On Thu, Nov 23, 2023 at 01:12:08AM +0000, Al Viro wrote:
> On Wed, Nov 22, 2023 at 09:19:01PM +0000, Al Viro wrote:
> > On Tue, Nov 21, 2023 at 02:27:34AM +0000, Al Viro wrote:
> >
> > > I will review that series; my impression from the previous iterations
> > > had been fairly unpleasant, TBH, but I hadn't rechecked since April
> > > or so.
> >
> > The serious gap, AFAICS, is the interplay with open-by-fhandle.
> > It's not unfixable, but we need to figure out what to do when
> > lookup runs into a disconnected directory alias. d_splice_alias()
> > will move it in place, all right, but any state ->lookup() has
> > hung off the dentry that had been passed to it will be lost.
> >
> > And I seriously suspect that we want to combine that state
> > propagation with d_splice_alias() (or its variant to be used in
> > such cases), rather than fixing the things up afterwards.
> >
> > In particular, propagating ->d_op is really not trivial at that
> > point; it is safe to do to ->lookup() argument prior to d_splice_alias()
> > (even though that's too subtle and brittle, IMO), but after
> > d_splice_alias() has succeeded, the damn thing is live and can
> > be hit by hash lookups, revalidate, etc.
> >
> > The only things that can't happen to it are ->d_delete(), ->d_prune(),
> > ->d_iput() and ->d_init(). Everything else is fair game.
> >
> > And then there's an interesting question about the interplay with
> > reparenting. It's OK to return an error rather than reparent,
> > but we need some way to tell if we need to do so.
>
> Hmm... int (*d_transfer)(struct dentry *alias, struct dentry *new)?
> Called if d_splice_alias() picks that sucker, under rename_lock,
> before the call of __d_move(). Can check IS_ROOT(alias) (due to
> rename_lock), so can tell attaching from reparenting, returning
> an error - failed d_splice_alias().
>
> Perhaps would be even better inside __d_move(), once all ->d_lock
> are taken... Turn the current bool exchange in there into honest
> enum (exchange/move/splice) and call ->d_transfer() on splice.
> In case of failure it's still not too late to back out - __d_move()
> would return an int, ignored in d_move() and d_exchange() and
> treated as "fail in unlikely case it's non-zero" in d_splice_alias()
> and __d_unalias()...
>
> Comments? Note that e.g.
> res = d_splice_alias(inode, dentry);
> if (!IS_ERR(fid)) {
> if (!res)
> v9fs_fid_add(dentry, &fid);
> else if (!IS_ERR(res))
> v9fs_fid_add(res, &fid);
> else
> p9_fid_put(fid);
> }
>
> in 9p ->lookup() would turn into
>
> v9fs_fid_add(dentry, &fid);
> return d_splice_alias(inode, dentry);
>
> with ->d_transfer(alias, new) being simply
>
> struct hlist_node *p = new->d_fsdata;
> hlist_del_init(p);
> __add_fid(alias, hlist_entry(p, struct p9_fid, dlist));
> return 0;
>
> assuming the call from __d_move()...
Incidentally, 9p and this one would not be the only places that could use it -
affs - alias->d_fsdata = new->d_fsdata
afs - ditto
ocfs2 - smells like another possible benefitiary (attaching locks, etc.
would be saner if done before d_splice_alias(), with ->d_transfer()
moving the lock to the alias)...
