Re: [PATCH v2 2/4] net: emac: implement TCP segmentation offload (TSO)

[David Miller <davem@davemloft.net>]

From: Christian Lamparter <chunkeey@gmail.com>
Date: Mon, 22 Oct 2018 13:04:12 +0200

> @@ -1452,8 +1509,49 @@ static inline u16 emac_tx_vlan(struct emac_instance *dev, struct sk_buff *skb)
>  	return 0;
>  }
>  
> +static netdev_tx_t
> +emac_start_xmit_sg(struct sk_buff *skb, struct net_device *ndev);
> +
> +static netdev_tx_t
> +emac_sw_tso(struct sk_buff *skb, struct net_device *ndev)
> +{
> +	struct emac_instance *dev = netdev_priv(ndev);
> +	struct sk_buff *segs, *curr;
> +	unsigned int i, frag_slots;
> +
> +	/* make sure to not overflow the tx ring */
> +	frag_slots = dev->tx_cnt;
> +	for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
> +		struct skb_frag_struct *frag = &skb_shinfo(skb)->frags[i];
> +
> +		frag_slots += mal_tx_chunks(skb_frag_size(frag));
> +
> +		if (frag_slots >= NUM_TX_BUFF)
> +			return NETDEV_TX_BUSY;
> +	};
> +
> +	segs = skb_gso_segment(skb, ndev->features &
> +					~(NETIF_F_TSO | NETIF_F_TSO6));

This NETDEV_TX_BUSY isn't going to work.

Your TX queue is awake.  So there won't be any guaranteed event to "wake up"
the queue and try sending this SKB again.

Please take a look at how the tg3.c driver handles this situation.  You have
to first stop the queue, do you overflow test, and then you can return
NETDEV_TX_BUSY if necessary.