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_vlan.h>
  18 #include <net/inet_ecn.h>
  19 #include <net/ip.h>
  20 #include <net/checksum.h>
  21
  22 #include "actions.h"
  23 #include "datapath.h"
  24 #include "openvswitch/xflow.h"
  25 #include "vport.h"
  26
  27 static struct sk_buff *
  28 make_writable(struct sk_buff *skb, unsigned min_headroom, gfp_t gfp)
  29 {
  30         if (skb_shared(skb) || 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 void set_tunnel(struct sk_buff *skb, struct xflow_key *key,
  51                        __be32 tun_id)
  52 {
  53         OVS_CB(skb)->tun_id = key->tun_id = tun_id;
  54 }
  55
  56 static struct sk_buff *
  57 vlan_pull_tag(struct sk_buff *skb)
  58 {
  59         struct vlan_ethhdr *vh = vlan_eth_hdr(skb);
  60         struct ethhdr *eh;
  61
  62         /* Verify we were given a vlan packet */
  63         if (vh->h_vlan_proto != htons(ETH_P_8021Q))
  64                 return skb;
  65
  66         if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE)
  67                 skb->csum = csum_sub(skb->csum, csum_partial(skb->data
  68                                         + ETH_HLEN, VLAN_HLEN, 0));
  69
  70         memmove(skb->data + VLAN_HLEN, skb->data, 2 * VLAN_ETH_ALEN);
  71
  72         eh = (struct ethhdr *)skb_pull(skb, VLAN_HLEN);
  73
  74         skb->protocol = eh->h_proto;
  75         skb->mac_header += VLAN_HLEN;
  76
  77         return skb;
  78 }
  79
  80
  81 static struct sk_buff *
  82 modify_vlan_tci(struct datapath *dp, struct sk_buff *skb,
  83                 struct xflow_key *key, const union xflow_action *a,
  84                 int n_actions, gfp_t gfp)
  85 {
  86         __be16 mask = a->dl_tci.mask;
  87         __be16 tci = a->dl_tci.tci;
  88
  89         key->dl_tci = (key->dl_tci & ~(mask | VLAN_TAG_PRESENT)) | tci;
  90
  91         skb = make_writable(skb, VLAN_HLEN, gfp);
  92         if (!skb)
  93                 return ERR_PTR(-ENOMEM);
  94
  95         if (skb->protocol == htons(ETH_P_8021Q)) {
  96                 /* Modify vlan id, but maintain other TCI values */
  97                 struct vlan_ethhdr *vh = vlan_eth_hdr(skb);
  98                 __be16 old_tci = vh->h_vlan_TCI;
  99
 100                 vh->h_vlan_TCI = (vh->h_vlan_TCI & ~mask) | tci;
 101
 102                 if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE) {
 103                         __be16 diff[] = { ~old_tci, vh->h_vlan_TCI };
 104
 105                         skb->csum = ~csum_partial((char *)diff, sizeof(diff),
 106                                                 ~skb->csum);
 107                 }
 108         } else {
 109                 /* Add vlan header */
 110
 111                 /* Set up checksumming pointers for checksum-deferred packets
 112                  * on Xen.  Otherwise, dev_queue_xmit() will try to do this
 113                  * when we send the packet out on the wire, and it will fail at
 114                  * that point because skb_checksum_setup() will not look inside
 115                  * an 802.1Q header. */
 116                 vswitch_skb_checksum_setup(skb);
 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                                         struct xflow_key segkey = *key;
 154                                         err = execute_actions(dp, segs,
 155                                                               &segkey, 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,
 196                                   struct xflow_key *key, gfp_t gfp)
 197 {
 198         skb = make_writable(skb, 0, gfp);
 199         if (skb) {
 200                 vlan_pull_tag(skb);
 201                 key->dl_tci = htons(0);
 202         }
 203         return skb;
 204 }
 205
 206 static struct sk_buff *set_dl_addr(struct sk_buff *skb,
 207                                    struct xflow_key *key,
 208                                    const struct xflow_action_dl_addr *a,
 209                                    gfp_t gfp)
 210 {
 211         skb = make_writable(skb, 0, gfp);
 212         if (skb) {
 213                 struct ethhdr *eh = eth_hdr(skb);
 214                 if (a->type == XFLOWAT_SET_DL_SRC) {
 215                         memcpy(eh->h_source, a->dl_addr, ETH_ALEN);
 216                         memcpy(key->dl_src, a->dl_addr, ETH_ALEN);
 217                 } else {
 218                         memcpy(eh->h_dest, a->dl_addr, ETH_ALEN);
 219                         memcpy(key->dl_dst, a->dl_addr, ETH_ALEN);
 220                 }
 221         }
 222         return skb;
 223 }
 224
 225 /* Updates 'sum', which is a field in 'skb''s data, given that a 4-byte field
 226  * covered by the sum has been changed from 'from' to 'to'.  If set,
 227  * 'pseudohdr' indicates that the field is in the TCP or UDP pseudo-header.
