Re: [PATCH 1/2] VFS: change kern_path_locked() and user_path_locked_at() to never return negative dentry

["NeilBrown" <neilb@suse.de>]

On Fri, 07 Feb 2025, Kent Overstreet wrote:
> On Fri, Feb 07, 2025 at 02:36:47PM +1100, NeilBrown wrote:
> > No callers of kern_path_locked() or user_path_locked_at() want a
> > negative dentry.  So change them to return -ENOENT instead.  This
> > simplifies callers.
> > 
> > This results in a subtle change to bcachefs in that an ioctl will now
> > return -ENOENT in preference to -EXDEV.  I believe this restores the
> > behaviour to what it was prior to
> 
> I'm not following how the code change matches the commit message?

Maybe it doesn't.  Let me checked.

Two of the possible error returns from bch2_ioctl_subvolume_destroy(),
which implements the BCH_IOCTL_SUBVOLUME_DESTROY ioctl, are -ENOENT and
-EXDEV.

-ENOENT is returned if the path named in arg.dst_ptr cannot be found.
-EXDEV is returned if the filesystem on which that path exists is not
 the one that the ioctl is called on.

If the target filesystem is "/foo" and the path given is "/bar/baz" and
/bar exists but /bar/baz does not, then user_path_locked_at or
user_path_at will return a negative dentry corresponding to the
(non-existent) name "baz" in /bar.

In this case the dentry exists so the filesystem on which it was found
can be tested, but the dentry is negative.  So both -ENOENT and -EXDEV
are credible return values.


- before bbe6a7c899e7 the -EXDEV is tested immediately after the call
  to user_path_att() so there is no chance that ENOENT will be returned.
  I cannot actually find where ENOENT could be returned ...  but that
  doesn't really matter now.

- after that patch .... again the -EXDEV test comes first. That isn't
  what I remember.  I must have misread it.

- after my patch user_path_locked_at() will return -ENOENT if the whole
  name cannot be found.  So now you get -ENOENT instead of -EXDEV.

So with my patch, ENOENT always wins, and it was never like that before.
Thanks for challenging me!

Do you think there could be a problem with changing the error returned
in this circumstance? i.e. if you try to destroy a subvolume with a
non-existant name on a different filesystem could getting -ENOENT
instead of -EXDEV be noticed?

Thanks,
NeilBrown