drivers/net/ethernet/mellanox/mlxsw/spectrum_ipip.c

   1 // SPDX-License-Identifier: BSD-3-Clause OR GPL-2.0
   2 /* Copyright (c) 2017-2018 Mellanox Technologies. All rights reserved */
   3
   4 #include <net/ip_tunnels.h>
   5 #include <net/ip6_tunnel.h>
   6
   7 #include "spectrum_ipip.h"
   8
   9 struct ip_tunnel_parm
  10 mlxsw_sp_ipip_netdev_parms4(const struct net_device *ol_dev)
  11 {
  12         struct ip_tunnel *tun = netdev_priv(ol_dev);
  13
  14         return tun->parms;
  15 }
  16
  17 struct __ip6_tnl_parm
  18 mlxsw_sp_ipip_netdev_parms6(const struct net_device *ol_dev)
  19 {
  20         struct ip6_tnl *tun = netdev_priv(ol_dev);
  21
  22         return tun->parms;
  23 }
  24
  25 static bool mlxsw_sp_ipip_parms4_has_ikey(struct ip_tunnel_parm parms)
  26 {
  27         return !!(parms.i_flags & TUNNEL_KEY);
  28 }
  29
  30 static bool mlxsw_sp_ipip_parms4_has_okey(struct ip_tunnel_parm parms)
  31 {
  32         return !!(parms.o_flags & TUNNEL_KEY);
  33 }
  34
  35 static u32 mlxsw_sp_ipip_parms4_ikey(struct ip_tunnel_parm parms)
  36 {
  37         return mlxsw_sp_ipip_parms4_has_ikey(parms) ?
  38                 be32_to_cpu(parms.i_key) : 0;
  39 }
  40
  41 static u32 mlxsw_sp_ipip_parms4_okey(struct ip_tunnel_parm parms)
  42 {
  43         return mlxsw_sp_ipip_parms4_has_okey(parms) ?
  44                 be32_to_cpu(parms.o_key) : 0;
  45 }
  46
  47 static union mlxsw_sp_l3addr
  48 mlxsw_sp_ipip_parms4_saddr(struct ip_tunnel_parm parms)
  49 {
  50         return (union mlxsw_sp_l3addr) { .addr4 = parms.iph.saddr };
  51 }
  52
  53 static union mlxsw_sp_l3addr
  54 mlxsw_sp_ipip_parms6_saddr(struct __ip6_tnl_parm parms)
  55 {
  56         return (union mlxsw_sp_l3addr) { .addr6 = parms.laddr };
  57 }
  58
  59 static union mlxsw_sp_l3addr
  60 mlxsw_sp_ipip_parms4_daddr(struct ip_tunnel_parm parms)
  61 {
  62         return (union mlxsw_sp_l3addr) { .addr4 = parms.iph.daddr };
  63 }
  64
  65 static union mlxsw_sp_l3addr
  66 mlxsw_sp_ipip_parms6_daddr(struct __ip6_tnl_parm parms)
  67 {
  68         return (union mlxsw_sp_l3addr) { .addr6 = parms.raddr };
  69 }
  70
  71 union mlxsw_sp_l3addr
  72 mlxsw_sp_ipip_netdev_saddr(enum mlxsw_sp_l3proto proto,
  73                            const struct net_device *ol_dev)
  74 {
  75         struct ip_tunnel_parm parms4;
  76         struct __ip6_tnl_parm parms6;
  77
  78         switch (proto) {
  79         case MLXSW_SP_L3_PROTO_IPV4:
  80                 parms4 = mlxsw_sp_ipip_netdev_parms4(ol_dev);
  81                 return mlxsw_sp_ipip_parms4_saddr(parms4);
  82         case MLXSW_SP_L3_PROTO_IPV6:
  83                 parms6 = mlxsw_sp_ipip_netdev_parms6(ol_dev);
  84                 return mlxsw_sp_ipip_parms6_saddr(parms6);
  85         }
  86
  87         WARN_ON(1);
  88         return (union mlxsw_sp_l3addr) {0};
  89 }
  90
  91 static __be32 mlxsw_sp_ipip_netdev_daddr4(const struct net_device *ol_dev)
  92 {
  93
  94         struct ip_tunnel_parm parms4 = mlxsw_sp_ipip_netdev_parms4(ol_dev);
  95
  96         return mlxsw_sp_ipip_parms4_daddr(parms4).addr4;
