Duplicate artifact tracking issue.

[ehuss]

#6308 added a check if multiple jobs produce the same files. To be safe, it is currently a warning. At some point in the future it should be turned into a hard error if no problems are reported. Collisions are almost certainly bad behavior that will cause problems, so rejecting it is probably the right thing to do.

Related issues:

• Warn on duplicate artifact creation #5524: Request to warn on duplicate artifact creation — wherein .rmeta files started to collide because bin targets started producing them, and were corrupting one another if built simultaneously. .rmeta collisions fixed in Don't hardlink rmeta files. #6292.
• What to do about duplicate targets in a workspace? #6293: Handling multiple targets in a workspace with the same name.
• Race condition when trying to compile a lib multiple times for different purposes. #5444: Hardlink collisions when target is built multiple times (like panic/no-panic). Fixed by Be more conservative about which files are linked to the output dir. #5460 by only hardlinking what was requested.

Known situations where collisions may occur:

• Multiple binary/example/lib targets in a workspace with the same name. (Or reusing a shared target directory.)
• Using --out-dir with an example and binary with the same name.
• Multiple path dependencies with dylibs with the same name.
• Multiple dependencies with the same name and you select both of them on the command-line, for example: cargo build -p rand:0.4.3 -p rand:0.5.5.
• rustdoc in a workspace where multiple crates have the same name. This can arise from a variety of situations (renamed dependencies, multiple versions of a package, different packages with the same crate names, etc.). (Multiple crates with the same name lead to conflicting target/doc/crate directories rust#56169, rustdoc: builds including multiple crates with the same name are always dirty rust#61378)
• panic="abort" and cdylibs and tests: Create a project with a lib (cdylib crate type), binary, and an integration test, with panic="abort" in the profile. When cargo test runs, the cdylib is built twice (once with panic=abort for the binary, and once without for the test), with the same filename. Building the lib for the test should probably skip the cdylib crate type (assuming rlib is also available), but implementing this is very difficult. See Duplicate artifact tracking issue. #6313 (comment).
• Multiple targets (particularly executables) that differ only by case on case-insensitive filesystems.
• PDB collisions on Windows. If the package has a binary and a dylib library of the same name, then the .pdb file for each target will have the same name. This is easy to hit if you have a proc-macro package with a binary.
• Dylib built multiple times with different features
  • for example with new feature resolver, see Undefined symbols for x86_64 when using resolver="2" #9278.
  • and with workspace members, see Fingerprint of dependency in workspace changes when running cargo build and cargo build -p <crate> #12345
• Shared cdylib built with different profile settings, such as a build script and release mode Duplicate artifact tracking issue. #6313 (comment)

Notes on implementation issues:

• Cargo is hard-coded with the outputs that rustc produces. If those outputs change, it will not be able to catch those changes. In particular, Warn on duplicate artifact creation #5524 would not have been caught by these checks.
• OutputFile is currently not calculated correctly in some cases. Known issues:
  • Debug files (like .dSYM) are not tracked in some cases because TargetInfo::file_types is making decisions about what should be hardlinked too early. Fixed in Rework rustc output file tracking. #8210
  • It is also very likely that it is missing certain platform differences.
  • Doc outputs are not tracked correctly (it generates incorrect paths). Fixed in Catch filename output collisions in rustdoc.  #6998.
  • Doctests generate incorrect paths. This should be fixed. Fixed in Catch filename output collisions in rustdoc.  #6998, doc tests do not have any output paths.
• Not all outputs are checked. Such as:
  • The .d dep info file. This uses the same hash as the main artifact, so is unlikely to be a problem.
  • Incremental files. One hopes that the hashes used in rustc are good enough?
  • Other temp files created by rustc, like rcgu files, which should always include the hash.
  • Anything done by build scripts.
• cargo doc has a dedicated path for detecting collisions. If cargo doc is ever updated to support multiple crates with the same names, this code path can be removed.

[kamathba]

I have an interesting scenario where we are getting this warning that feels a little like a false positive.

