eBPF Summit keynote and lightning talk videos are available. Watch Now!
November 20, 2020

Title iconeBPF Updates #1: eBPF Summit Coverage, libbpf 0.2, BTF Developments, Bulk API for XDP, Local Task Storage for eBPF LSM

Foreword

Curious people, eBPF aficionados, kernel hackers, welcome! The eBPF Updates aim to be a regular publication providing news on the latest resources and developments in the eBPF world. This report is the first of the series, although it is in fact a reviving of the former BPF Updates that once came out on Cilium's blog.

We will not attempt to close the gap with the former series, the eBPF community has been way too active. We will instead focus on the latest news over October, November 2020, there is plenty to list. From new tutorials to academic research and kernel discussions, everyone should find a good read or two!

Important News

  • The ebpf.io website, hosting this post, launched this summer. It is a gateway to the eBPF world. Browse and learn about the basics, the use cases, the existing projects, get started or dive deeper into eBPF. If you landed on the current page without browsing the rest of the website, remember to plan a visit!
  • The eBPF Summit, first of its kind, took place at the end of October 2020. The agenda and the links to the videos are available on the page of the Summit. There are also recaps available for day 1 and day 2. The bee mascot, representing all use cases for eBPF, was baptized during the Summit: Please meet eBee!

    eBee, the eBPF mascot

  • Isovalent, the company behind Cilium, home of eBPF co-maintainer Daniel Borkmann, and one of the biggest users of and contributors to eBPF, announced its $29M Series A financing round backed by Andreessen Horowitz, Google, and Cisco. Here are the announcement from the Cilium community, the TechCrunch interview, or the post on a16z's blog.
  • The same day, Cilium itself released version 1.9 of its software. The prominent new features include Maglev load-balancing, deny policies, support for virtual or baremetal machines, OpenShift support, mutual TLS (mTLS) automation for Hubble, a bandwidth manager, and eBPF node-local redirect policies. It also brings a number of datapath optimizations, and many other improvements.

New Resources

Blog Posts, Presentations

  • How eBPF Turns Linux into a Programmable Kernel, from Joab Jackson.
    This is a high-level introduction to eBPF and to the role it plays in the Linux ecosystem, based after talks from KubeCon and DockerCon.
  • BPF, XDP, Packet Filters and UDP, from Thomas Ptacek.
    This post is another introduction to eBPF and XDP, from a very different angle. It provides historical context and “more than you wanted to know” on packet filters lineage, before concluding with some programming advice.
  • Inspecting python tools using eBPF, from Michał Wcisło.
    To trace Python-based programs such as Ansible, this tutorial introduces a number of tools from the bcc project.
  • BPF ring buffer, from Andrii Nakryiko.
    eBPF programs have been able to use the perf buffer to stream data to user space for some time. A new structure, the BPF ring buffer, addresses some shortcomings of the perf buffer (memory overhead, data ordering, wasted work and extra data copying) and provides increased performance. Andrii, who implemented this new buffer, describes its concepts, its advantages and its usage in this article.
  • BPF Tales, or Why Did I Recompile the Kernel to Average Some Numbers?, from Giulia Frascaria.
    eBPF programs often aggregate data in the kernel before sending metrics to user space, to help users trace programs. In this work, eBPF runs instead at the boundary between compute and storage nodes in a datacenter, to apply filter-reduce operations on the data to transfer, and to overcome the gap between storage hardware and network bandwidths. This work was also the object of a lightning talk during the eBPF Summit.
  • Full body HTTP request/responses tracing using eBPF, from Zain Asgar.
    This is the second part of a series of posts explaining how to trace applications with eBPF, and this one focuses on extracting the body of HTTP traffic with programs attached to kprobes and uprobes.
  • Fuzzing for eBPF JIT bugs in the Linux kernel, from Simon Scannell.
    Fuzzing tools are very helpful to discover bugs in software, and the eBPF subsystem is no exception. This write-up details the architecture and the generator used to build such a fuzzer, and explains how it let to the discovery of a kernel vulnerability. The kernel has been patched since then.
  • BPF binaries: BTF, CO-RE, and the future of BPF perf tools, from Brendan Gregg.
    Without entering too much into the technical details, Brendan exposes the principles of BTF (BPF Type Format) and CO-RE (Compile-Once, Run-Everywhere), which are two eBPF features enabling a far greater portability for eBPF programs between different kernels. Beyond the increased stability for tracing programs, these features are becoming essential in a growing number of use cases.
  • What is eBPF and How Does it Work?, from Virag Mody.
    Part technical introduction, part tutorial, this article explains how eBPF works and how to leverage some of the bcc tools to monitor SSH sessions.
  • eBPF: Put the Kubernetes Data Plane in the Kernel, from Joab Jackson.
    Built on talks from the eBPF Summit, this post presents the motivations for implementing the Kubernetes dataplane with eBPF. Performance and scalability, network policies and observability are the most prominent gains.
  • eBPF - The Future of Networking & Security, from Thomas Graf.
    A summary of the recent evolution of networking: SDN, containers and Kubernetes, and the rise of eBPF. Cilium is emphasized as the ideal solution to bring eBPF-based networking, observability, and security to the cloud-native environments.

