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                 int err;
 110
 111                 /* Add vlan header */
 112
 113                 /* Set up checksumming pointers for checksum-deferred packets
 114                  * on Xen.  Otherwise, dev_queue_xmit() will try to do this
 115                  * when we send the packet out on the wire, and it will fail at
 116                  * that point because skb_checksum_setup() will not look inside
 117                  * an 802.1Q header. */
 118                 err = vswitch_skb_checksum_setup(skb);
 119                 if (unlikely(err)) {
 120                         kfree_skb(skb);
 121                         return ERR_PTR(err);
 122                 }
 123
 124                 /* GSO is not implemented for packets with an 802.1Q header, so
 125                  * we have to do segmentation before we add that header.
 126                  *
 127                  * GSO does work with hardware-accelerated VLAN tagging, but we
 128                  * can't use hardware-accelerated VLAN tagging since it
 129                  * requires the device to have a VLAN group configured (with
 130                  * e.g. vconfig(8)) and we don't do that.
 131                  *
 132                  * Having to do this here may be a performance loss, since we
 133                  * can't take advantage of TSO hardware support, although it
 134                  * does not make a measurable network performance difference
 135                  * for 1G Ethernet.  Fixing that would require patching the
 136                  * kernel (either to add GSO support to the VLAN protocol or to
 137                  * support hardware-accelerated VLAN tagging without VLAN
 138                  * groups configured). */
 139                 if (skb_is_gso(skb)) {
 140                         struct sk_buff *segs;
 141
 142                         segs = skb_gso_segment(skb, 0);
 143                         kfree_skb(skb);
 144                         if (unlikely(IS_ERR(segs)))
 145                                 return ERR_CAST(segs);
 146
 147                         do {
 148                                 struct sk_buff *nskb = segs->next;
 149                                 int err;
 150
 151                                 segs->next = NULL;
 152
 153                                 /* GSO can change the checksum type so update.*/
 154                                 compute_ip_summed(segs, true);
 155
 156                                 segs = __vlan_put_tag(segs, ntohs(tci));
 157                                 err = -ENOMEM;
 158                                 if (segs) {
 159                                         struct xflow_key segkey = *key;
 160                                         err = execute_actions(dp, segs,
 161                                                               &segkey, a + 1,
 162                                                               n_actions - 1,
 163                                                               gfp);
 164                                 }
 165
 166                                 if (unlikely(err)) {
 167                                         while ((segs = nskb)) {
 168                                                 nskb = segs->next;
 169                                                 segs->next = NULL;
 170                                                 kfree_skb(segs);
 171                                         }
 172                                         return ERR_PTR(err);
 173                                 }
 174
 175                                 segs = nskb;
 176                         } while (segs->next);
 177
 178                         skb = segs;
 179                         compute_ip_summed(skb, true);
 180                 }
 181
 182                 /* The hardware-accelerated version of vlan_put_tag() works
 183                  * only for a device that has a VLAN group configured (with
 184                  * e.g. vconfig(8)), so call the software-only version
 185                  * __vlan_put_tag() directly instead.
 186                  */
 187                 skb = __vlan_put_tag(skb, ntohs(tci));
 188                 if (!skb)
 189                         return ERR_PTR(-ENOMEM);
 190
 191                 /* GSO doesn't fix up the hardware computed checksum so this
 192                  * will only be hit in the non-GSO case. */
 193                 if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE)
 194                         skb->csum = csum_add(skb->csum, csum_partial(skb->data
 195                                                 + ETH_HLEN, VLAN_HLEN, 0));
 196         }
 197
 198         return skb;
 199 }
 200
 201 static struct sk_buff *strip_vlan(struct sk_buff *skb,
 202                                   struct xflow_key *key, gfp_t gfp)
 203 {
 204         skb = make_writable(skb, 0, gfp);
 205         if (skb) {
 206                 vlan_pull_tag(skb);
 207                 key->dl_tci = htons(0);
 208         }
 209         return skb;
 210 }
 211
 212 static struct sk_buff *set_dl_addr(struct sk_buff *skb,
 213                                    struct xflow_key *key,
 214                                    const struct xflow_action_dl_addr *a,
 215                                    gfp_t gfp)
 216 {
 217         skb = make_writable(skb, 0, gfp);
 218         if (skb) {
 219                 struct ethhdr *eh = eth_hdr(skb);
 220                 if (a->type == XFLOWAT_SET_DL_SRC) {
 221                         memcpy(eh->h_source, a->dl_addr, ETH_ALEN);
 222                         memcpy(key->dl_src, a->dl_addr, ETH_ALEN);
 223                 } else {
 224                         memcpy(eh->h_dest, a->dl_addr, ETH_ALEN);
 225                         memcpy(key->dl_dst, a->dl_addr, ETH_ALEN);
 226                 }
 227         }
 228         return skb;
 229 }
 230
 231 /* Updates 'sum', which is a field in 'skb''s data, given that a 4-byte field
 232  * covered by the sum has been changed from 'from' to 'to'.  If set,
 233  * 'pseudohdr' indicates that the field is in the TCP or UDP pseudo-header.
