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_shared(skb) || 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 void set_tunnel(struct sk_buff *skb, struct odp_flow_key *key,
  50                        __be32 tun_id)
  51 {
  52         OVS_CB(skb)->tun_id = key->tun_id = tun_id;
  53 }
  54
  55 static struct sk_buff *vlan_pull_tag(struct sk_buff *skb)
  56 {
  57         struct vlan_ethhdr *vh = vlan_eth_hdr(skb);
  58         struct ethhdr *eh;
  59
  60         /* Verify we were given a vlan packet */
  61         if (vh->h_vlan_proto != htons(ETH_P_8021Q))
  62                 return skb;
  63
  64         if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE)
  65                 skb->csum = csum_sub(skb->csum, csum_partial(skb->data
  66                                         + ETH_HLEN, VLAN_HLEN, 0));
  67
  68         memmove(skb->data + VLAN_HLEN, skb->data, 2 * VLAN_ETH_ALEN);
  69
  70         eh = (struct ethhdr *)skb_pull(skb, VLAN_HLEN);
  71
  72         skb->protocol = eh->h_proto;
  73         skb->mac_header += VLAN_HLEN;
  74
  75         return skb;
  76 }
  77
  78
  79 static struct sk_buff *modify_vlan_tci(struct datapath *dp, struct sk_buff *skb,
  80                                        struct odp_flow_key *key, const union odp_action *a,
  81                                        int n_actions, gfp_t gfp)
  82 {
  83         u16 tci, mask;
  84
  85         if (a->type == ODPAT_SET_VLAN_VID) {
  86                 tci = ntohs(a->vlan_vid.vlan_vid);
  87                 mask = VLAN_VID_MASK;
  88                 key->dl_vlan = a->vlan_vid.vlan_vid;
  89         } else {
  90                 tci = a->vlan_pcp.vlan_pcp << VLAN_PCP_SHIFT;
  91                 mask = VLAN_PCP_MASK;
  92                 key->dl_vlan_pcp = a->vlan_pcp.vlan_pcp;
  93         }
  94
  95         skb = make_writable(skb, VLAN_HLEN, gfp);
  96         if (!skb)
  97                 return ERR_PTR(-ENOMEM);
  98
  99         if (skb->protocol == htons(ETH_P_8021Q)) {
 100                 /* Modify vlan id, but maintain other TCI values */
 101                 struct vlan_ethhdr *vh = vlan_eth_hdr(skb);
 102                 __be16 old_tci = vh->h_vlan_TCI;
 103
 104                 vh->h_vlan_TCI = htons((ntohs(vh->h_vlan_TCI) & ~mask) | tci);
 105
 106                 if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE) {
 107                         __be16 diff[] = { ~old_tci, vh->h_vlan_TCI };
 108
 109                         skb->csum = ~csum_partial((char *)diff, sizeof(diff),
 110                                                 ~skb->csum);
 111                 }
 112         } else {
 113                 int err;
 114
 115                 /* Add vlan header */
 116
 117                 /* Set up checksumming pointers for checksum-deferred packets
 118                  * on Xen.  Otherwise, dev_queue_xmit() will try to do this
 119                  * when we send the packet out on the wire, and it will fail at
 120                  * that point because skb_checksum_setup() will not look inside
 121                  * an 802.1Q header. */
 122                 err = vswitch_skb_checksum_setup(skb);
 123                 if (unlikely(err)) {
 124                         kfree_skb(skb);
 125                         return ERR_PTR(err);
 126                 }
 127
 128                 /* GSO is not implemented for packets with an 802.1Q header, so
 129                  * we have to do segmentation before we add that header.
 130                  *
 131                  * GSO does work with hardware-accelerated VLAN tagging, but we
 132                  * can't use hardware-accelerated VLAN tagging since it
 133                  * requires the device to have a VLAN group configured (with
 134                  * e.g. vconfig(8)) and we don't do that.
