MPTCP 源码分析(五) 接收端窗口值

简述：

     在TCP协议中影响数据发送的三个因素分别为：发送端窗口值、接收端窗口值和拥塞窗口值。

本文主要分析MPTCP中各个子路径对接收端窗口值rcv_wnd的处理。

 

接收端窗口值的初始化

     根据《MPTCP 源码分析(二) 建立子路径》中描述服务端在发送完SYN/ACK并接收到ACK的时候建立新的sock。

在内核实现中，针对连接请求分为两个步骤处理：

1. SYN队列处理：当服务端收到SYN的时候，此连接请求request_sock将被存放于listening socket的SYN队列，服务端发送SYN/ACK并等待相应的ACK。
2. accept队列处理：一旦等待的ACK收到，服务端将会创建新的socket，并将连接请求从listening socket的SYN队列移到其accept队列。

当服务端进入LINSTEN状态后，收到第一个SYN包后的处理流程如下：

详细的函数调用为：

tcp_v4_rcv 

               =》 tcp_v4_do_rcv

                     =》 tcp_rcv_state_process 

                         =》mptcp_conn_request

                              =》tcp_v4_conn_request

                                   =》tcp_conn_request

                                        =》tcp_openreq_init

在函数tcp_conn_request中对连接请求request_sock进行了分配内存。

"net/ipv4/tcp_input.c" line  of
     req = inet_reqsk_alloc(rsk_ops);
     if (!req)
         goto drop;

在函数tcp_openreq_init中对request_sock进行了初始化操作。

 static inline void tcp_openreq_init(struct request_sock *req,
                     struct tcp_options_received *rx_opt,
                     struct sk_buff *skb)
 {
     struct inet_request_sock *ireq = inet_rsk(req);

     req->rcv_wnd = ;       /* So that tcp_send_synack() knows! */
     req->cookie_ts = ;
     tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
     tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + ;
     tcp_rsk(req)->snt_synack = ;
     req->mss = rx_opt->mss_clamp;
     req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : ;
     ireq->tstamp_ok = rx_opt->tstamp_ok;
     ireq->sack_ok = rx_opt->sack_ok;
     ireq->snd_wscale = rx_opt->snd_wscale;
     ireq->wscale_ok = rx_opt->wscale_ok;
     ireq->acked = ;
     ireq->ecn_ok = ;
     ireq->mptcp_rqsk = ;
     ireq->saw_mpc = ;
     ireq->ir_rmt_port = tcp_hdr(skb)->source;
     ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
 }

第1232行对request_sock的rcv_wnd进行了初始化为0。

 

当服务端收到ACK的时候就会建立相应的socket。将会调用tcp_create_openreq_child函数实现，定义如下：

"include/net/tcp.h" line  of
struct sock *tcp_create_openreq_child(struct sock *sk,
                      struct request_sock *req,
                      struct sk_buff *skb);

对于rcv_wnd的处理具体如下：

"net/ipv4/tcp_minisocks.c" line  of
         newtp->window_clamp = req->window_clamp;
         newtp->rcv_ssthresh = req->rcv_wnd;
         newtp->rcv_wnd = req->rcv_wnd;
         newtp->rx_opt.wscale_ok = ireq->wscale_ok; 

这个阶段为MPTCP的第一条子路径建立情况的三次握手，因此此时创建的socket的属性为master而非slave.

 

下面的情景为创建一条子路径的情况，当服务端收到第一个SYN包的函数调用情况如下：

函数mptcp_v4_join_request将会对连接请求request_sock进行内存分配并初始化。具体的调用如下：

mptcp_v4_join_request

                                   =》tcp_conn_request

                                        =》inet_reqsk_alloc

                                        =》tcp_openreq_init

当客户端的ACK到达后，内核会将此连接请求request_sock的rcv_wnd赋值给slave subsocket.

