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_arp.h>
  18 #include <linux/if_vlan.h>
  19 #include <net/inet_ecn.h>
  20 #include <net/ip.h>
  21 #include <net/checksum.h>
  22
  23 #include "actions.h"
  24 #include "datapath.h"
  25 #include "openvswitch/datapath-protocol.h"
  26 #include "vport.h"
  27
  28 static struct sk_buff *make_writable(struct sk_buff *skb, unsigned min_headroom, gfp_t gfp)
  29 {
  30         if (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                         set_skb_csum_bits(skb, nskb);
  37                         kfree_skb(skb);
  38                         return nskb;
  39                 }
  40         } else {
  41                 unsigned int hdr_len = (skb_transport_offset(skb)
  42                                         + sizeof(struct tcphdr));
  43                 if (pskb_may_pull(skb, min(hdr_len, skb->len)))
  44                         return skb;
  45         }
  46         kfree_skb(skb);
  47         return NULL;
  48 }
  49
  50 static struct sk_buff *vlan_pull_tag(struct sk_buff *skb)
  51 {
  52         struct vlan_ethhdr *vh = vlan_eth_hdr(skb);
  53         struct ethhdr *eh;
  54
  55         /* Verify we were given a vlan packet */
  56         if (vh->h_vlan_proto != htons(ETH_P_8021Q))
  57                 return skb;
  58
  59         if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE)
  60                 skb->csum = csum_sub(skb->csum, csum_partial(skb->data
  61                                         + ETH_HLEN, VLAN_HLEN, 0));
  62
  63         memmove(skb->data + VLAN_HLEN, skb->data, 2 * VLAN_ETH_ALEN);
  64
  65         eh = (struct ethhdr *)skb_pull(skb, VLAN_HLEN);
  66
  67         skb->protocol = eh->h_proto;
  68         skb->mac_header += VLAN_HLEN;
  69
  70         return skb;
  71 }
  72
  73 static struct sk_buff *modify_vlan_tci(struct datapath *dp, struct sk_buff *skb,
  74                                        const struct odp_flow_key *key,
  75                                        const union odp_action *a, int n_actions,
  76                                        gfp_t gfp)
  77 {
  78         u16 tci, mask;
  79
  80         if (a->type == ODPAT_SET_VLAN_VID) {
  81                 tci = ntohs(a->vlan_vid.vlan_vid);
  82                 mask = VLAN_VID_MASK;
  83         } else {
  84                 tci = a->vlan_pcp.vlan_pcp << VLAN_PCP_SHIFT;
  85                 mask = VLAN_PCP_MASK;
  86         }
  87
  88         skb = make_writable(skb, VLAN_HLEN, gfp);
  89         if (!skb)
  90                 return ERR_PTR(-ENOMEM);
  91
  92         if (skb->protocol == htons(ETH_P_8021Q)) {
  93                 /* Modify vlan id, but maintain other TCI values */
  94                 struct vlan_ethhdr *vh = vlan_eth_hdr(skb);
  95                 __be16 old_tci = vh->h_vlan_TCI;
  96
  97                 vh->h_vlan_TCI = htons((ntohs(vh->h_vlan_TCI) & ~mask) | tci);
  98
  99                 if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE) {
 100                         __be16 diff[] = { ~old_tci, vh->h_vlan_TCI };
 101
 102                         skb->csum = ~csum_partial((char *)diff, sizeof(diff),
 103                                                 ~skb->csum);
 104                 }
 105         } else {
 106                 int err;
 107
 108                 /* Add vlan header */
 109
 110                 /* Set up checksumming pointers for checksum-deferred packets
 111                  * on Xen.  Otherwise, dev_queue_xmit() will try to do this
 112                  * when we send the packet out on the wire, and it will fail at
 113                  * that point because skb_checksum_setup() will not look inside
 114                  * an 802.1Q header. */
 115                 err = vswitch_skb_checksum_setup(skb);
 116                 if (unlikely(err)) {
 117                         kfree_skb(skb);
 118                         return ERR_PTR(err);
 119                 }
 120
 121                 /* GSO is not implemented for packets with an 802.1Q header, so
 122                  * we have to do segmentation before we add that header.
