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