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 "checksum.h"
  25 #include "datapath.h"
  26 #include "openvswitch/datapath-protocol.h"
  27 #include "vport.h"
  28
  29 static int do_execute_actions(struct datapath *, struct sk_buff *,
  30                               const struct sw_flow_key *,
  31                               const struct nlattr *actions, u32 actions_len);
  32
  33 static struct sk_buff *make_writable(struct sk_buff *skb, unsigned min_headroom)
  34 {
  35         if (skb_cloned(skb)) {
  36                 struct sk_buff *nskb;
  37                 unsigned headroom = max(min_headroom, skb_headroom(skb));
  38
  39                 nskb = skb_copy_expand(skb, headroom, skb_tailroom(skb), GFP_ATOMIC);
  40                 if (nskb) {
  41                         set_skb_csum_bits(skb, nskb);
  42                         kfree_skb(skb);
  43                         return nskb;
  44                 }
  45         } else {
  46                 unsigned int hdr_len = (skb_transport_offset(skb)
  47                                         + sizeof(struct tcphdr));
  48                 if (pskb_may_pull(skb, min(hdr_len, skb->len)))
  49                         return skb;
  50         }
  51         kfree_skb(skb);
  52         return NULL;
  53 }
  54
  55 static struct sk_buff *vlan_pull_tag(struct sk_buff *skb)
  56 {
  57         struct vlan_ethhdr *vh = vlan_eth_hdr(skb);
  58         struct ethhdr *eh;
  59
  60         /* Verify we were given a vlan packet */
  61         if (vh->h_vlan_proto != htons(ETH_P_8021Q) || skb->len < VLAN_ETH_HLEN)
  62                 return skb;
  63
  64         if (get_ip_summed(skb) == OVS_CSUM_COMPLETE)
  65                 skb->csum = csum_sub(skb->csum, csum_partial(skb->data
  66                                         + ETH_HLEN, VLAN_HLEN, 0));
  67
  68         memmove(skb->data + VLAN_HLEN, skb->data, 2 * VLAN_ETH_ALEN);
  69
  70         eh = (struct ethhdr *)skb_pull(skb, VLAN_HLEN);
  71
  72         skb->protocol = eh->h_proto;
  73         skb->mac_header += VLAN_HLEN;
  74
  75         return skb;
  76 }
  77
  78 static struct sk_buff *modify_vlan_tci(struct datapath *dp, struct sk_buff *skb,
  79                                        const struct sw_flow_key *key,
  80                                        const struct nlattr *a, u32 actions_len)
  81 {
  82         __be16 tci = nla_get_be16(a);
  83
  84         skb = make_writable(skb, VLAN_HLEN);
  85         if (!skb)
  86                 return ERR_PTR(-ENOMEM);
  87
  88         if (skb->protocol == htons(ETH_P_8021Q)) {
  89                 /* Modify vlan id, but maintain other TCI values */
  90                 struct vlan_ethhdr *vh;
  91                 __be16 old_tci;
  92
  93                 if (skb->len < VLAN_ETH_HLEN)
  94                         return skb;
  95
  96                 vh = vlan_eth_hdr(skb);
  97                 old_tci = vh->h_vlan_TCI;
  98
  99                 vh->h_vlan_TCI = tci;
 100
 101                 if (get_ip_summed(skb) == OVS_CSUM_COMPLETE) {
 102                         __be16 diff[] = { ~old_tci, vh->h_vlan_TCI };
 103
 104                         skb->csum = ~csum_partial((char *)diff, sizeof(diff),
 105                                                 ~skb->csum);
 106                 }
 107         } else {
 108                 int err;
 109
 110                 /* Add vlan header */
 111
 112                 /* Set up checksumming pointers for checksum-deferred packets
 113                  * on Xen.  Otherwise, dev_queue_xmit() will try to do this
 114                  * when we send the packet out on the wire, and it will fail at
 115                  * that point because skb_checksum_setup() will not look inside
 116                  * an 802.1Q header. */
 117                 err = vswitch_skb_checksum_setup(skb);
 118                 if (unlikely(err)) {
 119                         kfree_skb(skb);
 120                         return ERR_PTR(err);
 121                 }
 122
 123                 /* GSO is not implemented for packets with an 802.1Q header, so
 124                  * we have to do segmentation before we add that header.
 125                  *
 126                  * GSO does work with hardware-accelerated VLAN tagging, but we
 127                  * can't use hardware-accelerated VLAN tagging since it
 128                  * requires the device to have a VLAN group configured (with
 129                  * e.g. vconfig(8)) and we don't do that.
