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

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

[David Miller]

It just xor hashes over IPv4/IPv6 addresses and ports of transport.

The only assumption it makes is that skb_network_header() is set
correctly.

With bug fixes from Eric Dumazet.

Signed-off-by: David S. Miller <davem@davemloft.net>
---
 net/core/dev.c |   52 ++++++++++++++++++++++++++++++++++++++++++++++++++++
 1 files changed, 52 insertions(+), 0 deletions(-)

diff --git a/net/core/dev.c b/net/core/dev.c
index 7ca9564..467bfb3 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -121,6 +121,9 @@
 #include <linux/ctype.h>
 #include <linux/if_arp.h>
 #include <linux/if_vlan.h>
+#include <linux/ip.h>
+#include <linux/ipv6.h>
+#include <linux/in.h>
 
 #include "net-sysfs.h"
 
@@ -1665,6 +1668,53 @@ out_kfree_skb:
  *          --BLG
  */
 
+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;
+	}
+
+	ports = (u32 *) (skb_network_header(skb) + (ihl * 4));
+
+	hash = 0;
+	while (alen--)
+		hash ^= *addr++;
+
+	switch (ip_proto) {
+	case IPPROTO_TCP:
+	case IPPROTO_UDP:
+	case IPPROTO_DCCP:
+	case IPPROTO_ESP:
+	case IPPROTO_AH:
+	case IPPROTO_SCTP:
+	case IPPROTO_UDPLITE:
+		hash ^= *ports;
+		break;
+
+	default:
+		break;
+	}
+
+	return hash % dev->real_num_tx_queues;
+}
+
 static struct netdev_queue *dev_pick_tx(struct net_device *dev,
 					struct sk_buff *skb)
 {
@@ -1672,6 +1722,8 @@ static struct netdev_queue *dev_pick_tx(struct net_device *dev,
 
 	if (dev->select_queue)
 		queue_index = dev->select_queue(dev, skb);
+	else if (dev->real_num_tx_queues > 1)
+		queue_index = simple_tx_hash(dev, skb);
 
 	skb_set_queue_mapping(skb, queue_index);
 	return netdev_get_tx_queue(dev, queue_index);
-- 
1.5.6.2.255.gbed62

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

[Andi Kleen]

David Miller <davem-fT/PcQaiUtIeIZ0/mPfg9Q@public.gmane.org> 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

--
To unsubscribe from this list: send the line "unsubscribe linux-wireless" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

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

[Brice Goglin]

David Miller wrote:
> It just xor hashes over IPv4/IPv6 addresses and ports of transport.
>
> The only assumption it makes is that skb_network_header() is set
> correctly.
>   

Hey David,

Sorry for the late reply, I didn't have time to look at multiqueue tx
details before. I have some curiosity-questions about your hashing.

> +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;
> +	}
>   

What's your plan for other protocols here? Should we add an optional
tx_hash() method in struct packet_type?

> @@ -1672,6 +1722,8 @@ static struct netdev_queue *dev_pick_tx(struct net_device *dev,
>  
>  	if (dev->select_queue)
>  		queue_index = dev->select_queue(dev, skb);
> +	else if (dev->real_num_tx_queues > 1)
> +		queue_index = simple_tx_hash(dev, skb);
>   

Some devices might will add some protocol specific hashing (IP and IPv6)
that could look like what simple_tx_hash() does above, right? But if we
add some new protocols in sample_tx_hash(), it will be ignored by these
devices, and we might have to update the hashing in all select_queue()
methods as well.

So I wonder if we shouldn't change all this into:
1) simple_tx_hash(skb) returns a big protocol-independent integer (like
a concatenation or basic xor of IP proto/addr/...). It basically just
converts the skb header in something flat and proto-independent.
2.a) if the device provides select_queue(), we pass this integer to it.
select_queue() either uses the skb headers if it really wants to hash IP
or IPv6 precisely, or uses our protocol-independent integer and just
hash it in a dumb way without caring about the protocol hidden behind it.
2.b) if there's no select_queue(), we hash the big integer depending on
the number of tx queues

This way, we can easily add some protocol specific hashing to all
drivers without having to modify select_queue() everywhere. Of course,
if we don't care about hashing anything but IP/IPv6, it doesn't matter :)

Brice

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

[Brice Goglin]

Brice Goglin wrote:
> So I wonder if we shouldn't change all this into:
> 1) simple_tx_hash(skb) returns a big protocol-independent integer (like
> a concatenation or basic xor of IP proto/addr/...). It basically just
> converts the skb header in something flat and proto-independent.
> 2.a) if the device provides select_queue(), we pass this integer to it.
> select_queue() either uses the skb headers if it really wants to hash IP
> or IPv6 precisely, or uses our protocol-independent integer and just
> hash it in a dumb way without caring about the protocol hidden behind it.
> 2.b) if there's no select_queue(), we hash the big integer depending on
> the number of tx queues
>   

Or (easier) just export simple_tx_hash() to modules so that drivers can
rely on it to hash protocols that they have no clue about?

Brice