On Tue, 2013-02-05 at 17:00 +0530, ganesanr@xxxxxxxxxxxx wrote:
> From: Ganesan Ramalingam <ganesanr@xxxxxxxxxxxx>
>
> Add support for the Network Accelerator Engine on Netlogic XLR/XLS
> MIPS SoCs. The XLR/XLS NAE blocks can be configured as one 10G
> interface or four 1G interfaces. This driver supports blocks
> with 1G ports.
>
> Signed-off-by: Ganesan Ramalingam <ganesanr@xxxxxxxxxxxx>
> ---
> This patch has to be merged through netdev tree.
>
> Changes since v1:
> * Implementing ethtool using phylib
> * Free skb buf incase of nlm_fmn_send() fail
> * Use of ndo_get_stats64 instead of ndo_get_stats
> * Watchdog timeo fix
[...]
> +static void xlr_stats(struct net_device *ndev, struct rtnl_link_stats64 *stats)
> +{
> + struct xlr_net_priv *priv = netdev_priv(ndev);
> +
> + stats->rx_packets = xlr_nae_rdreg(priv->base_addr, RX_PACKET_COUNTER);
> + stats->tx_packets = xlr_nae_rdreg(priv->base_addr, TX_PACKET_COUNTER);
> + stats->rx_bytes = xlr_nae_rdreg(priv->base_addr, RX_BYTE_COUNTER);
> + stats->tx_bytes = xlr_nae_rdreg(priv->base_addr, TX_BYTE_COUNTER);
> + stats->tx_errors = xlr_nae_rdreg(priv->base_addr, TX_FCS_ERROR_COUNTER);
> + stats->rx_dropped = xlr_nae_rdreg(priv->base_addr,
> + RX_DROP_PACKET_COUNTER);
> + stats->tx_dropped = xlr_nae_rdreg(priv->base_addr,
> + TX_DROP_FRAME_COUNTER);
[...]
But these are still 32-bit counters in hardware, yes? So I think you
need to arrange to poll at least the byte counters every few seconds and
maintain the higher bits in software e.g.:
struct xlr_stats {
spinlock_t lock;
struct timer timer;
u64 rx_bytes;
u64 tx_bytes;
...
};
static void xlr_poll_stats(struct xlr_net_priv *priv)
{
struct xlr_stats = &priv->stats;
u32 counter;
spin_lock_bh(&stats->lock);
counter = xlr_nae_rdreg(priv->base_addr, RX_BYTE_COUNTER);
stats->rx_bytes += (u32)(counter - stats->rx_bytes);
counter = xlr_nae_rdreg(priv->base_addr, TX_BYTE_COUNTER);
stats->tx_bytes += (u32)(counter - stats->tx_bytes);
...
spin_unlock_bh(&stats->lock);
}
static void xlr_poll_stats_timer(void *data)
{
struct xlr_net_priv *priv = data;
xlr_poll_stats(priv);
mod_timer(&priv->stats.timer, jiffies + 2 * HZ);
}
static void xlr_stats(struct net_device *ndev, struct rtnl_link_stats64 *stats)
{
struct xlr_net_priv *priv = netdev_priv(ndev);
xlr_poll_stats(priv);
...
}
static int xlr_net_probe(struct platform_device *pdev)
{
...
spin_lock_init(&priv->stats.lock);
setup_timer(&priv->stats.timer, xlr_poll_stats_timer, priv);
...
}
Plus add_timer() and del_timer_sync() when starting/stopping the
interface. Also you need to clear the software stats if you have to
reset the hardware. All of which is a fair amount of work but you can't
avoid it if the hardware counters just aren't wide enough.
Ben.
--
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.
