net/sched/cls_flower.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * net/sched/cls_flower.c               Flower classifier
   4  *
   5  * Copyright (c) 2015 Jiri Pirko <jiri@resnulli.us>
   6  */
   7
   8 #include <linux/kernel.h>
   9 #include <linux/init.h>
  10 #include <linux/module.h>
  11 #include <linux/rhashtable.h>
  12 #include <linux/workqueue.h>
  13 #include <linux/refcount.h>
  14 #include <linux/bitfield.h>
  15
  16 #include <linux/if_ether.h>
  17 #include <linux/in6.h>
  18 #include <linux/ip.h>
  19 #include <linux/mpls.h>
  20 #include <linux/ppp_defs.h>
  21
  22 #include <net/sch_generic.h>
  23 #include <net/pkt_cls.h>
  24 #include <net/pkt_sched.h>
  25 #include <net/ip.h>
  26 #include <net/flow_dissector.h>
  27 #include <net/geneve.h>
  28 #include <net/vxlan.h>
  29 #include <net/erspan.h>
  30 #include <net/gtp.h>
  31 #include <net/pfcp.h>
  32 #include <net/tc_wrapper.h>
  33
  34 #include <net/dst.h>
  35 #include <net/dst_metadata.h>
  36
  37 #include <uapi/linux/netfilter/nf_conntrack_common.h>
  38
  39 #define TCA_FLOWER_KEY_CT_FLAGS_MAX \
  40                 ((__TCA_FLOWER_KEY_CT_FLAGS_MAX - 1) << 1)
  41 #define TCA_FLOWER_KEY_CT_FLAGS_MASK \
  42                 (TCA_FLOWER_KEY_CT_FLAGS_MAX - 1)
  43
  44 #define TCA_FLOWER_KEY_FLAGS_POLICY_MASK \
  45                 (TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT | \
  46                 TCA_FLOWER_KEY_FLAGS_FRAG_IS_FIRST)
  47
  48 #define TCA_FLOWER_KEY_ENC_FLAGS_POLICY_MASK \
  49                 (TCA_FLOWER_KEY_FLAGS_TUNNEL_CSUM | \
  50                 TCA_FLOWER_KEY_FLAGS_TUNNEL_DONT_FRAGMENT | \
  51                 TCA_FLOWER_KEY_FLAGS_TUNNEL_OAM | \
  52                 TCA_FLOWER_KEY_FLAGS_TUNNEL_CRIT_OPT)
  53
  54 struct fl_flow_key {
  55         struct flow_dissector_key_meta meta;
  56         struct flow_dissector_key_control control;
  57         struct flow_dissector_key_control enc_control;
  58         struct flow_dissector_key_basic basic;
  59         struct flow_dissector_key_eth_addrs eth;
  60         struct flow_dissector_key_vlan vlan;
  61         struct flow_dissector_key_vlan cvlan;
  62         union {
  63                 struct flow_dissector_key_ipv4_addrs ipv4;
  64                 struct flow_dissector_key_ipv6_addrs ipv6;
  65         };
  66         struct flow_dissector_key_ports tp;
  67         struct flow_dissector_key_icmp icmp;
  68         struct flow_dissector_key_arp arp;
  69         struct flow_dissector_key_keyid enc_key_id;
  70         union {
  71                 struct flow_dissector_key_ipv4_addrs enc_ipv4;
  72                 struct flow_dissector_key_ipv6_addrs enc_ipv6;
  73         };
  74         struct flow_dissector_key_ports enc_tp;
  75         struct flow_dissector_key_mpls mpls;
  76         struct flow_dissector_key_tcp tcp;
  77         struct flow_dissector_key_ip ip;
  78         struct flow_dissector_key_ip enc_ip;
  79         struct flow_dissector_key_enc_opts enc_opts;
  80         struct flow_dissector_key_ports_range tp_range;
  81         struct flow_dissector_key_ct ct;
  82         struct flow_dissector_key_hash hash;
  83         struct flow_dissector_key_num_of_vlans num_of_vlans;
  84         struct flow_dissector_key_pppoe pppoe;
  85         struct flow_dissector_key_l2tpv3 l2tpv3;
  86         struct flow_dissector_key_ipsec ipsec;
  87         struct flow_dissector_key_cfm cfm;
  88 } __aligned(BITS_PER_LONG / 8); /* Ensure that we can do comparisons as longs. */
  89
  90 struct fl_flow_mask_range {
  91         unsigned short int start;
  92         unsigned short int end;
  93 };
  94
  95 struct fl_flow_mask {
  96         struct fl_flow_key key;
  97         struct fl_flow_mask_range range;
  98         u32 flags;
  99         struct rhash_head ht_node;
 100         struct rhashtable ht;
 101         struct rhashtable_params filter_ht_params;
 102         struct flow_dissector dissector;
 103         struct list_head filters;
 104         struct rcu_work rwork;
 105         struct list_head list;
 106         refcount_t refcnt;
 107 };
 108
 109 struct fl_flow_tmplt {
 110         struct fl_flow_key dummy_key;
 111         struct fl_flow_key mask;
 112         struct flow_dissector dissector;
 113         struct tcf_chain *chain;
 114 };
 115
 116 struct cls_fl_head {
 117         struct rhashtable ht;
 118         spinlock_t masks_lock; /* Protect masks list */
 119         struct list_head masks;
 120         struct list_head hw_filters;
 121         struct rcu_work rwork;
 122         struct idr handle_idr;
 123 };
 124
 125 struct cls_fl_filter {
 126         struct fl_flow_mask *mask;
 127         struct rhash_head ht_node;
 128         struct fl_flow_key mkey;
 129         struct tcf_exts exts;
 130         struct tcf_result res;
 131         struct fl_flow_key key;
 132         struct list_head list;
 133         struct list_head hw_list;
 134         u32 handle;
 135         u32 flags;
 136         u32 in_hw_count;
 137         u8 needs_tc_skb_ext:1;
 138         struct rcu_work rwork;
 139         struct net_device *hw_dev;
 140         /* Flower classifier is unlocked, which means that its reference counter
 141          * can be changed concurrently without any kind of external
 142          * synchronization. Use atomic reference counter to be concurrency-safe.
 143          */
 144         refcount_t refcnt;
 145         bool deleted;
 146 };
 147
 148 static const struct rhashtable_params mask_ht_params = {
 149         .key_offset = offsetof(struct fl_flow_mask, key),
 150         .key_len = sizeof(struct fl_flow_key),
 151         .head_offset = offsetof(struct fl_flow_mask, ht_node),
 152         .automatic_shrinking = true,
 153 };
 154
 155 static unsigned short int fl_mask_range(const struct fl_flow_mask *mask)
 156 {
 157         return mask->range.end - mask->range.start;
 158 }
 159
 160 static void fl_mask_update_range(struct fl_flow_mask *mask)
 161 {
 162         const u8 *bytes = (const u8 *) &mask->key;
 163         size_t size = sizeof(mask->key);
 164         size_t i, first = 0, last;
 165
 166         for (i = 0; i < size; i++) {
 167                 if (bytes[i]) {
 168                         first = i;
 169                         break;
 170                 }
 171         }
 172         last = first;
 173         for (i = size - 1; i != first; i--) {
 174                 if (bytes[i]) {
 175                         last = i;
 176                         break;
 177                 }
 178         }
 179         mask->range.start = rounddown(first, sizeof(long));
 180         mask->range.end = roundup(last + 1, sizeof(long));
 181 }
 182
 183 static void *fl_key_get_start(struct fl_flow_key *key,
 184                               const struct fl_flow_mask *mask)
 185 {
 186         return (u8 *) key + mask->range.start;
 187 }
 188
 189 static void fl_set_masked_key(struct fl_flow_key *mkey, struct fl_flow_key *key,
 190                               struct fl_flow_mask *mask)
 191 {
 192         const long *lkey = fl_key_get_start(key, mask);
 193         const long *lmask = fl_key_get_start(&mask->key, mask);
 194         long *lmkey = fl_key_get_start(mkey, mask);
 195         int i;
 196
 197         for (i = 0; i < fl_mask_range(mask); i += sizeof(long))
 198                 *lmkey++ = *lkey++ & *lmask++;
 199 }
 200
 201 static bool fl_mask_fits_tmplt(struct fl_flow_tmplt *tmplt,
 202                                struct fl_flow_mask *mask)
 203 {
 204         const long *lmask = fl_key_get_start(&mask->key, mask);
 205         const long *ltmplt;
 206         int i;
 207
 208         if (!tmplt)
 209                 return true;
 210         ltmplt = fl_key_get_start(&tmplt->mask, mask);
 211         for (i = 0; i < fl_mask_range(mask); i += sizeof(long)) {
 212                 if (~*ltmplt++ & *lmask++)
 213                         return false;
 214         }
 215         return true;
 216 }
 217
 218 static void fl_clear_masked_range(struct fl_flow_key *key,
 219                                   struct fl_flow_mask *mask)
 220 {
 221         memset(fl_key_get_start(key, mask), 0, fl_mask_range(mask));
 222 }
 223
 224 static bool fl_range_port_dst_cmp(struct cls_fl_filter *filter,
 225                                   struct fl_flow_key *key,
 226                                   struct fl_flow_key *mkey)
 227 {
 228         u16 min_mask, max_mask, min_val, max_val;
 229
 230         min_mask = ntohs(filter->mask->key.tp_range.tp_min.dst);
 231         max_mask = ntohs(filter->mask->key.tp_range.tp_max.dst);
 232         min_val = ntohs(filter->key.tp_range.tp_min.dst);
 233         max_val = ntohs(filter->key.tp_range.tp_max.dst);
 234
 235         if (min_mask && max_mask) {
 236                 if (ntohs(key->tp_range.tp.dst) < min_val ||
 237                     ntohs(key->tp_range.tp.dst) > max_val)
 238                         return false;
 239
 240                 /* skb does not have min and max values */
 241                 mkey->tp_range.tp_min.dst = filter->mkey.tp_range.tp_min.dst;
 242                 mkey->tp_range.tp_max.dst = filter->mkey.tp_range.tp_max.dst;
 243         }
 244         return true;
 245 }
 246
 247 static bool fl_range_port_src_cmp(struct cls_fl_filter *filter,
 248                                   struct fl_flow_key *key,
 249                                   struct fl_flow_key *mkey)
 250 {
 251         u16 min_mask, max_mask, min_val, max_val;
 252
 253         min_mask = ntohs(filter->mask->key.tp_range.tp_min.src);
 254         max_mask = ntohs(filter->mask->key.tp_range.tp_max.src);
 255         min_val = ntohs(filter->key.tp_range.tp_min.src);
 256         max_val = ntohs(filter->key.tp_range.tp_max.src);
 257
 258         if (min_mask && max_mask) {
 259                 if (ntohs(key->tp_range.tp.src) < min_val ||
 260                     ntohs(key->tp_range.tp.src) > max_val)
 261                         return false;
 262
 263                 /* skb does not have min and max values */
 264                 mkey->tp_range.tp_min.src = filter->mkey.tp_range.tp_min.src;
 265                 mkey->tp_range.tp_max.src = filter->mkey.tp_range.tp_max.src;
 266         }
 267         return true;
 268 }
 269
 270 static struct cls_fl_filter *__fl_lookup(struct fl_flow_mask *mask,
 271                                          struct fl_flow_key *mkey)
 272 {
 273         return rhashtable_lookup_fast(&mask->ht, fl_key_get_start(mkey, mask),
 274                                       mask->filter_ht_params);
 275 }
 276
 277 static struct cls_fl_filter *fl_lookup_range(struct fl_flow_mask *mask,
 278                                              struct fl_flow_key *mkey,
 279                                              struct fl_flow_key *key)
 280 {
 281         struct cls_fl_filter *filter, *f;
 282
 283         list_for_each_entry_rcu(filter, &mask->filters, list) {
 284                 if (!fl_range_port_dst_cmp(filter, key, mkey))
 285                         continue;
 286
 287                 if (!fl_range_port_src_cmp(filter, key, mkey))
 288                         continue;
 289
 290                 f = __fl_lookup(mask, mkey);
 291                 if (f)
 292                         return f;
 293         }
 294         return NULL;
 295 }
 296
 297 static noinline_for_stack
 298 struct cls_fl_filter *fl_mask_lookup(struct fl_flow_mask *mask, struct fl_flow_key *key)
 299 {
 300         struct fl_flow_key mkey;
 301
 302         fl_set_masked_key(&mkey, key, mask);
 303         if ((mask->flags & TCA_FLOWER_MASK_FLAGS_RANGE))
 304                 return fl_lookup_range(mask, &mkey, key);
 305
 306         return __fl_lookup(mask, &mkey);
 307 }
 308
 309 static u16 fl_ct_info_to_flower_map[] = {
 310         [IP_CT_ESTABLISHED] =           TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 311                                         TCA_FLOWER_KEY_CT_FLAGS_ESTABLISHED,
 312         [IP_CT_RELATED] =               TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 313                                         TCA_FLOWER_KEY_CT_FLAGS_RELATED,
 314         [IP_CT_ESTABLISHED_REPLY] =     TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 315                                         TCA_FLOWER_KEY_CT_FLAGS_ESTABLISHED |
 316                                         TCA_FLOWER_KEY_CT_FLAGS_REPLY,
 317         [IP_CT_RELATED_REPLY] =         TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 318                                         TCA_FLOWER_KEY_CT_FLAGS_RELATED |
 319                                         TCA_FLOWER_KEY_CT_FLAGS_REPLY,
 320         [IP_CT_NEW] =                   TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 321                                         TCA_FLOWER_KEY_CT_FLAGS_NEW,
 322 };
 323
 324 TC_INDIRECT_SCOPE int fl_classify(struct sk_buff *skb,
 325                                   const struct tcf_proto *tp,
 326                                   struct tcf_result *res)
 327 {
 328         struct cls_fl_head *head = rcu_dereference_bh(tp->root);
 329         bool post_ct = tc_skb_cb(skb)->post_ct;
 330         u16 zone = tc_skb_cb(skb)->zone;
 331         struct fl_flow_key skb_key;
 332         struct fl_flow_mask *mask;
 333         struct cls_fl_filter *f;
 334
 335         list_for_each_entry_rcu(mask, &head->masks, list) {
 336                 flow_dissector_init_keys(&skb_key.control, &skb_key.basic);
 337                 fl_clear_masked_range(&skb_key, mask);
 338
 339                 skb_flow_dissect_meta(skb, &mask->dissector, &skb_key);
 340                 /* skb_flow_dissect() does not set n_proto in case an unknown
 341                  * protocol, so do it rather here.
 342                  */
 343                 skb_key.basic.n_proto = skb_protocol(skb, false);
 344                 skb_flow_dissect_tunnel_info(skb, &mask->dissector, &skb_key);
 345                 skb_flow_dissect_ct(skb, &mask->dissector, &skb_key,
 346                                     fl_ct_info_to_flower_map,
 347                                     ARRAY_SIZE(fl_ct_info_to_flower_map),
 348                                     post_ct, zone);
 349                 skb_flow_dissect_hash(skb, &mask->dissector, &skb_key);
 350                 skb_flow_dissect(skb, &mask->dissector, &skb_key,
 351                                  FLOW_DISSECTOR_F_STOP_BEFORE_ENCAP);
 352
 353                 f = fl_mask_lookup(mask, &skb_key);
 354                 if (f && !tc_skip_sw(f->flags)) {
 355                         *res = f->res;
 356                         return tcf_exts_exec(skb, &f->exts, res);
 357                 }
 358         }
 359         return -1;
 360 }
 361
 362 static int fl_init(struct tcf_proto *tp)
 363 {
 364         struct cls_fl_head *head;
 365
 366         head = kzalloc(sizeof(*head), GFP_KERNEL);
 367         if (!head)
 368                 return -ENOBUFS;
 369
 370         spin_lock_init(&head->masks_lock);
 371         INIT_LIST_HEAD_RCU(&head->masks);
 372         INIT_LIST_HEAD(&head->hw_filters);
 373         rcu_assign_pointer(tp->root, head);
 374         idr_init(&head->handle_idr);
 375
 376         return rhashtable_init(&head->ht, &mask_ht_params);
 377 }
 378
 379 static void fl_mask_free(struct fl_flow_mask *mask, bool mask_init_done)
 380 {
 381         /* temporary masks don't have their filters list and ht initialized */
 382         if (mask_init_done) {
 383                 WARN_ON(!list_empty(&mask->filters));
 384                 rhashtable_destroy(&mask->ht);
 385         }
 386         kfree(mask);
 387 }
 388
 389 static void fl_mask_free_work(struct work_struct *work)
 390 {
 391         struct fl_flow_mask *mask = container_of(to_rcu_work(work),
 392                                                  struct fl_flow_mask, rwork);
 393
 394         fl_mask_free(mask, true);
 395 }
 396
 397 static void fl_uninit_mask_free_work(struct work_struct *work)
 398 {
 399         struct fl_flow_mask *mask = container_of(to_rcu_work(work),
 400                                                  struct fl_flow_mask, rwork);
 401
 402         fl_mask_free(mask, false);
 403 }
 404
 405 static bool fl_mask_put(struct cls_fl_head *head, struct fl_flow_mask *mask)
 406 {
 407         if (!refcount_dec_and_test(&mask->refcnt))
 408                 return false;
 409
 410         rhashtable_remove_fast(&head->ht, &mask->ht_node, mask_ht_params);
 411
 412         spin_lock(&head->masks_lock);
 413         list_del_rcu(&mask->list);
 414         spin_unlock(&head->masks_lock);
 415
 416         tcf_queue_work(&mask->rwork, fl_mask_free_work);
 417
 418         return true;
 419 }
 420
 421 static struct cls_fl_head *fl_head_dereference(struct tcf_proto *tp)
 422 {
 423         /* Flower classifier only changes root pointer during init and destroy.
 424          * Users must obtain reference to tcf_proto instance before calling its
 425          * API, so tp->root pointer is protected from concurrent call to
 426          * fl_destroy() by reference counting.
