Stack Unwinding
When a Rust program encounters an unrecoverable error, it panic!()
s. The behavior of a panic is intentionally not defined in Rust, and may vary based on the platform and strategy.
In general, Rust defines two "panic strategies":
abort
: After printing a stacktrace (defined by a#[panic_handler]
), the program will exit immediately. No memory cleanup is performed. This form of panic is truly unrecoverable.unwind
: Rather than exiting immediately, the program's stack memory will be unwound, allowing for a more graceful exit. Every active struct instance'sDrop
implementation will be called, and panics created outside of the main thread can be caught and handled.
no_std
and Panics: The current approach.
Panic behavior is platform-dependent. On a more "traditional" platform target, we have the luxury of an operating system with I/O utilities and a well-defined allocator. On an embedded target such as armv7a-VEXos-eabi
(the platform target defined by vexide
), Rust takes no assumptions and leaves the panic implementation up to us.
In order to build a barebones Rust program on bare metal, we must define a panic handler:
#[panic_handler]
fn panic(info: &PanicInfo) -> ! {
loop {}
}
In this (extremely barebones) example, if panic!()
were to be called, the panic handler would simply spinlock the CPU indefinitely. If this were an enviornment where libc
was available (PROS has access to newlib libc) we could call libc::exit(1)
.
vexide currently defines a basic panic handler under the assumption of the abort
strategy:
- It prints a panic message over stdout through serial.
- It optionally displays a brief error message on the brain screen.
- Finally, it exits the user program using PROS' wrapper over over the
vexSystemExitRequest()
SDK call.
The panic implementation does not perform stack unwinding or any cleanup, as armv7a-VEXos-eabi
's panic_strategy is currently abort
.
Unwinding on ARM
As it turns out, it might be possible to support "panic_strategy": "unwind"
on our bare metal ARM target. PROS links against libgcc
, which defines primitives for dealing with exception handling, including the parts needed to implement an unwinding panic handler in Rust. Unfortunately, stack unwinding has a very complicated and advanced implementation. The best we have to go off of is the libpanic
implementation over libgcc
used internally by rustc. This piece of code is a pretty obscure part of Rust's runtime. It's not well documented and is supposedly reverse-engineered from other language implementations.
Essentially, in order to implement an unwinding panic, we must first override the eh_personality
language item which is called by Rust internally when panicking using unwind
. The #[eh_personality]
routine is then responsible for hooking into libgcc's ARM unwinding functions for actually unwinding the stack. These functions are actually part of the Itanium C++ Exception Handling ABI.
In a previous attempt to reimplement unwinding panics, that item looked something like this:
#![feature(lang_items)]
#[lang = "eh_personality"]
#[no_mangle]
unsafe extern "C" fn rust_eh_personality(
state: _Unwind_State,
exception_object: *mut _Unwind_Exception,
context: *mut _Unwind_Context,
) -> _Unwind_Reason_Code {
...
}
PROS Additionally has its own unwinding exception handler.
Other Alternatives
The unwinding
crate provides pure-rust implementations of exception handlers that we could use in place of a libgcc implementation. Unfortunately, ARM support isn't available in this crate yet.