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
  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 }
  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;
  50                 iph_to_flow_copy_addrs(flow, iph);
  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;
  63                 flow->src = (__force __be32)ipv6_addr_hash(&iph->saddr);
  64                 flow->dst = (__force __be32)ipv6_addr_hash(&iph->daddr);
  65                 nhoff += sizeof(struct ipv6hdr);
  66                 break;
  67         }
  68         case __constant_htons(ETH_P_8021AD):
  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;
 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                         }
 138                         goto again;
 139                 }
 140                 break;
 141         }
 142         case IPPROTO_IPIP:
 143                 goto again;
 144         default:
 145                 break;
 146         }
 147
 148         flow->ip_proto = ip_proto;
 149         poff = proto_ports_offset(ip_proto);
 150         if (poff >= 0) {
 151                 __be32 *ports, _ports;
 152
 153                 nhoff += poff;
 154                 ports = skb_header_pointer(skb, nhoff, sizeof(_ports), &_ports);
 155                 if (ports)
 156                         flow->ports = *ports;
 157         }
 158
 159         flow->thoff = (u16) nhoff;
 160
 161         return true;
 162 }
 163 EXPORT_SYMBOL(skb_flow_dissect);
 164
 165 static u32 hashrnd __read_mostly;
 166
 167 /*
 168  * __skb_get_rxhash: calculate a flow hash based on src/dst addresses
 169  * and src/dst port numbers.  Sets rxhash in skb to non-zero hash value
 170  * on success, zero indicates no valid hash.  Also, sets l4_rxhash in skb
 171  * if hash is a canonical 4-tuple hash over transport ports.
 172  */
 173 void __skb_get_rxhash(struct sk_buff *skb)
 174 {
 175         struct flow_keys keys;
 176         u32 hash;
 177
 178         if (!skb_flow_dissect(skb, &keys))
 179                 return;
 180
 181         if (keys.ports)
 182                 skb->l4_rxhash = 1;
 183
 184         /* get a consistent hash (same value on both flow directions) */
 185         if (((__force u32)keys.dst < (__force u32)keys.src) ||
 186             (((__force u32)keys.dst == (__force u32)keys.src) &&
 187              ((__force u16)keys.port16[1] < (__force u16)keys.port16[0]))) {
 188                 swap(keys.dst, keys.src);
 189                 swap(keys.port16[0], keys.port16[1]);
 190         }
 191
 192         hash = jhash_3words((__force u32)keys.dst,
 193                             (__force u32)keys.src,
 194                             (__force u32)keys.ports, hashrnd);
 195         if (!hash)
 196                 hash = 1;
 197
 198         skb->rxhash = hash;
 199 }
 200 EXPORT_SYMBOL(__skb_get_rxhash);
 201
 202 /*
 203  * Returns a Tx hash based on the given packet descriptor a Tx queues' number
 204  * to be used as a distribution range.
 205  */
 206 u16 __skb_tx_hash(const struct net_device *dev, const struct sk_buff *skb,
 207                   unsigned int num_tx_queues)
 208 {
 209         u32 hash;
 210         u16 qoffset = 0;
 211         u16 qcount = num_tx_queues;
 212
 213         if (skb_rx_queue_recorded(skb)) {
 214                 hash = skb_get_rx_queue(skb);
 215                 while (unlikely(hash >= num_tx_queues))
 216                         hash -= num_tx_queues;
 217                 return hash;
 218         }
 219
 220         if (dev->num_tc) {
 221                 u8 tc = netdev_get_prio_tc_map(dev, skb->priority);
 222                 qoffset = dev->tc_to_txq[tc].offset;
 223                 qcount = dev->tc_to_txq[tc].count;
 224         }
 225
 226         if (skb->sk && skb->sk->sk_hash)
 227                 hash = skb->sk->sk_hash;
 228         else
 229                 hash = (__force u16) skb->protocol;
 230         hash = jhash_1word(hash, hashrnd);
 231
 232         return (u16) (((u64) hash * qcount) >> 32) + qoffset;
 233 }
 234 EXPORT_SYMBOL(__skb_tx_hash);
 235
 236 /* __skb_get_poff() returns the offset to the payload as far as it could
 237  * be dissected. The main user is currently BPF, so that we can dynamically
 238  * truncate packets without needing to push actual payload to the user
 239  * space and can analyze headers only, instead.
