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 "datapath.h"
  25 #include "openvswitch/datapath-protocol.h"
  26 #include "vport.h"
  27
  28 static struct sk_buff *make_writable(struct sk_buff *skb, unsigned min_headroom, gfp_t gfp)
  29 {
  30         if (skb_cloned(skb)) {
  31                 struct sk_buff *nskb;
  32                 unsigned headroom = max(min_headroom, skb_headroom(skb));
  33
  34                 nskb = skb_copy_expand(skb, headroom, skb_tailroom(skb), gfp);
  35                 if (nskb) {
  36                         set_skb_csum_bits(skb, nskb);
  37                         kfree_skb(skb);
  38                         return nskb;
  39                 }
  40         } else {
  41                 unsigned int hdr_len = (skb_transport_offset(skb)
  42                                         + sizeof(struct tcphdr));
  43                 if (pskb_may_pull(skb, min(hdr_len, skb->len)))
  44                         return skb;
  45         }
  46         kfree_skb(skb);
  47         return NULL;
  48 }
  49
  50 static struct sk_buff *vlan_pull_tag(struct sk_buff *skb)
  51 {
  52         struct vlan_ethhdr *vh = vlan_eth_hdr(skb);
  53         struct ethhdr *eh;
  54
  55         /* Verify we were given a vlan packet */
  56         if (vh->h_vlan_proto != htons(ETH_P_8021Q) || skb->len < VLAN_ETH_HLEN)
  57                 return skb;
  58
  59         if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE)
  60                 skb->csum = csum_sub(skb->csum, csum_partial(skb->data
  61                                         + ETH_HLEN, VLAN_HLEN, 0));
  62
  63         memmove(skb->data + VLAN_HLEN, skb->data, 2 * VLAN_ETH_ALEN);
  64
  65         eh = (struct ethhdr *)skb_pull(skb, VLAN_HLEN);
  66
  67         skb->protocol = eh->h_proto;
  68         skb->mac_header += VLAN_HLEN;
  69
  70         return skb;
  71 }
  72
  73 static struct sk_buff *modify_vlan_tci(struct datapath *dp, struct sk_buff *skb,
  74                                        const struct odp_flow_key *key,
  75                                        const union odp_action *a, int n_actions,
  76                                        gfp_t gfp)
  77 {
  78         u16 tci, mask;
  79
  80         if (a->type == ODPAT_SET_VLAN_VID) {
  81                 tci = ntohs(a->vlan_vid.vlan_vid);
  82                 mask = VLAN_VID_MASK;
  83         } else {
  84                 tci = a->vlan_pcp.vlan_pcp << VLAN_PCP_SHIFT;
  85                 mask = VLAN_PCP_MASK;
  86         }
  87
  88         skb = make_writable(skb, VLAN_HLEN, gfp);
  89         if (!skb)
  90                 return ERR_PTR(-ENOMEM);
  91
  92         if (skb->protocol == htons(ETH_P_8021Q)) {
  93                 /* Modify vlan id, but maintain other TCI values */
  94                 struct vlan_ethhdr *vh;
  95                 __be16 old_tci;
  96
  97                 if (skb->len < VLAN_ETH_HLEN)
  98                         return skb;
  99
 100                 vh = vlan_eth_hdr(skb);
 101                 old_tci = vh->h_vlan_TCI;
 102
 103                 vh->h_vlan_TCI = htons((ntohs(vh->h_vlan_TCI) & ~mask) | tci);
 104
 105                 if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE) {
 106                         __be16 diff[] = { ~old_tci, vh->h_vlan_TCI };
 107
 108                         skb->csum = ~csum_partial((char *)diff, sizeof(diff),
 109                                                 ~skb->csum);
 110                 }
 111         } else {
 112                 int err;
 113
 114                 /* Add vlan header */
 115
 116                 /* Set up checksumming pointers for checksum-deferred packets
 117                  * on Xen.  Otherwise, dev_queue_xmit() will try to do this
 118                  * when we send the packet out on the wire, and it will fail at
 119                  * that point because skb_checksum_setup() will not look inside
 120                  * an 802.1Q header. */
 121                 err = vswitch_skb_checksum_setup(skb);
 122                 if (unlikely(err)) {
 123                         kfree_skb(skb);
 124                         return ERR_PTR(err);
 125                 }
 126
 127                 /* GSO is not implemented for packets with an 802.1Q header, so
 128                  * we have to do segmentation before we add that header.
