speice.io/_posts/2019-02-14-insane-allocators.md

117 lines
5.4 KiB
Markdown
Raw Normal View History

2019-02-16 16:10:58 -05:00
---
layout: post
title: "Insane Allocators: segfaults in safe Rust"
2019-02-16 20:46:25 -05:00
description: "...and what it means to be \"safe.\""
2019-02-16 16:10:58 -05:00
category: rust, memory
tags: []
---
Having recently spent a lot of time down rabbit holes looking at how
2019-02-16 16:10:58 -05:00
[Rust uses memory](/2019/02/understanding-allocations-in-rust.html),
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](https://doc.rust-lang.org/nomicon/meet-safe-and-unsafe.html)
...and subject to (relatively infrequent)
[borrow checker bugs](https://github.com/rust-lang/rust/labels/A-borrow-checker),
it's correct. There's ongoing work to [formalize](https://plv.mpi-sws.org/rustbelt/popl18/)
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"](https://en.wikipedia.org/wiki/Segmentation_fault)
when using an unmodified compiler:
2019-02-16 16:10:58 -05:00
<script id="asciicast-ENIpRYpdDazCkppanf3LSCetX" src="https://asciinema.org/a/ENIpRYpdDazCkppanf3LSCetX.js" async></script>
# Wait, wat?
[Wat indeed.](https://www.destroyallsoftware.com/talks/wat)
2019-02-16 16:10:58 -05:00
There are two tricks used to pull this off. First, I'm making
use of a special environment variable in Linux called
[`LD_PRELOAD`](https://blog.fpmurphy.com/2012/09/all-about-ld_preload.html).
Matt Godbolt goes into [way more detail](https://www.youtube.com/watch?v=dOfucXtyEsU)
than I can cover, but the important bit is this: I can insert my own code in place of
2019-02-16 16:10:58 -05:00
functions typically implemented by the [C standard library](https://www.gnu.org/software/libc/).
Second, there's a very special implementation of [`malloc`](https://linux.die.net/man/3/malloc)
that is being picked up by `LD_PRELOAD`:
```rust
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
2019-02-16 16:10:58 -05:00
ALLOC = libc::malloc(size)
}
// ...and then give that same section of memory to everyone
// for all subsequent allocations, corrupting the location.
2019-02-16 16:10:58 -05:00
return ALLOC;
}
2019-02-16 20:46:25 -05:00
// 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.
2019-02-16 16:10:58 -05:00
}
```
2019-02-16 20:46:25 -05:00
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?
2019-02-16 20:46:25 -05:00
`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
2019-02-16 20:46:25 -05:00
`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](https://blog.rust-lang.org/2019/01/17/Rust-1.32.0.html#jemalloc-is-removed-by-default).
In Rust 1.28 through 1.31, programs were statically compiled against
[jemalloc](http://jemalloc.net/) 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`](https://doc.rust-lang.org/std/alloc/struct.System.html)
global allocator. Rust programs prior to 1.28 aren't subject to this
`LD_PRELOAD` trick.
# So what?
2019-02-16 20:46:25 -05:00
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](http://www.stroustrup.com/JSF-AV-rules.pdf)
2019-02-16 20:46:25 -05:00
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](https://doc.rust-lang.org/std/alloc/trait.GlobalAlloc.html#unsafety),
safe Rust will always rely on unsafe Rust somewhere.
2019-02-16 20:46:25 -05:00
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](https://linux-kernel-labs.github.io/master/labs/kernel_api.html#memory-allocation),
2019-02-16 20:46:25 -05:00
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.