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_vlan.h>
  18 #include <net/inet_ecn.h>
  19 #include <net/ip.h>
  20 #include <net/checksum.h>
  21 #include "datapath.h"
  22 #include "dp_dev.h"
  23 #include "actions.h"
  24 #include "openvswitch/datapath-protocol.h"
  25
  26 static struct sk_buff *
  27 make_writable(struct sk_buff *skb, unsigned min_headroom, gfp_t gfp)
  28 {
  29         if (skb_shared(skb) || skb_cloned(skb)) {
  30                 struct sk_buff *nskb;
  31                 unsigned headroom = max(min_headroom, skb_headroom(skb));
  32
  33                 nskb = skb_copy_expand(skb, headroom, skb_tailroom(skb), gfp);
  34                 if (nskb) {
  35 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,6,24)
  36                         /* Before 2.6.24 these fields were not copied when
  37                          * doing an skb_copy_expand. */
  38                         nskb->ip_summed = skb->ip_summed;
  39                         nskb->csum = skb->csum;
  40 #endif
  41 #if defined(CONFIG_XEN) && defined(HAVE_PROTO_DATA_VALID)
  42                         /* These fields are copied in skb_clone but not in
  43                          * skb_copy or related functions.  We need to manually
  44                          * copy them over here. */
  45                         nskb->proto_data_valid = skb->proto_data_valid;
  46                         nskb->proto_csum_blank = skb->proto_csum_blank;
  47 #endif
  48                         kfree_skb(skb);
  49                         return nskb;
  50                 }
  51         } else {
  52                 unsigned int hdr_len = (skb_transport_offset(skb)
  53                                         + sizeof(struct tcphdr));
  54                 if (pskb_may_pull(skb, min(hdr_len, skb->len)))
  55                         return skb;
  56         }
  57         kfree_skb(skb);
  58         return NULL;
  59 }
  60
  61 static void set_tunnel(struct sk_buff *skb, struct odp_flow_key *key,
  62                        __be32 tun_id)
  63 {
  64         OVS_CB(skb)->tun_id = key->tun_id = tun_id;
  65 }
  66
  67 static struct sk_buff *
  68 vlan_pull_tag(struct sk_buff *skb)
  69 {
  70         struct vlan_ethhdr *vh = vlan_eth_hdr(skb);
  71         struct ethhdr *eh;
  72
  73         /* Verify we were given a vlan packet */
  74         if (vh->h_vlan_proto != htons(ETH_P_8021Q))
  75                 return skb;
  76
  77         if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE)
  78                 skb->csum = csum_sub(skb->csum, csum_partial(skb->data
  79                                         + ETH_HLEN, VLAN_HLEN, 0));
  80
  81         memmove(skb->data + VLAN_HLEN, skb->data, 2 * VLAN_ETH_ALEN);
  82
  83         eh = (struct ethhdr *)skb_pull(skb, VLAN_HLEN);
  84
  85         skb->protocol = eh->h_proto;
  86         skb->mac_header += VLAN_HLEN;
  87
  88         return skb;
  89 }
  90
  91
  92 static struct sk_buff *
  93 modify_vlan_tci(struct datapath *dp, struct sk_buff *skb,
  94                 struct odp_flow_key *key, const union odp_action *a,
  95                 int n_actions, gfp_t gfp)
  96 {
  97         u16 tci, mask;
  98
  99         if (a->type == ODPAT_SET_VLAN_VID) {
 100                 tci = ntohs(a->vlan_vid.vlan_vid);
 101                 mask = VLAN_VID_MASK;
 102                 key->dl_vlan = a->vlan_vid.vlan_vid;
 103         } else {
 104                 tci = a->vlan_pcp.vlan_pcp << VLAN_PCP_SHIFT;
 105                 mask = VLAN_PCP_MASK;
 106                 key->dl_vlan_pcp = a->vlan_pcp.vlan_pcp;
 107         }
 108
 109         skb = make_writable(skb, VLAN_HLEN, gfp);
 110         if (!skb)
 111                 return ERR_PTR(-ENOMEM);
 112
 113         if (skb->protocol == htons(ETH_P_8021Q)) {
 114                 /* Modify vlan id, but maintain other TCI values */
 115                 struct vlan_ethhdr *vh = vlan_eth_hdr(skb);
 116                 __be16 old_tci = vh->h_vlan_TCI;
 117
 118                 vh->h_vlan_TCI = htons((ntohs(vh->h_vlan_TCI) & ~mask) | tci);
 119
 120                 if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE) {
 121                         __be16 diff[] = { ~old_tci, vh->h_vlan_TCI };
 122
 123                         skb->csum = ~csum_partial((char *)diff, sizeof(diff),
 124                                                 ~skb->csum);
 125                 }
 126         } else {
 127                 /* Add vlan header */
 128
 129                 /* Set up checksumming pointers for checksum-deferred packets
 130                  * on Xen.  Otherwise, dev_queue_xmit() will try to do this
 131                  * when we send the packet out on the wire, and it will fail at
 132                  * that point because skb_checksum_setup() will not look inside
 133                  * an 802.1Q header. */
 134                 vswitch_skb_checksum_setup(skb);
 135
 136                 /* GSO is not implemented for packets with an 802.1Q header, so
 137                  * we have to do segmentation before we add that header.
