Commit | Line | Data |
---|---|---|
c9422999 | 1 | /* SPDX-License-Identifier: GPL-2.0-only */ |
ccb1352e | 2 | /* |
9dd7f890 | 3 | * Copyright (c) 2007-2017 Nicira, Inc. |
ccb1352e JG |
4 | */ |
5 | ||
6 | #ifndef FLOW_H | |
7 | #define FLOW_H 1 | |
8 | ||
e298e505 | 9 | #include <linux/cache.h> |
ccb1352e JG |
10 | #include <linux/kernel.h> |
11 | #include <linux/netlink.h> | |
12 | #include <linux/openvswitch.h> | |
13 | #include <linux/spinlock.h> | |
14 | #include <linux/types.h> | |
15 | #include <linux/rcupdate.h> | |
16 | #include <linux/if_ether.h> | |
17 | #include <linux/in6.h> | |
18 | #include <linux/jiffies.h> | |
19 | #include <linux/time.h> | |
c4b2bf6b | 20 | #include <linux/cpumask.h> |
ccb1352e | 21 | #include <net/inet_ecn.h> |
1d8fff90 | 22 | #include <net/ip_tunnels.h> |
34ae932a | 23 | #include <net/dst_metadata.h> |
b2d0f5d5 | 24 | #include <net/nsh.h> |
ccb1352e JG |
25 | |
26 | struct sk_buff; | |
ccb1352e | 27 | |
329f45bc JB |
28 | enum sw_flow_mac_proto { |
29 | MAC_PROTO_NONE = 0, | |
30 | MAC_PROTO_ETHERNET, | |
31 | }; | |
32 | #define SW_FLOW_KEY_INVALID 0x80 | |
fbdcdd78 | 33 | #define MPLS_LABEL_DEPTH 3 |
329f45bc | 34 | |
28a3f060 TA |
35 | /* Bit definitions for IPv6 Extension Header pseudo-field. */ |
36 | enum ofp12_ipv6exthdr_flags { | |
37 | OFPIEH12_NONEXT = 1 << 0, /* "No next header" encountered. */ | |
38 | OFPIEH12_ESP = 1 << 1, /* Encrypted Sec Payload header present. */ | |
39 | OFPIEH12_AUTH = 1 << 2, /* Authentication header present. */ | |
40 | OFPIEH12_DEST = 1 << 3, /* 1 or 2 dest headers present. */ | |
41 | OFPIEH12_FRAG = 1 << 4, /* Fragment header present. */ | |
42 | OFPIEH12_ROUTER = 1 << 5, /* Router header present. */ | |
43 | OFPIEH12_HOP = 1 << 6, /* Hop-by-hop header present. */ | |
44 | OFPIEH12_UNREP = 1 << 7, /* Unexpected repeats encountered. */ | |
45 | OFPIEH12_UNSEQ = 1 << 8 /* Unexpected sequencing encountered. */ | |
46 | }; | |
47 | ||
f5796684 JG |
48 | /* Store options at the end of the array if they are less than the |
49 | * maximum size. This allows us to get the benefits of variable length | |
50 | * matching for small options. | |
51 | */ | |
d91641d9 | 52 | #define TUN_METADATA_OFFSET(opt_len) \ |
c593642c | 53 | (sizeof_field(struct sw_flow_key, tun_opts) - opt_len) |
d91641d9 TG |
54 | #define TUN_METADATA_OPTS(flow_key, opt_len) \ |
55 | ((void *)((flow_key)->tun_opts + TUN_METADATA_OFFSET(opt_len))) | |
f5796684 | 56 | |
34ae932a TG |
57 | struct ovs_tunnel_info { |
58 | struct metadata_dst *tun_dst; | |
59 | }; | |
60 | ||
018c1dda EG |
61 | struct vlan_head { |
62 | __be16 tpid; /* Vlan type. Generally 802.1q or 802.1ad.*/ | |
9df46aef | 63 | __be16 tci; /* 0 if no VLAN, VLAN_CFI_MASK set otherwise. */ |
018c1dda EG |
64 | }; |
65 | ||
8f0aad6f WZ |
66 | #define OVS_SW_FLOW_KEY_METADATA_SIZE \ |
67 | (offsetof(struct sw_flow_key, recirc_id) + \ | |
c593642c | 68 | sizeof_field(struct sw_flow_key, recirc_id)) |
8f0aad6f | 69 | |
b2d0f5d5 YY |
70 | struct ovs_key_nsh { |
71 | struct ovs_nsh_key_base base; | |
