net/sched/act_mpls.c

   1 // SPDX-License-Identifier: (GPL-2.0-only OR BSD-2-Clause)
   2 /* Copyright (C) 2019 Netronome Systems, Inc. */
   3
   4 #include <linux/if_arp.h>
   5 #include <linux/init.h>
   6 #include <linux/kernel.h>
   7 #include <linux/module.h>
   8 #include <linux/mpls.h>
   9 #include <linux/rtnetlink.h>
  10 #include <linux/skbuff.h>
  11 #include <linux/tc_act/tc_mpls.h>
  12 #include <net/mpls.h>
  13 #include <net/netlink.h>
  14 #include <net/pkt_sched.h>
  15 #include <net/pkt_cls.h>
  16 #include <net/tc_act/tc_mpls.h>
  17
  18 static struct tc_action_ops act_mpls_ops;
  19
  20 #define ACT_MPLS_TTL_DEFAULT    255
  21
  22 static __be32 tcf_mpls_get_lse(struct mpls_shim_hdr *lse,
  23                                struct tcf_mpls_params *p, bool set_bos)
  24 {
  25         u32 new_lse = 0;
  26
  27         if (lse)
  28                 new_lse = be32_to_cpu(lse->label_stack_entry);
  29
  30         if (p->tcfm_label != ACT_MPLS_LABEL_NOT_SET) {
  31                 new_lse &= ~MPLS_LS_LABEL_MASK;
  32                 new_lse |= p->tcfm_label << MPLS_LS_LABEL_SHIFT;
  33         }
  34         if (p->tcfm_ttl) {
  35                 new_lse &= ~MPLS_LS_TTL_MASK;
  36                 new_lse |= p->tcfm_ttl << MPLS_LS_TTL_SHIFT;
  37         }
  38         if (p->tcfm_tc != ACT_MPLS_TC_NOT_SET) {
  39                 new_lse &= ~MPLS_LS_TC_MASK;
  40                 new_lse |= p->tcfm_tc << MPLS_LS_TC_SHIFT;
  41         }
  42         if (p->tcfm_bos != ACT_MPLS_BOS_NOT_SET) {
  43                 new_lse &= ~MPLS_LS_S_MASK;
  44                 new_lse |= p->tcfm_bos << MPLS_LS_S_SHIFT;
  45         } else if (set_bos) {
  46                 new_lse |= 1 << MPLS_LS_S_SHIFT;
  47         }
  48
  49         return cpu_to_be32(new_lse);
  50 }
  51
  52 static int tcf_mpls_act(struct sk_buff *skb, const struct tc_action *a,
  53                         struct tcf_result *res)
  54 {
  55         struct tcf_mpls *m = to_mpls(a);
  56         struct tcf_mpls_params *p;
  57         __be32 new_lse;
  58         int ret, mac_len;
  59
  60         tcf_lastuse_update(&m->tcf_tm);
  61         bstats_update(this_cpu_ptr(m->common.cpu_bstats), skb);
  62
  63         /* Ensure 'data' points at mac_header prior calling mpls manipulating
  64          * functions.
  65          */
  66         if (skb_at_tc_ingress(skb)) {
  67                 skb_push_rcsum(skb, skb->mac_len);
  68                 mac_len = skb->mac_len;
  69         } else {
  70                 mac_len = skb_network_header(skb) - skb_mac_header(skb);
  71         }
  72
  73         ret = READ_ONCE(m->tcf_action);
  74
  75         p = rcu_dereference_bh(m->mpls_p);
  76
  77         switch (p->tcfm_action) {
  78         case TCA_MPLS_ACT_POP:
  79                 if (skb_mpls_pop(skb, p->tcfm_proto, mac_len,
  80                                  skb->dev && skb->dev->type == ARPHRD_ETHER))
  81                         goto drop;
  82                 break;
  83         case TCA_MPLS_ACT_PUSH:
  84                 new_lse = tcf_mpls_get_lse(NULL, p, !eth_p_mpls(skb_protocol(skb, true)));
  85                 if (skb_mpls_push(skb, new_lse, p->tcfm_proto, mac_len,
  86                                   skb->dev && skb->dev->type == ARPHRD_ETHER))
  87                         goto drop;
  88                 break;
  89         case TCA_MPLS_ACT_MAC_PUSH:
  90                 if (skb_vlan_tag_present(skb)) {
  91                         if (__vlan_insert_inner_tag(skb, skb->vlan_proto,
  92                                                     skb_vlan_tag_get(skb),
  93                                                     ETH_HLEN) < 0)
  94                                 goto drop;
  95
  96                         skb->protocol = skb->vlan_proto;
  97                         __vlan_hwaccel_clear_tag(skb);
  98                 }
  99
