Programmer’s Guide

Release 2.0.0

April 24, 2015

 CONTENTS

1 Introduction 1
1.1 Documentation Roadmap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Related Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

2 Overview 3
2.1 Development Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 Environment Abstraction Layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.3 Core Components . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.4 Ethernet* Poll Mode Driver Architecture . . . . . . . . . . . . . . . . . . . . . . . 6
2.5 Packet Forwarding Algorithm Support . . . . . . . . . . . . . . . . . . . . . . . . 6
2.6 librte_net . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

3 Environment Abstraction Layer 8

3.1 EAL in a Linux-userland Execution Environment . . . . . . . . . . . . . . . . . . 8
3.2 Memory Segments and Memory Zones (memzone) . . . . . . . . . . . . . . . . 12
3.3 Multiple pthread . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

4 Malloc Library 16
4.1 Cookies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4.2 Alignment and NUMA Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4.3 Use Cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4.4 Internal Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

5 Ring Library 21
5.1 References for Ring Implementation in FreeBSD* . . . . . . . . . . . . . . . . . 22
5.2 Lockless Ring Buffer in Linux* . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
5.3 Additional Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
5.4 Use Cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
5.5 Anatomy of a Ring Buffer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
5.6 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

6 Mempool Library 33
6.1 Cookies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
6.2 Stats . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
6.3 Memory Alignment Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
6.4 Local Cache . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
6.5 Use Cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

7 Mbuf Library 36
7.1 Design of Packet Buffers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

i
 7.2 Buffers Stored in Memory Pools . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
7.3 Constructors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
7.4 Allocating and Freeing mbufs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
7.5 Manipulating mbufs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
7.6 Meta Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
7.7 Direct and Indirect Buffers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
7.8 Debug . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
7.9 Use Cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

8 Poll Mode Driver 42

8.1 Requirements and Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
8.2 Design Principles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
8.3 Logical Cores, Memory and NIC Queues Relationships . . . . . . . . . . . . . . 44
8.4 Device Identification and Configuration . . . . . . . . . . . . . . . . . . . . . . . 44
8.5 Poll Mode Driver API . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

9 IVSHMEM Library 48
9.1 IVHSHMEM Library API Overview . . . . . . . . . . . . . . . . . . . . . . . . . . 49
9.2 IVSHMEM Environment Configuration . . . . . . . . . . . . . . . . . . . . . . . . 49
9.3 Best Practices for Writing IVSHMEM Applications . . . . . . . . . . . . . . . . . 50
9.4 Best Practices for Running IVSHMEM Applications . . . . . . . . . . . . . . . . 50

10 Link Bonding Poll Mode Driver Library 51

10.1 Link Bonding Modes Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
10.2 Implementation Details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
10.3 Using Link Bonding Devices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

11 Timer Library 62
11.1 Implementation Details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
11.2 Use Cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
11.3 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

12 Hash Library 64
12.1 Hash API Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
12.2 Implementation Details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
12.3 Use Case: Flow Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
12.4 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

13 LPM Library 67
13.1 LPM API Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
13.2 Implementation Details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

14 LPM6 Library 71
14.1 LPM6 API Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
14.2 Use Case: IPv6 Forwarding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

15 Packet Distributor Library 76

15.1 Distributor Core Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
15.2 Worker Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

16 Reorder Library 79
16.1 Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
16.2 Implementation Details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

ii
 16.3 Use Case: Packet Distributor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

17 IP Fragmentation and Reassembly Library 81

17.1 Packet fragmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
17.2 Packet reassembly . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

18 Multi-process Support 84
18.1 Memory Sharing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
18.2 Deployment Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
18.3 Multi-process Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87

19 Kernel NIC Interface 88

19.1 The DPDK KNI Kernel Module . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
19.2 KNI Creation and Deletion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
19.3 DPDK mbuf Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
19.4 Use Case: Ingress . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
19.5 Use Case: Egress . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
19.6 Ethtool . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
19.7 Link state and MTU change . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
19.8 KNI Working as a Kernel vHost Backend . . . . . . . . . . . . . . . . . . . . . . 92

20 Thread Safety of DPDK Functions 95

20.1 Fast-Path APIs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
20.2 Performance Insensitive API . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
20.3 Library Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
20.4 Interrupt Thread . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96

21 Quality of Service (QoS) Framework 97

21.1 Packet Pipeline with QoS Support . . . . . . . . . . . . . . . . . . . . . . . . . . 97
21.2 Hierarchical Scheduler . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
21.3 Dropper . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
21.4 Traffic Metering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132

22 Power Management 134

22.1 CPU Frequency Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
22.2 Core-load Throttling through C-States . . . . . . . . . . . . . . . . . . . . . . . . 135
22.3 API Overview of the Power Library . . . . . . . . . . . . . . . . . . . . . . . . . . 135
22.4 User Cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
22.5 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

23 Packet Classification and Access Control 136

23.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
23.2 Application Programming Interface (API) Usage . . . . . . . . . . . . . . . . . . 141

24 Packet Framework 144

24.1 Design Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
24.2 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
24.3 Port Library Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
24.4 Table Library Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
24.5 Pipeline Library Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
24.6 Multicore Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
24.7 Interfacing with Accelerators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163

iii
25 Vhost Library 164
25.1 Vhost API Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
25.2 Vhost Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
25.3 Vhost supported vSwitch reference . . . . . . . . . . . . . . . . . . . . . . . . . 166

26 Port Hotplug Framework 167

26.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
26.2 Port Hotplug API overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
26.3 Reference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
26.4 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168

27 Source Organization 169

27.1 Makefiles and Config . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
27.2 Libraries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
27.3 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170

28 Development Kit Build System 171

28.1 Building the Development Kit Binary . . . . . . . . . . . . . . . . . . . . . . . . . 171
28.2 Building External Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
28.3 Makefile Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174

29 Development Kit Root Makefile Help 178

29.1 Configuration Targets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
29.2 Build Targets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
29.3 Install Targets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
29.4 Test Targets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
29.5 Documentation Targets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
29.6 Deps Targets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
29.7 Misc Targets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
29.8 Other Useful Command-line Variables . . . . . . . . . . . . . . . . . . . . . . . . 180
29.9 Make in a Build Directory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
29.10Compiling for Debug . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181

30 Extending the DPDK 182

30.1 Example: Adding a New Library libfoo . . . . . . . . . . . . . . . . . . . . . . . . 182

31 Building Your Own Application 184

31.1 Compiling a Sample Application in the Development Kit Directory . . . . . . . . 184
31.2 Build Your Own Application Outside the Development Kit . . . . . . . . . . . . . 184
31.3 Customizing Makefiles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185

32 External Application/Library Makefile help 186

32.1 Prerequisites . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
32.2 Build Targets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
32.3 Help Targets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
32.4 Other Useful Command-line Variables . . . . . . . . . . . . . . . . . . . . . . . . 187
32.5 Make from Another Directory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187

33 Performance Optimization Guidelines 188

33.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188

34 Writing Efficient Code 189

34.1 Memory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189

iv
 34.2 Communication Between lcores . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
34.3 PMD Driver . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
34.4 Locks and Atomic Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
34.5 Coding Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
34.6 Setting the Target CPU Type . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192

35 Profile Your Application 193

36 Glossary 194

v
 CHAPTER

ONE

INTRODUCTION

This document provides software architecture information, development environment informa-

tion and optimization guidelines.
For programming examples and for instructions on compiling and running each sample appli-
cation, see the DPDK Sample Applications User Guide for details.
For general information on compiling and running applications, see the DPDK Getting Started
Guide.

1.1 Documentation Roadmap

The following is a list of DPDK documents in the suggested reading order:

• Release Notes (this document): Provides release-specific information, including sup-
ported features, limitations, fixed issues, known issues and so on. Also, provides the
answers to frequently asked questions in FAQ format.
• Getting Started Guide : Describes how to install and configure the DPDK software;
designed to get users up and running quickly with the software.
• FreeBSD* Getting Started Guide : A document describing the use of the DPDK with
FreeBSD* has been added in DPDK Release 1.6.0. Refer to this guide for installation
and configuration instructions to get started using the DPDK with FreeBSD*.
• Programmer’s Guide (this document): Describes:
– The software architecture and how to use it (through examples), specifically in a
Linux* application (linuxapp) environment
– The content of the DPDK, the build system (including the commands that can be
used in the root DPDK Makefile to build the development kit and an application) and
guidelines for porting an application
– Optimizations used in the software and those that should be considered for new
development
A glossary of terms is also provided.
• API Reference : Provides detailed information about DPDK functions, data structures
and other programming constructs.
• Sample Applications User Guide: Describes a set of sample applications. Each chap-
ter describes a sample application that showcases specific functionality and provides
instructions on how to compile, run and use the sample application.

1
 Programmer’s Guide, Release 2.0.0

1.2 Related Publications

The following documents provide information that is relevant to the development of applications
using the DPDK:
• Intel® 64 and IA-32 Architectures Software Developer’s Manual Volume 3A: System Pro-
gramming Guide

1.2. Related Publications 2

 CHAPTER

TWO

OVERVIEW

This section gives a global overview of the architecture of Data Plane Development Kit (DPDK).
The main goal of the DPDK is to provide a simple, complete framework for fast packet process-
ing in data plane applications. Users may use the code to understand some of the techniques
employed, to build upon for prototyping or to add their own protocol stacks. Alternative ecosys-
tem options that use the DPDK are available.
The framework creates a set of libraries for specific environments through the creation of an
Environment Abstraction Layer (EAL), which may be specific to a mode of the Intel® architec-
ture (32-bit or 64-bit), Linux* user space compilers or a specific platform. These environments
are created through the use of make files and configuration files. Once the EAL library is cre-
ated, the user may link with the library to create their own applications. Other libraries, outside
of EAL, including the Hash, Longest Prefix Match (LPM) and rings libraries are also provided.
Sample applications are provided to help show the user how to use various features of the
DPDK.
The DPDK implements a run to completion model for packet processing, where all resources
must be allocated prior to calling Data Plane applications, running as execution units on logical
processing cores. The model does not support a scheduler and all devices are accessed by
polling. The primary reason for not using interrupts is the performance overhead imposed by
interrupt processing.
In addition to the run-to-completion model, a pipeline model may also be used by passing
packets or messages between cores via the rings. This allows work to be performed in stages
and may allow more efficient use of code on cores.

2.1 Development Environment

The DPDK project installation requires Linux and the associated toolchain, such as one or more
compilers, assembler, make utility, editor and various libraries to create the DPDK components
and libraries.
Once these libraries are created for the specific environment and architecture, they may then
be used to create the user’s data plane application.
When creating applications for the Linux user space, the glibc library is used. For DPDK
applications, two environmental variables (RTE_SDK and RTE_TARGET) must be configured
before compiling the applications. The following are examples of how the variables can be set:
export RTE_SDK=/home/user/DPDK
export RTE_TARGET=x86_64-native-linuxapp-gcc

3
 Programmer’s Guide, Release 2.0.0

See the DPDK Getting Started Guide for information on setting up the development environ-
ment.

2.2 Environment Abstraction Layer

The Environment Abstraction Layer (EAL) provides a generic interface that hides the environ-
ment specifics from the applications and libraries. The services provided by the EAL are:
• DPDK loading and launching
• Support for multi-process and multi-thread execution types
• Core affinity/assignment procedures
• System memory allocation/de-allocation
• Atomic/lock operations
• Time reference
• PCI bus access
• Trace and debug functions
• CPU feature identification
• Interrupt handling
• Alarm operations
The EAL is fully described in Environment Abstraction Layer .

2.3 Core Components

The core components are a set of libraries that provide all the elements needed for high-
performance packet processing applications.

2.3.1 Memory Manager (librte_malloc)

The librte_malloc library provides an API to allocate memory from the memzones created from
the hugepages instead of the heap. This helps when allocating large numbers of items that
may become susceptible to TLB misses when using typical 4k heap pages in the Linux user
space environment.
This memory allocator is fully described in Malloc Library .

2.3.2 Ring Manager (librte_ring)

The ring structure provides a lockless multi-producer, multi-consumer FIFO API in a finite size
table. It has some advantages over lockless queues; easier to implement, adapted to bulk
operations and faster. A ring is used by the Memory Pool Manager (librte_mempool) and
may be used as a general communication mechanism between cores and/or execution blocks
connected together on a logical core.

2.2. Environment Abstraction Layer 4

 Programmer’s Guide, Release 2.0.0

Manipulation of packet
buffers carrying network
X Y data.
X uses Y
Handle a pool of objects
using a ring to store rte_mbuf
them. Allow bulk
Timer facilities. Based enqueue/dequeue and
on HPET interface that per-CPU cache.
is provided by EAL.

rte_ring
rte_timer rte_mempool

Fixed-size lockless
FIFO for storing objects
in a table.
rte_malloc rte_eal + libc

Allocation of named Environment abstraction

memory zones using layer: RTE loading, memory rte_debug
libc's malloc() allocation, time reference,
PCI access, logging Provides debug helpers

Figure 2.1: Core Components Architecture

2.3. Core Components 5

 Programmer’s Guide, Release 2.0.0

This ring buffer and its usage are fully described in Ring Library .

2.3.3 Memory Pool Manager (librte_mempool)

The Memory Pool Manager is responsible for allocating pools of objects in memory. A pool
is identified by name and uses a ring to store free objects. It provides some other optional
services, such as a per-core object cache and an alignment helper to ensure that objects are
padded to spread them equally on all RAM channels.
This memory pool allocator is described in Mempool Library .

2.3.4 Network Packet Buffer Management (librte_mbuf)

The mbuf library provides the facility to create and destroy buffers that may be used by the
DPDK application to store message buffers. The message buffers are created at startup time
and stored in a mempool, using the DPDK mempool library.
This library provide an API to allocate/free mbufs, manipulate control message buffers (ctrlm-
buf) which are generic message buffers, and packet buffers (pktmbuf) which are used to carry
network packets.
Network Packet Buffer Management is described in Mbuf Library .

2.3.5 Timer Manager (librte_timer)

This library provides a timer service to DPDK execution units, providing the ability to execute
a function asynchronously. It can be periodic function calls, or just a one-shot call. It uses
the timer interface provided by the Environment Abstraction Layer (EAL) to get a precise time
reference and can be initiated on a per-core basis as required.
The library documentation is available in Timer Library .

2.4 Ethernet* Poll Mode Driver Architecture

The DPDK includes Poll Mode Drivers (PMDs) for 1 GbE, 10 GbE and 40GbE, and para virtu-
alized virtio Ethernet controllers which are designed to work without asynchronous, interrupt-
based signaling mechanisms.
See Poll Mode Driver .

2.5 Packet Forwarding Algorithm Support

The DPDK includes Hash (librte_hash) and Longest Prefix Match (LPM,librte_lpm) libraries to
support the corresponding packet forwarding algorithms.
See Hash Library and LPM Library for more information.

2.4. Ethernet* Poll Mode Driver Architecture 6

 Programmer’s Guide, Release 2.0.0

2.6 librte_net

The librte_net library is a collection of IP protocol definitions and convenience macros. It is

based on code from the FreeBSD* IP stack and contains protocol numbers (for use in IP
headers), IP-related macros, IPv4/IPv6 header structures and TCP, UDP and SCTP header
structures.

2.6. librte_net 7
 CHAPTER

THREE

ENVIRONMENT ABSTRACTION LAYER

The Environment Abstraction Layer (EAL) is responsible for gaining access to low-level re-
sources such as hardware and memory space. It provides a generic interface that hides the
environment specifics from the applications and libraries. It is the responsibility of the initial-
ization routine to decide how to allocate these resources (that is, memory space, PCI devices,
timers, consoles, and so on).
Typical services expected from the EAL are:
• DPDK Loading and Launching: The DPDK and its application are linked as a single
application and must be loaded by some means.
• Core Affinity/Assignment Procedures: The EAL provides mechanisms for assigning exe-
cution units to specific cores as well as creating execution instances.
• System Memory Reservation: The EAL facilitates the reservation of different memory
zones, for example, physical memory areas for device interactions.
• PCI Address Abstraction: The EAL provides an interface to access PCI address space.
• Trace and Debug Functions: Logs, dump_stack, panic and so on.
• Utility Functions: Spinlocks and atomic counters that are not provided in libc.
• CPU Feature Identification: Determine at runtime if a particular feature, for example,
Intel® AVX is supported. Determine if the current CPU supports the feature set that the
binary was compiled for.
• Interrupt Handling: Interfaces to register/unregister callbacks to specific interrupt
sources.
• Alarm Functions: Interfaces to set/remove callbacks to be run at a specific time.

3.1 EAL in a Linux-userland Execution Environment

In a Linux user space environment, the DPDK application runs as a user-space application
using the pthread library. PCI information about devices and address space is discovered
through the /sys kernel interface and through kernel modules such as uio_pci_generic, or
igb_uio. Refer to the UIO: User-space drivers documentation in the Linux kernel. This memory
is mmap’d in the application.
The EAL performs physical memory allocation using mmap() in hugetlbfs (using huge page
sizes to increase performance). This memory is exposed to DPDK service layers such as the
Mempool Library .

8
 Programmer’s Guide, Release 2.0.0

At this point, the DPDK services layer will be initialized, then through pthread setaffinity calls,
each execution unit will be assigned to a specific logical core to run as a user-level thread.
The time reference is provided by the CPU Time-Stamp Counter (TSC) or by the HPET kernel
API through a mmap() call.

3.1.1 Initialization and Core Launching

Part of the initialization is done by the start function of glibc. A check is also performed at
initialization time to ensure that the micro architecture type chosen in the config file is supported
by the CPU. Then, the main() function is called. The core initialization and launch is done
in rte_eal_init() (see the API documentation). It consist of calls to the pthread library (more
specifically, pthread_self(), pthread_create(), and pthread_setaffinity_np()).

Note: Initialization of objects, such as memory zones, rings, memory pools, lpm tables and
hash tables, should be done as part of the overall application initialization on the master lcore.
The creation and initialization functions for these objects are not multi-thread safe. However,
once initialized, the objects themselves can safely be used in multiple threads simultaneously.

3.1.2 Multi-process Support

The Linuxapp EAL allows a multi-process as well as a multi-threaded (pthread) deployment

model. See chapter 2.20 Multi-process Support for more details.

3.1.3 Memory Mapping Discovery and Memory Reservation

The allocation of large contiguous physical memory is done using the hugetlbfs kernel filesys-
tem. The EAL provides an API to reserve named memory zones in this contiguous memory.
The physical address of the reserved memory for that memory zone is also returned to the
user by the memory zone reservation API.

Note: Memory reservations done using the APIs provided by the rte_malloc library are also
backed by pages from the hugetlbfs filesystem. However, physical address information is not
available for the blocks of memory allocated in this way.

3.1.4 Xen Dom0 support without hugetbls

The existing memory management implementation is based on the Linux kernel hugepage
mechanism. However, Xen Dom0 does not support hugepages, so a new Linux kernel module
rte_dom0_mm is added to workaround this limitation.
The EAL uses IOCTL interface to notify the Linux kernel module rte_dom0_mm to allocate
memory of specified size, and get all memory segments information from the module, and
the EAL uses MMAP interface to map the allocated memory. For each memory segment,
the physical addresses are contiguous within it but actual hardware addresses are contiguous
within 2MB.

3.1. EAL in a Linux-userland Execution Environment 9

 Programmer’s Guide, Release 2.0.0

Master lcore lcore 1 lcore 2

main()

rte_eal_init()

rte_eal_memory_init()
rte_eal_logs_init()
rte_eal_pci_init()
...

pthread_create(1) per-thread init

pthread_create(2) per-thread init

wait

wait

wait all threads

other inits (libs, drivers)

rte_eal_remote_lauch(
per_lcore_app_init)

rte_eal_mp_wait_lcore()
per_lcore_ per_lcore_
app_init() app_init()

wait wait
rte_eal_remote_lauch(app)

application application application

... ... ...

Figure 3.1: EAL Initialization in a Linux Application Environment

3.1. EAL in a Linux-userland Execution Environment 10

 Programmer’s Guide, Release 2.0.0

3.1.5 PCI Access

The EAL uses the /sys/bus/pci utilities provided by the kernel to scan the content on the PCI
bus. To access PCI memory, a kernel module called uio_pci_generic provides a /dev/uioX
device file and resource files in /sys that can be mmap’d to obtain access to PCI address
space from the application. The DPDK-specific igb_uio module can also be used for this. Both
drivers use the uio kernel feature (userland driver).

3.1.6 Per-lcore and Shared Variables

Note: lcore refers to a logical execution unit of the processor, sometimes called a hardware
thread.

Shared variables are the default behavior. Per-lcore variables are implemented using Thread
Local Storage (TLS) to provide per-thread local storage.

3.1.7 Logs

A logging API is provided by EAL. By default, in a Linux application, logs are sent to syslog and
also to the console. However, the log function can be overridden by the user to use a different
logging mechanism.

Trace and Debug Functions

There are some debug functions to dump the stack in glibc. The rte_panic() function can
voluntarily provoke a SIG_ABORT, which can trigger the generation of a core file, readable by
gdb.

3.1.8 CPU Feature Identification

The EAL can query the CPU at runtime (using the rte_cpu_get_feature() function) to determine
which CPU features are available.

3.1.9 User Space Interrupt and Alarm Handling

The EAL creates a host thread to poll the UIO device file descriptors to detect the interrupts.
Callbacks can be registered or unregistered by the EAL functions for a specific interrupt event
and are called in the host thread asynchronously. The EAL also allows timed callbacks to be
used in the same way as for NIC interrupts.

Note: The only interrupts supported by the DPDK Poll-Mode Drivers are those for link status
change, i.e. link up and link down notification.

3.1. EAL in a Linux-userland Execution Environment 11

 Programmer’s Guide, Release 2.0.0

3.1.10 Blacklisting

The EAL PCI device blacklist functionality can be used to mark certain NIC ports as blacklisted,
so they are ignored by the DPDK. The ports to be blacklisted are identified using the PCIe*
description (Domain:Bus:Device.Function).

3.1.11 Misc Functions

Locks and atomic operations are per-architecture (i686 and x86_64).

3.2 Memory Segments and Memory Zones (memzone)

The mapping of physical memory is provided by this feature in the EAL. As physical memory
can have gaps, the memory is described in a table of descriptors, and each descriptor (called
rte_memseg ) describes a contiguous portion of memory.
On top of this, the memzone allocator’s role is to reserve contiguous portions of physical mem-
ory. These zones are identified by a unique name when the memory is reserved.
The rte_memzone descriptors are also located in the configuration structure. This structure is
accessed using rte_eal_get_configuration(). The lookup (by name) of a memory zone returns
a descriptor containing the physical address of the memory zone.
Memory zones can be reserved with specific start address alignment by supplying the align
parameter (by default, they are aligned to cache line size). The alignment value should be a
power of two and not less than the cache line size (64 bytes). Memory zones can also be
reserved from either 2 MB or 1 GB hugepages, provided that both are available on the system.

3.3 Multiple pthread

DPDK usually pins one pthread per core to avoid the overhead of task switching. This allows
for significant performance gains, but lacks flexibility and is not always efficient.
Power management helps to improve the CPU efficiency by limiting the CPU runtime frequency.
However, alternately it is possible to utilize the idle cycles available to take advantage of the
full capability of the CPU.
By taking advantage of cgroup, the CPU utilization quota can be simply assigned. This gives
another way to improve the CPU efficiency, however, there is a prerequisite; DPDK must handle
the context switching between multiple pthreads per core.
For further flexibility, it is useful to set pthread affinity not only to a CPU but to a CPU set.

3.3.1 EAL pthread and lcore Affinity

The term “lcore” refers to an EAL thread, which is really a Linux/FreeBSD pthread. “EAL
pthreads” are created and managed by EAL and execute the tasks issued by remote_launch.
In each EAL pthread, there is a TLS (Thread Local Storage) called _lcore_id for unique identi-
fication. As EAL pthreads usually bind 1:1 to the physical CPU, the _lcore_id is typically equal
to the CPU ID.

3.2. Memory Segments and Memory Zones (memzone) 12

 Programmer’s Guide, Release 2.0.0

When using multiple pthreads, however, the binding is no longer always 1:1 between an EAL
pthread and a specified physical CPU. The EAL pthread may have affinity to a CPU set, and
as such the _lcore_id will not be the same as the CPU ID. For this reason, there is an EAL
long option ‘–lcores’ defined to assign the CPU affinity of lcores. For a specified lcore ID or ID
group, the option allows setting the CPU set for that EAL pthread.
The format pattern: –lcores=’<lcore_set>[@cpu_set][,<lcore_set>[@cpu_set],...]’
‘lcore_set’ and ‘cpu_set’ can be a single number, range or a group.
A number is a “digit([0-9]+)”; a range is “<number>-<number>”; a group is “(<num-
ber|range>[,<number|range>,...])”.
If a ‘@cpu_set’ value is not supplied, the value of ‘cpu_set’ will default to the value of ‘lcore_set’.
For example, "--lcores='1,2@(5-7),(3-5)@(0,2),(0,6),7-8'"
which means start 9 EAL thread;
lcore 0 runs on cpuset 0x41 (cpu 0,6);
lcore 1 runs on cpuset 0x2 (cpu 1);
lcore 2 runs on cpuset 0xe0 (cpu 5,6,7);
lcore 3,4,5 runs on cpuset 0x5 (cpu 0,2);
lcore 6 runs on cpuset 0x41 (cpu 0,6);
lcore 7 runs on cpuset 0x80 (cpu 7);
lcore 8 runs on cpuset 0x100 (cpu 8).

Using this option, for each given lcore ID, the associated CPUs can be assigned. It’s also
compatible with the pattern of corelist(‘-l’) option.

3.3.2 non-EAL pthread support

It is possible to use the DPDK execution context with any user pthread (aka. Non-EAL
pthreads). In a non-EAL pthread, the _lcore_id is always LCORE_ID_ANY which identifies
that it is not an EAL thread with a valid, unique, _lcore_id. Some libraries will use an alter-
native unique ID (e.g. TID), some will not be impacted at all, and some will work but with
limitations (e.g. timer and mempool libraries).
All these impacts are mentioned in Known Issues section.

3.3.3 Public Thread API

There are two public APIs rte_thread_set_affinity() and

rte_pthread_get_affinity() introduced for threads. When they’re used in any
pthread context, the Thread Local Storage(TLS) will be set/get.
Those TLS include _cpuset and _socket_id:
• _cpuset stores the CPUs bitmap to which the pthread is affinitized.
• _socket_id stores the NUMA node of the CPU set. If the CPUs in CPU set belong to
different NUMA node, the _socket_id will be set to SOCKET_ID_ANY.

3.3.4 Known Issues

• rte_mempool

3.3. Multiple pthread 13

 Programmer’s Guide, Release 2.0.0

The rte_mempool uses a per-lcore cache inside the mempool. For non-EAL pthreads,
rte_lcore_id() will not return a valid number. So for now, when rte_mempool is used
with non-EAL pthreads, the put/get operations will bypass the mempool cache and there
is a performance penalty because of this bypass. Support for non-EAL mempool cache
is currently being enabled.
• rte_ring
rte_ring supports multi-producer enqueue and multi-consumer dequeue. However, it is
non-preemptive, this has a knock on effect of making rte_mempool non-preemptable.

Note: The “non-preemptive” constraint means:

– a pthread doing multi-producers enqueues on a given ring must not be preempted
by another pthread doing a multi-producer enqueue on the same ring.
– a pthread doing multi-consumers dequeues on a given ring must not be preempted
by another pthread doing a multi-consumer dequeue on the same ring.
Bypassing this constraint it may cause the 2nd pthread to spin until the 1st one is sched-
uled again. Moreover, if the 1st pthread is preempted by a context that has an higher
priority, it may even cause a dead lock.

This does not mean it cannot be used, simply, there is a need to narrow down the situation
when it is used by multi-pthread on the same core.
1. It CAN be used for any single-producer or single-consumer situation.
2. It MAY be used by multi-producer/consumer pthread whose scheduling policy are all
SCHED_OTHER(cfs). User SHOULD be aware of the performance penalty before
using it.
3. It MUST not be used by multi-producer/consumer pthreads, whose scheduling poli-
cies are SCHED_FIFO or SCHED_RR.
RTE_RING_PAUSE_REP_COUNT is defined for rte_ring to reduce contention. It’s mainly
for case 2, a yield is issued after number of times pause repeat.
It adds a sched_yield() syscall if the thread spins for too long while waiting on the other
thread to finish its operations on the ring. This gives the preempted thread a chance to
proceed and finish with the ring enqueue/dequeue operation.
• rte_timer
Running rte_timer_manager() on a non-EAL pthread is not allowed. However, re-
setting/stopping the timer from a non-EAL pthread is allowed.
• rte_log
In non-EAL pthreads, there is no per thread loglevel and logtype, global loglevels are
used.
• misc
The debug statistics of rte_ring, rte_mempool and rte_timer are not supported in a non-
EAL pthread.

3.3. Multiple pthread 14

 Programmer’s Guide, Release 2.0.0

3.3.5 cgroup control

The following is a simple example of cgroup control usage, there are two pthreads(t0 and t1)
doing packet I/O on the same core ($CPU). We expect only 50% of CPU spend on packet IO.
mkdir /sys/fs/cgroup/cpu/pkt_io
mkdir /sys/fs/cgroup/cpuset/pkt_io

echo $cpu > /sys/fs/cgroup/cpuset/cpuset.cpus

echo $t0 > /sys/fs/cgroup/cpu/pkt_io/tasks

echo $t0 > /sys/fs/cgroup/cpuset/pkt_io/tasks

echo $t1 > /sys/fs/cgroup/cpu/pkt_io/tasks

echo $t1 > /sys/fs/cgroup/cpuset/pkt_io/tasks

cd /sys/fs/cgroup/cpu/pkt_io
echo 100000 > pkt_io/cpu.cfs_period_us
echo 50000 > pkt_io/cpu.cfs_quota_us

3.3. Multiple pthread 15

 CHAPTER

FOUR

MALLOC LIBRARY

The librte_malloc library provides an API to allocate any-sized memory.

The objective of this library is to provide malloc-like functions to allow allocation from hugepage
memory and to facilitate application porting. The DPDK API Reference manual describes the
available functions.
Typically, these kinds of allocations should not be done in data plane processing because they
are slower than pool-based allocation and make use of locks within the allocation and free
paths. However, they can be used in configuration code.
Refer to the rte_malloc() function description in the DPDK API Reference manual for more
information.

4.1 Cookies

When CONFIG_RTE_MALLOC_DEBUG is enabled, the allocated memory contains overwrite

protection fields to help identify buffer overflows.

4.2 Alignment and NUMA Constraints

The rte_malloc() takes an align argument that can be used to request a memory area that is
aligned on a multiple of this value (which must be a power of two).
On systems with NUMA support, a call to the rte_malloc() function will return memory that has
been allocated on the NUMA socket of the core which made the call. A set of APIs is also
provided, to allow memory to be explicitly allocated on a NUMA socket directly, or by allocated
on the NUMA socket where another core is located, in the case where the memory is to be
used by a logical core other than on the one doing the memory allocation.

4.3 Use Cases

This library is needed by an application that requires malloc-like functions at initialization time,
and does not require the physical address information for the individual memory blocks.
For allocating/freeing data at runtime, in the fast-path of an application, the memory pool library
should be used instead.

16
 Programmer’s Guide, Release 2.0.0

If a block of memory with a known physical address is needed, e.g. for use by a hardware
device, a memory zone should be used.

4.4 Internal Implementation

4.4.1 Data Structures

There are two data structure types used internally in the malloc library:
• struct malloc_heap - used to track free space on a per-socket basis
• struct malloc_elem - the basic element of allocation and free-space tracking inside the
library.

Structure: malloc_heap

The malloc_heap structure is used in the library to manage free space on a per-socket basis.
Internally in the library, there is one heap structure per NUMA node, which allows us to allocate
memory to a thread based on the NUMA node on which this thread runs. While this does not
guarantee that the memory will be used on that NUMA node, it is no worse than a scheme
where the memory is always allocated on a fixed or random node.
The key fields of the heap structure and their function are described below (see also diagram
above):
• mz_count - field to count the number of memory zones which have been allocated for
heap memory on this NUMA node. The sole use of this value is, in combination with the
numa_socket value, to generate a suitable, unique name for each memory zone.
• lock - the lock field is needed to synchronize access to the heap. Given that the free
space in the heap is tracked using a linked list, we need a lock to prevent two threads
manipulating the list at the same time.
• free_head - this points to the first element in the list of free nodes for this malloc heap.

Note: The malloc_heap structure does not keep track of either the memzones allocated, since
there is little point as they cannot be freed. Neither does it track the in-use blocks of memory,
since these are never touched except when they are to be freed again - at which point the
pointer to the block is an input to the free() function.

Structure: malloc_elem

The malloc_elem structure is used as a generic header structure for various blocks of memory
in a memzone. It is used in three different ways - all shown in the diagram above:
1. As a header on a block of free or allocated memory - normal case
2. As a padding header inside a block of memory
3. As an end-of-memzone marker

4.4. Internal Implementation 17

 Programmer’s Guide, Release 2.0.0

Figure 4.1: Example of a malloc heap and malloc elements within the malloc library

4.4. Internal Implementation 18

 Programmer’s Guide, Release 2.0.0

The most important fields in the structure and how they are used are described below.

Note: If the usage of a particular field in one of the above three usages is not described, the
field can be assumed to have an undefined value in that situation, for example, for padding
headers only the “state” and “pad” fields have valid values.

• heap - this pointer is a reference back to the heap structure from which this block was
allocated. It is used for normal memory blocks when they are being freed, to add the
newly-freed block to the heap’s free-list.
• prev - this pointer points to the header element/block in the memzone immediately behind
the current one. When freeing a block, this pointer is used to reference the previous block
to check if that block is also free. If so, then the two free blocks are merged to form a
single larger block.
• next_free - this pointer is used to chain the free-list of unallocated memory blocks to-
gether. Again, it is only used in normal memory blocks - on malloc() to find a suitable free
block to allocate, and on free() to add the newly freed element to the free-list.
• state - This field can have one of three values: “Free”, “Busy” or “Pad”. The former two,
are to indicate the allocation state of a normal memory block, and the latter is to indicate
that the element structure is a dummy structure at the end of the start-of-block padding
(i.e. where the start of the data within a block is not at the start of the block itself, due to
alignment constraints). In this case, the pad header is used to locate the actual malloc
element header for the block. For the end-of-memzone structure, this is always a “busy”
value, which ensures that no element, on being freed, searches beyond the end of the
memzone for other blocks to merge with into a larger free area.
• pad - this holds the length of the padding present at the start of the block. In the case
of a normal block header, it is added to the address of the end of the header to give the
address of the start of the data area i.e. the value passed back to the application on
a malloc. Within a dummy header inside the padding, this same value is stored, and is
subtracted from the address of the dummy header to yield the address of the actual block
header.
• size - the size of the data block, including the header itself. For end-of-memzone struc-
tures, this size is given as zero, though it is never actually checked. For normal blocks
which are being freed, this size value is used in place of a “next” pointer to identify the
location of the next block of memory (so that if it too is free, the two free blocks can be
merged into one).

4.4.2 Memory Allocation

When an application makes a call to a malloc-like function, the malloc function will first index
the lcore_config structure for the calling thread, and determine the NUMA node idea of that
thread. That is used to index the array of malloc_heap structures, and the heap_alloc () func-
tion is called with that heap as parameter, along with the requested size, type and alignment
parameters.
The heap_alloc() function will scan the free_list for the heap, and attempt to find a free block
suitable for storing data of the requested size, with the requested alignment constraints. If no
suitable block is found - for example, the first time malloc is called for a node, and the free-
list is NULL - a new memzone is reserved and set up as heap elements. The setup involves

4.4. Internal Implementation 19

 Programmer’s Guide, Release 2.0.0

placing a dummy structure at the end of the memzone to act as a sentinel to prevent accesses
beyond the end (as the sentinel is marked as BUSY, the malloc library code will never attempt
to reference it further), and a proper element header at the start of the memzone. This latter
header identifies all space in the memzone, bar the sentinel value at the end, as a single free
heap element, and it is then added to the free_list for the heap.
Once the new memzone has been set up, the scan of the free-list for the heap is redone,
and on this occasion should find the newly created, suitable element as the size of memory
reserved in the memzone is set to be at least the size of the requested data block plus the
alignment - subject to a minimum size specified in the DPDK compile-time configuration.
When a suitable, free element has been identified, the pointer to be returned to the user is
calculated, with the space to be provided to the user being at the end of the free block. The
cache-line of memory immediately preceding this space is filled with a struct malloc_elem
header: if the remaining space within the block is small e.g. <=128 bytes, then a pad header
is used, and the remaining space is wasted. If, however, the remaining space is greater than
this, then the single free element block is split into two, and a new, proper, malloc_elem header
is put before the returned data space. [The advantage of allocating the memory from the end
of the existing element is that in this case no adjustment of the free list needs to take place -
the existing element on the free list just has its size pointer adjusted, and the following element
has its “prev” pointer redirected to the newly created element].

4.4.3 Freeing Memory

To free an area of memory, the pointer to the start of the data area is passed to the free function.
The size of the malloc_elem structure is subtracted from this pointer to get the element header
for the block. If this header is of type “PAD” then the pad length is further subtracted from the
pointer to get the proper element header for the entire block.
From this element header, we get pointers to the heap from which the block came – and to
where it must be freed, as well as the pointer to the previous element, and, via the size field,
we can calculate the pointer to the next element. These next and previous elements are then
checked to see if they too are free, and if so, they are merged with the current elements. This
means that we can never have two free memory blocks adjacent to one another, they are
always merged into a single block.

4.4. Internal Implementation 20

 CHAPTER

FIVE

RING LIBRARY

The ring allows the management of queues. Instead of having a linked list of infinite size, the
rte_ring has the following properties:
• FIFO
• Maximum size is fixed, the pointers are stored in a table
• Lockless implementation
• Multi-consumer or single-consumer dequeue
• Multi-producer or single-producer enqueue
• Bulk dequeue - Dequeues the specified count of objects if successful; otherwise fails
• Bulk enqueue - Enqueues the specified count of objects if successful; otherwise fails
• Burst dequeue - Dequeue the maximum available objects if the specified count cannot
be fulfilled
• Burst enqueue - Enqueue the maximum available objects if the specified count cannot
be fulfilled
The advantages of this data structure over a linked list queue are as follows:
• Faster; only requires a single Compare-And-Swap instruction of sizeof(void *) instead of
several double-Compare-And-Swap instructions.
• Simpler than a full lockless queue.
• Adapted to bulk enqueue/dequeue operations. As pointers are stored in a table, a de-
queue of several objects will not produce as many cache misses as in a linked queue.
Also, a bulk dequeue of many objects does not cost more than a dequeue of a simple
object.
The disadvantages:
• Size is fixed
• Having many rings costs more in terms of memory than a linked list queue. An empty
ring contains at least N pointers.
A simplified representation of a Ring is shown in with consumer and producer head and tail
pointers to objects stored in the data structure.

