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