datapath/vport-gre.c

   1 /*
   2  * Copyright (c) 2007-2012 Nicira, Inc.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of version 2 of the GNU General Public
   6  * License as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful, but
   9  * WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public License
  14  * along with this program; if not, write to the Free Software
  15  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA
  16  * 02110-1301, USA
  17  */
  18
  19 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  20
  21 #include <linux/if.h>
  22 #include <linux/skbuff.h>
  23 #include <linux/ip.h>
  24 #include <linux/if_tunnel.h>
  25 #include <linux/if_vlan.h>
  26 #include <linux/in.h>
  27
  28 #include <net/icmp.h>
  29 #include <net/ip.h>
  30 #include <net/protocol.h>
  31
  32 #include "datapath.h"
  33 #include "tunnel.h"
  34 #include "vport.h"
  35 #include "vport-generic.h"
  36
  37 /*
  38  * The GRE header is composed of a series of sections: a base and then a variable
  39  * number of options.
  40  */
  41 #define GRE_HEADER_SECTION 4
  42
  43 struct gre_base_hdr {
  44         __be16 flags;
  45         __be16 protocol;
  46 };
  47
  48 static int gre_hdr_len(const struct tnl_mutable_config *mutable,
  49                        const struct ovs_key_ipv4_tunnel *tun_key)
  50 {
  51         int len;
  52         u32 flags;
  53         __be64 out_key;
  54
  55         tnl_get_param(mutable, tun_key, &flags, &out_key);
  56         len = GRE_HEADER_SECTION;
  57
  58         if (flags & TNL_F_CSUM)
  59                 len += GRE_HEADER_SECTION;
  60
  61         /* Set key for GRE64 tunnels, even when key if is zero. */
  62         if (out_key ||
  63             mutable->key.tunnel_type & TNL_T_PROTO_GRE64 ||
  64             flags & TNL_F_OUT_KEY_ACTION) {
  65
  66                 len += GRE_HEADER_SECTION;
  67                 if (mutable->key.tunnel_type & TNL_T_PROTO_GRE64)
  68                         len += GRE_HEADER_SECTION;
  69         }
  70         return len;
  71 }
  72
  73
  74 /* Returns the least-significant 32 bits of a __be64. */
  75 static __be32 be64_get_low32(__be64 x)
  76 {
  77 #ifdef __BIG_ENDIAN
  78         return (__force __be32)x;
  79 #else
  80         return (__force __be32)((__force u64)x >> 32);
  81 #endif
  82 }
  83
  84 static __be32 be64_get_high32(__be64 x)
  85 {
  86 #ifdef __BIG_ENDIAN
  87         return (__force __be32)((__force u64)x >> 32);
  88 #else
  89         return (__force __be32)x;
  90 #endif
  91 }
  92
  93 static struct sk_buff *gre_build_header(const struct vport *vport,
  94                                          const struct tnl_mutable_config *mutable,
  95                                          struct dst_entry *dst,
  96                                          struct sk_buff *skb,
  97                                          int tunnel_hlen)
  98 {
  99         u32 flags;
 100         __be64 out_key;
 101         const struct ovs_key_ipv4_tunnel *tun_key = OVS_CB(skb)->tun_key;
 102         __be32 *options = (__be32 *)(skb_network_header(skb) + tunnel_hlen
 103                                                - GRE_HEADER_SECTION);
 104         struct gre_base_hdr *greh = (struct gre_base_hdr *) skb_transport_header(skb);
 105
 106         tnl_get_param(mutable, tun_key, &flags, &out_key);
 107
 108         greh->protocol = htons(ETH_P_TEB);
 109         greh->flags = 0;
 110
 111         /* Work backwards over the options so the checksum is last. */
 112         if (out_key || flags & TNL_F_OUT_KEY_ACTION ||
 113             mutable->key.tunnel_type & TNL_T_PROTO_GRE64) {
 114                 greh->flags |= GRE_KEY;
 115                 if (mutable->key.tunnel_type & TNL_T_PROTO_GRE64) {
 116                         /* Set higher 32 bits to seq. */
 117                         *options = be64_get_high32(out_key);
 118                         options--;
 119                         greh->flags |= GRE_SEQ;
 120                 }
 121                 *options = be64_get_low32(out_key);
 122                 options--;
 123         }
 124
 125         if (flags & TNL_F_CSUM) {
 126                 greh->flags |= GRE_CSUM;
 127                 *options = 0;
 128                 *(__sum16 *)options = csum_fold(skb_checksum(skb,
 129                                                 skb_transport_offset(skb),
 130                                                 skb->len - skb_transport_offset(skb),
 131                                                 0));
 132         }
 133         /*
 134          * Allow our local IP stack to fragment the outer packet even if the
 135          * DF bit is set as a last resort.  We also need to force selection of
 136          * an IP ID here because Linux will otherwise leave it at 0 if the
 137          * packet originally had DF set.
