Re: [PATCH 4.19 43/92] do_last(): fetch directory ->i_mode and ->i_uid before its too late

[Rantala, Tommi T. (Nokia - FI/Espoo)]

On Fri, 2020-01-31 at 12:20 +0000, Al Viro wrote:
> On Fri, Jan 31, 2020 at 10:08:37AM +0000, Rantala, Tommi T. (Nokia -
> FI/Espoo) wrote:
> > On Tue, 2020-01-28 at 15:08 +0100, Greg Kroah-Hartman wrote:
> > > From: Al Viro <viro@xxxxxxxxxxxxxxxxxx>
> > > 
> > > @@ -3258,6 +3259,8 @@ static int do_last(struct nameidata *nd,
> > >  		   struct file *file, const struct open_flags *op)
> > >  {
> > >  	struct dentry *dir = nd->path.dentry;
> > > +	kuid_t dir_uid = dir->d_inode->i_uid;
> > 
> > I hit the following oops in 4.19.100 while running kselftests.
> > 
> > fs/namei.c:3262 matches the line above.
> > 
> > Any ideas?
> 
> Yes.  Make those two line
> 	kuid_t dir_uid = nd->inode->i_uid;
> 	umode_t dir_mode = nd->inode->i_mode;
> 
> I'm pretty sure that I know which way I'd fucked up there; we can
> get here in RCU mode with stale nd->path.dentry (that would make
> the thing fail with -ECHILD. with retry in non-RCU mode).  In
> non-stale case nd->inode is the same as nd->path.dentry->d_inode
> and it's always pointing to a struct inode that hadn't been
> freed yet.

The oops does not seem to reproduce easily with kselftests, I've only hit
it once with 4.19.100.

Made the change in fs/namei.c, so far no ill effects.
I'll continue running the kselftests...

-Tommi