 129                  *
 130                  * GSO does work with hardware-accelerated VLAN tagging, but we
 131                  * can't use hardware-accelerated VLAN tagging since it
 132                  * requires the device to have a VLAN group configured (with
 133                  * e.g. vconfig(8)) and we don't do that.
 134                  *
 135                  * Having to do this here may be a performance loss, since we
 136                  * can't take advantage of TSO hardware support, although it
 137                  * does not make a measurable network performance difference
 138                  * for 1G Ethernet.  Fixing that would require patching the
 139                  * kernel (either to add GSO support to the VLAN protocol or to
 140                  * support hardware-accelerated VLAN tagging without VLAN
 141                  * groups configured). */
 142                 if (skb_is_gso(skb)) {
 143                         struct sk_buff *segs;
 144
 145                         segs = skb_gso_segment(skb, 0);
 146                         kfree_skb(skb);
 147                         if (unlikely(IS_ERR(segs)))
 148                                 return ERR_CAST(segs);
 149
 150                         do {
 151                                 struct sk_buff *nskb = segs->next;
 152                                 int err;
 153
 154                                 segs->next = NULL;
 155
 156                                 /* GSO can change the checksum type so update.*/
 157                                 compute_ip_summed(segs, true);
 158
 159                                 segs = __vlan_put_tag(segs, tci);
 160                                 err = -ENOMEM;
 161                                 if (segs) {
 162                                         err = execute_actions(dp, segs,
 163                                                               key, a + 1,
 164                                                               n_actions - 1,
 165                                                               gfp);
 166                                 }
 167
 168                                 if (unlikely(err)) {
 169                                         while ((segs = nskb)) {
 170                                                 nskb = segs->next;
 171                                                 segs->next = NULL;
 172                                                 kfree_skb(segs);
 173                                         }
 174                                         return ERR_PTR(err);
 175                                 }
 176
 177                                 segs = nskb;
 178                         } while (segs->next);
 179
 180                         skb = segs;
 181                         compute_ip_summed(skb, true);
 182                 }
 183
 184                 /* The hardware-accelerated version of vlan_put_tag() works
 185                  * only for a device that has a VLAN group configured (with
 186                  * e.g. vconfig(8)), so call the software-only version
 187                  * __vlan_put_tag() directly instead.
 188                  */
 189                 skb = __vlan_put_tag(skb, tci);
 190                 if (!skb)
 191                         return ERR_PTR(-ENOMEM);
 192
 193                 /* GSO doesn't fix up the hardware computed checksum so this
 194                  * will only be hit in the non-GSO case. */
 195                 if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE)
 196                         skb->csum = csum_add(skb->csum, csum_partial(skb->data
 197                                                 + ETH_HLEN, VLAN_HLEN, 0));
 198         }
 199
 200         return skb;
 201 }
 202
 203 static struct sk_buff *strip_vlan(struct sk_buff *skb, gfp_t gfp)
 204 {
 205         skb = make_writable(skb, 0, gfp);
 206         if (skb)
 207                 vlan_pull_tag(skb);
 208
 209         return skb;
 210 }
 211
 212 static struct sk_buff *set_dl_addr(struct sk_buff *skb,
 213                                    const struct odp_action_dl_addr *a,
 214                                    gfp_t gfp)
 215 {
 216         skb = make_writable(skb, 0, gfp);
 217         if (skb) {
 218                 struct ethhdr *eh = eth_hdr(skb);
 219                 if (a->type == ODPAT_SET_DL_SRC)
 220                         memcpy(eh->h_source, a->dl_addr, ETH_ALEN);
 221                 else
 222                         memcpy(eh->h_dest, a->dl_addr, ETH_ALEN);
 223         }
 224         return skb;
 225 }
 226
 227 /* Updates 'sum', which is a field in 'skb''s data, given that a 4-byte field
 228  * covered by the sum has been changed from 'from' to 'to'.  If set,
 229  * 'pseudohdr' indicates that the field is in the TCP or UDP pseudo-header.
