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 static bool is_ip(struct sk_buff *skb, const struct odp_flow_key *key)
 217 {
 218         return (key->dl_type == htons(ETH_P_IP) &&
 219                 skb->transport_header > skb->network_header);
 220 }
 221
 222 static __sum16 *get_l4_checksum(struct sk_buff *skb, const struct odp_flow_key *key)
 223 {
 224         int transport_len = skb->len - skb_transport_offset(skb);
 225         if (key->nw_proto == IPPROTO_TCP) {
 226                 if (likely(transport_len >= sizeof(struct tcphdr)))
 227                         return &tcp_hdr(skb)->check;
 228         } else if (key->nw_proto == IPPROTO_UDP) {
 229                 if (likely(transport_len >= sizeof(struct udphdr)))
 230                         return &udp_hdr(skb)->check;
 231         }
 232         return NULL;
 233 }
 234
 235 static struct sk_buff *set_nw_addr(struct sk_buff *skb,
 236                                    const struct odp_flow_key *key,
 237                                    const struct odp_action_nw_addr *a)
 238 {
 239         struct iphdr *nh;
 240         __sum16 *check;
 241         __be32 *nwaddr;
 242
 243         if (unlikely(!is_ip(skb, key)))
 244                 return skb;
 245
 246         skb = make_writable(skb, 0);
 247         if (unlikely(!skb))
 248                 return NULL;
 249
 250         nh = ip_hdr(skb);
 251         nwaddr = a->type == ODPAT_SET_NW_SRC ? &nh->saddr : &nh->daddr;
 252
 253         check = get_l4_checksum(skb, key);
 254         if (likely(check))
 255                 inet_proto_csum_replace4(check, skb, *nwaddr, a->nw_addr, 1);
 256         csum_replace4(&nh->check, *nwaddr, a->nw_addr);
 257
 258         *nwaddr = a->nw_addr;
 259
 260         return skb;
 261 }
 262
 263 static struct sk_buff *set_nw_tos(struct sk_buff *skb,
 264                                    const struct odp_flow_key *key,
 265                                    const struct odp_action_nw_tos *a)
 266 {
 267         if (unlikely(!is_ip(skb, key)))
 268                 return skb;
 269
 270         skb = make_writable(skb, 0);
 271         if (skb) {
 272                 struct iphdr *nh = ip_hdr(skb);
 273                 u8 *f = &nh->tos;
 274                 u8 old = *f;
 275                 u8 new;
 276
 277                 /* Set the DSCP bits and preserve the ECN bits. */
 278                 new = a->nw_tos | (nh->tos & INET_ECN_MASK);
 279                 csum_replace4(&nh->check, (__force __be32)old,
 280                                           (__force __be32)new);
 281                 *f = new;
 282         }
 283         return skb;
 284 }
 285
 286 static struct sk_buff *set_tp_port(struct sk_buff *skb,
 287                                    const struct odp_flow_key *key,
 288                                    const struct odp_action_tp_port *a)
 289 {
 290         struct udphdr *th;
 291         __sum16 *check;
 292         __be16 *port;
 293
 294         if (unlikely(!is_ip(skb, key)))
 295                 return skb;
 296
 297         skb = make_writable(skb, 0);
 298         if (unlikely(!skb))
 299                 return NULL;
 300
 301         /* Must follow make_writable() since that can move the skb data. */
 302         check = get_l4_checksum(skb, key);
 303         if (unlikely(!check))
 304                 return skb;
 305
 306         /*
 307          * Update port and checksum.
 308          *
 309          * This is OK because source and destination port numbers are at the
 310          * same offsets in both UDP and TCP headers, and get_l4_checksum() only
 311          * supports those protocols.
 312          */
 313         th = udp_hdr(skb);
 314         port = a->type == ODPAT_SET_TP_SRC ? &th->source : &th->dest;
 315         inet_proto_csum_replace2(check, skb, *port, a->tp_port, 0);
 316         *port = a->tp_port;
 317
 318         return skb;
 319 }
 320
 321 /**
 322  * is_spoofed_arp - check for invalid ARP packet
 323  *
 324  * @skb: skbuff containing an Ethernet packet, with network header pointing
 325  * just past the Ethernet and optional 802.1Q header.
 326  * @key: flow key extracted from @skb by flow_extract()
 327  *
 328  * Returns true if @skb is an invalid Ethernet+IPv4 ARP packet: one with screwy
 329  * or truncated header fields or one whose inner and outer Ethernet address
 330  * differ.
