# BPF and XDP Reference Guide¶

Note

This documentation section is targeted at developers and users who want to understand BPF and XDP in great technical depth. While reading this reference guide may help broaden your understanding of Cilium, it is not a requirement to use Cilium. Please refer to the Getting Started Guides and Architecture for a higher level introduction.

BPF is a highly flexible and efficient virtual machine-like construct in the Linux kernel allowing to execute bytecode at various hook points in a safe manner. It is used in a number of Linux kernel subsystems, most prominently networking, tracing and security (e.g. sandboxing).

Although BPF exists since 1992, this document covers the extended Berkeley Packet Filter (eBPF) version which has first appeared in Kernel 3.18 and renders the original version which is being referred to as “classic” BPF (cBPF) these days mostly obsolete. cBPF is known to many as being the packet filter language used by tcpdump. Nowadays, the Linux kernel runs eBPF only and loaded cBPF bytecode is transparently translated into an eBPF representation in the kernel before program execution. This documentation will generally refer to the term BPF unless explicit differences between eBPF and cBPF are being pointed out.

Even though the name Berkeley Packet Filter hints at a packet filtering specific purpose, the instruction set is generic and flexible enough these days that there are many use cases for BPF apart from networking. See Further Reading for a list of projects which use BPF.

Cilium uses BPF heavily in its data path, see Architecture for further information. The goal of this chapter is to provide a BPF reference guide in order to gain understanding of BPF, its networking specific use including loading BPF programs with tc (traffic control) and XDP (eXpress Data Path), and to aid with developing Cilium’s BPF templates.

## BPF Architecture¶

BPF does not define itself by only providing its instruction set, but also by offering further infrastructure around it such as maps which act as efficient key / value stores, helper functions to interact with and leverage kernel functionality, tail calls for calling into other BPF programs, security hardening primitives, a pseudo file system for pinning objects (maps, programs), and infrastructure for allowing BPF to be offloaded, for example, to a network card.

LLVM provides a BPF back end, so that tools like clang can be used to compile C into a BPF object file, which can then be loaded into the kernel. BPF is deeply tied to the Linux kernel and allows for full programmability without sacrificing native kernel performance.

Last but not least, also the kernel subsystems making use of BPF are part of BPF’s infrastructure. The two main subsystems discussed throughout this document are tc and XDP where BPF programs can be attached to. XDP BPF programs are attached at the earliest networking driver stage and trigger a run of the BPF program upon packet reception. By definition, this achieves the best possible packet processing performance since packets cannot get processed at an even earlier point in software. However, since this processing occurs so early in the networking stack, the stack has not yet extracted metadata out of the packet. On the other hand, tc BPF programs are executed later in the kernel stack, so they have access to more metadata and core kernel functionality. Apart from tc and XDP programs, there are various other kernel subsystems as well which use BPF such as tracing (kprobes, uprobes, tracepoints, etc).

The following subsections provide further details on individual aspects of the BPF architecture.

### Instruction Set¶

BPF is a general purpose RISC instruction set and was originally designed for the purpose of writing programs in a subset of C which can be compiled into BPF instructions through a compiler back end (e.g. LLVM), so that the kernel can later on map them through an in-kernel JIT compiler into native opcodes for optimal execution performance inside the kernel.

The advantages for pushing these instructions into the kernel include:

• Making the kernel programmable without having to cross kernel / user space boundaries. For example, BPF programs related to networking, as in the case of Cilium, can implement flexible container policies, load balancing and other means without having to move packets to user space and back into the kernel. State between BPF programs and kernel / user space can still be shared through maps whenever needed.
• Given the flexibility of a programmable data path, programs can be heavily optimized for performance also by compiling out features that are not required for the use cases the program solves. For example, if a container does not require IPv4, then the BPF program can be built to only deal with IPv6 in order to save resources in the fast-path.
• In case of networking (e.g. tc and XDP), BPF programs can be updated atomically without having to restart the kernel, system services or containers, and without traffic interruptions. Furthermore, any program state can also be maintained throughout updates via BPF maps.
• BPF provides a stable ABI towards user space, and does not require any third party kernel modules. BPF is a core part of the Linux kernel that is shipped everywhere, and guarantees that existing BPF programs keep running with newer kernel versions. This guarantee is the same guarantee that the kernel provides for system calls with regard to user space applications. Moreover, BPF programs are portable across different architectures.
• BPF programs work in concert with the kernel, they make use of existing kernel infrastructure (e.g. drivers, netdevices, tunnels, protocol stack, sockets) and tooling (e.g. iproute2) as well as the safety guarantees which the kernel provides. Unlike kernel modules, BPF programs are verified through an in-kernel verifier in order to ensure that they cannot crash the kernel, always terminate, etc. XDP programs, for example, reuse the existing in-kernel drivers and operate on the provided DMA buffers containing the packet frames without exposing them or an entire driver to user space as in other models. Moreover, XDP programs reuse the existing stack instead of bypassing it. BPF can be considered a generic “glue code” to kernel facilities for crafting programs to solve specific use cases.

The execution of a BPF program inside the kernel is always event-driven! Examples:

• A networking device which has a BPF program attached on its ingress path will trigger the execution of the program once a packet is received.
• A kernel address which has a kprobe with a BPF program attached will trap once the code at that address gets executed, which will then invoke the kprobe’s callback function for instrumentation, subsequently triggering the execution of the attached BPF program.

BPF consists of eleven 64 bit registers with 32 bit subregisters, a program counter and a 512 byte large BPF stack space. Registers are named r0 - r10. The operating mode is 64 bit by default, the 32 bit subregisters can only be accessed through special ALU (arithmetic logic unit) operations. The 32 bit lower subregisters zero-extend into 64 bit when they are being written to.

Register r10 is the only register which is read-only and contains the frame pointer address in order to access the BPF stack space. The remaining r0 - r9 registers are general purpose and of read/write nature.

A BPF program can call into a predefined helper function, which is defined by the core kernel (never by modules). The BPF calling convention is defined as follows:

• r0 contains the return value of a helper function call.
• r1 - r5 hold arguments from the BPF program to the kernel helper function.
• r6 - r9 are callee saved registers that will be preserved on helper function call.

The BPF calling convention is generic enough to map directly to x86_64, arm64 and other ABIs, thus all BPF registers map one to one to HW CPU registers, so that a JIT only needs to issue a call instruction, but no additional extra moves for placing function arguments. This calling convention was modeled to cover common call situations without having a performance penalty. Calls with 6 or more arguments are currently not supported. The helper functions in the kernel which are dedicated to BPF (BPF_CALL_0() to BPF_CALL_5() functions) are specifically designed with this convention in mind.

Register r0 is also the register containing the exit value for the BPF program. The semantics of the exit value are defined by the type of program. Furthermore, when handing execution back to the kernel, the exit value is passed as a 32 bit value.

Registers r1 - r5 are scratch registers, meaning the BPF program needs to either spill them to the BPF stack or move them to callee saved registers if these arguments are to be reused across multiple helper function calls. Spilling means that the variable in the register is moved to the BPF stack. The reverse operation of moving the variable from the BPF stack to the register is called filling. The reason for spilling/filling is due to the limited number of registers.

Upon entering execution of a BPF program, register r1 initially contains the context for the program. The context is the input argument for the program (similar to argc/argv pair for a typical C program). BPF is restricted to work on a single context. The context is defined by the program type, for example, a networking program can have a kernel representation of the network packet (skb) as the input argument.

The general operation of BPF is 64 bit to follow the natural model of 64 bit architectures in order to perform pointer arithmetics, pass pointers but also pass 64 bit values into helper functions, and to allow for 64 bit atomic operations.

The maximum instruction limit per program is restricted to 4096 BPF instructions, which, by design, means that any program will terminate quickly. For kernel newer than 5.1 this limit was lifted to 1 million BPF instructions. Although the instruction set contains forward as well as backward jumps, the in-kernel BPF verifier will forbid loops so that termination is always guaranteed. Since BPF programs run inside the kernel, the verifier’s job is to make sure that these are safe to run, not affecting the system’s stability. This means that from an instruction set point of view, loops can be implemented, but the verifier will restrict that. However, there is also a concept of tail calls that allows for one BPF program to jump into another one. This, too, comes with an upper nesting limit of 32 calls, and is usually used to decouple parts of the program logic, for example, into stages.

The instruction format is modeled as two operand instructions, which helps mapping BPF instructions to native instructions during JIT phase. The instruction set is of fixed size, meaning every instruction has 64 bit encoding. Currently, 87 instructions have been implemented and the encoding also allows to extend the set with further instructions when needed. The instruction encoding of a single 64 bit instruction on a big-endian machine is defined as a bit sequence from most significant bit (MSB) to least significant bit (LSB) of op:8, dst_reg:4, src_reg:4, off:16, imm:32. off and imm is of signed type. The encodings are part of the kernel headers and defined in linux/bpf.h header, which also includes linux/bpf_common.h.

op defines the actual operation to be performed. Most of the encoding for op has been reused from cBPF. The operation can be based on register or immediate operands. The encoding of op itself provides information on which mode to use (BPF_X for denoting register-based operations, and BPF_K for immediate-based operations respectively). In the latter case, the destination operand is always a register. Both dst_reg and src_reg provide additional information about the register operands to be used (e.g. r0 - r9) for the operation. off is used in some instructions to provide a relative offset, for example, for addressing the stack or other buffers available to BPF (e.g. map values, packet data, etc), or jump targets in jump instructions. imm contains a constant / immediate value.

The available op instructions can be categorized into various instruction classes. These classes are also encoded inside the op field. The op field is divided into (from MSB to LSB) code:4, source:1 and class:3. class is the more generic instruction class, code denotes a specific operational code inside that class, and source tells whether the source operand is a register or an immediate value. Possible instruction classes include:

• BPF_LD, BPF_LDX: Both classes are for load operations. BPF_LD is used for loading a double word as a special instruction spanning two instructions due to the imm:32 split, and for byte / half-word / word loads of packet data. The latter was carried over from cBPF mainly in order to keep cBPF to BPF translations efficient, since they have optimized JIT code. For native BPF these packet load instructions are less relevant nowadays. BPF_LDX class holds instructions for byte / half-word / word / double-word loads out of memory. Memory in this context is generic and could be stack memory, map value data, packet data, etc.
• BPF_ST, BPF_STX: Both classes are for store operations. Similar to BPF_LDX the BPF_STX is the store counterpart and is used to store the data from a register into memory, which, again, can be stack memory, map value, packet data, etc. BPF_STX also holds special instructions for performing word and double-word based atomic add operations, which can be used for counters, for example. The BPF_ST class is similar to BPF_STX by providing instructions for storing data into memory only that the source operand is an immediate value.
• BPF_ALU, BPF_ALU64: Both classes contain ALU operations. Generally, BPF_ALU operations are in 32 bit mode and BPF_ALU64 in 64 bit mode. Both ALU classes have basic operations with source operand which is register-based and an immediate-based counterpart. Supported by both are add (+), sub (-), and (&), or (|), left shift (<<), right shift (>>), xor (^), mul (*), div (/), mod (%), neg (~) operations. Also mov (<X> := <Y>) was added as a special ALU operation for both classes in both operand modes. BPF_ALU64 also contains a signed right shift. BPF_ALU additionally contains endianness conversion instructions for half-word / word / double-word on a given source register.
• BPF_JMP: This class is dedicated to jump operations. Jumps can be unconditional and conditional. Unconditional jumps simply move the program counter forward, so that the next instruction to be executed relative to the current instruction is off + 1, where off is the constant offset encoded in the instruction. Since off is signed, the jump can also be performed backwards as long as it does not create a loop and is within program bounds. Conditional jumps operate on both, register-based and immediate-based source operands. If the condition in the jump operations results in true, then a relative jump to off + 1 is performed, otherwise the next instruction (0 + 1) is performed. This fall-through jump logic differs compared to cBPF and allows for better branch prediction as it fits the CPU branch predictor logic more naturally. Available conditions are jeq (==), jne (!=), jgt (>), jge (>=), jsgt (signed >), jsge (signed >=), jlt (<), jle (<=), jslt (signed <), jsle (signed <=) and jset (jump if DST & SRC). Apart from that, there are three special jump operations within this class: the exit instruction which will leave the BPF program and return the current value in r0 as a return code, the call instruction, which will issue a function call into one of the available BPF helper functions, and a hidden tail call instruction, which will jump into a different BPF program.

The Linux kernel is shipped with a BPF interpreter which executes programs assembled in BPF instructions. Even cBPF programs are translated into eBPF programs transparently in the kernel, except for architectures that still ship with a cBPF JIT and have not yet migrated to an eBPF JIT.

Currently x86_64, arm64, ppc64, s390x, mips64, sparc64 and arm architectures come with an in-kernel eBPF JIT compiler.

All BPF handling such as loading of programs into the kernel or creation of BPF maps is managed through a central bpf() system call. It is also used for managing map entries (lookup / update / delete), and making programs as well as maps persistent in the BPF file system through pinning.

### Helper Functions¶

Helper functions are a concept which enables BPF programs to consult a core kernel defined set of function calls in order to retrieve / push data from / to the kernel. Available helper functions may differ for each BPF program type, for example, BPF programs attached to sockets are only allowed to call into a subset of helpers compared to BPF programs attached to the tc layer. Encapsulation and decapsulation helpers for lightweight tunneling constitute an example of functions which are only available to lower tc layers, whereas event output helpers for pushing notifications to user space are available to tc and XDP programs.

Each helper function is implemented with a commonly shared function signature similar to system calls. The signature is defined as:

u64 fn(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)


The calling convention as described in the previous section applies to all BPF helper functions.

The kernel abstracts helper functions into macros BPF_CALL_0() to BPF_CALL_5() which are similar to those of system calls. The following example is an extract from a helper function which updates map elements by calling into the corresponding map implementation callbacks:

BPF_CALL_4(bpf_map_update_elem, struct bpf_map *, map, void *, key,
void *, value, u64, flags)
{
WARN_ON_ONCE(!rcu_read_lock_held());
return map->ops->map_update_elem(map, key, value, flags);
}

const struct bpf_func_proto bpf_map_update_elem_proto = {
.func           = bpf_map_update_elem,
.gpl_only       = false,
.ret_type       = RET_INTEGER,
.arg1_type      = ARG_CONST_MAP_PTR,
.arg2_type      = ARG_PTR_TO_MAP_KEY,
.arg3_type      = ARG_PTR_TO_MAP_VALUE,
.arg4_type      = ARG_ANYTHING,
};


There are various advantages of this approach: while cBPF overloaded its load instructions in order to fetch data at an impossible packet offset to invoke auxiliary helper functions, each cBPF JIT needed to implement support for such a cBPF extension. In case of eBPF, each newly added helper function will be JIT compiled in a transparent and efficient way, meaning that the JIT compiler only needs to emit a call instruction since the register mapping is made in such a way that BPF register assignments already match the underlying architecture’s calling convention. This allows for easily extending the core kernel with new helper functionality. All BPF helper functions are part of the core kernel and cannot be extended or added through kernel modules.

The aforementioned function signature also allows the verifier to perform type checks. The above struct bpf_func_proto is used to hand all the necessary information which need to be known about the helper to the verifier, so that the verifier can make sure that the expected types from the helper match the current contents of the BPF program’s analyzed registers.

Argument types can range from passing in any kind of value up to restricted contents such as a pointer / size pair for the BPF stack buffer, which the helper should read from or write to. In the latter case, the verifier can also perform additional checks, for example, whether the buffer was previously initialized.

The list of available BPF helper functions is rather long and constantly growing, for example, at the time of this writing, tc BPF programs can choose from 38 different BPF helpers. The kernel’s struct bpf_verifier_ops contains a get_func_proto callback function that provides the mapping of a specific enum bpf_func_id to one of the available helpers for a given BPF program type.

### Maps¶

Maps are efficient key / value stores that reside in kernel space. They can be accessed from a BPF program in order to keep state among multiple BPF program invocations. They can also be accessed through file descriptors from user space and can be arbitrarily shared with other BPF programs or user space applications.

BPF programs which share maps with each other are not required to be of the same program type, for example, tracing programs can share maps with networking programs. A single BPF program can currently access up to 64 different maps directly.

Map implementations are provided by the core kernel. There are generic maps with per-CPU and non-per-CPU flavor that can read / write arbitrary data, but there are also a few non-generic maps that are used along with helper functions.

Generic maps currently available are BPF_MAP_TYPE_HASH, BPF_MAP_TYPE_ARRAY, BPF_MAP_TYPE_PERCPU_HASH, BPF_MAP_TYPE_PERCPU_ARRAY, BPF_MAP_TYPE_LRU_HASH, BPF_MAP_TYPE_LRU_PERCPU_HASH and BPF_MAP_TYPE_LPM_TRIE. They all use the same common set of BPF helper functions in order to perform lookup, update or delete operations while implementing a different backend with differing semantics and performance characteristics.

Non-generic maps that are currently in the kernel are BPF_MAP_TYPE_PROG_ARRAY, BPF_MAP_TYPE_PERF_EVENT_ARRAY, BPF_MAP_TYPE_CGROUP_ARRAY, BPF_MAP_TYPE_STACK_TRACE, BPF_MAP_TYPE_ARRAY_OF_MAPS, BPF_MAP_TYPE_HASH_OF_MAPS. For example, BPF_MAP_TYPE_PROG_ARRAY is an array map which holds other BPF programs, BPF_MAP_TYPE_ARRAY_OF_MAPS and BPF_MAP_TYPE_HASH_OF_MAPS both hold pointers to other maps such that entire BPF maps can be atomically replaced at runtime. These types of maps tackle a specific issue which was unsuitable to be implemented solely through a BPF helper function since additional (non-data) state is required to be held across BPF program invocations.

### Object Pinning¶

BPF maps and programs act as a kernel resource and can only be accessed through file descriptors, backed by anonymous inodes in the kernel. Advantages, but also a number of disadvantages come along with them:

User space applications can make use of most file descriptor related APIs, file descriptor passing for Unix domain sockets work transparently, etc, but at the same time, file descriptors are limited to a processes’ lifetime, which makes options like map sharing rather cumbersome to carry out.

Thus, it brings a number of complications for certain use cases such as iproute2, where tc or XDP sets up and loads the program into the kernel and terminates itself eventually. With that, also access to maps is unavailable from user space side, where it could otherwise be useful, for example, when maps are shared between ingress and egress locations of the data path. Also, third party applications may wish to monitor or update map contents during BPF program runtime.

To overcome this limitation, a minimal kernel space BPF file system has been implemented, where BPF map and programs can be pinned to, a process called object pinning. The BPF system call has therefore been extended with two new commands which can pin (BPF_OBJ_PIN) or retrieve (BPF_OBJ_GET) a previously pinned object.

For instance, tools such as tc make use of this infrastructure for sharing maps on ingress and egress. The BPF related file system is not a singleton, it does support multiple mount instances, hard and soft links, etc.

### Tail Calls¶

Another concept that can be used with BPF is called tail calls. Tail calls can be seen as a mechanism that allows one BPF program to call another, without returning back to the old program. Such a call has minimal overhead as unlike function calls, it is implemented as a long jump, reusing the same stack frame.

Such programs are verified independently of each other, thus for transferring state, either per-CPU maps as scratch buffers or in case of tc programs, skb fields such as the cb[] area must be used.

Only programs of the same type can be tail called, and they also need to match in terms of JIT compilation, thus either JIT compiled or only interpreted programs can be invoked, but not mixed together.

There are two components involved for carrying out tail calls: the first part needs to setup a specialized map called program array (BPF_MAP_TYPE_PROG_ARRAY) that can be populated by user space with key / values, where values are the file descriptors of the tail called BPF programs, the second part is a bpf_tail_call() helper where the context, a reference to the program array and the lookup key is passed to. Then the kernel inlines this helper call directly into a specialized BPF instruction. Such a program array is currently write-only from user space side.

The kernel looks up the related BPF program from the passed file descriptor and atomically replaces program pointers at the given map slot. When no map entry has been found at the provided key, the kernel will just “fall through” and continue execution of the old program with the instructions following after the bpf_tail_call(). Tail calls are a powerful utility, for example, parsing network headers could be structured through tail calls. During runtime, functionality can be added or replaced atomically, and thus altering the BPF program’s execution behavior.

### BPF to BPF Calls¶

Aside from BPF helper calls and BPF tail calls, a more recent feature that has been added to the BPF core infrastructure is BPF to BPF calls. Before this feature was introduced into the kernel, a typical BPF C program had to declare any reusable code that, for example, resides in headers as always_inline such that when LLVM compiles and generates the BPF object file all these functions were inlined and therefore duplicated many times in the resulting object file, artificially inflating its code size:

#include <linux/bpf.h>

#ifndef __section
# define __section(NAME)                  \
__attribute__((section(NAME), used))
#endif

#ifndef __inline
# define __inline                         \
inline __attribute__((always_inline))
#endif

static __inline int foo(void)
{
return XDP_DROP;
}

__section("prog")
int xdp_drop(struct xdp_md *ctx)
{
return foo();
}

char __license[] __section("license") = "GPL";


The main reason why this was necessary was due to lack of function call support in the BPF program loader as well as verifier, interpreter and JITs. Starting with Linux kernel 4.16 and LLVM 6.0 this restriction got lifted and BPF programs no longer need to use always_inline everywhere. Thus, the prior shown BPF example code can then be rewritten more naturally as:

#include <linux/bpf.h>

#ifndef __section
# define __section(NAME)                  \
__attribute__((section(NAME), used))
#endif

static int foo(void)
{
return XDP_DROP;
}

__section("prog")
int xdp_drop(struct xdp_md *ctx)
{
return foo();
}

char __license[] __section("license") = "GPL";


Mainstream BPF JIT compilers like x86_64 and arm64 support BPF to BPF calls today with others following in near future. BPF to BPF call is an important performance optimization since it heavily reduces the generated BPF code size and therefore becomes friendlier to a CPU’s instruction cache.