 138                  *
 139                  * GSO does work with hardware-accelerated VLAN tagging, but we
 140                  * can't use hardware-accelerated VLAN tagging since it
 141                  * requires the device to have a VLAN group configured (with
 142                  * e.g. vconfig(8)) and we don't do that.
 143                  *
 144                  * Having to do this here may be a performance loss, since we
 145                  * can't take advantage of TSO hardware support, although it
 146                  * does not make a measurable network performance difference
 147                  * for 1G Ethernet.  Fixing that would require patching the
 148                  * kernel (either to add GSO support to the VLAN protocol or to
 149                  * support hardware-accelerated VLAN tagging without VLAN
 150                  * groups configured). */
 151                 if (skb_is_gso(skb)) {
 152                         struct sk_buff *segs;
 153
 154                         segs = skb_gso_segment(skb, 0);
 155                         kfree_skb(skb);
 156                         if (unlikely(IS_ERR(segs)))
 157                                 return ERR_CAST(segs);
 158
 159                         do {
 160                                 struct sk_buff *nskb = segs->next;
 161                                 int err;
 162
 163                                 segs->next = NULL;
 164
 165                                 /* GSO can change the checksum type so update.*/
 166                                 compute_ip_summed(segs, true);
 167
 168                                 segs = __vlan_put_tag(segs, tci);
 169                                 err = -ENOMEM;
 170                                 if (segs) {
 171                                         struct odp_flow_key segkey = *key;
 172                                         err = execute_actions(dp, segs,
 173                                                               &segkey, a + 1,
 174                                                               n_actions - 1,
 175                                                               gfp);
 176                                 }
 177
 178                                 if (unlikely(err)) {
 179                                         while ((segs = nskb)) {
 180                                                 nskb = segs->next;
 181                                                 segs->next = NULL;
 182                                                 kfree_skb(segs);
 183                                         }
 184                                         return ERR_PTR(err);
 185                                 }
 186
 187                                 segs = nskb;
 188                         } while (segs->next);
 189
 190                         skb = segs;
 191                         compute_ip_summed(skb, true);
 192                 }
 193
 194                 /* The hardware-accelerated version of vlan_put_tag() works
 195                  * only for a device that has a VLAN group configured (with
 196                  * e.g. vconfig(8)), so call the software-only version
 197                  * __vlan_put_tag() directly instead.
 198                  */
 199                 skb = __vlan_put_tag(skb, tci);
 200                 if (!skb)
 201                         return ERR_PTR(-ENOMEM);
 202
 203                 /* GSO doesn't fix up the hardware computed checksum so this
 204                  * will only be hit in the non-GSO case. */
 205                 if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE)
 206                         skb->csum = csum_add(skb->csum, csum_partial(skb->data
 207                                                 + ETH_HLEN, VLAN_HLEN, 0));
 208         }
 209
 210         return skb;
 211 }
 212
 213 static struct sk_buff *strip_vlan(struct sk_buff *skb,
 214                                   struct odp_flow_key *key, gfp_t gfp)
 215 {
 216         skb = make_writable(skb, 0, gfp);
 217         if (skb) {
 218                 vlan_pull_tag(skb);
 219                 key->dl_vlan = htons(ODP_VLAN_NONE);
 220         }
 221         return skb;
 222 }
 223
 224 static struct sk_buff *set_dl_addr(struct sk_buff *skb,
 225                                    struct odp_flow_key *key,
 226                                    const struct odp_action_dl_addr *a,
 227                                    gfp_t gfp)
 228 {
 229         skb = make_writable(skb, 0, gfp);
 230         if (skb) {
 231                 struct ethhdr *eh = eth_hdr(skb);
 232                 if (a->type == ODPAT_SET_DL_SRC) {
 233                         memcpy(eh->h_source, a->dl_addr, ETH_ALEN);
 234                         memcpy(key->dl_src, a->dl_addr, ETH_ALEN);
 235                 } else {
 236                         memcpy(eh->h_dest, a->dl_addr, ETH_ALEN);
 237                         memcpy(key->dl_dst, a->dl_addr, ETH_ALEN);
 238                 }
 239         }
 240         return skb;
 241 }
 242
 243 /* Updates 'sum', which is a field in 'skb''s data, given that a 4-byte field
 244  * covered by the sum has been changed from 'from' to 'to'.  If set,
 245  * 'pseudohdr' indicates that the field is in the TCP or UDP pseudo-header.
