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
  15 #include <linux/if_ether.h>
  16 #include <linux/in6.h>
  17 #include <linux/ip.h>
  18 #include <linux/mpls.h>
  19
  20 #include <net/sch_generic.h>
  21 #include <net/pkt_cls.h>
  22 #include <net/ip.h>
  23 #include <net/flow_dissector.h>
  24 #include <net/geneve.h>
  25 #include <net/vxlan.h>
  26 #include <net/erspan.h>
  27
  28 #include <net/dst.h>
  29 #include <net/dst_metadata.h>
  30
  31 #include <uapi/linux/netfilter/nf_conntrack_common.h>
  32
  33 #define TCA_FLOWER_KEY_CT_FLAGS_MAX \
  34                 ((__TCA_FLOWER_KEY_CT_FLAGS_MAX - 1) << 1)
  35 #define TCA_FLOWER_KEY_CT_FLAGS_MASK \
  36                 (TCA_FLOWER_KEY_CT_FLAGS_MAX - 1)
  37
  38 struct fl_flow_key {
  39         struct flow_dissector_key_meta meta;
  40         struct flow_dissector_key_control control;
  41         struct flow_dissector_key_control enc_control;
  42         struct flow_dissector_key_basic basic;
  43         struct flow_dissector_key_eth_addrs eth;
  44         struct flow_dissector_key_vlan vlan;
  45         struct flow_dissector_key_vlan cvlan;
  46         union {
  47                 struct flow_dissector_key_ipv4_addrs ipv4;
  48                 struct flow_dissector_key_ipv6_addrs ipv6;
  49         };
  50         struct flow_dissector_key_ports tp;
  51         struct flow_dissector_key_icmp icmp;
  52         struct flow_dissector_key_arp arp;
  53         struct flow_dissector_key_keyid enc_key_id;
  54         union {
  55                 struct flow_dissector_key_ipv4_addrs enc_ipv4;
  56                 struct flow_dissector_key_ipv6_addrs enc_ipv6;
  57         };
  58         struct flow_dissector_key_ports enc_tp;
  59         struct flow_dissector_key_mpls mpls;
  60         struct flow_dissector_key_tcp tcp;
  61         struct flow_dissector_key_ip ip;
  62         struct flow_dissector_key_ip enc_ip;
  63         struct flow_dissector_key_enc_opts enc_opts;
  64         union {
  65                 struct flow_dissector_key_ports tp;
  66                 struct {
  67                         struct flow_dissector_key_ports tp_min;
  68                         struct flow_dissector_key_ports tp_max;
  69                 };
  70         } tp_range;
  71         struct flow_dissector_key_ct ct;
  72         struct flow_dissector_key_hash hash;
  73 } __aligned(BITS_PER_LONG / 8); /* Ensure that we can do comparisons as longs. */
  74
  75 struct fl_flow_mask_range {
  76         unsigned short int start;
  77         unsigned short int end;
  78 };
  79
  80 struct fl_flow_mask {
  81         struct fl_flow_key key;
  82         struct fl_flow_mask_range range;
  83         u32 flags;
  84         struct rhash_head ht_node;
  85         struct rhashtable ht;
  86         struct rhashtable_params filter_ht_params;
  87         struct flow_dissector dissector;
  88         struct list_head filters;
  89         struct rcu_work rwork;
  90         struct list_head list;
  91         refcount_t refcnt;
  92 };
  93
  94 struct fl_flow_tmplt {
  95         struct fl_flow_key dummy_key;
  96         struct fl_flow_key mask;
  97         struct flow_dissector dissector;
  98         struct tcf_chain *chain;
  99 };
 100
 101 struct cls_fl_head {
 102         struct rhashtable ht;
 103         spinlock_t masks_lock; /* Protect masks list */
 104         struct list_head masks;
 105         struct list_head hw_filters;
 106         struct rcu_work rwork;
 107         struct idr handle_idr;
 108 };
 109
 110 struct cls_fl_filter {
 111         struct fl_flow_mask *mask;
 112         struct rhash_head ht_node;
 113         struct fl_flow_key mkey;
 114         struct tcf_exts exts;
 115         struct tcf_result res;
 116         struct fl_flow_key key;
 117         struct list_head list;
 118         struct list_head hw_list;
 119         u32 handle;
 120         u32 flags;
 121         u32 in_hw_count;
 122         struct rcu_work rwork;
 123         struct net_device *hw_dev;
 124         /* Flower classifier is unlocked, which means that its reference counter
 125          * can be changed concurrently without any kind of external
 126          * synchronization. Use atomic reference counter to be concurrency-safe.