 135                  *
 136                  * Having to do this here may be a performance loss, since we
 137                  * can't take advantage of TSO hardware support, although it
 138                  * does not make a measurable network performance difference
 139                  * for 1G Ethernet.  Fixing that would require patching the
 140                  * kernel (either to add GSO support to the VLAN protocol or to
 141                  * support hardware-accelerated VLAN tagging without VLAN
 142                  * groups configured). */
 143                 if (skb_is_gso(skb)) {
 144                         struct sk_buff *segs;
 145
 146                         segs = skb_gso_segment(skb, 0);
 147                         kfree_skb(skb);
 148                         if (unlikely(IS_ERR(segs)))
 149                                 return ERR_CAST(segs);
 150
 151                         do {
 152                                 struct sk_buff *nskb = segs->next;
 153                                 int err;
 154
 155                                 segs->next = NULL;
 156
 157                                 /* GSO can change the checksum type so update.*/
 158                                 compute_ip_summed(segs, true);
 159
 160                                 segs = __vlan_put_tag(segs, tci);
 161                                 err = -ENOMEM;
 162                                 if (segs) {
 163                                         struct odp_flow_key segkey = *key;
 164                                         err = execute_actions(dp, segs,
 165                                                               &segkey, a + 1,
 166                                                               n_actions - 1,
 167                                                               gfp);
 168                                 }
 169
 170                                 if (unlikely(err)) {
 171                                         while ((segs = nskb)) {
 172                                                 nskb = segs->next;
 173                                                 segs->next = NULL;
 174                                                 kfree_skb(segs);
 175                                         }
 176                                         return ERR_PTR(err);
 177                                 }
 178
 179                                 segs = nskb;
 180                         } while (segs->next);
 181
 182                         skb = segs;
 183                         compute_ip_summed(skb, true);
 184                 }
 185
 186                 /* The hardware-accelerated version of vlan_put_tag() works
 187                  * only for a device that has a VLAN group configured (with
 188                  * e.g. vconfig(8)), so call the software-only version
 189                  * __vlan_put_tag() directly instead.
 190                  */
 191                 skb = __vlan_put_tag(skb, tci);
 192                 if (!skb)
 193                         return ERR_PTR(-ENOMEM);
 194
 195                 /* GSO doesn't fix up the hardware computed checksum so this
 196                  * will only be hit in the non-GSO case. */
 197                 if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE)
 198                         skb->csum = csum_add(skb->csum, csum_partial(skb->data
 199                                                 + ETH_HLEN, VLAN_HLEN, 0));
 200         }
 201
 202         return skb;
 203 }
 204
 205 static struct sk_buff *strip_vlan(struct sk_buff *skb,
 206                                   struct odp_flow_key *key, gfp_t gfp)
 207 {
 208         skb = make_writable(skb, 0, gfp);
 209         if (skb) {
 210                 vlan_pull_tag(skb);
 211                 key->dl_vlan = htons(ODP_VLAN_NONE);
 212         }
 213         return skb;
 214 }
 215
 216 static struct sk_buff *set_dl_addr(struct sk_buff *skb,
 217                                    struct odp_flow_key *key,
 218                                    const struct odp_action_dl_addr *a,
 219                                    gfp_t gfp)
 220 {
 221         skb = make_writable(skb, 0, gfp);
 222         if (skb) {
 223                 struct ethhdr *eh = eth_hdr(skb);
 224                 if (a->type == ODPAT_SET_DL_SRC) {
 225                         memcpy(eh->h_source, a->dl_addr, ETH_ALEN);
 226                         memcpy(key->dl_src, a->dl_addr, ETH_ALEN);
 227                 } else {
 228                         memcpy(eh->h_dest, a->dl_addr, ETH_ALEN);
 229                         memcpy(key->dl_dst, a->dl_addr, ETH_ALEN);
 230                 }
 231         }
 232         return skb;
 233 }
 234
 235 /* Updates 'sum', which is a field in 'skb''s data, given that a 4-byte field
 236  * covered by the sum has been changed from 'from' to 'to'.  If set,
 237  * 'pseudohdr' indicates that the field is in the TCP or UDP pseudo-header.