 138          */
 139         skb->local_df = 1;
 140         __ip_select_ident(ip_hdr(skb), dst, 0);
 141
 142         return skb;
 143 }
 144
 145 static __be64 key_to_tunnel_id(__be32 key, __be32 seq)
 146 {
 147 #ifdef __BIG_ENDIAN
 148         return (__force __be64)((__force u64)seq << 32 | (__force u32)key);
 149 #else
 150         return (__force __be64)((__force u64)key << 32 | (__force u32)seq);
 151 #endif
 152 }
 153
 154 static int parse_header(struct iphdr *iph, __be16 *flags, __be64 *tun_id,
 155                         u32 *tunnel_type)
 156 {
 157         /* IP and ICMP protocol handlers check that the IHL is valid. */
 158         struct gre_base_hdr *greh = (struct gre_base_hdr *)((u8 *)iph + (iph->ihl << 2));
 159         __be32 *options = (__be32 *)(greh + 1);
 160         int hdr_len;
 161
 162         *flags = greh->flags;
 163
 164         if (unlikely(greh->flags & (GRE_VERSION | GRE_ROUTING)))
 165                 return -EINVAL;
 166
 167         if (unlikely(greh->protocol != htons(ETH_P_TEB)))
 168                 return -EINVAL;
 169
 170         hdr_len = GRE_HEADER_SECTION;
 171
 172         if (greh->flags & GRE_CSUM) {
 173                 hdr_len += GRE_HEADER_SECTION;
 174                 options++;
 175         }
 176
 177         if (greh->flags & GRE_KEY) {
 178                 __be32 seq;
 179                 __be32 gre_key;
 180
 181                 gre_key = *options;
 182                 hdr_len += GRE_HEADER_SECTION;
 183                 options++;
 184
 185                 if (greh->flags & GRE_SEQ) {
 186                         seq = *options;
 187                         *tunnel_type = TNL_T_PROTO_GRE64;
 188                 } else {
 189                         seq = 0;
 190                         *tunnel_type = TNL_T_PROTO_GRE;
 191                 }
 192                 *tun_id = key_to_tunnel_id(gre_key, seq);
 193         } else {
 194                 *tun_id = 0;
 195                 /* Ignore GRE seq if there is no key present. */
 196                 *tunnel_type = TNL_T_PROTO_GRE;
 197         }
 198
 199         if (greh->flags & GRE_SEQ)
 200                 hdr_len += GRE_HEADER_SECTION;
 201
 202         return hdr_len;
 203 }
 204
 205 /* Called with rcu_read_lock and BH disabled. */
 206 static void gre_err(struct sk_buff *skb, u32 info)
 207 {
 208         struct vport *vport;
 209         const struct tnl_mutable_config *mutable;
 210         const int type = icmp_hdr(skb)->type;
 211         const int code = icmp_hdr(skb)->code;
 212         int mtu = ntohs(icmp_hdr(skb)->un.frag.mtu);
 213         u32 tunnel_type;
 214
 215         struct iphdr *iph;
 216         __be16 flags;
 217         __be64 key;
 218         int tunnel_hdr_len, tot_hdr_len;
 219         unsigned int orig_mac_header;
 220         unsigned int orig_nw_header;
 221
 222         if (type != ICMP_DEST_UNREACH || code != ICMP_FRAG_NEEDED)
 223                 return;
 224
 225         /*
 226          * The mimimum size packet that we would actually be able to process:
 227          * encapsulating IP header, minimum GRE header, Ethernet header,
 228          * inner IPv4 header.