21
 Programmer’s Guide, Release 2.0.0

obj1 obj2 obj3

cons_head prod_head
cons_tail prod_tail

Figure 5.1: Ring Structure

5.1 References for Ring Implementation in FreeBSD*

The following code was added in FreeBSD 8.0, and is used in some network device drivers (at
least in Intel drivers):
• bufring.h in FreeBSD
• bufring.c in FreeBSD

5.2 Lockless Ring Buffer in Linux*

The following is a link describing the Linux Lockless Ring Buffer Design.

5.3 Additional Features

5.3.1 Name

A ring is identified by a unique name. It is not possible to create two rings with the same name
(rte_ring_create() returns NULL if this is attempted).

5.3.2 Water Marking

The ring can have a high water mark (threshold). Once an enqueue operation reaches the high
water mark, the producer is notified, if the water mark is configured.
This mechanism can be used, for example, to exert a back pressure on I/O to inform the LAN
to PAUSE.

5.3.3 Debug

When debug is enabled (CONFIG_RTE_LIBRTE_RING_DEBUG is set), the library stores

some per-ring statistic counters about the number of enqueues/dequeues. These statistics
are per-core to avoid concurrent accesses or atomic operations.

5.1. References for Ring Implementation in FreeBSD* 22

 Programmer’s Guide, Release 2.0.0

5.4 Use Cases

Use cases for the Ring library include:

• Communication between applications in the DPDK
• Used by memory pool allocator

5.5 Anatomy of a Ring Buffer

This section explains how a ring buffer operates. The ring structure is composed of two head
and tail couples; one is used by producers and one is used by the consumers. The figures of
the following sections refer to them as prod_head, prod_tail, cons_head and cons_tail.
Each figure represents a simplified state of the ring, which is a circular buffer. The content
of the function local variables is represented on the top of the figure, and the content of ring
structure is represented on the bottom of the figure.

5.5.1 Single Producer Enqueue

This section explains what occurs when a producer adds an object to the ring. In this example,
only the producer head and tail (prod_head and prod_tail) are modified, and there is only one
producer.
The initial state is to have a prod_head and prod_tail pointing at the same location.

Enqueue First Step

First, ring->prod_head and ring->cons_tail are copied in local variables. The prod_next lo-
cal variable points to the next element of the table, or several elements after in case of bulk
enqueue.
If there is not enough room in the ring (this is detected by checking cons_tail), it returns an
error.

Enqueue Second Step

The second step is to modify ring->prod_head in ring structure to point to the same location
as prod_next.
A pointer to the added object is copied in the ring (obj4).

Enqueue Last Step

Once the object is added in the ring, ring->prod_tail in the ring structure is modified to point to
the same location as ring->prod_head. The enqueue operation is finished.

5.4. Use Cases 23

 Programmer’s Guide, Release 2.0.0

local variables

cons_tail prod_head prod_next

obj1 obj2 obj3

cons_head prod_head
cons_tail prod_tail

structure state

Figure 5.2: Enqueue first step

local variables

cons_tail prod_head prod_next

obj1 obj2 obj3 obj4

cons_head prod_tail prod_head

cons_tail

structure state

Figure 5.3: Enqueue second step

5.5. Anatomy of a Ring Buffer 24

 Programmer’s Guide, Release 2.0.0

local variables

cons_tail prod_head prod_next

obj1 obj2 obj3 obj4

cons_head prod_tail
cons_tail prod_head

structure state

Figure 5.4: Enqueue last step

5.5.2 Single Consumer Dequeue

This section explains what occurs when a consumer dequeues an object from the ring. In this
example, only the consumer head and tail (cons_head and cons_tail) are modified and there
is only one consumer.
The initial state is to have a cons_head and cons_tail pointing at the same location.

Dequeue First Step

First, ring->cons_head and ring->prod_tail are copied in local variables. The cons_next local
variable points to the next element of the table, or several elements after in the case of bulk
dequeue.
If there are not enough objects in the ring (this is detected by checking prod_tail), it returns an
error.

Dequeue Second Step

The second step is to modify ring->cons_head in the ring structure to point to the same location
as cons_next.
The pointer to the dequeued object (obj1) is copied in the pointer given by the user.

5.5. Anatomy of a Ring Buffer 25

 Programmer’s Guide, Release 2.0.0

local variables

cons_head cons_next prod_tail

obj1 obj2 obj3 obj4

cons_head prod_tail
cons_tail prod_head

structure state

Figure 5.5: Dequeue last step

local variables

cons_head cons_next prod_tail

obj2 obj3 obj4

cons_tail cons_head prod_tail

prod_head

structure state
Figure 5.6: Dequeue second step

5.5. Anatomy of a Ring Buffer 26

 Programmer’s Guide, Release 2.0.0

Dequeue Last Step

Finally, ring->cons_tail in the ring structure is modified to point to the same location as ring-
>cons_head. The dequeue operation is finished.

local variables

cons_head cons_next prod_tail

obj2 obj3 obj4

cons_head prod_tail
cons_tail prod_head

structure state

Figure 5.7: Dequeue last step

5.5.3 Multiple Producers Enqueue

This section explains what occurs when two producers concurrently add an object to the ring.
In this example, only the producer head and tail (prod_head and prod_tail) are modified.
The initial state is to have a prod_head and prod_tail pointing at the same location.

Multiple Consumer Enqueue First Step

On both cores, ring->prod_head and ring->cons_tail are copied in local variables. The
prod_next local variable points to the next element of the table, or several elements after in
the case of bulk enqueue.
If there is not enough room in the ring (this is detected by checking cons_tail), it returns an
error.

5.5. Anatomy of a Ring Buffer 27

 Programmer’s Guide, Release 2.0.0

local variables cons_tail prod_head prod_next

core 2

local variables cons_tail prod_head prod_next

core 1

obj1 obj2 obj3

cons_head prod_head
cons_tail prod_tail

structure state

Figure 5.8: Multiple consumer enqueue first step

Multiple Consumer Enqueue Second Step

The second step is to modify ring->prod_head in the ring structure to point to the same location
as prod_next. This operation is done using a Compare And Swap (CAS) instruction, which
does the following operations atomically:
• If ring->prod_head is different to local variable prod_head, the CAS operation fails, and
the code restarts at first step.
• Otherwise, ring->prod_head is set to local prod_next, the CAS operation is successful,
and processing continues.
In the figure, the operation succeeded on core 1, and step one restarted on core 2.

Multiple Consumer Enqueue Third Step

The CAS operation is retried on core 2 with success.

The core 1 updates one element of the ring(obj4), and the core 2 updates another one (obj5).

Multiple Consumer Enqueue Fourth Step

Each core now wants to update ring->prod_tail. A core can only update it if ring->prod_tail is
equal to the prod_head local variable. This is only true on core 1. The operation is finished on
core 1.

5.5. Anatomy of a Ring Buffer 28

 Programmer’s Guide, Release 2.0.0

compare and swap succeeds

on core 1 and fails on core 2
local variables cons_tail prod_head prod_next
core 2

local variables cons_tail prod_head prod_next

core 1

obj1 obj2 obj3

cons_head prod_head
cons_tail prod_tail

structure state

Figure 5.9: Multiple consumer enqueue second step

Multiple Consumer Enqueue Last Step

Once ring->prod_tail is updated by core 1, core 2 is allowed to update it too. The operation is
also finished on core 2.

5.5.4 Modulo 32-bit Indexes

In the preceding figures, the prod_head, prod_tail, cons_head and cons_tail indexes are repre-
sented by arrows. In the actual implementation, these values are not between 0 and size(ring)-
1 as would be assumed. The indexes are between 0 and 2^32 -1, and we mask their value
when we access the pointer table (the ring itself). 32-bit modulo also implies that operations
on indexes (such as, add/subtract) will automatically do 2^32 modulo if the result overflows the
32-bit number range.
The following are two examples that help to explain how indexes are used in a ring.

Note: To simplify the explanation, operations with modulo 16-bit are used instead of modulo
32-bit. In addition, the four indexes are defined as unsigned 16-bit integers, as opposed to
unsigned 32-bit integers in the more realistic case.

This ring contains 11000 entries.

This ring contains 12536 entries.

5.5. Anatomy of a Ring Buffer 29

 Programmer’s Guide, Release 2.0.0

compare and swap succeeds

on core 2
local variables cons_tail prod_head prod_next
core 2

local variables cons_tail prod_head prod_next

core 1

obj1 obj2 obj3 obj4 obj5

cons_head prod_head
cons_tail prod_tail

structure state

Figure 5.10: Multiple consumer enqueue third step

Note: For ease of understanding, we use modulo 65536 operations in the above examples.
In real execution cases, this is redundant for low efficiency, but is done automatically when the
result overflows.

The code always maintains a distance between producer and consumer between 0 and
size(ring)-1. Thanks to this property, we can do subtractions between 2 index values in a
modulo-32bit base: that’s why the overflow of the indexes is not a problem.
At any time, entries and free_entries are between 0 and size(ring)-1, even if only the first term
of subtraction has overflowed:
uint32_t entries = (prod_tail - cons_head);
uint32_t free_entries = (mask + cons_tail -prod_head);

5.6 References

• bufring.h in FreeBSD (version 8)

• bufring.c in FreeBSD (version 8)
• Linux Lockless Ring Buffer Design

5.6. References 30
 Programmer’s Guide, Release 2.0.0

core 2 is waiting for

r->prod_tail == prod_head
local variables cons_tail prod_head prod_next
core 2

local variables cons_tail prod_head prod_next

core 1

obj1 obj2 obj3 obj4 obj5

cons_head prod_head
cons_tail prod_tail

structure state

Figure 5.11: Multiple consumer enqueue fourth step

local variables cons_tail prod_head prod_next

core 2

obj1 obj2 obj3 obj4 obj5

cons_head prod_head
cons_tail prod_tail

structure state

Figure 5.12: Multiple consumer enqueue last step

5.6. References 31
 Programmer’s Guide, Release 2.0.0

0 0
0 16384 32768 49152 65536 16384 32768 49152 65536 16384

value for
ring indexes
(prod_head,
used_entries prod_tail, ...)
size = 16384
ch ph
mask = 16383
ct pt
ph = pt = 14000
ct = ch = 3000 used entries in ring
used_entries = (pt - ch) % 65536 = 11000
free_entries = (mask + ct - ph) % 65536 = 5383

Figure 5.13: Modulo 32-bit indexes - Example 1

0 0
0 16384 32768 49152 65536 16384 32768 49152 65536 16384

value for
ring indexes
(prod_head,
used_entries prod_tail, ...)
size = 16384
ch ph
mask = 16383
ct pt
ph = pt = 6000
ct = ch = 59000 used entries in ring
used_entries = (pt - ch) % 65536 = 12536
free_entries = (mask + ct - ph) % 65536 = 3847

Figure 5.14: Modulo 32-bit indexes - Example 2

5.6. References 32
 CHAPTER

SIX

MEMPOOL LIBRARY

A memory pool is an allocator of a fixed-sized object. In the DPDK, it is identified by name

and uses a ring to store free objects. It provides some other optional services such as a per-
core object cache and an alignment helper to ensure that objects are padded to spread them
equally on all DRAM or DDR3 channels.
This library is used by the Mbuf Library and the Environment Abstraction Layer (for logging
history).

6.1 Cookies

In debug mode (CONFIG_RTE_LIBRTE_MEMPOOL_DEBUG is enabled), cookies are added

at the beginning and end of allocated blocks. The allocated objects then contain overwrite
protection fields to help debugging buffer overflows.

6.2 Stats

In debug mode (CONFIG_RTE_LIBRTE_MEMPOOL_DEBUG is enabled), statistics about get

from/put in the pool are stored in the mempool structure. Statistics are per-lcore to avoid
concurrent access to statistics counters.

6.3 Memory Alignment Constraints

Depending on hardware memory configuration, performance can be greatly improved by

adding a specific padding between objects. The objective is to ensure that the beginning of
each object starts on a different channel and rank in memory so that all channels are equally
loaded.
This is particularly true for packet buffers when doing L3 forwarding or flow classification. Only
the first 64 bytes are accessed, so performance can be increased by spreading the start ad-
dresses of objects among the different channels.
The number of ranks on any DIMM is the number of independent sets of DRAMs that can be
accessed for the full data bit-width of the DIMM. The ranks cannot be accessed simultaneously
since they share the same data path. The physical layout of the DRAM chips on the DIMM itself
does not necessarily relate to the number of ranks.

33
 Programmer’s Guide, Release 2.0.0

When running an application, the EAL command line options provide the ability to add the
number of memory channels and ranks.

Note: The command line must always have the number of memory channels specified for the
processor.

Examples of alignment for different DIMM architectures are shown in Fig. 6.1 and Fig. 6.2.
memory addresses 64 bytes wide

Block num 0 1 2 3 4 5 6 7 8 9 A B C D E F 10 11 12 13 14 15 ...

Channel 0 1 0 1 0 1 0 1
Rank 0 1 2 3 0 1 2 3 0 1 2

0 1 2 3 4 5 6 7 8 9 A B C D E F 0 1 2 3
packet 1 padding packet 2

pkt1 starts at pkt2 starts at

channel 0, rank 0 channel 1, rank 1

Figure 6.1: Two Channels and Quad-ranked DIMM Example

In this case, the assumption is that a packet is 16 blocks of 64 bytes, which is not true.
The Intel® 5520 chipset has three channels, so in most cases, no padding is required between
objects (except for objects whose size are n x 3 x 64 bytes blocks).
memory addresses 64 bytes wide

Block num 0 1 2 3 4 5 6 7 8 9 A B C D E F 10 11 12 13 14 15 ...

Channel 0 1 2 0 1 1 2 0 1
Rank 0 1 0 1 0 1 0 1 0 1 0
DIMM 0 1 0 1 0 1
0 1 2 3 4 5 6 7 8 9 A B C D E F 0 1 2 3
packet 1 packet 2

pkt0 starts at pkt2 starts at

channel 0, rank 1 channel 1, rank 0
(no padding needed)

Figure 6.2: Three Channels and Two Dual-ranked DIMM Example

When creating a new pool, the user can specify to use this feature or not.

6.4 Local Cache

In terms of CPU usage, the cost of multiple cores accessing a memory pool’s ring of free
buffers may be high since each access requires a compare-and-set (CAS) operation. To avoid
having too many access requests to the memory pool’s ring, the memory pool allocator can
maintain a per-core cache and do bulk requests to the memory pool’s ring, via the cache with
many fewer locks on the actual memory pool structure. In this way, each core has full access
to its own cache (with locks) of free objects and only when the cache fills does the core need to
shuffle some of the free objects back to the pools ring or obtain more objects when the cache
is empty.
While this may mean a number of buffers may sit idle on some core’s cache, the speed at
which a core can access its own cache for a specific memory pool without locks provides
performance gains.

6.4. Local Cache 34

 Programmer’s Guide, Release 2.0.0

The cache is composed of a small, per-core table of pointers and its length (used as a stack).
This cache can be enabled or disabled at creation of the pool.
The maximum size of the cache is static and is defined at compilation time (CON-
FIG_RTE_MEMPOOL_CACHE_MAX_SIZE).
Fig. 6.3 shows a cache in operation.

Object caches for header trailer

Core 0
App A - ring
core 0 obj 0

Core 1 If cache empty get from ring

App B - ring core 1 if cache full move to ring elt_size

App C - ring obj 1

obj 2

obj n

rte_ring: stores memory pool's free objects

mempool

Figure 6.3: A mempool in Memory with its Associated Ring

6.5 Use Cases

All allocations that require a high level of performance should use a pool-based memory allo-
cator. Below are some examples:
• Mbuf Library
• Environment Abstraction Layer , for logging service
• Any application that needs to allocate fixed-sized objects in the data plane and that will
be continuously utilized by the system.

6.5. Use Cases 35

 CHAPTER

SEVEN

MBUF LIBRARY

The mbuf library provides the ability to allocate and free buffers (mbufs) that may be used by
the DPDK application to store message buffers. The message buffers are stored in a mempool,
using the Mempool Library .
A rte_mbuf struct can carry network packet buffers or generic control buffers (indicated by the
CTRL_MBUF_FLAG). This can be extended to other types. The rte_mbuf header structure is
kept as small as possible and currently uses just two cache lines, with the most frequently used
fields being on the first of the two cache lines.

7.1 Design of Packet Buffers

For the storage of the packet data (including protocol headers), two approaches were consid-
ered:
1. Embed metadata within a single memory buffer the structure followed by a fixed size area
for the packet data.
2. Use separate memory buffers for the metadata structure and for the packet data.
The advantage of the first method is that it only needs one operation to allocate/free the whole
memory representation of a packet. On the other hand, the second method is more flexible
and allows the complete separation of the allocation of metadata structures from the allocation
of packet data buffers.
The first method was chosen for the DPDK. The metadata contains control information such as
message type, length, offset to the start of the data and a pointer for additional mbuf structures
allowing buffer chaining.
Message buffers that are used to carry network packets can handle buffer chaining where
multiple buffers are required to hold the complete packet. This is the case for jumbo frames
that are composed of many mbufs linked together through their next field.
For a newly allocated mbuf, the area at which the data begins in the message buffer is
RTE_PKTMBUF_HEADROOM bytes after the beginning of the buffer, which is cache aligned.
Message buffers may be used to carry control information, packets, events, and so on between
different entities in the system. Message buffers may also use their buffer pointers to point to
other message buffer data sections or other structures.
Fig. 7.1 and Fig. 7.2 show some of these scenarios.
The Buffer Manager implements a fairly standard set of buffer access functions to manipulate
network packets.

36
 Programmer’s Guide, Release 2.0.0

rte_pktmbuf_mtod(m)

mbuf
struct
headroom tailroom

m->pkt.next = NULL rte_pktmbuf_pktlen(m)

or rte_pktmbuf_datalen(m)

m->buf_addr
(m->buf_physaddr is the
corresponding physical address)
struct rte_mbuf

Figure 7.1: An mbuf with One Segment

rte_pktmbuf_pktlen(m) = rte_pktmbuf_datalen(m) +
rte_pktmbuf_datalen(mseg2) + rte_pktmbuf_datalen(mseg3)

rte_pktmbuf_mtod(m)

m mseg2 mseg3

rte_pktmbuf_datalen(m) rte_pktmbuf_datalen(m) rte_pktmbuf_datalen(m)

m->pkt.next = mseg2 m->pkt.next = mseg3 m->pkt.next = NULL

multi-segmented rte_mbuf

Figure 7.2: An mbuf with Three Segments

7.1. Design of Packet Buffers 37

 Programmer’s Guide, Release 2.0.0

7.2 Buffers Stored in Memory Pools

The Buffer Manager uses the Mempool Library to allocate buffers. Therefore, it ensures
that the packet header is interleaved optimally across the channels and ranks for L3 pro-
cessing. An mbuf contains a field indicating the pool that it originated from. When calling
rte_ctrlmbuf_free(m) or rte_pktmbuf_free(m), the mbuf returns to its original pool.

7.3 Constructors

Packet and control mbuf constructors are provided by the API. The rte_pktmbuf_init() and
rte_ctrlmbuf_init() functions initialize some fields in the mbuf structure that are not modified by
the user once created (mbuf type, origin pool, buffer start address, and so on). This function is
given as a callback function to the rte_mempool_create() function at pool creation time.

7.4 Allocating and Freeing mbufs

Allocating a new mbuf requires the user to specify the mempool from which the mbuf
should be taken. For any newly-allocated mbuf, it contains one segment, with a length
of 0. The offset to data is initialized to have some bytes of headroom in the buffer
(RTE_PKTMBUF_HEADROOM).
Freeing a mbuf means returning it into its original mempool. The content of an mbuf is not
modified when it is stored in a pool (as a free mbuf). Fields initialized by the constructor do not
need to be re-initialized at mbuf allocation.
When freeing a packet mbuf that contains several segments, all of them are freed and returned
to their original mempool.

7.5 Manipulating mbufs

This library provides some functions for manipulating the data in a packet mbuf. For instance:
• Get data length
• Get a pointer to the start of data
• Prepend data before data
• Append data after data
• Remove data at the beginning of the buffer (rte_pktmbuf_adj())
• Remove data at the end of the buffer (rte_pktmbuf_trim()) Refer to the DPDK API Refer-
ence for details.

7.6 Meta Information

Some information is retrieved by the network driver and stored in an mbuf to make process-
ing easier. For instance, the VLAN, the RSS hash result (see Poll Mode Driver ) and a flag
indicating that the checksum was computed by hardware.

7.2. Buffers Stored in Memory Pools 38

 Programmer’s Guide, Release 2.0.0

An mbuf also contains the input port (where it comes from), and the number of segment mbufs
in the chain.
For chained buffers, only the first mbuf of the chain stores this meta information.
For instance, this is the case on RX side for the IEEE1588 packet timestamp mechanism, the
VLAN tagging and the IP checksum computation.
On TX side, it is also possible for an application to delegate some processing to the hardware
if it supports it. For instance, the PKT_TX_IP_CKSUM flag allows to offload the computation
of the IPv4 checksum.
The following examples explain how to configure different TX offloads on a vxlan-encapsulated
tcp packet: out_eth/out_ip/out_udp/vxlan/in_eth/in_ip/in_tcp/payload
• calculate checksum of out_ip:
mb->l2_len = len(out_eth)
mb->l3_len = len(out_ip)
mb->ol_flags |= PKT_TX_IPV4 | PKT_TX_IP_CSUM
set out_ip checksum to 0 in the packet

This is supported on hardware advertising DEV_TX_OFFLOAD_IPV4_CKSUM.

• calculate checksum of out_ip and out_udp:
mb->l2_len = len(out_eth)
mb->l3_len = len(out_ip)
mb->ol_flags |= PKT_TX_IPV4 | PKT_TX_IP_CSUM | PKT_TX_UDP_CKSUM
set out_ip checksum to 0 in the packet
set out_udp checksum to pseudo header using rte_ipv4_phdr_cksum()

This is supported on hardware advertising DEV_TX_OFFLOAD_IPV4_CKSUM

and DEV_TX_OFFLOAD_UDP_CKSUM.

• calculate checksum of in_ip:

mb->l2_len = len(out_eth + out_ip + out_udp + vxlan + in_eth)
mb->l3_len = len(in_ip)
mb->ol_flags |= PKT_TX_IPV4 | PKT_TX_IP_CSUM
set in_ip checksum to 0 in the packet

This is similar to case 1), but l2_len is different. It is supported on hardware advertising
DEV_TX_OFFLOAD_IPV4_CKSUM. Note that it can only work if outer L4 checksum is
0.
• calculate checksum of in_ip and in_tcp:
mb->l2_len = len(out_eth + out_ip + out_udp + vxlan + in_eth)
mb->l3_len = len(in_ip)
mb->ol_flags |= PKT_TX_IPV4 | PKT_TX_IP_CSUM | PKT_TX_TCP_CKSUM
set in_ip checksum to 0 in the packet
set in_tcp checksum to pseudo header using rte_ipv4_phdr_cksum()

This is similar to case 2), but l2_len is different. It is supported on hardware advertising
DEV_TX_OFFLOAD_IPV4_CKSUM and DEV_TX_OFFLOAD_TCP_CKSUM. Note that
it can only work if outer L4 checksum is 0.
• segment inner TCP:

7.6. Meta Information 39

 Programmer’s Guide, Release 2.0.0

mb->l2_len = len(out_eth + out_ip + out_udp + vxlan + in_eth)

mb->l3_len = len(in_ip)
mb->l4_len = len(in_tcp)
mb->ol_flags |= PKT_TX_IPV4 | PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM |
PKT_TX_TCP_SEG;
set in_ip checksum to 0 in the packet
set in_tcp checksum to pseudo header without including the IP
payload length using rte_ipv4_phdr_cksum()

This is supported on hardware advertising DEV_TX_OFFLOAD_TCP_TSO. Note that it

can only work if outer L4 checksum is 0.
• calculate checksum of out_ip, in_ip, in_tcp:
mb->outer_l2_len = len(out_eth)
mb->outer_l3_len = len(out_ip)
mb->l2_len = len(out_udp + vxlan + in_eth)
mb->l3_len = len(in_ip)
mb->ol_flags |= PKT_TX_OUTER_IPV4 | PKT_TX_OUTER_IP_CKSUM | \
PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM;
set out_ip checksum to 0 in the packet
set in_ip checksum to 0 in the packet
set in_tcp checksum to pseudo header using rte_ipv4_phdr_cksum()

This is supported on hardware advertising DEV_TX_OFFLOAD_IPV4_CKSUM,

DEV_TX_OFFLOAD_UDP_CKSUM and DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM.

The list of flags and their precise meaning is described in the mbuf API documentation
(rte_mbuf.h). Also refer to the testpmd source code (specifically the csumonly.c file) for de-
tails.

7.7 Direct and Indirect Buffers

A direct buffer is a buffer that is completely separate and self-contained. An indirect buffer
behaves like a direct buffer but for the fact that the buffer pointer and data offset in it refer to
data in another direct buffer. This is useful in situations where packets need to be duplicated
or fragmented, since indirect buffers provide the means to reuse the same packet data across
multiple buffers.
A buffer becomes indirect when it is “attached” to a direct buffer using the rte_pktmbuf_attach()
function. Each buffer has a reference counter field and whenever an indirect buffer is attached
to the direct buffer, the reference counter on the direct buffer is incremented. Similarly, when-
ever the indirect buffer is detached, the reference counter on the direct buffer is decremented.
If the resulting reference counter is equal to 0, the direct buffer is freed since it is no longer in
use.
There are a few things to remember when dealing with indirect buffers. First of all, it is not
possible to attach an indirect buffer to another indirect buffer. Secondly, for a buffer to become
indirect, its reference counter must be equal to 1, that is, it must not be already referenced by
another indirect buffer. Finally, it is not possible to reattach an indirect buffer to the direct buffer
(unless it is detached first).
While the attach/detach operations can be invoked directly using the recommended
rte_pktmbuf_attach() and rte_pktmbuf_detach() functions, it is suggested to use the higher-
level rte_pktmbuf_clone() function, which takes care of the correct initialization of an indirect

7.7. Direct and Indirect Buffers 40

 Programmer’s Guide, Release 2.0.0

buffer and can clone buffers with multiple segments.

Since indirect buffers are not supposed to actually hold any data, the memory pool for indirect
buffers should be configured to indicate the reduced memory consumption. Examples of the
initialization of a memory pool for indirect buffers (as well as use case examples for indirect
buffers) can be found in several of the sample applications, for example, the IPv4 Multicast
sample application.

7.8 Debug

In debug mode (CONFIG_RTE_MBUF_DEBUG is enabled), the functions of the mbuf library

perform sanity checks before any operation (such as, buffer corruption, bad type, and so on).

7.9 Use Cases

All networking application should use mbufs to transport network packets.

7.8. Debug 41
 CHAPTER

EIGHT

POLL MODE DRIVER

The DPDK includes 1 Gigabit, 10 Gigabit and 40 Gigabit and para virtualized virtio Poll Mode
Drivers.
A Poll Mode Driver (PMD) consists of APIs, provided through the BSD driver running in user
space, to configure the devices and their respective queues. In addition, a PMD accesses the
RX and TX descriptors directly without any interrupts (with the exception of Link Status Change
interrupts) to quickly receive, process and deliver packets in the user’s application. This section
describes the requirements of the PMDs, their global design principles and proposes a high-
level architecture and a generic external API for the Ethernet PMDs.

8.1 Requirements and Assumptions

The DPDK environment for packet processing applications allows for two models, run-to-
completion and pipe-line:
• In the run-to-completion model, a specific port’s RX descriptor ring is polled for packets
through an API. Packets are then processed on the same core and placed on a port’s TX
descriptor ring through an API for transmission.
• In the pipe-line model, one core polls one or more port’s RX descriptor ring through
an API. Packets are received and passed to another core via a ring. The other core
continues to process the packet which then may be placed on a port’s TX descriptor ring
through an API for transmission.
In a synchronous run-to-completion model, each logical core assigned to the DPDK executes
a packet processing loop that includes the following steps:
• Retrieve input packets through the PMD receive API
• Process each received packet one at a time, up to its forwarding
• Send pending output packets through the PMD transmit API
Conversely, in an asynchronous pipe-line model, some logical cores may be dedicated to the
retrieval of received packets and other logical cores to the processing of previously received
packets. Received packets are exchanged between logical cores through rings. The loop for
packet retrieval includes the following steps:
• Retrieve input packets through the PMD receive API
• Provide received packets to processing lcores through packet queues
The loop for packet processing includes the following steps:

42
 Programmer’s Guide, Release 2.0.0

• Retrieve the received packet from the packet queue

• Process the received packet, up to its retransmission if forwarded
To avoid any unnecessary interrupt processing overhead, the execution environment must not
use any asynchronous notification mechanisms. Whenever needed and appropriate, asyn-
chronous communication should be introduced as much as possible through the use of rings.
Avoiding lock contention is a key issue in a multi-core environment. To address this issue,
PMDs are designed to work with per-core private resources as much as possible. For example,
a PMD maintains a separate transmit queue per-core, per-port. In the same way, every receive
queue of a port is assigned to and polled by a single logical core (lcore).
To comply with Non-Uniform Memory Access (NUMA), memory management is designed to
assign to each logical core a private buffer pool in local memory to minimize remote memory
access. The configuration of packet buffer pools should take into account the underlying physi-
cal memory architecture in terms of DIMMS, channels and ranks. The application must ensure
that appropriate parameters are given at memory pool creation time. See Mempool Library .

8.2 Design Principles

The API and architecture of the Ethernet* PMDs are designed with the following guidelines in
mind.
PMDs must help global policy-oriented decisions to be enforced at the upper application level.
Conversely, NIC PMD functions should not impede the benefits expected by upper-level global
policies, or worse prevent such policies from being applied.
For instance, both the receive and transmit functions of a PMD have a maximum number of
packets/descriptors to poll. This allows a run-to-completion processing stack to statically fix or
to dynamically adapt its overall behavior through different global loop policies, such as:
• Receive, process immediately and transmit packets one at a time in a piecemeal fashion.
• Receive as many packets as possible, then process all received packets, transmitting
them immediately.
• Receive a given maximum number of packets, process the received packets, accumulate
them and finally send all accumulated packets to transmit.
To achieve optimal performance, overall software design choices and pure software optimiza-
tion techniques must be considered and balanced against available low-level hardware-based
optimization features (CPU cache properties, bus speed, NIC PCI bandwidth, and so on). The
case of packet transmission is an example of this software/hardware tradeoff issue when opti-
mizing burst-oriented network packet processing engines. In the initial case, the PMD could ex-
port only an rte_eth_tx_one function to transmit one packet at a time on a given queue. On top
of that, one can easily build an rte_eth_tx_burst function that loops invoking the rte_eth_tx_one
function to transmit several packets at a time. However, an rte_eth_tx_burst function is effec-
tively implemented by the PMD to minimize the driver-level transmit cost per packet through
the following optimizations:
• Share among multiple packets the un-amortized cost of invoking the rte_eth_tx_one func-
tion.
• Enable the rte_eth_tx_burst function to take advantage of burst-oriented hardware fea-
tures (prefetch data in cache, use of NIC head/tail registers) to minimize the number of

8.2. Design Principles 43

 Programmer’s Guide, Release 2.0.0

CPU cycles per packet, for example by avoiding unnecessary read memory accesses
to ring transmit descriptors, or by systematically using arrays of pointers that exactly fit
cache line boundaries and sizes.
• Apply burst-oriented software optimization techniques to remove operations that would
otherwise be unavoidable, such as ring index wrap back management.
Burst-oriented functions are also introduced via the API for services that are intensively used
by the PMD. This applies in particular to buffer allocators used to populate NIC rings, which
provide functions to allocate/free several buffers at a time. For example, an mbuf_multiple_alloc
function returning an array of pointers to rte_mbuf buffers which speeds up the receive poll
function of the PMD when replenishing multiple descriptors of the receive ring.

8.3 Logical Cores, Memory and NIC Queues Relationships

The DPDK supports NUMA allowing for better performance when a processor’s logical cores
and interfaces utilize its local memory. Therefore, mbuf allocation associated with local PCIe*
interfaces should be allocated from memory pools created in the local memory. The buffers
should, if possible, remain on the local processor to obtain the best performance results and RX
and TX buffer descriptors should be populated with mbufs allocated from a mempool allocated
from local memory.
The run-to-completion model also performs better if packet or data manipulation is in local
memory instead of a remote processors memory. This is also true for the pipe-line model
provided all logical cores used are located on the same processor.
Multiple logical cores should never share receive or transmit queues for interfaces since this
would require global locks and hinder performance.

8.4 Device Identification and Configuration

8.4.1 Device Identification

Each NIC port is uniquely designated by its (bus/bridge, device, function) PCI identifiers as-
signed by the PCI probing/enumeration function executed at DPDK initialization. Based on
their PCI identifier, NIC ports are assigned two other identifiers:
• A port index used to designate the NIC port in all functions exported by the PMD API.
• A port name used to designate the port in console messages, for administration or de-
bugging purposes. For ease of use, the port name includes the port index.

8.4.2 Device Configuration

The configuration of each NIC port includes the following operations:

• Allocate PCI resources
• Reset the hardware (issue a Global Reset) to a well-known default state
• Set up the PHY and the link
• Initialize statistics counters

8.3. Logical Cores, Memory and NIC Queues Relationships 44

 Programmer’s Guide, Release 2.0.0

The PMD API must also export functions to start/stop the all-multicast feature of a port and
functions to set/unset the port in promiscuous mode.
Some hardware offload features must be individually configured at port initialization through
specific configuration parameters. This is the case for the Receive Side Scaling (RSS) and
Data Center Bridging (DCB) features for example.

8.4.3 On-the-Fly Configuration

All device features that can be started or stopped “on the fly” (that is, without stopping the
device) do not require the PMD API to export dedicated functions for this purpose.
All that is required is the mapping address of the device PCI registers to implement the config-
uration of these features in specific functions outside of the drivers.
For this purpose, the PMD API exports a function that provides all the information associated
with a device that can be used to set up a given device feature outside of the driver. This
includes the PCI vendor identifier, the PCI device identifier, the mapping address of the PCI
device registers, and the name of the driver.
The main advantage of this approach is that it gives complete freedom on the choice of the
API used to configure, to start, and to stop such features.
As an example, refer to the configuration of the IEEE1588 feature for the Intel® 82576 Giga-
bit Ethernet Controller and the Intel® 82599 10 Gigabit Ethernet Controller controllers in the
testpmd application.
Other features such as the L3/L4 5-Tuple packet filtering feature of a port can be configured in
the same way. Ethernet* flow control (pause frame) can be configured on the individual port.
Refer to the testpmd source code for details. Also, L4 (UDP/TCP/ SCTP) checksum offload by
the NIC can be enabled for an individual packet as long as the packet mbuf is set up correctly.
See Hardware Offload for details.

8.4.4 Configuration of Transmit and Receive Queues

Each transmit queue is independently configured with the following information:

• The number of descriptors of the transmit ring
• The socket identifier used to identify the appropriate DMA memory zone from which to
allocate the transmit ring in NUMA architectures
• The values of the Prefetch, Host and Write-Back threshold registers of the transmit queue
• The minimum transmit packets to free threshold (tx_free_thresh). When the number of
descriptors used to transmit packets exceeds this threshold, the network adaptor should
be checked to see if it has written back descriptors. A value of 0 can be passed during
the TX queue configuration to indicate the default value should be used. The default
value for tx_free_thresh is 32. This ensures that the PMD does not search for completed
descriptors until at least 32 have been processed by the NIC for this queue.
• The minimum RS bit threshold. The minimum number of transmit descriptors to use be-
fore setting the Report Status (RS) bit in the transmit descriptor. Note that this parameter
may only be valid for Intel 10 GbE network adapters. The RS bit is set on the last de-
scriptor used to transmit a packet if the number of descriptors used since the last RS bit

8.4. Device Identification and Configuration 45

 Programmer’s Guide, Release 2.0.0

setting, up to the first descriptor used to transmit the packet, exceeds the transmit RS
bit threshold (tx_rs_thresh). In short, this parameter controls which transmit descriptors
are written back to host memory by the network adapter. A value of 0 can be passed
during the TX queue configuration to indicate that the default value should be used. The
default value for tx_rs_thresh is 32. This ensures that at least 32 descriptors are used
before the network adapter writes back the most recently used descriptor. This saves
upstream PCIe* bandwidth resulting from TX descriptor write-backs. It is important to
note that the TX Write-back threshold (TX wthresh) should be set to 0 when tx_rs_thresh
is greater than 1. Refer to the Intel® 82599 10 Gigabit Ethernet Controller Datasheet for
more details.
The following constraints must be satisfied for tx_free_thresh and tx_rs_thresh:
• tx_rs_thresh must be greater than 0.
• tx_rs_thresh must be less than the size of the ring minus 2.
• tx_rs_thresh must be less than or equal to tx_free_thresh.
• tx_free_thresh must be greater than 0.
• tx_free_thresh must be less than the size of the ring minus 3.
• For optimal performance, TX wthresh should be set to 0 when tx_rs_thresh is greater
than 1.
One descriptor in the TX ring is used as a sentinel to avoid a hardware race condition, hence
the maximum threshold constraints.

Note: When configuring for DCB operation, at port initialization, both the number of transmit
queues and the number of receive queues must be set to 128.

8.4.5 Hardware Offload

Depending on driver capabilities advertised by rte_eth_dev_info_get(), the PMD may

support hardware offloading feature like checksumming, TCP segmentation or VLAN insertion.
The support of these offload features implies the addition of dedicated status bit(s) and value
field(s) into the rte_mbuf data structure, along with their appropriate handling by the re-
ceive/transmit functions exported by each PMD. The list of flags and their precise meaning
is described in the mbuf API documentation and in the in Mbuf Library , section “Meta Informa-
tion”.

8.5 Poll Mode Driver API

8.5.1 Generalities

By default, all functions exported by a PMD are lock-free functions that are assumed not to be
invoked in parallel on different logical cores to work on the same target object. For instance,
a PMD receive function cannot be invoked in parallel on two logical cores to poll the same RX
queue of the same port. Of course, this function can be invoked in parallel by different logical
cores on different RX queues. It is the responsibility of the upper-level application to enforce
this rule.

8.5. Poll Mode Driver API 46

 Programmer’s Guide, Release 2.0.0

If needed, parallel accesses by multiple logical cores to shared queues can be explicitly pro-
tected by dedicated inline lock-aware functions built on top of their corresponding lock-free
functions of the PMD API.

8.5.2 Generic Packet Representation

A packet is represented by an rte_mbuf structure, which is a generic metadata structure con-

taining all necessary housekeeping information. This includes fields and status bits corre-
sponding to offload hardware features, such as checksum computation of IP headers or VLAN
tags.
The rte_mbuf data structure includes specific fields to represent, in a generic way, the offload
features provided by network controllers. For an input packet, most fields of the rte_mbuf
structure are filled in by the PMD receive function with the information contained in the receive
descriptor. Conversely, for output packets, most fields of rte_mbuf structures are used by the
PMD transmit function to initialize transmit descriptors.
The mbuf structure is fully described in the Mbuf Library chapter.

