Re: [PATCH v3 04/18] nvmet-fcloop: refactor fcloop_nport_alloc

[Daniel Wagner]

On Tue, Mar 18, 2025 at 12:02:48PM +0100, Hannes Reinecke wrote:
> > -	list_for_each_entry(tmplport, &fcloop_lports, lport_list) {
> > -		if (tmplport->localport->node_name == opts->wwnn &&
> > -		    tmplport->localport->port_name == opts->wwpn)
> > -			goto out_invalid_opts;
> > +		INIT_LIST_HEAD(&nport->nport_list);
> > +		nport->node_name = opts->wwnn;
> > +		nport->port_name = opts->wwpn;
> > +		refcount_set(&nport->ref, 1);
> > -		if (tmplport->localport->node_name == opts->lpwwnn &&
> > -		    tmplport->localport->port_name == opts->lpwwpn)
> > -			lport = tmplport;
> > +		spin_lock_irqsave(&fcloop_lock, flags);
> > +		list_add_tail(&nport->nport_list, &fcloop_nports);
> > +		spin_unlock_irqrestore(&fcloop_lock, flags);
> >   	}
> Hmm. I don't really like this pattern; there is a race condition
> between lookup and allocation leading to possibly duplicate entries
> on the list.

Yes, that's not a good thing.

> Lookup and allocation really need to be under the same lock.

This means the new entry has always to be allocated first and then we
either free it again or insert into the list, because it's not possible
to allocate under the spinlock. Not that beautiful but correctness wins.