 238  * Based on nf_proto_csum_replace4. */
 239 static void update_csum(__sum16 *sum, struct sk_buff *skb,
 240                         __be32 from, __be32 to, int pseudohdr)
 241 {
 242         __be32 diff[] = { ~from, to };
 243
 244         if (OVS_CB(skb)->ip_summed != OVS_CSUM_PARTIAL) {
 245                 *sum = csum_fold(csum_partial((char *)diff, sizeof(diff),
 246                                 ~csum_unfold(*sum)));
 247                 if (OVS_CB(skb)->ip_summed == OVS_CSUM_COMPLETE && pseudohdr)
 248                         skb->csum = ~csum_partial((char *)diff, sizeof(diff),
 249                                                 ~skb->csum);
 250         } else if (pseudohdr)
 251                 *sum = ~csum_fold(csum_partial((char *)diff, sizeof(diff),
 252                                 csum_unfold(*sum)));
 253 }
 254
 255 static struct sk_buff *set_nw_addr(struct sk_buff *skb,
 256                                    struct odp_flow_key *key,
 257                                    const struct odp_action_nw_addr *a,
 258                                    gfp_t gfp)
 259 {
 260         if (key->dl_type != htons(ETH_P_IP))
 261                 return skb;
 262
 263         skb = make_writable(skb, 0, gfp);
 264         if (skb) {
 265                 struct iphdr *nh = ip_hdr(skb);
 266                 u32 *f = a->type == ODPAT_SET_NW_SRC ? &nh->saddr : &nh->daddr;
 267                 u32 old = *f;
 268                 u32 new = a->nw_addr;
 269
 270                 if (key->nw_proto == IPPROTO_TCP) {
 271                         struct tcphdr *th = tcp_hdr(skb);
 272                         update_csum(&th->check, skb, old, new, 1);
 273                 } else if (key->nw_proto == IPPROTO_UDP) {
 274                         struct udphdr *th = udp_hdr(skb);
 275                         update_csum(&th->check, skb, old, new, 1);
 276                 }
 277                 update_csum(&nh->check, skb, old, new, 0);
 278                 *f = new;
 279
 280                 if (a->type == ODPAT_SET_NW_SRC)
 281                         key->nw_src = a->nw_addr;
 282                 else
 283                         key->nw_dst = a->nw_addr;
 284         }
 285         return skb;
 286 }
 287
 288 static struct sk_buff *set_nw_tos(struct sk_buff *skb,
 289                                    struct odp_flow_key *key,
 290                                    const struct odp_action_nw_tos *a,
 291                                    gfp_t gfp)
 292 {
 293         if (key->dl_type != htons(ETH_P_IP))
 294                 return skb;
 295
 296         skb = make_writable(skb, 0, gfp);
 297         if (skb) {
 298                 struct iphdr *nh = ip_hdr(skb);
 299                 u8 *f = &nh->tos;
 300                 u8 old = *f;
 301                 u8 new;
 302
 303                 /* Set the DSCP bits and preserve the ECN bits. */
 304                 new = a->nw_tos | (nh->tos & INET_ECN_MASK);
 305                 update_csum(&nh->check, skb, htons((uint16_t)old),
 306                                 htons((uint16_t)new), 0);
 307                 *f = new;
 308                 key->nw_tos = a->nw_tos;
 309         }
 310         return skb;
 311 }
 312
 313 static struct sk_buff *set_tp_port(struct sk_buff *skb, struct odp_flow_key *key,
 314                                    const struct odp_action_tp_port *a, gfp_t gfp)
 315 {
 316         int check_ofs;
 317
 318         if (key->dl_type != htons(ETH_P_IP))
 319                 return skb;
 320
 321         if (key->nw_proto == IPPROTO_TCP)
 322                 check_ofs = offsetof(struct tcphdr, check);
 323         else if (key->nw_proto == IPPROTO_UDP)
 324                 check_ofs = offsetof(struct udphdr, check);
 325         else
 326                 return skb;
 327
 328         skb = make_writable(skb, 0, gfp);
