Armin Ronacher's Thoughts and Writings

How to WASM DWARF

written on Monday, November 30, 2020

So you're excited about WebAssembly? You're not alone, many are. WebAssembly has huge opportunities that extend way past just compiling non JavaScript code to make it run in the browser. The reason for this is that it is quickly becoming a widely supported compilation target for a range of runtimes. This lets one ship a custom WebAssembly runtime to run your stuff and then update the code that is running on it trivially across a variety of environments. It's a more modern take of Java's “compile once run everywhere” philosophy.

Not excited about WebAssembly yet? You really should be though. There are so many projects now that show the power of it. From emulating Flash to edge computing, WebAssembly is showing up everywhere. WebAssembly thus also makes a perfect tool for custom plugin systems in cross platform applications.

There is one catch though: if you start distributing more and more complex WebAssembly targets to the edge (or browser) and you want to do fast iteration, you are going to experience crashes and errors in production that haven't shown up during development. So you need a crash reporting tool. If we emphasize the “web” in WebAssembly for a second then one instantly starts thinking about source maps. Source maps are what enables crash reporting of minified JavaScript in production. Source maps take the location in a minified JavaScript file and let us figure out where that pointed to in the original source file.

With WebAssembly that simplified approach breaks down. I don't want to explain for too long while source maps are not appropriate and why DWARF is what we should all be interested in but gladly I don't have to. DWARF is now slowly being adopted by more and more and we're about to support it for WebAssembly on Sentry. A lot of what went into this post is extracted from notes I took while working on WebAssembly support for our crash reporting system.

Short WebAssembly Primer

Before we can dive into debugging WebAssembly we need to talk about how WebAssembly actually works. A WebAssembly runtime is a stack machine (kinda) that can execute WebAssembly instructions at near native speeds. It does this in a way that is optimized to allow interoperability with the outside world, even if the outside world is a JavaScript browser environment.

If you have never looked at the internals of WebAssembly you can find lots of explanations online about how it works. However one of the probably most interesting ways to get excited about it is a talk titled “A talk Near the Future of Python” by David Beazley where he live codes a WASM interpreter in Python. The reason the talk is a great way to dive into WASM is because it shows really well the basics of what it's all about.

WebAssembly while being quite simple comes in a few flavors. The first flavor dimension is the address size. Today all of WebAssembly you will encounter is dubbed "wasm32". That's the flavor of WebAssembly with 32 bit pointer address size. There is also a 64bit variant but basically nobody supports it and because JavaScript has no 64bit integers there will be likely some restrictions in interoperability with the JavaScript world. In addition we have a second dimension which is the representation of WebAssembly on disk / in source: text (also called WAT) and binary (normally called WASM). Today when you open a browser console and try to debug WebAssembly most likely you will be presented with the textual representation of WebAssembly most of the time as the browser will "decompile" the binary into the textual representation if it has no other debug information available.

The difference between WASM and WAT is going to play a pretty significant role later when we talk about debugging. For now you need to know that WAT does not play a significant role when working with WASM for anything other than as an intermediate format and for debugging.

To run WebAssembly you need a runtime. The runtime executes the instructions and provides an interface to the outside world. The most popular runtimes are node and your browser. Whenever you want to do something fancy in your application (like IO or really anything that sounds like a syscall) the runtime needs to "inject" that into your module. This sounds like wild west and in a way it is. To allow interoperability of different runtimes some standards are emerging. The most prominent one is called “WASI”.

How to WebAssembly

So how do I get a WebAssembly thing compiled? At the moment the language of choice appears to be Rust because of the excellent tooling around it that makes starting with WebAssembly pretty straightforward. You just need to add the wasm32-unknown-unknown target and you're ready to go:

$ rustup target add wasm32-unknown-unknown
$ cargo new --lib wasm-example

Then you need to change the library target to cdylib in the generated Cargo.toml and you're ready to roll:

$ cargo build --target=wasm32-unknown-unknown
   Compiling wasm-example v0.1.0 (/Users/mitsuhiko/Development/wasm-example)
    Finished dev [unoptimized + debuginfo] target(s) in 0.63s

This produces a file called target/wasm32-unknown-unknown/debug/wasm-example.wasm. If you were to use a tool that can dump out contents of .wasm files you would see that it consists of all the general WASM sections (like Code) but also custom sections called .debug_frame. That's because rustc like many other tools now can emit DWARF for WASM. Point being: we're at the point where the out of the box tooling experience can produce DWARF debug information.

DWARF is a debug format that can help us make sense of the binary. In particular it lets us figure out where a specific instruction in our binary points to in the source files. So for instance it might tell us that at 0x53 there is a function called foobar declared in file example/src/foobar.rs in line 42. More importantly it can also tell is if a function call was "virtual" in the sense that it was inlined. Inline information is particularly important in release builds where small function calls are typically fully inlined. Without that information we would not see where the actual crash happened.

What the DWARF?

