Eric Dumazet | 0744dd0 | 2011-11-28 05:22:18 +0000 | [diff] [blame] | 1 | #include <linux/skbuff.h> |
Jesper Dangaard Brouer | c452ed7 | 2012-01-24 16:03:33 -0500 | [diff] [blame] | 2 | #include <linux/export.h> |
Eric Dumazet | 0744dd0 | 2011-11-28 05:22:18 +0000 | [diff] [blame] | 3 | #include <linux/ip.h> |
| 4 | #include <linux/ipv6.h> |
| 5 | #include <linux/if_vlan.h> |
| 6 | #include <net/ip.h> |
Eric Dumazet | ddbe503 | 2012-07-18 08:11:12 +0000 | [diff] [blame] | 7 | #include <net/ipv6.h> |
Daniel Borkmann | f77668d | 2013-03-19 06:39:30 +0000 | [diff] [blame] | 8 | #include <linux/igmp.h> |
| 9 | #include <linux/icmp.h> |
| 10 | #include <linux/sctp.h> |
| 11 | #include <linux/dccp.h> |
Eric Dumazet | 0744dd0 | 2011-11-28 05:22:18 +0000 | [diff] [blame] | 12 | #include <linux/if_tunnel.h> |
| 13 | #include <linux/if_pppox.h> |
| 14 | #include <linux/ppp_defs.h> |
| 15 | #include <net/flow_keys.h> |
| 16 | |
Eric Dumazet | 4d77d2b | 2011-11-28 20:30:35 +0000 | [diff] [blame] | 17 | /* copy saddr & daddr, possibly using 64bit load/store |
| 18 | * Equivalent to : flow->src = iph->saddr; |
| 19 | * flow->dst = iph->daddr; |
| 20 | */ |
| 21 | static void iph_to_flow_copy_addrs(struct flow_keys *flow, const struct iphdr *iph) |
| 22 | { |
| 23 | BUILD_BUG_ON(offsetof(typeof(*flow), dst) != |
| 24 | offsetof(typeof(*flow), src) + sizeof(flow->src)); |
| 25 | memcpy(&flow->src, &iph->saddr, sizeof(flow->src) + sizeof(flow->dst)); |
| 26 | } |
Eric Dumazet | 0744dd0 | 2011-11-28 05:22:18 +0000 | [diff] [blame] | 27 | |
| 28 | bool skb_flow_dissect(const struct sk_buff *skb, struct flow_keys *flow) |
| 29 | { |
| 30 | int poff, nhoff = skb_network_offset(skb); |
| 31 | u8 ip_proto; |
| 32 | __be16 proto = skb->protocol; |
| 33 | |
| 34 | memset(flow, 0, sizeof(*flow)); |
| 35 | |
| 36 | again: |
| 37 | switch (proto) { |
| 38 | case __constant_htons(ETH_P_IP): { |
| 39 | const struct iphdr *iph; |
| 40 | struct iphdr _iph; |
| 41 | ip: |
| 42 | iph = skb_header_pointer(skb, nhoff, sizeof(_iph), &_iph); |
| 43 | if (!iph) |
| 44 | return false; |
| 45 | |
| 46 | if (ip_is_fragment(iph)) |
| 47 | ip_proto = 0; |
| 48 | else |
| 49 | ip_proto = iph->protocol; |
Eric Dumazet | 4d77d2b | 2011-11-28 20:30:35 +0000 | [diff] [blame] | 50 | iph_to_flow_copy_addrs(flow, iph); |
Eric Dumazet | 0744dd0 | 2011-11-28 05:22:18 +0000 | [diff] [blame] | 51 | nhoff += iph->ihl * 4; |
| 52 | break; |
| 53 | } |
| 54 | case __constant_htons(ETH_P_IPV6): { |
| 55 | const struct ipv6hdr *iph; |
| 56 | struct ipv6hdr _iph; |
| 57 | ipv6: |
| 58 | iph = skb_header_pointer(skb, nhoff, sizeof(_iph), &_iph); |
| 59 | if (!iph) |
| 60 | return false; |
| 61 | |
| 62 | ip_proto = iph->nexthdr; |
Eric Dumazet | ddbe503 | 2012-07-18 08:11:12 +0000 | [diff] [blame] | 63 | flow->src = (__force __be32)ipv6_addr_hash(&iph->saddr); |