 427          */
 428         return rcu_dereference_raw(tp->root);
 429 }
 430
 431 static void __fl_destroy_filter(struct cls_fl_filter *f)
 432 {
 433         if (f->needs_tc_skb_ext)
 434                 tc_skb_ext_tc_disable();
 435         tcf_exts_destroy(&f->exts);
 436         tcf_exts_put_net(&f->exts);
 437         kfree(f);
 438 }
 439
 440 static void fl_destroy_filter_work(struct work_struct *work)
 441 {
 442         struct cls_fl_filter *f = container_of(to_rcu_work(work),
 443                                         struct cls_fl_filter, rwork);
 444
 445         __fl_destroy_filter(f);
 446 }
 447
 448 static void fl_hw_destroy_filter(struct tcf_proto *tp, struct cls_fl_filter *f,
 449                                  bool rtnl_held, struct netlink_ext_ack *extack)
 450 {
 451         struct tcf_block *block = tp->chain->block;
 452         struct flow_cls_offload cls_flower = {};
 453
 454         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, extack);
 455         cls_flower.command = FLOW_CLS_DESTROY;
 456         cls_flower.cookie = (unsigned long) f;
 457
 458         tc_setup_cb_destroy(block, tp, TC_SETUP_CLSFLOWER, &cls_flower, false,
 459                             &f->flags, &f->in_hw_count, rtnl_held);
 460
 461 }
 462
 463 static int fl_hw_replace_filter(struct tcf_proto *tp,
 464                                 struct cls_fl_filter *f, bool rtnl_held,
 465                                 struct netlink_ext_ack *extack)
 466 {
 467         struct tcf_block *block = tp->chain->block;
 468         struct flow_cls_offload cls_flower = {};
 469         bool skip_sw = tc_skip_sw(f->flags);
 470         int err = 0;
 471
 472         cls_flower.rule = flow_rule_alloc(tcf_exts_num_actions(&f->exts));
 473         if (!cls_flower.rule)
 474                 return -ENOMEM;
 475
 476         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, extack);
 477         cls_flower.command = FLOW_CLS_REPLACE;
 478         cls_flower.cookie = (unsigned long) f;
 479         cls_flower.rule->match.dissector = &f->mask->dissector;
 480         cls_flower.rule->match.mask = &f->mask->key;
 481         cls_flower.rule->match.key = &f->mkey;
 482         cls_flower.classid = f->res.classid;
 483
 484         err = tc_setup_offload_action(&cls_flower.rule->action, &f->exts,
 485                                       cls_flower.common.extack);
 486         if (err) {
 487                 kfree(cls_flower.rule);
 488
 489                 return skip_sw ? err : 0;
 490         }
 491
 492         err = tc_setup_cb_add(block, tp, TC_SETUP_CLSFLOWER, &cls_flower,
 493                               skip_sw, &f->flags, &f->in_hw_count, rtnl_held);
 494         tc_cleanup_offload_action(&cls_flower.rule->action);
 495         kfree(cls_flower.rule);
 496
 497         if (err) {
 498                 fl_hw_destroy_filter(tp, f, rtnl_held, NULL);
 499                 return err;
 500         }
 501
 502         if (skip_sw && !(f->flags & TCA_CLS_FLAGS_IN_HW))
 503                 return -EINVAL;
 504
 505         return 0;
 506 }
 507
 508 static void fl_hw_update_stats(struct tcf_proto *tp, struct cls_fl_filter *f,
 509                                bool rtnl_held)
 510 {
 511         struct tcf_block *block = tp->chain->block;
 512         struct flow_cls_offload cls_flower = {};
 513
 514         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, NULL);
 515         cls_flower.command = FLOW_CLS_STATS;
 516         cls_flower.cookie = (unsigned long) f;
 517         cls_flower.classid = f->res.classid;
 518
 519         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false,
 520                          rtnl_held);
 521
 522         tcf_exts_hw_stats_update(&f->exts, &cls_flower.stats, cls_flower.use_act_stats);
 523 }
 524
 525 static void __fl_put(struct cls_fl_filter *f)
 526 {
 527         if (!refcount_dec_and_test(&f->refcnt))
 528                 return;
 529
 530         if (tcf_exts_get_net(&f->exts))
 531                 tcf_queue_work(&f->rwork, fl_destroy_filter_work);
 532         else
 533                 __fl_destroy_filter(f);
 534 }
 535
 536 static struct cls_fl_filter *__fl_get(struct cls_fl_head *head, u32 handle)
 537 {
 538         struct cls_fl_filter *f;
 539
 540         rcu_read_lock();
 541         f = idr_find(&head->handle_idr, handle);
 542         if (f && !refcount_inc_not_zero(&f->refcnt))
 543                 f = NULL;
 544         rcu_read_unlock();
 545
 546         return f;
 547 }
 548
 549 static struct tcf_exts *fl_get_exts(const struct tcf_proto *tp, u32 handle)
 550 {
 551         struct cls_fl_head *head = rcu_dereference_bh(tp->root);
 552         struct cls_fl_filter *f;
 553
 554         f = idr_find(&head->handle_idr, handle);
 555         return f ? &f->exts : NULL;
 556 }
 557
 558 static int __fl_delete(struct tcf_proto *tp, struct cls_fl_filter *f,
 559                        bool *last, bool rtnl_held,
 560                        struct netlink_ext_ack *extack)
 561 {
 562         struct cls_fl_head *head = fl_head_dereference(tp);
 563
 564         *last = false;
 565
 566         spin_lock(&tp->lock);
 567         if (f->deleted) {
 568                 spin_unlock(&tp->lock);
 569                 return -ENOENT;
 570         }
 571
 572         f->deleted = true;
 573         rhashtable_remove_fast(&f->mask->ht, &f->ht_node,
 574                                f->mask->filter_ht_params);
 575         idr_remove(&head->handle_idr, f->handle);
 576         list_del_rcu(&f->list);
 577         spin_unlock(&tp->lock);
 578
 579         *last = fl_mask_put(head, f->mask);
 580         if (!tc_skip_hw(f->flags))
 581                 fl_hw_destroy_filter(tp, f, rtnl_held, extack);
 582         tcf_unbind_filter(tp, &f->res);
 583         __fl_put(f);
 584
 585         return 0;
 586 }
 587
 588 static void fl_destroy_sleepable(struct work_struct *work)
 589 {
 590         struct cls_fl_head *head = container_of(to_rcu_work(work),
 591                                                 struct cls_fl_head,
 592                                                 rwork);
 593
 594         rhashtable_destroy(&head->ht);
 595         kfree(head);
 596         module_put(THIS_MODULE);
 597 }
 598
 599 static void fl_destroy(struct tcf_proto *tp, bool rtnl_held,
 600                        struct netlink_ext_ack *extack)
 601 {
 602         struct cls_fl_head *head = fl_head_dereference(tp);
 603         struct fl_flow_mask *mask, *next_mask;
 604         struct cls_fl_filter *f, *next;
 605         bool last;
 606
 607         list_for_each_entry_safe(mask, next_mask, &head->masks, list) {
 608                 list_for_each_entry_safe(f, next, &mask->filters, list) {
 609                         __fl_delete(tp, f, &last, rtnl_held, extack);
 610                         if (last)
 611                                 break;
 612                 }
 613         }
 614         idr_destroy(&head->handle_idr);
 615
 616         __module_get(THIS_MODULE);
 617         tcf_queue_work(&head->rwork, fl_destroy_sleepable);
 618 }
 619
 620 static void fl_put(struct tcf_proto *tp, void *arg)
 621 {
 622         struct cls_fl_filter *f = arg;
 623
 624         __fl_put(f);
 625 }
 626
 627 static void *fl_get(struct tcf_proto *tp, u32 handle)
 628 {
 629         struct cls_fl_head *head = fl_head_dereference(tp);
 630
 631         return __fl_get(head, handle);
 632 }
 633
 634 static const struct nla_policy fl_policy[TCA_FLOWER_MAX + 1] = {
 635         [TCA_FLOWER_UNSPEC]             = { .strict_start_type =
 636                                                 TCA_FLOWER_L2_MISS },
 637         [TCA_FLOWER_CLASSID]            = { .type = NLA_U32 },
 638         [TCA_FLOWER_INDEV]              = { .type = NLA_STRING,
 639                                             .len = IFNAMSIZ },
 640         [TCA_FLOWER_KEY_ETH_DST]        = { .len = ETH_ALEN },
 641         [TCA_FLOWER_KEY_ETH_DST_MASK]   = { .len = ETH_ALEN },
 642         [TCA_FLOWER_KEY_ETH_SRC]        = { .len = ETH_ALEN },
 643         [TCA_FLOWER_KEY_ETH_SRC_MASK]   = { .len = ETH_ALEN },
 644         [TCA_FLOWER_KEY_ETH_TYPE]       = { .type = NLA_U16 },
 645         [TCA_FLOWER_KEY_IP_PROTO]       = { .type = NLA_U8 },
 646         [TCA_FLOWER_KEY_IPV4_SRC]       = { .type = NLA_U32 },
 647         [TCA_FLOWER_KEY_IPV4_SRC_MASK]  = { .type = NLA_U32 },
 648         [TCA_FLOWER_KEY_IPV4_DST]       = { .type = NLA_U32 },
 649         [TCA_FLOWER_KEY_IPV4_DST_MASK]  = { .type = NLA_U32 },
 650         [TCA_FLOWER_KEY_IPV6_SRC]       = { .len = sizeof(struct in6_addr) },
 651         [TCA_FLOWER_KEY_IPV6_SRC_MASK]  = { .len = sizeof(struct in6_addr) },
 652         [TCA_FLOWER_KEY_IPV6_DST]       = { .len = sizeof(struct in6_addr) },
 653         [TCA_FLOWER_KEY_IPV6_DST_MASK]  = { .len = sizeof(struct in6_addr) },
 654         [TCA_FLOWER_KEY_TCP_SRC]        = { .type = NLA_U16 },
 655         [TCA_FLOWER_KEY_TCP_DST]        = { .type = NLA_U16 },
 656         [TCA_FLOWER_KEY_UDP_SRC]        = { .type = NLA_U16 },
 657         [TCA_FLOWER_KEY_UDP_DST]        = { .type = NLA_U16 },
 658         [TCA_FLOWER_KEY_VLAN_ID]        = { .type = NLA_U16 },
 659         [TCA_FLOWER_KEY_VLAN_PRIO]      = { .type = NLA_U8 },
 660         [TCA_FLOWER_KEY_VLAN_ETH_TYPE]  = { .type = NLA_U16 },
 661         [TCA_FLOWER_KEY_ENC_KEY_ID]     = { .type = NLA_U32 },
 662         [TCA_FLOWER_KEY_ENC_IPV4_SRC]   = { .type = NLA_U32 },
 663         [TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK] = { .type = NLA_U32 },
 664         [TCA_FLOWER_KEY_ENC_IPV4_DST]   = { .type = NLA_U32 },
 665         [TCA_FLOWER_KEY_ENC_IPV4_DST_MASK] = { .type = NLA_U32 },
 666         [TCA_FLOWER_KEY_ENC_IPV6_SRC]   = { .len = sizeof(struct in6_addr) },
 667         [TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK] = { .len = sizeof(struct in6_addr) },
 668         [TCA_FLOWER_KEY_ENC_IPV6_DST]   = { .len = sizeof(struct in6_addr) },
 669         [TCA_FLOWER_KEY_ENC_IPV6_DST_MASK] = { .len = sizeof(struct in6_addr) },
 670         [TCA_FLOWER_KEY_TCP_SRC_MASK]   = { .type = NLA_U16 },
 671         [TCA_FLOWER_KEY_TCP_DST_MASK]   = { .type = NLA_U16 },
 672         [TCA_FLOWER_KEY_UDP_SRC_MASK]   = { .type = NLA_U16 },
 673         [TCA_FLOWER_KEY_UDP_DST_MASK]   = { .type = NLA_U16 },
 674         [TCA_FLOWER_KEY_SCTP_SRC_MASK]  = { .type = NLA_U16 },
 675         [TCA_FLOWER_KEY_SCTP_DST_MASK]  = { .type = NLA_U16 },
 676         [TCA_FLOWER_KEY_SCTP_SRC]       = { .type = NLA_U16 },
 677         [TCA_FLOWER_KEY_SCTP_DST]       = { .type = NLA_U16 },
 678         [TCA_FLOWER_KEY_ENC_UDP_SRC_PORT]       = { .type = NLA_U16 },
 679         [TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK]  = { .type = NLA_U16 },
 680         [TCA_FLOWER_KEY_ENC_UDP_DST_PORT]       = { .type = NLA_U16 },
 681         [TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK]  = { .type = NLA_U16 },
 682         [TCA_FLOWER_KEY_FLAGS]          = NLA_POLICY_MASK(NLA_BE32,
 683                                                           TCA_FLOWER_KEY_FLAGS_POLICY_MASK),
 684         [TCA_FLOWER_KEY_FLAGS_MASK]     = NLA_POLICY_MASK(NLA_BE32,
 685                                                           TCA_FLOWER_KEY_FLAGS_POLICY_MASK),
 686         [TCA_FLOWER_KEY_ICMPV4_TYPE]    = { .type = NLA_U8 },
 687         [TCA_FLOWER_KEY_ICMPV4_TYPE_MASK] = { .type = NLA_U8 },
 688         [TCA_FLOWER_KEY_ICMPV4_CODE]    = { .type = NLA_U8 },
 689         [TCA_FLOWER_KEY_ICMPV4_CODE_MASK] = { .type = NLA_U8 },
 690         [TCA_FLOWER_KEY_ICMPV6_TYPE]    = { .type = NLA_U8 },
 691         [TCA_FLOWER_KEY_ICMPV6_TYPE_MASK] = { .type = NLA_U8 },
 692         [TCA_FLOWER_KEY_ICMPV6_CODE]    = { .type = NLA_U8 },
 693         [TCA_FLOWER_KEY_ICMPV6_CODE_MASK] = { .type = NLA_U8 },
 694         [TCA_FLOWER_KEY_ARP_SIP]        = { .type = NLA_U32 },
 695         [TCA_FLOWER_KEY_ARP_SIP_MASK]   = { .type = NLA_U32 },
 696         [TCA_FLOWER_KEY_ARP_TIP]        = { .type = NLA_U32 },
 697         [TCA_FLOWER_KEY_ARP_TIP_MASK]   = { .type = NLA_U32 },
 698         [TCA_FLOWER_KEY_ARP_OP]         = { .type = NLA_U8 },
 699         [TCA_FLOWER_KEY_ARP_OP_MASK]    = { .type = NLA_U8 },
 700         [TCA_FLOWER_KEY_ARP_SHA]        = { .len = ETH_ALEN },
 701         [TCA_FLOWER_KEY_ARP_SHA_MASK]   = { .len = ETH_ALEN },
 702         [TCA_FLOWER_KEY_ARP_THA]        = { .len = ETH_ALEN },
 703         [TCA_FLOWER_KEY_ARP_THA_MASK]   = { .len = ETH_ALEN },
 704         [TCA_FLOWER_KEY_MPLS_TTL]       = { .type = NLA_U8 },
 705         [TCA_FLOWER_KEY_MPLS_BOS]       = { .type = NLA_U8 },
 706         [TCA_FLOWER_KEY_MPLS_TC]        = { .type = NLA_U8 },
 707         [TCA_FLOWER_KEY_MPLS_LABEL]     = { .type = NLA_U32 },
 708         [TCA_FLOWER_KEY_MPLS_OPTS]      = { .type = NLA_NESTED },
 709         [TCA_FLOWER_KEY_TCP_FLAGS]      = { .type = NLA_U16 },
 710         [TCA_FLOWER_KEY_TCP_FLAGS_MASK] = { .type = NLA_U16 },
 711         [TCA_FLOWER_KEY_IP_TOS]         = { .type = NLA_U8 },
 712         [TCA_FLOWER_KEY_IP_TOS_MASK]    = { .type = NLA_U8 },
 713         [TCA_FLOWER_KEY_IP_TTL]         = { .type = NLA_U8 },
 714         [TCA_FLOWER_KEY_IP_TTL_MASK]    = { .type = NLA_U8 },
 715         [TCA_FLOWER_KEY_CVLAN_ID]       = { .type = NLA_U16 },
 716         [TCA_FLOWER_KEY_CVLAN_PRIO]     = { .type = NLA_U8 },
 717         [TCA_FLOWER_KEY_CVLAN_ETH_TYPE] = { .type = NLA_U16 },
 718         [TCA_FLOWER_KEY_ENC_IP_TOS]     = { .type = NLA_U8 },
 719         [TCA_FLOWER_KEY_ENC_IP_TOS_MASK] = { .type = NLA_U8 },
 720         [TCA_FLOWER_KEY_ENC_IP_TTL]      = { .type = NLA_U8 },
 721         [TCA_FLOWER_KEY_ENC_IP_TTL_MASK] = { .type = NLA_U8 },
 722         [TCA_FLOWER_KEY_ENC_OPTS]       = { .type = NLA_NESTED },
 723         [TCA_FLOWER_KEY_ENC_OPTS_MASK]  = { .type = NLA_NESTED },
 724         [TCA_FLOWER_KEY_CT_STATE]       =
 725                 NLA_POLICY_MASK(NLA_U16, TCA_FLOWER_KEY_CT_FLAGS_MASK),
 726         [TCA_FLOWER_KEY_CT_STATE_MASK]  =
 727                 NLA_POLICY_MASK(NLA_U16, TCA_FLOWER_KEY_CT_FLAGS_MASK),
 728         [TCA_FLOWER_KEY_CT_ZONE]        = { .type = NLA_U16 },
 729         [TCA_FLOWER_KEY_CT_ZONE_MASK]   = { .type = NLA_U16 },
 730         [TCA_FLOWER_KEY_CT_MARK]        = { .type = NLA_U32 },
 731         [TCA_FLOWER_KEY_CT_MARK_MASK]   = { .type = NLA_U32 },
 732         [TCA_FLOWER_KEY_CT_LABELS]      = { .type = NLA_BINARY,
 733                                             .len = 128 / BITS_PER_BYTE },
 734         [TCA_FLOWER_KEY_CT_LABELS_MASK] = { .type = NLA_BINARY,
 735                                             .len = 128 / BITS_PER_BYTE },
 736         [TCA_FLOWER_FLAGS]              = { .type = NLA_U32 },
 737         [TCA_FLOWER_KEY_HASH]           = { .type = NLA_U32 },
 738         [TCA_FLOWER_KEY_HASH_MASK]      = { .type = NLA_U32 },
 739         [TCA_FLOWER_KEY_NUM_OF_VLANS]   = { .type = NLA_U8 },
 740         [TCA_FLOWER_KEY_PPPOE_SID]      = { .type = NLA_U16 },
 741         [TCA_FLOWER_KEY_PPP_PROTO]      = { .type = NLA_U16 },
 742         [TCA_FLOWER_KEY_L2TPV3_SID]     = { .type = NLA_U32 },
 743         [TCA_FLOWER_KEY_SPI]            = { .type = NLA_U32 },
 744         [TCA_FLOWER_KEY_SPI_MASK]       = { .type = NLA_U32 },
 745         [TCA_FLOWER_L2_MISS]            = NLA_POLICY_MAX(NLA_U8, 1),
 746         [TCA_FLOWER_KEY_CFM]            = { .type = NLA_NESTED },
 747         [TCA_FLOWER_KEY_ENC_FLAGS]      = NLA_POLICY_MASK(NLA_BE32,
 748                                                           TCA_FLOWER_KEY_ENC_FLAGS_POLICY_MASK),
 749         [TCA_FLOWER_KEY_ENC_FLAGS_MASK] = NLA_POLICY_MASK(NLA_BE32,
 750                                                           TCA_FLOWER_KEY_ENC_FLAGS_POLICY_MASK),
 751 };
 752
 753 static const struct nla_policy
 754 enc_opts_policy[TCA_FLOWER_KEY_ENC_OPTS_MAX + 1] = {
 755         [TCA_FLOWER_KEY_ENC_OPTS_UNSPEC]        = {
 756                 .strict_start_type = TCA_FLOWER_KEY_ENC_OPTS_VXLAN },
 757         [TCA_FLOWER_KEY_ENC_OPTS_GENEVE]        = { .type = NLA_NESTED },
 758         [TCA_FLOWER_KEY_ENC_OPTS_VXLAN]         = { .type = NLA_NESTED },
 759         [TCA_FLOWER_KEY_ENC_OPTS_ERSPAN]        = { .type = NLA_NESTED },
 760         [TCA_FLOWER_KEY_ENC_OPTS_GTP]           = { .type = NLA_NESTED },
 761         [TCA_FLOWER_KEY_ENC_OPTS_PFCP]          = { .type = NLA_NESTED },
 762 };
 763
 764 static const struct nla_policy
 765 geneve_opt_policy[TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX + 1] = {
 766         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS]      = { .type = NLA_U16 },
 767         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE]       = { .type = NLA_U8 },
 768         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA]       = { .type = NLA_BINARY,
 769                                                        .len = 127 },
 770 };
 771
 772 static const struct nla_policy
 773 vxlan_opt_policy[TCA_FLOWER_KEY_ENC_OPT_VXLAN_MAX + 1] = {
 774         [TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]         = { .type = NLA_U32 },
 775 };
 776
 777 static const struct nla_policy
 778 erspan_opt_policy[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_MAX + 1] = {
 779         [TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER]        = { .type = NLA_U8 },
 780         [TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX]      = { .type = NLA_U32 },
 781         [TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR]        = { .type = NLA_U8 },
 782         [TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID]       = { .type = NLA_U8 },
 783 };
 784
 785 static const struct nla_policy
 786 gtp_opt_policy[TCA_FLOWER_KEY_ENC_OPT_GTP_MAX + 1] = {
 787         [TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE]      = { .type = NLA_U8 },
 788         [TCA_FLOWER_KEY_ENC_OPT_GTP_QFI]           = { .type = NLA_U8 },
 789 };
 790
 791 static const struct nla_policy
 792 pfcp_opt_policy[TCA_FLOWER_KEY_ENC_OPT_PFCP_MAX + 1] = {
 793         [TCA_FLOWER_KEY_ENC_OPT_PFCP_TYPE]         = { .type = NLA_U8 },
 794         [TCA_FLOWER_KEY_ENC_OPT_PFCP_SEID]         = { .type = NLA_U64 },
 795 };
 796
 797 static const struct nla_policy
 798 mpls_stack_entry_policy[TCA_FLOWER_KEY_MPLS_OPT_LSE_MAX + 1] = {
 799         [TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH]    = { .type = NLA_U8 },
 800         [TCA_FLOWER_KEY_MPLS_OPT_LSE_TTL]      = { .type = NLA_U8 },
 801         [TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS]      = { .type = NLA_U8 },
 802         [TCA_FLOWER_KEY_MPLS_OPT_LSE_TC]       = { .type = NLA_U8 },
 803         [TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL]    = { .type = NLA_U32 },
 804 };
 805
 806 static const struct nla_policy
 807 cfm_opt_policy[TCA_FLOWER_KEY_CFM_OPT_MAX + 1] = {
 808         [TCA_FLOWER_KEY_CFM_MD_LEVEL]   = NLA_POLICY_MAX(NLA_U8,
 809                                                 FLOW_DIS_CFM_MDL_MAX),
 810         [TCA_FLOWER_KEY_CFM_OPCODE]     = { .type = NLA_U8 },
 811 };
 812
 813 static void fl_set_key_val(struct nlattr **tb,
 814                            void *val, int val_type,
 815                            void *mask, int mask_type, int len)
 816 {
 817         if (!tb[val_type])
 818                 return;
 819         nla_memcpy(val, tb[val_type], len);
 820         if (mask_type == TCA_FLOWER_UNSPEC || !tb[mask_type])
 821                 memset(mask, 0xff, len);
 822         else
 823                 nla_memcpy(mask, tb[mask_type], len);
 824 }
 825
 826 static int fl_set_key_spi(struct nlattr **tb, struct fl_flow_key *key,
 827                           struct fl_flow_key *mask,
 828                           struct netlink_ext_ack *extack)
 829 {
 830         if (key->basic.ip_proto != IPPROTO_ESP &&
 831             key->basic.ip_proto != IPPROTO_AH) {
 832                 NL_SET_ERR_MSG(extack,
 833                                "Protocol must be either ESP or AH");
 834                 return -EINVAL;
 835         }
 836
 837         fl_set_key_val(tb, &key->ipsec.spi,
 838                        TCA_FLOWER_KEY_SPI,
 839                        &mask->ipsec.spi, TCA_FLOWER_KEY_SPI_MASK,
 840                        sizeof(key->ipsec.spi));
 841         return 0;
 842 }
 843
 844 static int fl_set_key_port_range(struct nlattr **tb, struct fl_flow_key *key,
 845                                  struct fl_flow_key *mask,
 846                                  struct netlink_ext_ack *extack)
 847 {
 848         fl_set_key_val(tb, &key->tp_range.tp_min.dst,
 849                        TCA_FLOWER_KEY_PORT_DST_MIN, &mask->tp_range.tp_min.dst,
 850                        TCA_FLOWER_UNSPEC, sizeof(key->tp_range.tp_min.dst));
 851         fl_set_key_val(tb, &key->tp_range.tp_max.dst,
 852                        TCA_FLOWER_KEY_PORT_DST_MAX, &mask->tp_range.tp_max.dst,
 853                        TCA_FLOWER_UNSPEC, sizeof(key->tp_range.tp_max.dst));
 854         fl_set_key_val(tb, &key->tp_range.tp_min.src,
 855                        TCA_FLOWER_KEY_PORT_SRC_MIN, &mask->tp_range.tp_min.src,
 856                        TCA_FLOWER_UNSPEC, sizeof(key->tp_range.tp_min.src));
 857         fl_set_key_val(tb, &key->tp_range.tp_max.src,
 858                        TCA_FLOWER_KEY_PORT_SRC_MAX, &mask->tp_range.tp_max.src,
 859                        TCA_FLOWER_UNSPEC, sizeof(key->tp_range.tp_max.src));
 860
 861         if (mask->tp_range.tp_min.dst != mask->tp_range.tp_max.dst) {
 862                 NL_SET_ERR_MSG(extack,
 863                                "Both min and max destination ports must be specified");
 864                 return -EINVAL;
 865         }
 866         if (mask->tp_range.tp_min.src != mask->tp_range.tp_max.src) {
 867                 NL_SET_ERR_MSG(extack,
 868                                "Both min and max source ports must be specified");
 869                 return -EINVAL;
 870         }
 871         if (mask->tp_range.tp_min.dst && mask->tp_range.tp_max.dst &&
 872             ntohs(key->tp_range.tp_max.dst) <=
 873             ntohs(key->tp_range.tp_min.dst)) {
 874                 NL_SET_ERR_MSG_ATTR(extack,
 875                                     tb[TCA_FLOWER_KEY_PORT_DST_MIN],
 876                                     "Invalid destination port range (min must be strictly smaller than max)");
 877                 return -EINVAL;
 878         }
 879         if (mask->tp_range.tp_min.src && mask->tp_range.tp_max.src &&
 880             ntohs(key->tp_range.tp_max.src) <=
 881             ntohs(key->tp_range.tp_min.src)) {
 882                 NL_SET_ERR_MSG_ATTR(extack,
 883                                     tb[TCA_FLOWER_KEY_PORT_SRC_MIN],
 884                                     "Invalid source port range (min must be strictly smaller than max)");
 885                 return -EINVAL;
 886         }
 887
 888         return 0;
 889 }
 890
 891 static int fl_set_key_mpls_lse(const struct nlattr *nla_lse,
 892                                struct flow_dissector_key_mpls *key_val,
 893                                struct flow_dissector_key_mpls *key_mask,
 894                                struct netlink_ext_ack *extack)
 895 {
 896         struct nlattr *tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_MAX + 1];
 897         struct flow_dissector_mpls_lse *lse_mask;
 898         struct flow_dissector_mpls_lse *lse_val;
 899         u8 lse_index;
 900         u8 depth;
 901         int err;
 902
 903         err = nla_parse_nested(tb, TCA_FLOWER_KEY_MPLS_OPT_LSE_MAX, nla_lse,
 904                                mpls_stack_entry_policy, extack);
 905         if (err < 0)
 906                 return err;
 907
 908         if (!tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH]) {
 909                 NL_SET_ERR_MSG(extack, "Missing MPLS option \"depth\"");
 910                 return -EINVAL;
 911         }
 912
 913         depth = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH]);
 914
 915         /* LSE depth starts at 1, for consistency with terminology used by
 916          * RFC 3031 (section 3.9), where depth 0 refers to unlabeled packets.