 234  * Based on nf_proto_csum_replace4. */
 235 static void update_csum(__sum16 *sum, struct sk_buff *skb,
 236                         __be32 from, __be32 to, int pseudohdr)
 237 {
 238         __be32 diff[] = { ~from, to };
 239
 240         if (OVS_CB(skb)->ip_summed != OVS_CSUM_PARTIAL) {
 241                 *sum = csum_fold(csum_partial((char *)diff, sizeof(diff),
 242                                 ~csum_unfold(*sum)));
 243                 if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE && pseudohdr)
 244                         skb->csum = ~csum_partial((char *)diff, sizeof(diff),
 245                                                 ~skb->csum);
 246         } else if (pseudohdr)
 247                 *sum = ~csum_fold(csum_partial((char *)diff, sizeof(diff),
 248                                 csum_unfold(*sum)));
 249 }
 250
 251 static struct sk_buff *set_nw_addr(struct sk_buff *skb,
 252                                    struct xflow_key *key,
 253                                    const struct xflow_action_nw_addr *a,
 254                                    gfp_t gfp)
 255 {
 256         if (key->dl_type != htons(ETH_P_IP))
 257                 return skb;
 258
 259         skb = make_writable(skb, 0, gfp);
 260         if (skb) {
 261                 struct iphdr *nh = ip_hdr(skb);
 262                 u32 *f = a->type == XFLOWAT_SET_NW_SRC ? &nh->saddr : &nh->daddr;
 263                 u32 old = *f;
 264                 u32 new = a->nw_addr;
 265
 266                 if (key->nw_proto == IPPROTO_TCP) {
 267                         struct tcphdr *th = tcp_hdr(skb);
 268                         update_csum(&th->check, skb, old, new, 1);
 269                 } else if (key->nw_proto == IPPROTO_UDP) {
 270                         struct udphdr *th = udp_hdr(skb);
 271                         update_csum(&th->check, skb, old, new, 1);
 272                 }
 273                 update_csum(&nh->check, skb, old, new, 0);
 274                 *f = new;
 275
 276                 if (a->type == XFLOWAT_SET_NW_SRC)
 277                         key->nw_src = a->nw_addr;
 278                 else
 279                         key->nw_dst = a->nw_addr;
 280         }
 281         return skb;
 282 }
 283
 284 static struct sk_buff *set_nw_tos(struct sk_buff *skb,
 285                                    struct xflow_key *key,
 286                                    const struct xflow_action_nw_tos *a,
 287                                    gfp_t gfp)
 288 {
 289         if (key->dl_type != htons(ETH_P_IP))
 290                 return skb;
 291
 292         skb = make_writable(skb, 0, gfp);
 293         if (skb) {
 294                 struct iphdr *nh = ip_hdr(skb);
 295                 u8 *f = &nh->tos;
 296                 u8 old = *f;
 297                 u8 new;
 298
 299                 /* Set the DSCP bits and preserve the ECN bits. */
 300                 new = a->nw_tos | (nh->tos & INET_ECN_MASK);
 301                 update_csum(&nh->check, skb, htons((uint16_t)old),
 302                                 htons((uint16_t)new), 0);
 303                 *f = new;
 304                 key->nw_tos = a->nw_tos;
 305         }
 306         return skb;
 307 }
 308
 309 static struct sk_buff *
 310 set_tp_port(struct sk_buff *skb, struct xflow_key *key,
 311             const struct xflow_action_tp_port *a,
 312             gfp_t gfp)
 313 {
 314         int check_ofs;
 315
 316         if (key->dl_type != htons(ETH_P_IP))
 317                 return skb;
 318
 319         if (key->nw_proto == IPPROTO_TCP)
 320                 check_ofs = offsetof(struct tcphdr, check);
 321         else if (key->nw_proto == IPPROTO_UDP)
 322                 check_ofs = offsetof(struct udphdr, check);
 323         else
 324                 return skb;
 325
 326         skb = make_writable(skb, 0, gfp);
 327         if (skb) {
 328                 struct udphdr *th = udp_hdr(skb);
 329                 u16 *f = a->type == XFLOWAT_SET_TP_SRC ? &th->source : &th->dest;
