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
  69         /* Verify we were given a vlan packet */
  70         if (vh->h_vlan_proto != htons(ETH_P_8021Q))
  71                 return skb;
  72
  73         memmove(skb->data + VLAN_HLEN, skb->data, 2 * VLAN_ETH_ALEN);
  74
  75         eh = (struct ethhdr *)skb_pull(skb, VLAN_HLEN);
  76
  77         skb->protocol = eh->h_proto;
  78         skb->mac_header += VLAN_HLEN;
  79
  80         return skb;
  81 }
  82
  83
  84 static struct sk_buff *
  85 modify_vlan_tci(struct datapath *dp, struct sk_buff *skb,
  86                 struct odp_flow_key *key, const union odp_action *a,
  87                 int n_actions, gfp_t gfp)
  88 {
  89         u16 tci, mask;
  90
  91         if (a->type == ODPAT_SET_VLAN_VID) {
  92                 tci = ntohs(a->vlan_vid.vlan_vid);
  93                 mask = VLAN_VID_MASK;
  94                 key->dl_vlan = htons(tci & mask);
  95         } else {
  96                 tci = a->vlan_pcp.vlan_pcp << 13;
  97                 mask = VLAN_PCP_MASK;
  98         }
  99
 100         skb = make_writable(skb, VLAN_HLEN, gfp);
 101         if (!skb)
 102                 return ERR_PTR(-ENOMEM);
 103
 104         if (skb->protocol == htons(ETH_P_8021Q)) {
 105                 /* Modify vlan id, but maintain other TCI values */
 106                 struct vlan_ethhdr *vh = vlan_eth_hdr(skb);
 107                 vh->h_vlan_TCI = htons((ntohs(vh->h_vlan_TCI) & ~mask) | tci);
 108         } else {
 109                 /* Add vlan header */
 110
 111                 /* Set up checksumming pointers for checksum-deferred packets
 112                  * on Xen.  Otherwise, dev_queue_xmit() will try to do this
 113                  * when we send the packet out on the wire, and it will fail at
 114                  * that point because skb_checksum_setup() will not look inside
 115                  * an 802.1Q header. */
 116                 vswitch_skb_checksum_setup(skb);
 117
 118                 /* GSO is not implemented for packets with an 802.1Q header, so
 119                  * we have to do segmentation before we add that header.
 120                  *
 121                  * GSO does work with hardware-accelerated VLAN tagging, but we
 122                  * can't use hardware-accelerated VLAN tagging since it
 123                  * requires the device to have a VLAN group configured (with
 124                  * e.g. vconfig(8)) and we don't do that.
 125                  *
 126                  * Having to do this here may be a performance loss, since we
 127                  * can't take advantage of TSO hardware support, although it
 128                  * does not make a measurable network performance difference
 129                  * for 1G Ethernet.  Fixing that would require patching the
 130                  * kernel (either to add GSO support to the VLAN protocol or to
 131                  * support hardware-accelerated VLAN tagging without VLAN
 132                  * groups configured). */
 133                 if (skb_is_gso(skb)) {
 134                         struct sk_buff *segs;
 135
 136                         segs = skb_gso_segment(skb, 0);
 137                         kfree_skb(skb);
 138                         if (unlikely(IS_ERR(segs)))
 139                                 return ERR_CAST(segs);
 140
 141                         do {
 142                                 struct sk_buff *nskb = segs->next;
 143                                 int err;
 144
 145                                 segs->next = NULL;
 146
 147                                 segs = __vlan_put_tag(segs, tci);
 148                                 err = -ENOMEM;
 149                                 if (segs) {
 150                                         struct odp_flow_key segkey = *key;
 151                                         err = execute_actions(dp, segs,
 152                                                               &segkey, a + 1,
 153                                                               n_actions - 1,
 154                                                               gfp);
 155                                 }
 156
 157                                 if (unlikely(err)) {
 158                                         while ((segs = nskb)) {
 159                                                 nskb = segs->next;
 160                                                 segs->next = NULL;
 161                                                 kfree_skb(segs);
 162                                         }
 163                                         return ERR_PTR(err);
 164                                 }
 165
 166                                 segs = nskb;
 167                         } while (segs->next);
 168
 169                         skb = segs;
 170                 }
 171
 172                 /* The hardware-accelerated version of vlan_put_tag() works
 173                  * only for a device that has a VLAN group configured (with
 174                  * e.g. vconfig(8)), so call the software-only version
 175                  * __vlan_put_tag() directly instead.