 228  * Based on nf_proto_csum_replace4. */
 229 static void update_csum(__sum16 *sum, struct sk_buff *skb,
 230                         __be32 from, __be32 to, int pseudohdr)
 231 {
 232         __be32 diff[] = { ~from, to };
 233
 234         if (OVS_CB(skb)->ip_summed != OVS_CSUM_PARTIAL) {
 235                 *sum = csum_fold(csum_partial((char *)diff, sizeof(diff),
 236                                 ~csum_unfold(*sum)));
 237                 if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE && pseudohdr)
 238                         skb->csum = ~csum_partial((char *)diff, sizeof(diff),
 239                                                 ~skb->csum);
 240         } else if (pseudohdr)
 241                 *sum = ~csum_fold(csum_partial((char *)diff, sizeof(diff),
 242                                 csum_unfold(*sum)));
 243 }
 244
 245 static struct sk_buff *set_nw_addr(struct sk_buff *skb,
 246                                    struct xflow_key *key,
 247                                    const struct xflow_action_nw_addr *a,
 248                                    gfp_t gfp)
 249 {
 250         if (key->dl_type != htons(ETH_P_IP))
 251                 return skb;
 252
 253         skb = make_writable(skb, 0, gfp);
 254         if (skb) {
 255                 struct iphdr *nh = ip_hdr(skb);
 256                 u32 *f = a->type == XFLOWAT_SET_NW_SRC ? &nh->saddr : &nh->daddr;
 257                 u32 old = *f;
 258                 u32 new = a->nw_addr;
 259
 260                 if (key->nw_proto == IPPROTO_TCP) {
 261                         struct tcphdr *th = tcp_hdr(skb);
 262                         update_csum(&th->check, skb, old, new, 1);
 263                 } else if (key->nw_proto == IPPROTO_UDP) {
 264                         struct udphdr *th = udp_hdr(skb);
 265                         update_csum(&th->check, skb, old, new, 1);
 266                 }
 267                 update_csum(&nh->check, skb, old, new, 0);
 268                 *f = new;
 269
 270                 if (a->type == XFLOWAT_SET_NW_SRC)
 271                         key->nw_src = a->nw_addr;
 272                 else
 273                         key->nw_dst = a->nw_addr;
 274         }
 275         return skb;
 276 }
 277
 278 static struct sk_buff *set_nw_tos(struct sk_buff *skb,
 279                                    struct xflow_key *key,
 280                                    const struct xflow_action_nw_tos *a,
 281                                    gfp_t gfp)
 282 {
 283         if (key->dl_type != htons(ETH_P_IP))
 284                 return skb;
 285
 286         skb = make_writable(skb, 0, gfp);
 287         if (skb) {
 288                 struct iphdr *nh = ip_hdr(skb);
 289                 u8 *f = &nh->tos;
 290                 u8 old = *f;
 291                 u8 new;
 292
 293                 /* Set the DSCP bits and preserve the ECN bits. */
 294                 new = a->nw_tos | (nh->tos & INET_ECN_MASK);
 295                 update_csum(&nh->check, skb, htons((uint16_t)old),
 296                                 htons((uint16_t)new), 0);
 297                 *f = new;
 298                 key->nw_tos = a->nw_tos;
 299         }
 300         return skb;
 301 }
 302
 303 static struct sk_buff *
 304 set_tp_port(struct sk_buff *skb, struct xflow_key *key,
 305             const struct xflow_action_tp_port *a,
 306             gfp_t gfp)
 307 {
 308         int check_ofs;
 309
 310         if (key->dl_type != htons(ETH_P_IP))
 311                 return skb;
 312
 313         if (key->nw_proto == IPPROTO_TCP)
 314                 check_ofs = offsetof(struct tcphdr, check);
 315         else if (key->nw_proto == IPPROTO_UDP)
 316                 check_ofs = offsetof(struct udphdr, check);
 317         else
 318                 return skb;
 319
 320         skb = make_writable(skb, 0, gfp);
 321         if (skb) {