 229          */
 230         if (!pskb_may_pull(skb, sizeof(struct iphdr) + GRE_HEADER_SECTION +
 231                                 ETH_HLEN + sizeof(struct iphdr)))
 232                 return;
 233
 234         iph = (struct iphdr *)skb->data;
 235         if (ipv4_is_multicast(iph->daddr))
 236                 return;
 237
 238         tunnel_hdr_len = parse_header(iph, &flags, &key, &tunnel_type);
 239         if (tunnel_hdr_len < 0)
 240                 return;
 241
 242         vport = ovs_tnl_find_port(dev_net(skb->dev), iph->saddr, iph->daddr, key,
 243                                   tunnel_type, &mutable);
 244         if (!vport)
 245                 return;
 246
 247         /*
 248          * Packets received by this function were previously sent by us, so
 249          * any comparisons should be to the output values, not the input.
 250          * However, it's not really worth it to have a hash table based on
 251          * output keys (especially since ICMP error handling of tunneled packets
 252          * isn't that reliable anyways).  Therefore, we do a lookup based on the
 253          * out key as if it were the in key and then check to see if the input
 254          * and output keys are the same.
 255          */
 256         if (mutable->key.in_key != mutable->out_key)
 257                 return;
 258
 259         if (!!(mutable->flags & TNL_F_IN_KEY_MATCH) !=
 260             !!(mutable->flags & TNL_F_OUT_KEY_ACTION))
 261                 return;
 262
 263         if ((mutable->flags & TNL_F_CSUM) && !(flags & GRE_CSUM))
 264                 return;
 265
 266         tunnel_hdr_len += iph->ihl << 2;
 267
 268         orig_mac_header = skb_mac_header(skb) - skb->data;
 269         orig_nw_header = skb_network_header(skb) - skb->data;
 270         skb_set_mac_header(skb, tunnel_hdr_len);
 271
 272         tot_hdr_len = tunnel_hdr_len + ETH_HLEN;
 273
 274         skb->protocol = eth_hdr(skb)->h_proto;
 275         if (skb->protocol == htons(ETH_P_8021Q)) {
 276                 tot_hdr_len += VLAN_HLEN;
 277                 skb->protocol = vlan_eth_hdr(skb)->h_vlan_encapsulated_proto;
 278         }
 279
 280         skb_set_network_header(skb, tot_hdr_len);
 281         mtu -= tot_hdr_len;
 282
 283         if (skb->protocol == htons(ETH_P_IP))
 284                 tot_hdr_len += sizeof(struct iphdr);
 285 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
 286         else if (skb->protocol == htons(ETH_P_IPV6))
 287                 tot_hdr_len += sizeof(struct ipv6hdr);
 288 #endif
 289         else
 290                 goto out;
 291
 292         if (!pskb_may_pull(skb, tot_hdr_len))
 293                 goto out;
 294
 295         if (skb->protocol == htons(ETH_P_IP)) {
 296                 if (mtu < IP_MIN_MTU) {
 297                         if (ntohs(ip_hdr(skb)->tot_len) >= IP_MIN_MTU)
 298                                 mtu = IP_MIN_MTU;
 299                         else
 300                                 goto out;
 301                 }
 302
 303         }
 304 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
 305         else if (skb->protocol == htons(ETH_P_IPV6)) {
 306                 if (mtu < IPV6_MIN_MTU) {
 307                         unsigned int packet_length = sizeof(struct ipv6hdr) +
 308                                               ntohs(ipv6_hdr(skb)->payload_len);
 309
 310                         if (packet_length >= IPV6_MIN_MTU
 311                             || ntohs(ipv6_hdr(skb)->payload_len) == 0)
 312                                 mtu = IPV6_MIN_MTU;
 313                         else
 314                                 goto out;
 315                 }
 316         }
 317 #endif
 318
 319         __skb_pull(skb, tunnel_hdr_len);
 320         ovs_tnl_frag_needed(vport, mutable, skb, mtu);
 321         __skb_push(skb, tunnel_hdr_len);
 322
