Swatinem Blog Resume

Files need Identity

— 11 min

Interestingly, the same theme has come up multiple times recently within Sentry. I myself recently wrote a Sentry RFC about SourceMap DebugIds. And at the same time, I was supporting and advising other teams working on Java Source Context, and Flutter Obfuscation.

All these different initiatives have the following in common: You have multiple build artifacts for a single application build. These artifacts together form a tight unit.

Whereas SourceBundles are a Sentry invention, the other two use-cases are being implemented by external tools. And they lack a strong association of the different artifacts / assets that form one final build output.

A SourceMap is just some JSON, so is the Flutter obfuscation mapping, though a little different which makes it harder to deal with, more in a minute.

We need the obfuscation mapping to be able to de-obfuscate, so far so good. But with a few different versions of apps being installed and used by end users, how do we know which obfuscation mapping we need?

That is where file identity comes in. Each group of tightly coupled build artifacts needs to be uniquely identified somehow, so we are able to find the matching file we need, no matter if that is a SourceMap, a SourceBundle, or an obfuscation mapping.

To achieve that, each artifact needs to have a unique identifier. It is also very beneficial if that unique identifier is embedded in that file, so it becomes self identifying.

This is the problem with the Flutter obfuscation mapping. It is a JSON file, but with an array at its root. There is no way to extend that file with another field at the root that includes this identifier. Well, too bad I guess :-(

Lets say we have not only two tightly coupled build artifacts but more. To stick with the Flutter example, it might be the case that a Flutter-web build outputs both a minified JS file, a corresponding SourceMap, and an obfuscation mapping.

Two of those files are just JSON. Our proposal for SourceMap DebugIds I linked above proposes to add a new field to the SourceMap with its unique identifier. It is pretty much impossible to extend the obfuscation mapping however. But lets ignore that problem for now. In the end we have two JSON files. How do we tell them apart then?

Each file needs to have some form of marker in it that tells us what kind of file it is. For JSON files, the JSON Schema "$schema" field naturally presents itself. Authoring a full JSON Schema might not be everyones cup of tea, and that is not the point here. The point is that this unique "$schema" field tells us what kind of file we are looking at. By having such a field, the file becomes self describing.

A SourceMap just happens to be a SourceMap if it has a "version": 3 field, and a "mappings" field. It might be very unlikely, but any random JSON file could potentially have these fields and then be wrongly interpreted as a SourceMap.

To summarize this section, every file should be self identifying, by embedding some kind of unique identifier, and it should also be self describing by embedding some kind of marker that describes the kind (or format) of the file.

With these two pieces of information, we can upload any file to any dumb storage service and look it up.


But how do we know which file to look up? Let us come back to the example from before. Lets assume we have an obfuscated Flutter app running on some customer device, and it produces an obfuscated stack trace that is uploaded to Sentry or any other service. How will Sentry know which obfuscation mapping to use?

To be able to do so, the report that has the obfuscated stack trace also has to provide the unique identifier of the obfuscation mapping. We can then look up the mapping using that unique identifier and correctly deobfuscate the stack trace.

So we need a way to get access to that unique identifier at runtime. Surprisingly, this is the most complex part of our Flutter example, as well as the most controversial thing about our SourceMap proposal. Ideally, the Platform (whatever it is) offers a programmatic API that provides this unique identifier.

It is totally possible to have a different unique identifier for each accompanying artifact, for example a different identifier for an associated SourceMap, and obfuscation mapping. Though I strongly advise to have one unique identifier that is shared among these tightly coupled artifacts.

To summarize, we have some self identifying and self describing artifacts that we will just stash away on some dumb storage service, and we need a way at runtime to query that unique identifier.

# Native Inspiration

The native ecosystem has most of this figured out to various degrees, lets take a look.

To start this off, binary file formats are usually self describing by starting off with a magic-byte sequence that identifies the file format. Our native platforms each have their own executable formats for example.

On macOS, we have Mach-O files which pretty consistently have a unique identifier called LC_UUID (for load command). The executables are also commonly split into a main executable, and an associated debug file called dSYM. Both share the same unique identifier. However, both have the Mach-O format.

As this first example shows, the file format on its own is not enough to identify the file kind / purpose. However by looking at the presence of various sections in that file, one can quite confidently say if it is an executable, or the corresponding debug file.