 123                  *
 124                  * GSO does work with hardware-accelerated VLAN tagging, but we
 125                  * can't use hardware-accelerated VLAN tagging since it
 126                  * requires the device to have a VLAN group configured (with
 127                  * e.g. vconfig(8)) and we don't do that.
 128                  *
 129                  * Having to do this here may be a performance loss, since we
 130                  * can't take advantage of TSO hardware support, although it
 131                  * does not make a measurable network performance difference
 132                  * for 1G Ethernet.  Fixing that would require patching the
 133                  * kernel (either to add GSO support to the VLAN protocol or to
 134                  * support hardware-accelerated VLAN tagging without VLAN
 135                  * groups configured). */
 136                 if (skb_is_gso(skb)) {
 137                         struct sk_buff *segs;
 138
 139                         segs = skb_gso_segment(skb, 0);
 140                         kfree_skb(skb);
 141                         if (unlikely(IS_ERR(segs)))
 142                                 return ERR_CAST(segs);
 143
 144                         do {
 145                                 struct sk_buff *nskb = segs->next;
 146                                 int err;
 147
 148                                 segs->next = NULL;
 149
 150                                 /* GSO can change the checksum type so update.*/
 151                                 compute_ip_summed(segs, true);
 152
 153                                 segs = __vlan_put_tag(segs, tci);
 154                                 err = -ENOMEM;
 155                                 if (segs) {
 156                                         err = execute_actions(dp, segs,
 157                                                               key, a + 1,
 158                                                               n_actions - 1,
 159                                                               gfp);
 160                                 }
 161
 162                                 if (unlikely(err)) {
 163                                         while ((segs = nskb)) {
 164                                                 nskb = segs->next;
 165                                                 segs->next = NULL;
 166                                                 kfree_skb(segs);
 167                                         }
 168                                         return ERR_PTR(err);
 169                                 }
 170
 171                                 segs = nskb;
 172                         } while (segs->next);
 173
 174                         skb = segs;
 175                         compute_ip_summed(skb, true);
 176                 }
 177
 178                 /* The hardware-accelerated version of vlan_put_tag() works
 179                  * only for a device that has a VLAN group configured (with
 180                  * e.g. vconfig(8)), so call the software-only version
 181                  * __vlan_put_tag() directly instead.
 182                  */
 183                 skb = __vlan_put_tag(skb, tci);
 184                 if (!skb)
 185                         return ERR_PTR(-ENOMEM);
 186
 187                 /* GSO doesn't fix up the hardware computed checksum so this
 188                  * will only be hit in the non-GSO case. */
 189                 if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE)
 190                         skb->csum = csum_add(skb->csum, csum_partial(skb->data
 191                                                 + ETH_HLEN, VLAN_HLEN, 0));
 192         }
 193
 194         return skb;
 195 }
 196
 197 static struct sk_buff *strip_vlan(struct sk_buff *skb, gfp_t gfp)
 198 {
 199         skb = make_writable(skb, 0, gfp);
 200         if (skb)
 201                 vlan_pull_tag(skb);
 202
 203         return skb;
 204 }
 205
 206 static struct sk_buff *set_dl_addr(struct sk_buff *skb,
 207                                    const struct odp_action_dl_addr *a,
 208                                    gfp_t gfp)
 209 {
 210         skb = make_writable(skb, 0, gfp);
 211         if (skb) {
 212                 struct ethhdr *eh = eth_hdr(skb);
 213                 if (a->type == ODPAT_SET_DL_SRC)
 214                         memcpy(eh->h_source, a->dl_addr, ETH_ALEN);
 215                 else
 216                         memcpy(eh->h_dest, a->dl_addr, ETH_ALEN);
 217         }
 218         return skb;
 219 }
 220
 221 /* Updates 'sum', which is a field in 'skb''s data, given that a 4-byte field
 222  * covered by the sum has been changed from 'from' to 'to'.  If set,
 223  * 'pseudohdr' indicates that the field is in the TCP or UDP pseudo-header.
