On Sat, 27 Aug 2022, Al Viro wrote:
> On Fri, Aug 26, 2022 at 12:10:43PM +1000, NeilBrown wrote:
> > Allow object can now be renamed from or to a directory in which a create
> > or unlink is concurrently happening.
> >
> > Two or more renames with the one directory can also be concurrent.
> > s_vfs_rename_mutex still serialises lookups for cross-directory renames,
> > but the renames themselves can proceed concurrently.
>
> Wha...? <checks>
> Not true, fortunately - you *do* hold ->s_vfs_rename_mutex over the
> rename itself. If not for that, it would be utterly broken.
> And I don't care for NFS server rejecting that - we are *NOT* taking
> loop prevention logics into every filesystem. It's highly non-local
> and trying to handle it with your per-dentry flags is going to be
> painful as hell, if at all possible.
>
I don't know what happened there - I let myself get confused somewhere
in the process. You are of course right that s_vfs_rename_mutex is held
the whole time. I wasn't intending to try to change that.
> > + if (d1 < d2) {
> > + ok1 = d_lock_update_nested(d1, p1, last1, I_MUTEX_PARENT);
> > + ok2 = d_lock_update_nested(d2, p2, last2, I_MUTEX_PARENT2);
> > + } else {
> > + ok2 = d_lock_update_nested(d2, p2, last2, I_MUTEX_PARENT);
> > + ok1 = d_lock_update_nested(d1, p1, last1, I_MUTEX_PARENT2);
> > + }
>
> Explain, please. What's that ordering about?
>
Deadlock avoidance, just like in the same-directory case.
But I guess as s_vfs_rename_mutex is held, ordering cannot matter.
I'll remove the ordering.
Thanks,
NeilBrown
