Updated 2017-03-21 ~

What is BPF?

BPF, as in Berkeley Packet Filter, was initially conceived in 1992 so as to provide a way to filter packets and to avoid useless packet copies from kernel to userspace. It initially consisted in a simple bytecode that is injected from userspace into the kernel, where it is checked by a verifier—to prevent kernel crashes or security issues—and attached to a socket, then run on each received packet. It was ported to Linux a couple of years later, and used for a small number of applications (tcpdump for example). The simplicity of the language as well as the existence of an in-kernel Just-In-Time (JIT) compiling machine for BPF were factors for the excellent performances of this tool.

Then in 2013, Alexei Starovoitov completely reshaped it, started to add new functionalities and to improve the performances of BPF. This new version is designated as eBPF (for “extended BPF”), while the former becomes cBPF (“classic” BPF). New features such as maps and tail calls appeared. The JIT machines were rewritten. The new language is even closer to native machine language than cBPF was. And also, new attach points in the kernel have been created.

Thanks to those new hooks, eBPF programs can be designed for a variety of use cases, that divide into two fields of applications. One of them is the domain of kernel tracing and event monitoring. BPF programs can be attached to kprobes and they compare with other tracing methods, with many advantages (and sometimes some drawbacks).

The other application domain remains network programming. In addition to socket filter, eBPF programs can be attached to tc (Linux traffic control tool) ingress or egress interfaces and perform a variety of packet processing tasks, in an efficient way. This opens new perspectives in the domain.

And eBPF performances are further leveraged through the technologies developed for the IO Visor project: new hooks have also been added for XDP (“eXpress Data Path”), a new fast path recently added to the kernel. XDP works in conjunction with the Linux stack, and relies on BPF to perform very fast packet processing.

Even some projects such as P4, Open vSwitch, consider or started to approach BPF. Some others, such as CETH, Cilium, are entirely based on it. BPF is buzzing, so we can expect a lot of tools and projects to orbit around it soon…

Dive into the bytecode

As for me: some of my work (including for BEBA) is closely related to eBPF, and several future articles on this site will focus on this topic. Logically, I wanted to somehow introduce BPF on this blog before going down to the details—I mean, a real introduction, more developed on BPF functionalities that the brief abstract provided in first section: What are BPF maps? Tail calls? What do the internals look like? And so on. But there are a lot of presentations on this topic available on the web already, and I do not wish to create “yet another BPF introduction” that would come as a duplicate of existing documents.

So instead, here is what we will do. After all, I spent some time reading and learning about BPF, and while doing so, I gathered a fair amount of material about BPF: introductions, documentation, but also tutorials or examples. There is a lot to read, but in order to read it, one has to find it first. Therefore, as an attempt to help people who wish to learn and use BPF, the present article introduces a list of resources. These are various kinds of readings, that hopefully will help you dive into the mechanics of this kernel bytecode.

Resources

Generic presentations

The documents linked below provide a generic overview of BPF, or of some closely related topics. If you are very new to BPF, you can try picking a couple of presentation among the first ones and reading the ones you like most. If you know eBPF already, you probably want to target specific topics instead, lower down in the list.

About BPF

Generic presentations about eBPF:

BPF internals:

The IO Visor blog has some interesting technical articles about BPF. Some of them contain a bit of marketing talks.

Kernel tracing: summing up all existing methods, including BPF:

Regarding event tracing and monitoring, Brendan Gregg uses eBPF a lot and does an excellent job at documenting some of his use cases. If you are in kernel tracing, you should see his blog articles related to eBPF or to flame graphs. Most of it are accessible from this article or by browsing his blog.

Introducing BPF, but also presenting generic concepts of Linux networking:

Hardware offload:

  • eBPF with tc or XDP supports hardware offload, starting with Linux kernel version 4.9 and introduced by Netronome. Here is a presentation about this feature:
    eBPF/XDP hardware offload to SmartNICs (Jakub Kicinski and Nic Viljoen, netdev 1.2, Tokyo, October 2016)

About cBPF:

About XDP

  • XDP overview on the IO Visor website.

  • eXpress Data Path (XDP) (Tom Herbert, Alexei Starovoitov, March 2016):
    The first presentation about XDP.

  • BoF - What Can BPF Do For You? (Brenden Blanco, LinuxCon, Toronto, August 2016).