 230  * Based on nf_proto_csum_replace4. */
 231 static void update_csum(__sum16 *sum, struct sk_buff *skb,
 232                         __be32 from, __be32 to, int pseudohdr)
 233 {
 234         __be32 diff[] = { ~from, to };
 235
 236         if (OVS_CB(skb)->ip_summed != OVS_CSUM_PARTIAL) {
 237                 *sum = csum_fold(csum_partial((char *)diff, sizeof(diff),
 238                                 ~csum_unfold(*sum)));
 239                 if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE && pseudohdr)
 240                         skb->csum = ~csum_partial((char *)diff, sizeof(diff),
 241                                                 ~skb->csum);
 242         } else if (pseudohdr)
 243                 *sum = ~csum_fold(csum_partial((char *)diff, sizeof(diff),
 244                                 csum_unfold(*sum)));
 245 }
 246
 247 static bool is_ip(struct sk_buff *skb, const struct odp_flow_key *key)
 248 {
 249         return (key->dl_type == htons(ETH_P_IP) &&
 250                 skb->transport_header > skb->network_header);
 251 }
 252
 253 static __sum16 *get_l4_checksum(struct sk_buff *skb, const struct odp_flow_key *key)
 254 {
 255         int transport_len = skb->len - skb_transport_offset(skb);
 256         if (key->nw_proto == IPPROTO_TCP) {
 257                 if (likely(transport_len >= sizeof(struct tcphdr)))
 258                         return &tcp_hdr(skb)->check;
 259         } else if (key->nw_proto == IPPROTO_UDP) {
 260                 if (likely(transport_len >= sizeof(struct udphdr)))
 261                         return &udp_hdr(skb)->check;
 262         }
 263         return NULL;
 264 }
 265
 266 static struct sk_buff *set_nw_addr(struct sk_buff *skb,
 267                                    const struct odp_flow_key *key,
 268                                    const struct odp_action_nw_addr *a,
 269                                    gfp_t gfp)
 270 {
 271         struct iphdr *nh;
 272         __sum16 *check;
 273         __be32 *nwaddr;
 274
 275         if (unlikely(!is_ip(skb, key)))
 276                 return skb;
 277
 278         skb = make_writable(skb, 0, gfp);
 279         if (unlikely(!skb))
 280                 return NULL;
 281
 282         nh = ip_hdr(skb);
 283         nwaddr = a->type == ODPAT_SET_NW_SRC ? &nh->saddr : &nh->daddr;
 284
 285         check = get_l4_checksum(skb, key);
 286         if (likely(check))
 287                 update_csum(check, skb, *nwaddr, a->nw_addr, 1);
 288         update_csum(&nh->check, skb, *nwaddr, a->nw_addr, 0);
 289
 290         *nwaddr = a->nw_addr;
 291
 292         return skb;
 293 }
 294
 295 static struct sk_buff *set_nw_tos(struct sk_buff *skb,
 296                                    const struct odp_flow_key *key,
 297                                    const struct odp_action_nw_tos *a,
 298                                    gfp_t gfp)
 299 {
 300         if (unlikely(!is_ip(skb, key)))
 301                 return skb;
 302
 303         skb = make_writable(skb, 0, gfp);
 304         if (skb) {
 305                 struct iphdr *nh = ip_hdr(skb);
 306                 u8 *f = &nh->tos;
 307                 u8 old = *f;
 308                 u8 new;
 309
 310                 /* Set the DSCP bits and preserve the ECN bits. */
 311                 new = a->nw_tos | (nh->tos & INET_ECN_MASK);
 312                 update_csum(&nh->check, skb, htons((u16)old),
 313                             htons((u16)new), 0);
 314                 *f = new;
 315         }
 316         return skb;
 317 }
 318
 319 static struct sk_buff *set_tp_port(struct sk_buff *skb,
 320                                    const struct odp_flow_key *key,
 321                                    const struct odp_action_tp_port *a, gfp_t gfp)
 322 {
 323         struct udphdr *th;
 324         __sum16 *check;
 325         __be16 *port;
 326
 327         if (unlikely(!is_ip(skb, key)))
 328                 return skb;
 329
 330         skb = make_writable(skb, 0, gfp);
 331         if (unlikely(!skb))
 332                 return NULL;
 333
 334         /* Must follow make_writable() since that can move the skb data. */
 335         check = get_l4_checksum(skb, key);
 336         if (unlikely(!check))
 337                 return skb;
 338
 339         /*
 340          * Update port and checksum.
 341          *
 342          * This is OK because source and destination port numbers are at the
 343          * same offsets in both UDP and TCP headers, and get_l4_checksum() only
 344          * supports those protocols.
