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) || skb->len < VLAN_ETH_HLEN)
  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         __be16 mask = a->dl_tci.mask;
  79         __be16 tci = a->dl_tci.tci;
  80
  81         skb = make_writable(skb, VLAN_HLEN, gfp);
  82         if (!skb)
  83                 return ERR_PTR(-ENOMEM);
  84
  85         if (skb->protocol == htons(ETH_P_8021Q)) {
  86                 /* Modify vlan id, but maintain other TCI values */
  87                 struct vlan_ethhdr *vh;
  88                 __be16 old_tci;
  89
  90                 if (skb->len < VLAN_ETH_HLEN)
  91                         return skb;
  92
  93                 vh = vlan_eth_hdr(skb);
  94                 old_tci = vh->h_vlan_TCI;
  95
  96                 vh->h_vlan_TCI = (vh->h_vlan_TCI & ~mask) | tci;
  97
  98                 if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE) {
  99                         __be16 diff[] = { ~old_tci, vh->h_vlan_TCI };
 100
 101                         skb->csum = ~csum_partial((char *)diff, sizeof(diff),
 102                                                 ~skb->csum);
 103                 }
 104         } else {
 105                 int err;
 106
 107                 /* Add vlan header */
 108
 109                 /* Set up checksumming pointers for checksum-deferred packets
 110                  * on Xen.  Otherwise, dev_queue_xmit() will try to do this
 111                  * when we send the packet out on the wire, and it will fail at
 112                  * that point because skb_checksum_setup() will not look inside
 113                  * an 802.1Q header. */
 114                 err = vswitch_skb_checksum_setup(skb);
 115                 if (unlikely(err)) {
 116                         kfree_skb(skb);
 117                         return ERR_PTR(err);
 118                 }
 119
 120                 /* GSO is not implemented for packets with an 802.1Q header, so
 121                  * we have to do segmentation before we add that header.
 122                  *
 123                  * GSO does work with hardware-accelerated VLAN tagging, but we
 124                  * can't use hardware-accelerated VLAN tagging since it
 125                  * requires the device to have a VLAN group configured (with
 126                  * e.g. vconfig(8)) and we don't do that.
 127                  *
 128                  * Having to do this here may be a performance loss, since we
 129                  * can't take advantage of TSO hardware support, although it
 130                  * does not make a measurable network performance difference
 131                  * for 1G Ethernet.  Fixing that would require patching the
 132                  * kernel (either to add GSO support to the VLAN protocol or to
 133                  * support hardware-accelerated VLAN tagging without VLAN
 134                  * groups configured). */
 135                 if (skb_is_gso(skb)) {
 136                         struct sk_buff *segs;
 137
 138                         segs = skb_gso_segment(skb, 0);
 139                         kfree_skb(skb);
 140                         if (unlikely(IS_ERR(segs)))
 141                                 return ERR_CAST(segs);
 142
 143                         do {
 144                                 struct sk_buff *nskb = segs->next;
 145                                 int err;
 146
 147                                 segs->next = NULL;
 148
 149                                 /* GSO can change the checksum type so update.*/
 150                                 compute_ip_summed(segs, true);
 151
 152                                 segs = __vlan_put_tag(segs, ntohs(tci));
 153                                 err = -ENOMEM;
 154                                 if (segs) {
 155                                         err = execute_actions(dp, segs,
 156                                                               key, a + 1,
 157                                                               n_actions - 1,
 158                                                               gfp);
 159                                 }
 160
 161                                 if (unlikely(err)) {
 162                                         while ((segs = nskb)) {
 163                                                 nskb = segs->next;
 164                                                 segs->next = NULL;
 165                                                 kfree_skb(segs);
 166                                         }
 167                                         return ERR_PTR(err);
 168                                 }
 169
 170                                 segs = nskb;
 171                         } while (segs->next);
 172
 173                         skb = segs;
 174                         compute_ip_summed(skb, true);
 175                 }
 176
 177                 /* The hardware-accelerated version of vlan_put_tag() works
 178                  * only for a device that has a VLAN group configured (with
 179                  * e.g. vconfig(8)), so call the software-only version
 180                  * __vlan_put_tag() directly instead.
