1#include <linux/skbuff.h> 2#include <linux/export.h> 3#include <linux/ip.h> 4#include <linux/ipv6.h> 5#include <linux/if_vlan.h> 6#include <net/ip.h> 7#include <net/ipv6.h> 8#include <linux/igmp.h> 9#include <linux/icmp.h> 10#include <linux/sctp.h> 11#include <linux/dccp.h> 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#include <scsi/fc/fc_fcoe.h> 17 18/* copy saddr & daddr, possibly using 64bit load/store 19 * Equivalent to : flow->src = iph->saddr; 20 * flow->dst = iph->daddr; 21 */ 22static void iph_to_flow_copy_addrs(struct flow_keys *flow, const struct iphdr *iph) 23{ 24 BUILD_BUG_ON(offsetof(typeof(*flow), dst) != 25 offsetof(typeof(*flow), src) + sizeof(flow->src)); 26 memcpy(&flow->src, &iph->saddr, sizeof(flow->src) + sizeof(flow->dst)); 27} 28 29/** 30 * __skb_flow_get_ports - extract the upper layer ports and return them 31 * @skb: sk_buff to extract the ports from 32 * @thoff: transport header offset 33 * @ip_proto: protocol for which to get port offset 34 * @data: raw buffer pointer to the packet, if NULL use skb->data 35 * @hlen: packet header length, if @data is NULL use skb_headlen(skb) 36 * 37 * The function will try to retrieve the ports at offset thoff + poff where poff 38 * is the protocol port offset returned from proto_ports_offset 39 */ 40__be32 __skb_flow_get_ports(const struct sk_buff *skb, int thoff, u8 ip_proto, 41 void *data, int hlen) 42{ 43 int poff = proto_ports_offset(ip_proto); 44 45 if (!data) { 46 data = skb->data; 47 hlen = skb_headlen(skb); 48 } 49 50 if (poff >= 0) { 51 __be32 *ports, _ports; 52 53 ports = __skb_header_pointer(skb, thoff + poff, 54 sizeof(_ports), data, hlen, &_ports); 55 if (ports) 56 return *ports; 57 } 58 59 return 0; 60} 61EXPORT_SYMBOL(__skb_flow_get_ports); 62 63/** 64 * __skb_flow_dissect - extract the flow_keys struct and return it 65 * @skb: sk_buff to extract the flow from, can be NULL if the rest are specified 66 * @data: raw buffer pointer to the packet, if NULL use skb->data 67 * @proto: protocol for which to get the flow, if @data is NULL use skb->protocol 68 * @nhoff: network header offset, if @data is NULL use skb_network_offset(skb) 69 * @hlen: packet header length, if @data is NULL use skb_headlen(skb) 70 * 71 * The function will try to retrieve the struct flow_keys from either the skbuff 72 * or a raw buffer specified by the rest parameters 73 */ 74bool __skb_flow_dissect(const struct sk_buff *skb, struct flow_keys *flow, 75 void *data, __be16 proto, int nhoff, int hlen) 76{ 77 u8 ip_proto; 78 79 if (!data) { 80 data = skb->data; 81 proto = skb->protocol; 82 nhoff = skb_network_offset(skb); 83 hlen = skb_headlen(skb); 84 } 85 86 memset(flow, 0, sizeof(*flow)); 87 88again: 89 switch (proto) { 90 case htons(ETH_P_IP): { 91 const struct iphdr *iph; 92 struct iphdr _iph; 93ip: 94 iph = __skb_header_pointer(skb, nhoff, sizeof(_iph), data, hlen, &_iph); 95 if (!iph || iph->ihl < 5) 96 return false; 97 nhoff += iph->ihl * 4; 98 99 ip_proto = iph->protocol; 100 if (ip_is_fragment(iph)) 101 ip_proto = 0; 102 103 /* skip the address processing if skb is NULL. The assumption 104 * here is that if there is no skb we are not looking for flow 105 * info but lengths and protocols. 