datapath/vport-gre.c

   1 /*
   2  * Copyright (c) 2010, 2011 Nicira Networks.
   3  * Distributed under the terms of the GNU GPL version 2.
   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 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  10
  11 #include <linux/if.h>
  12 #include <linux/skbuff.h>
  13 #include <linux/ip.h>
  14 #include <linux/if_tunnel.h>
  15 #include <linux/if_vlan.h>
  16 #include <linux/in.h>
  17
  18 #include <net/icmp.h>
  19 #include <net/ip.h>
  20 #include <net/protocol.h>
  21
  22 #include "tunnel.h"
  23 #include "vport.h"
  24 #include "vport-generic.h"
  25
  26 /*
  27  * The GRE header is composed of a series of sections: a base and then a variable
  28  * number of options.
  29  */
  30 #define GRE_HEADER_SECTION 4
  31
  32 struct gre_base_hdr {
  33         __be16 flags;
  34         __be16 protocol;
  35 };
  36
  37 static int gre_hdr_len(const struct tnl_mutable_config *mutable)
  38 {
  39         int len;
  40
  41         len = GRE_HEADER_SECTION;
  42
  43         if (mutable->flags & TNL_F_CSUM)
  44                 len += GRE_HEADER_SECTION;
  45
  46         if (mutable->out_key || mutable->flags & TNL_F_OUT_KEY_ACTION)
  47                 len += GRE_HEADER_SECTION;
  48
  49         return len;
  50 }
  51
  52 /* Returns the least-significant 32 bits of a __be64. */
  53 static __be32 be64_get_low32(__be64 x)
  54 {
  55 #ifdef __BIG_ENDIAN
  56         return (__force __be32)x;
  57 #else
  58         return (__force __be32)((__force u64)x >> 32);
  59 #endif
  60 }
  61
  62 static void gre_build_header(const struct vport *vport,
  63                              const struct tnl_mutable_config *mutable,
  64                              void *header)
  65 {
  66         struct gre_base_hdr *greh = header;
  67         __be32 *options = (__be32 *)(greh + 1);
  68
  69         greh->protocol = htons(ETH_P_TEB);
  70         greh->flags = 0;
  71
  72         if (mutable->flags & TNL_F_CSUM) {
  73                 greh->flags |= GRE_CSUM;
  74                 *options = 0;
  75                 options++;
  76         }
  77
  78         if (mutable->out_key || mutable->flags & TNL_F_OUT_KEY_ACTION)
  79                 greh->flags |= GRE_KEY;
  80
  81         if (mutable->out_key)
  82                 *options = be64_get_low32(mutable->out_key);
  83 }
  84
  85 static struct sk_buff *gre_update_header(const struct vport *vport,
  86                                          const struct tnl_mutable_config *mutable,
  87                                          struct dst_entry *dst,
  88                                          struct sk_buff *skb)
  89 {
  90         __be32 *options = (__be32 *)(skb_network_header(skb) + mutable->tunnel_hlen
  91                                                - GRE_HEADER_SECTION);
  92
  93         /* Work backwards over the options so the checksum is last. */
  94         if (mutable->flags & TNL_F_OUT_KEY_ACTION) {
  95                 *options = be64_get_low32(OVS_CB(skb)->tun_id);
  96                 options--;
  97         }
  98
  99         if (mutable->flags & TNL_F_CSUM)
 100                 *(__sum16 *)options = csum_fold(skb_checksum(skb,
 101                                                 skb_transport_offset(skb),
 102                                                 skb->len - skb_transport_offset(skb),
 103                                                 0));
 104         /*
 105          * Allow our local IP stack to fragment the outer packet even if the
 106          * DF bit is set as a last resort.
 107          */
 108         skb->local_df = 1;
 109
 110         return skb;
 111 }
 112
 113 /* Zero-extends a __be32 into the least-significant 32 bits of a __be64. */
 114 static __be64 be32_extend_to_be64(__be32 x)
 115 {
 116 #ifdef __BIG_ENDIAN
 117         return (__force __be64)x;
 118 #else
 119         return (__force __be64)((__force u64)x << 32);
 120 #endif
 121 }
 122
 123 static int parse_header(struct iphdr *iph, __be16 *flags, __be64 *key)
 124 {
 125         /* IP and ICMP protocol handlers check that the IHL is valid. */
 126         struct gre_base_hdr *greh = (struct gre_base_hdr *)((u8 *)iph + (iph->ihl << 2));
 127         __be32 *options = (__be32 *)(greh + 1);
 128         int hdr_len;
 129
 130         *flags = greh->flags;
 131
 132         if (unlikely(greh->flags & (GRE_VERSION | GRE_ROUTING)))
 133                 return -EINVAL;
 134
 135         if (unlikely(greh->protocol != htons(ETH_P_TEB)))
 136                 return -EINVAL;
 137
 138         hdr_len = GRE_HEADER_SECTION;
 139
 140         if (greh->flags & GRE_CSUM) {
 141                 hdr_len += GRE_HEADER_SECTION;
 142                 options++;
 143         }
 144
 145         if (greh->flags & GRE_KEY) {
 146                 hdr_len += GRE_HEADER_SECTION;
 147
 148                 *key = be32_extend_to_be64(*options);
 149                 options++;
 150         } else
 151                 *key = 0;
 152
 153         if (unlikely(greh->flags & GRE_SEQ))
 154                 hdr_len += GRE_HEADER_SECTION;
 155
 156         return hdr_len;
 157 }
 158
 159 /* Called with rcu_read_lock and BH disabled. */
 160 static void gre_err(struct sk_buff *skb, u32 info)
 161 {
 162         struct vport *vport;
 163         const struct tnl_mutable_config *mutable;
 164         const int type = icmp_hdr(skb)->type;
 165         const int code = icmp_hdr(skb)->code;
 166         int mtu = ntohs(icmp_hdr(skb)->un.frag.mtu);
 167
 168         struct iphdr *iph;
 169         __be16 flags;
 170         __be64 key;
 171         int tunnel_hdr_len, tot_hdr_len;
 172         unsigned int orig_mac_header;
 173         unsigned int orig_nw_header;
 174
 175         if (type != ICMP_DEST_UNREACH || code != ICMP_FRAG_NEEDED)
 176                 return;
 177
 178         /*
 179          * The mimimum size packet that we would actually be able to process:
 180          * encapsulating IP header, minimum GRE header, Ethernet header,
 181          * inner IPv4 header.