| 64 | flow->dst = (__force __be32)ipv6_addr_hash(&iph->daddr); |
Eric Dumazet | 0744dd0 | 2011-11-28 05:22:18 +0000 | [diff] [blame] | 65 | nhoff += sizeof(struct ipv6hdr); |
| 66 | break; |
| 67 | } |
Eric Dumazet | e11aada | 2013-08-06 04:35:06 -0700 | [diff] [blame] | 68 | case __constant_htons(ETH_P_8021AD): |
Eric Dumazet | 0744dd0 | 2011-11-28 05:22:18 +0000 | [diff] [blame] | 69 | case __constant_htons(ETH_P_8021Q): { |
| 70 | const struct vlan_hdr *vlan; |
| 71 | struct vlan_hdr _vlan; |
| 72 | |
| 73 | vlan = skb_header_pointer(skb, nhoff, sizeof(_vlan), &_vlan); |
| 74 | if (!vlan) |
| 75 | return false; |
| 76 | |
| 77 | proto = vlan->h_vlan_encapsulated_proto; |
| 78 | nhoff += sizeof(*vlan); |
| 79 | goto again; |
| 80 | } |
| 81 | case __constant_htons(ETH_P_PPP_SES): { |
| 82 | struct { |
| 83 | struct pppoe_hdr hdr; |
| 84 | __be16 proto; |
| 85 | } *hdr, _hdr; |
| 86 | hdr = skb_header_pointer(skb, nhoff, sizeof(_hdr), &_hdr); |
| 87 | if (!hdr) |
| 88 | return false; |
| 89 | proto = hdr->proto; |
| 90 | nhoff += PPPOE_SES_HLEN; |
| 91 | switch (proto) { |
| 92 | case __constant_htons(PPP_IP): |
| 93 | goto ip; |
| 94 | case __constant_htons(PPP_IPV6): |
| 95 | goto ipv6; |
| 96 | default: |
| 97 | return false; |
| 98 | } |
| 99 | } |
| 100 | default: |
| 101 | return false; |
| 102 | } |
| 103 | |
| 104 | switch (ip_proto) { |
| 105 | case IPPROTO_GRE: { |
| 106 | struct gre_hdr { |
| 107 | __be16 flags; |
| 108 | __be16 proto; |
| 109 | } *hdr, _hdr; |
| 110 | |
| 111 | hdr = skb_header_pointer(skb, nhoff, sizeof(_hdr), &_hdr); |
| 112 | if (!hdr) |
| 113 | return false; |
| 114 | /* |
| 115 | * Only look inside GRE if version zero and no |
| 116 | * routing |
| 117 | */ |
| 118 | if (!(hdr->flags & (GRE_VERSION|GRE_ROUTING))) { |
| 119 | proto = hdr->proto; |
| 120 | nhoff += 4; |
| 121 | if (hdr->flags & GRE_CSUM) |
| 122 | nhoff += 4; |
| 123 | if (hdr->flags & GRE_KEY) |
| 124 | nhoff += 4; |
| 125 | if (hdr->flags & GRE_SEQ) |
| 126 | nhoff += 4; |
Michael Dalton | e1733de | 2013-03-11 06:52:28 +0000 | [diff] [blame] | 127 | if (proto == htons(ETH_P_TEB)) { |
| 128 | const struct ethhdr *eth; |
| 129 | struct ethhdr _eth; |
| 130 | |
| 131 | eth = skb_header_pointer(skb, nhoff, |
| 132 | sizeof(_eth), &_eth); |
| 133 | if (!eth) |
| 134 | return false; |
| 135 | proto = eth->h_proto; |
| 136 | nhoff += sizeof(*eth); |
| 137 | } |
Eric Dumazet | 0744dd0 | 2011-11-28 05:22:18 +0000 | [diff] [blame] | 138 | goto again; |
| 139 | } |
| 140 | break; |
| 141 | } |
| 142 | case IPPROTO_IPIP: |
Tom Herbert | fca4189 | 2013-07-29 11:07:36 -0700 | [diff] [blame] | 143 | proto = htons(ETH_P_IP); |
| 144 | goto ip; |
Tom Herbert | b438f94 | 2013-07-29 11:07:42 -0700 | [diff] [blame] | 145 | case IPPROTO_IPV6: |
| 146 | proto = htons(ETH_P_IPV6); |
