Re: Linux 2.6.38-rc5

[Al Viro <viro@ZenIV.linux.org.uk>]

On Wed, Feb 16, 2011 at 07:46:20AM -0800, Linus Torvalds wrote:
> On Wed, Feb 16, 2011 at 3:14 AM, Eric Dumazet <eric.dumazet@gmail.com> wrote:
> >
> > Using this kernel on my dev machine (2x4x2 cpus), I hit BUG_ON() in
> > fs/namei.c:1461 on my kernel build (make -j16)
> 
> Uhhuh. We replaced one BUG_ON() with another.
> 
> And I think it's a really silly problem too: when Al moved the
> 
>         /* We drop rcu-walk here */
>         if (nameidata_dentry_drop_rcu_maybe(nd, path->dentry))
>                 return -ECHILD;
> 
> test into do_follow_link(), he _should_ have moved the BUG_ON() in
> there too, methinks. He didn't, and as a result the BUG_ON() is now
> before the "drop_rcu_maybe".
> 
> This patch should fix it. Al? Comments?

Sigh...  I see what's going on.  We'd got inode from dentry that is getting
crapped under us.  We will *not* survive dropping RCU - it's bad enough for
full restart in normal mode.  So right after we'd seen that (already wrong)
inode has ->follow_link(), we decide to drop RCU.  Originally this BUG_ON
hadn't been reached in that case - we had already failed with -ECHILD before
we got to it.  Now we don't...

_However_, I don't like passing inode to do_follow_link().  I'd rather set
nd->inode to inode first and use it there.  Let me think a bit and see if
it's feasible...