Linux内核--网络栈实现分析（七）--数据包的传递过程（下）

本文分析基于Linux Kernel 1.2.13

原创作品，转载请标明http://blog.csdn.net/yming0221/article/details/7545855

更多请查看专栏，地址http://blog.csdn.net/column/details/linux-kernel-net.html

作者：闫明

注：标题中的”（上）“，”（下）“表示分析过程基于数据包的传递方向：”（上）“表示分析是从底层向上分析、”（下）“表示分析是从上向下分析。

在博文Linux内核--网络栈实现分析（二）--数据包的传递过程（上）中分析了数据包从网卡设备经过驱动链路层，网络层，传输层到应用层的过程。

本文就分析一下本机产生数据是如何通过传输层，网络层到达物理层的。

综述来说，数据流程图如下：

一、应用层

应用层可以通过系统调用或文件操作来调用内核函数，BSD层的sock_write()函数会调用INET层的inet_wirte()函数。

/*
*	Write data to a socket. We verify that the user area ubuf..ubuf+size-1 is
*	readable by the user process.
*/

static int sock_write(struct inode *inode, struct file *file, char *ubuf, int size)
{
struct socket *sock;
int err;

if (!(sock = socki_lookup(inode)))
{
printk("NET: sock_write: can't find socket for inode!\n");
return(-EBADF);
}

if (sock->flags & SO_ACCEPTCON)
return(-EINVAL);

if(size<0)
return -EINVAL;
if(size==0)
return 0;

if ((err=verify_area(VERIFY_READ,ubuf,size))<0)
return err;
return(sock->ops->write(sock, ubuf, size,(file->f_flags & O_NONBLOCK)));
}

INET层会调用具体传输层协议的write函数，该函数是通过调用本层的inet_send()函数实现功能的，inet_send()函数的UDP协议对应的函数为udp_write()

static int inet_send(struct socket *sock, void *ubuf, int size, int noblock,
unsigned flags)
{
struct sock *sk = (struct sock *) sock->data;
if (sk->shutdown & SEND_SHUTDOWN)
{
send_sig(SIGPIPE, current, 1);
return(-EPIPE);
}
if(sk->err)
return inet_error(sk);
/* We may need to bind the socket. */
if(inet_autobind(sk)!=0)
return(-EAGAIN);
return(sk->prot->write(sk, (unsigned char *) ubuf, size, noblock, flags));
}

static int inet_write(struct socket *sock, char *ubuf, int size, int noblock)
{
return inet_send(sock,ubuf,size,noblock,0);
}

二、传输层

在传输层udp_write()函数调用本层的udp_sendto()函数完成功能。

/*
*	In BSD SOCK_DGRAM a write is just like a send.
*/

static int udp_write(struct sock *sk, unsigned char *buff, int len, int noblock,
unsigned flags)
{
return(udp_sendto(sk, buff, len, noblock, flags, NULL, 0));
}

udp_send()函数完成sk_buff结构相应的设置和报头的填写后会调用udp_send()来发送数据。具体的实现过程后面会详细分析。

而在udp_send()函数中，最后会调用ip_queue_xmit()函数，将数据包下放的网络层。

下面是udp_prot定义：

struct proto udp_prot = {
sock_wmalloc,
sock_rmalloc,
sock_wfree,
sock_rfree,
sock_rspace,
sock_wspace,
udp_close,
udp_read,
udp_write,
udp_sendto,
udp_recvfrom,
ip_build_header,
udp_connect,
NULL,
ip_queue_xmit,
NULL,
NULL,
NULL,
udp_rcv,
datagram_select,
udp_ioctl,
NULL,
NULL,
ip_setsockopt,
ip_getsockopt,
128,
0,
{NULL,},
"UDP",
0, 0
};

static int udp_send(struct sock *sk, struct sockaddr_in *sin,
unsigned char *from, int len, int rt)
{
struct sk_buff *skb;
struct device *dev;
struct udphdr *uh;
unsigned char *buff;
unsigned long saddr;
int size, tmp;
int ttl;

/*
*	Allocate an sk_buff copy of the packet.
*/

........................

/*
*	Now build the IP and MAC header.
*/

..........................
/*
*	Fill in the UDP header.
*/

..............................