  • eXpress Data Path (Brenden Blanco, Linux Meetup at Santa Clara, July 2016):
    Contains some (somewhat marketing?) benchmark results! With a single core:
    • ip routing drop: ~3.6 million packets per second (Mpps)
    • tc (with clsact qdisc) drop using BPF: ~4.2 Mpps
    • XDP drop using BPF: 20 Mpps (<10 % CPU utilization)
    • XDP forward (on port on which the packet was received) with rewrite: 10 Mpps

    (Tests performed with the mlx4 driver).

  • Jesper Dangaard Brouer has several excellent sets of slides, that are essential to fully understand the internals of XDP.

    (Jesper also created and tries to extend some documentation about eBPF and XDP, see related section.)

  • XDP workshop — Introduction, experience, and future development (Tom Herbert, netdev 1.2, Tokyo, October 2016) — as of this writing, only the video is available, I don’t know if the slides will be added.
  • P4 on the Edge (John Fastabend, May 2016):
    Presents the use of P4, a description language for packet processing, with BPF to create high-performance programmable switches.

  • If you like audio presentations, there is an associated OvS Orbit episode (#11), called P4 on the Edge, dating from August 2016. OvS Orbit are interviews realized by Ben Pfaff, who is one of the core maintainers of Open vSwitch. In this case, John Fastabend is interviewed.

  • P4, EBPF and Linux TC Offload (Dinan Gunawardena and Jakub Kicinski, August 2016):
    Another presentation on P4, with some elements related to eBPF hardware offload on Netronome’s NFP (Network Flow Processor) architecture.

  • Cilium is a technology initiated by Cisco and relying on BPF and XDP to provide “fast in-kernel networking and security policy enforcement for containers based on eBPF programs generated on the fly”. The code of this project is available on GitHub. Thomas Graf has been performing a number of presentations of this topic:

    A good deal of contents is repeated between the different presentations; if in doubt, just pick the most recent one. Daniel Borkmann has also written a generic introduction to Cilium as a guest author on Google Open Source blog.

  • There are also podcasts about Cilium: an OvS Orbit episode (#4), in which Ben Pfaff interviews Thomas Graf (May 2016), and another podcast by Ivan Pepelnjak, still with Thomas Graf about eBPF, P4, XDP and Cilium (October 2016).

  • Open vSwitch (OvS), and its related project Open Virtual Network (OVN, an open source network virtualization solution) are considering to use eBPF at various level, with several proof-of-concept prototypes already implemented:

    These use cases for eBPF seem to be only at the stage of proposals (nothing merge to OvS main branch) as far as I know, but it will be very interesting to see what comes out of it.

  • CETH for XDP (Yan Chan and Yunsong Lu, Linux Meetup, Santa Clara, July 2016):
    CETH stands for Common Ethernet Driver Framework for faster network I/O, a technology initiated by Mellanox.

  • Suricata, an open source intrusion detection system, seems to rely on eBPF components for its “capture bypass” features:
    The adventures of a Suricate in eBPF land (Éric Leblond, netdev 1.2, Tokyo, October 2016)

  • InKeV: In-Kernel Distributed Network Virtualization for DCN (Z. Ahmed, M. H. Alizai and A. A. Syed, SIGCOMM, August 2016):
    InKeV is an eBPF-based datapath architecture for virtual networks, targeting data center networks. It was initiated by PLUMgrid, and claims to achieve better performances than OvS-based OpenStack solutions.

  • gobpf - utilizing eBPF from Go (Michael Schubert, fosdem17, Brussels, Belgium, February 2017):
    A “library to create, load and use eBPF programs from Go”

  • ply is a small but flexible open source dynamic tracer for Linux, with some features similar to the bcc tools, but with a simpler language inspired by awk and dtrace, written by Tobias Waldekranz.

  • If you read my previous article, you might be interested in this talk I gave about implementing the OpenState interface with eBPF, for stateful packet processing, at fosdem17.

Documentation

Once you managed to get a broad idea of what BPF is, you can put aside generic presentations and start diving into the documentation. Below are the most complete documents about BPF specifications and functioning. Pick the one you need and read them carefully!

About BPF

  • The specification of BPF (both classic and extended versions) can be found within the documentation of the Linux kernel, and in particular in filelinux/Documentation/networking/filter.txt. The use of BPF as well as its internals are documented there. Also, this is where you can find information about errors thrown by the verifier when loading BPF code fails. Can be helpful to troubleshoot obscure error messages.