 323 out:
 324         skb_set_mac_header(skb, orig_mac_header);
 325         skb_set_network_header(skb, orig_nw_header);
 326         skb->protocol = htons(ETH_P_IP);
 327 }
 328
 329 static bool check_checksum(struct sk_buff *skb)
 330 {
 331         struct iphdr *iph = ip_hdr(skb);
 332         struct gre_base_hdr *greh = (struct gre_base_hdr *)(iph + 1);
 333         __sum16 csum = 0;
 334
 335         if (greh->flags & GRE_CSUM) {
 336                 switch (skb->ip_summed) {
 337                 case CHECKSUM_COMPLETE:
 338                         csum = csum_fold(skb->csum);
 339
 340                         if (!csum)
 341                                 break;
 342                         /* Fall through. */
 343
 344                 case CHECKSUM_NONE:
 345                         skb->csum = 0;
 346                         csum = __skb_checksum_complete(skb);
 347                         skb->ip_summed = CHECKSUM_COMPLETE;
 348                         break;
 349                 }
 350         }
 351
 352         return (csum == 0);
 353 }
 354
 355 static u32 gre_flags_to_tunnel_flags(const struct tnl_mutable_config *mutable,
 356                                      __be16 gre_flags, __be64 *key)
 357 {
 358         u32 tunnel_flags = 0;
 359
 360         if (gre_flags & GRE_KEY) {
 361                 if (mutable->flags & TNL_F_IN_KEY_MATCH ||
 362                     !mutable->key.daddr)
 363                         tunnel_flags = OVS_TNL_F_KEY;
 364                 else
 365                         *key = 0;
 366         }
 367
 368         if (gre_flags & GRE_CSUM)
 369                 tunnel_flags |= OVS_TNL_F_CSUM;
 370
 371         return tunnel_flags;
 372 }
 373
 374 /* Called with rcu_read_lock and BH disabled. */
 375 static int gre_rcv(struct sk_buff *skb)
 376 {
 377         struct vport *vport;
 378         const struct tnl_mutable_config *mutable;
 379         int hdr_len;
 380         struct iphdr *iph;
 381         struct ovs_key_ipv4_tunnel tun_key;
 382         __be16 gre_flags;
 383         u32 tnl_flags;
 384         __be64 key;
 385         u32 tunnel_type;
 386
 387         if (unlikely(!pskb_may_pull(skb, sizeof(struct gre_base_hdr) + ETH_HLEN)))
 388                 goto error;
 389         if (unlikely(!check_checksum(skb)))
 390                 goto error;
 391
 392         hdr_len = parse_header(ip_hdr(skb), &gre_flags, &key, &tunnel_type);
 393         if (unlikely(hdr_len < 0))
 394                 goto error;
 395
 396         if (unlikely(!pskb_may_pull(skb, hdr_len + ETH_HLEN)))
 397                 goto error;
 398
 399         iph = ip_hdr(skb);
 400         vport = ovs_tnl_find_port(dev_net(skb->dev), iph->daddr, iph->saddr, key,
 401                                   tunnel_type, &mutable);
 402         if (unlikely(!vport)) {
 403                 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_PORT_UNREACH, 0);
 404                 goto error;
 405         }
 406
 407         tnl_flags = gre_flags_to_tunnel_flags(mutable, gre_flags, &key);
 408         tnl_tun_key_init(&tun_key, iph, key, tnl_flags);
 409         OVS_CB(skb)->tun_key = &tun_key;
 410
 411         __skb_pull(skb, hdr_len);
 412         skb_postpull_rcsum(skb, skb_transport_header(skb), hdr_len + ETH_HLEN);
 413
 414         ovs_tnl_rcv(vport, skb);
 415         return 0;
 416
 417 error:
 418         kfree_skb(skb);
 419         return 0;
 420 }
 421
 422 static const struct tnl_ops gre_tnl_ops = {
 423         .tunnel_type    = TNL_T_PROTO_GRE,
 424         .ipproto        = IPPROTO_GRE,
 425         .hdr_len        = gre_hdr_len,
 426         .build_header   = gre_build_header,
 427 };
 428
 429 static struct vport *gre_create(const struct vport_parms *parms)
 430 {
 431         return ovs_tnl_create(parms, &ovs_gre_vport_ops, &gre_tnl_ops);
 432 }
 433
