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 struct sk_buff *make_writable(struct sk_buff *skb, unsigned min_headroom)
  30 {
  31         if (skb_cloned(skb)) {
  32                 struct sk_buff *nskb;
  33                 unsigned headroom = max(min_headroom, skb_headroom(skb));
  34
  35                 nskb = skb_copy_expand(skb, headroom, skb_tailroom(skb), GFP_ATOMIC);
  36                 if (nskb) {
  37                         set_skb_csum_bits(skb, nskb);
  38                         kfree_skb(skb);
  39                         return nskb;
  40                 }
  41         } else {
  42                 unsigned int hdr_len = (skb_transport_offset(skb)
  43                                         + sizeof(struct tcphdr));
  44                 if (pskb_may_pull(skb, min(hdr_len, skb->len)))
  45                         return skb;
  46         }
  47         kfree_skb(skb);
  48         return NULL;
  49 }
  50
  51 static struct sk_buff *vlan_pull_tag(struct sk_buff *skb)
  52 {
  53         struct vlan_ethhdr *vh = vlan_eth_hdr(skb);
  54         struct ethhdr *eh;
  55
  56         /* Verify we were given a vlan packet */
  57         if (vh->h_vlan_proto != htons(ETH_P_8021Q) || skb->len < VLAN_ETH_HLEN)
  58                 return skb;
  59
  60         if (get_ip_summed(skb) == OVS_CSUM_COMPLETE)
  61                 skb->csum = csum_sub(skb->csum, csum_partial(skb->data
  62                                         + ETH_HLEN, VLAN_HLEN, 0));
  63
  64         memmove(skb->data + VLAN_HLEN, skb->data, 2 * VLAN_ETH_ALEN);
  65
  66         eh = (struct ethhdr *)skb_pull(skb, VLAN_HLEN);
  67
  68         skb->protocol = eh->h_proto;
  69         skb->mac_header += VLAN_HLEN;
  70
  71         return skb;
  72 }
  73
  74 static struct sk_buff *modify_vlan_tci(struct datapath *dp, struct sk_buff *skb,
  75                                        const struct odp_flow_key *key,
  76                                        const union odp_action *a, int n_actions)
  77 {
  78         __be16 tci = a->dl_tci.tci;
  79
  80         skb = make_writable(skb, VLAN_HLEN);
  81         if (!skb)
  82                 return ERR_PTR(-ENOMEM);
  83
  84         if (skb->protocol == htons(ETH_P_8021Q)) {
  85                 /* Modify vlan id, but maintain other TCI values */
  86                 struct vlan_ethhdr *vh;
  87                 __be16 old_tci;
  88
  89                 if (skb->len < VLAN_ETH_HLEN)
  90                         return skb;
  91
  92                 vh = vlan_eth_hdr(skb);
  93                 old_tci = vh->h_vlan_TCI;
  94
  95                 vh->h_vlan_TCI = tci;
  96
  97                 if (get_ip_summed(skb) == OVS_CSUM_COMPLETE) {
  98                         __be16 diff[] = { ~old_tci, vh->h_vlan_TCI };
  99
 100                         skb->csum = ~csum_partial((char *)diff, sizeof(diff),
 101                                                 ~skb->csum);
 102                 }
 103         } else {
 104                 int err;
 105
 106                 /* Add vlan header */
 107
 108                 /* Set up checksumming pointers for checksum-deferred packets
 109                  * on Xen.  Otherwise, dev_queue_xmit() will try to do this
 110                  * when we send the packet out on the wire, and it will fail at
 111                  * that point because skb_checksum_setup() will not look inside
 112                  * an 802.1Q header. */
 113                 err = vswitch_skb_checksum_setup(skb);
 114                 if (unlikely(err)) {
 115                         kfree_skb(skb);
 116                         return ERR_PTR(err);
 117                 }
 118
 119                 /* GSO is not implemented for packets with an 802.1Q header, so
 120                  * we have to do segmentation before we add that header.
