datapath/forward.c

   1 /*
   2  * Distributed under the terms of the GNU GPL version 2.
   3  * Copyright (c) 2007, 2008 The Board of Trustees of The Leland
   4  * Stanford Junior University
   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 int make_writable(struct sk_buff **);
  25
  26 static struct sk_buff *retrieve_skb(uint32_t id);
  27 static void discard_skb(uint32_t id);
  28
  29 /* 'skb' was received on 'in_port', a physical switch port between 0 and
  30  * OFPP_MAX.  Process it according to 'chain'.  Returns 0 if successful, in
  31  * which case 'skb' is destroyed, or -ESRCH if there is no matching flow, in
  32  * which case 'skb' still belongs to the caller. */
  33 int run_flow_through_tables(struct sw_chain *chain, struct sk_buff *skb,
  34                             int in_port)
  35 {
  36         struct sw_flow_key key;
  37         struct sw_flow *flow;
  38
  39         if (flow_extract(skb, in_port, &key)
  40             && (chain->dp->flags & OFPC_FRAG_MASK) == OFPC_FRAG_DROP) {
  41                 /* Drop fragment. */
  42                 kfree_skb(skb);
  43                 return 0;
  44         }
  45
  46         flow = chain_lookup(chain, &key);
  47         if (likely(flow != NULL)) {
  48                 flow_used(flow, skb);
  49                 execute_actions(chain->dp, skb, &key,
  50                                 flow->actions, flow->n_actions);
  51                 return 0;
  52         } else {
  53                 return -ESRCH;
  54         }
  55 }
  56
  57 /* 'skb' was received on 'in_port', a physical switch port between 0 and
  58  * OFPP_MAX.  Process it according to 'chain', sending it up to the controller
  59  * if no flow matches.  Takes ownership of 'skb'. */
  60 void fwd_port_input(struct sw_chain *chain, struct sk_buff *skb, int in_port)
  61 {
  62         if (run_flow_through_tables(chain, skb, in_port))
  63                 dp_output_control(chain->dp, skb, fwd_save_skb(skb),
  64                                   chain->dp->miss_send_len,
  65                                   OFPR_NO_MATCH);
  66 }
  67
  68 static int do_output(struct datapath *dp, struct sk_buff *skb, size_t max_len,
  69                         int out_port)
  70 {
  71         if (!skb)
  72                 return -ENOMEM;
  73         return (likely(out_port != OFPP_CONTROLLER)
  74                 ? dp_output_port(dp, skb, out_port)
  75                 : dp_output_control(dp, skb, fwd_save_skb(skb),
  76                                          max_len, OFPR_ACTION));
  77 }
  78
  79 void execute_actions(struct datapath *dp, struct sk_buff *skb,
  80                                 const struct sw_flow_key *key,
  81                                 const struct ofp_action *actions, int n_actions)
  82 {
  83         /* Every output action needs a separate clone of 'skb', but the common
  84          * case is just a single output action, so that doing a clone and
  85          * then freeing the original skbuff is wasteful.  So the following code
  86          * is slightly obscure just to avoid that. */
  87         int prev_port;
  88         size_t max_len=0;        /* Initialze to make compiler happy */
  89         uint16_t eth_proto;
  90         int i;
  91
  92         prev_port = -1;
  93         eth_proto = ntohs(key->dl_type);
  94
  95         for (i = 0; i < n_actions; i++) {
  96                 const struct ofp_action *a = &actions[i];
  97
  98                 if (prev_port != -1) {
  99                         do_output(dp, skb_clone(skb, GFP_ATOMIC),
 100                                   max_len, prev_port);
 101                         prev_port = -1;
 102                 }
 103
 104                 if (likely(a->type == htons(OFPAT_OUTPUT))) {
 105                         prev_port = ntohs(a->arg.output.port);
 106                         max_len = ntohs(a->arg.output.max_len);
 107                 } else {
 108                         if (!make_writable(&skb)) {
 109                                 if (net_ratelimit())
 110                                     printk("make_writable failed\n");
 111                                 break;
 112                         }
 113                         skb = execute_setter(skb, eth_proto, key, a);
 114                         if (!skb) {
 115                                 if (net_ratelimit())
 116                                         printk("execute_setter lost skb\n");
 117                                 return;
 118                         }
 119                 }
 120         }
 121         if (prev_port != -1)
 122                 do_output(dp, skb, max_len, prev_port);
 123         else
 124                 kfree_skb(skb);
 125 }
 126
 127 /* Updates 'sum', which is a field in 'skb''s data, given that a 4-byte field
 128  * covered by the sum has been changed from 'from' to 'to'.  If set,
 129  * 'pseudohdr' indicates that the field is in the TCP or UDP pseudo-header.