 322                 struct udphdr *th = udp_hdr(skb);
 323                 u16 *f = a->type == XFLOWAT_SET_TP_SRC ? &th->source : &th->dest;
 324                 u16 old = *f;
 325                 u16 new = a->tp_port;
 326                 update_csum((u16*)(skb_transport_header(skb) + check_ofs),
 327                                 skb, old, new, 0);
 328                 *f = new;
 329                 if (a->type == XFLOWAT_SET_TP_SRC)
 330                         key->tp_src = a->tp_port;
 331                 else
 332                         key->tp_dst = a->tp_port;
 333         }
 334         return skb;
 335 }
 336
 337 static inline unsigned packet_length(const struct sk_buff *skb)
 338 {
 339         unsigned length = skb->len - ETH_HLEN;
 340         if (skb->protocol == htons(ETH_P_8021Q))
 341                 length -= VLAN_HLEN;
 342         return length;
 343 }
 344
 345 static void
 346 do_output(struct datapath *dp, struct sk_buff *skb, int out_port)
 347 {
 348         struct dp_port *p;
 349         int mtu;
 350
 351         if (!skb)
 352                 goto error;
 353
 354         p = rcu_dereference(dp->ports[out_port]);
 355         if (!p)
 356                 goto error;
 357
 358         mtu = vport_get_mtu(p->vport);
 359         if (packet_length(skb) > mtu && !skb_is_gso(skb)) {
 360                 printk(KERN_WARNING "%s: dropped over-mtu packet: %d > %d\n",
 361                        dp_name(dp), packet_length(skb), mtu);
 362                 goto error;
 363         }
 364
 365         vport_send(p->vport, skb);
 366         return;
 367
 368 error:
 369         kfree_skb(skb);
 370 }
 371
 372 /* Never consumes 'skb'.  Returns a port that 'skb' should be sent to, -1 if
 373  * none.  */
 374 static int output_group(struct datapath *dp, __u16 group,
 375                         struct sk_buff *skb, gfp_t gfp)
 376 {
 377         struct dp_port_group *g = rcu_dereference(dp->groups[group]);
 378         int prev_port = -1;
 379         int i;
 380
 381         if (!g)
 382                 return -1;
 383         for (i = 0; i < g->n_ports; i++) {
 384                 struct dp_port *p = rcu_dereference(dp->ports[g->ports[i]]);
 385                 if (!p || OVS_CB(skb)->dp_port == p)
 386                         continue;
 387                 if (prev_port != -1) {
 388                         struct sk_buff *clone = skb_clone(skb, gfp);
 389                         if (!clone)
 390                                 return -1;
 391                         do_output(dp, clone, prev_port);
 392                 }
 393                 prev_port = p->port_no;
 394         }
 395         return prev_port;
 396 }
 397
 398 static int
 399 output_control(struct datapath *dp, struct sk_buff *skb, u32 arg, gfp_t gfp)
 400 {
 401         skb = skb_clone(skb, gfp);
 402         if (!skb)
 403                 return -ENOMEM;
 404         return dp_output_control(dp, skb, _XFLOWL_ACTION_NR, arg);
 405 }
 406
 407 /* Send a copy of this packet up to the sFlow agent, along with extra
 408  * information about what happened to it. */
 409 static void sflow_sample(struct datapath *dp, struct sk_buff *skb,
 410                          const union xflow_action *a, int n_actions,
 411                          gfp_t gfp, struct dp_port *dp_port)
 412 {
 413         struct xflow_sflow_sample_header *hdr;
 414         unsigned int actlen = n_actions * sizeof(union xflow_action);
 415         unsigned int hdrlen = sizeof(struct xflow_sflow_sample_header);
 416         struct sk_buff *nskb;
 417
 418         nskb = skb_copy_expand(skb, actlen + hdrlen, 0, gfp);
 419         if (!nskb)
 420                 return;
 421
 422         memcpy(__skb_push(nskb, actlen), a, actlen);
 423         hdr = (struct xflow_sflow_sample_header*)__skb_push(nskb, hdrlen);
 424         hdr->n_actions = n_actions;
 425         hdr->sample_pool = atomic_read(&dp_port->sflow_pool);
