speice.io/_posts/2019-02-14-insane-allocators.md
2019-02-16 20:46:25 -05:00

5.4 KiB

layout title description category tags
post Insane Allocators: segfaults in safe Rust ...and what it means to be "safe." rust, memory

Having recently spent a lot of time down rabbit holes looking at how Rust uses memory, I like to think I finally understand the rules well enough to break them. See, Rust will go so far as to claim:

If all you do is write Safe Rust, you will never have to worry about type-safety or memory-safety. You will never endure a dangling pointer, a use-after-free, or any other kind of Undefined Behavior.

-- The Nomicon

...and subject to (relatively infrequent) borrow checker bugs, it's correct. There's ongoing work to formalize the rules and prove that Rust is safe, but for our purposes it's a reasonable assumption.

Until it isn't. It's totally possible for "safe" Rust programs (under contrived circumstances) to encounter memory corruption. It's even possible for these programs to "segfault" when using an unmodified compiler:

Wait, wat?

Wat indeed.

There are two tricks used to pull this off. First, I'm making use of a special environment variable in Linux called LD_PRELOAD. Matt Godbolt goes into way more detail than I can cover, but the important bit is this: I can insert my own code in place of functions typically implemented by the C standard library.

Second, there's a very special implementation of malloc that is being picked up by LD_PRELOAD:

use std::ffi::c_void;
use std::ptr::null_mut;

// Start off with an empty pointer
static mut ALLOC: *mut c_void = null_mut();

#[no_mangle]
pub extern "C" fn malloc(size: usize) -> *mut c_void {
    unsafe { 
        // If we've never allocated anything, ask the operating system
        // for some memory...
        if ALLOC == null_mut() {
            // Use a `libc` binding to avoid recursive malloc calls
            ALLOC = libc::malloc(size)
        }
        // ...and then give that same section of memory to everyone
        // for all subsequent allocations, corrupting the location.
        return ALLOC;
    }
    // Note that we don't ever handle `free`; if the first object
    // we allocate gets freed, the memory address being given
    // to everyone becomes a "use-after-free" bug.
}

Because this implementation of malloc is intentionally broken, every program run using this library will crash. And I mean every program; if you use dynamic memory, you're going down.

So how is it possible to run the Rust compiler in this environment? LD_PRELOAD applies to all programs, so the compiler should also encounter memory corruption and crash, right? The answer is that sudo deletes environment variables like LD_PRELOAD and LD_LIBRARY_PATH when running commands. While it is technically possible to crash sudo in the same way using our evil malloc implementation, the default policy is to delete these variables because of security concerns.

Finally, why does Rust 1.31 work, and 1.32 fail? The answer is in the release notes: jemalloc is removed by default. In Rust 1.28 through 1.31, programs were statically compiled against jemalloc by default; our evil malloc implementation never gets invoked because the program goes straight to the operating system to request memory. However, it's still possible to trigger segfaults in Rust programs from 1.28 - 1.31 by using the System global allocator. Rust programs prior to 1.28 aren't subject to this LD_PRELOAD trick.

So what?

I do want to clarify: the code demonstrated here isn't a security issue, and doesn't call into question Rust's definition of "safe." The code demonstrated here crashes because the memory allocator is lying to it. Even in mission critical systems, there are a lot of concerns beyond allocators; the F-35 Joint Strike Fighter coding standards give memory allocation about 10 sentences total.

But this example does highlight an assumption of Rust's memory model that I haven't seen discussed much: safe Rust is safe if, and only if, the allocator it relies on is "correct". And because writing a non-trivial allocator is fundamentally unsafe, safe Rust will always rely on unsafe Rust somewhere.

That all said, know that "safe" Rust can claim to be safe because it stands on the shoulders of incredible developers working on jemalloc, kmalloc, and others. Without being able to trust the allocators, we wouldn't be able to trust the promise of safe Rust. So to all the people who make the safety promises of Rust possible - thanks.