 121                  *
 122                  * GSO does work with hardware-accelerated VLAN tagging, but we
 123                  * can't use hardware-accelerated VLAN tagging since it
 124                  * requires the device to have a VLAN group configured (with
 125                  * e.g. vconfig(8)) and we don't do that.
 126                  *
 127                  * Having to do this here may be a performance loss, since we
 128                  * can't take advantage of TSO hardware support, although it
 129                  * does not make a measurable network performance difference
 130                  * for 1G Ethernet.  Fixing that would require patching the
 131                  * kernel (either to add GSO support to the VLAN protocol or to
 132                  * support hardware-accelerated VLAN tagging without VLAN
 133                  * groups configured). */
 134                 if (skb_is_gso(skb)) {
 135                         struct sk_buff *segs;
 136
 137                         segs = skb_gso_segment(skb, 0);
 138                         kfree_skb(skb);
 139                         if (unlikely(IS_ERR(segs)))
 140                                 return ERR_CAST(segs);
 141
 142                         do {
 143                                 struct sk_buff *nskb = segs->next;
 144                                 int err;
 145
 146                                 segs->next = NULL;
 147
 148                                 /* GSO can change the checksum type so update.*/
 149                                 compute_ip_summed(segs, true);
 150
 151                                 segs = __vlan_put_tag(segs, ntohs(tci));
 152                                 err = -ENOMEM;
 153                                 if (segs) {
 154                                         err = execute_actions(dp, segs,
 155                                                               key, a + 1,
 156                                                               n_actions - 1);
 157                                 }
 158
 159                                 if (unlikely(err)) {
 160                                         while ((segs = nskb)) {
 161                                                 nskb = segs->next;
 162                                                 segs->next = NULL;
 163                                                 kfree_skb(segs);
 164                                         }
 165                                         return ERR_PTR(err);
 166                                 }
 167
 168                                 segs = nskb;
 169                         } while (segs->next);
 170
 171                         skb = segs;
 172                         compute_ip_summed(skb, true);
 173                 }
 174
 175                 /* The hardware-accelerated version of vlan_put_tag() works
 176                  * only for a device that has a VLAN group configured (with
 177                  * e.g. vconfig(8)), so call the software-only version
 178                  * __vlan_put_tag() directly instead.
 179                  */
 180                 skb = __vlan_put_tag(skb, ntohs(tci));
 181                 if (!skb)
 182                         return ERR_PTR(-ENOMEM);
 183
 184                 /* GSO doesn't fix up the hardware computed checksum so this
 185                  * will only be hit in the non-GSO case. */
 186                 if (get_ip_summed(skb) == OVS_CSUM_COMPLETE)
 187                         skb->csum = csum_add(skb->csum, csum_partial(skb->data
 188                                                 + ETH_HLEN, VLAN_HLEN, 0));
 189         }
 190
 191         return skb;
 192 }
 193
 194 static struct sk_buff *strip_vlan(struct sk_buff *skb)
 195 {
 196         skb = make_writable(skb, 0);
 197         if (skb)
 198                 vlan_pull_tag(skb);
 199         return skb;
 200 }
 201
 202 static struct sk_buff *set_dl_addr(struct sk_buff *skb,
 203                                    const struct odp_action_dl_addr *a)
 204 {
 205         skb = make_writable(skb, 0);
 206         if (skb) {
 207                 struct ethhdr *eh = eth_hdr(skb);
 208                 if (a->type == ODPAT_SET_DL_SRC)
 209                         memcpy(eh->h_source, a->dl_addr, ETH_ALEN);
 210                 else
 211                         memcpy(eh->h_dest, a->dl_addr, ETH_ALEN);
 212         }
 213         return skb;
 214 }
 215
 216 /* Updates 'sum', which is a field in 'skb''s data, given that a 4-byte field
 217  * covered by the sum has been changed from 'from' to 'to'.  If set,
 218  * 'pseudohdr' indicates that the field is in the TCP or UDP pseudo-header.
