git.openwrt.org Git - openwrt/staging/xback.git/blob

   1 From c9d1d23e5239f41700be69133a5769ac5ebc88a8 Mon Sep 17 00:00:00 2001
   2 From: Felix Fietkau <nbd@nbd.name>
   3 Date: Thu, 2 May 2024 10:44:47 +0200
   4 Subject: [PATCH 6/6] net: add heuristic for enabling TCP fraglist GRO
   5
   6 When forwarding TCP after GRO, software segmentation is very expensive,
   7 especially when the checksum needs to be recalculated.
   8 One case where that's currently unavoidable is when routing packets over
   9 PPPoE. Performance improves significantly when using fraglist GRO
  10 implemented in the same way as for UDP.
  11
  12 When NETIF_F_GRO_FRAGLIST is enabled, perform a lookup for an established
  13 socket in the same netns as the receiving device. While this may not
  14 cover all relevant use cases in multi-netns configurations, it should be
  15 good enough for most configurations that need this.
  16
  17 Here's a measurement of running 2 TCP streams through a MediaTek MT7622
  18 device (2-core Cortex-A53), which runs NAT with flow offload enabled from
  19 one ethernet port to PPPoE on another ethernet port + cake qdisc set to
  20 1Gbps.
  21
  22 rx-gro-list off: 630 Mbit/s, CPU 35% idle
  23 rx-gro-list on:  770 Mbit/s, CPU 40% idle
  24
  25 Acked-by: Paolo Abeni <pabeni@redhat.com>
  26 Reviewed-by: Eric Dumazet <edumazet@google.com>
  27 Signed-off-by: Felix Fietkau <nbd@nbd.name>
  28 Reviewed-by: David Ahern <dsahern@kernel.org>
  29 Reviewed-by: Willem de Bruijn <willemb@google.com>
  30 Signed-off-by: Paolo Abeni <pabeni@redhat.com>
  31 ---
  32  net/ipv4/tcp_offload.c   | 32 ++++++++++++++++++++++++++++++++
  33  net/ipv6/tcpv6_offload.c | 35 +++++++++++++++++++++++++++++++++++
  34  2 files changed, 67 insertions(+)
  35
  36 --- a/net/ipv4/tcp_offload.c
  37 +++ b/net/ipv4/tcp_offload.c
  38 @@ -413,6 +413,36 @@ void tcp_gro_complete(struct sk_buff *sk
  39  }
  40  EXPORT_SYMBOL(tcp_gro_complete);
  41
  42 +static void tcp4_check_fraglist_gro(struct list_head *head, struct sk_buff *skb,
  43 +                                   struct tcphdr *th)
  44 +{
  45 +       const struct iphdr *iph;
  46 +       struct sk_buff *p;
  47 +       struct sock *sk;
  48 +       struct net *net;
  49 +       int iif, sdif;
  50 +
  51 +       if (likely(!(skb->dev->features & NETIF_F_GRO_FRAGLIST)))
  52 +               return;
  53 +
  54 +       p = tcp_gro_lookup(head, th);
  55 +       if (p) {
  56 +               NAPI_GRO_CB(skb)->is_flist = NAPI_GRO_CB(p)->is_flist;
  57 +               return;
  58 +       }
  59 +
  60 +       inet_get_iif_sdif(skb, &iif, &sdif);
  61 +       iph = skb_gro_network_header(skb);
  62 +       net = dev_net(skb->dev);
  63 +       sk = __inet_lookup_established(net, net->ipv4.tcp_death_row.hashinfo,
  64 +                                      iph->saddr, th->source,
  65 +                                      iph->daddr, ntohs(th->dest),
  66 +                                      iif, sdif);
  67 +       NAPI_GRO_CB(skb)->is_flist = !sk;
  68 +       if (sk)
  69 +               sock_put(sk);
  70 +}
  71 +
  72  INDIRECT_CALLABLE_SCOPE
  73  struct sk_buff *tcp4_gro_receive(struct list_head *head, struct sk_buff *skb)
  74  {
  75 @@ -428,6 +458,8 @@ struct sk_buff *tcp4_gro_receive(struct
  76         if (!th)
  77                 goto flush;
  78
  79 +       tcp4_check_fraglist_gro(head, skb, th);
  80 +
  81         return tcp_gro_receive(head, skb, th);
  82
  83  flush:
  84 --- a/net/ipv6/tcpv6_offload.c
  85 +++ b/net/ipv6/tcpv6_offload.c
  86 @@ -7,12 +7,45 @@
  87   */
  88  #include <linux/indirect_call_wrapper.h>
  89  #include <linux/skbuff.h>
  90 +#include <net/inet6_hashtables.h>
  91  #include <net/gro.h>
  92  #include <net/protocol.h>
  93  #include <net/tcp.h>
  94  #include <net/ip6_checksum.h>
  95  #include "ip6_offload.h"
  96
  97 +static void tcp6_check_fraglist_gro(struct list_head *head, struct sk_buff *skb,
  98 +                                   struct tcphdr *th)
  99 +{
 100 +#if IS_ENABLED(CONFIG_IPV6)
 101 +       const struct ipv6hdr *hdr;
 102 +       struct sk_buff *p;
 103 +       struct sock *sk;
 104 +       struct net *net;
 105 +       int iif, sdif;
 106 +
 107 +       if (likely(!(skb->dev->features & NETIF_F_GRO_FRAGLIST)))
 108 +               return;
 109 +
 110 +       p = tcp_gro_lookup(head, th);
 111 +       if (p) {
 112 +               NAPI_GRO_CB(skb)->is_flist = NAPI_GRO_CB(p)->is_flist;
 113 +               return;
 114 +       }
 115 +
 116 +       inet6_get_iif_sdif(skb, &iif, &sdif);
 117 +       hdr = skb_gro_network_header(skb);
 118 +       net = dev_net(skb->dev);
 119 +       sk = __inet6_lookup_established(net, net->ipv4.tcp_death_row.hashinfo,
 120 +                                       &hdr->saddr, th->source,
 121 +                                       &hdr->daddr, ntohs(th->dest),
 122 +                                       iif, sdif);
 123 +       NAPI_GRO_CB(skb)->is_flist = !sk;
 124 +       if (sk)
 125 +               sock_put(sk);
 126 +#endif /* IS_ENABLED(CONFIG_IPV6) */
 127 +}
 128 +
 129  INDIRECT_CALLABLE_SCOPE
 130  struct sk_buff *tcp6_gro_receive(struct list_head *head, struct sk_buff *skb)
 131  {
 132 @@ -28,6 +61,8 @@ struct sk_buff *tcp6_gro_receive(struct
 133         if (!th)
 134                 goto flush;
 135
 136 +       tcp6_check_fraglist_gro(head, skb, th);
 137 +
 138         return tcp_gro_receive(head, skb, th);
 139
 140  flush: