kernel/net/core/flow_dissector.c

   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  */
  22 static 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 }
  61 EXPORT_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  */
  74 bool __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
  88 again:
  89         switch (proto) {
  90         case htons(ETH_P_IP): {
  91                 const struct iphdr *iph;
  92                 struct iphdr _iph;
  93 ip:
  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                 __be32 flow_label;
 117
 118 ipv6:
 119                 iph = __skb_header_pointer(skb, nhoff, sizeof(_iph), data, hlen, &_iph);
 120                 if (!iph)
 121                         return false;
 122
 123                 ip_proto = iph->nexthdr;
 124                 nhoff += sizeof(struct ipv6hdr);
 125
 126                 /* see comment above in IPv4 section */
 127                 if (!skb)
 128                         break;
 129
 130                 flow->src = (__force __be32)ipv6_addr_hash(&iph->saddr);
 131                 flow->dst = (__force __be32)ipv6_addr_hash(&iph->daddr);
 132
 133                 flow_label = ip6_flowlabel(iph);
 134                 if (flow_label) {
 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                         goto again;
 238                 }
 239                 break;
 240         }
 241         case IPPROTO_IPIP:
 242                 proto = htons(ETH_P_IP);
 243                 goto ip;
 244         case IPPROTO_IPV6:
 245                 proto = htons(ETH_P_IPV6);
 246                 goto ipv6;
 247         default:
 248                 break;
 249         }
 250
 251         flow->n_proto = proto;
 252         flow->ip_proto = ip_proto;
 253         flow->thoff = (u16) nhoff;
 254
 255         /* unless skb is set we don't need to record port info */
 256         if (skb)
 257                 flow->ports = __skb_flow_get_ports(skb, nhoff, ip_proto,
 258                                                    data, hlen);
 259
 260         return true;
 261 }
 262 EXPORT_SYMBOL(__skb_flow_dissect);
 263
 264 static u32 hashrnd __read_mostly;
 265 static __always_inline void __flow_hash_secret_init(void)
 266 {
 267         net_get_random_once(&hashrnd, sizeof(hashrnd));
 268 }
 269
 270 static __always_inline u32 __flow_hash_3words(u32 a, u32 b, u32 c)
 271 {
 272         __flow_hash_secret_init();
 273         return jhash_3words(a, b, c, hashrnd);
 274 }
 275
 276 static inline u32 __flow_hash_from_keys(struct flow_keys *keys)
 277 {
 278         u32 hash;
 279
 280         /* get a consistent hash (same value on both flow directions) */
 281         if (((__force u32)keys->dst < (__force u32)keys->src) ||
 282             (((__force u32)keys->dst == (__force u32)keys->src) &&
 283              ((__force u16)keys->port16[1] < (__force u16)keys->port16[0]))) {
 284                 swap(keys->dst, keys->src);
 285                 swap(keys->port16[0], keys->port16[1]);
 286         }
 287
 288         hash = __flow_hash_3words((__force u32)keys->dst,
 289                                   (__force u32)keys->src,
 290                                   (__force u32)keys->ports);
 291         if (!hash)
 292                 hash = 1;
 293
 294         return hash;
 295 }
 296
 297 u32 flow_hash_from_keys(struct flow_keys *keys)
 298 {
 299         return __flow_hash_from_keys(keys);
 300 }
 301 EXPORT_SYMBOL(flow_hash_from_keys);
 302
 303 /*
 304  * __skb_get_hash: calculate a flow hash based on src/dst addresses
 305  * and src/dst port numbers.  Sets hash in skb to non-zero hash value
 306  * on success, zero indicates no valid hash.  Also, sets l4_hash in skb
 307  * if hash is a canonical 4-tuple hash over transport ports.
 308  */
 309 void __skb_get_hash(struct sk_buff *skb)
 310 {
 311         struct flow_keys keys;
 312
 313         if (!skb_flow_dissect(skb, &keys))
 314                 return;
 315
 316         if (keys.ports)
 317                 skb->l4_hash = 1;
 318
 319         skb->sw_hash = 1;
 320
 321         skb->hash = __flow_hash_from_keys(&keys);
 322 }
 323 EXPORT_SYMBOL(__skb_get_hash);
 324
 325 /*
 326  * Returns a Tx hash based on the given packet descriptor a Tx queues' number
 327  * to be used as a distribution range.