The calling convention known from BPF helper function applies to BPF to BPF calls just as well, meaning r1 up to r5 are for passing arguments to the callee and the result is returned in r0. r1 to r5 are scratch registers whereas r6 to r9 preserved across calls the usual way. The maximum number of nesting calls respectively allowed call frames is 8. A caller can pass pointers (e.g. to the caller’s stack frame) down to the callee, but never vice versa.

BPF to BPF calls are currently incompatible with the use of BPF tail calls, since the latter requires to reuse the current stack setup as-is, whereas the former adds additional stack frames and thus changes the expected layout for tail calls.

BPF JIT compilers emit separate images for each function body and later fix up the function call addresses in the image in a final JIT pass. This has proven to require minimal changes to the JITs in that they can treat BPF to BPF calls as conventional BPF helper calls.

### JIT¶

The 64 bit x86_64, arm64, ppc64, s390x, mips64, sparc64 and 32 bit arm, x86_32 architectures are all shipped with an in-kernel eBPF JIT compiler, also all of them are feature equivalent and can be enabled through:

# echo 1 > /proc/sys/net/core/bpf_jit_enable


The 32 bit mips, ppc and sparc architectures currently have a cBPF JIT compiler. The mentioned architectures still having a cBPF JIT as well as all remaining architectures supported by the Linux kernel which do not have a BPF JIT compiler at all need to run eBPF programs through the in-kernel interpreter.

In the kernel’s source tree, eBPF JIT support can be easily determined through issuing a grep for HAVE_EBPF_JIT:

# git grep HAVE_EBPF_JIT arch/
arch/arm/Kconfig:       select HAVE_EBPF_JIT   if !CPU_ENDIAN_BE32
arch/arm64/Kconfig:     select HAVE_EBPF_JIT
arch/powerpc/Kconfig:   select HAVE_EBPF_JIT   if PPC64
arch/mips/Kconfig:      select HAVE_EBPF_JIT   if (64BIT && !CPU_MICROMIPS)
arch/s390/Kconfig:      select HAVE_EBPF_JIT   if PACK_STACK && HAVE_MARCH_Z196_FEATURES
arch/sparc/Kconfig:     select HAVE_EBPF_JIT   if SPARC64
arch/x86/Kconfig:       select HAVE_EBPF_JIT   if X86_64


JIT compilers speed up execution of the BPF program significantly since they reduce the per instruction cost compared to the interpreter. Often instructions can be mapped 1:1 with native instructions of the underlying architecture. This also reduces the resulting executable image size and is therefore more instruction cache friendly to the CPU. In particular in case of CISC instruction sets such as x86, the JITs are optimized for emitting the shortest possible opcodes for a given instruction to shrink the total necessary size for the program translation.

### Hardening¶

BPF locks the entire BPF interpreter image (struct bpf_prog) as well as the JIT compiled image (struct bpf_binary_header) in the kernel as read-only during the program’s lifetime in order to prevent the code from potential corruptions. Any corruption happening at that point, for example, due to some kernel bugs will result in a general protection fault and thus crash the kernel instead of allowing the corruption to happen silently.

Architectures that support setting the image memory as read-only can be determined through:

$git grep ARCH_HAS_SET_MEMORY | grep select arch/arm/Kconfig: select ARCH_HAS_SET_MEMORY arch/arm64/Kconfig: select ARCH_HAS_SET_MEMORY arch/s390/Kconfig: select ARCH_HAS_SET_MEMORY arch/x86/Kconfig: select ARCH_HAS_SET_MEMORY  The option CONFIG_ARCH_HAS_SET_MEMORY is not configurable, thanks to which this protection is always built-in. Other architectures might follow in the future. In case of the x86_64 JIT compiler, the JITing of the indirect jump from the use of tail calls is realized through a retpoline in case CONFIG_RETPOLINE has been set which is the default at the time of writing in most modern Linux distributions. In case of /proc/sys/net/core/bpf_jit_harden set to 1 additional hardening steps for the JIT compilation take effect for unprivileged users. This effectively trades off their performance slightly by decreasing a (potential) attack surface in case of untrusted users operating on the system. The decrease in program execution still results in better performance compared to switching to interpreter entirely. Currently, enabling hardening will blind all user provided 32 bit and 64 bit constants from the BPF program when it gets JIT compiled in order to prevent JIT spraying attacks which inject native opcodes as immediate values. This is problematic as these immediate values reside in executable kernel memory, therefore a jump that could be triggered from some kernel bug would jump to the start of the immediate value and then execute these as native instructions. JIT constant blinding prevents this due to randomizing the actual instruction, which means the operation is transformed from an immediate based source operand to a register based one through rewriting the instruction by splitting the actual load of the value into two steps: 1) load of a blinded immediate value rnd ^ imm into a register, 2) xoring that register with rnd such that the original imm immediate then resides in the register and can be used for the actual operation. The example was provided for a load operation, but really all generic operations are blinded. Example of JITing a program with hardening disabled: # echo 0 > /proc/sys/net/core/bpf_jit_harden ffffffffa034f5e9 + <x>: [...] 39: mov$0xa8909090,%eax
3e:   mov    $0xa8909090,%eax 43: mov$0xa8ff3148,%eax
48:   mov    $0xa89081b4,%eax 4d: mov$0xa8900bb0,%eax
52:   mov    $0xa810e0c1,%eax 57: mov$0xa8908eb4,%eax
5c:   mov    $0xa89020b0,%eax [...]  The same program gets constant blinded when loaded through BPF as an unprivileged user in the case hardening is enabled: # echo 1 > /proc/sys/net/core/bpf_jit_harden ffffffffa034f1e5 + <x>: [...] 39: mov$0xe1192563,%r10d
3f:   xor    $0x4989b5f3,%r10d 46: mov %r10d,%eax 49: mov$0xb8296d93,%r10d
4f:   xor    $0x10b9fd03,%r10d 56: mov %r10d,%eax 59: mov$0x8c381146,%r10d
5f:   xor    $0x24c7200e,%r10d 66: mov %r10d,%eax 69: mov$0xeb2a830e,%r10d
6f:   xor    $0x43ba02ba,%r10d 76: mov %r10d,%eax 79: mov$0xd9730af,%r10d
7f:   xor    $0xa5073b1f,%r10d 86: mov %r10d,%eax 89: mov$0x9a45662b,%r10d
8f:   xor    $0x325586ea,%r10d 96: mov %r10d,%eax [...]  Both programs are semantically the same, only that none of the original immediate values are visible anymore in the disassembly of the second program. At the same time, hardening also disables any JIT kallsyms exposure for privileged users, preventing that JIT image addresses are not exposed to /proc/kallsyms anymore. Moreover, the Linux kernel provides the option CONFIG_BPF_JIT_ALWAYS_ON which removes the entire BPF interpreter from the kernel and permanently enables the JIT compiler. This has been developed as part of a mitigation in the context of Spectre v2 such that when used in a VM-based setting, the guest kernel is not going to reuse the host kernel’s BPF interpreter when mounting an attack anymore. For container-based environments, the CONFIG_BPF_JIT_ALWAYS_ON configuration option is optional, but in case JITs are enabled there anyway, the interpreter may as well be compiled out to reduce the kernel’s complexity. Thus, it is also generally recommended for widely used JITs in case of main stream architectures such as x86_64 and arm64. Last but not least, the kernel offers an option to disable the use of the bpf(2) system call for unprivileged users through the /proc/sys/kernel/unprivileged_bpf_disabled sysctl knob. This is on purpose a one-time kill switch, meaning once set to 1, there is no option to reset it back to 0 until a new kernel reboot. When set only CAP_SYS_ADMIN privileged processes out of the initial namespace are allowed to use the bpf(2) system call from that point onwards. Upon start, Cilium sets this knob to 1 as well. # echo 1 > /proc/sys/kernel/unprivileged_bpf_disabled  ### Offloads¶ Networking programs in BPF, in particular for tc and XDP do have an offload-interface to hardware in the kernel in order to execute BPF code directly on the NIC. Currently, the nfp driver from Netronome has support for offloading BPF through a JIT compiler which translates BPF instructions to an instruction set implemented against the NIC. This includes offloading of BPF maps to the NIC as well, thus the offloaded BPF program can perform map lookups, updates and deletions. ## Toolchain¶ Current user space tooling, introspection facilities and kernel control knobs around BPF are discussed in this section. Note, the tooling and infrastructure around BPF is still rapidly evolving and thus may not provide a complete picture of all available tools. ### Development Environment¶ A step by step guide for setting up a development environment for BPF can be found below for both Fedora and Ubuntu. This will guide you through building, installing and testing a development kernel as well as building and installing iproute2. The step of manually building iproute2 and Linux kernel is usually not necessary given that major distributions already ship recent enough kernels by default, but would be needed for testing bleeding edge versions or contributing BPF patches to iproute2 and to the Linux kernel, respectively. Similarly, for debugging and introspection purposes building bpftool is optional, but recommended. #### Fedora¶ The following applies to Fedora 25 or later: $ sudo dnf install -y git gcc ncurses-devel elfutils-libelf-devel bc \
openssl-devel libcap-devel clang llvm graphviz bison flex glibc-static


Note

If you are running some other Fedora derivative and dnf is missing, try using yum instead.

#### Ubuntu¶

The following applies to Ubuntu 17.04 or later:

$sudo apt-get install -y make gcc libssl-dev bc libelf-dev libcap-dev \ clang gcc-multilib llvm libncurses5-dev git pkg-config libmnl-dev bison flex \ graphviz  #### openSUSE Tumbleweed¶ The following applies to openSUSE Tumbleweed and openSUSE Leap 15.0 or later: $ sudo  zypper install -y git gcc ncurses-devel libelf-devel bc libopenssl-devel \
libcap-devel clang llvm graphviz bison flex glibc-devel-static


#### Compiling the Kernel¶

Development of new BPF features for the Linux kernel happens inside the net-next git tree, latest BPF fixes in the net tree. The following command will obtain the kernel source for the net-next tree through git:

$git clone git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-next.git  If the git commit history is not of interest, then --depth 1 will clone the tree much faster by truncating the git history only to the most recent commit. In case the net tree is of interest, it can be cloned from this url: $ git clone git://git.kernel.org/pub/scm/linux/kernel/git/davem/net.git


There are dozens of tutorials in the Internet on how to build Linux kernels, one good resource is the Kernel Newbies website (https://kernelnewbies.org/KernelBuild) that can be followed with one of the two git trees mentioned above.

Make sure that the generated .config file contains the following CONFIG_* entries for running BPF. These entries are also needed for Cilium.

CONFIG_CGROUP_BPF=y
CONFIG_BPF=y
CONFIG_BPF_SYSCALL=y
CONFIG_NET_SCH_INGRESS=m
CONFIG_NET_CLS_BPF=m
CONFIG_NET_CLS_ACT=y
CONFIG_BPF_JIT=y
CONFIG_LWTUNNEL_BPF=y
CONFIG_HAVE_EBPF_JIT=y
CONFIG_BPF_EVENTS=y
CONFIG_TEST_BPF=m


Some of the entries cannot be adjusted through make menuconfig. For example, CONFIG_HAVE_EBPF_JIT is selected automatically if a given architecture does come with an eBPF JIT. In this specific case, CONFIG_HAVE_EBPF_JIT is optional but highly recommended. An architecture not having an eBPF JIT compiler will need to fall back to the in-kernel interpreter with the cost of being less efficient executing BPF instructions.

#### Verifying the Setup¶

After you have booted into the newly compiled kernel, navigate to the BPF selftest suite in order to test BPF functionality (current working directory points to the root of the cloned git tree):

$cd tools/testing/selftests/bpf/$ make
$sudo ./test_verifier  The verifier tests print out all the current checks being performed. The summary at the end of running all tests will dump information of test successes and failures: Summary: 847 PASSED, 0 SKIPPED, 0 FAILED  Note For kernel releases 4.16+ the BPF selftest has a dependency on LLVM 6.0+ caused by the BPF function calls which do not need to be inlined anymore. See section BPF to BPF Calls or the cover letter mail from the kernel patch (https://lwn.net/Articles/741773/) for more information. Not every BPF program has a dependency on LLVM 6.0+ if it does not use this new feature. If your distribution does not provide LLVM 6.0+ you may compile it by following the instruction in the LLVM section. In order to run through all BPF selftests, the following command is needed: $ sudo make run_tests


If you see any failures, please contact us on Slack with the full test output.

#### Compiling iproute2¶

Similar to the net (fixes only) and net-next (new features) kernel trees, the iproute2 git tree has two branches, namely master and net-next. The master branch is based on the net tree and the net-next branch is based against the net-next kernel tree. This is necessary, so that changes in header files can be synchronized in the iproute2 tree.

In order to clone the iproute2 master branch, the following command can be used:

$git clone git://git.kernel.org/pub/scm/linux/kernel/git/iproute2/iproute2.git  Similarly, to clone into mentioned net-next branch of iproute2, run the following: $ git clone -b net-next git://git.kernel.org/pub/scm/linux/kernel/git/iproute2/iproute2.git


After that, proceed with the build and installation:

$cd iproute2/$ ./configure --prefix=/usr
TC schedulers
ATM    no

libc has setns: yes
SELinux support: yes
ELF support: yes
libmnl support: no
Berkeley DB: no

docs: latex: no
WARNING: no docs can be built from LaTeX files
sgml2html: no
WARNING: no HTML docs can be built from SGML
$make [...]$ sudo make install


Ensure that the configure script shows ELF support: yes, so that iproute2 can process ELF files from LLVM’s BPF back end. libelf was listed in the instructions for installing the dependencies in case of Fedora and Ubuntu earlier.

#### Compiling bpftool¶

bpftool is an essential tool around debugging and introspection of BPF programs and maps. It is part of the kernel tree and available under tools/bpf/bpftool/.

Make sure to have cloned either the net or net-next kernel tree as described earlier. In order to build and install bpftool, the following steps are required:

$cd <kernel-tree>/tools/bpf/bpftool/$ make
Auto-detecting system features:
...                        libbfd: [ on  ]
...        disassembler-four-args: [ OFF ]

CC       xlated_dumper.o
CC       prog.o
CC       common.o
CC       cgroup.o
CC       main.o
CC       json_writer.o
CC       cfg.o
CC       map.o
CC       jit_disasm.o
CC       disasm.o
make[1]: Entering directory '/home/foo/trees/net/tools/lib/bpf'

Auto-detecting system features:
...                        libelf: [ on  ]
...                           bpf: [ on  ]

CC       libbpf.o
CC       bpf.o
CC       nlattr.o
LD       libbpf-in.o
LINK     libbpf.a
make[1]: Leaving directory '/home/foo/trees/bpf/tools/lib/bpf'
LINK     bpftool
$sudo make install  ### LLVM¶ LLVM is currently the only compiler suite providing a BPF back end. gcc does not support BPF at this point. The BPF back end was merged into LLVM’s 3.7 release. Major distributions enable the BPF back end by default when they package LLVM, therefore installing clang and llvm is sufficient on most recent distributions to start compiling C into BPF object files. The typical workflow is that BPF programs are written in C, compiled by LLVM into object / ELF files, which are parsed by user space BPF ELF loaders (such as iproute2 or others), and pushed into the kernel through the BPF system call. The kernel verifies the BPF instructions and JITs them, returning a new file descriptor for the program, which then can be attached to a subsystem (e.g. networking). If supported, the subsystem could then further offload the BPF program to hardware (e.g. NIC). For LLVM, BPF target support can be checked, for example, through the following: $ llc --version
LLVM (http://llvm.org/):
LLVM version 3.8.1
Optimized build.
Default target: x86_64-unknown-linux-gnu
Host CPU: skylake

Registered Targets:
[...]
bpf        - BPF (host endian)
bpfeb      - BPF (big endian)
bpfel      - BPF (little endian)
[...]


By default, the bpf target uses the endianness of the CPU it compiles on, meaning that if the CPU’s endianness is little endian, the program is represented in little endian format as well, and if the CPU’s endianness is big endian, the program is represented in big endian. This also matches the runtime behavior of BPF, which is generic and uses the CPU’s endianness it runs on in order to not disadvantage architectures in any of the format.

For cross-compilation, the two targets bpfeb and bpfel were introduced, thanks to that BPF programs can be compiled on a node running in one endianness (e.g. little endian on x86) and run on a node in another endianness format (e.g. big endian on arm). Note that the front end (clang) needs to run in the target endianness as well.

Using bpf as a target is the preferred way in situations where no mixture of endianness applies. For example, compilation on x86_64 results in the same output for the targets bpf and bpfel due to being little endian, therefore scripts triggering a compilation also do not have to be endian aware.

A minimal, stand-alone XDP drop program might look like the following example (xdp-example.c):

#include <linux/bpf.h>

#ifndef __section
# define __section(NAME)                  \
__attribute__((section(NAME), used))
#endif

__section("prog")
int xdp_drop(struct xdp_md *ctx)
{
return XDP_DROP;
}

char __license[] __section("license") = "GPL";


It can then be compiled and loaded into the kernel as follows:

$clang -O2 -Wall -target bpf -c xdp-example.c -o xdp-example.o # ip link set dev em1 xdp obj xdp-example.o  Note Attaching an XDP BPF program to a network device as above requires Linux 4.11 with a device that supports XDP, or Linux 4.12 or later. For the generated object file LLVM (>= 3.9) uses the official BPF machine value, that is, EM_BPF (decimal: 247 / hex: 0xf7). In this example, the program has been compiled with bpf target under x86_64, therefore LSB (as opposed to MSB) is shown regarding endianness: $ file xdp-example.o
xdp-example.o: ELF 64-bit LSB relocatable, *unknown arch 0xf7* version 1 (SYSV), not stripped


readelf -a xdp-example.o will dump further information about the ELF file, which can sometimes be useful for introspecting generated section headers, relocation entries and the symbol table.

In the unlikely case where clang and LLVM need to be compiled from scratch, the following commands can be used:

$git clone http://llvm.org/git/llvm.git$ cd llvm/tools
$git clone --depth 1 http://llvm.org/git/clang.git$ cd ..; mkdir build; cd build
$cmake .. -DLLVM_TARGETS_TO_BUILD="BPF;X86" -DBUILD_SHARED_LIBS=OFF -DCMAKE_BUILD_TYPE=Release -DLLVM_BUILD_RUNTIME=OFF$ make -j $(getconf _NPROCESSORS_ONLN)$ ./bin/llc --version
LLVM (http://llvm.org/):
LLVM version x.y.zsvn
Optimized build.
Default target: x86_64-unknown-linux-gnu
Host CPU: skylake

Registered Targets:
bpf    - BPF (host endian)
bpfeb  - BPF (big endian)
bpfel  - BPF (little endian)
x86    - 32-bit X86: Pentium-Pro and above
x86-64 - 64-bit X86: EM64T and AMD64

$export PATH=$PWD/bin:$PATH # add to ~/.bashrc  Make sure that --version mentions Optimized build., otherwise the compilation time for programs when having LLVM in debugging mode will significantly increase (e.g. by 10x or more). For debugging, clang can generate the assembler output as follows: $ clang -O2 -S -Wall -target bpf -c xdp-example.c -o xdp-example.S
cat xdp-example.S .text .section prog,"ax",@progbits .globl xdp_drop .p2align 3 xdp_drop: # @xdp_drop # BB#0: r0 = 1 exit .section license,"aw",@progbits .globl __license # @__license __license: .asciz "GPL"  Starting from LLVM’s release 6.0, there is also assembler parser support. You can program using BPF assembler directly, then use llvm-mc to assemble it into an object file. For example, you can assemble the xdp-example.S listed above back into object file using:  llvm-mc -triple bpf -filetype=obj -o xdp-example.o xdp-example.S


Furthermore, more recent LLVM versions (>= 4.0) can also store debugging information in dwarf format into the object file. This can be done through the usual workflow by adding -g for compilation.

$clang -O2 -g -Wall -target bpf -c xdp-example.c -o xdp-example.o$ llvm-objdump -S -no-show-raw-insn xdp-example.o

xdp-example.o:        file format ELF64-BPF

Disassembly of section prog:
xdp_drop:
; {
0:        r0 = 1
; return XDP_DROP;
1:        exit


The llvm-objdump tool can then annotate the assembler output with the original C code used in the compilation. The trivial example in this case does not contain much C code, however, the line numbers shown as 0: and 1: correspond directly to the kernel’s verifier log.

This means that in case BPF programs get rejected by the verifier, llvm-objdump can help to correlate the instructions back to the original C code, which is highly useful for analysis.

# ip link set dev em1 xdp obj xdp-example.o verb

Prog section 'prog' loaded (5)!
- Type:         6
- Instructions: 2 (0 over limit)
- License:      GPL

Verifier analysis:

0: (b7) r0 = 1
1: (95) exit
processed 2 insns


As it can be seen in the verifier analysis, the llvm-objdump output dumps the same BPF assembler code as the kernel.

Leaving out the -no-show-raw-insn option will also dump the raw struct bpf_insn as hex in front of the assembly:

$llvm-objdump -S xdp-example.o xdp-example.o: file format ELF64-BPF Disassembly of section prog: xdp_drop: ; { 0: b7 00 00 00 01 00 00 00 r0 = 1 ; return foo(); 1: 95 00 00 00 00 00 00 00 exit  For LLVM IR debugging, the compilation process for BPF can be split into two steps, generating a binary LLVM IR intermediate file xdp-example.bc, which can later on be passed to llc: $ clang -O2 -Wall -target bpf -emit-llvm -c xdp-example.c -o xdp-example.bc
$llc xdp-example.bc -march=bpf -filetype=obj -o xdp-example.o  The generated LLVM IR can also be dumped in human readable format through: $ clang -O2 -Wall -emit-llvm -S -c xdp-example.c -o -


LLVM is able to attach debug information such as the description of used data types in the program to the generated BPF object file. By default this is in DWARF format.

