fix: copy_dir and copy_file to preserve symlinks instead of following them

[cachebag]

This PR employs symlink preservation in copy_file and adds it to copy_dir.

PR #1504 added symlink preservation to fix crashes in lldb-preview, which contains internal symlinks (liblldb was loading twice due to broken symlinks). Then, PR #1521 later changed copy_file to create symlinks pointing to the source path (for Homebrew integration), which may have regressed #1504.

This PR restores #1504's behavior: read the symlink target and preserve it. What I'm curious about is what the intended behavior for #1521 was because shouldn't users be able to run rustup-init --no-self-update? Is that not the recommended way to manage rustup via external package managers?

If this PR steers out of scope of the goals of rustup I'm happy to close it but I feel like the former behavior is preferable for most use cases.

[rami3l]

Thanks for making this patch! Reasonable support for external packagers are still expected, and whether a manual setup is required is largely based on the packagers' opinion: https://rust-lang.github.io/rustup/installation/other.html#using-a-package-manager

I have the feeling that we need to somehow support both cases based on whether that's copying the original binary or not...

[rami3l]

Do you have a specific use case where rustup is not working as you'd expect, which this PR allows to fix?

If you can find such a case, I agree with using this new implementation except for this:

rustup/src/cli/self_update.rs

Line 845 in 3382f6f

utils::copy_file(&this_exe_path, &rustup_path)?;

#1521 is probably too widespread as a fix, and it makes perfect sense to use different copy implementations during self installation and during toolchain installation.

Suggest refining the test cases accordingly if going this way.

[cachebag]

Do you have a specific use case where rustup is not working as you'd expect, which this PR allows to fix?

I don't, and I only encountered this looking at the code. It's my understanding that copying symlinks breaks their identity and can cause software to load libraries twice (like the crash that was fixed ad-hoc above). It's also just correct behavior afaik; cp -R preserves symlinks by default.

#1521 is probably too widespread as a fix, and it makes perfect sense to use different copy implementations during self installation and during toolchain installation.

I would agree with this and will adjust, so long as you think it's a reasonable patch. I figured at the least I could flag this to note the variance in functionality for copy's.

[rami3l]

@cachebag Thanks for your clarification. I think although it has no direct influence now, judging by what I'd expect from the function API then indeed, the old behavior feels more correct to me, and leaving it uncorrected might lead to misuse in the long run.

Looking back there is still a slight possibility that we could be missing other cases apart from self installation, so I'd say let's go on polishing this patch by carefully analyzing the call sites. Please don't hesitate to ask me if you don't feel certain about anything!

[cachebag]

@cachebag Thanks for your clarification. I think although it has no direct influence now, judging by what I'd expect from the function API then indeed, the old behavior feels more correct to me, and leaving it uncorrected might lead to misuse in the long run.

Looking back there is still a slight possibility that we could be missing other cases apart from self installation, so I'd say let's go on polishing this patch by carefully analyzing the call sites. Please don't hesitate to ask me if you don't feel certain about anything!

That makes sense to me, thank you!

[rami3l]

LGTM modulo a few lints, many thanks!