72 | __be32 context[NSH_MD1_CONTEXT_SIZE]; | |
73 | }; | |
74 | ||
ccb1352e | 75 | struct sw_flow_key { |
fca5fdf6 | 76 | u8 tun_opts[IP_TUNNEL_OPTS_MAX]; |
f5796684 | 77 | u8 tun_opts_len; |
1d8fff90 | 78 | struct ip_tunnel_key tun_key; /* Encapsulating tunnel key. */ |
ccb1352e JG |
79 | struct { |
80 | u32 priority; /* Packet QoS priority. */ | |
39c7caeb | 81 | u32 skb_mark; /* SKB mark. */ |
15eac2a7 | 82 | u16 in_port; /* Input switch port (or DP_MAX_PORTS). */ |
1139e241 | 83 | } __packed phy; /* Safe when right after 'tun_key'. */ |
329f45bc | 84 | u8 mac_proto; /* MAC layer protocol (e.g. Ethernet). */ |
00a93bab | 85 | u8 tun_proto; /* Protocol of encapsulating tunnel. */ |
971427f3 AZ |
86 | u32 ovs_flow_hash; /* Datapath computed hash value. */ |
87 | u32 recirc_id; /* Recirculation ID. */ | |
ccb1352e JG |
88 | struct { |
89 | u8 src[ETH_ALEN]; /* Ethernet source address. */ | |
90 | u8 dst[ETH_ALEN]; /* Ethernet destination address. */ | |
018c1dda EG |
91 | struct vlan_head vlan; |
92 | struct vlan_head cvlan; | |
ccb1352e JG |
93 | __be16 type; /* Ethernet frame type. */ |
94 | } eth; | |
316d4d78 JR |
95 | /* Filling a hole of two bytes. */ |
96 | u8 ct_state; | |
97 | u8 ct_orig_proto; /* CT original direction tuple IP | |
98 | * protocol. | |
99 | */ | |
25cd9ba0 | 100 | union { |
25cd9ba0 SH |
101 | struct { |
102 | u8 proto; /* IP protocol or lower 8 bits of ARP opcode. */ | |
103 | u8 tos; /* IP ToS. */ | |
104 | u8 ttl; /* IP TTL/hop limit. */ | |
105 | u8 frag; /* One of OVS_FRAG_TYPE_*. */ | |
106 | } ip; | |
107 | }; | |
316d4d78 | 108 | u16 ct_zone; /* Conntrack zone. */ |
1139e241 JR |
109 | struct { |
110 | __be16 src; /* TCP/UDP/SCTP source port. */ | |
111 | __be16 dst; /* TCP/UDP/SCTP destination port. */ | |
112 | __be16 flags; /* TCP flags. */ | |
113 | } tp; | |
ccb1352e JG |
114 | union { |
115 | struct { | |
116 | struct { | |
117 | __be32 src; /* IP source address. */ | |
118 | __be32 dst; /* IP destination address. */ | |
119 | } addr; | |
9dd7f890 JR |
120 | union { |
121 | struct { | |
122 | __be32 src; | |
123 | __be32 dst; | |
124 | } ct_orig; /* Conntrack original direction fields. */ | |
125 | struct { | |
126 | u8 sha[ETH_ALEN]; /* ARP source hardware address. */ | |
127 | u8 tha[ETH_ALEN]; /* ARP target hardware address. */ | |
128 | } arp; | |
129 | }; | |
ccb1352e JG |
130 | } ipv4; |
131 | struct { | |
132 | struct { | |
133 | struct in6_addr src; /* IPv6 source address. */ | |
134 | struct in6_addr dst; /* IPv6 destination address. */ | |
135 | } addr; | |
136 | __be32 label; /* IPv6 flow label. */ | |
28a3f060 | 137 | u16 exthdrs; /* IPv6 extension header flags */ |
9dd7f890 JR |
138 | union { |
139 | struct { | |
140 | struct in6_addr src; | |
141 | struct in6_addr dst; | |
142 | } ct_orig; /* Conntrack original direction fields. */ | |
143 | struct { | |
144 | struct in6_addr target; /* ND target address. */ | |
145 | u8 sll[ETH_ALEN]; /* ND source link layer address. */ | |
146 | u8 tll[ETH_ALEN]; /* ND target link layer address. */ | |
147 | } nd; | |