 100                 new_lse = tcf_mpls_get_lse(NULL, p, mac_len ||
 101                                            !eth_p_mpls(skb->protocol));
 102
 103                 if (skb_mpls_push(skb, new_lse, p->tcfm_proto, 0, false))
 104                         goto drop;
 105                 break;
 106         case TCA_MPLS_ACT_MODIFY:
 107                 if (!pskb_may_pull(skb,
 108                                    skb_network_offset(skb) + MPLS_HLEN))
 109                         goto drop;
 110                 new_lse = tcf_mpls_get_lse(mpls_hdr(skb), p, false);
 111                 if (skb_mpls_update_lse(skb, new_lse))
 112                         goto drop;
 113                 break;
 114         case TCA_MPLS_ACT_DEC_TTL:
 115                 if (skb_mpls_dec_ttl(skb))
 116                         goto drop;
 117                 break;
 118         }
 119
 120         if (skb_at_tc_ingress(skb))
 121                 skb_pull_rcsum(skb, skb->mac_len);
 122
 123         return ret;
 124
 125 drop:
 126         qstats_drop_inc(this_cpu_ptr(m->common.cpu_qstats));
 127         return TC_ACT_SHOT;
 128 }
 129
 130 static int valid_label(const struct nlattr *attr,
 131                        struct netlink_ext_ack *extack)
 132 {
 133         const u32 *label = nla_data(attr);
 134
 135         if (*label & ~MPLS_LABEL_MASK || *label == MPLS_LABEL_IMPLNULL) {
 136                 NL_SET_ERR_MSG_MOD(extack, "MPLS label out of range");
 137                 return -EINVAL;
 138         }
 139
 140         return 0;
 141 }
 142
 143 static const struct nla_policy mpls_policy[TCA_MPLS_MAX + 1] = {
 144         [TCA_MPLS_PARMS]        = NLA_POLICY_EXACT_LEN(sizeof(struct tc_mpls)),
 145         [TCA_MPLS_PROTO]        = { .type = NLA_U16 },
 146         [TCA_MPLS_LABEL]        = NLA_POLICY_VALIDATE_FN(NLA_U32, valid_label),
 147         [TCA_MPLS_TC]           = NLA_POLICY_RANGE(NLA_U8, 0, 7),
 148         [TCA_MPLS_TTL]          = NLA_POLICY_MIN(NLA_U8, 1),
 149         [TCA_MPLS_BOS]          = NLA_POLICY_RANGE(NLA_U8, 0, 1),
 150 };
 151
 152 static int tcf_mpls_init(struct net *net, struct nlattr *nla,
 153                          struct nlattr *est, struct tc_action **a,
 154                          struct tcf_proto *tp, u32 flags,
 155                          struct netlink_ext_ack *extack)
 156 {
 157         struct tc_action_net *tn = net_generic(net, act_mpls_ops.net_id);
 158         bool bind = flags & TCA_ACT_FLAGS_BIND;
 159         struct nlattr *tb[TCA_MPLS_MAX + 1];
 160         struct tcf_chain *goto_ch = NULL;
 161         struct tcf_mpls_params *p;
 162         struct tc_mpls *parm;
 163         bool exists = false;
 164         struct tcf_mpls *m;
 165         int ret = 0, err;
 166         u8 mpls_ttl = 0;
 167         u32 index;
 168
 169         if (!nla) {
 170                 NL_SET_ERR_MSG_MOD(extack, "Missing netlink attributes");
 171                 return -EINVAL;
 172         }
 173
 174         err = nla_parse_nested(tb, TCA_MPLS_MAX, nla, mpls_policy, extack);
 175         if (err < 0)
 176                 return err;
 177
 178         if (!tb[TCA_MPLS_PARMS]) {
 179                 NL_SET_ERR_MSG_MOD(extack, "No MPLS params");
 180                 return -EINVAL;
 181         }
 182         parm = nla_data(tb[TCA_MPLS_PARMS]);
 183         index = parm->index;
 184
 185         /* Verify parameters against action type. */
 186         switch (parm->m_action) {
 187         case TCA_MPLS_ACT_POP:
 188                 if (!tb[TCA_MPLS_PROTO]) {
 189                         NL_SET_ERR_MSG_MOD(extack, "Protocol must be set for MPLS pop");
 190                         return -EINVAL;
 191                 }
 192                 if (!eth_proto_is_802_3(nla_get_be16(tb[TCA_MPLS_PROTO]))) {
 193                         NL_SET_ERR_MSG_MOD(extack, "Invalid protocol type for MPLS pop");
 194                         return -EINVAL;