 917          */
 918         if (depth < 1 || depth > FLOW_DIS_MPLS_MAX) {
 919                 NL_SET_ERR_MSG_ATTR(extack,
 920                                     tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH],
 921                                     "Invalid MPLS depth");
 922                 return -EINVAL;
 923         }
 924         lse_index = depth - 1;
 925
 926         dissector_set_mpls_lse(key_val, lse_index);
 927         dissector_set_mpls_lse(key_mask, lse_index);
 928
 929         lse_val = &key_val->ls[lse_index];
 930         lse_mask = &key_mask->ls[lse_index];
 931
 932         if (tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_TTL]) {
 933                 lse_val->mpls_ttl = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_TTL]);
 934                 lse_mask->mpls_ttl = MPLS_TTL_MASK;
 935         }
 936         if (tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS]) {
 937                 u8 bos = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS]);
 938
 939                 if (bos & ~MPLS_BOS_MASK) {
 940                         NL_SET_ERR_MSG_ATTR(extack,
 941                                             tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS],
 942                                             "Bottom Of Stack (BOS) must be 0 or 1");
 943                         return -EINVAL;
 944                 }
 945                 lse_val->mpls_bos = bos;
 946                 lse_mask->mpls_bos = MPLS_BOS_MASK;
 947         }
 948         if (tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_TC]) {
 949                 u8 tc = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_TC]);
 950
 951                 if (tc & ~MPLS_TC_MASK) {
 952                         NL_SET_ERR_MSG_ATTR(extack,
 953                                             tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_TC],
 954                                             "Traffic Class (TC) must be between 0 and 7");
 955                         return -EINVAL;
 956                 }
 957                 lse_val->mpls_tc = tc;
 958                 lse_mask->mpls_tc = MPLS_TC_MASK;
 959         }
 960         if (tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL]) {
 961                 u32 label = nla_get_u32(tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL]);
 962
 963                 if (label & ~MPLS_LABEL_MASK) {
 964                         NL_SET_ERR_MSG_ATTR(extack,
 965                                             tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL],
 966                                             "Label must be between 0 and 1048575");
 967                         return -EINVAL;
 968                 }
 969                 lse_val->mpls_label = label;
 970                 lse_mask->mpls_label = MPLS_LABEL_MASK;
 971         }
 972
 973         return 0;
 974 }
 975
 976 static int fl_set_key_mpls_opts(const struct nlattr *nla_mpls_opts,
 977                                 struct flow_dissector_key_mpls *key_val,
 978                                 struct flow_dissector_key_mpls *key_mask,
 979                                 struct netlink_ext_ack *extack)
 980 {
 981         struct nlattr *nla_lse;
 982         int rem;
 983         int err;
 984
 985         if (!(nla_mpls_opts->nla_type & NLA_F_NESTED)) {
 986                 NL_SET_ERR_MSG_ATTR(extack, nla_mpls_opts,
 987                                     "NLA_F_NESTED is missing");
 988                 return -EINVAL;
 989         }
 990
 991         nla_for_each_nested(nla_lse, nla_mpls_opts, rem) {
 992                 if (nla_type(nla_lse) != TCA_FLOWER_KEY_MPLS_OPTS_LSE) {
 993                         NL_SET_ERR_MSG_ATTR(extack, nla_lse,
 994                                             "Invalid MPLS option type");
 995                         return -EINVAL;
 996                 }
 997
 998                 err = fl_set_key_mpls_lse(nla_lse, key_val, key_mask, extack);
 999                 if (err < 0)
1000                         return err;
1001         }
1002         if (rem) {
1003                 NL_SET_ERR_MSG(extack,
1004                                "Bytes leftover after parsing MPLS options");
1005                 return -EINVAL;
1006         }
1007
1008         return 0;
1009 }
1010
1011 static int fl_set_key_mpls(struct nlattr **tb,
1012                            struct flow_dissector_key_mpls *key_val,
1013                            struct flow_dissector_key_mpls *key_mask,
1014                            struct netlink_ext_ack *extack)
1015 {
1016         struct flow_dissector_mpls_lse *lse_mask;
1017         struct flow_dissector_mpls_lse *lse_val;
1018
1019         if (tb[TCA_FLOWER_KEY_MPLS_OPTS]) {
1020                 if (tb[TCA_FLOWER_KEY_MPLS_TTL] ||
1021                     tb[TCA_FLOWER_KEY_MPLS_BOS] ||
1022                     tb[TCA_FLOWER_KEY_MPLS_TC] ||
1023                     tb[TCA_FLOWER_KEY_MPLS_LABEL]) {
1024                         NL_SET_ERR_MSG_ATTR(extack,
1025                                             tb[TCA_FLOWER_KEY_MPLS_OPTS],
1026                                             "MPLS label, Traffic Class, Bottom Of Stack and Time To Live must be encapsulated in the MPLS options attribute");
1027                         return -EBADMSG;
1028                 }
1029
1030                 return fl_set_key_mpls_opts(tb[TCA_FLOWER_KEY_MPLS_OPTS],
1031                                             key_val, key_mask, extack);
1032         }
1033
1034         lse_val = &key_val->ls[0];
1035         lse_mask = &key_mask->ls[0];
1036
1037         if (tb[TCA_FLOWER_KEY_MPLS_TTL]) {
1038                 lse_val->mpls_ttl = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_TTL]);
1039                 lse_mask->mpls_ttl = MPLS_TTL_MASK;
1040                 dissector_set_mpls_lse(key_val, 0);
1041                 dissector_set_mpls_lse(key_mask, 0);
1042         }
1043         if (tb[TCA_FLOWER_KEY_MPLS_BOS]) {
1044                 u8 bos = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_BOS]);
1045
1046                 if (bos & ~MPLS_BOS_MASK) {
1047                         NL_SET_ERR_MSG_ATTR(extack,
1048                                             tb[TCA_FLOWER_KEY_MPLS_BOS],
1049                                             "Bottom Of Stack (BOS) must be 0 or 1");
1050                         return -EINVAL;
1051                 }
1052                 lse_val->mpls_bos = bos;
1053                 lse_mask->mpls_bos = MPLS_BOS_MASK;
1054                 dissector_set_mpls_lse(key_val, 0);
1055                 dissector_set_mpls_lse(key_mask, 0);
1056         }
1057         if (tb[TCA_FLOWER_KEY_MPLS_TC]) {
1058                 u8 tc = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_TC]);
1059
1060                 if (tc & ~MPLS_TC_MASK) {
1061                         NL_SET_ERR_MSG_ATTR(extack,
1062                                             tb[TCA_FLOWER_KEY_MPLS_TC],
1063                                             "Traffic Class (TC) must be between 0 and 7");
1064                         return -EINVAL;
1065                 }
1066                 lse_val->mpls_tc = tc;
1067                 lse_mask->mpls_tc = MPLS_TC_MASK;
1068                 dissector_set_mpls_lse(key_val, 0);
1069                 dissector_set_mpls_lse(key_mask, 0);
1070         }
1071         if (tb[TCA_FLOWER_KEY_MPLS_LABEL]) {
1072                 u32 label = nla_get_u32(tb[TCA_FLOWER_KEY_MPLS_LABEL]);
1073
1074                 if (label & ~MPLS_LABEL_MASK) {
1075                         NL_SET_ERR_MSG_ATTR(extack,
1076                                             tb[TCA_FLOWER_KEY_MPLS_LABEL],
1077                                             "Label must be between 0 and 1048575");
1078                         return -EINVAL;
1079                 }
1080                 lse_val->mpls_label = label;
1081                 lse_mask->mpls_label = MPLS_LABEL_MASK;
1082                 dissector_set_mpls_lse(key_val, 0);
1083                 dissector_set_mpls_lse(key_mask, 0);
1084         }
1085         return 0;
1086 }
1087
1088 static void fl_set_key_vlan(struct nlattr **tb,
1089                             __be16 ethertype,
1090                             int vlan_id_key, int vlan_prio_key,
1091                             int vlan_next_eth_type_key,
1092                             struct flow_dissector_key_vlan *key_val,
1093                             struct flow_dissector_key_vlan *key_mask)
1094 {
1095 #define VLAN_PRIORITY_MASK      0x7
1096
1097         if (tb[vlan_id_key]) {
1098                 key_val->vlan_id =
1099                         nla_get_u16(tb[vlan_id_key]) & VLAN_VID_MASK;
1100                 key_mask->vlan_id = VLAN_VID_MASK;
1101         }
1102         if (tb[vlan_prio_key]) {
1103                 key_val->vlan_priority =
1104                         nla_get_u8(tb[vlan_prio_key]) &
1105                         VLAN_PRIORITY_MASK;
1106                 key_mask->vlan_priority = VLAN_PRIORITY_MASK;
1107         }
1108         if (ethertype) {
1109                 key_val->vlan_tpid = ethertype;
1110                 key_mask->vlan_tpid = cpu_to_be16(~0);
1111         }
1112         if (tb[vlan_next_eth_type_key]) {
1113                 key_val->vlan_eth_type =
1114                         nla_get_be16(tb[vlan_next_eth_type_key]);
1115                 key_mask->vlan_eth_type = cpu_to_be16(~0);
1116         }
1117 }
1118
1119 static void fl_set_key_pppoe(struct nlattr **tb,
1120                              struct flow_dissector_key_pppoe *key_val,
1121                              struct flow_dissector_key_pppoe *key_mask,
1122                              struct fl_flow_key *key,
1123                              struct fl_flow_key *mask)
1124 {
1125         /* key_val::type must be set to ETH_P_PPP_SES
1126          * because ETH_P_PPP_SES was stored in basic.n_proto
1127          * which might get overwritten by ppp_proto
1128          * or might be set to 0, the role of key_val::type
1129          * is similar to vlan_key::tpid
1130          */
1131         key_val->type = htons(ETH_P_PPP_SES);
1132         key_mask->type = cpu_to_be16(~0);
1133
1134         if (tb[TCA_FLOWER_KEY_PPPOE_SID]) {
1135                 key_val->session_id =
1136                         nla_get_be16(tb[TCA_FLOWER_KEY_PPPOE_SID]);
1137                 key_mask->session_id = cpu_to_be16(~0);
1138         }
1139         if (tb[TCA_FLOWER_KEY_PPP_PROTO]) {
1140                 key_val->ppp_proto =
1141                         nla_get_be16(tb[TCA_FLOWER_KEY_PPP_PROTO]);
1142                 key_mask->ppp_proto = cpu_to_be16(~0);
1143
1144                 if (key_val->ppp_proto == htons(PPP_IP)) {
1145                         key->basic.n_proto = htons(ETH_P_IP);
1146                         mask->basic.n_proto = cpu_to_be16(~0);
1147                 } else if (key_val->ppp_proto == htons(PPP_IPV6)) {
1148                         key->basic.n_proto = htons(ETH_P_IPV6);
1149                         mask->basic.n_proto = cpu_to_be16(~0);
1150                 } else if (key_val->ppp_proto == htons(PPP_MPLS_UC)) {
1151                         key->basic.n_proto = htons(ETH_P_MPLS_UC);
1152                         mask->basic.n_proto = cpu_to_be16(~0);
1153                 } else if (key_val->ppp_proto == htons(PPP_MPLS_MC)) {
1154                         key->basic.n_proto = htons(ETH_P_MPLS_MC);
1155                         mask->basic.n_proto = cpu_to_be16(~0);
1156                 }
1157         } else {
1158                 key->basic.n_proto = 0;
1159                 mask->basic.n_proto = cpu_to_be16(0);
1160         }
1161 }
1162
1163 static void fl_set_key_flag(u32 flower_key, u32 flower_mask,
1164                             u32 *dissector_key, u32 *dissector_mask,
1165                             u32 flower_flag_bit, u32 dissector_flag_bit)
1166 {
1167         if (flower_mask & flower_flag_bit) {
1168                 *dissector_mask |= dissector_flag_bit;
1169                 if (flower_key & flower_flag_bit)
1170                         *dissector_key |= dissector_flag_bit;
1171         }
1172 }
1173
1174 static int fl_set_key_flags(struct nlattr *tca_opts, struct nlattr **tb,
1175                             bool encap, u32 *flags_key, u32 *flags_mask,
1176                             struct netlink_ext_ack *extack)
1177 {
1178         int fl_key, fl_mask;
1179         u32 key, mask;
1180
1181         if (encap) {
1182                 fl_key = TCA_FLOWER_KEY_ENC_FLAGS;
1183                 fl_mask = TCA_FLOWER_KEY_ENC_FLAGS_MASK;
1184         } else {
1185                 fl_key = TCA_FLOWER_KEY_FLAGS;
1186                 fl_mask = TCA_FLOWER_KEY_FLAGS_MASK;
1187         }
1188
1189         /* mask is mandatory for flags */
1190         if (NL_REQ_ATTR_CHECK(extack, tca_opts, tb, fl_mask)) {
1191                 NL_SET_ERR_MSG(extack, "Missing flags mask");
1192                 return -EINVAL;
1193         }
1194
1195         key = be32_to_cpu(nla_get_be32(tb[fl_key]));
1196         mask = be32_to_cpu(nla_get_be32(tb[fl_mask]));
1197
1198         *flags_key  = 0;
1199         *flags_mask = 0;
1200
1201         fl_set_key_flag(key, mask, flags_key, flags_mask,
1202                         TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT, FLOW_DIS_IS_FRAGMENT);
1203         fl_set_key_flag(key, mask, flags_key, flags_mask,
1204                         TCA_FLOWER_KEY_FLAGS_FRAG_IS_FIRST,
1205                         FLOW_DIS_FIRST_FRAG);
1206
1207         fl_set_key_flag(key, mask, flags_key, flags_mask,
1208                         TCA_FLOWER_KEY_FLAGS_TUNNEL_CSUM,
1209                         FLOW_DIS_F_TUNNEL_CSUM);
1210
1211         fl_set_key_flag(key, mask, flags_key, flags_mask,
1212                         TCA_FLOWER_KEY_FLAGS_TUNNEL_DONT_FRAGMENT,
1213                         FLOW_DIS_F_TUNNEL_DONT_FRAGMENT);
1214
1215         fl_set_key_flag(key, mask, flags_key, flags_mask,
1216                         TCA_FLOWER_KEY_FLAGS_TUNNEL_OAM, FLOW_DIS_F_TUNNEL_OAM);
1217
1218         fl_set_key_flag(key, mask, flags_key, flags_mask,
1219                         TCA_FLOWER_KEY_FLAGS_TUNNEL_CRIT_OPT,
1220                         FLOW_DIS_F_TUNNEL_CRIT_OPT);
1221
1222         return 0;
1223 }
1224
1225 static void fl_set_key_ip(struct nlattr **tb, bool encap,
1226                           struct flow_dissector_key_ip *key,
1227                           struct flow_dissector_key_ip *mask)
1228 {
1229         int tos_key = encap ? TCA_FLOWER_KEY_ENC_IP_TOS : TCA_FLOWER_KEY_IP_TOS;
1230         int ttl_key = encap ? TCA_FLOWER_KEY_ENC_IP_TTL : TCA_FLOWER_KEY_IP_TTL;
1231         int tos_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TOS_MASK : TCA_FLOWER_KEY_IP_TOS_MASK;
1232         int ttl_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TTL_MASK : TCA_FLOWER_KEY_IP_TTL_MASK;
1233
1234         fl_set_key_val(tb, &key->tos, tos_key, &mask->tos, tos_mask, sizeof(key->tos));
1235         fl_set_key_val(tb, &key->ttl, ttl_key, &mask->ttl, ttl_mask, sizeof(key->ttl));
1236 }
1237
1238 static int fl_set_geneve_opt(const struct nlattr *nla, struct fl_flow_key *key,
1239                              int depth, int option_len,
1240                              struct netlink_ext_ack *extack)
1241 {
1242         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX + 1];
1243         struct nlattr *class = NULL, *type = NULL, *data = NULL;
1244         struct geneve_opt *opt;
1245         int err, data_len = 0;
1246
1247         if (option_len > sizeof(struct geneve_opt))
1248                 data_len = option_len - sizeof(struct geneve_opt);
1249
1250         if (key->enc_opts.len > FLOW_DIS_TUN_OPTS_MAX - 4)
1251                 return -ERANGE;
1252
1253         opt = (struct geneve_opt *)&key->enc_opts.data[key->enc_opts.len];
1254         memset(opt, 0xff, option_len);
1255         opt->length = data_len / 4;
1256         opt->r1 = 0;
1257         opt->r2 = 0;
1258         opt->r3 = 0;
1259
1260         /* If no mask has been prodived we assume an exact match. */
1261         if (!depth)
1262                 return sizeof(struct geneve_opt) + data_len;
1263
1264         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_GENEVE) {
1265                 NL_SET_ERR_MSG(extack, "Non-geneve option type for mask");
1266                 return -EINVAL;
1267         }
1268
1269         err = nla_parse_nested_deprecated(tb,
1270                                           TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX,
1271                                           nla, geneve_opt_policy, extack);
1272         if (err < 0)
1273                 return err;
1274
1275         /* We are not allowed to omit any of CLASS, TYPE or DATA
1276          * fields from the key.
1277          */
1278         if (!option_len &&
1279             (!tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS] ||
1280              !tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE] ||
1281              !tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA])) {
1282                 NL_SET_ERR_MSG(extack, "Missing tunnel key geneve option class, type or data");
1283                 return -EINVAL;
1284         }
1285
1286         /* Omitting any of CLASS, TYPE or DATA fields is allowed
1287          * for the mask.