  97 }
  98
  99 static union mlxsw_sp_l3addr
 100 mlxsw_sp_ipip_netdev_daddr(enum mlxsw_sp_l3proto proto,
 101                            const struct net_device *ol_dev)
 102 {
 103         struct ip_tunnel_parm parms4;
 104         struct __ip6_tnl_parm parms6;
 105
 106         switch (proto) {
 107         case MLXSW_SP_L3_PROTO_IPV4:
 108                 parms4 = mlxsw_sp_ipip_netdev_parms4(ol_dev);
 109                 return mlxsw_sp_ipip_parms4_daddr(parms4);
 110         case MLXSW_SP_L3_PROTO_IPV6:
 111                 parms6 = mlxsw_sp_ipip_netdev_parms6(ol_dev);
 112                 return mlxsw_sp_ipip_parms6_daddr(parms6);
 113         }
 114
 115         WARN_ON(1);
 116         return (union mlxsw_sp_l3addr) {0};
 117 }
 118
 119 bool mlxsw_sp_l3addr_is_zero(union mlxsw_sp_l3addr addr)
 120 {
 121         union mlxsw_sp_l3addr naddr = {0};
 122
 123         return !memcmp(&addr, &naddr, sizeof(naddr));
 124 }
 125
 126 static int
 127 mlxsw_sp_ipip_nexthop_update_gre4(struct mlxsw_sp *mlxsw_sp, u32 adj_index,
 128                                   struct mlxsw_sp_ipip_entry *ipip_entry)
 129 {
 130         u16 rif_index = mlxsw_sp_ipip_lb_rif_index(ipip_entry->ol_lb);
 131         __be32 daddr4 = mlxsw_sp_ipip_netdev_daddr4(ipip_entry->ol_dev);
 132         char ratr_pl[MLXSW_REG_RATR_LEN];
 133
 134         mlxsw_reg_ratr_pack(ratr_pl, MLXSW_REG_RATR_OP_WRITE_WRITE_ENTRY,
 135                             true, MLXSW_REG_RATR_TYPE_IPIP,
 136                             adj_index, rif_index);
 137         mlxsw_reg_ratr_ipip4_entry_pack(ratr_pl, be32_to_cpu(daddr4));
 138
 139         return mlxsw_reg_write(mlxsw_sp->core, MLXSW_REG(ratr), ratr_pl);
 140 }
 141
 142 static int
 143 mlxsw_sp_ipip_fib_entry_op_gre4_rtdp(struct mlxsw_sp *mlxsw_sp,
 144                                      u32 tunnel_index,
 145                                      struct mlxsw_sp_ipip_entry *ipip_entry)
 146 {
 147         u16 rif_index = mlxsw_sp_ipip_lb_rif_index(ipip_entry->ol_lb);
 148         char rtdp_pl[MLXSW_REG_RTDP_LEN];
 149         struct ip_tunnel_parm parms;
 150         unsigned int type_check;
 151         bool has_ikey;
 152         u32 daddr4;
 153         u32 ikey;
 154
 155         parms = mlxsw_sp_ipip_netdev_parms4(ipip_entry->ol_dev);
 156         has_ikey = mlxsw_sp_ipip_parms4_has_ikey(parms);
 157         ikey = mlxsw_sp_ipip_parms4_ikey(parms);
 158
 159         mlxsw_reg_rtdp_pack(rtdp_pl, MLXSW_REG_RTDP_TYPE_IPIP, tunnel_index);
 160
 161         type_check = has_ikey ?
 162                 MLXSW_REG_RTDP_IPIP_TYPE_CHECK_ALLOW_GRE_KEY :
 163                 MLXSW_REG_RTDP_IPIP_TYPE_CHECK_ALLOW_GRE;
 164
 165         /* Linux demuxes tunnels based on packet SIP (which must match tunnel
 166          * remote IP). Thus configure decap so that it filters out packets that
 167          * are not IPv4 or have the wrong SIP. IPIP_DECAP_ERROR trap is
 168          * generated for packets that fail this criterion. Linux then handles
 169          * such packets in slow path and generates ICMP destination unreachable.
