Make permissions and ownership? configurable while persisting

[xaverdh]

Related issues: #30, #114

It might not be possible to support on all platforms, but it would still be nice to have it at least on unix. This issue is specifically about permissions / ownership of the persisted file rather than the temporary file, but I would also be quite happy with an option to configure permissions on the temporary, if they get carried over to the persisted file.

My use case is to atomically write files with given permissions / ownership by first writing to a temporary file and then persisting it (relying on rename likely being atomic), so just changing these after the fact will slightly defeat the purpose..

[xaverdh]

Implementation wise, I think the most realistic approach is to make permissions on the temporary file configurable and then make persist keep those. Ownership could be configured through the Builder, but would only come into effect when persisting.

On Linux this would correspond to passing O_TMPFILE, but without O_EXCL and creating it with given mode. Then just before persisting it with linkat you would call fchmod to adjust ownership. I don't know how well this fits with other platforms though..

[Stebalien]

As far as I know, persist should keep the permissions. So you should be able to:

1. Create a named temporary file.
2. Chmod the temporary file.
3. Call persist.

On Linux this would correspond to passing O_TMPFILE, but without O_EXCL and creating it with given mode. Then just before persisting it with linkat you would call fchmod to adjust ownership.

This library does not support persisting temporary files created with O_TMPFILE, unfortunately. I'd like to implement this, but I have a policy of only providing features that are truly cross-platform (linux, windows, macos, freebsd), including sharing equivalent security guarantees.

[D1plo1d]

This library does not support persisting temporary files created with O_TMPFILE, unfortunately. I'd like to implement this, but I have a policy of only providing features that are truly cross-platform (linux, windows, macos, freebsd), including sharing equivalent security guarantees.

Would it be possible to remove the O_EXCL flag entirely? Is there a security difference from eg. windows or some other OS if it is removed? It appears that according to the gnu libc docs O_EXCL | O_TMPFILE only serve to prevent persisting tempfiles and it's removal would allow developers to persist tempfiles externally to tempfile.rs using nix.

For example without O_EXCL the following works:

tmpfile::tmpfile();
let tmp_file_path = format!("/proc/self/fd/{}", tmp_file.as_raw_fd());

nix::unistd::linkat(
    None,
    &tmp_file_path[..],
    None,
    &file_path[..],
    nix::unistd::LinkatFlags::SymlinkFollow,
)?;

[Stebalien]

IIRC, the problem is MacOS and the BDSs. On those platforms, temporary files are implemented by creating then immediately unlinking, and there's no way to "relink" (last I checked).

[D1plo1d]

Ok yeah, so I'm wondering if removing the flag preventing the file from being copied on platforms where it can be - is that a feature or a secruity concern?

I don't know if the absense of a flag is a feature - maybe this is pedantic but it feels like a move to be more inline with the OS defaults IMHO.

I'm trying to imagine the malicious use of this but I haven't been able to come up with anything yet.

So I don't know: If we only removed that flag what would we be potentially disrupting for users of tmpfile?

[Stebalien]

Ok yeah, so I'm wondering if removing the flag preventing the file from being copied on platforms where it can be - is that a feature or a secruity concern?

Ah, no. I'm happy to accept a patch to remove that (~assuming it works cross-platform...~ linux only so it should be fine).

[Byron]

For gitoxide it would be most beneficial if the mode could be set for the temporary file at creation time to avoid a follow-up chmod call for increased performance by reducing the amount of syscalls.

Git does the same by default.

As per the policy of providing cross-platform features only, implementing this would be out of scope though. Would you be open to a contribution if this was an exceptional case?

Thank you 🙏