1288          */
1289         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA]) {
1290                 int new_len = key->enc_opts.len;
1291
1292                 data = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA];
1293                 data_len = nla_len(data);
1294                 if (data_len < 4) {
1295                         NL_SET_ERR_MSG(extack, "Tunnel key geneve option data is less than 4 bytes long");
1296                         return -ERANGE;
1297                 }
1298                 if (data_len % 4) {
1299                         NL_SET_ERR_MSG(extack, "Tunnel key geneve option data is not a multiple of 4 bytes long");
1300                         return -ERANGE;
1301                 }
1302
1303                 new_len += sizeof(struct geneve_opt) + data_len;
1304                 BUILD_BUG_ON(FLOW_DIS_TUN_OPTS_MAX != IP_TUNNEL_OPTS_MAX);
1305                 if (new_len > FLOW_DIS_TUN_OPTS_MAX) {
1306                         NL_SET_ERR_MSG(extack, "Tunnel options exceeds max size");
1307                         return -ERANGE;
1308                 }
1309                 opt->length = data_len / 4;
1310                 memcpy(opt->opt_data, nla_data(data), data_len);
1311         }
1312
1313         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS]) {
1314                 class = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS];
1315                 opt->opt_class = nla_get_be16(class);
1316         }
1317
1318         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE]) {
1319                 type = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE];
1320                 opt->type = nla_get_u8(type);
1321         }
1322
1323         return sizeof(struct geneve_opt) + data_len;
1324 }
1325
1326 static int fl_set_vxlan_opt(const struct nlattr *nla, struct fl_flow_key *key,
1327                             int depth, int option_len,
1328                             struct netlink_ext_ack *extack)
1329 {
1330         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_MAX + 1];
1331         struct vxlan_metadata *md;
1332         int err;
1333
1334         md = (struct vxlan_metadata *)&key->enc_opts.data[key->enc_opts.len];
1335         memset(md, 0xff, sizeof(*md));
1336
1337         if (!depth)
1338                 return sizeof(*md);
1339
1340         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_VXLAN) {
1341                 NL_SET_ERR_MSG(extack, "Non-vxlan option type for mask");
1342                 return -EINVAL;
1343         }
1344
1345         err = nla_parse_nested(tb, TCA_FLOWER_KEY_ENC_OPT_VXLAN_MAX, nla,
1346                                vxlan_opt_policy, extack);
1347         if (err < 0)
1348                 return err;
1349
1350         if (!option_len && !tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]) {
1351                 NL_SET_ERR_MSG(extack, "Missing tunnel key vxlan option gbp");
1352                 return -EINVAL;
1353         }
1354
1355         if (tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]) {
1356                 md->gbp = nla_get_u32(tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]);
1357                 md->gbp &= VXLAN_GBP_MASK;
1358         }
1359
1360         return sizeof(*md);
1361 }
1362
1363 static int fl_set_erspan_opt(const struct nlattr *nla, struct fl_flow_key *key,
1364                              int depth, int option_len,
1365                              struct netlink_ext_ack *extack)
1366 {
1367         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_MAX + 1];
1368         struct erspan_metadata *md;
1369         int err;
1370
1371         md = (struct erspan_metadata *)&key->enc_opts.data[key->enc_opts.len];
1372         md->version = 1;
1373
1374         if (!depth)
1375                 return sizeof(*md);
1376
1377         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_ERSPAN) {
1378                 NL_SET_ERR_MSG(extack, "Non-erspan option type for mask");
1379                 return -EINVAL;
1380         }
1381
1382         err = nla_parse_nested(tb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_MAX, nla,
1383                                erspan_opt_policy, extack);
1384         if (err < 0)
1385                 return err;
1386
1387         if (!option_len && !tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER]) {
1388                 NL_SET_ERR_MSG(extack, "Missing tunnel key erspan option ver");
1389                 return -EINVAL;
1390         }
1391
1392         if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER])
1393                 md->version = nla_get_u8(tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER]);
1394
1395         if (md->version == 1) {
1396                 if (!option_len && !tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX]) {
1397                         NL_SET_ERR_MSG(extack, "Missing tunnel key erspan option index");
1398                         return -EINVAL;
1399                 }
1400                 memset(&md->u.index, 0xff, sizeof(md->u.index));
1401                 if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX]) {
1402                         nla = tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX];
1403                         md->u.index = nla_get_be32(nla);
1404                 }
1405         } else if (md->version == 2) {
1406                 if (!option_len && (!tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR] ||
1407                                     !tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID])) {
1408                         NL_SET_ERR_MSG(extack, "Missing tunnel key erspan option dir or hwid");
1409                         return -EINVAL;
1410                 }
1411                 md->u.md2.dir = 1;
1412                 if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR]) {
1413                         nla = tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR];
1414                         md->u.md2.dir = nla_get_u8(nla);
1415                 }
1416                 set_hwid(&md->u.md2, 0xff);
1417                 if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID]) {
1418                         nla = tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID];
1419                         set_hwid(&md->u.md2, nla_get_u8(nla));
1420                 }
1421         } else {
1422                 NL_SET_ERR_MSG(extack, "Tunnel key erspan option ver is incorrect");
1423                 return -EINVAL;
1424         }
1425
1426         return sizeof(*md);
1427 }
1428
1429 static int fl_set_gtp_opt(const struct nlattr *nla, struct fl_flow_key *key,
1430                           int depth, int option_len,
1431                           struct netlink_ext_ack *extack)
1432 {
1433         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_GTP_MAX + 1];
1434         struct gtp_pdu_session_info *sinfo;
1435         u8 len = key->enc_opts.len;
1436         int err;
1437
1438         sinfo = (struct gtp_pdu_session_info *)&key->enc_opts.data[len];
1439         memset(sinfo, 0xff, option_len);
1440
1441         if (!depth)
1442                 return sizeof(*sinfo);
1443
1444         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_GTP) {
1445                 NL_SET_ERR_MSG_MOD(extack, "Non-gtp option type for mask");
1446                 return -EINVAL;
1447         }
1448
1449         err = nla_parse_nested(tb, TCA_FLOWER_KEY_ENC_OPT_GTP_MAX, nla,
1450                                gtp_opt_policy, extack);
1451         if (err < 0)
1452                 return err;
1453
1454         if (!option_len &&
1455             (!tb[TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE] ||
1456              !tb[TCA_FLOWER_KEY_ENC_OPT_GTP_QFI])) {
1457                 NL_SET_ERR_MSG_MOD(extack,
1458                                    "Missing tunnel key gtp option pdu type or qfi");
1459                 return -EINVAL;
1460         }
1461
1462         if (tb[TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE])
1463                 sinfo->pdu_type =
1464                         nla_get_u8(tb[TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE]);
1465
1466         if (tb[TCA_FLOWER_KEY_ENC_OPT_GTP_QFI])
1467                 sinfo->qfi = nla_get_u8(tb[TCA_FLOWER_KEY_ENC_OPT_GTP_QFI]);
1468
1469         return sizeof(*sinfo);
1470 }
1471
1472 static int fl_set_pfcp_opt(const struct nlattr *nla, struct fl_flow_key *key,
1473                            int depth, int option_len,
1474                            struct netlink_ext_ack *extack)
1475 {
1476         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_PFCP_MAX + 1];
1477         struct pfcp_metadata *md;
1478         int err;
1479
1480         md = (struct pfcp_metadata *)&key->enc_opts.data[key->enc_opts.len];
1481         memset(md, 0xff, sizeof(*md));
1482
1483         if (!depth)
1484                 return sizeof(*md);
1485
1486         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_PFCP) {
1487                 NL_SET_ERR_MSG_MOD(extack, "Non-pfcp option type for mask");
1488                 return -EINVAL;
1489         }
1490
1491         err = nla_parse_nested(tb, TCA_FLOWER_KEY_ENC_OPT_PFCP_MAX, nla,
1492                                pfcp_opt_policy, extack);
1493         if (err < 0)
1494                 return err;
1495
1496         if (!option_len && !tb[TCA_FLOWER_KEY_ENC_OPT_PFCP_TYPE]) {
1497                 NL_SET_ERR_MSG_MOD(extack, "Missing tunnel key pfcp option type");
1498                 return -EINVAL;
1499         }
1500
1501         if (tb[TCA_FLOWER_KEY_ENC_OPT_PFCP_TYPE])
1502                 md->type = nla_get_u8(tb[TCA_FLOWER_KEY_ENC_OPT_PFCP_TYPE]);
1503
1504         if (tb[TCA_FLOWER_KEY_ENC_OPT_PFCP_SEID])
1505                 md->seid = nla_get_be64(tb[TCA_FLOWER_KEY_ENC_OPT_PFCP_SEID]);
1506
1507         return sizeof(*md);
1508 }
1509
1510 static int fl_set_enc_opt(struct nlattr **tb, struct fl_flow_key *key,
1511                           struct fl_flow_key *mask,
1512                           struct netlink_ext_ack *extack)
1513 {
1514         const struct nlattr *nla_enc_key, *nla_opt_key, *nla_opt_msk = NULL;
1515         int err, option_len, key_depth, msk_depth = 0;
1516
1517         err = nla_validate_nested_deprecated(tb[TCA_FLOWER_KEY_ENC_OPTS],
1518                                              TCA_FLOWER_KEY_ENC_OPTS_MAX,
1519                                              enc_opts_policy, extack);
1520         if (err)
1521                 return err;
1522
1523         nla_enc_key = nla_data(tb[TCA_FLOWER_KEY_ENC_OPTS]);
1524
1525         if (tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]) {
1526                 err = nla_validate_nested_deprecated(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK],
1527                                                      TCA_FLOWER_KEY_ENC_OPTS_MAX,
1528                                                      enc_opts_policy, extack);
1529                 if (err)
1530                         return err;
1531
1532                 nla_opt_msk = nla_data(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]);
1533                 msk_depth = nla_len(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]);
1534                 if (!nla_ok(nla_opt_msk, msk_depth)) {
1535                         NL_SET_ERR_MSG(extack, "Invalid nested attribute for masks");
1536                         return -EINVAL;
1537                 }
1538         }
1539
1540         nla_for_each_attr(nla_opt_key, nla_enc_key,
1541                           nla_len(tb[TCA_FLOWER_KEY_ENC_OPTS]), key_depth) {
1542                 switch (nla_type(nla_opt_key)) {
1543                 case TCA_FLOWER_KEY_ENC_OPTS_GENEVE:
1544                         if (key->enc_opts.dst_opt_type &&
1545                             key->enc_opts.dst_opt_type !=
1546                             IP_TUNNEL_GENEVE_OPT_BIT) {
1547                                 NL_SET_ERR_MSG(extack, "Duplicate type for geneve options");
1548                                 return -EINVAL;
1549                         }
1550                         option_len = 0;
1551                         key->enc_opts.dst_opt_type = IP_TUNNEL_GENEVE_OPT_BIT;
1552                         option_len = fl_set_geneve_opt(nla_opt_key, key,
1553                                                        key_depth, option_len,
1554                                                        extack);
1555                         if (option_len < 0)
1556                                 return option_len;
1557
1558                         key->enc_opts.len += option_len;
1559                         /* At the same time we need to parse through the mask
1560                          * in order to verify exact and mask attribute lengths.
1561                          */
1562                         mask->enc_opts.dst_opt_type = IP_TUNNEL_GENEVE_OPT_BIT;
1563                         option_len = fl_set_geneve_opt(nla_opt_msk, mask,
1564                                                        msk_depth, option_len,
1565                                                        extack);
1566                         if (option_len < 0)
1567                                 return option_len;
1568
1569                         mask->enc_opts.len += option_len;
1570                         if (key->enc_opts.len != mask->enc_opts.len) {
1571                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
1572                                 return -EINVAL;
1573                         }
1574                         break;
1575                 case TCA_FLOWER_KEY_ENC_OPTS_VXLAN:
1576                         if (key->enc_opts.dst_opt_type) {
1577                                 NL_SET_ERR_MSG(extack, "Duplicate type for vxlan options");
1578                                 return -EINVAL;
1579                         }
1580                         option_len = 0;
1581                         key->enc_opts.dst_opt_type = IP_TUNNEL_VXLAN_OPT_BIT;
1582                         option_len = fl_set_vxlan_opt(nla_opt_key, key,
1583                                                       key_depth, option_len,
1584                                                       extack);
1585                         if (option_len < 0)
1586                                 return option_len;
1587
1588                         key->enc_opts.len += option_len;
1589                         /* At the same time we need to parse through the mask
1590                          * in order to verify exact and mask attribute lengths.
1591                          */
1592                         mask->enc_opts.dst_opt_type = IP_TUNNEL_VXLAN_OPT_BIT;
1593                         option_len = fl_set_vxlan_opt(nla_opt_msk, mask,
1594                                                       msk_depth, option_len,
1595                                                       extack);
1596                         if (option_len < 0)
1597                                 return option_len;
1598
1599                         mask->enc_opts.len += option_len;
1600                         if (key->enc_opts.len != mask->enc_opts.len) {
1601                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
1602                                 return -EINVAL;
1603                         }
1604                         break;
1605                 case TCA_FLOWER_KEY_ENC_OPTS_ERSPAN:
1606                         if (key->enc_opts.dst_opt_type) {
1607                                 NL_SET_ERR_MSG(extack, "Duplicate type for erspan options");
1608                                 return -EINVAL;
1609                         }
1610                         option_len = 0;
1611                         key->enc_opts.dst_opt_type = IP_TUNNEL_ERSPAN_OPT_BIT;
1612                         option_len = fl_set_erspan_opt(nla_opt_key, key,
1613                                                        key_depth, option_len,
1614                                                        extack);
1615                         if (option_len < 0)
1616                                 return option_len;
1617
1618                         key->enc_opts.len += option_len;
1619                         /* At the same time we need to parse through the mask
1620                          * in order to verify exact and mask attribute lengths.
1621                          */
1622                         mask->enc_opts.dst_opt_type = IP_TUNNEL_ERSPAN_OPT_BIT;
1623                         option_len = fl_set_erspan_opt(nla_opt_msk, mask,
1624                                                        msk_depth, option_len,
1625                                                        extack);
1626                         if (option_len < 0)
1627                                 return option_len;
1628
1629                         mask->enc_opts.len += option_len;
1630                         if (key->enc_opts.len != mask->enc_opts.len) {
1631                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
1632                                 return -EINVAL;
1633                         }
1634                         break;
1635                 case TCA_FLOWER_KEY_ENC_OPTS_GTP:
1636                         if (key->enc_opts.dst_opt_type) {
1637                                 NL_SET_ERR_MSG_MOD(extack,
1638                                                    "Duplicate type for gtp options");
1639                                 return -EINVAL;
1640                         }
1641                         option_len = 0;
1642                         key->enc_opts.dst_opt_type = IP_TUNNEL_GTP_OPT_BIT;
1643                         option_len = fl_set_gtp_opt(nla_opt_key, key,
1644                                                     key_depth, option_len,
1645                                                     extack);
1646                         if (option_len < 0)
1647                                 return option_len;
1648
1649                         key->enc_opts.len += option_len;
1650                         /* At the same time we need to parse through the mask
1651                          * in order to verify exact and mask attribute lengths.
1652                          */
1653                         mask->enc_opts.dst_opt_type = IP_TUNNEL_GTP_OPT_BIT;
1654                         option_len = fl_set_gtp_opt(nla_opt_msk, mask,
1655                                                     msk_depth, option_len,
1656                                                     extack);
1657                         if (option_len < 0)
1658                                 return option_len;
1659
1660                         mask->enc_opts.len += option_len;
1661                         if (key->enc_opts.len != mask->enc_opts.len) {
1662                                 NL_SET_ERR_MSG_MOD(extack,
1663                                                    "Key and mask miss aligned");
1664                                 return -EINVAL;
1665                         }
1666                         break;
1667                 case TCA_FLOWER_KEY_ENC_OPTS_PFCP:
1668                         if (key->enc_opts.dst_opt_type) {
1669                                 NL_SET_ERR_MSG_MOD(extack, "Duplicate type for pfcp options");
1670                                 return -EINVAL;
1671                         }
1672                         option_len = 0;
1673                         key->enc_opts.dst_opt_type = IP_TUNNEL_PFCP_OPT_BIT;
1674                         option_len = fl_set_pfcp_opt(nla_opt_key, key,
1675                                                      key_depth, option_len,
1676                                                      extack);
1677                         if (option_len < 0)
1678                                 return option_len;
1679
1680                         key->enc_opts.len += option_len;
1681                         /* At the same time we need to parse through the mask
1682                          * in order to verify exact and mask attribute lengths.
