Re: [PATCH v10 5/8] fuse: Revalidate positive entries in fuse_atomic_open

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

On Mon, Oct 23, 2023 at 08:30:32PM +0200, Bernd Schubert wrote:

> +		if (!switched && !d_in_lookup(entry)) {
> +			d_drop(entry);
> +			new = d_alloc_parallel(entry->d_parent, &entry->d_name,
> +					       wq);
> +			if (IS_ERR(new))
> +				return new;
> +
> +			if (unlikely(!d_in_lookup(new))) {
> +				dput(new);
> +				new = ERR_PTR(-EIO);
> +				return new;

Again, huh?  You call d_drop().  Then another thread tries to look
the same thing up and gets there while d_alloc_parallel() is
allocating a new dentry.  d_alloc_parallel() sees that dentry
is already in hash (if lookup has managed to complete) or
in in-lookup hash (if lookup is in progress).  In the former
case it returns the dentry it had found (and frees the one
it intended to put in); in the latter it waits for lookup to
complete and checks if dentry is hashed, has the same name and
the same parent.  If it is, same as if we'd found it hashed
in the first place - we return an additional reference to it.

It's perfectly valid and I really don't understand what are you
trying to achieve here.