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/module.h>
  35 #include <linux/workqueue.h>
  36 #include <linux/rculist.h>
  37 #include <net/net_namespace.h>
  38 #include <net/netns/generic.h>
  39 #include <net/route.h>
  40 #include <net/xfrm.h>
  41
  42 #include <net/icmp.h>
  43 #include <net/ip.h>
  44 #include <net/ip_tunnels.h>
  45 #include <net/gre.h>
  46 #include <net/protocol.h>
  47
  48 #include "datapath.h"
  49 #include "vport.h"
  50
  51 static struct vport_ops ovs_gre_vport_ops;
  52 static struct vport_ops ovs_gre64_vport_ops;
  53
  54 /* Returns the least-significant 32 bits of a __be64. */
  55 static __be32 be64_get_low32(__be64 x)
  56 {
  57 #ifdef __BIG_ENDIAN
  58         return (__force __be32)x;
  59 #else
  60         return (__force __be32)((__force u64)x >> 32);
  61 #endif
  62 }
  63
  64 static __be16 filter_tnl_flags(__be16 flags)
  65 {
  66         return flags & (TUNNEL_CSUM | TUNNEL_KEY);
  67 }
  68
  69 static struct sk_buff *__build_header(struct sk_buff *skb,
  70                                       int tunnel_hlen,
  71                                       __be32 seq, __be16 gre64_flag)
  72 {
  73         const struct ovs_key_ipv4_tunnel *tun_key;
  74         struct tnl_ptk_info tpi;
  75
  76         tun_key = &OVS_CB(skb)->egress_tun_info->tunnel;
  77         skb = gre_handle_offloads(skb, !!(tun_key->tun_flags & TUNNEL_CSUM));
  78         if (IS_ERR(skb))
  79                 return skb;
  80
  81         tpi.flags = filter_tnl_flags(tun_key->tun_flags) | gre64_flag;
  82
  83         tpi.proto = htons(ETH_P_TEB);
  84         tpi.key = be64_get_low32(tun_key->tun_id);
  85         tpi.seq = seq;
  86         gre_build_header(skb, &tpi, tunnel_hlen);
  87
  88         return skb;
  89 }
  90
  91 static __be64 key_to_tunnel_id(__be32 key, __be32 seq)
  92 {
  93 #ifdef __BIG_ENDIAN
  94         return (__force __be64)((__force u64)seq << 32 | (__force u32)key);
  95 #else
  96         return (__force __be64)((__force u64)key << 32 | (__force u32)seq);
  97 #endif
  98 }
  99
 100 /* Called with rcu_read_lock and BH disabled. */
 101 static int gre_rcv(struct sk_buff *skb,
 102                    const struct tnl_ptk_info *tpi)
 103 {
 104         struct ovs_tunnel_info tun_info;
 105         struct ovs_net *ovs_net;
 106         struct vport *vport;
 107         __be64 key;
 108
 109         ovs_net = net_generic(dev_net(skb->dev), ovs_net_id);
 110         if ((tpi->flags & TUNNEL_KEY) && (tpi->flags & TUNNEL_SEQ))
 111                 vport = rcu_dereference(ovs_net->vport_net.gre64_vport);
 112         else
 113                 vport = rcu_dereference(ovs_net->vport_net.gre_vport);
 114         if (unlikely(!vport))
 115                 return PACKET_REJECT;
 116
 117         key = key_to_tunnel_id(tpi->key, tpi->seq);
 118         ovs_flow_tun_info_init(&tun_info, ip_hdr(skb), 0, 0, key,
 119                                filter_tnl_flags(tpi->flags), NULL, 0);
 120
 121         ovs_vport_receive(vport, skb, &tun_info);
 122         return PACKET_RCVD;
 123 }
 124
 125 /* Called with rcu_read_lock and BH disabled. */
 126 static int gre_err(struct sk_buff *skb, u32 info,
 127                    const struct tnl_ptk_info *tpi)
 128 {
 129         struct ovs_net *ovs_net;
 130         struct vport *vport;
 131
 132         ovs_net = net_generic(dev_net(skb->dev), ovs_net_id);
 133         if ((tpi->flags & TUNNEL_KEY) && (tpi->flags & TUNNEL_SEQ))
