On Thu, Apr 25, 2024 at 5:04 PM Felix Fietkau <[email protected]> wrote:
>
> When forwarding TCP after GRO, software segmentation is very expensive,
> especially when the checksum needs to be recalculated.
> One case where that's currently unavoidable is when routing packets over
> PPPoE. Performance improves significantly when using fraglist GRO
> implemented in the same way as for UDP.
>
> When NETIF_F_GRO_FRAGLIST is enabled, perform a lookup for an established
> socket in the same netns as the receiving device. While this may not
> cover all relevant use cases in multi-netns configurations, it should be
> good enough for most configurations that need this.
>
> Here's a measurement of running 2 TCP streams through a MediaTek MT7622
> device (2-core Cortex-A53), which runs NAT with flow offload enabled from
> one ethernet port to PPPoE on another ethernet port + cake qdisc set to
> 1Gbps.
>
> rx-gro-list off: 630 Mbit/s, CPU 35% idle
> rx-gro-list on: 770 Mbit/s, CPU 40% idle
>
> Signe-off-by: Felix Fietkau <[email protected]>
> ---
> net/ipv4/tcp_offload.c | 48 +++++++++++++++++++++++++++++++++++++
> net/ipv6/tcpv6_offload.c | 51 ++++++++++++++++++++++++++++++++++++++++
> 2 files changed, 99 insertions(+)
>
> diff --git a/net/ipv4/tcp_offload.c b/net/ipv4/tcp_offload.c
> index 520fd425ab19..3bb96a110402 100644
> --- a/net/ipv4/tcp_offload.c
> +++ b/net/ipv4/tcp_offload.c
> @@ -405,6 +405,52 @@ void tcp_gro_complete(struct sk_buff *skb)
> }
> EXPORT_SYMBOL(tcp_gro_complete);
>
> +static void tcp4_check_fraglist_gro(struct list_head *head, struct sk_buff *skb)
> +{
> + const struct iphdr *iph = skb_gro_network_header(skb);
I do not think loading iph before all skb_gro_header() and
skb_gro_header_slow() calls is wise.
pskb_may_pull() can re-allocate skb->head
> + struct net *net = dev_net(skb->dev);
> + unsigned int off, hlen, thlen;
> + struct sk_buff *p;
> + struct tcphdr *th;
> + struct sock *sk;
> + int iif, sdif;
> +
> + if (!(skb->dev->features & NETIF_F_GRO_FRAGLIST))
> + return;
> +
> + off = skb_gro_offset(skb);
> + hlen = off + sizeof(*th);
> + th = skb_gro_header(skb, hlen, off);
> + if (unlikely(!th))
> + return;
> +
> + thlen = th->doff * 4;
> + if (thlen < sizeof(*th))
> + return;
> +
> + hlen = off + thlen;
> + if (!skb_gro_may_pull(skb, hlen)) {
> + th = skb_gro_header_slow(skb, hlen, off);
> + if (unlikely(!th))
> + return;
> + }
> +
> + p = tcp_gro_lookup(head, th);
> + if (p) {
> + NAPI_GRO_CB(skb)->is_flist = NAPI_GRO_CB(p)->is_flist;
> + return;
> + }
> +
> + inet_get_iif_sdif(skb, &iif, &sdif);
> + sk = __inet_lookup_established(net, net->ipv4.tcp_death_row.hashinfo,
> + iph->saddr, th->source,
> + iph->daddr, ntohs(th->dest),
> + iif, sdif);
> + NAPI_GRO_CB(skb)->is_flist = !sk;
> + if (sk)
> + sock_put(sk);
> +}
> +
> INDIRECT_CALLABLE_SCOPE
> struct sk_buff *tcp4_gro_receive(struct list_head *head, struct sk_buff *skb)
> {
> @@ -416,6 +462,8 @@ struct sk_buff *tcp4_gro_receive(struct list_head *head, struct sk_buff *skb)
> return NULL;
> }
>
I would probably pull the whole TCP header here, before calling
tcp4_check_fraglist_gro(head, skb)
and no longer do this twice from tcp4_check_fraglist_gro() and tcp_gro_receive()
Perhaps define a new inline helper, that will be called from
tcp4_gro_receive() and tcp6_gro_receive(),
and not anymore from tcp_gro_receive()
static inline struct tcphdr *tcp_gro_pull_header(...)
