ebpf-rust — Claude Code skill: what it does, how to install
Guidance for developing Linux eBPF programs in Rust via the Aya framework. It helps write kernel probes, stream events into Tokio, and fix verifier issues.
npx skills add mohitmishra786/low-level-dev-skills --skill ebpf-rust -a claude-codeWhat ebpf-rust does
The skill directs AI assistants on building eBPF code with Aya without C or libbpf dependencies. It structures a multi-crate workspace separating the #![no_std] kernel artifact from the host runner. Code patterns cover probe macros like #[tracepoint], #[kprobe], #[xdp], and #[tc] alongside structured logging with aya-log.
Data sharing relies on kernel maps such as HashMap, RingBuf, and LruHashMap. To ingest high-throughput events on the host side, it wires the kernel ring buffer to Tokio via AsyncFd. It also covers generating kernel BTF definitions with aya-tool and accessing fields with bpf_core_read!.
Loading programs into the kernel requires elevated privileges. Without CAP_BPF or CAP_SYS_ADMIN, operations fail with permission errors. When the BPF verifier rejects bytecode, running RUST_LOG=debug cargo xtask run and inspecting state with bpftool surfaces the underlying check failures.
What it helps with
- scaffold a Rust eBPF workspace using Aya project templates
- write
#![no_std]kernel probes and attach them to tracepoints - stream kernel events into async Tokio runtimes with RingBuf
- troubleshoot BPF verifier errors and BTF type mismatches
Example prompts for the agent
- How do I write an eBPF tracepoint program in Rust with aya-bpf?
- Show an example of streaming RingBuf events to tokio in userspace.
- My Aya program fails to pass the BPF verifier, how do I debug it?
How to install ebpf-rust
1Skills CLI (Vercel)
npx skills add mohitmishra786/low-level-dev-skills --skill ebpf-rust -a claude-code2Manually: put the files in place
~/.claude/skills/ebpf-rust/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/ebpf-rust/
1Skills CLI (Vercel)
npx skills add mohitmishra786/low-level-dev-skills --skill ebpf-rust -a codex2Manually: put the files in place
.agents/skills/ebpf-rust/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add mohitmishra786/low-level-dev-skills --skill ebpf-rust -a cursor2Manually: put the files in place
.cursor/skills/ebpf-rust/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install mohitmishra786/low-level-dev-skills ebpf-rust2Manually: put the files in place
.github/skills/ebpf-rust/SKILL.md1Skills CLI (Vercel)
npx skills add mohitmishra786/low-level-dev-skills --skill ebpf-rust -a gemini-cli2Platform CLI
gemini skills install https://github.com/mohitmishra786/low-level-dev-skills --path skills/observability/ebpf-rustManually: put the files in place
.devin/skills/ebpf-rust/SKILL.mdThe legacy .windsurf/skills/ path still works
Skills CLI (Vercel)
npx skills add mohitmishra786/low-level-dev-skills --skill ebpf-rust -a opencodeClaude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about ebpf-rust
How do I install ebpf-rust in Claude Code?
Run npx skills add mohitmishra786/low-level-dev-skills --skill ebpf-rust -a claude-code. It creates the skill definition in ~/.claude/skills/ebpf-rust/SKILL.md.
What does the ebpf-rust skill do?
It provides step-by-step guidance for writing, compiling, and loading Rust-based eBPF programs using Aya, aya-log, and async Tokio.
What Linux permissions are required to run Aya applications?
Loading BPF objects into the kernel requires root access or CAP_BPF and CAP_SYS_ADMIN capabilities, meaning host loaders must execute with sudo.
Source text
SKILL.md · MIT
---
name: ebpf-rust
description: Rust eBPF skill using the Aya framework. Use when writing eBPF programs in Rust with aya-bpf and aya-log, defining BPF map types, integrating with tokio userspace, sharing maps between kernel and userspace, or debugging Aya compilation and loading errors. Activates on queries about Aya, aya-bpf, Rust eBPF, aya-log, eBPF in Rust, or BPF programs with tokio.
---
# eBPF with Rust (Aya)
## Purpose
Guide agents through building production eBPF programs in Rust using the Aya framework: writing kernel-side BPF code with `aya-bpf`, structured logging with `aya-log`, sharing maps between BPF and userspace, and integrating with async tokio.
## Triggers
- "How do I write an eBPF program in Rust?"
- "How do I use the Aya framework?"
- "How do I share a BPF map between kernel and userspace in Rust?"
- "How do I log from a BPF program in Rust?"
- "My Aya program fails to load — how do I debug it?"
- "How do I integrate an eBPF program with tokio?"