 330                 u16 old = *f;
 331                 u16 new = a->tp_port;
 332                 update_csum((u16*)(skb_transport_header(skb) + check_ofs),
 333                                 skb, old, new, 0);
 334                 *f = new;
 335                 if (a->type == XFLOWAT_SET_TP_SRC)
 336                         key->tp_src = a->tp_port;
 337                 else
 338                         key->tp_dst = a->tp_port;
 339         }
 340         return skb;
 341 }
 342
 343 static inline unsigned packet_length(const struct sk_buff *skb)
 344 {
 345         unsigned length = skb->len - ETH_HLEN;
 346         if (skb->protocol == htons(ETH_P_8021Q))
 347                 length -= VLAN_HLEN;
 348         return length;
 349 }
 350
 351 static void
 352 do_output(struct datapath *dp, struct sk_buff *skb, int out_port)
 353 {
 354         struct dp_port *p;
 355         int mtu;
 356
 357         if (!skb)
 358                 goto error;
 359
 360         p = rcu_dereference(dp->ports[out_port]);
 361         if (!p)
 362                 goto error;
 363
 364         mtu = vport_get_mtu(p->vport);
 365         if (packet_length(skb) > mtu && !skb_is_gso(skb)) {
 366                 printk(KERN_WARNING "%s: dropped over-mtu packet: %d > %d\n",
 367                        dp_name(dp), packet_length(skb), mtu);
 368                 goto error;
 369         }
 370
 371         vport_send(p->vport, skb);
 372         return;
 373
 374 error:
 375         kfree_skb(skb);
 376 }
 377
 378 /* Never consumes 'skb'.  Returns a port that 'skb' should be sent to, -1 if
 379  * none.  */
 380 static int output_group(struct datapath *dp, __u16 group,
 381                         struct sk_buff *skb, gfp_t gfp)
 382 {
 383         struct dp_port_group *g = rcu_dereference(dp->groups[group]);
 384         int prev_port = -1;
 385         int i;
 386
 387         if (!g)
 388                 return -1;
 389         for (i = 0; i < g->n_ports; i++) {
 390                 struct dp_port *p = rcu_dereference(dp->ports[g->ports[i]]);
 391                 if (!p || OVS_CB(skb)->dp_port == p)
 392                         continue;
 393                 if (prev_port != -1) {
 394                         struct sk_buff *clone = skb_clone(skb, gfp);
 395                         if (!clone)
 396                                 return -1;
 397                         do_output(dp, clone, prev_port);
 398                 }
 399                 prev_port = p->port_no;
 400         }
 401         return prev_port;
 402 }
 403
 404 static int
 405 output_control(struct datapath *dp, struct sk_buff *skb, u32 arg, gfp_t gfp)
 406 {
 407         skb = skb_clone(skb, gfp);
 408         if (!skb)
 409                 return -ENOMEM;
 410         return dp_output_control(dp, skb, _XFLOWL_ACTION_NR, arg);
 411 }
 412
 413 /* Send a copy of this packet up to the sFlow agent, along with extra
 414  * information about what happened to it. */
 415 static void sflow_sample(struct datapath *dp, struct sk_buff *skb,
 416                          const union xflow_action *a, int n_actions,
 417                          gfp_t gfp, struct dp_port *dp_port)
 418 {
 419         struct xflow_sflow_sample_header *hdr;
 420         unsigned int actlen = n_actions * sizeof(union xflow_action);
 421         unsigned int hdrlen = sizeof(struct xflow_sflow_sample_header);
 422         struct sk_buff *nskb;
 423
 424         nskb = skb_copy_expand(skb, actlen + hdrlen, 0, gfp);
 425         if (!nskb)
 426                 return;
 427
 428         memcpy(__skb_push(nskb, actlen), a, actlen);
 429         hdr = (struct xflow_sflow_sample_header*)__skb_push(nskb, hdrlen);
 430         hdr->n_actions = n_actions;
 431         hdr->sample_pool = atomic_read(&dp_port->sflow_pool);
 432         dp_output_control(dp, nskb, _XFLOWL_SFLOW_NR, 0);
 433 }
 434
 435 /* Execute a list of actions against 'skb'. */
 436 int execute_actions(struct datapath *dp, struct sk_buff *skb,
 437                     struct xflow_key *key,