A heavily simplified version used by BPF is called BTF (BPF Type Format). The resulting DWARF can be converted into BTF and is later on loaded into the kernel through BPF object loaders. The kernel will then verify the BTF data for correctness and keeps track of the data types the BTF data is containing.

BPF maps can then be annotated with key and value types out of the BTF data such that a later dump of the map exports the map data along with the related type information. This allows for better introspection, debugging and value pretty printing. Note that BTF data is a generic debugging data format and as such any DWARF to BTF converted data can be loaded (e.g. kernel’s vmlinux DWARF data could be converted to BTF and loaded). Latter is in particular useful for BPF tracing in the future.

In order to generate BTF from DWARF debugging information, elfutils (>= 0.173) is needed. If that is not available, then adding the -mattr=dwarfris option to the llc command is required during compilation:

$llc -march=bpf -mattr=help |& grep dwarfris dwarfris - Disable MCAsmInfo DwarfUsesRelocationsAcrossSections. [...]  The reason using -mattr=dwarfris is because the flag dwarfris (dwarf relocation in section) disables DWARF cross-section relocations between DWARF and the ELF’s symbol table since libdw does not have proper BPF relocation support, and therefore tools like pahole would otherwise not be able to properly dump structures from the object. elfutils (>= 0.173) implements proper BPF relocation support and therefore the same can be achieved without the -mattr=dwarfris option. Dumping the structures from the object file could be done from either DWARF or BTF information. pahole uses the LLVM emitted DWARF information at this point, however, future pahole versions could rely on BTF if available. For converting DWARF into BTF, a recent pahole version (>= 1.12) is required. A recent pahole version can also be obtained from its official git repository if not available from one of the distribution packages: $ git clone https://git.kernel.org/pub/scm/devel/pahole/pahole.git


pahole comes with the option -J to convert DWARF into BTF from an object file. pahole can be probed for BTF support as follows (note that the llvm-objcopy tool is required for pahole as well, so check its presence, too):

$pahole --help | grep BTF -J, --btf_encode Encode as BTF  Generating debugging information also requires the front end to generate source level debug information by passing -g to the clang command line. Note that -g is needed independently of whether llc’s dwarfris option is used. Full example for generating the object file: $ clang -O2 -g -Wall -target bpf -emit-llvm -c xdp-example.c -o xdp-example.bc
$llc xdp-example.bc -march=bpf -mattr=dwarfris -filetype=obj -o xdp-example.o  Alternatively, by using clang only to build a BPF program with debugging information (again, the dwarfris flag can be omitted when having proper elfutils version): $ clang -target bpf -O2 -g -c -Xclang -target-feature -Xclang +dwarfris -c xdp-example.c -o xdp-example.o


After successful compilation pahole can be used to properly dump structures of the BPF program based on the DWARF information:

$pahole xdp-example.o struct xdp_md { __u32 data; /* 0 4 */ __u32 data_end; /* 4 4 */ __u32 data_meta; /* 8 4 */ /* size: 12, cachelines: 1, members: 3 */ /* last cacheline: 12 bytes */ };  Through the option -J pahole can eventually generate the BTF from DWARF. In the object file DWARF data will still be retained alongside the newly added BTF data. Full clang and pahole example combined: $ clang -target bpf -O2 -Wall -g -c -Xclang -target-feature -Xclang +dwarfris -c xdp-example.c -o xdp-example.o
$pahole -J xdp-example.o  The presence of a .BTF section can be seen through readelf tool: $ readelf -a xdp-example.o
[...]
[18] .BTF              PROGBITS         0000000000000000  00000671
[...]


BPF loaders such as iproute2 will detect and load the BTF section, so that BPF maps can be annotated with type information.

LLVM by default uses the BPF base instruction set for generating code in order to make sure that the generated object file can also be loaded with older kernels such as long-term stable kernels (e.g. 4.9+).

However, LLVM has a -mcpu selector for the BPF back end in order to select different versions of the BPF instruction set, namely instruction set extensions on top of the BPF base instruction set in order to generate more efficient and smaller code.

Available -mcpu options can be queried through:

$llc -march bpf -mcpu=help Available CPUs for this target: generic - Select the generic processor. probe - Select the probe processor. v1 - Select the v1 processor. v2 - Select the v2 processor. [...]  The generic processor is the default processor, which is also the base instruction set v1 of BPF. Options v1 and v2 are typically useful in an environment where the BPF program is being cross compiled and the target host where the program is loaded differs from the one where it is compiled (and thus available BPF kernel features might differ as well). The recommended -mcpu option which is also used by Cilium internally is -mcpu=probe! Here, the LLVM BPF back end queries the kernel for availability of BPF instruction set extensions and when found available, LLVM will use them for compiling the BPF program whenever appropriate. A full command line example with llc’s -mcpu=probe: $ clang -O2 -Wall -target bpf -emit-llvm -c xdp-example.c -o xdp-example.bc
$llc xdp-example.bc -march=bpf -mcpu=probe -filetype=obj -o xdp-example.o  Generally, LLVM IR generation is architecture independent. There are however a few differences when using clang -target bpf versus leaving -target bpf out and thus using clang’s default target which, depending on the underlying architecture, might be x86_64, arm64 or others. Quoting from the kernel’s Documentation/bpf/bpf_devel_QA.txt: • BPF programs may recursively include header file(s) with file scope inline assembly codes. The default target can handle this well, while bpf target may fail if bpf backend assembler does not understand these assembly codes, which is true in most cases. • When compiled without -g, additional elf sections, e.g., .eh_frame and .rela.eh_frame, may be present in the object file with default target, but not with bpf target. • The default target may turn a C switch statement into a switch table lookup and jump operation. Since the switch table is placed in the global read-only section, the bpf program will fail to load. The bpf target does not support switch table optimization. The clang option -fno-jump-tables can be used to disable switch table generation. • For clang -target bpf, it is guaranteed that pointer or long / unsigned long types will always have a width of 64 bit, no matter whether underlying clang binary or default target (or kernel) is 32 bit. However, when native clang target is used, then it will compile these types based on the underlying architecture’s conventions, meaning in case of 32 bit architecture, pointer or long / unsigned long types e.g. in BPF context structure will have width of 32 bit while the BPF LLVM back end still operates in 64 bit. The native target is mostly needed in tracing for the case of walking the kernel’s struct pt_regs that maps CPU registers, or other kernel structures where CPU’s register width matters. In all other cases such as networking, the use of clang -target bpf is the preferred choice. Also, LLVM started to support 32-bit subregisters and BPF ALU32 instructions since LLVM’s release 7.0. A new code generation attribute alu32 is added. When it is enabled, LLVM will try to use 32-bit subregisters whenever possible, typically when there are operations on 32-bit types. The associated ALU instructions with 32-bit subregisters will become ALU32 instructions. For example, for the following sample code: $ cat 32-bit-example.c
void cal(unsigned int *a, unsigned int *b, unsigned int *c)
{
unsigned int sum = *a + *b;
*c = sum;
}


At default code generation, the assembler will looks like:

$clang -target bpf -emit-llvm -S 32-bit-example.c$ llc -march=bpf 32-bit-example.ll
$cat 32-bit-example.s cal: r1 = *(u32 *)(r1 + 0) r2 = *(u32 *)(r2 + 0) r2 += r1 *(u32 *)(r3 + 0) = r2 exit  64-bit registers are used, hence the addition means 64-bit addition. Now, if you enable the new 32-bit subregisters support by specifying -mattr=+alu32, then the assembler will looks like: $ llc -march=bpf -mattr=+alu32 32-bit-example.ll
$cat 32-bit-example.s cal: w1 = *(u32 *)(r1 + 0) w2 = *(u32 *)(r2 + 0) w2 += w1 *(u32 *)(r3 + 0) = w2 exit  w register, meaning 32-bit subregister, will be used instead of 64-bit r register. Enable 32-bit subregisters might help reducing type extension instruction sequences. It could also help kernel eBPF JIT compiler for 32-bit architectures for which registers pairs are used to model the 64-bit eBPF registers and extra instructions are needed for manipulating the high 32-bit. Given read from 32-bit subregister is guaranteed to read from low 32-bit only even though write still needs to clear the high 32-bit, if the JIT compiler has known the definition of one register only has subregister reads, then instructions for setting the high 32-bit of the destination could be eliminated. When writing C programs for BPF, there are a couple of pitfalls to be aware of, compared to usual application development with C. The following items describe some of the differences for the BPF model: 1. Everything needs to be inlined, there are no function calls (on older LLVM versions) or shared library calls available. Shared libraries, etc cannot be used with BPF. However, common library code used in BPF programs can be placed into header files and included in the main programs. For example, Cilium makes heavy use of it (see bpf/lib/). However, this still allows for including header files, for example, from the kernel or other libraries and reuse their static inline functions or macros / definitions. Unless a recent kernel (4.16+) and LLVM (6.0+) is used where BPF to BPF function calls are supported, then LLVM needs to compile and inline the entire code into a flat sequence of BPF instructions for a given program section. In such case, best practice is to use an annotation like __inline for every library function as shown below. The use of always_inline is recommended, since the compiler could still decide to uninline large functions that are only annotated as inline. In case the latter happens, LLVM will generate a relocation entry into the ELF file, which BPF ELF loaders such as iproute2 cannot resolve and will thus produce an error since only BPF maps are valid relocation entries which loaders can process. #include <linux/bpf.h> #ifndef __section # define __section(NAME) \ __attribute__((section(NAME), used)) #endif #ifndef __inline # define __inline \ inline __attribute__((always_inline)) #endif static __inline int foo(void) { return XDP_DROP; } __section("prog") int xdp_drop(struct xdp_md *ctx) { return foo(); } char __license[] __section("license") = "GPL";  2. Multiple programs can reside inside a single C file in different sections. C programs for BPF make heavy use of section annotations. A C file is typically structured into 3 or more sections. BPF ELF loaders use these names to extract and prepare the relevant information in order to load the programs and maps through the bpf system call. For example, iproute2 uses maps and license as default section name to find metadata needed for map creation and the license for the BPF program, respectively. On program creation time the latter is pushed into the kernel as well, and enables some of the helper functions which are exposed as GPL only in case the program also holds a GPL compatible license, for example bpf_ktime_get_ns(), bpf_probe_read() and others. The remaining section names are specific for BPF program code, for example, the below code has been modified to contain two program sections, ingress and egress. The toy example code demonstrates that both can share a map and common static inline helpers such as the account_data() function. The xdp-example.c example has been modified to a tc-example.c example that can be loaded with tc and attached to a netdevice’s ingress and egress hook. It accounts the transferred bytes into a map called acc_map, which has two map slots, one for traffic accounted on the ingress hook, one on the egress hook. #include <linux/bpf.h> #include <linux/pkt_cls.h> #include <stdint.h> #include <iproute2/bpf_elf.h> #ifndef __section # define __section(NAME) \ __attribute__((section(NAME), used)) #endif #ifndef __inline # define __inline \ inline __attribute__((always_inline)) #endif #ifndef lock_xadd # define lock_xadd(ptr, val) \ ((void)__sync_fetch_and_add(ptr, val)) #endif #ifndef BPF_FUNC # define BPF_FUNC(NAME, ...) \ (*NAME)(__VA_ARGS__) = (void *)BPF_FUNC_##NAME #endif static void *BPF_FUNC(map_lookup_elem, void *map, const void *key); struct bpf_elf_map acc_map __section("maps") = { .type = BPF_MAP_TYPE_ARRAY, .size_key = sizeof(uint32_t), .size_value = sizeof(uint32_t), .pinning = PIN_GLOBAL_NS, .max_elem = 2, }; static __inline int account_data(struct __sk_buff *skb, uint32_t dir) { uint32_t *bytes; bytes = map_lookup_elem(&acc_map, &dir); if (bytes) lock_xadd(bytes, skb->len); return TC_ACT_OK; } __section("ingress") int tc_ingress(struct __sk_buff *skb) { return account_data(skb, 0); } __section("egress") int tc_egress(struct __sk_buff *skb) { return account_data(skb, 1); } char __license[] __section("license") = "GPL";  The example also demonstrates a couple of other things which are useful to be aware of when developing programs. The code includes kernel headers, standard C headers and an iproute2 specific header containing the definition of struct bpf_elf_map. iproute2 has a common BPF ELF loader and as such the definition of struct bpf_elf_map is the very same for XDP and tc typed programs. A struct bpf_elf_map entry defines a map in the program and contains all relevant information (such as key / value size, etc) needed to generate a map which is used from the two BPF programs. The structure must be placed into the maps section, so that the loader can find it. There can be multiple map declarations of this type with different variable names, but all must be annotated with __section("maps"). The struct bpf_elf_map is specific to iproute2. Different BPF ELF loaders can have different formats, for example, the libbpf in the kernel source tree, which is mainly used by perf, has a different specification. iproute2 guarantees backwards compatibility for struct bpf_elf_map. Cilium follows the iproute2 model. The example also demonstrates how BPF helper functions are mapped into the C code and being used. Here, map_lookup_elem() is defined by mapping this function into the BPF_FUNC_map_lookup_elem enum value which is exposed as a helper in uapi/linux/bpf.h. When the program is later loaded into the kernel, the verifier checks whether the passed arguments are of the expected type and re-points the helper call into a real function call. Moreover, map_lookup_elem() also demonstrates how maps can be passed to BPF helper functions. Here, &acc_map from the maps section is passed as the first argument to map_lookup_elem(). Since the defined array map is global, the accounting needs to use an atomic operation, which is defined as lock_xadd(). LLVM maps __sync_fetch_and_add() as a built-in function to the BPF atomic add instruction, that is, BPF_STX | BPF_XADD | BPF_W for word sizes. Last but not least, the struct bpf_elf_map tells that the map is to be pinned as PIN_GLOBAL_NS. This means that tc will pin the map into the BPF pseudo file system as a node. By default, it will be pinned to /sys/fs/bpf/tc/globals/acc_map for the given example. Due to the PIN_GLOBAL_NS, the map will be placed under /sys/fs/bpf/tc/globals/. globals acts as a global namespace that spans across object files. If the example used PIN_OBJECT_NS, then tc would create a directory that is local to the object file. For example, different C files with BPF code could have the same acc_map definition as above with a PIN_GLOBAL_NS pinning. In that case, the map will be shared among BPF programs originating from various object files. PIN_NONE would mean that the map is not placed into the BPF file system as a node, and as a result will not be accessible from user space after tc quits. It would also mean that tc creates two separate map instances for each program, since it cannot retrieve a previously pinned map under that name. The acc_map part from the mentioned path is the name of the map as specified in the source code. Thus, upon loading of the ingress program, tc will find that no such map exists in the BPF file system and creates a new one. On success, the map will also be pinned, so that when the egress program is loaded through tc, it will find that such map already exists in the BPF file system and will reuse that for the egress program. The loader also makes sure in case maps exist with the same name that also their properties (key / value size, etc) match. Just like tc can retrieve the same map, also third party applications can use the BPF_OBJ_GET command from the bpf system call in order to create a new file descriptor pointing to the same map instance, which can then be used to lookup / update / delete map elements. The code can be compiled and loaded via iproute2 as follows: $ clang -O2 -Wall -target bpf -c tc-example.c -o tc-example.o

# tc qdisc add dev em1 clsact
# tc filter add dev em1 ingress bpf da obj tc-example.o sec ingress
# tc filter add dev em1 egress bpf da obj tc-example.o sec egress

# tc filter show dev em1 ingress
filter protocol all pref 49152 bpf
filter protocol all pref 49152 bpf handle 0x1 tc-example.o:[ingress] direct-action id 1 tag c5f7825e5dac396f

# tc filter show dev em1 egress
filter protocol all pref 49152 bpf
filter protocol all pref 49152 bpf handle 0x1 tc-example.o:[egress] direct-action id 2 tag b2fd5adc0f262714

# mount | grep bpf
sysfs on /sys/fs/bpf type sysfs (rw,nosuid,nodev,noexec,relatime,seclabel)
bpf on /sys/fs/bpf type bpf (rw,relatime,mode=0700)

# tree /sys/fs/bpf/
/sys/fs/bpf/
+-- ip -> /sys/fs/bpf/tc/
+-- tc
|   +-- globals
|       +-- acc_map
+-- xdp -> /sys/fs/bpf/tc/

4 directories, 1 file


As soon as packets pass the em1 device, counters from the BPF map will be increased.

1. There are no global variables allowed.

For the reasons already mentioned in point 1, BPF cannot have global variables as often used in normal C programs.

However, there is a work-around in that the program can simply use a BPF map of type BPF_MAP_TYPE_PERCPU_ARRAY with just a single slot of arbitrary value size. This works, because during execution, BPF programs are guaranteed to never get preempted by the kernel and therefore can use the single map entry as a scratch buffer for temporary data, for example, to extend beyond the stack limitation. This also functions across tail calls, since it has the same guarantees with regards to preemption.

Otherwise, for holding state across multiple BPF program runs, normal BPF maps can be used.

1. There are no const strings or arrays allowed.

Defining const strings or other arrays in the BPF C program does not work for the same reasons as pointed out in sections 1 and 3, which is, that relocation entries will be generated in the ELF file which will be rejected by loaders due to not being part of the ABI towards loaders (loaders also cannot fix up such entries as it would require large rewrites of the already compiled BPF sequence).

In the future, LLVM might detect these occurrences and early throw an error to the user.

Helper functions such as trace_printk() can be worked around as follows:

static void BPF_FUNC(trace_printk, const char *fmt, int fmt_size, ...);

#ifndef printk
# define printk(fmt, ...)                                      \
({                                                         \
char ____fmt[] = fmt;                                  \
trace_printk(____fmt, sizeof(____fmt), ##__VA_ARGS__); \
})
#endif


The program can then use the macro naturally like printk("skb len:%u\n", skb->len);. The output will then be written to the trace pipe. tc exec bpf dbg can be used to retrieve the messages from there.

The use of the trace_printk() helper function has a couple of disadvantages and thus is not recommended for production usage. Constant strings like the "skb len:%u\n" need to be loaded into the BPF stack each time the helper function is called, but also BPF helper functions are limited to a maximum of 5 arguments. This leaves room for only 3 additional variables which can be passed for dumping.

Therefore, despite being helpful for quick debugging, it is recommended (for networking programs) to use the skb_event_output() or the xdp_event_output() helper, respectively. They allow for passing custom structs from the BPF program to the perf event ring buffer along with an optional packet sample. For example, Cilium’s monitor makes use of these helpers in order to implement a debugging framework, notifications for network policy violations, etc. These helpers pass the data through a lockless memory mapped per-CPU perf ring buffer, and is thus significantly faster than trace_printk().

1. Use of LLVM built-in functions for memset()/memcpy()/memmove()/memcmp().

Since BPF programs cannot perform any function calls other than those to BPF helpers, common library code needs to be implemented as inline functions. In addition, also LLVM provides some built-ins that the programs can use for constant sizes (here: n) which will then always get inlined:

#ifndef memset
# define memset(dest, chr, n)   __builtin_memset((dest), (chr), (n))
#endif

#ifndef memcpy
# define memcpy(dest, src, n)   __builtin_memcpy((dest), (src), (n))
#endif

#ifndef memmove
# define memmove(dest, src, n)  __builtin_memmove((dest), (src), (n))
#endif


The memcmp() built-in had some corner cases where inlining did not take place due to an LLVM issue in the back end, and is therefore not recommended to be used until the issue is fixed.

1. There are no loops available (yet).

The BPF verifier in the kernel checks that a BPF program does not contain loops by performing a depth first search of all possible program paths besides other control flow graph validations. The purpose is to make sure that the program is always guaranteed to terminate.

A very limited form of looping is available for constant upper loop bounds by using #pragma unroll directive. Example code that is compiled to BPF:

#pragma unroll
for (i = 0; i < IPV6_MAX_HEADERS; i++) {
switch (nh) {
case NEXTHDR_NONE:
return DROP_INVALID_EXTHDR;
case NEXTHDR_FRAGMENT:
return DROP_FRAG_NOSUPPORT;
case NEXTHDR_HOP:
case NEXTHDR_ROUTING:
case NEXTHDR_AUTH:
case NEXTHDR_DEST:
if (skb_load_bytes(skb, l3_off + len, &opthdr, sizeof(opthdr)) < 0)
return DROP_INVALID;

nh = opthdr.nexthdr;
if (nh == NEXTHDR_AUTH)
len += ipv6_authlen(&opthdr);
else
len += ipv6_optlen(&opthdr);
break;
default:
*nexthdr = nh;
return len;
}
}


Another possibility is to use tail calls by calling into the same program again and using a BPF_MAP_TYPE_PERCPU_ARRAY map for having a local scratch space. While being dynamic, this form of looping however is limited to a maximum of 32 iterations.

In the future, BPF may have some native, but limited form of implementing loops.

1. Partitioning programs with tail calls.

Tail calls provide the flexibility to atomically alter program behavior during runtime by jumping from one BPF program into another. In order to select the next program, tail calls make use of program array maps (BPF_MAP_TYPE_PROG_ARRAY), and pass the map as well as the index to the next program to jump to. There is no return to the old program after the jump has been performed, and in case there was no program present at the given map index, then execution continues on the original program.

For example, this can be used to implement various stages of a parser, where such stages could be updated with new parsing features during runtime.

Another use case are event notifications, for example, Cilium can opt in packet drop notifications during runtime, where the skb_event_output() call is located inside the tail called program. Thus, during normal operations, the fall-through path will always be executed unless a program is added to the related map index, where the program then prepares the metadata and triggers the event notification to a user space daemon.

