datapath/forward.c

   1 /*
   2  * Distributed under the terms of the GNU GPL version 2.
   3  * Copyright (c) 2007 The Board of Trustees of The Leland Stanford Junior Univer
   4 sity
   5  */
   6
   7 #include <linux/if_ether.h>
   8 #include <linux/if_vlan.h>
   9 #include <linux/in.h>
  10 #include <linux/ip.h>
  11 #include <linux/tcp.h>
  12 #include <linux/udp.h>
  13 #include <linux/in6.h>
  14 #include <asm/uaccess.h>
  15 #include <linux/types.h>
  16 #include <net/checksum.h>
  17 #include "forward.h"
  18 #include "datapath.h"
  19 #include "chain.h"
  20 #include "flow.h"
  21
  22 /* FIXME: do we need to use GFP_ATOMIC everywhere here? */
  23
  24 static void execute_actions(struct datapath *, struct sk_buff *,
  25                                 const struct sw_flow_key *,
  26                                 const struct ofp_action *, int n_actions);
  27 static int make_writable(struct sk_buff **);
  28
  29 static struct sk_buff *retrieve_skb(uint32_t id);
  30 static void discard_skb(uint32_t id);
  31
  32 /* 'skb' was received on 'in_port', a physical switch port between 0 and
  33  * OFPP_MAX.  Process it according to 'chain'. */
  34 void fwd_port_input(struct sw_chain *chain, struct sk_buff *skb, int in_port)
  35 {
  36         struct sw_flow_key key;
  37         struct sw_flow *flow;
  38
  39         flow_extract(skb, in_port, &key);
  40         flow = chain_lookup(chain, &key);
  41         if (likely(flow != NULL)) {
  42                 flow_used(flow, skb);
  43                 execute_actions(chain->dp, skb, &key,
  44                                 flow->actions, flow->n_actions);
  45         } else {
  46                 dp_output_control(chain->dp, skb, fwd_save_skb(skb),
  47                                 chain->dp->miss_send_len, OFPR_NO_MATCH);
  48         }
  49 }
  50
  51 static int do_output(struct datapath *dp, struct sk_buff *skb, size_t max_len,
  52                         int out_port)
  53 {
  54         if (!skb)
  55                 return -ENOMEM;
  56         return (likely(out_port != OFPP_CONTROLLER)
  57                 ? dp_output_port(dp, skb, out_port)
  58                 : dp_output_control(dp, skb, fwd_save_skb(skb),
  59                                          max_len, OFPR_ACTION));
  60 }
  61
  62 static void execute_actions(struct datapath *dp, struct sk_buff *skb,
  63                                 const struct sw_flow_key *key,
  64                                 const struct ofp_action *actions, int n_actions)
  65 {
  66         /* Every output action needs a separate clone of 'skb', but the common
  67          * case is just a single output action, so that doing a clone and
  68          * then freeing the original skbuff is wasteful.  So the following code
  69          * is slightly obscure just to avoid that. */
  70         int prev_port;
  71         size_t max_len=0;        /* Initialze to make compiler happy */
  72         uint16_t eth_proto;
  73         int i;
  74
  75         prev_port = -1;
  76         eth_proto = ntohs(key->dl_type);
  77
  78         for (i = 0; i < n_actions; i++) {
  79                 const struct ofp_action *a = &actions[i];
  80
  81                 if (prev_port != -1) {
  82                         do_output(dp, skb_clone(skb, GFP_ATOMIC),
  83                                   max_len, prev_port);
  84                         prev_port = -1;
  85                 }
  86
  87                 if (likely(a->type == ntohs(OFPAT_OUTPUT))) {
  88                         prev_port = ntohs(a->arg.output.port);
  89                         max_len = ntohs(a->arg.output.max_len);
  90                 } else {
  91                         if (!make_writable(&skb)) {
  92                                 printk("make_writable failed\n");
  93                                 break;
  94                         }
  95                         skb = execute_setter(skb, eth_proto, key, a);
  96                 }
  97         }
  98         if (prev_port != -1)
  99                 do_output(dp, skb, max_len, prev_port);
 100         else
 101                 kfree_skb(skb);
 102 }
 103
 104 /* Updates 'sum', which is a field in 'skb''s data, given that a 4-byte field
 105  * covered by the sum has been changed from 'from' to 'to'.  If set,
 106  * 'pseudohdr' indicates that the field is in the TCP or UDP pseudo-header.