106 */ 107 if (!skb) 108 break; 109 110 iph_to_flow_copy_addrs(flow, iph); 111 break; 112 } 113 case htons(ETH_P_IPV6): { 114 const struct ipv6hdr *iph; 115 struct ipv6hdr _iph; 116 117ipv6: 118 iph = __skb_header_pointer(skb, nhoff, sizeof(_iph), data, hlen, &_iph); 119 if (!iph) 120 return false; 121 122 ip_proto = iph->nexthdr; 123 nhoff += sizeof(struct ipv6hdr); 124 125 /* see comment above in IPv4 section */ 126 if (!skb) 127 break; 128 129 flow->src = (__force __be32)ipv6_addr_hash(&iph->saddr); 130 flow->dst = (__force __be32)ipv6_addr_hash(&iph->daddr); 131 132 if (skb && ip6_flowlabel(iph)) { 133 __be32 flow_label = ip6_flowlabel(iph); 134 135 /* Awesome, IPv6 packet has a flow label so we can 136 * use that to represent the ports without any 137 * further dissection. 138 */ 139 flow->n_proto = proto; 140 flow->ip_proto = ip_proto; 141 flow->ports = flow_label; 142 flow->thoff = (u16)nhoff; 143 144 return true; 145 } 146 147 break; 148 } 149 case htons(ETH_P_8021AD): 150 case htons(ETH_P_8021Q): { 151 const struct vlan_hdr *vlan; 152 struct vlan_hdr _vlan; 153 154 vlan = __skb_header_pointer(skb, nhoff, sizeof(_vlan), data, hlen, &_vlan); 155 if (!vlan) 156 return false; 157 158 proto = vlan->h_vlan_encapsulated_proto; 159 nhoff += sizeof(*vlan); 160 goto again; 161 } 162 case htons(ETH_P_PPP_SES): { 163 struct { 164 struct pppoe_hdr hdr; 165 __be16 proto; 166 } *hdr, _hdr; 167 hdr = __skb_header_pointer(skb, nhoff, sizeof(_hdr), data, hlen, &_hdr); 168 if (!hdr) 169 return false; 170 proto = hdr->proto; 171 nhoff += PPPOE_SES_HLEN; 172 switch (proto) { 173 case htons(PPP_IP): 174 goto ip; 175 case htons(PPP_IPV6): 176 goto ipv6; 177 default: 178 return false; 179 } 180 } 181 case htons(ETH_P_TIPC): { 182 struct { 183 __be32 pre[3]; 184 __be32 srcnode; 185 } *hdr, _hdr; 186 hdr = __skb_header_pointer(skb, nhoff, sizeof(_hdr), data, hlen, &_hdr); 187 if (!hdr) 188 return false; 189 flow->src = hdr->srcnode; 190 flow->dst = 0; 191 flow->n_proto = proto; 192 flow->thoff = (u16)nhoff; 193 return true; 194 } 195 case htons(ETH_P_FCOE): 196 flow->thoff = (u16)(nhoff + FCOE_HEADER_LEN); 197 /* fall through */ 198 default: 199 return false; 200 } 201 202 switch (ip_proto) { 203 case IPPROTO_GRE: { 204 struct gre_hdr { 205 __be16 flags; 206 __be16 proto; 207 } *hdr, _hdr; 208 209 hdr = __skb_header_pointer(skb, nhoff, sizeof(_hdr), data, hlen, &_hdr); 210 if (!hdr) 211 return false; 212 /* 213 * Only look inside GRE if version zero and no 214 * routing 215 */ 216 if (!(hdr->flags & (GRE_VERSION|GRE_ROUTING))) { 217 proto = hdr->proto; 218 nhoff += 4; 219 if (hdr->flags & GRE_CSUM) 220 nhoff += 4; 221 if (hdr->flags & GRE_KEY) 222 nhoff += 4; 223 if (hdr->flags & GRE_SEQ) 224 nhoff += 4; 225 if (proto == htons(ETH_P_TEB)) { 226 const struct ethhdr *eth; 227 struct ethhdr _eth; 228 229 eth = __skb_header_pointer(skb, nhoff, 230 sizeof(_eth), 231 data, hlen, &_eth); 232 if (!eth) 233 return false; 234 proto = eth->h_proto; 235 nhoff += sizeof(*eth); 236 237 /* Cap headers that we access via pointers at the 238 * end of the Ethernet header as our maximum alignment 239 * at that point is only 2 bytes. 