Add workaround for sporadic kills when building on Macos

[charlesroussel]

This is the workaround for the issue #10060

[rust-highfive]

Thanks for the pull request, and welcome! The Rust team is excited to review your changes, and you should hear from @alexcrichton (or someone else) soon.

Please see the contribution instructions for more information.

[ehuss]

Thanks for the PR! Can you say why this is specific to arm? I have been experiencing this on x86_64 myself.

[charlesroussel]

Thanks for the PR! Can you say why this is specific to arm? I have been experiencing this on x86_64 myself.

I thought it was linked to ARM and the fact that all applications requiring to being signed. Apparently it is only due to APFS.
I'll update the review to include macos version from 11+ instead of arm only.

[alexcrichton]

This seems like a particularly heavy hammer for this issue since file copying can have a pretty significant performance impact on Cargo. It sounds like this is just intentionally slowing down Cargo to let other processes in the system settle which also would mean that this isn't actually solving the issue, it's just making it more rare. I don't know what a "proper" fix would look like since I'm not sure what we could synchronize on, but otherwise I think it would be useful to quantify what kind of a slowdown this will look like.

[hkratz]

@alexcrichton fs::copy() tries to use the CoW fclonefileat on macOS, which appears to be not slower than hardlinking on APFS filesystems.

[alexcrichton]

Oh nice! To confirm, have you benchmarked a bit before and after and seen no difference?

[hkratz]

On a Macbook Air M1, Monterey 12.1, cargo 1.58.0-nightly (40dc281 2021-12-06) I tested it with the already existing __CARGO_COPY_DONT_LINK_DO_NOT_USE_THIS env variable on cargo itself - difference not measurable.

With hardlinks:

$ cargo clean; cargo -q build
# confirm paths
$ ls -i target/debug/cargo target/debug/deps/cargo-90dea0042da6ad93
4196983 target/debug/cargo
4196983 target/debug/deps/cargo-90dea0042da6ad93
# confirm test methodology
$ rm target/debug/cargo; cargo -q build
$ ls -i target/debug/cargo target/debug/deps/cargo-90dea0042da6ad93
4196983 target/debug/cargo
4196983 target/debug/deps/cargo-90dea0042da6ad93
# !! Note same inodes -> hardlinking used

# benchmark
$ hyperfine -w 3 "rm target/debug/cargo; cargo -q build"
Benchmark #1: rm target/debug/cargo; cargo -q build
  Time (mean ± σ):      98.1 ms ±   0.4 ms    [User: 65.3 ms, System: 31.3 ms]
  Range (min … max):    97.3 ms …  99.3 ms    29 runs

With CoW copies:

# 2) CoW copy performance
$ export __CARGO_COPY_DONT_LINK_DO_NOT_USE_THIS=1
# confirm test methodology
$ rm target/debug/cargo; cargo -q build
$ ls -i target/debug/cargo target/debug/deps/cargo-90dea0042da6ad93
4197238 target/debug/cargo
4196983 target/debug/deps/cargo-90dea0042da6ad93
# !! Note different inodes -> clone/copy used

# benchmark
$ hyperfine -w 3 "rm target/debug/cargo; cargo -q build"
Benchmark #1: rm target/debug/cargo; cargo -q build
  Time (mean ± σ):      98.5 ms ±   0.6 ms    [User: 65.4 ms, System: 31.5 ms]
  Range (min … max):    97.7 ms …  99.8 ms    29 runs

I regularly clone multiple GB files with cp -c and it is instantaneous.

[alexcrichton]

I forget precisely where and when we use hard links vs copies, but once we invoke rustc all bets are off because rustc is by far always going to be slower than Cargo. Could you try benchmarking builds where Cargo doesn't actually do anything? (where it basically just determines that the project is up-to-date). I think projects with a number of medium-sized binaries would be good to benchmark for that.

[charlesroussel]

I tried on rust-analyser project

Clean build without and with removing target binary. The first one is a clean build to have everything built.

unset __CARGO_COPY_DONT_LINK_DO_NOT_USE_THIS
cargo build
hyperfine -w 3 "cargo build"
Benchmark 1: cargo build
  Time (mean ± σ):     130.7 ms ±   4.3 ms    [User: 97.7 ms, System: 30.9 ms]
  Range (min … max):   127.1 ms … 143.7 ms    22 runs

hyperfine -w 3 "rm target/debug/rust-analyser;cargo build"
Benchmark 1: rm target/debug/rust-analyser;cargo build
  Time (mean ± σ):     130.9 ms ±   2.2 ms    [User: 97.7 ms, System: 31.4 ms]
  Range (min … max):   128.3 ms … 136.6 ms    22 runs

export __CARGO_COPY_DONT_LINK_DO_NOT_USE_THIS=1
cargo build
hyperfine -w 3 "cargo build"
Benchmark 1: cargo build
  Time (mean ± σ):     129.8 ms ±   3.8 ms    [User: 96.8 ms, System: 30.4 ms]
  Range (min … max):   126.9 ms … 145.3 ms    22 runs

hyperfine -w 3 "rm target/debug/rust-analyser;cargo build"
Benchmark 1: rm target/debug/rust-analyser;cargo build
  Time (mean ± σ):     130.3 ms ±   0.9 ms    [User: 96.9 ms, System: 31.5 ms]
  Range (min … max):   128.2 ms … 131.5 ms    22 runs

There is no apparent difference when cargo "does nothing"

[alexcrichton]

Awesome, thanks!

Could you update the comment a bit with the performance implication where using the CoW-enabled syscall this should theoretically be just as fast as the hard linking version?

[charlesroussel]

Awesome, thanks!

Could you update the comment a bit with the performance implication where using the CoW-enabled syscall this should theoretically be just as fast as the hard linking version?

Done

[alexcrichton]

@bors: r+

👍

[bors]

📌 Commit c9c67c0 has been approved by alexcrichton

[bors]

⌛ Testing commit c9c67c0 with merge 180f599...

[bors]

☀️ Test successful - checks-actions
Approved by: alexcrichton
Pushing 180f599 to master...