 107  * Based on nf_proto_csum_replace4. */
 108 static void update_csum(__sum16 *sum, struct sk_buff *skb,
 109                         __be32 from, __be32 to, int pseudohdr)
 110 {
 111         __be32 diff[] = { ~from, to };
 112         if (skb->ip_summed != CHECKSUM_PARTIAL) {
 113                 *sum = csum_fold(csum_partial((char *)diff, sizeof(diff),
 114                                 ~csum_unfold(*sum)));
 115                 if (skb->ip_summed == CHECKSUM_COMPLETE && pseudohdr)
 116                         skb->csum = ~csum_partial((char *)diff, sizeof(diff),
 117                                                 ~skb->csum);
 118         } else if (pseudohdr)
 119                 *sum = ~csum_fold(csum_partial((char *)diff, sizeof(diff),
 120                                 csum_unfold(*sum)));
 121 }
 122
 123 static void modify_nh(struct sk_buff *skb, uint16_t eth_proto,
 124                         uint8_t nw_proto, const struct ofp_action *a)
 125 {
 126         if (eth_proto == ETH_P_IP) {
 127                 struct iphdr *nh = ip_hdr(skb);
 128                 uint32_t new, *field;
 129
 130                 new = a->arg.nw_addr;
 131
 132                 if (a->type == OFPAT_SET_NW_SRC)
 133                         field = &nh->saddr;
 134                 else
 135                         field = &nh->daddr;
 136
 137                 if (nw_proto == IPPROTO_TCP) {
 138                         struct tcphdr *th = tcp_hdr(skb);
 139                         update_csum(&th->check, skb, *field, new, 1);
 140                 } else if (nw_proto == IPPROTO_UDP) {
 141                         struct udphdr *th = udp_hdr(skb);
 142                         update_csum(&th->check, skb, *field, new, 1);
 143                 }
 144                 update_csum(&nh->check, skb, *field, new, 0);
 145                 *field = new;
 146         }
 147 }
 148
 149 static void modify_th(struct sk_buff *skb, uint16_t eth_proto,
 150                         uint8_t nw_proto, const struct ofp_action *a)
 151 {
 152         if (eth_proto == ETH_P_IP) {
 153                 uint16_t new, *field;
 154
 155                 new = a->arg.tp;
 156
 157                 if (nw_proto == IPPROTO_TCP) {
 158                         struct tcphdr *th = tcp_hdr(skb);
 159
 160                         if (a->type == OFPAT_SET_TP_SRC)
 161                                 field = &th->source;
 162                         else
 163                                 field = &th->dest;
 164
 165                         update_csum(&th->check, skb, *field, new, 1);
 166                         *field = new;
 167                 } else if (nw_proto == IPPROTO_UDP) {
 168                         struct udphdr *th = udp_hdr(skb);
 169
 170                         if (a->type == OFPAT_SET_TP_SRC)
 171                                 field = &th->source;
 172                         else
 173                                 field = &th->dest;
 174
 175                         update_csum(&th->check, skb, *field, new, 1);
 176                         *field = new;
 177                 }
 178         }
 179 }
 180
 181 static struct sk_buff *vlan_pull_tag(struct sk_buff *skb)
 182 {
 183         struct vlan_ethhdr *vh = vlan_eth_hdr(skb);
 184         struct ethhdr *eh;
 185
 186
 187         /* Verify we were given a vlan packet */
 188         if (vh->h_vlan_proto != __constant_htons(ETH_P_8021Q))
 189                 return skb;
 190
 191         memmove(skb->data + VLAN_HLEN, skb->data, 2 * VLAN_ETH_ALEN);