 182          */
 183         if (!pskb_may_pull(skb, sizeof(struct iphdr) + GRE_HEADER_SECTION +
 184                                 ETH_HLEN + sizeof(struct iphdr)))
 185                 return;
 186
 187         iph = (struct iphdr *)skb->data;
 188
 189         tunnel_hdr_len = parse_header(iph, &flags, &key);
 190         if (tunnel_hdr_len < 0)
 191                 return;
 192
 193         vport = tnl_find_port(iph->saddr, iph->daddr, key,
 194                               TNL_T_PROTO_GRE | TNL_T_KEY_EITHER, &mutable);
 195         if (!vport)
 196                 return;
 197
 198         /*
 199          * Packets received by this function were previously sent by us, so
 200          * any comparisons should be to the output values, not the input.
 201          * However, it's not really worth it to have a hash table based on
 202          * output keys (especially since ICMP error handling of tunneled packets
 203          * isn't that reliable anyways).  Therefore, we do a lookup based on the
 204          * out key as if it were the in key and then check to see if the input
 205          * and output keys are the same.
 206          */
 207         if (mutable->in_key != mutable->out_key)
 208                 return;
 209
 210         if (!!(mutable->flags & TNL_F_IN_KEY_MATCH) !=
 211             !!(mutable->flags & TNL_F_OUT_KEY_ACTION))
 212                 return;
 213
 214         if ((mutable->flags & TNL_F_CSUM) && !(flags & GRE_CSUM))
 215                 return;
 216
 217         tunnel_hdr_len += iph->ihl << 2;
 218
 219         orig_mac_header = skb_mac_header(skb) - skb->data;
 220         orig_nw_header = skb_network_header(skb) - skb->data;
 221         skb_set_mac_header(skb, tunnel_hdr_len);
 222
 223         tot_hdr_len = tunnel_hdr_len + ETH_HLEN;
 224
 225         skb->protocol = eth_hdr(skb)->h_proto;
 226         if (skb->protocol == htons(ETH_P_8021Q)) {
 227                 tot_hdr_len += VLAN_HLEN;
 228                 skb->protocol = vlan_eth_hdr(skb)->h_vlan_encapsulated_proto;
 229         }
 230
 231         skb_set_network_header(skb, tot_hdr_len);
 232         mtu -= tot_hdr_len;
 233
 234         if (skb->protocol == htons(ETH_P_IP))
 235                 tot_hdr_len += sizeof(struct iphdr);
 236 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
 237         else if (skb->protocol == htons(ETH_P_IPV6))
 238                 tot_hdr_len += sizeof(struct ipv6hdr);
 239 #endif
 240         else
 241                 goto out;
 242
 243         if (!pskb_may_pull(skb, tot_hdr_len))
 244                 goto out;
 245
 246         if (skb->protocol == htons(ETH_P_IP)) {
 247                 if (mtu < IP_MIN_MTU) {
 248                         if (ntohs(ip_hdr(skb)->tot_len) >= IP_MIN_MTU)
 249                                 mtu = IP_MIN_MTU;
 250                         else
 251                                 goto out;
 252                 }
 253
 254         }
 255 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
 256         else if (skb->protocol == htons(ETH_P_IPV6)) {
 257                 if (mtu < IPV6_MIN_MTU) {
 258                         unsigned int packet_length = sizeof(struct ipv6hdr) +
 259                                               ntohs(ipv6_hdr(skb)->payload_len);
 260
 261                         if (packet_length >= IPV6_MIN_MTU
 262                             || ntohs(ipv6_hdr(skb)->payload_len) == 0)
 263                                 mtu = IPV6_MIN_MTU;
 264                         else
 265                                 goto out;
 266                 }
 267         }
 268 #endif
 269
 270         __skb_pull(skb, tunnel_hdr_len);
 271         tnl_frag_needed(vport, mutable, skb, mtu, key);
 272         __skb_push(skb, tunnel_hdr_len);
 273
 274 out:
 275         skb_set_mac_header(skb, orig_mac_header);
 276         skb_set_network_header(skb, orig_nw_header);
 277         skb->protocol = htons(ETH_P_IP);
 278 }
 279
 280 static bool check_checksum(struct sk_buff *skb)
 281 {
 282         struct iphdr *iph = ip_hdr(skb);
 283         struct gre_base_hdr *greh = (struct gre_base_hdr *)(iph + 1);