 127          */
 128         refcount_t refcnt;
 129         bool deleted;
 130 };
 131
 132 static const struct rhashtable_params mask_ht_params = {
 133         .key_offset = offsetof(struct fl_flow_mask, key),
 134         .key_len = sizeof(struct fl_flow_key),
 135         .head_offset = offsetof(struct fl_flow_mask, ht_node),
 136         .automatic_shrinking = true,
 137 };
 138
 139 static unsigned short int fl_mask_range(const struct fl_flow_mask *mask)
 140 {
 141         return mask->range.end - mask->range.start;
 142 }
 143
 144 static void fl_mask_update_range(struct fl_flow_mask *mask)
 145 {
 146         const u8 *bytes = (const u8 *) &mask->key;
 147         size_t size = sizeof(mask->key);
 148         size_t i, first = 0, last;
 149
 150         for (i = 0; i < size; i++) {
 151                 if (bytes[i]) {
 152                         first = i;
 153                         break;
 154                 }
 155         }
 156         last = first;
 157         for (i = size - 1; i != first; i--) {
 158                 if (bytes[i]) {
 159                         last = i;
 160                         break;
 161                 }
 162         }
 163         mask->range.start = rounddown(first, sizeof(long));
 164         mask->range.end = roundup(last + 1, sizeof(long));
 165 }
 166
 167 static void *fl_key_get_start(struct fl_flow_key *key,
 168                               const struct fl_flow_mask *mask)
 169 {
 170         return (u8 *) key + mask->range.start;
 171 }
 172
 173 static void fl_set_masked_key(struct fl_flow_key *mkey, struct fl_flow_key *key,
 174                               struct fl_flow_mask *mask)
 175 {
 176         const long *lkey = fl_key_get_start(key, mask);
 177         const long *lmask = fl_key_get_start(&mask->key, mask);
 178         long *lmkey = fl_key_get_start(mkey, mask);
 179         int i;
 180
 181         for (i = 0; i < fl_mask_range(mask); i += sizeof(long))
 182                 *lmkey++ = *lkey++ & *lmask++;
 183 }
 184
 185 static bool fl_mask_fits_tmplt(struct fl_flow_tmplt *tmplt,
 186                                struct fl_flow_mask *mask)
 187 {
 188         const long *lmask = fl_key_get_start(&mask->key, mask);
 189         const long *ltmplt;
 190         int i;
 191
 192         if (!tmplt)
 193                 return true;
 194         ltmplt = fl_key_get_start(&tmplt->mask, mask);
 195         for (i = 0; i < fl_mask_range(mask); i += sizeof(long)) {
 196                 if (~*ltmplt++ & *lmask++)
 197                         return false;
 198         }
 199         return true;
 200 }
 201
 202 static void fl_clear_masked_range(struct fl_flow_key *key,
 203                                   struct fl_flow_mask *mask)
 204 {
 205         memset(fl_key_get_start(key, mask), 0, fl_mask_range(mask));
 206 }
 207
 208 static bool fl_range_port_dst_cmp(struct cls_fl_filter *filter,
 209                                   struct fl_flow_key *key,
 210                                   struct fl_flow_key *mkey)
 211 {
 212         u16 min_mask, max_mask, min_val, max_val;
 213
 214         min_mask = ntohs(filter->mask->key.tp_range.tp_min.dst);
 215         max_mask = ntohs(filter->mask->key.tp_range.tp_max.dst);
 216         min_val = ntohs(filter->key.tp_range.tp_min.dst);
 217         max_val = ntohs(filter->key.tp_range.tp_max.dst);
 218
 219         if (min_mask && max_mask) {
 220                 if (ntohs(key->tp_range.tp.dst) < min_val ||
 221                     ntohs(key->tp_range.tp.dst) > max_val)
 222                         return false;
 223
 224                 /* skb does not have min and max values */
 225                 mkey->tp_range.tp_min.dst = filter->mkey.tp_range.tp_min.dst;
 226                 mkey->tp_range.tp_max.dst = filter->mkey.tp_range.tp_max.dst;
 227         }
 228         return true;
 229 }
 230
 231 static bool fl_range_port_src_cmp(struct cls_fl_filter *filter,
 232                                   struct fl_flow_key *key,
 233                                   struct fl_flow_key *mkey)
 234 {
 235         u16 min_mask, max_mask, min_val, max_val;
 236
 237         min_mask = ntohs(filter->mask->key.tp_range.tp_min.src);
 238         max_mask = ntohs(filter->mask->key.tp_range.tp_max.src);
 239         min_val = ntohs(filter->key.tp_range.tp_min.src);
 240         max_val = ntohs(filter->key.tp_range.tp_max.src);
 241
 242         if (min_mask && max_mask) {
 243                 if (ntohs(key->tp_range.tp.src) < min_val ||
 244                     ntohs(key->tp_range.tp.src) > max_val)
 245                         return false;
 246
 247                 /* skb does not have min and max values */
 248                 mkey->tp_range.tp_min.src = filter->mkey.tp_range.tp_min.src;
 249                 mkey->tp_range.tp_max.src = filter->mkey.tp_range.tp_max.src;
 250         }
 251         return true;
 252 }
 253
 254 static struct cls_fl_filter *__fl_lookup(struct fl_flow_mask *mask,
 255                                          struct fl_flow_key *mkey)
 256 {
 257         return rhashtable_lookup_fast(&mask->ht, fl_key_get_start(mkey, mask),
 258                                       mask->filter_ht_params);
 259 }
 260
 261 static struct cls_fl_filter *fl_lookup_range(struct fl_flow_mask *mask,
 262                                              struct fl_flow_key *mkey,
 263                                              struct fl_flow_key *key)
 264 {
 265         struct cls_fl_filter *filter, *f;
 266
 267         list_for_each_entry_rcu(filter, &mask->filters, list) {
 268                 if (!fl_range_port_dst_cmp(filter, key, mkey))
 269                         continue;
 270
 271                 if (!fl_range_port_src_cmp(filter, key, mkey))
 272                         continue;
 273
 274                 f = __fl_lookup(mask, mkey);
 275                 if (f)
 276                         return f;
 277         }
 278         return NULL;
 279 }
 280
 281 static noinline_for_stack
 282 struct cls_fl_filter *fl_mask_lookup(struct fl_flow_mask *mask, struct fl_flow_key *key)
 283 {
 284         struct fl_flow_key mkey;
 285
 286         fl_set_masked_key(&mkey, key, mask);
 287         if ((mask->flags & TCA_FLOWER_MASK_FLAGS_RANGE))
 288                 return fl_lookup_range(mask, &mkey, key);
 289
 290         return __fl_lookup(mask, &mkey);
 291 }
 292
 293 static u16 fl_ct_info_to_flower_map[] = {
 294         [IP_CT_ESTABLISHED] =           TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 295                                         TCA_FLOWER_KEY_CT_FLAGS_ESTABLISHED,
 296         [IP_CT_RELATED] =               TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 297                                         TCA_FLOWER_KEY_CT_FLAGS_RELATED,
 298         [IP_CT_ESTABLISHED_REPLY] =     TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 299                                         TCA_FLOWER_KEY_CT_FLAGS_ESTABLISHED,
 300         [IP_CT_RELATED_REPLY] =         TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 301                                         TCA_FLOWER_KEY_CT_FLAGS_RELATED,
 302         [IP_CT_NEW] =                   TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 303                                         TCA_FLOWER_KEY_CT_FLAGS_NEW,
 304 };
 305
 306 static int fl_classify(struct sk_buff *skb, const struct tcf_proto *tp,
 307                        struct tcf_result *res)
 308 {
 309         struct cls_fl_head *head = rcu_dereference_bh(tp->root);
 310         struct fl_flow_key skb_key;
 311         struct fl_flow_mask *mask;
 312         struct cls_fl_filter *f;
 313
 314         list_for_each_entry_rcu(mask, &head->masks, list) {
 315                 flow_dissector_init_keys(&skb_key.control, &skb_key.basic);
 316                 fl_clear_masked_range(&skb_key, mask);
 317
 318                 skb_flow_dissect_meta(skb, &mask->dissector, &skb_key);
 319                 /* skb_flow_dissect() does not set n_proto in case an unknown
 320                  * protocol, so do it rather here.