8.5.3 Ethernet Device API

The Ethernet device API exported by the Ethernet PMDs is described in the DPDK API Refer-
ence.

8.5. Poll Mode Driver API 47

 CHAPTER

NINE

IVSHMEM LIBRARY

The DPDK IVSHMEM library facilitates fast zero-copy data sharing among virtual machines
(host-to-guest or guest-to-guest) by means of QEMU’s IVSHMEM mechanism.
The library works by providing a command line for QEMU to map several hugepages into a
single IVSHMEM device. For the guest to know what is inside any given IVSHMEM device
(and to distinguish between DPDK and non-DPDK IVSHMEM devices), a metadata file is also
mapped into the IVSHMEM segment. No work needs to be done by the guest application to
map IVSHMEM devices into memory; they are automatically recognized by the DPDK Envi-
ronment Abstraction Layer (EAL).
A typical DPDK IVSHMEM use case looks like the following.

Figure 9.1: Typical Ivshmem use case

The same could work with several virtual machines, providing host-to-VM or VM-to-VM com-

48
 Programmer’s Guide, Release 2.0.0

munication. The maximum number of metadata files is 32 (by default) and each metadata file
can contain different (or even the same) hugepages. The only constraint is that each VM has
to have access to the memory it is sharing with other entities (be it host or another VM). For
example, if the user wants to share the same memzone across two VMs, each VM must have
that memzone in its metadata file.

9.1 IVHSHMEM Library API Overview

The following is a simple guide to using the IVSHMEM Library API:

• Call rte_ivshmem_metadata_create() to create a new metadata file. The metadata name
is used to distinguish between multiple metadata files.
• Populate each metadata file with DPDK data structures. This can be done using the
following API calls:
– rte_ivhshmem_metadata_add_memzone() to add rte_memzone to metadata file
– rte_ivshmem_metadata_add_ring() to add rte_ring to metadata file
– rte_ivshmem_metadata_add_mempool() to add rte_mempool to metadata file
• Finally, call rte_ivshmem_metadata_cmdline_generate() to generate the command line
for QEMU. Multiple metadata files (and thus multiple command lines) can be supplied to
a single VM.

Note: Only data structures fully residing in DPDK hugepage memory work correctly. Sup-
ported data structures created by malloc(), mmap() or otherwise using non-DPDK memory
cause undefined behavior and even a segmentation fault.

9.2 IVSHMEM Environment Configuration

The steps needed to successfully run IVSHMEM applications are the following:
• Compile a special version of QEMU from sources.
The source code can be found on the QEMU website (currently, version 1.4.x is sup-
ported, but version 1.5.x is known to work also), however, the source code will need to
be patched to support using regular files as the IVSHMEM memory backend. The patch
is not included in the DPDK package, but is available on the Intel®DPDK-vswitch project
webpage (either separately or in a DPDK vSwitch package).
• Enable IVSHMEM library in the DPDK build configuration.
In the default configuration, IVSHMEM library is not compiled. To compile the IVSH-
MEM library, one has to either use one of the provided IVSHMEM targets (for example,
x86_64-ivshmem-linuxapp-gcc), or set CONFIG_RTE_LIBRTE_IVSHMEM to “y” in the
build configuration.
• Set up hugepage memory on the virtual machine.
The guest applications run as regular DPDK (primary) processes and thus need their own
hugepage memory set up inside the VM. The process is identical to the one described in
the DPDK Getting Started Guide.

9.1. IVHSHMEM Library API Overview 49

 Programmer’s Guide, Release 2.0.0

9.3 Best Practices for Writing IVSHMEM Applications

When considering the use of IVSHMEM for sharing memory, security implications need to be
carefully evaluated. IVSHMEM is not suitable for untrusted guests, as IVSHMEM is essentially
a window into the host process memory. This also has implications for the multiple VM scenar-
ios. While the IVSHMEM library tries to share as little memory as possible, it is quite probable
that data designated for one VM might also be present in an IVSMHMEM device designated
for another VM. Consequently, any shared memory corruption will affect both host and all VMs
sharing that particular memory.
IVSHMEM applications essentially behave like multi-process applications, so it is important to
implement access serialization to data and thread safety. DPDK ring structures are already
thread-safe, however, any custom data structures that the user might need would have to be
thread-safe also.
Similar to regular DPDK multi-process applications, it is not recommended to use function
pointers as functions might have different memory addresses in different processes.
It is best to avoid freeing the rte_mbuf structure on a different machine from where it was
allocated, that is, if the mbuf was allocated on the host, the host should free it. Consequently,
any packet transmission and reception should also happen on the same machine (whether
virtual or physical). Failing to do so may lead to data corruption in the mempool cache.
Despite the IVSHMEM mechanism being zero-copy and having good performance, it is still
desirable to do processing in batches and follow other procedures described in Performance
Optimization.

9.4 Best Practices for Running IVSHMEM Applications

For performance reasons, it is best to pin host processes and QEMU processes to different
cores so that they do not interfere with each other. If NUMA support is enabled, it is also
desirable to keep host process’ hugepage memory and QEMU process on the same NUMA
node.
For the best performance across all NUMA nodes, each QEMU core should be pinned to host
CPU core on the appropriate NUMA node. QEMU’s virtual NUMA nodes should also be set
up to correspond to physical NUMA nodes. More on how to set up DPDK and QEMU NUMA
support can be found in DPDK Getting Started Guide and QEMU documentation respectively.
A script called cpu_layout.py is provided with the DPDK package (in the tools directory) that
can be used to identify which CPU cores correspond to which NUMA node.
The QEMU IVSHMEM command line creation should be considered the last step before start-
ing the virtual machine. Currently, there is no hot plug support for QEMU IVSHMEM devices,
so one cannot add additional memory to an IVSHMEM device once it has been created. There-
fore, the correct sequence to run an IVSHMEM application is to run host application first, obtain
the command lines for each IVSHMEM device and then run all QEMU instances with guest ap-
plications afterwards.
It is important to note that once QEMU is started, it holds on to the hugepages it uses for
IVSHMEM devices. As a result, if the user wishes to shut down or restart the IVSHMEM host
application, it is not enough to simply shut the application down. The virtual machine must also
be shut down (if not, it will hold onto outdated host data).

9.3. Best Practices for Writing IVSHMEM Applications 50

 CHAPTER

TEN

LINK BONDING POLL MODE DRIVER LIBRARY

In addition to Poll Mode Drivers (PMDs) for physical and virtual hardware, DPDK also includes
a pure-software library that allows physical PMD’s to be bonded together to create a single
logical PMD.

User Application

DPDK
bonded ethdev

ethdev ethdev ethdev ethdev ethdev

port port port port port

Figure 10.1: Bonded PMDs

The Link Bonding PMD library(librte_pmd_bond) supports bonding of groups of rte_eth_dev

ports of the same speed and duplex to provide similar the capabilities to that found in Linux
bonding driver to allow the aggregation of multiple (slave) NICs into a single logical interface
between a server and a switch. The new bonded PMD will then process these interfaces based
on the mode of operation specified to provide support for features such as redundant links, fault
tolerance and/or load balancing.
The librte_pmd_bond library exports a C API which provides an API for the creation of bonded
devices as well as the configuration and management of the bonded device and its slave
devices.

Note: The Link Bonding PMD Library is enabled by default in the build configuration files, the
library can be disabled by setting CONFIG_RTE_LIBRTE_PMD_BOND=n and recompiling the
DPDK.

10.1 Link Bonding Modes Overview

Currently the Link Bonding PMD library supports 4 modes of operation:

• Round-Robin (Mode 0):

51
 Programmer’s Guide, Release 2.0.0

User Application

5
4
3
2
DPDK
1
bonded ethdev

ethdev port ethdev port ethdev port

4 5
1 2 3

Figure 10.2: Round-Robin (Mode 0)

This mode provides load balancing and fault tolerance by transmission of packets in se-
quential order from the first available slave device through the last. Packets are bulk de-
queued from devices then serviced in a round-robin manner. This mode does not guaran-
tee in order reception of packets and down stream should be able to handle out of order
packets.

10.1. Link Bonding Modes Overview 52

 Programmer’s Guide, Release 2.0.0

• Active Backup (Mode 1):

User Application

3
2
DPDK
1
bonded ethdev

ethdev port ethdev port ethdev port

3
2
1

Figure 10.3: Active Backup (Mode 1)

In this mode only one slave in the bond is active at any time, a different slave becomes
active if, and only if, the primary active slave fails, thereby providing fault tolerance to slave
failure. The single logical bonded interface’s MAC address is externally visible on only one
NIC (port) to avoid confusing the network switch.

• Balance XOR (Mode 2):

Note: The coloring differences of the packets are used to identify different flow classification
calculated by the selected transmit policy

• Broadcast (Mode 3):

• Link Aggregation 802.3AD (Mode 4):
• Transmit Load Balancing (Mode 5):

10.2 Implementation Details

The librte_pmd_bond bonded device are compatible with the Ethernet device API exported by
the Ethernet PMDs described in the DPDK API Reference.
The Link Bonding Library supports the creation of bonded devices at application startup time
during EAL initialization using the --vdev option as well as programmatically via the C API
rte_eth_bond_create function.
Bonded devices support the dynamical addition and removal of slave devices using the
rte_eth_bond_slave_add / rte_eth_bond_slave_remove APIs.
After a slave device is added to a bonded device slave is stopped using rte_eth_dev_stop
and then reconfigured using rte_eth_dev_configure the RX and TX queues are also

10.2. Implementation Details 53

 Programmer’s Guide, Release 2.0.0

User Application
6
5
4
3
2 DPDK
1
bonded ethdev

ethdev port ethdev port ethdev port

5
3 4
1 2 6

Figure 10.4: Balance XOR (Mode 2)

This mode provides transmit load balancing (based on the selected transmission policy)
and fault tolerance. The default policy (layer2) uses a simple calculation based on the
packet flow source and destination MAC addresses as well as the number of active slaves
available to the bonded device to classify the packet to a specific slave to transmit on. Alter-
nate transmission policies supported are layer 2+3, this takes the IP source and destination
addresses into the calculation of the transmit slave port and the final supported policy is
layer 3+4, this uses IP source and destination addresses as well as the TCP/UDP source
and destination port.

10.2. Implementation Details 54

 Programmer’s Guide, Release 2.0.0

User Application

3
2
DPDK
1
bonded ethdev

ethdev port ethdev port ethdev port

3 3 3
2 2 2
1 1 1

Figure 10.5: Broadcast (Mode 3)

This mode provides fault tolerance by transmission of packets on all slave ports.

reconfigured using rte_eth_tx_queue_setup / rte_eth_rx_queue_setup with the pa-

rameters use to configure the bonding device.

10.2.1 Link Status Change Interrupts / Polling

Link bonding devices support the registration of a link status change callback, using the
rte_eth_dev_callback_register API, this will be called when the status of the bond-
ing device changes. For example in the case of a bonding device which has 3 slaves, the link
status will change to up when one slave becomes active or change to down when all slaves
become inactive. There is no callback notification when a single slave changes state and the
previous conditions are not met. If a user wishes to monitor individual slaves then they must
register callbacks with that slave directly.
The link bonding library also supports devices which do not implement link status change
interrupts, this is achieve by polling the devices link status at a defined period which is
set using the rte_eth_bond_link_monitoring_set API, the default polling interval is
10ms. When a device is added as a slave to a bonding device it is determined using the
RTE_PCI_DRV_INTR_LSC flag whether the device supports interrupts or whether the link sta-
tus should be monitored by polling it.

10.2.2 Requirements / Limitations

The current implementation only supports devices that support the same speed and duplex to
be added as a slaves to the same bonded device. The bonded device inherits these attributes
from the first active slave added to the bonded device and then all further slaves added to the
bonded device must support these parameters.

10.2. Implementation Details 55

 Programmer’s Guide, Release 2.0.0

User Application
6
5
4
3
2 DPDK
1
bonded ethdev

ethdev port ethdev port ethdev port

5
3 O
O 4 6
1 2 O

Figure 10.6: Link Aggregation 802.3AD (Mode 4)

This mode provides dynamic link aggregation according to the 802.3ad specification. It
negotiates and monitors aggregation groups that share the same speed and duplex settings
using the selected balance transmit policy for balancing outgoing traffic.
DPDK implementation of this mode provide some additional requirements of the applica-
tion.
1. It needs to call rte_eth_tx_burst and rte_eth_rx_burst with intervals period
of less than 100ms.
2. Calls to rte_eth_tx_burst must have a buffer size of at least 2xN, where N is
the number of slaves. This is a space required for LACP frames. Additionally LACP
packets are included in the statistics, but they are not returned to the application.

10.2. Implementation Details 56

 Programmer’s Guide, Release 2.0.0

User Application

12003

5006
5005

0002
0001 DPDK

bonded ethdev

ethdev port ethdev port ethdev port

0002 5006
0001 5005 12003

Figure 10.7: Transmit Load Balancing (Mode 5)

This mode provides an adaptive transmit load balancing. It dynamically changes the trans-
mitting slave, according to the computed load. Statistics are collected in 100ms intervals
and scheduled every 10ms.

10.2. Implementation Details 57

 Programmer’s Guide, Release 2.0.0

A bonding device must have a minimum of one slave before the bonding device itself can be
started.
Like all other PMD, all functions exported by a PMD are lock-free functions that are assumed
not to be invoked in parallel on different logical cores to work on the same target object.
It should also be noted that the PMD receive function should not be invoked directly on a slave
devices after they have been to a bonded device since packets read directly from the slave
device will no longer be available to the bonded device to read.

10.2.3 Configuration

Link bonding devices are created using the rte_eth_bond_create API which requires a
unique device name, the bonding mode, and the socket Id to allocate the bonding device’s
resources on. The other configurable parameters for a bonded device are its slave devices, its
primary slave, a user defined MAC address and transmission policy to use if the device is in
balance XOR mode.

Slave Devices

Bonding devices support up to a maximum of RTE_MAX_ETHPORTS slave devices of the same

speed and duplex. Ethernet devices can be added as a slave to a maximum of one bonded
device. Slave devices are reconfigured with the configuration of the bonded device on being
added to a bonded device.
The bonded also guarantees to return the MAC address of the slave device to its original value
of removal of a slave from it.

Primary Slave

The primary slave is used to define the default port to use when a bonded device is in active
backup mode. A different port will only be used if, and only if, the current primary port goes
down. If the user does not specify a primary port it will default to being the first port added to
the bonded device.

MAC Address

The bonded device can be configured with a user specified MAC address, this address will be
inherited by the some/all slave devices depending on the operating mode. If the device is in
active backup mode then only the primary device will have the user specified MAC, all other
slaves will retain their original MAC address. In mode 0, 2, 3, 4 all slaves devices are configure
with the bonded devices MAC address.
If a user defined MAC address is not defined then the bonded device will default to using the
primary slaves MAC address.

Balance XOR Transmit Policies

There are 3 supported transmission policies for bonded device running in Balance XOR mode.
Layer 2, Layer 2+3, Layer 3+4.

10.2. Implementation Details 58

 Programmer’s Guide, Release 2.0.0

• Layer 2: Ethernet MAC address based balancing is the default transmission policy for
Balance XOR bonding mode. It uses a simple XOR calculation on the source MAC
address and destination MAC address of the packet and then calculate the modulus of
this value to calculate the slave device to transmit the packet on.
• Layer 2 + 3: Ethernet MAC address & IP Address based balancing uses a combination of
source/destination MAC addresses and the source/destination IP addresses of the data
packet to decide which slave port the packet will be transmitted on.
• Layer 3 + 4: IP Address & UDP Port based balancing uses a combination of
source/destination IP Address and the source/destination UDP ports of the packet of
the data packet to decide which slave port the packet will be transmitted on.
All these policies support 802.1Q VLAN Ethernet packets, as well as IPv4, IPv6 and UDP
protocols for load balancing.

10.3 Using Link Bonding Devices

The librte_pmd_bond library support two modes of device creation, the libraries export full C
API or using the EAL command line to statically configure link bonding devices at application
startup. Using the EAL option it is possible to use link bonding functionality transparently
without specific knowledge of the libraries API, this can be used, for example, to add bonding
functionality, such as active backup, to an existing application which has no knowledge of the
link bonding C API.

10.3.1 Using the Poll Mode Driver from an Application

Using the librte_pmd_bond libraries API it is possible to dynamically create and manage
link bonding device from within any application. Link bonding device are created using the
rte_eth_bond_create API which requires a unique device name, the link bonding mode
to initial the device in and finally the socket Id which to allocate the devices resources onto.
After successful creation of a bonding device it must be configured using the generic Ethernet
device configure API rte_eth_dev_configure and then the RX and TX queues which will
be used must be setup using rte_eth_tx_queue_setup / rte_eth_rx_queue_setup.
Slave devices can be dynamically added and removed from a link bonding device us-
ing the rte_eth_bond_slave_add / rte_eth_bond_slave_remove APIs but at least
one slave device must be added to the link bonding device before it can be started using
rte_eth_dev_start.
The link status of a bonded device is dictated by that of its slaves, if all slave device link status
are down or if all slaves are removed from the link bonding device then the link status of the
bonding device will go down.
It is also possible to configure / query the configuration of the control param-
eters of a bonded device using the provided APIs rte_eth_bond_mode_set/
get, rte_eth_bond_primary_set/get, rte_eth_bond_mac_set/reset and
rte_eth_bond_xmit_policy_set/get.

10.3. Using Link Bonding Devices 59

 Programmer’s Guide, Release 2.0.0

10.3.2 Using Link Bonding Devices from the EAL Command Line

Link bonding devices can be created at application startup time using the --vdev EAL com-
mand line option. The device name must start with the eth_bond prefix followed by numbers
or letters. The name must be unique for each device. Each device can have multiple options
arranged in a comma separated list. Multiple devices definitions can be arranged by calling the
--vdev option multiple times.
Device names and bonding options must be separated by commas as shown below:
$RTE_TARGET/app/testpmd -c f -n 4 \
--vdev 'eth_bond0,bond_opt0=..,bond opt1=..' \
--vdev 'eth_bond1,bond _opt0=..,bond_opt1=..'

Link Bonding EAL Options

There are multiple ways of definitions that can be assessed and combined as long as the
following two rules are respected:
• A unique device name, in the format of eth_bondX is provided, where X can be any
combination of numbers and/or letters, and the name is no greater than 32 characters
long.
• A least one slave device is provided with for each bonded device definition.
• The operation mode of the bonded device being created is provided.
The different options are:
• mode: Integer value defining the bonding mode of the device. Currently supports modes
0,1,2,3,4,5 (round-robin, active backup, balance, broadcast, link aggregation, transmit
load balancing).
mode=2

• slave: Defines the PMD device which will be added as slave to the bonded de-
vice. This option can be selected multiple time, for each device to be added as a
slave. Physical devices should be specified using their PCI address, in the format do-
main:bus:devid.function
slave=0000:0a:00.0,slave=0000:0a:00.1

• primary: Optional parameter which defines the primary slave port, is used in active
backup mode to select the primary slave for data TX/RX if it is available. The primary
port also is used to select the MAC address to use when it is not defined by the user.
This defaults to the first slave added to the device if it is specified. The primary device
must be a slave of the bonded device.
primary=0000:0a:00.0

• socket_id: Optional parameter used to select which socket on a NUMA device the bonded
devices resources will be allocated on.
socket_id=0

• mac: Optional parameter to select a MAC address for link bonding device, this overrides
the value of the primary slave device.

10.3. Using Link Bonding Devices 60

 Programmer’s Guide, Release 2.0.0

mac=00:1e:67:1d:fd:1d

• xmit_policy: Optional parameter which defines the transmission policy when the bonded
device is in balance mode. If not user specified this defaults to l2 (layer 2) forwarding, the
other transmission policies available are l23 (layer 2+3) and l34 (layer 3+4)
xmit_policy=l23

• lsc_poll_period_ms: Optional parameter which defines the polling interval in milli-

seconds at which devices which don’t support lsc interrupts are checked for a change
in the devices link status
lsc_poll_period_ms=100

• up_delay: Optional parameter which adds a delay in milli-seconds to the propagation of

a devices link status changing to up, by default this parameter is zero.
up_delay=10

• down_delay: Optional parameter which adds a delay in milli-seconds to the propagation

of a devices link status changing to down, by default this parameter is zero.
down_delay=50

Examples of Usage

Create a bonded device in round robin mode with two slaves specified by their PCI address:
$RTE_TARGET/app/testpmd -c '0xf' -n 4 \
--vdev 'eth_bond0,mode=0, slave=0000:00a:00.01,slave=0000:004:00.00' \
-- --port-topology=chained

Create a bonded device in round robin mode with two slaves specified by their PCI address
and an overriding MAC address:
$RTE_TARGET/app/testpmd -c '0xf' -n 4 \
--vdev 'eth_bond0,mode=0, slave=0000:00a:00.01,slave=0000:004:00.00,mac=00:1e:67:1d:fd:1d' \
-- --port-topology=chained

Create a bonded device in active backup mode with two slaves specified, and a primary slave
specified by their PCI addresses:
$RTE_TARGET/app/testpmd -c '0xf' -n 4 \
--vdev 'eth_bond0,mode=1, slave=0000:00a:00.01,slave=0000:004:00.00,primary=0000:00a:00.01' \
-- --port-topology=chained

Create a bonded device in balance mode with two slaves specified by their PCI addresses,
and a transmission policy of layer 3 + 4 forwarding:
$RTE_TARGET/app/testpmd -c '0xf' -n 4 \
--vdev 'eth_bond0,mode=2, slave=0000:00a:00.01,slave=0000:004:00.00,xmit_policy=l34' \
-- --port-topology=chained

10.3. Using Link Bonding Devices 61

 CHAPTER

ELEVEN

TIMER LIBRARY

The Timer library provides a timer service to DPDK execution units to enable execution of
callback functions asynchronously. Features of the library are:
• Timers can be periodic (multi-shot) or single (one-shot).
• Timers can be loaded from one core and executed on another. It has to be specified in
the call to rte_timer_reset().
• Timers provide high precision (depends on the call frequency to rte_timer_manage() that
checks timer expiration for the local core).
• If not required in the application, timers can be disabled at compilation time by not calling
the rte_timer_manage() to increase performance.
The timer library uses the rte_get_timer_cycles() function that uses the High Precision Event
Timer (HPET) or the CPUs Time Stamp Counter (TSC) to provide a reliable time reference.
This library provides an interface to add, delete and restart a timer. The API is based on BSD
callout() with a few differences. Refer to the callout manual.

11.1 Implementation Details

Timers are tracked on a per-lcore basis, with all pending timers for a core being maintained
in order of timer expiry in a skiplist data structure. The skiplist used has ten levels and each
entry in the table appears in each level with probability ¼^level. This means that all entries are
present in level 0, 1 in every 4 entries is present at level 1, one in every 16 at level 2 and so on
up to level 9. This means that adding and removing entries from the timer list for a core can be
done in log(n) time, up to 4^10 entries, that is, approximately 1,000,000 timers per lcore.
A timer structure contains a special field called status, which is a union of a timer state
(stopped, pending, running, config) and an owner (lcore id). Depending on the timer state,
we know if a timer is present in a list or not:
• STOPPED: no owner, not in a list
• CONFIG: owned by a core, must not be modified by another core, maybe in a list or not,
depending on previous state
• PENDING: owned by a core, present in a list
• RUNNING: owned by a core, must not be modified by another core, present in a list

62
 Programmer’s Guide, Release 2.0.0

Resetting or stopping a timer while it is in a CONFIG or RUNNING state is not allowed. When
modifying the state of a timer, a Compare And Swap instruction should be used to guarantee
that the status (state+owner) is modified atomically.
Inside the rte_timer_manage() function, the skiplist is used as a regular list by iterating along
the level 0 list, which contains all timer entries, until an entry which has not yet expired has
been encountered. To improve performance in the case where there are entries in the timer
list but none of those timers have yet expired, the expiry time of the first list entry is maintained
within the per-core timer list structure itself. On 64-bit platforms, this value can be checked
without the need to take a lock on the overall structure. (Since expiry times are maintained
as 64-bit values, a check on the value cannot be done on 32-bit platforms without using either
a compare-and-swap (CAS) instruction or using a lock, so this additional check is skipped in
favor of checking as normal once the lock has been taken.) On both 64-bit and 32-bit platforms,
a call to rte_timer_manage() returns without taking a lock in the case where the timer list for
the calling core is empty.

11.2 Use Cases

The timer library is used for periodic calls, such as garbage collectors, or some state machines
(ARP, bridging, and so on).

11.3 References

• callout manual - The callout facility that provides timers with a mechanism to execute a
function at a given time.
• HPET - Information about the High Precision Event Timer (HPET).

11.2. Use Cases 63

 CHAPTER

TWELVE

HASH LIBRARY

The DPDK provides a Hash Library for creating hash table for fast lookup. The hash table is
a data structure optimized for searching through a set of entries that are each identified by a
unique key. For increased performance the DPDK Hash requires that all the keys have the
same number of bytes which is set at the hash creation time.

12.1 Hash API Overview

The main configuration parameters for the hash are:

• Total number of hash entries
• Size of the key in bytes
The hash also allows the configuration of some low-level implementation related parameters
such as:
• Hash function to translate the key into a bucket index
• Number of entries per bucket
The main methods exported by the hash are:
• Add entry with key: The key is provided as input. If a new entry is successfully added to
the hash for the specified key, or there is already an entry in the hash for the specified
key, then the position of the entry is returned. If the operation was not successful, for
example due to lack of free entries in the hash, then a negative value is returned;
• Delete entry with key: The key is provided as input. If an entry with the specified key is
found in the hash, then the entry is removed from the hash and the position where the
entry was found in the hash is returned. If no entry with the specified key exists in the
hash, then a negative value is returned
• Lookup for entry with key: The key is provided as input. If an entry with the specified
key is found in the hash (lookup hit), then the position of the entry is returned, otherwise
(lookup miss) a negative value is returned.
The current hash implementation handles the key management only. The actual data associ-
ated with each key has to be managed by the user using a separate table that mirrors the hash
in terms of number of entries and position of each entry, as shown in the Flow Classification
use case describes in the following sections.
The example hash tables in the L2/L3 Forwarding sample applications defines which port to
forward a packet to based on a packet flow identified by the five-tuple lookup. However, this

64
 Programmer’s Guide, Release 2.0.0

table could also be used for more sophisticated features and provide many other functions and
actions that could be performed on the packets and flows.

12.2 Implementation Details

The hash table is implemented as an array of entries which is further divided into buckets, with
the same number of consecutive array entries in each bucket. For any input key, there is always
a single bucket where that key can be stored in the hash, therefore only the entries within that
bucket need to be examined when the key is looked up. The lookup speed is achieved by
reducing the number of entries to be scanned from the total number of hash entries down to
the number of entries in a hash bucket, as opposed to the basic method of linearly scanning all
the entries in the array. The hash uses a hash function (configurable) to translate the input key
into a 4-byte key signature. The bucket index is the key signature modulo the number of hash
buckets. Once the bucket is identified, the scope of the hash add, delete and lookup operations
is reduced to the entries in that bucket.
To speed up the search logic within the bucket, each hash entry stores the 4-byte key signa-
ture together with the full key for each hash entry. For large key sizes, comparing the input key
against a key from the bucket can take significantly more time than comparing the 4-byte sig-
nature of the input key against the signature of a key from the bucket. Therefore, the signature
comparison is done first and the full key comparison done only when the signatures matches.
The full key comparison is still necessary, as two input keys from the same bucket can still
potentially have the same 4-byte hash signature, although this event is relatively rare for hash
functions providing good uniform distributions for the set of input keys.

12.3 Use Case: Flow Classification

Flow classification is used to map each input packet to the connection/flow it belongs to. This
operation is necessary as the processing of each input packet is usually done in the context
of their connection, so the same set of operations is applied to all the packets from the same
flow.
Applications using flow classification typically have a flow table to manage, with each separate
flow having an entry associated with it in this table. The size of the flow table entry is application
specific, with typical values of 4, 16, 32 or 64 bytes.
Each application using flow classification typically has a mechanism defined to uniquely iden-
tify a flow based on a number of fields read from the input packet that make up the flow key.
One example is to use the DiffServ 5-tuple made up of the following fields of the IP and trans-
port layer packet headers: Source IP Address, Destination IP Address, Protocol, Source Port,
Destination Port.
The DPDK hash provides a generic method to implement an application specific flow classifi-
cation mechanism. Given a flow table implemented as an array, the application should create
a hash object with the same number of entries as the flow table and with the hash key size set
to the number of bytes in the selected flow key.
The flow table operations on the application side are described below:
• Add flow: Add the flow key to hash. If the returned position is valid, use it to access the
flow entry in the flow table for adding a new flow or updating the information associated

12.2. Implementation Details 65

 Programmer’s Guide, Release 2.0.0

with an existing flow. Otherwise, the flow addition failed, for example due to lack of free
entries for storing new flows.
• Delete flow: Delete the flow key from the hash. If the returned position is valid, use it to
access the flow entry in the flow table to invalidate the information associated with the
flow.
• Lookup flow: Lookup for the flow key in the hash. If the returned position is valid (flow
lookup hit), use the returned position to access the flow entry in the flow table. Otherwise
(flow lookup miss) there is no flow registered for the current packet.

12.4 References

• Donald E. Knuth, The Art of Computer Programming, Volume 3: Sorting and Searching
(2nd Edition), 1998, Addison-Wesley Professional

12.4. References 66
 CHAPTER

THIRTEEN

LPM LIBRARY

The DPDK LPM library component implements the Longest Prefix Match (LPM) table search
method for 32-bit keys that is typically used to find the best route match in IP forwarding appli-
cations.

13.1 LPM API Overview

The main configuration parameter for LPM component instances is the maximum number of
rules to support. An LPM prefix is represented by a pair of parameters (32- bit key, depth), with
depth in the range of 1 to 32. An LPM rule is represented by an LPM prefix and some user
data associated with the prefix. The prefix serves as the unique identifier of the LPM rule. In
this implementation, the user data is 1-byte long and is called next hop, in correlation with its
main use of storing the ID of the next hop in a routing table entry.
The main methods exported by the LPM component are:
• Add LPM rule: The LPM rule is provided as input. If there is no rule with the same prefix
present in the table, then the new rule is added to the LPM table. If a rule with the same
prefix is already present in the table, the next hop of the rule is updated. An error is
returned when there is no available rule space left.
• Delete LPM rule: The prefix of the LPM rule is provided as input. If a rule with the
specified prefix is present in the LPM table, then it is removed.
• Lookup LPM key: The 32-bit key is provided as input. The algorithm selects the rule that
represents the best match for the given key and returns the next hop of that rule. In the
case that there are multiple rules present in the LPM table that have the same 32-bit key,
the algorithm picks the rule with the highest depth as the best match rule, which means
that the rule has the highest number of most significant bits matching between the input
key and the rule key.

13.2 Implementation Details

The current implementation uses a variation of the DIR-24-8 algorithm that trades memory
usage for improved LPM lookup speed. The algorithm allows the lookup operation to be per-
formed with typically a single memory read access. In the statistically rare case when the best
match rule is having a depth bigger than 24, the lookup operation requires two memory read
accesses. Therefore, the performance of the LPM lookup operation is greatly influenced by
whether the specific memory location is present in the processor cache or not.

67
 Programmer’s Guide, Release 2.0.0

The main data structure is built using the following elements:

• A table with 2^24 entries.
• A number of tables (RTE_LPM_TBL8_NUM_GROUPS) with 2^8 entries.
The first table, called tbl24, is indexed using the first 24 bits of the IP address to be looked up,
while the second table(s), called tbl8, is indexed using the last 8 bits of the IP address. This
means that depending on the outcome of trying to match the IP address of an incoming packet
to the rule stored in the tbl24 we might need to continue the lookup process in the second level.
Since every entry of the tbl24 can potentially point to a tbl8, ideally, we would have 2^24 tbl8s,
which would be the same as having a single table with 2^32 entries. This is not feasible due
to resource restrictions. Instead, this approach takes advantage of the fact that rules longer
than 24 bits are very rare. By splitting the process in two different tables/levels and limiting the
number of tbl8s, we can greatly reduce memory consumption while maintaining a very good
lookup speed (one memory access, most of the times).

Figure 13.1: Table split into different levels

An entry in tbl24 contains the following fields:

• next hop / index to the tbl8
• valid flag
• external entry flag
• depth of the rule (length)
The first field can either contain a number indicating the tbl8 in which the lookup process should
continue or the next hop itself if the longest prefix match has already been found. The two flags
are used to determine whether the entry is valid or not and whether the search process have

13.2. Implementation Details 68

 Programmer’s Guide, Release 2.0.0

finished or not respectively. The depth or length of the rule is the number of bits of the rule that
is stored in a specific entry.
An entry in a tbl8 contains the following fields:
• next hop
• valid
• valid group
• depth
Next hop and depth contain the same information as in the tbl24. The two flags show whether
the entry and the table are valid respectively.
The other main data structure is a table containing the main information about the rules (IP
and next hop). This is a higher level table, used for different things:
• Check whether a rule already exists or not, prior to addition or deletion, without having to
actually perform a lookup.
• When deleting, to check whether there is a rule containing the one that is to be deleted.
This is important, since the main data structure will have to be updated accordingly.

13.2.1 Addition

When adding a rule, there are different possibilities. If the rule’s depth is exactly 24 bits, then:
• Use the rule (IP address) as an index to the tbl24.
• If the entry is invalid (i.e. it doesn’t already contain a rule) then set its next hop to its value,
the valid flag to 1 (meaning this entry is in use), and the external entry flag to 0 (meaning
the lookup process ends at this point, since this is the longest prefix that matches).
If the rule’s depth is exactly 32 bits, then:
• Use the first 24 bits of the rule as an index to the tbl24.
• If the entry is invalid (i.e. it doesn’t already contain a rule) then look for a free tbl8, set
the index to the tbl8 to this value, the valid flag to 1 (meaning this entry is in use), and the
external entry flag to 1 (meaning the lookup process must continue since the rule hasn’t
been explored completely).
If the rule’s depth is any other value, prefix expansion must be performed. This means the rule
is copied to all the entries (as long as they are not in use) which would also cause a match.
As a simple example, let’s assume the depth is 20 bits. This means that there are 2^(24 -
20) = 16 different combinations of the first 24 bits of an IP address that would cause a match.
Hence, in this case, we copy the exact same entry to every position indexed by one of these
combinations.
By doing this we ensure that during the lookup process, if a rule matching the IP address exists,
it is found in either one or two memory accesses, depending on whether we need to move to
the next table or not. Prefix expansion is one of the keys of this algorithm, since it improves the
speed dramatically by adding redundancy.

13.2. Implementation Details 69

 Programmer’s Guide, Release 2.0.0

13.2.2 Lookup

The lookup process is much simpler and quicker. In this case:

• Use the first 24 bits of the IP address as an index to the tbl24. If the entry is not in use,
then it means we don’t have a rule matching this IP. If it is valid and the external entry
flag is set to 0, then the next hop is returned.
• If it is valid and the external entry flag is set to 1, then we use the tbl8 index to find out
the tbl8 to be checked, and the last 8 bits of the IP address as an index to this table.
Similarly, if the entry is not in use, then we don’t have a rule matching this IP address. If
it is valid then the next hop is returned.

13.2.3 Limitations in the Number of Rules

There are different things that limit the number of rules that can be added. The first one is the
maximum number of rules, which is a parameter passed through the API. Once this number is
reached, it is not possible to add any more rules to the routing table unless one or more are
removed.
The second reason is an intrinsic limitation of the algorithm. As explained before, to avoid high
memory consumption, the number of tbl8s is limited in compilation time (this value is by default
256). If we exhaust tbl8s, we won’t be able to add any more rules. How many of them are
necessary for a specific routing table is hard to determine in advance.
A tbl8 is consumed whenever we have a new rule with depth bigger than 24, and the first 24
bits of this rule are not the same as the first 24 bits of a rule previously added. If they are, then
the new rule will share the same tbl8 than the previous one, since the only difference between
the two rules is within the last byte.
With the default value of 256, we can have up to 256 rules longer than 24 bits that differ on
their first three bytes. Since routes longer than 24 bits are unlikely, this shouldn’t be a problem
in most setups. Even if it is, however, the number of tbl8s can be modified.

13.2.4 Use Case: IPv4 Forwarding

The LPM algorithm is used to implement Classless Inter-Domain Routing (CIDR) strategy used
by routers implementing IPv4 forwarding.

13.2.5 References

• RFC1519 Classless Inter-Domain Routing (CIDR): an Address Assignment and Aggre-

gation Strategy, http://www.ietf.org/rfc/rfc1519
• Pankaj Gupta, Algorithms for Routing Lookups and Packet Classification, PhD Thesis,
Stanford University, 2000 (http://klamath.stanford.edu/~pankaj/thesis/ thesis_1sided.pdf
)

13.2. Implementation Details 70

 CHAPTER

FOURTEEN

LPM6 LIBRARY

The LPM6 (LPM for IPv6) library component implements the Longest Prefix Match (LPM) ta-
ble search method for 128-bit keys that is typically used to find the best match route in IPv6
forwarding applications.

14.1 LPM6 API Overview

The main configuration parameters for the LPM6 library are:

• Maximum number of rules: This defines the size of the table that holds the rules, and
therefore the maximum number of rules that can be added.
• Number of tbl8s: A tbl8 is a node of the trie that the LPM6 algorithm is based on.
This parameter is related to the number of rules you can have, but there is no way to accurately
predict the number needed to hold a specific number of rules, since it strongly depends on the
depth and IP address of every rule. One tbl8 consumes 1 kb of memory. As a recommendation,
65536 tbl8s should be sufficient to store several thousand IPv6 rules, but the number can vary
depending on the case.
An LPM prefix is represented by a pair of parameters (128-bit key, depth), with depth in the
range of 1 to 128. An LPM rule is represented by an LPM prefix and some user data associated
with the prefix. The prefix serves as the unique identifier for the LPM rule. In this implementa-
tion, the user data is 1-byte long and is called “next hop”, which corresponds to its main use of
storing the ID of the next hop in a routing table entry.
The main methods exported for the LPM component are:
• Add LPM rule: The LPM rule is provided as input. If there is no rule with the same prefix
present in the table, then the new rule is added to the LPM table. If a rule with the same
prefix is already present in the table, the next hop of the rule is updated. An error is
returned when there is no available space left.
• Delete LPM rule: The prefix of the LPM rule is provided as input. If a rule with the
specified prefix is present in the LPM table, then it is removed.
• Lookup LPM key: The 128-bit key is provided as input. The algorithm selects the rule
that represents the best match for the given key and returns the next hop of that rule. In
the case that there are multiple rules present in the LPM table that have the same 128-bit
value, the algorithm picks the rule with the highest depth as the best match rule, which
means the rule has the highest number of most significant bits matching between the
input key and the rule key.