 438                     const union xflow_action *a, int n_actions,
 439                     gfp_t gfp)
 440 {
 441         /* Every output action needs a separate clone of 'skb', but the common
 442          * case is just a single output action, so that doing a clone and
 443          * then freeing the original skbuff is wasteful.  So the following code
 444          * is slightly obscure just to avoid that. */
 445         int prev_port = -1;
 446         u32 priority = skb->priority;
 447         int err;
 448
 449         if (dp->sflow_probability) {
 450                 struct dp_port *p = OVS_CB(skb)->dp_port;
 451                 if (p) {
 452                         atomic_inc(&p->sflow_pool);
 453                         if (dp->sflow_probability == UINT_MAX ||
 454                             net_random() < dp->sflow_probability)
 455                                 sflow_sample(dp, skb, a, n_actions, gfp, p);
 456                 }
 457         }
 458
 459         OVS_CB(skb)->tun_id = 0;
 460
 461         for (; n_actions > 0; a++, n_actions--) {
 462                 WARN_ON_ONCE(skb_shared(skb));
 463                 if (prev_port != -1) {
 464                         do_output(dp, skb_clone(skb, gfp), prev_port);
 465                         prev_port = -1;
 466                 }
 467
 468                 switch (a->type) {
 469                 case XFLOWAT_OUTPUT:
 470                         prev_port = a->output.port;
 471                         break;
 472
 473                 case XFLOWAT_OUTPUT_GROUP:
 474                         prev_port = output_group(dp, a->output_group.group,
 475                                                  skb, gfp);
 476                         break;
 477
 478                 case XFLOWAT_CONTROLLER:
 479                         err = output_control(dp, skb, a->controller.arg, gfp);
 480                         if (err) {
 481                                 kfree_skb(skb);
 482                                 return err;
 483                         }
 484                         break;
 485
 486                 case XFLOWAT_SET_TUNNEL:
 487                         set_tunnel(skb, key, a->tunnel.tun_id);
 488                         break;
 489
 490                 case XFLOWAT_SET_DL_TCI:
 491                         skb = modify_vlan_tci(dp, skb, key, a, n_actions, gfp);
 492                         if (IS_ERR(skb))
 493                                 return PTR_ERR(skb);
 494                         break;
 495
 496                 case XFLOWAT_STRIP_VLAN:
 497                         skb = strip_vlan(skb, key, gfp);
 498                         break;
 499
 500                 case XFLOWAT_SET_DL_SRC:
 501                 case XFLOWAT_SET_DL_DST:
 502                         skb = set_dl_addr(skb, key, &a->dl_addr, gfp);
 503                         break;
 504
 505                 case XFLOWAT_SET_NW_SRC:
 506                 case XFLOWAT_SET_NW_DST:
 507                         skb = set_nw_addr(skb, key, &a->nw_addr, gfp);
 508                         break;
 509
 510                 case XFLOWAT_SET_NW_TOS:
 511                         skb = set_nw_tos(skb, key, &a->nw_tos, gfp);
 512                         break;
 513
 514                 case XFLOWAT_SET_TP_SRC:
 515                 case XFLOWAT_SET_TP_DST:
 516                         skb = set_tp_port(skb, key, &a->tp_port, gfp);
 517                         break;
 518
 519                 case XFLOWAT_SET_PRIORITY:
 520                         skb->priority = a->priority.priority;
 521                         break;
 522
 523                 case XFLOWAT_POP_PRIORITY:
 524                         skb->priority = priority;
 525                         break;
 526                 }
 527                 if (!skb)
 528                         return -ENOMEM;
 529         }
 530         if (prev_port != -1)
 531                 do_output(dp, skb, prev_port);
 532         else
 533                 kfree_skb(skb);
 534         return 0;
 535 }