Re: [dpdk-dev] [PATCH v2] net/mlx5: support device removal event

["Nélio Laranjeiro" <nelio.laranjeiro@6wind.com>]

Hi Matan,

Please see comments below,

On Tue, Aug 29, 2017 at 11:30:44AM +0300, Matan Azrad wrote:
> Extend the LSC event handling to support the device removal as well.
> The Verbs library may send several related events, which are
> different from LSC event.
> 
> The mlx5 event handling has been made capable of receiving and
> signaling several event types at once.
> 
> This support includes next:
> 1. Removal event detection according to the user configuration.
> 2. Calling to all registered mlx5 removal callbacks.
> 3. Capabilities extension to include removal interrupt handling.
> 
> Signed-off-by: Matan Azrad <matan@mellanox.com>
> ---
>  drivers/net/mlx5/mlx5.c        |   2 +-
>  drivers/net/mlx5/mlx5_ethdev.c | 108 ++++++++++++++++++++++++++++-------------
>  2 files changed, 76 insertions(+), 34 deletions(-)
> 
> diff --git a/drivers/net/mlx5/mlx5.c b/drivers/net/mlx5/mlx5.c
> index bd66a7c..1a3d7f1 100644
> --- a/drivers/net/mlx5/mlx5.c
> +++ b/drivers/net/mlx5/mlx5.c
> @@ -865,7 +865,7 @@ static struct rte_pci_driver mlx5_driver = {
>  	},
>  	.id_table = mlx5_pci_id_map,
>  	.probe = mlx5_pci_probe,
> -	.drv_flags = RTE_PCI_DRV_INTR_LSC,
> +	.drv_flags = RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV,
>  };
>  
>  /**
> diff --git a/drivers/net/mlx5/mlx5_ethdev.c b/drivers/net/mlx5/mlx5_ethdev.c
> index 57f6237..2cf7726 100644
> --- a/drivers/net/mlx5/mlx5_ethdev.c
> +++ b/drivers/net/mlx5/mlx5_ethdev.c
> @@ -1112,47 +1112,83 @@ mlx5_ibv_device_to_pci_addr(const struct ibv_device *device,
>  }
>  
>  /**
> - * Link status handler.
> + * Update the link status.
>   *
>   * @param priv
>   *   Pointer to private structure.
> - * @param dev
> - *   Pointer to the rte_eth_dev structure.
>   *
>   * @return
> - *   Nonzero if the callback process can be called immediately.
> + *   Zero if the callback process can be called immediately.
>   */
>  static int
> -priv_dev_link_status_handler(struct priv *priv, struct rte_eth_dev *dev)
> +priv_link_status_update(struct priv *priv)
> +{
> +	struct rte_eth_link *link = &priv->dev->data->dev_link;
> +
> +	mlx5_link_update(priv->dev, 0);
> +	if (((link->link_speed == 0) && link->link_status) ||
> +		((link->link_speed != 0) && !link->link_status)) {
> +		/*
> +		 * Inconsistent status.
> +		 * The link status is read from Ethtool through an IOCTL,
> +		 * but as the PMD is working in polling mode it gets the port
> +		 * event before the Kernel driver had time to process it.
> +		 * PMD then request the link from the kernel but the event is
> +		 * still not processed (due to more urgent interrupts) and
> +		 * finally the PMD may get an inconsistent link.
> +		 * Setting alarm for later checking.
> +		 */

This comment is not totally correct, it is not the PMD which handles the LCS
event, but the application.  In testpmd the LSC event are polled in loop,
there is no guarantee that another application will do the same.
Remaining part is correct.

> +		if (!priv->pending_alarm) {
> +			priv->pending_alarm = 1;
> +			rte_eal_alarm_set(MLX5_ALARM_TIMEOUT_US,
> +				mlx5_dev_link_status_handler,
> +				priv->dev);

Indentation.