71
 Programmer’s Guide, Release 2.0.0

14.1.1 Implementation Details

This is a modification of the algorithm used for IPv4 (see IPv4 LPM Implementation Details). In
this case, instead of using two levels, one with a tbl24 and a second with a tbl8, 14 levels are
used.
The implementation can be seen as a multi-bit trie where the stride or number of bits inspected
on each level varies from level to level. Specifically, 24 bits are inspected on the root node, and
the remaining 104 bits are inspected in groups of 8 bits. This effectively means that the trie
has 14 levels at the most, depending on the rules that are added to the table.
The algorithm allows the lookup operation to be performed with a number of memory accesses
that directly depends on the length of the rule and whether there are other rules with bigger
depths and the same key in the data structure. It can vary from 1 to 14 memory accesses, with
5 being the average value for the lengths that are most commonly used in IPv6.
The main data structure is built using the following elements:
• A table with 224 entries
• A number of tables, configurable by the user through the API, with 28 entries
The first table, called tbl24, is indexed using the first 24 bits of the IP address be looked up,
while the rest of the tables, called tbl8s, are indexed using the rest of the bytes of the IP
address, in chunks of 8 bits. This means that depending on the outcome of trying to match
the IP address of an incoming packet to the rule stored in the tbl24 or the subsequent tbl8s we
might need to continue the lookup process in deeper levels of the tree.
Similar to the limitation presented in the algorithm for IPv4, to store every possible IPv6 rule,
we would need a table with 2^128 entries. This is not feasible due to resource restrictions.
By splitting the process in different tables/levels and limiting the number of tbl8s, we can greatly
reduce memory consumption while maintaining a very good lookup speed (one memory ac-
cess per level).
An entry in a table contains the following fields:
• next hop / index to the tbl8
• depth of the rule (length)
• valid flag
• valid group flag
• external entry flag
The first field can either contain a number indicating the tbl8 in which the lookup process should
continue or the next hop itself if the longest prefix match has already been found. The depth
or length of the rule is the number of bits of the rule that is stored in a specific entry. The flags
are used to determine whether the entry/table is valid or not and whether the search process
have finished or not respectively.
Both types of tables share the same structure.
The other main data structure is a table containing the main information about the rules (IP,
next hop and depth). This is a higher level table, used for different things:
• Check whether a rule already exists or not, prior to addition or deletion, without having to
actually perform a lookup.

14.1. LPM6 API Overview 72

 Programmer’s Guide, Release 2.0.0

Figure 14.1: Table split into different levels

When deleting, to check whether there is a rule containing the one that is to be deleted. This
is important, since the main data structure will have to be updated accordingly.

14.1.2 Addition

When adding a rule, there are different possibilities. If the rule’s depth is exactly 24 bits, then:
• Use the rule (IP address) as an index to the tbl24.
• If the entry is invalid (i.e. it doesn’t already contain a rule) then set its next hop to its value,
the valid flag to 1 (meaning this entry is in use), and the external entry flag to 0 (meaning
the lookup process ends at this point, since this is the longest prefix that matches).
If the rule’s depth is bigger than 24 bits but a multiple of 8, then:
• Use the first 24 bits of the rule as an index to the tbl24.
• If the entry is invalid (i.e. it doesn’t already contain a rule) then look for a free tbl8, set
the index to the tbl8 to this value, the valid flag to 1 (meaning this entry is in use), and the
external entry flag to 1 (meaning the lookup process must continue since the rule hasn’t
been explored completely).
• Use the following 8 bits of the rule as an index to the next tbl8.
• Repeat the process until the tbl8 at the right level (depending on the depth) has been
reached and fill it with the next hop, setting the next entry flag to 0.
If the rule’s depth is any other value, prefix expansion must be performed. This means the rule
is copied to all the entries (as long as they are not in use) which would also cause a match.

14.1. LPM6 API Overview 73

 Programmer’s Guide, Release 2.0.0

As a simple example, let’s assume the depth is 20 bits. This means that there are 2^(24-20)
= 16 different combinations of the first 24 bits of an IP address that would cause a match.
Hence, in this case, we copy the exact same entry to every position indexed by one of these
combinations.
By doing this we ensure that during the lookup process, if a rule matching the IP address exists,
it is found in, at the most, 14 memory accesses, depending on how many times we need to
move to the next table. Prefix expansion is one of the keys of this algorithm, since it improves
the speed dramatically by adding redundancy.
Prefix expansion can be performed at any level. So, for example, is the depth is 34 bits, it will
be performed in the third level (second tbl8-based level).

14.1.3 Lookup

The lookup process is much simpler and quicker. In this case:

• Use the first 24 bits of the IP address as an index to the tbl24. If the entry is not in use,
then it means we don’t have a rule matching this IP. If it is valid and the external entry
flag is set to 0, then the next hop is returned.
• If it is valid and the external entry flag is set to 1, then we use the tbl8 index to find out
the tbl8 to be checked, and the next 8 bits of the IP address as an index to this table.
Similarly, if the entry is not in use, then we don’t have a rule matching this IP address. If
it is valid then check the external entry flag for a new tbl8 to be inspected.
• Repeat the process until either we find an invalid entry (lookup miss) or a valid entry with
the external entry flag set to 0. Return the next hop in the latter case.

14.1.4 Limitations in the Number of Rules

There are different things that limit the number of rules that can be added. The first one is the
maximum number of rules, which is a parameter passed through the API. Once this number is
reached, it is not possible to add any more rules to the routing table unless one or more are
removed.
The second limitation is in the number of tbl8s available. If we exhaust tbl8s, we won’t be able
to add any more rules. How to know how many of them are necessary for a specific routing
table is hard to determine in advance.
In this algorithm, the maximum number of tbl8s a single rule can consume is 13, which is the
number of levels minus one, since the first three bytes are resolved in the tbl24. However:
• Typically, on IPv6, routes are not longer than 48 bits, which means rules usually take up
to 3 tbl8s.
As explained in the LPM for IPv4 algorithm, it is possible and very likely that several rules will
share one or more tbl8s, depending on what their first bytes are. If they share the same first 24
bits, for instance, the tbl8 at the second level will be shared. This might happen again in deeper
levels, so, effectively, two 48 bit-long rules may use the same three tbl8s if the only difference
is in their last byte.
The number of tbl8s is a parameter exposed to the user through the API in this version of the
algorithm, due to its impact in memory consumption and the number or rules that can be added
to the LPM table. One tbl8 consumes 1 kilobyte of memory.

14.1. LPM6 API Overview 74

 Programmer’s Guide, Release 2.0.0

14.2 Use Case: IPv6 Forwarding

The LPM algorithm is used to implement the Classless Inter-Domain Routing (CIDR) strategy
used by routers implementing IP forwarding.

14.2. Use Case: IPv6 Forwarding 75

 CHAPTER

FIFTEEN

PACKET DISTRIBUTOR LIBRARY

The DPDK Packet Distributor library is a library designed to be used for dynamic load balancing
of traffic while supporting single packet at a time operation. When using this library, the logical
cores in use are to be considered in two roles: firstly a distributor lcore, which is responsible
for load balancing or distributing packets, and a set of worker lcores which are responsible for
receiving the packets from the distributor and operating on them. The model of operation is
shown in the diagram below.

Figure 15.1: Packet Distributor mode of operation

76
 Programmer’s Guide, Release 2.0.0

15.1 Distributor Core Operation

The distributor core does the majority of the processing for ensuring that packets are fairly
shared among workers. The operation of the distributor is as follows:
1. Packets are passed to the distributor component by having the distributor lcore thread
call the “rte_distributor_process()” API
2. The worker lcores all share a single cache line with the distributor core in order to pass
messages and packets to and from the worker. The process API call will poll all the
worker cache lines to see what workers are requesting packets.
3. As workers request packets, the distributor takes packets from the set of packets passed
in and distributes them to the workers. As it does so, it examines the “tag” – stored in the
RSS hash field in the mbuf – for each packet and records what tags are being processed
by each worker.
4. If the next packet in the input set has a tag which is already being processed by a worker,
then that packet will be queued up for processing by that worker and given to it in prefer-
ence to other packets when that work next makes a request for work. This ensures that
no two packets with the same tag are processed in parallel, and that all packets with the
same tag are processed in input order.
5. Once all input packets passed to the process API have either been distributed to workers
or been queued up for a worker which is processing a given tag, then the process API
returns to the caller.
Other functions which are available to the distributor lcore are:
• rte_distributor_returned_pkts()
• rte_distributor_flush()
• rte_distributor_clear_returns()
Of these the most important API call is “rte_distributor_returned_pkts()” which should only be
called on the lcore which also calls the process API. It returns to the caller all packets which
have finished processing by all worker cores. Within this set of returned packets, all packets
sharing the same tag will be returned in their original order.
NOTE: If worker lcores buffer up packets internally for transmission in bulk afterwards, the
packets sharing a tag will likely get out of order. Once a worker lcore requests a new packet,
the distributor assumes that it has completely finished with the previous packet and therefore
that additional packets with the same tag can safely be distributed to other workers – who may
then flush their buffered packets sooner and cause packets to get out of order.
NOTE: No packet ordering guarantees are made about packets which do not share a common
packet tag.
Using the process and returned_pkts API, the following application workflow can be used, while
allowing packet order within a packet flow – identified by a tag – to be maintained.
The flush and clear_returns API calls, mentioned previously, are likely of less use that the
process and returned_pkts APIS, and are principally provided to aid in unit testing of the li-
brary. Descriptions of these functions and their use can be found in the DPDK API Reference
document.

15.1. Distributor Core Operation 77

 Programmer’s Guide, Release 2.0.0

Figure 15.2: Application workflow

15.2 Worker Operation

Worker cores are the cores which do the actual manipulation of the packets distributed by the
packet distributor. Each worker calls “rte_distributor_get_pkt()” API to request a new packet
when it has finished processing the previous one. [The previous packet should be returned to
the distributor component by passing it as the final parameter to this API call.]
Since it may be desirable to vary the number of worker cores, depending on the traffic load i.e.
to save power at times of lighter load, it is possible to have a worker stop processing packets
by calling “rte_distributor_return_pkt()” to indicate that it has finished the current packet and
does not want a new one.

15.2. Worker Operation 78

 CHAPTER

SIXTEEN

REORDER LIBRARY

The Reorder Library provides a mechanism for reordering mbufs based on their sequence
number.

16.1 Operation

The reorder library is essentially a buffer that reorders mbufs. The user inserts out of order
mbufs into the reorder buffer and pulls in-order mbufs from it.
At a given time, the reorder buffer contains mbufs whose sequence number are inside the
sequence window. The sequence window is determined by the minimum sequence number
and the number of entries that the buffer was configured to hold. For example, given a reorder
buffer with 200 entries and a minimum sequence number of 350, the sequence window has
low and high limits of 350 and 550 respectively.
When inserting mbufs, the reorder library differentiates between valid, early and late mbufs
depending on the sequence number of the inserted mbuf:
• valid: the sequence number is inside the window.
• late: the sequence number is outside the window and less than the low limit.
• early: the sequence number is outside the window and greater than the high limit.
The reorder buffer directly returns late mbufs and tries to accommodate early mbufs.

16.2 Implementation Details

The reorder library is implemented as a pair of buffers, which referred to as the Order buffer
and the Ready buffer.
On an insert call, valid mbufs are inserted directly into the Order buffer and late mbufs are
returned to the user with an error.
In the case of early mbufs, the reorder buffer will try to move the window (incrementing the
minimum sequence number) so that the mbuf becomes a valid one. To that end, mbufs in the
Order buffer are moved into the Ready buffer. Any mbufs that have not arrived yet are ignored
and therefore will become late mbufs. This means that as long as there is room in the Ready
buffer, the window will be moved to accommodate early mbufs that would otherwise be outside
the reordering window.

79
 Programmer’s Guide, Release 2.0.0

For example, assuming that we have a buffer of 200 entries with a 350 minimum sequence
number, and we need to insert an early mbuf with 565 sequence number. That means that we
would need to move the windows at least 15 positions to accommodate the mbuf. The reorder
buffer would try to move mbufs from at least the next 15 slots in the Order buffer to the Ready
buffer, as long as there is room in the Ready buffer. Any gaps in the Order buffer at that point
are skipped, and those packet will be reported as late packets when they arrive. The process
of moving packets to the Ready buffer continues beyond the minimum required until a gap, i.e.
missing mbuf, in the Order buffer is encountered.
When draining mbufs, the reorder buffer would return mbufs in the Ready buffer first and then
from the Order buffer until a gap is found (mbufs that have not arrived yet).

16.3 Use Case: Packet Distributor

An application using the DPDK packet distributor could make use of the reorder library to
transmit packets in the same order they were received.
A basic packet distributor use case would consist of a distributor with multiple workers cores.
The processing of packets by the workers is not guaranteed to be in order, hence a reorder
buffer can be used to order as many packets as possible.
In such a scenario, the distributor assigns a sequence number to mbufs before delivering them
to the workers. As the workers finish processing the packets, the distributor inserts those mbufs
into the reorder buffer and finally transmit drained mbufs.
NOTE: Currently the reorder buffer is not thread safe so the same thread is responsible for
inserting and draining mbufs.

16.3. Use Case: Packet Distributor 80

 CHAPTER

SEVENTEEN

IP FRAGMENTATION AND REASSEMBLY LIBRARY

The IP Fragmentation and Reassembly Library implements IPv4 and IPv6 packet fragmenta-
tion and reassembly.

17.1 Packet fragmentation

Packet fragmentation routines divide input packet into number of fragments. Both
rte_ipv4_fragment_packet() and rte_ipv6_fragment_packet() functions assume that input mbuf
data points to the start of the IP header of the packet (i.e. L2 header is already stripped out).
To avoid copying of the actual packet’s data zero-copy technique is used (rte_pktmbuf_attach).
For each fragment two new mbufs are created:
• Direct mbuf – mbuf that will contain L3 header of the new fragment.
• Indirect mbuf – mbuf that is attached to the mbuf with the original packet. It’s data field
points to the start of the original packets data plus fragment offset.
Then L3 header is copied from the original mbuf into the ‘direct’ mbuf and updated to reflect
new fragmented status. Note that for IPv4, header checksum is not recalculated and is set to
zero.
Finally ‘direct’ and ‘indirect’ mbufs for each fragment are linked together via mbuf’s next filed to
compose a packet for the new fragment.
The caller has an ability to explicitly specify which mempools should be used to allocate ‘direct’
and ‘indirect’ mbufs from.
For more information about direct and indirect mbufs, refer to the DPDK Programmers guide
7.7 Direct and Indirect Buffers.

17.2 Packet reassembly

17.2.1 IP Fragment Table

Fragment table maintains information about already received fragments of the packet.
Each IP packet is uniquely identified by triple <Source IP address>, <Destination IP address>,
<ID>.
Note that all update/lookup operations on Fragment Table are not thread safe. So if different
execution contexts (threads/processes) will access the same table simultaneously, then some
external syncing mechanism have to be provided.

81
 Programmer’s Guide, Release 2.0.0

Each table entry can hold information about packets consisting of up to

RTE_LIBRTE_IP_FRAG_MAX (by default: 4) fragments.
Code example, that demonstrates creation of a new Fragment table:
frag_cycles = (rte_get_tsc_hz() + MS_PER_S - 1) / MS_PER_S * max_flow_ttl;
bucket_num = max_flow_num + max_flow_num / 4;
frag_tbl = rte_ip_frag_table_create(max_flow_num, bucket_entries,
max_flow_num, frag_cycles, socket_id);

Internally Fragment table is a simple hash table. The basic idea is to use two hash functions
and <bucket_entries> * associativity. This provides 2 * <bucket_entries> possible locations
in the hash table for each key. When the collision occurs and all 2 * <bucket_entries> are
occupied, instead of reinserting existing keys into alternative locations, ip_frag_tbl_add() just
returns a failure.
Also, entries that resides in the table longer then <max_cycles> are considered as invalid, and
could be removed/replaced by the new ones.
Note that reassembly demands a lot of mbuf’s to be allocated. At any given time up to (2 *
bucket_entries * RTE_LIBRTE_IP_FRAG_MAX * <maximum number of mbufs per packet>)
can be stored inside Fragment Table waiting for remaining fragments.

17.2.2 Packet Reassembly

Fragmented packets processing and reassembly is done by the

rte_ipv4_frag_reassemble_packet()/rte_ipv6_frag_reassemble_packet. Functions. They
either return a pointer to valid mbuf that contains reassembled packet, or NULL (if the packet
can’t be reassembled for some reason).
These functions are responsible for:
1. Search the Fragment Table for entry with packet’s <IPv4 Source Address, IPv4 Destina-
tion Address, Packet ID>.
2. If the entry is found, then check if that entry already timed-out. If yes, then free all
previously received fragments, and remove information about them from the entry.
3. If no entry with such key is found, then try to create a new one by one of two ways:
(a) Use as empty entry.
(b) Delete a timed-out entry, free mbufs associated with it mbufs and store a new entry
with specified key in it.
4. Update the entry with new fragment information and check if a packet can be reassem-
bled (the packet’s entry contains all fragments).
(a) If yes, then, reassemble the packet, mark table’s entry as empty and return the
reassembled mbuf to the caller.
(b) If no, then return a NULL to the caller.
If at any stage of packet processing an error is encountered (e.g: can’t insert new entry into the
Fragment Table, or invalid/timed-out fragment), then the function will free all associated with
the packet fragments, mark the table entry as invalid and return NULL to the caller.

17.2. Packet reassembly 82

 Programmer’s Guide, Release 2.0.0

17.2.3 Debug logging and Statistics Collection

The RTE_LIBRTE_IP_FRAG_TBL_STAT config macro controls statistics collection for the

Fragment Table. This macro is not enabled by default.
The RTE_LIBRTE_IP_FRAG_DEBUG controls debug logging of IP fragments processing and
reassembling. This macro is disabled by default. Note that while logging contains a lot of
detailed information, it slows down packet processing and might cause the loss of a lot of
packets.

17.2. Packet reassembly 83

 CHAPTER

EIGHTEEN

MULTI-PROCESS SUPPORT

In the DPDK, multi-process support is designed to allow a group of DPDK processes to work
together in a simple transparent manner to perform packet processing, or other workloads, on
Intel® architecture hardware. To support this functionality, a number of additions have been
made to the core DPDK Environment Abstraction Layer (EAL).
The EAL has been modified to allow different types of DPDK processes to be spawned, each
with different permissions on the hugepage memory used by the applications. For now, there
are two types of process specified:
• primary processes, which can initialize and which have full permissions on shared mem-
ory
• secondary processes, which cannot initialize shared memory, but can attach to pre- ini-
tialized shared memory and create objects in it.
Standalone DPDK processes are primary processes, while secondary processes can only run
alongside a primary process or after a primary process has already configured the hugepage
shared memory for them.
To support these two process types, and other multi-process setups described later, two addi-
tional command-line parameters are available to the EAL:
• –proc-type: for specifying a given process instance as the primary or secondary DPDK
instance
• –file-prefix: to allow processes that do not want to co-operate to have different memory
regions
A number of example applications are provided that demonstrate how multiple DPDK pro-
cesses can be used together. These are more fully documented in the “Multi- process Sample
Application” chapter in the DPDK Sample Application’s User Guide.

18.1 Memory Sharing

The key element in getting a multi-process application working using the DPDK is to ensure that
memory resources are properly shared among the processes making up the multi-process ap-
plication. Once there are blocks of shared memory available that can be accessed by multiple
processes, then issues such as inter-process communication (IPC) becomes much simpler.
On application start-up in a primary or standalone process, the DPDK records to memory-
mapped files the details of the memory configuration it is using - hugepages in use, the virtual

84
 Programmer’s Guide, Release 2.0.0

addresses they are mapped at, the number of memory channels present, etc. When a sec-
ondary process is started, these files are read and the EAL recreates the same memory con-
figuration in the secondary process so that all memory zones are shared between processes
and all pointers to that memory are valid, and point to the same objects, in both processes.

Note: Refer to Section Multi-process Limitations for details of how Linux kernel Address-
Space Layout Randomization (ASLR) can affect memory sharing.

Figure 18.1: Memory Sharing in the DPDK Multi-process Sample Application

The EAL also supports an auto-detection mode (set by EAL –proc-type=auto flag ), whereby
an DPDK process is started as a secondary instance if a primary instance is already running.

18.2 Deployment Models

18.2.1 Symmetric/Peer Processes

DPDK multi-process support can be used to create a set of peer processes where each pro-
cess performs the same workload. This model is equivalent to having multiple threads each
running the same main-loop function, as is done in most of the supplied DPDK sample ap-
plications. In this model, the first of the processes spawned should be spawned using the
–proc-type=primary EAL flag, while all subsequent instances should be spawned using the
–proc-type=secondary flag.
The simple_mp and symmetric_mp sample applications demonstrate this usage model. They

18.2. Deployment Models 85

 Programmer’s Guide, Release 2.0.0

are described in the “Multi-process Sample Application” chapter in the DPDK Sample Applica-
tion’s User Guide.

18.2.2 Asymmetric/Non-Peer Processes

An alternative deployment model that can be used for multi-process applications is to have
a single primary process instance that acts as a load-balancer or server distributing received
packets among worker or client threads, which are run as secondary processes. In this case,
extensive use of rte_ring objects is made, which are located in shared hugepage memory.
The client_server_mp sample application shows this usage model. It is described in the “Multi-
process Sample Application” chapter in the DPDK Sample Application’s User Guide.

18.2.3 Running Multiple Independent DPDK Applications

In addition to the above scenarios involving multiple DPDK processes working together, it is
possible to run multiple DPDK processes side-by-side, where those processes are all working
independently. Support for this usage scenario is provided using the –file-prefix parameter to
the EAL.
By default, the EAL creates hugepage files on each hugetlbfs filesystem using the rtemap_X
filename, where X is in the range 0 to the maximum number of hugepages -1. Similarly, it cre-
ates shared configuration files, memory mapped in each process, using the /var/run/.rte_config
filename, when run as root (or $HOME/.rte_config when run as a non-root user; if filesystem
and device permissions are set up to allow this). The rte part of the filenames of each of the
above is configurable using the file-prefix parameter.
In addition to specifying the file-prefix parameter, any DPDK applications that are to be run
side-by-side must explicitly limit their memory use. This is done by passing the -m flag to each
process to specify how much hugepage memory, in megabytes, each process can use (or
passing –socket-mem to specify how much hugepage memory on each socket each process
can use).

Note: Independent DPDK instances running side-by-side on a single machine cannot share
any network ports. Any network ports being used by one process should be blacklisted in every
other process.

18.2.4 Running Multiple Independent Groups of DPDK Applications

In the same way that it is possible to run independent DPDK applications side- by-side on
a single system, this can be trivially extended to multi-process groups of DPDK applications
running side-by-side. In this case, the secondary processes must use the same –file-prefix
parameter as the primary process whose shared memory they are connecting to.

Note: All restrictions and issues with multiple independent DPDK processes running side-by-
side apply in this usage scenario also.

18.2. Deployment Models 86

 Programmer’s Guide, Release 2.0.0

18.3 Multi-process Limitations

There are a number of limitations to what can be done when running DPDK multi-process
applications. Some of these are documented below:
• The multi-process feature requires that the exact same hugepage memory mappings be
present in all applications. The Linux security feature - Address-Space Layout Random-
ization (ASLR) can interfere with this mapping, so it may be necessary to disable this
feature in order to reliably run multi-process applications.

Warning: Disabling Address-Space Layout Randomization (ASLR) may have security im-
plications, so it is recommended that it be disabled only when absolutely necessary, and
only when the implications of this change have been understood.

• All DPDK processes running as a single application and using shared memory must
have distinct coremask arguments. It is not possible to have a primary and secondary
instance, or two secondary instances, using any of the same logical cores. Attempting to
do so can cause corruption of memory pool caches, among other issues.
• The delivery of interrupts, such as Ethernet* device link status interrupts, do not work
in secondary processes. All interrupts are triggered inside the primary process only.
Any application needing interrupt notification in multiple processes should provide its
own mechanism to transfer the interrupt information from the primary process to any
secondary process that needs the information.
• The use of function pointers between multiple processes running based of different com-
piled binaries is not supported, since the location of a given function in one process may
be different to its location in a second. This prevents the librte_hash library from behav-
ing properly as in a multi-threaded instance, since it uses a pointer to the hash function
internally.
To work around this issue, it is recommended that multi-process applications perform the
hash calculations by directly calling the hashing function from the code and then using the
rte_hash_add_with_hash()/rte_hash_lookup_with_hash() functions instead of the functions
which do the hashing internally, such as rte_hash_add()/rte_hash_lookup().
• Depending upon the hardware in use, and the number of DPDK processes used, it may
not be possible to have HPET timers available in each DPDK instance. The minimum
number of HPET comparators available to Linux* userspace can be just a single com-
parator, which means that only the first, primary DPDK process instance can open and
mmap /dev/hpet. If the number of required DPDK processes exceeds that of the number
of available HPET comparators, the TSC (which is the default timer in this release) must
be used as a time source across all processes instead of the HPET.

18.3. Multi-process Limitations 87

 CHAPTER

NINETEEN

KERNEL NIC INTERFACE

The DPDK Kernel NIC Interface (KNI) allows userspace applications access to the Linux*
control plane.
The benefits of using the DPDK KNI are:
• Faster than existing Linux TUN/TAP interfaces (by eliminating system calls and
copy_to_user()/copy_from_user() operations.
• Allows management of DPDK ports using standard Linux net tools such as ethtool, ifcon-
fig and tcpdump.
• Allows an interface with the kernel network stack.
The components of an application using the DPDK Kernel NIC Interface are shown in Fig.
19.1.

19.1 The DPDK KNI Kernel Module

The KNI kernel loadable module provides support for two types of devices:
• A Miscellaneous device (/dev/kni) that:
– Creates net devices (via ioctl calls).
– Maintains a kernel thread context shared by all KNI instances (simulating the RX
side of the net driver).
– For single kernel thread mode, maintains a kernel thread context shared by all KNI
instances (simulating the RX side of the net driver).
– For multiple kernel thread mode, maintains a kernel thread context for each KNI
instance (simulating the RX side of the new driver).
• Net device:
– Net functionality provided by implementing several operations such as netdev_ops,
header_ops, ethtool_ops that are defined by struct net_device, including support for
DPDK mbufs and FIFOs.
– The interface name is provided from userspace.
– The MAC address can be the real NIC MAC address or random.

88
 Programmer’s Guide, Release 2.0.0

Figure 19.1: Components of a DPDK KNI Application

19.1. The DPDK KNI Kernel Module 89

 Programmer’s Guide, Release 2.0.0

19.2 KNI Creation and Deletion

The KNI interfaces are created by a DPDK application dynamically. The interface name and
FIFO details are provided by the application through an ioctl call using the rte_kni_device_info
struct which contains:
• The interface name.
• Physical addresses of the corresponding memzones for the relevant FIFOs.
• Mbuf mempool details, both physical and virtual (to calculate the offset for mbuf pointers).
• PCI information.
• Core affinity.
Refer to rte_kni_common.h in the DPDK source code for more details.
The physical addresses will be re-mapped into the kernel address space and stored in separate
KNI contexts.
Once KNI interfaces are created, the KNI context information can be queried by calling the
rte_kni_info_get() function.
The KNI interfaces can be deleted by a DPDK application dynamically after being created.
Furthermore, all those KNI interfaces not deleted will be deleted on the release operation of
the miscellaneous device (when the DPDK application is closed).

19.3 DPDK mbuf Flow

To minimize the amount of DPDK code running in kernel space, the mbuf mempool is managed
in userspace only. The kernel module will be aware of mbufs, but all mbuf allocation and free
operations will be handled by the DPDK application only.
Fig. 19.2 shows a typical scenario with packets sent in both directions.

19.4 Use Case: Ingress

On the DPDK RX side, the mbuf is allocated by the PMD in the RX thread context. This thread
will enqueue the mbuf in the rx_q FIFO. The KNI thread will poll all KNI active devices for the
rx_q. If an mbuf is dequeued, it will be converted to a sk_buff and sent to the net stack via
netif_rx(). The dequeued mbuf must be freed, so the same pointer is sent back in the free_q
FIFO.
The RX thread, in the same main loop, polls this FIFO and frees the mbuf after dequeuing it.

19.5 Use Case: Egress

For packet egress the DPDK application must first enqueue several mbufs to create an mbuf
cache on the kernel side.

19.2. KNI Creation and Deletion 90

 Programmer’s Guide, Release 2.0.0

Figure 19.2: Packet Flow via mbufs in the DPDK KNI

The packet is received from the Linux net stack, by calling the kni_net_tx() callback. The mbuf
is dequeued (without waiting due the cache) and filled with data from sk_buff. The sk_buff is
then freed and the mbuf sent in the tx_q FIFO.
The DPDK TX thread dequeues the mbuf and sends it to the PMD (via rte_eth_tx_burst()). It
then puts the mbuf back in the cache.

19.6 Ethtool

Ethtool is a Linux-specific tool with corresponding support in the kernel where each net device
must register its own callbacks for the supported operations. The current implementation uses
the igb/ixgbe modified Linux drivers for ethtool support. Ethtool is not supported in i40e and
VMs (VF or EM devices).

19.7 Link state and MTU change

Link state and MTU change are network interface specific operations usually done via ifconfig.
The request is initiated from the kernel side (in the context of the ifconfig process) and handled
by the user space DPDK application. The application polls the request, calls the application
handler and returns the response back into the kernel space.
The application handlers can be registered upon interface creation or explicitly regis-
tered/unregistered in runtime. This provides flexibility in multiprocess scenarios (where the
KNI is created in the primary process but the callbacks are handled in the secondary one).
The constraint is that a single process can register and handle the requests.

19.6. Ethtool 91
 Programmer’s Guide, Release 2.0.0

19.8 KNI Working as a Kernel vHost Backend

vHost is a kernel module usually working as the backend of virtio (a para- virtualization driver
framework) to accelerate the traffic from the guest to the host. The DPDK Kernel NIC interface
provides the ability to hookup vHost traffic into userspace DPDK application. Together with
the DPDK PMD virtio, it significantly improves the throughput between guest and host. In the
scenario where DPDK is running as fast path in the host, kni-vhost is an efficient path for the
traffic.

19.8.1 Overview

vHost-net has three kinds of real backend implementations. They are: 1) tap, 2) macvtap and
3) RAW socket. The main idea behind kni-vhost is making the KNI work as a RAW socket,
attaching it as the backend instance of vHost-net. It is using the existing interface with vHost-
net, so it does not require any kernel hacking, and is fully-compatible with the kernel vhost
module. As vHost is still taking responsibility for communicating with the front-end virtio, it
naturally supports both legacy virtio -net and the DPDK PMD virtio. There is a little penalty that
comes from the non-polling mode of vhost. However, it scales throughput well when using KNI
in multi-thread mode.

Figure 19.3: vHost-net Architecture Overview

19.8.2 Packet Flow

There is only a minor difference from the original KNI traffic flows. On transmit side, vhost
kthread calls the RAW socket’s ops sendmsg and it puts the packets into the KNI transmit
FIFO. On the receive side, the kni kthread gets packets from the KNI receive FIFO, puts them
into the queue of the raw socket, and wakes up the task in vhost kthread to begin receiving. All
the packet copying, irrespective of whether it is on the transmit or receive side, happens in the

19.8. KNI Working as a Kernel vHost Backend 92

 Programmer’s Guide, Release 2.0.0

context of vhost kthread. Every vhost-net device is exposed to a front end virtio device in the
guest.

Figure 19.4: KNI Traffic Flow

19.8.3 Sample Usage

Before starting to use KNI as the backend of vhost, the CONFIG_RTE_KNI_VHOST configu-
ration option must be turned on. Otherwise, by default, KNI will not enable its backend support
capability.
Of course, as a prerequisite, the vhost/vhost-net kernel CONFIG should be chosen before
compiling the kernel.
1. Compile the DPDK and insert uio_pci_generic/igb_uio kernel modules as normal.
2. Insert the KNI kernel module:
insmod ./rte_kni.ko

If using KNI in multi-thread mode, use the following command line:

insmod ./rte_kni.ko kthread_mode=multiple

3. Running the KNI sample application:

./kni -c -0xf0 -n 4 -- -p 0x3 -P -config="(0,4,6),(1,5,7)"

This command runs the kni sample application with two physical ports. Each port pins
two forwarding cores (ingress/egress) in user space.

19.8. KNI Working as a Kernel vHost Backend 93

 Programmer’s Guide, Release 2.0.0

4. Assign a raw socket to vhost-net during qemu-kvm startup. The DPDK does not provide
a script to do this since it is easy for the user to customize. The following shows the key
steps to launch qemu-kvm with kni-vhost:
#!/bin/bash
echo 1 > /sys/class/net/vEth0/sock_en
fd=cat /sys/class/net/vEth0/sock_fd
qemu-kvm \
-name vm1 -cpu host -m 2048 -smp 1 -hda /opt/vm-fc16.img \
-netdev tap,fd=$fd,id=hostnet1,vhost=on \
-device virti-net-pci,netdev=hostnet1,id=net1,bus=pci.0,addr=0x4

It is simple to enable raw socket using sysfs sock_en and get raw socket fd using sock_fd under
the KNI device node.
Then, using the qemu-kvm command with the -netdev option to assign such raw socket fd as
vhost’s backend.

Note: The key word tap must exist as qemu-kvm now only supports vhost with a tap backend,
so here we cheat qemu-kvm by an existing fd.

19.8.4 Compatibility Configure Option

There is a CONFIG_RTE_KNI_VHOST_VNET_HDR_EN configuration option in DPDK config-

uration file. By default, it set to n, which means do not turn on the virtio net header, which is
used to support additional features (such as, csum offload, vlan offload, generic-segmentation
and so on), since the kni-vhost does not yet support those features.
Even if the option is turned on, kni-vhost will ignore the information that the header contains.
When working with legacy virtio on the guest, it is better to turn off unsupported offload features
using ethtool -K. Otherwise, there may be problems such as an incorrect L4 checksum error.

19.8. KNI Working as a Kernel vHost Backend 94

 CHAPTER

TWENTY

THREAD SAFETY OF DPDK FUNCTIONS

The DPDK is comprised of several libraries. Some of the functions in these libraries can be
safely called from multiple threads simultaneously, while others cannot. This section allows the
developer to take these issues into account when building their own application.
The run-time environment of the DPDK is typically a single thread per logical core. In some
cases, it is not only multi-threaded, but multi-process. Typically, it is best to avoid sharing data
structures between threads and/or processes where possible. Where this is not possible, then
the execution blocks must access the data in a thread- safe manner. Mechanisms such as
atomics or locking can be used that will allow execution blocks to operate serially. However,
this can have an effect on the performance of the application.

20.1 Fast-Path APIs

Applications operating in the data plane are performance sensitive but certain functions within
those libraries may not be safe to call from multiple threads simultaneously. The hash, LPM
and mempool libraries and RX/TX in the PMD are examples of this.
The hash and LPM libraries are, by design, thread unsafe in order to maintain performance.
However, if required the developer can add layers on top of these libraries to provide thread
safety. Locking is not needed in all situations, and in both the hash and LPM libraries, lookups
of values can be performed in parallel in multiple threads. Adding, removing or modifying
values, however, cannot be done in multiple threads without using locking when a single hash
or LPM table is accessed. Another alternative to locking would be to create multiple instances
of these tables allowing each thread its own copy.
The RX and TX of the PMD are the most critical aspects of a DPDK application and it is
recommended that no locking be used as it will impact performance. Note, however, that these
functions can safely be used from multiple threads when each thread is performing I/O on a
different NIC queue. If multiple threads are to use the same hardware queue on the same NIC
port, then locking, or some other form of mutual exclusion, is necessary.
The ring library is based on a lockless ring-buffer algorithm that maintains its original de-
sign for thread safety. Moreover, it provides high performance for either multi- or single-
consumer/producer enqueue/dequeue operations. The mempool library is based on the DPDK
lockless ring library and therefore is also multi-thread safe.

95
 Programmer’s Guide, Release 2.0.0

20.2 Performance Insensitive API

Outside of the performance sensitive areas described in the previous section, the DPDK pro-
vides a thread-safe API for most other libraries. For example, malloc(librte_malloc) and mem-
zone functions are safe for use in multi-threaded and multi-process environments.
The setup and configuration of the PMD is not performance sensitive, but is not thread safe
either. It is possible that the multiple read/writes during PMD setup and configuration could be
corrupted in a multi-thread environment. Since this is not performance sensitive, the developer
can choose to add their own layer to provide thread-safe setup and configuration. It is expected
that, in most applications, the initial configuration of the network ports would be done by a
single thread at startup.

20.3 Library Initialization

It is recommended that DPDK libraries are initialized in the main thread at application startup
rather than subsequently in the forwarding threads. However, the DPDK performs checks to
ensure that libraries are only initialized once. If initialization is attempted more than once, an
error is returned.
In the multi-process case, the configuration information of shared memory will only be initialized
by the master process. Thereafter, both master and secondary processes can allocate/release
any objects of memory that finally rely on rte_malloc or memzones.

20.4 Interrupt Thread

The DPDK works almost entirely in Linux user space in polling mode. For certain infrequent
operations, such as receiving a PMD link status change notification, callbacks may be called
in an additional thread outside the main DPDK processing threads. These function callbacks
should avoid manipulating DPDK objects that are also managed by the normal DPDK threads,
and if they need to do so, it is up to the application to provide the appropriate locking or mutual
exclusion restrictions around those objects.

20.2. Performance Insensitive API 96

 CHAPTER

TWENTYONE

QUALITY OF SERVICE (QOS) FRAMEWORK

This chapter describes the DPDK Quality of Service (QoS) framework.

21.1 Packet Pipeline with QoS Support

An example of a complex packet processing pipeline with QoS support is shown in the following
figure.

Figure 21.1: Complex Packet Processing Pipeline with QoS Support

This pipeline can be built using reusable DPDK software libraries. The main blocks implement-
ing QoS in this pipeline are: the policer, the dropper and the scheduler. A functional description
of each block is provided in the following table.

97
 Programmer’s Guide, Release 2.0.0

Table 21.1: Packet Processing Pipeline Implementing QoS

# Block Functional Description
1 Packet I/O Packet reception/ transmission from/to multiple NIC ports. Poll mode
RX & TX drivers (PMDs) for Intel 1 GbE/10 GbE NICs.
2 Packet Identify the protocol stack of the input packet. Check the integrity of the
parser packet headers.
3 Flow clas- Map the input packet to one of the known traffic flows. Exact match table
sification lookup using configurable hash function (jhash, CRC and so on) and
bucket logic to handle collisions.
4 Policer Packet metering using srTCM (RFC 2697) or trTCM (RFC2698)
algorithms.
5 Load Distribute the input packets to the application workers. Provide uniform
Balancer load to each worker. Preserve the affinity of traffic flows to workers and
the packet order within each flow.
6 Worker Placeholders for the customer specific application workload (for
threads example, IP stack and so on).
7 Dropper Congestion management using the Random Early Detection (RED)
algorithm (specified by the Sally Floyd - Van Jacobson paper) or
Weighted RED (WRED). Drop packets based on the current scheduler
queue load level and packet priority. When congestion is experienced,
lower priority packets are dropped first.
8 Hierarchi- 5-level hierarchical scheduler (levels are: output port, subport, pipe,
cal traffic class and queue) with thousands (typically 64K) leaf nodes
Scheduler (queues). Implements traffic shaping (for subport and pipe levels), strict
priority (for traffic class level) and Weighted Round Robin (WRR) (for
queues within each pipe traffic class).
The infrastructure blocks used throughout the packet processing pipeline are listed in the fol-
lowing table.

