Re: [PATCH 11/31]: net: Implement simple sw TX hashing.

[Andi Kleen <andi@firstfloor.org>]

David Miller <davem@davemloft.net> writes:

> +static u16 simple_tx_hash(struct net_device *dev, struct sk_buff *skb)
> +{
> +	u32 *addr, *ports, hash, ihl;
> +	u8 ip_proto;
> +	int alen;
> +
> +	switch (skb->protocol) {
> +	case __constant_htons(ETH_P_IP):
> +		ip_proto = ip_hdr(skb)->protocol;
> +		addr = &ip_hdr(skb)->saddr;
> +		ihl = ip_hdr(skb)->ihl;
> +		alen = 2;
> +		break;
> +	case __constant_htons(ETH_P_IPV6):
> +		ip_proto = ipv6_hdr(skb)->nexthdr;
> +		addr = &ipv6_hdr(skb)->saddr.s6_addr32[0];
> +		ihl = (40 >> 2);
> +		alen = 8;
> +		break;
> +	default:
> +		return 0;

Could you add a linuxmib counter for this default event? Just so that 
people can diagnose this more easily.

> +
> +	ports = (u32 *) (skb_network_header(skb) + (ihl * 4));
> +
> +	hash = 0;
> +	while (alen--)
> +		hash ^= *addr++;

Are you sure the hash behaves the same between big endian and little
endian? Perhaps it would be safer to always convert endian even if
that complicates the code a bit.

> +		hash ^= *ports;
> +		break;
> +
> +	default:
> +		break;

And also a linuxmib counter here.

> +	}
> +
> +	return hash % dev->real_num_tx_queues;

I haven't rechecked it in detail, but I suspect some more
folding of higher bits would be better again for endian safety.

-Andi