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/xflow.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 xflow_key *key, const union xflow_action *a,
  75                                        int n_actions, gfp_t gfp)
  76 {
  77         __be16 mask = a->dl_tci.mask;
  78         __be16 tci = a->dl_tci.tci;
  79
  80         skb = make_writable(skb, VLAN_HLEN, gfp);
  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 = (vh->h_vlan_TCI & ~mask) | tci;
  96
  97                 if (OVS_CB(skb)->ip_summed == 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                                                               gfp);
 158                                 }
 159
 160                                 if (unlikely(err)) {
 161                                         while ((segs = nskb)) {
 162                                                 nskb = segs->next;
 163                                                 segs->next = NULL;
 164                                                 kfree_skb(segs);
 165                                         }
 166                                         return ERR_PTR(err);
 167                                 }
 168
 169                                 segs = nskb;
 170                         } while (segs->next);
 171
 172                         skb = segs;
 173                         compute_ip_summed(skb, true);
 174                 }
 175
 176                 /* The hardware-accelerated version of vlan_put_tag() works
 177                  * only for a device that has a VLAN group configured (with
 178                  * e.g. vconfig(8)), so call the software-only version
 179                  * __vlan_put_tag() directly instead.
 180                  */
 181                 skb = __vlan_put_tag(skb, ntohs(tci));
 182                 if (!skb)
 183                         return ERR_PTR(-ENOMEM);
 184
 185                 /* GSO doesn't fix up the hardware computed checksum so this
 186                  * will only be hit in the non-GSO case. */
 187                 if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE)
 188                         skb->csum = csum_add(skb->csum, csum_partial(skb->data
 189                                                 + ETH_HLEN, VLAN_HLEN, 0));
 190         }
 191
 192         return skb;
 193 }
 194
 195 static struct sk_buff *strip_vlan(struct sk_buff *skb, gfp_t gfp)
 196 {
 197         skb = make_writable(skb, 0, gfp);
 198         if (skb)
 199                 vlan_pull_tag(skb);
 200
 201         return skb;
 202 }
 203
 204 static struct sk_buff *set_dl_addr(struct sk_buff *skb,
 205                                    const struct xflow_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 == XFLOWAT_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 xflow_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 xflow_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 xflow_key *key,
 260                                    const struct xflow_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 == XFLOWAT_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 xflow_key *key,
 289                                    const struct xflow_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 xflow_key *key,
 313                                    const struct xflow_action_tp_port *a,
 314                                    gfp_t gfp)
 315 {
 316         struct udphdr *th;
 317         __sum16 *check;
 318         __be16 *port;
 319
 320         if (unlikely(!is_ip(skb, key)))
 321                 return skb;
 322
 323         skb = make_writable(skb, 0, gfp);
 324         if (unlikely(!skb))
 325                 return NULL;
 326
 327         /* Must follow make_writable() since that can move the skb data. */
 328         check = get_l4_checksum(skb, key);
 329         if (unlikely(!check))
 330                 return skb;
 331
 332         /*
 333          * Update port and checksum.
 334          *
 335          * This is OK because source and destination port numbers are at the
 336          * same offsets in both UDP and TCP headers, and get_l4_checksum() only
 337          * supports those protocols.
 338          */
 339         th = udp_hdr(skb);
 340         port = a->type == XFLOWAT_SET_TP_SRC ? &th->source : &th->dest;
 341         update_csum(check, skb, *port, a->tp_port, 0);
 342         *port = a->tp_port;
 343
 344         return skb;
 345 }
 346
 347 /**
 348  * is_spoofed_arp - check for invalid ARP packet
 349  *
 350  * @skb: skbuff containing an Ethernet packet, with network header pointing
 351  * just past the Ethernet and optional 802.1Q header.
 352  * @key: flow key extracted from @skb by flow_extract()
 353  *
 354  * Returns true if @skb is an invalid Ethernet+IPv4 ARP packet: one with screwy
 355  * or truncated header fields or one whose inner and outer Ethernet address
 356  * differ.