240 */ 241 if (NET_IP_ALIGN) 242 hlen = nhoff; 243 } 244 goto again; 245 } 246 break; 247 } 248 case IPPROTO_IPIP: 249 proto = htons(ETH_P_IP); 250 goto ip; 251 case IPPROTO_IPV6: 252 proto = htons(ETH_P_IPV6); 253 goto ipv6; 254 default: 255 break; 256 } 257 258 flow->n_proto = proto; 259 flow->ip_proto = ip_proto; 260 flow->thoff = (u16) nhoff; 261 262 /* unless skb is set we don't need to record port info */ 263 if (skb) 264 flow->ports = __skb_flow_get_ports(skb, nhoff, ip_proto, 265 data, hlen); 266 267 return true; 268} 269EXPORT_SYMBOL(__skb_flow_dissect); 270 271static u32 hashrnd __read_mostly; 272static __always_inline void __flow_hash_secret_init(void) 273{ 274 net_get_random_once(&hashrnd, sizeof(hashrnd)); 275} 276 277static __always_inline u32 __flow_hash_3words(u32 a, u32 b, u32 c) 278{ 279 __flow_hash_secret_init(); 280 return jhash_3words(a, b, c, hashrnd); 281} 282 283static inline u32 __flow_hash_from_keys(struct flow_keys *keys) 284{ 285 u32 hash; 286 287 /* get a consistent hash (same value on both flow directions) */ 288 if (((__force u32)keys->dst < (__force u32)keys->src) || 289 (((__force u32)keys->dst == (__force u32)keys->src) && 290 ((__force u16)keys->port16[1] < (__force u16)keys->port16[0]))) { 291 swap(keys->dst, keys->src); 292 swap(keys->port16[0], keys->port16[1]); 293 } 294 295 hash = __flow_hash_3words((__force u32)keys->dst, 296 (__force u32)keys->src, 297 (__force u32)keys->ports); 298 if (!hash) 299 hash = 1; 300 301 return hash; 302} 303 304u32 flow_hash_from_keys(struct flow_keys *keys) 305{ 306 return __flow_hash_from_keys(keys); 307} 308EXPORT_SYMBOL(flow_hash_from_keys); 309 310/* 311 * __skb_get_hash: calculate a flow hash based on src/dst addresses 312 * and src/dst port numbers. Sets hash in skb to non-zero hash value 313 * on success, zero indicates no valid hash. Also, sets l4_hash in skb 314 * if hash is a canonical 4-tuple hash over transport ports. 315 */ 316void __skb_get_hash(struct sk_buff *skb) 317{ 318 struct flow_keys keys; 319 320 if (!skb_flow_dissect(skb, &keys)) 321 return; 322 323 if (keys.ports) 324 skb->l4_hash = 1; 325 326 skb->sw_hash = 1; 327 328 skb->hash = __flow_hash_from_keys(&keys); 329} 330EXPORT_SYMBOL(__skb_get_hash); 331 332/* 333 * Returns a Tx hash based on the given packet descriptor a Tx queues' number 334 * to be used as a distribution range. 335 */ 336u16 __skb_tx_hash(const struct net_device *dev, struct sk_buff *skb, 337 unsigned int num_tx_queues) 338{ 339 u32 hash; 340 u16 qoffset = 0; 341 u16 qcount = num_tx_queues; 342 343 if (skb_rx_queue_recorded(skb)) { 344 hash = skb_get_rx_queue(skb); 345 while (unlikely(hash >= num_tx_queues)) 346 hash -= num_tx_queues; 347 return hash; 348 } 349 350 if (dev->num_tc) { 351 u8 tc = netdev_get_prio_tc_map(dev, skb->priority); 352 qoffset = dev->tc_to_txq[tc].offset; 353 qcount = dev->tc_to_txq[tc].count; 354 } 355 356 return (u16) reciprocal_scale(skb_get_hash(skb), qcount) + qoffset; 357} 358EXPORT_SYMBOL(__skb_tx_hash); 359 360u32 __skb_get_poff(const struct sk_buff *skb, void *data, 361 const struct flow_keys *keys, int hlen) 362{ 363 u32 poff = keys->thoff; 364 365 switch (keys->ip_proto) { 366 case IPPROTO_TCP: { 367 /* access doff as u8 to avoid unaligned access */ 368 const u8 *doff; 369 u8 _doff; 370 371 doff = __skb_header_pointer(skb, poff + 