Program array maps are quite flexible, enabling also individual actions to be implemented for programs located in each map index. For example, the root program attached to XDP or tc could perform an initial tail call to index 0 of the program array map, performing traffic sampling, then jumping to index 1 of the program array map, where firewalling policy is applied and the packet either dropped or further processed in index 2 of the program array map, where it is mangled and sent out of an interface again. Jumps in the program array map can, of course, be arbitrary. The kernel will eventually execute the fall-through path when the maximum tail call limit has been reached.

Minimal example extract of using tail calls:

[...]

#ifndef __stringify
# define __stringify(X)   #X
#endif

#ifndef __section
# define __section(NAME)                  \
__attribute__((section(NAME), used))
#endif

#ifndef __section_tail
# define __section_tail(ID, KEY)          \
__section(__stringify(ID) "/" __stringify(KEY))
#endif

#ifndef BPF_FUNC
# define BPF_FUNC(NAME, ...)              \
(*NAME)(__VA_ARGS__) = (void *)BPF_FUNC_##NAME
#endif

#define BPF_JMP_MAP_ID   1

static void BPF_FUNC(tail_call, struct __sk_buff *skb, void *map,
uint32_t index);

struct bpf_elf_map jmp_map __section("maps") = {
.type           = BPF_MAP_TYPE_PROG_ARRAY,
.id             = BPF_JMP_MAP_ID,
.size_key       = sizeof(uint32_t),
.size_value     = sizeof(uint32_t),
.pinning        = PIN_GLOBAL_NS,
.max_elem       = 1,
};

__section_tail(JMP_MAP_ID, 0)
int looper(struct __sk_buff *skb)
{
printk("skb cb: %u\n", skb->cb[0]++);
tail_call(skb, &jmp_map, 0);
return TC_ACT_OK;
}

__section("prog")
int entry(struct __sk_buff *skb)
{
skb->cb[0] = 0;
tail_call(skb, &jmp_map, 0);
return TC_ACT_OK;
}

char __license[] __section("license") = "GPL";


When loading this toy program, tc will create the program array and pin it to the BPF file system in the global namespace under jmp_map. Also, the BPF ELF loader in iproute2 will also recognize sections that are marked as __section_tail(). The provided id in struct bpf_elf_map will be matched against the id marker in the __section_tail(), that is, JMP_MAP_ID, and the program therefore loaded at the user specified program array map index, which is 0 in this example. As a result, all provided tail call sections will be populated by the iproute2 loader to the corresponding maps. This mechanism is not specific to tc, but can be applied with any other BPF program type that iproute2 supports (such as XDP, lwt).

The generated elf contains section headers describing the map id and the entry within that map:

llvm-objdump -S --no-show-raw-insn prog_array.o | less prog_array.o: file format ELF64-BPF Disassembly of section 1/0: looper: 0: r6 = r1 1: r2 = *(u32 *)(r6 + 48) 2: r1 = r2 3: r1 += 1 4: *(u32 *)(r6 + 48) = r1 5: r1 = 0 ll 7: call -1 8: r1 = r6 9: r2 = 0 ll 11: r3 = 0 12: call 12 13: r0 = 0 14: exit Disassembly of section prog: entry: 0: r2 = 0 1: *(u32 *)(r1 + 48) = r2 2: r2 = 0 ll 4: r3 = 0 5: call 12 6: r0 = 0 7: exi  In this case, the section 1/0 indicates that the looper() function resides in the map id 1 at position 0. The pinned map can be retrieved by a user space applications (e.g. Cilium daemon), but also by tc itself in order to update the map with new programs. Updates happen atomically, the initial entry programs that are triggered first from the various subsystems are also updated atomically. Example for tc to perform tail call map updates: # tc exec bpf graft m:globals/jmp_map key 0 obj new.o sec foo  In case iproute2 would update the pinned program array, the graft command can be used. By pointing it to globals/jmp_map, tc will update the map at index / key 0 with a new program residing in the object file new.o under section foo. 1. Limited stack space of maximum 512 bytes. Stack space in BPF programs is limited to only 512 bytes, which needs to be taken into careful consideration when implementing BPF programs in C. However, as mentioned earlier in point 3, a BPF_MAP_TYPE_PERCPU_ARRAY map with a single entry can be used in order to enlarge scratch buffer space. 1. Use of BPF inline assembly possible. LLVM 6.0 or later allows use of inline assembly for BPF for the rare cases where it might be needed. The following (nonsense) toy example shows a 64 bit atomic add. Due to lack of documentation, LLVM source code in lib/Target/BPF/BPFInstrInfo.td as well as test/CodeGen/BPF/ might be helpful for providing some additional examples. Test code: #include <linux/bpf.h> #ifndef __section # define __section(NAME) \ __attribute__((section(NAME), used)) #endif __section("prog") int xdp_test(struct xdp_md *ctx) { __u64 a = 2, b = 3, *c = &a; /* just a toy xadd example to show the syntax */ asm volatile("lock *(u64 *)(%0+0) += %1" : "=r"(c) : "r"(b), "0"(c)); return a; } char __license[] __section("license") = "GPL";  The above program is compiled into the following sequence of BPF instructions: Verifier analysis: 0: (b7) r1 = 2 1: (7b) *(u64 *)(r10 -8) = r1 2: (b7) r1 = 3 3: (bf) r2 = r10 4: (07) r2 += -8 5: (db) lock *(u64 *)(r2 +0) += r1 6: (79) r0 = *(u64 *)(r10 -8) 7: (95) exit processed 8 insns (limit 131072), stack depth 8  1. Remove struct padding with aligning members by using #pragma pack. In modern compilers, data structures are aligned by default to access memory efficiently. Structure members are aligned to memory address that multiples their size, and padding is added for the proper alignment. Because of this, the size of struct may often grow larger than expected. struct called_info { u64 start; // 8-byte u64 end; // 8-byte u32 sector; // 4-byte }; // size of 20-byte ? printf("size of %d-byte\n", sizeof(struct called_info)); // size of 24-byte // Actual compiled composition of struct called_info // 0x0(0) 0x8(8) // ↓________________________↓ // | start (8) | // |________________________| // | end (8) | // |________________________| // | sector(4) | PADDING | <= address aligned to 8 // |____________|___________| with 4-byte PADDING.  The BPF verifier in the kernel checks the stack boundary that a BPF program does not access outside of boundary or uninitialized stack area. Using struct with the padding as a map value, will cause invalid indirect read from stack failure on bpf_prog_load(). Example code: struct called_info { u64 start; u64 end; u32 sector; }; struct bpf_map_def SEC("maps") called_info_map = { .type = BPF_MAP_TYPE_HASH, .key_size = sizeof(long), .value_size = sizeof(struct called_info), .max_entries = 4096, }; SEC("kprobe/submit_bio") int submit_bio_entry(struct pt_regs *ctx) { char fmt[] = "submit_bio(bio=0x%lx) called: %llu\n"; u64 start_time = bpf_ktime_get_ns(); long bio_ptr = PT_REGS_PARM1(ctx); struct called_info called_info = { .start = start_time, .end = 0, .bi_sector = 0 }; bpf_map_update_elem(&called_info_map, &bio_ptr, &called_info, BPF_ANY); bpf_trace_printk(fmt, sizeof(fmt), bio_ptr, start_time); return 0; } // On bpf_load_program bpf_load_program() err=13 0: (bf) r6 = r1 ... 19: (b7) r1 = 0 20: (7b) *(u64 *)(r10 -72) = r1 21: (7b) *(u64 *)(r10 -80) = r7 22: (63) *(u32 *)(r10 -64) = r1 ... 30: (85) call bpf_map_update_elem#2 invalid indirect read from stack off -80+20 size 24  At bpf_prog_load(), an eBPF verifier bpf_check() is called, and it’ll check stack boundary by calling check_func_arg() -> check_stack_boundary(). From the upper error shows, struct called_info is compiled to 24-byte size, and the message says reading a data from +20 is an invalid indirect read. And as we discussed earlier, the address 0x14(20) is the place where PADDING is. // Actual compiled composition of struct called_info // 0x10(16) 0x14(20) 0x18(24) // ↓____________↓___________↓ // | sector(4) | PADDING | <= address aligned to 8 // |____________|___________| with 4-byte PADDING.  The check_stack_boundary() internally loops through the every access_size (24) byte from the start pointer to make sure that it’s within stack boundary and all elements of the stack are initialized. Since the padding isn’t supposed to be used, it gets the ‘invalid indirect read from stack’ failure. To avoid this kind of failure, remove the padding from the struct is necessary. Removing the padding by using #pragma pack(n) directive: #pragma pack(4) struct called_info { u64 start; // 8-byte u64 end; // 8-byte u32 sector; // 4-byte }; // size of 20-byte ? printf("size of %d-byte\n", sizeof(struct called_info)); // size of 20-byte // Actual compiled composition of packed struct called_info // 0x0(0) 0x8(8) // ↓________________________↓ // | start (8) | // |________________________| // | end (8) | // |________________________| // | sector(4) | <= address aligned to 4 // |____________| with no PADDING.  By locating #pragma pack(4) before of struct called_info, compiler will align members of a struct to the least of 4-byte and their natural alignment. As you can see, the size of struct called_info has been shrunk to 20-byte and the padding is no longer exist. But, removing the padding have downsides either. For example, compiler will generate less optimized code. Since we’ve removed the padding, processors will conduct unaligned access to the structure and this might lead to performance degradation. And also, unaligned access might get rejected by verifier on some architectures. However, there is a way to avoid downsides of packed structure. By simply adding the explicit padding u32 pad member at the end will resolve the same problem without packing of the structure. struct called_info { u64 start; // 8-byte u64 end; // 8-byte u32 sector; // 4-byte u32 pad; // 4-byte }; // size of 24-byte ? printf("size of %d-byte\n", sizeof(struct called_info)); // size of 24-byte // Actual compiled composition of struct called_info with explicit padding // 0x0(0) 0x8(8) // ↓________________________↓ // | start (8) | // |________________________| // | end (8) | // |________________________| // | sector(4) | pad (4) | <= address aligned to 8 // |____________|___________| with explicit PADDING.  1. Accessing packet data via invalidated references Some networking BPF helper functions such as bpf_skb_store_bytes might change the size of a packet data. As verifier is not able to track such changes, any a priori reference to the data will be invalidated by verifier. Therefore, the reference needs to be updated before accessing the data to avoid verifier rejecting a program. To illustrate this, consider the following snippet: struct iphdr *ip4 = (struct iphdr *) skb->data + ETH_HLEN; skb_store_bytes(skb, l3_off + offsetof(struct iphdr, saddr), &new_saddr, 4, 0); if (ip4->protocol == IPPROTO_TCP) { // do something }  Verifier will reject the snippet due to dereference of the invalidated ip4->protocol: R1=pkt_end(id=0,off=0,imm=0) R2=pkt(id=0,off=34,r=34,imm=0) R3=inv0 R6=ctx(id=0,off=0,imm=0) R7=inv(id=0,umax_value=4294967295,var_off=(0x0; 0xffffffff)) R8=inv4294967162 R9=pkt(id=0,off=0,r=34,imm=0) R10=fp0,call_-1 ... 18: (85) call bpf_skb_store_bytes#9 19: (7b) *(u64 *)(r10 -56) = r7 R0=inv(id=0) R6=ctx(id=0,off=0,imm=0) R7=inv(id=0,umax_value=2,var_off=(0x0; 0x3)) R8=inv4294967162 R9=inv(id=0) R10=fp0,call_-1 fp-48=mmmm???? fp-56=mmmmmmmm 21: (61) r1 = *(u32 *)(r9 +23) R9 invalid mem access 'inv'  To fix this, the reference to ip4 has to be updated: struct iphdr *ip4 = (struct iphdr *) skb->data + ETH_HLEN; skb_store_bytes(skb, l3_off + offsetof(struct iphdr, saddr), &new_saddr, 4, 0); ip4 = (struct iphdr *) skb->data + ETH_HLEN; if (ip4->protocol == IPPROTO_TCP) { // do something }  ### iproute2¶ There are various front ends for loading BPF programs into the kernel such as bcc, perf, iproute2 and others. The Linux kernel source tree also provides a user space library under tools/lib/bpf/, which is mainly used and driven by perf for loading BPF tracing programs into the kernel. However, the library itself is generic and not limited to perf only. bcc is a toolkit providing many useful BPF programs mainly for tracing that are loaded ad-hoc through a Python interface embedding the BPF C code. Syntax and semantics for implementing BPF programs slightly differ among front ends in general, though. Additionally, there are also BPF samples in the kernel source tree (samples/bpf/) which parse the generated object files and load the code directly through the system call interface. This and previous sections mainly focus on the iproute2 suite’s BPF front end for loading networking programs of XDP, tc or lwt type, since Cilium’s programs are implemented against this BPF loader. In future, Cilium will be equipped with a native BPF loader, but programs will still be compatible to be loaded through iproute2 suite in order to facilitate development and debugging. All BPF program types supported by iproute2 share the same BPF loader logic due to having a common loader back end implemented as a library (lib/bpf.c in iproute2 source tree). The previous section on LLVM also covered some iproute2 parts related to writing BPF C programs, and later sections in this document are related to tc and XDP specific aspects when writing programs. Therefore, this section will rather focus on usage examples for loading object files with iproute2 as well as some of the generic mechanics of the loader. It does not try to provide a complete coverage of all details, but enough for getting started. 1. Loading of XDP BPF object files. Given a BPF object file prog.o has been compiled for XDP, it can be loaded through ip to a XDP-supported netdevice called em1 with the following command: # ip link set dev em1 xdp obj prog.o  The above command assumes that the program code resides in the default section which is called prog in XDP case. Should this not be the case, and the section is named differently, for example, foobar, then the program needs to be loaded as: # ip link set dev em1 xdp obj prog.o sec foobar  Note that it is also possible to load the program out of the .text section. Changing the minimal, stand-alone XDP drop program by removing the __section() annotation from the xdp_drop entry point would look like the following: #include <linux/bpf.h> #ifndef __section # define __section(NAME) \ __attribute__((section(NAME), used)) #endif int xdp_drop(struct xdp_md *ctx) { return XDP_DROP; } char __license[] __section("license") = "GPL";  And can be loaded as follows: # ip link set dev em1 xdp obj prog.o sec .text  By default, ip will throw an error in case a XDP program is already attached to the networking interface, to prevent it from being overridden by accident. In order to replace the currently running XDP program with a new one, the -force option must be used: # ip -force link set dev em1 xdp obj prog.o  Most XDP-enabled drivers today support an atomic replacement of the existing program with a new one without traffic interruption. There is always only a single program attached to an XDP-enabled driver due to performance reasons, hence a chain of programs is not supported. However, as described in the previous section, partitioning of programs can be performed through tail calls to achieve a similar use case when necessary. The ip link command will display an xdp flag if the interface has an XDP program attached. ip link | grep xdp can thus be used to find all interfaces that have XDP running. Further introspection facilities are provided through the detailed view with ip -d link and bpftool can be used to retrieve information about the attached program based on the BPF program ID shown in the ip link dump. In order to remove the existing XDP program from the interface, the following command must be issued: # ip link set dev em1 xdp off  In the case of switching a driver’s operation mode from non-XDP to native XDP and vice versa, typically the driver needs to reconfigure its receive (and transmit) rings in order to ensure received packet are set up linearly within a single page for BPF to read and write into. However, once completed, then most drivers only need to perform an atomic replacement of the program itself when a BPF program is requested to be swapped. In total, XDP supports three operation modes which iproute2 implements as well: xdpdrv, xdpoffload and xdpgeneric. xdpdrv stands for native XDP, meaning the BPF program is run directly in the driver’s receive path at the earliest possible point in software. This is the normal / conventional XDP mode and requires driver’s to implement XDP support, which all major 10G/40G/+ networking drivers in the upstream Linux kernel already provide. xdpgeneric stands for generic XDP and is intended as an experimental test bed for drivers which do not yet support native XDP. Given the generic XDP hook in the ingress path comes at a much later point in time when the packet already enters the stack’s main receive path as a skb, the performance is significantly less than with processing in xdpdrv mode. xdpgeneric therefore is for the most part only interesting for experimenting, less for production environments. Last but not least, the xdpoffload mode is implemented by SmartNICs such as those supported by Netronome’s nfp driver and allow for offloading the entire BPF/XDP program into hardware, thus the program is run on each packet reception directly on the card. This provides even higher performance than running in native XDP although not all BPF map types or BPF helper functions are available for use compared to native XDP. The BPF verifier will reject the program in such case and report to the user what is unsupported. Other than staying in the realm of supported BPF features and helper functions, no special precautions have to be taken when writing BPF C programs. When a command like ip link set dev em1 xdp obj [...] is used, then the kernel will attempt to load the program first as native XDP, and in case the driver does not support native XDP, it will automatically fall back to generic XDP. Thus, for example, using explicitly xdpdrv instead of xdp, the kernel will only attempt to load the program as native XDP and fail in case the driver does not support it, which provides a guarantee that generic XDP is avoided altogether. Example for enforcing a BPF/XDP program to be loaded in native XDP mode, dumping the link details and unloading the program again: # ip -force link set dev em1 xdpdrv obj prog.o # ip link show [...] 6: em1: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 xdp qdisc mq state UP mode DORMANT group default qlen 1000 link/ether be:08:4d:b6:85:65 brd ff:ff:ff:ff:ff:ff prog/xdp id 1 tag 57cd311f2e27366b [...] # ip link set dev em1 xdpdrv off  Same example now for forcing generic XDP, even if the driver would support native XDP, and additionally dumping the BPF instructions of the attached dummy program through bpftool: # ip -force link set dev em1 xdpgeneric obj prog.o # ip link show [...] 6: em1: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 xdpgeneric qdisc mq state UP mode DORMANT group default qlen 1000 link/ether be:08:4d:b6:85:65 brd ff:ff:ff:ff:ff:ff prog/xdp id 4 tag 57cd311f2e27366b <-- BPF program ID 4 [...] # bpftool prog dump xlated id 4 <-- Dump of instructions running on em1 0: (b7) r0 = 1 1: (95) exit # ip link set dev em1 xdpgeneric off  And last but not least offloaded XDP, where we additionally dump program information via bpftool for retrieving general metadata: # ip -force link set dev em1 xdpoffload obj prog.o # ip link show [...] 6: em1: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 xdpoffload qdisc mq state UP mode DORMANT group default qlen 1000 link/ether be:08:4d:b6:85:65 brd ff:ff:ff:ff:ff:ff prog/xdp id 8 tag 57cd311f2e27366b [...] # bpftool prog show id 8 8: xdp tag 57cd311f2e27366b dev em1 <-- Also indicates a BPF program offloaded to em1 loaded_at Apr 11/20:38 uid 0 xlated 16B not jited memlock 4096B # ip link set dev em1 xdpoffload off  Note that it is not possible to use xdpdrv and xdpgeneric or other modes at the same time, meaning only one of the XDP operation modes must be picked. A switch between different XDP modes e.g. from generic to native or vice versa is not atomically possible. Only switching programs within a specific operation mode is: # ip -force link set dev em1 xdpgeneric obj prog.o # ip -force link set dev em1 xdpoffload obj prog.o RTNETLINK answers: File exists # ip -force link set dev em1 xdpdrv obj prog.o RTNETLINK answers: File exists # ip -force link set dev em1 xdpgeneric obj prog.o <-- Succeeds due to xdpgeneric #  Switching between modes requires to first leave the current operation mode in order to then enter the new one: # ip -force link set dev em1 xdpgeneric obj prog.o # ip -force link set dev em1 xdpgeneric off # ip -force link set dev em1 xdpoffload obj prog.o # ip l [...] 6: em1: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 xdpoffload qdisc mq state UP mode DORMANT group default qlen 1000 link/ether be:08:4d:b6:85:65 brd ff:ff:ff:ff:ff:ff prog/xdp id 17 tag 57cd311f2e27366b [...] # ip -force link set dev em1 xdpoffload off  2. Loading of tc BPF object files. Given a BPF object file prog.o has been compiled for tc, it can be loaded through the tc command to a netdevice. Unlike XDP, there is no driver dependency for supporting attaching BPF programs to the device. Here, the netdevice is called em1, and with the following command the program can be attached to the networking ingress path of em1: # tc qdisc add dev em1 clsact # tc filter add dev em1 ingress bpf da obj prog.o  The first step is to set up a clsact qdisc (Linux queueing discipline). clsact is a dummy qdisc similar to the ingress qdisc, which can only hold classifier and actions, but does not perform actual queueing. It is needed in order to attach the bpf classifier. The clsact qdisc provides two special hooks called ingress and egress, where the classifier can be attached to. Both ingress and egress hooks are located in central receive and transmit locations in the networking data path, where every packet on the device passes through. The ingress hook is called from __netif_receive_skb_core() -> sch_handle_ingress() in the kernel and the egress hook from __dev_queue_xmit() -> sch_handle_egress(). The equivalent for attaching the program to the egress hook looks as follows: # tc filter add dev em1 egress bpf da obj prog.o  The clsact qdisc is processed lockless from ingress and egress direction and can also be attached to virtual, queue-less devices such as veth devices connecting containers. Next to the hook, the tc filter command selects bpf to be used in da (direct-action) mode. da mode is recommended and should always be specified. It basically means that the bpf classifier does not need to call into external tc action modules, which are not necessary for bpf anyway, since all packet mangling, forwarding or other kind of actions can already be performed inside the single BPF program which is to be attached, and is therefore significantly faster. At this point, the program has been attached and is executed once packets traverse the device. Like in XDP, should the default section name not be used, then it can be specified during load, for example, in case of section foobar: # tc filter add dev em1 egress bpf da obj prog.o sec foobar  iproute2’s BPF loader allows for using the same command line syntax across program types, hence the obj prog.o sec foobar is the same syntax as with XDP mentioned earlier. The attached programs can be listed through the following commands: # tc filter show dev em1 ingress filter protocol all pref 49152 bpf filter protocol all pref 49152 bpf handle 0x1 prog.o:[ingress] direct-action id 1 tag c5f7825e5dac396f # tc filter show dev em1 egress filter protocol all pref 49152 bpf filter protocol all pref 49152 bpf handle 0x1 prog.o:[egress] direct-action id 2 tag b2fd5adc0f262714  The output of prog.o:[ingress] tells that program section ingress was loaded from the file prog.o, and bpf operates in direct-action mode. The program id and tag is appended for each case, where the latter denotes a hash over the instruction stream which can be correlated with the object file or perf reports with stack traces, etc. Last but not least, the id represents the system-wide unique BPF program identifier that can be used along with bpftool to further inspect or dump the attached BPF program. tc can attach more than just a single BPF program, it provides various other classifiers which can be chained together. However, attaching a single BPF program is fully sufficient since all packet operations can be contained in the program itself thanks to da (direct-action) mode, meaning the BPF program itself will already return the tc action verdict such as TC_ACT_OK, TC_ACT_SHOT and others. For optimal performance and flexibility, this is the recommended usage. In the above show command, tc also displays pref 49152 and handle 0x1 next to the BPF related output. Both are auto-generated in case they are not explicitly provided through the command line. pref denotes a priority number, which means that in case multiple classifiers are attached, they will be executed based on ascending priority, and handle represents an identifier in case multiple instances of the same classifier have been loaded under the same pref. Since in case of BPF, a single program is fully sufficient, pref and handle can typically be ignored. Only in the case where it is planned to atomically replace the attached BPF programs, it would be recommended to explicitly specify pref and handle a priori on initial load, so that they do not have to be queried at a later point in time for the replace operation. Thus, creation becomes: # tc filter add dev em1 ingress pref 1 handle 1 bpf da obj prog.o sec foobar # tc filter show dev em1 ingress filter protocol all pref 1 bpf filter protocol all pref 1 bpf handle 0x1 prog.o:[foobar] direct-action id 1 tag c5f7825e5dac396f  And for the atomic replacement, the following can be issued for updating the existing program at ingress hook with the new BPF program from the file prog.o in section foobar: # tc filter replace dev em1 ingress pref 1 handle 1 bpf da obj prog.o sec foobar  Last but not least, in order to remove all attached programs from the ingress respectively egress hook, the following can be used: # tc filter del dev em1 ingress # tc filter del dev em1 egress  For removing the entire clsact qdisc from the netdevice, which implicitly also removes all attached programs from the ingress and egress hooks, the below command is provided: # tc qdisc del dev em1 clsact  tc BPF programs can also be offloaded if the NIC and driver has support for it similarly as with XDP BPF programs. Netronome’s nfp supported NICs offer both types of BPF offload. # tc qdisc add dev em1 clsact # tc filter replace dev em1 ingress pref 1 handle 1 bpf skip_sw da obj prog.o Error: TC offload is disabled on net device. We have an error talking to the kernel  If the above error is shown, then tc hardware offload first needs to be enabled for the device through ethtool’s hw-tc-offload setting: # ethtool -K em1 hw-tc-offload on # tc qdisc add dev em1 clsact # tc filter replace dev em1 ingress pref 1 handle 1 bpf skip_sw da obj prog.o # tc filter show dev em1 ingress filter protocol all pref 1 bpf filter protocol all pref 1 bpf handle 0x1 prog.o:[classifier] direct-action skip_sw in_hw id 19 tag 57cd311f2e27366b  The in_hw flag confirms that the program has been offloaded to the NIC. Note that BPF offloads for both tc and XDP cannot be loaded at the same time, either the tc or XDP offload option must be selected. 3. Testing BPF offload interface via netdevsim driver. The netdevsim driver which is part of the Linux kernel provides a dummy driver which implements offload interfaces for XDP BPF and tc BPF programs and facilitates testing kernel changes or low-level user space programs implementing a control plane directly against the kernel’s UAPI. A netdevsim device can be created as follows: # modprobe netdevsim // [ID] [PORT_COUNT] # echo "1 1" > /sys/bus/netdevsim/new_device # devlink dev netdevsim/netdevsim1 # devlink port netdevsim/netdevsim1/0: type eth netdev eth0 flavour physical # ip l [...] 4: eth0: <BROADCAST,NOARP,UP,LOWER_UP> mtu 1500 qdisc noqueue state UNKNOWN mode DEFAULT group default qlen 1000 link/ether 2a:d5:cd:08:d1:3f brd ff:ff:ff:ff:ff:ff  After that step, XDP BPF or tc BPF programs can be test loaded as shown in the various examples earlier: # ip -force link set dev eth0 xdpoffload obj prog.o # ip l [...] 4: eth0: <BROADCAST,NOARP,UP,LOWER_UP> mtu 1500 xdpoffload qdisc noqueue state UNKNOWN mode DEFAULT group default qlen 1000 link/ether 2a:d5:cd:08:d1:3f brd ff:ff:ff:ff:ff:ff prog/xdp id 16 tag a04f5eef06a7f555  These two workflows are the basic operations to load XDP BPF respectively tc BPF programs with iproute2. There are other various advanced options for the BPF loader that apply both to XDP and tc, some of them are listed here. In the examples only XDP is presented for simplicity. 1. Verbose log output even on success. The option verb can be appended for loading programs in order to dump the verifier log, even if no error occurred: # ip link set dev em1 xdp obj xdp-example.o verb Prog section 'prog' loaded (5)! - Type: 6 - Instructions: 2 (0 over limit) - License: GPL Verifier analysis: 0: (b7) r0 = 1 1: (95) exit processed 2 insns  2. Load program that is already pinned in BPF file system. Instead of loading a program from an object file, iproute2 can also retrieve the program from the BPF file system in case some external entity pinned it there and attach it to the device: # ip link set dev em1 xdp pinned /sys/fs/bpf/prog  iproute2 can also use the short form that is relative to the detected mount point of the BPF file system: # ip link set dev em1 xdp pinned m:prog  When loading BPF programs, iproute2 will automatically detect the mounted file system instance in order to perform pinning of nodes. In case no mounted BPF file system instance was found, then tc will automatically mount it to the default location under /sys/fs/bpf/. In case an instance has already been found, then it will be used and no additional mount will be performed: # mkdir /var/run/bpf # mount --bind /var/run/bpf /var/run/bpf # mount -t bpf bpf /var/run/bpf # tc filter add dev em1 ingress bpf da obj tc-example.o sec prog # tree /var/run/bpf /var/run/bpf +-- ip -> /run/bpf/tc/ +-- tc | +-- globals | +-- jmp_map +-- xdp -> /run/bpf/tc/ 4 directories, 1 file  By default tc will create an initial directory structure as shown above, where all subsystem users will point to the same location through symbolic links for the globals namespace, so that pinned BPF maps can be reused among various BPF program types in iproute2. In case the file system instance has already been mounted and an existing structure already exists, then tc will not override it. This could be the case for separating lwt, tc and xdp maps in order to not share globals among all. As briefly covered in the previous LLVM section, iproute2 will install a header file upon installation which can be included through the standard include path by BPF programs: #include <iproute2/bpf_elf.h>  The purpose of this header file is to provide an API for maps and default section names used by programs. It’s a stable contract between iproute2 and BPF programs. The map definition for iproute2 is struct bpf_elf_map. Its members have been covered earlier in the LLVM section of this document. When parsing the BPF object file, the iproute2 loader will walk through all ELF sections. It initially fetches ancillary sections like maps and license. For maps, the struct bpf_elf_map array will be checked for validity and whenever needed, compatibility workarounds are performed. Subsequently all maps are created with the user provided information, either retrieved as a pinned object, or newly created and then pinned into the BPF file system. Next the loader will handle all program sections that contain ELF relocation entries for maps, meaning that BPF instructions loading map file descriptors into registers are rewritten so that the corresponding map file descriptors are encoded into the instructions immediate value, in order for the kernel to be able to convert them later on into map kernel pointers. After that all the programs themselves are created through the BPF system call, and tail called maps, if present, updated with the program’s file descriptors. ### bpftool¶ bpftool is the main introspection and debugging tool around BPF and developed and shipped along with the Linux kernel tree under tools/bpf/bpftool/. The tool can dump all BPF programs and maps that are currently loaded in the system, or list and correlate all BPF maps used by a specific program. Furthermore, it allows to dump the entire map’s key / value pairs, or lookup, update, delete individual ones as well as retrieve a key’s neighbor key in the map. Such operations can be performed based on BPF program or map IDs or by specifying the location of a BPF file system pinned program or map. The tool additionally also offers an option to pin maps or programs into the BPF file system. For a quick overview of all BPF programs currently loaded on the host invoke the following command: # bpftool prog 398: sched_cls tag 56207908be8ad877 loaded_at Apr 09/16:24 uid 0 xlated 8800B jited 6184B memlock 12288B map_ids 18,5,17,14 399: sched_cls tag abc95fb4835a6ec9 loaded_at Apr 09/16:24 uid 0 xlated 344B jited 223B memlock 4096B map_ids 18 400: sched_cls tag afd2e542b30ff3ec loaded_at Apr 09/16:24 uid 0 xlated 1720B jited 1001B memlock 4096B map_ids 17 401: sched_cls tag 2dbbd74ee5d51cc8 loaded_at Apr 09/16:24 uid 0 xlated 3728B jited 2099B memlock 4096B map_ids 17 [...]  Similarly, to get an overview of all active maps: # bpftool map 5: hash flags 0x0 key 20B value 112B max_entries 65535 memlock 13111296B 6: hash flags 0x0 key 20B value 20B max_entries 65536 memlock 7344128B 7: hash flags 0x0 key 10B value 16B max_entries 8192 memlock 790528B 8: hash flags 0x0 key 22B value 28B max_entries 8192 memlock 987136B 9: hash flags 0x0 key 20B value 8B max_entries 512000 memlock 49352704B [...]  Note that for each command, bpftool also supports json based output by appending --json at the end of the command line. An additional --pretty improves the output to be more human readable. # bpftool prog --json --pretty  For dumping the post-verifier BPF instruction image of a specific BPF program, one starting point could be to inspect a specific program, e.g. attached to the tc ingress hook: # tc filter show dev cilium_host egress filter protocol all pref 1 bpf chain 0 filter protocol all pref 1 bpf chain 0 handle 0x1 bpf_host.o:[from-netdev] \ direct-action not_in_hw id 406 tag e0362f5bd9163a0a jited  The program from the object file bpf_host.o, section from-netdev has a BPF program ID of 406 as denoted in id 406. Based on this information bpftool can provide some high-level metadata specific to the program: # bpftool prog show id 406 406: sched_cls tag e0362f5bd9163a0a loaded_at Apr 09/16:24 uid 0 xlated 11144B jited 7721B memlock 12288B map_ids 18,20,8,5,6,14  The program of ID 406 is of type sched_cls (BPF_PROG_TYPE_SCHED_CLS), has a tag of e0362f5bd9163a0a (SHA sum over the instruction sequence), it was loaded by root uid 0 on Apr 09/16:24. The BPF instruction sequence is 11,144 bytes long and the JITed image 7,721 bytes. The program itself (excluding maps) consumes 12,288 bytes that are accounted / charged against user uid 0. And the BPF program uses the BPF maps with IDs 18, 20, 8, 5, 6 and 14. The latter IDs can further be used to get information or dump the map themselves. Additionally, bpftool can issue a dump request of the BPF instructions the program runs: # bpftool prog dump xlated id 406 0: (b7) r7 = 0 1: (63) *(u32 *)(r1 +60) = r7 2: (63) *(u32 *)(r1 +56) = r7 3: (63) *(u32 *)(r1 +52) = r7 [...] 47: (bf) r4 = r10 48: (07) r4 += -40 49: (79) r6 = *(u64 *)(r10 -104) 50: (bf) r1 = r6 51: (18) r2 = map[id:18] <-- BPF map id 18 53: (b7) r5 = 32 54: (85) call bpf_skb_event_output#5656112 <-- BPF helper call 55: (69) r1 = *(u16 *)(r6 +192) [...]  bpftool correlates BPF map IDs into the instruction stream as shown above as well as calls to BPF helpers or other BPF programs. The instruction dump reuses the same ‘pretty-printer’ as the kernel’s BPF verifier. Since the program was JITed and therefore the actual JIT image that was generated out of above xlated instructions is executed, it can be dumped as well through bpftool: # bpftool prog dump jited id 406 0: push %rbp 1: mov %rsp,%rbp 4: sub0x228,%rsp
b:        sub    $0x28,%rbp f: mov %rbx,0x0(%rbp) 13: mov %r13,0x8(%rbp) 17: mov %r14,0x10(%rbp) 1b: mov %r15,0x18(%rbp) 1f: xor %eax,%eax 21: mov %rax,0x20(%rbp) 25: mov 0x80(%rdi),%r9d [...]  Mainly for BPF JIT developers, the option also exists to interleave the disassembly with the actual native opcodes: # bpftool prog dump jited id 406 opcodes 0: push %rbp 55 1: mov %rsp,%rbp 48 89 e5 4: sub$0x228,%rsp
48 81 ec 28 02 00 00
b:        sub    $0x28,%rbp 48 83 ed 28 f: mov %rbx,0x0(%rbp) 48 89 5d 00 13: mov %r13,0x8(%rbp) 4c 89 6d 08 17: mov %r14,0x10(%rbp) 4c 89 75 10 1b: mov %r15,0x18(%rbp) 4c 89 7d 18 [...]  The same interleaving can be done for the normal BPF instructions which can sometimes be useful for debugging in the kernel: # bpftool prog dump xlated id 406 opcodes 0: (b7) r7 = 0 b7 07 00 00 00 00 00 00 1: (63) *(u32 *)(r1 +60) = r7 63 71 3c 00 00 00 00 00 2: (63) *(u32 *)(r1 +56) = r7 63 71 38 00 00 00 00 00 3: (63) *(u32 *)(r1 +52) = r7 63 71 34 00 00 00 00 00 4: (63) *(u32 *)(r1 +48) = r7 63 71 30 00 00 00 00 00 5: (63) *(u32 *)(r1 +64) = r7 63 71 40 00 00 00 00 00 [...]  The basic blocks of a program can also be visualized with the help of graphviz. For this purpose bpftool has a visual dump mode that generates a dot file instead of the plain BPF xlated instruction dump that can later be converted to a png file: # bpftool prog dump xlated id 406 visual &> output.dot$ dot -Tpng output.dot -o output.png


