tcp tso

根據Linux的協定棧，在發送tcp資料時，最終會調用如下函數：

檔案 net/ipv4/tcp.c 

int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg, size_t size)

在此函數中，要根據mss來将大資料拆分成小封包發送到網絡上。

mss是通過如下函數擷取的：

mss_now = tcp_send_mss(sk, &size_goal, flags);

static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)

{

         int mss_now;

         mss_now = tcp_current_mss(sk);

         *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));

         return mss_now;

}

tcp_current_mss傳回TCP連結建立時協商的MSS，考慮協商的TCP選項。

tcp_xmit_size_goal考慮網卡是否支援gso，如果網卡支援gso，傳回網卡的gso max size最為mss。

static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,

                                            int large_allowed)

{

         …

if (large_allowed && sk_can_gso(sk)) {

                   xmit_size_goal = ((sk->sk_gso_max_size - 1) -

                                      inet_csk(sk)->icsk_af_ops->net_header_len -

                                      inet_csk(sk)->icsk_ext_hdr_len -

                                      tp->tcp_header_len);

         …

         return max(xmit_size_goal, mss_now);

}

socket建立時對gso的設定：

void sk_setup_caps(struct sock *sk, struct dst_entry *dst)

{

         __sk_dst_set(sk, dst);

         sk->sk_route_caps = dst->dev->features;

         if (sk->sk_route_caps & NETIF_F_GSO)

                   sk->sk_route_caps |= NETIF_F_GSO_SOFTWARE;

         sk->sk_route_caps &= ~sk->sk_route_nocaps;

         if (sk_can_gso(sk)) {

                   if (dst->header_len) {

                            sk->sk_route_caps &= ~NETIF_F_GSO_MASK;

                   } else {

                            sk->sk_route_caps |= NETIF_F_SG | NETIF_F_HW_CSUM;

                            sk->sk_gso_max_size = dst->dev->gso_max_size;

                   }

         }

}

//完整的看一下gso對tcp發包過程的影響

int tcp_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,

                   size_t size)

{

         mss_now = tcp_send_mss(sk, &size_goal, flags); //考慮gso後的mss

         sg = sk->sk_route_caps & NETIF_F_SG;

   ...

if (merge) {

          skb_shinfo(skb)->frags[i - 1].size +=      copy;

   } else {

          skb_fill_page_desc(skb, i, page, off, copy); //安排skb資料的存儲方式

          if (TCP_PAGE(sk)) {

               get_page(page);

          } else if (off + copy < PAGE_SIZE) {

                get_page(page);

                TCP_PAGE(sk) = page;

          }

    }

......

}

//填充tcp資料分片結構：

skb_fill_page_desc（）

{

         skb_frag_t *frag = &skb_shinfo(skb)->frags[i];

         frag->page                  = page;

         frag->page_offset    = off;

         frag->size                    = size;

         skb_shinfo(skb)->nr_frags = i + 1;

}