Re: mv: move files across bind mounts using the rename syscall

[Pádraig Brady]

On 06/12/2024 12:07, Milan Hauth wrote:
> currently mv fails to detect an identical filesystem
> of source and destination file across bind mounts
>
> instead of using the rename syscall
> mv copies the file (to the same filesystem)
> and then deletes the source file
>
> obviously, this is not ideal:
>
> - slow
> - waste of disk space
> - change of inode number
> - unnecessary disk write operations (destructive)
>
> > currently mv fails to detect an identical filesystem
> > of source and destination file across bind mounts
>
> this is possible using /proc/self/mountinfo
> proof of concept:
> https://github.com/milahu/move-files-across-bind-mounts
>
> related:
> https://serverfault.com/questions/327447
> https://unix.stackexchange.com/questions/406351
> https://unix.stackexchange.com/questions/380025

Thanks for taking the time to explore a solution.
It does seem a bit messy though, and I'd be worried about edge cases,
so I'm not sure it's appropriate for mv to do this.

Perhaps another approach might be for you to use BTRFS subvolumes,
which are similar to bind mounts.  When the rename() fails there,
mv will then try to reflink,unlink which should succeed (cheaply).

thanks,
Pádraig.