Table 21.2: Infrastructure Blocks Used by the Packet Processing Pipeline

# Block Functional Description
1 Buffer manager Support for global buffer pools and private per-thread buffer caches.
2 Queue manager Support for message passing between pipeline blocks.
3 Power saving Support for power saving during low activity periods.
The mapping of pipeline blocks to CPU cores is configurable based on the performance level
required by each specific application and the set of features enabled for each block. Some
blocks might consume more than one CPU core (with each CPU core running a different
instance of the same block on different input packets), while several other blocks could be
mapped to the same CPU core.

21.2 Hierarchical Scheduler

The hierarchical scheduler block, when present, usually sits on the TX side just before the
transmission stage. Its purpose is to prioritize the transmission of packets from different users
and different traffic classes according to the policy specified by the Service Level Agreements
(SLAs) of each network node.

21.2. Hierarchical Scheduler 98

 Programmer’s Guide, Release 2.0.0

21.2.1 Overview

The hierarchical scheduler block is similar to the traffic manager block used by network proces-
sors that typically implement per flow (or per group of flows) packet queuing and scheduling. It
typically acts like a buffer that is able to temporarily store a large number of packets just before
their transmission (enqueue operation); as the NIC TX is requesting more packets for trans-
mission, these packets are later on removed and handed over to the NIC TX with the packet
selection logic observing the predefined SLAs (dequeue operation).

Figure 21.2: Hierarchical Scheduler Block Internal Diagram

The hierarchical scheduler is optimized for a large number of packet queues. When only a
small number of queues are needed, message passing queues should be used instead of
this block. See the section on Worst Case Scenarios for Performance for a more detailed
discussion.

21.2.2 Scheduling Hierarchy

The scheduling hierarchy is shown in Fig. 21.3. The first level of the hierarchy is the Ethernet
TX port 1/10/40 GbE, with subsequent hierarchy levels defined as subport, pipe, traffic class
and queue.
Typically, each subport represents a predefined group of users, while each pipe represents an
individual user/subscriber. Each traffic class is the representation of a different traffic type with
specific loss rate, delay and jitter requirements, such as voice, video or data transfers. Each
queue hosts packets from one or multiple connections of the same type belonging to the same
user.

21.2. Hierarchical Scheduler 99

 Programmer’s Guide, Release 2.0.0

Figure 21.3: Scheduling Hierarchy per Port

21.2. Hierarchical Scheduler 100

 Programmer’s Guide, Release 2.0.0

The functionality of each hierarchical level is detailed in the following table.

21.2. Hierarchical Scheduler 101

 Programmer’s Guide, Release 2.0.0

Table 21.3: Port Scheduling Hierarchy

# Level Siblings per Parent Functional Descrip-
tion
1 Port
• 1. Output Ethernet
port 1/10/40
GbE.
2. Multiple ports
are scheduled
in round robin
order with all
ports having
equal priority.

2 Subport Configurable (default:

1. Traffic shaping
8)
using token
bucket algo-
rithm (one token
bucket per
subport).
2. Upper limit en-
forced per Traf-
fic Class (TC)
at the subport
level.
3. Lower priority
TCs able to
reuse subport
bandwidth cur-
rently unused by
higher priority
TCs.

3 Pipe Configurable (default:

1. Traffic shaping
4K)
using the token
bucket algo-
rithm (one token
bucket per pipe.

4 Traffic Class (TC) 4

1. TCs of the same
pipe handled in
strict priority or-
der.
2. Upper limit en-
forced per TC at
the pipe level.
3. Lower prior-
ity TCs able
to reuse pipe
bandwidth cur-
rently unused by
21.2. Hierarchical Scheduler 102
higher priority
TCs.
4. When subport
 Programmer’s Guide, Release 2.0.0

21.2.3 Application Programming Interface (API)

Port Scheduler Configuration API

The rte_sched.h file contains configuration functions for port, subport and pipe.

Port Scheduler Enqueue API

The port scheduler enqueue API is very similar to the API of the DPDK PMD TX function.
int rte_sched_port_enqueue(struct rte_sched_port *port,
struct rte_mbuf **pkts,
uint32_t n_pkts);

Port Scheduler Dequeue API

The port scheduler dequeue API is very similar to the API of the DPDK PMD RX function.
int rte_sched_port_dequeue(struct rte_sched_port *port,
struct rte_mbuf **pkts,
uint32_t n_pkts);

Usage Example

/* File "application.c" */

#define N_PKTS_RX 64
#define N_PKTS_TX 48
#define NIC_RX_PORT 0
#define NIC_RX_QUEUE 0
#define NIC_TX_PORT 1
#define NIC_TX_QUEUE 0

struct rte_sched_port *port = NULL;

struct rte_mbuf *pkts_rx[N_PKTS_RX], *pkts_tx[N_PKTS_TX];
uint32_t n_pkts_rx, n_pkts_tx;

/* Initialization */

<initialization code>

/* Runtime */
while (1) {
/* Read packets from NIC RX queue */

n_pkts_rx = rte_eth_rx_burst(NIC_RX_PORT, NIC_RX_QUEUE, pkts_rx, N_PKTS_RX);

/* Hierarchical scheduler enqueue */

rte_sched_port_enqueue(port, pkts_rx, n_pkts_rx);

/* Hierarchical scheduler dequeue */

n_pkts_tx = rte_sched_port_dequeue(port, pkts_tx, N_PKTS_TX);

/* Write packets to NIC TX queue */

21.2. Hierarchical Scheduler 103

 Programmer’s Guide, Release 2.0.0

rte_eth_tx_burst(NIC_TX_PORT, NIC_TX_QUEUE, pkts_tx, n_pkts_tx);

}

21.2.4 Implementation

Internal Data Structures per Port

A schematic of the internal data structures in shown in with details in.

21.2. Hierarchical Scheduler 104

 Programmer’s Guide, Release 2.0.0

Figure 21.4: Internal Data Structures per Port

21.2. Hierarchical Scheduler 105

 Programmer’s Guide, Release 2.0.0

Table 21.4: Scheduler Internal Data Structures per Port

Access type Description
# Data structure Size (bytes) # per port
Enq Deq
1 Subport ta- 64 # subports Rd, Wr Persistent
•
ble entry per port subport data
(credits,
etc).
2 Pipe table 64 # pipes per Rd, Wr Persistent
•
entry port data for
pipe, its
TCs and
its queues
(credits, etc)
that is up-
dated during
run-time.
The pipe
configura-
tion param-
eters do not
change dur-
ing run-time.
The same
pipe con-
figuration
parameters
are shared
by multiple
pipes, there-
fore they are
not part of
pipe table
entry.
3 Queue table 4 #queues per Rd, Wr Rd, Wr Persistent
entry port queue data
(read and
write point-
ers). The
queue size
is the same
per TC for
all queues,
allowing
the queue
base ad-
dress to be
computed
using a fast
formula, so
these two
parameters
are not part
21.2. Hierarchical Scheduler 106of queue
table entry.
The queue
 Programmer’s Guide, Release 2.0.0

Multicore Scaling Strategy

The multicore scaling strategy is:

1. Running different physical ports on different threads. The enqueue and dequeue of the
same port are run by the same thread.
2. Splitting the same physical port to different threads by running different sets of subports
of the same physical port (virtual ports) on different threads. Similarly, a subport can
be split into multiple subports that are each run by a different thread. The enqueue
and dequeue of the same port are run by the same thread. This is only required if, for
performance reasons, it is not possible to handle a full port with a single core.

Enqueue and Dequeue for the Same Output Port

Running enqueue and dequeue operations for the same output port from different cores is likely
to cause significant impact on scheduler’s performance and it is therefore not recommended.
The port enqueue and dequeue operations share access to the following data structures:
1. Packet descriptors
2. Queue table
3. Queue storage area
4. Bitmap of active queues
The expected drop in performance is due to:
1. Need to make the queue and bitmap operations thread safe, which requires either using
locking primitives for access serialization (for example, spinlocks/ semaphores) or using
atomic primitives for lockless access (for example, Test and Set, Compare And Swap, an
so on). The impact is much higher in the former case.
2. Ping-pong of cache lines storing the shared data structures between the cache hierar-
chies of the two cores (done transparently by the MESI protocol cache coherency CPU
hardware).
Therefore, the scheduler enqueue and dequeue operations have to be run from the same
thread, which allows the queues and the bitmap operations to be non-thread safe and keeps
the scheduler data structures internal to the same core.

Performance Scaling

Scaling up the number of NIC ports simply requires a proportional increase in the number of
CPU cores to be used for traffic scheduling.

Enqueue Pipeline

The sequence of steps per packet:

1. Access the mbuf to read the data fields required to identify the destination queue for the
packet. These fields are: port, subport, traffic class and queue within traffic class, and
are typically set by the classification stage.

21.2. Hierarchical Scheduler 107

 Programmer’s Guide, Release 2.0.0

2. Access the queue structure to identify the write location in the queue array. If the queue
is full, then the packet is discarded.
3. Access the queue array location to store the packet (i.e. write the mbuf pointer).
It should be noted the strong data dependency between these steps, as steps 2 and 3 cannot
start before the result from steps 1 and 2 becomes available, which prevents the processor out
of order execution engine to provide any significant performance optimizations.
Given the high rate of input packets and the large amount of queues, it is expected that the
data structures accessed to enqueue the current packet are not present in the L1 or L2 data
cache of the current core, thus the above 3 memory accesses would result (on average) in L1
and L2 data cache misses. A number of 3 L1/L2 cache misses per packet is not acceptable for
performance reasons.
The workaround is to prefetch the required data structures in advance. The prefetch operation
has an execution latency during which the processor should not attempt to access the data
structure currently under prefetch, so the processor should execute other work. The only other
work available is to execute different stages of the enqueue sequence of operations on other
input packets, thus resulting in a pipelined implementation for the enqueue operation.
Fig. 21.5 illustrates a pipelined implementation for the enqueue operation with 4 pipeline stages
and each stage executing 2 different input packets. No input packet can be part of more than
one pipeline stage at a given time.

Figure 21.5: Prefetch Pipeline for the Hierarchical Scheduler Enqueue Operation

The congestion management scheme implemented by the enqueue pipeline described above
is very basic: packets are enqueued until a specific queue becomes full, then all the packets
destined to the same queue are dropped until packets are consumed (by the dequeue oper-
ation). This can be improved by enabling RED/WRED as part of the enqueue pipeline which
looks at the queue occupancy and packet priority in order to yield the enqueue/drop decision for
a specific packet (as opposed to enqueuing all packets / dropping all packets indiscriminately).

Dequeue State Machine

The sequence of steps to schedule the next packet from the current pipe is:
1. Identify the next active pipe using the bitmap scan operation, prefetch pipe.
2. Read pipe data structure. Update the credits for the current pipe and its subport. Identify
the first active traffic class within the current pipe, select the next queue using WRR,
prefetch queue pointers for all the 16 queues of the current pipe.
3. Read next element from the current WRR queue and prefetch its packet descriptor.
4. Read the packet length from the packet descriptor (mbuf structure). Based on the packet
length and the available credits (of current pipe, pipe traffic class, subport and subport
traffic class), take the go/no go scheduling decision for the current packet.

21.2. Hierarchical Scheduler 108

 Programmer’s Guide, Release 2.0.0

To avoid the cache misses, the above data structures (pipe, queue, queue array, mbufs) are
prefetched in advance of being accessed. The strategy of hiding the latency of the prefetch
operations is to switch from the current pipe (in grinder A) to another pipe (in grinder B) imme-
diately after a prefetch is issued for the current pipe. This gives enough time to the prefetch
operation to complete before the execution switches back to this pipe (in grinder A).
The dequeue pipe state machine exploits the data presence into the processor cache, therefore
it tries to send as many packets from the same pipe TC and pipe as possible (up to the available
packets and credits) before moving to the next active TC from the same pipe (if any) or to
another active pipe.

Figure 21.6: Pipe Prefetch State Machine for the Hierarchical Scheduler Dequeue Operation

21.2. Hierarchical Scheduler 109

 Programmer’s Guide, Release 2.0.0

Timing and Synchronization

The output port is modeled as a conveyor belt of byte slots that need to be filled by the sched-
uler with data for transmission. For 10 GbE, there are 1.25 billion byte slots that need to be
filled by the port scheduler every second. If the scheduler is not fast enough to fill the slots, pro-
vided that enough packets and credits exist, then some slots will be left unused and bandwidth
will be wasted.
In principle, the hierarchical scheduler dequeue operation should be triggered by NIC TX.
Usually, once the occupancy of the NIC TX input queue drops below a predefined threshold,
the port scheduler is woken up (interrupt based or polling based, by continuously monitoring
the queue occupancy) to push more packets into the queue.

Internal Time Reference

The scheduler needs to keep track of time advancement for the credit logic, which requires
credit updates based on time (for example, subport and pipe traffic shaping, traffic class upper
limit enforcement, and so on).
Every time the scheduler decides to send a packet out to the NIC TX for transmission, the
scheduler will increment its internal time reference accordingly. Therefore, it is convenient
to keep the internal time reference in units of bytes, where a byte signifies the time duration
required by the physical interface to send out a byte on the transmission medium. This way,
as a packet is scheduled for transmission, the time is incremented with (n + h), where n is the
packet length in bytes and h is the number of framing overhead bytes per packet.

Internal Time Reference Re-synchronization

The scheduler needs to align its internal time reference to the pace of the port conveyor belt.
The reason is to make sure that the scheduler does not feed the NIC TX with more bytes than
the line rate of the physical medium in order to prevent packet drop (by the scheduler, due to
the NIC TX input queue being full, or later on, internally by the NIC TX).
The scheduler reads the current time on every dequeue invocation. The CPU time stamp can
be obtained by reading either the Time Stamp Counter (TSC) register or the High Precision
Event Timer (HPET) register. The current CPU time stamp is converted from number of CPU
clocks to number of bytes: time_bytes = time_cycles / cycles_per_byte, where cycles_per_byte
is the amount of CPU cycles that is equivalent to the transmission time for one byte on the wire
(e.g. for a CPU frequency of 2 GHz and a 10GbE port,*cycles_per_byte = 1.6*).
The scheduler maintains an internal time reference of the NIC time. Whenever a packet is
scheduled, the NIC time is incremented with the packet length (including framing overhead).
On every dequeue invocation, the scheduler checks its internal reference of the NIC time
against the current time:
1. If NIC time is in the future (NIC time >= current time), no adjustment of NIC time is
needed. This means that scheduler is able to schedule NIC packets before the NIC
actually needs those packets, so the NIC TX is well supplied with packets;
2. If NIC time is in the past (NIC time < current time), then NIC time should be adjusted by
setting it to the current time. This means that the scheduler is not able to keep up with
the speed of the NIC byte conveyor belt, so NIC bandwidth is wasted due to poor packet
supply to the NIC TX.

21.2. Hierarchical Scheduler 110

 Programmer’s Guide, Release 2.0.0

Scheduler Accuracy and Granularity

The scheduler round trip delay (SRTD) is the time (number of CPU cycles) between two con-
secutive examinations of the same pipe by the scheduler.
To keep up with the output port (that is, avoid bandwidth loss), the scheduler should be able to
schedule n packets faster than the same n packets are transmitted by NIC TX.
The scheduler needs to keep up with the rate of each individual pipe, as configured for the pipe
token bucket, assuming that no port oversubscription is taking place. This means that the size
of the pipe token bucket should be set high enough to prevent it from overflowing due to big
SRTD, as this would result in credit loss (and therefore bandwidth loss) for the pipe.

Credit Logic

Scheduling Decision

The scheduling decision to send next packet from (subport S, pipe P, traffic class TC, queue
Q) is favorable (packet is sent) when all the conditions below are met:
• Pipe P of subport S is currently selected by one of the port grinders;
• Traffic class TC is the highest priority active traffic class of pipe P;
• Queue Q is the next queue selected by WRR within traffic class TC of pipe P;
• Subport S has enough credits to send the packet;
• Subport S has enough credits for traffic class TC to send the packet;
• Pipe P has enough credits to send the packet;
• Pipe P has enough credits for traffic class TC to send the packet.
If all the above conditions are met, then the packet is selected for transmission and the nec-
essary credits are subtracted from subport S, subport S traffic class TC, pipe P, pipe P traffic
class TC.

Framing Overhead

As the greatest common divisor for all packet lengths is one byte, the unit of credit is selected
as one byte. The number of credits required for the transmission of a packet of n bytes is equal
to (n+h), where h is equal to the number of framing overhead bytes per packet.

21.2. Hierarchical Scheduler 111

 Programmer’s Guide, Release 2.0.0

Table 21.5: Ethernet Frame Overhead Fields

# Packet field Length Comments
(bytes)
1 Preamble 7
2 Start of Frame 1
Delimiter (SFD)
3 Frame Check 4 Considered overhead only if not included in the
Sequence (FCS) mbuf packet length field.
4 Inter Frame Gap 12
(IFG)
5 Total 24

Traffic Shaping

The traffic shaping for subport and pipe is implemented using a token bucket per subport/per
pipe. Each token bucket is implemented using one saturated counter that keeps track of the
number of available credits.
The token bucket generic parameters and operations are presented in Table 6 and Table 7.

Table 21.6: Token Bucket Generic Operations

# Token Bucket Unit Description
Parameter
1 bucket_rate Credits per Rate of adding credits to the bucket.
second
2 bucket_size Credits Max number of credits that can be stored in
the bucket.
Table 21.7: Token Bucket Generic Parameters
# Token Description
Bucket
Operation
1 Initialization Bucket set to a predefined value, e.g. zero or half of the bucket size.
2 Credit Credits are added to the bucket on top of existing ones, either
update periodically or on demand, based on the bucket_rate. Credits cannot
exceed the upper limit defined by the bucket_size, so any credits to be
added to the bucket while the bucket is full are dropped.
3 Credit con- As result of packet scheduling, the necessary number of credits is
sumption removed from the bucket. The packet can only be sent if enough credits
are in the bucket to send the full packet (packet bytes and framing
overhead for the packet).
To implement the token bucket generic operations described above, the current design uses the
persistent data structure presented in, while the implementation of the token bucket operations
is described in Table 9.

21.2. Hierarchical Scheduler 112

 Programmer’s Guide, Release 2.0.0

Table 21.8: Token Bucket Persistent Data Structure

# Token Unit Description
bucket
field
1 tb_time BytesTime of the last credit update. Measured in bytes instead of seconds
or CPU cycles for ease of credit consumption operation (as the
current time is also maintained in bytes).
See the section on Internal Time Reference for an explanation of
why the time is maintained in byte units.
2 tb_period BytesTime period that should elapse since the last credit update in order
for the bucket to be awarded tb_credits_per_period worth or credits.
3 tb_credits_per_period
BytesCredit allowance per tb_period.
4 tb_size BytesBucket size, i.e. upper limit for the tb_credits.
5 tb_credits BytesNumber of credits currently in the bucket.
The bucket rate (in bytes per second) can be computed with the following formula:
bucket_rate = (tb_credits_per_period / tb_period) * r
where, r = port line rate (in bytes per second).

21.2. Hierarchical Scheduler 113

 Programmer’s Guide, Release 2.0.0

Table 21.9: Token Bucket Operations

# Token bucket operation Description
1 Initialization tb_credits = 0; or tb_credits =
tb_size / 2;
2 Credit update Credit update options:
• Every time a packet is
sent for a port, update
the credits of all the the
subports and pipes of
that port. Not feasible.
• Every time a packet is
sent, update the cred-
its for the pipe and sub-
port. Very accurate, but
not needed (a lot of cal-
culations).
• Every time a pipe is se-
lected (that is, picked by
one of the grinders), up-
date the credits for the
pipe and its subport.
The current implementation
is using option 3. Accord-
ing to Section Dequeue State
Machine, the pipe and sub-
port credits are updated every
time a pipe is selected by the
dequeue process before the
pipe and subport credits are
actually used.
The implementation uses a
tradeoff between accuracy
and speed by updating the
bucket credits only when at
least a full tb_period has
elapsed since the last update.
• Full accuracy can
be achieved by se-
lecting the value for
tb_period for which
tb_credits_per_period =
1.
• When full accuracy is
not required, better per-
formance is achieved by
setting tb_credits to a
larger value.
Update operations:
• n_periods = (time -
tb_time) / tb_period;
• tb_credits += n_periods
* tb_credits_per_period;
21.2. Hierarchical Scheduler • tb_credits 114 =
min(tb_credits, tb_size);
• tb_time += n_periods *
 Programmer’s Guide, Release 2.0.0

Traffic Classes

Implementation of Strict Priority Scheduling Strict priority scheduling of traffic classes

within the same pipe is implemented by the pipe dequeue state machine, which selects the
queues in ascending order. Therefore, queues 0..3 (associated with TC 0, highest priority TC)
are handled before queues 4..7 (TC 1, lower priority than TC 0), which are handled before
queues 8..11 (TC 2), which are handled before queues 12..15 (TC 3, lowest priority TC).

Upper Limit Enforcement The traffic classes at the pipe and subport levels are not traffic
shaped, so there is no token bucket maintained in this context. The upper limit for the traffic
classes at the subport and pipe levels is enforced by periodically refilling the subport / pipe
traffic class credit counter, out of which credits are consumed every time a packet is scheduled
for that subport / pipe, as described in Table 10 and Table 11.

Table 21.10: Subport/Pipe Traffic Class Upper Limit Enforcement Persistent Data Structure
# Subport Unit Description
or pipe
field
1 tc_time BytesTime of the next update (upper limit refill) for the 4 TCs of the current
subport / pipe.
See the section on Internal Time Reference for the explanation of
why the time is maintained in byte units.
2 tc_period BytesTime between two consecutive updates for the 4 TCs of the current
subport / pipe. This is expected to be many times bigger than the
typical value of the token bucket tb_period.
3 tc_credits_per_period
BytesUpper limit for the number of credits allowed to be consumed by the
current TC during each enforcement period tc_period.
4 tc_credits BytesCurrent upper limit for the number of credits that can be consumed
by the current traffic class for the remainder of the current
enforcement period.
Table 21.11: Subport/Pipe Traffic Class Upper Limit Enforcement Operations
# Traffic Class Description
Operation
1 Initialization tc_credits = tc_credits_per_period;
tc_time = tc_period;
2 Credit update Update operations:
if (time >= tc_time) {
tc_credits = tc_credits_per_period;
tc_time = time + tc_period;
}
3 Credit As result of packet scheduling, the TC limit is decreased with the
consumption (on necessary number of credits. The packet can only be sent if
packet scheduling) enough credits are currently available in the TC limit to send the
full packet (packet bytes and framing overhead for the packet).
Scheduling operations:
pkt_credits = pk_len + frame_overhead;
if (tc_credits >= pkt_credits) {tc_credits -= pkt_credits;}

21.2. Hierarchical Scheduler 115

 Programmer’s Guide, Release 2.0.0

Weighted Round Robin (WRR)

The evolution of the WRR design solution from simple to complex is shown in Table 12.

21.2. Hierarchical Scheduler 116

 Programmer’s Guide, Release 2.0.0

Table 21.12: Weighted Round Robin (WRR)

# All Equal All Strategy
Queues Weights for Pack-
Ac- All ets
tive? Queues? Equal?
1 Yes Yes Yes Byte level round robin
Next queue queue #i, i = (i + 1) % n
2 Yes Yes No Packet level round robin
Consuming one byte from queue #i requires
consuming exactly one token for queue #i.
T(i) = Accumulated number of tokens previously
consumed from queue #i. Every time a packet is
consumed from queue #i, T(i) is updated as: T(i) +=
pkt_len.
Next queue : queue with the smallest T.
3 Yes No No Packet level weighted round robin
This case can be reduced to the previous case by
introducing a cost per byte that is different for each
queue. Queues with lower weights have a higher
cost per byte. This way, it is still meaningful to
compare the consumption among different queues in
order to select the next queue.
w(i) = Weight of queue #i
t(i) = Tokens per byte for queue #i, defined as the
inverse weight of queue #i. For example, if w[0..3] =
[1:2:4:8], then t[0..3] = [8:4:2:1]; if w[0..3] =
[1:4:15:20], then t[0..3] = [60:15:4:3]. Consuming one
byte from queue #i requires consuming t(i) tokens for
queue #i.
T(i) = Accumulated number of tokens previously
consumed from queue #i. Every time a packet is
consumed from queue #i, T(i) is updated as: T(i) +=
pkt_len * t(i). Next queue : queue with the smallest T.
4 No No No Packet level weighted round robin with variable
queue status
Reduce this case to the previous case by setting the
consumption of inactive queues to a high number, so
that the inactive queues will never be selected by the
smallest T logic.
To prevent T from overflowing as result of successive
accumulations, T(i) is truncated after each packet
consumption for all queues. For example, T[0..3] =
[1000, 1100, 1200, 1300] is truncated to T[0..3] = [0,
100, 200, 300] by subtracting the min T from T(i), i =
0..n.
This requires having at least one active queue in the
set of input queues, which is guaranteed by the
dequeue state machine never selecting an inactive
traffic class.
mask(i) = Saturation mask for queue #i, defined as:
mask(i) = (queue #i is active)? 0 : 0xFFFFFFFF;
w(i) = Weight of queue #i
21.2. Hierarchical Scheduler t(i) = Tokens per byte for queue #i, defined as the 117
inverse weight of queue #i.
T(i) = Accumulated numbers of tokens previously
 Programmer’s Guide, Release 2.0.0

Subport Traffic Class Oversubscription

Problem Statement Oversubscription for subport traffic class X is a configuration-time event

that occurs when more bandwidth is allocated for traffic class X at the level of subport member
pipes than allocated for the same traffic class at the parent subport level.
The existence of the oversubscription for a specific subport and traffic class is solely the result
of pipe and subport-level configuration as opposed to being created due to dynamic evolution
of the traffic load at run-time (as congestion is).
When the overall demand for traffic class X for the current subport is low, the existence of
the oversubscription condition does not represent a problem, as demand for traffic class X is
completely satisfied for all member pipes. However, this can no longer be achieved when the
aggregated demand for traffic class X for all subport member pipes exceeds the limit configured
at the subport level.

Solution Space summarizes some of the possible approaches for handling this problem,
with the third approach selected for implementation.

Table 21.13: Subport Traffic Class Oversubscription

No. Approach Description
1 Don’t care First come, first served.
This approach is not fair among subport member pipes, as pipes that
are served first will use up as much bandwidth for TC X as they need,
while pipes that are served later will receive poor service due to
bandwidth for TC X at the subport level being scarce.
2 Scale down All pipes within the subport have their bandwidth limit for TC X scaled
all pipes down by the same factor.
This approach is not fair among subport member pipes, as the low end
pipes (that is, pipes configured with low bandwidth) can potentially
experience severe service degradation that might render their service
unusable (if available bandwidth for these pipes drops below the
minimum requirements for a workable service), while the service
degradation for high end pipes might not be noticeable at all.
3 Cap the high Each subport member pipe receives an equal share of the bandwidth
demand available at run-time for TC X at the subport level. Any bandwidth left
pipes unused by the low-demand pipes is redistributed in equal portions to
the high-demand pipes. This way, the high-demand pipes are
truncated while the low-demand pipes are not impacted.
Typically, the subport TC oversubscription feature is enabled only for the lowest priority traffic
class (TC 3), which is typically used for best effort traffic, with the management plane prevent-
ing this condition from occurring for the other (higher priority) traffic classes.
To ease implementation, it is also assumed that the upper limit for subport TC 3 is set to 100%
of the subport rate, and that the upper limit for pipe TC 3 is set to 100% of pipe rate for all
subport member pipes.

Implementation Overview The algorithm computes a watermark, which is periodically up-

dated based on the current demand experienced by the subport member pipes, whose purpose
is to limit the amount of traffic that each pipe is allowed to send for TC 3. The watermark is

21.2. Hierarchical Scheduler 118

 Programmer’s Guide, Release 2.0.0

computed at the subport level at the beginning of each traffic class upper limit enforcement
period and the same value is used by all the subport member pipes throughout the current
enforcement period. illustrates how the watermark computed as subport level at the beginning
of each period is propagated to all subport member pipes.
At the beginning of the current enforcement period (which coincides with the end of the pre-
vious enforcement period), the value of the watermark is adjusted based on the amount of
bandwidth allocated to TC 3 at the beginning of the previous period that was not left unused
by the subport member pipes at the end of the previous period.
If there was subport TC 3 bandwidth left unused, the value of the watermark for the current
period is increased to encourage the subport member pipes to consume more bandwidth. Oth-
erwise, the value of the watermark is decreased to enforce equality of bandwidth consumption
among subport member pipes for TC 3.
The increase or decrease in the watermark value is done in small increments, so several
enforcement periods might be required to reach the equilibrium state. This state can change
at any moment due to variations in the demand experienced by the subport member pipes for
TC 3, for example, as a result of demand increase (when the watermark needs to be lowered)
or demand decrease (when the watermark needs to be increased).
When demand is low, the watermark is set high to prevent it from impeding the subport member
pipes from consuming more bandwidth. The highest value for the watermark is picked as
the highest rate configured for a subport member pipe. Table 15 illustrates the watermark
operation.

21.2. Hierarchical Scheduler 119

 Programmer’s Guide, Release 2.0.0

Table 21.14: Watermark Propagation from Subport Level to Member Pipes at the Beginning of
Each Traffic Class Upper Limit Enforcement Period
No. Subport Traffic Class Opera- Description
tion
1 Initialization Subport level: sub-
port_period_id= 0
Pipe level: pipe_period_id =
0
2 Credit update Subport Level:
if (time>=subport_tc_time)
{ subport_wm = wa-
ter_mark_update();
subport_tc_time = time
+ subport_tc_period;
subport_period_id++;
}
Pipelevel:
if(pipe_period_id != sub-
port_period_id)
{
pipe_ov_credits
= subport_wm *
pipe_weight;
pipe_period_id
= sub-
port_period_id;
}
3 Credit consumption (on Pipe level:
packet scheduling) pkt_credits = pk_len +
frame_overhead;
if(pipe_ov_credits >=
pkt_credits{
pipe_ov_credits -
= pkt_credits;
}

21.2. Hierarchical Scheduler 120

 Programmer’s Guide, Release 2.0.0

Table 21.15: Watermark Calculation

No. Subport Traffic Class Opera- Description
tion
1 Initialization Subport level:
wm = WM_MAX
2 Credit update Subport level (wa-
ter_mark_update):
tc0_cons = sub-
port_tc0_credits_per_period -
subport_tc0_credits;
tc1_cons = sub-
port_tc1_credits_per_period -
subport_tc1_credits;
tc2_cons = sub-
port_tc2_credits_per_period -
subport_tc2_credits;
tc3_cons = sub-
port_tc3_credits_per_period -
subport_tc3_credits;
tc3_cons_max = sub-
port_tc3_credits_per_period
- (tc0_cons + tc1_cons +
tc2_cons);
if(tc3_consumption >
(tc3_consumption_max -
MTU)){
wm -= wm >> 7;
if(wm < WM_MIN)
wm = WM_MIN;
} else {
wm += (wm >> 7)
+ 1;
if(wm >
WM_MAX) wm =
WM_MAX;
}

21.2.5 Worst Case Scenarios for Performance

Lots of Active Queues with Not Enough Credits

The more queues the scheduler has to examine for packets and credits in order to select one
packet, the lower the performance of the scheduler is.
The scheduler maintains the bitmap of active queues, which skips the non-active queues, but
in order to detect whether a specific pipe has enough credits, the pipe has to be drilled down
using the pipe dequeue state machine, which consumes cycles regardless of the scheduling
result (no packets are produced or at least one packet is produced).
This scenario stresses the importance of the policer for the scheduler performance: if the pipe

21.2. Hierarchical Scheduler 121

 Programmer’s Guide, Release 2.0.0

does not have enough credits, its packets should be dropped as soon as possible (before they
reach the hierarchical scheduler), thus rendering the pipe queues as not active, which allows
the dequeue side to skip that pipe with no cycles being spent on investigating the pipe credits
that would result in a “not enough credits” status.

Single Queue with 100% Line Rate

The port scheduler performance is optimized for a large number of queues. If the number of
queues is small, then the performance of the port scheduler for the same level of active traffic
is expected to be worse than the performance of a small set of message passing queues.

21.3 Dropper

The purpose of the DPDK dropper is to drop packets arriving at a packet scheduler to avoid
congestion. The dropper supports the Random Early Detection (RED), Weighted Random
Early Detection (WRED) and tail drop algorithms. Fig. 21.7 illustrates how the dropper inte-
grates with the scheduler. The DPDK currently does not support congestion management so
the dropper provides the only method for congestion avoidance.

Figure 21.7: High-level Block Diagram of the DPDK Dropper

The dropper uses the Random Early Detection (RED) congestion avoidance algorithm as doc-
umented in the reference publication. The purpose of the RED algorithm is to monitor a packet
queue, determine the current congestion level in the queue and decide whether an arriving
packet should be enqueued or dropped. The RED algorithm uses an Exponential Weighted

21.3. Dropper 122

 Programmer’s Guide, Release 2.0.0

Moving Average (EWMA) filter to compute average queue size which gives an indication of the
current congestion level in the queue.
For each enqueue operation, the RED algorithm compares the average queue size to minimum
and maximum thresholds. Depending on whether the average queue size is below, above or in
between these thresholds, the RED algorithm calculates the probability that an arriving packet
should be dropped and makes a random decision based on this probability.
The dropper also supports Weighted Random Early Detection (WRED) by allowing the sched-
uler to select different RED configurations for the same packet queue at run-time. In the case
of severe congestion, the dropper resorts to tail drop. This occurs when a packet queue has
reached maximum capacity and cannot store any more packets. In this situation, all arriving
packets are dropped.
The flow through the dropper is illustrated in Fig. 21.8. The RED/WRED algorithm is exercised
first and tail drop second.
The use cases supported by the dropper are:
• – Initialize configuration data
• – Initialize run-time data
• – Enqueue (make a decision to enqueue or drop an arriving packet)
• – Mark empty (record the time at which a packet queue becomes empty)
The configuration use case is explained in Configuration, the enqueue operation is explained
in Enqueue Operation and the mark empty operation is explained in Queue Empty Operation.

21.3.1 Configuration

A RED configuration contains the parameters given in Table 16.

Table 21.16: RED Configuration Parameters

Parameter Minimum Maximum Typical
Minimum Threshold 0 1022 1/4 x queue size
Maximum Threshold 1 1023 1/2 x queue size
Inverse Mark Probability 1 255 10
EWMA Filter Weight 1 12 9
The meaning of these parameters is explained in more detail in the following sections. The
format of these parameters as specified to the dropper module API corresponds to the format
used by Cisco* in their RED implementation. The minimum and maximum threshold parame-
ters are specified to the dropper module in terms of number of packets. The mark probability
parameter is specified as an inverse value, for example, an inverse mark probability parameter
value of 10 corresponds to a mark probability of 1/10 (that is, 1 in 10 packets will be dropped).
The EWMA filter weight parameter is specified as an inverse log value, for example, a filter
weight parameter value of 9 corresponds to a filter weight of 1/29.

21.3.2 Enqueue Operation

In the example shown in Fig. 21.9, q (actual queue size) is the input value, avg (average
queue size) and count (number of packets since the last drop) are run-time values, decision is

21.3. Dropper 123

 Programmer’s Guide, Release 2.0.0

Figure 21.8: Flow Through the Dropper

21.3. Dropper 124

 Programmer’s Guide, Release 2.0.0

the output value and the remaining values are configuration parameters.

Figure 21.9: Example Data Flow Through Dropper

EWMA Filter Microblock

The purpose of the EWMA Filter microblock is to filter queue size values to smooth out transient
changes that result from “bursty” traffic. The output value is the average queue size which gives
a more stable view of the current congestion level in the queue.
The EWMA filter has one configuration parameter, filter weight, which determines how quickly
or slowly the average queue size output responds to changes in the actual queue size input.
Higher values of filter weight mean that the average queue size responds more quickly to
changes in actual queue size.

Average Queue Size Calculation when the Queue is not Empty

The definition of the EWMA filter is given in the following equation.

Where:
• avg = average queue size
• wq = filter weight
• q = actual queue size

Note:
The filter weight, wq = 1/2^n, where n is the filter weight parameter value passed to the dropper modu
on configuration (see Configuration ).

21.3. Dropper 125

 Programmer’s Guide, Release 2.0.0

Average Queue Size Calculation when the Queue is Empty

The EWMA filter does not read time stamps and instead assumes that enqueue operations
will happen quite regularly. Special handling is required when the queue becomes empty as
the queue could be empty for a short time or a long time. When the queue becomes empty,
average queue size should decay gradually to zero instead of dropping suddenly to zero or
remaining stagnant at the last computed value. When a packet is enqueued on an empty
queue, the average queue size is computed using the following formula:

Where:
• m = the number of enqueue operations that could have occurred on this queue while the
queue was empty
In the dropper module, m is defined as:

Where:
• time = current time
• qtime = time the queue became empty
• s = typical time between successive enqueue operations on this queue
The time reference is in units of bytes, where a byte signifies the time duration required by
the physical interface to send out a byte on the transmission medium (see the section on
Internal Time Reference). The parameter s is defined in the dropper module as a constant
with the value: s=2^22. This corresponds to the time required by every leaf node in a hierarchy
with 64K leaf nodes to transmit one 64-byte packet onto the wire and represents the worst
case scenario. For much smaller scheduler hierarchies, it may be necessary to reduce the
parameter s, which is defined in the red header source file (rte_red.h) as:
#define RTE_RED_S

Since the time reference is in bytes, the port speed is implied in the expression: time-qtime.
The dropper does not have to be configured with the actual port speed. It adjusts automatically
to low speed and high speed links.

Implementation

A numerical method is used to compute the factor (1-wq)^m that appears in Equation 2.
This method is based on the following identity:

This allows us to express the following:

21.3. Dropper 126

 Programmer’s Guide, Release 2.0.0

In the dropper module, a look-up table is used to compute log2(1-wq) for each value of wq
supported by the dropper module. The factor (1-wq)^m can then be obtained by multiplying
the table value by m and applying shift operations. To avoid overflow in the multiplication, the
value, m, and the look-up table values are limited to 16 bits. The total size of the look-up table
is 56 bytes. Once the factor (1-wq)^m is obtained using this method, the average queue size
can be calculated from Equation 2.