 130  * Based on nf_proto_csum_replace4. */
 131 static void update_csum(__sum16 *sum, struct sk_buff *skb,
 132                         __be32 from, __be32 to, int pseudohdr)
 133 {
 134         __be32 diff[] = { ~from, to };
 135         if (skb->ip_summed != CHECKSUM_PARTIAL) {
 136                 *sum = csum_fold(csum_partial((char *)diff, sizeof(diff),
 137                                 ~csum_unfold(*sum)));
 138                 if (skb->ip_summed == CHECKSUM_COMPLETE && pseudohdr)
 139                         skb->csum = ~csum_partial((char *)diff, sizeof(diff),
 140                                                 ~skb->csum);
 141         } else if (pseudohdr)
 142                 *sum = ~csum_fold(csum_partial((char *)diff, sizeof(diff),
 143                                 csum_unfold(*sum)));
 144 }
 145
 146 static void modify_nh(struct sk_buff *skb, uint16_t eth_proto,
 147                         uint8_t nw_proto, const struct ofp_action *a)
 148 {
 149         if (eth_proto == ETH_P_IP) {
 150                 struct iphdr *nh = ip_hdr(skb);
 151                 uint32_t new, *field;
 152
 153                 new = a->arg.nw_addr;
 154
 155                 if (a->type == htons(OFPAT_SET_NW_SRC))
 156                         field = &nh->saddr;
 157                 else
 158                         field = &nh->daddr;
 159
 160                 if (nw_proto == IPPROTO_TCP) {
 161                         struct tcphdr *th = tcp_hdr(skb);
 162                         update_csum(&th->check, skb, *field, new, 1);
 163                 } else if (nw_proto == IPPROTO_UDP) {
 164                         struct udphdr *th = udp_hdr(skb);
 165                         update_csum(&th->check, skb, *field, new, 1);
 166                 }
 167                 update_csum(&nh->check, skb, *field, new, 0);
 168                 *field = new;
 169         }
 170 }
 171
 172 static void modify_th(struct sk_buff *skb, uint16_t eth_proto,
 173                         uint8_t nw_proto, const struct ofp_action *a)
 174 {
 175         if (eth_proto == ETH_P_IP) {
 176                 uint16_t new, *field;
 177
 178                 new = a->arg.tp;
 179
 180                 if (nw_proto == IPPROTO_TCP) {
 181                         struct tcphdr *th = tcp_hdr(skb);
 182
 183                         if (a->type == htons(OFPAT_SET_TP_SRC))
 184                                 field = &th->source;
 185                         else
 186                                 field = &th->dest;
 187
 188                         update_csum(&th->check, skb, *field, new, 1);
 189                         *field = new;
 190                 } else if (nw_proto == IPPROTO_UDP) {
 191                         struct udphdr *th = udp_hdr(skb);
 192
 193                         if (a->type == htons(OFPAT_SET_TP_SRC))
 194                                 field = &th->source;
 195                         else
 196                                 field = &th->dest;
 197
 198                         update_csum(&th->check, skb, *field, new, 1);
 199                         *field = new;
 200                 }
 201         }
 202 }
 203
 204 static struct sk_buff *vlan_pull_tag(struct sk_buff *skb)
 205 {
 206         struct vlan_ethhdr *vh = vlan_eth_hdr(skb);
 207         struct ethhdr *eh;
 208
 209
 210         /* Verify we were given a vlan packet */
 211         if (vh->h_vlan_proto != htons(ETH_P_8021Q))
 212                 return skb;
 213
 214         memmove(skb->data + VLAN_HLEN, skb->data, 2 * VLAN_ETH_ALEN);
 215