 321                  */
 322                 skb_key.basic.n_proto = skb_protocol(skb, false);
 323                 skb_flow_dissect_tunnel_info(skb, &mask->dissector, &skb_key);
 324                 skb_flow_dissect_ct(skb, &mask->dissector, &skb_key,
 325                                     fl_ct_info_to_flower_map,
 326                                     ARRAY_SIZE(fl_ct_info_to_flower_map));
 327                 skb_flow_dissect_hash(skb, &mask->dissector, &skb_key);
 328                 skb_flow_dissect(skb, &mask->dissector, &skb_key, 0);
 329
 330                 f = fl_mask_lookup(mask, &skb_key);
 331                 if (f && !tc_skip_sw(f->flags)) {
 332                         *res = f->res;
 333                         return tcf_exts_exec(skb, &f->exts, res);
 334                 }
 335         }
 336         return -1;
 337 }
 338
 339 static int fl_init(struct tcf_proto *tp)
 340 {
 341         struct cls_fl_head *head;
 342
 343         head = kzalloc(sizeof(*head), GFP_KERNEL);
 344         if (!head)
 345                 return -ENOBUFS;
 346
 347         spin_lock_init(&head->masks_lock);
 348         INIT_LIST_HEAD_RCU(&head->masks);
 349         INIT_LIST_HEAD(&head->hw_filters);
 350         rcu_assign_pointer(tp->root, head);
 351         idr_init(&head->handle_idr);
 352
 353         return rhashtable_init(&head->ht, &mask_ht_params);
 354 }
 355
 356 static void fl_mask_free(struct fl_flow_mask *mask, bool mask_init_done)
 357 {
 358         /* temporary masks don't have their filters list and ht initialized */
 359         if (mask_init_done) {
 360                 WARN_ON(!list_empty(&mask->filters));
 361                 rhashtable_destroy(&mask->ht);
 362         }
 363         kfree(mask);
 364 }
 365
 366 static void fl_mask_free_work(struct work_struct *work)
 367 {
 368         struct fl_flow_mask *mask = container_of(to_rcu_work(work),
 369                                                  struct fl_flow_mask, rwork);
 370
 371         fl_mask_free(mask, true);
 372 }
 373
 374 static void fl_uninit_mask_free_work(struct work_struct *work)
 375 {
 376         struct fl_flow_mask *mask = container_of(to_rcu_work(work),
 377                                                  struct fl_flow_mask, rwork);
 378
 379         fl_mask_free(mask, false);
 380 }
 381
 382 static bool fl_mask_put(struct cls_fl_head *head, struct fl_flow_mask *mask)
 383 {
 384         if (!refcount_dec_and_test(&mask->refcnt))
 385                 return false;
 386
 387         rhashtable_remove_fast(&head->ht, &mask->ht_node, mask_ht_params);
 388
 389         spin_lock(&head->masks_lock);
 390         list_del_rcu(&mask->list);
 391         spin_unlock(&head->masks_lock);
 392
 393         tcf_queue_work(&mask->rwork, fl_mask_free_work);
 394
 395         return true;
 396 }
 397
 398 static struct cls_fl_head *fl_head_dereference(struct tcf_proto *tp)
 399 {
 400         /* Flower classifier only changes root pointer during init and destroy.
 401          * Users must obtain reference to tcf_proto instance before calling its
 402          * API, so tp->root pointer is protected from concurrent call to
 403          * fl_destroy() by reference counting.
 404          */
 405         return rcu_dereference_raw(tp->root);
 406 }
 407
 408 static void __fl_destroy_filter(struct cls_fl_filter *f)
 409 {
 410         tcf_exts_destroy(&f->exts);
 411         tcf_exts_put_net(&f->exts);
 412         kfree(f);
 413 }
 414
 415 static void fl_destroy_filter_work(struct work_struct *work)
 416 {
 417         struct cls_fl_filter *f = container_of(to_rcu_work(work),
 418                                         struct cls_fl_filter, rwork);
 419
 420         __fl_destroy_filter(f);
 421 }
 422
 423 static void fl_hw_destroy_filter(struct tcf_proto *tp, struct cls_fl_filter *f,
 424                                  bool rtnl_held, struct netlink_ext_ack *extack)
 425 {
 426         struct tcf_block *block = tp->chain->block;
 427         struct flow_cls_offload cls_flower = {};
 428
 429         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, extack);
 430         cls_flower.command = FLOW_CLS_DESTROY;
 431         cls_flower.cookie = (unsigned long) f;
 432
 433         tc_setup_cb_destroy(block, tp, TC_SETUP_CLSFLOWER, &cls_flower, false,
 434                             &f->flags, &f->in_hw_count, rtnl_held);
 435
 436 }
 437
 438 static int fl_hw_replace_filter(struct tcf_proto *tp,
 439                                 struct cls_fl_filter *f, bool rtnl_held,
 440                                 struct netlink_ext_ack *extack)
 441 {
 442         struct tcf_block *block = tp->chain->block;
 443         struct flow_cls_offload cls_flower = {};
 444         bool skip_sw = tc_skip_sw(f->flags);
 445         int err = 0;
 446
 447         cls_flower.rule = flow_rule_alloc(tcf_exts_num_actions(&f->exts));
 448         if (!cls_flower.rule)
 449                 return -ENOMEM;
 450
 451         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, extack);
 452         cls_flower.command = FLOW_CLS_REPLACE;
 453         cls_flower.cookie = (unsigned long) f;
 454         cls_flower.rule->match.dissector = &f->mask->dissector;
 455         cls_flower.rule->match.mask = &f->mask->key;
 456         cls_flower.rule->match.key = &f->mkey;
 457         cls_flower.classid = f->res.classid;
 458
 459         err = tc_setup_flow_action(&cls_flower.rule->action, &f->exts);
 460         if (err) {
 461                 kfree(cls_flower.rule);
 462                 if (skip_sw) {
 463                         NL_SET_ERR_MSG_MOD(extack, "Failed to setup flow action");
 464                         return err;
 465                 }
 466                 return 0;
 467         }
 468
