datapath/actions.c

   1 /*
   2  * Distributed under the terms of the GNU GPL version 2.
   3  * Copyright (c) 2007, 2008, 2009, 2010 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 "vport.h"
  28
  29 static struct sk_buff *make_writable(struct sk_buff *skb, unsigned min_headroom)
  30 {
  31         if (skb_cloned(skb)) {
  32                 struct sk_buff *nskb;
  33                 unsigned headroom = max(min_headroom, skb_headroom(skb));
  34
  35                 nskb = skb_copy_expand(skb, headroom, skb_tailroom(skb), GFP_ATOMIC);
  36                 if (nskb) {
  37                         set_skb_csum_bits(skb, nskb);
  38                         kfree_skb(skb);
  39                         return nskb;
  40                 }
  41         } else {
  42                 unsigned int hdr_len = (skb_transport_offset(skb)
  43                                         + sizeof(struct tcphdr));
  44                 if (pskb_may_pull(skb, min(hdr_len, skb->len)))
  45                         return skb;
  46         }
  47         kfree_skb(skb);
  48         return NULL;
  49 }
  50
  51 static struct sk_buff *vlan_pull_tag(struct sk_buff *skb)
  52 {
  53         struct vlan_ethhdr *vh = vlan_eth_hdr(skb);
  54         struct ethhdr *eh;
  55
  56         /* Verify we were given a vlan packet */
  57         if (vh->h_vlan_proto != htons(ETH_P_8021Q) || skb->len < VLAN_ETH_HLEN)
  58                 return skb;
  59
  60         if (get_ip_summed(skb) == OVS_CSUM_COMPLETE)
  61                 skb->csum = csum_sub(skb->csum, csum_partial(skb->data
  62                                         + ETH_HLEN, VLAN_HLEN, 0));
  63
  64         memmove(skb->data + VLAN_HLEN, skb->data, 2 * VLAN_ETH_ALEN);
  65
  66         eh = (struct ethhdr *)skb_pull(skb, VLAN_HLEN);
  67
  68         skb->protocol = eh->h_proto;
  69         skb->mac_header += VLAN_HLEN;
  70
  71         return skb;
  72 }
  73
  74 static struct sk_buff *modify_vlan_tci(struct datapath *dp, struct sk_buff *skb,
  75                                        const struct odp_flow_key *key,
  76                                        const struct nlattr *a, u32 actions_len)
  77 {
  78         __be16 tci = nla_get_be16(a);
  79
  80         skb = make_writable(skb, VLAN_HLEN);
  81         if (!skb)
  82                 return ERR_PTR(-ENOMEM);
  83
  84         if (skb->protocol == htons(ETH_P_8021Q)) {
  85                 /* Modify vlan id, but maintain other TCI values */
  86                 struct vlan_ethhdr *vh;
  87                 __be16 old_tci;
  88
  89                 if (skb->len < VLAN_ETH_HLEN)
  90                         return skb;
  91
  92                 vh = vlan_eth_hdr(skb);
  93                 old_tci = vh->h_vlan_TCI;
  94
  95                 vh->h_vlan_TCI = tci;
  96
  97                 if (get_ip_summed(skb) == OVS_CSUM_COMPLETE) {
  98                         __be16 diff[] = { ~old_tci, vh->h_vlan_TCI };
  99
 100                         skb->csum = ~csum_partial((char *)diff, sizeof(diff),
 101                                                 ~skb->csum);
 102                 }
 103         } else {
 104                 int err;
 105
 106                 /* Add vlan header */
 107
 108                 /* Set up checksumming pointers for checksum-deferred packets
 109                  * on Xen.  Otherwise, dev_queue_xmit() will try to do this
 110                  * when we send the packet out on the wire, and it will fail at
 111                  * that point because skb_checksum_setup() will not look inside
 112                  * an 802.1Q header. */
 113                 err = vswitch_skb_checksum_setup(skb);
 114                 if (unlikely(err)) {
 115                         kfree_skb(skb);
 116                         return ERR_PTR(err);
 117                 }
 118
 119                 /* GSO is not implemented for packets with an 802.1Q header, so
 120                  * we have to do segmentation before we add that header.
 121                  *
 122                  * GSO does work with hardware-accelerated VLAN tagging, but we
 123                  * can't use hardware-accelerated VLAN tagging since it
 124                  * requires the device to have a VLAN group configured (with
 125                  * e.g. vconfig(8)) and we don't do that.