Another option would be to pass the dot file to dotty as a viewer, that is dotty output.dot, where the result for the bpf_host.o program looks as follows (small extract):

Note that the xlated instruction dump provides the post-verifier BPF instruction image which means that it dumps the instructions as if they were to be run through the BPF interpreter. In the kernel, the verifier performs various rewrites of the original instructions provided by the BPF loader.

One example of rewrites is the inlining of helper functions in order to improve runtime performance, here in the case of a map lookup for hash tables:

# bpftool prog dump xlated id 3
0: (b7) r1 = 2
1: (63) *(u32 *)(r10 -4) = r1
2: (bf) r2 = r10
3: (07) r2 += -4
4: (18) r1 = map[id:2]                      <-- BPF map id 2
6: (85) call __htab_map_lookup_elem#77408   <-+ BPF helper inlined rewrite
7: (15) if r0 == 0x0 goto pc+2                |
8: (07) r0 += 56                              |
9: (79) r0 = *(u64 *)(r0 +0)                <-+
10: (15) if r0 == 0x0 goto pc+24
11: (bf) r2 = r10
12: (07) r2 += -4
[...]


bpftool correlates calls to helper functions or BPF to BPF calls through kallsyms. Therefore, make sure that JITed BPF programs are exposed to kallsyms (bpf_jit_kallsyms) and that kallsyms addresses are not obfuscated (calls are otherwise shown as call bpf_unspec#0):

# echo 0 > /proc/sys/kernel/kptr_restrict
# echo 1 > /proc/sys/net/core/bpf_jit_kallsyms


BPF to BPF calls are correlated as well for both, interpreter as well as JIT case. In the latter, the tag of the subprogram is shown as call target. In each case, the pc+2 is the pc-relative offset of the call target, which denotes the subprogram.

# bpftool prog dump xlated id 1
0: (85) call pc+2#__bpf_prog_run_args32
1: (b7) r0 = 1
2: (95) exit
3: (b7) r0 = 2
4: (95) exit


JITed variant of the dump:

# bpftool prog dump xlated id 1
0: (85) call pc+2#bpf_prog_3b185187f1855c4c_F
1: (b7) r0 = 1
2: (95) exit
3: (b7) r0 = 2
4: (95) exit


In the case of tail calls, the kernel maps them into a single instruction internally, bpftool will still correlate them as a helper call for ease of debugging:

# bpftool prog dump xlated id 2
[...]
10: (b7) r2 = 8
11: (85) call bpf_trace_printk#-41312
12: (bf) r1 = r6
13: (18) r2 = map[id:1]
15: (b7) r3 = 0
16: (85) call bpf_tail_call#12
17: (b7) r1 = 42
18: (6b) *(u16 *)(r6 +46) = r1
19: (b7) r0 = 0
20: (95) exit

# bpftool map show id 1
1: prog_array  flags 0x0
key 4B  value 4B  max_entries 1  memlock 4096B


Dumping an entire map is possible through the map dump subcommand which iterates through all present map elements and dumps the key / value pairs.

If no BTF (BPF Type Format) data is available for a given map, then the key / value pairs are dumped as hex:

# bpftool map dump id 5
key:
f0 0d 00 00 00 00 00 00  0a 66 00 00 00 00 8a d6
02 00 00 00
value:
00 00 00 00 00 00 00 00  01 00 00 00 00 00 00 00
00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00
00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00
00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00
00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00
00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00
00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00
key:
0a 66 1c ee 00 00 00 00  00 00 00 00 00 00 00 00
01 00 00 00
value:
00 00 00 00 00 00 00 00  01 00 00 00 00 00 00 00
00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00
[...]
Found 6 elements


However, with BTF, the map also holds debugging information about the key and value structures. For example, BTF in combination with BPF maps and the BPF_ANNOTATE_KV_PAIR() macro from iproute2 will result in the following dump (test_xdp_noinline.o from kernel selftests):

# cat tools/testing/selftests/bpf/test_xdp_noinline.c
[...]
struct ctl_value {
union {
__u64 value;
__u32 ifindex;
__u8 mac[6];
};
};

struct bpf_map_def __attribute__ ((section("maps"), used)) ctl_array = {
.type            = BPF_MAP_TYPE_ARRAY,
.key_size        = sizeof(__u32),
.value_size      = sizeof(struct ctl_value),
.max_entries     = 16,
.map_flags       = 0,
};
BPF_ANNOTATE_KV_PAIR(ctl_array, __u32, struct ctl_value);

[...]


The BPF_ANNOTATE_KV_PAIR() macro forces a map-specific ELF section containing an empty key and value, this enables the iproute2 BPF loader to correlate BTF data with that section and thus allows to choose the corresponding types out of the BTF for loading the map.

Compiling through LLVM and generating BTF through debugging information by pahole:

# clang [...] -O2 -target bpf -g -emit-llvm -c test_xdp_noinline.c -o - |
llc -march=bpf -mcpu=probe -mattr=dwarfris -filetype=obj -o test_xdp_noinline.o
# pahole -J test_xdp_noinline.o


Now loading into kernel and dumping the map via bpftool:

# ip -force link set dev lo xdp obj test_xdp_noinline.o sec xdp-test
# ip a
1: lo: <LOOPBACK,UP,LOWER_UP> mtu 65536 xdpgeneric/id:227 qdisc noqueue state UNKNOWN group default qlen 1000
link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00
inet 127.0.0.1/8 scope host lo
valid_lft forever preferred_lft forever
inet6 ::1/128 scope host
valid_lft forever preferred_lft forever
[...]
# bpftool prog show id 227
227: xdp  tag a85e060c275c5616  gpl
loaded_at 2018-07-17T14:41:29+0000  uid 0
xlated 8152B  not jited  memlock 12288B  map_ids 381,385,386,382,384,383
# bpftool map dump id 386
[{
"key": 0,
"value": {
"": {
"value": 0,
"ifindex": 0,
"mac": []
}
}
},{
"key": 1,
"value": {
"": {
"value": 0,
"ifindex": 0,
"mac": []
}
}
},{
[...]


Lookup, update, delete, and ‘get next key’ operations on the map for specific keys can be performed through bpftool as well.

### BPF sysctls¶

The Linux kernel provides few sysctls that are BPF related and covered in this section.

• /proc/sys/net/core/bpf_jit_enable: Enables or disables the BPF JIT compiler.

 Value Description 0 Disable the JIT and use only interpreter (kernel’s default value) 1 Enable the JIT compiler 2 Enable the JIT and emit debugging traces to the kernel log

As described in subsequent sections, bpf_jit_disasm tool can be used to process debugging traces when the JIT compiler is set to debugging mode (option 2).

• /proc/sys/net/core/bpf_jit_harden: Enables or disables BPF JIT hardening. Note that enabling hardening trades off performance, but can mitigate JIT spraying by blinding out the BPF program’s immediate values. For programs processed through the interpreter, blinding of immediate values is not needed / performed.

 Value Description 0 Disable JIT hardening (kernel’s default value) 1 Enable JIT hardening for unprivileged users only 2 Enable JIT hardening for all users
• /proc/sys/net/core/bpf_jit_kallsyms: Enables or disables export of JITed programs as kernel symbols to /proc/kallsyms so that they can be used together with perf tooling as well as making these addresses aware to the kernel for stack unwinding, for example, used in dumping stack traces. The symbol names contain the BPF program tag (bpf_prog_<tag>). If bpf_jit_harden is enabled, then this feature is disabled.