I already talked a bit about DWARF but in practice this requires some explanation. DWARF is a debug standard that has been embraced on most platforms. The notable exception for that is the Microsoft ecosystem where two competing file formats prevail: native PDBs with embedded CodeView and portable PDBs for .NET. Neither of those file formats use DWARF. It's not clear to me right now if Microsoft will support DWARF once they start supporting WebAssembly on their toolchain so there is hoping. That said, everywhere else we have been exposed to the different versions of DWARF for many years now so tooling is pretty good.

DWARF is not a file format in itself, it's a standard that defines a lot of different aspects of debugging. Because it's not a file format it requires a container to put this information in. On Linux for instance DWARF information is embedded in ELF files, on iOS/macOS and other Apple platforms it's embedded in Mach-O binaries. On all those platforms it's also common to split these files in two. That often leaves one ELF file behind with the code you run and a separate ELF debug file just containing the DWARF information.

One added complexity here is that you often need access to both files if you want to do certain types of debugging happen. For instance to produce a stack trace out of a memory dump you don't just need the DWARF data, you also need the executable. The reason for this is that that the process of creating a stack trace is also something that the executable itself needs for a lot of languages. So for instance C++ has exceptions and in order to throw them, it needs to “unwind” the stack. For that it uses on some platforms a derived version of DWARF embedded in the binary as eh_frame. Since often that information is not retained in the debug files we typically need both.

To match those two files together the concept of “build IDs” (also called “debug IDs” and/or “code IDs”) has been established. In Mach-O binaries they are prominently stored as a header in the Mach-O file and are called LC_UUID. In ELF binaries two systems are used: the more modern the NT_GNU_BUILD_ID ELF note in the program headers or the more legacy .note.gnu.build-id ELF section. The same concept also exists for PDBs on Windows executables contain an ID that uniquely defines the PDB that goes with it.

So how do we do this on WASM? It turns out for WASM there is no standard yet. I proposed one which is what we currently support on Sentry. Basically we embed a custom section called build_id into the WASM file containing a UUID. When the binary is "stripped" (that is, the debug data is removed for size or intellectual property concerns) the build_id section remains in both files so we can match them together. This is particularly important when debug files are stored in a central location like a symbol server. Sentry for instance will look up the debug data exclusively with the build_id at any symbol server configured or in a customer's uploaded symbol repository. Due to a quirk in the WASM spec it's important that the debug file made available to Sentry or other crash reporting tools retains all original data including the Code sections. More about that later.

Stack Traces and Instruction Addresses

So earlier we talked about WebAssembly being a stack machine. Unlike what you encounter in the real world most of the time WebAssembly does not have registers and it does not have a unified virtual address space. This poses some challenges to DWARF but not insurmountable ones. To understand the problem let's look at how a normal binary works and then how WebAssembly works.

When you're on Linux and you compile a program you typically end up linking in some other code too. Your binary might thus once it's loaded into memory also refer to other dynamic libraries. Every function that exists has a unique address in the same address space as your variables. This is typically referred to as a von Neumann architecture. One of the effects of this is that I can normally take the address of a function and then figure out based on the address of the function from which module the function came. For instance I might see that from 0x1000 to 0x5000 all functions come from a library called utils.dylib. Simplified speaking if I see that my CPU crashed in 0x1024 I can just look into the debug information for utils.dylib and look for 0x1024 - 0x1000 and see what it tells me about.

With WebAssembly we have two immediate problems. First of all code and data live are separated. This is generally called a Harvard architecture. Functions in WASM are as far as the runtime is concerned referenced by name or index. The “address” of a function is not a thing that WASM understands. It might be something that would be nice to have for the language that compiles down to it though. For instance it's very common to take the address of a function in C++ and put it into a variable. The other place where function addresses show up is typically in stack traces. When you generate a stack trace in most native languages and operating systems you end up with something that looks like a list of instruction addresses that point directly into functions. Since everything is in a huge memory space no issues here. In WebAssembly we might be offset within a function so we need to know which function index we're in and how far we managed to execute with in that function.

Generally it's currently not possible to generate stack traces in most WebAssembly environments unless a custom implementation was made. What is possible in Browsers is to register a function with the WebAssembly module that is implemented in JavaScript and raises an exception just to catch the stringified stack trace:

function getStackTrace() {
    try {
        throw new Error();
    } catch (e) {
        return e.stack;
    }
}

When this function is passed to WebAssembly the target module written in Rust or C++ can parse that string to figure out what the stack observed by the web assembly runtime looks like. So what does such a stack look like?

Typically it looks something like this (in fact at least the WebAssembly frames are standardized across browsers):

getStackTrace@http://localhost:8002/:23:13
example@http://localhost:8002/wasm-example.wasm:wasm-function[1]:0x8c
@http://localhost:8002/:37:9

Here we can see that our wasm-example was modified to call into the getStackTrace function from above. What's important here is that the WebAssembly code tells us a) the name of the function when available, the URL of the WebAssembly module, the index of the function and a hexadecimal address. The latter is particularly important and we will get to it shortly but first let's think about the URL a bit. Remember the example above from where I talked about having a dynamic library mapped into your executable? Now imagine we do the same in WebAssembly. We have a WebAssembly module linked to another one at runtime. This could result in a stack trace where two different WebAssembly modules show up in the stack trace. The only thing that tells them apart is the URL. In both files we will find a function with the index 1 and in both files we are likely to find some code at address 0x8c.

