netdev

qdisc study part1: qdisc_base

Qdisc_ops is the core of a Qdisc.

All kinds of the Qdisc_ops are linked in a list by qdisc_base.
The key item of different Qdisc_ops is id[IFNAMSIZ].

Note: the list is a Singly-linked list, not a common list of kernel.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
158 struct Qdisc_ops {
159         struct Qdisc_ops        *next;
160         const struct Qdisc_class_ops    *cl_ops;
161         char                    id[IFNAMSIZ];
162         int                     priv_size;
163
164         int                     (*enqueue)(struct sk_buff *, struct Qdisc *);
165         struct sk_buff *        (*dequeue)(struct Qdisc *);
166         struct sk_buff *        (*peek)(struct Qdisc *);
167         unsigned int            (*drop)(struct Qdisc *);
168
169         int                     (*init)(struct Qdisc *, struct nlattr *arg);
170         void                    (*reset)(struct Qdisc *);
171         void                    (*destroy)(struct Qdisc *);
172         int                     (*change)(struct Qdisc *, struct nlattr *arg);
173         void                    (*attach)(struct Qdisc *);
174
175         int                     (*dump)(struct Qdisc *, struct sk_buff *);
176         int                     (*dump_stats)(struct Qdisc *, struct gnet_dump *);
177
178         struct module           *owner;
179 };

qdisc_base

1
2
3
134 /* The list of all installed queueing disciplines. */
135
136 static struct Qdisc_ops *qdisc_base;

Read More

how to remember ip/tcp/udp header

IP HEADER

以前找工作时，常被问到IP头都有哪些字段？
现在觉得真的理解了记起来没那么难。

上路时总要记得终点和起点(src/dst ip),
要倒几次车，大体也知道（ttl).

路有大路有小路，有高速路和土路。
大路到小路要分片，小路到大路可能重组。
上高速路需要通行证（qos/tos).

路上有警察，查你是不是非法(csum, header len, ip option)，
查你装的什么货(protocol)？

Read More

how does tcp server accept a new connection request

summary

Two packets will be proessed by tcp server side:

  1. SYN pakcet: For the first packet(syn) of handshake, it first lookup the listen socket,
    and create a req socket as temporary. Send SYN+ACK packet to client.
  2. ACK packet: For the third packet(ack) of handshake, it will lookup the req socket created
    in previous steps.

SYN 报文calltrace

内核版本 v6.14

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
=> tcp_v4_rcv(struct sk_buff *skb)
=> => sk = __inet_lookup_skb(&tcp_hashinfo, skb, th->source, th->dest);
=> => tcp_v4_do_rcv(sk, skb);
=> => => struct sock *nsk = tcp_v4_hnd_req(sk, skb)
=> => => tcp_rcv_state_process
=> => => => if (icsk->icsk_af_ops->conn_request(sk, skb) < 0) 相当 tcp_v4_conn_request
=> => => => tcp_v4_conn_request(&tcp_request_sock_ops, &tcp_request_sock_ipv4_ops, sk, skb);
=> => => => => inet_reqsk_alloc
=> => => => => => struct request_sock *req = reqsk_alloc(ops, sk_listener, attach_listener);
=> => => => => => ireq->ireq_state = TCP_NEW_SYN_RECV; //#define ireq_state req.__req_common.skc_state
=> => => => => tcp_openreq_init(req, &tmp_opt, skb);
=> => => => => inet_csk_reqsk_queue_hash_add
=> => => => => => reqsk_queue_hash_req
=> => => => => => => inet_ehash_insert  //将 req sock放到 establish hash链表里
=> => => => => => inet_csk_reqsk_queue_added
=> => => => => => => reqsk_queue_added(&inet_csk(sk)->icsk_accept_queue); //半链接队列，真实名字叫 accept。 功能也只有计数能力了，没有队列了。
=> => => => => => => => atomic_inc(&queue->young);
=> => => => => => => => atomic_inc(&queue->qlen);
=> => => => => af_ops->send_synack  //发送synack报文
=> => => => return // tcp_v4_conn_request
=> => => return; // tcp_rcv_state_process
注意
  • tcp_request_sock_ops： 是一个结构体的名字，同时又是一个变量的名字。
  • icsk_accept_queue: 在结构体struct inet_connection_sock里的这个队列并不在保存半链接队列的 req socket，而是计数。

Read More

How does IPV4/6 process input tcp/udp packet

For the packet to localhost, ip_local_deliver_finish will be the
last funciton called by network layer.

In ip_local_deliver_finish, it will be process the protocol
hander of a array element of inet_protos according the protocol value
in IPv4 header.

IPv6 is very similar vs IPv4 except the name is a bit different.

Read More

inetsw table

Data struct.

1
2
3
4
5
123 /* The inetsw table contains everything that inet_create needs to
124  * build a new socket.
125  */
126 static struct list_head inetsw[SOCK_MAX];
127 static DEFINE_SPINLOCK(inetsw_lock);

Read More

how to create a inet socket

Data Structure

1
2
3
4
5
1019 static const struct net_proto_family inet_family_ops = {
1020         .family = PF_INET,
1021         .create = inet_create,
1022         .owner  = THIS_MODULE,
1023 };

call trace

1
2
3
4
5
6
> inet_create
> > search the right inetsw array element with type/protocol
> > sock->ops = answer->ops;
> > k = sk_alloc(net, PF_INET, GFP_KERNEL, answer_prot);
> > sock_init_data(sock, sk);
> > init the sk

Read More

system call socket

Summary

System call socket will do two things:

  1. create a struct socket *sock.
    Mainly done by __sock_create, which alloc a struct socket *sock,
    then init it with the creating method of net_families[family].
  2. map the sock to a file descriptor by sock_map_fd.
    TODO…..

call trace:

1
2
3
4
5
6
7
8
9
10
11
12
> socket
> > sock_create
> > > __sock_create
> > > > sock = sock_alloc();
> > > > sock->type = type;
> > > > rcu_read_lock();
> > > > pf = rcu_dereference(net_families[family]);
> > > > try_module_get(pf->owner))
> > > > rcu_read_unlock();
> > > > pf->create(net, sock, protocol, kern);
> > > > module_put(pf->owner)
> > socket_map_fd

Read More

socket net_proto_family

summary

Each family has a corresponding array element of struct net_proto_family,
which will be called in system call socket.

Data Structure

1
2
3
4
5
6
181 struct net_proto_family {
182         int             family;
183         int             (*create)(struct net *net, struct socket *sock,
184                                   int protocol, int kern);
185         struct module   *owner;
186 };

The create is important, which is first and basic function during
system call socket.

1
2
164 static DEFINE_SPINLOCK(net_family_lock);
165 static const struct net_proto_family __rcu *net_families[NPROTO] __read_mostly;

Read More