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 #include <linux/kconfig.h>
  20 #if IS_ENABLED(CONFIG_NET_IPGRE_DEMUX)
  21 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  22
  23 #include <linux/if.h>
  24 #include <linux/skbuff.h>
  25 #include <linux/ip.h>
  26 #include <linux/if_tunnel.h>
  27 #include <linux/if_vlan.h>
  28 #include <linux/in.h>
  29 #include <linux/in_route.h>
  30 #include <linux/inetdevice.h>
  31 #include <linux/jhash.h>
  32 #include <linux/list.h>
  33 #include <linux/kernel.h>
  34 #include <linux/workqueue.h>
  35 #include <linux/rculist.h>
  36 #include <net/net_namespace.h>
  37 #include <net/netns/generic.h>
  38 #include <net/route.h>
  39 #include <net/xfrm.h>
  40
  41 #include <net/icmp.h>
  42 #include <net/ip.h>
  43 #include <net/ip_tunnels.h>
  44 #include <net/gre.h>
  45 #include <net/protocol.h>
  46
  47 #include "datapath.h"
  48 #include "vport.h"
  49
  50 /* Returns the least-significant 32 bits of a __be64. */
  51 static __be32 be64_get_low32(__be64 x)
  52 {
  53 #ifdef __BIG_ENDIAN
  54         return (__force __be32)x;
  55 #else
  56         return (__force __be32)((__force u64)x >> 32);
  57 #endif
  58 }
  59
  60 static __be16 filter_tnl_flags(__be16 flags)
  61 {
  62         return flags & (TUNNEL_CSUM | TUNNEL_KEY);
  63 }
  64
  65 static struct sk_buff *__build_header(struct sk_buff *skb,
  66                                       int tunnel_hlen,
  67                                       __be32 seq, __be16 gre64_flag)
  68 {
  69         const struct ovs_key_ipv4_tunnel *tun_key;
  70         struct tnl_ptk_info tpi;
  71
  72         tun_key = &OVS_CB(skb)->egress_tun_info->tunnel;
  73         skb = gre_handle_offloads(skb, !!(tun_key->tun_flags & TUNNEL_CSUM));
  74         if (IS_ERR(skb))
  75                 return skb;
  76
  77         tpi.flags = filter_tnl_flags(tun_key->tun_flags) | gre64_flag;
  78
  79         tpi.proto = htons(ETH_P_TEB);
  80         tpi.key = be64_get_low32(tun_key->tun_id);
  81         tpi.seq = seq;
  82         gre_build_header(skb, &tpi, tunnel_hlen);
  83
  84         return skb;
  85 }
  86
  87 static __be64 key_to_tunnel_id(__be32 key, __be32 seq)
  88 {
  89 #ifdef __BIG_ENDIAN
  90         return (__force __be64)((__force u64)seq << 32 | (__force u32)key);
  91 #else
  92         return (__force __be64)((__force u64)key << 32 | (__force u32)seq);
  93 #endif
  94 }
  95
  96 /* Called with rcu_read_lock and BH disabled. */
  97 static int gre_rcv(struct sk_buff *skb,
  98                    const struct tnl_ptk_info *tpi)
  99 {
 100         struct ovs_tunnel_info tun_info;
 101         struct ovs_net *ovs_net;
 102         struct vport *vport;
 103         __be64 key;
 104
 105         ovs_net = net_generic(dev_net(skb->dev), ovs_net_id);
 106         if ((tpi->flags & TUNNEL_KEY) && (tpi->flags & TUNNEL_SEQ))
 107                 vport = rcu_dereference(ovs_net->vport_net.gre64_vport);
 108         else
 109                 vport = rcu_dereference(ovs_net->vport_net.gre_vport);
 110         if (unlikely(!vport))
 111                 return PACKET_REJECT;
 112
 113         key = key_to_tunnel_id(tpi->key, tpi->seq);
 114         ovs_flow_tun_info_init(&tun_info, ip_hdr(skb), 0, 0, key,
 115                                filter_tnl_flags(tpi->flags), NULL, 0);
 116
 117         ovs_vport_receive(vport, skb, &tun_info);
 118         return PACKET_RCVD;
 119 }
 120
 121 /* Called with rcu_read_lock and BH disabled. */
 122 static int gre_err(struct sk_buff *skb, u32 info,
 123                    const struct tnl_ptk_info *tpi)
 124 {
 125         struct ovs_net *ovs_net;
 126         struct vport *vport;
 127
 128         ovs_net = net_generic(dev_net(skb->dev), ovs_net_id);
 129         if ((tpi->flags & TUNNEL_KEY) && (tpi->flags & TUNNEL_SEQ))
 130                 vport = rcu_dereference(ovs_net->vport_net.gre64_vport);
 131         else
 132                 vport = rcu_dereference(ovs_net->vport_net.gre_vport);
 133
 134         if (unlikely(!vport))
 135                 return PACKET_REJECT;
 136         else
 137                 return PACKET_RCVD;
 138 }
 139
 140 static int __send(struct vport *vport, struct sk_buff *skb,
 141                   int tunnel_hlen,
 142                   __be32 seq, __be16 gre64_flag)
 143 {
 144         struct ovs_key_ipv4_tunnel *tun_key;
 145         struct rtable *rt;
