datapath/actions.c

   1 /*
   2  * Distributed under the terms of the GNU GPL version 2.
   3  * Copyright (c) 2007, 2008, 2009, 2010, 2011 Nicira Networks.
   4  *
   5  * Significant portions of this file may be copied from parts of the Linux
   6  * kernel, by Linus Torvalds and others.
   7  */
   8
   9 /* Functions for executing flow actions. */
  10
  11 #include <linux/skbuff.h>
  12 #include <linux/in.h>
  13 #include <linux/ip.h>
  14 #include <linux/tcp.h>
  15 #include <linux/udp.h>
  16 #include <linux/in6.h>
  17 #include <linux/if_arp.h>
  18 #include <linux/if_vlan.h>
  19 #include <net/inet_ecn.h>
  20 #include <net/ip.h>
  21 #include <net/checksum.h>
  22
  23 #include "actions.h"
  24 #include "checksum.h"
  25 #include "datapath.h"
  26 #include "openvswitch/datapath-protocol.h"
  27 #include "vlan.h"
  28 #include "vport.h"
  29
  30 static int do_execute_actions(struct datapath *, struct sk_buff *,
  31                               const struct sw_flow_key *,
  32                               const struct nlattr *actions, u32 actions_len);
  33
  34 static struct sk_buff *make_writable(struct sk_buff *skb, unsigned min_headroom)
  35 {
  36         if (skb_cloned(skb)) {
  37                 struct sk_buff *nskb;
  38                 unsigned headroom = max(min_headroom, skb_headroom(skb));
  39
  40                 nskb = skb_copy_expand(skb, headroom, skb_tailroom(skb), GFP_ATOMIC);
  41                 if (nskb) {
  42                         set_skb_csum_bits(skb, nskb);
  43                         kfree_skb(skb);
  44                         return nskb;
  45                 }
  46         } else {
  47                 unsigned int hdr_len = (skb_transport_offset(skb)
  48                                         + sizeof(struct tcphdr));
  49                 if (pskb_may_pull(skb, min(hdr_len, skb->len)))
  50                         return skb;
  51         }
  52         kfree_skb(skb);
  53         return NULL;
  54 }
  55
  56 static struct sk_buff *strip_vlan(struct sk_buff *skb)
  57 {
  58         struct ethhdr *eh;
  59
  60         if (vlan_tx_tag_present(skb)) {
  61                 vlan_set_tci(skb, 0);
  62                 return skb;
  63         }
  64
  65         if (unlikely(vlan_eth_hdr(skb)->h_vlan_proto != htons(ETH_P_8021Q) ||
  66             skb->len < VLAN_ETH_HLEN))
  67                 return skb;
  68
  69         skb = make_writable(skb, 0);
  70         if (unlikely(!skb))
  71                 return NULL;
  72
  73         if (get_ip_summed(skb) == OVS_CSUM_COMPLETE)
  74                 skb->csum = csum_sub(skb->csum, csum_partial(skb->data
  75                                         + ETH_HLEN, VLAN_HLEN, 0));
  76
  77         memmove(skb->data + VLAN_HLEN, skb->data, 2 * ETH_ALEN);
  78
  79         eh = (struct ethhdr *)skb_pull(skb, VLAN_HLEN);
  80
  81         skb->protocol = eh->h_proto;
  82         skb->mac_header += VLAN_HLEN;
  83
  84         return skb;
  85 }
  86
  87 static struct sk_buff *modify_vlan_tci(struct datapath *dp, struct sk_buff *skb,
  88                                        const struct sw_flow_key *key,
  89                                        const struct nlattr *a, u32 actions_len)
  90 {
  91         __be16 tci = nla_get_be16(a);
  92         struct vlan_ethhdr *vh;
  93         __be16 old_tci;
  94
  95         if (vlan_tx_tag_present(skb) || skb->protocol != htons(ETH_P_8021Q))
  96                 return __vlan_hwaccel_put_tag(skb, ntohs(tci));
  97
  98         skb = make_writable(skb, 0);
  99         if (unlikely(!skb))
 100                 return NULL;
 101
 102         if (unlikely(skb->len < VLAN_ETH_HLEN))
 103                 return skb;
 104
 105         vh = vlan_eth_hdr(skb);
 106
 107         old_tci = vh->h_vlan_TCI;
 108         vh->h_vlan_TCI = tci;
 109
 110         if (get_ip_summed(skb) == OVS_CSUM_COMPLETE) {