In the same way as stack traces are not defined in WebAssembly there is also no concept of dynamic linking or working with multiple modules. While it is possible to dynamically link there is no API to work with it. This is not too different from how dynamic linking wasn't really ever standardized in C either. For instance at Sentry we need to implement custom code for all platforms to figure out which dynamic library sits at what address. On macOS we need to work with the Apple dynamic linker and parse Mach-O files, on Linux we need to parse /proc/maps and reach ELF files etc. The only advantage on those platforms is that because all code ends up in the same address space it is pretty trivial to accomplish this.

In WebAssembly the situation is a lot more complex. First of all it looks like the only way we can keep these modules apart is the file name. So how do we map from file name (or URL) to the handle of our WebAssembly object? We effectively need to establish such a mapping in our runtime. When we want to do this in the browser we quickly run into the limitation that no such API exists. The way we work around this currently is to monkey patch the WebAssembly.instanciateStreaming function and hope that the caller passes a fetch result which contains a URL. Once we have a handle for the WebAssembly module we can also access the build_id custom section to read the build ID.

The next thing we need to figure out is this instruction address. As we talked about before WebAssembly doesn't really have this address space for instructions either. In fact there are two formats for WebAssembly: text (wat) and binary (wasm). The DWARF standard for it came later and thus the desire to talk about addresses appeared for the first time. To make it easier for DWARF tools the debug information is encoded in byte offsets in the original source file. This means that if you serialize a WASM module to binary or text, the DWARF offsets would have to be different. To make matters a bit more confusing, DWARF is specified to encode the addresses in offsets within the ``Code`` section. This creates again a bit of a problem because browsers (and other runtimes) report the offset within the entire WASM file instead. Since the Code section never sits at the beginning of the file, there is always going to be an offset between them. In the same way as browsers do not provide an API to access the URL of a WebAssembly module, they also do not provide a way to access the offset of the Code section. This means that crash reporting tools are required to be able to operate with the absolute offset in the WASM file instead. This in turn means that the WASM debug file must not remove the Code section or other sections or we would have to add a second section that holds the original code offset.

For for instance if the address 0x8c is reported, but code starts at 0x80 the actual address reported in the DWARF file is 0xc.

Wasm-Bindgen and Friends

So cool, now we walked through all of this. Unfortunately there is more. When you start out with just rustc or another compiler you can get a binary with DWARF data and all is well. However once you work with tools which open WASM files and serialize them back you're in a bit of a pickle because they either destroy the debug info or remove it. For instance a common crate to use in the Rust ecosystem is wasm-bindgen which can help work with browser APIs. The issue with it is that wasm-bindgen completely destroys the DWARF debug info as it's based on walrus which does not yet implement debug information tracking. If your project involves a tool like that, you're currently out of luck getting DWARF debug information going.

Custom Hacks

So what do we do at Sentry do now if you want WASM debugging going?

  • We provide you with a tool called wasm-split which can add a build_id section if missing and splits debug information into a separate file and strips it from the release file leaving you with two files: a debug file containing everything and a library file without debug sections.
  • We monkey-patch the browser's WebAssembly.instanciateStreaming functions (Pending integration into the JavaScript SDK) to keep track of URLs to build IDs.
  • We rebase all pointers on the server to be relative within the DWARF file and back.

Wishlist and Future Direction

So what does this leave us with? Really overall we're getting there! Sentry can now give you proper stack traces with source code even:

Unfortunately there are many restrictions I wish we would not have to live with:

  • When a WebAssembly module is loaded dynamically there is no way for us to work with it. That's because on the one hand stack traces in browsers do not provide the build_id directly so we have no way to uniquely identify the file, secondly because there is also no way to get the URL of a WebAssembly.Module. I really wish there was an API for that.
  • Speaking of build_ids: I wish everybody would embrace them and they would become a standard.
  • The fact that we have code section relative offsets within DWARF files but no way to get the offset to the Code section in browsers seems suboptimal for people who want to use separate debug files.
  • I wish wasm-bindgen and the libraries it's build on would start supporting DWARF.
  • I wish there was a standardized API to get stack traces even from within WebAssembly (WASI etc.) so Rust code like the backtrace crate could provide stack traces for error logging and more.

To end on a positive note: the ecosystem around WebAssembly — particularly in the Rust world — is amazing. In fact in general the ecosystem for working with DWARF data in Rust is top notch and always a pleasure to work with. There are a lot of people working on making everybody's debugging experience as good as possible and that work is rarely honored. Every developer knows and wants to have stack traces, yet very few people are comparatively working on enabling this functionality.

This entry was tagged rust and wasm