 134                 vport = rcu_dereference(ovs_net->vport_net.gre64_vport);
 135         else
 136                 vport = rcu_dereference(ovs_net->vport_net.gre_vport);
 137
 138         if (unlikely(!vport))
 139                 return PACKET_REJECT;
 140         else
 141                 return PACKET_RCVD;
 142 }
 143
 144 static int __send(struct vport *vport, struct sk_buff *skb,
 145                   int tunnel_hlen,
 146                   __be32 seq, __be16 gre64_flag)
 147 {
 148         struct ovs_key_ipv4_tunnel *tun_key;
 149         struct rtable *rt;
 150         int min_headroom;
 151         __be16 df;
 152         __be32 saddr;
 153         int err;
 154
 155         /* Route lookup */
 156         tun_key = &OVS_CB(skb)->egress_tun_info->tunnel;
 157         saddr = tun_key->ipv4_src;
 158         rt = find_route(ovs_dp_get_net(vport->dp),
 159                         &saddr, tun_key->ipv4_dst,
 160                         IPPROTO_GRE, tun_key->ipv4_tos,
 161                         skb->mark);
 162         if (IS_ERR(rt)) {
 163                 err = PTR_ERR(rt);
 164                 goto error;
 165         }
 166
 167         min_headroom = LL_RESERVED_SPACE(rt_dst(rt).dev) + rt_dst(rt).header_len
 168                         + tunnel_hlen + sizeof(struct iphdr)
 169                         + (skb_vlan_tag_present(skb) ? VLAN_HLEN : 0);
 170
 171         if (skb_headroom(skb) < min_headroom || skb_header_cloned(skb)) {
 172                 int head_delta = SKB_DATA_ALIGN(min_headroom -
 173                                                 skb_headroom(skb) +
 174                                                 16);
 175                 err = pskb_expand_head(skb, max_t(int, head_delta, 0),
 176                                         0, GFP_ATOMIC);
 177                 if (unlikely(err))
 178                         goto err_free_rt;
 179         }
 180
 181         if (skb_vlan_tag_present(skb)) {
 182                 if (unlikely(!vlan_insert_tag_set_proto(skb,
 183                                                         skb->vlan_proto,
 184                                                         skb_vlan_tag_get(skb)))) {
 185                         err = -ENOMEM;
 186                         skb = NULL;
 187                         goto err_free_rt;
 188                 }
 189                 vlan_set_tci(skb, 0);
 190         }
 191
 192         /* Push Tunnel header. */
 193         skb = __build_header(skb, tunnel_hlen, seq, gre64_flag);
 194         if (IS_ERR(skb)) {
 195                 err = PTR_ERR(skb);
 196                 skb = NULL;
 197                 goto err_free_rt;
 198         }
 199
 200         df = tun_key->tun_flags & TUNNEL_DONT_FRAGMENT ? htons(IP_DF) : 0;
 201         skb->ignore_df = 1;
 202
 203         return iptunnel_xmit(skb->sk, rt, skb, saddr,
 204                              tun_key->ipv4_dst, IPPROTO_GRE,
 205                              tun_key->ipv4_tos,
 206                              tun_key->ipv4_ttl, df, false);
 207 err_free_rt:
 208         ip_rt_put(rt);
 209 error:
 210         kfree_skb(skb);
 211         return err;
 212 }
 213
 214 static struct gre_cisco_protocol gre_protocol = {
 215         .handler        = gre_rcv,
 216         .err_handler    = gre_err,
 217         .priority       = 1,
 218 };
 219
 220 static int gre_ports;
 221 static int gre_init(void)
 222 {
 223         int err;
 224
