Berkeley Packet Filter

The original paper was written by Steven McCanne and Van Jacobson in 1992 while at Lawrence Berkeley Laboratory. == Raw data-link interface ==

BPF provides pseudo-devices that can be bound to a network interface; reads from the device will read buffers full of packets received on the network interface, and writes to the device will inject packets on the network interface. In 2007, Robert Watson and Christian Peron added zero-copy buffer extensions to the BPF implementation in the FreeBSD operating system, allowing kernel packet capture in the device driver interrupt handler to write directly to user process memory in order to avoid the requirement for two copies for all packet data received via the BPF device. While one copy remains in the receipt path for user processes, this preserves the independence of different BPF device consumers, as well as allowing the packing of headers into the BPF buffer rather than copying complete packet data. == Filtering ==

BPF's filtering capabilities are implemented as an interpreter for a machine language for the BPF virtual machine, a 32-bit machine with fixed-length instructions, one accumulator, and one index register. Programs in that language can fetch data from the packet, perform arithmetic operations on data from the packet, and compare the results against constants or against data in the packet or test bits in the results, accepting or rejecting the packet based on the results of those tests. BPF is often extended by "overloading" the load (ld) and store (str) instructions. Traditional Unix-like BPF implementations can be used in userspace, despite being written for kernel-space. This is accomplished using preprocessor conditions. Since version 3.18, the Linux kernel includes an extended BPF virtual machine with ten 64-bit registers, termed eBPF. It can be used for non-networking purposes, such as for attaching eBPF programs to various tracepoints. Since kernel version 3.19, eBPF filters can be attached to sockets, and, since kernel version 4.1, to traffic control classifiers for the ingress and egress networking data path. The original and obsolete version has been retroactively renamed to classic BPF (cBPF). Nowadays, the Linux kernel runs eBPF only and loaded cBPF bytecode is transparently translated into an eBPF representation in the kernel before program execution. All bytecode is verified before running to prevent denial-of-service attacks. Until Linux 5.3, the verifier prohibited the use of loops, to prevent potentially unbounded execution times; loops with bounded execution time are now permitted in more recent kernels. == Extensions and optimizations ==

Some projects use BPF instruction sets or execution techniques different from the originals. Some platforms, including FreeBSD, NetBSD, and WinPcap, use a just-in-time compiler (JIT) to convert BPF instructions into native code in order to improve performance. Linux includes a BPF JIT compiler which is disabled by default. Kernel-mode interpreters for that same virtual machine language are used in raw data link layer mechanisms in other operating systems, such as Tru64 Unix, and for socket filters in the Linux kernel and in the WinPcap and Npcap packet capture mechanism. == Implementations ==

A user-mode interpreter for BPF is provided with the libpcap/WinPcap/Npcap implementation of the pcap API, so that, when capturing packets on systems without kernel-mode support for that filtering mechanism, packets can be filtered in user mode; code using the pcap API will work on both types of systems, although, on systems where the filtering is done in user mode, all packets, including those that will be filtered out, are copied from the kernel to user space. That interpreter can also be used when reading a file containing packets captured using pcap. Another user-mode interpreter is uBPF, which supports JIT and eBPF (without cBPF). Its code has been reused to provide eBPF support in non-Linux systems. Microsoft's eBPF on Windows builds on uBPF and the PREVAIL formal verifier. rBPF, a Rust rewrite of uBPF, is used by the Solana blockchain platform as the execution engine. == Programming ==

Classic BPF is generally emitted by a program from some very high-level textual rule describing the pattern to match. One such representation is found in libpcap. Classic BPF and eBPF can also be written either directly as machine code, or using an assembly language for a textual representation. Notable assemblers include Linux kernel's tool (cBPF), (cBPF), and the assembler (eBPF). The command can also act as a disassembler for both flavors of BPF. The assembly languages are not necessarily compatible with each other. eBPF bytecode has recently become a target of higher-level languages. LLVM added eBPF support in 2014, and GCC followed in 2019. Both toolkits allow compiling C and other supported languages to eBPF. A subset of P4 can also be compiled into eBPF using BCC, an LLVM-based compiler kit. == Security ==

The Spectre attack could leverage the Linux kernel's eBPF interpreter or JIT compiler to extract data from other kernel processes. A JIT hardening feature in the kernel mitigates this vulnerability. Chinese computer security group Pangu Lab said the NSA used BPF to conceal network communications as part of a complex Linux backdoor. == See also ==