Openvswitch原理与代码分析(5): 内核中的流表flow table操作

 

 

当一个数据包到达网卡的时候，首先要经过内核Openvswitch.ko，流表Flow Table在内核中有一份，通过key查找内核中的flow table，即可以得到action，然后执行action之后，直接发送这个包，只有在内核无法查找到流表项的时候，才会到用户态查找用户态的流表。仅仅查找内核中flow table的情况被称为fast path.

第一步：从数据包中提取出key

实现函数为int ovs_flow_key_extract(const struct ip_tunnel_info *tun_info, struct sk_buff *skb, struct sw_flow_key *key)

在这个函数中，首先提取的是物理层的信息，主要是从哪个网口进入的。

1. key->phy.priority = skb->priority;
2. key->phy.in_port = OVS_CB(skb)->input_vport->port_no;
3. key->phy.skb_mark = skb->mark;
4. ovs_ct_fill_key(skb, key);
5. key->ovs_flow_hash = 0;
6. key->recirc_id = 0;

然后调用函数static int key_extract(struct sk_buff *skb, struct sw_flow_key *key)提取其他的key

提取MAC层

1. /* Link layer. We are guaranteed to have at least the 14 byte Ethernet
2.  * header in the linear data area.
3.  */
4. eth = eth_hdr(skb);
5. ether_addr_copy(key->eth.src, eth->h_source);
6. ether_addr_copy(key->eth.dst, eth->h_dest);
7. __skb_pull(skb, 2 * ETH_ALEN);
8. /* We are going to push all headers that we pull, so no need to
9.  * update skb->csum here.
10.  */
11. key->eth.tci = 0;
12. if (skb_vlan_tag_present(skb))
13.    key->eth.tci = htons(vlan_get_tci(skb));
14. else if (eth->h_proto == htons(ETH_P_8021Q))
15.    if (unlikely(parse_vlan(skb, key)))
16.       return -ENOMEM;
17. key->eth.type = parse_ethertype(skb);

提取网络层

1. struct iphdr *nh;
2. __be16 offset;
3. error = check_iphdr(skb);
4. if (unlikely(error)) {
5.    memset(&key->ip, 0, sizeof(key->ip));
6.    memset(&key->ipv4, 0, sizeof(key->ipv4));
7.    if (error == -EINVAL) {
8.       skb->transport_header = skb->network_header;
9.       error = 0;
10.    }
11.    return error;
12. }
13. nh = ip_hdr(skb);
14. key->ipv4.addr.src = nh->saddr;
15. key->ipv4.addr.dst = nh->daddr;
16. key->ip.proto = nh->protocol;
17. key->ip.tos = nh->tos;
18. key->ip.ttl = nh->ttl;
19. offset = nh->frag_off & htons(IP_OFFSET);
20. if (offset) {
21.    key->ip.frag = OVS_FRAG_TYPE_LATER;
22.    return 0;
23. }
24. if (nh->frag_off & htons(IP_MF) ||
25.    skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
26.    key->ip.frag = OVS_FRAG_TYPE_FIRST;
27. else
28.    key->ip.frag = OVS_FRAG_TYPE_NONE;

提取传输层

1. /* Transport layer. */
2. if (key->ip.proto == IPPROTO_TCP) {
3.    if (tcphdr_ok(skb)) {
4.       struct tcphdr *tcp = tcp_hdr(skb);
5.       key->tp.src = tcp->source;
6.       key->tp.dst = tcp->dest;
7.       key->tp.flags = TCP_FLAGS_BE16(tcp);
8.    } else {
9.       memset(&key->tp, 0, sizeof(key->tp));
10.    }
11. } else if (key->ip.proto == IPPROTO_UDP) {
12.    if (udphdr_ok(skb)) {
13.       struct udphdr *udp = udp_hdr(skb);
14.       key->tp.src = udp->source;
15.       key->tp.dst = udp->dest;
16.    } else {
17.       memset(&key->tp, 0, sizeof(key->tp));
18.    }
19. } else if (key->ip.proto == IPPROTO_SCTP) {
20.    if (sctphdr_ok(skb)) {
21.       struct sctphdr *sctp = sctp_hdr(skb);
22.       key->tp.src = sctp->source;
23.       key->tp.dst = sctp->dest;
24.    } else {
25.       memset(&key->tp, 0, sizeof(key->tp));
26.    }
27. } else if (key->ip.proto == IPPROTO_ICMP) {
28.    if (icmphdr_ok(skb)) {
29.       struct icmphdr *icmp = icmp_hdr(skb);
30.       /* The ICMP type and code fields use the 16-bit
31.        * transport port fields, so we need to store
32.        * them in 16-bit network byte order.
33.        */
34.       key->tp.src = htons(icmp->type);
35.       key->tp.dst = htons(icmp->code);
36.    } else {
37.       memset(&key->tp, 0, sizeof(key->tp));
38.    }
39. }

第二步：根据key查找flow table

调用struct sw_flow *ovs_flow_tbl_lookup_stats(struct flow_table *tbl, const struct sw_flow_key *key, u32 skb_hash, u32 *n_mask_hit)进行查找。

在内核中，flow table的数据结构如上图所示。

每个虚拟交换机对应一个datapath，每个datapath有一个flow table，每个flow table分成N个桶，根据key进行哈希，不同的key分布在不同的桶里面。

每个桶的大小是一个内存页的大小，在内存页的头部保存了保存了多少个元素，每个元素的大小。每个元素都是sw_flow，里面有key，也有action。

ovs_flow_tbl_lookup_stats会调用static struct sw_flow *flow_lookup(struct flow_table *tbl, struct table_instance *ti, const struct mask_array *ma, const struct sw_flow_key *key, u32 *n_mask_hit, u32 *index)