 216         eh = (struct ethhdr *)skb_pull(skb, VLAN_HLEN);
 217
 218         skb->protocol = eh->h_proto;
 219         skb->mac_header += VLAN_HLEN;
 220
 221         return skb;
 222 }
 223
 224 static struct sk_buff *modify_vlan(struct sk_buff *skb,
 225                 const struct sw_flow_key *key, const struct ofp_action *a)
 226 {
 227         uint16_t new_id = ntohs(a->arg.vlan_id);
 228
 229         if (new_id != OFP_VLAN_NONE) {
 230                 if (key->dl_vlan != htons(OFP_VLAN_NONE)) {
 231                         /* Modify vlan id, but maintain other TCI values */
 232                         struct vlan_ethhdr *vh = vlan_eth_hdr(skb);
 233                         vh->h_vlan_TCI = (vh->h_vlan_TCI
 234                                         & ~(htons(VLAN_VID_MASK))) | a->arg.vlan_id;
 235                 } else  {
 236                         /* Add vlan header */
 237
 238                         /* xxx The vlan_put_tag function, doesn't seem to work
 239                          * xxx reliably when it attempts to use the hardware-accelerated
 240                          * xxx version.  We'll directly use the software version
 241                          * xxx until the problem can be diagnosed.
 242                          */
 243                         skb = __vlan_put_tag(skb, new_id);
 244                 }
 245         } else  {
 246                 /* Remove an existing vlan header if it exists */
 247                 vlan_pull_tag(skb);
 248         }
 249
 250         return skb;
 251 }
 252
 253 struct sk_buff *execute_setter(struct sk_buff *skb, uint16_t eth_proto,
 254                         const struct sw_flow_key *key, const struct ofp_action *a)
 255 {
 256         switch (ntohs(a->type)) {
 257         case OFPAT_SET_DL_VLAN:
 258                 skb = modify_vlan(skb, key, a);
 259                 break;
 260
 261         case OFPAT_SET_DL_SRC: {
 262                 struct ethhdr *eh = eth_hdr(skb);
 263                 memcpy(eh->h_source, a->arg.dl_addr, sizeof eh->h_source);
 264                 break;
 265         }
 266         case OFPAT_SET_DL_DST: {
 267                 struct ethhdr *eh = eth_hdr(skb);
 268                 memcpy(eh->h_dest, a->arg.dl_addr, sizeof eh->h_dest);
 269                 break;
 270         }
 271
 272         case OFPAT_SET_NW_SRC:
 273         case OFPAT_SET_NW_DST:
 274                 modify_nh(skb, eth_proto, key->nw_proto, a);
 275                 break;
 276
 277         case OFPAT_SET_TP_SRC:
 278         case OFPAT_SET_TP_DST:
 279                 modify_th(skb, eth_proto, key->nw_proto, a);
 280                 break;
 281
 282         default:
 283                 if (net_ratelimit())
 284                         printk("execute_setter: unknown action: %d\n", ntohs(a->type));
 285         }
 286
 287         return skb;
 288 }
 289
 290 static int
 291 recv_features_request(struct sw_chain *chain, const struct sender *sender,
 292                       const void *msg)
 293 {
 294         return dp_send_features_reply(chain->dp, sender);
 295 }
 296
 297 static int
 298 recv_get_config_request(struct sw_chain *chain, const struct sender *sender,
 299                         const void *msg)
 300 {
 301         return dp_send_config_reply(chain->dp, sender);
 302 }
 303
 304 static int
 305 recv_set_config(struct sw_chain *chain, const struct sender *sender,
 306                 const void *msg)
 307 {
 308         const struct ofp_switch_config *osc = msg;
 309         int flags;
 310