 224  * Based on nf_proto_csum_replace4. */
 225 static void update_csum(__sum16 *sum, struct sk_buff *skb,
 226                         __be32 from, __be32 to, int pseudohdr)
 227 {
 228         __be32 diff[] = { ~from, to };
 229
 230         if (OVS_CB(skb)->ip_summed != OVS_CSUM_PARTIAL) {
 231                 *sum = csum_fold(csum_partial((char *)diff, sizeof(diff),
 232                                 ~csum_unfold(*sum)));
 233                 if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE && pseudohdr)
 234                         skb->csum = ~csum_partial((char *)diff, sizeof(diff),
 235                                                 ~skb->csum);
 236         } else if (pseudohdr)
 237                 *sum = ~csum_fold(csum_partial((char *)diff, sizeof(diff),
 238                                 csum_unfold(*sum)));
 239 }
 240
 241 static struct sk_buff *set_nw_addr(struct sk_buff *skb,
 242                                    const struct odp_flow_key *key,
 243                                    const struct odp_action_nw_addr *a,
 244                                    gfp_t gfp)
 245 {
 246         if (key->dl_type != htons(ETH_P_IP))
 247                 return skb;
 248
 249         skb = make_writable(skb, 0, gfp);
 250         if (skb) {
 251                 struct iphdr *nh = ip_hdr(skb);
 252                 u32 *f = a->type == ODPAT_SET_NW_SRC ? &nh->saddr : &nh->daddr;
 253                 u32 old = *f;
 254                 u32 new = a->nw_addr;
 255
 256                 if (key->nw_proto == IPPROTO_TCP) {
 257                         struct tcphdr *th = tcp_hdr(skb);
 258                         update_csum(&th->check, skb, old, new, 1);
 259                 } else if (key->nw_proto == IPPROTO_UDP) {
 260                         struct udphdr *th = udp_hdr(skb);
 261                         update_csum(&th->check, skb, old, new, 1);
 262                 }
 263                 update_csum(&nh->check, skb, old, new, 0);
 264                 *f = new;
 265         }
 266         return skb;
 267 }
 268
 269 static struct sk_buff *set_nw_tos(struct sk_buff *skb,
 270                                    const struct odp_flow_key *key,
 271                                    const struct odp_action_nw_tos *a,
 272                                    gfp_t gfp)
 273 {
 274         if (key->dl_type != htons(ETH_P_IP))
 275                 return skb;
 276
 277         skb = make_writable(skb, 0, gfp);
 278         if (skb) {
 279                 struct iphdr *nh = ip_hdr(skb);
 280                 u8 *f = &nh->tos;
 281                 u8 old = *f;
 282                 u8 new;
 283
 284                 /* Set the DSCP bits and preserve the ECN bits. */
 285                 new = a->nw_tos | (nh->tos & INET_ECN_MASK);
 286                 update_csum(&nh->check, skb, htons((uint16_t)old),
 287                                 htons((uint16_t)new), 0);
 288                 *f = new;
 289         }
 290         return skb;
 291 }
 292
 293 static struct sk_buff *set_tp_port(struct sk_buff *skb,
 294                                    const struct odp_flow_key *key,
 295                                    const struct odp_action_tp_port *a, gfp_t gfp)
 296 {
 297         int check_ofs;
 298
 299         if (key->dl_type != htons(ETH_P_IP))
 300                 return skb;
 301
 302         if (key->nw_proto == IPPROTO_TCP)
 303                 check_ofs = offsetof(struct tcphdr, check);
 304         else if (key->nw_proto == IPPROTO_UDP)
 305                 check_ofs = offsetof(struct udphdr, check);
 306         else
 307                 return skb;
 308
 309         skb = make_writable(skb, 0, gfp);
 310         if (skb) {
 311                 struct udphdr *th = udp_hdr(skb);
 312                 u16 *f = a->type == ODPAT_SET_TP_SRC ? &th->source : &th->dest;
 313                 u16 old = *f;
 314                 u16 new = a->tp_port;
 315                 update_csum((u16*)(skb_transport_header(skb) + check_ofs),
 316                                 skb, old, new, 0);
 317                 *f = new;
 318         }
 319         return skb;
 320 }
 321
 322 /**
 323  * is_spoofed_arp - check for invalid ARP packet
 324  *
 325  * @skb: skbuff containing an Ethernet packet, with network header pointing
 326  * just past the Ethernet and optional 802.1Q header.