 331  */
 332 static bool is_spoofed_arp(struct sk_buff *skb, const struct odp_flow_key *key)
 333 {
 334         struct arp_eth_header *arp;
 335
 336         if (key->dl_type != htons(ETH_P_ARP))
 337                 return false;
 338
 339         if (skb_network_offset(skb) + sizeof(struct arp_eth_header) > skb->len)
 340                 return true;
 341
 342         arp = (struct arp_eth_header *)skb_network_header(skb);
 343         return (arp->ar_hrd != htons(ARPHRD_ETHER) ||
 344                 arp->ar_pro != htons(ETH_P_IP) ||
 345                 arp->ar_hln != ETH_ALEN ||
 346                 arp->ar_pln != 4 ||
 347                 compare_ether_addr(arp->ar_sha, eth_hdr(skb)->h_source));
 348 }
 349
 350 static void do_output(struct datapath *dp, struct sk_buff *skb, int out_port)
 351 {
 352         struct vport *p;
 353
 354         if (!skb)
 355                 goto error;
 356
 357         p = rcu_dereference(dp->ports[out_port]);
 358         if (!p)
 359                 goto error;
 360
 361         vport_send(p, skb);
 362         return;
 363
 364 error:
 365         kfree_skb(skb);
 366 }
 367
 368 static int output_control(struct datapath *dp, struct sk_buff *skb, u32 arg)
 369 {
 370         skb = skb_clone(skb, GFP_ATOMIC);
 371         if (!skb)
 372                 return -ENOMEM;
 373         return dp_output_control(dp, skb, _ODPL_ACTION_NR, arg);
 374 }
 375
 376 /* Send a copy of this packet up to the sFlow agent, along with extra
 377  * information about what happened to it. */
 378 static void sflow_sample(struct datapath *dp, struct sk_buff *skb,
 379                          const union odp_action *a, int n_actions,
 380                          struct vport *vport)
 381 {
 382         struct odp_sflow_sample_header *hdr;
 383         unsigned int actlen = n_actions * sizeof(union odp_action);
 384         unsigned int hdrlen = sizeof(struct odp_sflow_sample_header);
 385         struct sk_buff *nskb;
 386
 387         nskb = skb_copy_expand(skb, actlen + hdrlen, 0, GFP_ATOMIC);
 388         if (!nskb)
 389                 return;
 390
 391         memcpy(__skb_push(nskb, actlen), a, actlen);
 392         hdr = (struct odp_sflow_sample_header*)__skb_push(nskb, hdrlen);
 393         hdr->n_actions = n_actions;
 394         hdr->sample_pool = atomic_read(&vport->sflow_pool);
 395         dp_output_control(dp, nskb, _ODPL_SFLOW_NR, 0);
 396 }
 397
 398 /* Execute a list of actions against 'skb'. */
 399 int execute_actions(struct datapath *dp, struct sk_buff *skb,
 400                     const struct odp_flow_key *key,
 401                     const union odp_action *a, int n_actions)
 402 {
 403         /* Every output action needs a separate clone of 'skb', but the common
 404          * case is just a single output action, so that doing a clone and
 405          * then freeing the original skbuff is wasteful.  So the following code
 406          * is slightly obscure just to avoid that. */
 407         int prev_port = -1;
 408         u32 priority = skb->priority;
 409         int err;
 410
 411         if (dp->sflow_probability) {
 412                 struct vport *p = OVS_CB(skb)->vport;
 413                 if (p) {
 414                         atomic_inc(&p->sflow_pool);
 415                         if (dp->sflow_probability == UINT_MAX ||
 416                             net_random() < dp->sflow_probability)
 417                                 sflow_sample(dp, skb, a, n_actions, p);
 418                 }
 419         }
 420
 421         OVS_CB(skb)->tun_id = 0;
 422
 423         for (; n_actions > 0; a++, n_actions--) {
 424                 if (prev_port != -1) {
 425                         do_output(dp, skb_clone(skb, GFP_ATOMIC), prev_port);
 426                         prev_port = -1;
 427                 }
 428
 429                 switch (a->type) {
 430                 case ODPAT_OUTPUT:
 431                         prev_port = a->output.port;
 432                         break;
 433
 434                 case ODPAT_CONTROLLER:
 435                         err = output_control(dp, skb, a->controller.arg);
 436                         if (err) {
 437                                 kfree_skb(skb);
 438                                 return err;
 439                         }
 440                         break;
 441
 442                 case ODPAT_SET_TUNNEL:
 443                         OVS_CB(skb)->tun_id = a->tunnel.tun_id;
 444                         break;
 445
 446                 case ODPAT_SET_DL_TCI:
 447                         skb = modify_vlan_tci(dp, skb, key, a, n_actions);
 448                         if (IS_ERR(skb))
 449                                 return PTR_ERR(skb);
 450                         break;
 451
 452                 case ODPAT_STRIP_VLAN:
 453                         skb = strip_vlan(skb);
 454                         break;
 455
 456                 case ODPAT_SET_DL_SRC:
 457                 case ODPAT_SET_DL_DST:
 458                         skb = set_dl_addr(skb, &a->dl_addr);
 459                         break;
 460
 461                 case ODPAT_SET_NW_SRC:
 462                 case ODPAT_SET_NW_DST:
 463                         skb = set_nw_addr(skb, key, &a->nw_addr);
 464                         break;
 465
 466                 case ODPAT_SET_NW_TOS:
 467                         skb = set_nw_tos(skb, key, &a->nw_tos);
 468                         break;
 469
 470                 case ODPAT_SET_TP_SRC:
 471                 case ODPAT_SET_TP_DST:
 472                         skb = set_tp_port(skb, key, &a->tp_port);
 473                         break;
 474
 475                 case ODPAT_SET_PRIORITY:
 476                         skb->priority = a->priority.priority;
 477                         break;
 478
 479                 case ODPAT_POP_PRIORITY:
 480                         skb->priority = priority;
 481                         break;
 482
 483                 case ODPAT_DROP_SPOOFED_ARP:
 484                         if (unlikely(is_spoofed_arp(skb, key)))
 485                                 goto exit;
 486                         break;
 487                 }
 488                 if (!skb)
 489                         return -ENOMEM;
 490         }
 491 exit:
 492         if (prev_port != -1)
 493                 do_output(dp, skb, prev_port);
 494         else
 495                 kfree_skb(skb);
 496         return 0;
 497 }