 146         int min_headroom;
 147         __be16 df;
 148         __be32 saddr;
 149         int err;
 150
 151         /* Route lookup */
 152         tun_key = &OVS_CB(skb)->egress_tun_info->tunnel;
 153         saddr = tun_key->ipv4_src;
 154         rt = find_route(ovs_dp_get_net(vport->dp),
 155                         &saddr, tun_key->ipv4_dst,
 156                         IPPROTO_GRE, tun_key->ipv4_tos,
 157                         skb->mark);
 158         if (IS_ERR(rt)) {
 159                 err = PTR_ERR(rt);
 160                 goto error;
 161         }
 162
 163         min_headroom = LL_RESERVED_SPACE(rt_dst(rt).dev) + rt_dst(rt).header_len
 164                         + tunnel_hlen + sizeof(struct iphdr)
 165                         + (skb_vlan_tag_present(skb) ? VLAN_HLEN : 0);
 166
 167         if (skb_headroom(skb) < min_headroom || skb_header_cloned(skb)) {
 168                 int head_delta = SKB_DATA_ALIGN(min_headroom -
 169                                                 skb_headroom(skb) +
 170                                                 16);
 171                 err = pskb_expand_head(skb, max_t(int, head_delta, 0),
 172                                         0, GFP_ATOMIC);
 173                 if (unlikely(err))
 174                         goto err_free_rt;
 175         }
 176
 177         if (skb_vlan_tag_present(skb)) {
 178                 if (unlikely(!vlan_insert_tag_set_proto(skb,
 179                                                         skb->vlan_proto,
 180                                                         skb_vlan_tag_get(skb)))) {
 181                         err = -ENOMEM;
 182                         skb = NULL;
 183                         goto err_free_rt;
 184                 }
 185                 vlan_set_tci(skb, 0);
 186         }
 187
 188         /* Push Tunnel header. */
 189         skb = __build_header(skb, tunnel_hlen, seq, gre64_flag);
 190         if (IS_ERR(skb)) {
 191                 err = PTR_ERR(skb);
 192                 skb = NULL;
 193                 goto err_free_rt;
 194         }
 195
 196         df = tun_key->tun_flags & TUNNEL_DONT_FRAGMENT ? htons(IP_DF) : 0;
 197         skb->ignore_df = 1;
 198
 199         return iptunnel_xmit(skb->sk, rt, skb, saddr,
 200                              tun_key->ipv4_dst, IPPROTO_GRE,
 201                              tun_key->ipv4_tos,
 202                              tun_key->ipv4_ttl, df, false);
 203 err_free_rt:
 204         ip_rt_put(rt);
 205 error:
 206         kfree_skb(skb);
 207         return err;
 208 }
 209
 210 static struct gre_cisco_protocol gre_protocol = {
 211         .handler        = gre_rcv,
 212         .err_handler    = gre_err,
 213         .priority       = 1,
 214 };
 215
 216 static int gre_ports;
 217 static int gre_init(void)
 218 {
 219         int err;
 220
 221         gre_ports++;
 222         if (gre_ports > 1)
 223                 return 0;
 224
 225         err = gre_cisco_register(&gre_protocol);
 226         if (err)
 227                 pr_warn("cannot register gre protocol handler\n");
 228
 229         return err;
 230 }
 231
 232 static void gre_exit(void)
 233 {
 234         gre_ports--;
 235         if (gre_ports > 0)
 236                 return;
 237
 238         gre_cisco_unregister(&gre_protocol);
 239 }
 240
 241 static const char *gre_get_name(const struct vport *vport)
 242 {
 243         return vport_priv(vport);
 244 }
 245
 246 static struct vport *gre_create(const struct vport_parms *parms)
 247 {
 248         struct net *net = ovs_dp_get_net(parms->dp);
 249         struct ovs_net *ovs_net;
 250         struct vport *vport;
 251         int err;
 252
 253         err = gre_init();
 254         if (err)
 255                 return ERR_PTR(err);
 256
 257         ovs_net = net_generic(net, ovs_net_id);
 258         if (ovsl_dereference(ovs_net->vport_net.gre_vport)) {
 259                 vport = ERR_PTR(-EEXIST);
 260                 goto error;
 261         }
 262
 263         vport = ovs_vport_alloc(IFNAMSIZ, &ovs_gre_vport_ops, parms);
 264         if (IS_ERR(vport))
 265                 goto error;
 266
 267         strncpy(vport_priv(vport), parms->name, IFNAMSIZ);
 268         rcu_assign_pointer(ovs_net->vport_net.gre_vport, vport);
 269         return vport;
 270
 271 error:
 272         gre_exit();
 273         return vport;