 170          */
 171         daddr4 = be32_to_cpu(mlxsw_sp_ipip_netdev_daddr4(ipip_entry->ol_dev));
 172         mlxsw_reg_rtdp_ipip4_pack(rtdp_pl, rif_index,
 173                                   MLXSW_REG_RTDP_IPIP_SIP_CHECK_FILTER_IPV4,
 174                                   type_check, has_ikey, daddr4, ikey);
 175
 176         return mlxsw_reg_write(mlxsw_sp->core, MLXSW_REG(rtdp), rtdp_pl);
 177 }
 178
 179 static int
 180 mlxsw_sp_ipip_fib_entry_op_gre4_ralue(struct mlxsw_sp *mlxsw_sp,
 181                                       u32 dip, u8 prefix_len, u16 ul_vr_id,
 182                                       enum mlxsw_reg_ralue_op op,
 183                                       u32 tunnel_index)
 184 {
 185         char ralue_pl[MLXSW_REG_RALUE_LEN];
 186
 187         mlxsw_reg_ralue_pack4(ralue_pl, MLXSW_REG_RALXX_PROTOCOL_IPV4, op,
 188                               ul_vr_id, prefix_len, dip);
 189         mlxsw_reg_ralue_act_ip2me_tun_pack(ralue_pl, tunnel_index);
 190         return mlxsw_reg_write(mlxsw_sp->core, MLXSW_REG(ralue), ralue_pl);
 191 }
 192
 193 static int mlxsw_sp_ipip_fib_entry_op_gre4(struct mlxsw_sp *mlxsw_sp,
 194                                         struct mlxsw_sp_ipip_entry *ipip_entry,
 195                                         enum mlxsw_reg_ralue_op op,
 196                                         u32 tunnel_index)
 197 {
 198         u16 ul_vr_id = mlxsw_sp_ipip_lb_ul_vr_id(ipip_entry->ol_lb);
 199         __be32 dip;
 200         int err;
 201
 202         err = mlxsw_sp_ipip_fib_entry_op_gre4_rtdp(mlxsw_sp, tunnel_index,
 203                                                    ipip_entry);
 204         if (err)
 205                 return err;
 206
 207         dip = mlxsw_sp_ipip_netdev_saddr(MLXSW_SP_L3_PROTO_IPV4,
 208                                          ipip_entry->ol_dev).addr4;
 209         return mlxsw_sp_ipip_fib_entry_op_gre4_ralue(mlxsw_sp, be32_to_cpu(dip),
 210                                                      32, ul_vr_id, op,
 211                                                      tunnel_index);
 212 }
 213
 214 static bool mlxsw_sp_ipip_tunnel_complete(enum mlxsw_sp_l3proto proto,
 215                                           const struct net_device *ol_dev)
 216 {
 217         union mlxsw_sp_l3addr saddr = mlxsw_sp_ipip_netdev_saddr(proto, ol_dev);
 218         union mlxsw_sp_l3addr daddr = mlxsw_sp_ipip_netdev_daddr(proto, ol_dev);
 219
 220         /* Tunnels with unset local or remote address are valid in Linux and
 221          * used for lightweight tunnels (LWT) and Non-Broadcast Multi-Access
 222          * (NBMA) tunnels. In principle these can be offloaded, but the driver
 223          * currently doesn't support this. So punt.
 224          */
 225         return !mlxsw_sp_l3addr_is_zero(saddr) &&
 226                !mlxsw_sp_l3addr_is_zero(daddr);
 227 }
 228
 229 static bool mlxsw_sp_ipip_can_offload_gre4(const struct mlxsw_sp *mlxsw_sp,
 230                                            const struct net_device *ol_dev,
 231                                            enum mlxsw_sp_l3proto ol_proto)
 232 {
 233         struct ip_tunnel *tunnel = netdev_priv(ol_dev);
 234         __be16 okflags = TUNNEL_KEY; /* We can't offload any other features. */
 235         bool inherit_ttl = tunnel->parms.iph.ttl == 0;
 236         bool inherit_tos = tunnel->parms.iph.tos & 0x1;
 237
 238         return (tunnel->parms.i_flags & ~okflags) == 0 &&
 239                (tunnel->parms.o_flags & ~okflags) == 0 &&
 240                inherit_ttl && inherit_tos &&
 241                mlxsw_sp_ipip_tunnel_complete(MLXSW_SP_L3_PROTO_IPV4, ol_dev);
 242 }
 243
 244 static struct mlxsw_sp_rif_ipip_lb_config
 245 mlxsw_sp_ipip_ol_loopback_config_gre4(struct mlxsw_sp *mlxsw_sp,
 246                                       const struct net_device *ol_dev)
 247 {
 248         struct ip_tunnel_parm parms = mlxsw_sp_ipip_netdev_parms4(ol_dev);
 249         enum mlxsw_reg_ritr_loopback_ipip_type lb_ipipt;
 250
 251         lb_ipipt = mlxsw_sp_ipip_parms4_has_okey(parms) ?