 130                  *
 131                  * Having to do this here may be a performance loss, since we
 132                  * can't take advantage of TSO hardware support, although it
 133                  * does not make a measurable network performance difference
 134                  * for 1G Ethernet.  Fixing that would require patching the
 135                  * kernel (either to add GSO support to the VLAN protocol or to
 136                  * support hardware-accelerated VLAN tagging without VLAN
 137                  * groups configured). */
 138                 if (skb_is_gso(skb)) {
 139                         const struct nlattr *actions_left;
 140                         int actions_len_left;
 141                         struct sk_buff *segs;
 142
 143                         segs = skb_gso_segment(skb, 0);
 144                         kfree_skb(skb);
 145                         if (IS_ERR(segs))
 146                                 return ERR_CAST(segs);
 147
 148                         actions_len_left = actions_len;
 149                         actions_left = nla_next(a, &actions_len_left);
 150
 151                         do {
 152                                 struct sk_buff *nskb = segs->next;
 153
 154                                 segs->next = NULL;
 155
 156                                 /* GSO can change the checksum type so update.*/
 157                                 compute_ip_summed(segs, true);
 158
 159                                 segs = __vlan_put_tag(segs, ntohs(tci));
 160                                 err = -ENOMEM;
 161                                 if (segs) {
 162                                         err = do_execute_actions(
 163                                                 dp, segs, key, actions_left,
 164                                                 actions_len_left);
 165                                 }
 166
 167                                 if (unlikely(err)) {
 168                                         while ((segs = nskb)) {
 169                                                 nskb = segs->next;
 170                                                 segs->next = NULL;
 171                                                 kfree_skb(segs);
 172                                         }
 173                                         return ERR_PTR(err);
 174                                 }
 175
 176                                 segs = nskb;
 177                         } while (segs->next);
 178
 179                         skb = segs;
 180                         compute_ip_summed(skb, true);
 181                 }
 182
 183                 /* The hardware-accelerated version of vlan_put_tag() works
 184                  * only for a device that has a VLAN group configured (with
 185                  * e.g. vconfig(8)), so call the software-only version
 186                  * __vlan_put_tag() directly instead.
 187                  */
 188                 skb = __vlan_put_tag(skb, ntohs(tci));
 189                 if (!skb)
 190                         return ERR_PTR(-ENOMEM);
 191
 192                 /* GSO doesn't fix up the hardware computed checksum so this
 193                  * will only be hit in the non-GSO case. */
 194                 if (get_ip_summed(skb) == OVS_CSUM_COMPLETE)
 195                         skb->csum = csum_add(skb->csum, csum_partial(skb->data
 196                                                 + ETH_HLEN, VLAN_HLEN, 0));
 197         }
 198
 199         return skb;
 200 }
 201
 202 static struct sk_buff *strip_vlan(struct sk_buff *skb)
 203 {
 204         skb = make_writable(skb, 0);
 205         if (skb)
 206                 vlan_pull_tag(skb);
 207         return skb;
 208 }
 209
 210 static bool is_ip(struct sk_buff *skb, const struct sw_flow_key *key)
 211 {
 212         return (key->dl_type == htons(ETH_P_IP) &&
 213                 skb->transport_header > skb->network_header);
 214 }
 215
 216 static __sum16 *get_l4_checksum(struct sk_buff *skb, const struct sw_flow_key *key)
 217 {
 218         int transport_len = skb->len - skb_transport_offset(skb);
 219         if (key->nw_proto == IPPROTO_TCP) {
 220                 if (likely(transport_len >= sizeof(struct tcphdr)))
 221                         return &tcp_hdr(skb)->check;
 222         } else if (key->nw_proto == IPPROTO_UDP) {
 223                 if (likely(transport_len >= sizeof(struct udphdr)))
 224                         return &udp_hdr(skb)->check;
 225         }
 226         return NULL;
 227 }
 228
 229 static struct sk_buff *set_nw_addr(struct sk_buff *skb,
 230                                    const struct sw_flow_key *key,
 231                                    const struct nlattr *a)
 232 {
 233         __be32 new_nwaddr = nla_get_be32(a);
 234         struct iphdr *nh;
 235         __sum16 *check;
 236         __be32 *nwaddr;
 237
 238         if (unlikely(!is_ip(skb, key)))
 239                 return skb;
 240
 241         skb = make_writable(skb, 0);
 242         if (unlikely(!skb))
 243                 return NULL;
 244
 245         nh = ip_hdr(skb);
 246         nwaddr = nla_type(a) == ODPAT_SET_NW_SRC ? &nh->saddr : &nh->daddr;
 247
 248         check = get_l4_checksum(skb, key);
 249         if (likely(check))
 250                 inet_proto_csum_replace4(check, skb, *nwaddr, new_nwaddr, 1);
 251         csum_replace4(&nh->check, *nwaddr, new_nwaddr);
 252
 253         *nwaddr = new_nwaddr;
 254
 255         return skb;
 256 }
 257
 258 static struct sk_buff *set_nw_tos(struct sk_buff *skb,
 259                                   const struct sw_flow_key *key,
 260                                   u8 nw_tos)
 261 {
 262         if (unlikely(!is_ip(skb, key)))
 263                 return skb;
 264
 265         skb = make_writable(skb, 0);
 266         if (skb) {
 267                 struct iphdr *nh = ip_hdr(skb);
 268                 u8 *f = &nh->tos;
 269                 u8 old = *f;
 270                 u8 new;
 271
 272                 /* Set the DSCP bits and preserve the ECN bits. */
 273                 new = nw_tos | (nh->tos & INET_ECN_MASK);
 274                 csum_replace4(&nh->check, (__force __be32)old,
 275                                           (__force __be32)new);
 276                 *f = new;
 277         }
 278         return skb;
 279 }
 280
 281 static struct sk_buff *set_tp_port(struct sk_buff *skb,
 282                                    const struct sw_flow_key *key,
 283                                    const struct nlattr *a)
 284 {
 285         struct udphdr *th;
 286         __sum16 *check;
 287         __be16 *port;
 288
 289         if (unlikely(!is_ip(skb, key)))
 290                 return skb;
 291
 292         skb = make_writable(skb, 0);
 293         if (unlikely(!skb))
 294                 return NULL;
 295
 296         /* Must follow make_writable() since that can move the skb data. */
 297         check = get_l4_checksum(skb, key);
 298         if (unlikely(!check))
 299                 return skb;
 300
 301         /*
 302          * Update port and checksum.
