[RFC PATCH v1 1/7] fs: Add inode_get_ino() and implement get_ino() for NFS

[Mickaël Salaün]

On Mon, Oct 14, 2024 at 04:45:00PM +0200, Christian Brauner wrote:
> On Sun, Oct 13, 2024 at 06:17:43AM -0400, Jeff Layton wrote:
> > On Fri, 2024-10-11 at 17:30 +0200, Mickaël Salaün wrote:
> > > On Fri, Oct 11, 2024 at 07:39:33AM -0700, Christoph Hellwig wrote:
> > > > On Fri, Oct 11, 2024 at 03:52:42PM +0200, Mickaël Salaün wrote:
> > > > > > > Yes, but how do you call getattr() without a path?
> > > > > > 
> > > > > > You don't because inode numbers are irrelevant without the path.
> > > > > 
> > > > > They are for kernel messages and audit logs.  Please take a look at the
> > > > > use cases with the other patches.
> > > > 
> > > > It still is useless.  E.g. btrfs has duplicate inode numbers due to
> > > > subvolumes.
> > > 
> > > At least it reflects what users see.
> > > 
> > > > 
> > > > If you want a better pretty but not useful value just work on making
> > > > i_ino 64-bits wide, which is long overdue.
> > > 
> > > That would require too much work for me, and this would be a pain to
> > > backport to all stable kernels.
> > > 
> > 
> > Would it though? Adding this new inode operation seems sub-optimal.
> 
> I agree.

Of course it would be better to fix the root of the issue.  Who is up
for the challenge?

> 
> > Inode numbers are static information. Once an inode number is set on an
> > inode it basically never changes.  This patchset will turn all of those
> > direct inode->i_ino fetches into a pointer chase for the new inode
> > operation, which will then almost always just result in a direct fetch.
> 
> Yup.
> 
> > A better solution here would be to make inode->i_ino a u64, and just
> > fix up all of the places that touch it to expect that. Then, just
> 
> I would like us to try and see to make this happen. I really dislike
> that struct inode is full of non-explicity types.

Also, it would be OK to backport it, right?