Upgrade addr2line and properly handle split DWARF on Linux

[khuey]

This doesn't actually implement loading for split DWARF but it will make backtrace continue to compile with the existing functionality once addr2line cuts a new release.

Cc @philipc

[philipc]

addr2line 0.20.0 has been released.

[khuey]

Ok the latest csets I've pushed support packed split DWARF on Linux. They don't yet support unpacked split DWARF (that's harder because we need to load the DWOs and presumably stash them somewhere). And bumping addr2line causes the MSRV test to start failing.

[philipc]

Is it possible to add a CI test for this?

They don't yet support unpacked split DWARF (that's harder because we need to load the DWOs and presumably stash them somewhere).

We could hash by DWO id, but an alternative would be to change SplitDwarfLoad to include the ResUnit index, and add an API to query the number of ResUnits so that we can preallocate a Vec for lazy loading (similar to unpacked Mach-O).

[khuey]

There is a CI test for this.

[philipc]

There is a CI test for this.

Oh, so there is, but I must be missing something still. How was it passing before this PR?

[khuey]

What you're missing is that it's not passing before this PR. These tests are failing on the tip of this repo.

[khuey]

so that we can preallocate a Vec for lazy loading (similar to unpacked Mach-O).

What's the point of preallocating the Vec here? search_object_map still takes &mut self to modify it. It looks like Mach-O gives indexes to tell you which object file to use, so having something you can index is helpful?

DWARF just gives us file paths and DWO ids. I think we'd want a HashMap here, but that's not available in alloc. We could do a BTreeMap on the paths, or just linearly search a Vec. Regardless preallocating doesn't seem to do any good with DWARF, unless I'm missing something.

[philipc]

The difference from Mach-O that I overlooked is that addr2line is internally doing the indexing itself. I don't think we actually need to look up anything. addr2line isn't going to ask for the same DWO twice. All we need is somewhere to stash it before returning it to addr2line, so we could just append to a Vec.

[khuey]

Hmm, yes, addr2line's ResUnit will cache the gimli::Dwarf for the DWO, and backtrace's Context contains addr2line's Context, so that does sound right.

Perhaps instead of having mmap_aux and mmap_dwp in Stash we should just have an append-only Vec<Mmap> like we do for Stash::buffers.

[khuey]

And that adds support for unpacked split DWARF, passing all the relevant tests.

[khuey]

Does @philipc review changes for this repo too?

[philipc]

I don't approve PRs in this repo. LGTM though.

[khuey]

It would be nice to get this merged and a new release cut. Split DWARF inlining was turned off by default in 1.68.

[JohnTitor]

r? @rust-lang/crate-maintainers could someone review this? I think this is kinda notable change but I'm not familiar enough to review exhaustively.

[khuey]

Realistically I don't know that you're going to find a reviewer with more subject matter expertise than me and @philipc here.

[Amanieu]

LGTM to my limited understanding of DWARF. I trust @philipc to have reviewed this properly.

[khuey]

Can we get a new release with this?

[workingjubilee]

It would be nice if we passed the test suite on the commit we cut our release from. I will make sure a release is cut once #531 is resolved.

[workingjubilee]

backtrace 0.3.68 is out, let us know if there are any bugs to be squashed!