 311         flags = ntohs(osc->flags) & ~(OFPC_SEND_FLOW_EXP | OFPC_FRAG_MASK);
 312         if ((flags & OFPC_FRAG_MASK) != OFPC_FRAG_NORMAL
 313             && (flags & OFPC_FRAG_MASK) != OFPC_FRAG_DROP) {
 314                 flags = (flags & ~OFPC_FRAG_MASK) | OFPC_FRAG_DROP;
 315         }
 316         chain->dp->flags = flags;
 317
 318         chain->dp->miss_send_len = ntohs(osc->miss_send_len);
 319
 320         return 0;
 321 }
 322
 323 static int
 324 recv_packet_out(struct sw_chain *chain, const struct sender *sender,
 325                 const void *msg)
 326 {
 327         const struct ofp_packet_out *opo = msg;
 328         struct sk_buff *skb;
 329         struct vlan_ethhdr *mac;
 330         int nh_ofs;
 331
 332         if (ntohl(opo->buffer_id) == (uint32_t) -1) {
 333                 int data_len = ntohs(opo->header.length) - sizeof *opo;
 334
 335                 /* FIXME: there is likely a way to reuse the data in msg. */
 336                 skb = alloc_skb(data_len, GFP_ATOMIC);
 337                 if (!skb)
 338                         return -ENOMEM;
 339
 340                 /* FIXME?  We don't reserve NET_IP_ALIGN or NET_SKB_PAD since
 341                  * we're just transmitting this raw without examining anything
 342                  * at those layers. */
 343                 memcpy(skb_put(skb, data_len), opo->u.data, data_len);
 344                 dp_set_origin(chain->dp, ntohs(opo->in_port), skb);
 345
 346                 skb_set_mac_header(skb, 0);
 347                 mac = vlan_eth_hdr(skb);
 348                 if (likely(mac->h_vlan_proto != htons(ETH_P_8021Q)))
 349                         nh_ofs = sizeof(struct ethhdr);
 350                 else
 351                         nh_ofs = sizeof(struct vlan_ethhdr);
 352                 skb_set_network_header(skb, nh_ofs);
 353
 354                 dp_output_port(chain->dp, skb, ntohs(opo->out_port));
 355         } else {
 356                 struct sw_flow_key key;
 357                 int n_acts;
 358
 359                 skb = retrieve_skb(ntohl(opo->buffer_id));
 360                 if (!skb)
 361                         return -ESRCH;
 362                 dp_set_origin(chain->dp, ntohs(opo->in_port), skb);
 363
 364                 n_acts = (ntohs(opo->header.length) - sizeof *opo)
 365                                 / sizeof *opo->u.actions;
 366                 flow_extract(skb, ntohs(opo->in_port), &key);
 367                 execute_actions(chain->dp, skb, &key, opo->u.actions, n_acts);
 368         }
 369         return 0;
 370 }
 371
 372 static int
 373 recv_port_mod(struct sw_chain *chain, const struct sender *sender,
 374               const void *msg)
 375 {
 376         const struct ofp_port_mod *opm = msg;
 377
 378         dp_update_port_flags(chain->dp, &opm->desc);
 379
 380         return 0;
 381 }
 382
 383 static int
 384 recv_echo_request(struct sw_chain *chain, const struct sender *sender,
 385                   const void *msg)
 386 {
 387         return dp_send_echo_reply(chain->dp, sender, msg);
 388 }
 389
 390 static int
 391 recv_echo_reply(struct sw_chain *chain, const struct sender *sender,
 392                   const void *msg)
 393 {
 394         return 0;
 395 }
 396
 397 static int
 398 add_flow(struct sw_chain *chain, const struct ofp_flow_mod *ofm)
 399 {
 400         int error = -ENOMEM;
 401         int i;
 402         int n_acts;
 403         struct sw_flow *flow;
 404
 405
 406         /* To prevent loops, make sure there's no action to send to the
 407          * OFP_TABLE virtual port.