 192
 193         eh = (struct ethhdr *)skb_pull(skb, VLAN_HLEN);
 194
 195         skb->protocol = eh->h_proto;
 196         skb->mac_header += VLAN_HLEN;
 197
 198         return skb;
 199 }
 200
 201 static struct sk_buff *modify_vlan(struct sk_buff *skb,
 202                 const struct sw_flow_key *key, const struct ofp_action *a)
 203 {
 204         uint16_t new_id = a->arg.vlan_id;
 205
 206         if (new_id != OFP_VLAN_NONE) {
 207                 if (key->dl_vlan != __constant_htons(OFP_VLAN_NONE)) {
 208                         /* Modify vlan id, but maintain other TCI values */
 209                         struct vlan_ethhdr *vh = vlan_eth_hdr(skb);
 210                         vh->h_vlan_TCI = (vh->h_vlan_TCI
 211                                         & ~(__constant_htons(VLAN_VID_MASK))) | htons(new_id);
 212                 } else  {
 213                         /* Add vlan header */
 214                         skb = vlan_put_tag(skb, new_id);
 215                 }
 216         } else  {
 217                 /* Remove an existing vlan header if it exists */
 218                 vlan_pull_tag(skb);
 219         }
 220
 221         return skb;
 222 }
 223
 224 struct sk_buff *execute_setter(struct sk_buff *skb, uint16_t eth_proto,
 225                         const struct sw_flow_key *key, const struct ofp_action *a)
 226 {
 227         switch (a->type) {
 228         case OFPAT_SET_DL_VLAN:
 229                 skb = modify_vlan(skb, key, a);
 230                 break;
 231
 232         case OFPAT_SET_DL_SRC: {
 233                 struct ethhdr *eh = eth_hdr(skb);
 234                 memcpy(eh->h_source, a->arg.dl_addr, sizeof eh->h_source);
 235                 break;
 236         }
 237         case OFPAT_SET_DL_DST: {
 238                 struct ethhdr *eh = eth_hdr(skb);
 239                 memcpy(eh->h_dest, a->arg.dl_addr, sizeof eh->h_dest);
 240                 break;
 241         }
 242
 243         case OFPAT_SET_NW_SRC:
 244         case OFPAT_SET_NW_DST:
 245                 modify_nh(skb, eth_proto, key->nw_proto, a);
 246                 break;
 247
 248         case OFPAT_SET_TP_SRC:
 249         case OFPAT_SET_TP_DST:
 250                 modify_th(skb, eth_proto, key->nw_proto, a);
 251                 break;
 252
 253         default:
 254                 BUG();
 255         }
 256
 257         return skb;
 258 }
 259
 260 static int
 261 recv_control_hello(struct sw_chain *chain, const void *msg)
 262 {
 263         const struct ofp_control_hello *och = msg;
 264
 265         printk("control_hello(version=%d)\n", ntohl(och->version));
 266
 267         if (ntohs(och->miss_send_len) != OFP_MISS_SEND_LEN_UNCHANGED) {
 268                 chain->dp->miss_send_len = ntohs(och->miss_send_len);
 269         }
 270
 271         chain->dp->hello_flags = ntohs(och->flags);
 272
 273         dp_send_hello(chain->dp);
 274
 275         return 0;
 276 }
 277
 278 static int
 279 recv_packet_out(struct sw_chain *chain, const void *msg)
 280 {
 281         const struct ofp_packet_out *opo = msg;
 282         struct sk_buff *skb;
 283         struct vlan_ethhdr *mac;
 284         int nh_ofs;
 285
 286         if (ntohl(opo->buffer_id) == (uint32_t) -1) {
 287                 int data_len = ntohs(opo->header.length) - sizeof *opo;
 288
 289                 /* FIXME: there is likely a way to reuse the data in msg. */
 290                 skb = alloc_skb(data_len, GFP_ATOMIC);
 291                 if (!skb)