 195                 }
 196                 if (tb[TCA_MPLS_LABEL] || tb[TCA_MPLS_TTL] || tb[TCA_MPLS_TC] ||
 197                     tb[TCA_MPLS_BOS]) {
 198                         NL_SET_ERR_MSG_MOD(extack, "Label, TTL, TC or BOS cannot be used with MPLS pop");
 199                         return -EINVAL;
 200                 }
 201                 break;
 202         case TCA_MPLS_ACT_DEC_TTL:
 203                 if (tb[TCA_MPLS_PROTO] || tb[TCA_MPLS_LABEL] ||
 204                     tb[TCA_MPLS_TTL] || tb[TCA_MPLS_TC] || tb[TCA_MPLS_BOS]) {
 205                         NL_SET_ERR_MSG_MOD(extack, "Label, TTL, TC, BOS or protocol cannot be used with MPLS dec_ttl");
 206                         return -EINVAL;
 207                 }
 208                 break;
 209         case TCA_MPLS_ACT_PUSH:
 210         case TCA_MPLS_ACT_MAC_PUSH:
 211                 if (!tb[TCA_MPLS_LABEL]) {
 212                         NL_SET_ERR_MSG_MOD(extack, "Label is required for MPLS push");
 213                         return -EINVAL;
 214                 }
 215                 if (tb[TCA_MPLS_PROTO] &&
 216                     !eth_p_mpls(nla_get_be16(tb[TCA_MPLS_PROTO]))) {
 217                         NL_SET_ERR_MSG_MOD(extack, "Protocol must be an MPLS type for MPLS push");
 218                         return -EPROTONOSUPPORT;
 219                 }
 220                 /* Push needs a TTL - if not specified, set a default value. */
 221                 if (!tb[TCA_MPLS_TTL]) {
 222 #if IS_ENABLED(CONFIG_MPLS)
 223                         mpls_ttl = net->mpls.default_ttl ?
 224                                    net->mpls.default_ttl : ACT_MPLS_TTL_DEFAULT;
 225 #else
 226                         mpls_ttl = ACT_MPLS_TTL_DEFAULT;
 227 #endif
 228                 }
 229                 break;
 230         case TCA_MPLS_ACT_MODIFY:
 231                 if (tb[TCA_MPLS_PROTO]) {
 232                         NL_SET_ERR_MSG_MOD(extack, "Protocol cannot be used with MPLS modify");
 233                         return -EINVAL;
 234                 }
 235                 break;
 236         default:
 237                 NL_SET_ERR_MSG_MOD(extack, "Unknown MPLS action");
 238                 return -EINVAL;
 239         }
 240
 241         err = tcf_idr_check_alloc(tn, &index, a, bind);
 242         if (err < 0)
 243                 return err;
 244         exists = err;
 245         if (exists && bind)
 246                 return 0;
 247
 248         if (!exists) {
 249                 ret = tcf_idr_create(tn, index, est, a,
 250                                      &act_mpls_ops, bind, true, flags);
 251                 if (ret) {
 252                         tcf_idr_cleanup(tn, index);
 253                         return ret;
 254                 }
 255
 256                 ret = ACT_P_CREATED;
 257         } else if (!(flags & TCA_ACT_FLAGS_REPLACE)) {
 258                 tcf_idr_release(*a, bind);
 259                 return -EEXIST;
 260         }
 261
 262         err = tcf_action_check_ctrlact(parm->action, tp, &goto_ch, extack);
 263         if (err < 0)
 264                 goto release_idr;
 265
 266         m = to_mpls(*a);
 267
 268         p = kzalloc(sizeof(*p), GFP_KERNEL);
 269         if (!p) {
 270                 err = -ENOMEM;
 271                 goto put_chain;
 272         }
 273
 274         p->tcfm_action = parm->m_action;
 275         p->tcfm_label = tb[TCA_MPLS_LABEL] ? nla_get_u32(tb[TCA_MPLS_LABEL]) :
 276                                              ACT_MPLS_LABEL_NOT_SET;
 277         p->tcfm_tc = tb[TCA_MPLS_TC] ? nla_get_u8(tb[TCA_MPLS_TC]) :
 278                                        ACT_MPLS_TC_NOT_SET;
 279         p->tcfm_ttl = tb[TCA_MPLS_TTL] ? nla_get_u8(tb[TCA_MPLS_TTL]) :
 280                                          mpls_ttl;
 281         p->tcfm_bos = tb[TCA_MPLS_BOS] ? nla_get_u8(tb[TCA_MPLS_BOS]) :