{
....
off = skb_gro_offset(skb);
hlen = off + sizeof(*th);
th = skb_gro_header(skb, hlen, off);
if (unlikely(!th))
return NULL;
thlen = th->doff * 4;
if (thlen < sizeof(*th))
return NULL;
hlen = off + thlen;
if (!skb_gro_may_pull(skb, hlen))
th = skb_gro_header_slow(skb, hlen, off);
return th;
}
> + tcp4_check_fraglist_gro(head, skb);
> +
> return tcp_gro_receive(head, skb);
> }
>
> diff --git a/net/ipv6/tcpv6_offload.c b/net/ipv6/tcpv6_offload.c
> index c97d55cf036f..7948420dcad0 100644
> --- a/net/ipv6/tcpv6_offload.c
> +++ b/net/ipv6/tcpv6_offload.c
> @@ -7,12 +7,61 @@
> */
> #include <linux/indirect_call_wrapper.h>
> #include <linux/skbuff.h>
> +#include <net/inet6_hashtables.h>
> #include <net/gro.h>
> #include <net/protocol.h>
> #include <net/tcp.h>
> #include <net/ip6_checksum.h>
> #include "ip6_offload.h"
>
> +static void tcp6_check_fraglist_gro(struct list_head *head, struct sk_buff *skb)
> +{
> +#if IS_ENABLED(CONFIG_IPV6)
> + const struct ipv6hdr *hdr = skb_gro_network_header(skb);
> + struct net *net = dev_net(skb->dev);
> + unsigned int off, hlen, thlen;
> + struct sk_buff *p;
> + struct tcphdr *th;
> + struct sock *sk;
> + int iif, sdif;
> +
> + if (!(skb->dev->features & NETIF_F_GRO_FRAGLIST))
> + return;
> +
> + off = skb_gro_offset(skb);
> + hlen = off + sizeof(*th);
> + th = skb_gro_header(skb, hlen, off);
> + if (unlikely(!th))
> + return;
> +
> + thlen = th->doff * 4;
> + if (thlen < sizeof(*th))
> + return;
> +
> + hlen = off + thlen;
> + if (!skb_gro_may_pull(skb, hlen)) {
> + th = skb_gro_header_slow(skb, hlen, off);
> + if (unlikely(!th))
> + return;
> + }
> +
> + p = tcp_gro_lookup(head, th);
> + if (p) {
> + NAPI_GRO_CB(skb)->is_flist = NAPI_GRO_CB(p)->is_flist;
> + return;
> + }
> +
> + inet6_get_iif_sdif(skb, &iif, &sdif);
> + sk = __inet6_lookup_established(net, net->ipv4.tcp_death_row.hashinfo,
> + &hdr->saddr, th->source,
> + &hdr->daddr, ntohs(th->dest),
> + iif, sdif);
> + NAPI_GRO_CB(skb)->is_flist = !sk;
> + if (sk)
> + sock_put(sk);
> +#endif /* IS_ENABLED(CONFIG_IPV6) */
> +}
> +
> INDIRECT_CALLABLE_SCOPE
> struct sk_buff *tcp6_gro_receive(struct list_head *head, struct sk_buff *skb)
> {
> @@ -24,6 +73,8 @@ struct sk_buff *tcp6_gro_receive(struct list_head *head, struct sk_buff *skb)
> return NULL;
> }
>
> + tcp6_check_fraglist_gro(head, skb);
> +
> return tcp_gro_receive(head, skb);
> }
>
> --
> 2.44.0
>
On 25.04.24 17:26, Eric Dumazet wrote:
> On Thu, Apr 25, 2024 at 5:04 PM Felix Fietkau <[email protected]> wrote:
>>
>> When forwarding TCP after GRO, software segmentation is very expensive,
>> especially when the checksum needs to be recalculated.
>> One case where that's currently unavoidable is when routing packets over
>> PPPoE. Performance improves significantly when using fraglist GRO
>> implemented in the same way as for UDP.
>>
>> When NETIF_F_GRO_FRAGLIST is enabled, perform a lookup for an established
>> socket in the same netns as the receiving device. While this may not
>> cover all relevant use cases in multi-netns configurations, it should be
>> good enough for most configurations that need this.
>>
>> Here's a measurement of running 2 TCP streams through a MediaTek MT7622
>> device (2-core Cortex-A53), which runs NAT with flow offload enabled from
>> one ethernet port to PPPoE on another ethernet port + cake qdisc set to
>> 1Gbps.
