Re: [PATCH] NFS: Fix directory delegation verifier checks

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

On Sat, Apr 04, 2026 at 07:32:47PM +0100, Al Viro wrote:
> [cc to fsdevel added]
> 
> On Wed, Dec 31, 2025 at 09:52:35PM +0000, Trond Myklebust wrote:
> 
> > +static void nfs_clear_verifier_directory(struct inode *dir)
> > +{
> > +	struct dentry *this_parent;
> > +	struct dentry *dentry;
> > +	struct inode *inode;
> > +
> > +	if (hlist_empty(&dir->i_dentry))
> > +		return;
> > +	this_parent =
> > +		hlist_entry(dir->i_dentry.first, struct dentry, d_u.d_alias);
> > +
> > +	spin_lock(&this_parent->d_lock);
> > +	nfs_unset_verifier_delegated(&this_parent->d_time);
> > +	dentry = d_first_child(this_parent);
> > +	hlist_for_each_entry_from(dentry, d_sib) {
> > +		if (unlikely(dentry->d_flags & DCACHE_DENTRY_CURSOR))
> > +			continue;
> > +		inode = d_inode_rcu(dentry);
> > +		if (inode &&
> > +		    NFS_PROTO(inode)->have_delegation(inode, FMODE_READ, 0))
> > +			continue;
> 
> What's to stop the inode from being freed right under you?  You are *not*
> guaranteed to be holding rcu_read_lock(), unless I'm missing something
> in the call chain, so... what's going on here?

Incidentally, nfs_clear_verifier_file() doesn't need to bother with
checking for NULL dir - alias is found in ->i_dentry of inode with
->i_lock held, which means it can't have progressed through
__dentry_kill() to dropping the reference to parent.  So holding
alias->d_lock (which stabilizes ->d_parent) is enough - ->d_parent
is still pinned by alias and it can't have become negative.

"Can't have progressed through __dentry_kill()" is critical here -
that's the difference from nfs_set_verifier_locked() case.

Folks, this stuff can get seriously subtle; _please_ ask around on
fsdevel if you are doing something non-trivial with it.