 225         gre_ports++;
 226         if (gre_ports > 1)
 227                 return 0;
 228
 229         err = gre_cisco_register(&gre_protocol);
 230         if (err)
 231                 pr_warn("cannot register gre protocol handler\n");
 232
 233         return err;
 234 }
 235
 236 static void gre_exit(void)
 237 {
 238         gre_ports--;
 239         if (gre_ports > 0)
 240                 return;
 241
 242         gre_cisco_unregister(&gre_protocol);
 243 }
 244
 245 static const char *gre_get_name(const struct vport *vport)
 246 {
 247         return vport_priv(vport);
 248 }
 249
 250 static struct vport *gre_create(const struct vport_parms *parms)
 251 {
 252         struct net *net = ovs_dp_get_net(parms->dp);
 253         struct ovs_net *ovs_net;
 254         struct vport *vport;
 255         int err;
 256
 257         err = gre_init();
 258         if (err)
 259                 return ERR_PTR(err);
 260
 261         ovs_net = net_generic(net, ovs_net_id);
 262         if (ovsl_dereference(ovs_net->vport_net.gre_vport)) {
 263                 vport = ERR_PTR(-EEXIST);
 264                 goto error;
 265         }
 266
 267         vport = ovs_vport_alloc(IFNAMSIZ, &ovs_gre_vport_ops, parms);
 268         if (IS_ERR(vport))
 269                 goto error;
 270
 271         strncpy(vport_priv(vport), parms->name, IFNAMSIZ);
 272         rcu_assign_pointer(ovs_net->vport_net.gre_vport, vport);
 273         return vport;
 274
 275 error:
 276         gre_exit();
 277         return vport;
 278 }
 279
 280 static void gre_tnl_destroy(struct vport *vport)
 281 {
 282         struct net *net = ovs_dp_get_net(vport->dp);
 283         struct ovs_net *ovs_net;
 284
 285         ovs_net = net_generic(net, ovs_net_id);
 286
 287         RCU_INIT_POINTER(ovs_net->vport_net.gre_vport, NULL);
 288         ovs_vport_deferred_free(vport);
 289         gre_exit();
 290 }
 291
 292 static int gre_send(struct vport *vport, struct sk_buff *skb)
 293 {
 294         int hlen;
 295
 296         if (unlikely(!OVS_CB(skb)->egress_tun_info)) {
 297                 kfree_skb(skb);
 298                 return -EINVAL;
 299         }
 300
 301         hlen = ip_gre_calc_hlen(OVS_CB(skb)->egress_tun_info->tunnel.tun_flags);
 302
 303         return __send(vport, skb, hlen, 0, 0);
 304 }
 305
 306 static int gre_get_egress_tun_info(struct vport *vport, struct sk_buff *skb,
 307                                    struct ovs_tunnel_info *egress_tun_info)
 308 {
 309         return ovs_tunnel_get_egress_info(egress_tun_info,
 310                                           ovs_dp_get_net(vport->dp),
 311                                           OVS_CB(skb)->egress_tun_info,
 312                                           IPPROTO_GRE, skb->mark, 0, 0);
 313 }
 314
 315 static struct vport_ops ovs_gre_vport_ops = {
 316         .type                   = OVS_VPORT_TYPE_GRE,
 317         .create                 = gre_create,
 318         .destroy                = gre_tnl_destroy,
 319         .get_name               = gre_get_name,
 320         .send                   = gre_send,
 321         .get_egress_tun_info    = gre_get_egress_tun_info,
 322         .owner                  = THIS_MODULE,
 323 };
 324
 325 /* GRE64 vport. */
 326 static struct vport *gre64_create(const struct vport_parms *parms)
 327 {
 328         struct net *net = ovs_dp_get_net(parms->dp);
 329         struct ovs_net *ovs_net;
