[PATCH net] net/mlx4_en: mlx4_en_netpoll shouldn't call napi_schedule when port is down

[PATCH net] net/mlx4_en: mlx4_en_netpoll shouldn't call napi_schedule when port is down

[Amir Vadai]

From: Ido Shamay <idos@mellanox.com>

mlx4_en_netpoll, which is mlx4_en ndo_poll_controller callback,
might be called when port is down, causing a napi_schedule when
napi->poll callback in NULL. mutex_trylock is needed to acquire
the port_state lock, since other threads may grab it and stop
the port while we are in napi scheduling. Using trylock since in atomic
context.

Signed-off-by: Ido Shamay <idos@mellanox.com>
Signed-off-by: Amir Vadai <amirv@mellanox.com>
---

Hi Dave,

Please push this commit to -stable

Thanks,
Amir

 drivers/net/ethernet/mellanox/mlx4/en_netdev.c | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/drivers/net/ethernet/mellanox/mlx4/en_netdev.c b/drivers/net/ethernet/mellanox/mlx4/en_netdev.c
index abddcf8..e243f1c 100644
--- a/drivers/net/ethernet/mellanox/mlx4/en_netdev.c
+++ b/drivers/net/ethernet/mellanox/mlx4/en_netdev.c
@@ -1272,13 +1272,23 @@ out:
 static void mlx4_en_netpoll(struct net_device *dev)
 {
 	struct mlx4_en_priv *priv = netdev_priv(dev);
+	struct mlx4_en_dev *mdev = priv->mdev;
 	struct mlx4_en_cq *cq;
 	int i;
 
+	if (!mutex_trylock(&mdev->state_lock))
+		return;
+
+	if (!priv->port_up)
+		goto out;
+
 	for (i = 0; i < priv->rx_ring_num; i++) {
 		cq = priv->rx_cq[i];
 		napi_schedule(&cq->napi);
 	}
+
+out:
+	mutex_unlock(&mdev->state_lock);
 }
 #endif
 
-- 
1.8.3.4

Re: [PATCH net] net/mlx4_en: mlx4_en_netpoll shouldn't call napi_schedule when port is down

[Cong Wang]

On Mon, Sep 29, 2014 at 4:04 AM, Amir Vadai <amirv@mellanox.com> wrote:
> From: Ido Shamay <idos@mellanox.com>
>
> mlx4_en_netpoll, which is mlx4_en ndo_poll_controller callback,
> might be called when port is down, causing a napi_schedule when
> napi->poll callback in NULL. mutex_trylock is needed to acquire
> the port_state lock, since other threads may grab it and stop
> the port while we are in napi scheduling. Using trylock since in atomic
> context.

Are you sure it's safe? Its comment says:

 * This function must not be used in interrupt context. The
 * mutex must be released by the same task that acquired it.

Did you test it with LOCKDEP enabled?

Re: [PATCH net] net/mlx4_en: mlx4_en_netpoll shouldn't call napi_schedule when port is down

[David Miller]

From: Amir Vadai <amirv@mellanox.com>
Date: Mon, 29 Sep 2014 14:04:55 +0300

> From: Ido Shamay <idos@mellanox.com>
> 
> mlx4_en_netpoll, which is mlx4_en ndo_poll_controller callback,
> might be called when port is down, causing a napi_schedule when
> napi->poll callback in NULL. mutex_trylock is needed to acquire
> the port_state lock, since other threads may grab it and stop
> the port while we are in napi scheduling. Using trylock since in atomic
> context.
> 
> Signed-off-by: Ido Shamay <idos@mellanox.com>
> Signed-off-by: Amir Vadai <amirv@mellanox.com>

I don't think netpoll should 'variably succeed' on poll controller's
ability to immediately take a trylock.

You need to find a way to make this code always process the queues
when it is called, the approach you are taking here is not a correct.

If you've designed things in such a way that the locking here is
difficult, that's unfortunate but you still have to make this
function behave properly.

Re: [PATCH net] net/mlx4_en: mlx4_en_netpoll shouldn't call napi_schedule when port is down

[Ido Shamay]

On 9/29/2014 8:47 PM, David Miller wrote:
> From: Amir Vadai <amirv@mellanox.com>
> Date: Mon, 29 Sep 2014 14:04:55 +0300
>
>> From: Ido Shamay <idos@mellanox.com>
>>
>> mlx4_en_netpoll, which is mlx4_en ndo_poll_controller callback,
>> might be called when port is down, causing a napi_schedule when
>> napi->poll callback in NULL. mutex_trylock is needed to acquire
>> the port_state lock, since other threads may grab it and stop
>> the port while we are in napi scheduling. Using trylock since in atomic
>> context.
>>
>> Signed-off-by: Ido Shamay <idos@mellanox.com>
>> Signed-off-by: Amir Vadai <amirv@mellanox.com>
>
> I don't think netpoll should 'variably succeed' on poll controller's
> ability to immediately take a trylock.
>
> You need to find a way to make this code always process the queues
> when it is called, the approach you are taking here is not a correct.
>
> If you've designed things in such a way that the locking here is
> difficult, that's unfortunate but you still have to make this
> function behave properly.

Sorry about that, aborting this commit since the problem was already 
fixed in commit 3484aac1, which added netif_device_detach to driver port 
stop flow and that made netconsole not to call ndo_poll_controller callback.

Thanks.
> --
> To unsubscribe from this list: send the line "unsubscribe netdev" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>