 181                  */
 182                 skb = __vlan_put_tag(skb, ntohs(tci));
 183                 if (!skb)
 184                         return ERR_PTR(-ENOMEM);
 185
 186                 /* GSO doesn't fix up the hardware computed checksum so this
 187                  * will only be hit in the non-GSO case. */
 188                 if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE)
 189                         skb->csum = csum_add(skb->csum, csum_partial(skb->data
 190                                                 + ETH_HLEN, VLAN_HLEN, 0));
 191         }
 192
 193         return skb;
 194 }
 195
 196 static struct sk_buff *strip_vlan(struct sk_buff *skb, gfp_t gfp)
 197 {
 198         skb = make_writable(skb, 0, gfp);
 199         if (skb)
 200                 vlan_pull_tag(skb);
 201         return skb;
 202 }
 203
 204 static struct sk_buff *set_dl_addr(struct sk_buff *skb,
 205                                    const struct odp_action_dl_addr *a,
 206                                    gfp_t gfp)
 207 {
 208         skb = make_writable(skb, 0, gfp);
 209         if (skb) {
 210                 struct ethhdr *eh = eth_hdr(skb);
 211                 if (a->type == ODPAT_SET_DL_SRC)
 212                         memcpy(eh->h_source, a->dl_addr, ETH_ALEN);
 213                 else
 214                         memcpy(eh->h_dest, a->dl_addr, ETH_ALEN);
 215         }
 216         return skb;
 217 }
 218
 219 /* Updates 'sum', which is a field in 'skb''s data, given that a 4-byte field
 220  * covered by the sum has been changed from 'from' to 'to'.  If set,
 221  * 'pseudohdr' indicates that the field is in the TCP or UDP pseudo-header.
 222  * Based on nf_proto_csum_replace4. */
 223 static void update_csum(__sum16 *sum, struct sk_buff *skb,
 224                         __be32 from, __be32 to, int pseudohdr)
 225 {
 226         __be32 diff[] = { ~from, to };
 227
 228         if (OVS_CB(skb)->ip_summed != OVS_CSUM_PARTIAL) {
 229                 *sum = csum_fold(csum_partial((char *)diff, sizeof(diff),
 230                                 ~csum_unfold(*sum)));
 231                 if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE && pseudohdr)
 232                         skb->csum = ~csum_partial((char *)diff, sizeof(diff),
 233                                                 ~skb->csum);
 234         } else if (pseudohdr)
 235                 *sum = ~csum_fold(csum_partial((char *)diff, sizeof(diff),
 236                                 csum_unfold(*sum)));
 237 }
 238
 239 static bool is_ip(struct sk_buff *skb, const struct odp_flow_key *key)
 240 {
 241         return (key->dl_type == htons(ETH_P_IP) &&
 242                 skb->transport_header > skb->network_header);
 243 }
 244
 245 static __sum16 *get_l4_checksum(struct sk_buff *skb, const struct odp_flow_key *key)
 246 {
 247         int transport_len = skb->len - skb_transport_offset(skb);
 248         if (key->nw_proto == IPPROTO_TCP) {
 249                 if (likely(transport_len >= sizeof(struct tcphdr)))
 250                         return &tcp_hdr(skb)->check;
 251         } else if (key->nw_proto == IPPROTO_UDP) {
 252                 if (likely(transport_len >= sizeof(struct udphdr)))
 253                         return &udp_hdr(skb)->check;
 254         }
 255         return NULL;
 256 }
 257
 258 static struct sk_buff *set_nw_addr(struct sk_buff *skb,
 259                                    const struct odp_flow_key *key,
 260                                    const struct odp_action_nw_addr *a,
 261                                    gfp_t gfp)
 262 {
 263         struct iphdr *nh;
 264         __sum16 *check;
 265         __be32 *nwaddr;
 266
 267         if (unlikely(!is_ip(skb, key)))
 268                 return skb;
 269
 270         skb = make_writable(skb, 0, gfp);
 271         if (unlikely(!skb))
 272                 return NULL;
 273
 274         nh = ip_hdr(skb);
 275         nwaddr = a->type == ODPAT_SET_NW_SRC ? &nh->saddr : &nh->daddr;
 276
 277         check = get_l4_checksum(skb, key);
 278         if (likely(check))
 279                 update_csum(check, skb, *nwaddr, a->nw_addr, 1);
 280         update_csum(&nh->check, skb, *nwaddr, a->nw_addr, 0);
 281
 282         *nwaddr = a->nw_addr;
 283
 284         return skb;
 285 }
 286
 287 static struct sk_buff *set_nw_tos(struct sk_buff *skb,
 288                                    const struct odp_flow_key *key,
 289                                    const struct odp_action_nw_tos *a,
