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/datapath-protocol.h"
  25 #include "vport.h"
  26
  27 static struct sk_buff *make_writable(struct sk_buff *skb, unsigned min_headroom, gfp_t gfp)
  28 {
  29         if (skb_cloned(skb)) {
  30                 struct sk_buff *nskb;
  31                 unsigned headroom = max(min_headroom, skb_headroom(skb));
  32
  33                 nskb = skb_copy_expand(skb, headroom, skb_tailroom(skb), gfp);
  34                 if (nskb) {
  35                         set_skb_csum_bits(skb, nskb);
  36                         kfree_skb(skb);
  37                         return nskb;
  38                 }
  39         } else {
  40                 unsigned int hdr_len = (skb_transport_offset(skb)
  41                                         + sizeof(struct tcphdr));
  42                 if (pskb_may_pull(skb, min(hdr_len, skb->len)))
  43                         return skb;
  44         }
  45         kfree_skb(skb);
  46         return NULL;
  47 }
  48
  49 static struct sk_buff *vlan_pull_tag(struct sk_buff *skb)
  50 {
  51         struct vlan_ethhdr *vh = vlan_eth_hdr(skb);
  52         struct ethhdr *eh;
  53
  54         /* Verify we were given a vlan packet */
  55         if (vh->h_vlan_proto != htons(ETH_P_8021Q))
  56                 return skb;
  57
  58         if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE)
  59                 skb->csum = csum_sub(skb->csum, csum_partial(skb->data
  60                                         + ETH_HLEN, VLAN_HLEN, 0));
  61
  62         memmove(skb->data + VLAN_HLEN, skb->data, 2 * VLAN_ETH_ALEN);
  63
  64         eh = (struct ethhdr *)skb_pull(skb, VLAN_HLEN);
  65
  66         skb->protocol = eh->h_proto;
  67         skb->mac_header += VLAN_HLEN;
  68
  69         return skb;
  70 }
  71
  72 static struct sk_buff *modify_vlan_tci(struct datapath *dp, struct sk_buff *skb,
  73                                        const struct odp_flow_key *key,
  74                                        const union odp_action *a, int n_actions,
  75                                        gfp_t gfp)
  76 {
  77         u16 tci, mask;
  78
  79         if (a->type == ODPAT_SET_VLAN_VID) {
  80                 tci = ntohs(a->vlan_vid.vlan_vid);
  81                 mask = VLAN_VID_MASK;
  82         } else {
  83                 tci = a->vlan_pcp.vlan_pcp << VLAN_PCP_SHIFT;
  84                 mask = VLAN_PCP_MASK;
  85         }
  86
  87         skb = make_writable(skb, VLAN_HLEN, gfp);
  88         if (!skb)
  89                 return ERR_PTR(-ENOMEM);
  90
  91         if (skb->protocol == htons(ETH_P_8021Q)) {
  92                 /* Modify vlan id, but maintain other TCI values */
  93                 struct vlan_ethhdr *vh = vlan_eth_hdr(skb);
  94                 __be16 old_tci = vh->h_vlan_TCI;
  95
  96                 vh->h_vlan_TCI = htons((ntohs(vh->h_vlan_TCI) & ~mask) | tci);
  97
  98                 if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE) {
  99                         __be16 diff[] = { ~old_tci, vh->h_vlan_TCI };
 100
 101                         skb->csum = ~csum_partial((char *)diff, sizeof(diff),
 102                                                 ~skb->csum);
 103                 }
 104         } else {
 105                 int err;
 106
 107                 /* Add vlan header */
 108
 109                 /* Set up checksumming pointers for checksum-deferred packets
 110                  * on Xen.  Otherwise, dev_queue_xmit() will try to do this
 111                  * when we send the packet out on the wire, and it will fail at
 112                  * that point because skb_checksum_setup() will not look inside
 113                  * an 802.1Q header. */
 114                 err = vswitch_skb_checksum_setup(skb);
 115                 if (unlikely(err)) {
 116                         kfree_skb(skb);
 117                         return ERR_PTR(err);
 118                 }
 119
 120                 /* GSO is not implemented for packets with an 802.1Q header, so
 121                  * we have to do segmentation before we add that header.
 122                  *
 123                  * GSO does work with hardware-accelerated VLAN tagging, but we
 124                  * can't use hardware-accelerated VLAN tagging since it
 125                  * requires the device to have a VLAN group configured (with
 126                  * e.g. vconfig(8)) and we don't do that.