 330         struct vport *vport;
 331         int err;
 332
 333         pr_warn_once("GRE64 tunnel protocol is deprecated.");
 334         err = gre_init();
 335         if (err)
 336                 return ERR_PTR(err);
 337
 338         ovs_net = net_generic(net, ovs_net_id);
 339         if (ovsl_dereference(ovs_net->vport_net.gre64_vport)) {
 340                 vport = ERR_PTR(-EEXIST);
 341                 goto error;
 342         }
 343
 344         vport = ovs_vport_alloc(IFNAMSIZ, &ovs_gre64_vport_ops, parms);
 345         if (IS_ERR(vport))
 346                 goto error;
 347
 348         strncpy(vport_priv(vport), parms->name, IFNAMSIZ);
 349         rcu_assign_pointer(ovs_net->vport_net.gre64_vport, vport);
 350         return vport;
 351 error:
 352         gre_exit();
 353         return vport;
 354 }
 355
 356 static void gre64_tnl_destroy(struct vport *vport)
 357 {
 358         struct net *net = ovs_dp_get_net(vport->dp);
 359         struct ovs_net *ovs_net;
 360
 361         ovs_net = net_generic(net, ovs_net_id);
 362
 363         rcu_assign_pointer(ovs_net->vport_net.gre64_vport, NULL);
 364         ovs_vport_deferred_free(vport);
 365         gre_exit();
 366 }
 367
 368 static __be32 be64_get_high32(__be64 x)
 369 {
 370 #ifdef __BIG_ENDIAN
 371         return (__force __be32)((__force u64)x >> 32);
 372 #else
 373         return (__force __be32)x;
 374 #endif
 375 }
 376
 377 static int gre64_send(struct vport *vport, struct sk_buff *skb)
 378 {
 379         int hlen = GRE_HEADER_SECTION +         /* GRE Hdr */
 380                    GRE_HEADER_SECTION +         /* GRE Key */
 381                    GRE_HEADER_SECTION;          /* GRE SEQ */
 382         __be32 seq;
 383
 384         if (unlikely(!OVS_CB(skb)->egress_tun_info)) {
 385                 kfree_skb(skb);
 386                 return -EINVAL;
 387         }
 388
 389         if (OVS_CB(skb)->egress_tun_info->tunnel.tun_flags & TUNNEL_CSUM)
 390                 hlen += GRE_HEADER_SECTION;
 391
 392         seq = be64_get_high32(OVS_CB(skb)->egress_tun_info->tunnel.tun_id);
 393         return __send(vport, skb, hlen, seq, (TUNNEL_KEY|TUNNEL_SEQ));
 394 }
 395
 396 static struct vport_ops ovs_gre64_vport_ops = {
 397         .type                   = OVS_VPORT_TYPE_GRE64,
 398         .create                 = gre64_create,
 399         .destroy                = gre64_tnl_destroy,
 400         .get_name               = gre_get_name,
 401         .send                   = gre64_send,
 402         .get_egress_tun_info    = gre_get_egress_tun_info,
 403         .owner                  = THIS_MODULE,
 404 };
 405
 406 static int __init ovs_gre_tnl_init(void)
 407 {
 408         int err;
 409
 410         err = ovs_vport_ops_register(&ovs_gre_vport_ops);
 411         if (err < 0)
 412                 return err;
 413
 414         err = ovs_vport_ops_register(&ovs_gre64_vport_ops);
 415         if (err < 0)
 416                 ovs_vport_ops_unregister(&ovs_gre_vport_ops);
 417
 418         return err;
 419 }
 420
 421 static void __exit ovs_gre_tnl_exit(void)
 422 {
 423         ovs_vport_ops_unregister(&ovs_gre64_vport_ops);
 424         ovs_vport_ops_unregister(&ovs_gre_vport_ops);
 425 }
 426
 427 module_init(ovs_gre_tnl_init);
 428 module_exit(ovs_gre_tnl_exit);
 429
 430 MODULE_DESCRIPTION("OVS: GRE switching port");
 431 MODULE_LICENSE("GPL");
 432 MODULE_ALIAS("vport-type-3");
 433 MODULE_ALIAS("vport-type-104");
 434 #endif