 Value Description 0 Disable JIT kallsyms export (kernel’s default value) 1 Enable JIT kallsyms export for privileged users only
• /proc/sys/kernel/unprivileged_bpf_disabled: Enables or disable unprivileged use of the bpf(2) system call. The Linux kernel has unprivileged use of bpf(2) enabled by default, but once the switch is flipped, unprivileged use will be permanently disabled until the next reboot. This sysctl knob is a one-time switch, meaning if once set, then neither an application nor an admin can reset the value anymore. This knob does not affect any cBPF programs such as seccomp or traditional socket filters that do not use the bpf(2) system call for loading the program into the kernel.

 Value Description 0 Unprivileged use of bpf syscall enabled (kernel’s default value) 1 Unprivileged use of bpf syscall disabled

### Kernel Testing¶

The Linux kernel ships a BPF selftest suite, which can be found in the kernel source tree under tools/testing/selftests/bpf/.

$cd tools/testing/selftests/bpf/$ make
# make run_tests


The test suite contains test cases against the BPF verifier, program tags, various tests against the BPF map interface and map types. It contains various runtime tests from C code for checking LLVM back end, and eBPF as well as cBPF asm code that is run in the kernel for testing the interpreter and JITs.

### JIT Debugging¶

For JIT developers performing audits or writing extensions, each compile run can output the generated JIT image into the kernel log through:

# echo 2 > /proc/sys/net/core/bpf_jit_enable


Whenever a new BPF program is loaded, the JIT compiler will dump the output, which can then be inspected with dmesg, for example:

[ 3389.935842] flen=6 proglen=70 pass=3 image=ffffffffa0069c8f from=tcpdump pid=20583
[ 3389.935847] JIT code: 00000000: 55 48 89 e5 48 83 ec 60 48 89 5d f8 44 8b 4f 68
[ 3389.935849] JIT code: 00000010: 44 2b 4f 6c 4c 8b 87 d8 00 00 00 be 0c 00 00 00
[ 3389.935850] JIT code: 00000020: e8 1d 94 ff e0 3d 00 08 00 00 75 16 be 17 00 00
[ 3389.935851] JIT code: 00000030: 00 e8 28 94 ff e0 83 f8 01 75 07 b8 ff ff 00 00
[ 3389.935852] JIT code: 00000040: eb 02 31 c0 c9 c3


flen is the length of the BPF program (here, 6 BPF instructions), and proglen tells the number of bytes generated by the JIT for the opcode image (here, 70 bytes in size). pass means that the image was generated in 3 compiler passes, for example, x86_64 can have various optimization passes to further reduce the image size when possible. image contains the address of the generated JIT image, from and pid the user space application name and PID respectively, which triggered the compilation process. The dump output for eBPF and cBPF JITs is the same format.

In the kernel tree under tools/bpf/, there is a tool called bpf_jit_disasm. It reads out the latest dump and prints the disassembly for further inspection:

# ./bpf_jit_disasm
70 bytes emitted from JIT compiler (pass:3, flen:6)
ffffffffa0069c8f + <x>:
0:       push   %rbp
1:       mov    %rsp,%rbp
4:       sub    $0x60,%rsp 8: mov %rbx,-0x8(%rbp) c: mov 0x68(%rdi),%r9d 10: sub 0x6c(%rdi),%r9d 14: mov 0xd8(%rdi),%r8 1b: mov$0xc,%esi
20:       callq  0xffffffffe0ff9442
25:       cmp    $0x800,%eax 2a: jne 0x0000000000000042 2c: mov$0x17,%esi
31:       callq  0xffffffffe0ff945e
36:       cmp    $0x1,%eax 39: jne 0x0000000000000042 3b: mov$0xffff,%eax
40:       jmp    0x0000000000000044
42:       xor    %eax,%eax
44:       leaveq
45:       retq


Alternatively, the tool can also dump related opcodes along with the disassembly.

# ./bpf_jit_disasm -o
70 bytes emitted from JIT compiler (pass:3, flen:6)
ffffffffa0069c8f + <x>:
0:       push   %rbp
55
1:       mov    %rsp,%rbp
48 89 e5
4:       sub    $0x60,%rsp 48 83 ec 60 8: mov %rbx,-0x8(%rbp) 48 89 5d f8 c: mov 0x68(%rdi),%r9d 44 8b 4f 68 10: sub 0x6c(%rdi),%r9d 44 2b 4f 6c 14: mov 0xd8(%rdi),%r8 4c 8b 87 d8 00 00 00 1b: mov$0xc,%esi
be 0c 00 00 00
20:       callq  0xffffffffe0ff9442
e8 1d 94 ff e0
25:       cmp    $0x800,%eax 3d 00 08 00 00 2a: jne 0x0000000000000042 75 16 2c: mov$0x17,%esi
be 17 00 00 00
31:       callq  0xffffffffe0ff945e
e8 28 94 ff e0
36:       cmp    $0x1,%eax 83 f8 01 39: jne 0x0000000000000042 75 07 3b: mov$0xffff,%eax
b8 ff ff 00 00
40:       jmp    0x0000000000000044
eb 02
42:       xor    %eax,%eax
31 c0
44:       leaveq
c9
45:       retq
c3


More recently, bpftool adapted the same feature of dumping the BPF JIT image based on a given BPF program ID already loaded in the system (see bpftool section).

For performance analysis of JITed BPF programs, perf can be used as usual. As a prerequisite, JITed programs need to be exported through kallsyms infrastructure.

# echo 1 > /proc/sys/net/core/bpf_jit_enable
# echo 1 > /proc/sys/net/core/bpf_jit_kallsyms


Enabling or disabling bpf_jit_kallsyms does not require a reload of the related BPF programs. Next, a small workflow example is provided for profiling BPF programs. A crafted tc BPF program is used for demonstration purposes, where perf records a failed allocation inside bpf_clone_redirect() helper. Due to the use of direct write, bpf_try_make_head_writable() failed, which would then release the cloned skb again and return with an error message. perf thus records all kfree_skb events.

# tc qdisc add dev em1 clsact
# tc filter add dev em1 ingress bpf da obj prog.o sec main
# tc filter show dev em1 ingress
filter protocol all pref 49152 bpf
filter protocol all pref 49152 bpf handle 0x1 prog.o:[main] direct-action id 1 tag 8227addf251b7543

# cat /proc/kallsyms
[...]
ffffffffc00349e0 t fjes_hw_init_command_registers    [fjes]
ffffffffc003e2e0 d __tracepoint_fjes_hw_stop_debug_err    [fjes]
ffffffffc0036190 t fjes_hw_epbuf_tx_pkt_send    [fjes]
ffffffffc004b000 t bpf_prog_8227addf251b7543

# perf record -a -g -e skb:kfree_skb sleep 60
# perf script --kallsyms=/proc/kallsyms
[...]
ksoftirqd/0     6 [000]  1004.578402:    skb:kfree_skb: skbaddr=0xffff9d4161f20a00 protocol=2048 location=0xffffffffc004b52c
7fffb8745961 bpf_clone_redirect (/lib/modules/4.10.0+/build/vmlinux)
7fffc004e52c bpf_prog_8227addf251b7543 (/lib/modules/4.10.0+/build/vmlinux)
7fffc05b6283 cls_bpf_classify (/lib/modules/4.10.0+/build/vmlinux)
7fffb875957a tc_classify (/lib/modules/4.10.0+/build/vmlinux)
7fffb8729840 __netif_receive_skb_core (/lib/modules/4.10.0+/build/vmlinux)
7fffb8729e38 __netif_receive_skb (/lib/modules/4.10.0+/build/vmlinux)
7fffb872ae05 process_backlog (/lib/modules/4.10.0+/build/vmlinux)
7fffb872a43e net_rx_action (/lib/modules/4.10.0+/build/vmlinux)
7fffb886176c __do_softirq (/lib/modules/4.10.0+/build/vmlinux)
7fffb80ac5b9 run_ksoftirqd (/lib/modules/4.10.0+/build/vmlinux)
7fffb80ca7fa smpboot_thread_fn (/lib/modules/4.10.0+/build/vmlinux)
7fffb80c6831 kthread (/lib/modules/4.10.0+/build/vmlinux)
7fffb885e09c ret_from_fork (/lib/modules/4.10.0+/build/vmlinux)


The stack trace recorded by perf will then show the bpf_prog_8227addf251b7543() symbol as part of the call trace, meaning that the BPF program with the tag 8227addf251b7543 was related to the kfree_skb event, and such program was attached to netdevice em1 on the ingress hook as shown by tc.

### Introspection¶

The Linux kernel provides various tracepoints around BPF and XDP which can be used for additional introspection, for example, to trace interactions of user space programs with the bpf system call.

Tracepoints for BPF:

# perf list | grep bpf:
bpf:bpf_map_create                                 [Tracepoint event]
bpf:bpf_map_delete_elem                            [Tracepoint event]
bpf:bpf_map_lookup_elem                            [Tracepoint event]
bpf:bpf_map_next_key                               [Tracepoint event]
bpf:bpf_map_update_elem                            [Tracepoint event]
bpf:bpf_obj_get_map                                [Tracepoint event]
bpf:bpf_obj_get_prog                               [Tracepoint event]
bpf:bpf_obj_pin_map                                [Tracepoint event]
bpf:bpf_obj_pin_prog                               [Tracepoint event]
bpf:bpf_prog_get_type                              [Tracepoint event]
bpf:bpf_prog_load                                  [Tracepoint event]
bpf:bpf_prog_put_rcu                               [Tracepoint event]


Example usage with perf (alternatively to sleep example used here, a specific application like tc could be used here instead, of course):

# perf record -a -e bpf:* sleep 10
# perf script
sock_example  6197 [005]   283.980322:      bpf:bpf_map_create: map type=ARRAY ufd=4 key=4 val=8 max=256 flags=0
sock_example  6197 [005]   283.980721:       bpf:bpf_prog_load: prog=a5ea8fa30ea6849c type=SOCKET_FILTER ufd=5
sock_example  6197 [005]   283.988423:   bpf:bpf_prog_get_type: prog=a5ea8fa30ea6849c type=SOCKET_FILTER
sock_example  6197 [005]   283.988443: bpf:bpf_map_lookup_elem: map type=ARRAY ufd=4 key=[06 00 00 00] val=[00 00 00 00 00 00 00 00]
[...]
sock_example  6197 [005]   288.990868: bpf:bpf_map_lookup_elem: map type=ARRAY ufd=4 key=[01 00 00 00] val=[14 00 00 00 00 00 00 00]
swapper     0 [005]   289.338243:    bpf:bpf_prog_put_rcu: prog=a5ea8fa30ea6849c type=SOCKET_FILTER


For the BPF programs, their individual program tag is displayed.

For debugging, XDP also has a tracepoint that is triggered when exceptions are raised:

# perf list | grep xdp:
xdp:xdp_exception                                  [Tracepoint event]


Exceptions are triggered in the following scenarios:

• The BPF program returned an invalid / unknown XDP action code.
• The BPF program returned with XDP_ABORTED indicating a non-graceful exit.
• The BPF program returned with XDP_TX, but there was an error on transmit, for example, due to the port not being up, due to the transmit ring being full, due to allocation failures, etc.

Both tracepoint classes can also be inspected with a BPF program itself attached to one or more tracepoints, collecting further information in a map or punting such events to a user space collector through the bpf_perf_event_output() helper, for example.

### Tracing pipe¶

When a BPF program makes a call to bpf_trace_printk(), the output is sent to the kernel tracing pipe. Users may read from this file to consume events that are traced to this buffer:

# tail -f /sys/kernel/debug/tracing/trace_pipe
...


### Miscellaneous¶

BPF programs and maps are memory accounted against RLIMIT_MEMLOCK similar to perf. The currently available size in unit of system pages which may be locked into memory can be inspected through ulimit -l. The setrlimit system call man page provides further details.

The default limit is usually insufficient to load more complex programs or larger BPF maps, so that the BPF system call will return with errno of EPERM. In such situations a workaround with ulimit -l unlimited or with a sufficiently large limit could be performed. The RLIMIT_MEMLOCK is mainly enforcing limits for unprivileged users. Depending on the setup, setting a higher limit for privileged users is often acceptable.

## Program Types¶

At the time of this writing, there are eighteen different BPF program types available, two of the main types for networking are further explained in below subsections, namely XDP BPF programs as well as tc BPF programs. Extensive usage examples for the two program types for LLVM, iproute2 or other tools are spread throughout the toolchain section and not covered here. Instead, this section focuses on their architecture, concepts and use cases.

### XDP¶

XDP stands for eXpress Data Path and provides a framework for BPF that enables high-performance programmable packet processing in the Linux kernel. It runs the BPF program at the earliest possible point in software, namely at the moment the network driver receives the packet.

At this point in the fast-path the driver just picked up the packet from its receive rings, without having done any expensive operations such as allocating an skb for pushing the packet further up the networking stack, without having pushed the packet into the GRO engine, etc. Thus, the XDP BPF program is executed at the earliest point when it becomes available to the CPU for processing.

XDP works in concert with the Linux kernel and its infrastructure, meaning the kernel is not bypassed as in various networking frameworks that operate in user space only. Keeping the packet in kernel space has several major advantages:

• XDP is able to reuse all the upstream developed kernel networking drivers, user space tooling, or even other available in-kernel infrastructure such as routing tables, sockets, etc in BPF helper calls itself.
• Residing in kernel space, XDP has the same security model as the rest of the kernel for accessing hardware.
• There is no need for crossing kernel / user space boundaries since the processed packet already resides in the kernel and can therefore flexibly forward packets into other in-kernel entities like namespaces used by containers or the kernel’s networking stack itself. This is particularly relevant in times of Meltdown and Spectre.
• Punting packets from XDP to the kernel’s robust, widely used and efficient TCP/IP stack is trivially possible, allows for full reuse and does not require maintaining a separate TCP/IP stack as with user space frameworks.
• The use of BPF allows for full programmability, keeping a stable ABI with the same ‘never-break-user-space’ guarantees as with the kernel’s system call ABI and compared to modules it also provides safety measures thanks to the BPF verifier that ensures the stability of the kernel’s operation.
• XDP trivially allows for atomically swapping programs during runtime without any network traffic interruption or even kernel / system reboot.
• XDP allows for flexible structuring of workloads integrated into the kernel. For example, it can operate in “busy polling” or “interrupt driven” mode. Explicitly dedicating CPUs to XDP is not required. There are no special hardware requirements and it does not rely on hugepages.
• XDP does not require any third party kernel modules or licensing. It is a long-term architectural solution, a core part of the Linux kernel, and developed by the kernel community.
• XDP is already enabled and shipped everywhere with major distributions running a kernel equivalent to 4.8 or higher and supports most major 10G or higher networking drivers.

As a framework for running BPF in the driver, XDP additionally ensures that packets are laid out linearly and fit into a single DMA’ed page which is readable and writable by the BPF program. XDP also ensures that additional headroom of 256 bytes is available to the program for implementing custom encapsulation headers with the help of the bpf_xdp_adjust_head() BPF helper or adding custom metadata in front of the packet through bpf_xdp_adjust_meta().

The framework contains XDP action codes further described in the section below which a BPF program can return in order to instruct the driver how to proceed with the packet, and it enables the possibility to atomically replace BPF programs running at the XDP layer. XDP is tailored for high-performance by design. BPF allows to access the packet data through ‘direct packet access’ which means that the program holds data pointers directly in registers, loads the content into registers, respectively writes from there into the packet.

The packet representation in XDP that is passed to the BPF program as the BPF context looks as follows:

struct xdp_buff {
void *data;
void *data_end;
void *data_meta;
void *data_hard_start;
struct xdp_rxq_info *rxq;
};


data points to the start of the packet data in the page, and as the name suggests, data_end points to the end of the packet data. Since XDP allows for a headroom, data_hard_start points to the maximum possible headroom start in the page, meaning, when the packet should be encapsulated, then data is moved closer towards data_hard_start via bpf_xdp_adjust_head(). The same BPF helper function also allows for decapsulation in which case data is moved further away from data_hard_start.

data_meta initially points to the same location as data but bpf_xdp_adjust_meta() is able to move the pointer towards data_hard_start as well in order to provide room for custom metadata which is invisible to the normal kernel networking stack but can be read by tc BPF programs since it is transferred from XDP to the skb. Vice versa, it can remove or reduce the size of the custom metadata through the same BPF helper function by moving data_meta away from data_hard_start again. data_meta can also be used solely for passing state between tail calls similarly to the skb->cb[] control block case that is accessible in tc BPF programs.

This gives the following relation respectively invariant for the struct xdp_buff packet pointers: data_hard_start <= data_meta <= data < data_end.

The rxq field points to some additional per receive queue metadata which is populated at ring setup time (not at XDP runtime):

struct xdp_rxq_info {
struct net_device *dev;
u32 queue_index;
u32 reg_state;
} ____cacheline_aligned;


The BPF program can retrieve queue_index as well as additional data from the netdevice itself such as ifindex, etc.

BPF program return codes

After running the XDP BPF program, a verdict is returned from the program in order to tell the driver how to process the packet next. In the linux/bpf.h system header file all available return verdicts are enumerated:

enum xdp_action {
XDP_ABORTED = 0,
XDP_DROP,
XDP_PASS,
XDP_TX,
XDP_REDIRECT,
};


XDP_DROP as the name suggests will drop the packet right at the driver level without wasting any further resources. This is in particular useful for BPF programs implementing DDoS mitigation mechanisms or firewalling in general. The XDP_PASS return code means that the packet is allowed to be passed up to the kernel’s networking stack. Meaning, the current CPU that was processing this packet now allocates a skb, populates it, and passes it onwards into the GRO engine. This would be equivalent to the default packet handling behavior without XDP. With XDP_TX the BPF program has an efficient option to transmit the network packet out of the same NIC it just arrived on again. This is typically useful when few nodes are implementing, for example, firewalling with subsequent load balancing in a cluster and thus act as a hairpinned load balancer pushing the incoming packets back into the switch after rewriting them in XDP BPF. XDP_REDIRECT is similar to XDP_TX in that it is able to transmit the XDP packet, but through another NIC. Another option for the XDP_REDIRECT case is to redirect into a BPF cpumap, meaning, the CPUs serving XDP on the NIC’s receive queues can continue to do so and push the packet for processing the upper kernel stack to a remote CPU. This is similar to XDP_PASS, but with the ability that the XDP BPF program can keep serving the incoming high load as opposed to temporarily spend work on the current packet for pushing into upper layers. Last but not least, XDP_ABORTED which serves denoting an exception like state from the program and has the same behavior as XDP_DROP only that XDP_ABORTED passes the trace_xdp_exception tracepoint which can be additionally monitored to detect misbehavior.

Use cases for XDP

Some of the main use cases for XDP are presented in this subsection. The list is non-exhaustive and given the programmability and efficiency XDP and BPF enables, it can easily be adapted to solve very specific use cases.

• DDoS mitigation, firewalling

One of the basic XDP BPF features is to tell the driver to drop a packet with XDP_DROP at this early stage which allows for any kind of efficient network policy enforcement with having an extremely low per-packet cost. This is ideal in situations when needing to cope with any sort of DDoS attacks, but also more general allows to implement any sort of firewalling policies with close to no overhead in BPF e.g. in either case as stand alone appliance (e.g. scrubbing ‘clean’ traffic through XDP_TX) or widely deployed on nodes protecting end hosts themselves (via XDP_PASS or cpumap XDP_REDIRECT for good traffic). Offloaded XDP takes this even one step further by moving the already small per-packet cost entirely into the NIC with processing at line-rate.

• Forwarding and load-balancing

Another major use case of XDP is packet forwarding and load-balancing through either XDP_TX or XDP_REDIRECT actions. The packet can be arbitrarily mangled by the BPF program running in the XDP layer, even BPF helper functions are available for increasing or decreasing the packet’s headroom in order to arbitrarily encapsulate respectively decapsulate the packet before sending it out again. With XDP_TX hairpinned load-balancers can be implemented that push the packet out of the same networking device it originally arrived on, or with the XDP_REDIRECT action it can be forwarded to another NIC for transmission. The latter return code can also be used in combination with BPF’s cpumap to load-balance packets for passing up the local stack, but on remote, non-XDP processing CPUs.

• Pre-stack filtering / processing

Besides policy enforcement, XDP can also be used for hardening the kernel’s networking stack with the help of XDP_DROP case, meaning, it can drop irrelevant packets for a local node right at the earliest possible point before the networking stack sees them e.g. given we know that a node only serves TCP traffic, any UDP, SCTP or other L4 traffic can be dropped right away. This has the advantage that packets do not need to traverse various entities like GRO engine, the kernel’s flow dissector and others before it can be determined to drop them and thus this allows for reducing the kernel’s attack surface. Thanks to XDP’s early processing stage, this effectively ‘pretends’ to the kernel’s networking stack that these packets have never been seen by the networking device. Additionally, if a potential bug in the stack’s receive path got uncovered and would cause a ‘ping of death’ like scenario, XDP can be utilized to drop such packets right away without having to reboot the kernel or restart any services. Due to the ability to atomically swap such programs to enforce a drop of bad packets, no network traffic is even interrupted on a host.

Another use case for pre-stack processing is that given the kernel has not yet allocated an skb for the packet, the BPF program is free to modify the packet and, again, have it ‘pretend’ to the stack that it was received by the networking device this way. This allows for cases such as having custom packet mangling and encapsulation protocols where the packet can be decapsulated prior to entering GRO aggregation in which GRO otherwise would not be able to perform any sort of aggregation due to not being aware of the custom protocol. XDP also allows to push metadata (non-packet data) in front of the packet. This is ‘invisible’ to the normal kernel stack, can be GRO aggregated (for matching metadata) and later on processed in coordination with a tc ingress BPF program where it has the context of a skb available for e.g. setting various skb fields.