 111                 __be16 diff[] = { ~old_tci, vh->h_vlan_TCI };
 112                 skb->csum = ~csum_partial((char *)diff, sizeof(diff), ~skb->csum);
 113         }
 114
 115         return skb;
 116 }
 117
 118 static bool is_ip(struct sk_buff *skb, const struct sw_flow_key *key)
 119 {
 120         return (key->dl_type == htons(ETH_P_IP) &&
 121                 skb->transport_header > skb->network_header);
 122 }
 123
 124 static __sum16 *get_l4_checksum(struct sk_buff *skb, const struct sw_flow_key *key)
 125 {
 126         int transport_len = skb->len - skb_transport_offset(skb);
 127         if (key->nw_proto == IPPROTO_TCP) {
 128                 if (likely(transport_len >= sizeof(struct tcphdr)))
 129                         return &tcp_hdr(skb)->check;
 130         } else if (key->nw_proto == IPPROTO_UDP) {
 131                 if (likely(transport_len >= sizeof(struct udphdr)))
 132                         return &udp_hdr(skb)->check;
 133         }
 134         return NULL;
 135 }
 136
 137 static struct sk_buff *set_nw_addr(struct sk_buff *skb,
 138                                    const struct sw_flow_key *key,
 139                                    const struct nlattr *a)
 140 {
 141         __be32 new_nwaddr = nla_get_be32(a);
 142         struct iphdr *nh;
 143         __sum16 *check;
 144         __be32 *nwaddr;
 145
 146         if (unlikely(!is_ip(skb, key)))
 147                 return skb;
 148
 149         skb = make_writable(skb, 0);
 150         if (unlikely(!skb))
 151                 return NULL;
 152
 153         nh = ip_hdr(skb);
 154         nwaddr = nla_type(a) == ODP_ACTION_ATTR_SET_NW_SRC ? &nh->saddr : &nh->daddr;
 155
 156         check = get_l4_checksum(skb, key);
 157         if (likely(check))
 158                 inet_proto_csum_replace4(check, skb, *nwaddr, new_nwaddr, 1);
 159         csum_replace4(&nh->check, *nwaddr, new_nwaddr);
 160
 161         skb_clear_rxhash(skb);
 162
 163         *nwaddr = new_nwaddr;
 164
 165         return skb;
 166 }
 167
 168 static struct sk_buff *set_nw_tos(struct sk_buff *skb,
 169                                   const struct sw_flow_key *key,
 170                                   u8 nw_tos)
 171 {
 172         if (unlikely(!is_ip(skb, key)))
 173                 return skb;
 174
 175         skb = make_writable(skb, 0);
 176         if (skb) {
 177                 struct iphdr *nh = ip_hdr(skb);
 178                 u8 *f = &nh->tos;
 179                 u8 old = *f;
 180                 u8 new;
 181
 182                 /* Set the DSCP bits and preserve the ECN bits. */
 183                 new = nw_tos | (nh->tos & INET_ECN_MASK);
 184                 csum_replace4(&nh->check, (__force __be32)old,
 185                                           (__force __be32)new);
 186                 *f = new;
 187         }
 188         return skb;
 189 }
 190
 191 static struct sk_buff *set_tp_port(struct sk_buff *skb,
 192                                    const struct sw_flow_key *key,
 193                                    const struct nlattr *a)
 194 {
 195         struct udphdr *th;
 196         __sum16 *check;
 197         __be16 *port;
 198
 199         if (unlikely(!is_ip(skb, key)))
 200                 return skb;
 201
 202         skb = make_writable(skb, 0);
 203         if (unlikely(!skb))
 204                 return NULL;
 205
 206         /* Must follow make_writable() since that can move the skb data. */
 207         check = get_l4_checksum(skb, key);
 208         if (unlikely(!check))
 209                 return skb;
 210
 211         /*
 212          * Update port and checksum.
 213          *
 214          * This is OK because source and destination port numbers are at the
 215          * same offsets in both UDP and TCP headers, and get_l4_checksum() only
 216          * supports those protocols.
