On Fri, Oct 06, 2017 at 10:52:38AM +0200, Nélio Laranjeiro wrote:
> On Thu, Oct 05, 2017 at 08:26:38PM -0700, Yongseok Koh wrote:
> > On Thu, Oct 05, 2017 at 02:49:42PM +0200, Nelio Laranjeiro wrote:
> > [...]
> > > +struct mlx5_rxq_ibv*
> > > +mlx5_priv_rxq_ibv_get(struct priv *priv, uint16_t idx)
> > > +{
> > > + struct mlx5_rxq_data *rxq_data = (*priv->rxqs)[idx];
> > > + struct mlx5_rxq_ctrl *rxq_ctrl;
> > > +
> > > + if (idx >= priv->rxqs_n)
> > > + return NULL;
> > > + if (!rxq_data)
> > > + return NULL;
> > > + rxq_ctrl = container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
> > > + if (rxq_ctrl->ibv) {
> > > + priv_mr_get(priv, rxq_data->mp);
> >
> > One rxq_ibv has one mr as one rxq has one mp. As long as rxq_ibv exist, the
> > mr
> > can't be released. So, it looks unnecessary to increase refcnt of the mr
> > here.
> >[...]
>
> But on MP can be shared among several queues, (see eth_rx_queue_setup()
> and eth_tx_queue_setup()), which means that a queue is not the single
> owner of the Memory pool and thus the memory region.
>
> As the Memory region can be shared among several queues, it is necessary
> to increase/decrease the ref count accordingly.
Here again, as mr->refcnt is increased when it firstly referenced/created in
mlx5_priv_rxq_ibv_new(), I thought it was redundant. If so, priv_mr_release()
can also be called once when rxq_ibv is really destroyed.
Thanks,
Yongseok