 469         err = tc_setup_cb_add(block, tp, TC_SETUP_CLSFLOWER, &cls_flower,
 470                               skip_sw, &f->flags, &f->in_hw_count, rtnl_held);
 471         tc_cleanup_flow_action(&cls_flower.rule->action);
 472         kfree(cls_flower.rule);
 473
 474         if (err) {
 475                 fl_hw_destroy_filter(tp, f, rtnl_held, NULL);
 476                 return err;
 477         }
 478
 479         if (skip_sw && !(f->flags & TCA_CLS_FLAGS_IN_HW))
 480                 return -EINVAL;
 481
 482         return 0;
 483 }
 484
 485 static void fl_hw_update_stats(struct tcf_proto *tp, struct cls_fl_filter *f,
 486                                bool rtnl_held)
 487 {
 488         struct tcf_block *block = tp->chain->block;
 489         struct flow_cls_offload cls_flower = {};
 490
 491         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, NULL);
 492         cls_flower.command = FLOW_CLS_STATS;
 493         cls_flower.cookie = (unsigned long) f;
 494         cls_flower.classid = f->res.classid;
 495
 496         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false,
 497                          rtnl_held);
 498
 499         tcf_exts_stats_update(&f->exts, cls_flower.stats.bytes,
 500                               cls_flower.stats.pkts,
 501                               cls_flower.stats.drops,
 502                               cls_flower.stats.lastused,
 503                               cls_flower.stats.used_hw_stats,
 504                               cls_flower.stats.used_hw_stats_valid);
 505 }
 506
 507 static void __fl_put(struct cls_fl_filter *f)
 508 {
 509         if (!refcount_dec_and_test(&f->refcnt))
 510                 return;
 511
 512         if (tcf_exts_get_net(&f->exts))
 513                 tcf_queue_work(&f->rwork, fl_destroy_filter_work);
 514         else
 515                 __fl_destroy_filter(f);
 516 }
 517
 518 static struct cls_fl_filter *__fl_get(struct cls_fl_head *head, u32 handle)
 519 {
 520         struct cls_fl_filter *f;
 521
 522         rcu_read_lock();
 523         f = idr_find(&head->handle_idr, handle);
 524         if (f && !refcount_inc_not_zero(&f->refcnt))
 525                 f = NULL;
 526         rcu_read_unlock();
 527
 528         return f;
 529 }
 530
 531 static int __fl_delete(struct tcf_proto *tp, struct cls_fl_filter *f,
 532                        bool *last, bool rtnl_held,
 533                        struct netlink_ext_ack *extack)
 534 {
 535         struct cls_fl_head *head = fl_head_dereference(tp);
 536
 537         *last = false;
 538
 539         spin_lock(&tp->lock);
 540         if (f->deleted) {
 541                 spin_unlock(&tp->lock);
 542                 return -ENOENT;
 543         }
 544
 545         f->deleted = true;
 546         rhashtable_remove_fast(&f->mask->ht, &f->ht_node,
 547                                f->mask->filter_ht_params);
 548         idr_remove(&head->handle_idr, f->handle);
 549         list_del_rcu(&f->list);
 550         spin_unlock(&tp->lock);
 551
 552         *last = fl_mask_put(head, f->mask);
 553         if (!tc_skip_hw(f->flags))
 554                 fl_hw_destroy_filter(tp, f, rtnl_held, extack);
 555         tcf_unbind_filter(tp, &f->res);
 556         __fl_put(f);
 557
 558         return 0;
 559 }
 560
 561 static void fl_destroy_sleepable(struct work_struct *work)
 562 {
 563         struct cls_fl_head *head = container_of(to_rcu_work(work),
 564                                                 struct cls_fl_head,
 565                                                 rwork);
 566
 567         rhashtable_destroy(&head->ht);
 568         kfree(head);
 569         module_put(THIS_MODULE);
 570 }
 571
 572 static void fl_destroy(struct tcf_proto *tp, bool rtnl_held,
 573                        struct netlink_ext_ack *extack)
 574 {
 575         struct cls_fl_head *head = fl_head_dereference(tp);
 576         struct fl_flow_mask *mask, *next_mask;
 577         struct cls_fl_filter *f, *next;
 578         bool last;
 579
 580         list_for_each_entry_safe(mask, next_mask, &head->masks, list) {
 581                 list_for_each_entry_safe(f, next, &mask->filters, list) {
 582                         __fl_delete(tp, f, &last, rtnl_held, extack);
 583                         if (last)
 584                                 break;
 585                 }
 586         }
 587         idr_destroy(&head->handle_idr);
 588
 589         __module_get(THIS_MODULE);
 590         tcf_queue_work(&head->rwork, fl_destroy_sleepable);
 591 }
 592
 593 static void fl_put(struct tcf_proto *tp, void *arg)
 594 {
 595         struct cls_fl_filter *f = arg;
 596
 597         __fl_put(f);
 598 }
 599
 600 static void *fl_get(struct tcf_proto *tp, u32 handle)
 601 {
 602         struct cls_fl_head *head = fl_head_dereference(tp);
 603
 604         return __fl_get(head, handle);
 605 }
 606
 607 static const struct nla_policy fl_policy[TCA_FLOWER_MAX + 1] = {
 608         [TCA_FLOWER_UNSPEC]             = { .type = NLA_UNSPEC },
 609         [TCA_FLOWER_CLASSID]            = { .type = NLA_U32 },
 610         [TCA_FLOWER_INDEV]              = { .type = NLA_STRING,
 611                                             .len = IFNAMSIZ },
 612         [TCA_FLOWER_KEY_ETH_DST]        = { .len = ETH_ALEN },
 613         [TCA_FLOWER_KEY_ETH_DST_MASK]   = { .len = ETH_ALEN },
 614         [TCA_FLOWER_KEY_ETH_SRC]        = { .len = ETH_ALEN },
 615         [TCA_FLOWER_KEY_ETH_SRC_MASK]   = { .len = ETH_ALEN },
 616         [TCA_FLOWER_KEY_ETH_TYPE]       = { .type = NLA_U16 },
 617         [TCA_FLOWER_KEY_IP_PROTO]       = { .type = NLA_U8 },
 618         [TCA_FLOWER_KEY_IPV4_SRC]       = { .type = NLA_U32 },
 619         [TCA_FLOWER_KEY_IPV4_SRC_MASK]  = { .type = NLA_U32 },