 292                         return -ENOMEM;
 293
 294                 /* FIXME?  We don't reserve NET_IP_ALIGN or NET_SKB_PAD since
 295                  * we're just transmitting this raw without examining anything
 296                  * at those layers. */
 297                 memcpy(skb_put(skb, data_len), opo->u.data, data_len);
 298                 dp_set_origin(chain->dp, ntohs(opo->in_port), skb);
 299
 300                 skb_set_mac_header(skb, 0);
 301                 mac = vlan_eth_hdr(skb);
 302                 if (likely(mac->h_vlan_proto != htons(ETH_P_8021Q)))
 303                         nh_ofs = sizeof(struct ethhdr);
 304                 else
 305                         nh_ofs = sizeof(struct vlan_ethhdr);
 306                 skb_set_network_header(skb, nh_ofs);
 307
 308                 dp_output_port(chain->dp, skb, ntohs(opo->out_port));
 309         } else {
 310                 struct sw_flow_key key;
 311                 int n_acts;
 312
 313                 skb = retrieve_skb(ntohl(opo->buffer_id));
 314                 if (!skb)
 315                         return -ESRCH;
 316                 dp_set_origin(chain->dp, ntohs(opo->in_port), skb);
 317
 318                 n_acts = (ntohs(opo->header.length) - sizeof *opo)
 319                                 / sizeof *opo->u.actions;
 320                 flow_extract(skb, ntohs(opo->in_port), &key);
 321                 execute_actions(chain->dp, skb, &key, opo->u.actions, n_acts);
 322         }
 323         return 0;
 324 }
 325
 326 static int
 327 recv_port_mod(struct sw_chain *chain, const void *msg)
 328 {
 329         const struct ofp_port_mod *opm = msg;
 330
 331         dp_update_port_flags(chain->dp, &opm->desc);
 332
 333         return 0;
 334 }
 335
 336 static int
 337 add_flow(struct sw_chain *chain, const struct ofp_flow_mod *ofm)
 338 {
 339         int error = -ENOMEM;
 340         int n_acts;
 341         struct sw_flow *flow;
 342
 343
 344         /* Check number of actions. */
 345         n_acts = (ntohs(ofm->header.length) - sizeof *ofm) / sizeof *ofm->actions;
 346         if (n_acts > MAX_ACTIONS) {
 347                 error = -E2BIG;
 348                 goto error;
 349         }
 350
 351         /* Allocate memory. */
 352         flow = flow_alloc(n_acts, GFP_ATOMIC);
 353         if (flow == NULL)
 354                 goto error;
 355
 356         /* Fill out flow. */
 357         flow_extract_match(&flow->key, &ofm->match);
 358         flow->group_id = ntohl(ofm->group_id);
 359         flow->max_idle = ntohs(ofm->max_idle);
 360         flow->timeout = jiffies + flow->max_idle * HZ;
 361         flow->n_actions = n_acts;
 362         flow->init_time = jiffies;
 363         flow->byte_count = 0;
 364         flow->packet_count = 0;
 365         atomic_set(&flow->deleted, 0);
 366         spin_lock_init(&flow->lock);
 367         memcpy(flow->actions, ofm->actions, n_acts * sizeof *flow->actions);
 368
 369         /* Act. */
 370         error = chain_insert(chain, flow);
 371         if (error)
 372                 goto error_free_flow;
 373         error = 0;
 374         if (ntohl(ofm->buffer_id) != (uint32_t) -1) {
 375                 struct sk_buff *skb = retrieve_skb(ntohl(ofm->buffer_id));
 376                 if (skb) {
 377                         struct sw_flow_key key;
 378                         flow_used(flow, skb);
 379                         flow_extract(skb, ntohs(ofm->match.in_port), &key);