 246  * Based on nf_proto_csum_replace4. */
 247 static void update_csum(__sum16 *sum, struct sk_buff *skb,
 248                         __be32 from, __be32 to, int pseudohdr)
 249 {
 250         __be32 diff[] = { ~from, to };
 251
 252         if (OVS_CB(skb)->ip_summed != OVS_CSUM_PARTIAL) {
 253                 *sum = csum_fold(csum_partial((char *)diff, sizeof(diff),
 254                                 ~csum_unfold(*sum)));
 255                 if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE && pseudohdr)
 256                         skb->csum = ~csum_partial((char *)diff, sizeof(diff),
 257                                                 ~skb->csum);
 258         } else if (pseudohdr)
 259                 *sum = ~csum_fold(csum_partial((char *)diff, sizeof(diff),
 260                                 csum_unfold(*sum)));
 261 }
 262
 263 static struct sk_buff *set_nw_addr(struct sk_buff *skb,
 264                                    struct odp_flow_key *key,
 265                                    const struct odp_action_nw_addr *a,
 266                                    gfp_t gfp)
 267 {
 268         if (key->dl_type != htons(ETH_P_IP))
 269                 return skb;
 270
 271         skb = make_writable(skb, 0, gfp);
 272         if (skb) {
 273                 struct iphdr *nh = ip_hdr(skb);
 274                 u32 *f = a->type == ODPAT_SET_NW_SRC ? &nh->saddr : &nh->daddr;
 275                 u32 old = *f;
 276                 u32 new = a->nw_addr;
 277
 278                 if (key->nw_proto == IPPROTO_TCP) {
 279                         struct tcphdr *th = tcp_hdr(skb);
 280                         update_csum(&th->check, skb, old, new, 1);
 281                 } else if (key->nw_proto == IPPROTO_UDP) {
 282                         struct udphdr *th = udp_hdr(skb);
 283                         update_csum(&th->check, skb, old, new, 1);
 284                 }
 285                 update_csum(&nh->check, skb, old, new, 0);
 286                 *f = new;
 287
 288                 if (a->type == ODPAT_SET_NW_SRC)
 289                         key->nw_src = a->nw_addr;
 290                 else
 291                         key->nw_dst = a->nw_addr;
 292         }
 293         return skb;
 294 }
 295
 296 static struct sk_buff *set_nw_tos(struct sk_buff *skb,
 297                                    struct odp_flow_key *key,
 298                                    const struct odp_action_nw_tos *a,
 299                                    gfp_t gfp)
 300 {
 301         if (key->dl_type != htons(ETH_P_IP))
 302                 return skb;
 303
 304         skb = make_writable(skb, 0, gfp);
 305         if (skb) {
 306                 struct iphdr *nh = ip_hdr(skb);
 307                 u8 *f = &nh->tos;
 308                 u8 old = *f;
 309                 u8 new;
 310
 311                 /* Set the DSCP bits and preserve the ECN bits. */
 312                 new = a->nw_tos | (nh->tos & INET_ECN_MASK);
 313                 update_csum(&nh->check, skb, htons((uint16_t)old),
 314                                 htons((uint16_t)new), 0);
 315                 *f = new;
 316                 key->nw_tos = a->nw_tos;
 317         }
 318         return skb;
 319 }
 320
 321 static struct sk_buff *
 322 set_tp_port(struct sk_buff *skb, struct odp_flow_key *key,
 323             const struct odp_action_tp_port *a,
 324             gfp_t gfp)
 325 {
 326         int check_ofs;
 327
 328         if (key->dl_type != htons(ETH_P_IP))
 329                 return skb;
 330
 331         if (key->nw_proto == IPPROTO_TCP)
 332                 check_ofs = offsetof(struct tcphdr, check);
 333         else if (key->nw_proto == IPPROTO_UDP)
 334                 check_ofs = offsetof(struct udphdr, check);
 335         else
 336                 return skb;
 337
 338         skb = make_writable(skb, 0, gfp);
 339         if (skb) {
 340                 struct udphdr *th = udp_hdr(skb);
 341                 u16 *f = a->type == ODPAT_SET_TP_SRC ? &th->source : &th->dest;
 342                 u16 old = *f;
