because such things are not needed for the majority of packet processing applications. P4 implementations also need not implement multiplication, division, or modulo operations, again given that such operations are often not needed for most packet processing applications.
- Note that integer arithmetic is sufficient for implementing fixed point operations, because fixed point add/subtract/shift can be viewed as the corresponding operations on integers that are in smaller units, e.g. an integer could be used to represent time in multiples of 0.8 nanoseconds, as opposed to in units of nanoseconds, which would be convenient if an implementation had a clock speed of 1.25 GHz and measured time in integer number of clock cycles (1 cycle / 0.8 nanosec = 1.25 GHz).
Very brief comparison of P4_14 and P4_16 languages
Some advantages of P4_16 over P4_14:
- You can write assignments that look like C/C++/Java, rather than
modify_field(dst, src);
all over the place, and you can have arithmetic expressions on the right-hand side instead ofadd_to_field
/subtract_from_field
/etc. This is not additional power in the language, but it is a nice convenience for those familiar with those other languages. - You may call controls from other controls in P4_14, but there are
no parameters or return values. All side effects must be done via
access to global variables. In P4_16, there are no global
variables -- you may pass parameters with directionality
in
,out
,inout
. - Tables must be, and externs may be, defined within the scope of a control, and are then accessible only from that control, which can be useful for keeping the code that accesses those objects closer to it, and knowing where they can be accessed from. Externs are used for things like counters, meters, and registers that were part of the base P4_14 language, but in P4_16 are defined as extern add-ons in the Portable Switch Architecture specification).
Disadvantages of P4_16 vs P4_14:
- Tool and vendor support may not have been as good for P4_16 in 2018 as it was for P4_14, but as of 2021, P4_16 is supported for most P4-programmable target devices.