1683                          */
1684                         mask->enc_opts.dst_opt_type = IP_TUNNEL_PFCP_OPT_BIT;
1685                         option_len = fl_set_pfcp_opt(nla_opt_msk, mask,
1686                                                      msk_depth, option_len,
1687                                                      extack);
1688                         if (option_len < 0)
1689                                 return option_len;
1690
1691                         mask->enc_opts.len += option_len;
1692                         if (key->enc_opts.len != mask->enc_opts.len) {
1693                                 NL_SET_ERR_MSG_MOD(extack, "Key and mask miss aligned");
1694                                 return -EINVAL;
1695                         }
1696                         break;
1697                 default:
1698                         NL_SET_ERR_MSG(extack, "Unknown tunnel option type");
1699                         return -EINVAL;
1700                 }
1701
1702                 if (!msk_depth)
1703                         continue;
1704
1705                 if (!nla_ok(nla_opt_msk, msk_depth)) {
1706                         NL_SET_ERR_MSG(extack, "A mask attribute is invalid");
1707                         return -EINVAL;
1708                 }
1709                 nla_opt_msk = nla_next(nla_opt_msk, &msk_depth);
1710         }
1711
1712         return 0;
1713 }
1714
1715 static int fl_validate_ct_state(u16 state, struct nlattr *tb,
1716                                 struct netlink_ext_ack *extack)
1717 {
1718         if (state && !(state & TCA_FLOWER_KEY_CT_FLAGS_TRACKED)) {
1719                 NL_SET_ERR_MSG_ATTR(extack, tb,
1720                                     "no trk, so no other flag can be set");
1721                 return -EINVAL;
1722         }
1723
1724         if (state & TCA_FLOWER_KEY_CT_FLAGS_NEW &&
1725             state & TCA_FLOWER_KEY_CT_FLAGS_ESTABLISHED) {
1726                 NL_SET_ERR_MSG_ATTR(extack, tb,
1727                                     "new and est are mutually exclusive");
1728                 return -EINVAL;
1729         }
1730
1731         if (state & TCA_FLOWER_KEY_CT_FLAGS_INVALID &&
1732             state & ~(TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
1733                       TCA_FLOWER_KEY_CT_FLAGS_INVALID)) {
1734                 NL_SET_ERR_MSG_ATTR(extack, tb,
1735                                     "when inv is set, only trk may be set");
1736                 return -EINVAL;
1737         }
1738
1739         if (state & TCA_FLOWER_KEY_CT_FLAGS_NEW &&
1740             state & TCA_FLOWER_KEY_CT_FLAGS_REPLY) {
1741                 NL_SET_ERR_MSG_ATTR(extack, tb,
1742                                     "new and rpl are mutually exclusive");
1743                 return -EINVAL;
1744         }
1745
1746         return 0;
1747 }
1748
1749 static int fl_set_key_ct(struct nlattr **tb,
1750                          struct flow_dissector_key_ct *key,
1751                          struct flow_dissector_key_ct *mask,
1752                          struct netlink_ext_ack *extack)
1753 {
1754         if (tb[TCA_FLOWER_KEY_CT_STATE]) {
1755                 int err;
1756
1757                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK)) {
1758                         NL_SET_ERR_MSG(extack, "Conntrack isn't enabled");
1759                         return -EOPNOTSUPP;
1760                 }
1761                 fl_set_key_val(tb, &key->ct_state, TCA_FLOWER_KEY_CT_STATE,
1762                                &mask->ct_state, TCA_FLOWER_KEY_CT_STATE_MASK,
1763                                sizeof(key->ct_state));
1764
1765                 err = fl_validate_ct_state(key->ct_state & mask->ct_state,
1766                                            tb[TCA_FLOWER_KEY_CT_STATE_MASK],
1767                                            extack);
1768                 if (err)
1769                         return err;
1770
1771         }
1772         if (tb[TCA_FLOWER_KEY_CT_ZONE]) {
1773                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK_ZONES)) {
1774                         NL_SET_ERR_MSG(extack, "Conntrack zones isn't enabled");
1775                         return -EOPNOTSUPP;
1776                 }
1777                 fl_set_key_val(tb, &key->ct_zone, TCA_FLOWER_KEY_CT_ZONE,
1778                                &mask->ct_zone, TCA_FLOWER_KEY_CT_ZONE_MASK,
1779                                sizeof(key->ct_zone));
1780         }
1781         if (tb[TCA_FLOWER_KEY_CT_MARK]) {
1782                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK_MARK)) {
1783                         NL_SET_ERR_MSG(extack, "Conntrack mark isn't enabled");
1784                         return -EOPNOTSUPP;
1785                 }
1786                 fl_set_key_val(tb, &key->ct_mark, TCA_FLOWER_KEY_CT_MARK,
1787                                &mask->ct_mark, TCA_FLOWER_KEY_CT_MARK_MASK,
1788                                sizeof(key->ct_mark));
1789         }
1790         if (tb[TCA_FLOWER_KEY_CT_LABELS]) {
1791                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK_LABELS)) {
1792                         NL_SET_ERR_MSG(extack, "Conntrack labels aren't enabled");
1793                         return -EOPNOTSUPP;
1794                 }
1795                 fl_set_key_val(tb, key->ct_labels, TCA_FLOWER_KEY_CT_LABELS,
1796                                mask->ct_labels, TCA_FLOWER_KEY_CT_LABELS_MASK,
1797                                sizeof(key->ct_labels));
1798         }
1799
1800         return 0;
1801 }
1802
1803 static bool is_vlan_key(struct nlattr *tb, __be16 *ethertype,
1804                         struct fl_flow_key *key, struct fl_flow_key *mask,
1805                         int vthresh)
1806 {
1807         const bool good_num_of_vlans = key->num_of_vlans.num_of_vlans > vthresh;
1808
1809         if (!tb) {
1810                 *ethertype = 0;
1811                 return good_num_of_vlans;
1812         }
1813
1814         *ethertype = nla_get_be16(tb);
1815         if (good_num_of_vlans || eth_type_vlan(*ethertype))
1816                 return true;
1817
1818         key->basic.n_proto = *ethertype;
1819         mask->basic.n_proto = cpu_to_be16(~0);
1820         return false;
1821 }
1822
1823 static void fl_set_key_cfm_md_level(struct nlattr **tb,
1824                                     struct fl_flow_key *key,
1825                                     struct fl_flow_key *mask,
1826                                     struct netlink_ext_ack *extack)
1827 {
1828         u8 level;
1829
1830         if (!tb[TCA_FLOWER_KEY_CFM_MD_LEVEL])
1831                 return;
1832
1833         level = nla_get_u8(tb[TCA_FLOWER_KEY_CFM_MD_LEVEL]);
1834         key->cfm.mdl_ver = FIELD_PREP(FLOW_DIS_CFM_MDL_MASK, level);
1835         mask->cfm.mdl_ver = FLOW_DIS_CFM_MDL_MASK;
1836 }
1837
1838 static void fl_set_key_cfm_opcode(struct nlattr **tb,
1839                                   struct fl_flow_key *key,
1840                                   struct fl_flow_key *mask,
1841                                   struct netlink_ext_ack *extack)
1842 {
1843         fl_set_key_val(tb, &key->cfm.opcode, TCA_FLOWER_KEY_CFM_OPCODE,
1844                        &mask->cfm.opcode, TCA_FLOWER_UNSPEC,
1845                        sizeof(key->cfm.opcode));
1846 }
1847
1848 static int fl_set_key_cfm(struct nlattr **tb,
1849                           struct fl_flow_key *key,
1850                           struct fl_flow_key *mask,
1851                           struct netlink_ext_ack *extack)
1852 {
1853         struct nlattr *nla_cfm_opt[TCA_FLOWER_KEY_CFM_OPT_MAX + 1];
1854         int err;
1855
1856         if (!tb[TCA_FLOWER_KEY_CFM])
1857                 return 0;
1858
1859         err = nla_parse_nested(nla_cfm_opt, TCA_FLOWER_KEY_CFM_OPT_MAX,
1860                                tb[TCA_FLOWER_KEY_CFM], cfm_opt_policy, extack);
1861         if (err < 0)
1862                 return err;
1863
1864         fl_set_key_cfm_opcode(nla_cfm_opt, key, mask, extack);
1865         fl_set_key_cfm_md_level(nla_cfm_opt, key, mask, extack);
1866
1867         return 0;
1868 }
1869
1870 static int fl_set_key(struct net *net, struct nlattr *tca_opts,
1871                       struct nlattr **tb, struct fl_flow_key *key,
1872                       struct fl_flow_key *mask, struct netlink_ext_ack *extack)
1873 {
1874         __be16 ethertype;
1875         int ret = 0;
1876
1877         if (tb[TCA_FLOWER_INDEV]) {
1878                 int err = tcf_change_indev(net, tb[TCA_FLOWER_INDEV], extack);
1879                 if (err < 0)
1880                         return err;
1881                 key->meta.ingress_ifindex = err;
1882                 mask->meta.ingress_ifindex = 0xffffffff;
1883         }
1884
1885         fl_set_key_val(tb, &key->meta.l2_miss, TCA_FLOWER_L2_MISS,
1886                        &mask->meta.l2_miss, TCA_FLOWER_UNSPEC,
1887                        sizeof(key->meta.l2_miss));
1888
1889         fl_set_key_val(tb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
1890                        mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
1891                        sizeof(key->eth.dst));
1892         fl_set_key_val(tb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
1893                        mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
1894                        sizeof(key->eth.src));
1895         fl_set_key_val(tb, &key->num_of_vlans,
1896                        TCA_FLOWER_KEY_NUM_OF_VLANS,
1897                        &mask->num_of_vlans,
1898                        TCA_FLOWER_UNSPEC,
1899                        sizeof(key->num_of_vlans));
1900
1901         if (is_vlan_key(tb[TCA_FLOWER_KEY_ETH_TYPE], &ethertype, key, mask, 0)) {
1902                 fl_set_key_vlan(tb, ethertype, TCA_FLOWER_KEY_VLAN_ID,
1903                                 TCA_FLOWER_KEY_VLAN_PRIO,
1904                                 TCA_FLOWER_KEY_VLAN_ETH_TYPE,
1905                                 &key->vlan, &mask->vlan);
1906
1907                 if (is_vlan_key(tb[TCA_FLOWER_KEY_VLAN_ETH_TYPE],
1908                                 &ethertype, key, mask, 1)) {
1909                         fl_set_key_vlan(tb, ethertype,
1910                                         TCA_FLOWER_KEY_CVLAN_ID,
1911                                         TCA_FLOWER_KEY_CVLAN_PRIO,
1912                                         TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
1913                                         &key->cvlan, &mask->cvlan);
1914                         fl_set_key_val(tb, &key->basic.n_proto,
1915                                        TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
1916                                        &mask->basic.n_proto,
1917                                        TCA_FLOWER_UNSPEC,
1918                                        sizeof(key->basic.n_proto));
1919                 }
1920         }
1921
1922         if (key->basic.n_proto == htons(ETH_P_PPP_SES))
1923                 fl_set_key_pppoe(tb, &key->pppoe, &mask->pppoe, key, mask);
1924
1925         if (key->basic.n_proto == htons(ETH_P_IP) ||
1926             key->basic.n_proto == htons(ETH_P_IPV6)) {
1927                 fl_set_key_val(tb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
1928                                &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
1929                                sizeof(key->basic.ip_proto));
1930                 fl_set_key_ip(tb, false, &key->ip, &mask->ip);
1931         }
1932
1933         if (tb[TCA_FLOWER_KEY_IPV4_SRC] || tb[TCA_FLOWER_KEY_IPV4_DST]) {
1934                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
1935                 mask->control.addr_type = ~0;
1936                 fl_set_key_val(tb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
1937                                &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
1938                                sizeof(key->ipv4.src));
1939                 fl_set_key_val(tb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
1940                                &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
1941                                sizeof(key->ipv4.dst));
1942         } else if (tb[TCA_FLOWER_KEY_IPV6_SRC] || tb[TCA_FLOWER_KEY_IPV6_DST]) {
1943                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
1944                 mask->control.addr_type = ~0;
1945                 fl_set_key_val(tb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
1946                                &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
1947                                sizeof(key->ipv6.src));
1948                 fl_set_key_val(tb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
1949                                &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
1950                                sizeof(key->ipv6.dst));
1951         }
1952
1953         if (key->basic.ip_proto == IPPROTO_TCP) {
1954                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
1955                                &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
1956                                sizeof(key->tp.src));
1957                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
1958                                &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
1959                                sizeof(key->tp.dst));
1960                 fl_set_key_val(tb, &key->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS,
1961                                &mask->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS_MASK,
1962                                sizeof(key->tcp.flags));
1963         } else if (key->basic.ip_proto == IPPROTO_UDP) {
1964                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
1965                                &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
1966                                sizeof(key->tp.src));
1967                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
1968                                &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
1969                                sizeof(key->tp.dst));
1970         } else if (key->basic.ip_proto == IPPROTO_SCTP) {
1971                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_SCTP_SRC,
1972                                &mask->tp.src, TCA_FLOWER_KEY_SCTP_SRC_MASK,
1973                                sizeof(key->tp.src));
1974                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_SCTP_DST,
1975                                &mask->tp.dst, TCA_FLOWER_KEY_SCTP_DST_MASK,
1976                                sizeof(key->tp.dst));
1977         } else if (key->basic.n_proto == htons(ETH_P_IP) &&
1978                    key->basic.ip_proto == IPPROTO_ICMP) {
1979                 fl_set_key_val(tb, &key->icmp.type, TCA_FLOWER_KEY_ICMPV4_TYPE,
1980                                &mask->icmp.type,
1981                                TCA_FLOWER_KEY_ICMPV4_TYPE_MASK,
1982                                sizeof(key->icmp.type));
1983                 fl_set_key_val(tb, &key->icmp.code, TCA_FLOWER_KEY_ICMPV4_CODE,
1984                                &mask->icmp.code,
1985                                TCA_FLOWER_KEY_ICMPV4_CODE_MASK,
1986                                sizeof(key->icmp.code));
1987         } else if (key->basic.n_proto == htons(ETH_P_IPV6) &&
1988                    key->basic.ip_proto == IPPROTO_ICMPV6) {
1989                 fl_set_key_val(tb, &key->icmp.type, TCA_FLOWER_KEY_ICMPV6_TYPE,
1990                                &mask->icmp.type,
1991                                TCA_FLOWER_KEY_ICMPV6_TYPE_MASK,
1992                                sizeof(key->icmp.type));
1993                 fl_set_key_val(tb, &key->icmp.code, TCA_FLOWER_KEY_ICMPV6_CODE,
1994                                &mask->icmp.code,
1995                                TCA_FLOWER_KEY_ICMPV6_CODE_MASK,
1996                                sizeof(key->icmp.code));
1997         } else if (key->basic.n_proto == htons(ETH_P_MPLS_UC) ||
1998                    key->basic.n_proto == htons(ETH_P_MPLS_MC)) {
1999                 ret = fl_set_key_mpls(tb, &key->mpls, &mask->mpls, extack);
2000                 if (ret)
2001                         return ret;
2002         } else if (key->basic.n_proto == htons(ETH_P_ARP) ||
2003                    key->basic.n_proto == htons(ETH_P_RARP)) {
2004                 fl_set_key_val(tb, &key->arp.sip, TCA_FLOWER_KEY_ARP_SIP,
2005                                &mask->arp.sip, TCA_FLOWER_KEY_ARP_SIP_MASK,
2006                                sizeof(key->arp.sip));
2007                 fl_set_key_val(tb, &key->arp.tip, TCA_FLOWER_KEY_ARP_TIP,
2008                                &mask->arp.tip, TCA_FLOWER_KEY_ARP_TIP_MASK,
2009                                sizeof(key->arp.tip));
2010                 fl_set_key_val(tb, &key->arp.op, TCA_FLOWER_KEY_ARP_OP,
2011                                &mask->arp.op, TCA_FLOWER_KEY_ARP_OP_MASK,
2012                                sizeof(key->arp.op));
2013                 fl_set_key_val(tb, key->arp.sha, TCA_FLOWER_KEY_ARP_SHA,
2014                                mask->arp.sha, TCA_FLOWER_KEY_ARP_SHA_MASK,
2015                                sizeof(key->arp.sha));
2016                 fl_set_key_val(tb, key->arp.tha, TCA_FLOWER_KEY_ARP_THA,
2017                                mask->arp.tha, TCA_FLOWER_KEY_ARP_THA_MASK,
2018                                sizeof(key->arp.tha));
2019         } else if (key->basic.ip_proto == IPPROTO_L2TP) {
2020                 fl_set_key_val(tb, &key->l2tpv3.session_id,
2021                                TCA_FLOWER_KEY_L2TPV3_SID,
2022                                &mask->l2tpv3.session_id, TCA_FLOWER_UNSPEC,
2023                                sizeof(key->l2tpv3.session_id));
2024         } else if (key->basic.n_proto  == htons(ETH_P_CFM)) {
2025                 ret = fl_set_key_cfm(tb, key, mask, extack);
2026                 if (ret)
2027                         return ret;
2028         }
2029
2030         if (key->basic.ip_proto == IPPROTO_TCP ||
2031             key->basic.ip_proto == IPPROTO_UDP ||
2032             key->basic.ip_proto == IPPROTO_SCTP) {
2033                 ret = fl_set_key_port_range(tb, key, mask, extack);
2034                 if (ret)
2035                         return ret;
2036         }
2037
2038         if (tb[TCA_FLOWER_KEY_SPI]) {
2039                 ret = fl_set_key_spi(tb, key, mask, extack);
2040                 if (ret)
2041                         return ret;
2042         }
2043
2044         if (tb[TCA_FLOWER_KEY_ENC_IPV4_SRC] ||
2045             tb[TCA_FLOWER_KEY_ENC_IPV4_DST]) {
2046                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
2047                 mask->enc_control.addr_type = ~0;
2048                 fl_set_key_val(tb, &key->enc_ipv4.src,
2049                                TCA_FLOWER_KEY_ENC_IPV4_SRC,
2050                                &mask->enc_ipv4.src,
2051                                TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
2052                                sizeof(key->enc_ipv4.src));
2053                 fl_set_key_val(tb, &key->enc_ipv4.dst,
2054                                TCA_FLOWER_KEY_ENC_IPV4_DST,
2055                                &mask->enc_ipv4.dst,
2056                                TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
2057                                sizeof(key->enc_ipv4.dst));
2058         }
2059
2060         if (tb[TCA_FLOWER_KEY_ENC_IPV6_SRC] ||
2061             tb[TCA_FLOWER_KEY_ENC_IPV6_DST]) {
2062                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
2063                 mask->enc_control.addr_type = ~0;
2064                 fl_set_key_val(tb, &key->enc_ipv6.src,
2065                                TCA_FLOWER_KEY_ENC_IPV6_SRC,
2066                                &mask->enc_ipv6.src,
2067                                TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
2068                                sizeof(key->enc_ipv6.src));
2069                 fl_set_key_val(tb, &key->enc_ipv6.dst,
2070                                TCA_FLOWER_KEY_ENC_IPV6_DST,
2071                                &mask->enc_ipv6.dst,
2072                                TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
2073                                sizeof(key->enc_ipv6.dst));
2074         }
2075
2076         fl_set_key_val(tb, &key->enc_key_id.keyid, TCA_FLOWER_KEY_ENC_KEY_ID,
2077                        &mask->enc_key_id.keyid, TCA_FLOWER_UNSPEC,
2078                        sizeof(key->enc_key_id.keyid));
2079
2080         fl_set_key_val(tb, &key->enc_tp.src, TCA_FLOWER_KEY_ENC_UDP_SRC_PORT,
2081                        &mask->enc_tp.src, TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK,
2082                        sizeof(key->enc_tp.src));
2083
2084         fl_set_key_val(tb, &key->enc_tp.dst, TCA_FLOWER_KEY_ENC_UDP_DST_PORT,
2085                        &mask->enc_tp.dst, TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK,
2086                        sizeof(key->enc_tp.dst));
2087
2088         fl_set_key_ip(tb, true, &key->enc_ip, &mask->enc_ip);
2089
2090         fl_set_key_val(tb, &key->hash.hash, TCA_FLOWER_KEY_HASH,
2091                        &mask->hash.hash, TCA_FLOWER_KEY_HASH_MASK,
2092                        sizeof(key->hash.hash));
2093
2094         if (tb[TCA_FLOWER_KEY_ENC_OPTS]) {
2095                 ret = fl_set_enc_opt(tb, key, mask, extack);
2096                 if (ret)
2097                         return ret;
2098         }
2099
2100         ret = fl_set_key_ct(tb, &key->ct, &mask->ct, extack);
2101         if (ret)
2102                 return ret;
2103
2104         if (tb[TCA_FLOWER_KEY_FLAGS]) {
2105                 ret = fl_set_key_flags(tca_opts, tb, false,
2106                                        &key->control.flags,
2107                                        &mask->control.flags, extack);
2108                 if (ret)
2109                         return ret;
2110         }
2111
2112         if (tb[TCA_FLOWER_KEY_ENC_FLAGS])
2113                 ret = fl_set_key_flags(tca_opts, tb, true,
2114                                        &key->enc_control.flags,
2115                                        &mask->enc_control.flags, extack);
2116
2117         return ret;
2118 }
2119
2120 static void fl_mask_copy(struct fl_flow_mask *dst,
2121                          struct fl_flow_mask *src)
2122 {
2123         const void *psrc = fl_key_get_start(&src->key, src);
2124         void *pdst = fl_key_get_start(&dst->key, src);
2125
2126         memcpy(pdst, psrc, fl_mask_range(src));
2127         dst->range = src->range;
2128 }
2129
2130 static const struct rhashtable_params fl_ht_params = {
2131         .key_offset = offsetof(struct cls_fl_filter, mkey), /* base offset */
2132         .head_offset = offsetof(struct cls_fl_filter, ht_node),
2133         .automatic_shrinking = true,
2134 };
2135
2136 static int fl_init_mask_hashtable(struct fl_flow_mask *mask)
2137 {
2138         mask->filter_ht_params = fl_ht_params;
2139         mask->filter_ht_params.key_len = fl_mask_range(mask);
2140         mask->filter_ht_params.key_offset += mask->range.start;
2141
2142         return rhashtable_init(&mask->ht, &mask->filter_ht_params);
2143 }
2144
2145 #define FL_KEY_MEMBER_OFFSET(member) offsetof(struct fl_flow_key, member)
2146 #define FL_KEY_MEMBER_SIZE(member) sizeof_field(struct fl_flow_key, member)
2147
2148 #define FL_KEY_IS_MASKED(mask, member)                                          \
2149         memchr_inv(((char *)mask) + FL_KEY_MEMBER_OFFSET(member),               \
2150                    0, FL_KEY_MEMBER_SIZE(member))                               \
2151
2152 #define FL_KEY_SET(keys, cnt, id, member)                                       \
2153         do {                                                                    \
2154                 keys[cnt].key_id = id;                                          \
2155                 keys[cnt].offset = FL_KEY_MEMBER_OFFSET(member);                \
2156                 cnt++;                                                          \
2157         } while(0);
2158
2159 #define FL_KEY_SET_IF_MASKED(mask, keys, cnt, id, member)                       \
2160         do {                                                                    \
2161                 if (FL_KEY_IS_MASKED(mask, member))                             \
2162                         FL_KEY_SET(keys, cnt, id, member);                      \
2163         } while(0);
2164
2165 static void fl_init_dissector(struct flow_dissector *dissector,
2166                               struct fl_flow_key *mask)
2167 {
2168         struct flow_dissector_key keys[FLOW_DISSECTOR_KEY_MAX];
2169         size_t cnt = 0;
2170
2171         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
2172                              FLOW_DISSECTOR_KEY_META, meta);
2173         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_CONTROL, control);
2174         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_BASIC, basic);
2175         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
2176                              FLOW_DISSECTOR_KEY_ETH_ADDRS, eth);
2177         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
2178                              FLOW_DISSECTOR_KEY_IPV4_ADDRS, ipv4);
2179         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
2180                              FLOW_DISSECTOR_KEY_IPV6_ADDRS, ipv6);
2181         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
2182                              FLOW_DISSECTOR_KEY_PORTS, tp);
2183         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
2184                              FLOW_DISSECTOR_KEY_PORTS_RANGE, tp_range);
2185         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
2186                              FLOW_DISSECTOR_KEY_IP, ip);
2187         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
2188                              FLOW_DISSECTOR_KEY_TCP, tcp);
2189         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
2190                              FLOW_DISSECTOR_KEY_ICMP, icmp);
2191         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
2192                              FLOW_DISSECTOR_KEY_ARP, arp);
2193         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
2194                              FLOW_DISSECTOR_KEY_MPLS, mpls);
2195         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
2196                              FLOW_DISSECTOR_KEY_VLAN, vlan);
2197         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
2198                              FLOW_DISSECTOR_KEY_CVLAN, cvlan);
2199         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
2200                              FLOW_DISSECTOR_KEY_ENC_KEYID, enc_key_id);
2201         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
2202                              FLOW_DISSECTOR_KEY_ENC_IPV4_ADDRS, enc_ipv4);
2203         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
2204                              FLOW_DISSECTOR_KEY_ENC_IPV6_ADDRS, enc_ipv6);
2205         if (FL_KEY_IS_MASKED(mask, enc_ipv4) ||
2206             FL_KEY_IS_MASKED(mask, enc_ipv6) ||
2207             FL_KEY_IS_MASKED(mask, enc_control))
2208                 FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_ENC_CONTROL,
2209                            enc_control);
2210         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
2211                              FLOW_DISSECTOR_KEY_ENC_PORTS, enc_tp);
2212         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
2213                              FLOW_DISSECTOR_KEY_ENC_IP, enc_ip);
2214         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
2215                              FLOW_DISSECTOR_KEY_ENC_OPTS, enc_opts);
2216         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
2217                              FLOW_DISSECTOR_KEY_CT, ct);
2218         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
2219                              FLOW_DISSECTOR_KEY_HASH, hash);
2220         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
2221                              FLOW_DISSECTOR_KEY_NUM_OF_VLANS, num_of_vlans);
2222         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
2223                              FLOW_DISSECTOR_KEY_PPPOE, pppoe);
2224         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
2225                              FLOW_DISSECTOR_KEY_L2TPV3, l2tpv3);
2226         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
2227                              FLOW_DISSECTOR_KEY_IPSEC, ipsec);
2228         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
2229                              FLOW_DISSECTOR_KEY_CFM, cfm);
2230
2231         skb_flow_dissector_init(dissector, keys, cnt);
2232 }
2233
2234 static struct fl_flow_mask *fl_create_new_mask(struct cls_fl_head *head,
2235                                                struct fl_flow_mask *mask)
2236 {
2237         struct fl_flow_mask *newmask;
2238         int err;
2239
2240         newmask = kzalloc(sizeof(*newmask), GFP_KERNEL);
2241         if (!newmask)
2242                 return ERR_PTR(-ENOMEM);
2243
2244         fl_mask_copy(newmask, mask);
2245
2246         if ((newmask->key.tp_range.tp_min.dst &&
2247              newmask->key.tp_range.tp_max.dst) ||
2248             (newmask->key.tp_range.tp_min.src &&
2249              newmask->key.tp_range.tp_max.src))
2250                 newmask->flags |= TCA_FLOWER_MASK_FLAGS_RANGE;
2251
2252         err = fl_init_mask_hashtable(newmask);
2253         if (err)
2254                 goto errout_free;
2255
2256         fl_init_dissector(&newmask->dissector, &newmask->key);
2257
2258         INIT_LIST_HEAD_RCU(&newmask->filters);
2259
2260         refcount_set(&newmask->refcnt, 1);
2261         err = rhashtable_replace_fast(&head->ht, &mask->ht_node,
2262                                       &newmask->ht_node, mask_ht_params);
2263         if (err)
2264                 goto errout_destroy;
2265
2266         spin_lock(&head->masks_lock);
2267         list_add_tail_rcu(&newmask->list, &head->masks);
2268         spin_unlock(&head->masks_lock);
2269
2270         return newmask;
2271
2272 errout_destroy:
2273         rhashtable_destroy(&newmask->ht);
2274 errout_free:
2275         kfree(newmask);
2276
2277         return ERR_PTR(err);
2278 }
2279
2280 static int fl_check_assign_mask(struct cls_fl_head *head,
2281                                 struct cls_fl_filter *fnew,
2282                                 struct cls_fl_filter *fold,
2283                                 struct fl_flow_mask *mask)
2284 {
2285         struct fl_flow_mask *newmask;
2286         int ret = 0;
2287
2288         rcu_read_lock();
2289
2290         /* Insert mask as temporary node to prevent concurrent creation of mask
2291          * with same key. Any concurrent lookups with same key will return
2292          * -EAGAIN because mask's refcnt is zero.