 176                  */
 177                 skb = __vlan_put_tag(skb, tci);
 178                 if (!skb)
 179                         return ERR_PTR(-ENOMEM);
 180         }
 181
 182         return skb;
 183 }
 184
 185 static struct sk_buff *strip_vlan(struct sk_buff *skb,
 186                                   struct odp_flow_key *key, gfp_t gfp)
 187 {
 188         skb = make_writable(skb, 0, gfp);
 189         if (skb) {
 190                 vlan_pull_tag(skb);
 191                 key->dl_vlan = htons(ODP_VLAN_NONE);
 192         }
 193         return skb;
 194 }
 195
 196 static struct sk_buff *set_dl_addr(struct sk_buff *skb,
 197                                    const struct odp_action_dl_addr *a,
 198                                    gfp_t gfp)
 199 {
 200         skb = make_writable(skb, 0, gfp);
 201         if (skb) {
 202                 struct ethhdr *eh = eth_hdr(skb);
 203                 memcpy(a->type == ODPAT_SET_DL_SRC ? eh->h_source : eh->h_dest,
 204                        a->dl_addr, ETH_ALEN);
 205         }
 206         return skb;
 207 }
 208
 209 /* Updates 'sum', which is a field in 'skb''s data, given that a 4-byte field
 210  * covered by the sum has been changed from 'from' to 'to'.  If set,
 211  * 'pseudohdr' indicates that the field is in the TCP or UDP pseudo-header.
 212  * Based on nf_proto_csum_replace4. */
 213 static void update_csum(__sum16 *sum, struct sk_buff *skb,
 214                         __be32 from, __be32 to, int pseudohdr)
 215 {
 216         __be32 diff[] = { ~from, to };
 217
 218 /* On older kernels, CHECKSUM_PARTIAL and CHECKSUM_COMPLETE are both defined
 219  * as CHECKSUM_HW.  However, we can make some inferences so that we can update
 220  * the checksums appropriately. */
 221         enum {
 222                 CSUM_PARTIAL,   /* Partial checksum, skb->csum undefined. */
 223                 CSUM_PACKET,    /* In-packet checksum, skb->csum undefined. */
 224                 CSUM_COMPLETE,  /* In-packet checksum, skb->csum valid. */
 225         } csum_type;
 226
 227         csum_type = CSUM_PACKET;
 228 #ifndef CHECKSUM_HW
 229         /* Newer kernel, just map between kernel types and ours. */
 230         if (skb->ip_summed == CHECKSUM_PARTIAL)
 231                 csum_type = CSUM_PARTIAL;
 232         else if (skb->ip_summed == CHECKSUM_COMPLETE)
 233                 csum_type = CSUM_COMPLETE;
 234 #else
 235         /* In theory this could be either CHECKSUM_PARTIAL or CHECKSUM_COMPLETE.
 236          * However, we should only get CHECKSUM_PARTIAL packets from Xen, which
 237          * uses some special fields to represent this (see below).  Since we
 238          * can only make one type work, pick the one that actually happens in
 239          * practice. */
 240         if (skb->ip_summed == CHECKSUM_HW)
 241                 csum_type = CSUM_COMPLETE;
 242 #endif
 243 #if defined(CONFIG_XEN) && defined(HAVE_PROTO_DATA_VALID)
 244         /* Xen has a special way of representing CHECKSUM_PARTIAL on older
 245          * kernels. */
 246         if (skb->proto_csum_blank)
 247                 csum_type = CSUM_PARTIAL;
 248 #endif
 249
 250         if (csum_type != CSUM_PARTIAL) {
 251                 *sum = csum_fold(csum_partial((char *)diff, sizeof(diff),
 252                                 ~csum_unfold(*sum)));
 253                 if (csum_type == CSUM_COMPLETE && pseudohdr)
 254                         skb->csum = ~csum_partial((char *)diff, sizeof(diff),
 255                                                 ~skb->csum);
 256         } else if (pseudohdr)
 257                 *sum = ~csum_fold(csum_partial((char *)diff, sizeof(diff),
 258                                 csum_unfold(*sum)));
 259 }
 260
 261 static struct sk_buff *set_nw_addr(struct sk_buff *skb,
 262                                    struct odp_flow_key *key,
 263                                    const struct odp_action_nw_addr *a,
 264                                    gfp_t gfp)
 265 {
 266         if (key->dl_type != htons(ETH_P_IP))
 267                 return skb;
 268
 269         skb = make_writable(skb, 0, gfp);
 270         if (skb) {
 271                 struct iphdr *nh = ip_hdr(skb);
 272                 u32 *f = a->type == ODPAT_SET_NW_SRC ? &nh->saddr : &nh->daddr;
 273                 u32 old = *f;
 274                 u32 new = a->nw_addr;
 275
 276                 if (key->nw_proto == IPPROTO_TCP) {