 329         if (skb) {
 330                 struct udphdr *th = udp_hdr(skb);
 331                 u16 *f = a->type == ODPAT_SET_TP_SRC ? &th->source : &th->dest;
 332                 u16 old = *f;
 333                 u16 new = a->tp_port;
 334                 update_csum((u16*)(skb_transport_header(skb) + check_ofs),
 335                                 skb, old, new, 0);
 336                 *f = new;
 337                 if (a->type == ODPAT_SET_TP_SRC)
 338                         key->tp_src = a->tp_port;
 339                 else
 340                         key->tp_dst = a->tp_port;
 341         }
 342         return skb;
 343 }
 344
 345 static void do_output(struct datapath *dp, struct sk_buff *skb, int out_port)
 346 {
 347         struct dp_port *p;
 348
 349         if (!skb)
 350                 goto error;
 351
 352         p = rcu_dereference(dp->ports[out_port]);
 353         if (!p)
 354                 goto error;
 355
 356         vport_send(p->vport, skb);
 357         return;
 358
 359 error:
 360         kfree_skb(skb);
 361 }
 362
 363 /* Never consumes 'skb'.  Returns a port that 'skb' should be sent to, -1 if
 364  * none.  */
 365 static int output_group(struct datapath *dp, __u16 group,
 366                         struct sk_buff *skb, gfp_t gfp)
 367 {
 368         struct dp_port_group *g = rcu_dereference(dp->groups[group]);
 369         int prev_port = -1;
 370         int i;
 371
 372         if (!g)
 373                 return -1;
 374         for (i = 0; i < g->n_ports; i++) {
 375                 struct dp_port *p = rcu_dereference(dp->ports[g->ports[i]]);
 376                 if (!p || OVS_CB(skb)->dp_port == p)
 377                         continue;
 378                 if (prev_port != -1) {
 379                         struct sk_buff *clone = skb_clone(skb, gfp);
 380                         if (!clone)
 381                                 return -1;
 382                         do_output(dp, clone, prev_port);
 383                 }
 384                 prev_port = p->port_no;
 385         }
 386         return prev_port;
 387 }
 388
 389 static int output_control(struct datapath *dp, struct sk_buff *skb, u32 arg,
 390                           gfp_t gfp)
 391 {
 392         skb = skb_clone(skb, gfp);
 393         if (!skb)
 394                 return -ENOMEM;
 395         return dp_output_control(dp, skb, _ODPL_ACTION_NR, arg);
 396 }
 397
 398 /* Send a copy of this packet up to the sFlow agent, along with extra
 399  * information about what happened to it. */
 400 static void sflow_sample(struct datapath *dp, struct sk_buff *skb,
 401                          const union odp_action *a, int n_actions,
 402                          gfp_t gfp, struct dp_port *dp_port)
 403 {
 404         struct odp_sflow_sample_header *hdr;
 405         unsigned int actlen = n_actions * sizeof(union odp_action);
 406         unsigned int hdrlen = sizeof(struct odp_sflow_sample_header);
 407         struct sk_buff *nskb;
 408
 409         nskb = skb_copy_expand(skb, actlen + hdrlen, 0, gfp);
 410         if (!nskb)
 411                 return;
 412
 413         memcpy(__skb_push(nskb, actlen), a, actlen);
 414         hdr = (struct odp_sflow_sample_header*)__skb_push(nskb, hdrlen);
 415         hdr->n_actions = n_actions;
 416         hdr->sample_pool = atomic_read(&dp_port->sflow_pool);
 417         dp_output_control(dp, nskb, _ODPL_SFLOW_NR, 0);
 418 }
 419
 420 /* Execute a list of actions against 'skb'. */
 421 int execute_actions(struct datapath *dp, struct sk_buff *skb,
 422                     struct odp_flow_key *key,
 423                     const union odp_action *a, int n_actions,
 424                     gfp_t gfp)
 425 {
 426         /* Every output action needs a separate clone of 'skb', but the common