 408          */
 409         n_acts = (ntohs(ofm->header.length) - sizeof *ofm) / sizeof *ofm->actions;
 410         for (i=0; i<n_acts; i++) {
 411                 const struct ofp_action *a = &ofm->actions[i];
 412
 413                 if (a->type == htons(OFPAT_OUTPUT)
 414                                         && (a->arg.output.port == htons(OFPP_TABLE)
 415                                                 || a->arg.output.port == htons(OFPP_NONE))) {
 416                         /* xxx Send fancy new error message? */
 417                         goto error;
 418                 }
 419         }
 420
 421         /* Allocate memory. */
 422         flow = flow_alloc(n_acts, GFP_ATOMIC);
 423         if (flow == NULL)
 424                 goto error;
 425
 426         /* Fill out flow. */
 427         flow_extract_match(&flow->key, &ofm->match);
 428         flow->max_idle = ntohs(ofm->max_idle);
 429         flow->priority = flow->key.wildcards ? ntohs(ofm->priority) : -1;
 430         flow->timeout = jiffies + flow->max_idle * HZ;
 431         flow->n_actions = n_acts;
 432         flow->init_time = jiffies;
 433         flow->byte_count = 0;
 434         flow->packet_count = 0;
 435         spin_lock_init(&flow->lock);
 436         memcpy(flow->actions, ofm->actions, n_acts * sizeof *flow->actions);
 437
 438         /* Act. */
 439         error = chain_insert(chain, flow);
 440         if (error)
 441                 goto error_free_flow;
 442         error = 0;
 443         if (ntohl(ofm->buffer_id) != (uint32_t) -1) {
 444                 struct sk_buff *skb = retrieve_skb(ntohl(ofm->buffer_id));
 445                 if (skb) {
 446                         struct sw_flow_key key;
 447                         flow_used(flow, skb);
 448                         flow_extract(skb, ntohs(ofm->match.in_port), &key);
 449                         execute_actions(chain->dp, skb, &key,
 450                                         ofm->actions, n_acts);
 451                 }
 452                 else
 453                         error = -ESRCH;
 454         }
 455         return error;
 456
 457 error_free_flow:
 458         flow_free(flow);
 459 error:
 460         if (ntohl(ofm->buffer_id) != (uint32_t) -1)
 461                 discard_skb(ntohl(ofm->buffer_id));
 462         return error;
 463 }
 464
 465 static int
 466 recv_flow(struct sw_chain *chain, const struct sender *sender, const void *msg)
 467 {
 468         const struct ofp_flow_mod *ofm = msg;
 469         uint16_t command = ntohs(ofm->command);
 470
 471         if (command == OFPFC_ADD) {
 472                 return add_flow(chain, ofm);
 473         }  else if (command == OFPFC_DELETE) {
 474                 struct sw_flow_key key;
 475                 flow_extract_match(&key, &ofm->match);
 476                 return chain_delete(chain, &key, 0, 0) ? 0 : -ESRCH;
 477         } else if (command == OFPFC_DELETE_STRICT) {
 478                 struct sw_flow_key key;
 479                 uint16_t priority;
 480                 flow_extract_match(&key, &ofm->match);
 481                 priority = key.wildcards ? ntohs(ofm->priority) : -1;
 482                 return chain_delete(chain, &key, priority, 1) ? 0 : -ESRCH;
 483         } else {
 484                 return -ENOTSUPP;
 485         }
 486 }
 487
 488 /* 'msg', which is 'length' bytes long, was received across Netlink from
 489  * 'sender'.  Apply it to 'chain'. */
 490 int
 491 fwd_control_input(struct sw_chain *chain, const struct sender *sender,
 492                   const void *msg, size_t length)