 357  */
 358 static bool is_spoofed_arp(struct sk_buff *skb, const struct xflow_key *key)
 359 {
 360         struct arp_eth_header *arp;
 361
 362         if (key->dl_type != htons(ETH_P_ARP))
 363                 return false;
 364
 365         if (skb_network_offset(skb) + sizeof(struct arp_eth_header) > skb->len)
 366                 return true;
 367
 368         arp = (struct arp_eth_header *)skb_network_header(skb);
 369         return (arp->ar_hrd != htons(ARPHRD_ETHER) ||
 370                 arp->ar_pro != htons(ETH_P_IP) ||
 371                 arp->ar_hln != ETH_ALEN ||
 372                 arp->ar_pln != 4 ||
 373                 compare_ether_addr(arp->ar_sha, eth_hdr(skb)->h_source));
 374 }
 375
 376 static void do_output(struct datapath *dp, struct sk_buff *skb, int out_port)
 377 {
 378         struct dp_port *p;
 379
 380         if (!skb)
 381                 goto error;
 382
 383         p = rcu_dereference(dp->ports[out_port]);
 384         if (!p)
 385                 goto error;
 386
 387         vport_send(p->vport, skb);
 388         return;
 389
 390 error:
 391         kfree_skb(skb);
 392 }
 393
 394 /* Never consumes 'skb'.  Returns a port that 'skb' should be sent to, -1 if
 395  * none.  */
 396 static int output_group(struct datapath *dp, __u16 group,
 397                         struct sk_buff *skb, gfp_t gfp)
 398 {
 399         struct dp_port_group *g = rcu_dereference(dp->groups[group]);
 400         int prev_port = -1;
 401         int i;
 402
 403         if (!g)
 404                 return -1;
 405         for (i = 0; i < g->n_ports; i++) {
 406                 struct dp_port *p = rcu_dereference(dp->ports[g->ports[i]]);
 407                 if (!p || OVS_CB(skb)->dp_port == p)
 408                         continue;
 409                 if (prev_port != -1) {
 410                         struct sk_buff *clone = skb_clone(skb, gfp);
 411                         if (!clone)
 412                                 return -1;
 413                         do_output(dp, clone, prev_port);
 414                 }
 415                 prev_port = p->port_no;
 416         }
 417         return prev_port;
 418 }
 419
 420 static int output_control(struct datapath *dp, struct sk_buff *skb, u32 arg,
 421                           gfp_t gfp)
 422 {
 423         skb = skb_clone(skb, gfp);
 424         if (!skb)
 425                 return -ENOMEM;
 426         return dp_output_control(dp, skb, _XFLOWL_ACTION_NR, arg);
 427 }
 428
 429 /* Send a copy of this packet up to the sFlow agent, along with extra
 430  * information about what happened to it. */
 431 static void sflow_sample(struct datapath *dp, struct sk_buff *skb,
 432                          const union xflow_action *a, int n_actions,
 433                          gfp_t gfp, struct dp_port *dp_port)
 434 {
 435         struct xflow_sflow_sample_header *hdr;
 436         unsigned int actlen = n_actions * sizeof(union xflow_action);
 437         unsigned int hdrlen = sizeof(struct xflow_sflow_sample_header);
 438         struct sk_buff *nskb;
 439
 440         nskb = skb_copy_expand(skb, actlen + hdrlen, 0, gfp);
 441         if (!nskb)
 442                 return;
 443
 444         memcpy(__skb_push(nskb, actlen), a, actlen);
 445         hdr = (struct xflow_sflow_sample_header*)__skb_push(nskb, hdrlen);
 446         hdr->n_actions = n_actions;
 447         hdr->sample_pool = atomic_read(&dp_port->sflow_pool);
 448         dp_output_control(dp, nskb, _XFLOWL_SFLOW_NR, 0);
 449 }
 450
 451 /* Execute a list of actions against 'skb'. */
 452 int execute_actions(struct datapath *dp, struct sk_buff *skb,
 453                     const struct xflow_key *key,
 454                     const union xflow_action *a, int n_actions,
 455                     gfp_t gfp)
 456 {
 457         /* Every output action needs a separate clone of 'skb', but the common
 458          * case is just a single output action, so that doing a clone and