 284         __sum16 csum = 0;
 285
 286         if (greh->flags & GRE_CSUM) {
 287                 switch (skb->ip_summed) {
 288                 case CHECKSUM_COMPLETE:
 289                         csum = csum_fold(skb->csum);
 290
 291                         if (!csum)
 292                                 break;
 293                         /* Fall through. */
 294
 295                 case CHECKSUM_NONE:
 296                         skb->csum = 0;
 297                         csum = __skb_checksum_complete(skb);
 298                         skb->ip_summed = CHECKSUM_COMPLETE;
 299                         break;
 300                 }
 301         }
 302
 303         return (csum == 0);
 304 }
 305
 306 /* Called with rcu_read_lock and BH disabled. */
 307 static int gre_rcv(struct sk_buff *skb)
 308 {
 309         struct vport *vport;
 310         const struct tnl_mutable_config *mutable;
 311         int hdr_len;
 312         struct iphdr *iph;
 313         __be16 flags;
 314         __be64 key;
 315
 316         if (unlikely(!pskb_may_pull(skb, sizeof(struct gre_base_hdr) + ETH_HLEN)))
 317                 goto error;
 318
 319         if (unlikely(!check_checksum(skb)))
 320                 goto error;
 321
 322         hdr_len = parse_header(ip_hdr(skb), &flags, &key);
 323         if (unlikely(hdr_len < 0))
 324                 goto error;
 325
 326         if (unlikely(!pskb_may_pull(skb, hdr_len + ETH_HLEN)))
 327                 goto error;
 328
 329         iph = ip_hdr(skb);
 330         vport = tnl_find_port(iph->daddr, iph->saddr, key,
 331                               TNL_T_PROTO_GRE | TNL_T_KEY_EITHER, &mutable);
 332         if (unlikely(!vport)) {
 333                 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_PORT_UNREACH, 0);
 334                 goto error;
 335         }
 336
 337         if (mutable->flags & TNL_F_IN_KEY_MATCH)
 338                 OVS_CB(skb)->tun_id = key;
 339         else
 340                 OVS_CB(skb)->tun_id = 0;
 341
 342         __skb_pull(skb, hdr_len);
 343         skb_postpull_rcsum(skb, skb_transport_header(skb), hdr_len + ETH_HLEN);
 344
 345         tnl_rcv(vport, skb, iph->tos);
 346         return 0;
 347
 348 error:
 349         kfree_skb(skb);
 350         return 0;
 351 }
 352
 353 static const struct tnl_ops gre_tnl_ops = {
 354         .tunnel_type    = TNL_T_PROTO_GRE,
 355         .ipproto        = IPPROTO_GRE,
 356         .hdr_len        = gre_hdr_len,
 357         .build_header   = gre_build_header,
 358         .update_header  = gre_update_header,
 359 };
 360
 361 static struct vport *gre_create(const struct vport_parms *parms)
 362 {
 363         return tnl_create(parms, &gre_vport_ops, &gre_tnl_ops);
 364 }
 365
 366 static const struct net_protocol gre_protocol_handlers = {
 367         .handler        =       gre_rcv,
 368         .err_handler    =       gre_err,
 369 };
 370
 371 static int gre_init(void)
 372 {
 373         int err;
 374
 375         err = inet_add_protocol(&gre_protocol_handlers, IPPROTO_GRE);
 376         if (err)
 377                 pr_warn("cannot register gre protocol handler\n");
 378
 379         return err;
 380 }
 381
 382 static void gre_exit(void)
 383 {
 384         inet_del_protocol(&gre_protocol_handlers, IPPROTO_GRE);
 385 }
 386
 387 const struct vport_ops gre_vport_ops = {
 388         .type           = ODP_VPORT_TYPE_GRE,
 389         .flags          = VPORT_F_GEN_STATS | VPORT_F_TUN_ID,
 390         .init           = gre_init,
 391         .exit           = gre_exit,
 392         .create         = gre_create,
 393         .destroy        = tnl_destroy,
 394         .set_addr       = tnl_set_addr,
 395         .get_name       = tnl_get_name,
 396         .get_addr       = tnl_get_addr,
 397         .get_options    = tnl_get_options,
 398         .set_options    = tnl_set_options,
 399         .get_dev_flags  = vport_gen_get_dev_flags,
 400         .is_running     = vport_gen_is_running,
 401         .get_operstate  = vport_gen_get_operstate,
 402         .send           = tnl_send,
 403 };