[dpdk-dev] [PATCH 2/3] net/mlx4: adjust removal error

[Adrien Mazarguil]

On Thu, Nov 02, 2017 at 03:42:03PM +0000, Matan Azrad wrote:
> Fail-safe PMD expects to get -ENODEV error value if sub PMD control
> command fails because of device removal.
> 
> Make control callbacks return with -ENODEV when the device has
> disappeared.
> 
> Signed-off-by: Matan Azrad <matan at mellanox.com>

I think there are a several inconsistencies regarding the places where
mlx4_removed() is used, this could lead to mistakes or redundant calls to
this function later on.

You have to choose between low-level internal functions
(e.g. mlx4_set_sysfs_ulong()) or user-facing ones from the eth_dev_ops
interface (e.g. mlx4_dev_set_link_up()), but neither intermediate functions
nor a mix of all approaches.

Standardizing on low-level functions is not practical as it means you'd have
to check for a device removal after each ibv_*() call. Therefore my
suggestion is to check it at the highest level, in all functions exposed
though mlx4_dev_ops in case of error, even innocuous one like
mlx4_stats_get() and those returning void (rte_errno can still be set), all
in the name of consistency.

The mlx4_removed() documentation should be updated to reflect the places
it's supposed to be called as well. All this means a larger patch is
necessary.

See below for coding style issues.

> ---
>  drivers/net/mlx4/mlx4.h        |  1 +
>  drivers/net/mlx4/mlx4_ethdev.c | 38 ++++++++++++++++++++++++++++++++++----
>  drivers/net/mlx4/mlx4_flow.c   |  2 ++
>  drivers/net/mlx4/mlx4_intr.c   |  5 ++++-
>  drivers/net/mlx4/mlx4_rxq.c    |  1 +
>  drivers/net/mlx4/mlx4_txq.c    |  1 +
>  6 files changed, 43 insertions(+), 5 deletions(-)
> 
> diff --git a/drivers/net/mlx4/mlx4.h b/drivers/net/mlx4/mlx4.h
> index e0a9853..cac9654 100644
> --- a/drivers/net/mlx4/mlx4.h
> +++ b/drivers/net/mlx4/mlx4.h
> @@ -149,6 +149,7 @@ int mlx4_flow_ctrl_get(struct rte_eth_dev *dev,
>  		       struct rte_eth_fc_conf *fc_conf);
>  int mlx4_flow_ctrl_set(struct rte_eth_dev *dev,
>  		       struct rte_eth_fc_conf *fc_conf);
> +int mlx4_removed(const struct priv *priv);
>  
>  /* mlx4_intr.c */
>  
> diff --git a/drivers/net/mlx4/mlx4_ethdev.c b/drivers/net/mlx4/mlx4_ethdev.c
> index b0acd12..76914b0 100644
> --- a/drivers/net/mlx4/mlx4_ethdev.c
> +++ b/drivers/net/mlx4/mlx4_ethdev.c
> @@ -312,6 +312,8 @@
>  
>  	ret = mlx4_sysfs_write(priv, name, value_str, (sizeof(value_str) - 1));
>  	if (ret < 0) {
> +		if (mlx4_removed(priv))
> +			ret = -ENODEV;
>  		DEBUG("cannot write %s `%s' (%lu) to sysfs: %s",
>  		      name, value_str, value, strerror(rte_errno));
>  		return ret;
> @@ -340,15 +342,19 @@
>  
>  	if (sock == -1) {
>  		rte_errno = errno;
> -		return -rte_errno;
> +		goto error;
>  	}
>  	ret = mlx4_get_ifname(priv, &ifr->ifr_name);
>  	if (!ret && ioctl(sock, req, ifr) == -1) {
>  		rte_errno = errno;
> -		ret = -rte_errno;
> +		close(sock);
> +		goto error;
>  	}
>  	close(sock);
>  	return ret;
> +error:
> +	mlx4_removed(priv);
> +	return -rte_errno;
>  }
>  
>  /**
> @@ -473,13 +479,17 @@
>  	if (up) {
>  		err = mlx4_set_flags(priv, ~IFF_UP, IFF_UP);
>  		if (err)
> -			return err;
> +			goto error;
>  	} else {
>  		err = mlx4_set_flags(priv, ~IFF_UP, ~IFF_UP);
>  		if (err)
> -			return err;
> +			goto error;
>  	}
>  	return 0;
> +error:
> +	if (mlx4_removed(priv))
> +		return -ENODEV;
> +	return err;
>  }
>  
>  /**
> @@ -947,6 +957,7 @@ enum rxmode_toggle {
>  
>  	ifr.ifr_data = (void *)ðpause;
>  	if (mlx4_ifreq(priv, SIOCETHTOOL, &ifr)) {
> +		mlx4_removed(priv);
>  		ret = rte_errno;
>  		WARN("ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM)"
>  		     " failed: %s",
> @@ -1002,6 +1013,7 @@ enum rxmode_toggle {
>  	else
>  		ethpause.tx_pause = 0;
>  	if (mlx4_ifreq(priv, SIOCETHTOOL, &ifr)) {
> +		mlx4_removed(priv);
>  		ret = rte_errno;
>  		WARN("ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
>  		     " failed: %s",
> @@ -1013,3 +1025,21 @@ enum rxmode_toggle {
>  	assert(ret >= 0);
>  	return -ret;
>  }