 126                  *
 127                  * Having to do this here may be a performance loss, since we
 128                  * can't take advantage of TSO hardware support, although it
 129                  * does not make a measurable network performance difference
 130                  * for 1G Ethernet.  Fixing that would require patching the
 131                  * kernel (either to add GSO support to the VLAN protocol or to
 132                  * support hardware-accelerated VLAN tagging without VLAN
 133                  * groups configured). */
 134                 if (skb_is_gso(skb)) {
 135                         const struct nlattr *actions_left;
 136                         u32 actions_len_left;
 137                         struct sk_buff *segs;
 138
 139                         segs = skb_gso_segment(skb, 0);
 140                         kfree_skb(skb);
 141                         if (unlikely(IS_ERR(segs)))
 142                                 return ERR_CAST(segs);
 143
 144                         actions_len_left = actions_len;
 145                         actions_left = nla_next(a, &actions_len_left);
 146
 147                         do {
 148                                 struct sk_buff *nskb = segs->next;
 149                                 int err;
 150
 151                                 segs->next = NULL;
 152
 153                                 /* GSO can change the checksum type so update.*/
 154                                 compute_ip_summed(segs, true);
 155
 156                                 segs = __vlan_put_tag(segs, ntohs(tci));
 157                                 err = -ENOMEM;
 158                                 if (segs) {
 159                                         err = execute_actions(
 160                                                 dp, segs, key, actions_left,
 161                                                 actions_len_left);
 162                                 }
 163
 164                                 if (unlikely(err)) {
 165                                         while ((segs = nskb)) {
 166                                                 nskb = segs->next;
 167                                                 segs->next = NULL;
 168                                                 kfree_skb(segs);
 169                                         }
 170                                         return ERR_PTR(err);
 171                                 }
 172
 173                                 segs = nskb;
 174                         } while (segs->next);
 175
 176                         skb = segs;
 177                         compute_ip_summed(skb, true);
 178                 }
 179
 180                 /* The hardware-accelerated version of vlan_put_tag() works
 181                  * only for a device that has a VLAN group configured (with
 182                  * e.g. vconfig(8)), so call the software-only version
 183                  * __vlan_put_tag() directly instead.
 184                  */
 185                 skb = __vlan_put_tag(skb, ntohs(tci));
 186                 if (!skb)
 187                         return ERR_PTR(-ENOMEM);
 188
 189                 /* GSO doesn't fix up the hardware computed checksum so this
 190                  * will only be hit in the non-GSO case. */
 191                 if (get_ip_summed(skb) == OVS_CSUM_COMPLETE)
 192                         skb->csum = csum_add(skb->csum, csum_partial(skb->data
 193                                                 + ETH_HLEN, VLAN_HLEN, 0));
 194         }
 195
 196         return skb;
 197 }
 198
 199 static struct sk_buff *strip_vlan(struct sk_buff *skb)
 200 {
 201         skb = make_writable(skb, 0);
 202         if (skb)
 203                 vlan_pull_tag(skb);
 204         return skb;
 205 }
 206
 207 static bool is_ip(struct sk_buff *skb, const struct odp_flow_key *key)
 208 {
 209         return (key->dl_type == htons(ETH_P_IP) &&
 210                 skb->transport_header > skb->network_header);
 211 }
 212
 213 static __sum16 *get_l4_checksum(struct sk_buff *skb, const struct odp_flow_key *key)
 214 {
 215         int transport_len = skb->len - skb_transport_offset(skb);
 216         if (key->nw_proto == IPPROTO_TCP) {
 217                 if (likely(transport_len >= sizeof(struct tcphdr)))
 218                         return &tcp_hdr(skb)->check;
 219         } else if (key->nw_proto == IPPROTO_UDP) {
 220                 if (likely(transport_len >= sizeof(struct udphdr)))
 221                         return &udp_hdr(skb)->check;
 222         }
 223         return NULL;
 224 }
 225
 226 static struct sk_buff *set_nw_addr(struct sk_buff *skb,
 227                                    const struct odp_flow_key *key,
 228                                    const struct nlattr *a)
 229 {
 230         __be32 new_nwaddr = nla_get_be32(a);
 231         struct iphdr *nh;
 232         __sum16 *check;
 233         __be32 *nwaddr;
 234
 235         if (unlikely(!is_ip(skb, key)))