 219  * Based on nf_proto_csum_replace4. */
 220 static void update_csum(__sum16 *sum, struct sk_buff *skb,
 221                         __be32 from, __be32 to, int pseudohdr)
 222 {
 223         __be32 diff[] = { ~from, to };
 224
 225         if (get_ip_summed(skb) != OVS_CSUM_PARTIAL) {
 226                 *sum = csum_fold(csum_partial((char *)diff, sizeof(diff),
 227                                 ~csum_unfold(*sum)));
 228                 if (get_ip_summed(skb) == OVS_CSUM_COMPLETE && pseudohdr)
 229                         skb->csum = ~csum_partial((char *)diff, sizeof(diff),
 230                                                 ~skb->csum);
 231         } else if (pseudohdr)
 232                 *sum = ~csum_fold(csum_partial((char *)diff, sizeof(diff),
 233                                 csum_unfold(*sum)));
 234 }
 235
 236 static bool is_ip(struct sk_buff *skb, const struct odp_flow_key *key)
 237 {
 238         return (key->dl_type == htons(ETH_P_IP) &&
 239                 skb->transport_header > skb->network_header);
 240 }
 241
 242 static __sum16 *get_l4_checksum(struct sk_buff *skb, const struct odp_flow_key *key)
 243 {
 244         int transport_len = skb->len - skb_transport_offset(skb);
 245         if (key->nw_proto == IPPROTO_TCP) {
 246                 if (likely(transport_len >= sizeof(struct tcphdr)))
 247                         return &tcp_hdr(skb)->check;
 248         } else if (key->nw_proto == IPPROTO_UDP) {
 249                 if (likely(transport_len >= sizeof(struct udphdr)))
 250                         return &udp_hdr(skb)->check;
 251         }
 252         return NULL;
 253 }
 254
 255 static struct sk_buff *set_nw_addr(struct sk_buff *skb,
 256                                    const struct odp_flow_key *key,
 257                                    const struct odp_action_nw_addr *a)
 258 {
 259         struct iphdr *nh;
 260         __sum16 *check;
 261         __be32 *nwaddr;
 262
 263         if (unlikely(!is_ip(skb, key)))
 264                 return skb;
 265
 266         skb = make_writable(skb, 0);
 267         if (unlikely(!skb))
 268                 return NULL;
 269
 270         nh = ip_hdr(skb);
 271         nwaddr = a->type == ODPAT_SET_NW_SRC ? &nh->saddr : &nh->daddr;
 272
 273         check = get_l4_checksum(skb, key);
 274         if (likely(check))
 275                 update_csum(check, skb, *nwaddr, a->nw_addr, 1);
 276         update_csum(&nh->check, skb, *nwaddr, a->nw_addr, 0);
 277
 278         *nwaddr = a->nw_addr;
 279
 280         return skb;
 281 }
 282
 283 static struct sk_buff *set_nw_tos(struct sk_buff *skb,
 284                                    const struct odp_flow_key *key,
 285                                    const struct odp_action_nw_tos *a)
 286 {
 287         if (unlikely(!is_ip(skb, key)))
 288                 return skb;
 289
 290         skb = make_writable(skb, 0);
 291         if (skb) {
 292                 struct iphdr *nh = ip_hdr(skb);
 293                 u8 *f = &nh->tos;
 294                 u8 old = *f;
 295                 u8 new;
 296
 297                 /* Set the DSCP bits and preserve the ECN bits. */
 298                 new = a->nw_tos | (nh->tos & INET_ECN_MASK);
 299                 update_csum(&nh->check, skb, htons((u16)old),
 300                             htons((u16)new), 0);
 301                 *f = new;
 302         }
 303         return skb;
 304 }
 305
 306 static struct sk_buff *set_tp_port(struct sk_buff *skb,
 307                                    const struct odp_flow_key *key,
 308                                    const struct odp_action_tp_port *a)
 309 {
 310         struct udphdr *th;
 311         __sum16 *check;
 312         __be16 *port;
 313
 314         if (unlikely(!is_ip(skb, key)))
 315                 return skb;
 316
 317         skb = make_writable(skb, 0);
 318         if (unlikely(!skb))
 319                 return NULL;
 320
 321         /* Must follow make_writable() since that can move the skb data. */
 322         check = get_l4_checksum(skb, key);
 323         if (unlikely(!check))
 324                 return skb;
 325
 326         /*
 327          * Update port and checksum.
