Re: [PATCH v4 13/18] xen/pvcalls: implement release command

[Stefano Stabellini]

On Tue, 20 Jun 2017, Boris Ostrovsky wrote:
> > +
> > +static int pvcalls_back_release_passive(struct xenbus_device *dev,
> > +					struct pvcalls_fedata *fedata,
> > +					struct sockpass_mapping *mappass)
> > +{
> > +	if (mappass->sock->sk != NULL) {
> > +		write_lock_bh(&mappass->sock->sk->sk_callback_lock);
> > +		mappass->sock->sk->sk_user_data = NULL;
> > +		mappass->sock->sk->sk_data_ready = mappass->saved_data_ready;
> > +		write_unlock_bh(&mappass->sock->sk->sk_callback_lock);
> > +	}
> > +	down(&fedata->socket_lock);
> > +	radix_tree_delete(&fedata->socketpass_mappings, mappass->id);
> > +	sock_release(mappass->sock);
> > +	flush_workqueue(mappass->wq);
> > +	destroy_workqueue(mappass->wq);
> > +	kfree(mappass);
> > +	up(&fedata->socket_lock);
> 
> Can you raise the semaphore earlier, once the mapping is deleted from
> the tree?

Yes, I think it can.

> Also, why are you not locking the tree in pvcalls_back_accept()?

Good point! Although socket_lock is used to protect insertions and
deletions to socketpass_mappings and socket_mappings, many of the sites
that only access socket_mappings and socketpass_mappings without making
modifications are left unprotected at the moment. I'll fix that, and to
do it I'll move radix_tree_delete out of pvcalls_back_release_passive.