Alternative Approaches

Other methods for calculating the factor (1-wq)^m in the expression for computing average
queue size when the queue is empty (Equation 2) were considered. These approaches include:
• Floating-point evaluation
• Fixed-point evaluation using a small look-up table (512B) and up to 16 multiplications
(this is the approach used in the FreeBSD* ALTQ RED implementation)
• Fixed-point evaluation using a small look-up table (512B) and 16 SSE multiplications
(SSE optimized version of the approach used in the FreeBSD* ALTQ RED implementa-
tion)
• Large look-up table (76 KB)
The method that was finally selected (described in the “Implementation” section above) out
performs all of these approaches in terms of run-time performance and memory requirements
and also achieves accuracy comparable to floating-point evaluation. Table 17 lists the per-
formance of each of these alternative approaches relative to the method that is used in the
dropper. As can be seen, the floating-point implementation achieved the worst performance.

Table 21.17: Relative Performance of Alternative Approaches

Method Relative Performance
Current dropper method (see Dropper ) 100%
Fixed-point method with small (512B) look-up table 148%
SSE method with small (512B) look-up table 114%
Large (76KB) look-up table 118%
Floating-point 595%
Note: In this case, since performance is expressed as time spent executing the operation in a
specific condition, any relative performance value above 100% runs slower than the reference
method.

Drop Decision Block

The Drop Decision block:

• Compares the average queue size with the minimum and maximum thresholds
• Calculates a packet drop probability
• Makes a random decision to enqueue or drop an arriving packet

21.3. Dropper 127

 Programmer’s Guide, Release 2.0.0

The calculation of the drop probability occurs in two stages. An initial drop probability is calcu-
lated based on the average queue size, the minimum and maximum thresholds and the mark
probability. An actual drop probability is then computed from the initial drop probability. The
actual drop probability takes the count run-time value into consideration so that the actual drop
probability increases as more packets arrive to the packet queue since the last packet was
dropped.

Initial Packet Drop Probability

The initial drop probability is calculated using the following equation.

Where:
• maxp = mark probability
• avg = average queue size
• minth = minimum threshold
• maxth = maximum threshold
The calculation of the packet drop probability using Equation 3 is illustrated in Fig. 21.10. If
the average queue size is below the minimum threshold, an arriving packet is enqueued. If the
average queue size is at or above the maximum threshold, an arriving packet is dropped. If
the average queue size is between the minimum and maximum thresholds, a drop probability
is calculated to determine if the packet should be enqueued or dropped.

Actual Drop Probability

If the average queue size is between the minimum and maximum thresholds, then the actual
drop probability is calculated from the following equation.

Where:
• Pb = initial drop probability (from Equation 3)
• count = number of packets that have arrived since the last drop
The constant 2, in Equation 4 is the only deviation from the drop probability formulae given in
the reference document where a value of 1 is used instead. It should be noted that the value pa
computed from can be negative or greater than 1. If this is the case, then a value of 1 should
be used instead.
The initial and actual drop probabilities are shown in Fig. 21.11. The actual drop probabil-
ity is shown for the case where the formula given in the reference document1 is used (blue
curve) and also for the case where the formula implemented in the dropper module, is used
(red curve). The formula in the reference document results in a significantly higher drop rate
compared to the mark probability configuration parameter specified by the user. The choice to

21.3. Dropper 128

 Programmer’s Guide, Release 2.0.0

Figure 21.10: Packet Drop Probability for a Given RED Configuration

21.3. Dropper 129

 Programmer’s Guide, Release 2.0.0

deviate from the reference document is simply a design decision and one that has been taken
by other RED implementations, for example, FreeBSD* ALTQ RED.

Figure 21.11: Initial Drop Probability (pb), Actual Drop probability (pa) Computed Using a
Factor 1 (Blue Curve) and a Factor 2 (Red Curve)

21.3.3 Queue Empty Operation

The time at which a packet queue becomes empty must be recorded and saved with the RED
run-time data so that the EWMA filter block can calculate the average queue size on the next
enqueue operation. It is the responsibility of the calling application to inform the dropper mod-
ule through the API that a queue has become empty.

21.3.4 Source Files Location

The source files for the DPDK dropper are located at:
• DPDK/lib/librte_sched/rte_red.h
• DPDK/lib/librte_sched/rte_red.c

21.3.5 Integration with the DPDK QoS Scheduler

RED functionality in the DPDK QoS scheduler is disabled by default. To enable it, use the
DPDK configuration parameter:
CONFIG_RTE_SCHED_RED=y

This parameter must be set to y. The parameter is found in the build configuration files in
the DPDK/config directory, for example, DPDK/config/common_linuxapp. RED configuration
parameters are specified in the rte_red_params structure within the rte_sched_port_params

21.3. Dropper 130

 Programmer’s Guide, Release 2.0.0

structure that is passed to the scheduler on initialization. RED parameters are specified sep-
arately for four traffic classes and three packet colors (green, yellow and red) allowing the
scheduler to implement Weighted Random Early Detection (WRED).

21.3.6 Integration with the DPDK QoS Scheduler Sample Application

The DPDK QoS Scheduler Application reads a configuration file on start-up. The configura-
tion file includes a section containing RED parameters. The format of these parameters is
described in Configuration. A sample RED configuration is shown below. In this example, the
queue size is 64 packets.

Note: For correct operation, the same EWMA filter weight parameter (wred weight) should be
used for each packet color (green, yellow, red) in the same traffic class (tc).

; RED params per traffic class and color (Green / Yellow / Red)

[red]
tc 0 wred min = 28 22 16
tc 0 wred max = 32 32 32
tc 0 wred inv prob = 10 10 10
tc 0 wred weight = 9 9 9

tc 1 wred min = 28 22 16
tc 1 wred max = 32 32 32
tc 1 wred inv prob = 10 10 10
tc 1 wred weight = 9 9 9

tc 2 wred min = 28 22 16
tc 2 wred max = 32 32 32
tc 2 wred inv prob = 10 10 10
tc 2 wred weight = 9 9 9

tc 3 wred min = 28 22 16
tc 3 wred max = 32 32 32
tc 3 wred inv prob = 10 10 10
tc 3 wred weight = 9 9 9

With this configuration file, the RED configuration that applies to green, yellow and red packets
in traffic class 0 is shown in Table 18.

Table 21.18: RED Configuration Corresponding to RED Configuration

File
RED Parameter Configuration Name Green Yellow Red
Minimum Threshold tc 0 wred min 28 22 16
Maximum Threshold tc 0 wred max 32 32 32
Mark Probability tc 0 wred inv prob 10 10 10
EWMA Filter Weight tc 0 wred weight 9 9 9

21.3.7 Application Programming Interface (API)

Enqueue API

The syntax of the enqueue API is as follows:

21.3. Dropper 131

 Programmer’s Guide, Release 2.0.0

int rte_red_enqueue(const struct rte_red_config *red_cfg,

struct rte_red *red,
const unsigned q,
const uint64_t time)

The arguments passed to the enqueue API are configuration data, run-time data, the current
size of the packet queue (in packets) and a value representing the current time. The time
reference is in units of bytes, where a byte signifies the time duration required by the physical
interface to send out a byte on the transmission medium (see the section on Internal Time
Reference). The dropper reuses the scheduler time stamps for performance reasons.

Empty API

The syntax of the empty API is as follows:

void rte_red_mark_queue_empty(struct rte_red *red, const uint64_t time)

The arguments passed to the empty API are run-time data and the current time in bytes.

21.4 Traffic Metering

The traffic metering component implements the Single Rate Three Color Marker (srTCM) and
Two Rate Three Color Marker (trTCM) algorithms, as defined by IETF RFC 2697 and 2698
respectively. These algorithms meter the stream of incoming packets based on the allowance
defined in advance for each traffic flow. As result, each incoming packet is tagged as green,
yellow or red based on the monitored consumption of the flow the packet belongs to.

21.4.1 Functional Overview

The srTCM algorithm defines two token buckets for each traffic flow, with the two buckets
sharing the same token update rate:
• Committed (C) bucket: fed with tokens at the rate defined by the Committed Information
Rate (CIR) parameter (measured in IP packet bytes per second). The size of the C bucket
is defined by the Committed Burst Size (CBS) parameter (measured in bytes);
• Excess (E) bucket: fed with tokens at the same rate as the C bucket. The size of the E
bucket is defined by the Excess Burst Size (EBS) parameter (measured in bytes).
The trTCM algorithm defines two token buckets for each traffic flow, with the two buckets being
updated with tokens at independent rates:
• Committed (C) bucket: fed with tokens at the rate defined by the Committed Information
Rate (CIR) parameter (measured in bytes of IP packet per second). The size of the C
bucket is defined by the Committed Burst Size (CBS) parameter (measured in bytes);
• Peak (P) bucket: fed with tokens at the rate defined by the Peak Information Rate (PIR)
parameter (measured in IP packet bytes per second). The size of the P bucket is defined
by the Peak Burst Size (PBS) parameter (measured in bytes).
Please refer to RFC 2697 (for srTCM) and RFC 2698 (for trTCM) for details on how tokens are
consumed from the buckets and how the packet color is determined.

21.4. Traffic Metering 132

 Programmer’s Guide, Release 2.0.0

Color Blind and Color Aware Modes

For both algorithms, the color blind mode is functionally equivalent to the color aware mode
with input color set as green. For color aware mode, a packet with red input color can only get
the red output color, while a packet with yellow input color can only get the yellow or red output
colors.
The reason why the color blind mode is still implemented distinctly than the color aware mode
is that color blind mode can be implemented with fewer operations than the color aware mode.

21.4.2 Implementation Overview

For each input packet, the steps for the srTCM / trTCM algorithms are:
• Update the C and E / P token buckets. This is done by reading the current time (from
the CPU timestamp counter), identifying the amount of time since the last bucket update
and computing the associated number of tokens (according to the pre-configured bucket
rate). The number of tokens in the bucket is limited by the pre-configured bucket size;
• Identify the output color for the current packet based on the size of the IP packet and the
amount of tokens currently available in the C and E / P buckets; for color aware mode
only, the input color of the packet is also considered. When the output color is not red, a
number of tokens equal to the length of the IP packet are subtracted from the C or E /P
or both buckets, depending on the algorithm and the output color of the packet.

21.4. Traffic Metering 133

 CHAPTER

TWENTYTWO

POWER MANAGEMENT

The DPDK Power Management feature allows users space applications to save power by dy-
namically adjusting CPU frequency or entering into different C-States.
• Adjusting the CPU frequency dynamically according to the utilization of RX queue.
• Entering into different deeper C-States according to the adaptive algorithms to speculate
brief periods of time suspending the application if no packets are received.
The interfaces for adjusting the operating CPU frequency are in the power management library.
C-State control is implemented in applications according to the different use cases.

22.1 CPU Frequency Scaling

The Linux kernel provides a cpufreq module for CPU frequency scaling for each lcore. For
example, for cpuX, /sys/devices/system/cpu/cpuX/cpufreq/ has the following sys files for fre-
quency scaling:
• affected_cpus
• bios_limit
• cpuinfo_cur_freq
• cpuinfo_max_freq
• cpuinfo_min_freq
• cpuinfo_transition_latency
• related_cpus
• scaling_available_frequencies
• scaling_available_governors
• scaling_cur_freq
• scaling_driver
• scaling_governor
• scaling_max_freq
• scaling_min_freq
• scaling_setspeed

134
 Programmer’s Guide, Release 2.0.0

In the DPDK, scaling_governor is configured in user space. Then, a user space application
can prompt the kernel by writing scaling_setspeed to adjust the CPU frequency according to
the strategies defined by the user space application.

22.2 Core-load Throttling through C-States

Core state can be altered by speculative sleeps whenever the specified lcore has nothing to
do. In the DPDK, if no packet is received after polling, speculative sleeps can be triggered
according the strategies defined by the user space application.

22.3 API Overview of the Power Library

The main methods exported by power library are for CPU frequency scaling and include the
following:
• Freq up: Prompt the kernel to scale up the frequency of the specific lcore.
• Freq down: Prompt the kernel to scale down the frequency of the specific lcore.
• Freq max: Prompt the kernel to scale up the frequency of the specific lcore to the maxi-
mum.
• Freq min: Prompt the kernel to scale down the frequency of the specific lcore to the
minimum.
• Get available freqs: Read the available frequencies of the specific lcore from the sys
file.
• Freq get: Get the current frequency of the specific lcore.
• Freq set: Prompt the kernel to set the frequency for the specific lcore.

22.4 User Cases

The power management mechanism is used to save power when performing L3 forwarding.

22.5 References

• l3fwd-power: The sample application in DPDK that performs L3 forwarding with power
management.
• The “L3 Forwarding with Power Management Sample Application” chapter in the DPDK
Sample Application’s User Guide.

22.2. Core-load Throttling through C-States 135

 CHAPTER

TWENTYTHREE

PACKET CLASSIFICATION AND ACCESS CONTROL

The DPDK provides an Access Control library that gives the ability to classify an input packet
based on a set of classification rules.
The ACL library is used to perform an N-tuple search over a set of rules with multiple categories
and find the best match (highest priority) for each category. The library API provides the
following basic operations:
• Create a new Access Control (AC) context.
• Add rules into the context.
• For all rules in the context, build the runtime structures necessary to perform packet
classification.
• Perform input packet classifications.
• Destroy an AC context and its runtime structures and free the associated memory.

23.1 Overview

23.1.1 Rule definition

The current implementation allows the user for each AC context to specify its own rule (set of
fields) over which packet classification will be performed. Though there are few restrictions on
the rule fields layout:
• First field in the rule definition has to be one byte long.
• All subsequent fields has to be grouped into sets of 4 consecutive bytes.
This is done mainly for performance reasons - search function processes the first input byte as
part of the flow setup and then the inner loop of the search function is unrolled to process four
input bytes at a time.
To define each field inside an AC rule, the following structure is used:
struct rte_acl_field_def {
uint8_t type; /*< type - ACL_FIELD_TYPE. */
uint8_t size; /*< size of field 1,2,4, or 8. */
uint8_t field_index; /*< index of field inside the rule. */
uint8_t input_index; /*< 0-N input index. */
uint32_t offset; /*< offset to start of field. */
};

136
 Programmer’s Guide, Release 2.0.0

• type The field type is one of three choices:

– _MASK - for fields such as IP addresses that have a value and a mask defining the
number of relevant bits.
– _RANGE - for fields such as ports that have a lower and upper value for the field.
– _BITMASK - for fields such as protocol identifiers that have a value and a bit mask.
• size The size parameter defines the length of the field in bytes. Allowable values are 1,
2, 4, or 8 bytes. Note that due to the grouping of input bytes, 1 or 2 byte fields must be
defined as consecutive fields that make up 4 consecutive input bytes. Also, it is best to
define fields of 8 or more bytes as 4 byte fields so that the build processes can eliminate
fields that are all wild.
• field_index A zero-based value that represents the position of the field inside the rule; 0
to N-1 for N fields.
• input_index As mentioned above, all input fields, except the very first one, must be in
groups of 4 consecutive bytes. The input index specifies to which input group that field
belongs to.
• offset The offset field defines the offset for the field. This is the offset from the beginning
of the buffer parameter for the search.
For example, to define classification for the following IPv4 5-tuple structure:
struct ipv4_5tuple {
uint8_t proto;
uint32_t ip_src;
uint32_t ip_dst;
uint16_t port_src;
uint16_t port_dst;
};

The following array of field definitions can be used:

struct rte_acl_field_def ipv4_defs[5] = {
/* first input field - always one byte long. */
{
.type = RTE_ACL_FIELD_TYPE_BITMASK,
.size = sizeof (uint8_t),
.field_index = 0,
.input_index = 0,
.offset = offsetof (struct ipv4_5tuple, proto),
},

/* next input field (IPv4 source address) - 4 consecutive bytes. */

{
.type = RTE_ACL_FIELD_TYPE_MASK,
.size = sizeof (uint32_t),
.field_index = 1,
.input_index = 1,
.offset = offsetof (struct ipv4_5tuple, ip_src),
},

/* next input field (IPv4 destination address) - 4 consecutive bytes. */

{
.type = RTE_ACL_FIELD_TYPE_MASK,
.size = sizeof (uint32_t),
.field_index = 2,
.input_index = 2,

23.1. Overview 137

 Programmer’s Guide, Release 2.0.0

.offset = offsetof (struct ipv4_5tuple, ip_dst),

},

/*
* Next 2 fields (src & dst ports) form 4 consecutive bytes.
* They share the same input index.
*/
{
.type = RTE_ACL_FIELD_TYPE_RANGE,
.size = sizeof (uint16_t),
.field_index = 3,
.input_index = 3,
.offset = offsetof (struct ipv4_5tuple, port_src),
},

{
.type = RTE_ACL_FIELD_TYPE_RANGE,
.size = sizeof (uint16_t),
.field_index = 4,
.input_index = 3,
.offset = offsetof (struct ipv4_5tuple, port_dst),
},
};

A typical example of such an IPv4 5-tuple rule is a follows:

source addr/mask destination addr/mask source ports dest ports protocol/mask
192.168.1.0/24 192.168.2.31/32 0:65535 1234:1234 17/0xff

Any IPv4 packets with protocol ID 17 (UDP), source address 192.168.1.[0-255], destination
address 192.168.2.31, source port [0-65535] and destination port 1234 matches the above
rule.
To define classification for the IPv6 2-tuple: <protocol, IPv6 source address> over the following
IPv6 header structure:
struct struct ipv6_hdr {
uint32_t vtc_flow; /* IP version, traffic class & flow label. */
uint16_t payload_len; /* IP packet length - includes sizeof(ip_header). */
uint8_t proto; /* Protocol, next header. */
uint8_t hop_limits; /* Hop limits. */
uint8_t src_addr[16]; /* IP address of source host. */
uint8_t dst_addr[16]; /* IP address of destination host(s). */
} __attribute__((__packed__));

The following array of field definitions can be used:

struct struct rte_acl_field_def ipv6_2tuple_defs[5] = {
{
.type = RTE_ACL_FIELD_TYPE_BITMASK,
.size = sizeof (uint8_t),
.field_index = 0,
.input_index = 0,
.offset = offsetof (struct ipv6_hdr, proto),
},

{
.type = RTE_ACL_FIELD_TYPE_MASK,
.size = sizeof (uint32_t),
.field_index = 1,
.input_index = 1,
.offset = offsetof (struct ipv6_hdr, src_addr[0]),

23.1. Overview 138

 Programmer’s Guide, Release 2.0.0

},

{
.type = RTE_ACL_FIELD_TYPE_MASK,
.size = sizeof (uint32_t),
.field_index = 2,
.input_index = 2,
.offset = offsetof (struct ipv6_hdr, src_addr[4]),
},

{
.type = RTE_ACL_FIELD_TYPE_MASK,
.size = sizeof (uint32_t),
.field_index = 3,
.input_index = 3,
.offset = offsetof (struct ipv6_hdr, src_addr[8]),
},

{
.type = RTE_ACL_FIELD_TYPE_MASK,
.size = sizeof (uint32_t),
.field_index = 4,
.input_index = 4,
.offset = offsetof (struct ipv6_hdr, src_addr[12]),
},
};

A typical example of such an IPv6 2-tuple rule is a follows:

source addr/mask protocol/mask
2001:db8:1234:0000:0000:0000:0000:0000/48 6/0xff

Any IPv6 packets with protocol ID 6 (TCP), and source address inside the range
[2001:db8:1234:0000:0000:0000:0000:0000 - 2001:db8:1234:ffff:ffff:ffff:ffff:ffff] matches the
above rule.
When creating a set of rules, for each rule, additional information must be supplied also:
• priority: A weight to measure the priority of the rules (higher is better). If the input tuple
matches more than one rule, then the rule with the higher priority is returned. Note that
if the input tuple matches more than one rule and these rules have equal priority, it is
undefined which rule is returned as a match. It is recommended to assign a unique
priority for each rule.
• category_mask: Each rule uses a bit mask value to select the relevant category(s) for
the rule. When a lookup is performed, the result for each category is returned. This ef-
fectively provides a “parallel lookup” by enabling a single search to return multiple results
if, for example, there were four different sets of ACL rules, one for access control, one for
routing, and so on. Each set could be assigned its own category and by combining them
into a single database, one lookup returns a result for each of the four sets.
• userdata: A user-defined field that could be any value except zero. For each category, a
successful match returns the userdata field of the highest priority matched rule.

Note: When adding new rules into an ACL context, all fields must be in host byte order (LSB).
When the search is performed for an input tuple, all fields in that tuple must be in network byte
order (MSB).

23.1. Overview 139

 Programmer’s Guide, Release 2.0.0

23.1.2 RT memory size limit

Build phase (rte_acl_build()) creates for a given set of rules internal structure for further run-
time traversal. With current implementation it is a set of multi-bit tries (with stride == 8).
Depending on the rules set, that could consume significant amount of memory. In attempt
to conserve some space ACL build process tries to split the given rule-set into several non-
intersecting subsets and construct a separate trie for each of them. Depending on the rule-set,
it might reduce RT memory requirements but might increase classification time. There is a
possibility at build-time to specify maximum memory limit for internal RT structures for given
AC context. It could be done via max_size field of the rte_acl_config structure. Setting it to
the value greater than zero, instructs rte_acl_build() to:
• attempt to minimize number of tries in the RT table, but
• make sure that size of RT table wouldn’t exceed given value.
Setting it to zero makes rte_acl_build() to use the default behavior: try to minimize size of the
RT structures, but doesn’t expose any hard limit on it.
That gives the user the ability to decisions about performance/space trade-off. For example:
struct rte_acl_ctx * acx;
struct rte_acl_config cfg;
int ret;

/*
* assuming that acx points to already created and
* populated with rules AC context and cfg filled properly.
*/

/* try to build AC context, with RT structures less then 8MB. */

cfg.max_size = 0x800000;
ret = rte_acl_build(acx, &cfg);

/*
* RT structures can't fit into 8MB for given context.
* Try to build without exposing any hard limit.
*/
if (ret == -ERANGE) {
cfg.max_size = 0;
ret = rte_acl_build(acx, &cfg);
}

23.1.3 Classification methods

After rte_acl_build() over given AC context has finished successfully, it can be used to perform
classification - search for a rule with highest priority over the input data. There are several
implementations of classify algorithm:
• RTE_ACL_CLASSIFY_SCALAR: generic implementation, doesn’t require any specific
HW support.
• RTE_ACL_CLASSIFY_SSE: vector implementation, can process up to 8 flows in paral-
lel. Requires SSE 4.1 support.
• RTE_ACL_CLASSIFY_AVX2: vector implementation, can process up to 16 flows in par-
allel. Requires AVX2 support.

23.1. Overview 140

 Programmer’s Guide, Release 2.0.0

It is purely a runtime decision which method to choose, there is no build-time difference. All
implementations operates over the same internal RT structures and use similar principles. The
main difference is that vector implementations can manually exploit IA SIMD instructions and
process several input data flows in parallel. At startup ACL library determines the highest
available classify method for the given platform and sets it as default one. Though the user has
an ability to override the default classifier function for a given ACL context or perform particular
search using non-default classify method. In that case it is user responsibility to make sure
that given platform supports selected classify implementation.

23.2 Application Programming Interface (API) Usage

Note: For more details about the Access Control API, please refer to the DPDK API Refer-
ence.

The following example demonstrates IPv4, 5-tuple classification for rules defined above with
multiple categories in more detail.

23.2.1 Classify with Multiple Categories

struct rte_acl_ctx * acx;

struct rte_acl_config cfg;
int ret;

/* define a structure for the rule with up to 5 fields. */

RTE_ACL_RULE_DEF(acl_ipv4_rule, RTE_DIM(ipv4_defs));

/* AC context creation parameters. */

struct rte_acl_param prm = {

.name = "ACL_example",
.socket_id = SOCKET_ID_ANY,
.rule_size = RTE_ACL_RULE_SZ(RTE_DIM(ipv4_defs)),

/* number of fields per rule. */

.max_rule_num = 8, /* maximum number of rules in the AC context. */

};

struct acl_ipv4_rule acl_rules[] = {

/* matches all packets traveling to 192.168.0.0/16,

* applies for categories: 0,1 */
{
.data = {.userdata = 1, .category_mask = 3, .priority = 1},

/* destination IPv4 */
.field[2] = {.value.u32 = IPv4(192,168,0,0),. mask_range.u32 = 16,},

/* source port */
.field[3] = {.value.u16 = 0, .mask_range.u16 = 0xffff,},

/* destination port */
.field[4] = {.value.u16 = 0, .mask_range.u16 = 0xffff,},
},

23.2. Application Programming Interface (API) Usage 141

 Programmer’s Guide, Release 2.0.0

/* matches all packets traveling to 192.168.1.0/24,

* applies for categories: 0 */
{
.data = {.userdata = 2, .category_mask = 1, .priority = 2},

/* destination IPv4 */
.field[2] = {.value.u32 = IPv4(192,168,1,0),. mask_range.u32 = 24,},

/* source port */
.field[3] = {.value.u16 = 0, .mask_range.u16 = 0xffff,},

/* destination port */
.field[4] = {.value.u16 = 0, .mask_range.u16 = 0xffff,},
},

/* matches all packets traveling from 10.1.1.1,

* applies for categories: 1 */
{
.data = {.userdata = 3, .category_mask = 2, .priority = 3},

/* source IPv4 */
.field[1] = {.value.u32 = IPv4(10,1,1,1),. mask_range.u32 = 32,},

/* source port */
.field[3] = {.value.u16 = 0, .mask_range.u16 = 0xffff,},

/* destination port */
.field[4] = {.value.u16 = 0, .mask_range.u16 = 0xffff,},
},

};

/* create an empty AC context */

if ((acx = rte_acl_create(&prm)) == NULL) {

/* handle context create failure. */

/* add rules to the context */

ret = rte_acl_add_rules(acx, acl_rules, RTE_DIM(acl_rules));

if (ret != 0) {
/* handle error at adding ACL rules. */
}

/* prepare AC build config. */

cfg.num_categories = 2;
cfg.num_fields = RTE_DIM(ipv4_defs);

memcpy(cfg.defs, ipv4_defs, sizeof (ipv4_defs));

/* build the runtime structures for added rules, with 2 categories. */

ret = rte_acl_build(acx, &cfg);

if (ret != 0) {
/* handle error at build runtime structures for ACL context. */
}

23.2. Application Programming Interface (API) Usage 142

 Programmer’s Guide, Release 2.0.0

For a tuple with source IP address: 10.1.1.1 and destination IP address: 192.168.1.15, once
the following lines are executed:
uint32_t results[4]; /* make classify for 4 categories. */

rte_acl_classify(acx, data, results, 1, 4);

then the results[] array contains:

results[4] = {2, 3, 0, 0};

• For category 0, both rules 1 and 2 match, but rule 2 has higher priority, therefore results[0]
contains the userdata for rule 2.
• For category 1, both rules 1 and 3 match, but rule 3 has higher priority, therefore results[1]
contains the userdata for rule 3.
• For categories 2 and 3, there are no matches, so results[2] and results[3] contain zero,
which indicates that no matches were found for those categories.
For a tuple with source IP address: 192.168.1.1 and destination IP address: 192.168.2.11,
once the following lines are executed:
uint32_t results[4]; /* make classify by 4 categories. */

rte_acl_classify(acx, data, results, 1, 4);

the results[] array contains:

results[4] = {1, 1, 0, 0};

• For categories 0 and 1, only rule 1 matches.

• For categories 2 and 3, there are no matches.
For a tuple with source IP address: 10.1.1.1 and destination IP address: 201.212.111.12, once
the following lines are executed:
uint32_t results[4]; /* make classify by 4 categories. */
rte_acl_classify(acx, data, results, 1, 4);

the results[] array contains:

results[4] = {0, 3, 0, 0};

• For category 1, only rule 3 matches.

• For categories 0, 2 and 3, there are no matches.

23.2. Application Programming Interface (API) Usage 143

 CHAPTER

TWENTYFOUR

PACKET FRAMEWORK

24.1 Design Objectives

The main design objectives for the DPDK Packet Framework are:
• Provide standard methodology to build complex packet processing pipelines. Provide
reusable and extensible templates for the commonly used pipeline functional blocks;
• Provide capability to switch between pure software and hardware-accelerated implemen-
tations for the same pipeline functional block;
• Provide the best trade-off between flexibility and performance. Hardcoded pipelines usu-
ally provide the best performance, but are not flexible, while developing flexible frame-
works is never a problem, but performance is usually low;
• Provide a framework that is logically similar to Open Flow.

24.2 Overview

Packet processing applications are frequently structured as pipelines of multiple stages, with
the logic of each stage glued around a lookup table. For each incoming packet, the table
defines the set of actions to be applied to the packet, as well as the next stage to send the
packet to.
The DPDK Packet Framework minimizes the development effort required to build packet pro-
cessing pipelines by defining a standard methodology for pipeline development, as well as
providing libraries of reusable templates for the commonly used pipeline blocks.
The pipeline is constructed by connecting the set of input ports with the set of output ports
through the set of tables in a tree-like topology. As result of lookup operation for the current
packet in the current table, one of the table entries (on lookup hit) or the default table entry (on
lookup miss) provides the set of actions to be applied on the current packet, as well as the next
hop for the packet, which can be either another table, an output port or packet drop.
An example of packet processing pipeline is presented in Fig. 24.1:

144
 Programmer’s Guide, Release 2.0.0

Figure 24.1: Example of Packet Processing Pipeline where Input Ports 0 and 1 are Connected
with Output Ports 0, 1 and 2 through Tables 0 and 1

24.3 Port Library Design

24.3.1 Port Types

Table 19 is a non-exhaustive list of ports that can be implemented with the Packet Framework.

Table 24.1: Port Types

# Port Description
type
1 SW ring SW circular buffer used for message passing between the application
threads. Uses the DPDK rte_ring primitive. Expected to be the most
commonly used type of port.
2 HW ring Queue of buffer descriptors used to interact with NIC, switch or accelerator
ports. For NIC ports, it uses the DPDK rte_eth_rx_queue or
rte_eth_tx_queue primitives.
3 IP re- Input packets are either IP fragments or complete IP datagrams. Output
assem- packets are complete IP datagrams.
bly
4 IP frag- Input packets are jumbo (IP datagrams with length bigger than MTU) or
menta- non-jumbo packets. Output packets are non-jumbo packets.
tion
5 Traffic Traffic manager attached to a specific NIC output port, performing
man- congestion management and hierarchical scheduling according to
ager pre-defined SLAs.
6 KNI Send/receive packets to/from Linux kernel space.
7 Source Input port used as packet generator. Similar to Linux kernel /dev/zero
character device.
8 Sink Output port used to drop all input packets. Similar to Linux kernel /dev/null
character device.

24.3. Port Library Design 145

 Programmer’s Guide, Release 2.0.0

24.3.2 Port Interface

Each port is unidirectional, i.e. either input port or output port. Each input/output port is
required to implement an abstract interface that defines the initialization and run-time operation
of the port. The port abstract interface is described in.

Table 24.2: 20 Port Abstract Interface

# Port Description
Operation
1 Create Create the low-level port object (e.g. queue). Can internally allocate
memory.
2 Free Free the resources (e.g. memory) used by the low-level port object.
3 RX Read a burst of input packets. Non-blocking operation. Only defined for
input ports.
4 TX Write a burst of input packets. Non-blocking operation. Only defined for
output ports.
5 Flush Flush the output buffer. Only defined for output ports.

24.4 Table Library Design

24.4.1 Table Types

Table 21 is a non-exhaustive list of types of tables that can be implemented with the Packet
Framework.

24.4. Table Library Design 146

 Programmer’s Guide, Release 2.0.0

Table 24.3: Table Types

# Table Type Description
1 Hash table Lookup key is n-tuple based.
Typically, the lookup key is hashed to produce a signature that is used
to identify a bucket of entries where the lookup key is searched next.
The signature associated with the lookup key of each input packet is
either read from the packet descriptor (pre-computed signature) or
computed at table lookup time.
The table lookup, add entry and delete entry operations, as well as
any other pipeline block that pre-computes the signature all have to
use the same hashing algorithm to generate the signature.
Typically used to implement flow classification tables, ARP caches,
routing table for tunnelling protocols, etc.
2 Longest Lookup key is the IP address.
Prefix Match Each table entries has an associated IP prefix (IP and depth).
(LPM) The table lookup operation selects the IP prefix that is matched by the
lookup key; in case of multiple matches, the entry with the longest
prefix depth wins.
Typically used to implement IP routing tables.
3 Access Lookup key is 7-tuple of two VLAN/MPLS labels, IP destination
Control List address, IP source addresses, L4 protocol, L4 destination port, L4
(ACLs) source port.
Each table entry has an associated ACL and priority. The ACL
contains bit masks for the VLAN/MPLS labels, IP prefix for IP
destination address, IP prefix for IP source addresses, L4 protocol and
bitmask, L4 destination port and bit mask, L4 source port and bit
mask.
The table lookup operation selects the ACL that is matched by the
lookup key; in case of multiple matches, the entry with the highest
priority wins.
Typically used to implement rule databases for firewalls, etc.
4 Pattern Lookup key is the packet payload.
matching Table is a database of patterns, with each pattern having a priority
search assigned.
The table lookup operation selects the patterns that is matched by the
input packet; in case of multiple matches, the matching pattern with
the highest priority wins.
5 Array Lookup key is the table entry index itself.

24.4.2 Table Interface

Each table is required to implement an abstract interface that defines the initialization and
run-time operation of the table. The table abstract interface is described in Table 29.

24.4. Table Library Design 147

 Programmer’s Guide, Release 2.0.0

Table 24.4: Table Abstract Interface

# Table Description
opera-
tion
1 Create Create the low-level data structures of the lookup table. Can internally
allocate memory.
2 Free Free up all the resources used by the lookup table.
3 Add Add new entry to the lookup table.
entry
4 Delete Delete specific entry from the lookup table.
entry
5 Lookup Look up a burst of input packets and return a bit mask specifying the result
of the lookup operation for each packet: a set bit signifies lookup hit for the
corresponding packet, while a cleared bit a lookup miss.
For each lookup hit packet, the lookup operation also returns a pointer to
the table entry that was hit, which contains the actions to be applied on the
packet and any associated metadata.
For each lookup miss packet, the actions to be applied on the packet and
any associated metadata are specified by the default table entry
preconfigured for lookup miss.

24.4.3 Hash Table Design

Hash Table Overview

Hash tables are important because the key lookup operation is optimized for speed: instead of
having to linearly search the lookup key through all the keys in the table, the search is limited
to only the keys stored in a single table bucket.
Associative Arrays
An associative array is a function that can be specified as a set of (key, value) pairs, with each
key from the possible set of input keys present at most once. For a given associative array, the
possible operations are:
1. add (key, value): When no value is currently associated with key, then the (key, value ) as-
sociation is created. When key is already associated value value0, then the association
(key, value0) is removed and association (key, value) is created;
2. delete key : When no value is currently associated with key, this operation has no effect.
When key is already associated value, then association (key, value) is removed;
3. lookup key : When no value is currently associated with key, then this operation returns
void value (lookup miss). When key is associated with value, then this operation returns
value. The (key, value) association is not changed.
The matching criterion used to compare the input key against the keys in the associative array
is exact match, as the key size (number of bytes) and the key value (array of bytes) have to
match exactly for the two keys under comparison.
Hash Function

24.4. Table Library Design 148

 Programmer’s Guide, Release 2.0.0

A hash function deterministically maps data of variable length (key) to data of fixed size (hash
value or key signature). Typically, the size of the key is bigger than the size of the key signature.
The hash function basically compresses a long key into a short signature. Several keys can
share the same signature (collisions).
High quality hash functions have uniform distribution. For large number of keys, when dividing
the space of signature values into a fixed number of equal intervals (buckets), it is desirable
to have the key signatures evenly distributed across these intervals (uniform distribution), as
opposed to most of the signatures going into only a few of the intervals and the rest of the
intervals being largely unused (non-uniform distribution).
Hash Table
A hash table is an associative array that uses a hash function for its operation. The reason for
using a hash function is to optimize the performance of the lookup operation by minimizing the
number of table keys that have to be compared against the input key.
Instead of storing the (key, value) pairs in a single list, the hash table maintains multiple lists
(buckets). For any given key, there is a single bucket where that key might exist, and this bucket
is uniquely identified based on the key signature. Once the key signature is computed and the
hash table bucket identified, the key is either located in this bucket or it is not present in the
hash table at all, so the key search can be narrowed down from the full set of keys currently in
the table to just the set of keys currently in the identified table bucket.
The performance of the hash table lookup operation is greatly improved, provided that the table
keys are evenly distributed among the hash table buckets, which can be achieved by using a
hash function with uniform distribution. The rule to map a key to its bucket can simply be to
use the key signature (modulo the number of table buckets) as the table bucket ID:
bucket_id = f_hash(key) % n_buckets;
By selecting the number of buckets to be a power of two, the modulo operator can be replaced
by a bitwise AND logical operation:
bucket_id = f_hash(key) & (n_buckets - 1);
considering n_bits as the number of bits set in bucket_mask = n_buckets - 1, this means that
all the keys that end up in the same hash table bucket have the lower n_bits of their signature
identical. In order to reduce the number of keys in the same bucket (collisions), the number of
hash table buckets needs to be increased.
In packet processing context, the sequence of operations involved in hash table operations is
described in Fig. 24.2:

Figure 24.2: Sequence of Steps for Hash Table Operations in a Packet Processing Context

Hash Table Use Cases

Flow Classification

24.4. Table Library Design 149

 Programmer’s Guide, Release 2.0.0

Description: The flow classification is executed at least once for each input packet. This oper-
ation maps each incoming packet against one of the known traffic flows in the flow database
that typically contains millions of flows.
Hash table name: Flow classification table
Number of keys: Millions
Key format: n-tuple of packet fields that uniquely identify a traffic flow/connection. Example:
DiffServ 5-tuple of (Source IP address, Destination IP address, L4 protocol, L4 protocol source
port, L4 protocol destination port). For IPv4 protocol and L4 protocols like TCP, UDP or SCTP,
the size of the DiffServ 5-tuple is 13 bytes, while for IPv6 it is 37 bytes.
Key value (key data): actions and action meta-data describing what processing to be applied
for the packets of the current flow. The size of the data associated with each traffic flow can
vary from 8 bytes to kilobytes.
Address Resolution Protocol (ARP)
Description: Once a route has been identified for an IP packet (so the output interface and
the IP address of the next hop station are known), the MAC address of the next hop station is
needed in order to send this packet onto the next leg of the journey towards its destination (as
identified by its destination IP address). The MAC address of the next hop station becomes
the destination MAC address of the outgoing Ethernet frame.
Hash table name: ARP table
Number of keys: Thousands
Key format: The pair of (Output interface, Next Hop IP address), which is typically 5 bytes for
IPv4 and 17 bytes for IPv6.
Key value (key data): MAC address of the next hop station (6 bytes).

Hash Table Types

Table 22 lists the hash table configuration parameters shared by all different hash table types.

Table 24.5: Configuration Parameters Common for All Hash Table Types
# Parameter Details
1 Key size Measured as number of bytes. All keys have the same size.
2 Key value (key Measured as number of bytes.
data) size
3 Number of Needs to be a power of two.
buckets
4 Maximum number Needs to be a power of two.
of keys
5 Hash function Examples: jhash, CRC hash, etc.
6 Hash function Parameter to be passed to the hash function.
seed
7 Key offset Offset of the lookup key byte array within the packet meta-data
stored in the packet buffer.