 217          */
 218         th = udp_hdr(skb);
 219         port = nla_type(a) == ODP_ACTION_ATTR_SET_TP_SRC ? &th->source : &th->dest;
 220         inet_proto_csum_replace2(check, skb, *port, nla_get_be16(a), 0);
 221         *port = nla_get_be16(a);
 222         skb_clear_rxhash(skb);
 223
 224         return skb;
 225 }
 226
 227 /**
 228  * is_spoofed_arp - check for invalid ARP packet
 229  *
 230  * @skb: skbuff containing an Ethernet packet, with network header pointing
 231  * just past the Ethernet and optional 802.1Q header.
 232  * @key: flow key extracted from @skb by flow_extract()
 233  *
 234  * Returns true if @skb is an invalid Ethernet+IPv4 ARP packet: one with screwy
 235  * or truncated header fields or one whose inner and outer Ethernet address
 236  * differ.
 237  */
 238 static bool is_spoofed_arp(struct sk_buff *skb, const struct sw_flow_key *key)
 239 {
 240         struct arp_eth_header *arp;
 241
 242         if (key->dl_type != htons(ETH_P_ARP))
 243                 return false;
 244
 245         if (skb_network_offset(skb) + sizeof(struct arp_eth_header) > skb->len)
 246                 return true;
 247
 248         arp = (struct arp_eth_header *)skb_network_header(skb);
 249         return (arp->ar_hrd != htons(ARPHRD_ETHER) ||
 250                 arp->ar_pro != htons(ETH_P_IP) ||
 251                 arp->ar_hln != ETH_ALEN ||
 252                 arp->ar_pln != 4 ||
 253                 compare_ether_addr(arp->ar_sha, eth_hdr(skb)->h_source));
 254 }
 255
 256 static void do_output(struct datapath *dp, struct sk_buff *skb, int out_port)
 257 {
 258         struct vport *p;
 259
 260         if (!skb)
 261                 goto error;
 262
 263         p = rcu_dereference(dp->ports[out_port]);
 264         if (!p)
 265                 goto error;
 266
 267         vport_send(p, skb);
 268         return;
 269
 270 error:
 271         kfree_skb(skb);
 272 }
 273
 274 static int output_control(struct datapath *dp, struct sk_buff *skb, u64 arg,
 275                           const struct sw_flow_key *key)
 276 {
 277         struct dp_upcall_info upcall;
 278
 279         skb = skb_clone(skb, GFP_ATOMIC);
 280         if (!skb)
 281                 return -ENOMEM;
 282
 283         upcall.cmd = ODP_PACKET_CMD_ACTION;
 284         upcall.key = key;
 285         upcall.userdata = arg;
 286         upcall.sample_pool = 0;
 287         upcall.actions = NULL;
 288         upcall.actions_len = 0;
 289         return dp_upcall(dp, skb, &upcall);
 290 }
 291
 292 /* Execute a list of actions against 'skb'. */
 293 static int do_execute_actions(struct datapath *dp, struct sk_buff *skb,
 294                               const struct sw_flow_key *key,
 295                               const struct nlattr *actions, u32 actions_len)
 296 {
 297         /* Every output action needs a separate clone of 'skb', but the common
 298          * case is just a single output action, so that doing a clone and
 299          * then freeing the original skbuff is wasteful.  So the following code
 300          * is slightly obscure just to avoid that. */
 301         int prev_port = -1;
 302         u32 priority = skb->priority;
 303         const struct nlattr *a;
 304         int rem, err;
 305
 306         for (a = actions, rem = actions_len; rem > 0; a = nla_next(a, &rem)) {
 307                 if (prev_port != -1) {
 308                         do_output(dp, skb_clone(skb, GFP_ATOMIC), prev_port);
 309                         prev_port = -1;
 310                 }
 311
 312                 switch (nla_type(a)) {
 313                 case ODP_ACTION_ATTR_OUTPUT:
 314                         prev_port = nla_get_u32(a);
 315                         break;
 316
 317                 case ODP_ACTION_ATTR_CONTROLLER:
 318                         err = output_control(dp, skb, nla_get_u64(a), key);
 319                         if (err) {
 320                                 kfree_skb(skb);
 321                                 return err;
 322                         }
