Re: Should sys_unmount() FOLLOW?

["J. Hannken-Illjes" <hannken%eis.cs.tu-bs.de@localhost>]

On 23 Nov 2014, at 17:47, Emmanuel Dreyfus <manu%netbsd.org@localhost> wrote:

> Hi
> 
> I ran into this strange bug with glusterfs NFS server, which is possible
> because it allows the mounted filesystem root vnode to be VLNK (NetBSD's
> native mountd prevents this situation and therefore the bug does not
> happen with our native NFS server):
> 
> bacasel# mount
> bacasel.net.espci.fr:/patchy/symlink1 on /mnt/nfs/0 type nfs
> 
> bacasel# ls -l /mnt/nfs
> lrwxrwxrwx  1 root  wheel    4 Nov 23 10:03 0 -> dir1
> 
> That is possible because on the exported filesystem, symlink1 is a
> symlink to dir1.
> 
> It looks funny and harmless, at least until one try to unmount while the
> NFS server is down. I do this using the most reliable way: umount -f -R
> /mnt/nfs/0
>  
> 
> umount(8) will quickly call unmount(2), passing the /mnt/nfs/0 path
> without trying anything fancy. But at the beginning of sys_unmount() in
> the kernel, we can find:
> 
>        NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | TRYEMULROOT, pb);
>        if ((error = namei(&nd)) != 0) {
>                pathbuf_destroy(pb);
>                return error;
>        }
> 
> The FOLLOW flag will cause the symlink to be resolved before
> sys_unmount() can proceed. Since the NFS server is down, the namei call
> will never return, and umount(8) get stuck in kernel with this
> backtrace:
> sleepq_block
> kpause
> nfs_reconnect
> nfs_request
> nfs_readlinkrpc
> nfs_readlink
> VOP_READLINK
> namei_tryemulroot
> namei
> sys_unmount
> syscall
> 
> This is annoying because umount -f should really unmount. Moreover,
> stuck process will crop on the system because umount(8) holds a vnode
> lock forever.
> 
> I see two way of fixing this:
> 1) remove FOLLOW flag in sys_unmount(). After all, unless the -R flag is
> given, umount(8) should have resolved the patch before calling
> unmount(2).

Did you try -- last time I tried a forced unmount with NFS server down
it didn't work even with root being a directory because the namei() call
would hang in VOP_LOOKUP().  Does it work these days?

> 2) Desfine a new MNT_RAW mount flag. If umount(8) is called with -R,
> pass that flag to unmount(8), and in sys_unmount(), do not use FOLLOW if
> MNT_RAW is set.

--
J. Hannken-Illjes - hannken%eis.cs.tu-bs.de@localhost - TU Braunschweig (Germany)