 282                                          ACT_MPLS_BOS_NOT_SET;
 283         p->tcfm_proto = tb[TCA_MPLS_PROTO] ? nla_get_be16(tb[TCA_MPLS_PROTO]) :
 284                                              htons(ETH_P_MPLS_UC);
 285
 286         spin_lock_bh(&m->tcf_lock);
 287         goto_ch = tcf_action_set_ctrlact(*a, parm->action, goto_ch);
 288         p = rcu_replace_pointer(m->mpls_p, p, lockdep_is_held(&m->tcf_lock));
 289         spin_unlock_bh(&m->tcf_lock);
 290
 291         if (goto_ch)
 292                 tcf_chain_put_by_act(goto_ch);
 293         if (p)
 294                 kfree_rcu(p, rcu);
 295
 296         return ret;
 297 put_chain:
 298         if (goto_ch)
 299                 tcf_chain_put_by_act(goto_ch);
 300 release_idr:
 301         tcf_idr_release(*a, bind);
 302         return err;
 303 }
 304
 305 static void tcf_mpls_cleanup(struct tc_action *a)
 306 {
 307         struct tcf_mpls *m = to_mpls(a);
 308         struct tcf_mpls_params *p;
 309
 310         p = rcu_dereference_protected(m->mpls_p, 1);
 311         if (p)
 312                 kfree_rcu(p, rcu);
 313 }
 314
 315 static int tcf_mpls_dump(struct sk_buff *skb, struct tc_action *a,
 316                          int bind, int ref)
 317 {
 318         unsigned char *b = skb_tail_pointer(skb);
 319         struct tcf_mpls *m = to_mpls(a);
 320         struct tcf_mpls_params *p;
 321         struct tc_mpls opt = {
 322                 .index    = m->tcf_index,
 323                 .refcnt   = refcount_read(&m->tcf_refcnt) - ref,
 324                 .bindcnt  = atomic_read(&m->tcf_bindcnt) - bind,
 325         };
 326         struct tcf_t t;
 327
 328         spin_lock_bh(&m->tcf_lock);
 329         opt.action = m->tcf_action;
 330         p = rcu_dereference_protected(m->mpls_p, lockdep_is_held(&m->tcf_lock));
 331         opt.m_action = p->tcfm_action;
 332
 333         if (nla_put(skb, TCA_MPLS_PARMS, sizeof(opt), &opt))
 334                 goto nla_put_failure;
 335
 336         if (p->tcfm_label != ACT_MPLS_LABEL_NOT_SET &&
 337             nla_put_u32(skb, TCA_MPLS_LABEL, p->tcfm_label))
 338                 goto nla_put_failure;
 339
 340         if (p->tcfm_tc != ACT_MPLS_TC_NOT_SET &&
 341             nla_put_u8(skb, TCA_MPLS_TC, p->tcfm_tc))
 342                 goto nla_put_failure;
 343
 344         if (p->tcfm_ttl && nla_put_u8(skb, TCA_MPLS_TTL, p->tcfm_ttl))
 345                 goto nla_put_failure;
 346
 347         if (p->tcfm_bos != ACT_MPLS_BOS_NOT_SET &&
 348             nla_put_u8(skb, TCA_MPLS_BOS, p->tcfm_bos))
 349                 goto nla_put_failure;
 350
 351         if (nla_put_be16(skb, TCA_MPLS_PROTO, p->tcfm_proto))
 352                 goto nla_put_failure;
 353
 354         tcf_tm_dump(&t, &m->tcf_tm);
 355
 356         if (nla_put_64bit(skb, TCA_MPLS_TM, sizeof(t), &t, TCA_MPLS_PAD))
 357                 goto nla_put_failure;
 358
 359         spin_unlock_bh(&m->tcf_lock);
 360
 361         return skb->len;
 362
 363 nla_put_failure:
 364         spin_unlock_bh(&m->tcf_lock);
 365         nlmsg_trim(skb, b);
 366         return -EMSGSIZE;
 367 }
 368
 369 static int tcf_mpls_offload_act_setup(struct tc_action *act, void *entry_data,
 370                                       u32 *index_inc, bool bind,
 371                                       struct netlink_ext_ack *extack)
 372 {
 373         if (bind) {
 374                 struct flow_action_entry *entry = entry_data;
 375
 376                 switch (tcf_mpls_action(act)) {
 377                 case TCA_MPLS_ACT_PUSH:
 378                         entry->id = FLOW_ACTION_MPLS_PUSH;
 379                         entry->mpls_push.proto = tcf_mpls_proto(act);
 380                         entry->mpls_push.label = tcf_mpls_label(act);
 381                         entry->mpls_push.tc = tcf_mpls_tc(act);
 382                         entry->mpls_push.bos = tcf_mpls_bos(act);
