[PATCH net-next] vlan: Add GRO support for non hardware accelerated vlan

[PATCH net-next] vlan: Add GRO support for non hardware accelerated vlan

[Toshiaki Makita]

Currently packets with non-hardware-accelerated vlan cannot be handled
by GRO. This causes low performance for 802.1ad and stacked vlan, as their
vlan tags are currently not stripped by hardware.

This patch adds GRO support for non-hardware-accelerated vlan and
improves receive performance of them.

Test Environment:
 vlan device (.1Q) on vlan device (.1ad) on ixgbe (82599)

Result:

- Before

$ netperf -t TCP_STREAM -H 192.168.20.2 -l 60
Recv   Send    Send
Socket Socket  Message  Elapsed
Size   Size    Size     Time     Throughput
bytes  bytes   bytes    secs.    10^6bits/sec

 87380  16384  16384    60.00    5233.17

Rx side CPU usage:
  %usr      %sys      %irq     %soft     %idle
  0.27     58.03      0.00     41.70      0.00

- After

$ netperf -t TCP_STREAM -H 192.168.20.2 -l 60
Recv   Send    Send
Socket Socket  Message  Elapsed
Size   Size    Size     Time     Throughput
bytes  bytes   bytes    secs.    10^6bits/sec

 87380  16384  16384    60.00    7586.85

Rx side CPU usage:
  %usr      %sys      %irq     %soft     %idle
  0.50     25.83      0.00     59.53     14.14

Signed-off-by: Toshiaki Makita <makita.toshiaki@lab.ntt.co.jp>
---
 net/8021q/vlan.c | 94 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 94 insertions(+)

diff --git a/net/8021q/vlan.c b/net/8021q/vlan.c
index 59555f0..0a9e8e1 100644
--- a/net/8021q/vlan.c
+++ b/net/8021q/vlan.c
@@ -618,6 +618,90 @@ out:
 	return err;
 }
 
+static struct sk_buff **vlan_gro_receive(struct sk_buff **head,
+					 struct sk_buff *skb)
+{
+	struct sk_buff *p, **pp = NULL;
+	struct vlan_hdr *vhdr;
+	unsigned int hlen, off_vlan;
+	const struct packet_offload *ptype;
+	__be16 type;
+	int flush = 1;
+
+	off_vlan = skb_gro_offset(skb);
+	hlen = off_vlan + sizeof(*vhdr);
+	vhdr = skb_gro_header_fast(skb, off_vlan);
+	if (skb_gro_header_hard(skb, hlen)) {
+		vhdr = skb_gro_header_slow(skb, hlen, off_vlan);
+		if (unlikely(!vhdr))
+			goto out;
+	}
+
+	type = vhdr->h_vlan_encapsulated_proto;
+
+	rcu_read_lock();
+	ptype = gro_find_receive_by_type(type);
+	if (!ptype)
+		goto out_unlock;
+
+	flush = 0;
+
+	for (p = *head; p; p = p->next) {
+		struct vlan_hdr *vhdr2;
+
+		if (!NAPI_GRO_CB(p)->same_flow)
+			continue;
+
+		vhdr2 = (struct vlan_hdr *)(p->data + off_vlan);
+		if (memcmp(vhdr, vhdr2, VLAN_HLEN))
+			NAPI_GRO_CB(p)->same_flow = 0;
+	}
+
+	skb_gro_pull(skb, sizeof(*vhdr));
+	skb_gro_postpull_rcsum(skb, vhdr, sizeof(*vhdr));
+	pp = ptype->callbacks.gro_receive(head, skb);
+
+out_unlock:
+	rcu_read_unlock();
+out:
+	NAPI_GRO_CB(skb)->flush |= flush;
+
+	return pp;
+}
+
+static int vlan_gro_complete(struct sk_buff *skb, int nhoff)
+{
+	struct vlan_hdr *vhdr = (struct vlan_hdr *)(skb->data + nhoff);
+	__be16 type = vhdr->h_vlan_encapsulated_proto;
+	struct packet_offload *ptype;
+	int err = -ENOENT;
+
+	rcu_read_lock();
+	ptype = gro_find_complete_by_type(type);
+	if (ptype)
+		err = ptype->callbacks.gro_complete(skb, nhoff + sizeof(*vhdr));
+
+	rcu_read_unlock();
+	return err;
+}
+
+static struct packet_offload vlan_packet_offloads[] __read_mostly = {
+	{
+		.type = cpu_to_be16(ETH_P_8021Q),
+		.callbacks = {
+			.gro_receive = vlan_gro_receive,
+			.gro_complete = vlan_gro_complete,
+		},
+	},
+	{
+		.type = cpu_to_be16(ETH_P_8021AD),
+		.callbacks = {
+			.gro_receive = vlan_gro_receive,
+			.gro_complete = vlan_gro_complete,
+		},
+	},
+};
+
 static int __net_init vlan_init_net(struct net *net)
 {
 	struct vlan_net *vn = net_generic(net, vlan_net_id);
@@ -645,6 +729,7 @@ static struct pernet_operations vlan_net_ops = {
 static int __init vlan_proto_init(void)
 {
 	int err;
+	unsigned int i;
 
 	pr_info("%s v%s\n", vlan_fullname, vlan_version);
 