2293          */
2294         fnew->mask = rhashtable_lookup_get_insert_fast(&head->ht,
2295                                                        &mask->ht_node,
2296                                                        mask_ht_params);
2297         if (!fnew->mask) {
2298                 rcu_read_unlock();
2299
2300                 if (fold) {
2301                         ret = -EINVAL;
2302                         goto errout_cleanup;
2303                 }
2304
2305                 newmask = fl_create_new_mask(head, mask);
2306                 if (IS_ERR(newmask)) {
2307                         ret = PTR_ERR(newmask);
2308                         goto errout_cleanup;
2309                 }
2310
2311                 fnew->mask = newmask;
2312                 return 0;
2313         } else if (IS_ERR(fnew->mask)) {
2314                 ret = PTR_ERR(fnew->mask);
2315         } else if (fold && fold->mask != fnew->mask) {
2316                 ret = -EINVAL;
2317         } else if (!refcount_inc_not_zero(&fnew->mask->refcnt)) {
2318                 /* Mask was deleted concurrently, try again */
2319                 ret = -EAGAIN;
2320         }
2321         rcu_read_unlock();
2322         return ret;
2323
2324 errout_cleanup:
2325         rhashtable_remove_fast(&head->ht, &mask->ht_node,
2326                                mask_ht_params);
2327         return ret;
2328 }
2329
2330 static bool fl_needs_tc_skb_ext(const struct fl_flow_key *mask)
2331 {
2332         return mask->meta.l2_miss;
2333 }
2334
2335 static int fl_ht_insert_unique(struct cls_fl_filter *fnew,
2336                                struct cls_fl_filter *fold,
2337                                bool *in_ht)
2338 {
2339         struct fl_flow_mask *mask = fnew->mask;
2340         int err;
2341
2342         err = rhashtable_lookup_insert_fast(&mask->ht,
2343                                             &fnew->ht_node,
2344                                             mask->filter_ht_params);
2345         if (err) {
2346                 *in_ht = false;
2347                 /* It is okay if filter with same key exists when
2348                  * overwriting.
2349                  */
2350                 return fold && err == -EEXIST ? 0 : err;
2351         }
2352
2353         *in_ht = true;
2354         return 0;
2355 }
2356
2357 static int fl_change(struct net *net, struct sk_buff *in_skb,
2358                      struct tcf_proto *tp, unsigned long base,
2359                      u32 handle, struct nlattr **tca,
2360                      void **arg, u32 flags,
2361                      struct netlink_ext_ack *extack)
2362 {
2363         struct cls_fl_head *head = fl_head_dereference(tp);
2364         bool rtnl_held = !(flags & TCA_ACT_FLAGS_NO_RTNL);
2365         struct nlattr *tca_opts = tca[TCA_OPTIONS];
2366         struct cls_fl_filter *fold = *arg;
2367         bool bound_to_filter = false;
2368         struct cls_fl_filter *fnew;
2369         struct fl_flow_mask *mask;
2370         struct nlattr **tb;
2371         bool in_ht;
2372         int err;
2373
2374         if (!tca_opts) {
2375                 err = -EINVAL;
2376                 goto errout_fold;
2377         }
2378
2379         mask = kzalloc(sizeof(struct fl_flow_mask), GFP_KERNEL);
2380         if (!mask) {
2381                 err = -ENOBUFS;
2382                 goto errout_fold;
2383         }
2384
2385         tb = kcalloc(TCA_FLOWER_MAX + 1, sizeof(struct nlattr *), GFP_KERNEL);
2386         if (!tb) {
2387                 err = -ENOBUFS;
2388                 goto errout_mask_alloc;
2389         }
2390
2391         err = nla_parse_nested_deprecated(tb, TCA_FLOWER_MAX,
2392                                           tca_opts, fl_policy, NULL);
2393         if (err < 0)
2394                 goto errout_tb;
2395
2396         if (fold && handle && fold->handle != handle) {
2397                 err = -EINVAL;
2398                 goto errout_tb;
2399         }
2400
2401         fnew = kzalloc(sizeof(*fnew), GFP_KERNEL);
2402         if (!fnew) {
2403                 err = -ENOBUFS;
2404                 goto errout_tb;
2405         }
2406         INIT_LIST_HEAD(&fnew->hw_list);
2407         refcount_set(&fnew->refcnt, 1);
2408
2409         if (tb[TCA_FLOWER_FLAGS]) {
2410                 fnew->flags = nla_get_u32(tb[TCA_FLOWER_FLAGS]);
2411
2412                 if (!tc_flags_valid(fnew->flags)) {
2413                         kfree(fnew);
2414                         err = -EINVAL;
2415                         goto errout_tb;
2416                 }
2417         }
2418
2419         if (!fold) {
2420                 spin_lock(&tp->lock);
2421                 if (!handle) {
2422                         handle = 1;
2423                         err = idr_alloc_u32(&head->handle_idr, NULL, &handle,
2424                                             INT_MAX, GFP_ATOMIC);
2425                 } else {
2426                         err = idr_alloc_u32(&head->handle_idr, NULL, &handle,
2427                                             handle, GFP_ATOMIC);
2428
2429                         /* Filter with specified handle was concurrently
2430                          * inserted after initial check in cls_api. This is not
2431                          * necessarily an error if NLM_F_EXCL is not set in
2432                          * message flags. Returning EAGAIN will cause cls_api to
2433                          * try to update concurrently inserted rule.
2434                          */
2435                         if (err == -ENOSPC)
2436                                 err = -EAGAIN;
2437                 }
2438                 spin_unlock(&tp->lock);
2439
2440                 if (err) {
2441                         kfree(fnew);
2442                         goto errout_tb;
2443                 }
2444         }
2445         fnew->handle = handle;
2446
2447         err = tcf_exts_init_ex(&fnew->exts, net, TCA_FLOWER_ACT, 0, tp, handle,
2448                                !tc_skip_hw(fnew->flags));
2449         if (err < 0)
2450                 goto errout_idr;
2451
2452         err = tcf_exts_validate_ex(net, tp, tb, tca[TCA_RATE],
2453                                    &fnew->exts, flags, fnew->flags,
2454                                    extack);
2455         if (err < 0)
2456                 goto errout_idr;
2457
2458         if (tb[TCA_FLOWER_CLASSID]) {
2459                 fnew->res.classid = nla_get_u32(tb[TCA_FLOWER_CLASSID]);
2460                 if (flags & TCA_ACT_FLAGS_NO_RTNL)
2461                         rtnl_lock();
2462                 tcf_bind_filter(tp, &fnew->res, base);
2463                 if (flags & TCA_ACT_FLAGS_NO_RTNL)
2464                         rtnl_unlock();
2465                 bound_to_filter = true;
2466         }
2467
2468         err = fl_set_key(net, tca_opts, tb, &fnew->key, &mask->key, extack);
2469         if (err)
2470                 goto unbind_filter;
2471
2472         fl_mask_update_range(mask);
2473         fl_set_masked_key(&fnew->mkey, &fnew->key, mask);
2474
2475         if (!fl_mask_fits_tmplt(tp->chain->tmplt_priv, mask)) {
2476                 NL_SET_ERR_MSG_MOD(extack, "Mask does not fit the template");
2477                 err = -EINVAL;
2478                 goto unbind_filter;
2479         }
2480
2481         /* Enable tc skb extension if filter matches on data extracted from
2482          * this extension.
2483          */
2484         if (fl_needs_tc_skb_ext(&mask->key)) {
2485                 fnew->needs_tc_skb_ext = 1;
2486                 tc_skb_ext_tc_enable();
2487         }
2488
2489         err = fl_check_assign_mask(head, fnew, fold, mask);
2490         if (err)
2491                 goto unbind_filter;
2492
2493         err = fl_ht_insert_unique(fnew, fold, &in_ht);
2494         if (err)
2495                 goto errout_mask;
2496
2497         if (!tc_skip_hw(fnew->flags)) {
2498                 err = fl_hw_replace_filter(tp, fnew, rtnl_held, extack);
2499                 if (err)
2500                         goto errout_ht;
2501         }
2502
2503         if (!tc_in_hw(fnew->flags))
2504                 fnew->flags |= TCA_CLS_FLAGS_NOT_IN_HW;
2505
2506         tcf_proto_update_usesw(tp, fnew->flags);
2507
2508         spin_lock(&tp->lock);
2509
2510         /* tp was deleted concurrently. -EAGAIN will cause caller to lookup
2511          * proto again or create new one, if necessary.
2512          */
2513         if (tp->deleting) {
2514                 err = -EAGAIN;
2515                 goto errout_hw;
2516         }
2517
2518         if (fold) {
2519                 /* Fold filter was deleted concurrently. Retry lookup. */
2520                 if (fold->deleted) {
2521                         err = -EAGAIN;
2522                         goto errout_hw;
2523                 }
2524
2525                 fnew->handle = handle;
2526
2527                 if (!in_ht) {
2528                         struct rhashtable_params params =
2529                                 fnew->mask->filter_ht_params;
2530
2531                         err = rhashtable_insert_fast(&fnew->mask->ht,
2532                                                      &fnew->ht_node,
2533                                                      params);
2534                         if (err)
2535                                 goto errout_hw;
2536                         in_ht = true;
2537                 }
2538
2539                 refcount_inc(&fnew->refcnt);
2540                 rhashtable_remove_fast(&fold->mask->ht,
2541                                        &fold->ht_node,
2542                                        fold->mask->filter_ht_params);
2543                 idr_replace(&head->handle_idr, fnew, fnew->handle);
2544                 list_replace_rcu(&fold->list, &fnew->list);
2545                 fold->deleted = true;
2546
2547                 spin_unlock(&tp->lock);
2548
2549                 fl_mask_put(head, fold->mask);
2550                 if (!tc_skip_hw(fold->flags))
2551                         fl_hw_destroy_filter(tp, fold, rtnl_held, NULL);
2552                 tcf_unbind_filter(tp, &fold->res);
2553                 /* Caller holds reference to fold, so refcnt is always > 0
2554                  * after this.
2555                  */
2556                 refcount_dec(&fold->refcnt);
2557                 __fl_put(fold);
2558         } else {
2559                 idr_replace(&head->handle_idr, fnew, fnew->handle);
2560
2561                 refcount_inc(&fnew->refcnt);
2562                 list_add_tail_rcu(&fnew->list, &fnew->mask->filters);
2563                 spin_unlock(&tp->lock);
2564         }
2565
2566         *arg = fnew;
2567
2568         kfree(tb);
2569         tcf_queue_work(&mask->rwork, fl_uninit_mask_free_work);
2570         return 0;
2571
2572 errout_ht:
2573         spin_lock(&tp->lock);
2574 errout_hw:
2575         fnew->deleted = true;
2576         spin_unlock(&tp->lock);
2577         if (!tc_skip_hw(fnew->flags))
2578                 fl_hw_destroy_filter(tp, fnew, rtnl_held, NULL);
2579         if (in_ht)
2580                 rhashtable_remove_fast(&fnew->mask->ht, &fnew->ht_node,
2581                                        fnew->mask->filter_ht_params);
2582 errout_mask:
2583         fl_mask_put(head, fnew->mask);
2584
2585 unbind_filter:
2586         if (bound_to_filter) {
2587                 if (flags & TCA_ACT_FLAGS_NO_RTNL)
2588                         rtnl_lock();
2589                 tcf_unbind_filter(tp, &fnew->res);
2590                 if (flags & TCA_ACT_FLAGS_NO_RTNL)
2591                         rtnl_unlock();
2592         }
2593
2594 errout_idr:
2595         if (!fold) {
2596                 spin_lock(&tp->lock);
2597                 idr_remove(&head->handle_idr, fnew->handle);
2598                 spin_unlock(&tp->lock);
2599         }
2600         __fl_put(fnew);
2601 errout_tb:
2602         kfree(tb);
2603 errout_mask_alloc:
2604         tcf_queue_work(&mask->rwork, fl_uninit_mask_free_work);
2605 errout_fold:
2606         if (fold)
2607                 __fl_put(fold);
2608         return err;
2609 }
2610
2611 static int fl_delete(struct tcf_proto *tp, void *arg, bool *last,
2612                      bool rtnl_held, struct netlink_ext_ack *extack)
2613 {
2614         struct cls_fl_head *head = fl_head_dereference(tp);
2615         struct cls_fl_filter *f = arg;
2616         bool last_on_mask;
2617         int err = 0;
2618
2619         err = __fl_delete(tp, f, &last_on_mask, rtnl_held, extack);
2620         *last = list_empty(&head->masks);
2621         __fl_put(f);
2622
2623         return err;
2624 }
2625
2626 static void fl_walk(struct tcf_proto *tp, struct tcf_walker *arg,
2627                     bool rtnl_held)
2628 {
2629         struct cls_fl_head *head = fl_head_dereference(tp);
2630         unsigned long id = arg->cookie, tmp;
2631         struct cls_fl_filter *f;
2632
2633         arg->count = arg->skip;
2634
2635         rcu_read_lock();
2636         idr_for_each_entry_continue_ul(&head->handle_idr, f, tmp, id) {
2637                 /* don't return filters that are being deleted */
2638                 if (!f || !refcount_inc_not_zero(&f->refcnt))
2639                         continue;
2640                 rcu_read_unlock();
2641
2642                 if (arg->fn(tp, f, arg) < 0) {
2643                         __fl_put(f);
2644                         arg->stop = 1;
2645                         rcu_read_lock();
2646                         break;
2647                 }
2648                 __fl_put(f);
2649                 arg->count++;
2650                 rcu_read_lock();
2651         }
2652         rcu_read_unlock();
2653         arg->cookie = id;
2654 }
2655
2656 static struct cls_fl_filter *
2657 fl_get_next_hw_filter(struct tcf_proto *tp, struct cls_fl_filter *f, bool add)
2658 {
2659         struct cls_fl_head *head = fl_head_dereference(tp);
2660
2661         spin_lock(&tp->lock);
2662         if (list_empty(&head->hw_filters)) {
2663                 spin_unlock(&tp->lock);
2664                 return NULL;
2665         }
2666
2667         if (!f)
2668                 f = list_entry(&head->hw_filters, struct cls_fl_filter,
2669                                hw_list);
2670         list_for_each_entry_continue(f, &head->hw_filters, hw_list) {
2671                 if (!(add && f->deleted) && refcount_inc_not_zero(&f->refcnt)) {
2672                         spin_unlock(&tp->lock);
2673                         return f;
2674                 }
2675         }
2676
2677         spin_unlock(&tp->lock);
2678         return NULL;
2679 }
2680
2681 static int fl_reoffload(struct tcf_proto *tp, bool add, flow_setup_cb_t *cb,
2682                         void *cb_priv, struct netlink_ext_ack *extack)
2683 {
2684         struct tcf_block *block = tp->chain->block;
2685         struct flow_cls_offload cls_flower = {};
2686         struct cls_fl_filter *f = NULL;
2687         int err;
2688
2689         /* hw_filters list can only be changed by hw offload functions after
2690          * obtaining rtnl lock. Make sure it is not changed while reoffload is
2691          * iterating it.
2692          */
2693         ASSERT_RTNL();
2694
2695         while ((f = fl_get_next_hw_filter(tp, f, add))) {
2696                 cls_flower.rule =
2697                         flow_rule_alloc(tcf_exts_num_actions(&f->exts));
2698                 if (!cls_flower.rule) {
2699                         __fl_put(f);
2700                         return -ENOMEM;
2701                 }
2702
2703                 tc_cls_common_offload_init(&cls_flower.common, tp, f->flags,
2704                                            extack);
2705                 cls_flower.command = add ?
2706                         FLOW_CLS_REPLACE : FLOW_CLS_DESTROY;
2707                 cls_flower.cookie = (unsigned long)f;
2708                 cls_flower.rule->match.dissector = &f->mask->dissector;
2709                 cls_flower.rule->match.mask = &f->mask->key;
2710                 cls_flower.rule->match.key = &f->mkey;
2711
2712                 err = tc_setup_offload_action(&cls_flower.rule->action, &f->exts,
2713                                               cls_flower.common.extack);
2714                 if (err) {
2715                         kfree(cls_flower.rule);
2716                         if (tc_skip_sw(f->flags)) {
2717                                 __fl_put(f);
2718                                 return err;
2719                         }
2720                         goto next_flow;
2721                 }
2722
2723                 cls_flower.classid = f->res.classid;
2724
2725                 err = tc_setup_cb_reoffload(block, tp, add, cb,
2726                                             TC_SETUP_CLSFLOWER, &cls_flower,
2727                                             cb_priv, &f->flags,
2728                                             &f->in_hw_count);
2729                 tc_cleanup_offload_action(&cls_flower.rule->action);
2730                 kfree(cls_flower.rule);
2731
2732                 if (err) {
2733                         __fl_put(f);
2734                         return err;
2735                 }
2736 next_flow:
2737                 __fl_put(f);
2738         }
2739
2740         return 0;
2741 }
2742
2743 static void fl_hw_add(struct tcf_proto *tp, void *type_data)
2744 {
2745         struct flow_cls_offload *cls_flower = type_data;
2746         struct cls_fl_filter *f =
2747                 (struct cls_fl_filter *) cls_flower->cookie;
2748         struct cls_fl_head *head = fl_head_dereference(tp);
2749
2750         spin_lock(&tp->lock);
2751         list_add(&f->hw_list, &head->hw_filters);
2752         spin_unlock(&tp->lock);
2753 }
2754
2755 static void fl_hw_del(struct tcf_proto *tp, void *type_data)
2756 {
2757         struct flow_cls_offload *cls_flower = type_data;
2758         struct cls_fl_filter *f =
2759                 (struct cls_fl_filter *) cls_flower->cookie;
2760
2761         spin_lock(&tp->lock);
2762         if (!list_empty(&f->hw_list))
2763                 list_del_init(&f->hw_list);
2764         spin_unlock(&tp->lock);
2765 }
2766
2767 static int fl_hw_create_tmplt(struct tcf_chain *chain,
2768                               struct fl_flow_tmplt *tmplt)
2769 {
2770         struct flow_cls_offload cls_flower = {};
2771         struct tcf_block *block = chain->block;
2772
2773         cls_flower.rule = flow_rule_alloc(0);
2774         if (!cls_flower.rule)
2775                 return -ENOMEM;
2776
2777         cls_flower.common.chain_index = chain->index;
2778         cls_flower.command = FLOW_CLS_TMPLT_CREATE;
2779         cls_flower.cookie = (unsigned long) tmplt;
2780         cls_flower.rule->match.dissector = &tmplt->dissector;
2781         cls_flower.rule->match.mask = &tmplt->mask;
2782         cls_flower.rule->match.key = &tmplt->dummy_key;
2783
2784         /* We don't care if driver (any of them) fails to handle this
2785          * call. It serves just as a hint for it.
