Re: [PATCH 3/3] VFS: close race between getcwd() and d_move()

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

On Fri, Nov 10, 2017 at 03:45:41PM +1100, NeilBrown wrote:
> -void __d_drop(struct dentry *dentry)
> +static void ___d_drop(struct dentry *dentry)
>  {
>  	if (!d_unhashed(dentry)) {
>  		struct hlist_bl_head *b;
> @@ -486,12 +488,15 @@ void __d_drop(struct dentry *dentry)
>  
>  		hlist_bl_lock(b);
>  		__hlist_bl_del(&dentry->d_hash);
> -		dentry->d_hash.pprev = NULL;
>  		hlist_bl_unlock(b);
>  		/* After this call, in-progress rcu-walk path lookup will fail. */
>  		write_seqcount_invalidate(&dentry->d_seq);
>  	}
>  }
> +void __d_drop(struct dentry *dentry) {
> +	___d_drop(dentry);
> +	dentry->d_hash.pprev = NULL;

Umm...  That reordering (unhashed vs. ->d_seq) might be a problem
on the RCU side.  I'm not sure it is, we might get away with that,
actually, but I want to finish digging through the pathwalk-related
code.  Cursing it for being too subtle for its own good, as usual...