Re: [PATCH] fs: allow rename across bind mounts on same superblock

[Al Viro <viro@zeniv.linux.org.uk>]

On Tue, Dec 23, 2025 at 08:27:58PM +0200, Andrei Topala wrote:
> Rename would not work if the user doesn't have the right permissions. The
> patch calls `mnt_want_write()` on both mounts and uses each side's idmap
> for permission checks.
> 
> This would make renames faster in this case. Currently, users who have
> write permission on both mounts get EXDEV and must fall back to copying the
> data, which can be slow for large files.
> 
> I initially attempted to handle this in userspace (in coreutils mv), but
> it's more complex there. It requires parsing /proc/self/mountinfo to
> translate paths through bind mount roots, and it only benefits mv. Other
> tools that call rename() directly would hit the slow path.
> 
> Is there a specific isolation scenario where allowing rename between
> same-superblock mounts would be problematic? I'd like to understand the
> concern better.

Consider fun with moving a subdirectory of a mounted subtree to another
mounted subtree, while some process has been chdired into it, for
starters...