  • … But the kernel documentation is dense and not especially easy to read. If you look for a simple description of eBPF language, head for its summarized description on the IO Visor GitHub repository instead.

  • By the way, the IO Visor project gathered a lot of resources about BPF. Mostly, it is split betweenthe documentation directory of its bcc repository, and the whole content of the bpf-docs repository, both on GitHub. Note the existence of this excellent BPF reference guide containing a detailed description of BPF C and bcc Python helpers.

  • To hack with BPF, there are some essential Linux manual pages. The first one is the bpf(2) man page about the bpf() system call, which is used to manage BPF programs and maps from userspace. It also contains a description of BPF advanced features (program types, maps and so on). The second one is mostly addressed to people wanting to attach BPF programs to tc interface: it is the tc-bpf(8) man page, which is a reference for using BPF with tc, and includes some example commands and samples of code.

  • Jesper Dangaard Brouer initiated an attempt to update eBPF documentation, including the different kinds of maps. He has a draft to which contributions are welcome. Once ready, this document should be merged into the man pages and into kernel documentation.

  • list of BPF features per kernel version is available in bcc repository. Useful is you want to know the minimal kernel version that is required to run a given feature. I contributed and added the links to the commits that introduced each feature, so you can also easily access the commit logs from there.

About tc

When using BPF for networking purposes in conjunction with tc, the Linux tool for traffic control, one may wish to gather information about tc’s generic functioning. Here are a couple of resources about it.

About XDP

About P4 and BPF

P4 is a language used to specify the behavior of a switch. It can be compiled for a number of hardware or software targets. As you may have guessed, one of these targets is BPF… The support is only partial: some P4 features cannot be translated towards BPF, and in a similar way there are things that BPF can do but that would not be possible to express with P4. Anyway, the documentation related to P4 use with BPF used to be hidden in bcc repository. This changed with P4_16 version, the p4c reference compiler including a backend for eBPF.

Tutorials

Brendan Gregg has produced excellent tutorials intended for people who want to use bcc tools for tracing and monitoring events in the kernel. The first tutorial about using bcc itself comes with eleven steps (as of today) to understand how to use the existing tools, while the one intended for Python developers focuses on developing new tools, across seventeen “lessons”.

Sasha Goldshtein also has some Linux Tracing Workshops Materials involving the use of several BPF tools for tracing.

Sadly, as of this writing, there are no tutorials yet on the networking part.

Examples

It is always nice to have examples. To see how things really work. But BPF program samples are scattered across several projects, so I listed all the ones I know of. The examples do not always use the same helpers (for instance, tc and bcc both have their own set of helpers to make it easier to write BPF programs in C language).

From the kernel

The kernel contains examples for most types of program: filters to bind to sockets or to tc interfaces, event tracing/monitoring, and even XDP. You can find these examples under the linux/samples/bpf/directory.

Also do not forget to have a look to the logs related to the (git) commits that introduced a particular feature, they may contain some detailed example of the feature.

From package iproute2

The iproute2 package provide several examples as well. They are obviously oriented towards network programming, since the programs are to be attached to tc ingress or egress interfaces. The examples dwell under the iproute2/examples/bpf/ directory.

From bcc set of tools

Many examples are provided with bcc:

  • Some are networking example programs, under the associated directory. They include socket filters, tc filters, and a XDP program.

  • The tracing directory include a lot of example tracing programs. The tutorials mentioned earlier are based on these. These programs cover a wide range of event monitoring functions, and some of them are production-oriented. Note that on certain Linux distributions (at least for Debian, Ubuntu, Fedora, Arch Linux), these programs have been packaged and can be “easily” installed by typing e.g. # apt install bcc-tools, but as of this writing (and except for Arch Linux), this first requires to set up IO Visor’s own package repository.

  • There are also some examples using Lua as a different BPF back-end (that is, BPF programs are written with Lua instead of a subset of C, allowing to use the same language for front-end and back-end), in the third directory.

Manual pages

While bcc is generally the easiest way to inject and run a BPF program in the kernel, attaching programs to tc interfaces can also be performed by the tc tool itself. So if you intend to use BPF with tc, you can find some example invocations in the tc-bpf(8) manual page.

The code

Sometimes, BPF documentation or examples are not enough, and you may have no other solution that to display the code in your favorite text editor (which should be Vim of course) and to read it. Or you may want to hack into the code so as to patch or add features to the machine. So here are a few pointers to the relevant files, finding the functions you want is up to you!