24.4. Table Library Design 150

 Programmer’s Guide, Release 2.0.0

Bucket Full Problem

On initialization, each hash table bucket is allocated space for exactly 4 keys. As keys are
added to the table, it can happen that a given bucket already has 4 keys when a new key has
to be added to this bucket. The possible options are:
1. Least Recently Used (LRU) Hash Table. One of the existing keys in the bucket is
deleted and the new key is added in its place. The number of keys in each bucket never
grows bigger than 4. The logic to pick the key to be dropped from the bucket is LRU. The
hash table lookup operation maintains the order in which the keys in the same bucket are
hit, so every time a key is hit, it becomes the new Most Recently Used (MRU) key, i.e.
the last candidate for drop. When a key is added to the bucket, it also becomes the new
MRU key. When a key needs to be picked and dropped, the first candidate for drop, i.e.
the current LRU key, is always picked. The LRU logic requires maintaining specific data
structures per each bucket.
2. Extendable Bucket Hash Table. The bucket is extended with space for 4 more keys.
This is done by allocating additional memory at table initialization time, which is used to
create a pool of free keys (the size of this pool is configurable and always a multiple of 4).
On key add operation, the allocation of a group of 4 keys only happens successfully within
the limit of free keys, otherwise the key add operation fails. On key delete operation, a
group of 4 keys is freed back to the pool of free keys when the key to be deleted is the
only key that was used within its group of 4 keys at that time. On key lookup operation,
if the current bucket is in extended state and a match is not found in the first group of 4
keys, the search continues beyond the first group of 4 keys, potentially until all keys in
this bucket are examined. The extendable bucket logic requires maintaining specific data
structures per table and per each bucket.

Table 24.6: Configuration Parameters Specific to Extendable Bucket Hash Table

# Parameter Details
1 Number of additional keys Needs to be a power of two, at least equal to 4.

Signature Computation

The possible options for key signature computation are:

1. Pre-computed key signature. The key lookup operation is split between two CPU cores.
The first CPU core (typically the CPU core that performs packet RX) extracts the key
from the input packet, computes the key signature and saves both the key and the key
signature in the packet buffer as packet meta-data. The second CPU core reads both
the key and the key signature from the packet meta-data and performs the bucket search
step of the key lookup operation.
2. Key signature computed on lookup (“do-sig” version). The same CPU core reads the
key from the packet meta-data, uses it to compute the key signature and also performs
the bucket search step of the key lookup operation.

Table 24.7: Configuration Parameters Specific to Pre-computed Key Signature Hash Table
# Parameter Details
1 Signature Offset of the pre-computed key signature within the packet
offset meta-data.

24.4. Table Library Design 151

 Programmer’s Guide, Release 2.0.0

Key Size Optimized Hash Tables

For specific key sizes, the data structures and algorithm of key lookup operation can be spe-
cially handcrafted for further performance improvements, so following options are possible:
1. Implementation supporting configurable key size.
2. Implementation supporting a single key size. Typical key sizes are 8 bytes and 16
bytes.

Bucket Search Logic for Configurable Key Size Hash Tables

The performance of the bucket search logic is one of the main factors influencing the perfor-
mance of the key lookup operation. The data structures and algorithm are designed to make
the best use of Intel CPU architecture resources like: cache memory space, cache memory
bandwidth, external memory bandwidth, multiple execution units working in parallel, out of
order instruction execution, special CPU instructions, etc.
The bucket search logic handles multiple input packets in parallel. It is built as a pipeline of
several stages (3 or 4), with each pipeline stage handling two different packets from the burst
of input packets. On each pipeline iteration, the packets are pushed to the next pipeline stage:
for the 4-stage pipeline, two packets (that just completed stage 3) exit the pipeline, two packets
(that just completed stage 2) are now executing stage 3, two packets (that just completed stage
1) are now executing stage 2, two packets (that just completed stage 0) are now executing
stage 1 and two packets (next two packets to read from the burst of input packets) are entering
the pipeline to execute stage 0. The pipeline iterations continue until all packets from the burst
of input packets execute the last stage of the pipeline.
The bucket search logic is broken into pipeline stages at the boundary of the next memory
access. Each pipeline stage uses data structures that are stored (with high probability) into
the L1 or L2 cache memory of the current CPU core and breaks just before the next memory
access required by the algorithm. The current pipeline stage finalizes by prefetching the data
structures required by the next pipeline stage, so given enough time for the prefetch to com-
plete, when the next pipeline stage eventually gets executed for the same packets, it will read
the data structures it needs from L1 or L2 cache memory and thus avoid the significant penalty
incurred by L2 or L3 cache memory miss.
By prefetching the data structures required by the next pipeline stage in advance (before they
are used) and switching to executing another pipeline stage for different packets, the number of
L2 or L3 cache memory misses is greatly reduced, hence one of the main reasons for improved
performance. This is because the cost of L2/L3 cache memory miss on memory read accesses
is high, as usually due to data dependency between instructions, the CPU execution units have
to stall until the read operation is completed from L3 cache memory or external DRAM memory.
By using prefetch instructions, the latency of memory read accesses is hidden, provided that it
is preformed early enough before the respective data structure is actually used.
By splitting the processing into several stages that are executed on different packets (the pack-
ets from the input burst are interlaced), enough work is created to allow the prefetch instruc-
tions to complete successfully (before the prefetched data structures are actually accessed)
and also the data dependency between instructions is loosened. For example, for the 4-stage
pipeline, stage 0 is executed on packets 0 and 1 and then, before same packets 0 and 1 are
used (i.e. before stage 1 is executed on packets 0 and 1), different packets are used: packets

24.4. Table Library Design 152

 Programmer’s Guide, Release 2.0.0

2 and 3 (executing stage 1), packets 4 and 5 (executing stage 2) and packets 6 and 7 (exe-
cuting stage 3). By executing useful work while the data structures are brought into the L1 or
L2 cache memory, the latency of the read memory accesses is hidden. By increasing the gap
between two consecutive accesses to the same data structure, the data dependency between
instructions is loosened; this allows making the best use of the super-scalar and out-of-order
execution CPU architecture, as the number of CPU core execution units that are active (rather
than idle or stalled due to data dependency constraints between instructions) is maximized.
The bucket search logic is also implemented without using any branch instructions. This avoids
the important cost associated with flushing the CPU core execution pipeline on every instance
of branch misprediction.

Configurable Key Size Hash Table

Fig. 24.3, Table 25 and Table 26 detail the main data structures used to implement configurable
key size hash tables (either LRU or extendable bucket, either with pre-computed signature or
“do-sig”).

Figure 24.3: Data Structures for Configurable Key Size Hash Tables

Table 24.8: Main Large Data Structures (Arrays) used for Configurable Key Size Hash Tables
# Array name Number of Entry size Description
entries (bytes)
1 Bucket array n_buckets 32 Buckets of the hash table.
(configurable)
2 Bucket n_buckets_ext 32 This array is only created for
extensions (configurable) extendable bucket tables.
array
3 Key array n_keys key_size Keys added to the hash table.
(configurable)
4 Data array n_keys entry_size Key values (key data) associated
(configurable) with the hash table keys.

24.4. Table Library Design 153

 Programmer’s Guide, Release 2.0.0

Table 24.9: Field Description for Bucket Array Entry (Configurable Key Size Hash Tables)
# Field Field Description
name size
(bytes)
1 Next 8 For LRU tables, this fields represents the LRU list for the current
Ptr/LRU bucket stored as array of 4 entries of 2 bytes each. Entry 0 stores
the index (0 .. 3) of the MRU key, while entry 3 stores the index of
the LRU key.
For extendable bucket tables, this field represents the next pointer
(i.e. the pointer to the next group of 4 keys linked to the current
bucket). The next pointer is not NULL if the bucket is currently
extended or NULL otherwise. To help the branchless
implementation, bit 0 (least significant bit) of this field is set to 1 if
the next pointer is not NULL and to 0 otherwise.
2 Sig[0 4x2 If key X (X = 0 .. 3) is valid, then sig X bits 15 .. 1 store the most
.. 3] significant 15 bits of key X signature and sig X bit 0 is set to 1.
If key X is not valid, then sig X is set to zero.
3 Key 4x4 If key X is valid (X = 0 .. 3), then Key Pos X represents the index
Pos [0 into the key array where key X is stored, as well as the index into
.. 3] the data array where the value associated with key X is stored.
If key X is not valid, then the value of Key Pos X is undefined.
Fig. 24.4 and Table 27 detail the bucket search pipeline stages (either LRU or extendable
bucket, either with pre-computed signature or “do-sig”). For each pipeline stage, the described
operations are applied to each of the two packets handled by that stage.

Figure 24.4: Bucket Search Pipeline for Key Lookup Operation (Configurable Key Size Hash
Tables)

24.4. Table Library Design 154

 Programmer’s Guide, Release 2.0.0

Table 24.10: Description of the Bucket Search Pipeline Stages (Configurable Key Size Hash
Tables)
# Stage name Description
0 Prefetch Select next two packets from the burst of input packets.
packet Prefetch packet meta-data containing the key and key signature.
meta-data
1 Prefetch Read the key signature from the packet meta-data (for extendable
table bucket bucket hash tables) or read the key from the packet meta-data and
compute key signature (for LRU tables).
Identify the bucket ID using the key signature.
Set bit 0 of the signature to 1 (to match only signatures of valid keys
from the table).
Prefetch the bucket.
2 Prefetch Read the key signatures from the bucket.
table key Compare the signature of the input key against the 4 key signatures
from the packet. As result, the following is obtained:
match = equal to TRUE if there was at least one signature match and
to FALSE in the case of no signature match;
match_many = equal to TRUE is there were more than one signature
matches (can be up to 4 signature matches in the worst case scenario)
and to FALSE otherwise;
match_pos = the index of the first key that produced signature match
(only valid if match is true).
For extendable bucket hash tables only, set match_many to TRUE if
next pointer is valid.
Prefetch the bucket key indicated by match_pos (even if match_pos
does not point to valid key valid).
3 Prefetch Read the bucket key indicated by match_pos.
table data Compare the bucket key against the input key. As result, the following
is obtained: match_key = equal to TRUE if the two keys match and to
FALSE otherwise.
Report input key as lookup hit only when both match and match_key
are equal to TRUE and as lookup miss otherwise.
For LRU tables only, use branchless logic to update the bucket LRU list
(the current key becomes the new MRU) only on lookup hit.
Prefetch the key value (key data) associated with the current key (to
avoid branches, this is done on both lookup hit and miss).
Additional notes:
1. The pipelined version of the bucket search algorithm is executed only if there are at least
7 packets in the burst of input packets. If there are less than 7 packets in the burst of input
packets, a non-optimized implementation of the bucket search algorithm is executed.
2. Once the pipelined version of the bucket search algorithm has been executed for all the
packets in the burst of input packets, the non-optimized implementation of the bucket
search algorithm is also executed for any packets that did not produce a lookup hit, but
have the match_many flag set. As result of executing the non-optimized version, some
of these packets may produce a lookup hit or lookup miss. This does not impact the
performance of the key lookup operation, as the probability of matching more than one
signature in the same group of 4 keys or of having the bucket in extended state (for
extendable bucket hash tables only) is relatively small.

24.4. Table Library Design 155

 Programmer’s Guide, Release 2.0.0

Key Signature Comparison Logic

The key signature comparison logic is described in Table 28.

Table 24.11: Lookup Tables for Match, Match_Many and Match_Pos

# mask match (1 bit) match_many (1 bit) match_pos (2 bits)
0 0000 0 0 00
1 0001 1 0 00
2 0010 1 0 01
3 0011 1 1 00
4 0100 1 0 10
5 0101 1 1 00
6 0110 1 1 01
7 0111 1 1 00
8 1000 1 0 11
9 1001 1 1 00
10 1010 1 1 01
11 1011 1 1 00
12 1100 1 1 10
13 1101 1 1 00
14 1110 1 1 01
15 1111 1 1 00
The input mask hash bit X (X = 0 .. 3) set to 1 if input signature is equal to bucket signature X
and set to 0 otherwise. The outputs match, match_many and match_pos are 1 bit, 1 bit and 2
bits in size respectively and their meaning has been explained above.
As displayed in Table 29, the lookup tables for match and match_many can be collapsed into
a single 32-bit value and the lookup table for match_pos can be collapsed into a 64-bit value.
Given the input mask, the values for match, match_many and match_pos can be obtained by
indexing their respective bit array to extract 1 bit, 1 bit and 2 bits respectively with branchless
logic.

Table 24.12: Collapsed Lookup Tables for Match, Match_Many and Match_Pos
Bit array Hexadecimal value
match 1111_1111_1111_1110 0xFFFELLU
match_many 1111_1110_1110_1000 0xFEE8LLU
match_pos 0001_0010_0001_0011__0001_0010_0001_0000 0x12131210LLU
The pseudo-code for match, match_many and match_pos is:
match = (0xFFFELLU >> mask) & 1;

match_many = (0xFEE8LLU >> mask) & 1;

match_pos = (0x12131210LLU >> (mask << 1)) & 3;

Single Key Size Hash Tables

Fig. 24.5, Fig. 24.6, Table 30 and 31 detail the main data structures used to implement 8-
byte and 16-byte key hash tables (either LRU or extendable bucket, either with pre-computed
signature or “do-sig”).

24.4. Table Library Design 156

 Programmer’s Guide, Release 2.0.0

Figure 24.5: Data Structures for 8-byte Key Hash Tables

Figure 24.6: Data Structures for 16-byte Key Hash Tables

Table 24.13: Main Large Data Structures (Arrays) used for 8-byte and 16-byte Key Size Hash
Tables
# Array Number of Entry size (bytes) Description
name entries
1 Bucket n_buckets 8-byte key size: Buckets of the hash
array (config- 64 + 4 x entry_size table.
urable) 16-byte key size:
128 + 4 x entry_size
2 Bucket n_buckets_ext 8-byte key size: This array is only
exten- (config- 64 + 4 x entry_size created for extendable
sions urable) 16-byte key size: bucket tables.
array 128 + 4 x entry_size

24.4. Table Library Design 157

 Programmer’s Guide, Release 2.0.0

Table 24.14: Field Description for Bucket Array Entry (8-byte and 16-byte Key Hash Tables)
# Field Field Description
name size
(bytes)
1 Valid 8 Bit X (X = 0 .. 3) is set to 1 if key X is valid or to 0 otherwise.
Bit 4 is only used for extendable bucket tables to help with the
implementation of the branchless logic. In this case, bit 4 is set to 1
if next pointer is valid (not NULL) or to 0 otherwise.
2 Next 8 For LRU tables, this fields represents the LRU list for the current
Ptr/LRU bucket stored as array of 4 entries of 2 bytes each. Entry 0 stores
the index (0 .. 3) of the MRU key, while entry 3 stores the index of
the LRU key.
For extendable bucket tables, this field represents the next pointer
(i.e. the pointer to the next group of 4 keys linked to the current
bucket). The next pointer is not NULL if the bucket is currently
extended or NULL otherwise.
3 Key 4x Full keys.
[0 .. key_size
3]
4 Data 4 x en- Full key values (key data) associated with keys 0 .. 3.
[0 .. try_size
3]
and detail the bucket search pipeline used to implement 8-byte and 16-byte key hash tables
(either LRU or extendable bucket, either with pre-computed signature or “do-sig”). For each
pipeline stage, the described operations are applied to each of the two packets handled by that
stage.

Figure 24.7: Bucket Search Pipeline for Key Lookup Operation (Single Key Size Hash Tables)

24.4. Table Library Design 158

 Programmer’s Guide, Release 2.0.0

Table 24.15: Description of the Bucket Search Pipeline Stages (8-byte and 16-byte Key Hash
Tables)
# Stage name Description
0 Prefetch packet meta-data
1. Select next two packets
from the burst of input
packets.
2. Prefetch packet meta-
data containing the key
and key signature.

1 Prefetch table bucket

1. Read the key signa-
ture from the packet
meta-data (for extend-
able bucket hash tables)
or read the key from the
packet meta-data and
compute key signature
(for LRU tables).
2. Identify the bucket ID
using the key signature.
3. Prefetch the bucket.

2 Prefetch table data

1. Read the bucket.
2. Compare all 4 bucket
keys against the input
key.
3. Report input key as
lookup hit only when
a match is identified
(more than one key
match is not possible)
4. For LRU tables only, use
branchless logic to up-
date the bucket LRU
list (the current key be-
comes the new MRU)
only on lookup hit.
5. Prefetch the key value
(key data) associated
with the matched key (to
avoid branches, this is
done on both lookup hit
and miss).

Additional notes:
1. The pipelined version of the bucket search algorithm is executed only if there are at least
5 packets in the burst of input packets. If there are less than 5 packets in the burst of input

24.4. Table Library Design 159

 Programmer’s Guide, Release 2.0.0

packets, a non-optimized implementation of the bucket search algorithm is executed.

2. For extendable bucket hash tables only, once the pipelined version of the bucket search
algorithm has been executed for all the packets in the burst of input packets, the non-
optimized implementation of the bucket search algorithm is also executed for any packets
that did not produce a lookup hit, but have the bucket in extended state. As result of
executing the non-optimized version, some of these packets may produce a lookup hit or
lookup miss. This does not impact the performance of the key lookup operation, as the
probability of having the bucket in extended state is relatively small.

24.5 Pipeline Library Design

A pipeline is defined by:

1. The set of input ports;
2. The set of output ports;
3. The set of tables;
4. The set of actions.
The input ports are connected with the output ports through tree-like topologies of intercon-
nected tables. The table entries contain the actions defining the operations to be executed on
the input packets and the packet flow within the pipeline.

24.5.1 Connectivity of Ports and Tables

To avoid any dependencies on the order in which pipeline elements are created, the connec-
tivity of pipeline elements is defined after all the pipeline input ports, output ports and tables
have been created.
General connectivity rules:
1. Each input port is connected to a single table. No input port should be left unconnected;
2. The table connectivity to other tables or to output ports is regulated by the next hop
actions of each table entry and the default table entry. The table connectivity is fluid, as
the table entries and the default table entry can be updated during run-time.
• A table can have multiple entries (including the default entry) connected to the same
output port. A table can have different entries connected to different output ports.
Different tables can have entries (including default table entry) connected to the
same output port.
• A table can have multiple entries (including the default entry) connected to another
table, in which case all these entries have to point to the same table. This constraint
is enforced by the API and prevents tree-like topologies from being created (allow-
ing table chaining only), with the purpose of simplifying the implementation of the
pipeline run-time execution engine.

24.5. Pipeline Library Design 160

 Programmer’s Guide, Release 2.0.0

24.5.2 Port Actions

Port Action Handler

An action handler can be assigned to each input/output port to define actions to be executed
on each input packet that is received by the port. Defining the action handler for a specific
input/output port is optional (i.e. the action handler can be disabled).
For input ports, the action handler is executed after RX function. For output ports, the action
handler is executed before the TX function.
The action handler can decide to drop packets.

24.5.3 Table Actions

Table Action Handler

An action handler to be executed on each input packet can be assigned to each table. Defining
the action handler for a specific table is optional (i.e. the action handler can be disabled).
The action handler is executed after the table lookup operation is performed and the table
entry associated with each input packet is identified. The action handler can only handle the
user-defined actions, while the reserved actions (e.g. the next hop actions) are handled by the
Packet Framework. The action handler can decide to drop the input packet.

Reserved Actions

The reserved actions are handled directly by the Packet Framework without the user being able
to change their meaning through the table action handler configuration. A special category of
the reserved actions is represented by the next hop actions, which regulate the packet flow
between input ports, tables and output ports through the pipeline. Table 33 lists the next hop
actions.

Table 24.16: Next Hop Actions (Reserved)

# Next hop Description
action
1 Drop Drop the current packet.
2 Send to Send the current packet to specified output port. The output port ID is
output port metadata stored in the same table entry.
3 Send to Send the current packet to specified table. The table ID is metadata
table stored in the same table entry.

User Actions

For each table, the meaning of user actions is defined through the configuration of the table
action handler. Different tables can be configured with different action handlers, therefore the
meaning of the user actions and their associated meta-data is private to each table. Within
the same table, all the table entries (including the table default entry) share the same definition
for the user actions and their associated meta-data, with each table entry having its own set

24.5. Pipeline Library Design 161

 Programmer’s Guide, Release 2.0.0

of enabled user actions and its own copy of the action meta-data. Table 34 contains a non-
exhaustive list of user action examples.

Table 24.17: User Action Examples

# User action Description
1 Metering Per flow traffic metering using the srTCM and trTCM algorithms.
2 Statistics Update the statistics counters maintained per flow.
3 App ID Per flow state machine fed by variable length sequence of packets
at the flow initialization with the purpose of identifying the traffic
type and application.
4 Push/pop labels Push/pop VLAN/MPLS labels to/from the current packet.
5 Network Translate between the internal (LAN) and external (WAN) IP
Address destination/source address and/or L4 protocol destination/source
Translation port.
(NAT)
6 TTL update Decrement IP TTL and, in case of IPv4 packets, update the IP
checksum.

24.6 Multicore Scaling

A complex application is typically split across multiple cores, with cores communicating through
SW queues. There is usually a performance limit on the number of table lookups and actions
that can be fitted on the same CPU core due to HW constraints like: available CPU cycles,
cache memory size, cache transfer BW, memory transfer BW, etc.
As the application is split across multiple CPU cores, the Packet Framework facilitates the
creation of several pipelines, the assignment of each such pipeline to a different CPU core
and the interconnection of all CPU core-level pipelines into a single application-level complex
pipeline. For example, if CPU core A is assigned to run pipeline P1 and CPU core B pipeline
P2, then the interconnection of P1 with P2 could be achieved by having the same set of SW
queues act like output ports for P1 and input ports for P2.
This approach enables the application development using the pipeline, run-to-completion (clus-
tered) or hybrid (mixed) models.
It is allowed for the same core to run several pipelines, but it is not allowed for several cores to
run the same pipeline.

24.6.1 Shared Data Structures

The threads performing table lookup are actually table writers rather than just readers. Even if
the specific table lookup algorithm is thread-safe for multiple readers (e. g. read-only access
of the search algorithm data structures is enough to conduct the lookup operation), once the
table entry for the current packet is identified, the thread is typically expected to update the
action meta-data stored in the table entry (e.g. increment the counter tracking the number of
packets that hit this table entry), and thus modify the table entry. During the time this thread
is accessing this table entry (either writing or reading; duration is application specific), for data
consistency reasons, no other threads (threads performing table lookup or entry add/delete
operations) are allowed to modify this table entry.
Mechanisms to share the same table between multiple threads:

24.6. Multicore Scaling 162

 Programmer’s Guide, Release 2.0.0

1. Multiple writer threads. Threads need to use synchronization primitives like

semaphores (distinct semaphore per table entry) or atomic instructions. The cost of
semaphores is usually high, even when the semaphore is free. The cost of atomic in-
structions is normally higher than the cost of regular instructions.
2. Multiple writer threads, with single thread performing table lookup operations and
multiple threads performing table entry add/delete operations. The threads perform-
ing table entry add/delete operations send table update requests to the reader (typically
through message passing queues), which does the actual table updates and then sends
the response back to the request initiator.
3. Single writer thread performing table entry add/delete operations and multiple
reader threads that perform table lookup operations with read-only access to the
table entries. The reader threads use the main table copy while the writer is updating
the mirror copy. Once the writer update is done, the writer can signal to the readers and
busy wait until all readers swaps between the mirror copy (which now becomes the main
copy) and the mirror copy (which now becomes the main copy).

24.7 Interfacing with Accelerators

The presence of accelerators is usually detected during the initialization phase by inspecting
the HW devices that are part of the system (e.g. by PCI bus enumeration). Typical devices
with acceleration capabilities are:
• Inline accelerators: NICs, switches, FPGAs, etc;
• Look-aside accelerators: chipsets, FPGAs, etc.
Usually, to support a specific functional block, specific implementation of Packet Framework
tables and/or ports and/or actions has to be provided for each accelerator, with all the imple-
mentations sharing the same API: pure SW implementation (no acceleration), implementation
using accelerator A, implementation using accelerator B, etc. The selection between these
implementations could be done at build time or at run-time (recommended), based on which
accelerators are present in the system, with no application changes required.

24.7. Interfacing with Accelerators 163

 CHAPTER

TWENTYFIVE

VHOST LIBRARY

The vhost library implements a user space vhost driver. It supports both vhost-cuse (cuse: user
space character device) and vhost-user(user space socket server). It also creates, manages
and destroys vhost devices for corresponding virtio devices in the guest. Vhost supported
vSwitch could register callbacks to this library, which will be called when a vhost device is
activated or deactivated by guest virtual machine.

25.1 Vhost API Overview

• Vhost driver registration

rte_vhost_driver_register registers the vhost driver into the system. For vhost-
cuse, character device file will be created under the /dev directory. Character
device name is specified as the parameter. For vhost-user, a Unix domain
socket server will be created with the parameter as the local socket path.
• Vhost session start
rte_vhost_driver_session_start starts the vhost session loop. Vhost session is
an infinite blocking loop. Put the session in a dedicate DPDK thread.
• Callback register
Vhost supported vSwitch could call rte_vhost_driver_callback_register to reg-
ister two callbacks, new_destory and destroy_device. When virtio device is
activated or deactivated by guest virtual machine, the callback will be called,
then vSwitch could put the device onto data core or remove the device from
data core by setting or unsetting VIRTIO_DEV_RUNNING on the device flags.
• Read/write packets from/to guest virtual machine
rte_vhost_enqueue_burst transmit host packets to guest.
rte_vhost_dequeue_burst receives packets from guest.
• Feature enable/disable
Now one negotiate-able feature in vhost is merge-able. vSwitch could en-
able/disable this feature for performance consideration.

164
 Programmer’s Guide, Release 2.0.0

25.2 Vhost Implementation

25.2.1 Vhost cuse implementation

When vSwitch registers the vhost driver, it will register a cuse device driver into the system
and creates a character device file. This cuse driver will receive vhost open/release/IOCTL
message from QEMU simulator.
When the open call is received, vhost driver will create a vhost device for the virtio device in
the guest.
When VHOST_SET_MEM_TABLE IOCTL is received, vhost searches the memory region to
find the starting user space virtual address that maps the memory of guest virtual machine.
Through this virtual address and the QEMU pid, vhost could find the file QEMU uses to map
the guest memory. Vhost maps this file into its address space, in this way vhost could fully
access the guest physical memory, which means vhost could access the shared virtio ring and
the guest physical address specified in the entry of the ring.
The guest virtual machine tells the vhost whether the virtio device is ready for processing or is
de-activated through VHOST_NET_SET_BACKEND message. The registered callback from
vSwitch will be called.
When the release call is released, vhost will destroy the device.

25.2.2 Vhost user implementation

When vSwitch registers a vhost driver, it will create a Unix domain socket server into the
system. This server will listen for a connection and process the vhost message from QEMU
simulator.
When there is a new socket connection, it means a new virtio device has been created in the
guest virtual machine, and the vhost driver will create a vhost device for this virtio device.
For messages with a file descriptor, the file descriptor could be directly used in the vhost
process as it is already installed by Unix domain socket.
• VHOST_SET_MEM_TABLE
• VHOST_SET_VRING_KICK
• VHOST_SET_VRING_CALL
• VHOST_SET_LOG_FD
• VHOST_SET_VRING_ERR
For VHOST_SET_MEM_TABLE message, QEMU will send us information for each memory
region and its file descriptor in the ancillary data of the message. The fd is used to map that
region.
There is no VHOST_NET_SET_BACKEND message as in vhost cuse to signal us whether
virtio device is ready or should be stopped. VHOST_SET_VRING_KICK is used as the signal
to put the vhost device onto data plane. VHOST_GET_VRING_BASE is used as the signal to
remove vhost device from data plane.
When the socket connection is closed, vhost will destroy the device.

25.2. Vhost Implementation 165

 Programmer’s Guide, Release 2.0.0

25.3 Vhost supported vSwitch reference

For more vhost details and how to support vhost in vSwitch, please refer to vhost example in
the DPDK Sample Applications Guide.

25.3. Vhost supported vSwitch reference 166

 CHAPTER

TWENTYSIX

PORT HOTPLUG FRAMEWORK

The Port Hotplug Framework provides DPDK applications with the ability to attach and detach
ports at runtime. Because the framework depends on PMD implementation, the ports that
PMDs cannot handle are out of scope of this framework. Furthermore, after detaching a port
from a DPDK application, the framework doesn’t provide a way for removing the devices from
the system. For the ports backed by a physical NIC, the kernel will need to support PCI Hotplug
feature.

26.1 Overview

The basic requirements of the Port Hotplug Framework are:

• DPDK applications that use the Port Hotplug Framework must manage their own ports.
The Port Hotplug Framework is implemented to allow DPDK applications to manage
ports. For example, when DPDK applications call the port attach function, the attached
port number is returned. DPDK applications can also detach the port by port number.
• Kernel support is needed for attaching or detaching physical device ports.
To attach new physical device ports, the device will be recognized by userspace driver I/O
framework in kernel at first. Then DPDK applications can call the Port Hotplug functions
to attach the ports. For detaching, steps are vice versa.
• Before detaching, they must be stopped and closed.
DPDK applications must call “rte_eth_dev_stop()” and “rte_eth_dev_close()” APIs before
detaching ports. These functions will start finalization sequence of the PMDs.
• The framework doesn’t affect legacy DPDK applications behavior.
If the Port Hotplug functions aren’t called, all legacy DPDK apps can still work without
modifications.

26.2 Port Hotplug API overview

• Attaching a port
“rte_eth_dev_attach()” API attaches a port to DPDK application, and returns the attached
port number. Before calling the API, the device should be recognized by an userspace
driver I/O framework. The API receives a pci address like “0000:01:00.0” or a virtual

167
 Programmer’s Guide, Release 2.0.0

device name like “eth_pcap0,iface=eth0”. In the case of virtual device name, the format
is the same as the general “–vdev” option of DPDK.
• Detaching a port
“rte_eth_dev_detach()” API detaches a port from DPDK application, and returns a pci
address of the detached device or a virtual device name of the device.

26.3 Reference

“testpmd” supports the Port Hotplug Framework.

26.4 Limitations

• The Port Hotplug APIs are not thread safe.

• The framework can only be enabled with Linux. BSD is not supported.
• To detach a port, the port should be backed by a device that igb_uio manages. VFIO is
not supported.
• Not all PMDs support detaching feature. To know whether a PMD can support detaching,
search for the “RTE_PCI_DRV_DETACHABLE” flag in PMD implementation. If the flag
is defined in the PMD, detaching is supported.

26.3. Reference 168

 CHAPTER

TWENTYSEVEN

SOURCE ORGANIZATION

This section describes the organization of sources in the DPDK framework.

27.1 Makefiles and Config

Note: In the following descriptions, RTE_SDK is the environment variable that points to the
base directory into which the tarball was extracted. See Useful Variables Provided by the Build
System for descriptions of other variables.

Makefiles that are provided by the DPDK libraries and applications are located in
$(RTE_SDK)/mk.
Config templates are located in $(RTE_SDK)/config. The templates describe the options that
are enabled for each target. The config file also contains items that can be enabled and
disabled for many of the DPDK libraries, including debug options. The user should look at the
config file and become familiar with the options. The config file is also used to create a header
file, which will be located in the new build directory.

27.2 Libraries

Libraries are located in subdirectories of $(RTE_SDK)/lib. By convention, we call a library any

code that provides an API to an application. Typically, it generates an archive file (.a), but a
kernel module should also go in the same directory.
The lib directory contains:
lib
+-- librte_cmdline # command line interface helper
+-- librte_distributor # packet distributor
+-- librte_eal # environment abstraction layer
+-- librte_ether # generic interface to poll mode driver
+-- librte_hash # hash library
+-- librte_ip_frag # IP fragmentation library
+-- librte_ivshmem # QEMU IVSHMEM library
+-- librte_kni # kernel NIC interface
+-- librte_kvargs # argument parsing library
+-- librte_lpm # longest prefix match library
+-- librte_malloc # malloc-like functions
+-- librte_mbuf # packet and control mbuf manipulation library
+-- librte_mempool # memory pool manager (fixedsized objects)
+-- librte_meter # QoS metering library

169
 Programmer’s Guide, Release 2.0.0

+-- librte_net # various IP-related headers

+-- librte_pmd_bond # bonding poll mode driver
+-- librte_pmd_e1000 # 1GbE poll mode drivers (igb and em)
+-- librte_pmd_fm10k # Host interface PMD driver for FM10000 Series
+-- librte_pmd_ixgbe # 10GbE poll mode driver
+-- librte_pmd_i40e # 40GbE poll mode driver
+-- librte_pmd_mlx4 # Mellanox ConnectX-3 poll mode driver
+-- librte_pmd_pcap # PCAP poll mode driver
+-- librte_pmd_ring # ring poll mode driver
+-- librte_pmd_virtio # virtio poll mode driver
+-- librte_pmd_vmxnet3 # VMXNET3 poll mode driver
+-- librte_pmd_xenvirt # Xen virtio poll mode driver
+-- librte_power # power management library
+-- librte_ring # software rings (act as lockless FIFOs)
+-- librte_sched # QoS scheduler and dropper library
+-- librte_timer # timer library

27.3 Applications

Applications are sources that contain a main() function. They are located in the
$(RTE_SDK)/app and $(RTE_SDK)/examples directories.
The app directory contains sample applications that are used to test the DPDK (autotests).
The examples directory contains sample applications that show how libraries can be used.
app
+-- chkincs # test prog to check include depends
+-- test # autotests, to validate DPDK features
+-- test-pmd # test and bench poll mode driver examples

examples
+-- cmdline # Example of using cmdline library
+-- dpdk_qat # Example showing integration with Intel QuickAssist
+-- exception_path # Sending packets to and from Linux Ethernet device (TAP)
+-- helloworld # Helloworld basic example
+-- ip_reassembly # Example showing IP Reassembly
+-- ip_fragmentation # Example showing IPv4 Fragmentation
+-- ipv4_multicast # Example showing IPv4 Multicast
+-- kni # Kernel NIC Interface example
+-- l2fwd # L2 Forwarding example with and without SR-IOV
+-- l3fwd # L3 Forwarding example
+-- l3fwd-power # L3 Forwarding example with power management
+-- l3fwd-vf # L3 Forwarding example with SR-IOV
+-- link_status_interrupt # Link status change interrupt example
+-- load_balancer # Load balancing across multiple cores/sockets
+-- multi_process # Example applications with multiple DPDK processes
+-- qos_meter # QoS metering example
+-- qos_sched # QoS scheduler and dropper example
+-- timer # Example of using librte_timer library
+-- vmdq_dcb # Intel 82599 Ethernet Controller VMDQ and DCB receiving
+-- vmdq # Example of VMDQ receiving
+-- vhost # Example of userspace vhost and switch

Note: The actual examples directory may contain additional sample applications to those
shown above. Check the latest DPDK source files for details.

27.3. Applications 170

 CHAPTER

TWENTYEIGHT

DEVELOPMENT KIT BUILD SYSTEM

The DPDK requires a build system for compilation activities and so on. This section describes
the constraints and the mechanisms used in the DPDK framework.
There are two use-cases for the framework:
• Compilation of the DPDK libraries and sample applications; the framework generates
specific binary libraries, include files and sample applications
• Compilation of an external application or library, using an installed binary DPDK

28.1 Building the Development Kit Binary

The following provides details on how to build the DPDK binary.

28.1.1 Build Directory Concept

After installation, a build directory structure is created. Each build directory contains include
files, libraries, and applications:
~/DPDK$ ls
app MAINTAINERS
config Makefile
COPYRIGHT mk
doc scripts
examples lib
tools x86_64-native-linuxapp-gcc
x86_64-native-linuxapp-icc i686-native-linuxapp-gcc
i686-native-linuxapp-icc

...
~/DEV/DPDK$ ls i686-native-linuxapp-gcc

app build hostapp include kmod lib Makefile

~/DEV/DPDK$ ls i686-native-linuxapp-gcc/app/
cmdline_test dump_cfg test testpmd
cmdline_test.map dump_cfg.map test.map
testpmd.map

~/DEV/DPDK$ ls i686-native-linuxapp-gcc/lib/

libethdev.a librte_hash.a librte_mbuf.a librte_pmd_ixgbe.a

171
 Programmer’s Guide, Release 2.0.0

librte_cmdline.a librte_lpm.a librte_mempool.a librte_ring.a

librte_eal.a librte_malloc.a librte_pmd_e1000.a librte_timer.a

~/DEV/DPDK$ ls i686-native-linuxapp-gcc/include/
arch rte_cpuflags.h rte_memcpy.h
cmdline_cirbuf.h rte_cycles.h rte_memory.h
cmdline.h rte_debug.h rte_mempool.h
cmdline_parse_etheraddr.h rte_eal.h rte_memzone.h
cmdline_parse.h rte_errno.h rte_pci_dev_ids.h
cmdline_parse_ipaddr.h rte_ethdev.h rte_pci.h
cmdline_parse_num.h rte_ether.h rte_per_lcore.h
cmdline_parse_portlist.h rte_fbk_hash.h rte_prefetch.h
cmdline_parse_string.h rte_hash_crc.h rte_random.h
cmdline_rdline.h rte_hash.h rte_ring.h
cmdline_socket.h rte_interrupts.h rte_rwlock.h
cmdline_vt100.h rte_ip.h rte_sctp.h
exec-env rte_jhash.h rte_spinlock.h
rte_alarm.h rte_launch.h rte_string_fns.h
rte_atomic.h rte_lcore.h rte_tailq.h
rte_branch_prediction.h rte_log.h rte_tcp.h
rte_byteorder.h rte_lpm.h rte_timer.h
rte_common.h rte_malloc.h rte_udp.h
rte_config.h rte_mbuf.h

A build directory is specific to a configuration that includes architecture + execution environ-

ment + toolchain. It is possible to have several build directories sharing the same sources with
different configurations.
For instance, to create a new build directory called my_sdk_build_dir using the default config-
uration template config/defconfig_x86_64-linuxapp, we use:
cd ${RTE_SDK}
make config T=x86_64-native-linuxapp-gcc O=my_sdk_build_dir

This creates a new my_sdk_build_dir directory. After that, we can compile by doing:
cd my_sdk_build_dir
make

which is equivalent to:

make O=my_sdk_build_dir

The content of the my_sdk_build_dir is then:

-- .config # used configuration

-- Makefile # wrapper that calls head Makefile

# with $PWD as build directory

-- build #All temporary files used during build

+--app # process, including . o, .d, and .cmd files.
| +-- test # For libraries, we have the .a file.
| +-- test.o # For applications, we have the elf file.
| -- ...
+-- lib
+-- librte_eal

28.1. Building the Development Kit Binary 172

 Programmer’s Guide, Release 2.0.0

| -- ...
+-- librte_mempool
| +-- mempool-file1.o
| +-- .mempool-file1.o.cmd
| +-- .mempool-file1.o.d
| +-- mempool-file2.o
| +-- .mempool-file2.o.cmd
| +-- .mempool-file2.o.d
| -- mempool.a
-- ...