Linux has ELF (executable and linker format) files. These files can have a unique identifier called NT_GNU_BUILD_ID (NT for note), though it is sadly frequently missing. The executables are not split by default as they are produced by build tools, but developers frequently split them apart manually. Again, the two files have the same file format, but it is possible to tell their purpose apart by looking at the various sections. When splitting those files apart, both retain the same unique identifier.

The situation on Windows is slightly different. An executable in PE (portable executable) format has its own identifying which is the combination of the Timestamp and SizeOfImage header values. This can hardly be called unique though. This file can then reference a PDB (program database) file via a DebugDirectoryEntry which contains the unique identifier of the PDB file. One thing here that tools frequently get wrong is that one executable can have multiple DebugDirectoryEntry entries, referencing more than one debug files. I wrote about that previously in a post titled Format Ossification, because most tools got so used to only ever seeing zero or one DebugDirectoryEntrys, the fact that there can be in fact more than one got completely lost.

In summary, the native formats are pretty good at self identifying.

# Symbol Lookup

One thing I mentioned before is being able to easily find and download these debug files from any dumb storage service. The native ecosystem offers mainly two possibilities here.

In the Linux ecosystem, we have debuginfod which defines a simple lookup scheme. One can simply download the /buildid/{BUILDID}/debuginfo file and get the debuginfo for a uniquely identified executable. There is public debuginfod servers for every major Linux distribution as well.

Then there is the symstore Server and accompanying SSQP (simple symbol query protocol), which is primarily used for the Windows ecosystem, but does support other ecosystems as well. The lookup scheme has support for a ton of formats, including lookup for ELF and Mach-O files using their corresponding unique identifiers.

One problem with symstore though becomes obvious looking at the scheme for PE files: <filename>/<Timestamp><SizeOfImage>/<filename> As I mentioned, the Timestamp and SizeOfImage combination might not be unique enough. So just combine it with the filename, problem solved, right? Well this creates new problems all on its own. For example Electron hosts its own symbol server. But what happens if you ship an electron app and rename the main electron.exe file? Well too bad, you can’t find that symbol anymore. This is indeed a real pain for Sentry customers.

For macOS, the situation is pure sadness. Apple does not host any public symbol server, and the licensing around these things is also unclear. Sentry goes through great pain to maintain its own internal symbol server for Apple symbols, but it is a frequent source of problems, with a brittle pipeline for scraping the symbols, and frequent problems with symbols missing.

# Programmatic API

This is another source of sadness. None of the native platforms have builtin platform support to get at these unique identifiers. Also getting at the list of all the loaded libraries is a huge pain on some platforms.

For each platforms, getting at the unique identifiers involves manually reading the platform native file format headers and chasing references around, which can be unsafe as it involves a lot of pointer arithmetic. The problem with these file formats is also that they are extremely badly documented. I wonder how it is possible that they are so well understood, although the DebugDirectoryEntry situation makes me doubtful.

The formats and the structures you have to read are not documented publicly on the internet. They are defined in some platform specific headers that are primarily only available on that platform. For example on Windows, the PE definitions are part of the Windows SDK. For macOS, I believe the Mach-O headers are provided by Xcode. One critical bit to get the unique identifier of a PDB file is also famously missing from the Windows SDK headers. The CodeView record is not defined anywhere. All the tools just copy-paste the definition into their own code from somewhere. The ELF format however is reasonably well specified and documented in various man pages, though those are far from easily usable.

And even if we have a couple of header definitions that we won’t find on the public web, it is C headers. Can someone tell me again how many bits a C long has? No? Thought so.

Well, I’m not going on a ranting spree about how un-portable C is. The point I’m trying to make is that reading the unique identifiers of files at runtime is a huge pain and involves unsafe code. It would be so much nicer if we had built-in platform APIs that let us easily enumerate all the loaded libraries, and easily get their unique identifiers.

# In summary

And with that I am at the end of todays post. What I would love to have from a tool developers perspective is:

The SourceMap RFC I mentioned in the beginning tries to solve two of these problems, and I would love to get feedback on it.

One thing from the RFC that still needs discussion is how to generate these unique identifiers. Our current draft implementation generates a new random UUID each time, which I argue is a bad idea.

I would like these identifiers, and the files themselves to be bit-for-bit deterministic / reproducible given the same inputs. The Portable PDB specification mentions explicitly how to create the unique identifier:

the checksum is calculated by hashing the entire content of the PDB file with the PDB ID set to 0 (20 zeroed bytes).

So there is precedent in other ecosystems for reproducible and deterministic unique identifiers.