BPF code in the kernel

  • The file linux/include/linux/bpf.h and its counterpart linux/include/uapi/bpf.h contain definitions related to eBPF, to be used respectively in the kernel and to interface with userspace programs.

  • On the same pattern, files linux/include/linux/filter.h and linux/include/uapi/filter.h contain information used to run the BPF programs.

  • The main pieces of code related to BPF are under linux/kernel/bpf/ directory. The different operations permitted by the system call, such as program loading or map management, are implemented in file syscall.c, while core.c contains the interpreter. The other files have self-explanatory names: verifier.c contains the verifier (no kidding), arraymap.c the code used to interact with maps of type array, and so on.

  • The helpers, as well as several functions related to networking (with tc, XDP…) and available to the user, are implemented in linux/net/core/filter.c. It also contains the code to migrate cBPF bytecode to eBPF (since all cBPF programs are now translated to eBPF in the kernel before being run).

  • The JIT compilers are under the directory of their respective architectures, such as filelinux/arch/x86/net/bpf_jit_comp.c for x86.

  • You will find the code related to the BPF components of tc in the linux/net/sched/ directory, and in particular in files act_bpf.c (action) and cls_bpf.c (filter).

  • I have not hacked with event tracing in BPF, so I do not really know about the hooks for such programs. There is some stuff in linux/kernel/trace/bpf_trace.c. If you are interested in this and want to know more, you may dig on the side of Brendan Gregg’s presentations or blog posts.

  • Nor have I used seccomp-BPF. But the code is in linux/kernel/seccomp.c, and some example use cases can be found in linux/tools/testing/selftests/seccomp/seccomp_bpf.c.

XDP hooks code

Once loaded into the in-kernel BPF virtual machine, XDP programs are hooked from userspace into the kernel network path thanks to a Netlink command. On reception, the function dev_change_xdp_fd() in file linux/net/core/dev.c is called and sets a XDP hook. Such hooks are located in the drivers of supported NICs. For example, the mlx4 driver used for some Mellanox hardware has hooks implemented in files under the drivers/net/ethernet/mellanox/mlx4/ directory. File en_netdev.c receives Netlink commands and calls mlx4_xdp_set(), which in turns calls for instance mlx4_en_process_rx_cq() (for the RX side) implemented in file en_rx.c.

BPF logic in bcc

One can find the code for the bcc set of tools on the bcc GitHub repository. The Python code, including the BPF class, is initiated in file bcc/src/python/bcc/__init__.py. But most of the interesting stuff—to my opinion—such as loading the BPF program into the kernel, happens in the libbcc C library.

Code to manage BPF with tc

The code related to BPF in tc comes with the iproute2 package, of course. Some of it is under theiproute2/tc/ directory. The files f_bpf.c and m_bpf.c (and e_bpf.c) are used respectively to handle BPF filters and actions (and tc exec command, whatever this may be). File q_clsact.c defines the clsact qdisc especially created for BPF. But most of the BPF userspace logic is implemented iniproute2/lib/bpf.c library, so this is probably where you should head to if you want to mess up with BPF and tc (it was moved from file iproute2/tc/tc_bpf.c, where you may find the same code in older versions of the package).

BPF utilities

The kernel also ships the sources of three tools (bpf_asm.cbpf_dbg.cbpf_jit_disasm.c) related to BPF, under the linux/tools/net/ directory:

  • bpf_asm is a minimal cBPF assembler.
  • bpf_dbg is a small debugger for cBPF programs.
  • bpf_jit_disasm is generic for both BPF flavors and could be highly useful for JIT debugging.

Read the comments at the top of the source files to get an overview of their usage.

Other interesting chunks

If you are interested the use of less common languages with BPF, bcc contains P4 compiler for BPF targets as well as Lua front-end that can be used as alternatives to the C subset and (in the case of Lua) to the Python tools.

LLVM backend

The BPF backend used by clang / LLVM for compiling C into eBPF was added to the LLVM sources inthis commit (and can also be accessed on the GitHub mirror).

Running in userspace

As far as I know there are at least two eBPF userspace implementations. The first one, uBPF, is written in C. It contains an interpreter, a JIT compiler for x86_64 architecture, an assembler and a disassembler.

The second one is my own work: rbpf, based on uBPF, but written in Rust. The interpreter and JIT-compiler work, there may be more in the future.