 277                         struct tcphdr *th = tcp_hdr(skb);
 278                         update_csum(&th->check, skb, old, new, 1);
 279                 } else if (key->nw_proto == IPPROTO_UDP) {
 280                         struct udphdr *th = udp_hdr(skb);
 281                         update_csum(&th->check, skb, old, new, 1);
 282                 }
 283                 update_csum(&nh->check, skb, old, new, 0);
 284                 *f = new;
 285         }
 286         return skb;
 287 }
 288
 289 static struct sk_buff *set_nw_tos(struct sk_buff *skb,
 290                                    struct odp_flow_key *key,
 291                                    const struct odp_action_nw_tos *a,
 292                                    gfp_t gfp)
 293 {
 294         if (key->dl_type != htons(ETH_P_IP))
 295                 return skb;
 296
 297         skb = make_writable(skb, 0, gfp);
 298         if (skb) {
 299                 struct iphdr *nh = ip_hdr(skb);
 300                 u8 *f = &nh->tos;
 301                 u8 old = *f;
 302                 u8 new;
 303
 304                 /* Set the DSCP bits and preserve the ECN bits. */
 305                 new = (a->nw_tos & ~INET_ECN_MASK) | (nh->tos & INET_ECN_MASK);
 306                 update_csum(&nh->check, skb, htons((uint16_t)old),
 307                                 htons((uint16_t)new), 0);
 308                 *f = new;
 309         }
 310         return skb;
 311 }
 312
 313 static struct sk_buff *
 314 set_tp_port(struct sk_buff *skb, struct odp_flow_key *key,
 315             const struct odp_action_tp_port *a,
 316             gfp_t gfp)
 317 {
 318         int check_ofs;
 319
 320         if (key->dl_type != htons(ETH_P_IP))
 321                 return skb;
 322
 323         if (key->nw_proto == IPPROTO_TCP)
 324                 check_ofs = offsetof(struct tcphdr, check);
 325         else if (key->nw_proto == IPPROTO_UDP)
 326                 check_ofs = offsetof(struct udphdr, check);
 327         else
 328                 return skb;
 329
 330         skb = make_writable(skb, 0, gfp);
 331         if (skb) {
 332                 struct udphdr *th = udp_hdr(skb);
 333                 u16 *f = a->type == ODPAT_SET_TP_SRC ? &th->source : &th->dest;
 334                 u16 old = *f;
 335                 u16 new = a->tp_port;
 336                 update_csum((u16*)(skb_transport_header(skb) + check_ofs),
 337                                 skb, old, new, 0);
 338                 *f = new;
 339         }
 340         return skb;
 341 }
 342
 343 static inline unsigned packet_length(const struct sk_buff *skb)
 344 {
 345         unsigned length = skb->len - ETH_HLEN;
 346         if (skb->protocol == htons(ETH_P_8021Q))
 347                 length -= VLAN_HLEN;
 348         return length;
 349 }
 350
 351 int dp_xmit_skb(struct sk_buff *skb)
 352 {
 353         struct datapath *dp = skb->dev->br_port->dp;
 354         int len = skb->len;
 355
 356         if (packet_length(skb) > skb->dev->mtu && !skb_is_gso(skb)) {
 357                 printk(KERN_WARNING "%s: dropped over-mtu packet: %d > %d\n",
 358                        dp_name(dp), packet_length(skb), skb->dev->mtu);
 359                 kfree_skb(skb);
 360                 return -E2BIG;
 361         }
 362
 363         forward_ip_summed(skb);
 364         dev_queue_xmit(skb);
 365
 366         return len;
 367 }
 368
 369 static void
 370 do_output(struct datapath *dp, struct sk_buff *skb, int out_port)
 371 {
 372         struct net_bridge_port *p;
 373         struct net_device *dev;
 374
 375         if (!skb)
 376                 goto error;
 377
 378         p = dp->ports[out_port];
 379         if (!p)
 380                 goto error;
 381
 382         dev = skb->dev = p->dev;
 383         if (is_dp_dev(dev))
 384                 dp_dev_recv(dev, skb);
 385         else
 386                 dp_xmit_skb(skb);
 387         return;
 388
 389 error:
 390         kfree_skb(skb);
 391 }
 392
 393 /* Never consumes 'skb'.  Returns a port that 'skb' should be sent to, -1 if
 394  * none.  */
 395 static int output_group(struct datapath *dp, __u16 group,
 396                         struct sk_buff *skb, gfp_t gfp)
 397 {
 398         struct dp_port_group *g = rcu_dereference(dp->groups[group]);
 399         int prev_port = -1;
 400         int i;
 401
 402         if (!g)
 403                 return -1;
 404         for (i = 0; i < g->n_ports; i++) {
 405                 struct net_bridge_port *p = dp->ports[g->ports[i]];
