fs: don't fail __d_obtain_alias() for unlinked inodes

Commit 669bb5af15d4 ("fs: make sure to call fsnotify_inoderemove() only
once when inode is dead") did two things. It made dentry_unlink_inode()
call fsnotify_inoderemove() only once the inode has neither links nor
aliases left. And it made __d_obtain_alias() fail with -ESTALE for an
inode in that state so that open_by_handle_at() can't create a new alias
after FS_DELETE_SELF has been reported and the inode marks are gone.

The second part is wrong for network filesystems. __d_obtain_alias()
backs d_obtain_alias() and d_obtain_root(). Those are what the
->fh_to_dentry() and ->get_parent() methods use to turn the inode they
looked up into a dentry for exportfs_decode_fh_raw(), i.e., for
open_by_handle_at() and for nfsd file handle decoding.

For a local filesystem an in-core inode without links and aliases is an
orphan that goes away on the final iput(). For a network filesystem
that's not true. i_nlink is just the link count the server last
reported and whether the object is still reachable through its file
handle is up to the server. If a file is unlinked while it is still
open elsewhere, e.g., by another NFSv4 client or by a process on the
server, the server keeps the object alive and its file handle stays
valid. nfs_fh_to_dentry() fetches the attributes for such a handle,
nfs_fhget() sets i_nlink to 0 from them, the new inode has no aliases
yet and so __d_obtain_alias() now fails with -ESTALE. The same applies
to fuse_get_dentry() which takes i_nlink from the fuse server.
Filesystems that care about unlinked inodes already make that decision
themselves in ->fh_to_dentry(), e.g., ceph's __fh_to_dentry() fetches
the link count with CEPH_CAP_LINK_SHARED and only refuses an unlinked
inode if it isn't open.

Note that on a network filesystem i_nlink drops to zero through
attribute updates when another client unlinks the file. So the inode
might never have gone through dentry_unlink_inode() with i_nlink == 0,
i.e., fsnotify_inoderemove() might never have been called for it and
there's nothing for the -ESTALE to protect.

So drop the two checks from __d_obtain_alias() again and go back to the
unconditional alias lookup. While at it, update the comment above
inode_notify_dead() since it no longer blocks new aliases.

Amir sent this as v3 to replace 669bb5af15d4 but that commit was already
in mainline by then. So do it as a fix on top. The resulting code is
identical to v3.

This means open_by_handle_at() can again create a new alias for an
unlinked inode after FS_DELETE_SELF has been reported. Closing that file
reports a second FS_DELETE_SELF and the file has none of the inode marks
that were torn down. That's what Daehyeon's open_by_handle_at() LTP test
checks so it is expected to fail again. This needs to be handled
separately. A mark that is added through such a file while
fsnotify_destroy_marks() is still running is rejected with -ENOENT since
36ebddca56b9 ("fsnotify: prevent mark additions during connector
teardown") so it can't end up on a detached connector.

Christian Brauner <brauner@kernel.org> says:

Amir's v3 replaced 669bb5af15d4 as a whole. I turned it into an
incremental fix on top of mainline and wrote a new commit message.
Jan's Reviewed-by was given on v3 and the resulting code is identical.

Fixes: 669bb5af15d4 ("fs: make sure to call fsnotify_inoderemove() only once when inode is dead")
Cc: stable@vger.kernel.org
Reported-by: Daehyeon Ko <4ncienth@gmail.com>
Signed-off-by: Amir Goldstein <amir73il@gmail.com>
Link: https://patch.msgid.link/CAOQ4uxg7Qt436SiKhKEJKOjtCjeB_XyTjazYF7HMyQAGGcgx2A@mail.gmail.com
Reviewed-by: Jan Kara <jack@suse.cz>
Link: https://patch.msgid.link/20261030-work-dcache-obtain_alias-v1-1-a1f7e44536d2@kernel.org
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>