学步园

0. tcp proto

struct proto tcp_prot = {<br /> sock_wmalloc,<br /> sock_rmalloc,<br /> sock_wfree,<br /> sock_rfree,<br /> sock_rspace,<br /> sock_wspace,<br /> tcp_close,<br /> tcp_read,<br /> tcp_write,<br /> tcp_sendto,<br /> tcp_recvfrom,<br /> ip_build_header,<br /> tcp_connect,<br /> tcp_accept,<br /> ip_queue_xmit,<br /> tcp_retransmit,<br /> tcp_write_wakeup,<br /> tcp_read_wakeup,<br /> tcp_rcv,<br /> tcp_select,<br /> tcp_ioctl,<br /> NULL,<br /> tcp_shutdown,<br /> tcp_setsockopt,<br /> tcp_getsockopt,<br /> 128,<br /> 0,<br /> {NULL,},<br /> "TCP",<br /> 0, 0<br /> };

1. tcp_mib => snmp

struct tcp_mib<br /> {<br /> unsigned long TcpRtoAlgorithm;<br /> unsigned long TcpRtoMin;<br /> unsigned long TcpRtoMax;<br /> unsigned long TcpMaxConn;<br /> unsigned long TcpActiveOpens;<br /> unsigned long TcpPassiveOpens;<br /> unsigned long TcpAttemptFails;<br /> unsigned long TcpEstabResets;<br /> unsigned long TcpCurrEstab;<br /> unsigned long TcpInSegs;<br /> unsigned long TcpOutSegs;<br /> unsigned long TcpRetransSegs;<br /> };

2. tcp_select_window

/*<br /> * This routine picks a TCP windows for a socket based on<br /> * the following constraints<br /> *<br /> * 1. The window can never be shrunk once it is offered (RFC 793)<br /> * 2. We limit memory per socket<br /> *<br /> * For now we use NET2E3's heuristic of offering half the memory<br /> * we have handy. All is not as bad as this seems however because<br /> * of two things. Firstly we will bin packets even within the window<br /> * in order to get the data we are waiting for into the memory limit.<br /> * Secondly we bin common duplicate forms at receive time<br /> * Better heuristics welcome<br /> */<br /> int tcp_select_window(struct sock *sk)<br /> {<br /> int new_window = sk->prot->rspace(sk);<br /> if(sk->window_clamp)<br /> new_window=min(sk->window_clamp,new_window);<br /> /*<br /> * Two things are going on here. First, we don't ever offer a<br /> * window less than min(sk->mss, MAX_WINDOW/2). This is the<br /> * receiver side of SWS as specified in RFC1122.<br /> * Second, we always give them at least the window they<br /> * had before, in order to avoid retracting window. This<br /> * is technically allowed, but RFC1122 advises against it and<br /> * in practice it causes trouble.<br /> *<br /> * Fixme: This doesn't correctly handle the case where<br /> * new_window > sk->window but not by enough to allow for the<br /> * shift in sequence space.<br /> */<br /> if (new_window < min(sk->mss, MAX_WINDOW/2) || new_window < sk->window)<br /> return(sk->window);<br /> return(new_window);<br /> }

3. 对于重发的数据包，其IP 首部中用于标识数据包的id 字段将被赋予新值，以区别于之前发送的数据包。只要数据包中TCP 首部中本地序列号不发生改变，其表示的数据依然是原先发送的数据，不会在本地发生人为的数据包乱序问题。

/*<br /> * A socket has timed out on its send queue and wants to do a<br /> * little retransmitting. Currently this means TCP.<br /> */<br /> void tcp_do_retransmit(struct sock *sk, int all)<br /> {<br /> ...<br /> /*<br /> * In general it's OK just to use the old packet. However we<br /> * need to use the current ack and window fields. Urg and<br /> * urg_ptr could possibly stand to be updated as well, but we<br /> * don't keep the necessary data. That shouldn't be a problem,<br /> * if the other end is doing the right thing. Since we're<br /> * changing the packet, we have to issue a new IP identifier.<br /> */<br /> iph = (struct iphdr *)(skb->data + dev->hard_header_len);<br /> th = (struct tcphdr *)(((char *)iph) + (iph->ihl << 2));<br /> size = skb->len - (((unsigned char *) th) - skb->data);<br /> /*<br /> * Note: We ought to check for window limits here but<br /> * currently this is done (less efficiently) elsewhere.<br /> * We do need to check for a route change but can't handle<br /> * that until we have the new 1.3.x buffers in.<br /> *<br /> */<br /> iph->id = htons(ip_id_count++);<br /> ip_send_check(iph);<br /> ...<br /> }