Commit logs

As stated earlier, do not hesitate to have a look at the commit log that introduced a particular BPF feature if you want to have more information about it. You can search the logs in many places, such as on git.kernel.orgon GitHub, or on your local repository if you have cloned it. If you are not familiar with git, try things like git blame <file> to see what commit introduced a particular line of code, then git show <commit> to have details (or search by keyword in git log results, but this may be tedious). See also the list of eBPF features per kernel version on bcc repository, that links to relevant commits.

Troubleshooting

The enthusiasm about eBPF is quite recent, and so far I have not found a lot of resources intending to help with troubleshooting. So here are the few I have, augmented with my own recollection of pitfalls encountered while working with BPF.

Errors at compilation time

  • Make sure you have a recent enough version of the Linux kernel (see also this document).

  • If you compiled the kernel yourself: make sure you installed correctly all components, including kernel image, headers and libc.

  • When using the bcc shell function provided by tc-bpf man page (to compile C code into BPF): I once had to add includes to the header for the clang call:

    __bcc() {
            clang -O2 -I "/usr/src/linux-headers-$(uname -r)/include/" \
                      -I "/usr/src/linux-headers-$(uname -r)/arch/x86/include/" \
                    -emit-llvm -c $1 -o - | \
            llc -march=bpf -filetype=obj -o "`basename $1 .c`.o"
    }
    

    (seems fixed as of today).

  • For other problems with bcc, do not forget to have a look at the FAQ of the tool set.

  • If you downloaded the examples from the iproute2 package in a version that does not exactly match your kernel, some errors can be triggered by the headers included in the files. The example snippets indeed assume that the same version of iproute2 package and kernel headers are installed on the system. If this is not the case, download the correct version of iproute2, or edit the path of included files in the examples to point to the headers included in iproute2 (some problems may or may not occur at runtime, depending on the features in use).

Errors at load and run time

  • To load a program with tc, make sure you use a tc binary coming from an iproute2 version equivalent to the kernel in use.

  • To load a program with bcc, make sure you have bcc installed on the system (just downloading the sources to run the Python script is not enough).

  • With tc, if the BPF program does not return the expected values, check that you called it in the correct fashion: filter, or action, or filter with “direct-action” mode.

  • With tc still, note that actions cannot be attached directly to qdiscs or interfaces without the use of a filter.

  • The errors thrown by the in-kernel verifier may be hard to interpret. The kernel documentation may help, so may the reference guide or, as a last resort, the source code (see above) (good luck!). For this kind of errors it is also important to keep in mind that the verifier does not run the program. If you get an error about an invalid memory access or about uninitialized data, it does not mean that these problems actually occurred (or sometimes, that they can possibly occur at all). It means that your program is written in such a way that the verifier estimates that such errors could happen, and therefore it rejects the program.

  • Note that tc tool has a verbose mode, and that it works well with BPF: try appending verboseat the end of your command line.

  • bcc also has verbose options: the BPF class has a debug argument that can take any combination of the three flags DEBUG_LLVM_IRDEBUG_BPF and DEBUG_PREPROCESSOR (see details in the source file). It even embeds some facilities to print output messages for debugging the code.

  • LLVM v4.0+ embeds a disassembler for eBPF programs. So if you compile your program with clang, adding the -g flag for compiling enables you to later dump your program in the rather human-friendly format used by the kernel verifier. To proceed to the dump, use:

    $ llvm-objdump -S -no-show-raw-insn bpf_program.o
    
  • Working with maps? You want to have a look at bpf-map, a very userful tool in Go created for the Cilium project, that can be used to dump the contents of kernel eBPF maps. There also exists a clone in Rust.

  • There is an old bpf tag on StackOverflow, but as of this writing it has been hardly used—ever (and there is nearly nothing related to the new eBPF version). If you are a reader from the Future though, you may want to check whether there has been more activity on this side.

And still more!

  • In case you would like to easily test XDP, there is a Vagrant setup available. You can also test bccin a Docker container.

  • Wondering where the development and activities around BPF occur? Well, the kernel patches always end up on the netdev mailing list (related to the Linux kernel networking stack development): search for “BPF” or “XDP” keywords. Many discussions and debates also occur on the IO Visor mailing list, since BPF is at the heart of the project. If you only want to keep informed from time to time, there is also an @IOVisor Twitter account.

And come back on this blog from time to time to see if they are new articles about BPF!

Special thanks to Daniel Borkmann for the numerous additional documents he pointed to me so that I could complete this collection.