/*
*	Copy the user data.
*/

memcpy_fromfs(buff, from, len);

/*
*	Set up the UDP checksum.
*/

udp_send_check(uh, saddr, sin->sin_addr.s_addr, skb->len - tmp, sk);

/*
*	Send the datagram to the interface.
*/

udp_statistics.UdpOutDatagrams++;

sk->prot->queue_xmit(sk, dev, skb, 1);
return(len);
}

三、网络层

在网络层，函数ip_queue_xmit()的功能是将数据包进行一系列复杂的操作，比如是检查数据包是否需要分片，是否是多播等一系列检查，最后调用dev_queue_xmit()函数发送数据。

/*
* Queues a packet to be sent, and starts the transmitter
* if necessary.  if free = 1 then we free the block after
* transmit, otherwise we don't. If free==2 we not only
* free the block but also don't assign a new ip seq number.
* This routine also needs to put in the total length,
* and compute the checksum
*/

void ip_queue_xmit(struct sock *sk, struct device *dev,
struct sk_buff *skb, int free)
{
struct iphdr *iph;
unsigned char *ptr;

/* Sanity check */
............
/*
*	Do some book-keeping in the packet for later
*/

...........

/*
*	Find the IP header and set the length. This is bad
*	but once we get the skb data handling code in the
*	hardware will push its header sensibly and we will
*	set skb->ip_hdr to avoid this mess and the fixed
*	header length problem
*/

..............
/*
*	No reassigning numbers to fragments...
*/

if(free!=2)
iph->id      = htons(ip_id_count++);
else
free=1;

/* All buffers without an owner socket get freed */
if (sk == NULL)
free = 1;

skb->free = free;

/*
*	Do we need to fragment. Again this is inefficient.
*	We need to somehow lock the original buffer and use
*	bits of it.
*/

................

/*
*	Add an IP checksum
*/

ip_send_check(iph);

/*
*	Print the frame when debugging
*/

/*
*	More debugging. You cannot queue a packet already on a list
*	Spot this and moan loudly.
*/
.......................

/*
*	If a sender wishes the packet to remain unfreed
*	we add it to his send queue. This arguably belongs
*	in the TCP level since nobody else uses it. BUT
*	remember IPng might change all the rules.
*/

......................
/*
*	If the indicated interface is up and running, send the packet.
*/

ip_statistics.IpOutRequests++;
.............................
.............................
if((dev->flags&IFF_BROADCAST) && iph->daddr==dev->pa_brdaddr && !(dev->flags&IFF_LOOPBACK))
ip_loopback(dev,skb);

if (dev->flags & IFF_UP)
{
/*
*	If we have an owner use its priority setting,
*	otherwise use NORMAL
*/

if (sk != NULL)
{
dev_queue_xmit(skb, dev, sk->priority);
}
else
{
dev_queue_xmit(skb, dev, SOPRI_NORMAL);
}
}
else
{
ip_statistics.IpOutDiscards++;
if (free)
kfree_skb(skb, FREE_WRITE);
}
}

四、驱动层（链路层）

在函数中，函数调用会调用具体设备的发送函数来发送数据包

dev->hard_start_xmit(skb, dev);

具体设备的发送函数在网络初始化的时候已经设置了。

这里以8390网卡为例来说明驱动层的工作原理，在net/drivers/8390.c中函数ethdev_init()函数中设置如下：

/* Initialize the rest of the 8390 device structure. */
int ethdev_init(struct device *dev)
{
if (ei_debug > 1)
printk(version);

if (dev->priv == NULL) {//申请私有空间
struct ei_device *ei_local;//8390网卡设备的结构体

dev->priv = kmalloc(sizeof(struct ei_device), GFP_KERNEL);//申请内核内存空间
memset(dev->priv, 0, sizeof(struct ei_device));
ei_local = (struct ei_device *)dev->priv;
#ifndef NO_PINGPONG
ei_local->pingpong = 1;
#endif
}

/* The open call may be overridden by the card-specific code. */
if (dev->open == NULL)
dev->open = &ei_open;//设备的打开函数
/* We should have a dev->stop entry also. */
dev->hard_start_xmit = &ei_start_xmit;//设备的发送函数，定义在8390.c中
dev->get_stats	= get_stats;
#ifdef HAVE_MULTICAST
dev->set_multicast_list = &set_multicast_list;
#endif

ether_setup(dev);

return 0;
}

驱动中的发送函数比较复杂，和硬件关系紧密，这里不再详细分析。

这样就大体分析了下网络数据从应用层到物理层的数据通路，后面会详细分析。