 406                 if (!p || skb->dev == p->dev)
 407                         continue;
 408                 if (prev_port != -1) {
 409                         struct sk_buff *clone = skb_clone(skb, gfp);
 410                         if (!clone)
 411                                 return -1;
 412                         do_output(dp, clone, prev_port);
 413                 }
 414                 prev_port = p->port_no;
 415         }
 416         return prev_port;
 417 }
 418
 419 static int
 420 output_control(struct datapath *dp, struct sk_buff *skb, u32 arg, gfp_t gfp)
 421 {
 422         skb = skb_clone(skb, gfp);
 423         if (!skb)
 424                 return -ENOMEM;
 425         return dp_output_control(dp, skb, _ODPL_ACTION_NR, arg);
 426 }
 427
 428 /* Send a copy of this packet up to the sFlow agent, along with extra
 429  * information about what happened to it. */
 430 static void sflow_sample(struct datapath *dp, struct sk_buff *skb,
 431                          const union odp_action *a, int n_actions,
 432                          gfp_t gfp, struct net_bridge_port *nbp)
 433 {
 434         struct odp_sflow_sample_header *hdr;
 435         unsigned int actlen = n_actions * sizeof(union odp_action);
 436         unsigned int hdrlen = sizeof(struct odp_sflow_sample_header);
 437         struct sk_buff *nskb;
 438
 439         nskb = skb_copy_expand(skb, actlen + hdrlen, 0, gfp);
 440         if (!nskb)
 441                 return;
 442
 443         memcpy(__skb_push(nskb, actlen), a, actlen);
 444         hdr = (struct odp_sflow_sample_header*)__skb_push(nskb, hdrlen);
 445         hdr->n_actions = n_actions;
 446         hdr->sample_pool = atomic_read(&nbp->sflow_pool);
 447         dp_output_control(dp, nskb, _ODPL_SFLOW_NR, 0);
 448 }
 449
 450 /* Execute a list of actions against 'skb'. */
 451 int execute_actions(struct datapath *dp, struct sk_buff *skb,
 452                     struct odp_flow_key *key,
 453                     const union odp_action *a, int n_actions,
 454                     gfp_t gfp)
 455 {
 456         /* Every output action needs a separate clone of 'skb', but the common
 457          * case is just a single output action, so that doing a clone and
 458          * then freeing the original skbuff is wasteful.  So the following code
 459          * is slightly obscure just to avoid that. */
 460         int prev_port = -1;
 461         int err;
 462
 463         if (dp->sflow_probability) {
 464                 struct net_bridge_port *p = skb->dev->br_port;
 465                 if (p) {
 466                         atomic_inc(&p->sflow_pool);
 467                         if (dp->sflow_probability == UINT_MAX ||
 468                             net_random() < dp->sflow_probability)
 469                                 sflow_sample(dp, skb, a, n_actions, gfp, p);
 470                 }
 471         }
 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_VLAN_VID:
 499                 case ODPAT_SET_VLAN_PCP:
 500                         skb = modify_vlan_tci(dp, skb, key, a, n_actions, gfp);
 501                         if (IS_ERR(skb))
 502                                 return PTR_ERR(skb);
 503                         break;
 504
 505                 case ODPAT_STRIP_VLAN:
 506                         skb = strip_vlan(skb, key, gfp);
 507                         break;
 508
 509                 case ODPAT_SET_DL_SRC:
 510                 case ODPAT_SET_DL_DST:
 511                         skb = set_dl_addr(skb, &a->dl_addr, gfp);
 512                         break;
 513
 514                 case ODPAT_SET_NW_SRC:
 515                 case ODPAT_SET_NW_DST:
 516                         skb = set_nw_addr(skb, key, &a->nw_addr, gfp);
 517                         break;
 518
 519                 case ODPAT_SET_NW_TOS:
 520                         skb = set_nw_tos(skb, key, &a->nw_tos, gfp);
 521                         break;
 522
 523                 case ODPAT_SET_TP_SRC:
 524                 case ODPAT_SET_TP_DST:
 525                         skb = set_tp_port(skb, key, &a->tp_port, gfp);
 526                         break;
 527                 }
 528                 if (!skb)
 529                         return -ENOMEM;
 530         }
 531         if (prev_port != -1)
 532                 do_output(dp, skb, prev_port);
 533         else
 534                 kfree_skb(skb);
 535         return 0;
 536 }