 252                 MLXSW_REG_RITR_LOOPBACK_IPIP_TYPE_IP_IN_GRE_KEY_IN_IP :
 253                 MLXSW_REG_RITR_LOOPBACK_IPIP_TYPE_IP_IN_GRE_IN_IP;
 254         return (struct mlxsw_sp_rif_ipip_lb_config){
 255                 .lb_ipipt = lb_ipipt,
 256                 .okey = mlxsw_sp_ipip_parms4_okey(parms),
 257                 .ul_protocol = MLXSW_SP_L3_PROTO_IPV4,
 258                 .saddr = mlxsw_sp_ipip_netdev_saddr(MLXSW_SP_L3_PROTO_IPV4,
 259                                                     ol_dev),
 260         };
 261 }
 262
 263 static int
 264 mlxsw_sp_ipip_ol_netdev_change_gre4(struct mlxsw_sp *mlxsw_sp,
 265                                     struct mlxsw_sp_ipip_entry *ipip_entry,
 266                                     struct netlink_ext_ack *extack)
 267 {
 268         union mlxsw_sp_l3addr old_saddr, new_saddr;
 269         union mlxsw_sp_l3addr old_daddr, new_daddr;
 270         struct ip_tunnel_parm new_parms;
 271         bool update_tunnel = false;
 272         bool update_decap = false;
 273         bool update_nhs = false;
 274         int err = 0;
 275
 276         new_parms = mlxsw_sp_ipip_netdev_parms4(ipip_entry->ol_dev);
 277
 278         new_saddr = mlxsw_sp_ipip_parms4_saddr(new_parms);
 279         old_saddr = mlxsw_sp_ipip_parms4_saddr(ipip_entry->parms4);
 280         new_daddr = mlxsw_sp_ipip_parms4_daddr(new_parms);
 281         old_daddr = mlxsw_sp_ipip_parms4_daddr(ipip_entry->parms4);
 282
 283         if (!mlxsw_sp_l3addr_eq(&new_saddr, &old_saddr)) {
 284                 u16 ul_tb_id = mlxsw_sp_ipip_dev_ul_tb_id(ipip_entry->ol_dev);
 285
 286                 /* Since the local address has changed, if there is another
 287                  * tunnel with a matching saddr, both need to be demoted.
 288                  */
 289                 if (mlxsw_sp_ipip_demote_tunnel_by_saddr(mlxsw_sp,
 290                                                          MLXSW_SP_L3_PROTO_IPV4,
 291                                                          new_saddr, ul_tb_id,
 292                                                          ipip_entry)) {
 293                         mlxsw_sp_ipip_entry_demote_tunnel(mlxsw_sp, ipip_entry);
 294                         return 0;
 295                 }
 296
 297                 update_tunnel = true;
 298         } else if ((mlxsw_sp_ipip_parms4_okey(ipip_entry->parms4) !=
 299                     mlxsw_sp_ipip_parms4_okey(new_parms)) ||
 300                    ipip_entry->parms4.link != new_parms.link) {
 301                 update_tunnel = true;
 302         } else if (!mlxsw_sp_l3addr_eq(&new_daddr, &old_daddr)) {
 303                 update_nhs = true;
 304         } else if (mlxsw_sp_ipip_parms4_ikey(ipip_entry->parms4) !=
 305                    mlxsw_sp_ipip_parms4_ikey(new_parms)) {
 306                 update_decap = true;
 307         }
 308
 309         if (update_tunnel)
 310                 err = __mlxsw_sp_ipip_entry_update_tunnel(mlxsw_sp, ipip_entry,
 311                                                           true, true, true,
 312                                                           extack);
 313         else if (update_nhs)
 314                 err = __mlxsw_sp_ipip_entry_update_tunnel(mlxsw_sp, ipip_entry,
 315                                                           false, false, true,
 316                                                           extack);
 317         else if (update_decap)
 318                 err = __mlxsw_sp_ipip_entry_update_tunnel(mlxsw_sp, ipip_entry,
 319                                                           false, false, false,
 320                                                           extack);
 321
 322         ipip_entry->parms4 = new_parms;
 323         return err;
 324 }
 325
 326 static const struct mlxsw_sp_ipip_ops mlxsw_sp_ipip_gre4_ops = {
 327         .dev_type = ARPHRD_IPGRE,
 328         .ul_proto = MLXSW_SP_L3_PROTO_IPV4,
 329         .nexthop_update = mlxsw_sp_ipip_nexthop_update_gre4,
 330         .fib_entry_op = mlxsw_sp_ipip_fib_entry_op_gre4,
 331         .can_offload = mlxsw_sp_ipip_can_offload_gre4,
 332         .ol_loopback_config = mlxsw_sp_ipip_ol_loopback_config_gre4,
 333         .ol_netdev_change = mlxsw_sp_ipip_ol_netdev_change_gre4,
 334 };
 335
 336 const struct mlxsw_sp_ipip_ops *mlxsw_sp_ipip_ops_arr[] = {
 337         [MLXSW_SP_IPIP_TYPE_GRE4] = &mlxsw_sp_ipip_gre4_ops,
 338 };