12, sizeof(_doff), 372 data, hlen, &_doff); 373 if (!doff) 374 return poff; 375 376 poff += max_t(u32, sizeof(struct tcphdr), (*doff & 0xF0) >> 2); 377 break; 378 } 379 case IPPROTO_UDP: 380 case IPPROTO_UDPLITE: 381 poff += sizeof(struct udphdr); 382 break; 383 /* For the rest, we do not really care about header 384 * extensions at this point for now. 385 */ 386 case IPPROTO_ICMP: 387 poff += sizeof(struct icmphdr); 388 break; 389 case IPPROTO_ICMPV6: 390 poff += sizeof(struct icmp6hdr); 391 break; 392 case IPPROTO_IGMP: 393 poff += sizeof(struct igmphdr); 394 break; 395 case IPPROTO_DCCP: 396 poff += sizeof(struct dccp_hdr); 397 break; 398 case IPPROTO_SCTP: 399 poff += sizeof(struct sctphdr); 400 break; 401 } 402 403 return poff; 404} 405 406/* skb_get_poff() returns the offset to the payload as far as it could 407 * be dissected. The main user is currently BPF, so that we can dynamically 408 * truncate packets without needing to push actual payload to the user 409 * space and can analyze headers only, instead. 410 */ 411u32 skb_get_poff(const struct sk_buff *skb) 412{ 413 struct flow_keys keys; 414 415 if (!skb_flow_dissect(skb, &keys)) 416 return 0; 417 418 return __skb_get_poff(skb, skb->data, &keys, skb_headlen(skb)); 419} 420 421static inline int get_xps_queue(struct net_device *dev, struct sk_buff *skb) 422{ 423#ifdef CONFIG_XPS 424 struct xps_dev_maps *dev_maps; 425 struct xps_map *map; 426 int queue_index = -1; 427 428 rcu_read_lock(); 429 dev_maps = rcu_dereference(dev->xps_maps); 430 if (dev_maps) { 431 map = rcu_dereference( 432 dev_maps->cpu_map[skb->sender_cpu - 1]); 433 if (map) { 434 if (map->len == 1) 435 queue_index = map->queues[0]; 436 else 437 queue_index = map->queues[reciprocal_scale(skb_get_hash(skb), 438 map->len)]; 439 if (unlikely(queue_index >= dev->real_num_tx_queues)) 440 queue_index = -1; 441 } 442 } 443 rcu_read_unlock(); 444 445 return queue_index; 446#else 447 return -1; 448#endif 449} 450 451static u16 __netdev_pick_tx(struct net_device *dev, struct sk_buff *skb) 452{ 453 struct sock *sk = skb->sk; 454 int queue_index = sk_tx_queue_get(sk); 455 456 if (queue_index < 0 || skb->ooo_okay || 457 queue_index >= dev->real_num_tx_queues) { 458 int new_index = get_xps_queue(dev, skb); 459 if (new_index < 0) 460 new_index = skb_tx_hash(dev, skb); 461 462 if (queue_index != new_index && sk && 463 rcu_access_pointer(sk->sk_dst_cache)) 464 sk_tx_queue_set(sk, new_index); 465 466 queue_index = new_index; 467 } 468 469 return queue_index; 470} 471 472struct netdev_queue *netdev_pick_tx(struct net_device *dev, 473 struct sk_buff *skb, 474 void *accel_priv) 475{ 476 int queue_index = 0; 477 478#ifdef CONFIG_XPS 479 if (skb->sender_cpu == 0) 480 skb->sender_cpu = raw_smp_processor_id() + 1; 481#endif 482 483 if (dev->real_num_tx_queues != 1) { 484 const struct net_device_ops *ops = dev->netdev_ops; 485 if (ops->ndo_select_queue) 486 queue_index = ops->ndo_select_queue(dev, skb, accel_priv, 487 __netdev_pick_tx); 488 else 489 queue_index = __netdev_pick_tx(dev, skb); 490 491 if (!accel_priv) 492 queue_index = netdev_cap_txqueue(dev, queue_index); 493 } 494 495 skb_set_queue_mapping(skb, queue_index); 496 return netdev_get_tx_queue(dev, queue_index); 497} 498