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