 343                 u16 new = a->tp_port;
 344                 update_csum((u16*)(skb_transport_header(skb) + check_ofs),
 345                                 skb, old, new, 0);
 346                 *f = new;
 347                 if (a->type == ODPAT_SET_TP_SRC)
 348                         key->tp_src = a->tp_port;
 349                 else
 350                         key->tp_dst = a->tp_port;
 351         }
 352         return skb;
 353 }
 354
 355 static inline unsigned packet_length(const struct sk_buff *skb)
 356 {
 357         unsigned length = skb->len - ETH_HLEN;
 358         if (skb->protocol == htons(ETH_P_8021Q))
 359                 length -= VLAN_HLEN;
 360         return length;
 361 }
 362
 363 int dp_xmit_skb(struct sk_buff *skb)
 364 {
 365         struct datapath *dp = skb->dev->br_port->dp;
 366         int len = skb->len;
 367
 368         if (packet_length(skb) > skb->dev->mtu && !skb_is_gso(skb)) {
 369                 printk(KERN_WARNING "%s: dropped over-mtu packet: %d > %d\n",
 370                        dp_name(dp), packet_length(skb), skb->dev->mtu);
 371                 kfree_skb(skb);
 372                 return -E2BIG;
 373         }
 374
 375         forward_ip_summed(skb);
 376         dev_queue_xmit(skb);
 377
 378         return len;
 379 }
 380
 381 static void
 382 do_output(struct datapath *dp, struct sk_buff *skb, int out_port)
 383 {
 384         struct net_bridge_port *p;
 385         struct net_device *dev;
 386
 387         if (!skb)
 388                 goto error;
 389
 390         p = dp->ports[out_port];
 391         if (!p)
 392                 goto error;
 393
 394         dev = skb->dev = p->dev;
 395         if (is_dp_dev(dev))
 396                 dp_dev_recv(dev, skb);
 397         else
 398                 dp_xmit_skb(skb);
 399         return;
 400
 401 error:
 402         kfree_skb(skb);
 403 }
 404
 405 /* Never consumes 'skb'.  Returns a port that 'skb' should be sent to, -1 if
 406  * none.  */
 407 static int output_group(struct datapath *dp, __u16 group,
 408                         struct sk_buff *skb, gfp_t gfp)
 409 {
 410         struct dp_port_group *g = rcu_dereference(dp->groups[group]);
 411         int prev_port = -1;
 412         int i;
 413
 414         if (!g)
 415                 return -1;
 416         for (i = 0; i < g->n_ports; i++) {
 417                 struct net_bridge_port *p = dp->ports[g->ports[i]];
 418                 if (!p || skb->dev == p->dev)
 419                         continue;
 420                 if (prev_port != -1) {
 421                         struct sk_buff *clone = skb_clone(skb, gfp);
 422                         if (!clone)
 423                                 return -1;
 424                         do_output(dp, clone, prev_port);
 425                 }
 426                 prev_port = p->port_no;
 427         }
 428         return prev_port;
 429 }
 430
 431 static int
 432 output_control(struct datapath *dp, struct sk_buff *skb, u32 arg, gfp_t gfp)
 433 {
 434         skb = skb_clone(skb, gfp);
 435         if (!skb)
 436                 return -ENOMEM;
 437         return dp_output_control(dp, skb, _ODPL_ACTION_NR, arg);
 438 }
 439
 440 /* Send a copy of this packet up to the sFlow agent, along with extra
 441  * information about what happened to it. */
 442 static void sflow_sample(struct datapath *dp, struct sk_buff *skb,
 443                          const union odp_action *a, int n_actions,
 444                          gfp_t gfp, struct net_bridge_port *nbp)
 445 {
 446         struct odp_sflow_sample_header *hdr;
 447         unsigned int actlen = n_actions * sizeof(union odp_action);
 448         unsigned int hdrlen = sizeof(struct odp_sflow_sample_header);
 449         struct sk_buff *nskb;
 450
 451         nskb = skb_copy_expand(skb, actlen + hdrlen, 0, gfp);
 452         if (!nskb)
 453                 return;
 454
 455         memcpy(__skb_push(nskb, actlen), a, actlen);
 456         hdr = (struct odp_sflow_sample_header*)__skb_push(nskb, hdrlen);