2786          */
2787         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false, true);
2788         kfree(cls_flower.rule);
2789
2790         return 0;
2791 }
2792
2793 static void fl_hw_destroy_tmplt(struct tcf_chain *chain,
2794                                 struct fl_flow_tmplt *tmplt)
2795 {
2796         struct flow_cls_offload cls_flower = {};
2797         struct tcf_block *block = chain->block;
2798
2799         cls_flower.common.chain_index = chain->index;
2800         cls_flower.command = FLOW_CLS_TMPLT_DESTROY;
2801         cls_flower.cookie = (unsigned long) tmplt;
2802
2803         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false, true);
2804 }
2805
2806 static void *fl_tmplt_create(struct net *net, struct tcf_chain *chain,
2807                              struct nlattr **tca,
2808                              struct netlink_ext_ack *extack)
2809 {
2810         struct nlattr *tca_opts = tca[TCA_OPTIONS];
2811         struct fl_flow_tmplt *tmplt;
2812         struct nlattr **tb;
2813         int err;
2814
2815         if (!tca_opts)
2816                 return ERR_PTR(-EINVAL);
2817
2818         tb = kcalloc(TCA_FLOWER_MAX + 1, sizeof(struct nlattr *), GFP_KERNEL);
2819         if (!tb)
2820                 return ERR_PTR(-ENOBUFS);
2821         err = nla_parse_nested_deprecated(tb, TCA_FLOWER_MAX,
2822                                           tca_opts, fl_policy, NULL);
2823         if (err)
2824                 goto errout_tb;
2825
2826         tmplt = kzalloc(sizeof(*tmplt), GFP_KERNEL);
2827         if (!tmplt) {
2828                 err = -ENOMEM;
2829                 goto errout_tb;
2830         }
2831         tmplt->chain = chain;
2832         err = fl_set_key(net, tca_opts, tb, &tmplt->dummy_key,
2833                          &tmplt->mask, extack);
2834         if (err)
2835                 goto errout_tmplt;
2836
2837         fl_init_dissector(&tmplt->dissector, &tmplt->mask);
2838
2839         err = fl_hw_create_tmplt(chain, tmplt);
2840         if (err)
2841                 goto errout_tmplt;
2842
2843         kfree(tb);
2844         return tmplt;
2845
2846 errout_tmplt:
2847         kfree(tmplt);
2848 errout_tb:
2849         kfree(tb);
2850         return ERR_PTR(err);
2851 }
2852
2853 static void fl_tmplt_destroy(void *tmplt_priv)
2854 {
2855         struct fl_flow_tmplt *tmplt = tmplt_priv;
2856
2857         fl_hw_destroy_tmplt(tmplt->chain, tmplt);
2858         kfree(tmplt);
2859 }
2860
2861 static void fl_tmplt_reoffload(struct tcf_chain *chain, bool add,
2862                                flow_setup_cb_t *cb, void *cb_priv)
2863 {
2864         struct fl_flow_tmplt *tmplt = chain->tmplt_priv;
2865         struct flow_cls_offload cls_flower = {};
2866
2867         cls_flower.rule = flow_rule_alloc(0);
2868         if (!cls_flower.rule)
2869                 return;
2870
2871         cls_flower.common.chain_index = chain->index;
2872         cls_flower.command = add ? FLOW_CLS_TMPLT_CREATE :
2873                                    FLOW_CLS_TMPLT_DESTROY;
2874         cls_flower.cookie = (unsigned long) tmplt;
2875         cls_flower.rule->match.dissector = &tmplt->dissector;
2876         cls_flower.rule->match.mask = &tmplt->mask;
2877         cls_flower.rule->match.key = &tmplt->dummy_key;
2878
2879         cb(TC_SETUP_CLSFLOWER, &cls_flower, cb_priv);
2880         kfree(cls_flower.rule);
2881 }
2882
2883 static int fl_dump_key_val(struct sk_buff *skb,
2884                            void *val, int val_type,
2885                            void *mask, int mask_type, int len)
2886 {
2887         int err;
2888
2889         if (!memchr_inv(mask, 0, len))
2890                 return 0;
2891         err = nla_put(skb, val_type, len, val);
2892         if (err)
2893                 return err;
2894         if (mask_type != TCA_FLOWER_UNSPEC) {
2895                 err = nla_put(skb, mask_type, len, mask);
2896                 if (err)
2897                         return err;
2898         }
2899         return 0;
2900 }
2901
2902 static int fl_dump_key_port_range(struct sk_buff *skb, struct fl_flow_key *key,
2903                                   struct fl_flow_key *mask)
2904 {
2905         if (fl_dump_key_val(skb, &key->tp_range.tp_min.dst,
2906                             TCA_FLOWER_KEY_PORT_DST_MIN,
2907                             &mask->tp_range.tp_min.dst, TCA_FLOWER_UNSPEC,
2908                             sizeof(key->tp_range.tp_min.dst)) ||
2909             fl_dump_key_val(skb, &key->tp_range.tp_max.dst,
2910                             TCA_FLOWER_KEY_PORT_DST_MAX,
2911                             &mask->tp_range.tp_max.dst, TCA_FLOWER_UNSPEC,
2912                             sizeof(key->tp_range.tp_max.dst)) ||
2913             fl_dump_key_val(skb, &key->tp_range.tp_min.src,
2914                             TCA_FLOWER_KEY_PORT_SRC_MIN,
2915                             &mask->tp_range.tp_min.src, TCA_FLOWER_UNSPEC,
2916                             sizeof(key->tp_range.tp_min.src)) ||
2917             fl_dump_key_val(skb, &key->tp_range.tp_max.src,
2918                             TCA_FLOWER_KEY_PORT_SRC_MAX,
2919                             &mask->tp_range.tp_max.src, TCA_FLOWER_UNSPEC,
2920                             sizeof(key->tp_range.tp_max.src)))
2921                 return -1;
2922
2923         return 0;
2924 }
2925
2926 static int fl_dump_key_mpls_opt_lse(struct sk_buff *skb,
2927                                     struct flow_dissector_key_mpls *mpls_key,
2928                                     struct flow_dissector_key_mpls *mpls_mask,
2929                                     u8 lse_index)
2930 {
2931         struct flow_dissector_mpls_lse *lse_mask = &mpls_mask->ls[lse_index];
2932         struct flow_dissector_mpls_lse *lse_key = &mpls_key->ls[lse_index];
2933         int err;
2934
2935         err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH,
2936                          lse_index + 1);
2937         if (err)
2938                 return err;
2939
2940         if (lse_mask->mpls_ttl) {
2941                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_TTL,
2942                                  lse_key->mpls_ttl);
2943                 if (err)
2944                         return err;
2945         }
2946         if (lse_mask->mpls_bos) {
2947                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS,
2948                                  lse_key->mpls_bos);
2949                 if (err)
2950                         return err;
2951         }
2952         if (lse_mask->mpls_tc) {
2953                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_TC,
2954                                  lse_key->mpls_tc);
2955                 if (err)
2956                         return err;
2957         }
2958         if (lse_mask->mpls_label) {
2959                 err = nla_put_u32(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL,
2960                                   lse_key->mpls_label);
2961                 if (err)
2962                         return err;
2963         }
2964
2965         return 0;
2966 }
2967
2968 static int fl_dump_key_mpls_opts(struct sk_buff *skb,
2969                                  struct flow_dissector_key_mpls *mpls_key,
2970                                  struct flow_dissector_key_mpls *mpls_mask)
2971 {
2972         struct nlattr *opts;
2973         struct nlattr *lse;
2974         u8 lse_index;
2975         int err;
2976
2977         opts = nla_nest_start(skb, TCA_FLOWER_KEY_MPLS_OPTS);
2978         if (!opts)
2979                 return -EMSGSIZE;
2980
2981         for (lse_index = 0; lse_index < FLOW_DIS_MPLS_MAX; lse_index++) {
2982                 if (!(mpls_mask->used_lses & 1 << lse_index))
2983                         continue;
2984
2985                 lse = nla_nest_start(skb, TCA_FLOWER_KEY_MPLS_OPTS_LSE);
2986                 if (!lse) {
2987                         err = -EMSGSIZE;
2988                         goto err_opts;
2989                 }
2990
2991                 err = fl_dump_key_mpls_opt_lse(skb, mpls_key, mpls_mask,
2992                                                lse_index);
2993                 if (err)
2994                         goto err_opts_lse;
2995                 nla_nest_end(skb, lse);
2996         }
2997         nla_nest_end(skb, opts);
2998
2999         return 0;
3000
3001 err_opts_lse:
3002         nla_nest_cancel(skb, lse);
3003 err_opts:
3004         nla_nest_cancel(skb, opts);
3005
3006         return err;
3007 }
3008
3009 static int fl_dump_key_mpls(struct sk_buff *skb,
3010                             struct flow_dissector_key_mpls *mpls_key,
3011                             struct flow_dissector_key_mpls *mpls_mask)
3012 {
3013         struct flow_dissector_mpls_lse *lse_mask;
3014         struct flow_dissector_mpls_lse *lse_key;
3015         int err;
3016
3017         if (!mpls_mask->used_lses)
3018                 return 0;
3019
3020         lse_mask = &mpls_mask->ls[0];
3021         lse_key = &mpls_key->ls[0];
3022
3023         /* For backward compatibility, don't use the MPLS nested attributes if
3024          * the rule can be expressed using the old attributes.
3025          */
3026         if (mpls_mask->used_lses & ~1 ||
3027             (!lse_mask->mpls_ttl && !lse_mask->mpls_bos &&
3028              !lse_mask->mpls_tc && !lse_mask->mpls_label))
3029                 return fl_dump_key_mpls_opts(skb, mpls_key, mpls_mask);
3030
3031         if (lse_mask->mpls_ttl) {
3032                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_TTL,
3033                                  lse_key->mpls_ttl);
3034                 if (err)
3035                         return err;
3036         }
3037         if (lse_mask->mpls_tc) {
3038                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_TC,
3039                                  lse_key->mpls_tc);
3040                 if (err)
3041                         return err;
3042         }
3043         if (lse_mask->mpls_label) {
3044                 err = nla_put_u32(skb, TCA_FLOWER_KEY_MPLS_LABEL,
3045                                   lse_key->mpls_label);
3046                 if (err)
3047                         return err;
3048         }
3049         if (lse_mask->mpls_bos) {
3050                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_BOS,
3051                                  lse_key->mpls_bos);
3052                 if (err)
3053                         return err;
3054         }
3055         return 0;
3056 }
3057
3058 static int fl_dump_key_ip(struct sk_buff *skb, bool encap,
3059                           struct flow_dissector_key_ip *key,
3060                           struct flow_dissector_key_ip *mask)
3061 {
3062         int tos_key = encap ? TCA_FLOWER_KEY_ENC_IP_TOS : TCA_FLOWER_KEY_IP_TOS;
3063         int ttl_key = encap ? TCA_FLOWER_KEY_ENC_IP_TTL : TCA_FLOWER_KEY_IP_TTL;
3064         int tos_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TOS_MASK : TCA_FLOWER_KEY_IP_TOS_MASK;
3065         int ttl_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TTL_MASK : TCA_FLOWER_KEY_IP_TTL_MASK;
3066
3067         if (fl_dump_key_val(skb, &key->tos, tos_key, &mask->tos, tos_mask, sizeof(key->tos)) ||
3068             fl_dump_key_val(skb, &key->ttl, ttl_key, &mask->ttl, ttl_mask, sizeof(key->ttl)))
3069                 return -1;
3070
3071         return 0;
3072 }
3073
3074 static int fl_dump_key_vlan(struct sk_buff *skb,
3075                             int vlan_id_key, int vlan_prio_key,
3076                             struct flow_dissector_key_vlan *vlan_key,
3077                             struct flow_dissector_key_vlan *vlan_mask)
3078 {
3079         int err;
3080
3081         if (!memchr_inv(vlan_mask, 0, sizeof(*vlan_mask)))
3082                 return 0;
3083         if (vlan_mask->vlan_id) {
3084                 err = nla_put_u16(skb, vlan_id_key,
3085                                   vlan_key->vlan_id);
3086                 if (err)
3087                         return err;
3088         }
3089         if (vlan_mask->vlan_priority) {
3090                 err = nla_put_u8(skb, vlan_prio_key,
3091                                  vlan_key->vlan_priority);
3092                 if (err)
3093                         return err;
3094         }
3095         return 0;
3096 }
3097
3098 static void fl_get_key_flag(u32 dissector_key, u32 dissector_mask,
3099                             u32 *flower_key, u32 *flower_mask,
3100                             u32 flower_flag_bit, u32 dissector_flag_bit)
3101 {
3102         if (dissector_mask & dissector_flag_bit) {
3103                 *flower_mask |= flower_flag_bit;
3104                 if (dissector_key & dissector_flag_bit)
3105                         *flower_key |= flower_flag_bit;
3106         }
3107 }
3108
3109 static int fl_dump_key_flags(struct sk_buff *skb, bool encap,
3110                              u32 flags_key, u32 flags_mask)
3111 {
3112         int fl_key, fl_mask;
3113         __be32 _key, _mask;
3114         u32 key, mask;
3115         int err;
3116
3117         if (encap) {
3118                 fl_key = TCA_FLOWER_KEY_ENC_FLAGS;
3119                 fl_mask = TCA_FLOWER_KEY_ENC_FLAGS_MASK;
3120         } else {
3121                 fl_key = TCA_FLOWER_KEY_FLAGS;
3122                 fl_mask = TCA_FLOWER_KEY_FLAGS_MASK;
3123         }
3124
3125         if (!memchr_inv(&flags_mask, 0, sizeof(flags_mask)))
3126                 return 0;
3127
3128         key = 0;
3129         mask = 0;
3130
3131         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
3132                         TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT, FLOW_DIS_IS_FRAGMENT);
3133         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
3134                         TCA_FLOWER_KEY_FLAGS_FRAG_IS_FIRST,
3135                         FLOW_DIS_FIRST_FRAG);
3136
3137         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
3138                         TCA_FLOWER_KEY_FLAGS_TUNNEL_CSUM,
3139                         FLOW_DIS_F_TUNNEL_CSUM);
3140
3141         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
3142                         TCA_FLOWER_KEY_FLAGS_TUNNEL_DONT_FRAGMENT,
3143                         FLOW_DIS_F_TUNNEL_DONT_FRAGMENT);
3144
3145         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
3146                         TCA_FLOWER_KEY_FLAGS_TUNNEL_OAM, FLOW_DIS_F_TUNNEL_OAM);
3147
3148         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
3149                         TCA_FLOWER_KEY_FLAGS_TUNNEL_CRIT_OPT,
3150                         FLOW_DIS_F_TUNNEL_CRIT_OPT);
3151
3152         _key = cpu_to_be32(key);
3153         _mask = cpu_to_be32(mask);
3154
3155         err = nla_put(skb, fl_key, 4, &_key);
3156         if (err)
3157                 return err;
3158
3159         return nla_put(skb, fl_mask, 4, &_mask);
3160 }
3161
3162 static int fl_dump_key_geneve_opt(struct sk_buff *skb,
3163                                   struct flow_dissector_key_enc_opts *enc_opts)
3164 {
3165         struct geneve_opt *opt;
3166         struct nlattr *nest;
3167         int opt_off = 0;
3168
3169         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_GENEVE);
3170         if (!nest)
3171                 goto nla_put_failure;
3172
3173         while (enc_opts->len > opt_off) {
3174                 opt = (struct geneve_opt *)&enc_opts->data[opt_off];
3175
3176                 if (nla_put_be16(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS,
3177                                  opt->opt_class))
3178                         goto nla_put_failure;
3179                 if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE,
3180                                opt->type))
3181                         goto nla_put_failure;
3182                 if (nla_put(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA,
3183                             opt->length * 4, opt->opt_data))
3184                         goto nla_put_failure;
3185
3186                 opt_off += sizeof(struct geneve_opt) + opt->length * 4;
3187         }
3188         nla_nest_end(skb, nest);
3189         return 0;
3190
3191 nla_put_failure:
3192         nla_nest_cancel(skb, nest);
3193         return -EMSGSIZE;
3194 }
3195
3196 static int fl_dump_key_vxlan_opt(struct sk_buff *skb,
3197                                  struct flow_dissector_key_enc_opts *enc_opts)
3198 {
3199         struct vxlan_metadata *md;
3200         struct nlattr *nest;
3201
3202         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_VXLAN);
3203         if (!nest)
3204                 goto nla_put_failure;
3205
3206         md = (struct vxlan_metadata *)&enc_opts->data[0];
3207         if (nla_put_u32(skb, TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP, md->gbp))
3208                 goto nla_put_failure;
3209
3210         nla_nest_end(skb, nest);
3211         return 0;
3212
3213 nla_put_failure:
3214         nla_nest_cancel(skb, nest);
3215         return -EMSGSIZE;
3216 }
3217
3218 static int fl_dump_key_erspan_opt(struct sk_buff *skb,
3219                                   struct flow_dissector_key_enc_opts *enc_opts)
3220 {
3221         struct erspan_metadata *md;
3222         struct nlattr *nest;
3223
3224         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_ERSPAN);
3225         if (!nest)
3226                 goto nla_put_failure;
3227
3228         md = (struct erspan_metadata *)&enc_opts->data[0];
3229         if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER, md->version))
3230                 goto nla_put_failure;
3231
3232         if (md->version == 1 &&
3233             nla_put_be32(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX, md->u.index))
3234                 goto nla_put_failure;
3235
3236         if (md->version == 2 &&
3237             (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR,
3238                         md->u.md2.dir) ||
3239              nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID,
3240                         get_hwid(&md->u.md2))))
3241                 goto nla_put_failure;
3242
3243         nla_nest_end(skb, nest);
3244         return 0;
3245
3246 nla_put_failure:
3247         nla_nest_cancel(skb, nest);
3248         return -EMSGSIZE;
3249 }
3250
3251 static int fl_dump_key_gtp_opt(struct sk_buff *skb,
3252                                struct flow_dissector_key_enc_opts *enc_opts)
3253
3254 {
3255         struct gtp_pdu_session_info *session_info;
3256         struct nlattr *nest;
3257
3258         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_GTP);
3259         if (!nest)
3260                 goto nla_put_failure;
3261
3262         session_info = (struct gtp_pdu_session_info *)&enc_opts->data[0];
3263
3264         if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE,
3265                        session_info->pdu_type))
3266                 goto nla_put_failure;
3267
3268         if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_GTP_QFI, session_info->qfi))
3269                 goto nla_put_failure;
3270
3271         nla_nest_end(skb, nest);
3272         return 0;
3273
3274 nla_put_failure:
3275         nla_nest_cancel(skb, nest);
3276         return -EMSGSIZE;
3277 }
3278
3279 static int fl_dump_key_pfcp_opt(struct sk_buff *skb,
3280                                 struct flow_dissector_key_enc_opts *enc_opts)
3281 {
3282         struct pfcp_metadata *md;
3283         struct nlattr *nest;
3284
3285         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_PFCP);
3286         if (!nest)
3287                 goto nla_put_failure;
3288
3289         md = (struct pfcp_metadata *)&enc_opts->data[0];
3290         if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_PFCP_TYPE, md->type))
3291                 goto nla_put_failure;
3292
3293         if (nla_put_be64(skb, TCA_FLOWER_KEY_ENC_OPT_PFCP_SEID,
3294                          md->seid, 0))
3295                 goto nla_put_failure;
3296
3297         nla_nest_end(skb, nest);
3298         return 0;
3299
3300 nla_put_failure:
3301         nla_nest_cancel(skb, nest);
3302         return -EMSGSIZE;
3303 }
3304
3305 static int fl_dump_key_ct(struct sk_buff *skb,
3306                           struct flow_dissector_key_ct *key,
3307                           struct flow_dissector_key_ct *mask)
3308 {
3309         if (IS_ENABLED(CONFIG_NF_CONNTRACK) &&
3310             fl_dump_key_val(skb, &key->ct_state, TCA_FLOWER_KEY_CT_STATE,
3311                             &mask->ct_state, TCA_FLOWER_KEY_CT_STATE_MASK,
3312                             sizeof(key->ct_state)))
3313                 goto nla_put_failure;
3314
3315         if (IS_ENABLED(CONFIG_NF_CONNTRACK_ZONES) &&
3316             fl_dump_key_val(skb, &key->ct_zone, TCA_FLOWER_KEY_CT_ZONE,
3317                             &mask->ct_zone, TCA_FLOWER_KEY_CT_ZONE_MASK,
3318                             sizeof(key->ct_zone)))
3319                 goto nla_put_failure;
3320
3321         if (IS_ENABLED(CONFIG_NF_CONNTRACK_MARK) &&
3322             fl_dump_key_val(skb, &key->ct_mark, TCA_FLOWER_KEY_CT_MARK,
3323                             &mask->ct_mark, TCA_FLOWER_KEY_CT_MARK_MASK,
3324                             sizeof(key->ct_mark)))
3325                 goto nla_put_failure;
3326
3327         if (IS_ENABLED(CONFIG_NF_CONNTRACK_LABELS) &&
3328             fl_dump_key_val(skb, &key->ct_labels, TCA_FLOWER_KEY_CT_LABELS,
3329                             &mask->ct_labels, TCA_FLOWER_KEY_CT_LABELS_MASK,
3330                             sizeof(key->ct_labels)))
3331                 goto nla_put_failure;
3332
3333         return 0;
3334
3335 nla_put_failure:
3336         return -EMSGSIZE;
3337 }
3338
3339 static int fl_dump_key_cfm(struct sk_buff *skb,
3340                            struct flow_dissector_key_cfm *key,
3341                            struct flow_dissector_key_cfm *mask)
3342 {
3343         struct nlattr *opts;
3344         int err;
3345         u8 mdl;
3346
3347         if (!memchr_inv(mask, 0, sizeof(*mask)))
3348                 return 0;
3349
3350         opts = nla_nest_start(skb, TCA_FLOWER_KEY_CFM);
3351         if (!opts)
3352                 return -EMSGSIZE;
3353
3354         if (FIELD_GET(FLOW_DIS_CFM_MDL_MASK, mask->mdl_ver)) {
3355                 mdl = FIELD_GET(FLOW_DIS_CFM_MDL_MASK, key->mdl_ver);
3356                 err = nla_put_u8(skb, TCA_FLOWER_KEY_CFM_MD_LEVEL, mdl);
3357                 if (err)
3358                         goto err_cfm_opts;
3359         }
3360
3361         if (mask->opcode) {
3362                 err = nla_put_u8(skb, TCA_FLOWER_KEY_CFM_OPCODE, key->opcode);
3363                 if (err)
3364                         goto err_cfm_opts;
3365         }
3366
3367         nla_nest_end(skb, opts);
3368
