Re: [PATCH v3 net-next 1/2] ibmvnic: Use atomic64_t for queue stats

[Jakub Kicinski <kuba@kernel.org>]

On Mon, 14 Jul 2025 13:35:06 -0400 Mingming Cao wrote:
> diff --git a/drivers/net/ethernet/ibm/ibmvnic.c b/drivers/net/ethernet/ibm/ibmvnic.c
> index 92647e137cf8..79fdba4293a4 100644
> --- a/drivers/net/ethernet/ibm/ibmvnic.c
> +++ b/drivers/net/ethernet/ibm/ibmvnic.c
> @@ -2314,9 +2314,17 @@ static void ibmvnic_tx_scrq_clean_buffer(struct ibmvnic_adapter *adapter,
>  		tx_buff = &tx_pool->tx_buff[index];
>  		adapter->netdev->stats.tx_packets--;
>  		adapter->netdev->stats.tx_bytes -= tx_buff->skb->len;
> -		adapter->tx_stats_buffers[queue_num].batched_packets--;
> -		adapter->tx_stats_buffers[queue_num].bytes -=
> -						tx_buff->skb->len;
> +		atomic64_dec(&adapter->tx_stats_buffers[queue_num].batched_packets);
> +		if (atomic64_sub_return(tx_buff->skb->len,
> +					&adapter->tx_stats_buffers[queue_num].bytes) < 0) {
> +			atomic64_set(&adapter->tx_stats_buffers[queue_num].bytes, 0);
> +			netdev_warn(adapter->netdev,

Any datapath print needs to be rate limited, otherwise it may flood
the logs.

> +				    "TX stats underflow on queue %u: bytes (%lld) < skb->len (%u),\n"
> +				    "clamping to 0\n",
> +				    queue_num,
> +				    atomic64_read(&adapter->tx_stats_buffers[queue_num].bytes),
> +				    tx_buff->skb->len);
> +		}
>  		dev_kfree_skb_any(tx_buff->skb);
>  		tx_buff->skb = NULL;
>  		adapter->netdev->stats.tx_dropped++;
> @@ -2652,10 +2660,10 @@ static netdev_tx_t ibmvnic_xmit(struct sk_buff *skb, struct net_device *netdev)
>  	netdev->stats.tx_packets += tx_bpackets + tx_dpackets;
>  	adapter->tx_send_failed += tx_send_failed;
>  	adapter->tx_map_failed += tx_map_failed;
> -	adapter->tx_stats_buffers[queue_num].batched_packets += tx_bpackets;
> -	adapter->tx_stats_buffers[queue_num].direct_packets += tx_dpackets;
> -	adapter->tx_stats_buffers[queue_num].bytes += tx_bytes;
> -	adapter->tx_stats_buffers[queue_num].dropped_packets += tx_dropped;
> +	atomic64_add(tx_bpackets, &adapter->tx_stats_buffers[queue_num].batched_packets);
> +	atomic64_add(tx_dpackets, &adapter->tx_stats_buffers[queue_num].direct_packets);
> +	atomic64_add(tx_bytes, &adapter->tx_stats_buffers[queue_num].bytes);
> +	atomic64_add(tx_dropped, &adapter->tx_stats_buffers[queue_num].dropped_packets);

Are atomics really cheap on your platform? Normally having these many
atomic ops per packet would bring performance concerns.
I assume queue accesses are already protected by some sort of a lock
so isn't it enough to make the stats per queue without making them
atomic?
-- 
pw-bot: cr