Can't unwind wasm frames on aarch64-darwin with system libunwind

[bnjbvr]

On Mac aarch64 (e.g. M1 MBP), the system's libunwind doesn't behave as on other platforms, and doesn't properly unwind wasm frames using the information we provide it (CFI etc.). The PR that introduces the switch from using signal handlers to mach ports to handle software traps has something that worked, using a custom build of libunwind and custom directives for linking against it.

Some ways to fix this:

• poke more at the system's libunwind, to try to understand what it wants. At some point its source code might be released on https://opensource.apple.com/, and if we can reproduce the issues observed with the system's libunwind, then we could investigate them with a debugger.
• I've had some discussions with Apple engineers who said these particular functions were not particularly used and tested by Apple. So we could look into other functions exposed by libunwind and try to use those instead. This would be pretty invasive, since that would require coordination with backtrace-rs too, which uses libunwind's functions as well.
• ship a custom build of libunwind, and link against it. I don't know backtrace-rs would also require the custom link annotations.
• implement our own frame unwinder (after all, we don't need a lot), or find another existing implementation that works equally well. There was some talk on the Rust's zulip to reimplement libunwind in Rust at some point, I don't know what came out of that.

[fitzgen]

Related: #2459

[akirilov-arm]

@bnjbvr @cfallin @fitzgen Is this issue still relevant? It affects the tests in tests/all/traps.rs in particular.

[fitzgen]

This is probably still relevant for when we emit unwind info, which is a configuration we still want to support to integrate with external tools like GDB/profilers/etc, but it is no longer an issue that is critical for correctness (e.g. reclaiming GC garbage).

[bnjbvr]

This is also used to compute backtraces, right? I'll note to try it again on my side.

[fitzgen]

Backtraces are now captured by walking frame pointers since #4431

[bnjbvr]

Very sweet! I've just retried on my mac m1, and indeed these tests now pass after uncommenting the cfg_attr (I'll submit a PR), and in our embedding I can see a full stack trace \o/
Thanks @fitzgen for working on this!

[akirilov-arm]

FYI recently I did a little investigation for my back-edge CFI work, and the only remaining user I could find of the backtrace crate, and by extension the system unwinder, is the wasmtime-fiber crate. In fact, that is just a test dependency, and, more precisely, only one test needs it - backtrace_traces_to_host.

Of course, there is still the issue about integrating with external tools such as debuggers.

[fitzgen]

the only remaining user I could find of the backtrace crate, and by extension the system unwinder, is the wasmtime-fiber crate. In fact, that is just a test dependency, and, more precisely, only one test needs it - backtrace_traces_to_host.

FWIW, I think this is 100% okay. We shouldn't be concerned about dev deps in the same way we are about regular deps.

Of course, there is still the issue about integrating with external tools such as debuggers.

We do still emit native unwind info by default, but also have a wasmtime::Config knob to turn it off should an embedder choose to.

[fitzgen]

Can we close this issue now? @bnjbvr I can't remember: did you submit the PR you mentioned you'd make in #2808 (comment)? I don't see any cfg_attr in tests/all/traps.rs anymore so I assume so.

[bnjbvr]

Yep, I did submit this PR, so we could consider this closed. I didn't link this issue in the PR because the native unwinders would still require this to work properly, but maybe that can be deferred to a future issue.