 127                  *
 128                  * Having to do this here may be a performance loss, since we
 129                  * can't take advantage of TSO hardware support, although it
 130                  * does not make a measurable network performance difference
 131                  * for 1G Ethernet.  Fixing that would require patching the
 132                  * kernel (either to add GSO support to the VLAN protocol or to
 133                  * support hardware-accelerated VLAN tagging without VLAN
 134                  * groups configured). */
 135                 if (skb_is_gso(skb)) {
 136                         struct sk_buff *segs;
 137
 138                         segs = skb_gso_segment(skb, 0);
 139                         kfree_skb(skb);
 140                         if (unlikely(IS_ERR(segs)))
 141                                 return ERR_CAST(segs);
 142
 143                         do {
 144                                 struct sk_buff *nskb = segs->next;
 145                                 int err;
 146
 147                                 segs->next = NULL;
 148
 149                                 /* GSO can change the checksum type so update.*/
 150                                 compute_ip_summed(segs, true);
 151
 152                                 segs = __vlan_put_tag(segs, tci);
 153                                 err = -ENOMEM;
 154                                 if (segs) {
 155                                         err = execute_actions(dp, segs,
 156                                                               key, a + 1,
 157                                                               n_actions - 1,
 158                                                               gfp);
 159                                 }
 160
 161                                 if (unlikely(err)) {
 162                                         while ((segs = nskb)) {
 163                                                 nskb = segs->next;
 164                                                 segs->next = NULL;
 165                                                 kfree_skb(segs);
 166                                         }
 167                                         return ERR_PTR(err);
 168                                 }
 169
 170                                 segs = nskb;
 171                         } while (segs->next);
 172
 173                         skb = segs;
 174                         compute_ip_summed(skb, true);
 175                 }
 176
 177                 /* The hardware-accelerated version of vlan_put_tag() works
 178                  * only for a device that has a VLAN group configured (with
 179                  * e.g. vconfig(8)), so call the software-only version
 180                  * __vlan_put_tag() directly instead.
 181                  */
 182                 skb = __vlan_put_tag(skb, tci);
 183                 if (!skb)
 184                         return ERR_PTR(-ENOMEM);
 185
 186                 /* GSO doesn't fix up the hardware computed checksum so this
 187                  * will only be hit in the non-GSO case. */
 188                 if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE)
 189                         skb->csum = csum_add(skb->csum, csum_partial(skb->data
 190                                                 + ETH_HLEN, VLAN_HLEN, 0));
 191         }
 192
 193         return skb;
 194 }
 195
 196 static struct sk_buff *strip_vlan(struct sk_buff *skb, gfp_t gfp)
 197 {
 198         skb = make_writable(skb, 0, gfp);
 199         if (skb)
 200                 vlan_pull_tag(skb);
 201
 202         return skb;
 203 }
 204
 205 static struct sk_buff *set_dl_addr(struct sk_buff *skb,
 206                                    const struct odp_action_dl_addr *a,
 207                                    gfp_t gfp)
 208 {
 209         skb = make_writable(skb, 0, gfp);
 210         if (skb) {
 211                 struct ethhdr *eh = eth_hdr(skb);
 212                 if (a->type == ODPAT_SET_DL_SRC)
 213                         memcpy(eh->h_source, a->dl_addr, ETH_ALEN);
 214                 else
 215                         memcpy(eh->h_dest, a->dl_addr, ETH_ALEN);
 216         }
 217         return skb;
 218 }
 219
 220 /* Updates 'sum', which is a field in 'skb''s data, given that a 4-byte field
 221  * covered by the sum has been changed from 'from' to 'to'.  If set,
 222  * 'pseudohdr' indicates that the field is in the TCP or UDP pseudo-header.