@@ -668,6 +753,9 @@ static int __init vlan_proto_init(void)
 	if (err < 0)
 		goto err5;
 
+	for (i = 0; i < ARRAY_SIZE(vlan_packet_offloads); i++)
+		dev_add_offload(&vlan_packet_offloads[i]);
+
 	vlan_ioctl_set(vlan_ioctl_handler);
 	return 0;
 
@@ -685,7 +773,13 @@ err0:
 
 static void __exit vlan_cleanup_module(void)
 {
+	unsigned int i;
+
 	vlan_ioctl_set(NULL);
+
+	for (i = 0; i < ARRAY_SIZE(vlan_packet_offloads); i++)
+		dev_remove_offload(&vlan_packet_offloads[i]);
+
 	vlan_netlink_fini();
 
 	unregister_netdevice_notifier(&vlan_notifier_block);
-- 
1.8.1.2

Re: [PATCH net-next] vlan: Add GRO support for non hardware accelerated vlan

[Eric Dumazet]

On Thu, 2015-05-28 at 20:17 +0900, Toshiaki Makita wrote:
> Currently packets with non-hardware-accelerated vlan cannot be handled
> by GRO. This causes low performance for 802.1ad and stacked vlan, as their
> vlan tags are currently not stripped by hardware.
> 
> This patch adds GRO support for non-hardware-accelerated vlan and
> improves receive performance of them.

Very nice patch !

> 
> Signed-off-by: Toshiaki Makita <makita.toshiaki@lab.ntt.co.jp>
> ---
>  net/8021q/vlan.c | 94 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++
>  1 file changed, 94 insertions(+)
> 
> diff --git a/net/8021q/vlan.c b/net/8021q/vlan.c
> index 59555f0..0a9e8e1 100644
> --- a/net/8021q/vlan.c
> +++ b/net/8021q/vlan.c
> @@ -618,6 +618,90 @@ out:
>  	return err;
>  }
>  
> +		vhdr2 = (struct vlan_hdr *)(p->data + off_vlan);
> +		if (memcmp(vhdr, vhdr2, VLAN_HLEN))
> +			NAPI_GRO_CB(p)->same_flow = 0;
> +	}


This memcmp() is quite expensive, you better use a helper like :

/* vlan header only guaranteed to be 16bit aligned */
static bool vlan_hdr_compare(const struct vlan_hdr *h1, const struct vlan_hdr *h2)
{
#if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
	return *(u32 *)h1 != *(u32 *)h2;
#else
	return (((__force u32)h1->h_vlan_TCI ^ (__force u32)h2->h_vlan_TCI) |
		((__force u32)h1->h_vlan_encapsulated_proto ^
		 (__force u32)h2->h_vlan_encapsulated_proto)) != 0;
#endif
}

Re: [PATCH net-next] vlan: Add GRO support for non hardware accelerated vlan

[Toshiaki Makita]

On 15/05/28 (木) 21:02, Eric Dumazet wrote:
> On Thu, 2015-05-28 at 20:17 +0900, Toshiaki Makita wrote:
>> Currently packets with non-hardware-accelerated vlan cannot be handled
>> by GRO. This causes low performance for 802.1ad and stacked vlan, as their
>> vlan tags are currently not stripped by hardware.
>>
>> This patch adds GRO support for non-hardware-accelerated vlan and
>> improves receive performance of them.
>
> Very nice patch !
>
>>
>> Signed-off-by: Toshiaki Makita <makita.toshiaki@lab.ntt.co.jp>
>> ---
>>   net/8021q/vlan.c | 94 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++
>>   1 file changed, 94 insertions(+)
>>
>> diff --git a/net/8021q/vlan.c b/net/8021q/vlan.c
>> index 59555f0..0a9e8e1 100644
>> --- a/net/8021q/vlan.c
>> +++ b/net/8021q/vlan.c
>> @@ -618,6 +618,90 @@ out:
>>   	return err;
>>   }
>>
>> +		vhdr2 = (struct vlan_hdr *)(p->data + off_vlan);
>> +		if (memcmp(vhdr, vhdr2, VLAN_HLEN))
>> +			NAPI_GRO_CB(p)->same_flow = 0;
>> +	}
>
>
> This memcmp() is quite expensive, you better use a helper like :
>
> /* vlan header only guaranteed to be 16bit aligned */
> static bool vlan_hdr_compare(const struct vlan_hdr *h1, const struct vlan_hdr *h2)
> {
> #if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
> 	return *(u32 *)h1 != *(u32 *)h2;
> #else
> 	return (((__force u32)h1->h_vlan_TCI ^ (__force u32)h2->h_vlan_TCI) |
> 		((__force u32)h1->h_vlan_encapsulated_proto ^
> 		 (__force u32)h2->h_vlan_encapsulated_proto)) != 0;
> #endif
> }

Hi Eric,

Thank you for your reviewing.
Indeed, memcmp() is not good for performance.
I'll include your feedback in v2.

Thanks,
Toshiaki Makita