-- include # All include files installed by libraries

+-- librte_mempool.h # and applications are located in this
+-- rte_eal.h # directory. The installed files can depend
+-- rte_spinlock.h # on configuration if needed (environment,
+-- rte_atomic.h # architecture, ..)
-- \*.h ...

-- lib # all compiled libraries are copied in this

+-- librte_eal.a # directory
+-- librte_mempool.a
-- \*.a ...

-- app # All compiled applications are installed

+ --test # here. It includes the binary in elf format

Refer to Development Kit Root Makefile Help for details about make commands that can be
used from the root of DPDK.

28.2 Building External Applications

Since DPDK is in essence a development kit, the first objective of end users will be to create
an application using this SDK. To compile an application, the user must set the RTE_SDK and
RTE_TARGET environment variables.
export RTE_SDK=/opt/DPDK
export RTE_TARGET=x86_64-native-linuxapp-gcc
cd /path/to/my_app

For a new application, the user must create their own Makefile that includes some .mk files,
such as ${RTE_SDK}/mk/rte.vars.mk, and ${RTE_SDK}/mk/ rte.app.mk. This is described in
Building Your Own Application.
Depending on the chosen target (architecture, machine, executive environment, toolchain) de-
fined in the Makefile or as an environment variable, the applications and libraries will com-
pile using the appropriate .h files and will link with the appropriate .a files. These files are
located in ${RTE_SDK}/arch-machine-execenv-toolchain, which is referenced internally by
${RTE_BIN_SDK}.
To compile their application, the user just has to call make. The compilation result will be
located in /path/to/my_app/build directory.
Sample applications are provided in the examples directory.

28.2. Building External Applications 173

 Programmer’s Guide, Release 2.0.0

28.3 Makefile Description

28.3.1 General Rules For DPDK Makefiles

In the DPDK, Makefiles always follow the same scheme:

1. Include $(RTE_SDK)/mk/rte.vars.mk at the beginning.
2. Define specific variables for RTE build system.
3. Include a specific $(RTE_SDK)/mk/rte.XYZ.mk, where XYZ can be app, lib, extapp, extlib,
obj, gnuconfigure, and so on, depending on what kind of object you want to build. See
Makefile Types below.
4. Include user-defined rules and variables.
The following is a very simple example of an external application Makefile:
include $(RTE_SDK)/mk/rte.vars.mk

# binary name
APP = helloworld

# all source are stored in SRCS-y

SRCS-y := main.c

CFLAGS += -O3
CFLAGS += $(WERROR_FLAGS)

include $(RTE_SDK)/mk/rte.extapp.mk

28.3.2 Makefile Types

Depending on the .mk file which is included at the end of the user Makefile, the Makefile will
have a different role. Note that it is not possible to build a library and an application in the
same Makefile. For that, the user must create two separate Makefiles, possibly in two different
directories.
In any case, the rte.vars.mk file must be included in the user Makefile as soon as possible.

Application

These Makefiles generate a binary application.

• rte.app.mk: Application in the development kit framework
• rte.extapp.mk: External application
• rte.hostapp.mk: Host application in the development kit framework

Library

Generate a .a library.
• rte.lib.mk: Library in the development kit framework

28.3. Makefile Description 174

 Programmer’s Guide, Release 2.0.0

• rte.extlib.mk: external library

• rte.hostlib.mk: host library in the development kit framework

Install

• rte.install.mk: Does not build anything, it is only used to create links or copy files to the
installation directory. This is useful for including files in the development kit framework.

Kernel Module

• rte.module.mk: Build a kernel module in the development kit framework.

Objects

• rte.obj.mk: Object aggregation (merge several .o in one) in the development kit frame-
work.
• rte.extobj.mk: Object aggregation (merge several .o in one) outside the development kit
framework.

Misc

• rte.doc.mk: Documentation in the development kit framework

• rte.gnuconfigure.mk: Build an application that is configure-based.
• rte.subdir.mk: Build several directories in the development kit framework.

28.3.3 Useful Variables Provided by the Build System

• RTE_SDK: The absolute path to the DPDK sources. When compiling the development
kit, this variable is automatically set by the framework. It has to be defined by the user as
an environment variable if compiling an external application.
• RTE_SRCDIR: The path to the root of the sources. When compiling the development kit,
RTE_SRCDIR = RTE_SDK. When compiling an external application, the variable points
to the root of external application sources.
• RTE_OUTPUT: The path to which output files are written. Typically, it is
$(RTE_SRCDIR)/build, but it can be overridden by the O= option in the make command
line.
• RTE_TARGET: A string identifying the target for which we are building. The format is
arch-machine-execenv-toolchain. When compiling the SDK, the target is deduced by the
build system from the configuration (.config). When building an external application, it
must be specified by the user in the Makefile or as an environment variable.
• RTE_SDK_BIN: References $(RTE_SDK)/$(RTE_TARGET).
• RTE_ARCH: Defines the architecture (i686, x86_64). It is the same value as CON-
FIG_RTE_ARCH but without the double-quotes around the string.

28.3. Makefile Description 175

 Programmer’s Guide, Release 2.0.0

• RTE_MACHINE: Defines the machine. It is the same value as CONFIG_RTE_MACHINE

but without the double-quotes around the string.
• RTE_TOOLCHAIN: Defines the toolchain (gcc , icc). It is the same value as CON-
FIG_RTE_TOOLCHAIN but without the double-quotes around the string.
• RTE_EXEC_ENV: Defines the executive environment (linuxapp). It is the same value as
CONFIG_RTE_EXEC_ENV but without the double-quotes around the string.
• RTE_KERNELDIR: This variable contains the absolute path to the kernel sources that
will be used to compile the kernel modules. The kernel headers must be the same as the
ones that will be used on the target machine (the machine that will run the application).
By default, the variable is set to /lib/modules/$(shell uname -r)/build, which is correct
when the target machine is also the build machine.

28.3.4 Variables that Can be Set/Overridden in a Makefile Only

• VPATH: The path list that the build system will search for sources. By default,
RTE_SRCDIR will be included in VPATH.
• CFLAGS: Flags to use for C compilation. The user should use += to append data in this
variable.
• LDFLAGS: Flags to use for linking. The user should use += to append data in this vari-
able.
• ASFLAGS: Flags to use for assembly. The user should use += to append data in this
variable.
• CPPFLAGS: Flags to use to give flags to C preprocessor (only useful when assembling
.S files). The user should use += to append data in this variable.
• LDLIBS: In an application, the list of libraries to link with (for example, -L /path/to/libfoo
-lfoo ). The user should use += to append data in this variable.
• SRC-y: A list of source files (.c, .S, or .o if the source is a binary) in case of application,
library or object Makefiles. The sources must be available from VPATH.
• INSTALL-y-$(INSTPATH): A list of files to be installed in $(INSTPATH). The files must be
available from VPATH and will be copied in $(RTE_OUTPUT)/$(INSTPATH). Can be used
in almost any RTE Makefile.
• SYMLINK-y-$(INSTPATH): A list of files to be installed in $(INSTPATH). The files must be
available from VPATH and will be linked (symbolically) in $(RTE_OUTPUT)/$(INSTPATH).
This variable can be used in almost any DPDK Makefile.
• PREBUILD: A list of prerequisite actions to be taken before building. The user should
use += to append data in this variable.
• POSTBUILD: A list of actions to be taken after the main build. The user should use += to
append data in this variable.
• PREINSTALL: A list of prerequisite actions to be taken before installing. The user should
use += to append data in this variable.
• POSTINSTALL: A list of actions to be taken after installing. The user should use += to
append data in this variable.

28.3. Makefile Description 176

 Programmer’s Guide, Release 2.0.0

• PRECLEAN: A list of prerequisite actions to be taken before cleaning. The user should
use += to append data in this variable.
• POSTCLEAN: A list of actions to be taken after cleaning. The user should use += to
append data in this variable.
• DEPDIR-y: Only used in the development kit framework to specify if the build of the
current directory depends on build of another one. This is needed to support parallel
builds correctly.

28.3.5 Variables that can be Set/Overridden by the User on the Command Line
Only

Some variables can be used to configure the build system behavior. They are documented in
Development Kit Root Makefile Help and External Application/Library Makefile Help
• WERROR_CFLAGS: By default, this is set to a specific value that depends on the com-
piler. Users are encouraged to use this variable as follows:
CFLAGS += $(WERROR_CFLAGS)
This avoids the use of different cases depending on the compiler (icc or gcc). Also, this variable
can be overridden from the command line, which allows bypassing of the flags for testing
purposes.

28.3.6 Variables that Can be Set/Overridden by the User in a Makefile or Com-

mand Line

• CFLAGS_my_file.o: Specific flags to add for C compilation of my_file.c.

• LDFLAGS_my_app: Specific flags to add when linking my_app.
• NO_AUTOLIBS: If set, the libraries provided by the framework will not be included in the
LDLIBS variable automatically.
• EXTRA_CFLAGS: The content of this variable is appended after CFLAGS when compil-
ing.
• EXTRA_LDFLAGS: The content of this variable is appended after LDFLAGS when link-
ing.
• EXTRA_ASFLAGS: The content of this variable is appended after ASFLAGS when as-
sembling.
• EXTRA_CPPFLAGS: The content of this variable is appended after CPPFLAGS when
using a C preprocessor on assembly files.

28.3. Makefile Description 177

 CHAPTER

TWENTYNINE

DEVELOPMENT KIT ROOT MAKEFILE HELP

The DPDK provides a root level Makefile with targets for configuration, building, cleaning, test-
ing, installation and others. These targets are explained in the following sections.

29.1 Configuration Targets

The configuration target requires the name of the target, which is specified using T=mytarget
and it is mandatory. The list of available targets are in $(RTE_SDK)/config (remove the def-
config _ prefix).
Configuration targets also support the specification of the name of the output directory, using
O=mybuilddir. This is an optional parameter, the default output directory is build.
• Config
This will create a build directory, and generates a configuration from a template. A Make-
file is also created in the new build directory.
Example:
make config O=mybuild T=x86_64-native-linuxapp-gcc

29.2 Build Targets

Build targets support the optional specification of the name of the output directory, using
O=mybuilddir. The default output directory is build.
• all, build or just make
Build the DPDK in the output directory previously created by a make config.
Example:
make O=mybuild

• clean
Clean all objects created using make build.
Example:
make clean O=mybuild

178
 Programmer’s Guide, Release 2.0.0

• %_sub
Build a subdirectory only, without managing dependencies on other directories.
Example:
make lib/librte_eal_sub O=mybuild

• %_clean
Clean a subdirectory only.
Example:
make lib/librte_eal_clean O=mybuild

29.3 Install Targets

• Install
Build the DPDK binary. Actually, this builds each supported target in a separate directory.
The name of each directory is the name of the target. The name of the targets to install
can be optionally specified using T=mytarget. The target name can contain wildcard *
characters. The list of available targets are in $(RTE_SDK)/config (remove the defconfig_
prefix).
Example:
make install T=x86_64-*

• Uninstall
Remove installed target directories.

29.4 Test Targets

• test
Launch automatic tests for a build directory specified using O=mybuilddir. It is optional,
the default output directory is build.
Example:
make test O=mybuild

• testall
Launch automatic tests for all installed target directories (after a make install). The name
of the targets to test can be optionally specified using T=mytarget. The target name can
contain wildcard (*) characters. The list of available targets are in $(RTE_SDK)/config
(remove the defconfig_ prefix).
Examples:
make testall, make testall T=x86_64-*

29.3. Install Targets 179

 Programmer’s Guide, Release 2.0.0

29.5 Documentation Targets

• doc
Generate the Doxygen documentation (API, html and pdf).
• doc-api-html
Generate the Doxygen API documentation in html.
• doc-guides-html
Generate the guides documentation in html.
• doc-guides-pdf
Generate the guides documentation in pdf.

29.6 Deps Targets

• depdirs
This target is implicitly called by make config. Typically, there is no need for a user to call
it, except if DEPDIRS-y variables have been updated in Makefiles. It will generate the file
$(RTE_OUTPUT)/.depdirs.
Example:
make depdirs O=mybuild

• depgraph
This command generates a dot graph of dependencies. It can be displayed to debug
circular dependency issues, or just to understand the dependencies.
Example:
make depgraph O=mybuild > /tmp/graph.dot && dotty /tmp/ graph.dot

29.7 Misc Targets

• help
Show this help.

29.8 Other Useful Command-line Variables

The following variables can be specified on the command line:

• V=
Enable verbose build (show full compilation command line, and some intermediate com-
mands).

29.5. Documentation Targets 180

 Programmer’s Guide, Release 2.0.0

• D=
Enable dependency debugging. This provides some useful information about why a tar-
get is built or not.
• EXTRA_CFLAGS=, EXTRA_LDFLAGS=, EXTRA_ASFLAGS=, EXTRA_CPPFLAGS=
Append specific compilation, link or asm flags.
• CROSS=
Specify a cross toolchain header that will prefix all gcc/binutils applications. This only
works when using gcc.

29.9 Make in a Build Directory

All targets described above are called from the SDK root $(RTE_SDK). It is possible to run the
same Makefile targets inside the build directory. For instance, the following command:
cd $(RTE_SDK)
make config O=mybuild T=x86_64-native-linuxapp-gcc
make O=mybuild

is equivalent to:
cd $(RTE_SDK)
make config O=mybuild T=x86_64-native-linuxapp-gcc
cd mybuild

# no need to specify O= now

make

29.10 Compiling for Debug

To compile the DPDK and sample applications with debugging information included and the
optimization level set to 0, the EXTRA_CFLAGS environment variable should be set before
compiling as follows:
export EXTRA_CFLAGS='-O0 -g'

The DPDK and any user or sample applications can then be compiled in the usual way. For
example:
make install T=x86_64-native-linuxapp-gcc make -C examples/<theapp>

29.9. Make in a Build Directory 181

 CHAPTER

THIRTY

EXTENDING THE DPDK

This chapter describes how a developer can extend the DPDK to provide a new library, a new
target, or support a new target.

30.1 Example: Adding a New Library libfoo

To add a new library to the DPDK, proceed as follows:

1. Add a new configuration option:
for f in config/\*; do \
echo CONFIG_RTE_LIBFOO=y >> $f; done

1. Create a new directory with sources:

mkdir ${RTE_SDK}/lib/libfoo
touch ${RTE_SDK}/lib/libfoo/foo.c
touch ${RTE_SDK}/lib/libfoo/foo.h

1. Add a foo() function in libfoo.

Definition is in foo.c:
void foo(void)
{
}

Declaration is in foo.h:
extern void foo(void);

2. Update lib/Makefile:
vi ${RTE_SDK}/lib/Makefile
# add:
# DIRS-$(CONFIG_RTE_LIBFOO) += libfoo

3. Create a new Makefile for this library, for example, derived from mempool Makefile:
cp ${RTE_SDK}/lib/librte_mempool/Makefile ${RTE_SDK}/lib/libfoo/

vi ${RTE_SDK}/lib/libfoo/Makefile
# replace:
# librte_mempool -> libfoo
# rte_mempool -> foo

182
 Programmer’s Guide, Release 2.0.0

4. Update mk/DPDK.app.mk, and add -lfoo in LDLIBS variable when the option is enabled.
This will automatically add this flag when linking a DPDK application.
5. Build the DPDK with the new library (we only show a specific target here):
cd ${RTE_SDK}
make config T=x86_64-native-linuxapp-gcc
make

6. Check that the library is installed:

ls build/lib
ls build/include

30.1.1 Example: Using libfoo in the Test Application

The test application is used to validate all functionality of the DPDK. Once you have added a
library, a new test case should be added in the test application.
• A new test_foo.c file should be added, that includes foo.h and calls the foo() function from
test_foo(). When the test passes, the test_foo() function should return 0.
• Makefile, test.h and commands.c must be updated also, to handle the new test case.
• Test report generation: autotest.py is a script that is used to generate the test re-
port that is available in the ${RTE_SDK}/doc/rst/test_report/autotests directory. This
script must be updated also. If libfoo is in a new test family, the links in
${RTE_SDK}/doc/rst/test_report/test_report.rst must be updated.
• Build the DPDK with the updated test application (we only show a specific target here):
cd ${RTE_SDK}
make config T=x86_64-native-linuxapp-gcc
make

30.1. Example: Adding a New Library libfoo 183

 CHAPTER

THIRTYONE

BUILDING YOUR OWN APPLICATION

31.1 Compiling a Sample Application in the Development Kit Di-

rectory

When compiling a sample application (for example, hello world), the following variables must
be exported: RTE_SDK and RTE_TARGET.
~/DPDK$ cd examples/helloworld/
~/DPDK/examples/helloworld$ export RTE_SDK=/home/user/DPDK
~/DPDK/examples/helloworld$ export RTE_TARGET=x86_64-native-linuxapp-gcc
~/DPDK/examples/helloworld$ make
CC main.o
LD helloworld
INSTALL-APP helloworld
INSTALL-MAP helloworld.map

The binary is generated in the build directory by default:

~/DPDK/examples/helloworld$ ls build/app
helloworld helloworld.map

31.2 Build Your Own Application Outside the Development Kit

The sample application (Hello World) can be duplicated in a new directory as a starting point
for your development:
~$ cp -r DPDK/examples/helloworld my_rte_app
~$ cd my_rte_app/
~/my_rte_app$ export RTE_SDK=/home/user/DPDK
~/my_rte_app$ export RTE_TARGET=x86_64-native-linuxapp-gcc
~/my_rte_app$ make
CC main.o
LD helloworld
INSTALL-APP helloworld
INSTALL-MAP helloworld.map

184
 Programmer’s Guide, Release 2.0.0

31.3 Customizing Makefiles

31.3.1 Application Makefile

The default makefile provided with the Hello World sample application is a good starting point.
It includes:
• $(RTE_SDK)/mk/rte.vars.mk at the beginning
• $(RTE_SDK)/mk/rte.extapp.mk at the end
The user must define several variables:
• APP: Contains the name of the application.
• SRCS-y: List of source files (*.c, *.S).

31.3.2 Library Makefile

It is also possible to build a library in the same way:

• Include $(RTE_SDK)/mk/rte.vars.mk at the beginning.
• Include $(RTE_SDK)/mk/rte.extlib.mk at the end.
The only difference is that APP should be replaced by LIB, which contains the name of the
library. For example, libfoo.a.

31.3.3 Customize Makefile Actions

Some variables can be defined to customize Makefile actions. The most common are listed
below. Refer to Makefile Description section in Development Kit Build System
chapter for details.
• VPATH: The path list where the build system will search for sources. By default,
RTE_SRCDIR will be included in VPATH.
• CFLAGS_my_file.o: The specific flags to add for C compilation of my_file.c.
• CFLAGS: The flags to use for C compilation.
• LDFLAGS: The flags to use for linking.
• CPPFLAGS: The flags to use to provide flags to the C preprocessor (only useful when
assembling .S files)
• LDLIBS: A list of libraries to link with (for example, -L /path/to/libfoo - lfoo)
• NO_AUTOLIBS: If set, the libraries provided by the framework will not be included in the
LDLIBS variable automatically.

31.3. Customizing Makefiles 185

 CHAPTER

THIRTYTWO

EXTERNAL APPLICATION/LIBRARY MAKEFILE HELP

External applications or libraries should include specific Makefiles from RTE_SDK, located in
mk directory. These Makefiles are:
• ${RTE_SDK}/mk/rte.extapp.mk: Build an application
• ${RTE_SDK}/mk/rte.extlib.mk: Build a static library
• ${RTE_SDK}/mk/rte.extobj.mk: Build objects (.o)

32.1 Prerequisites

The following variables must be defined:

• ${RTE_SDK}: Points to the root directory of the DPDK.
• ${RTE_TARGET}: Reference the target to be used for compilation (for example, x86_64-
native-linuxapp-gcc).

32.2 Build Targets

Build targets support the specification of the name of the output directory, using O=mybuilddir.
This is optional; the default output directory is build.
• all, “nothing” (meaning just make)
Build the application or the library in the specified output directory.
Example:
make O=mybuild

• clean
Clean all objects created using make build.
Example:
make clean O=mybuild

186
 Programmer’s Guide, Release 2.0.0

32.3 Help Targets

• help
Show this help.

32.4 Other Useful Command-line Variables

The following variables can be specified at the command line:

• S=
Specify the directory in which the sources are located. By default, it is the current direc-
tory.
• M=
Specify the Makefile to call once the output directory is created. By default, it uses
$(S)/Makefile.
• V=
Enable verbose build (show full compilation command line and some intermediate com-
mands).
• D=
Enable dependency debugging. This provides some useful information about why a tar-
get must be rebuilt or not.
• EXTRA_CFLAGS=, EXTRA_LDFLAGS=, EXTRA_ASFLAGS=, EXTRA_CPPFLAGS=
Append specific compilation, link or asm flags.
• CROSS=
Specify a cross-toolchain header that will prefix all gcc/binutils applications. This only
works when using gcc.

32.5 Make from Another Directory

It is possible to run the Makefile from another directory, by specifying the output and the source
dir. For example:
export RTE_SDK=/path/to/DPDK
export RTE_TARGET=x86_64-native-linuxapp-icc
make -f /path/to/my_app/Makefile S=/path/to/my_app O=/path/to/build_dir

32.3. Help Targets 187

 CHAPTER

THIRTYTHREE

PERFORMANCE OPTIMIZATION GUIDELINES

33.1 Introduction

The following sections describe optimizations used in the DPDK and optimizations that should
be considered for a new applications.
They also highlight the performance-impacting coding techniques that should, and should not
be, used when developing an application using the DPDK.
And finally, they give an introduction to application profiling using a Performance Analyzer from
Intel to optimize the software.

188
 CHAPTER

THIRTYFOUR

WRITING EFFICIENT CODE

This chapter provides some tips for developing efficient code using the DPDK. For additional
and more general information, please refer to the Intel® 64 and IA-32 Architectures Optimiza-
tion Reference Manual which is a valuable reference to writing efficient code.

34.1 Memory

This section describes some key memory considerations when developing applications in the
DPDK environment.

34.1.1 Memory Copy: Do not Use libc in the Data Plane

Many libc functions are available in the DPDK, via the Linux* application environment. This
can ease the porting of applications and the development of the configuration plane. However,
many of these functions are not designed for performance. Functions such as memcpy() or
strcpy() should not be used in the data plane. To copy small structures, the preference is for
a simpler technique that can be optimized by the compiler. Refer to the VTune™ Performance
Analyzer Essentials publication from Intel Press for recommendations.
For specific functions that are called often, it is also a good idea to provide a self-made opti-
mized function, which should be declared as static inline.
The DPDK API provides an optimized rte_memcpy() function.

34.1.2 Memory Allocation

Other functions of libc, such as malloc(), provide a flexible way to allocate and free memory. In
some cases, using dynamic allocation is necessary, but it is really not advised to use malloc-
like functions in the data plane because managing a fragmented heap can be costly and the
allocator may not be optimized for parallel allocation.
If you really need dynamic allocation in the data plane, it is better to use a memory pool of
fixed-size objects. This API is provided by librte_mempool. This data structure provides several
services that increase performance, such as memory alignment of objects, lockless access to
objects, NUMA awareness, bulk get/put and per-lcore cache. The rte_malloc () function uses
a similar concept to mempools.

189
 Programmer’s Guide, Release 2.0.0

34.1.3 Concurrent Access to the Same Memory Area

Read-Write (RW) access operations by several lcores to the same memory area can generate
a lot of data cache misses, which are very costly. It is often possible to use per-lcore variables,
for example, in the case of statistics. There are at least two solutions for this:
• Use RTE_PER_LCORE variables. Note that in this case, data on lcore X is not available
to lcore Y.
• Use a table of structures (one per lcore). In this case, each structure must be cache-
aligned.
Read-mostly variables can be shared among lcores without performance losses if there are no
RW variables in the same cache line.

34.1.4 NUMA

On a NUMA system, it is preferable to access local memory since remote memory access
is slower. In the DPDK, the memzone, ring, rte_malloc and mempool APIs provide a way to
create a pool on a specific socket.
Sometimes, it can be a good idea to duplicate data to optimize speed. For read-mostly vari-
ables that are often accessed, it should not be a problem to keep them in one socket only,
since data will be present in cache.

34.1.5 Distribution Across Memory Channels

Modern memory controllers have several memory channels that can load or store data in par-
allel. Depending on the memory controller and its configuration, the number of channels and
the way the memory is distributed across the channels varies. Each channel has a bandwidth
limit, meaning that if all memory access operations are done on the first channel only, there is
a potential bottleneck.
By default, the Mempool Library spreads the addresses of objects among memory channels.

34.2 Communication Between lcores

To provide a message-based communication between lcores, it is advised to use the DPDK

ring API, which provides a lockless ring implementation.
The ring supports bulk and burst access, meaning that it is possible to read several elements
from the ring with only one costly atomic operation (see Chapter 5 “Ring Library”). Performance
is greatly improved when using bulk access operations.
The code algorithm that dequeues messages may be something similar to the following:
#define MAX_BULK 32

while (1) {
/* Process as many elements as can be dequeued. */
count = rte_ring_dequeue_burst(ring, obj_table, MAX_BULK);
if (unlikely(count == 0))
continue;

34.2. Communication Between lcores 190

 Programmer’s Guide, Release 2.0.0

my_process_bulk(obj_table, count);
}

34.3 PMD Driver

The DPDK Poll Mode Driver (PMD) is also able to work in bulk/burst mode, allowing the factor-
ization of some code for each call in the send or receive function.
Avoid partial writes. When PCI devices write to system memory through DMA, it costs less if
the write operation is on a full cache line as opposed to part of it. In the PMD code, actions
have been taken to avoid partial writes as much as possible.

34.3.1 Lower Packet Latency

Traditionally, there is a trade-off between throughput and latency. An application can be tuned
to achieve a high throughput, but the end-to-end latency of an average packet will typically
increase as a result. Similarly, the application can be tuned to have, on average, a low end-to-
end latency, at the cost of lower throughput.
In order to achieve higher throughput, the DPDK attempts to aggregate the cost of processing
each packet individually by processing packets in bursts.
Using the testpmd application as an example, the burst size can be set on the command line
to a value of 16 (also the default value). This allows the application to request 16 packets at
a time from the PMD. The testpmd application then immediately attempts to transmit all the
packets that were received, in this case, all 16 packets.
The packets are not transmitted until the tail pointer is updated on the corresponding TX queue
of the network port. This behavior is desirable when tuning for high throughput because the
cost of tail pointer updates to both the RX and TX queues can be spread across 16 packets,
effectively hiding the relatively slow MMIO cost of writing to the PCIe* device. However, this
is not very desirable when tuning for low latency because the first packet that was received
must also wait for another 15 packets to be received. It cannot be transmitted until the other
15 packets have also been processed because the NIC will not know to transmit the packets
until the TX tail pointer has been updated, which is not done until all 16 packets have been
processed for transmission.
To consistently achieve low latency, even under heavy system load, the application developer
should avoid processing packets in bunches. The testpmd application can be configured from
the command line to use a burst value of 1. This will allow a single packet to be processed at
a time, providing lower latency, but with the added cost of lower throughput.

34.4 Locks and Atomic Operations

Atomic operations imply a lock prefix before the instruction, causing the processor’s LOCK#
signal to be asserted during execution of the following instruction. This has a big impact on
performance in a multicore environment.
Performance can be improved by avoiding lock mechanisms in the data plane. It can often be
replaced by other solutions like per-lcore variables. Also, some locking techniques are more

34.3. PMD Driver 191

 Programmer’s Guide, Release 2.0.0

efficient than others. For instance, the Read-Copy-Update (RCU) algorithm can frequently
replace simple rwlocks.

34.5 Coding Considerations

34.5.1 Inline Functions

Small functions can be declared as static inline in the header file. This avoids the cost of a call
instruction (and the associated context saving). However, this technique is not always efficient;
it depends on many factors including the compiler.

34.5.2 Branch Prediction

The Intel® C/C++ Compiler (icc)/gcc built-in helper functions likely() and unlikely() allow the
developer to indicate if a code branch is likely to be taken or not. For instance:
if (likely(x > 1))
do_stuff();

34.6 Setting the Target CPU Type

The DPDK supports CPU microarchitecture-specific optimizations by means of CON-

FIG_RTE_MACHINE option in the DPDK configuration file. The degree of optimization de-
pends on the compiler’s ability to optimize for a specific microarchitecture, therefore it is prefer-
able to use the latest compiler versions whenever possible.
If the compiler version does not support the specific feature set (for example, the Intel® AVX in-
struction set), the build process gracefully degrades to whatever latest feature set is supported
by the compiler.
Since the build and runtime targets may not be the same, the resulting binary also contains
a platform check that runs before the main() function and checks if the current machine is
suitable for running the binary.
Along with compiler optimizations, a set of preprocessor defines are automatically added to the
build process (regardless of the compiler version). These defines correspond to the instruc-
tion sets that the target CPU should be able to support. For example, a binary compiled for
any SSE4.2-capable processor will have RTE_MACHINE_CPUFLAG_SSE4_2 defined, thus
enabling compile-time code path selection for different platforms.

34.5. Coding Considerations 192

 CHAPTER

THIRTYFIVE

PROFILE YOUR APPLICATION

Intel processors provide performance counters to monitor events. Some tools provided by Intel
can be used to profile and benchmark an application. See the VTune Performance Analyzer
Essentials publication from Intel Press for more information.
For a DPDK application, this can be done in a Linux* application environment only.
The main situations that should be monitored through event counters are:
• Cache misses
• Branch mis-predicts
• DTLB misses
• Long latency instructions and exceptions
Refer to the Intel Performance Analysis Guide for details about application profiling.

193
 CHAPTER

THIRTYSIX

GLOSSARY

ACL Access Control List

API Application Programming Interface
ASLR Linux* kernel Address-Space Layout Randomization
BSD Berkeley Software Distribution
Clr Clear
CIDR Classless Inter-Domain Routing
Control Plane The control plane is concerned with the routing of packets and with providing
a start or end point.
Core A core may include several lcores or threads if the processor supports hyperthreading.
Core Components A set of libraries provided by the DPDK, including eal, ring, mempool,
mbuf, timers, and so on.
CPU Central Processing Unit
CRC Cyclic Redundancy Check
ctrlmbuf An mbuf carrying control data.
Data Plane In contrast to the control plane, the data plane in a network architecture are the lay-
ers involved when forwarding packets. These layers must be highly optimized to achieve
good performance.
DIMM Dual In-line Memory Module
Doxygen A documentation generator used in the DPDK to generate the API reference.
DPDK Data Plane Development Kit
DRAM Dynamic Random Access Memory
EAL The Environment Abstraction Layer (EAL) provides a generic interface that hides the
environment specifics from the applications and libraries. The services expected from the
EAL are: development kit loading and launching, core affinity/ assignment procedures,
system memory allocation/description, PCI bus access, inter-partition communication.
FIFO First In First Out
FPGA Field Programmable Gate Array
GbE Gigabit Ethernet
HW Hardware

194
 Programmer’s Guide, Release 2.0.0

HPET High Precision Event Timer; a hardware timer that provides a precise time reference on
x86 platforms.
ID Identifier
IOCTL Input/Output Control
I/O Input/Output
IP Internet Protocol
IPv4 Internet Protocol version 4
IPv6 Internet Protocol version 6
lcore A logical execution unit of the processor, sometimes called a hardware thread.
KNI Kernel Network Interface
L1 Layer 1
L2 Layer 2
L3 Layer 3
L4 Layer 4
LAN Local Area Network
LPM Longest Prefix Match
master lcore The execution unit that executes the main() function and that launches other
lcores.
mbuf An mbuf is a data structure used internally to carry messages (mainly network packets).
The name is derived from BSD stacks. To understand the concepts of packet buffers or
mbuf, refer to TCP/IP Illustrated, Volume 2: The Implementation.
MESI Modified Exclusive Shared Invalid (CPU cache coherency protocol)
MTU Maximum Transfer Unit
NIC Network Interface Card
OOO Out Of Order (execution of instructions within the CPU pipeline)
NUMA Non-uniform Memory Access
PCI Peripheral Connect Interface
PHY An abbreviation for the physical layer of the OSI model.
pktmbuf An mbuf carrying a network packet.
PMD Poll Mode Driver
QoS Quality of Service
RCU Read-Copy-Update algorithm, an alternative to simple rwlocks.
Rd Read
RED Random Early Detection
RSS Receive Side Scaling

195
 Programmer’s Guide, Release 2.0.0

RTE Run Time Environment. Provides a fast and simple framework for fast packet processing,
in a lightweight environment as a Linux* application and using Poll Mode Drivers (PMDs)
to increase speed.
Rx Reception
Slave lcore Any lcore that is not the master lcore.
Socket A physical CPU, that includes several cores.
SLA Service Level Agreement
srTCM Single Rate Three Color Marking
SRTD Scheduler Round Trip Delay
SW Software
Target In the DPDK, the target is a combination of architecture, machine, executive environ-
ment and toolchain. For example: i686-native-linuxapp-gcc.
TCP Transmission Control Protocol
TC Traffic Class
TLB Translation Lookaside Buffer
TLS Thread Local Storage
trTCM Two Rate Three Color Marking
TSC Time Stamp Counter
Tx Transmission
TUN/TAP TUN and TAP are virtual network kernel devices.
VLAN Virtual Local Area Network
Wr Write
WRED Weighted Random Early Detection
WRR Weighted Round Robin
Figures
Fig. 2.1 Core Components Architecture
Fig. 3.1 EAL Initialization in a Linux Application Environment
Fig. 4.1 Example of a malloc heap and malloc elements within the malloc library
Fig. 5.1 Ring Structure
Fig. 5.2 Enqueue first step
Fig. 5.3 Enqueue second step
Fig. 5.4 Enqueue last step
Fig. 5.5 Dequeue last step
Fig. 5.6 Dequeue second step
Fig. 5.7 Dequeue last step

196
 Programmer’s Guide, Release 2.0.0

Fig. 5.8 Multiple consumer enqueue first step

Fig. 5.9 Multiple consumer enqueue second step
Fig. 5.10 Multiple consumer enqueue third step
Fig. 5.11 Multiple consumer enqueue fourth step
Fig. 5.12 Multiple consumer enqueue last step
Fig. 5.13 Modulo 32-bit indexes - Example 1
Fig. 5.14 Modulo 32-bit indexes - Example 2
Fig. 6.1 Two Channels and Quad-ranked DIMM Example
Fig. 6.2 Three Channels and Two Dual-ranked DIMM Example
Fig. 6.3 A mempool in Memory with its Associated Ring
Fig. 7.1 An mbuf with One Segment
Fig. 7.2 An mbuf with Three Segments
Fig. 18.1 Memory Sharing in the DPDK Multi-process Sample Application
Fig. 19.1 Components of a DPDK KNI Application
Fig. 19.2 Packet Flow via mbufs in the DPDK KNI
Fig. 19.3 vHost-net Architecture Overview
Fig. 19.4 KNI Traffic Flow
Fig. 21.1 Complex Packet Processing Pipeline with QoS Support
Fig. 21.2 Hierarchical Scheduler Block Internal Diagram
Fig. 21.3 Scheduling Hierarchy per Port
Fig. 21.4 Internal Data Structures per Port
Fig. 21.5 Prefetch Pipeline for the Hierarchical Scheduler Enqueue Operation
Fig. 21.6 Pipe Prefetch State Machine for the Hierarchical Scheduler Dequeue Operation
Fig. 21.7 High-level Block Diagram of the DPDK Dropper
Fig. 21.8 Flow Through the Dropper
Fig. 21.9 Example Data Flow Through Dropper
Fig. 21.10 Packet Drop Probability for a Given RED Configuration
Fig. 21.11 Initial Drop Probability (pb), Actual Drop probability (pa) Computed Using a Factor
1 (Blue Curve) and a Factor 2 (Red Curve)
Fig. 24.1 Example of Packet Processing Pipeline where Input Ports 0 and 1 are Connected
with Output Ports 0, 1 and 2 through Tables 0 and 1
Fig. 24.2 Sequence of Steps for Hash Table Operations in a Packet Processing Context
Fig. 24.3 Data Structures for Configurable Key Size Hash Tables
Fig. 24.4 Bucket Search Pipeline for Key Lookup Operation (Configurable Key Size Hash
Tables)

197
 Programmer’s Guide, Release 2.0.0

Fig. 24.5 Data Structures for 8-byte Key Hash Tables

Fig. 24.6 Data Structures for 16-byte Key Hash Tables
Fig. 24.7 Bucket Search Pipeline for Key Lookup Operation (Single Key Size Hash Tables)
Tables
Table 21.1 Packet Processing Pipeline Implementing QoS
Table 21.2 Infrastructure Blocks Used by the Packet Processing Pipeline
Table 21.3 Port Scheduling Hierarchy
Table 21.4 Scheduler Internal Data Structures per Port
Table 21.5 Ethernet Frame Overhead Fields
Table 21.6 Token Bucket Generic Operations
Table 21.7 Token Bucket Generic Parameters
Table 21.8 Token Bucket Persistent Data Structure
Table 21.9 Token Bucket Operations
Table 21.10 Subport/Pipe Traffic Class Upper Limit Enforcement Persistent Data Structure
Table 21.11 Subport/Pipe Traffic Class Upper Limit Enforcement Operations
Table 21.12 Weighted Round Robin (WRR)
Table 21.13 Subport Traffic Class Oversubscription
Table 21.14 Watermark Propagation from Subport Level to Member Pipes at the Beginning of
Each Traffic Class Upper Limit Enforcement Period
Table 21.15 Watermark Calculation
Table 21.16 RED Configuration Parameters
Table 21.17 Relative Performance of Alternative Approaches
Table 21.18 RED Configuration Corresponding to RED Configuration File
Table 24.1 Port Types
Table 24.2 20 Port Abstract Interface
Table 24.3 Table Types
Table 24.5 Configuration Parameters Common for All Hash Table Types
Table 24.6 Configuration Parameters Specific to Extendable Bucket Hash Table
Table 24.7 Configuration Parameters Specific to Pre-computed Key Signature Hash Table
Table 24.8 Main Large Data Structures (Arrays) used for Configurable Key Size Hash Tables
Table 24.9 Field Description for Bucket Array Entry (Configurable Key Size Hash Tables)
Table 24.10 Description of the Bucket Search Pipeline Stages (Configurable Key Size Hash
Tables)
Table 24.11 Lookup Tables for Match, Match_Many and Match_Pos
Table 24.12 Collapsed Lookup Tables for Match, Match_Many and Match_Pos

198
 Programmer’s Guide, Release 2.0.0

Table 24.13 Main Large Data Structures (Arrays) used for 8-byte and 16-byte Key Size Hash
Tables
Table 24.14 Field Description for Bucket Array Entry (8-byte and 16-byte Key Hash Tables)
Table 24.15 Description of the Bucket Search Pipeline Stages (8-byte and 16-byte Key Hash
Tables)
Table 24.16 Next Hop Actions (Reserved)
Table 24.17 User Action Examples

199