On Thu, May 31, 2018 at 08:19:55PM +0100, Al Viro wrote:
> On Fri, May 25, 2018 at 01:07:34AM +0100, David Howells wrote:
> > + if (unlikely(file->f_mode & FMODE_NEED_UNMOUNT))
> > + __detach_mounts(dentry);
> > +
>
> This is completely wrong. First of all, you want to dissolve the mount tree
> on file->f_path.mount, not every tree rooted at dentry equal to
> file->f_path.dentry.
> This is easily done - it would be a simple call of drop_collected_mounts(mnt)
> if not for one detail. You want it to happen only if the sucker isn't
> attached
> anywhere by that point. IOW,
> namespace_lock();
> lock_mount_hash();
> if (!real_mount(mnt)->mnt_ns)
> umount_tree(real_mount(mnt), UMOUNT_SYNC);
> unlock_mount_hash();
> namespace_unlock();
> and that's it. You don't need that magical mystery turd in move_mount() later
> in the series and all the infrastructure you grow for it.
>
> FWIW, I would've suggested this
> void drop_collected_mounts(struct vfsmount *mnt)
> {
> namespace_lock();
> lock_mount_hash();
> + if (!real_mount(mnt)->mnt_ns)
> + umount_tree(real_mount(mnt), UMOUNT_SYNC);
> - umount_tree(real_mount(mnt), UMOUNT_SYNC);
> unlock_mount_hash();
> namespace_unlock();
> }
>
> and in __fput()
> if (unlikely(file->f_mode & FMODE_NEED_UNMOUNT))
> drop_collected_mounts(mnt);
>
> All there is to it, AFAICS...
... and that eliminates #27 and #28 entirely, with #31 becoming simpler -
no move_mount_lookup(), no dfd_ref, the check in do_move_mount() becomes
+ if (!mnt_has_parent(old) && old->mnt_ns) {
+ /* We need to allow open(O_PATH|O_CLONE_MOUNT) or fsmount()
+ * followed by move_mount(), but mustn't allow "/" to be moved.
+ */
+ goto out1;
+ }
and I wouldn't be surprised if move_mount_old()/move_mount() split turns out
to be not needed at all, seeing that the whole "clear FMODE_NEED_UNMOUNT
on success" part goes away.