 274 }
 275
 276 static void gre_tnl_destroy(struct vport *vport)
 277 {
 278         struct net *net = ovs_dp_get_net(vport->dp);
 279         struct ovs_net *ovs_net;
 280
 281         ovs_net = net_generic(net, ovs_net_id);
 282
 283         RCU_INIT_POINTER(ovs_net->vport_net.gre_vport, NULL);
 284         ovs_vport_deferred_free(vport);
 285         gre_exit();
 286 }
 287
 288 static int gre_send(struct vport *vport, struct sk_buff *skb)
 289 {
 290         int hlen;
 291
 292         if (unlikely(!OVS_CB(skb)->egress_tun_info)) {
 293                 kfree_skb(skb);
 294                 return -EINVAL;
 295         }
 296
 297         hlen = ip_gre_calc_hlen(OVS_CB(skb)->egress_tun_info->tunnel.tun_flags);
 298
 299         return __send(vport, skb, hlen, 0, 0);
 300 }
 301
 302 static int gre_get_egress_tun_info(struct vport *vport, struct sk_buff *skb,
 303                                    struct ovs_tunnel_info *egress_tun_info)
 304 {
 305         return ovs_tunnel_get_egress_info(egress_tun_info,
 306                                           ovs_dp_get_net(vport->dp),
 307                                           OVS_CB(skb)->egress_tun_info,
 308                                           IPPROTO_GRE, skb->mark, 0, 0);
 309 }
 310
 311 const struct vport_ops ovs_gre_vport_ops = {
 312         .type                   = OVS_VPORT_TYPE_GRE,
 313         .create                 = gre_create,
 314         .destroy                = gre_tnl_destroy,
 315         .get_name               = gre_get_name,
 316         .send                   = gre_send,
 317         .get_egress_tun_info    = gre_get_egress_tun_info,
 318 };
 319
 320 /* GRE64 vport. */
 321 static struct vport *gre64_create(const struct vport_parms *parms)
 322 {
 323         struct net *net = ovs_dp_get_net(parms->dp);
 324         struct ovs_net *ovs_net;
 325         struct vport *vport;
 326         int err;
 327
 328         err = gre_init();
 329         if (err)
 330                 return ERR_PTR(err);
 331
 332         ovs_net = net_generic(net, ovs_net_id);
 333         if (ovsl_dereference(ovs_net->vport_net.gre64_vport)) {
 334                 vport = ERR_PTR(-EEXIST);
 335                 goto error;
 336         }
 337
 338         vport = ovs_vport_alloc(IFNAMSIZ, &ovs_gre64_vport_ops, parms);
 339         if (IS_ERR(vport))
 340                 goto error;
 341
 342         strncpy(vport_priv(vport), parms->name, IFNAMSIZ);
 343         rcu_assign_pointer(ovs_net->vport_net.gre64_vport, vport);
 344         return vport;
 345 error:
 346         gre_exit();
 347         return vport;
 348 }
 349
 350 static void gre64_tnl_destroy(struct vport *vport)
 351 {
 352         struct net *net = ovs_dp_get_net(vport->dp);
 353         struct ovs_net *ovs_net;
 354
 355         ovs_net = net_generic(net, ovs_net_id);
 356
 357         rcu_assign_pointer(ovs_net->vport_net.gre64_vport, NULL);
 358         ovs_vport_deferred_free(vport);
 359         gre_exit();
 360 }
 361
 362 static __be32 be64_get_high32(__be64 x)
 363 {
 364 #ifdef __BIG_ENDIAN
 365         return (__force __be32)((__force u64)x >> 32);
 366 #else
 367         return (__force __be32)x;
 368 #endif
 369 }
 370
 371 static int gre64_send(struct vport *vport, struct sk_buff *skb)
 372 {
 373         int hlen = GRE_HEADER_SECTION +         /* GRE Hdr */
 374                    GRE_HEADER_SECTION +         /* GRE Key */
 375                    GRE_HEADER_SECTION;          /* GRE SEQ */
 376         __be32 seq;
 377
 378         if (unlikely(!OVS_CB(skb)->egress_tun_info)) {
 379                 kfree_skb(skb);
 380                 return -EINVAL;
 381         }
 382
 383         if (OVS_CB(skb)->egress_tun_info->tunnel.tun_flags & TUNNEL_CSUM)
 384                 hlen += GRE_HEADER_SECTION;
 385
 386         seq = be64_get_high32(OVS_CB(skb)->egress_tun_info->tunnel.tun_id);
 387         return __send(vport, skb, hlen, seq, (TUNNEL_KEY|TUNNEL_SEQ));
 388 }
 389
 390 const struct vport_ops ovs_gre64_vport_ops = {
 391         .type                   = OVS_VPORT_TYPE_GRE64,
 392         .create                 = gre64_create,
 393         .destroy                = gre64_tnl_destroy,
 394         .get_name               = gre_get_name,
 395         .send                   = gre64_send,
 396         .get_egress_tun_info    = gre_get_egress_tun_info,
 397 };
 398 #endif