 303          *
 304          * This is OK because source and destination port numbers are at the
 305          * same offsets in both UDP and TCP headers, and get_l4_checksum() only
 306          * supports those protocols.
 307          */
 308         th = udp_hdr(skb);
 309         port = nla_type(a) == ODPAT_SET_TP_SRC ? &th->source : &th->dest;
 310         inet_proto_csum_replace2(check, skb, *port, nla_get_be16(a), 0);
 311         *port = nla_get_be16(a);
 312
 313         return skb;
 314 }
 315
 316 /**
 317  * is_spoofed_arp - check for invalid ARP packet
 318  *
 319  * @skb: skbuff containing an Ethernet packet, with network header pointing
 320  * just past the Ethernet and optional 802.1Q header.
 321  * @key: flow key extracted from @skb by flow_extract()
 322  *
 323  * Returns true if @skb is an invalid Ethernet+IPv4 ARP packet: one with screwy
 324  * or truncated header fields or one whose inner and outer Ethernet address
 325  * differ.
 326  */
 327 static bool is_spoofed_arp(struct sk_buff *skb, const struct sw_flow_key *key)
 328 {
 329         struct arp_eth_header *arp;
 330
 331         if (key->dl_type != htons(ETH_P_ARP))
 332                 return false;
 333
 334         if (skb_network_offset(skb) + sizeof(struct arp_eth_header) > skb->len)
 335                 return true;
 336
 337         arp = (struct arp_eth_header *)skb_network_header(skb);
 338         return (arp->ar_hrd != htons(ARPHRD_ETHER) ||
 339                 arp->ar_pro != htons(ETH_P_IP) ||
 340                 arp->ar_hln != ETH_ALEN ||
 341                 arp->ar_pln != 4 ||
 342                 compare_ether_addr(arp->ar_sha, eth_hdr(skb)->h_source));
 343 }
 344
 345 static void do_output(struct datapath *dp, struct sk_buff *skb, int out_port)
 346 {
 347         struct vport *p;
 348
 349         if (!skb)
 350                 goto error;
 351
 352         p = rcu_dereference(dp->ports[out_port]);
 353         if (!p)
 354                 goto error;
 355
 356         vport_send(p, skb);
 357         return;
 358
 359 error:
 360         kfree_skb(skb);
 361 }
 362
 363 static int output_control(struct datapath *dp, struct sk_buff *skb, u64 arg)
 364 {
 365         skb = skb_clone(skb, GFP_ATOMIC);
 366         if (!skb)
 367                 return -ENOMEM;
 368         return dp_output_control(dp, skb, _ODPL_ACTION_NR, arg);
 369 }
 370
 371 /* Execute a list of actions against 'skb'. */
 372 static int do_execute_actions(struct datapath *dp, struct sk_buff *skb,
 373                               const struct sw_flow_key *key,
 374                               const struct nlattr *actions, u32 actions_len)
 375 {
 376         /* Every output action needs a separate clone of 'skb', but the common
 377          * case is just a single output action, so that doing a clone and
 378          * then freeing the original skbuff is wasteful.  So the following code
 379          * is slightly obscure just to avoid that. */
 380         int prev_port = -1;
 381         u32 priority = skb->priority;
 382         const struct nlattr *a;
 383         int rem, err;
 384
 385         for (a = actions, rem = actions_len; rem > 0; a = nla_next(a, &rem)) {
 386                 if (prev_port != -1) {
 387                         do_output(dp, skb_clone(skb, GFP_ATOMIC), prev_port);
 388                         prev_port = -1;
 389                 }
 390
 391                 switch (nla_type(a)) {
 392                 case ODPAT_OUTPUT:
 393                         prev_port = nla_get_u32(a);
 394                         break;
 395
 396                 case ODPAT_CONTROLLER:
 397                         err = output_control(dp, skb, nla_get_u64(a));
 398                         if (err) {
 399                                 kfree_skb(skb);
 400                                 return err;
 401                         }
 402                         break;
 403
 404                 case ODPAT_SET_TUNNEL:
 405                         OVS_CB(skb)->tun_id = nla_get_be64(a);
 406                         break;
 407
 408                 case ODPAT_SET_DL_TCI:
