On Wed, Nov 16, 2016 at 5:55 PM, Daniel Borkmann <dan...@iogearbox.net> wrote:
> On 11/16/2016 04:45 PM, Daniel Borkmann wrote:
>>
>> On 11/16/2016 01:51 PM, Saeed Mahameed wrote:
>>>
>>> On Wed, Nov 16, 2016 at 2:04 AM, Daniel Borkmann <dan...@iogearbox.net>
>>> wrote:
>>>>
>>>> mlx5e_xdp_set() is currently the only place where we drop reference on
>>>> the
>>>> prog sitting in priv->xdp_prog when it's exchanged by a new one. We also
>>>> need to make sure that we eventually release that reference, for
>>>> example,
>>>> in case the netdev is dismantled.
>>>>
>>>> Fixes: 86994156c736 ("net/mlx5e: XDP fast RX drop bpf programs support")
>>>> Signed-off-by: Daniel Borkmann <dan...@iogearbox.net>
>>>> ---
>>>> drivers/net/ethernet/mellanox/mlx5/core/en_main.c | 3 +++
>>>> 1 file changed, 3 insertions(+)
>>>>
>>>> diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en_main.c
>>>> b/drivers/net/ethernet/mellanox/mlx5/core/en_main.c
>>>> index cf26672..60fe54c 100644
>>>> --- a/drivers/net/ethernet/mellanox/mlx5/core/en_main.c
>>>> +++ b/drivers/net/ethernet/mellanox/mlx5/core/en_main.c
>>>> @@ -3715,6 +3715,9 @@ static void mlx5e_nic_cleanup(struct mlx5e_priv
>>>> *priv)
>>>>
>>>> if (MLX5_CAP_GEN(mdev, vport_group_manager))
>>>> mlx5_eswitch_unregister_vport_rep(esw, 0);
>>>> +
>>>> + if (priv->xdp_prog)
>>>> + bpf_prog_put(priv->xdp_prog);
>>>> }
>>>
>>>
>>> I thought that on unregister_netdev ndo_xdp_set will be called with
>>> NULL prog to cleanup. like any other resources (Vlans/mac_lists/
>>> etc..), why xdp should be different ?
>>> Anyway if this is the case, I am ok with this fix, you can even send
>>> it to net (it looks like a serious leak).
>>
>>
>> The only interaction with ndo_xdp() right now is dev_change_xdp_fd()
>> and the currently a bit terse dump via rtnl_xdp_fill(). The latter
>> only tells whether something is actually attached and will have more
>> info in near future, but doesn't alter anything.
>>
>> dev_change_xdp_fd() is only triggered from user side via netlink when
>> IFLA_XDP container attr is around, so no automatic cleanup here. This
>> means as per documentation in enum xdp_netdev_command, that the driver
>> has full ownership, thus needs to bpf_prog_put().
>
>
> Note that without patch 2/4, just sending this one to net doesn't really
> solve anything, since there the mlx5e_xdp_set() still has the incorrect
> bpf_prog_add(prog, 1) around. So it's the whole series if so. I had it
> originally targeted at net, but Alexei suggested net-next; I don't really
> mind either way, so I agreed to go for net-next.
Ok, Thank you Daniel