 236                 return skb;
 237
 238         skb = make_writable(skb, 0);
 239         if (unlikely(!skb))
 240                 return NULL;
 241
 242         nh = ip_hdr(skb);
 243         nwaddr = nla_type(a) == ODPAT_SET_NW_SRC ? &nh->saddr : &nh->daddr;
 244
 245         check = get_l4_checksum(skb, key);
 246         if (likely(check))
 247                 inet_proto_csum_replace4(check, skb, *nwaddr, new_nwaddr, 1);
 248         csum_replace4(&nh->check, *nwaddr, new_nwaddr);
 249
 250         *nwaddr = new_nwaddr;
 251
 252         return skb;
 253 }
 254
 255 static struct sk_buff *set_nw_tos(struct sk_buff *skb,
 256                                   const struct odp_flow_key *key,
 257                                   u8 nw_tos)
 258 {
 259         if (unlikely(!is_ip(skb, key)))
 260                 return skb;
 261
 262         skb = make_writable(skb, 0);
 263         if (skb) {
 264                 struct iphdr *nh = ip_hdr(skb);
 265                 u8 *f = &nh->tos;
 266                 u8 old = *f;
 267                 u8 new;
 268
 269                 /* Set the DSCP bits and preserve the ECN bits. */
 270                 new = nw_tos | (nh->tos & INET_ECN_MASK);
 271                 csum_replace4(&nh->check, (__force __be32)old,
 272                                           (__force __be32)new);
 273                 *f = new;
 274         }
 275         return skb;
 276 }
 277
 278 static struct sk_buff *set_tp_port(struct sk_buff *skb,
 279                                    const struct odp_flow_key *key,
 280                                    const struct nlattr *a)
 281 {
 282         struct udphdr *th;
 283         __sum16 *check;
 284         __be16 *port;
 285
 286         if (unlikely(!is_ip(skb, key)))
 287                 return skb;
 288
 289         skb = make_writable(skb, 0);
 290         if (unlikely(!skb))
 291                 return NULL;
 292
 293         /* Must follow make_writable() since that can move the skb data. */
 294         check = get_l4_checksum(skb, key);
 295         if (unlikely(!check))
 296                 return skb;
 297
 298         /*
 299          * Update port and checksum.
 300          *
 301          * This is OK because source and destination port numbers are at the
 302          * same offsets in both UDP and TCP headers, and get_l4_checksum() only
 303          * supports those protocols.
 304          */
 305         th = udp_hdr(skb);
 306         port = nla_type(a) == ODPAT_SET_TP_SRC ? &th->source : &th->dest;
 307         inet_proto_csum_replace2(check, skb, *port, nla_get_be16(a), 0);
 308         *port = nla_get_be16(a);
 309
 310         return skb;
 311 }
 312
 313 /**
 314  * is_spoofed_arp - check for invalid ARP packet
 315  *
 316  * @skb: skbuff containing an Ethernet packet, with network header pointing
 317  * just past the Ethernet and optional 802.1Q header.
 318  * @key: flow key extracted from @skb by flow_extract()
 319  *
 320  * Returns true if @skb is an invalid Ethernet+IPv4 ARP packet: one with screwy
 321  * or truncated header fields or one whose inner and outer Ethernet address
 322  * differ.
 323  */
 324 static bool is_spoofed_arp(struct sk_buff *skb, const struct odp_flow_key *key)
 325 {
 326         struct arp_eth_header *arp;
 327
 328         if (key->dl_type != htons(ETH_P_ARP))
 329                 return false;
 330
 331         if (skb_network_offset(skb) + sizeof(struct arp_eth_header) > skb->len)
 332                 return true;
 333
 334         arp = (struct arp_eth_header *)skb_network_header(skb);
 335         return (arp->ar_hrd != htons(ARPHRD_ETHER) ||
 336                 arp->ar_pro != htons(ETH_P_IP) ||
 337                 arp->ar_hln != ETH_ALEN ||
 338                 arp->ar_pln != 4 ||
 339                 compare_ether_addr(arp->ar_sha, eth_hdr(skb)->h_source));
 340 }
 341
 342 static void do_output(struct datapath *dp, struct sk_buff *skb, int out_port)
 343 {
 344         struct vport *p;
 345
 346         if (!skb)
 347                 goto error;
 348
 349         p = rcu_dereference(dp->ports[out_port]);
 350         if (!p)
 351                 goto error;
 352
 353         vport_send(p, skb);
 354         return;
 355
 356 error:
 357         kfree_skb(skb);
 358 }
 359
 360 static int output_control(struct datapath *dp, struct sk_buff *skb, u32 arg)
 361 {
 362         skb = skb_clone(skb, GFP_ATOMIC);
 363         if (!skb)
 364                 return -ENOMEM;
 365         return dp_output_control(dp, skb, _ODPL_ACTION_NR, arg);
 366 }
 367
 368 /* Send a copy of this packet up to the sFlow agent, along with extra
 369  * information about what happened to it. */
 370 static void sflow_sample(struct datapath *dp, struct sk_buff *skb,
 371                          const struct nlattr *a, u32 actions_len,
 372                          struct vport *vport)
 373 {
 374         struct odp_sflow_sample_header *hdr;
 375         unsigned int hdrlen = sizeof(struct odp_sflow_sample_header);