3369         return 0;
3370
3371 err_cfm_opts:
3372         nla_nest_cancel(skb, opts);
3373         return err;
3374 }
3375
3376 static int fl_dump_key_options(struct sk_buff *skb, int enc_opt_type,
3377                                struct flow_dissector_key_enc_opts *enc_opts)
3378 {
3379         struct nlattr *nest;
3380         int err;
3381
3382         if (!enc_opts->len)
3383                 return 0;
3384
3385         nest = nla_nest_start_noflag(skb, enc_opt_type);
3386         if (!nest)
3387                 goto nla_put_failure;
3388
3389         switch (enc_opts->dst_opt_type) {
3390         case IP_TUNNEL_GENEVE_OPT_BIT:
3391                 err = fl_dump_key_geneve_opt(skb, enc_opts);
3392                 if (err)
3393                         goto nla_put_failure;
3394                 break;
3395         case IP_TUNNEL_VXLAN_OPT_BIT:
3396                 err = fl_dump_key_vxlan_opt(skb, enc_opts);
3397                 if (err)
3398                         goto nla_put_failure;
3399                 break;
3400         case IP_TUNNEL_ERSPAN_OPT_BIT:
3401                 err = fl_dump_key_erspan_opt(skb, enc_opts);
3402                 if (err)
3403                         goto nla_put_failure;
3404                 break;
3405         case IP_TUNNEL_GTP_OPT_BIT:
3406                 err = fl_dump_key_gtp_opt(skb, enc_opts);
3407                 if (err)
3408                         goto nla_put_failure;
3409                 break;
3410         case IP_TUNNEL_PFCP_OPT_BIT:
3411                 err = fl_dump_key_pfcp_opt(skb, enc_opts);
3412                 if (err)
3413                         goto nla_put_failure;
3414                 break;
3415         default:
3416                 goto nla_put_failure;
3417         }
3418         nla_nest_end(skb, nest);
3419         return 0;
3420
3421 nla_put_failure:
3422         nla_nest_cancel(skb, nest);
3423         return -EMSGSIZE;
3424 }
3425
3426 static int fl_dump_key_enc_opt(struct sk_buff *skb,
3427                                struct flow_dissector_key_enc_opts *key_opts,
3428                                struct flow_dissector_key_enc_opts *msk_opts)
3429 {
3430         int err;
3431
3432         err = fl_dump_key_options(skb, TCA_FLOWER_KEY_ENC_OPTS, key_opts);
3433         if (err)
3434                 return err;
3435
3436         return fl_dump_key_options(skb, TCA_FLOWER_KEY_ENC_OPTS_MASK, msk_opts);
3437 }
3438
3439 static int fl_dump_key(struct sk_buff *skb, struct net *net,
3440                        struct fl_flow_key *key, struct fl_flow_key *mask)
3441 {
3442         if (mask->meta.ingress_ifindex) {
3443                 struct net_device *dev;
3444
3445                 dev = __dev_get_by_index(net, key->meta.ingress_ifindex);
3446                 if (dev && nla_put_string(skb, TCA_FLOWER_INDEV, dev->name))
3447                         goto nla_put_failure;
3448         }
3449
3450         if (fl_dump_key_val(skb, &key->meta.l2_miss,
3451                             TCA_FLOWER_L2_MISS, &mask->meta.l2_miss,
3452                             TCA_FLOWER_UNSPEC, sizeof(key->meta.l2_miss)))
3453                 goto nla_put_failure;
3454
3455         if (fl_dump_key_val(skb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
3456                             mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
3457                             sizeof(key->eth.dst)) ||
3458             fl_dump_key_val(skb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
3459                             mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
3460                             sizeof(key->eth.src)) ||
3461             fl_dump_key_val(skb, &key->basic.n_proto, TCA_FLOWER_KEY_ETH_TYPE,
3462                             &mask->basic.n_proto, TCA_FLOWER_UNSPEC,
3463                             sizeof(key->basic.n_proto)))
3464                 goto nla_put_failure;
3465
3466         if (mask->num_of_vlans.num_of_vlans) {
3467                 if (nla_put_u8(skb, TCA_FLOWER_KEY_NUM_OF_VLANS, key->num_of_vlans.num_of_vlans))
3468                         goto nla_put_failure;
3469         }
3470
3471         if (fl_dump_key_mpls(skb, &key->mpls, &mask->mpls))
3472                 goto nla_put_failure;
3473
3474         if (fl_dump_key_vlan(skb, TCA_FLOWER_KEY_VLAN_ID,
3475                              TCA_FLOWER_KEY_VLAN_PRIO, &key->vlan, &mask->vlan))
3476                 goto nla_put_failure;
3477
3478         if (fl_dump_key_vlan(skb, TCA_FLOWER_KEY_CVLAN_ID,
3479                              TCA_FLOWER_KEY_CVLAN_PRIO,
3480                              &key->cvlan, &mask->cvlan) ||
3481             (mask->cvlan.vlan_tpid &&
3482              nla_put_be16(skb, TCA_FLOWER_KEY_VLAN_ETH_TYPE,
3483                           key->cvlan.vlan_tpid)))
3484                 goto nla_put_failure;
3485
3486         if (mask->basic.n_proto) {
3487                 if (mask->cvlan.vlan_eth_type) {
3488                         if (nla_put_be16(skb, TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
3489                                          key->basic.n_proto))
3490                                 goto nla_put_failure;
3491                 } else if (mask->vlan.vlan_eth_type) {
3492                         if (nla_put_be16(skb, TCA_FLOWER_KEY_VLAN_ETH_TYPE,
3493                                          key->vlan.vlan_eth_type))
3494                                 goto nla_put_failure;
3495                 }
3496         }
3497
3498         if ((key->basic.n_proto == htons(ETH_P_IP) ||
3499              key->basic.n_proto == htons(ETH_P_IPV6)) &&
3500             (fl_dump_key_val(skb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
3501                             &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
3502                             sizeof(key->basic.ip_proto)) ||
3503             fl_dump_key_ip(skb, false, &key->ip, &mask->ip)))
3504                 goto nla_put_failure;
3505
3506         if (mask->pppoe.session_id) {
3507                 if (nla_put_be16(skb, TCA_FLOWER_KEY_PPPOE_SID,
3508                                  key->pppoe.session_id))
3509                         goto nla_put_failure;
3510         }
3511         if (mask->basic.n_proto && mask->pppoe.ppp_proto) {
3512                 if (nla_put_be16(skb, TCA_FLOWER_KEY_PPP_PROTO,
3513                                  key->pppoe.ppp_proto))
3514                         goto nla_put_failure;
3515         }
3516
3517         if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
3518             (fl_dump_key_val(skb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
3519                              &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
3520                              sizeof(key->ipv4.src)) ||
3521              fl_dump_key_val(skb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
3522                              &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
3523                              sizeof(key->ipv4.dst))))
3524                 goto nla_put_failure;
3525         else if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
3526                  (fl_dump_key_val(skb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
3527                                   &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
3528                                   sizeof(key->ipv6.src)) ||
3529                   fl_dump_key_val(skb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
3530                                   &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
3531                                   sizeof(key->ipv6.dst))))
3532                 goto nla_put_failure;
3533
3534         if (key->basic.ip_proto == IPPROTO_TCP &&
3535             (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
3536                              &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
3537                              sizeof(key->tp.src)) ||
3538              fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
3539                              &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
3540                              sizeof(key->tp.dst)) ||
3541              fl_dump_key_val(skb, &key->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS,
3542                              &mask->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS_MASK,
3543                              sizeof(key->tcp.flags))))
3544                 goto nla_put_failure;
3545         else if (key->basic.ip_proto == IPPROTO_UDP &&
3546                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
3547                                   &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
3548                                   sizeof(key->tp.src)) ||
3549                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
3550                                   &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
3551                                   sizeof(key->tp.dst))))
3552                 goto nla_put_failure;
3553         else if (key->basic.ip_proto == IPPROTO_SCTP &&
3554                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_SCTP_SRC,
3555                                   &mask->tp.src, TCA_FLOWER_KEY_SCTP_SRC_MASK,
3556                                   sizeof(key->tp.src)) ||
3557                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_SCTP_DST,
3558                                   &mask->tp.dst, TCA_FLOWER_KEY_SCTP_DST_MASK,
3559                                   sizeof(key->tp.dst))))
3560                 goto nla_put_failure;
3561         else if (key->basic.n_proto == htons(ETH_P_IP) &&
3562                  key->basic.ip_proto == IPPROTO_ICMP &&
3563                  (fl_dump_key_val(skb, &key->icmp.type,
3564                                   TCA_FLOWER_KEY_ICMPV4_TYPE, &mask->icmp.type,
3565                                   TCA_FLOWER_KEY_ICMPV4_TYPE_MASK,
3566                                   sizeof(key->icmp.type)) ||
3567                   fl_dump_key_val(skb, &key->icmp.code,
3568                                   TCA_FLOWER_KEY_ICMPV4_CODE, &mask->icmp.code,
3569                                   TCA_FLOWER_KEY_ICMPV4_CODE_MASK,
3570                                   sizeof(key->icmp.code))))
3571                 goto nla_put_failure;
3572         else if (key->basic.n_proto == htons(ETH_P_IPV6) &&
3573                  key->basic.ip_proto == IPPROTO_ICMPV6 &&
3574                  (fl_dump_key_val(skb, &key->icmp.type,
3575                                   TCA_FLOWER_KEY_ICMPV6_TYPE, &mask->icmp.type,
3576                                   TCA_FLOWER_KEY_ICMPV6_TYPE_MASK,
3577                                   sizeof(key->icmp.type)) ||
3578                   fl_dump_key_val(skb, &key->icmp.code,
3579                                   TCA_FLOWER_KEY_ICMPV6_CODE, &mask->icmp.code,
3580                                   TCA_FLOWER_KEY_ICMPV6_CODE_MASK,
3581                                   sizeof(key->icmp.code))))
3582                 goto nla_put_failure;
3583         else if ((key->basic.n_proto == htons(ETH_P_ARP) ||
3584                   key->basic.n_proto == htons(ETH_P_RARP)) &&
3585                  (fl_dump_key_val(skb, &key->arp.sip,
3586                                   TCA_FLOWER_KEY_ARP_SIP, &mask->arp.sip,
3587                                   TCA_FLOWER_KEY_ARP_SIP_MASK,
3588                                   sizeof(key->arp.sip)) ||
3589                   fl_dump_key_val(skb, &key->arp.tip,
3590                                   TCA_FLOWER_KEY_ARP_TIP, &mask->arp.tip,
3591                                   TCA_FLOWER_KEY_ARP_TIP_MASK,
3592                                   sizeof(key->arp.tip)) ||
3593                   fl_dump_key_val(skb, &key->arp.op,
3594                                   TCA_FLOWER_KEY_ARP_OP, &mask->arp.op,
3595                                   TCA_FLOWER_KEY_ARP_OP_MASK,
3596                                   sizeof(key->arp.op)) ||
3597                   fl_dump_key_val(skb, key->arp.sha, TCA_FLOWER_KEY_ARP_SHA,
3598                                   mask->arp.sha, TCA_FLOWER_KEY_ARP_SHA_MASK,
3599                                   sizeof(key->arp.sha)) ||
3600                   fl_dump_key_val(skb, key->arp.tha, TCA_FLOWER_KEY_ARP_THA,
3601                                   mask->arp.tha, TCA_FLOWER_KEY_ARP_THA_MASK,
3602                                   sizeof(key->arp.tha))))
3603                 goto nla_put_failure;
3604         else if (key->basic.ip_proto == IPPROTO_L2TP &&
3605                  fl_dump_key_val(skb, &key->l2tpv3.session_id,
3606                                  TCA_FLOWER_KEY_L2TPV3_SID,
3607                                  &mask->l2tpv3.session_id,
3608                                  TCA_FLOWER_UNSPEC,
3609                                  sizeof(key->l2tpv3.session_id)))
3610                 goto nla_put_failure;
3611
3612         if (key->ipsec.spi &&
3613             fl_dump_key_val(skb, &key->ipsec.spi, TCA_FLOWER_KEY_SPI,
3614                             &mask->ipsec.spi, TCA_FLOWER_KEY_SPI_MASK,
3615                             sizeof(key->ipsec.spi)))
3616                 goto nla_put_failure;
3617
3618         if ((key->basic.ip_proto == IPPROTO_TCP ||
3619              key->basic.ip_proto == IPPROTO_UDP ||
3620              key->basic.ip_proto == IPPROTO_SCTP) &&
3621              fl_dump_key_port_range(skb, key, mask))
3622                 goto nla_put_failure;
3623
3624         if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
3625             (fl_dump_key_val(skb, &key->enc_ipv4.src,
3626                             TCA_FLOWER_KEY_ENC_IPV4_SRC, &mask->enc_ipv4.src,
3627                             TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
3628                             sizeof(key->enc_ipv4.src)) ||
3629              fl_dump_key_val(skb, &key->enc_ipv4.dst,
3630                              TCA_FLOWER_KEY_ENC_IPV4_DST, &mask->enc_ipv4.dst,
3631                              TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
3632                              sizeof(key->enc_ipv4.dst))))
3633                 goto nla_put_failure;
3634         else if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
3635                  (fl_dump_key_val(skb, &key->enc_ipv6.src,
3636                             TCA_FLOWER_KEY_ENC_IPV6_SRC, &mask->enc_ipv6.src,
3637                             TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
3638                             sizeof(key->enc_ipv6.src)) ||
3639                  fl_dump_key_val(skb, &key->enc_ipv6.dst,
3640                                  TCA_FLOWER_KEY_ENC_IPV6_DST,
3641                                  &mask->enc_ipv6.dst,
3642                                  TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
3643                             sizeof(key->enc_ipv6.dst))))
3644                 goto nla_put_failure;
3645
3646         if (fl_dump_key_val(skb, &key->enc_key_id, TCA_FLOWER_KEY_ENC_KEY_ID,
3647                             &mask->enc_key_id, TCA_FLOWER_UNSPEC,
3648                             sizeof(key->enc_key_id)) ||
3649             fl_dump_key_val(skb, &key->enc_tp.src,
3650                             TCA_FLOWER_KEY_ENC_UDP_SRC_PORT,
3651                             &mask->enc_tp.src,
3652                             TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK,
3653                             sizeof(key->enc_tp.src)) ||
3654             fl_dump_key_val(skb, &key->enc_tp.dst,
3655                             TCA_FLOWER_KEY_ENC_UDP_DST_PORT,
3656                             &mask->enc_tp.dst,
3657                             TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK,
3658                             sizeof(key->enc_tp.dst)) ||
3659             fl_dump_key_ip(skb, true, &key->enc_ip, &mask->enc_ip) ||
3660             fl_dump_key_enc_opt(skb, &key->enc_opts, &mask->enc_opts))
3661                 goto nla_put_failure;
3662
3663         if (fl_dump_key_ct(skb, &key->ct, &mask->ct))
3664                 goto nla_put_failure;
3665
3666         if (fl_dump_key_flags(skb, false, key->control.flags,
3667                               mask->control.flags))
3668                 goto nla_put_failure;
3669
3670         if (fl_dump_key_val(skb, &key->hash.hash, TCA_FLOWER_KEY_HASH,
3671                              &mask->hash.hash, TCA_FLOWER_KEY_HASH_MASK,
3672                              sizeof(key->hash.hash)))
3673                 goto nla_put_failure;
3674
3675         if (fl_dump_key_cfm(skb, &key->cfm, &mask->cfm))
3676                 goto nla_put_failure;
3677
3678         if (fl_dump_key_flags(skb, true, key->enc_control.flags,
3679                               mask->enc_control.flags))
3680                 goto nla_put_failure;
3681
3682         return 0;
3683
3684 nla_put_failure:
3685         return -EMSGSIZE;
3686 }
3687
3688 static int fl_dump(struct net *net, struct tcf_proto *tp, void *fh,
3689                    struct sk_buff *skb, struct tcmsg *t, bool rtnl_held)
3690 {
3691         struct cls_fl_filter *f = fh;
3692         struct nlattr *nest;
3693         struct fl_flow_key *key, *mask;
3694         bool skip_hw;
3695
3696         if (!f)
3697                 return skb->len;
3698
3699         t->tcm_handle = f->handle;
3700
3701         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
3702         if (!nest)
3703                 goto nla_put_failure;
3704
3705         spin_lock(&tp->lock);
3706
3707         if (f->res.classid &&
3708             nla_put_u32(skb, TCA_FLOWER_CLASSID, f->res.classid))
3709                 goto nla_put_failure_locked;
3710
3711         key = &f->key;
3712         mask = &f->mask->key;
3713         skip_hw = tc_skip_hw(f->flags);
3714
3715         if (fl_dump_key(skb, net, key, mask))
3716                 goto nla_put_failure_locked;
3717
3718         if (f->flags && nla_put_u32(skb, TCA_FLOWER_FLAGS, f->flags))
3719                 goto nla_put_failure_locked;
3720
3721         spin_unlock(&tp->lock);
3722
3723         if (!skip_hw)
3724                 fl_hw_update_stats(tp, f, rtnl_held);
3725
3726         if (nla_put_u32(skb, TCA_FLOWER_IN_HW_COUNT, f->in_hw_count))
3727                 goto nla_put_failure;
3728
3729         if (tcf_exts_dump(skb, &f->exts))
3730                 goto nla_put_failure;
3731
3732         nla_nest_end(skb, nest);
3733
3734         if (tcf_exts_dump_stats(skb, &f->exts) < 0)
3735                 goto nla_put_failure;
3736
3737         return skb->len;
3738
3739 nla_put_failure_locked:
3740         spin_unlock(&tp->lock);
3741 nla_put_failure:
3742         nla_nest_cancel(skb, nest);
3743         return -1;
3744 }
3745
3746 static int fl_terse_dump(struct net *net, struct tcf_proto *tp, void *fh,
3747                          struct sk_buff *skb, struct tcmsg *t, bool rtnl_held)
3748 {
3749         struct cls_fl_filter *f = fh;
3750         struct nlattr *nest;
3751         bool skip_hw;
3752
3753         if (!f)
3754                 return skb->len;
3755
3756         t->tcm_handle = f->handle;
3757
3758         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
3759         if (!nest)
3760                 goto nla_put_failure;
3761
3762         spin_lock(&tp->lock);
3763
3764         skip_hw = tc_skip_hw(f->flags);
3765
3766         if (f->flags && nla_put_u32(skb, TCA_FLOWER_FLAGS, f->flags))
3767                 goto nla_put_failure_locked;
3768
3769         spin_unlock(&tp->lock);
3770
3771         if (!skip_hw)
3772                 fl_hw_update_stats(tp, f, rtnl_held);
3773
3774         if (tcf_exts_terse_dump(skb, &f->exts))
3775                 goto nla_put_failure;
3776
3777         nla_nest_end(skb, nest);
3778
3779         return skb->len;
3780
3781 nla_put_failure_locked:
3782         spin_unlock(&tp->lock);
3783 nla_put_failure:
3784         nla_nest_cancel(skb, nest);
3785         return -1;
3786 }
3787
3788 static int fl_tmplt_dump(struct sk_buff *skb, struct net *net, void *tmplt_priv)
3789 {
3790         struct fl_flow_tmplt *tmplt = tmplt_priv;
3791         struct fl_flow_key *key, *mask;
3792         struct nlattr *nest;
3793
3794         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
3795         if (!nest)
3796                 goto nla_put_failure;
3797
3798         key = &tmplt->dummy_key;
3799         mask = &tmplt->mask;
3800
3801         if (fl_dump_key(skb, net, key, mask))
3802                 goto nla_put_failure;
3803
3804         nla_nest_end(skb, nest);
3805
3806         return skb->len;
3807
3808 nla_put_failure:
3809         nla_nest_cancel(skb, nest);
3810         return -EMSGSIZE;
3811 }
3812
3813 static void fl_bind_class(void *fh, u32 classid, unsigned long cl, void *q,
3814                           unsigned long base)
3815 {
3816         struct cls_fl_filter *f = fh;
3817
3818         tc_cls_bind_class(classid, cl, q, &f->res, base);
3819 }
3820
3821 static bool fl_delete_empty(struct tcf_proto *tp)
3822 {
3823         struct cls_fl_head *head = fl_head_dereference(tp);
3824
3825         spin_lock(&tp->lock);
3826         tp->deleting = idr_is_empty(&head->handle_idr);
3827         spin_unlock(&tp->lock);
3828
3829         return tp->deleting;
3830 }
3831
3832 static struct tcf_proto_ops cls_fl_ops __read_mostly = {
3833         .kind           = "flower",
3834         .classify       = fl_classify,
3835         .init           = fl_init,
3836         .destroy        = fl_destroy,
3837         .get            = fl_get,
3838         .put            = fl_put,
3839         .change         = fl_change,
3840         .delete         = fl_delete,
3841         .delete_empty   = fl_delete_empty,
3842         .walk           = fl_walk,
3843         .reoffload      = fl_reoffload,
3844         .hw_add         = fl_hw_add,
3845         .hw_del         = fl_hw_del,
3846         .dump           = fl_dump,
3847         .terse_dump     = fl_terse_dump,
3848         .bind_class     = fl_bind_class,
3849         .tmplt_create   = fl_tmplt_create,
3850         .tmplt_destroy  = fl_tmplt_destroy,
3851         .tmplt_reoffload = fl_tmplt_reoffload,
3852         .tmplt_dump     = fl_tmplt_dump,
3853         .get_exts       = fl_get_exts,
3854         .owner          = THIS_MODULE,
3855         .flags          = TCF_PROTO_OPS_DOIT_UNLOCKED,
3856 };
3857 MODULE_ALIAS_NET_CLS("flower");
3858
3859 static int __init cls_fl_init(void)
3860 {
3861         return register_tcf_proto_ops(&cls_fl_ops);
3862 }
3863
3864 static void __exit cls_fl_exit(void)
3865 {
3866         unregister_tcf_proto_ops(&cls_fl_ops);
3867 }
3868
3869 module_init(cls_fl_init);
3870 module_exit(cls_fl_exit);
3871
3872 MODULE_AUTHOR("Jiri Pirko <jiri@resnulli.us>");
3873 MODULE_DESCRIPTION("Flower classifier");
3874 MODULE_LICENSE("GPL v2");