From: Louis Peens <louis.peens@xxxxxxxxxxxx>
Date: Mon, 3 Jul 2023 14:01:16 +0200
> From: Yinjun Zhang <yinjun.zhang@xxxxxxxxxxxx>
>
> When moving devices from one namespace to another, mc addresses are
> cleaned in software while not removed from application firmware. Thus
> the mc addresses are remained and will cause resource leak.
>
> Now use `__dev_mc_unsync` to clean mc addresses when closing port.
>
> Fixes: e20aa071cd95 ("nfp: fix schedule in atomic context when sync mc address")
> Cc: stable@xxxxxxxxxxxxxxx
> Signed-off-by: Yinjun Zhang <yinjun.zhang@xxxxxxxxxxxx>
> Acked-by: Simon Horman <simon.horman@xxxxxxxxxxxx>
> Signed-off-by: Louis Peens <louis.peens@xxxxxxxxxxxx>
> ---
> Changes since v1:
>
> * Use __dev_mc_unsyc to clean mc addresses instead of tracking mc addresses by
> driver itself.
> * Clean mc addresses when closing port instead of driver exits,
> so that the issue of moving devices between namespaces can be fixed.
> * Modify commit message accordingly.
>
> .../ethernet/netronome/nfp/nfp_net_common.c | 171 +++++++++---------
> 1 file changed, 87 insertions(+), 84 deletions(-)
[...]
> +static int nfp_net_mc_sync(struct net_device *netdev, const unsigned char *addr)
> +{
> + struct nfp_net *nn = netdev_priv(netdev);
> +
> + if (netdev_mc_count(netdev) > NFP_NET_CFG_MAC_MC_MAX) {
> + nn_err(nn, "Requested number of MC addresses (%d) exceeds maximum (%d).\n",
> + netdev_mc_count(netdev), NFP_NET_CFG_MAC_MC_MAX);
> + return -EINVAL;
> + }
> +
> + return nfp_net_sched_mbox_amsg_work(nn, NFP_NET_CFG_MBOX_CMD_MULTICAST_ADD, addr,
> + NFP_NET_CFG_MULTICAST_SZ, nfp_net_mc_cfg);
> +}
> +
> +static int nfp_net_mc_unsync(struct net_device *netdev, const unsigned char *addr)
> +{
> + struct nfp_net *nn = netdev_priv(netdev);
> +
> + return nfp_net_sched_mbox_amsg_work(nn, NFP_NET_CFG_MBOX_CMD_MULTICAST_DEL, addr,
> + NFP_NET_CFG_MULTICAST_SZ, nfp_net_mc_cfg);
> +}
You can just declare nfp_net_mc_unsync()'s prototype here, so that it
will be visible to nfp_net_netdev_close(), without moving the whole set
of functions. Either way works, but that one would allow avoiding big
diffs not really related to fixing things going through the net-fixes tree.
> +
> /**
> * nfp_net_clear_config_and_disable() - Clear control BAR and disable NFP
> * @nn: NFP Net device to reconfigure
> @@ -1084,6 +1168,9 @@ static int nfp_net_netdev_close(struct net_device *netdev)
>
> /* Step 2: Tell NFP
> */
> + if (nn->cap_w1 & NFP_NET_CFG_CTRL_MCAST_FILTER)
> + __dev_mc_unsync(netdev, nfp_net_mc_unsync);
> +
> nfp_net_clear_config_and_disable(nn);
> nfp_port_configure(netdev, false);
[...]
Thanks,
Olek