Missing empty line.

> +/**
> + * Check if mlx4 device was removed.

"mlx4" is a somewhat redundant given PMD name.

A separate paragraph should describe where this function is supposed to be
called.

> + *
> + * @param priv
> + *   Pointer to private structure.
> + *
> + * @return
> + *   -ENODEV when device is removed and rte_errno is set, otherwise 0.
> + */
> +int
> +mlx4_removed(const struct priv *priv)
> +{
> +	struct ibv_device_attr device_attr;
> +
> +	if (ibv_query_device(priv->ctx, &device_attr) == EIO)
> +		return -(rte_errno = ENODEV);

Although a nice shortcut, coding rules don't allow this. You have to assign
rte_errno on its own separate line. My suggestion if you want to avoid a
block would be to return 0 directly when != EIO.

> +	return 0;
> +}
> diff --git a/drivers/net/mlx4/mlx4_flow.c b/drivers/net/mlx4/mlx4_flow.c
> index 8b87b29..606c888 100644
> --- a/drivers/net/mlx4/mlx4_flow.c
> +++ b/drivers/net/mlx4/mlx4_flow.c
> @@ -1069,6 +1069,8 @@ struct mlx4_drop {
>  	err = errno;
>  	msg = "flow rule rejected by device";
>  error:
> +	if (mlx4_removed(priv))
> +		err = ENODEV;
>  	return rte_flow_error_set
>  		(error, err, RTE_FLOW_ERROR_TYPE_HANDLE, flow, msg);
>  }
> diff --git a/drivers/net/mlx4/mlx4_intr.c b/drivers/net/mlx4/mlx4_intr.c
> index b17d109..0ebdb28 100644
> --- a/drivers/net/mlx4/mlx4_intr.c
> +++ b/drivers/net/mlx4/mlx4_intr.c
> @@ -359,7 +359,10 @@
>  			ret = EINVAL;
>  	}
>  	if (ret) {
> -		rte_errno = ret;
> +		if (mlx4_removed(dev->data->dev_private))
> +			ret = ENODEV;
> +		else
> +			rte_errno = ret;
>  		WARN("unable to disable interrupt on rx queue %d",
>  		     idx);
>  	} else {
> diff --git a/drivers/net/mlx4/mlx4_rxq.c b/drivers/net/mlx4/mlx4_rxq.c
> index 7fe21b6..43dad26 100644
> --- a/drivers/net/mlx4/mlx4_rxq.c
> +++ b/drivers/net/mlx4/mlx4_rxq.c
> @@ -832,6 +832,7 @@ void mlx4_rss_detach(struct mlx4_rss *rss)
>  	ret = rte_errno;
>  	mlx4_rx_queue_release(rxq);
>  	rte_errno = ret;
> +	mlx4_removed(priv);
>  	assert(rte_errno > 0);
>  	return -rte_errno;
>  }
> diff --git a/drivers/net/mlx4/mlx4_txq.c b/drivers/net/mlx4/mlx4_txq.c
> index a9c5bd2..09bdfd8 100644
> --- a/drivers/net/mlx4/mlx4_txq.c
> +++ b/drivers/net/mlx4/mlx4_txq.c
> @@ -372,6 +372,7 @@ struct txq_mp2mr_mbuf_check_data {
>  	ret = rte_errno;
>  	mlx4_tx_queue_release(txq);
>  	rte_errno = ret;
> +	mlx4_removed(priv);
>  	assert(rte_errno > 0);
>  	return -rte_errno;
>  }
> -- 
> 1.8.3.1
> 

-- 
Adrien Mazarguil
6WIND