 457         hdr->n_actions = n_actions;
 458         hdr->sample_pool = atomic_read(&nbp->sflow_pool);
 459         dp_output_control(dp, nskb, _ODPL_SFLOW_NR, 0);
 460 }
 461
 462 /* Execute a list of actions against 'skb'. */
 463 int execute_actions(struct datapath *dp, struct sk_buff *skb,
 464                     struct odp_flow_key *key,
 465                     const union odp_action *a, int n_actions,
 466                     gfp_t gfp)
 467 {
 468         /* Every output action needs a separate clone of 'skb', but the common
 469          * case is just a single output action, so that doing a clone and
 470          * then freeing the original skbuff is wasteful.  So the following code
 471          * is slightly obscure just to avoid that. */
 472         int prev_port = -1;
 473         int err;
 474
 475         if (dp->sflow_probability) {
 476                 struct net_bridge_port *p = skb->dev->br_port;
 477                 if (p) {
 478                         atomic_inc(&p->sflow_pool);
 479                         if (dp->sflow_probability == UINT_MAX ||
 480                             net_random() < dp->sflow_probability)
 481                                 sflow_sample(dp, skb, a, n_actions, gfp, p);
 482                 }
 483         }
 484
 485         OVS_CB(skb)->tun_id = 0;
 486
 487         for (; n_actions > 0; a++, n_actions--) {
 488                 WARN_ON_ONCE(skb_shared(skb));
 489                 if (prev_port != -1) {
 490                         do_output(dp, skb_clone(skb, gfp), prev_port);
 491                         prev_port = -1;
 492                 }
 493
 494                 switch (a->type) {
 495                 case ODPAT_OUTPUT:
 496                         prev_port = a->output.port;
 497                         break;
 498
 499                 case ODPAT_OUTPUT_GROUP:
 500                         prev_port = output_group(dp, a->output_group.group,
 501                                                  skb, gfp);
 502                         break;
 503
 504                 case ODPAT_CONTROLLER:
 505                         err = output_control(dp, skb, a->controller.arg, gfp);
 506                         if (err) {
 507                                 kfree_skb(skb);
 508                                 return err;
 509                         }
 510                         break;
 511
 512                 case ODPAT_SET_TUNNEL:
 513                         set_tunnel(skb, key, a->tunnel.tun_id);
 514                         break;
 515
 516                 case ODPAT_SET_VLAN_VID:
 517                 case ODPAT_SET_VLAN_PCP:
 518                         skb = modify_vlan_tci(dp, skb, key, a, n_actions, gfp);
 519                         if (IS_ERR(skb))
 520                                 return PTR_ERR(skb);
 521                         break;
 522
 523                 case ODPAT_STRIP_VLAN:
 524                         skb = strip_vlan(skb, key, gfp);
 525                         break;
 526
 527                 case ODPAT_SET_DL_SRC:
 528                 case ODPAT_SET_DL_DST:
 529                         skb = set_dl_addr(skb, key, &a->dl_addr, gfp);
 530                         break;
 531
 532                 case ODPAT_SET_NW_SRC:
 533                 case ODPAT_SET_NW_DST:
 534                         skb = set_nw_addr(skb, key, &a->nw_addr, gfp);
 535                         break;
 536
 537                 case ODPAT_SET_NW_TOS:
 538                         skb = set_nw_tos(skb, key, &a->nw_tos, gfp);
 539                         break;
 540
 541                 case ODPAT_SET_TP_SRC:
 542                 case ODPAT_SET_TP_DST:
 543                         skb = set_tp_port(skb, key, &a->tp_port, gfp);
 544                         break;
 545                 }
 546                 if (!skb)
 547                         return -ENOMEM;
 548         }
 549         if (prev_port != -1)
 550                 do_output(dp, skb, prev_port);
 551         else
 552                 kfree_skb(skb);
 553         return 0;
 554 }