 427          * case is just a single output action, so that doing a clone and
 428          * then freeing the original skbuff is wasteful.  So the following code
 429          * is slightly obscure just to avoid that. */
 430         int prev_port = -1;
 431         u32 priority = skb->priority;
 432         int err;
 433
 434         if (dp->sflow_probability) {
 435                 struct dp_port *p = OVS_CB(skb)->dp_port;
 436                 if (p) {
 437                         atomic_inc(&p->sflow_pool);
 438                         if (dp->sflow_probability == UINT_MAX ||
 439                             net_random() < dp->sflow_probability)
 440                                 sflow_sample(dp, skb, a, n_actions, gfp, p);
 441                 }
 442         }
 443
 444         OVS_CB(skb)->tun_id = 0;
 445
 446         for (; n_actions > 0; a++, n_actions--) {
 447                 WARN_ON_ONCE(skb_shared(skb));
 448                 if (prev_port != -1) {
 449                         do_output(dp, skb_clone(skb, gfp), prev_port);
 450                         prev_port = -1;
 451                 }
 452
 453                 switch (a->type) {
 454                 case ODPAT_OUTPUT:
 455                         prev_port = a->output.port;
 456                         break;
 457
 458                 case ODPAT_OUTPUT_GROUP:
 459                         prev_port = output_group(dp, a->output_group.group,
 460                                                  skb, gfp);
 461                         break;
 462
 463                 case ODPAT_CONTROLLER:
 464                         err = output_control(dp, skb, a->controller.arg, gfp);
 465                         if (err) {
 466                                 kfree_skb(skb);
 467                                 return err;
 468                         }
 469                         break;
 470
 471                 case ODPAT_SET_TUNNEL:
 472                         set_tunnel(skb, key, a->tunnel.tun_id);
 473                         break;
 474
 475                 case ODPAT_SET_VLAN_VID:
 476                 case ODPAT_SET_VLAN_PCP:
 477                         skb = modify_vlan_tci(dp, skb, key, a, n_actions, gfp);
 478                         if (IS_ERR(skb))
 479                                 return PTR_ERR(skb);
 480                         break;
 481
 482                 case ODPAT_STRIP_VLAN:
 483                         skb = strip_vlan(skb, key, gfp);
 484                         break;
 485
 486                 case ODPAT_SET_DL_SRC:
 487                 case ODPAT_SET_DL_DST:
 488                         skb = set_dl_addr(skb, key, &a->dl_addr, gfp);
 489                         break;
 490
 491                 case ODPAT_SET_NW_SRC:
 492                 case ODPAT_SET_NW_DST:
 493                         skb = set_nw_addr(skb, key, &a->nw_addr, gfp);
 494                         break;
 495
 496                 case ODPAT_SET_NW_TOS:
 497                         skb = set_nw_tos(skb, key, &a->nw_tos, gfp);
 498                         break;
 499
 500                 case ODPAT_SET_TP_SRC:
 501                 case ODPAT_SET_TP_DST:
 502                         skb = set_tp_port(skb, key, &a->tp_port, gfp);
 503                         break;
 504
 505                 case ODPAT_SET_PRIORITY:
 506                         skb->priority = a->priority.priority;
 507                         break;
 508
 509                 case ODPAT_POP_PRIORITY:
 510                         skb->priority = priority;
 511                         break;
 512                 }
 513                 if (!skb)
 514                         return -ENOMEM;
 515         }
 516         if (prev_port != -1)
 517                 do_output(dp, skb, prev_port);
 518         else
 519                 kfree_skb(skb);
 520         return 0;
 521 }