 345          */
 346         th = udp_hdr(skb);
 347         port = a->type == ODPAT_SET_TP_SRC ? &th->source : &th->dest;
 348         update_csum(check, skb, *port, a->tp_port, 0);
 349         *port = a->tp_port;
 350
 351         return skb;
 352 }
 353
 354 /**
 355  * is_spoofed_arp - check for invalid ARP packet
 356  *
 357  * @skb: skbuff containing an Ethernet packet, with network header pointing
 358  * just past the Ethernet and optional 802.1Q header.
 359  * @key: flow key extracted from @skb by flow_extract()
 360  *
 361  * Returns true if @skb is an invalid Ethernet+IPv4 ARP packet: one with screwy
 362  * or truncated header fields or one whose inner and outer Ethernet address
 363  * differ.
 364  */
 365 static bool is_spoofed_arp(struct sk_buff *skb, const struct odp_flow_key *key)
 366 {
 367         struct arp_eth_header *arp;
 368
 369         if (key->dl_type != htons(ETH_P_ARP))
 370                 return false;
 371
 372         if (skb_network_offset(skb) + sizeof(struct arp_eth_header) > skb->len)
 373                 return true;
 374
 375         arp = (struct arp_eth_header *)skb_network_header(skb);
 376         return (arp->ar_hrd != htons(ARPHRD_ETHER) ||
 377                 arp->ar_pro != htons(ETH_P_IP) ||
 378                 arp->ar_hln != ETH_ALEN ||
 379                 arp->ar_pln != 4 ||
 380                 compare_ether_addr(arp->ar_sha, eth_hdr(skb)->h_source));
 381 }
 382
 383 static void do_output(struct datapath *dp, struct sk_buff *skb, int out_port)
 384 {
 385         struct dp_port *p;
 386
 387         if (!skb)
 388                 goto error;
 389
 390         p = rcu_dereference(dp->ports[out_port]);
 391         if (!p)
 392                 goto error;
 393
 394         vport_send(p->vport, skb);
 395         return;
 396
 397 error:
 398         kfree_skb(skb);
 399 }
 400
 401 /* Never consumes 'skb'.  Returns a port that 'skb' should be sent to, -1 if
 402  * none.  */
 403 static int output_group(struct datapath *dp, __u16 group,
 404                         struct sk_buff *skb, gfp_t gfp)
 405 {
 406         struct dp_port_group *g = rcu_dereference(dp->groups[group]);
 407         int prev_port = -1;
 408         int i;
 409
 410         if (!g)
 411                 return -1;
 412         for (i = 0; i < g->n_ports; i++) {
 413                 struct dp_port *p = rcu_dereference(dp->ports[g->ports[i]]);
 414                 if (!p || OVS_CB(skb)->dp_port == p)
 415                         continue;
 416                 if (prev_port != -1) {
 417                         struct sk_buff *clone = skb_clone(skb, gfp);
 418                         if (!clone)
 419                                 return -1;
 420                         do_output(dp, clone, prev_port);
 421                 }
 422                 prev_port = p->port_no;
 423         }
 424         return prev_port;
 425 }
 426
 427 static int output_control(struct datapath *dp, struct sk_buff *skb, u32 arg,
 428                           gfp_t gfp)
 429 {
 430         skb = skb_clone(skb, gfp);
 431         if (!skb)
 432                 return -ENOMEM;
 433         return dp_output_control(dp, skb, _ODPL_ACTION_NR, arg);
 434 }
 435
 436 /* Send a copy of this packet up to the sFlow agent, along with extra
 437  * information about what happened to it. */
 438 static void sflow_sample(struct datapath *dp, struct sk_buff *skb,
 439                          const union odp_action *a, int n_actions,
 440                          gfp_t gfp, struct dp_port *dp_port)
 441 {
 442         struct odp_sflow_sample_header *hdr;
 443         unsigned int actlen = n_actions * sizeof(union odp_action);
 444         unsigned int hdrlen = sizeof(struct odp_sflow_sample_header);
 445         struct sk_buff *nskb;
 446
 447         nskb = skb_copy_expand(skb, actlen + hdrlen, 0, gfp);
 448         if (!nskb)
 449                 return;
 450
 451         memcpy(__skb_push(nskb, actlen), a, actlen);
 452         hdr = (struct odp_sflow_sample_header*)__skb_push(nskb, hdrlen);
 453         hdr->n_actions = n_actions;
 454         hdr->sample_pool = atomic_read(&dp_port->sflow_pool);
 455         dp_output_control(dp, nskb, _ODPL_SFLOW_NR, 0);
 456 }
 457
 458 /* Execute a list of actions against 'skb'. */