 290                                    gfp_t gfp)
 291 {
 292         if (unlikely(!is_ip(skb, key)))
 293                 return skb;
 294
 295         skb = make_writable(skb, 0, gfp);
 296         if (skb) {
 297                 struct iphdr *nh = ip_hdr(skb);
 298                 u8 *f = &nh->tos;
 299                 u8 old = *f;
 300                 u8 new;
 301
 302                 /* Set the DSCP bits and preserve the ECN bits. */
 303                 new = a->nw_tos | (nh->tos & INET_ECN_MASK);
 304                 update_csum(&nh->check, skb, htons((u16)old),
 305                             htons((u16)new), 0);
 306                 *f = new;
 307         }
 308         return skb;
 309 }
 310
 311 static struct sk_buff *set_tp_port(struct sk_buff *skb,
 312                                    const struct odp_flow_key *key,
 313                                    const struct odp_action_tp_port *a, gfp_t gfp)
 314 {
 315         struct udphdr *th;
 316         __sum16 *check;
 317         __be16 *port;
 318
 319         if (unlikely(!is_ip(skb, key)))
 320                 return skb;
 321
 322         skb = make_writable(skb, 0, gfp);
 323         if (unlikely(!skb))
 324                 return NULL;
 325
 326         /* Must follow make_writable() since that can move the skb data. */
 327         check = get_l4_checksum(skb, key);
 328         if (unlikely(!check))
 329                 return skb;
 330
 331         /*
 332          * Update port and checksum.
 333          *
 334          * This is OK because source and destination port numbers are at the
 335          * same offsets in both UDP and TCP headers, and get_l4_checksum() only
 336          * supports those protocols.
 337          */
 338         th = udp_hdr(skb);
 339         port = a->type == ODPAT_SET_TP_SRC ? &th->source : &th->dest;
 340         update_csum(check, skb, *port, a->tp_port, 0);
 341         *port = a->tp_port;
 342
 343         return skb;
 344 }
 345
 346 /**
 347  * is_spoofed_arp - check for invalid ARP packet
 348  *
 349  * @skb: skbuff containing an Ethernet packet, with network header pointing
 350  * just past the Ethernet and optional 802.1Q header.
 351  * @key: flow key extracted from @skb by flow_extract()
 352  *
 353  * Returns true if @skb is an invalid Ethernet+IPv4 ARP packet: one with screwy
 354  * or truncated header fields or one whose inner and outer Ethernet address
 355  * differ.
 356  */
 357 static bool is_spoofed_arp(struct sk_buff *skb, const struct odp_flow_key *key)
 358 {
 359         struct arp_eth_header *arp;
 360
 361         if (key->dl_type != htons(ETH_P_ARP))
 362                 return false;
 363
 364         if (skb_network_offset(skb) + sizeof(struct arp_eth_header) > skb->len)
 365                 return true;
 366
 367         arp = (struct arp_eth_header *)skb_network_header(skb);
 368         return (arp->ar_hrd != htons(ARPHRD_ETHER) ||
 369                 arp->ar_pro != htons(ETH_P_IP) ||
 370                 arp->ar_hln != ETH_ALEN ||
 371                 arp->ar_pln != 4 ||
 372                 compare_ether_addr(arp->ar_sha, eth_hdr(skb)->h_source));
 373 }
 374
 375 static void do_output(struct datapath *dp, struct sk_buff *skb, int out_port)
 376 {
 377         struct dp_port *p;
 378
 379         if (!skb)
 380                 goto error;
 381
 382         p = rcu_dereference(dp->ports[out_port]);
 383         if (!p)
 384                 goto error;
 385
 386         vport_send(p->vport, skb);
 387         return;
 388
 389 error:
 390         kfree_skb(skb);
 391 }
 392
 393 static int output_control(struct datapath *dp, struct sk_buff *skb, u32 arg,
 394                           gfp_t gfp)
 395 {
 396         skb = skb_clone(skb, gfp);
 397         if (!skb)
 398                 return -ENOMEM;
 399         return dp_output_control(dp, skb, _ODPL_ACTION_NR, arg);
 400 }
 401
 402 /* Send a copy of this packet up to the sFlow agent, along with extra
 403  * information about what happened to it. */
 404 static void sflow_sample(struct datapath *dp, struct sk_buff *skb,
 405                          const union odp_action *a, int n_actions,
 406                          gfp_t gfp, struct dp_port *dp_port)
 407 {
 408         struct odp_sflow_sample_header *hdr;
 409         unsigned int actlen = n_actions * sizeof(union odp_action);
 410         unsigned int hdrlen = sizeof(struct odp_sflow_sample_header);
 411         struct sk_buff *nskb;
 412