>>
>> rx-gro-list off: 630 Mbit/s, CPU 35% idle
>> rx-gro-list on: 770 Mbit/s, CPU 40% idle
>>
>> Signe-off-by: Felix Fietkau <[email protected]>
>> ---
>> net/ipv4/tcp_offload.c | 48 +++++++++++++++++++++++++++++++++++++
>> net/ipv6/tcpv6_offload.c | 51 ++++++++++++++++++++++++++++++++++++++++
>> 2 files changed, 99 insertions(+)
>>
>> diff --git a/net/ipv4/tcp_offload.c b/net/ipv4/tcp_offload.c
>> index 520fd425ab19..3bb96a110402 100644
>> --- a/net/ipv4/tcp_offload.c
>> +++ b/net/ipv4/tcp_offload.c
>> @@ -405,6 +405,52 @@ void tcp_gro_complete(struct sk_buff *skb)
>> }
>> EXPORT_SYMBOL(tcp_gro_complete);
>>
>> +static void tcp4_check_fraglist_gro(struct list_head *head, struct sk_buff *skb)
>> +{
>> + const struct iphdr *iph = skb_gro_network_header(skb);
>
> I do not think loading iph before all skb_gro_header() and
> skb_gro_header_slow() calls is wise.
>
> pskb_may_pull() can re-allocate skb->head
Will fix.
>> + struct net *net = dev_net(skb->dev);
>> + unsigned int off, hlen, thlen;
>> + struct sk_buff *p;
>> + struct tcphdr *th;
>> + struct sock *sk;
>> + int iif, sdif;
>> +
>> + if (!(skb->dev->features & NETIF_F_GRO_FRAGLIST))
>> + return;
>> +
>> + off = skb_gro_offset(skb);
>> + hlen = off + sizeof(*th);
>> + th = skb_gro_header(skb, hlen, off);
>> + if (unlikely(!th))
>> + return;
>> +
>> + thlen = th->doff * 4;
>> + if (thlen < sizeof(*th))
>> + return;
>> +
>> + hlen = off + thlen;
>> + if (!skb_gro_may_pull(skb, hlen)) {
>> + th = skb_gro_header_slow(skb, hlen, off);
>> + if (unlikely(!th))
>> + return;
>> + }
>> +
>> + p = tcp_gro_lookup(head, th);
>> + if (p) {
>> + NAPI_GRO_CB(skb)->is_flist = NAPI_GRO_CB(p)->is_flist;
>> + return;
>> + }
>> +
>> + inet_get_iif_sdif(skb, &iif, &sdif);
>> + sk = __inet_lookup_established(net, net->ipv4.tcp_death_row.hashinfo,
>> + iph->saddr, th->source,
>> + iph->daddr, ntohs(th->dest),
>> + iif, sdif);
>> + NAPI_GRO_CB(skb)->is_flist = !sk;
>> + if (sk)
>> + sock_put(sk);
>> +}
>> +
>> INDIRECT_CALLABLE_SCOPE
>> struct sk_buff *tcp4_gro_receive(struct list_head *head, struct sk_buff *skb)
>> {
>> @@ -416,6 +462,8 @@ struct sk_buff *tcp4_gro_receive(struct list_head *head, struct sk_buff *skb)
>> return NULL;
>> }
>>
>
> I would probably pull the whole TCP header here, before calling
> tcp4_check_fraglist_gro(head, skb)
> and no longer do this twice from tcp4_check_fraglist_gro() and tcp_gro_receive()
>
> Perhaps define a new inline helper, that will be called from
> tcp4_gro_receive() and tcp6_gro_receive(),
> and not anymore from tcp_gro_receive()
>
> static inline struct tcphdr *tcp_gro_pull_header(...)
> {
> ....
> off = skb_gro_offset(skb);
> hlen = off + sizeof(*th);
> th = skb_gro_header(skb, hlen, off);
> if (unlikely(!th))
> return NULL;
>
> thlen = th->doff * 4;
> if (thlen < sizeof(*th))
> return NULL;
>
> hlen = off + thlen;
> if (!skb_gro_may_pull(skb, hlen))
> th = skb_gro_header_slow(skb, hlen, off);
>
> return th;
> }
Makes sense
Thanks,
- Felix