 328          *
 329          * This is OK because source and destination port numbers are at the
 330          * same offsets in both UDP and TCP headers, and get_l4_checksum() only
 331          * supports those protocols.
 332          */
 333         th = udp_hdr(skb);
 334         port = a->type == ODPAT_SET_TP_SRC ? &th->source : &th->dest;
 335         update_csum(check, skb, *port, a->tp_port, 0);
 336         *port = a->tp_port;
 337
 338         return skb;
 339 }
 340
 341 /**
 342  * is_spoofed_arp - check for invalid ARP packet
 343  *
 344  * @skb: skbuff containing an Ethernet packet, with network header pointing
 345  * just past the Ethernet and optional 802.1Q header.
 346  * @key: flow key extracted from @skb by flow_extract()
 347  *
 348  * Returns true if @skb is an invalid Ethernet+IPv4 ARP packet: one with screwy
 349  * or truncated header fields or one whose inner and outer Ethernet address
 350  * differ.
 351  */
 352 static bool is_spoofed_arp(struct sk_buff *skb, const struct odp_flow_key *key)
 353 {
 354         struct arp_eth_header *arp;
 355
 356         if (key->dl_type != htons(ETH_P_ARP))
 357                 return false;
 358
 359         if (skb_network_offset(skb) + sizeof(struct arp_eth_header) > skb->len)
 360                 return true;
 361
 362         arp = (struct arp_eth_header *)skb_network_header(skb);
 363         return (arp->ar_hrd != htons(ARPHRD_ETHER) ||
 364                 arp->ar_pro != htons(ETH_P_IP) ||
 365                 arp->ar_hln != ETH_ALEN ||
 366                 arp->ar_pln != 4 ||
 367                 compare_ether_addr(arp->ar_sha, eth_hdr(skb)->h_source));
 368 }
 369
 370 static void do_output(struct datapath *dp, struct sk_buff *skb, int out_port)
 371 {
 372         struct vport *p;
 373
 374         if (!skb)
 375                 goto error;
 376
 377         p = rcu_dereference(dp->ports[out_port]);
 378         if (!p)
 379                 goto error;
 380
 381         vport_send(p, skb);
 382         return;
 383
 384 error:
 385         kfree_skb(skb);
 386 }
 387
 388 static int output_control(struct datapath *dp, struct sk_buff *skb, u32 arg)
 389 {
 390         skb = skb_clone(skb, GFP_ATOMIC);
 391         if (!skb)
 392                 return -ENOMEM;
 393         return dp_output_control(dp, skb, _ODPL_ACTION_NR, arg);
 394 }
 395
 396 /* Send a copy of this packet up to the sFlow agent, along with extra
 397  * information about what happened to it. */
 398 static void sflow_sample(struct datapath *dp, struct sk_buff *skb,
 399                          const union odp_action *a, int n_actions,
 400                          struct vport *vport)
 401 {
 402         struct odp_sflow_sample_header *hdr;
 403         unsigned int actlen = n_actions * sizeof(union odp_action);
 404         unsigned int hdrlen = sizeof(struct odp_sflow_sample_header);
 405         struct sk_buff *nskb;
 406
 407         nskb = skb_copy_expand(skb, actlen + hdrlen, 0, GFP_ATOMIC);
 408         if (!nskb)
 409                 return;
 410
 411         memcpy(__skb_push(nskb, actlen), a, actlen);
 412         hdr = (struct odp_sflow_sample_header*)__skb_push(nskb, hdrlen);
 413         hdr->n_actions = n_actions;
 414         hdr->sample_pool = atomic_read(&vport->sflow_pool);
 415         dp_output_control(dp, nskb, _ODPL_SFLOW_NR, 0);
 416 }
 417
 418 /* Execute a list of actions against 'skb'. */
 419 int execute_actions(struct datapath *dp, struct sk_buff *skb,
 420                     const struct odp_flow_key *key,
 421                     const union odp_action *a, int n_actions)
 422 {
 423         /* Every output action needs a separate clone of 'skb', but the common