| 147 | goto ipv6; |
Eric Dumazet | 0744dd0 | 2011-11-28 05:22:18 +0000 | [diff] [blame] | 148 | default: |
| 149 | break; |
| 150 | } |
| 151 | |
| 152 | flow->ip_proto = ip_proto; |
| 153 | poff = proto_ports_offset(ip_proto); |
| 154 | if (poff >= 0) { |
| 155 | __be32 *ports, _ports; |
| 156 | |
Eric Dumazet | b867835 | 2013-09-26 08:44:06 -0700 | [diff] [blame] | 157 | ports = skb_header_pointer(skb, nhoff + poff, |
| 158 | sizeof(_ports), &_ports); |
Eric Dumazet | 0744dd0 | 2011-11-28 05:22:18 +0000 | [diff] [blame] | 159 | if (ports) |
| 160 | flow->ports = *ports; |
| 161 | } |
| 162 | |
Daniel Borkmann | 8ed7816 | 2013-03-19 06:39:29 +0000 | [diff] [blame] | 163 | flow->thoff = (u16) nhoff; |
| 164 | |
Eric Dumazet | 0744dd0 | 2011-11-28 05:22:18 +0000 | [diff] [blame] | 165 | return true; |
| 166 | } |
| 167 | EXPORT_SYMBOL(skb_flow_dissect); |
Cong Wang | 441d9d3 | 2013-01-21 00:39:24 +0000 | [diff] [blame] | 168 | |
| 169 | static u32 hashrnd __read_mostly; |
| 170 | |
| 171 | /* |
| 172 | * __skb_get_rxhash: calculate a flow hash based on src/dst addresses |
| 173 | * and src/dst port numbers. Sets rxhash in skb to non-zero hash value |
| 174 | * on success, zero indicates no valid hash. Also, sets l4_rxhash in skb |
| 175 | * if hash is a canonical 4-tuple hash over transport ports. |
| 176 | */ |
| 177 | void __skb_get_rxhash(struct sk_buff *skb) |
| 178 | { |
| 179 | struct flow_keys keys; |
| 180 | u32 hash; |
| 181 | |
| 182 | if (!skb_flow_dissect(skb, &keys)) |
| 183 | return; |
| 184 | |
| 185 | if (keys.ports) |
| 186 | skb->l4_rxhash = 1; |
| 187 | |
| 188 | /* get a consistent hash (same value on both flow directions) */ |
| 189 | if (((__force u32)keys.dst < (__force u32)keys.src) || |
| 190 | (((__force u32)keys.dst == (__force u32)keys.src) && |
| 191 | ((__force u16)keys.port16[1] < (__force u16)keys.port16[0]))) { |
| 192 | swap(keys.dst, keys.src); |
| 193 | swap(keys.port16[0], keys.port16[1]); |
| 194 | } |
| 195 | |
| 196 | hash = jhash_3words((__force u32)keys.dst, |
| 197 | (__force u32)keys.src, |
| 198 | (__force u32)keys.ports, hashrnd); |
| 199 | if (!hash) |
| 200 | hash = 1; |
| 201 | |
| 202 | skb->rxhash = hash; |
| 203 | } |
| 204 | EXPORT_SYMBOL(__skb_get_rxhash); |
| 205 | |
| 206 | /* |
| 207 | * Returns a Tx hash based on the given packet descriptor a Tx queues' number |
| 208 | * to be used as a distribution range. |
| 209 | */ |
| 210 | u16 __skb_tx_hash(const struct net_device *dev, const struct sk_buff *skb, |
| 211 | unsigned int num_tx_queues) |
| 212 | { |
| 213 | u32 hash; |
| 214 | u16 qoffset = 0; |
| 215 | u16 qcount = num_tx_queues; |
| 216 | |
| 217 | if (skb_rx_queue_recorded(skb)) { |
| 218 | hash = skb_get_rx_queue(skb); |
| 219 | while (unlikely(hash >= num_tx_queues)) |
| 220 | hash -= num_tx_queues; |
| 221 | return hash; |
