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