af-xdp

AF_XDP

Purpose

Guide agents through AF_XDP sockets for high-performance packet I/O: socket creation, UMEM setup, fill/completion/RX/TX rings, XDP programs with XDP_REDIRECT to XSK, copy vs zero-copy modes, libbpf helpers, performance comparison with DPDK, and production use cases.

When to Use

• Building a userspace packet processor with lower overhead than raw sockets
• Redirecting XDP-filtered traffic to userspace without DPDK complexity
• Implementing a custom load balancer or IDS dataplane
• Comparing zero-copy vs copy mode on your NIC/driver
• Integrating with existing libbpf/XDP infrastructure
• Need kernel cooperation (firewall rules) plus userspace processing

Workflow

1. Architecture overview

NIC → XDP program (BPF) → XDP_REDIRECT → AF_XDP socket → userspace
                ↓
           XDP_DROP/PASS/TX

Components:

• UMEM — shared memory region for frames
• Fill ring — userspace provides empty frame addresses to kernel
• Completion ring — kernel returns completed TX frames
• RX ring — kernel delivers received packets
• TX ring — userspace submits packets for transmission

2. UMEM and XSK socket creation

#include <bpf/xsk.h>
#include <bpf/libbpf.h>
#include <xdp/xsk.h>

#define NUM_FRAMES     4096
#define FRAME_SIZE     XSK_UMEM__DEFAULT_FRAME_SIZE
#define RX_BATCH_SIZE  64

struct xsk_umem_info {
    struct xsk_ring_prod fill;
    struct xsk_ring_cons comp;
    struct xsk_umem *umem;
    void *buffer;
};

struct xsk_socket_info {
    struct xsk_ring_cons rx;
    struct xsk_ring_prod tx;
    struct xsk_socket *xsk;
};

int xsk_setup(struct xsk_umem_info *umem_info,
              struct xsk_socket_info *xsk_info,
              int ifindex, int queue_id, int xsk_flags)
{
    umem_info->buffer = mmap(NULL, NUM_FRAMES * FRAME_SIZE,
        PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);

    struct xsk_umem_config umem_cfg = {
        .fill_size = XSK_RING_PROD__DEFAULT_NUM_DESCS,
        .comp_size = XSK_RING_CONS__DEFAULT_NUM_DESCS,
        .frame_size = FRAME_SIZE,
        .frame_headroom = XSK_UMEM__DEFAULT_FRAME_HEADROOM,
        .flags = 0,
    };

    int ret = xsk_umem__create(&umem_info->umem, umem_info->buffer,
        NUM_FRAMES * FRAME_SIZE, &umem_info->fill, &umem_info->comp,
        &umem_cfg);
    if (ret)
        return ret;

    struct xsk_socket_config xsk_cfg = {
        .rx_size = XSK_RING_CONS__DEFAULT_NUM_DESCS,
        .tx_size = XSK_RING_PROD__DEFAULT_NUM_DESCS,
        .libbpf_flags = XSK_LIBBPF_FLAGS__INHIBIT_PROG_LOAD,
        .xdp_flags = XDP_FLAGS_UPDATE_IF_NOEXIST,
        .bind_flags = xsk_flags,  // XDP_ZEROCOPY or XDP_COPY
    };

    return xsk_socket__create(&xsk_info->xsk, "eth0", queue_id,
        umem_info->umem, &xsk_info->rx, &xsk_info->tx, &xsk_cfg);
}

3. Populate fill ring

void populate_fill_ring(struct xsk_umem_info *umem) {
    uint32_t idx;
    uint32_t ret = xsk_ring_prod__reserve(&umem->fill, RX_BATCH_SIZE, &idx);
    for (uint32_t i = 0; i < ret; i++)
        *xsk_ring_prod__fill_addr(&umem->fill, idx + i) = i * FRAME_SIZE;
    xsk_ring_prod__submit(&umem->fill, ret);
}

Must keep fill ring stocked — kernel drops packets if no buffers available.

