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