On Sun, Nov 24, 2013 at 06:04:13AM -0800, Christoph Hellwig wrote:
> Seems I can reproduce this by doing a full xfstests run and then
> shutting down the VM. Doesn't seem to happen with the XFS tree
> which is still based on 3.12-rc1.
may_lookup() with LOOKUP_RCU in nd->flags and NULL nd->inode, by
the look of it... Interesting.
AFAICS, path_init() initializes nd->inode on all success exits.
As for the places where we reassign nd->inode, we have the following:
fs/namei.c:681: nd->inode = nd->path.dentry->d_inode;
nd_jump_link(), we would've already left RCU mode
fs/namei.c:856: nd->inode = nd->path.dentry->d_inode;
follow_link() - ditto.
fs/namei.c:1160: nd->inode = nd->path.dentry->d_inode;
fs/namei.c:1257: nd->inode = nd->path.dentry->d_inode;
follow_dotdot() - not in RCU mode
fs/namei.c:1547: nd->inode = inode;
walk_component() - inode can't be NULL here, variable has
been explicitly compared with NULL a few lines prior
fs/namei.c:1825: nd->inode = inode;
path_init() with LOOKUP_ROOT; not from stat(2) and it would've
already oopsed with NULL inode anyway.
fs/namei.c:1892: nd->inode = nd->path.dentry->d_inode;
fs/namei.c:3006: nd->inode = inode;
do_last(), not from stat(2) and no may_lookup() done past that
fs/namei.c:3077: nd->inode = dir->d_inode;
do_last(), not from stat(2)
which seems to leave two candidates - follow_dotdot_rcu() and path_init(),
both setting nd->inode to nd->path.dentry->d_inode...
Could you try to reproduce it with something like
if (read_seqretry(&mount_lock, nd->m_seq))
slapped before the success exit in follow_dotdot_rcu(), just to see if
we are hitting some races with umount here?