 383                         entry->mpls_push.ttl = tcf_mpls_ttl(act);
 384                         break;
 385                 case TCA_MPLS_ACT_POP:
 386                         entry->id = FLOW_ACTION_MPLS_POP;
 387                         entry->mpls_pop.proto = tcf_mpls_proto(act);
 388                         break;
 389                 case TCA_MPLS_ACT_MODIFY:
 390                         entry->id = FLOW_ACTION_MPLS_MANGLE;
 391                         entry->mpls_mangle.label = tcf_mpls_label(act);
 392                         entry->mpls_mangle.tc = tcf_mpls_tc(act);
 393                         entry->mpls_mangle.bos = tcf_mpls_bos(act);
 394                         entry->mpls_mangle.ttl = tcf_mpls_ttl(act);
 395                         break;
 396                 case TCA_MPLS_ACT_DEC_TTL:
 397                         NL_SET_ERR_MSG_MOD(extack, "Offload not supported when \"dec_ttl\" option is used");
 398                         return -EOPNOTSUPP;
 399                 case TCA_MPLS_ACT_MAC_PUSH:
 400                         NL_SET_ERR_MSG_MOD(extack, "Offload not supported when \"mac_push\" option is used");
 401                         return -EOPNOTSUPP;
 402                 default:
 403                         NL_SET_ERR_MSG_MOD(extack, "Unsupported MPLS mode offload");
 404                         return -EOPNOTSUPP;
 405                 }
 406                 *index_inc = 1;
 407         } else {
 408                 struct flow_offload_action *fl_action = entry_data;
 409
 410                 switch (tcf_mpls_action(act)) {
 411                 case TCA_MPLS_ACT_PUSH:
 412                         fl_action->id = FLOW_ACTION_MPLS_PUSH;
 413                         break;
 414                 case TCA_MPLS_ACT_POP:
 415                         fl_action->id = FLOW_ACTION_MPLS_POP;
 416                         break;
 417                 case TCA_MPLS_ACT_MODIFY:
 418                         fl_action->id = FLOW_ACTION_MPLS_MANGLE;
 419                         break;
 420                 default:
 421                         return -EOPNOTSUPP;
 422                 }
 423         }
 424
 425         return 0;
 426 }
 427
 428 static struct tc_action_ops act_mpls_ops = {
 429         .kind           =       "mpls",
 430         .id             =       TCA_ID_MPLS,
 431         .owner          =       THIS_MODULE,
 432         .act            =       tcf_mpls_act,
 433         .dump           =       tcf_mpls_dump,
 434         .init           =       tcf_mpls_init,
 435         .cleanup        =       tcf_mpls_cleanup,
 436         .offload_act_setup =    tcf_mpls_offload_act_setup,
 437         .size           =       sizeof(struct tcf_mpls),
 438 };
 439
 440 static __net_init int mpls_init_net(struct net *net)
 441 {
 442         struct tc_action_net *tn = net_generic(net, act_mpls_ops.net_id);
 443
 444         return tc_action_net_init(net, tn, &act_mpls_ops);
 445 }
 446
 447 static void __net_exit mpls_exit_net(struct list_head *net_list)
 448 {
 449         tc_action_net_exit(net_list, act_mpls_ops.net_id);
 450 }
 451
 452 static struct pernet_operations mpls_net_ops = {
 453         .init = mpls_init_net,
 454         .exit_batch = mpls_exit_net,
 455         .id   = &act_mpls_ops.net_id,
 456         .size = sizeof(struct tc_action_net),
 457 };
 458
 459 static int __init mpls_init_module(void)
 460 {
 461         return tcf_register_action(&act_mpls_ops, &mpls_net_ops);
 462 }
 463
 464 static void __exit mpls_cleanup_module(void)
 465 {
 466         tcf_unregister_action(&act_mpls_ops, &mpls_net_ops);
 467 }
 468
 469 module_init(mpls_init_module);
 470 module_exit(mpls_cleanup_module);
 471
 472 MODULE_SOFTDEP("post: mpls_gso");
 473 MODULE_AUTHOR("Netronome Systems <oss-drivers@netronome.com>");
 474 MODULE_LICENSE("GPL");
 475 MODULE_DESCRIPTION("MPLS manipulation actions");