 459          * then freeing the original skbuff is wasteful.  So the following code
 460          * is slightly obscure just to avoid that. */
 461         int prev_port = -1;
 462         u32 priority = skb->priority;
 463         int err;
 464
 465         if (dp->sflow_probability) {
 466                 struct dp_port *p = OVS_CB(skb)->dp_port;
 467                 if (p) {
 468                         atomic_inc(&p->sflow_pool);
 469                         if (dp->sflow_probability == UINT_MAX ||
 470                             net_random() < dp->sflow_probability)
 471                                 sflow_sample(dp, skb, a, n_actions, gfp, p);
 472                 }
 473         }
 474
 475         OVS_CB(skb)->tun_id = 0;
 476
 477         for (; n_actions > 0; a++, n_actions--) {
 478                 if (prev_port != -1) {
 479                         do_output(dp, skb_clone(skb, gfp), prev_port);
 480                         prev_port = -1;
 481                 }
 482
 483                 switch (a->type) {
 484                 case XFLOWAT_OUTPUT:
 485                         prev_port = a->output.port;
 486                         break;
 487
 488                 case XFLOWAT_OUTPUT_GROUP:
 489                         prev_port = output_group(dp, a->output_group.group,
 490                                                  skb, gfp);
 491                         break;
 492
 493                 case XFLOWAT_CONTROLLER:
 494                         err = output_control(dp, skb, a->controller.arg, gfp);
 495                         if (err) {
 496                                 kfree_skb(skb);
 497                                 return err;
 498                         }
 499                         break;
 500
 501                 case XFLOWAT_SET_TUNNEL:
 502                         OVS_CB(skb)->tun_id = a->tunnel.tun_id;
 503                         break;
 504
 505                 case XFLOWAT_SET_DL_TCI:
 506                         skb = modify_vlan_tci(dp, skb, key, a, n_actions, gfp);
 507                         if (IS_ERR(skb))
 508                                 return PTR_ERR(skb);
 509                         break;
 510
 511                 case XFLOWAT_STRIP_VLAN:
 512                         skb = strip_vlan(skb, gfp);
 513                         break;
 514
 515                 case XFLOWAT_SET_DL_SRC:
 516                 case XFLOWAT_SET_DL_DST:
 517                         skb = set_dl_addr(skb, &a->dl_addr, gfp);
 518                         break;
 519
 520                 case XFLOWAT_SET_NW_SRC:
 521                 case XFLOWAT_SET_NW_DST:
 522                         skb = set_nw_addr(skb, key, &a->nw_addr, gfp);
 523                         break;
 524
 525                 case XFLOWAT_SET_NW_TOS:
 526                         skb = set_nw_tos(skb, key, &a->nw_tos, gfp);
 527                         break;
 528
 529                 case XFLOWAT_SET_TP_SRC:
 530                 case XFLOWAT_SET_TP_DST:
 531                         skb = set_tp_port(skb, key, &a->tp_port, gfp);
 532                         break;
 533
 534                 case XFLOWAT_SET_PRIORITY:
 535                         skb->priority = a->priority.priority;
 536                         break;
 537
 538                 case XFLOWAT_POP_PRIORITY:
 539                         skb->priority = priority;
 540                         break;
 541
 542                 case XFLOWAT_DROP_SPOOFED_ARP:
 543                         if (unlikely(is_spoofed_arp(skb, key)))
 544                                 goto exit;
 545                         break;
 546                 }
 547                 if (!skb)
 548                         return -ENOMEM;
 549         }
 550 exit:
 551         if (prev_port != -1)
 552                 do_output(dp, skb, prev_port);
 553         else
 554                 kfree_skb(skb);
 555         return 0;
 556 }