 323                         break;
 324
 325                 case ODP_ACTION_ATTR_SET_TUNNEL:
 326                         OVS_CB(skb)->tun_id = nla_get_be64(a);
 327                         break;
 328
 329                 case ODP_ACTION_ATTR_SET_DL_TCI:
 330                         skb = modify_vlan_tci(dp, skb, key, a, rem);
 331                         break;
 332
 333                 case ODP_ACTION_ATTR_STRIP_VLAN:
 334                         skb = strip_vlan(skb);
 335                         break;
 336
 337                 case ODP_ACTION_ATTR_SET_DL_SRC:
 338                         skb = make_writable(skb, 0);
 339                         if (!skb)
 340                                 return -ENOMEM;
 341                         memcpy(eth_hdr(skb)->h_source, nla_data(a), ETH_ALEN);
 342                         break;
 343
 344                 case ODP_ACTION_ATTR_SET_DL_DST:
 345                         skb = make_writable(skb, 0);
 346                         if (!skb)
 347                                 return -ENOMEM;
 348                         memcpy(eth_hdr(skb)->h_dest, nla_data(a), ETH_ALEN);
 349                         break;
 350
 351                 case ODP_ACTION_ATTR_SET_NW_SRC:
 352                 case ODP_ACTION_ATTR_SET_NW_DST:
 353                         skb = set_nw_addr(skb, key, a);
 354                         break;
 355
 356                 case ODP_ACTION_ATTR_SET_NW_TOS:
 357                         skb = set_nw_tos(skb, key, nla_get_u8(a));
 358                         break;
 359
 360                 case ODP_ACTION_ATTR_SET_TP_SRC:
 361                 case ODP_ACTION_ATTR_SET_TP_DST:
 362                         skb = set_tp_port(skb, key, a);
 363                         break;
 364
 365                 case ODP_ACTION_ATTR_SET_PRIORITY:
 366                         skb->priority = nla_get_u32(a);
 367                         break;
 368
 369                 case ODP_ACTION_ATTR_POP_PRIORITY:
 370                         skb->priority = priority;
 371                         break;
 372
 373                 case ODP_ACTION_ATTR_DROP_SPOOFED_ARP:
 374                         if (unlikely(is_spoofed_arp(skb, key)))
 375                                 goto exit;
 376                         break;
 377                 }
 378                 if (!skb)
 379                         return -ENOMEM;
 380         }
 381 exit:
 382         if (prev_port != -1)
 383                 do_output(dp, skb, prev_port);
 384         else
 385                 kfree_skb(skb);
 386         return 0;
 387 }
 388
 389 static void sflow_sample(struct datapath *dp, struct sk_buff *skb,
 390                          const struct sw_flow_key *key,
 391                          const struct nlattr *a, u32 actions_len)
 392 {
 393         struct sk_buff *nskb;
 394         struct vport *p = OVS_CB(skb)->vport;
 395         struct dp_upcall_info upcall;
 396
 397         if (unlikely(!p))
 398                 return;
 399
 400         atomic_inc(&p->sflow_pool);
 401         if (net_random() >= dp->sflow_probability)
 402                 return;
 403
 404         nskb = skb_clone(skb, GFP_ATOMIC);
 405         if (unlikely(!nskb))
 406                 return;
 407
 408         upcall.cmd = ODP_PACKET_CMD_SAMPLE;
 409         upcall.key = key;
 410         upcall.userdata = 0;
 411         upcall.sample_pool = atomic_read(&p->sflow_pool);
 412         upcall.actions = a;
 413         upcall.actions_len = actions_len;
 414         dp_upcall(dp, nskb, &upcall);
 415 }
 416
 417 /* Execute a list of actions against 'skb'. */
 418 int execute_actions(struct datapath *dp, struct sk_buff *skb,
 419                     const struct sw_flow_key *key,
 420                     const struct nlattr *actions, u32 actions_len)
 421 {
 422         if (dp->sflow_probability)
 423                 sflow_sample(dp, skb, key, actions, actions_len);
 424
 425         OVS_CB(skb)->tun_id = 0;
 426
 427         return do_execute_actions(dp, skb, key, actions, actions_len);
 428 }