 380                         execute_actions(chain->dp, skb, &key,
 381                                         ofm->actions, n_acts);
 382                 }
 383                 else
 384                         error = -ESRCH;
 385         }
 386         return error;
 387
 388 error_free_flow:
 389         flow_free(flow);
 390 error:
 391         if (ntohl(ofm->buffer_id) != (uint32_t) -1)
 392                 discard_skb(ntohl(ofm->buffer_id));
 393         return error;
 394 }
 395
 396 static int
 397 recv_flow(struct sw_chain *chain, const void *msg)
 398 {
 399         const struct ofp_flow_mod *ofm = msg;
 400         uint16_t command = ntohs(ofm->command);
 401
 402         if (command == OFPFC_ADD) {
 403                 return add_flow(chain, ofm);
 404         }  else if (command == OFPFC_DELETE) {
 405                 struct sw_flow_key key;
 406                 flow_extract_match(&key, &ofm->match);
 407                 return chain_delete(chain, &key, 0) ? 0 : -ESRCH;
 408         } else if (command == OFPFC_DELETE_STRICT) {
 409                 struct sw_flow_key key;
 410                 flow_extract_match(&key, &ofm->match);
 411                 return chain_delete(chain, &key, 1) ? 0 : -ESRCH;
 412         } else {
 413                 return -ENOTSUPP;
 414         }
 415 }
 416
 417 /* 'msg', which is 'length' bytes long, was received from the control path.
 418  * Apply it to 'chain'. */
 419 int
 420 fwd_control_input(struct sw_chain *chain, const void *msg, size_t length)
 421 {
 422
 423         struct openflow_packet {
 424                 size_t min_size;
 425                 int (*handler)(struct sw_chain *, const void *);
 426         };
 427
 428         static const struct openflow_packet packets[] = {
 429                 [OFPT_CONTROL_HELLO] = {
 430                         sizeof (struct ofp_control_hello),
 431                         recv_control_hello,
 432                 },
 433                 [OFPT_PACKET_OUT] = {
 434                         sizeof (struct ofp_packet_out),
 435                         recv_packet_out,
 436                 },
 437                 [OFPT_FLOW_MOD] = {
 438                         sizeof (struct ofp_flow_mod),
 439                         recv_flow,
 440                 },
 441                 [OFPT_PORT_MOD] = {
 442                         sizeof (struct ofp_port_mod),
 443                         recv_port_mod,
 444                 },
 445         };
 446
 447         const struct openflow_packet *pkt;
 448         struct ofp_header *oh;
 449
 450         if (length < sizeof(struct ofp_header))
 451                 return -EINVAL;
 452
 453         oh = (struct ofp_header *) msg;
 454         if (oh->version != 1 || oh->type >= ARRAY_SIZE(packets)
 455                 || ntohs(oh->length) > length)
 456                 return -EINVAL;
 457
 458         pkt = &packets[oh->type];
 459         if (!pkt->handler)
 460                 return -ENOSYS;
 461         if (length < pkt->min_size)
 462                 return -EFAULT;
 463
 464         return pkt->handler(chain, msg);
 465 }
 466
 467 /* Packet buffering. */
 468
 469 #define OVERWRITE_SECS  1
 470 #define OVERWRITE_JIFFIES (OVERWRITE_SECS * HZ)
 471
 472 struct packet_buffer {
 473         struct sk_buff *skb;
 474         uint32_t cookie;
 475         unsigned long exp_jiffies;
 476 };
 477
 478 static struct packet_buffer buffers[N_PKT_BUFFERS];
 479 static unsigned int buffer_idx;
 480 static DEFINE_SPINLOCK(buffer_lock);
 481
 482 uint32_t fwd_save_skb(struct sk_buff *skb)