4. XDP redirect program

// xdp_redirect.c
#include <linux/bpf.h>
#include <bpf/bpf_helpers.h>

struct {
    __uint(type, BPF_MAP_TYPE_XSKMAP);
    __uint(max_entries, 64);
    __type(key, int);
    __type(value, int);
} xsks_map SEC(".maps");

SEC("xdp")
int xdp_redirect_prog(struct xdp_md *ctx)
{
    int index = ctx->rx_queue_index;
    return bpf_redirect_map(&xsks_map, index, 0);
}

# Load and attach
bpftool prog load xdp_redirect.o /sys/fs/bpf/xdp_redirect
ip link set dev eth0 xdp obj xdp_redirect.o sec xdp

# Pin xsks_map and update with socket fd

5. RX processing loop

while (running) {
    uint32_t idx_rx = 0, rcvd;
    rcvd = xsk_ring_cons__peek(&xsk_info->rx, RX_BATCH_SIZE, &idx_rx);
    if (!rcvd)
        continue;

    for (uint32_t i = 0; i < rcvd; i++) {
        const struct xdp_desc *desc = xsk_ring_cons__rx_desc(&xsk_info->rx, idx_rx + i);
        uint64_t addr = desc->addr;
        uint32_t len = desc->len;
        uint8_t *pkt = (uint8_t *)xsk_umem__get_data(umem_info->buffer, addr);
        process_packet(pkt, len);
    }
    xsk_ring_cons__release(&xsk_info->rx, rcvd);

    // Return frames to fill ring
    refill_fill_ring(umem_info, rcvd);
}

6. Copy vs zero-copy

Mode	Flag	Requirements
Copy	XDP_COPY (default)	Any driver; kernel copies to UMEM
Zero-copy	XDP_ZEROCOPY	Driver support (i40e, ixgbe, mlx5, etc.)

# Check driver ZC support
ethtool -i eth0
# Kernel log on bind:
dmesg | grep xsk
# "Zero-copy enabled" or "Copy mode"

Zero-copy: NIC DMAs directly into UMEM frames — lowest latency. Copy mode: safer, universal.

7. libbpf xsk.h helpers

# Modern libbpf includes xsk API
pkg-config --libs libbpf
# -lbpf -lxdp (if separate libxdp installed)

Key functions:

• xsk_umem__create / xsk_umem__delete
• xsk_socket__create / xsk_socket__delete
• xsk_umem__get_data — pointer from frame address
• xsk_socket__fd — for epoll/poll integration

8. Performance vs DPDK

Factor	AF_XDP	DPDK
Setup complexity	Moderate	High (hugepages, EAL)
Kernel integration	XDP filter in kernel	Full bypass
Typical throughput	Near-DPDK with ZC	Highest
NIC binding	Stays on kernel driver	vfio/uio binding
Use case fit	Filter + selective userspace	Full dataplane takeover

9. Production patterns

Common deployments
├── CDN edge cache — XDP_DROP junk, redirect cacheable to XSK
├── DDoS mitigation — XDP_DROP attack patterns
├── Load balancer — XDP_TX hairpin or redirect to backend XSK
└── Observability — mirror subset to XSK for analysis

# Multi-queue: one XSK per RX queue, pinned to CPU
taskset -c 2 ./xsk_app --queue 2

Common Problems

Symptom	Cause	Fix
No packets in RX ring	XDP program not redirecting	Verify xsks_map entry for queue index
EBUSY on socket create	XDP already attached	ip link set dev eth0 xdp off first
Zero-copy fallback to copy	Driver lacks ZC	Check dmesg; use supported NIC
Packet drops	Fill ring empty	Aggressive refill; increase NUM_FRAMES
TX not working	Completion ring not polled	Process comp ring to recycle frames
Permission denied	CAP_NET_RAW needed	Run with appropriate capabilities

Related Skills

• skills/observability/ebpf — XDP/BPF program development
• skills/async-io/dpdk — full kernel bypass alternative
• skills/async-io/io-uring — async I/O for non-packet workloads
• skills/observability/ebpf-rust — Aya for XDP in Rust
• skills/profilers/linux-perf — profile XDP program CPU usage
• skills/allocators/numa-programming — NUMA-local UMEM allocation