Re: [patch net-next 4/4] mlxsw: Introduce Mellanox SwitchX-2 ASIC support

[Jiri Pirko <jiri@resnulli.us>]

Sun, Jul 26, 2015 at 04:45:07AM CEST, sfeldma@gmail.com wrote:
>On Thu, Jul 23, 2015 at 8:43 AM, Jiri Pirko <jiri@resnulli.us> wrote:
>> From: Jiri Pirko <jiri@mellanox.com>
>>
>> Benefit from the previously introduced Mellanox Switch infrastructure and
>> add driver for SwitchX-2 ASIC. Note that this driver is very simple now.
>> It implements bare minimum for getting device to work on slow-path.
>> Fast-path offload functionality is going to be added soon.
>>
>> Signed-off-by: Jiri Pirko <jiri@mellanox.com>
>> Signed-off-by: Ido Schimmel <idosch@mellanox.com>
>> Signed-off-by: Elad Raz <eladr@mellanox.com>
>
>[cut]
>
>> +static netdev_tx_t mlxsw_sx_port_xmit(struct sk_buff *skb,
>> +                                     struct net_device *dev)
>> +{
>> +       struct mlxsw_sx_port *mlxsw_sx_port = netdev_priv(dev);
>> +       struct mlxsw_sx *mlxsw_sx = mlxsw_sx_port->mlxsw_sx;
>> +       struct mlxsw_sx_port_pcpu_stats *pcpu_stats;
>> +       const struct mlxsw_tx_info tx_info = {
>> +               .local_port = mlxsw_sx_port->local_port,
>> +               .is_emad = false,
>> +       };
>> +       int err;
>> +
>> +       if (unlikely(skb_headroom(skb) < MLXSW_TXHDR_LEN)) {
>
>Does this happen at all since dev->hard_header_len was set in probe to
>add MLXSW_TXHDR_LEN?

This needs to be done for example for bridge forwarding case and other
forwarding cases.


>
>> +               struct sk_buff *skb_new;
>> +
>> +               skb_new = skb_realloc_headroom(skb, MLXSW_TXHDR_LEN);
>> +               dev_kfree_skb_any(skb);
>> +               if (!skb_new) {
>> +                       this_cpu_inc(mlxsw_sx_port->pcpu_stats->tx_dropped);
>> +                       return NETDEV_TX_OK;
>> +               }
>> +               skb = skb_new;
>> +       }
>> +       mlxsw_sx_txhdr_construct(skb, &tx_info);
>> +       err = mlxsw_core_skb_transmit(mlxsw_sx, skb, &tx_info);
>> +       if (err == -EAGAIN)
>> +               return NETDEV_TX_BUSY;
>
>I think there is a problem here when returning NETDEV_TX_BUSY when
>original skb might have been freed above in the headroom check. (ref
>Documentation/networking/driver.txt).

I have to check this out a bit more. Thanks for pointing that out.


>
>[cut]
>
>> +static int mlxsw_sx_port_dev_addr_get(struct mlxsw_sx_port *mlxsw_sx_port)
>> +{
>> +       struct mlxsw_sx *mlxsw_sx = mlxsw_sx_port->mlxsw_sx;
>> +       struct net_device *dev = mlxsw_sx_port->dev;
>> +       char ppad_pl[MLXSW_REG_PPAD_LEN];
>> +       int err;
>> +
>> +       mlxsw_reg_ppad_pack(ppad_pl, false, 0);
>> +       err = mlxsw_reg_query(mlxsw_sx->core, MLXSW_REG(ppad), ppad_pl);
>> +       if (err)
>> +               return err;
>> +       mlxsw_reg_ppad_mac_memcpy_from(ppad_pl, dev->dev_addr);
>> +       /* The last byte in base mac address is always 0 */
>> +       dev->dev_addr[ETH_ALEN - 1] += mlxsw_sx_port->local_port;
>
>If MLXSW_PORT_MAX_PORTS > 256, you'll wrap this.  Is dev_addr[ETH_ALEN
>- 2] available to carry into?

That will never happen. MLXSW_PORT_MAX_PORTS is 0x40 and address got
from ppad register always ends with 0


>
>> +       return 0;
>> +}
>> +
>
>[cut]
>
>> +static int mlxsw_sx_port_create(struct mlxsw_sx *mlxsw_sx, u8 local_port)
>> +{
>> +       struct mlxsw_sx_port *mlxsw_sx_port;
>> +       struct net_device *dev;
>> +       bool usable;
>> +       int err;
>> +
>> +       dev = alloc_etherdev(sizeof(struct mlxsw_sx_port));
>> +       if (!dev)
>> +               return -ENOMEM;
>> +       mlxsw_sx_port = netdev_priv(dev);
>> +       mlxsw_sx_port->dev = dev;
>> +       mlxsw_sx_port->mlxsw_sx = mlxsw_sx;
>> +       mlxsw_sx_port->local_port = local_port;
>> +
>> +       mlxsw_sx_port->pcpu_stats =
>> +               netdev_alloc_pcpu_stats(struct mlxsw_sx_port_pcpu_stats);
>> +       if (!mlxsw_sx_port->pcpu_stats) {
>> +               err = -ENOMEM;
>> +               goto err_alloc_stats;
>> +       }
>> +
>> +       dev->netdev_ops = &mlxsw_sx_port_netdev_ops;
>> +       dev->ethtool_ops = &mlxsw_sx_port_ethtool_ops;
>> +       dev->switchdev_ops = &mlxsw_sx_port_switchdev_ops;
>> +
>> +       err = mlxsw_sx_port_dev_addr_get(mlxsw_sx_port);
>> +       if (err) {
>> +               dev_err(mlxsw_sx->bus_info->dev, "Port %d: Unable to get port mac address\n",
>> +                       mlxsw_sx_port->local_port);
>> +               goto err_dev_addr_get;
>> +       }
>> +
>> +       netif_carrier_off(dev);
>> +
>> +       dev->features |= NETIF_F_NETNS_LOCAL | NETIF_F_LLTX |
>
>Not supposed to use LLTX in new drivers, according to
>include/linux/netdev_features.h.

In our case, wee need to use this. Since multiple port netdevs may use
the same send dataqueue, we need to do locking ourselves.

Thanks for review!