## Workflow
### 1. Project setup
```bash
# Install aya-tool (generates bindings from vmlinux BTF)
cargo install aya-tool
# Create new Aya project from template
cargo install cargo-generate
cargo generate https://github.com/aya-rs/aya-template
# Workspace layout (generated)
# my-ebpf/
# ├── my-ebpf-ebpf/ <- kernel-side crate (target: bpf)
# ├── my-ebpf/ <- userspace crate (runs on host)
# └── xtask/ <- build helper (cargo xtask build/run)
```
```bash
# Build both sides
cargo xtask build-ebpf # builds BPF object
cargo xtask run # builds + runs with sudo
```
### 2. Kernel-side BPF program
```rust
// my-ebpf-ebpf/src/main.rs
#![no_std]
#![no_main]
use aya_bpf::{
macros::{map, tracepoint},
maps::HashMap,
programs::TracePointContext,
helpers::bpf_get_current_pid_tgid,
};
use aya_log_ebpf::info;
#[map]
static CALL_COUNT: HashMap<u32, u64> = HashMap::with_max_entries(1024, 0);
#[tracepoint]
pub fn trace_read(ctx: TracePointContext) -> u32 {
let pid = (bpf_get_current_pid_tgid() >> 32) as u32;
// Lookup or insert
match unsafe { CALL_COUNT.get(&pid) } {
Some(count) => {
let _ = CALL_COUNT.insert(&pid, &(count + 1), 0);
}
None => {
let _ = CALL_COUNT.insert(&pid, &1u64, 0);
}
}
info!(&ctx, "read() called by pid {}", pid);
0
}
#[panic_handler]
fn panic(_info: &core::panic::PanicInfo) -> ! {
unsafe { core::hint::unreachable_unchecked() }
}
```
### 3. Userspace loader with tokio
```rust
// my-ebpf/src/main.rs
use aya::{Bpf, programs::TracePoint, maps::HashMap};
use aya_log::BpfLogger;
use tokio::signal;
#[tokio::main]
async fn main() -> anyhow::Result<()> {
// Load compiled BPF object (embedded at build time)
let mut bpf = Bpf::load(include_bytes_aligned!(
"../../target/bpf/my-ebpf-ebpf.bpf.o"
))?;
// Initialize structured logging from BPF programs
BpfLogger::init(&mut bpf)?;
// Attach tracepoint
let program: &mut TracePoint = bpf.program_mut("trace_read").unwrap().try_into()?;
program.load()?;
program.attach("syscalls", "sys_enter_read")?;
// Read from shared map
let map: HashMap<_, u32, u64> = HashMap::try_from(bpf.map("CALL_COUNT").unwrap())?;
// Wait for Ctrl+C
signal::ctrl_c().await?;
// Print final counts
for (pid, count) in map.iter().filter_map(|r| r.ok()) {
println!("PID {}: {} reads", pid, count);
}
Ok(())
}
```
### 4. Map types in Aya
```rust
use aya_bpf::maps::{HashMap, Array, RingBuf, PerfEventArray, LruHashMap};
// Hash map
#[map]
static MY_MAP: HashMap<u32, u64> = HashMap::with_max_entries(1024, 0);
// Ring buffer (preferred for events)
#[map]
static EVENTS: RingBuf = RingBuf::with_byte_size(256 * 1024, 0);
// LRU hash (connection tracking)
#[map]
static CONNS: LruHashMap<u32, ConnInfo> = LruHashMap::with_max_entries(10000, 0);
// Sending events via RingBuf (kernel side)
if let Ok(mut entry) = unsafe { EVENTS.reserve::<MyEvent>(0) } {
entry.write(MyEvent { pid, ts });
entry.submit(0);
}
```
```rust
// Reading RingBuf events (userspace)
use aya::maps::RingBuf;
use tokio::io::unix::AsyncFd;
let ring = RingBuf::try_from(bpf.take_map("EVENTS").unwrap())?;
let mut ring = AsyncFd::new(ring)?;
loop {
let mut guard = ring.readable_mut().await?;
let rb = guard.get_inner_mut();
while let Some(item) = rb.next() {
let event: &MyEvent = unsafe { &*(item.as_ptr() as *const MyEvent) };
println!("Event from PID {}", event.pid);
}
guard.clear_ready();
}
```
### 5. Supported program types
| Aya macro | Program type | Attach target |
|-----------|-------------|---------------|
| `#[tracepoint]` | Tracepoint | `"syscalls", "sys_enter_read"` |
| `#[kprobe]` | kprobe | function name |
| `#[kretprobe]` | kretprobe | function name |
| `#[uprobe]` | uprobe | userspace binary + offset |
| `#[xdp]` | XDP | network interface |
| `#[tc]` | TC (traffic control) | netdevice + direction |
| `#[socket_filter]` | Socket filter | raw socket fd |
| `#[perf_event]` | Perf event | perf_event fd |
| `#[lsm]` | LSM hook | security hook name |
| `#[sk_msg]` | Sockmap | socket map |
### 6. Generating kernel type bindings
```bash
# Generate bindings from running kernel's BTF
aya-tool generate task_struct > src/vmlinux.rs
# Or use btf_type_tag and CO-RE
# aya-bpf supports CO-RE via bpf_core_read! macro
```
```rust
// CO-RE field access in Aya
use aya_bpf::helpers::bpf_core_read;
let dport: u16 = unsafe {
bpf_core_read!(sk, __sk_common.skc_dport)?
};
```
### 7. Debugging load failures
```bash
# Check verifier errors (Aya surfaces them as Rust errors)
# Run with RUST_LOG=debug for verbose output
RUST_LOG=debug cargo xtask run 2>&1 | grep -A 20 "verifier"
# Check BTF info
bpftool btf dump file /sys/kernel/btf/vmlinux | grep task_struct
# Inspect loaded programs after load
bpftool prog list
bpftool prog dump xlated name trace_read
```
| Error | Cause | Fix |
|-------|-------|-----|
| `invalid mem access` | Unbounded pointer dereference | Add null check before reading |
| `Type not found` | BTF mismatch with kernel | Regenerate vmlinux bindings |
| `Permission denied` | No CAP_BPF or CAP_SYS_ADMIN | Run with `sudo` or set capability |
| `map already exists` | Map pinned, name collision | Unpin or rename map |
## Related skills
- Use `skills/observability/ebpf` for C-based eBPF with libbpf
- Use `skills/rust/rust-async-internals` for tokio async patterns used in userspace
- Use `skills/rust/rust-unsafe` for unsafe code patterns in BPF helpers
© mohitmishra786 · MIT · mohitmishra786/low-level-dev-skills