Re: [PATCH] root-hopping for pre-2.3.41-3

[Linus Torvalds]

On Tue, 25 Jan 2000 [e-mail] wrote:
> 
> I'm sending you a patch for pre-2.3.41-3 that allows the use of chroot to
> change the root directory from an initrd or similar to a mounted file
> system, and to cleanly unmount the former root. This is much cleaner than
> the current "change_root on exit from linuxrc" magic, and it also allows
> again the use of NFS-root with parameter selection at run time.

Hmm.

I like the thought, but I don't like much of the implementation.

I'd prefer:

 - your MNT_DETACH thing is a "umount" flag, and I think that is
   fundamentally flawed. It's a separate operation entirely. If we want to
   root-hop, what we should do is rather simple:

	- detach the trees and haev two trees. This is not the same as
	  "umount" at all, this is really just

		struct dentry *mnt = dentry_lookup("new_root_path");
		struct dentry *covered = mnt->d_covers;

		mnt->d_covers = mnt;		/* New root, no longer covering anything */
		covered->d_mounts = covered;	/* Old mount-point, no longer attached */
		dput(covered);
		dput(mnt);

	   and you're done. You now have two trees, neither of which has
	   really been unmounted.

 - Now that the old root has been detached, NOW you can just try to
   unmount it: it's a perfectly regular unmount() at this point, and it
   will fail if it is busy.

I don't think your "complicate the normal umount with new flags" approach
is needed at all.

Also, /proc/mounts should just use the dentry and the path lookup function
(d_path()) to create the /proc/mounts output on-the-fly, and that will get
the name right whatever happens (_including_ a rename of the whole
mount-point path, which neither the current approach nor your approach
gets right at all.

		Linus