MPTCP 源码分析(五) 接收端窗口值

来源：互联网发布：京东聊天软件编辑：程序博客网时间：2024/06/05 01:16

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：

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。
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 6097 of 61956097     req = inet_reqsk_alloc(rsk_ops);6098     if (!req)6099         goto drop;

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

1226 static inline void tcp_openreq_init(struct request_sock *req,1227                     struct tcp_options_received *rx_opt,1228                     struct sk_buff *skb)1229 {                   1230     struct inet_request_sock *ireq = inet_rsk(req);1231    1232     req->rcv_wnd = 0;       /* So that tcp_send_synack() knows! */1233     req->cookie_ts = 0;1234     tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;1235     tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;1236     tcp_rsk(req)->snt_synack = 0;1237     req->mss = rx_opt->mss_clamp;1238     req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;1239     ireq->tstamp_ok = rx_opt->tstamp_ok;1240     ireq->sack_ok = rx_opt->sack_ok;1241     ireq->snd_wscale = rx_opt->snd_wscale;1242     ireq->wscale_ok = rx_opt->wscale_ok;1243     ireq->acked = 0;1244     ireq->ecn_ok = 0;1245     ireq->mptcp_rqsk = 0;1246     ireq->saw_mpc = 0;1247     ireq->ir_rmt_port = tcp_hdr(skb)->source;1248     ireq->ir_num = ntohs(tcp_hdr(skb)->dest);1249 }

View Code

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 578 of 1787struct 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 441 of 872512         newtp->window_clamp = req->window_clamp;513         newtp->rcv_ssthresh = req->rcv_wnd;514         newtp->rcv_wnd = req->rcv_wnd;515         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 992 of 1667 992 u16 mptcp_select_window(struct sock *sk)993 {      994     u16 new_win     = tcp_select_window(sk);995     struct tcp_sock *tp = tcp_sk(sk);996     struct tcp_sock *meta_tp = mptcp_meta_tp(tp);997    998     meta_tp->rcv_wnd    = tp->rcv_wnd;999     meta_tp->rcv_wup    = meta_tp->rcv_nxt;1000        1001     return new_win;1002 }

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 275 of 3327275 u16 tcp_select_window(struct sock *sk)276 {277     struct tcp_sock *tp = tcp_sk(sk);278     /* The window must never shrink at the meta-level. At the subflow we279      * have to allow this. Otherwise we may announce a window too large280      * for the current meta-level sk_rcvbuf.281      */282     u32 cur_win = tcp_receive_window(mptcp(tp) ? tcp_sk(mptcp_meta_sk(sk)) : tp);283     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 771 of 1667771 u32 __mptcp_select_window(struct sock *sk)772 {773     struct inet_connection_sock *icsk = inet_csk(sk);774     struct tcp_sock *tp = tcp_sk(sk), *meta_tp = mptcp_meta_tp(tp);775     struct sock *meta_sk = mptcp_meta_sk(sk);776     int mss, free_space, full_space, window;777778     /* MSS for the peer's data.  Previous versions used mss_clamp779      * here.  I don't know if the value based on our guesses780      * of peer's MSS is better for the performance.  It's more correct781      * but may be worse for the performance because of rcv_mss782      * fluctuations.  --SAW  1998/11/1783      */784     mss = icsk->icsk_ack.rcv_mss;785     free_space = tcp_space(meta_sk);786     full_space = min_t(int, meta_tp->window_clamp,787             tcp_full_space(meta_sk));788789     if (mss > full_space)790         mss = full_space;791792     if (free_space < (full_space >> 1)) {793         icsk->icsk_ack.quick = 0;794795         if (tcp_memory_pressure)796             /* TODO this has to be adapted when we support different797              * MSS's among the subflows.798              */799             meta_tp->rcv_ssthresh = min(meta_tp->rcv_ssthresh,800                             4U * meta_tp->advmss);801802         if (free_space < mss)803             return 0;804     }805806     if (free_space > meta_tp->rcv_ssthresh)807         free_space = meta_tp->rcv_ssthresh;808809     /* Don't do rounding if we are using window scaling, since the810      * scaled window will not line up with the MSS boundary anyway.811      */812     window = meta_tp->rcv_wnd;813     if (tp->rx_opt.rcv_wscale) {814         window = free_space;815816         /* Advertise enough space so that it won't get scaled away.817          * Import case: prevent zero window announcement if818          * 1<<rcv_wscale > mss.819          */820         if (((window >> tp->rx_opt.rcv_wscale) << tp->821              rx_opt.rcv_wscale) != window)822             window = (((window >> tp->rx_opt.rcv_wscale) + 1)823                   << tp->rx_opt.rcv_wscale);824     } else {825         /* 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 the828          * free space we just keep it. This prevents the divide829          * and multiply from happening most of the time.830          * We also don't do any window rounding when the free space831          * is too small.832          */833         if (window <= free_space - mss || window > free_space)834             window = (free_space / mss) * mss;835         else if (mss == full_space &&836              free_space > window + (full_space >> 1))837             window = free_space;838     }839840     return window;841 }

View Code

影响window的计算的因素为：

收到的MSS( icsk->icsk_ack.rcv_mss)
套接字缓冲区总的空间(tcp_full_space)
套接字缓冲区的空闲空间(tcp_space)
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。

0 0