4. 发送数据包

/*<br /> * This is the main buffer sending routine. We queue the buffer<br /> * having checked it is sane seeming.<br /> */<br /> static void tcp_send_skb(struct sock *sk, struct sk_buff *skb)<br /> {<br /> int size;<br /> struct tcphdr * th = skb->h.th;<br /> /*<br /> * length of packet (not counting length of pre-tcp headers)<br /> */<br /> size = skb->len - ((unsigned char *) th - skb->data);<br /> /*<br /> * Sanity check it..<br /> */<br /> if (size < sizeof(struct tcphdr) || size > skb->len)<br /> {<br /> printk("tcp_send_skb: bad skb (skb = %p, data = %p, th = %p, len = %lu)/n",<br /> skb, skb->data, th, skb->len);<br /> kfree_skb(skb, FREE_WRITE);<br /> return;<br /> }<br /> /*<br /> * If we have queued a header size packet.. (these crash a few<br /> * tcp stacks if ack is not set)<br /> */<br /> if (size == sizeof(struct tcphdr))<br /> {<br /> /* If it's got a syn or fin it's notionally included in the size..*/<br /> if(!th->syn && !th->fin)<br /> {<br /> printk("tcp_send_skb: attempt to queue a bogon./n");<br /> kfree_skb(skb,FREE_WRITE);<br /> return;<br /> }<br /> }<br /> /*<br /> * Actual processing.<br /> */<br /> tcp_statistics.TcpOutSegs++;<br /> skb->h.seq = ntohl(th->seq) + size - 4*th->doff;<br /> /*<br /> * We must queue if<br /> *<br /> * a) The right edge of this frame exceeds the window<br /> * b) We are retransmitting (Nagle's rule)<br /> * c) We have too many packets 'in flight'<br /> */<br /> if (after(skb->h.seq, sk->window_seq) ||<br /> (sk->retransmits && sk->ip_xmit_timeout == TIME_WRITE) ||<br /> sk->packets_out >= sk->cong_window)<br /> {<br /> /* checksum will be supplied by tcp_write_xmit. So<br /> * we shouldn't need to set it at all. I'm being paranoid */<br /> th->check = 0;<br /> if (skb->next != NULL)<br /> {<br /> printk("tcp_send_partial: next != NULL/n");<br /> skb_unlink(skb);<br /> }<br /> skb_queue_tail(&sk->write_queue, skb);<br /> /*<br /> * If we don't fit we have to start the zero window<br /> * probes. This is broken - we really need to do a partial<br /> * send _first_ (This is what causes the Cisco and PC/TCP<br /> * grief).<br /> */<br /> if (before(sk->window_seq, sk->write_queue.next->h.seq) &&<br /> sk->send_head == NULL && sk->ack_backlog == 0)<br /> reset_xmit_timer(sk, TIME_PROBE0, sk->rto);<br /> }<br /> else<br /> {<br /> /*<br /> * This is going straight out<br /> */<br /> th->ack_seq = ntohl(sk->acked_seq);<br /> th->window = ntohs(tcp_select_window(sk));<br /> tcp_send_check(th, sk->saddr, sk->daddr, size, sk);<br /> sk->sent_seq = sk->write_seq;<br /> /*<br /> * This is mad. The tcp retransmit queue is put together<br /> * by the ip layer. This causes half the problems with<br /> * unroutable FIN's and other things.<br /> */<br /> sk->prot->queue_xmit(sk, skb->dev, skb, 0);<br /> /*<br /> * Set for next retransmit based on expected ACK time.<br /> * FIXME: We set this every time which means our<br /> * retransmits are really about a window behind.<br /> */<br /> reset_xmit_timer(sk, TIME_WRITE, sk->rto);<br /> }<br /> }

发送用户数据

/*<br /> * This routine copies from a user buffer into a socket,<br /> * and starts the transmit system.<br /> */<br /> static int tcp_write(struct sock *sk, unsigned char *from,<br /> int len, int nonblock, unsigned flags);

发送过程中的两个队列：
1〉写队列，对应sock 结构中write_queue 字段指向的队列，这是一个双向队列。该队列接收应用层发送的数据包（传输层将数据封装为数据包，将其挂接到write_queue 队列中）。该队列中数据包尚未发送出去。
2〉重发队列，对应 sock 结构中send_head, send_tail 字段指向的队列，这是一个单向队列，send_head 指向队列头部，send_tail 指向队列尾部。传输层（实际上是网络层）将数据包发送出去以后，将数据包缓存到该队列中，以防止发送的数据包可能丢失后的重发工作。

5. 接收数据

/*<br /> * This routine copies from a sock struct into the user buffer.<br /> */<br /> static int tcp_read(struct sock *sk, unsigned char *to,<br /> int len, int nonblock, unsigned flags);

6. 半关闭

/*<br /> * Shutdown the sending side of a connection. Much like close except<br /> * that we don't receive shut down or set sk->dead=1.<br /> */<br /> void tcp_shutdown(struct sock *sk, int how);

7. 处理一个新的连接请求

/*<br /> * This routine handles a connection request.<br /> * It should make sure we haven't already responded.<br /> * Because of the way BSD works, we have to send a syn/ack now.<br /> * This also means it will be harder to close a socket which is<br /> * listening.<br /> */<br /> static void tcp_conn_request(struct sock *sk, struct sk_buff *skb,<br /> unsigned long daddr, unsigned long saddr,<br /> struct options *opt, struct device *dev, unsigned long seq)