 223  * Based on nf_proto_csum_replace4. */
 224 static void update_csum(__sum16 *sum, struct sk_buff *skb,
 225                         __be32 from, __be32 to, int pseudohdr)
 226 {
 227         __be32 diff[] = { ~from, to };
 228
 229         if (OVS_CB(skb)->ip_summed != OVS_CSUM_PARTIAL) {
 230                 *sum = csum_fold(csum_partial((char *)diff, sizeof(diff),
 231                                 ~csum_unfold(*sum)));
 232                 if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE && pseudohdr)
 233                         skb->csum = ~csum_partial((char *)diff, sizeof(diff),
 234                                                 ~skb->csum);
 235         } else if (pseudohdr)
 236                 *sum = ~csum_fold(csum_partial((char *)diff, sizeof(diff),
 237                                 csum_unfold(*sum)));
 238 }
 239
 240 static struct sk_buff *set_nw_addr(struct sk_buff *skb,
 241                                    const struct odp_flow_key *key,
 242                                    const struct odp_action_nw_addr *a,
 243                                    gfp_t gfp)
 244 {
 245         if (key->dl_type != htons(ETH_P_IP))
 246                 return skb;
 247
 248         skb = make_writable(skb, 0, gfp);
 249         if (skb) {
 250                 struct iphdr *nh = ip_hdr(skb);
 251                 u32 *f = a->type == ODPAT_SET_NW_SRC ? &nh->saddr : &nh->daddr;
 252                 u32 old = *f;
 253                 u32 new = a->nw_addr;
 254
 255                 if (key->nw_proto == IPPROTO_TCP) {
 256                         struct tcphdr *th = tcp_hdr(skb);
 257                         update_csum(&th->check, skb, old, new, 1);
 258                 } else if (key->nw_proto == IPPROTO_UDP) {
 259                         struct udphdr *th = udp_hdr(skb);
 260                         update_csum(&th->check, skb, old, new, 1);
 261                 }
 262                 update_csum(&nh->check, skb, old, new, 0);
 263                 *f = new;
 264         }
 265         return skb;
 266 }
 267
 268 static struct sk_buff *set_nw_tos(struct sk_buff *skb,
 269                                    const struct odp_flow_key *key,
 270                                    const struct odp_action_nw_tos *a,
 271                                    gfp_t gfp)
 272 {
 273         if (key->dl_type != htons(ETH_P_IP))
 274                 return skb;
 275
 276         skb = make_writable(skb, 0, gfp);
 277         if (skb) {
 278                 struct iphdr *nh = ip_hdr(skb);
 279                 u8 *f = &nh->tos;
 280                 u8 old = *f;
 281                 u8 new;
 282
 283                 /* Set the DSCP bits and preserve the ECN bits. */
 284                 new = a->nw_tos | (nh->tos & INET_ECN_MASK);
 285                 update_csum(&nh->check, skb, htons((uint16_t)old),
 286                                 htons((uint16_t)new), 0);
 287                 *f = new;
 288         }
 289         return skb;
 290 }
 291
 292 static struct sk_buff *set_tp_port(struct sk_buff *skb,
 293                                    const struct odp_flow_key *key,
 294                                    const struct odp_action_tp_port *a, gfp_t gfp)
 295 {
 296         int check_ofs;
 297
 298         if (key->dl_type != htons(ETH_P_IP))
 299                 return skb;
 300
 301         if (key->nw_proto == IPPROTO_TCP)
 302                 check_ofs = offsetof(struct tcphdr, check);
 303         else if (key->nw_proto == IPPROTO_UDP)
 304                 check_ofs = offsetof(struct udphdr, check);
 305         else
 306                 return skb;
 307
 308         skb = make_writable(skb, 0, gfp);
 309         if (skb) {
 310                 struct udphdr *th = udp_hdr(skb);
 311                 u16 *f = a->type == ODPAT_SET_TP_SRC ? &th->source : &th->dest;
 312                 u16 old = *f;
 313                 u16 new = a->tp_port;
 314                 update_csum((u16*)(skb_transport_header(skb) + check_ofs),
 315                                 skb, old, new, 0);
 316                 *f = new;
 317         }
 318         return skb;
 319 }
 320
 321 static void do_output(struct datapath *dp, struct sk_buff *skb, int out_port)
 322 {
 323         struct dp_port *p;
 324
 325         if (!skb)
 326                 goto error;
 327
 328         p = rcu_dereference(dp->ports[out_port]);
 329         if (!p)
 330                 goto error;
 331
 332         vport_send(p->vport, skb);
 333         return;
 334
 335 error:
 336         kfree_skb(skb);
 337 }
 338
 339 /* Never consumes 'skb'.  Returns a port that 'skb' should be sent to, -1 if
 340  * none.  */
 341 static int output_group(struct datapath *dp, __u16 group,
 342                         struct sk_buff *skb, gfp_t gfp)
 343 {
 344         struct dp_port_group *g = rcu_dereference(dp->groups[group]);
 345         int prev_port = -1;
 346         int i;
 347
 348         if (!g)
 349                 return -1;
 350         for (i = 0; i < g->n_ports; i++) {
 351                 struct dp_port *p = rcu_dereference(dp->ports[g->ports[i]]);
 352                 if (!p || OVS_CB(skb)->dp_port == p)
 353                         continue;
 354                 if (prev_port != -1) {
 355                         struct sk_buff *clone = skb_clone(skb, gfp);
 356                         if (!clone)
 357                                 return -1;
 358                         do_output(dp, clone, prev_port);
 359                 }
 360                 prev_port = p->port_no;