| 222 | } |
| 223 | |
| 224 | if (dev->num_tc) { |
| 225 | u8 tc = netdev_get_prio_tc_map(dev, skb->priority); |
| 226 | qoffset = dev->tc_to_txq[tc].offset; |
| 227 | qcount = dev->tc_to_txq[tc].count; |
| 228 | } |
| 229 | |
| 230 | if (skb->sk && skb->sk->sk_hash) |
| 231 | hash = skb->sk->sk_hash; |
| 232 | else |
| 233 | hash = (__force u16) skb->protocol; |
| 234 | hash = jhash_1word(hash, hashrnd); |
| 235 | |
| 236 | return (u16) (((u64) hash * qcount) >> 32) + qoffset; |
| 237 | } |
| 238 | EXPORT_SYMBOL(__skb_tx_hash); |
| 239 | |
Daniel Borkmann | f77668d | 2013-03-19 06:39:30 +0000 | [diff] [blame] | 240 | /* __skb_get_poff() returns the offset to the payload as far as it could |
| 241 | * be dissected. The main user is currently BPF, so that we can dynamically |
| 242 | * truncate packets without needing to push actual payload to the user |
| 243 | * space and can analyze headers only, instead. |
| 244 | */ |
| 245 | u32 __skb_get_poff(const struct sk_buff *skb) |
| 246 | { |
| 247 | struct flow_keys keys; |
| 248 | u32 poff = 0; |
| 249 | |
| 250 | if (!skb_flow_dissect(skb, &keys)) |
| 251 | return 0; |
| 252 | |
| 253 | poff += keys.thoff; |
| 254 | switch (keys.ip_proto) { |
| 255 | case IPPROTO_TCP: { |
| 256 | const struct tcphdr *tcph; |
| 257 | struct tcphdr _tcph; |
| 258 | |
| 259 | tcph = skb_header_pointer(skb, poff, sizeof(_tcph), &_tcph); |
| 260 | if (!tcph) |
| 261 | return poff; |
| 262 | |
| 263 | poff += max_t(u32, sizeof(struct tcphdr), tcph->doff * 4); |
| 264 | break; |
| 265 | } |
| 266 | case IPPROTO_UDP: |
| 267 | case IPPROTO_UDPLITE: |
| 268 | poff += sizeof(struct udphdr); |
| 269 | break; |
| 270 | /* For the rest, we do not really care about header |
| 271 | * extensions at this point for now. |
| 272 | */ |
| 273 | case IPPROTO_ICMP: |
| 274 | poff += sizeof(struct icmphdr); |
| 275 | break; |
| 276 | case IPPROTO_ICMPV6: |
| 277 | poff += sizeof(struct icmp6hdr); |
| 278 | break; |
| 279 | case IPPROTO_IGMP: |
| 280 | poff += sizeof(struct igmphdr); |
| 281 | break; |
| 282 | case IPPROTO_DCCP: |
| 283 | poff += sizeof(struct dccp_hdr); |
| 284 | break; |
| 285 | case IPPROTO_SCTP: |
| 286 | poff += sizeof(struct sctphdr); |
| 287 | break; |
| 288 | } |
| 289 | |
| 290 | return poff; |
| 291 | } |
| 292 | |
Cong Wang | 441d9d3 | 2013-01-21 00:39:24 +0000 | [diff] [blame] | 293 | static inline u16 dev_cap_txqueue(struct net_device *dev, u16 queue_index) |
| 294 | { |
| 295 | if (unlikely(queue_index >= dev->real_num_tx_queues)) { |
| 296 | net_warn_ratelimited("%s selects TX queue %d, but real number of TX queues is %d\n", |
| 297 | dev->name, queue_index, |
| 298 | dev->real_num_tx_queues); |
| 299 | return 0; |
| 300 | } |
| 301 | return queue_index; |
| 302 | } |
| 303 | |
| 304 | static inline int get_xps_queue(struct net_device *dev, struct sk_buff *skb) |