 483 {
 484         struct packet_buffer *p;
 485         unsigned long int flags;
 486         uint32_t id;
 487
 488         spin_lock_irqsave(&buffer_lock, flags);
 489         buffer_idx = (buffer_idx + 1) & PKT_BUFFER_MASK;
 490         p = &buffers[buffer_idx];
 491         if (p->skb) {
 492                 /* Don't buffer packet if existing entry is less than
 493                  * OVERWRITE_SECS old. */
 494                 if (time_before(jiffies, p->exp_jiffies)) {
 495                         spin_unlock_irqrestore(&buffer_lock, flags);
 496                         return -1;
 497                 } else
 498                         kfree_skb(p->skb);
 499         }
 500         /* Don't use maximum cookie value since the all-bits-1 id is
 501          * special. */
 502         if (++p->cookie >= (1u << PKT_COOKIE_BITS) - 1)
 503                 p->cookie = 0;
 504         skb_get(skb);
 505         p->skb = skb;
 506         p->exp_jiffies = jiffies + OVERWRITE_JIFFIES;
 507         id = buffer_idx | (p->cookie << PKT_BUFFER_BITS);
 508         spin_unlock_irqrestore(&buffer_lock, flags);
 509
 510         return id;
 511 }
 512
 513 static struct sk_buff *retrieve_skb(uint32_t id)
 514 {
 515         unsigned long int flags;
 516         struct sk_buff *skb = NULL;
 517         struct packet_buffer *p;
 518
 519         spin_lock_irqsave(&buffer_lock, flags);
 520         p = &buffers[id & PKT_BUFFER_MASK];
 521         if (p->cookie == id >> PKT_BUFFER_BITS) {
 522                 skb = p->skb;
 523                 p->skb = NULL;
 524         } else {
 525                 printk("cookie mismatch: %x != %x\n",
 526                                 id >> PKT_BUFFER_BITS, p->cookie);
 527         }
 528         spin_unlock_irqrestore(&buffer_lock, flags);
 529
 530         return skb;
 531 }
 532
 533 static void discard_skb(uint32_t id)
 534 {
 535         unsigned long int flags;
 536         struct packet_buffer *p;
 537
 538         spin_lock_irqsave(&buffer_lock, flags);
 539         p = &buffers[id & PKT_BUFFER_MASK];
 540         if (p->cookie == id >> PKT_BUFFER_BITS) {
 541                 kfree_skb(p->skb);
 542                 p->skb = NULL;
 543         }
 544         spin_unlock_irqrestore(&buffer_lock, flags);
 545 }
 546
 547 void fwd_exit(void)
 548 {
 549         int i;
 550
 551         for (i = 0; i < N_PKT_BUFFERS; i++)
 552                 kfree_skb(buffers[i].skb);
 553 }
 554
 555 /* Utility functions. */
 556
 557 /* Makes '*pskb' writable, possibly copying it and setting '*pskb' to point to
 558  * the copy.
 559  * Returns 1 if successful, 0 on failure. */
 560 static int
 561 make_writable(struct sk_buff **pskb)
 562 {
 563         /* Based on skb_make_writable() in net/netfilter/core.c. */
 564         struct sk_buff *nskb;
 565
 566         /* Not exclusive use of packet?  Must copy. */
 567         if (skb_shared(*pskb) || skb_cloned(*pskb))
 568                 goto copy_skb;
 569
 570         return pskb_may_pull(*pskb, 64); /* FIXME? */
 571
 572 copy_skb:
 573         nskb = skb_copy(*pskb, GFP_ATOMIC);
 574         if (!nskb)
 575                 return 0;
 576         BUG_ON(skb_is_nonlinear(nskb));
 577
 578         /* Rest of kernel will get very unhappy if we pass it a
 579            suddenly-orphaned skbuff */
 580         if ((*pskb)->sk)
 581                 skb_set_owner_w(nskb, (*pskb)->sk);
 582         kfree_skb(*pskb);
 583         *pskb = nskb;
 584         return 1;
 585 }