RFC: case-insensitive rename support in the VFS

RFC: case-insensitive rename support in the VFS

[Barry Naujok]

As some of you know, I've been working on CI support for XFS.
Currently, the Linux VFS does not make this very easy and neat.

One issue I've encountered is the case of rename in a CI
filesystem where the destination is the same as the source
(both parent and child) differing only in case.

So, currently, do_rename() calls do_path_lookup() which
gets the source and destination dentries, etc. The third
parameter, flags, is stored in the nameidata structure
which is passed to the filesystem implementations of
lookup.

vfs_rename() checks to see if the two dentry inodes are
the same, and if so, just returns 0.

I propose adding a 7th LOOKUP_xxx flag for the rename case
where we indicate we are after an exact (non-CI) match for
the destination name in the lookup call in do_rename.

Filesystem support in lookup would check for this flag
in addition to LOOKUP_CONTINUE not existing (ie. at the
last component during lookup) and if so, do an exact
lookup rather than a CI lookup as it normally would do.
Traditional non-CI filesystems would ignore this flag.

If this is deemed acceptable, or a better idea is proposed,
I can write a patch for further scrutiny.

Regards,
Barry.

Re: RFC: case-insensitive rename support in the VFS

[Christoph Hellwig]

On Fri, May 02, 2008 at 02:14:56PM +1000, Barry Naujok wrote:
> As some of you know, I've been working on CI support for XFS.
> Currently, the Linux VFS does not make this very easy and neat.
>
> One issue I've encountered is the case of rename in a CI
> filesystem where the destination is the same as the source
> (both parent and child) differing only in case.
>
> So, currently, do_rename() calls do_path_lookup() which
> gets the source and destination dentries, etc. The third
> parameter, flags, is stored in the nameidata structure
> which is passed to the filesystem implementations of
> lookup.
>
> vfs_rename() checks to see if the two dentry inodes are
> the same, and if so, just returns 0.
>
> I propose adding a 7th LOOKUP_xxx flag for the rename case
> where we indicate we are after an exact (non-CI) match for
> the destination name in the lookup call in do_rename.
>
> Filesystem support in lookup would check for this flag
> in addition to LOOKUP_CONTINUE not existing (ie. at the
> last component during lookup) and if so, do an exact
> lookup rather than a CI lookup as it normally would do.
> Traditional non-CI filesystems would ignore this flag.

I think passing a flag to the filesystem for this is fine, but
flag in nameidata won't do it since lookup is going to lose the
nameidata argument soon.