会调用masked_flow_lookup如下

1. static struct sw_flow *masked_flow_lookup(struct table_instance *ti,
2.                  const struct sw_flow_key *unmasked,
3.                  const struct sw_flow_mask *mask,
4.                  u32 *n_mask_hit)
5. {
6.    struct sw_flow *flow;
7.    struct hlist_head *head;
8.    u32 hash;
9.    struct sw_flow_key masked_key;
10.  
11.    ovs_flow_mask_key(&masked_key, unmasked, false, mask);
12.    hash = flow_hash(&masked_key, &mask->range);
13.    head = find_bucket(ti, hash);
14.    (*n_mask_hit)++;
15.    hlist_for_each_entry_rcu(flow, head, flow_table.node[ti->node_ver]) {
16.       if (flow->mask == mask && flow->flow_table.hash == hash &&
17.           flow_cmp_masked_key(flow, &masked_key, &mask->range))
18.          return flow;
19.    }
20.    return NULL;
21. }

其中flow_hash计算哈希值，find_bucket根据哈希值查找桶，然后就是一个循环，逐个比较key是否相等，相等则返回flow。

第三步：执行action

调用int ovs_execute_actions(struct datapath *dp, struct sk_buff *skb, const struct sw_flow_actions *acts,struct sw_flow_key *key)

调用static int do_execute_actions(struct datapath *dp, struct sk_buff *skb, struct sw_flow_key *key, const struct nlattr *attr, int len)

在这个函数中，通过case语句，不同的action进行不同的操作。

1. static int do_execute_actions(struct datapath *dp, struct sk_buff *skb,
2.                struct sw_flow_key *key,
3.                const struct nlattr *attr, int len)
4. {
5.    /* Every output action needs a separate clone of 'skb', but the common
6.     * case is just a single output action, so that doing a clone and
7.     * then freeing the original skbuff is wasteful. So the following code
8.     * is slightly obscure just to avoid that.
9.     */
10.    int prev_port = -1;
11.    const struct nlattr *a;
12.    int rem;
13.  
14.    for (a = attr, rem = len; rem > 0;
15.         a = nla_next(a, &rem)) {
16.       int err = 0;
17.  
18.       if (unlikely(prev_port != -1)) {
19.          struct sk_buff *out_skb = skb_clone(skb, GFP_ATOMIC);
20.  
21.          if (out_skb)
22.             do_output(dp, out_skb, prev_port, key);
23.  
24.          prev_port = -1;
25.       }
26.  
27.       switch (nla_type(a)) {
28.       case OVS_ACTION_ATTR_OUTPUT:
29.          prev_port = nla_get_u32(a);
30.          break;
31.  
32.       case OVS_ACTION_ATTR_USERSPACE:
33.          output_userspace(dp, skb, key, a, attr, len);
34.          break;
35.  
36.       case OVS_ACTION_ATTR_HASH:
37.          execute_hash(skb, key, a);
38.          break;
39.  
40.       case OVS_ACTION_ATTR_PUSH_MPLS:
41.          err = push_mpls(skb, key, nla_data(a));
42.          break;
43.  
44.       case OVS_ACTION_ATTR_POP_MPLS:
45.          err = pop_mpls(skb, key, nla_get_be16(a));
46.          break;
47.  
48.       case OVS_ACTION_ATTR_PUSH_VLAN:
49.          err = push_vlan(skb, key, nla_data(a));
50.          break;
51.  
52.       case OVS_ACTION_ATTR_POP_VLAN:
53.          err = pop_vlan(skb, key);
54.          break;
55.  
56.       case OVS_ACTION_ATTR_RECIRC:
57.          err = execute_recirc(dp, skb, key, a, rem);
58.          if (nla_is_last(a, rem)) {
59.             /* If this is the last action, the skb has
60.              * been consumed or freed.
61.              * Return immediately.
62.              */
63.             return err;
64.          }
65.          break;
66.  
67.       case OVS_ACTION_ATTR_SET:
68.          err = execute_set_action(skb, key, nla_data(a));
69.          break;
70.  
71.       case OVS_ACTION_ATTR_SET_MASKED:
72.       case OVS_ACTION_ATTR_SET_TO_MASKED:
73.          err = execute_masked_set_action(skb, key, nla_data(a));
74.          break;
75.  
76.       case OVS_ACTION_ATTR_SAMPLE:
77.          err = sample(dp, skb, key, a, attr, len);
78.          break;
79.  
80.       case OVS_ACTION_ATTR_CT:
81.          if (!is_flow_key_valid(key)) {
82.             err = ovs_flow_key_update(skb, key);
83.             if (err)
84.                return err;
85.          }
86.  
87.          err = ovs_ct_execute(ovs_dp_get_net(dp), skb, key,
88.                     nla_data(a));
89.  
90.          /* Hide stolen IP fragments from user space. */
91.          if (err)
92.             return err == -EINPROGRESS ? 0 : err;
93.          break;
94.       }
95.  
96.       if (unlikely(err)) {
97.          kfree_skb(skb);
98.          return err;
99.       }
100.    }
101.  
102.    if (prev_port != -1)
103.       do_output(dp, skb, prev_port, key);
104.    else
105.       consume_skb(skb);
106.  
107.    return 0;
108. }

如果可以直接输出，则调用static void do_output(struct datapath *dp, struct sk_buff *skb, int out_port, struct sw_flow_key *key)他调用void ovs_vport_send(struct vport *vport, struct sk_buff *skb)进行发送。