 409                         skb = modify_vlan_tci(dp, skb, key, a, rem);
 410                         if (IS_ERR(skb))
 411                                 return PTR_ERR(skb);
 412                         break;
 413
 414                 case ODPAT_STRIP_VLAN:
 415                         skb = strip_vlan(skb);
 416                         break;
 417
 418                 case ODPAT_SET_DL_SRC:
 419                         skb = make_writable(skb, 0);
 420                         if (!skb)
 421                                 return -ENOMEM;
 422                         memcpy(eth_hdr(skb)->h_source, nla_data(a), ETH_ALEN);
 423                         break;
 424
 425                 case ODPAT_SET_DL_DST:
 426                         skb = make_writable(skb, 0);
 427                         if (!skb)
 428                                 return -ENOMEM;
 429                         memcpy(eth_hdr(skb)->h_dest, nla_data(a), ETH_ALEN);
 430                         break;
 431
 432                 case ODPAT_SET_NW_SRC:
 433                 case ODPAT_SET_NW_DST:
 434                         skb = set_nw_addr(skb, key, a);
 435                         break;
 436
 437                 case ODPAT_SET_NW_TOS:
 438                         skb = set_nw_tos(skb, key, nla_get_u8(a));
 439                         break;
 440
 441                 case ODPAT_SET_TP_SRC:
 442                 case ODPAT_SET_TP_DST:
 443                         skb = set_tp_port(skb, key, a);
 444                         break;
 445
 446                 case ODPAT_SET_PRIORITY:
 447                         skb->priority = nla_get_u32(a);
 448                         break;
 449
 450                 case ODPAT_POP_PRIORITY:
 451                         skb->priority = priority;
 452                         break;
 453
 454                 case ODPAT_DROP_SPOOFED_ARP:
 455                         if (unlikely(is_spoofed_arp(skb, key)))
 456                                 goto exit;
 457                         break;
 458                 }
 459                 if (!skb)
 460                         return -ENOMEM;
 461         }
 462 exit:
 463         if (prev_port != -1)
 464                 do_output(dp, skb, prev_port);
 465         else
 466                 kfree_skb(skb);
 467         return 0;
 468 }
 469
 470 /* Send a copy of this packet up to the sFlow agent, along with extra
 471  * information about what happened to it. */
 472 static void sflow_sample(struct datapath *dp, struct sk_buff *skb,
 473                          const struct nlattr *a, u32 actions_len,
 474                          struct vport *vport)
 475 {
 476         struct odp_sflow_sample_header *hdr;
 477         unsigned int hdrlen = sizeof(struct odp_sflow_sample_header);
 478         struct sk_buff *nskb;
 479
 480         nskb = skb_copy_expand(skb, actions_len + hdrlen, 0, GFP_ATOMIC);
 481         if (!nskb)
 482                 return;
 483
 484         memcpy(__skb_push(nskb, actions_len), a, actions_len);
 485         hdr = (struct odp_sflow_sample_header*)__skb_push(nskb, hdrlen);
 486         hdr->actions_len = actions_len;
 487         hdr->sample_pool = atomic_read(&vport->sflow_pool);
 488         dp_output_control(dp, nskb, _ODPL_SFLOW_NR, 0);
 489 }
 490
 491 /* Execute a list of actions against 'skb'. */
 492 int execute_actions(struct datapath *dp, struct sk_buff *skb,
 493                     const struct sw_flow_key *key,
 494                     const struct nlattr *actions, u32 actions_len)
 495 {
 496         if (dp->sflow_probability) {
 497                 struct vport *p = OVS_CB(skb)->vport;
 498                 if (p) {
 499                         atomic_inc(&p->sflow_pool);
 500                         if (dp->sflow_probability == UINT_MAX ||
 501                             net_random() < dp->sflow_probability)
 502                                 sflow_sample(dp, skb, actions, actions_len, p);
 503                 }
 504         }
 505
 506         OVS_CB(skb)->tun_id = 0;
 507
 508         return do_execute_actions(dp, skb, key, actions, actions_len);
 509 }