 240  */
 241 u32 __skb_get_poff(const struct sk_buff *skb)
 242 {
 243         struct flow_keys keys;
 244         u32 poff = 0;
 245
 246         if (!skb_flow_dissect(skb, &keys))
 247                 return 0;
 248
 249         poff += keys.thoff;
 250         switch (keys.ip_proto) {
 251         case IPPROTO_TCP: {
 252                 const struct tcphdr *tcph;
 253                 struct tcphdr _tcph;
 254
 255                 tcph = skb_header_pointer(skb, poff, sizeof(_tcph), &_tcph);
 256                 if (!tcph)
 257                         return poff;
 258
 259                 poff += max_t(u32, sizeof(struct tcphdr), tcph->doff * 4);
 260                 break;
 261         }
 262         case IPPROTO_UDP:
 263         case IPPROTO_UDPLITE:
 264                 poff += sizeof(struct udphdr);
 265                 break;
 266         /* For the rest, we do not really care about header
 267          * extensions at this point for now.
 268          */
 269         case IPPROTO_ICMP:
 270                 poff += sizeof(struct icmphdr);
 271                 break;
 272         case IPPROTO_ICMPV6:
 273                 poff += sizeof(struct icmp6hdr);
 274                 break;
 275         case IPPROTO_IGMP:
 276                 poff += sizeof(struct igmphdr);
 277                 break;
 278         case IPPROTO_DCCP:
 279                 poff += sizeof(struct dccp_hdr);
 280                 break;
 281         case IPPROTO_SCTP:
 282                 poff += sizeof(struct sctphdr);
 283                 break;
 284         }
 285
 286         return poff;
 287 }
 288
 289 static inline u16 dev_cap_txqueue(struct net_device *dev, u16 queue_index)
 290 {
 291         if (unlikely(queue_index >= dev->real_num_tx_queues)) {
 292                 net_warn_ratelimited("%s selects TX queue %d, but real number of TX queues is %d\n",
 293                                      dev->name, queue_index,
 294                                      dev->real_num_tx_queues);
 295                 return 0;
 296         }
 297         return queue_index;
 298 }
 299
 300 static inline int get_xps_queue(struct net_device *dev, struct sk_buff *skb)
 301 {
 302 #ifdef CONFIG_XPS
 303         struct xps_dev_maps *dev_maps;
 304         struct xps_map *map;
 305         int queue_index = -1;
 306
 307         rcu_read_lock();
 308         dev_maps = rcu_dereference(dev->xps_maps);
 309         if (dev_maps) {
 310                 map = rcu_dereference(
 311                     dev_maps->cpu_map[raw_smp_processor_id()]);
 312                 if (map) {
 313                         if (map->len == 1)
 314                                 queue_index = map->queues[0];
 315                         else {
 316                                 u32 hash;
 317                                 if (skb->sk && skb->sk->sk_hash)
 318                                         hash = skb->sk->sk_hash;
 319                                 else
 320                                         hash = (__force u16) skb->protocol ^
 321                                             skb->rxhash;
 322                                 hash = jhash_1word(hash, hashrnd);
 323                                 queue_index = map->queues[
 324                                     ((u64)hash * map->len) >> 32];
 325                         }
 326                         if (unlikely(queue_index >= dev->real_num_tx_queues))
 327                                 queue_index = -1;
 328                 }
 329         }
 330         rcu_read_unlock();
 331
 332         return queue_index;
 333 #else
 334         return -1;
 335 #endif
 336 }
 337
 338 u16 __netdev_pick_tx(struct net_device *dev, struct sk_buff *skb)
 339 {
 340         struct sock *sk = skb->sk;
 341         int queue_index = sk_tx_queue_get(sk);
 342
 343         if (queue_index < 0 || skb->ooo_okay ||
 344             queue_index >= dev->real_num_tx_queues) {
 345                 int new_index = get_xps_queue(dev, skb);
 346                 if (new_index < 0)
 347                         new_index = skb_tx_hash(dev, skb);
 348
 349                 if (queue_index != new_index && sk &&
 350                     rcu_access_pointer(sk->sk_dst_cache))
 351                         sk_tx_queue_set(sk, queue_index);
 352
 353                 queue_index = new_index;
 354         }
 355
 356         return queue_index;
 357 }
 358 EXPORT_SYMBOL(__netdev_pick_tx);
 359
 360 struct netdev_queue *netdev_pick_tx(struct net_device *dev,
 361                                     struct sk_buff *skb)
 362 {
 363         int queue_index = 0;
 364
 365         if (dev->real_num_tx_queues != 1) {
 366                 const struct net_device_ops *ops = dev->netdev_ops;
 367                 if (ops->ndo_select_queue)
 368                         queue_index = ops->ndo_select_queue(dev, skb);
 369                 else
 370                         queue_index = __netdev_pick_tx(dev, skb);
 371                 queue_index = dev_cap_txqueue(dev, queue_index);
 372         }
 373
 374         skb_set_queue_mapping(skb, queue_index);
 375         return netdev_get_tx_queue(dev, queue_index);
 376 }
 377
 378 static int __init initialize_hashrnd(void)
 379 {
 380         get_random_bytes(&hashrnd, sizeof(hashrnd));
 381         return 0;
 382 }
 383
 384 late_initcall_sync(initialize_hashrnd);