We have a project that we can build with x86_64-unknown-linux-gnu or x86_64-unknown-linux-musl. We use a docker container (rust-musl-cross) typically when building for the musl target, since the dependencies are a little harder to set up on everyone's machine. This is also how we build/package in CI. When building with the gnu taget, our output folders are typically target/debug and target/release. When building for the musl target, the output folders are namespaced under target/x86_64-unknown-linux-musl. This diagnostic considers it a warning if using the gnu target I run:

cargo build --release -Z unstable-options --out-dir target/release

Now the why for us doing this has to do more with either nuances or our misunderstandings of how the cargo-deb and cargo-rpm packages work. Since we only ever package from the musl toolchain, we could probably hard-code the paths to target/x86_64-unknown-linux-musl/release/foo inside our Cargo.toml files, but the above scenario seems like a detectable situation where you've selected the output directory that things were going to end up in anyways. It also manifests as a confusing warning:

(using x86_64-unknown-linux-gnu)
cargo build --release -Z unstable-options --out-dir target/release/ 
warning: `--out-dir` filename collision.
The lib target `foo` in package `foo v0.2.0 (/home/bkamath/foo/foo)` has the same output filename as the lib target `foo` in package `foo v0.2.0 (/home/bkamath/foo/foo)`.
Colliding filename is: /home/bkamath/foo/foo/target/release/libfoo.rlib
The exported filenames should be unique.

The same warnings happen for bin targets as well. The two things seemingly being compared in this case aren't created by different targets/build.

[ehuss]

@kamathba Can you clarify some things about your use case for me? I'm trying to understand why you are using --out-dir at all. Are you manually setting the list of files to be packaged? It looks like with cargo-deb you can just specify target/release paths and it rewrites them when cross compiling, so using --out-dir shouldn't be necessary. Can you explain more why you need it?

[kamathba]

@ehuss I think its mostly because of us using both the x86_64-unknown-linux-gnu and x86_64-unknown-linux-musl targets and we want everything to "just work". Also, we use cargo-rpm, which doesn't seem as smart about picking up cross-compile settings and has limited options (which we could probably fix).

All that said, I'm more wondering why cargo can't elide when the --out-dir is the same as the directory it was going to place things in anyways?

As I'm digging into this answer, I'm wondering why our builds default to building for x86_64-unknown-linux-musl even though we don't specify it in .cargo/config. We'd like to not use --out-dir, as its one of the only reasons we aren't on stable now that clippy/rustfmt landed.

[jjpe]

As described here I have a crate in which I'd like to produce both a dylib and a cdylib file.

This seems impossible at the moment. Specifically, Cargo writes both dylibs to the same filename i.e. it seems to write one and then it seems to overwrite it with the other (can't tell you which is which without digging into the final file though).
If it is possible to have Cargo produce both cdylib and dylib, I've yet to find out how, as adding
crate-type = ["cdylib", "dylib"] to the [lib] section of the crate gives a compile warning leading to this issue.

[Lokathor]

I'm trying to produce rlib, staticlib, and cdylib just to make sure that all modes are buildable (because I've had some tricky build problems once or twice already) and I'm getting a warning from cargo test that told me to come here.

Also, cargo test fails to run if I don't run cargo build first because it tries to link to target/debug/deps/my_lib.dll and fails to find it.

[ehuss]

@Lokathor Can you share a minimum example to demonstrate? I'm not sure how to reproduce your issue. Also, which platform are you on?

[Lokathor]

how about a build log?
here's Linux, https://travis-ci.org/Lokathor/thorium/jobs/515790083

the same general warning happens for the mac build too, which you can also find there

on windows, https://ci.appveyor.com/project/Lokathor/thorium/builds/23603097/job/sia07tfchtthym5a (im not 100% sure this is the exact same commit, but it should be)

I don't know how to make a more minimal example. I actually don't even care except for the scary warning that it might become an error later.

[ehuss]

@Lokathor OK, I see what is happening. It is a consequence of having cdylib and panic="abort" with an integration test and a binary. For reasons, cdylibs don't have a unique hash added to their filename. In this case, the "thorium" library needs to be built twice (once with panic="abort" for main.rs, and once without for the test which doesn't allow panic="abort"), and both of those end up with the same filename.

I can't think of an easy workaround for you, unfortunately.

I'm not sure what the best approach here is. Perhaps dylibs could be placed in unique directory names. Also, maybe Cargo could be more conservative what it builds for tests, since it is only linking against one lib type.

[Lokathor]

I only need the dylib for debug and/or release, and even then only if the dynamic_link feature is enabled. It's just a development ability, not meant to be shipped.

Unfortunately, cargo features are currently quite boring and cannot control much. What I would like is if features could control more.