 493 {
 494
 495         struct openflow_packet {
 496                 size_t min_size;
 497                 int (*handler)(struct sw_chain *, const struct sender *,
 498                                const void *);
 499         };
 500
 501         static const struct openflow_packet packets[] = {
 502                 [OFPT_FEATURES_REQUEST] = {
 503                         sizeof (struct ofp_header),
 504                         recv_features_request,
 505                 },
 506                 [OFPT_GET_CONFIG_REQUEST] = {
 507                         sizeof (struct ofp_header),
 508                         recv_get_config_request,
 509                 },
 510                 [OFPT_SET_CONFIG] = {
 511                         sizeof (struct ofp_switch_config),
 512                         recv_set_config,
 513                 },
 514                 [OFPT_PACKET_OUT] = {
 515                         sizeof (struct ofp_packet_out),
 516                         recv_packet_out,
 517                 },
 518                 [OFPT_FLOW_MOD] = {
 519                         sizeof (struct ofp_flow_mod),
 520                         recv_flow,
 521                 },
 522                 [OFPT_PORT_MOD] = {
 523                         sizeof (struct ofp_port_mod),
 524                         recv_port_mod,
 525                 },
 526                 [OFPT_ECHO_REQUEST] = {
 527                         sizeof (struct ofp_header),
 528                         recv_echo_request,
 529                 },
 530                 [OFPT_ECHO_REPLY] = {
 531                         sizeof (struct ofp_header),
 532                         recv_echo_reply,
 533                 },
 534         };
 535
 536         const struct openflow_packet *pkt;
 537         struct ofp_header *oh;
 538
 539         oh = (struct ofp_header *) msg;
 540         if (oh->version != OFP_VERSION || oh->type >= ARRAY_SIZE(packets)
 541                 || ntohs(oh->length) > length)
 542                 return -EINVAL;
 543
 544         pkt = &packets[oh->type];
 545         if (!pkt->handler)
 546                 return -ENOSYS;
 547         if (length < pkt->min_size)
 548                 return -EFAULT;
 549
 550         return pkt->handler(chain, sender, msg);
 551 }
 552
 553 /* Packet buffering. */
 554
 555 #define OVERWRITE_SECS  1
 556 #define OVERWRITE_JIFFIES (OVERWRITE_SECS * HZ)
 557
 558 struct packet_buffer {
 559         struct sk_buff *skb;
 560         uint32_t cookie;
 561         unsigned long exp_jiffies;
 562 };
 563
 564 static struct packet_buffer buffers[N_PKT_BUFFERS];
 565 static unsigned int buffer_idx;
 566 static DEFINE_SPINLOCK(buffer_lock);
 567
 568 uint32_t fwd_save_skb(struct sk_buff *skb)
 569 {
 570         struct sk_buff *old_skb = NULL;
 571         struct packet_buffer *p;
 572         unsigned long int flags;
 573         uint32_t id;
 574
 575         spin_lock_irqsave(&buffer_lock, flags);
 576         buffer_idx = (buffer_idx + 1) & PKT_BUFFER_MASK;
 577         p = &buffers[buffer_idx];
 578         if (p->skb) {
 579                 /* Don't buffer packet if existing entry is less than
 580                  * OVERWRITE_SECS old. */
 581                 if (time_before(jiffies, p->exp_jiffies)) {
 582                         spin_unlock_irqrestore(&buffer_lock, flags);
 583                         return -1;
 584                 } else {
 585                         /* Defer kfree_skb() until interrupts re-enabled. */
 586                         old_skb = p->skb;
 587                 }