148 | }; | |
ccb1352e | 149 | } ipv6; |
fbdcdd78 MV |
150 | struct { |
151 | u32 num_labels_mask; /* labels present bitmap of effective length MPLS_LABEL_DEPTH */ | |
152 | __be32 lse[MPLS_LABEL_DEPTH]; /* label stack entry */ | |
153 | } mpls; | |
154 | ||
b2d0f5d5 | 155 | struct ovs_key_nsh nsh; /* network service header */ |
ccb1352e | 156 | }; |
7f8a436e | 157 | struct { |
316d4d78 | 158 | /* Connection tracking fields not packed above. */ |
9dd7f890 JR |
159 | struct { |
160 | __be16 src; /* CT orig tuple tp src port. */ | |
161 | __be16 dst; /* CT orig tuple tp dst port. */ | |
162 | } orig_tp; | |
316d4d78 | 163 | u32 mark; |
33db4125 | 164 | struct ovs_key_ct_labels labels; |
7f8a436e JS |
165 | } ct; |
166 | ||
0d40f75b | 167 | } __aligned(BITS_PER_LONG/8); /* Ensure that we can do comparisons as longs. */ |
ccb1352e | 168 | |
9dd7f890 JR |
169 | static inline bool sw_flow_key_is_nd(const struct sw_flow_key *key) |
170 | { | |
171 | return key->eth.type == htons(ETH_P_IPV6) && | |
172 | key->ip.proto == NEXTHDR_ICMP && | |
173 | key->tp.dst == 0 && | |
174 | (key->tp.src == htons(NDISC_NEIGHBOUR_SOLICITATION) || | |
175 | key->tp.src == htons(NDISC_NEIGHBOUR_ADVERTISEMENT)); | |
176 | } | |
177 | ||
e6445719 | 178 | struct sw_flow_key_range { |
8f49ce11 BP |
179 | unsigned short int start; |
180 | unsigned short int end; | |
e6445719 PS |
181 | }; |
182 | ||
183 | struct sw_flow_mask { | |
184 | int ref_count; | |
185 | struct rcu_head rcu; | |
e6445719 PS |
186 | struct sw_flow_key_range range; |
187 | struct sw_flow_key key; | |
188 | }; | |
189 | ||
190 | struct sw_flow_match { | |
191 | struct sw_flow_key *key; | |
192 | struct sw_flow_key_range range; | |
193 | struct sw_flow_mask *mask; | |
194 | }; | |
195 | ||
74ed7ab9 JS |
196 | #define MAX_UFID_LENGTH 16 /* 128 bits */ |
197 | ||
198 | struct sw_flow_id { | |
199 | u32 ufid_len; | |
200 | union { | |
201 | u32 ufid[MAX_UFID_LENGTH / 4]; | |
202 | struct sw_flow_key *unmasked_key; | |
203 | }; | |
204 | }; | |
205 | ||
e6445719 PS |
206 | struct sw_flow_actions { |
207 | struct rcu_head rcu; | |
8e2fed1c | 208 | size_t orig_len; /* From flow_cmd_new netlink actions size */ |
e6445719 PS |
209 | u32 actions_len; |
210 | struct nlattr actions[]; | |
211 | }; | |
212 | ||
aef833c5 | 213 | struct sw_flow_stats { |
e298e505 PS |
214 | u64 packet_count; /* Number of packets matched. */ |
215 | u64 byte_count; /* Number of bytes matched. */ | |
216 | unsigned long used; /* Last used time (in jiffies). */ | |
217 | spinlock_t lock; /* Lock for atomic stats update. */ | |
218 | __be16 tcp_flags; /* Union of seen TCP flags. */ | |
219 | }; | |
220 | ||
ccb1352e JG |
221 | struct sw_flow { |
222 | struct rcu_head rcu; | |
74ed7ab9 JS |
223 | struct { |
224 | struct hlist_node node[2]; | |
225 | u32 hash; | |
226 | } flow_table, ufid_table; | |
db74a333 | 227 | int stats_last_writer; /* CPU id of the last writer on |
63e7959c JR |
228 | * 'stats[0]'. |
229 | */ | |
ccb1352e | 230 | struct sw_flow_key key; |
74ed7ab9 | 231 | struct sw_flow_id id; |