 620         [TCA_FLOWER_KEY_IPV4_DST]       = { .type = NLA_U32 },
 621         [TCA_FLOWER_KEY_IPV4_DST_MASK]  = { .type = NLA_U32 },
 622         [TCA_FLOWER_KEY_IPV6_SRC]       = { .len = sizeof(struct in6_addr) },
 623         [TCA_FLOWER_KEY_IPV6_SRC_MASK]  = { .len = sizeof(struct in6_addr) },
 624         [TCA_FLOWER_KEY_IPV6_DST]       = { .len = sizeof(struct in6_addr) },
 625         [TCA_FLOWER_KEY_IPV6_DST_MASK]  = { .len = sizeof(struct in6_addr) },
 626         [TCA_FLOWER_KEY_TCP_SRC]        = { .type = NLA_U16 },
 627         [TCA_FLOWER_KEY_TCP_DST]        = { .type = NLA_U16 },
 628         [TCA_FLOWER_KEY_UDP_SRC]        = { .type = NLA_U16 },
 629         [TCA_FLOWER_KEY_UDP_DST]        = { .type = NLA_U16 },
 630         [TCA_FLOWER_KEY_VLAN_ID]        = { .type = NLA_U16 },
 631         [TCA_FLOWER_KEY_VLAN_PRIO]      = { .type = NLA_U8 },
 632         [TCA_FLOWER_KEY_VLAN_ETH_TYPE]  = { .type = NLA_U16 },
 633         [TCA_FLOWER_KEY_ENC_KEY_ID]     = { .type = NLA_U32 },
 634         [TCA_FLOWER_KEY_ENC_IPV4_SRC]   = { .type = NLA_U32 },
 635         [TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK] = { .type = NLA_U32 },
 636         [TCA_FLOWER_KEY_ENC_IPV4_DST]   = { .type = NLA_U32 },
 637         [TCA_FLOWER_KEY_ENC_IPV4_DST_MASK] = { .type = NLA_U32 },
 638         [TCA_FLOWER_KEY_ENC_IPV6_SRC]   = { .len = sizeof(struct in6_addr) },
 639         [TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK] = { .len = sizeof(struct in6_addr) },
 640         [TCA_FLOWER_KEY_ENC_IPV6_DST]   = { .len = sizeof(struct in6_addr) },
 641         [TCA_FLOWER_KEY_ENC_IPV6_DST_MASK] = { .len = sizeof(struct in6_addr) },
 642         [TCA_FLOWER_KEY_TCP_SRC_MASK]   = { .type = NLA_U16 },
 643         [TCA_FLOWER_KEY_TCP_DST_MASK]   = { .type = NLA_U16 },
 644         [TCA_FLOWER_KEY_UDP_SRC_MASK]   = { .type = NLA_U16 },
 645         [TCA_FLOWER_KEY_UDP_DST_MASK]   = { .type = NLA_U16 },
 646         [TCA_FLOWER_KEY_SCTP_SRC_MASK]  = { .type = NLA_U16 },
 647         [TCA_FLOWER_KEY_SCTP_DST_MASK]  = { .type = NLA_U16 },
 648         [TCA_FLOWER_KEY_SCTP_SRC]       = { .type = NLA_U16 },
 649         [TCA_FLOWER_KEY_SCTP_DST]       = { .type = NLA_U16 },
 650         [TCA_FLOWER_KEY_ENC_UDP_SRC_PORT]       = { .type = NLA_U16 },
 651         [TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK]  = { .type = NLA_U16 },
 652         [TCA_FLOWER_KEY_ENC_UDP_DST_PORT]       = { .type = NLA_U16 },
 653         [TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK]  = { .type = NLA_U16 },
 654         [TCA_FLOWER_KEY_FLAGS]          = { .type = NLA_U32 },
 655         [TCA_FLOWER_KEY_FLAGS_MASK]     = { .type = NLA_U32 },
 656         [TCA_FLOWER_KEY_ICMPV4_TYPE]    = { .type = NLA_U8 },
 657         [TCA_FLOWER_KEY_ICMPV4_TYPE_MASK] = { .type = NLA_U8 },
 658         [TCA_FLOWER_KEY_ICMPV4_CODE]    = { .type = NLA_U8 },
 659         [TCA_FLOWER_KEY_ICMPV4_CODE_MASK] = { .type = NLA_U8 },
 660         [TCA_FLOWER_KEY_ICMPV6_TYPE]    = { .type = NLA_U8 },
 661         [TCA_FLOWER_KEY_ICMPV6_TYPE_MASK] = { .type = NLA_U8 },
 662         [TCA_FLOWER_KEY_ICMPV6_CODE]    = { .type = NLA_U8 },
 663         [TCA_FLOWER_KEY_ICMPV6_CODE_MASK] = { .type = NLA_U8 },
 664         [TCA_FLOWER_KEY_ARP_SIP]        = { .type = NLA_U32 },
 665         [TCA_FLOWER_KEY_ARP_SIP_MASK]   = { .type = NLA_U32 },
 666         [TCA_FLOWER_KEY_ARP_TIP]        = { .type = NLA_U32 },
 667         [TCA_FLOWER_KEY_ARP_TIP_MASK]   = { .type = NLA_U32 },
 668         [TCA_FLOWER_KEY_ARP_OP]         = { .type = NLA_U8 },
 669         [TCA_FLOWER_KEY_ARP_OP_MASK]    = { .type = NLA_U8 },
 670         [TCA_FLOWER_KEY_ARP_SHA]        = { .len = ETH_ALEN },
 671         [TCA_FLOWER_KEY_ARP_SHA_MASK]   = { .len = ETH_ALEN },
 672         [TCA_FLOWER_KEY_ARP_THA]        = { .len = ETH_ALEN },
 673         [TCA_FLOWER_KEY_ARP_THA_MASK]   = { .len = ETH_ALEN },
 674         [TCA_FLOWER_KEY_MPLS_TTL]       = { .type = NLA_U8 },
 675         [TCA_FLOWER_KEY_MPLS_BOS]       = { .type = NLA_U8 },
 676         [TCA_FLOWER_KEY_MPLS_TC]        = { .type = NLA_U8 },
 677         [TCA_FLOWER_KEY_MPLS_LABEL]     = { .type = NLA_U32 },
 678         [TCA_FLOWER_KEY_MPLS_OPTS]      = { .type = NLA_NESTED },
 679         [TCA_FLOWER_KEY_TCP_FLAGS]      = { .type = NLA_U16 },
 680         [TCA_FLOWER_KEY_TCP_FLAGS_MASK] = { .type = NLA_U16 },
 681         [TCA_FLOWER_KEY_IP_TOS]         = { .type = NLA_U8 },
 682         [TCA_FLOWER_KEY_IP_TOS_MASK]    = { .type = NLA_U8 },
 683         [TCA_FLOWER_KEY_IP_TTL]         = { .type = NLA_U8 },
 684         [TCA_FLOWER_KEY_IP_TTL_MASK]    = { .type = NLA_U8 },
 685         [TCA_FLOWER_KEY_CVLAN_ID]       = { .type = NLA_U16 },
 686         [TCA_FLOWER_KEY_CVLAN_PRIO]     = { .type = NLA_U8 },
 687         [TCA_FLOWER_KEY_CVLAN_ETH_TYPE] = { .type = NLA_U16 },
 688         [TCA_FLOWER_KEY_ENC_IP_TOS]     = { .type = NLA_U8 },
 689         [TCA_FLOWER_KEY_ENC_IP_TOS_MASK] = { .type = NLA_U8 },
 690         [TCA_FLOWER_KEY_ENC_IP_TTL]      = { .type = NLA_U8 },
 691         [TCA_FLOWER_KEY_ENC_IP_TTL_MASK] = { .type = NLA_U8 },
 692         [TCA_FLOWER_KEY_ENC_OPTS]       = { .type = NLA_NESTED },
 693         [TCA_FLOWER_KEY_ENC_OPTS_MASK]  = { .type = NLA_NESTED },
 694         [TCA_FLOWER_KEY_CT_STATE]       =
 695                 NLA_POLICY_MASK(NLA_U16, TCA_FLOWER_KEY_CT_FLAGS_MASK),
 696         [TCA_FLOWER_KEY_CT_STATE_MASK]  =
 697                 NLA_POLICY_MASK(NLA_U16, TCA_FLOWER_KEY_CT_FLAGS_MASK),
 698         [TCA_FLOWER_KEY_CT_ZONE]        = { .type = NLA_U16 },
 699         [TCA_FLOWER_KEY_CT_ZONE_MASK]   = { .type = NLA_U16 },
 700         [TCA_FLOWER_KEY_CT_MARK]        = { .type = NLA_U32 },
 701         [TCA_FLOWER_KEY_CT_MARK_MASK]   = { .type = NLA_U32 },