 327  * @key: flow key extracted from @skb by flow_extract()
 328  *
 329  * Returns true if @skb is an invalid Ethernet+IPv4 ARP packet: one with screwy
 330  * or truncated header fields or one whose inner and outer Ethernet address
 331  * differ.
 332  */
 333 static bool is_spoofed_arp(struct sk_buff *skb, const struct odp_flow_key *key)
 334 {
 335         struct arp_eth_header *arp;
 336
 337         if (key->dl_type != htons(ETH_P_ARP))
 338                 return false;
 339
 340         if (skb_network_offset(skb) + sizeof(struct arp_eth_header) > skb->len)
 341                 return true;
 342
 343         arp = (struct arp_eth_header *)skb_network_header(skb);
 344         return (arp->ar_hrd != htons(ARPHRD_ETHER) ||
 345                 arp->ar_pro != htons(ETH_P_IP) ||
 346                 arp->ar_hln != ETH_ALEN ||
 347                 arp->ar_pln != 4 ||
 348                 compare_ether_addr(arp->ar_sha, eth_hdr(skb)->h_source));
 349 }
 350
 351 static void do_output(struct datapath *dp, struct sk_buff *skb, int out_port)
 352 {
 353         struct dp_port *p;
 354
 355         if (!skb)
 356                 goto error;
 357
 358         p = rcu_dereference(dp->ports[out_port]);
 359         if (!p)
 360                 goto error;
 361
 362         vport_send(p->vport, skb);
 363         return;
 364
 365 error:
 366         kfree_skb(skb);
 367 }
 368
 369 /* Never consumes 'skb'.  Returns a port that 'skb' should be sent to, -1 if
 370  * none.  */
 371 static int output_group(struct datapath *dp, __u16 group,
 372                         struct sk_buff *skb, gfp_t gfp)
 373 {
 374         struct dp_port_group *g = rcu_dereference(dp->groups[group]);
 375         int prev_port = -1;
 376         int i;
 377
 378         if (!g)
 379                 return -1;
 380         for (i = 0; i < g->n_ports; i++) {
 381                 struct dp_port *p = rcu_dereference(dp->ports[g->ports[i]]);
 382                 if (!p || OVS_CB(skb)->dp_port == p)
 383                         continue;
 384                 if (prev_port != -1) {
 385                         struct sk_buff *clone = skb_clone(skb, gfp);
 386                         if (!clone)
 387                                 return -1;
 388                         do_output(dp, clone, prev_port);
 389                 }
 390                 prev_port = p->port_no;
 391         }
 392         return prev_port;
 393 }
 394
 395 static int output_control(struct datapath *dp, struct sk_buff *skb, u32 arg,
 396                           gfp_t gfp)
 397 {
 398         skb = skb_clone(skb, gfp);
 399         if (!skb)
 400                 return -ENOMEM;
 401         return dp_output_control(dp, skb, _ODPL_ACTION_NR, arg);
 402 }
 403
 404 /* Send a copy of this packet up to the sFlow agent, along with extra
 405  * information about what happened to it. */
 406 static void sflow_sample(struct datapath *dp, struct sk_buff *skb,
 407                          const union odp_action *a, int n_actions,
 408                          gfp_t gfp, struct dp_port *dp_port)
 409 {
 410         struct odp_sflow_sample_header *hdr;
 411         unsigned int actlen = n_actions * sizeof(union odp_action);
 412         unsigned int hdrlen = sizeof(struct odp_sflow_sample_header);
 413         struct sk_buff *nskb;
 414
 415         nskb = skb_copy_expand(skb, actlen + hdrlen, 0, gfp);
 416         if (!nskb)
 417                 return;
 418
 419         memcpy(__skb_push(nskb, actlen), a, actlen);
 420         hdr = (struct odp_sflow_sample_header*)__skb_push(nskb, hdrlen);
 421         hdr->n_actions = n_actions;
 422         hdr->sample_pool = atomic_read(&dp_port->sflow_pool);
 423         dp_output_control(dp, nskb, _ODPL_SFLOW_NR, 0);
 424 }
 425
 426 /* Execute a list of actions against 'skb'. */
 427 int execute_actions(struct datapath *dp, struct sk_buff *skb,
 428                     const struct odp_flow_key *key,
 429                     const union odp_action *a, int n_actions,
 430                     gfp_t gfp)
 431 {
