Nov 25, 2019 • RustEditsPermalink

How to Panic in Rust

What exactly happens when you panic!()? I recently spent a lot of time looking at the parts of the standard library concerned with this, and it turns out the answer is quite complicated! I have not been able to find docs explaining the high-level picture of panicking in Rust, so this feels worth writing down.

(Shameless plug: the reason I looked at this is that @Aaron1011 implemented unwinding support in Miri. I wanted to see that in Miri since forever and never had the time to implement it myself, so it was really great to see someone just submit PRs for that out of the blue. After a lot of review rounds, this landed just recently. There are still some rough edges, but the foundations are solid.)

The purpose of this post is to document the high-level structure and the relevant interfaces that come into play on the Rust side of this. The actual mechanism of unwinding is a totally different matter (and one that I am not qualified to speak about).

Note: This post describes panicking as of this commit. Many of the interfaces described here are unstable internal details of libstd, and subject to change any time.

High-level structure

When trying to figure out how panicking works by reading the code in libstd, one can easily get lost in the maze. There are multiple layers of indirection that are only put together by the linker, there is the #[panic_handler] attribute and the “panic runtime” (controlled by the panic strategy, which is set via -C panic) and “panic hooks”, and it turns out panicking in #[no_std] context takes an entirely different code path… there is just a lot going on. To make things worse, the RFC describing panic hooks calls them “panic handler”, but that term has since been re-purposed.

I think the best place to start are the interfaces controlling the two indirections:

Both of these interfaces are implemented through extern blocks: listd/libcore, respectively, just import some function that they delegate to, and somewhere entirely else in the crate tree, that function gets implemented. The import only gets resolved during linking; looking locally at that code, one cannot tell where the actual implementation of the respective interface lives. No wonder that I got lost several times along the way.

In the following, both of these interfaces will come up a lot; when you get confused, the first thing to check is if you just mixed up panic handler and panic runtime. (And remember there’s also panic hooks, we will get to those.) That happens to me all the time.

Moreover, core::panic! and std::panic! are not the same; as we will see, they take very different code paths. libcore and libstd each implement their own way to cause panics:

Let us now look at these pieces in a bit more detail.

Panic Runtime

The interface to the panic runtime (introduced by this RFC) is a function __rust_start_panic(payload: usize) -> u32 that gets imported by libstd, and that is later resolved by the linker.

The usize argument here actually is a *mut &mut dyn core::panic::BoxMeUp – this is where the “payload” of the panic (the information available when it gets caught) gets passed in. BoxMeUp is an unstable internal implementation detail, but looking at the trait we can see that all it really does is wrap a dyn Any + Send, which is the type of the panic payload as returned by catch_unwind and thread::spawn. BoxMeUp::take_box returns a Box<dyn Any + Send>, but as a raw pointer (because Box is not available in the context where this trait is defined); BoxMeUp::get just borrows the contents.

The two implementations of this interface Rust ships with are libpanic_unwind for -C panic=unwind (the default on most platforms) and libpanic_abort for -C panic=abort.

std::panic!

On top of the panic runtime interface, libstd implements the default Rust panic machinery in the internal std::panicking module.

rust_panic_with_hook

The key function that almost everything passes through is rust_panic_with_hook:

fn rust_panic_with_hook(
    payload: &mut dyn BoxMeUp,
    message: Option<&fmt::Arguments<'_>>,
    file_line_col: &(&str, u32, u32),
) -> !

This function takes a panic source location, an optional unformatted panic message (see the fmt::Arguments docs), and a payload.

Its main job is to call whatever the current panic hook is. Panic hooks have a PanicInfo argument, so we need a panic source location, format information for a panic message, and a payload. This matches rust_panic_with_hook’s arguments quite nicely! file_line_col and message can be used directly for the first two elements; payload gets turned into &(dyn Any + Send) through the BoxMeUp interface.

Interestingly, the default panic hook entirely ignores the message; what you actually see printed is the payload downcast to &str or String (whatever works). Supposedly, the caller should ensure that formatting message, if present, gives the same result. (And the ones we discuss below do ensure this.)

Finally, rust_panic_with_hook dispatches to the current panic runtime. At this point, only the payload is still relevant – and that is important: message (as the '_ lifetime indicates) may contain short-lived references, but the panic payload will be propagated up the stack and hence must be 'static. The 'static bound is quite well hidden there, but after a while I realized that Any implies 'static (and remember dyn BoxMeUp is just used to obtain a Box<dyn Any + Send>).

libstd panicking entry points

rust_panic_with_hook is a private function to std::panicking; the module provides three entry points on top of this central function, and one that circumvents it:

Panic Handler

All of the std::panic! machinery is really useful, but it relies on heap allocations through Box which is not always available. To give libcore a way to cause panics, panic handlers were introduced. As we have seen, if libstd is available, it provides an implementation of that interface to wire core::panic! into the libstd panic machinery.

The interface to the panic handler is a function fn panic(info: &core::panic::PanicInfo) -> ! that libcore imports and that is later resolved by the linker. The PanicInfo type is the same as for panic hooks: it contains a panic source location, a panic message, and a payload (a dyn Any + Send). The panic message is represented as fmt::Arguments, i.e., a format string with its arguments that has not been formatted yet.

core::panic!

On top of the panic handler interface, libcore provides a minimal panic API. The core::panic! macro creates a fmt::Arguments which is then passed to the panic handler. No formatting happens here as that would require heap allocations; this is why PanicInfo contains an “uninterpreted” format string with its arguments.

Curiously, the payload field of the PanicInfo that gets passed to the panic handler is always set to a dummy value. This explains why the libstd panic handler ignores the payload (and instead constructs a new payload from the message), but that makes me wonder why that field is part of the panic handler API in the first place. Another consequence of this is that core::panic!("message") and std::panic!("message") (the variants without any formatting) actually result in very different panics: the former gets turned into fmt::Arguments, passed through the panic handler interface, and then libstd creates a String payload by formatting it. The latter, however, directly uses the &str as a payload, and the message field remains None (as already mentioned).

Some elements of the libcore panic API are lang items because the compiler inserts calls to these functions during code generation:

Conclusion

We have walked through 4 layers of APIs, 2 of which are indirected through imported function calls and resolved by the linker. That’s quite a journey! But we have reached the end now. I hope you didn’t panic yourself along the way. ;)

I mentioned some things as being surprising. Turns out they all have to do with the fact that panic hooks and panic handlers share the PanicInfo struct in their interface, which contains both an optional not-yet-formatted message and a type-erased payload:

Reading the panic handler RFC, it seems like the plan was for core::panic! to also support arbitrary payloads, but so far that has not materialized. However, even with that future extension, I think we have the invariant that when message is Some, then either payload == &NoPayload (so the payload is redundant) or payload is the formatted message (so the message is redundant). I wonder if there is any case where both fields will be useful – and if not, couldn’t we encode that by making them two variants of an enum? There are probably good reasons against that proposal and for the current design; it would be great to get them documented somewhere. :)

There is a lot more to say, but at this point, I invite you to follow the links to the source code that I included above. With the high-level structure in mind, you should be able to follow that code. If people think this overview would be worth putting somewhere more permanently, I’d be happy to work this blog post into some form of docs – I am not sure what would be a good place for those, though. And if you find any mistakes in what I wrote, please let me know!

Posted on Ralf's Ramblings on Nov 25, 2019.
Comments? Drop me a mail or leave a note in the forum!