15ea59a0 | 232 | struct cpumask *cpu_used_mask; |
03f0d916 | 233 | struct sw_flow_mask *mask; |
ccb1352e | 234 | struct sw_flow_actions __rcu *sf_acts; |
aef833c5 | 235 | struct sw_flow_stats __rcu *stats[]; /* One for each CPU. First one |
63e7959c JR |
236 | * is allocated at flow creation time, |
237 | * the rest are allocated on demand | |
238 | * while holding the 'stats[0].lock'. | |
239 | */ | |
ccb1352e JG |
240 | }; |
241 | ||
242 | struct arp_eth_header { | |
243 | __be16 ar_hrd; /* format of hardware address */ | |
244 | __be16 ar_pro; /* format of protocol address */ | |
245 | unsigned char ar_hln; /* length of hardware address */ | |
246 | unsigned char ar_pln; /* length of protocol address */ | |
247 | __be16 ar_op; /* ARP opcode (command) */ | |
248 | ||
249 | /* Ethernet+IPv4 specific members. */ | |
250 | unsigned char ar_sha[ETH_ALEN]; /* sender hardware address */ | |
251 | unsigned char ar_sip[4]; /* sender IP address */ | |
252 | unsigned char ar_tha[ETH_ALEN]; /* target hardware address */ | |
253 | unsigned char ar_tip[4]; /* target IP address */ | |
254 | } __packed; | |
255 | ||
329f45bc JB |
256 | static inline u8 ovs_key_mac_proto(const struct sw_flow_key *key) |
257 | { | |
258 | return key->mac_proto & ~SW_FLOW_KEY_INVALID; | |
259 | } | |
260 | ||
261 | static inline u16 __ovs_mac_header_len(u8 mac_proto) | |
262 | { | |
263 | return mac_proto == MAC_PROTO_ETHERNET ? ETH_HLEN : 0; | |
264 | } | |
265 | ||
266 | static inline u16 ovs_mac_header_len(const struct sw_flow_key *key) | |
267 | { | |
268 | return __ovs_mac_header_len(ovs_key_mac_proto(key)); | |
269 | } | |
270 | ||
74ed7ab9 JS |
271 | static inline bool ovs_identifier_is_ufid(const struct sw_flow_id *sfid) |
272 | { | |
273 | return sfid->ufid_len; | |
274 | } | |
275 | ||
276 | static inline bool ovs_identifier_is_key(const struct sw_flow_id *sfid) | |
277 | { | |
278 | return !ovs_identifier_is_ufid(sfid); | |
279 | } | |
280 | ||
ad552007 | 281 | void ovs_flow_stats_update(struct sw_flow *, __be16 tcp_flags, |
12eb18f7 | 282 | const struct sk_buff *); |
86ec8dba | 283 | void ovs_flow_stats_get(const struct sw_flow *, struct ovs_flow_stats *, |
e298e505 | 284 | unsigned long *used, __be16 *tcp_flags); |
86ec8dba | 285 | void ovs_flow_stats_clear(struct sw_flow *); |
ccb1352e | 286 | u64 ovs_flow_used_time(unsigned long flow_jiffies); |
ccb1352e | 287 | |
971427f3 | 288 | int ovs_flow_key_update(struct sk_buff *skb, struct sw_flow_key *key); |
ad06a566 | 289 | int ovs_flow_key_update_l3l4(struct sk_buff *skb, struct sw_flow_key *key); |
1d8fff90 | 290 | int ovs_flow_key_extract(const struct ip_tunnel_info *tun_info, |
12eb18f7 | 291 | struct sk_buff *skb, |
f0b128c1 | 292 | struct sw_flow_key *key); |
83c8df26 | 293 | /* Extract key from packet coming from userspace. */ |
c2ac6673 | 294 | int ovs_flow_key_extract_userspace(struct net *net, const struct nlattr *attr, |
83c8df26 | 295 | struct sk_buff *skb, |
05da5898 | 296 | struct sw_flow_key *key, bool log); |
03f0d916 | 297 | |
ccb1352e | 298 | #endif /* flow.h */ |