| 305 | { |
| 306 | #ifdef CONFIG_XPS |
| 307 | struct xps_dev_maps *dev_maps; |
| 308 | struct xps_map *map; |
| 309 | int queue_index = -1; |
| 310 | |
| 311 | rcu_read_lock(); |
| 312 | dev_maps = rcu_dereference(dev->xps_maps); |
| 313 | if (dev_maps) { |
| 314 | map = rcu_dereference( |
| 315 | dev_maps->cpu_map[raw_smp_processor_id()]); |
| 316 | if (map) { |
| 317 | if (map->len == 1) |
| 318 | queue_index = map->queues[0]; |
| 319 | else { |
| 320 | u32 hash; |
| 321 | if (skb->sk && skb->sk->sk_hash) |
| 322 | hash = skb->sk->sk_hash; |
| 323 | else |
| 324 | hash = (__force u16) skb->protocol ^ |
| 325 | skb->rxhash; |
| 326 | hash = jhash_1word(hash, hashrnd); |
| 327 | queue_index = map->queues[ |
| 328 | ((u64)hash * map->len) >> 32]; |
| 329 | } |
| 330 | if (unlikely(queue_index >= dev->real_num_tx_queues)) |
| 331 | queue_index = -1; |
| 332 | } |
| 333 | } |
| 334 | rcu_read_unlock(); |
| 335 | |
| 336 | return queue_index; |
| 337 | #else |
| 338 | return -1; |
| 339 | #endif |
| 340 | } |
| 341 | |
| 342 | u16 __netdev_pick_tx(struct net_device *dev, struct sk_buff *skb) |
| 343 | { |
| 344 | struct sock *sk = skb->sk; |
| 345 | int queue_index = sk_tx_queue_get(sk); |
| 346 | |
| 347 | if (queue_index < 0 || skb->ooo_okay || |
| 348 | queue_index >= dev->real_num_tx_queues) { |
| 349 | int new_index = get_xps_queue(dev, skb); |
| 350 | if (new_index < 0) |
| 351 | new_index = skb_tx_hash(dev, skb); |
| 352 | |
Eric Dumazet | 702821f | 2013-08-28 18:10:43 -0700 | [diff] [blame] | 353 | if (queue_index != new_index && sk && |
| 354 | rcu_access_pointer(sk->sk_dst_cache)) |
Eric Dumazet | 50d1784 | 2013-09-07 12:02:57 -0700 | [diff] [blame] | 355 | sk_tx_queue_set(sk, new_index); |
Cong Wang | 441d9d3 | 2013-01-21 00:39:24 +0000 | [diff] [blame] | 356 | |
| 357 | queue_index = new_index; |
| 358 | } |
| 359 | |
| 360 | return queue_index; |
| 361 | } |
| 362 | EXPORT_SYMBOL(__netdev_pick_tx); |
| 363 | |
| 364 | struct netdev_queue *netdev_pick_tx(struct net_device *dev, |
| 365 | struct sk_buff *skb) |
| 366 | { |
| 367 | int queue_index = 0; |
| 368 | |
| 369 | if (dev->real_num_tx_queues != 1) { |
| 370 | const struct net_device_ops *ops = dev->netdev_ops; |
| 371 | if (ops->ndo_select_queue) |
| 372 | queue_index = ops->ndo_select_queue(dev, skb); |
| 373 | else |
| 374 | queue_index = __netdev_pick_tx(dev, skb); |
| 375 | queue_index = dev_cap_txqueue(dev, queue_index); |
| 376 | } |
| 377 | |
| 378 | skb_set_queue_mapping(skb, queue_index); |
| 379 | return netdev_get_tx_queue(dev, queue_index); |
| 380 | } |
| 381 | |
| 382 | static int __init initialize_hashrnd(void) |
| 383 | { |
| 384 | get_random_bytes(&hashrnd, sizeof(hashrnd)); |
| 385 | return 0; |
| 386 | } |
| 387 | |
| 388 | late_initcall_sync(initialize_hashrnd); |