 426         dp_output_control(dp, nskb, _XFLOWL_SFLOW_NR, 0);
 427 }
 428
 429 /* Execute a list of actions against 'skb'. */
 430 int execute_actions(struct datapath *dp, struct sk_buff *skb,
 431                     struct xflow_key *key,
 432                     const union xflow_action *a, int n_actions,
 433                     gfp_t gfp)
 434 {
 435         /* Every output action needs a separate clone of 'skb', but the common
 436          * case is just a single output action, so that doing a clone and
 437          * then freeing the original skbuff is wasteful.  So the following code
 438          * is slightly obscure just to avoid that. */
 439         int prev_port = -1;
 440         int err;
 441
 442         if (dp->sflow_probability) {
 443                 struct dp_port *p = OVS_CB(skb)->dp_port;
 444                 if (p) {
 445                         atomic_inc(&p->sflow_pool);
 446                         if (dp->sflow_probability == UINT_MAX ||
 447                             net_random() < dp->sflow_probability)
 448                                 sflow_sample(dp, skb, a, n_actions, gfp, p);
 449                 }
 450         }
 451
 452         OVS_CB(skb)->tun_id = 0;
 453
 454         for (; n_actions > 0; a++, n_actions--) {
 455                 WARN_ON_ONCE(skb_shared(skb));
 456                 if (prev_port != -1) {
 457                         do_output(dp, skb_clone(skb, gfp), prev_port);
 458                         prev_port = -1;
 459                 }
 460
 461                 switch (a->type) {
 462                 case XFLOWAT_OUTPUT:
 463                         prev_port = a->output.port;
 464                         break;
 465
 466                 case XFLOWAT_OUTPUT_GROUP:
 467                         prev_port = output_group(dp, a->output_group.group,
 468                                                  skb, gfp);
 469                         break;
 470
 471                 case XFLOWAT_CONTROLLER:
 472                         err = output_control(dp, skb, a->controller.arg, gfp);
 473                         if (err) {
 474                                 kfree_skb(skb);
 475                                 return err;
 476                         }
 477                         break;
 478
 479                 case XFLOWAT_SET_TUNNEL:
 480                         set_tunnel(skb, key, a->tunnel.tun_id);
 481                         break;
 482
 483                 case XFLOWAT_SET_DL_TCI:
 484                         skb = modify_vlan_tci(dp, skb, key, a, n_actions, gfp);
 485                         if (IS_ERR(skb))
 486                                 return PTR_ERR(skb);
 487                         break;
 488
 489                 case XFLOWAT_STRIP_VLAN:
 490                         skb = strip_vlan(skb, key, gfp);
 491                         break;
 492
 493                 case XFLOWAT_SET_DL_SRC:
 494                 case XFLOWAT_SET_DL_DST:
 495                         skb = set_dl_addr(skb, key, &a->dl_addr, gfp);
 496                         break;
 497
 498                 case XFLOWAT_SET_NW_SRC:
 499                 case XFLOWAT_SET_NW_DST:
 500                         skb = set_nw_addr(skb, key, &a->nw_addr, gfp);
 501                         break;
 502
 503                 case XFLOWAT_SET_NW_TOS:
 504                         skb = set_nw_tos(skb, key, &a->nw_tos, gfp);
 505                         break;
 506
 507                 case XFLOWAT_SET_TP_SRC:
 508                 case XFLOWAT_SET_TP_DST:
 509                         skb = set_tp_port(skb, key, &a->tp_port, gfp);
 510                         break;
 511                 }
 512                 if (!skb)
 513                         return -ENOMEM;
 514         }
 515         if (prev_port != -1)
 516                 do_output(dp, skb, prev_port);
 517         else
 518                 kfree_skb(skb);
 519         return 0;
 520 }