xref: /linux/samples/bpf/xdp_router_ipv4.bpf.c (revision 320fefa9e2edc67011e235ea1d50f0d00ddfe004)
1 /* Copyright (C) 2017 Cavium, Inc.
2  *
3  * This program is free software; you can redistribute it and/or modify it
4  * under the terms of version 2 of the GNU General Public License
5  * as published by the Free Software Foundation.
6  */
7 
8 #include "vmlinux.h"
9 #include "xdp_sample.bpf.h"
10 #include "xdp_sample_shared.h"
11 
12 #define ETH_ALEN	6
13 #define ETH_P_8021Q	0x8100
14 #define ETH_P_8021AD	0x88A8
15 
16 struct trie_value {
17 	__u8 prefix[4];
18 	__be64 value;
19 	int ifindex;
20 	int metric;
21 	__be32 gw;
22 };
23 
24 /* Key for lpm_trie */
25 union key_4 {
26 	u32 b32[2];
27 	u8 b8[8];
28 };
29 
30 struct arp_entry {
31 	__be64 mac;
32 	__be32 dst;
33 };
34 
35 struct direct_map {
36 	struct arp_entry arp;
37 	int ifindex;
38 	__be64 mac;
39 };
40 
41 /* Map for trie implementation */
42 struct {
43 	__uint(type, BPF_MAP_TYPE_LPM_TRIE);
44 	__uint(key_size, 8);
45 	__uint(value_size, sizeof(struct trie_value));
46 	__uint(max_entries, 50);
47 	__uint(map_flags, BPF_F_NO_PREALLOC);
48 } lpm_map SEC(".maps");
49 
50 /* Map for ARP table */
51 struct {
52 	__uint(type, BPF_MAP_TYPE_HASH);
53 	__type(key, __be32);
54 	__type(value, __be64);
55 	__uint(max_entries, 50);
56 } arp_table SEC(".maps");
57 
58 /* Map to keep the exact match entries in the route table */
59 struct {
60 	__uint(type, BPF_MAP_TYPE_HASH);
61 	__type(key, __be32);
62 	__type(value, struct direct_map);
63 	__uint(max_entries, 50);
64 } exact_match SEC(".maps");
65 
66 struct {
67 	__uint(type, BPF_MAP_TYPE_DEVMAP);
68 	__uint(key_size, sizeof(int));
69 	__uint(value_size, sizeof(int));
70 	__uint(max_entries, 100);
71 } tx_port SEC(".maps");
72 
73 SEC("xdp")
74 int xdp_router_ipv4_prog(struct xdp_md *ctx)
75 {
76 	void *data_end = (void *)(long)ctx->data_end;
77 	void *data = (void *)(long)ctx->data;
78 	struct ethhdr *eth = data;
79 	u64 nh_off = sizeof(*eth);
80 	struct datarec *rec;
81 	__be16 h_proto;
82 	u32 key = 0;
83 
84 	rec = bpf_map_lookup_elem(&rx_cnt, &key);
85 	if (rec)
86 		NO_TEAR_INC(rec->processed);
87 
88 	if (data + nh_off > data_end)
89 		goto drop;
90 
91 	h_proto = eth->h_proto;
92 	if (h_proto == bpf_htons(ETH_P_8021Q) ||
93 	    h_proto == bpf_htons(ETH_P_8021AD)) {
94 		struct vlan_hdr *vhdr;
95 
96 		vhdr = data + nh_off;
97 		nh_off += sizeof(struct vlan_hdr);
98 		if (data + nh_off > data_end)
99 			goto drop;
100 
101 		h_proto = vhdr->h_vlan_encapsulated_proto;
102 	}
103 
104 	switch (bpf_ntohs(h_proto)) {
105 	case ETH_P_ARP:
106 		if (rec)
107 			NO_TEAR_INC(rec->xdp_pass);
108 		return XDP_PASS;
109 	case ETH_P_IP: {
110 		struct iphdr *iph = data + nh_off;
111 		struct direct_map *direct_entry;
112 		__be64 *dest_mac, *src_mac;
113 		int forward_to;
114 
115 		if (iph + 1 > data_end)
116 			goto drop;
117 
118 		direct_entry = bpf_map_lookup_elem(&exact_match, &iph->daddr);
119 
120 		/* Check for exact match, this would give a faster lookup */
121 		if (direct_entry && direct_entry->mac &&
122 		    direct_entry->arp.mac) {
123 			src_mac = &direct_entry->mac;
124 			dest_mac = &direct_entry->arp.mac;
125 			forward_to = direct_entry->ifindex;
126 		} else {
127 			struct trie_value *prefix_value;
128 			union key_4 key4;
129 
130 			/* Look up in the trie for lpm */
131 			key4.b32[0] = 32;
132 			key4.b8[4] = iph->daddr & 0xff;
133 			key4.b8[5] = (iph->daddr >> 8) & 0xff;
134 			key4.b8[6] = (iph->daddr >> 16) & 0xff;
135 			key4.b8[7] = (iph->daddr >> 24) & 0xff;
136 
137 			prefix_value = bpf_map_lookup_elem(&lpm_map, &key4);
138 			if (!prefix_value)
139 				goto drop;
140 
141 			forward_to = prefix_value->ifindex;
142 			src_mac = &prefix_value->value;
143 			if (!src_mac)
144 				goto drop;
145 
146 			dest_mac = bpf_map_lookup_elem(&arp_table, &iph->daddr);
147 			if (!dest_mac) {
148 				if (!prefix_value->gw)
149 					goto drop;
150 
151 				dest_mac = bpf_map_lookup_elem(&arp_table,
152 							       &prefix_value->gw);
153 				if (!dest_mac) {
154 					/* Forward the packet to the kernel in
155 					 * order to trigger ARP discovery for
156 					 * the default gw.
157 					 */
158 					if (rec)
159 						NO_TEAR_INC(rec->xdp_pass);
160 					return XDP_PASS;
161 				}
162 			}
163 		}
164 
165 		if (src_mac && dest_mac) {
166 			int ret;
167 
168 			__builtin_memcpy(eth->h_dest, dest_mac, ETH_ALEN);
169 			__builtin_memcpy(eth->h_source, src_mac, ETH_ALEN);
170 
171 			ret = bpf_redirect_map(&tx_port, forward_to, 0);
172 			if (ret == XDP_REDIRECT) {
173 				if (rec)
174 					NO_TEAR_INC(rec->xdp_redirect);
175 				return ret;
176 			}
177 		}
178 	}
179 	default:
180 		break;
181 	}
182 drop:
183 	if (rec)
184 		NO_TEAR_INC(rec->xdp_drop);
185 
186 	return XDP_DROP;
187 }
188 
189 char _license[] SEC("license") = "GPL";
190