fakeroot invalidates build cache

[ishitatsuyuki]

In Arch Linux, we use fakeroot to do make install while packaging.

This however seems to invalidate built artifacts, building it again.

Please check the rules of invalidation. While bootstrap is preserved, anything other like stage0 is completely built from scratch.

Random notes: fakeroot preserves HOME, which should work well with cargo.

[Mark-Simulacrum]

Could you provide more details? A log of RUST_LOG=cargo::metadata=info would be helpful.

[ishitatsuyuki]

I don't think that logging will help, since the problem is probably inside rustbuild itself, resulting in cleaning the built artifacts.

[kennytm]

@ishitatsuyuki Could you provide details how to reproduce the problem? Can this PKGBUILD file reproduce it?

[ishitatsuyuki]

@kennytm this is confirmed to build twice, so probably using that helps.

I have no idea if the "official" AUR version does too. It's quite diverged.

[Mark-Simulacrum]

Okay, then a run with RUST_LOG=bootstrap=debug would be nice. I asked for the cargo metadata log since if it isn't inside bootstrap, it would be helpful; I don't think there are any assurances where the problem lies. I'll take a look at the script itself within a week or two probably, though I'm not sure if I have access to a platform that will work with it. Shouldn't be too hard to spin up a VM though.

[Mark-Simulacrum]

It seems that I was correct asking for a cargo fingerprint log. I reproduced the problem in a docker container with minimal trouble; the log is here: https://gist.github.com/Mark-Simulacrum/e089a2727a4f0f7ed290d17eb6330017#file-log-arch-39951-L468. The line selected is where the first stale fingerprint appears to be. I'm uncertain whether this is rustbuild's, Cargo's or the fakeroot's fault though. @alexcrichton may be able to clarify more based on the log though.

[ishitatsuyuki]

Thanks for the log.

I suspect Git submodule handling is trashing the timestamps. Will investigate why it happens.

[ishitatsuyuki]

It seems at least git -C reset hard ran between fakeroot rewrites (unlink and open(O_WRONLY)) every file. I feel it like file ownership, but running git with sudo doesn't yield such weird behavior. Any thoughts?

[Mark-Simulacrum]

INFO:cargo::ops::cargo_rustc::fingerprint: fingerprint error for compiler_builtins v0.0.0 (file:///home/mark/src/rust/src/libcompiler_builtins): mtime based components have changed: previously Some(FileTime { seconds: 1495761200, nanos: 142046176 }) now None, paths are "/home/mark/src/rust/build/x86_64-unknown-linux-gnu/stage0-std/x86_64-unknown-linux-gnu/release/build/compiler_builtins-7332fe221efc4ba8/output" and "/home/mark/src/rust/build/x86_64-unknown-linux-gnu/stage0-std/x86_64-unknown-linux-gnu/release/build/compiler_builtins-7332fe221efc4ba8/output"

That build output leads me to believe that maybe fakeroot maybe makes the stdlib's mtime commands not succeed? I don't really know. I can do some tests later...

[jyn514]

This seems likely to be an issue with fakeroot, not bootstrap or cargo, the timestamps really are different than before.

The issue tracker is not a support forum. For future questions please ask in one of these places:

https://discord.gg/rust-lang/
https://discord.gg/rust-lang-community/
https://users.rust-lang.org/
https://rust-lang.zulipchat.com/