 432         /* Every output action needs a separate clone of 'skb', but the common
 433          * case is just a single output action, so that doing a clone and
 434          * then freeing the original skbuff is wasteful.  So the following code
 435          * is slightly obscure just to avoid that. */
 436         int prev_port = -1;
 437         u32 priority = skb->priority;
 438         int err;
 439
 440         if (dp->sflow_probability) {
 441                 struct dp_port *p = OVS_CB(skb)->dp_port;
 442                 if (p) {
 443                         atomic_inc(&p->sflow_pool);
 444                         if (dp->sflow_probability == UINT_MAX ||
 445                             net_random() < dp->sflow_probability)
 446                                 sflow_sample(dp, skb, a, n_actions, gfp, p);
 447                 }
 448         }
 449
 450         OVS_CB(skb)->tun_id = 0;
 451
 452         for (; n_actions > 0; a++, n_actions--) {
 453                 if (prev_port != -1) {
 454                         do_output(dp, skb_clone(skb, gfp), prev_port);
 455                         prev_port = -1;
 456                 }
 457
 458                 switch (a->type) {
 459                 case ODPAT_OUTPUT:
 460                         prev_port = a->output.port;
 461                         break;
 462
 463                 case ODPAT_OUTPUT_GROUP:
 464                         prev_port = output_group(dp, a->output_group.group,
 465                                                  skb, gfp);
 466                         break;
 467
 468                 case ODPAT_CONTROLLER:
 469                         err = output_control(dp, skb, a->controller.arg, gfp);
 470                         if (err) {
 471                                 kfree_skb(skb);
 472                                 return err;
 473                         }
 474                         break;
 475
 476                 case ODPAT_SET_TUNNEL:
 477                         OVS_CB(skb)->tun_id = a->tunnel.tun_id;
 478                         break;
 479
 480                 case ODPAT_SET_VLAN_VID:
 481                 case ODPAT_SET_VLAN_PCP:
 482                         skb = modify_vlan_tci(dp, skb, key, a, n_actions, gfp);
 483                         if (IS_ERR(skb))
 484                                 return PTR_ERR(skb);
 485                         break;
 486
 487                 case ODPAT_STRIP_VLAN:
 488                         skb = strip_vlan(skb, gfp);
 489                         break;
 490
 491                 case ODPAT_SET_DL_SRC:
 492                 case ODPAT_SET_DL_DST:
 493                         skb = set_dl_addr(skb, &a->dl_addr, gfp);
 494                         break;
 495
 496                 case ODPAT_SET_NW_SRC:
 497                 case ODPAT_SET_NW_DST:
 498                         skb = set_nw_addr(skb, key, &a->nw_addr, gfp);
 499                         break;
 500
 501                 case ODPAT_SET_NW_TOS:
 502                         skb = set_nw_tos(skb, key, &a->nw_tos, gfp);
 503                         break;
 504
 505                 case ODPAT_SET_TP_SRC:
 506                 case ODPAT_SET_TP_DST:
 507                         skb = set_tp_port(skb, key, &a->tp_port, gfp);
 508                         break;
 509
 510                 case ODPAT_SET_PRIORITY:
 511                         skb->priority = a->priority.priority;
 512                         break;
 513
 514                 case ODPAT_POP_PRIORITY:
 515                         skb->priority = priority;
 516                         break;
 517
 518                 case ODPAT_DROP_SPOOFED_ARP:
 519                         if (unlikely(is_spoofed_arp(skb, key)))
 520                                 goto exit;
 521                         break;
 522                 }
 523                 if (!skb)
 524                         return -ENOMEM;
 525         }
 526 exit:
 527         if (prev_port != -1)
 528                 do_output(dp, skb, prev_port);
 529         else
 530                 kfree_skb(skb);
 531         return 0;
 532 }