 459 int execute_actions(struct datapath *dp, struct sk_buff *skb,
 460                     const struct odp_flow_key *key,
 461                     const union odp_action *a, int n_actions,
 462                     gfp_t gfp)
 463 {
 464         /* Every output action needs a separate clone of 'skb', but the common
 465          * case is just a single output action, so that doing a clone and
 466          * then freeing the original skbuff is wasteful.  So the following code
 467          * is slightly obscure just to avoid that. */
 468         int prev_port = -1;
 469         u32 priority = skb->priority;
 470         int err;
 471
 472         if (dp->sflow_probability) {
 473                 struct dp_port *p = OVS_CB(skb)->dp_port;
 474                 if (p) {
 475                         atomic_inc(&p->sflow_pool);
 476                         if (dp->sflow_probability == UINT_MAX ||
 477                             net_random() < dp->sflow_probability)
 478                                 sflow_sample(dp, skb, a, n_actions, gfp, p);
 479                 }
 480         }
 481
 482         OVS_CB(skb)->tun_id = 0;
 483
 484         for (; n_actions > 0; a++, n_actions--) {
 485                 if (prev_port != -1) {
 486                         do_output(dp, skb_clone(skb, gfp), prev_port);
 487                         prev_port = -1;
 488                 }
 489
 490                 switch (a->type) {
 491                 case ODPAT_OUTPUT:
 492                         prev_port = a->output.port;
 493                         break;
 494
 495                 case ODPAT_OUTPUT_GROUP:
 496                         prev_port = output_group(dp, a->output_group.group,
 497                                                  skb, gfp);
 498                         break;
 499
 500                 case ODPAT_CONTROLLER:
 501                         err = output_control(dp, skb, a->controller.arg, gfp);
 502                         if (err) {
 503                                 kfree_skb(skb);
 504                                 return err;
 505                         }
 506                         break;
 507
 508                 case ODPAT_SET_TUNNEL:
 509                         OVS_CB(skb)->tun_id = a->tunnel.tun_id;
 510                         break;
 511
 512                 case ODPAT_SET_VLAN_VID:
 513                 case ODPAT_SET_VLAN_PCP:
 514                         skb = modify_vlan_tci(dp, skb, key, a, n_actions, gfp);
 515                         if (IS_ERR(skb))
 516                                 return PTR_ERR(skb);
 517                         break;
 518
 519                 case ODPAT_STRIP_VLAN:
 520                         skb = strip_vlan(skb, gfp);
 521                         break;
 522
 523                 case ODPAT_SET_DL_SRC:
 524                 case ODPAT_SET_DL_DST:
 525                         skb = set_dl_addr(skb, &a->dl_addr, gfp);
 526                         break;
 527
 528                 case ODPAT_SET_NW_SRC:
 529                 case ODPAT_SET_NW_DST:
 530                         skb = set_nw_addr(skb, key, &a->nw_addr, gfp);
 531                         break;
 532
 533                 case ODPAT_SET_NW_TOS:
 534                         skb = set_nw_tos(skb, key, &a->nw_tos, gfp);
 535                         break;
 536
 537                 case ODPAT_SET_TP_SRC:
 538                 case ODPAT_SET_TP_DST:
 539                         skb = set_tp_port(skb, key, &a->tp_port, gfp);
 540                         break;
 541
 542                 case ODPAT_SET_PRIORITY:
 543                         skb->priority = a->priority.priority;
 544                         break;
 545
 546                 case ODPAT_POP_PRIORITY:
 547                         skb->priority = priority;
 548                         break;
 549
 550                 case ODPAT_DROP_SPOOFED_ARP:
 551                         if (unlikely(is_spoofed_arp(skb, key)))
 552                                 goto exit;
 553                         break;
 554                 }
 555                 if (!skb)
 556                         return -ENOMEM;
 557         }
 558 exit:
 559         if (prev_port != -1)
 560                 do_output(dp, skb, prev_port);
 561         else
 562                 kfree_skb(skb);
 563         return 0;
 564 }