> +		}
> +		return 1;
> +	} else if (unlikely(priv->pending_alarm)) {
> +		/* In case of link interrupt while link alarm was setting. */
> +		priv->pending_alarm = 0;
> +		rte_eal_alarm_cancel(mlx5_dev_link_status_handler, priv->dev);
> +	}
> +	return 0;
> +}
> +
> +/**
> + * Device status handler.
> + *
> + * @param priv
> + *   Pointer to private structure.
> + * @param events
> + *   Pointer to event flags holder.
> + *
> + * @return
> + *   Events bitmap of callback process which can be called immediately.
> + */
> +static uint32_t
> +priv_dev_status_handler(struct priv *priv)
>  {
>  	struct ibv_async_event event;
> -	struct rte_eth_link *link = &dev->data->dev_link;
> -	int ret = 0;
> +	uint32_t ret = 0;
>  
>  	/* Read all message and acknowledge them. */
>  	for (;;) {
>  		if (ibv_get_async_event(priv->ctx, &event))
>  			break;
> -
> -		if (event.event_type != IBV_EVENT_PORT_ACTIVE &&
> -		    event.event_type != IBV_EVENT_PORT_ERR)
> +		if ((event.event_type == IBV_EVENT_PORT_ACTIVE ||
> +			event.event_type == IBV_EVENT_PORT_ERR) &&
> +			(priv->dev->data->dev_conf.intr_conf.lsc == 1))
> +			ret |= (1 << RTE_ETH_EVENT_INTR_LSC);
> +		else if (event.event_type == IBV_EVENT_DEVICE_FATAL &&
> +			priv->dev->data->dev_conf.intr_conf.rmv == 1)
> +			ret |= (1 << RTE_ETH_EVENT_INTR_RMV);
> +		else
>  			DEBUG("event type %d on port %d not handled",
> -			      event.event_type, event.element.port_num);
> +				event.event_type, event.element.port_num);

Why is this line modified?, the indentation was correct and nothing has been
modified.

>  		ibv_ack_async_event(&event);
>  	}
> -	mlx5_link_update(dev, 0);
> -	if (((link->link_speed == 0) && link->link_status) ||
> -	    ((link->link_speed != 0) && !link->link_status)) {
> -		if (!priv->pending_alarm) {
> -			/* Inconsistent status, check again later. */
> -			priv->pending_alarm = 1;
> -			rte_eal_alarm_set(MLX5_ALARM_TIMEOUT_US,
> -					  mlx5_dev_link_status_handler,
> -					  dev);
> -		}
> -	} else {
> -		ret = 1;
> -	}
> +	if (ret & (1 << RTE_ETH_EVENT_INTR_LSC))
> +		if (priv_link_status_update(priv))
> +			ret &= ~(1 << RTE_ETH_EVENT_INTR_LSC);
>  	return ret;
>  }
>  
> @@ -1172,11 +1208,11 @@ mlx5_dev_link_status_handler(void *arg)
>  	priv_lock(priv);
>  	assert(priv->pending_alarm == 1);
>  	priv->pending_alarm = 0;
> -	ret = priv_dev_link_status_handler(priv, dev);
> +	ret = priv_link_status_update(priv);
>  	priv_unlock(priv);
> -	if (ret)
> +	if (!ret)
>  		_rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL,
> -					      NULL);
> +			NULL);

Same as above, why is this line modified?

>  }
>  
>  /**
> @@ -1192,14 +1228,17 @@ mlx5_dev_interrupt_handler(void *cb_arg)
>  {
>  	struct rte_eth_dev *dev = cb_arg;
>  	struct priv *priv = dev->data->dev_private;
> -	int ret;
> +	uint32_t events;
>  
>  	priv_lock(priv);
> -	ret = priv_dev_link_status_handler(priv, dev);
> +	events = priv_dev_status_handler(priv);
>  	priv_unlock(priv);
> -	if (ret)
> +	if (events & (1 << RTE_ETH_EVENT_INTR_LSC))
>  		_rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL,
> -					      NULL);
> +			NULL);

Same question here,

> +	if (events & (1 << RTE_ETH_EVENT_INTR_RMV))
> +		_rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_RMV, NULL,
> +			NULL);

and here.

Thanks,

-- 
Nélio Laranjeiro
6WIND