• Flow sampling, monitoring

XDP can also be used for cases such as packet monitoring, sampling or any other network analytics, for example, as part of an intermediate node in the path or on end hosts in combination also with prior mentioned use cases. For complex packet analysis, XDP provides a facility to efficiently push network packets (truncated or with full payload) and custom metadata into a fast lockless per CPU memory mapped ring buffer provided from the Linux perf infrastructure to an user space application. This also allows for cases where only a flow’s initial data can be analyzed and once determined as good traffic having the monitoring bypassed. Thanks to the flexibility brought by BPF, this allows for implementing any sort of custom monitoring or sampling.

One example of XDP BPF production usage is Facebook’s SHIV and Droplet infrastructure which implement their L4 load-balancing and DDoS countermeasures. Migrating their production infrastructure away from netfilter’s IPVS (IP Virtual Server) over to XDP BPF allowed for a 10x speedup compared to their previous IPVS setup. This was first presented at the netdev 2.1 conference:

Another example is the integration of XDP into Cloudflare’s DDoS mitigation pipeline, which originally was using cBPF instead of eBPF for attack signature matching through iptables’ xt_bpf module. Due to use of iptables this caused severe performance problems under attack where a user space bypass solution was deemed necessary but came with drawbacks as well such as needing to busy poll the NIC and expensive packet re-injection into the kernel’s stack. The migration over to eBPF and XDP combined best of both worlds by having high-performance programmable packet processing directly inside the kernel:

XDP operation modes

XDP has three operation modes where ‘native’ XDP is the default mode. When talked about XDP this mode is typically implied.

• Native XDP

This is the default mode where the XDP BPF program is run directly out of the networking driver’s early receive path. Most widespread used NICs for 10G and higher support native XDP already.

• Offloaded XDP

In the offloaded XDP mode the XDP BPF program is directly offloaded into the NIC instead of being executed on the host CPU. Thus, the already extremely low per-packet cost is pushed off the host CPU entirely and executed on the NIC, providing even higher performance than running in native XDP. This offload is typically implemented by SmartNICs containing multi-threaded, multicore flow processors where a in-kernel JIT compiler translates BPF into native instructions for the latter. Drivers supporting offloaded XDP usually also support native XDP for cases where some BPF helpers may not yet or only be available for the native mode.

• Generic XDP

For drivers not implementing native or offloaded XDP yet, the kernel provides an option for generic XDP which does not require any driver changes since run at a much later point out of the networking stack. This setting is primarily targeted at developers who want to write and test programs against the kernel’s XDP API, and will not operate at the performance rate of the native or offloaded modes. For XDP usage in a production environment either the native or offloaded mode is better suited and the recommended way to run XDP.

Driver support

Since BPF and XDP is evolving quickly in terms of feature and driver support, the following lists native and offloaded XDP drivers as of kernel 4.17.

Drivers supporting native XDP

• Broadcom
• bnxt
• Cavium
• thunderx
• Intel
• ixgbe
• ixgbevf
• i40e
• Mellanox
• mlx4
• mlx5
• Netronome
• nfp
• Others
• tun
• virtio_net
• Qlogic
• qede
• Solarflare

Drivers supporting offloaded XDP

• Netronome

Note that examples for writing and loading XDP programs are included in the toolchain section under the respective tools.

 [1] XDP for sfc available via out of tree driver as of kernel 4.17, but will be upstreamed soon.
 [2] (1, 2) Some BPF helper functions such as retrieving the current CPU number will not be available in an offloaded setting.

### tc (traffic control)¶

Aside from other program types such as XDP, BPF can also be used out of the kernel’s tc (traffic control) layer in the networking data path. On a high-level there are three major differences when comparing XDP BPF programs to tc BPF ones:

• The BPF input context is a sk_buff not a xdp_buff. When the kernel’s networking stack receives a packet, after the XDP layer, it allocates a buffer and parses the packet to store metadata about the packet. This representation is known as the sk_buff. This structure is then exposed in the BPF input context so that BPF programs from the tc ingress layer can use the metadata that the stack extracts from the packet. This can be useful, but comes with an associated cost of the stack performing this allocation and metadata extraction, and handling the packet until it hits the tc hook. By definition, the xdp_buff doesn’t have access to this metadata because the XDP hook is called before this work is done. This is a significant contributor to the performance difference between the XDP and tc hooks.

Therefore, BPF programs attached to the tc BPF hook can, for instance, read or write the skb’s mark, pkt_type, protocol, priority, queue_mapping, napi_id, cb[] array, hash, tc_classid or tc_index, vlan metadata, the XDP transferred custom metadata and various other information. All members of the struct __sk_buff BPF context used in tc BPF are defined in the linux/bpf.h system header.

Generally, the sk_buff is of a completely different nature than xdp_buff where both come with advantages and disadvantages. For example, the sk_buff case has the advantage that it is rather straight forward to mangle its associated metadata, however, it also contains a lot of protocol specific information (e.g. GSO related state) which makes it difficult to simply switch protocols by solely rewriting the packet data. This is due to the stack processing the packet based on the metadata rather than having the cost of accessing the packet contents each time. Thus, additional conversion is required from BPF helper functions taking care that sk_buff internals are properly converted as well. The xdp_buff case however does not face such issues since it comes at such an early stage where the kernel has not even allocated an sk_buff yet, thus packet rewrites of any kind can be realized trivially. However, the xdp_buff case has the disadvantage that sk_buff metadata is not available for mangling at this stage. The latter is overcome by passing custom metadata from XDP BPF to tc BPF, though. In this way, the limitations of each program type can be overcome by operating complementary programs of both types as the use case requires.

• Compared to XDP, tc BPF programs can be triggered out of ingress and also egress points in the networking data path as opposed to ingress only in the case of XDP.

The two hook points sch_handle_ingress() and sch_handle_egress() in the kernel are triggered out of __netif_receive_skb_core() and __dev_queue_xmit(), respectively. The latter two are the main receive and transmit functions in the data path that, setting XDP aside, are triggered for every network packet going in or coming out of the node allowing for full visibility for tc BPF programs at these hook points.

• The tc BPF programs do not require any driver changes since they are run at hook points in generic layers in the networking stack. Therefore, they can be attached to any type of networking device.

While this provides flexibility, it also trades off performance compared to running at the native XDP layer. However, tc BPF programs still come at the earliest point in the generic kernel’s networking data path after GRO has been run but before any protocol processing, traditional iptables firewalling such as iptables PREROUTING or nftables ingress hooks or other packet processing takes place. Likewise on egress, tc BPF programs execute at the latest point before handing the packet to the driver itself for transmission, meaning after traditional iptables firewalling hooks like iptables POSTROUTING, but still before handing the packet to the kernel’s GSO engine.

One exception which does require driver changes however are offloaded tc BPF programs, typically provided by SmartNICs in a similar way as offloaded XDP just with differing set of features due to the differences in the BPF input context, helper functions and verdict codes.

BPF programs run in the tc layer are run from the cls_bpf classifier. While the tc terminology describes the BPF attachment point as a “classifier”, this is a bit misleading since it under-represents what cls_bpf is capable of. That is to say, a fully programmable packet processor being able not only to read the skb metadata and packet data, but to also arbitrarily mangle both, and terminate the tc processing with an action verdict. cls_bpf can thus be regarded as a self-contained entity that manages and executes tc BPF programs.

cls_bpf can hold one or more tc BPF programs. In the case where Cilium deploys cls_bpf programs, it attaches only a single program for a given hook in direct-action mode. Typically, in the traditional tc scheme, there is a split between classifier and action modules, where the classifier has one or more actions attached to it that are triggered once the classifier has a match. In the modern world for using tc in the software data path this model does not scale well for complex packet processing. Given tc BPF programs attached to cls_bpf are fully self-contained, they effectively fuse the parsing and action process together into a single unit. Thanks to cls_bpf’s direct-action mode, it will just return the tc action verdict and terminate the processing pipeline immediately. This allows for implementing scalable programmable packet processing in the networking data path by avoiding linear iteration of actions. cls_bpf is the only such “classifier” module in the tc layer capable of such a fast-path.

Like XDP BPF programs, tc BPF programs can be atomically updated at runtime via cls_bpf without interrupting any network traffic or having to restart services.

Both the tc ingress and the egress hook where cls_bpf itself can be attached to is managed by a pseudo qdisc called sch_clsact. This is a drop-in replacement and proper superset of the ingress qdisc since it is able to manage both, ingress and egress tc hooks. For tc’s egress hook in __dev_queue_xmit() it is important to stress that it is not executed under the kernel’s qdisc root lock. Thus, both tc ingress and egress hooks are executed in a lockless manner in the fast-path. In either case, preemption is disabled and execution happens under RCU read side.

Typically on egress there are qdiscs attached to netdevices such as sch_mq, sch_fq, sch_fq_codel or sch_htb where some of them are classful qdiscs that contain subclasses and thus require a packet classification mechanism to determine a verdict where to demux the packet. This is handled by a call to tcf_classify() which calls into tc classifiers if present. cls_bpf can also be attached and used in such cases. Such operation usually happens under the qdisc root lock and can be subject to lock contention. The sch_clsact qdisc’s egress hook comes at a much earlier point however which does not fall under that and operates completely independent from conventional egress qdiscs. Thus for cases like sch_htb the sch_clsact qdisc could perform the heavy lifting packet classification through tc BPF outside of the qdisc root lock, setting the skb->mark or skb->priority from there such that sch_htb only requires a flat mapping without expensive packet classification under the root lock thus reducing contention.

Offloaded tc BPF programs are supported for the case of sch_clsact in combination with cls_bpf where the prior loaded BPF program was JITed from a SmartNIC driver to be run natively on the NIC. Only cls_bpf programs operating in direct-action mode are supported to be offloaded. cls_bpf only supports offloading a single program and cannot offload multiple programs. Furthermore only the ingress hook supports offloading BPF programs.

One cls_bpf instance is able to hold multiple tc BPF programs internally. If this is the case, then the TC_ACT_UNSPEC program return code will continue execution with the next tc BPF program in that list. However, this has the drawback that several programs would need to parse the packet over and over again resulting in degraded performance.

BPF program return codes

Both the tc ingress and egress hook share the same action return verdicts that tc BPF programs can use. They are defined in the linux/pkt_cls.h system header:

#define TC_ACT_UNSPEC         (-1)
#define TC_ACT_OK               0
#define TC_ACT_SHOT             2
#define TC_ACT_STOLEN           4
#define TC_ACT_REDIRECT         7


There are a few more action TC_ACT_* verdicts available in the system header file which are also used in the two hooks. However, they share the same semantics with the ones above. Meaning, from a tc BPF perspective, TC_ACT_OK and TC_ACT_RECLASSIFY have the same semantics, as well as the three TC_ACT_STOLEN, TC_ACT_QUEUED and TC_ACT_TRAP opcodes. Therefore, for these cases we only describe TC_ACT_OK and the TC_ACT_STOLEN opcode for the two groups.

Starting out with TC_ACT_UNSPEC. It has the meaning of “unspecified action” and is used in three cases, i) when an offloaded tc BPF program is attached and the tc ingress hook is run where the cls_bpf representation for the offloaded program will return TC_ACT_UNSPEC, ii) in order to continue with the next tc BPF program in cls_bpf for the multi-program case. The latter also works in combination with offloaded tc BPF programs from point i) where the TC_ACT_UNSPEC from there continues with a next tc BPF program solely running in non-offloaded case. Last but not least, iii) TC_ACT_UNSPEC is also used for the single program case to simply tell the kernel to continue with the skb without additional side-effects. TC_ACT_UNSPEC is very similar to the TC_ACT_OK action code in the sense that both pass the skb onwards either to upper layers of the stack on ingress or down to the networking device driver for transmission on egress, respectively. The only difference to TC_ACT_OK is that TC_ACT_OK sets skb->tc_index based on the classid the tc BPF program set. The latter is set out of the tc BPF program itself through skb->tc_classid from the BPF context.

TC_ACT_SHOT instructs the kernel to drop the packet, meaning, upper layers of the networking stack will never see the skb on ingress and similarly the packet will never be submitted for transmission on egress. TC_ACT_SHOT and TC_ACT_STOLEN are both similar in nature with few differences: TC_ACT_SHOT will indicate to the kernel that the skb was released through kfree_skb() and return NET_XMIT_DROP to the callers for immediate feedback, whereas TC_ACT_STOLEN will release the skb through consume_skb() and pretend to upper layers that the transmission was successful through NET_XMIT_SUCCESS. The perf’s drop monitor which records traces of kfree_skb() will therefore also not see any drop indications from TC_ACT_STOLEN since its semantics are such that the skb has been “consumed” or queued but certainly not “dropped”.

Last but not least the TC_ACT_REDIRECT action which is available for tc BPF programs as well. This allows to redirect the skb to the same or another’s device ingress or egress path together with the bpf_redirect() helper. Being able to inject the packet into another device’s ingress or egress direction allows for full flexibility in packet forwarding with BPF. There are no requirements on the target networking device other than being a networking device itself, there is no need to run another instance of cls_bpf on the target device or other such restrictions.

tc BPF FAQ

This section contains a few miscellaneous question and answer pairs related to tc BPF programs that are asked from time to time.

• Question: What about act_bpf as a tc action module, is it still relevant?
• Answer: Not really. Although cls_bpf and act_bpf share the same functionality for tc BPF programs, cls_bpf is more flexible since it is a proper superset of act_bpf. The way tc works is that tc actions need to be attached to tc classifiers. In order to achieve the same flexibility as cls_bpf, act_bpf would need to be attached to the cls_matchall classifier. As the name says, this will match on every packet in order to pass them through for attached tc action processing. For act_bpf, this is will result in less efficient packet processing than using cls_bpf in direct-action mode directly. If act_bpf is used in a setting with other classifiers than cls_bpf or cls_matchall then this will perform even worse due to the nature of operation of tc classifiers. Meaning, if classifier A has a mismatch, then the packet is passed to classifier B, reparsing the packet, etc, thus in the typical case there will be linear processing where the packet would need to traverse N classifiers in the worst case to find a match and execute act_bpf on that. Therefore, act_bpf has never been largely relevant. Additionally, act_bpf does not provide a tc offloading interface either compared to cls_bpf.
• Question: Is it recommended to use cls_bpf not in direct-action mode?
• Answer: No. The answer is similar to the one above in that this is otherwise unable to scale for more complex processing. tc BPF can already do everything needed by itself in an efficient manner and thus there is no need for anything other than direct-action mode.
• Question: Is there any performance difference in offloaded cls_bpf and offloaded XDP?
• Answer: No. Both are JITed through the same compiler in the kernel which handles the offloading to the SmartNIC and the loading mechanism for both is very similar as well. Thus, the BPF program gets translated into the same target instruction set in order to be able to run on the NIC natively. The two tc BPF and XDP BPF program types have a differing set of features, so depending on the use case one might be picked over the other due to availability of certain helper functions in the offload case, for example.

Use cases for tc BPF

Some of the main use cases for tc BPF programs are presented in this subsection. Also here, the list is non-exhaustive and given the programmability and efficiency of tc BPF, it can easily be tailored and integrated into orchestration systems in order to solve very specific use cases. While some use cases with XDP may overlap, tc BPF and XDP BPF are mostly complementary to each other and both can also be used at the same time or one over the other depending which is most suitable for a given problem to solve.

• Policy enforcement for containers

One application which tc BPF programs are suitable for is to implement policy enforcement, custom firewalling or similar security measures for containers or pods, respectively. In the conventional case, container isolation is implemented through network namespaces with veth networking devices connecting the host’s initial namespace with the dedicated container’s namespace. Since one end of the veth pair has been moved into the container’s namespace whereas the other end remains in the initial namespace of the host, all network traffic from the container has to pass through the host-facing veth device allowing for attaching tc BPF programs on the tc ingress and egress hook of the veth. Network traffic going into the container will pass through the host-facing veth’s tc egress hook whereas network traffic coming from the container will pass through the host-facing veth’s tc ingress hook.

For virtual devices like veth devices XDP is unsuitable in this case since the kernel operates solely on a skb here and generic XDP has a few limitations where it does not operate with cloned skb’s. The latter is heavily used from the TCP/IP stack in order to hold data segments for retransmission where the generic XDP hook would simply get bypassed instead. Moreover, generic XDP needs to linearize the entire skb resulting in heavily degraded performance. tc BPF on the other hand is more flexible as it specializes on the skb input context case and thus does not need to cope with the limitations from generic XDP.

• Forwarding and load-balancing

The forwarding and load-balancing use case is quite similar to XDP, although slightly more targeted towards east-west container workloads rather than north-south traffic (though both technologies can be used in either case). Since XDP is only available on ingress side, tc BPF programs allow for further use cases that apply in particular on egress, for example, container based traffic can already be NATed and load-balanced on the egress side through BPF out of the initial namespace such that this is done transparent to the container itself. Egress traffic is already based on the sk_buff structure due to the nature of the kernel’s networking stack, so packet rewrites and redirects are suitable out of tc BPF. By utilizing the bpf_redirect() helper function, BPF can take over the forwarding logic to push the packet either into the ingress or egress path of another networking device. Thus, any bridge-like devices become unnecessary to use as well by utilizing tc BPF as forwarding fabric.

• Flow sampling, monitoring

Like in XDP case, flow sampling and monitoring can be realized through a high-performance lockless per-CPU memory mapped perf ring buffer where the BPF program is able to push custom data, the full or truncated packet contents, or both up to a user space application. From the tc BPF program this is realized through the bpf_skb_event_output() BPF helper function which has the same function signature and semantics as bpf_xdp_event_output(). Given tc BPF programs can be attached to ingress and egress as opposed to only ingress in XDP BPF case plus the two tc hooks are at the lowest layer in the (generic) networking stack, this allows for bidirectional monitoring of all network traffic from a particular node. This might be somewhat related to the cBPF case which tcpdump and Wireshark makes use of, though, without having to clone the skb and with being a lot more flexible in terms of programmability where, for example, BPF can already perform in-kernel aggregation rather than pushing everything up to user space as well as custom annotations for packets pushed into the ring buffer. The latter is also heavily used in Cilium where packet drops can be further annotated to correlate container labels and reasons for why a given packet had to be dropped (such as due to policy violation) in order to provide a richer context.

• Packet scheduler pre-processing

The sch_clsact’s egress hook which is called sch_handle_egress() runs right before taking the kernel’s qdisc root lock, thus tc BPF programs can be utilized to perform all the heavy lifting packet classification and mangling before the packet is transmitted into a real full blown qdisc such as sch_htb. This type of interaction of sch_clsact with a real qdisc like sch_htb coming later in the transmission phase allows to reduce the lock contention on transmission since sch_clsact’s egress hook is executed without taking locks.

One concrete example user of tc BPF but also XDP BPF programs is Cilium. Cilium is open source software for transparently securing the network connectivity between application services deployed using Linux container management platforms like Docker and Kubernetes and operates at Layer 3/4 as well as Layer 7. At the heart of Cilium operates BPF in order to implement the policy enforcement as well as load balancing and monitoring.

Driver support

Since tc BPF programs are triggered from the kernel’s networking stack and not directly out of the driver, they do not require any extra driver modification and therefore can run on any networking device. The only exception listed below is for offloading tc BPF programs to the NIC.

Drivers supporting offloaded tc BPF

• Netronome

Note that also here examples for writing and loading tc BPF programs are included in the toolchain section under the respective tools.

## Further Reading¶

Mentioned lists of docs, projects, talks, papers, and further reading material are likely not complete. Thus, feel free to open pull requests to complete the list.

### Kernel Developer FAQ¶

Under Documentation/bpf/, the Linux kernel provides two FAQ files that are mainly targeted for kernel developers involved in the BPF subsystem.