 588         }
 589         /* Don't use maximum cookie value since the all-bits-1 id is
 590          * special. */
 591         if (++p->cookie >= (1u << PKT_COOKIE_BITS) - 1)
 592                 p->cookie = 0;
 593         skb_get(skb);
 594         p->skb = skb;
 595         p->exp_jiffies = jiffies + OVERWRITE_JIFFIES;
 596         id = buffer_idx | (p->cookie << PKT_BUFFER_BITS);
 597         spin_unlock_irqrestore(&buffer_lock, flags);
 598
 599         if (old_skb)
 600                 kfree_skb(old_skb);
 601
 602         return id;
 603 }
 604
 605 static struct sk_buff *retrieve_skb(uint32_t id)
 606 {
 607         unsigned long int flags;
 608         struct sk_buff *skb = NULL;
 609         struct packet_buffer *p;
 610
 611         spin_lock_irqsave(&buffer_lock, flags);
 612         p = &buffers[id & PKT_BUFFER_MASK];
 613         if (p->cookie == id >> PKT_BUFFER_BITS) {
 614                 skb = p->skb;
 615                 p->skb = NULL;
 616         } else {
 617                 printk("cookie mismatch: %x != %x\n",
 618                                 id >> PKT_BUFFER_BITS, p->cookie);
 619         }
 620         spin_unlock_irqrestore(&buffer_lock, flags);
 621
 622         return skb;
 623 }
 624
 625 void fwd_discard_all(void)
 626 {
 627         int i;
 628
 629         for (i = 0; i < N_PKT_BUFFERS; i++) {
 630                 struct sk_buff *skb;
 631                 unsigned long int flags;
 632
 633                 /* Defer kfree_skb() until interrupts re-enabled. */
 634                 spin_lock_irqsave(&buffer_lock, flags);
 635                 skb = buffers[i].skb;
 636                 buffers[i].skb = NULL;
 637                 spin_unlock_irqrestore(&buffer_lock, flags);
 638
 639                 kfree_skb(skb);
 640         }
 641 }
 642
 643 static void discard_skb(uint32_t id)
 644 {
 645         struct sk_buff *old_skb = NULL;
 646         unsigned long int flags;
 647         struct packet_buffer *p;
 648
 649         spin_lock_irqsave(&buffer_lock, flags);
 650         p = &buffers[id & PKT_BUFFER_MASK];
 651         if (p->cookie == id >> PKT_BUFFER_BITS) {
 652                 /* Defer kfree_skb() until interrupts re-enabled. */
 653                 old_skb = p->skb;
 654                 p->skb = NULL;
 655         }
 656         spin_unlock_irqrestore(&buffer_lock, flags);
 657
 658         if (old_skb)
 659                 kfree_skb(old_skb);
 660 }
 661
 662 void fwd_exit(void)
 663 {
 664         fwd_discard_all();
 665 }
 666
 667 /* Utility functions. */
 668
 669 /* Makes '*pskb' writable, possibly copying it and setting '*pskb' to point to
 670  * the copy.
 671  * Returns 1 if successful, 0 on failure. */
 672 static int
 673 make_writable(struct sk_buff **pskb)
 674 {
 675         /* Based on skb_make_writable() in net/netfilter/core.c. */
 676         struct sk_buff *nskb;
 677
 678         /* Not exclusive use of packet?  Must copy. */
 679         if (skb_shared(*pskb) || skb_cloned(*pskb))
 680                 goto copy_skb;
 681
 682         return pskb_may_pull(*pskb, 40); /* FIXME? */
 683
 684 copy_skb:
 685         nskb = skb_copy(*pskb, GFP_ATOMIC);
 686         if (!nskb)
 687                 return 0;
 688         BUG_ON(skb_is_nonlinear(nskb));
 689
 690         /* Rest of kernel will get very unhappy if we pass it a
 691            suddenly-orphaned skbuff */
 692         if ((*pskb)->sk)
 693                 skb_set_owner_w(nskb, (*pskb)->sk);
 694         kfree_skb(*pskb);
 695         *pskb = nskb;
 696         return 1;
 697 }