 702         [TCA_FLOWER_KEY_CT_LABELS]      = { .type = NLA_BINARY,
 703                                             .len = 128 / BITS_PER_BYTE },
 704         [TCA_FLOWER_KEY_CT_LABELS_MASK] = { .type = NLA_BINARY,
 705                                             .len = 128 / BITS_PER_BYTE },
 706         [TCA_FLOWER_FLAGS]              = { .type = NLA_U32 },
 707         [TCA_FLOWER_KEY_HASH]           = { .type = NLA_U32 },
 708         [TCA_FLOWER_KEY_HASH_MASK]      = { .type = NLA_U32 },
 709
 710 };
 711
 712 static const struct nla_policy
 713 enc_opts_policy[TCA_FLOWER_KEY_ENC_OPTS_MAX + 1] = {
 714         [TCA_FLOWER_KEY_ENC_OPTS_UNSPEC]        = {
 715                 .strict_start_type = TCA_FLOWER_KEY_ENC_OPTS_VXLAN },
 716         [TCA_FLOWER_KEY_ENC_OPTS_GENEVE]        = { .type = NLA_NESTED },
 717         [TCA_FLOWER_KEY_ENC_OPTS_VXLAN]         = { .type = NLA_NESTED },
 718         [TCA_FLOWER_KEY_ENC_OPTS_ERSPAN]        = { .type = NLA_NESTED },
 719 };
 720
 721 static const struct nla_policy
 722 geneve_opt_policy[TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX + 1] = {
 723         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS]      = { .type = NLA_U16 },
 724         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE]       = { .type = NLA_U8 },
 725         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA]       = { .type = NLA_BINARY,
 726                                                        .len = 128 },
 727 };
 728
 729 static const struct nla_policy
 730 vxlan_opt_policy[TCA_FLOWER_KEY_ENC_OPT_VXLAN_MAX + 1] = {
 731         [TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]         = { .type = NLA_U32 },
 732 };
 733
 734 static const struct nla_policy
 735 erspan_opt_policy[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_MAX + 1] = {
 736         [TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER]        = { .type = NLA_U8 },
 737         [TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX]      = { .type = NLA_U32 },
 738         [TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR]        = { .type = NLA_U8 },
 739         [TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID]       = { .type = NLA_U8 },
 740 };
 741
 742 static const struct nla_policy
 743 mpls_stack_entry_policy[TCA_FLOWER_KEY_MPLS_OPT_LSE_MAX + 1] = {
 744         [TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH]    = { .type = NLA_U8 },
 745         [TCA_FLOWER_KEY_MPLS_OPT_LSE_TTL]      = { .type = NLA_U8 },
 746         [TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS]      = { .type = NLA_U8 },
 747         [TCA_FLOWER_KEY_MPLS_OPT_LSE_TC]       = { .type = NLA_U8 },
 748         [TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL]    = { .type = NLA_U32 },
 749 };
 750
 751 static void fl_set_key_val(struct nlattr **tb,
 752                            void *val, int val_type,
 753                            void *mask, int mask_type, int len)
 754 {
 755         if (!tb[val_type])
 756                 return;
 757         nla_memcpy(val, tb[val_type], len);
 758         if (mask_type == TCA_FLOWER_UNSPEC || !tb[mask_type])
 759                 memset(mask, 0xff, len);
 760         else
 761                 nla_memcpy(mask, tb[mask_type], len);
 762 }
 763
 764 static int fl_set_key_port_range(struct nlattr **tb, struct fl_flow_key *key,
 765                                  struct fl_flow_key *mask,
 766                                  struct netlink_ext_ack *extack)
 767 {
 768         fl_set_key_val(tb, &key->tp_range.tp_min.dst,
 769                        TCA_FLOWER_KEY_PORT_DST_MIN, &mask->tp_range.tp_min.dst,
 770                        TCA_FLOWER_UNSPEC, sizeof(key->tp_range.tp_min.dst));
 771         fl_set_key_val(tb, &key->tp_range.tp_max.dst,
 772                        TCA_FLOWER_KEY_PORT_DST_MAX, &mask->tp_range.tp_max.dst,
 773                        TCA_FLOWER_UNSPEC, sizeof(key->tp_range.tp_max.dst));
 774         fl_set_key_val(tb, &key->tp_range.tp_min.src,
 775                        TCA_FLOWER_KEY_PORT_SRC_MIN, &mask->tp_range.tp_min.src,
 776                        TCA_FLOWER_UNSPEC, sizeof(key->tp_range.tp_min.src));
 777         fl_set_key_val(tb, &key->tp_range.tp_max.src,
 778                        TCA_FLOWER_KEY_PORT_SRC_MAX, &mask->tp_range.tp_max.src,
 779                        TCA_FLOWER_UNSPEC, sizeof(key->tp_range.tp_max.src));
 780
 781         if (mask->tp_range.tp_min.dst != mask->tp_range.tp_max.dst) {
 782                 NL_SET_ERR_MSG(extack,
 783                                "Both min and max destination ports must be specified");
 784                 return -EINVAL;
 785         }
 786         if (mask->tp_range.tp_min.src != mask->tp_range.tp_max.src) {
 787                 NL_SET_ERR_MSG(extack,
 788                                "Both min and max source ports must be specified");
 789                 return -EINVAL;
 790         }
 791         if (mask->tp_range.tp_min.dst && mask->tp_range.tp_max.dst &&
 792             ntohs(key->tp_range.tp_max.dst) <=
 793             ntohs(key->tp_range.tp_min.dst)) {
 794                 NL_SET_ERR_MSG_ATTR(extack,
 795                                     tb[TCA_FLOWER_KEY_PORT_DST_MIN],
 796                                     "Invalid destination port range (min must be strictly smaller than max)");
 797                 return -EINVAL;
 798         }
 799         if (mask->tp_range.tp_min.src && mask->tp_range.tp_max.src &&
 800             ntohs(key->tp_range.tp_max.src) <=
 801             ntohs(key->tp_range.tp_min.src)) {
 802                 NL_SET_ERR_MSG_ATTR(extack,
 803                                     tb[TCA_FLOWER_KEY_PORT_SRC_MIN],
 804                                     "Invalid source port range (min must be strictly smaller than max)");
