Re: [PATCH v3 09/18] nvmet-fcloop: prevent double port deletion

[Daniel Wagner <dwagner@suse.de>]

On Tue, Mar 18, 2025 at 12:15:04PM +0100, Hannes Reinecke wrote:
> >   fcloop_remoteport_delete(struct nvme_fc_remote_port *remoteport)
> >   {
> >   	struct fcloop_rport *rport = remoteport->private;
> > +	bool delete_port = true;
> >   	unsigned long flags;
> >   	flush_work(&rport->ls_work);
> >   	spin_lock_irqsave(&fcloop_lock, flags);
> > +	if (test_and_set_bit(PORT_DELETE, &rport->flags))
> > +		delete_port = false;
> >   	rport->nport->rport = NULL;
> >   	spin_unlock_irqrestore(&fcloop_lock, flags);
> Can't you just check for a NULL rport->nport->rport pointer
> and do away with the PORT_DELETE flag?

Unfortunately, nport->rport is also set to NULL in __unlink_remote_port
and __unlink_target_port. If we would just update the pointer here, it
would be possible.

I played a bit around when to clear the nport->rport pointer but it
didn't work. There were always some UAFs or NULL pointer accesses. With
the flags I was able to get it fixed. I am not insisting on this
solution, just trying to explain why I choosed it.