Academic Research

Software Projects

  • bpfbox is a policy enforcement engine confining process access to sensitive system resources. It leverages the ability of the recent kernels to attach eBPF programs to LSM (Linux Security Module) hooks.
  • ebpfbench is a Go framework to profile eBPF programs.
  • ebpfpub is a generic tracing library for kernel and user space functions, relying on eBPF.
  • Swoll relies on eBPF to monitor low-level kernel activity like system calls, in the context of a Kubernetes cluster.

Sandbox and Experiments

  • Liz Rice's Beginner's Guide to eBPF is a simple “Hello World” eBPF program demonstrated during conferences and designed to help newcomers familiarize themselves with eBPF.
  • greeter-bpf is an implementation of a gRPC GreeterServer in eBPF “just for fun”, but actually for benchmarks, too.
  • rbperf is an experimental eBPF-based Ruby Profiler.

The Kernel Side

Andrii Nakryiko has been nominated as the third eBPF maintainer to help out on the eBPF tooling side. Congratulations!

Here are some highlights from the first pull request for the bpf-next tree for the 5.11 cycle:

  • Extend BTF infrastructure in the kernel to support loading and validation of “split BTF objects”. This is a prerequisite for generating BTF for kernel modules. (Andrii Nakryiko, link)
  • Support pointers beyond packet end (ctx->data_end) to recognize LLVM generated patterns on inlined branch conditions and avoid rejecting some valid programs. (Alexei Starovoitov. link)
  • Implement a new map type BPF_MAP_TYPE_TASK_STORAGE, providing local storage for task_struct for the eBPF LSM. This also results in the addition of three new eBPF helpers: bpf_task_storage_get(), bpf_task_storage_delete(), and bpf_get_current_task_btf(). (KP Singh, link)
  • Enable FENTRY/FEXIT/RAW_TP tracing programs to use the infrastructure (map and helpers) for per-socket local storage. (Martin KaFai Lau, link)
  • Add XDP bulk APIs that introduce a defer/flush mechanism to optimize the XDP_REDIRECT path. (Lorenzo Bianconi, link)
  • Lift the key size limit for hash tables to accept keys larger than MAX_BPF_STACK (512 bytes). This improves some use cases where hash table keys are values from other per-CPU maps, for example to implement allow/deny lists for file paths. (Florian Lehner, link)
  • Fix cross and out-of-tree builds of bpftool and runqslower. This helps build these tools for different target architectures on the same source tree. (Jean-Philippe Brucker, link)

Libbpf

There is an effort in progress to update iproute2 to have it use libbpf instead of its own code to handle eBPF programs. Libbpf has more exhaustive and more up-to-date support for loading and managing eBPF objects (programs, but also BTF descriptions), and tools such as tc would benefit from it. However, there are disagreements on the best way to ship the library, as summarized in a LWN article.

Libbpf v0.2 came out a few weeks ago.

A New Patchwork

In other news, development for the Linux networking (netdev) and eBPF subsystems moved their Patchwork instance (for patch tracking) from Ozlabs to a new “Netdev + BPF” instance managed by the kernel community. The advantages obtained in the process include:

  • Automatic patch delegation into the netdev and bpf buckets, based on the subject lines of the emails.
  • Automatic bot replies to the mailing list when patches have been applied to the source trees.
  • CI integration for netdev and eBPF, with basic sanity checks.
  • Patchwork to link to the eBPF CI pipeline in the future.

Credits

eBPF Updates are brought to you by the Cilium project. This report was produced by Quentin Monnet (Isovalent). Thanks to Cilium engineering team for input and reviews.

If you would like to submit contributions for the next report, please submit them via the #ebpf-news channel on eBPF Slack.