 805                 return -EINVAL;
 806         }
 807
 808         return 0;
 809 }
 810
 811 static int fl_set_key_mpls_lse(const struct nlattr *nla_lse,
 812                                struct flow_dissector_key_mpls *key_val,
 813                                struct flow_dissector_key_mpls *key_mask,
 814                                struct netlink_ext_ack *extack)
 815 {
 816         struct nlattr *tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_MAX + 1];
 817         struct flow_dissector_mpls_lse *lse_mask;
 818         struct flow_dissector_mpls_lse *lse_val;
 819         u8 lse_index;
 820         u8 depth;
 821         int err;
 822
 823         err = nla_parse_nested(tb, TCA_FLOWER_KEY_MPLS_OPT_LSE_MAX, nla_lse,
 824                                mpls_stack_entry_policy, extack);
 825         if (err < 0)
 826                 return err;
 827
 828         if (!tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH]) {
 829                 NL_SET_ERR_MSG(extack, "Missing MPLS option \"depth\"");
 830                 return -EINVAL;
 831         }
 832
 833         depth = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH]);
 834
 835         /* LSE depth starts at 1, for consistency with terminology used by
 836          * RFC 3031 (section 3.9), where depth 0 refers to unlabeled packets.
 837          */
 838         if (depth < 1 || depth > FLOW_DIS_MPLS_MAX) {
 839                 NL_SET_ERR_MSG_ATTR(extack,
 840                                     tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH],
 841                                     "Invalid MPLS depth");
 842                 return -EINVAL;
 843         }
 844         lse_index = depth - 1;
 845
 846         dissector_set_mpls_lse(key_val, lse_index);
 847         dissector_set_mpls_lse(key_mask, lse_index);
 848
 849         lse_val = &key_val->ls[lse_index];
 850         lse_mask = &key_mask->ls[lse_index];
 851
 852         if (tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_TTL]) {
 853                 lse_val->mpls_ttl = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_TTL]);
 854                 lse_mask->mpls_ttl = MPLS_TTL_MASK;
 855         }
 856         if (tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS]) {
 857                 u8 bos = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS]);
 858
 859                 if (bos & ~MPLS_BOS_MASK) {
 860                         NL_SET_ERR_MSG_ATTR(extack,
 861                                             tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS],
 862                                             "Bottom Of Stack (BOS) must be 0 or 1");
 863                         return -EINVAL;
 864                 }
 865                 lse_val->mpls_bos = bos;
 866                 lse_mask->mpls_bos = MPLS_BOS_MASK;
 867         }
 868         if (tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_TC]) {
 869                 u8 tc = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_TC]);
 870
 871                 if (tc & ~MPLS_TC_MASK) {
 872                         NL_SET_ERR_MSG_ATTR(extack,
 873                                             tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_TC],
 874                                             "Traffic Class (TC) must be between 0 and 7");
 875                         return -EINVAL;
 876                 }
 877                 lse_val->mpls_tc = tc;
 878                 lse_mask->mpls_tc = MPLS_TC_MASK;
 879         }
 880         if (tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL]) {
 881                 u32 label = nla_get_u32(tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL]);
 882
 883                 if (label & ~MPLS_LABEL_MASK) {
 884                         NL_SET_ERR_MSG_ATTR(extack,
 885                                             tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL],
 886                                             "Label must be between 0 and 1048575");
 887                         return -EINVAL;
 888                 }
 889                 lse_val->mpls_label = label;
 890                 lse_mask->mpls_label = MPLS_LABEL_MASK;
 891         }
 892
 893         return 0;
 894 }
 895
 896 static int fl_set_key_mpls_opts(const struct nlattr *nla_mpls_opts,
 897                                 struct flow_dissector_key_mpls *key_val,
 898                                 struct flow_dissector_key_mpls *key_mask,
 899                                 struct netlink_ext_ack *extack)
 900 {
 901         struct nlattr *nla_lse;
 902         int rem;
 903         int err;
 904
 905         if (!(nla_mpls_opts->nla_type & NLA_F_NESTED)) {
 906                 NL_SET_ERR_MSG_ATTR(extack, nla_mpls_opts,
 907                                     "NLA_F_NESTED is missing");
 908                 return -EINVAL;
 909         }
 910
 911         nla_for_each_nested(nla_lse, nla_mpls_opts, rem) {
 912                 if (nla_type(nla_lse) != TCA_FLOWER_KEY_MPLS_OPTS_LSE) {
 913                         NL_SET_ERR_MSG_ATTR(extack, nla_lse,
 914                                             "Invalid MPLS option type");
 915                         return -EINVAL;
 916                 }
 917
 918                 err = fl_set_key_mpls_lse(nla_lse, key_val, key_mask, extack);
 919                 if (err < 0)
 920                         return err;
 921         }
 922         if (rem) {
 923                 NL_SET_ERR_MSG(extack,
 924                                "Bytes leftover after parsing MPLS options");
 925                 return -EINVAL;
 926         }
 927
 928         return 0;
 929 }
 930
 931 static int fl_set_key_mpls(struct nlattr **tb,
 932                            struct flow_dissector_key_mpls *key_val,