### Projects using BPF¶

The following list includes a selection of open source projects making use of BPF respectively provide tooling for BPF. In this context the eBPF instruction set is specifically meant instead of projects utilizing the legacy cBPF:

Tracing

• BCC

BCC stands for BPF Compiler Collection and its key feature is to provide a set of easy to use and efficient kernel tracing utilities all based upon BPF programs hooking into kernel infrastructure based upon kprobes, kretprobes, tracepoints, uprobes, uretprobes as well as USDT probes. The collection provides close to hundred tools targeting different layers across the stack from applications, system libraries, to the various different kernel subsystems in order to analyze a system’s performance characteristics or problems. Additionally, BCC provides an API in order to be used as a library for other projects.

https://github.com/iovisor/bcc

• bpftrace

bpftrace is a DTrace-style dynamic tracing tool for Linux and uses LLVM as a back end to compile scripts to BPF-bytecode and makes use of BCC for interacting with the kernel’s BPF tracing infrastructure. It provides a higher-level language for implementing tracing scripts compared to native BCC.

https://github.com/ajor/bpftrace

• perf

The perf tool which is developed by the Linux kernel community as part of the kernel source tree provides a way to load tracing BPF programs through the conventional perf record subcommand where the aggregated data from BPF can be retrieved and post processed in perf.data for example through perf script and other means.

https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/perf

• ply

ply is a tracing tool that follows the ‘Little Language’ approach of yore, and compiles ply scripts into Linux BPF programs that are attached to kprobes and tracepoints in the kernel. The scripts have a C-like syntax, heavily inspired by DTrace and by extension awk. ply keeps dependencies to very minimum and only requires flex and bison at build time, only libc at runtime.

https://github.com/wkz/ply

• systemtap

systemtap is a scripting language and tool for extracting, filtering and summarizing data in order to diagnose and analyze performance or functional problems. It comes with a BPF back end called stapbpf which translates the script directly into BPF without the need of an additional compiler and injects the probe into the kernel. Thus, unlike stap’s kernel modules this does neither have external dependencies nor requires to load kernel modules.

https://sourceware.org/git/gitweb.cgi?p=systemtap.git;a=summary

• PCP

Performance Co-Pilot (PCP) is a system performance and analysis framework which is able to collect metrics through a variety of agents as well as analyze collected systems’ performance metrics in real-time or by using historical data. With pmdabcc, PCP has a BCC based performance metrics domain agent which extracts data from the kernel via BPF and BCC.

https://github.com/performancecopilot/pcp

• Weave Scope

Weave Scope is a cloud monitoring tool collecting data about processes, networking connections or other system data by making use of BPF in combination with kprobes. Weave Scope works on top of the gobpf library in order to load BPF ELF files into the kernel, and comes with a tcptracer-bpf tool which monitors connect, accept and close calls in order to trace TCP events.

https://github.com/weaveworks/scope

Networking

• Cilium

Cilium provides and transparently secures network connectivity and load-balancing between application workloads such as application containers or processes. Cilium operates at Layer 3/4 to provide traditional networking and security services as well as Layer 7 to protect and secure use of modern application protocols such as HTTP, gRPC and Kafka. It is integrated into orchestration frameworks such as Kubernetes and Mesos, and BPF is the foundational part of Cilium that operates in the kernel’s networking data path.

https://github.com/cilium/cilium

• iproute2

iproute2 offers the ability to load BPF programs as LLVM generated ELF files into the kernel. iproute2 supports both, XDP BPF programs as well as tc BPF programs through a common BPF loader backend. The tc and ip command line utilities enable loader and introspection functionality for the user.

https://git.kernel.org/pub/scm/network/iproute2/iproute2.git/

• p4c-xdp

p4c-xdp presents a P4 compiler backend targeting BPF and XDP. P4 is a domain specific language describing how packets are processed by the data plane of a programmable network element such as NICs, appliances or switches, and with the help of p4c-xdp P4 programs can be translated into BPF C programs which can be compiled by clang / LLVM and loaded as BPF programs into the kernel at XDP layer for high performance packet processing.

https://github.com/vmware/p4c-xdp

Others

• LLVM

clang / LLVM provides the BPF back end in order to compile C BPF programs into BPF instructions contained in ELF files. The LLVM BPF back end is developed alongside with the BPF core infrastructure in the Linux kernel and maintained by the same community. clang / LLVM is a key part in the toolchain for developing BPF programs.

https://llvm.org/

• bpftool

bpftool is the main tool for introspecting and debugging BPF programs and BPF maps, and like libbpf is developed by the Linux kernel community. It allows for dumping all active BPF programs and maps in the system, dumping and disassembling BPF or JITed BPF instructions from a program as well as dumping and manipulating BPF maps in the system. bpftool supports interaction with the BPF filesystem, loading various program types from an object file into the kernel and much more.

https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/bpf/bpftool

• gobpf

gobpf provides go bindings for the bcc framework as well as low-level routines in order to load and use BPF programs from ELF files.

https://github.com/iovisor/gobpf

• ebpf_asm

ebpf_asm provides an assembler for BPF programs written in an Intel-like assembly syntax, and therefore offers an alternative for writing BPF programs directly in assembly for cases where programs are rather small and simple without needing the clang / LLVM toolchain.

https://github.com/solarflarecom/ebpf_asm

### XDP Newbies¶

There are a couple of walk-through posts by David S. Miller to the xdp-newbies mailing list (http://vger.kernel.org/vger-lists.html#xdp-newbies), which explain various parts of XDP and BPF:

1. May 2017,
BPF Verifier Overview, David S. Miller, https://www.spinics.net/lists/xdp-newbies/msg00185.html
1. May 2017,
Contextually speaking…, David S. Miller, https://www.spinics.net/lists/xdp-newbies/msg00181.html
1. May 2017,
bpf.h and you…, David S. Miller, https://www.spinics.net/lists/xdp-newbies/msg00179.html
1. Apr 2017,
XDP example of the day, David S. Miller, https://www.spinics.net/lists/xdp-newbies/msg00009.html

### BPF Newsletter¶

Alexander Alemayhu initiated a newsletter around BPF roughly once per week covering latest developments around BPF in Linux kernel land and its surrounding ecosystem in user space.

All BPF update newsletters (01 - 12) can be found here:

### Podcasts¶

There have been a number of technical podcasts partially covering BPF. Incomplete list:

1. Feb 2017,
Linux Networking Update from Netdev Conference, Thomas Graf, Software Gone Wild, Show 71, http://blog.ipspace.net/2017/02/linux-networking-update-from-netdev.html http://media.blubrry.com/ipspace/stream.ipspace.net/nuggets/podcast/Show_71-NetDev_Update.mp3
1. Jan 2017,
The IO Visor Project, Brenden Blanco, OVS Orbit, Episode 23, https://ovsorbit.org/#e23 https://ovsorbit.org/episode-23.mp3
1. Oct 2016,
Fast Linux Packet Forwarding, Thomas Graf, Software Gone Wild, Show 64, http://blog.ipspace.net/2016/10/fast-linux-packet-forwarding-with.html http://media.blubrry.com/ipspace/stream.ipspace.net/nuggets/podcast/Show_64-Cilium_with_Thomas_Graf.mp3
1. Aug 2016,
P4 on the Edge, John Fastabend, OVS Orbit, Episode 11, https://ovsorbit.org/#e11 https://ovsorbit.org/episode-11.mp3
1. May 2016,
Cilium, Thomas Graf, OVS Orbit, Episode 4, https://ovsorbit.org/#e4 https://ovsorbit.benpfaff.org/episode-4.mp3

### Blog posts¶

The following (incomplete) list includes blog posts around BPF, XDP and related projects:

1. May 2017,
An entertaining eBPF XDP adventure, Suchakra Sharma, https://suchakra.wordpress.com/2017/05/23/an-entertaining-ebpf-xdp-adventure/
1. May 2017,
eBPF, part 2: Syscall and Map Types, Ferris Ellis, https://ferrisellis.com/posts/ebpf_syscall_and_maps/
1. May 2017,
Monitoring the Control Plane, Gary Berger, http://firstclassfunc.com/2017/05/monitoring-the-control-plane/
1. Apr 2017,
USENIX/LISA 2016 Linux bcc/BPF Tools, Brendan Gregg, http://www.brendangregg.com/blog/2017-04-29/usenix-lisa-2016-bcc-bpf-tools.html
1. Apr 2017,
Liveblog: Cilium for Network and Application Security with BPF and XDP, Scott Lowe, http://blog.scottlowe.org//2017/04/18/black-belt-cilium/
1. Apr 2017,
eBPF, part 1: Past, Present, and Future, Ferris Ellis, https://ferrisellis.com/posts/ebpf_past_present_future/
1. Mar 2017,
Analyzing KVM Hypercalls with eBPF Tracing, Suchakra Sharma, https://suchakra.wordpress.com/2017/03/31/analyzing-kvm-hypercalls-with-ebpf-tracing/
1. Jan 2017,
Golang bcc/BPF Function Tracing, Brendan Gregg, http://www.brendangregg.com/blog/2017-01-31/golang-bcc-bpf-function-tracing.html
1. Dec 2016,
Give me 15 minutes and I’ll change your view of Linux tracing, Brendan Gregg, http://www.brendangregg.com/blog/2016-12-27/linux-tracing-in-15-minutes.html
1. Nov 2016,
Cilium: Networking and security for containers with BPF and XDP, Daniel Borkmann, https://opensource.googleblog.com/2016/11/cilium-networking-and-security.html
1. Nov 2016,
Linux bcc/BPF tcplife: TCP Lifespans, Brendan Gregg, http://www.brendangregg.com/blog/2016-11-30/linux-bcc-tcplife.html
1. Oct 2016,
DTrace for Linux 2016, Brendan Gregg, http://www.brendangregg.com/blog/2016-10-27/dtrace-for-linux-2016.html
1. Oct 2016,
Linux 4.9’s Efficient BPF-based Profiler, Brendan Gregg, http://www.brendangregg.com/blog/2016-10-21/linux-efficient-profiler.html
1. Oct 2016,
Linux bcc tcptop, Brendan Gregg, http://www.brendangregg.com/blog/2016-10-15/linux-bcc-tcptop.html
1. Oct 2016,
Linux bcc/BPF Node.js USDT Tracing, Brendan Gregg, http://www.brendangregg.com/blog/2016-10-12/linux-bcc-nodejs-usdt.html
1. Oct 2016,
Linux bcc/BPF Run Queue (Scheduler) Latency, Brendan Gregg, http://www.brendangregg.com/blog/2016-10-08/linux-bcc-runqlat.html
1. Oct 2016,
Linux bcc ext4 Latency Tracing, Brendan Gregg, http://www.brendangregg.com/blog/2016-10-06/linux-bcc-ext4dist-ext4slower.html
1. Oct 2016,
Linux MySQL Slow Query Tracing with bcc/BPF, Brendan Gregg, http://www.brendangregg.com/blog/2016-10-04/linux-bcc-mysqld-qslower.html
1. Oct 2016,
Linux bcc Tracing Security Capabilities, Brendan Gregg, http://www.brendangregg.com/blog/2016-10-01/linux-bcc-security-capabilities.html
1. Sep 2016,
Suricata bypass feature, Eric Leblond, https://www.stamus-networks.com/2016/09/28/suricata-bypass-feature/
1. Aug 2016,
Introducing the p0f BPF compiler, Gilberto Bertin, https://blog.cloudflare.com/introducing-the-p0f-bpf-compiler/
1. Jun 2016,
Ubuntu Xenial bcc/BPF, Brendan Gregg, http://www.brendangregg.com/blog/2016-06-14/ubuntu-xenial-bcc-bpf.html
1. Mar 2016,
Linux BPF/bcc Road Ahead, March 2016, Brendan Gregg, http://www.brendangregg.com/blog/2016-03-28/linux-bpf-bcc-road-ahead-2016.html
1. Mar 2016,
Linux BPF Superpowers, Brendan Gregg, http://www.brendangregg.com/blog/2016-03-05/linux-bpf-superpowers.html
1. Feb 2016,
Linux eBPF/bcc uprobes, Brendan Gregg, http://www.brendangregg.com/blog/2016-02-08/linux-ebpf-bcc-uprobes.html
1. Feb 2016,
Who is waking the waker? (Linux chain graph prototype), Brendan Gregg, http://www.brendangregg.com/blog/2016-02-05/ebpf-chaingraph-prototype.html
1. Feb 2016,
Linux Wakeup and Off-Wake Profiling, Brendan Gregg, http://www.brendangregg.com/blog/2016-02-01/linux-wakeup-offwake-profiling.html
1. Jan 2016,
Linux eBPF Off-CPU Flame Graph, Brendan Gregg, http://www.brendangregg.com/blog/2016-01-20/ebpf-offcpu-flame-graph.html
1. Jan 2016,
Linux eBPF Stack Trace Hack, Brendan Gregg, http://www.brendangregg.com/blog/2016-01-18/ebpf-stack-trace-hack.html
1. Sep 2015,
Linux Networking, Tracing and IO Visor, a New Systems Performance Tool for a Distributed World, Suchakra Sharma, https://thenewstack.io/comparing-dtrace-iovisor-new-systems-performance-platform-advance-linux-networking-virtualization/
1. Aug 2015,
BPF Internals - II, Suchakra Sharma, https://suchakra.wordpress.com/2015/08/12/bpf-internals-ii/
1. May 2015,
eBPF: One Small Step, Brendan Gregg, http://www.brendangregg.com/blog/2015-05-15/ebpf-one-small-step.html
1. May 2015,
BPF Internals - I, Suchakra Sharma, https://suchakra.wordpress.com/2015/05/18/bpf-internals-i/
1. Jul 2014,
Introducing the BPF Tools, Marek Majkowski, https://blog.cloudflare.com/introducing-the-bpf-tools/
1. May 2014,
BPF - the forgotten bytecode, Marek Majkowski, https://blog.cloudflare.com/bpf-the-forgotten-bytecode/

### Talks¶

The following (incomplete) list includes talks and conference papers related to BPF and XDP:

1. May 2017,
PyCon 2017, Portland, Executing python functions in the linux kernel by transpiling to bpf, Alex Gartrell, https://www.youtube.com/watch?v=CpqMroMBGP4
1. May 2017,
gluecon 2017, Denver, Cilium + BPF: Least Privilege Security on API Call Level for Microservices, Dan Wendlandt, http://gluecon.com/#agenda
1. May 2017,
Lund Linux Con, Lund, XDP - eXpress Data Path, Jesper Dangaard Brouer, http://people.netfilter.org/hawk/presentations/LLC2017/XDP_DDoS_protecting_LLC2017.pdf
1. May 2017,
Polytechnique Montreal, Trace Aggregation and Collection with eBPF, Suchakra Sharma, http://step.polymtl.ca/~suchakra/eBPF-5May2017.pdf
1. Apr 2017,
DockerCon, Austin, Cilium - Network and Application Security with BPF and XDP, Thomas Graf, https://www.slideshare.net/ThomasGraf5/dockercon-2017-cilium-network-and-application-security-with-bpf-and-xdp
1. Apr 2017,
NetDev 2.1, Montreal, XDP Mythbusters, David S. Miller, https://www.netdevconf.org/2.1/slides/apr7/miller-XDP-MythBusters.pdf
1. Apr 2017,
NetDev 2.1, Montreal, Droplet: DDoS countermeasures powered by BPF + XDP, Huapeng Zhou, Doug Porter, Ryan Tierney, Nikita Shirokov, https://www.netdevconf.org/2.1/slides/apr6/zhou-netdev-xdp-2017.pdf
1. Apr 2017,
NetDev 2.1, Montreal, XDP in practice: integrating XDP in our DDoS mitigation pipeline, Gilberto Bertin, https://www.netdevconf.org/2.1/slides/apr6/bertin_Netdev-XDP.pdf
1. Apr 2017,
NetDev 2.1, Montreal, XDP for the Rest of Us, Andy Gospodarek, Jesper Dangaard Brouer, https://www.netdevconf.org/2.1/slides/apr7/gospodarek-Netdev2.1-XDP-for-the-Rest-of-Us_Final.pdf
1. Mar 2017,
SCALE15x, Pasadena, Linux 4.x Tracing: Performance Analysis with bcc/BPF, Brendan Gregg, https://www.slideshare.net/brendangregg/linux-4x-tracing-performance-analysis-with-bccbpf
1. Mar 2017,
XDP Inside and Out, David S. Miller, https://github.com/iovisor/bpf-docs/raw/master/XDP_Inside_and_Out.pdf
1. Mar 2017,
OpenSourceDays, Copenhagen, XDP - eXpress Data Path, Used for DDoS protection, Jesper Dangaard Brouer, https://github.com/iovisor/bpf-docs/raw/master/XDP_Inside_and_Out.pdf
1. Mar 2017,
source{d}, Infrastructure 2017, Madrid, High-performance Linux monitoring with eBPF, Alfonso Acosta, https://www.youtube.com/watch?v=k4jqTLtdrxQ
1. Feb 2017,
FOSDEM 2017, Brussels, Stateful packet processing with eBPF, an implementation of OpenState interface, Quentin Monnet, https://fosdem.org/2017/schedule/event/stateful_ebpf/
1. Feb 2017,
FOSDEM 2017, Brussels, eBPF and XDP walkthrough and recent updates, Daniel Borkmann, http://borkmann.ch/talks/2017_fosdem.pdf
1. Feb 2017,
FOSDEM 2017, Brussels, Cilium - BPF & XDP for containers, Thomas Graf, https://fosdem.org/2017/schedule/event/cilium/
1. Jan 2017,
linuxconf.au, Hobart, BPF: Tracing and more, Brendan Gregg, https://www.slideshare.net/brendangregg/bpf-tracing-and-more
1. Dec 2016,
USENIX LISA 2016, Boston, Linux 4.x Tracing Tools: Using BPF Superpowers, Brendan Gregg, https://www.slideshare.net/brendangregg/linux-4x-tracing-tools-using-bpf-superpowers
1. Nov 2016,
Linux Plumbers, Santa Fe, Cilium: Networking & Security for Containers with BPF & XDP, Thomas Graf, http://www.slideshare.net/ThomasGraf5/clium-container-networking-with-bpf-xdp
1. Nov 2016,
OVS Conference, Santa Clara, Offloading OVS Flow Processing using eBPF, William (Cheng-Chun) Tu, http://openvswitch.org/support/ovscon2016/7/1120-tu.pdf
1. Oct 2016,
One.com, Copenhagen, XDP - eXpress Data Path, Intro and future use-cases, Jesper Dangaard Brouer, http://people.netfilter.org/hawk/presentations/xdp2016/xdp_intro_and_use_cases_sep2016.pdf
1. Oct 2016,
Docker Distributed Systems Summit, Berlin, Cilium: Networking & Security for Containers with BPF & XDP, Thomas Graf, http://www.slideshare.net/Docker/cilium-bpf-xdp-for-containers-66969823
1. Oct 2016,
NetDev 1.2, Tokyo, Data center networking stack, Tom Herbert, http://netdevconf.org/1.2/session.html?tom-herbert
1. Oct 2016,
NetDev 1.2, Tokyo, Fast Programmable Networks & Encapsulated Protocols, David S. Miller, http://netdevconf.org/1.2/session.html?david-miller-keynote
1. Oct 2016,
NetDev 1.2, Tokyo, XDP workshop - Introduction, experience, and future development, Tom Herbert, http://netdevconf.org/1.2/session.html?herbert-xdp-workshop
1. Oct 2016,
NetDev1.2, Tokyo, The adventures of a Suricate in eBPF land, Eric Leblond, http://netdevconf.org/1.2/slides/oct6/10_suricata_ebpf.pdf
1. Oct 2016,
NetDev1.2, Tokyo, cls_bpf/eBPF updates since netdev 1.1, Daniel Borkmann, http://borkmann.ch/talks/2016_tcws.pdf
1. Oct 2016,
NetDev1.2, Tokyo, Advanced programmability and recent updates with tc’s cls_bpf, Daniel Borkmann, http://borkmann.ch/talks/2016_netdev2.pdf http://www.netdevconf.org/1.2/papers/borkmann.pdf
1. Oct 2016,
NetDev 1.2, Tokyo, eBPF/XDP hardware offload to SmartNICs, Jakub Kicinski, Nic Viljoen, http://netdevconf.org/1.2/papers/eBPF_HW_OFFLOAD.pdf
1. Aug 2016,
LinuxCon, Toronto, What Can BPF Do For You?, Brenden Blanco, https://events.linuxfoundation.org/sites/events/files/slides/iovisor-lc-bof-2016.pdf
1. Aug 2016,
LinuxCon, Toronto, Cilium - Fast IPv6 Container Networking with BPF and XDP, Thomas Graf, https://www.slideshare.net/ThomasGraf5/cilium-fast-ipv6-container-networking-with-bpf-and-xdp
1. Aug 2016,
P4, EBPF and Linux TC Offload, Dinan Gunawardena, Jakub Kicinski, https://de.slideshare.net/Open-NFP/p4-epbf-and-linux-tc-offload
1. Jul 2016,
Linux Meetup, Santa Clara, eXpress Data Path, Brenden Blanco, http://www.slideshare.net/IOVisor/express-data-path-linux-meetup-santa-clara-july-2016
1. Jul 2016,
Linux Meetup, Santa Clara, CETH for XDP, Yan Chan, Yunsong Lu, http://www.slideshare.net/IOVisor/ceth-for-xdp-linux-meetup-santa-clara-july-2016
1. May 2016,
P4 workshop, Stanford, P4 on the Edge, John Fastabend, https://schd.ws/hosted_files/2016p4workshop/1d/Intel%20Fastabend-P4%20on%20the%20Edge.pdf
1. Mar 2016,
Performance @Scale 2016, Menlo Park, Linux BPF Superpowers, Brendan Gregg, https://www.slideshare.net/brendangregg/linux-bpf-superpowers
1. Mar 2016,
eXpress Data Path, Tom Herbert, Alexei Starovoitov, https://github.com/iovisor/bpf-docs/raw/master/Express_Data_Path.pdf
1. Feb 2016,
NetDev1.1, Seville, On getting tc classifier fully programmable with cls_bpf, Daniel Borkmann, http://borkmann.ch/talks/2016_netdev.pdf http://www.netdevconf.org/1.1/proceedings/papers/On-getting-tc-classifier-fully-programmable-with-cls-bpf.pdf
1. Jan 2016,
FOSDEM 2016, Brussels, Linux tc and eBPF, Daniel Borkmann, http://borkmann.ch/talks/2016_fosdem.pdf
1. Oct 2015,
LinuxCon Europe, Dublin, eBPF on the Mainframe, Michael Holzheu, https://events.linuxfoundation.org/sites/events/files/slides/ebpf_on_the_mainframe_lcon_2015.pdf
1. Aug 2015,
Tracing Summit, Seattle, LLTng’s Trace Filtering and beyond (with some eBPF goodness, of course!), Suchakra Sharma, https://github.com/iovisor/bpf-docs/raw/master/ebpf_excerpt_20Aug2015.pdf
1. Jun 2015,
LinuxCon Japan, Tokyo, Exciting Developments in Linux Tracing, Elena Zannoni, https://events.linuxfoundation.org/sites/events/files/slides/tracing-linux-ezannoni-linuxcon-ja-2015_0.pdf
1. Feb 2015,
Collaboration Summit, Santa Rosa, BPF: In-kernel Virtual Machine, Alexei Starovoitov, https://events.linuxfoundation.org/sites/events/files/slides/bpf_collabsummit_2015feb20.pdf
1. Feb 2015,
NetDev 0.1, Ottawa, BPF: In-kernel Virtual Machine, Alexei Starovoitov, http://netdevconf.org/0.1/sessions/15.html
1. Feb 2014,
DevConf.cz, Brno, tc and cls_bpf: lightweight packet classifying with BPF, Daniel Borkmann, http://borkmann.ch/talks/2014_devconf.pdf