 361         }
 362         return prev_port;
 363 }
 364
 365 static int output_control(struct datapath *dp, struct sk_buff *skb, u32 arg,
 366                           gfp_t gfp)
 367 {
 368         skb = skb_clone(skb, gfp);
 369         if (!skb)
 370                 return -ENOMEM;
 371         return dp_output_control(dp, skb, _ODPL_ACTION_NR, arg);
 372 }
 373
 374 /* Send a copy of this packet up to the sFlow agent, along with extra
 375  * information about what happened to it. */
 376 static void sflow_sample(struct datapath *dp, struct sk_buff *skb,
 377                          const union odp_action *a, int n_actions,
 378                          gfp_t gfp, struct dp_port *dp_port)
 379 {
 380         struct odp_sflow_sample_header *hdr;
 381         unsigned int actlen = n_actions * sizeof(union odp_action);
 382         unsigned int hdrlen = sizeof(struct odp_sflow_sample_header);
 383         struct sk_buff *nskb;
 384
 385         nskb = skb_copy_expand(skb, actlen + hdrlen, 0, gfp);
 386         if (!nskb)
 387                 return;
 388
 389         memcpy(__skb_push(nskb, actlen), a, actlen);
 390         hdr = (struct odp_sflow_sample_header*)__skb_push(nskb, hdrlen);
 391         hdr->n_actions = n_actions;
 392         hdr->sample_pool = atomic_read(&dp_port->sflow_pool);
 393         dp_output_control(dp, nskb, _ODPL_SFLOW_NR, 0);
 394 }
 395
 396 /* Execute a list of actions against 'skb'. */
 397 int execute_actions(struct datapath *dp, struct sk_buff *skb,
 398                     const struct odp_flow_key *key,
 399                     const union odp_action *a, int n_actions,
 400                     gfp_t gfp)
 401 {
 402         /* Every output action needs a separate clone of 'skb', but the common
 403          * case is just a single output action, so that doing a clone and
 404          * then freeing the original skbuff is wasteful.  So the following code
 405          * is slightly obscure just to avoid that. */
 406         int prev_port = -1;
 407         u32 priority = skb->priority;
 408         int err;
 409
 410         if (dp->sflow_probability) {
 411                 struct dp_port *p = OVS_CB(skb)->dp_port;
 412                 if (p) {
 413                         atomic_inc(&p->sflow_pool);
 414                         if (dp->sflow_probability == UINT_MAX ||
 415                             net_random() < dp->sflow_probability)
 416                                 sflow_sample(dp, skb, a, n_actions, gfp, p);
 417                 }
 418         }
 419
 420         OVS_CB(skb)->tun_id = 0;
 421
 422         for (; n_actions > 0; a++, n_actions--) {
 423                 if (prev_port != -1) {
 424                         do_output(dp, skb_clone(skb, gfp), prev_port);
 425                         prev_port = -1;
 426                 }
 427
 428                 switch (a->type) {
 429                 case ODPAT_OUTPUT:
 430                         prev_port = a->output.port;
 431                         break;
 432
 433                 case ODPAT_OUTPUT_GROUP:
 434                         prev_port = output_group(dp, a->output_group.group,
 435                                                  skb, gfp);
 436                         break;
 437
 438                 case ODPAT_CONTROLLER:
 439                         err = output_control(dp, skb, a->controller.arg, gfp);
 440                         if (err) {
 441                                 kfree_skb(skb);
 442                                 return err;
 443                         }
 444                         break;
 445
 446                 case ODPAT_SET_TUNNEL:
 447                         OVS_CB(skb)->tun_id = a->tunnel.tun_id;
 448                         break;
 449
 450                 case ODPAT_SET_VLAN_VID:
 451                 case ODPAT_SET_VLAN_PCP:
 452                         skb = modify_vlan_tci(dp, skb, key, a, n_actions, gfp);
 453                         if (IS_ERR(skb))
 454                                 return PTR_ERR(skb);
 455                         break;
 456
 457                 case ODPAT_STRIP_VLAN:
 458                         skb = strip_vlan(skb, gfp);
 459                         break;
 460
 461                 case ODPAT_SET_DL_SRC:
 462                 case ODPAT_SET_DL_DST:
 463                         skb = set_dl_addr(skb, &a->dl_addr, gfp);
 464                         break;
 465
 466                 case ODPAT_SET_NW_SRC:
 467                 case ODPAT_SET_NW_DST:
 468                         skb = set_nw_addr(skb, key, &a->nw_addr, gfp);
 469                         break;
 470
 471                 case ODPAT_SET_NW_TOS:
 472                         skb = set_nw_tos(skb, key, &a->nw_tos, gfp);
 473                         break;
 474
 475                 case ODPAT_SET_TP_SRC:
 476                 case ODPAT_SET_TP_DST:
 477                         skb = set_tp_port(skb, key, &a->tp_port, gfp);
 478                         break;
 479
 480                 case ODPAT_SET_PRIORITY:
 481                         skb->priority = a->priority.priority;
 482                         break;
 483
 484                 case ODPAT_POP_PRIORITY:
 485                         skb->priority = priority;
 486                         break;
 487                 }
 488                 if (!skb)
 489                         return -ENOMEM;
 490         }
 491         if (prev_port != -1)
 492                 do_output(dp, skb, prev_port);
 493         else
 494                 kfree_skb(skb);
 495         return 0;
 496 }