 328  */
 329 u16 __skb_tx_hash(const struct net_device *dev, struct sk_buff *skb,
 330                   unsigned int num_tx_queues)
 331 {
 332         u32 hash;
 333         u16 qoffset = 0;
 334         u16 qcount = num_tx_queues;
 335
 336         if (skb_rx_queue_recorded(skb)) {
 337                 hash = skb_get_rx_queue(skb);
 338                 while (unlikely(hash >= num_tx_queues))
 339                         hash -= num_tx_queues;
 340                 return hash;
 341         }
 342
 343         if (dev->num_tc) {
 344                 u8 tc = netdev_get_prio_tc_map(dev, skb->priority);
 345                 qoffset = dev->tc_to_txq[tc].offset;
 346                 qcount = dev->tc_to_txq[tc].count;
 347         }
 348
 349         return (u16) reciprocal_scale(skb_get_hash(skb), qcount) + qoffset;
 350 }
 351 EXPORT_SYMBOL(__skb_tx_hash);
 352
 353 u32 __skb_get_poff(const struct sk_buff *skb, void *data,
 354                    const struct flow_keys *keys, int hlen)
 355 {
 356         u32 poff = keys->thoff;
 357
 358         switch (keys->ip_proto) {
 359         case IPPROTO_TCP: {
 360                 /* access doff as u8 to avoid unaligned access */
 361                 const u8 *doff;
 362                 u8 _doff;
 363
 364                 doff = __skb_header_pointer(skb, poff + 12, sizeof(_doff),
 365                                             data, hlen, &_doff);
 366                 if (!doff)
 367                         return poff;
 368
 369                 poff += max_t(u32, sizeof(struct tcphdr), (*doff & 0xF0) >> 2);
 370                 break;
 371         }
 372         case IPPROTO_UDP:
 373         case IPPROTO_UDPLITE:
 374                 poff += sizeof(struct udphdr);
 375                 break;
 376         /* For the rest, we do not really care about header
 377          * extensions at this point for now.
 378          */
 379         case IPPROTO_ICMP:
 380                 poff += sizeof(struct icmphdr);
 381                 break;
 382         case IPPROTO_ICMPV6:
 383                 poff += sizeof(struct icmp6hdr);
 384                 break;
 385         case IPPROTO_IGMP:
 386                 poff += sizeof(struct igmphdr);
 387                 break;
 388         case IPPROTO_DCCP:
 389                 poff += sizeof(struct dccp_hdr);
 390                 break;
 391         case IPPROTO_SCTP:
 392                 poff += sizeof(struct sctphdr);
 393                 break;
 394         }
 395
 396         return poff;
 397 }
 398
 399 /* skb_get_poff() returns the offset to the payload as far as it could
 400  * be dissected. The main user is currently BPF, so that we can dynamically
 401  * truncate packets without needing to push actual payload to the user
 402  * space and can analyze headers only, instead.
 403  */
 404 u32 skb_get_poff(const struct sk_buff *skb)
 405 {
 406         struct flow_keys keys;
 407
 408         if (!skb_flow_dissect(skb, &keys))
 409                 return 0;
 410
 411         return __skb_get_poff(skb, skb->data, &keys, skb_headlen(skb));
 412 }
 413
 414 static inline int get_xps_queue(struct net_device *dev, struct sk_buff *skb)
 415 {
 416 #ifdef CONFIG_XPS
 417         struct xps_dev_maps *dev_maps;
 418         struct xps_map *map;
 419         int queue_index = -1;
 420
 421         rcu_read_lock();
 422         dev_maps = rcu_dereference(dev->xps_maps);
 423         if (dev_maps) {
 424                 map = rcu_dereference(
 425                     dev_maps->cpu_map[skb->sender_cpu - 1]);
 426                 if (map) {
 427                         if (map->len == 1)
 428                                 queue_index = map->queues[0];
 429                         else
 430                                 queue_index = map->queues[reciprocal_scale(skb_get_hash(skb),
 431                                                                            map->len)];
 432                         if (unlikely(queue_index >= dev->real_num_tx_queues))
 433                                 queue_index = -1;
 434                 }
 435         }
 436         rcu_read_unlock();
 437
 438         return queue_index;
 439 #else
 440         return -1;
 441 #endif
 442 }
 443
 444 static u16 __netdev_pick_tx(struct net_device *dev, struct sk_buff *skb)
 445 {
 446         struct sock *sk = skb->sk;
 447         int queue_index = sk_tx_queue_get(sk);
 448
 449         if (queue_index < 0 || skb->ooo_okay ||
 450             queue_index >= dev->real_num_tx_queues) {
 451                 int new_index = get_xps_queue(dev, skb);
 452                 if (new_index < 0)
 453                         new_index = skb_tx_hash(dev, skb);
 454
 455                 if (queue_index != new_index && sk &&
 456                     rcu_access_pointer(sk->sk_dst_cache))
 457                         sk_tx_queue_set(sk, new_index);
 458
 459                 queue_index = new_index;
 460         }
 461
 462         return queue_index;
 463 }
 464
 465 struct netdev_queue *netdev_pick_tx(struct net_device *dev,
 466                                     struct sk_buff *skb,
 467                                     void *accel_priv)
 468 {
 469         int queue_index = 0;
 470
 471 #ifdef CONFIG_XPS
 472         if (skb->sender_cpu == 0)
 473                 skb->sender_cpu = raw_smp_processor_id() + 1;
 474 #endif
 475
 476         if (dev->real_num_tx_queues != 1) {
 477                 const struct net_device_ops *ops = dev->netdev_ops;
 478                 if (ops->ndo_select_queue)
 479                         queue_index = ops->ndo_select_queue(dev, skb, accel_priv,
 480                                                             __netdev_pick_tx);
 481                 else
 482                         queue_index = __netdev_pick_tx(dev, skb);
 483
 484                 if (!accel_priv)
 485                         queue_index = netdev_cap_txqueue(dev, queue_index);
 486         }
 487
 488         skb_set_queue_mapping(skb, queue_index);
 489         return netdev_get_tx_queue(dev, queue_index);
 490 }