Relink: Rust ELF Loader and Runtime/JIT Linker

English | 简体中文 | Contributing | 贡献指南

Load, link, and rewrite ELF in Rust and no_std environments.

Relink is a Rust ELF loading and runtime linking library. It can load .so files, executables, and object files from disk or memory, then resolve dependencies, apply relocations, and look up symbols.

When dlopen is too rigid, Relink lets you decide how dependencies are found, how symbols are searched, how memory is mapped, and whether to scan and adjust layout before loading.

When To Use It

• Load plugins, JIT artifacts, or hot-reload modules at runtime.
• Control DT_NEEDED dependencies, symbol scopes, or relocation handling yourself.
• Load ELF from memory, or plug in your own mmap or memory-management backend.
• Scan dependencies and sections first, then reorder layout, pack hot code, use huge pages, or run custom handling.
• Keep ELF loading available in no_std, kernels, embedded systems, or non-standard runtimes.

What It Loads

• Shared objects / dynamic libraries (ET_DYN)
• Executables and PIE-style images (ET_EXEC, plus executable-style ET_DYN)
• Relocatable object files (ET_REL) when the object feature is enabled
• File-backed or in-memory inputs via &str, String, &[u8], Vec<u8>, ElfFile, and ElfBinary

Use Loader::load() when you want automatic ELF type detection. Use load_dylib(), load_exec(), or load_object() when you want strict type checks.

Core Capabilities

Compared With dlopen

Quick Start

The default feature set is suitable for loading dynamic libraries, executables, and handling TLS:

[dependencies]
elf_loader = "0.15.1"

To enable the common advanced features in one bundle:

[dependencies]
elf_loader = { version = "0.15.1", features = ["full"] }

Load a Dynamic Library and Call a Symbol

use elf_loader::{
    image::{SyntheticModule, SyntheticSymbol},
    Loader, Result,
};

extern "C" fn host_double(value: i32) -> i32 {
    value * 2
}

fn main() -> Result<()> {
    let lib = Loader::new()
        .load_dylib("path/to/plugin.so")?
        .relocator()
        .scope([SyntheticModule::new(
            "__host",
            [SyntheticSymbol::function("host_double", host_double as *const ())],
        )])
        .relocate()?;

    let run = unsafe {
        lib.get::<extern "C" fn(i32) -> i32>("run")
            .expect("symbol `run` not found")
    };
    assert_eq!(run(21), 42);

    Ok(())
}

Loading Paths

Advanced Capability Index

For section reordering or hot-code packing before loading, the target ELF usually needs to keep relocation information, for example by passing -Wl,--emit-relocs to the linker.

Benchmarks

The table below is a GitHub Actions performance snapshot. Use it only as a reference for the current test suite. Full environment details are in actions/runs/25632675040/job/75239090388. The fixture is the repository's libc -> libb -> liba test chain, not the system C library.

Lower is better for loading. scan_first includes dependency scanning and section planning, so it is not a direct dlopen replacement.

Symbol lookup was measured after both loaders had already loaded the fixture chain:

Feature Flags

Notes:

• The default features are tls + libc.
• Compiling with tls is not enough by itself for TLS-using modules. Start from Loader::new().with_default_tls_resolver() or provide your own TLS resolver when loading ELF objects that require TLS relocations.
• load_object() is feature-gated. cargo run --example load_object will fail under the default feature set unless you add --features object.

Examples

The examples/ directory covers the main extension points:

Platform Notes

Legend: ✅ supported, 🟡 basic support, ⏳ pending. Complex section-reorder repair and .o / ET_REL support are currently centered on x86_64 and riscv64 relocation handling; contributions for the other architectures are welcome.

Symbol lookup is name-based and does not perform Rust name mangling for you. Export C ABI symbols when you want stable runtime lookup names.

Contributing

Issues and pull requests are welcome. Star the project if it is useful in your work.

License

This project is dual-licensed under either of the following:

• MIT License
• Apache License 2.0

Contributors