 424          * case is just a single output action, so that doing a clone and
 425          * then freeing the original skbuff is wasteful.  So the following code
 426          * is slightly obscure just to avoid that. */
 427         int prev_port = -1;
 428         u32 priority = skb->priority;
 429         int err;
 430
 431         if (dp->sflow_probability) {
 432                 struct vport *p = OVS_CB(skb)->vport;
 433                 if (p) {
 434                         atomic_inc(&p->sflow_pool);
 435                         if (dp->sflow_probability == UINT_MAX ||
 436                             net_random() < dp->sflow_probability)
 437                                 sflow_sample(dp, skb, a, n_actions, p);
 438                 }
 439         }
 440
 441         OVS_CB(skb)->tun_id = 0;
 442
 443         for (; n_actions > 0; a++, n_actions--) {
 444                 if (prev_port != -1) {
 445                         do_output(dp, skb_clone(skb, GFP_ATOMIC), prev_port);
 446                         prev_port = -1;
 447                 }
 448
 449                 switch (a->type) {
 450                 case ODPAT_OUTPUT:
 451                         prev_port = a->output.port;
 452                         break;
 453
 454                 case ODPAT_CONTROLLER:
 455                         err = output_control(dp, skb, a->controller.arg);
 456                         if (err) {
 457                                 kfree_skb(skb);
 458                                 return err;
 459                         }
 460                         break;
 461
 462                 case ODPAT_SET_TUNNEL:
 463                         OVS_CB(skb)->tun_id = a->tunnel.tun_id;
 464                         break;
 465
 466                 case ODPAT_SET_DL_TCI:
 467                         skb = modify_vlan_tci(dp, skb, key, a, n_actions);
 468                         if (IS_ERR(skb))
 469                                 return PTR_ERR(skb);
 470                         break;
 471
 472                 case ODPAT_STRIP_VLAN:
 473                         skb = strip_vlan(skb);
 474                         break;
 475
 476                 case ODPAT_SET_DL_SRC:
 477                 case ODPAT_SET_DL_DST:
 478                         skb = set_dl_addr(skb, &a->dl_addr);
 479                         break;
 480
 481                 case ODPAT_SET_NW_SRC:
 482                 case ODPAT_SET_NW_DST:
 483                         skb = set_nw_addr(skb, key, &a->nw_addr);
 484                         break;
 485
 486                 case ODPAT_SET_NW_TOS:
 487                         skb = set_nw_tos(skb, key, &a->nw_tos);
 488                         break;
 489
 490                 case ODPAT_SET_TP_SRC:
 491                 case ODPAT_SET_TP_DST:
 492                         skb = set_tp_port(skb, key, &a->tp_port);
 493                         break;
 494
 495                 case ODPAT_SET_PRIORITY:
 496                         skb->priority = a->priority.priority;
 497                         break;
 498
 499                 case ODPAT_POP_PRIORITY:
 500                         skb->priority = priority;
 501                         break;
 502
 503                 case ODPAT_DROP_SPOOFED_ARP:
 504                         if (unlikely(is_spoofed_arp(skb, key)))
 505                                 goto exit;
 506                         break;
 507                 }
 508                 if (!skb)
 509                         return -ENOMEM;
 510         }
 511 exit:
 512         if (prev_port != -1)
 513                 do_output(dp, skb, prev_port);
 514         else
 515                 kfree_skb(skb);
 516         return 0;
 517 }