 434 static struct vport *gre_create_ft(const struct vport_parms *parms)
 435 {
 436         return ovs_tnl_create(parms, &ovs_gre_ft_vport_ops, &gre_tnl_ops);
 437 }
 438
 439 static const struct tnl_ops gre64_tnl_ops = {
 440         .tunnel_type    = TNL_T_PROTO_GRE64,
 441         .ipproto        = IPPROTO_GRE,
 442         .hdr_len        = gre_hdr_len,
 443         .build_header   = gre_build_header,
 444 };
 445
 446 static struct vport *gre_create64(const struct vport_parms *parms)
 447 {
 448         return ovs_tnl_create(parms, &ovs_gre64_vport_ops, &gre64_tnl_ops);
 449 }
 450
 451 static const struct net_protocol gre_protocol_handlers = {
 452         .handler        =       gre_rcv,
 453         .err_handler    =       gre_err,
 454 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,32)
 455         .netns_ok       =       1,
 456 #endif
 457 };
 458
 459 static bool inited;
 460
 461 static int gre_init(void)
 462 {
 463         int err;
 464
 465         if (inited)
 466                 return 0;
 467
 468         inited = true;
 469         err = inet_add_protocol(&gre_protocol_handlers, IPPROTO_GRE);
 470         if (err)
 471                 pr_warn("cannot register gre protocol handler\n");
 472
 473         return err;
 474 }
 475
 476 static void gre_exit(void)
 477 {
 478         if (!inited)
 479                 return;
 480
 481         inited = false;
 482
 483         inet_del_protocol(&gre_protocol_handlers, IPPROTO_GRE);
 484 }
 485
 486 const struct vport_ops ovs_gre_ft_vport_ops = {
 487         .type           = OVS_VPORT_TYPE_FT_GRE,
 488         .flags          = VPORT_F_TUN_ID,
 489         .init           = gre_init,
 490         .exit           = gre_exit,
 491         .create         = gre_create_ft,
 492         .destroy        = ovs_tnl_destroy,
 493         .set_addr       = ovs_tnl_set_addr,
 494         .get_name       = ovs_tnl_get_name,
 495         .get_addr       = ovs_tnl_get_addr,
 496         .get_options    = ovs_tnl_get_options,
 497         .set_options    = ovs_tnl_set_options,
 498         .get_dev_flags  = ovs_vport_gen_get_dev_flags,
 499         .is_running     = ovs_vport_gen_is_running,
 500         .get_operstate  = ovs_vport_gen_get_operstate,
 501         .send           = ovs_tnl_send,
 502 };
 503
 504 const struct vport_ops ovs_gre_vport_ops = {
 505         .type           = OVS_VPORT_TYPE_GRE,
 506         .flags          = VPORT_F_TUN_ID,
 507         .init           = gre_init,
 508         .exit           = gre_exit,
 509         .create         = gre_create,
 510         .destroy        = ovs_tnl_destroy,
 511         .set_addr       = ovs_tnl_set_addr,
 512         .get_name       = ovs_tnl_get_name,
 513         .get_addr       = ovs_tnl_get_addr,
 514         .get_options    = ovs_tnl_get_options,
 515         .set_options    = ovs_tnl_set_options,
 516         .get_dev_flags  = ovs_vport_gen_get_dev_flags,
 517         .is_running     = ovs_vport_gen_is_running,
 518         .get_operstate  = ovs_vport_gen_get_operstate,
 519         .send           = ovs_tnl_send,
 520 };
 521
 522 const struct vport_ops ovs_gre64_vport_ops = {
 523         .type           = OVS_VPORT_TYPE_GRE64,
 524         .flags          = VPORT_F_TUN_ID,
 525         .init           = gre_init,
 526         .exit           = gre_exit,
 527         .create         = gre_create64,
 528         .destroy        = ovs_tnl_destroy,
 529         .set_addr       = ovs_tnl_set_addr,
 530         .get_name       = ovs_tnl_get_name,
 531         .get_addr       = ovs_tnl_get_addr,
 532         .get_options    = ovs_tnl_get_options,
 533         .set_options    = ovs_tnl_set_options,
 534         .get_dev_flags  = ovs_vport_gen_get_dev_flags,
 535         .is_running     = ovs_vport_gen_is_running,
 536         .get_operstate  = ovs_vport_gen_get_operstate,
 537         .send           = ovs_tnl_send,
 538 };