 413         nskb = skb_copy_expand(skb, actlen + hdrlen, 0, gfp);
 414         if (!nskb)
 415                 return;
 416
 417         memcpy(__skb_push(nskb, actlen), a, actlen);
 418         hdr = (struct odp_sflow_sample_header*)__skb_push(nskb, hdrlen);
 419         hdr->n_actions = n_actions;
 420         hdr->sample_pool = atomic_read(&dp_port->sflow_pool);
 421         dp_output_control(dp, nskb, _ODPL_SFLOW_NR, 0);
 422 }
 423
 424 /* Execute a list of actions against 'skb'. */
 425 int execute_actions(struct datapath *dp, struct sk_buff *skb,
 426                     const struct odp_flow_key *key,
 427                     const union odp_action *a, int n_actions,
 428                     gfp_t gfp)
 429 {
 430         /* Every output action needs a separate clone of 'skb', but the common
 431          * case is just a single output action, so that doing a clone and
 432          * then freeing the original skbuff is wasteful.  So the following code
 433          * is slightly obscure just to avoid that. */
 434         int prev_port = -1;
 435         u32 priority = skb->priority;
 436         int err;
 437
 438         if (dp->sflow_probability) {
 439                 struct dp_port *p = OVS_CB(skb)->dp_port;
 440                 if (p) {
 441                         atomic_inc(&p->sflow_pool);
 442                         if (dp->sflow_probability == UINT_MAX ||
 443                             net_random() < dp->sflow_probability)
 444                                 sflow_sample(dp, skb, a, n_actions, gfp, p);
 445                 }
 446         }
 447
 448         OVS_CB(skb)->tun_id = 0;
 449
 450         for (; n_actions > 0; a++, n_actions--) {
 451                 if (prev_port != -1) {
 452                         do_output(dp, skb_clone(skb, gfp), prev_port);
 453                         prev_port = -1;
 454                 }
 455
 456                 switch (a->type) {
 457                 case ODPAT_OUTPUT:
 458                         prev_port = a->output.port;
 459                         break;
 460
 461                 case ODPAT_CONTROLLER:
 462                         err = output_control(dp, skb, a->controller.arg, gfp);
 463                         if (err) {
 464                                 kfree_skb(skb);
 465                                 return err;
 466                         }
 467                         break;
 468
 469                 case ODPAT_SET_TUNNEL:
 470                         OVS_CB(skb)->tun_id = a->tunnel.tun_id;
 471                         break;
 472
 473                 case ODPAT_SET_DL_TCI:
 474                         skb = modify_vlan_tci(dp, skb, key, a, n_actions, gfp);
 475                         if (IS_ERR(skb))
 476                                 return PTR_ERR(skb);
 477                         break;
 478
 479                 case ODPAT_STRIP_VLAN:
 480                         skb = strip_vlan(skb, gfp);
 481                         break;
 482
 483                 case ODPAT_SET_DL_SRC:
 484                 case ODPAT_SET_DL_DST:
 485                         skb = set_dl_addr(skb, &a->dl_addr, gfp);
 486                         break;
 487
 488                 case ODPAT_SET_NW_SRC:
 489                 case ODPAT_SET_NW_DST:
 490                         skb = set_nw_addr(skb, key, &a->nw_addr, gfp);
 491                         break;
 492
 493                 case ODPAT_SET_NW_TOS:
 494                         skb = set_nw_tos(skb, key, &a->nw_tos, gfp);
 495                         break;
 496
 497                 case ODPAT_SET_TP_SRC:
 498                 case ODPAT_SET_TP_DST:
 499                         skb = set_tp_port(skb, key, &a->tp_port, gfp);
 500                         break;
 501
 502                 case ODPAT_SET_PRIORITY:
 503                         skb->priority = a->priority.priority;
 504                         break;
 505
 506                 case ODPAT_POP_PRIORITY:
 507                         skb->priority = priority;
 508                         break;
 509
 510                 case ODPAT_DROP_SPOOFED_ARP:
 511                         if (unlikely(is_spoofed_arp(skb, key)))
 512                                 goto exit;
 513                         break;
 514                 }
 515                 if (!skb)
 516                         return -ENOMEM;
 517         }
 518 exit:
 519         if (prev_port != -1)
 520                 do_output(dp, skb, prev_port);
 521         else
 522                 kfree_skb(skb);
 523         return 0;
 524 }