 

  

master sock 和 slave sock之间接收端窗口值的关系

     TCP在发包的时候会告诉对方自身的接收端窗口值。MPTCP的实现如下：

"net/mptcp/mptcp_output.c" line  of
  u16 mptcp_select_window(struct sock *sk)
 {
     u16 new_win     = tcp_select_window(sk);
     struct tcp_sock *tp = tcp_sk(sk);
     struct tcp_sock *meta_tp = mptcp_meta_tp(tp);

     meta_tp->rcv_wnd    = tp->rcv_wnd;
     meta_tp->rcv_wup    = meta_tp->rcv_nxt;

     return new_win;
 }  

第994获得最新的窗口值并返回。第998行将slave sock的rcv_wnd赋值给master sock。

 

第994行的函数tcp_select_window的实现如下：

"net/ipv4/tcp_output.c" line  of
 u16 tcp_select_window(struct sock *sk)
 {
     struct tcp_sock *tp = tcp_sk(sk);
     /* The window must never shrink at the meta-level. At the subflow we
279      * have to allow this. Otherwise we may announce a window too large
280      * for the current meta-level sk_rcvbuf.
281      */
     u32 cur_win = tcp_receive_window(mptcp(tp) ? tcp_sk(mptcp_meta_sk(sk)) : tp);
     u32 new_win = tp->__select_window(sk);

对于第283行的__select_window()函数，MPTCP的内核实现如下：

 

"net/mptcp/mptcp_output.c" line  of
 u32 __mptcp_select_window(struct sock *sk)
 {
     struct inet_connection_sock *icsk = inet_csk(sk);
     struct tcp_sock *tp = tcp_sk(sk), *meta_tp = mptcp_meta_tp(tp);
     struct sock *meta_sk = mptcp_meta_sk(sk);
     int mss, free_space, full_space, window;

     /* MSS for the peer's data.  Previous versions used mss_clamp
779      * here.  I don't know if the value based on our guesses
780      * of peer's MSS is better for the performance.  It's more correct
781      * but may be worse for the performance because of rcv_mss
782      * fluctuations.  --SAW  1998/11/1
783      */
     mss = icsk->icsk_ack.rcv_mss;
     free_space = tcp_space(meta_sk);
     full_space = min_t(int, meta_tp->window_clamp,
             tcp_full_space(meta_sk));

     if (mss > full_space)
         mss = full_space;

     if (free_space < (full_space >> )) {
         icsk->icsk_ack.quick = ;

         if (tcp_memory_pressure)
             /* TODO this has to be adapted when we support different
797              * MSS's among the subflows.
798              */
             meta_tp->rcv_ssthresh = min(meta_tp->rcv_ssthresh,
                             4U * meta_tp->advmss);

         if (free_space < mss)
             return ;
     }

     if (free_space > meta_tp->rcv_ssthresh)
         free_space = meta_tp->rcv_ssthresh;

     /* Don't do rounding if we are using window scaling, since the
810      * scaled window will not line up with the MSS boundary anyway.
811      */
     window = meta_tp->rcv_wnd;
     if (tp->rx_opt.rcv_wscale) {
         window = free_space;

         /* Advertise enough space so that it won't get scaled away.
817          * Import case: prevent zero window announcement if
818          * 1<<rcv_wscale > mss.
819          */
         if (((window >> tp->rx_opt.rcv_wscale) << tp->
              rx_opt.rcv_wscale) != window)
             window = (((window >> tp->rx_opt.rcv_wscale) + )
                   << tp->rx_opt.rcv_wscale);
     } else {
         /* Get the largest window that is a nice multiple of mss.
826          * Window clamp already applied above.
827          * If our current window offering is within 1 mss of the
828          * free space we just keep it. This prevents the divide
829          * and multiply from happening most of the time.
830          * We also don't do any window rounding when the free space
831          * is too small.
832          */
         if (window <= free_space - mss || window > free_space)
             window = (free_space / mss) * mss;
         else if (mss == full_space &&
              free_space > window + (full_space >> ))
             window = free_space;
     }

     return window;
 }

影响window的计算的因素为：

1. 收到的MSS( icsk->icsk_ack.rcv_mss)
2. 套接字缓冲区总的空间(tcp_full_space)
3. 套接字缓冲区的空闲空间(tcp_space)
4. meta_tp->rcv_ssthresh  /* Current window clamp */

观察上面的代码可以知道MPTCP的实现和__tcp_select_window的区别是都是依据meta_tp，而非tp。这说明

master sock 和 其余slave sock使用相同的 rcv_wnd。

 

结论：

1.master sock 和 其余slave sock使用相同的接收缓冲区和 rcv_wnd。