Skip to content

Latest commit

 

History

History
123 lines (100 loc) · 5.72 KB

README.md

File metadata and controls

123 lines (100 loc) · 5.72 KB
Raw

ELF Loader

A dead simple crate for ELF64 parsing and loading.

Features

  • This crate is #[no_std], you do not even need the alloc crate to start having fun.
  • It performs a lot of bounds, alignment and other sanity checks. Don't just blindly trust a buffer containing code. =)
  • There is no recursion in the code and the call graph is rather flat. Put simply: This crate won't eat much stack space.
  • Errors are very descriptive and very compact error codes. No wasted register space on your happy path.
  • This crate does its job in a quite small amount of code, despite all the error checking.
  • No dependencies, except for libcore.

TODOs

  • Currently, only page-aligned re-locatable x86_64 executables are supported. However, at least support for AArch64 and RISC-V is planned.
  • An other "not yet implemented" feature is dynamic linking. This is required to eventually make this crate a minimal drop-in replacement for dlopen. You cannot currently look up symbols, so all you get from loading an ELF is its entry point.
  • Currently, custom linker scripts have to be used that page-align all loadable sections. To relax this requirement, I'd need help finding and understanding the source code of ld.so from glibc. I.e. this crate does not currently act as a program interpreter.
  • Just Ctrl+F this crate for TODO and FIXME. ಥ‿ಥ
  • Guarantee 100% that no panic!s will occur.

Getting Started

To use it, you need an in-memory buffer containing valid ELF data. Parsing and loading an ELF is as easy as following these few steps:

  1. Have a byte slice containing all of your ELF data. This might originate from loading a file or from invoking include_bytes!("path/to/elf"). Note that shared objects (.so) are also ELF files.
  2. Call Elf::try_parse with your ELF slice. On success it will return a small parsed Elf struct.
  3. Call the Elf's mem_len and mem_align functions. Those will give you the layout information needed to allocate a buffer for the next step.
  4. Call Elf::try_load with the Elf struct and a mutable slice of your newly allocated memory. This will copy all necessary segments into the new memory region after zero-filling it first. On success, the result is a LoadedElf struct which holds a mutable borrow to your allocated memory.
  5. Call LoadedElf::try_reloc with a chosen virtual base address and an optional memory protection callback. The base address is where the final running program will think its first memory page is located. This allows you to re-locate an ELF from within a different address space. If you don't change the memory mapping of the loaded ELF, then the base address is the pointer of your allocated memory block's slice. You can get this pointer from LoadedElf::loader_base.
  6. The memory protection function receives base addresses, a slice, and the requested memory protection level. You can use this callback to actually apply memory protection flags as specified by the ELF data. Do not assume that protection regions won't overlap and just blindly handle each request in order.
  7. On success, the LoadedElf::try_reloc function returns a ReadyElf. This struct provides functions needed to run the ELF or grab its memory range.

Examples

// Grab a buffer containing ELF data.
let elf_data = get_aligned_buffer();

// Try parsing the buffer as ELF data.
let elf = Elf::try_parse(elf_data)?;

// For the next step, we need to allocate a bunch of page-aligned memory.
// You might as well use a pre-allocated buffer from your `.bss` section.
let align = elf.mem_align() as usize;
let size  = elf.mem_len()   as usize;
let mem   = alloc_aligned(size, align);

// Now, load the ELF into our allocated memory. After that, you are free to throw
// `elf_data` and `elf` out of the window.
let mut loaded_elf = elf.try_load(mem)?;
drop(elf);
drop(elf_data);

// The next step is to re-locate and memory-protect our ELF. To do that we first need
// a base address. If you intend to run the loaded ELF as a plugin in your own address
// space, you can use `loader_base` as a base address, which is just `mem.as_ptr()`.
// Otherwise, you need a base address within the loaded ELF's address space.
let base  = loaded_elf.loader_base();
let ready = match loaded_elf.try_reloc(base, Some(protection_fn)) {
    Ok(r) => r,

    // In case of an error, you get back your memory slice to de-allocate or inspect
    // it. You don't get a `LoadedElf` back for a retry, that one is consumed. The
    // reason for this is that `m` might already have been partially modified until
    // the error occurred. You can, however, just re-load `elf` if you did not yet
    // throw it away.
    Err((m, e)) => {
        dealloc(m);
        return Err(e.into());
    },
};

// Now you can grab an entry function pointer for whichever address space.
// Go on and have fun!
let main: fn() = unsafe { mem::transmute(ready.p_entry()) };
unsafe { (main)() };

// Done? Better not leak all the precious memory. Only you have control
// over all the allocations.
dealloc(ready.p_mem());

B-but why?!

I have two personal goals for this crate.

One is to use it in my toy OS to unpack the kernels etc. from the OS loader, without needing some fancy virtual init file system. Eventually, in a century, it might even load user-space applications.

The other goal is to use it as an embeddable dlopen-replacement for plugins with a minimum amount of dependencies. You could load the same ELF plugin into a Windows or Linux build of your application. Just exchange some v-tables on plugin entry.

In both cases, possibly illformed ELF data should not cause any kind of undefined or undesired behaviour, from parsing to re-locating. Your only security risk should be calling the entry function.