 376         struct sk_buff *nskb;
 377
 378         nskb = skb_copy_expand(skb, actions_len + hdrlen, 0, GFP_ATOMIC);
 379         if (!nskb)
 380                 return;
 381
 382         memcpy(__skb_push(nskb, actions_len), a, actions_len);
 383         hdr = (struct odp_sflow_sample_header*)__skb_push(nskb, hdrlen);
 384         hdr->actions_len = actions_len;
 385         hdr->sample_pool = atomic_read(&vport->sflow_pool);
 386         dp_output_control(dp, nskb, _ODPL_SFLOW_NR, 0);
 387 }
 388
 389 /* Execute a list of actions against 'skb'. */
 390 int execute_actions(struct datapath *dp, struct sk_buff *skb,
 391                     const struct odp_flow_key *key,
 392                     const struct nlattr *actions, u32 actions_len)
 393 {
 394         /* Every output action needs a separate clone of 'skb', but the common
 395          * case is just a single output action, so that doing a clone and
 396          * then freeing the original skbuff is wasteful.  So the following code
 397          * is slightly obscure just to avoid that. */
 398         int prev_port = -1;
 399         u32 priority = skb->priority;
 400         const struct nlattr *a;
 401         int rem, err;
 402
 403         if (dp->sflow_probability) {
 404                 struct vport *p = OVS_CB(skb)->vport;
 405                 if (p) {
 406                         atomic_inc(&p->sflow_pool);
 407                         if (dp->sflow_probability == UINT_MAX ||
 408                             net_random() < dp->sflow_probability)
 409                                 sflow_sample(dp, skb, actions, actions_len, p);
 410                 }
 411         }
 412
 413         OVS_CB(skb)->tun_id = 0;
 414
 415         for (a = actions, rem = actions_len; rem > 0; a = nla_next(a, &rem)) {
 416                 if (prev_port != -1) {
 417                         do_output(dp, skb_clone(skb, GFP_ATOMIC), prev_port);
 418                         prev_port = -1;
 419                 }
 420
 421                 switch (nla_type(a)) {
 422                 case ODPAT_OUTPUT:
 423                         prev_port = nla_get_u32(a);
 424                         break;
 425
 426                 case ODPAT_CONTROLLER:
 427                         err = output_control(dp, skb, nla_get_u32(a));
 428                         if (err) {
 429                                 kfree_skb(skb);
 430                                 return err;
 431                         }
 432                         break;
 433
 434                 case ODPAT_SET_TUNNEL:
 435                         OVS_CB(skb)->tun_id = nla_get_be32(a);
 436                         break;
 437
 438                 case ODPAT_SET_DL_TCI:
 439                         skb = modify_vlan_tci(dp, skb, key, a, rem);
 440                         if (IS_ERR(skb))
 441                                 return PTR_ERR(skb);
 442                         break;
 443
 444                 case ODPAT_STRIP_VLAN:
 445                         skb = strip_vlan(skb);
 446                         break;
 447
 448                 case ODPAT_SET_DL_SRC:
 449                         skb = make_writable(skb, 0);
 450                         if (!skb)
 451                                 return -ENOMEM;
 452                         memcpy(eth_hdr(skb)->h_source, nla_data(a), ETH_ALEN);
 453                         break;
 454
 455                 case ODPAT_SET_DL_DST:
 456                         skb = make_writable(skb, 0);
 457                         if (!skb)
 458                                 return -ENOMEM;
 459                         memcpy(eth_hdr(skb)->h_dest, nla_data(a), ETH_ALEN);
 460                         break;
 461
 462                 case ODPAT_SET_NW_SRC:
 463                 case ODPAT_SET_NW_DST:
 464                         skb = set_nw_addr(skb, key, a);
 465                         break;
 466
 467                 case ODPAT_SET_NW_TOS:
 468                         skb = set_nw_tos(skb, key, nla_get_u8(a));
 469                         break;
 470
 471                 case ODPAT_SET_TP_SRC:
 472                 case ODPAT_SET_TP_DST:
 473                         skb = set_tp_port(skb, key, a);
 474                         break;
 475
 476                 case ODPAT_SET_PRIORITY:
 477                         skb->priority = nla_get_u32(a);
 478                         break;
 479
 480                 case ODPAT_POP_PRIORITY:
 481                         skb->priority = priority;
 482                         break;
 483
 484                 case ODPAT_DROP_SPOOFED_ARP:
 485                         if (unlikely(is_spoofed_arp(skb, key)))
 486                                 goto exit;
 487                         break;
 488                 }
 489                 if (!skb)
 490                         return -ENOMEM;
 491         }
 492 exit:
 493         if (prev_port != -1)
 494                 do_output(dp, skb, prev_port);
 495         else
 496                 kfree_skb(skb);
 497         return 0;
 498 }