 933                            struct flow_dissector_key_mpls *key_mask,
 934                            struct netlink_ext_ack *extack)
 935 {
 936         struct flow_dissector_mpls_lse *lse_mask;
 937         struct flow_dissector_mpls_lse *lse_val;
 938
 939         if (tb[TCA_FLOWER_KEY_MPLS_OPTS]) {
 940                 if (tb[TCA_FLOWER_KEY_MPLS_TTL] ||
 941                     tb[TCA_FLOWER_KEY_MPLS_BOS] ||
 942                     tb[TCA_FLOWER_KEY_MPLS_TC] ||
 943                     tb[TCA_FLOWER_KEY_MPLS_LABEL]) {
 944                         NL_SET_ERR_MSG_ATTR(extack,
 945                                             tb[TCA_FLOWER_KEY_MPLS_OPTS],
 946                                             "MPLS label, Traffic Class, Bottom Of Stack and Time To Live must be encapsulated in the MPLS options attribute");
 947                         return -EBADMSG;
 948                 }
 949
 950                 return fl_set_key_mpls_opts(tb[TCA_FLOWER_KEY_MPLS_OPTS],
 951                                             key_val, key_mask, extack);
 952         }
 953
 954         lse_val = &key_val->ls[0];
 955         lse_mask = &key_mask->ls[0];
 956
 957         if (tb[TCA_FLOWER_KEY_MPLS_TTL]) {
 958                 lse_val->mpls_ttl = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_TTL]);
 959                 lse_mask->mpls_ttl = MPLS_TTL_MASK;
 960                 dissector_set_mpls_lse(key_val, 0);
 961                 dissector_set_mpls_lse(key_mask, 0);
 962         }
 963         if (tb[TCA_FLOWER_KEY_MPLS_BOS]) {
 964                 u8 bos = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_BOS]);
 965
 966                 if (bos & ~MPLS_BOS_MASK) {
 967                         NL_SET_ERR_MSG_ATTR(extack,
 968                                             tb[TCA_FLOWER_KEY_MPLS_BOS],
 969                                             "Bottom Of Stack (BOS) must be 0 or 1");
 970                         return -EINVAL;
 971                 }
 972                 lse_val->mpls_bos = bos;
 973                 lse_mask->mpls_bos = MPLS_BOS_MASK;
 974                 dissector_set_mpls_lse(key_val, 0);
 975                 dissector_set_mpls_lse(key_mask, 0);
 976         }
 977         if (tb[TCA_FLOWER_KEY_MPLS_TC]) {
 978                 u8 tc = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_TC]);
 979
 980                 if (tc & ~MPLS_TC_MASK) {
 981                         NL_SET_ERR_MSG_ATTR(extack,
 982                                             tb[TCA_FLOWER_KEY_MPLS_TC],
 983                                             "Traffic Class (TC) must be between 0 and 7");
 984                         return -EINVAL;
 985                 }
 986                 lse_val->mpls_tc = tc;
 987                 lse_mask->mpls_tc = MPLS_TC_MASK;
 988                 dissector_set_mpls_lse(key_val, 0);
 989                 dissector_set_mpls_lse(key_mask, 0);
 990         }
 991         if (tb[TCA_FLOWER_KEY_MPLS_LABEL]) {
 992                 u32 label = nla_get_u32(tb[TCA_FLOWER_KEY_MPLS_LABEL]);
 993
 994                 if (label & ~MPLS_LABEL_MASK) {
 995                         NL_SET_ERR_MSG_ATTR(extack,
 996                                             tb[TCA_FLOWER_KEY_MPLS_LABEL],
 997                                             "Label must be between 0 and 1048575");
 998                         return -EINVAL;
 999                 }
1000                 lse_val->mpls_label = label;
1001                 lse_mask->mpls_label = MPLS_LABEL_MASK;
1002                 dissector_set_mpls_lse(key_val, 0);
1003                 dissector_set_mpls_lse(key_mask, 0);
1004         }
1005         return 0;
1006 }
1007
1008 static void fl_set_key_vlan(struct nlattr **tb,
1009                             __be16 ethertype,
1010                             int vlan_id_key, int vlan_prio_key,
1011                             int vlan_next_eth_type_key,
1012                             struct flow_dissector_key_vlan *key_val,
1013                             struct flow_dissector_key_vlan *key_mask)
1014 {
1015 #define VLAN_PRIORITY_MASK      0x7
1016
1017         if (tb[vlan_id_key]) {
1018                 key_val->vlan_id =
1019                         nla_get_u16(tb[vlan_id_key]) & VLAN_VID_MASK;
1020                 key_mask->vlan_id = VLAN_VID_MASK;
1021         }
1022         if (tb[vlan_prio_key]) {
1023                 key_val->vlan_priority =
1024                         nla_get_u8(tb[vlan_prio_key]) &
1025                         VLAN_PRIORITY_MASK;
1026                 key_mask->vlan_priority = VLAN_PRIORITY_MASK;
1027         }
1028         key_val->vlan_tpid = ethertype;
1029         key_mask->vlan_tpid = cpu_to_be16(~0);
1030         if (tb[vlan_next_eth_type_key]) {
1031                 key_val->vlan_eth_type =
1032                         nla_get_be16(tb[vlan_next_eth_type_key]);
1033                 key_mask->vlan_eth_type = cpu_to_be16(~0);
1034         }
1035 }
1036
1037 static void fl_set_key_flag(u32 flower_key, u32 flower_mask,
1038                             u32 *dissector_key, u32 *dissector_mask,
1039                             u32 flower_flag_bit, u32 dissector_flag_bit)
1040 {
1041         if (flower_mask & flower_flag_bit) {
1042                 *dissector_mask |= dissector_flag_bit;
1043                 if (flower_key & flower_flag_bit)
1044                         *dissector_key |= dissector_flag_bit;
1045         }
1046 }
1047
1048 static int fl_set_key_flags(struct nlattr **tb, u32 *flags_key,
1049                             u32 *flags_mask, struct netlink_ext_ack *extack)
1050 {
1051         u32 key, mask;
1052
1053         /* mask is mandatory for flags */
1054         if (!tb[TCA_FLOWER_KEY_FLAGS_MASK]) {
1055                 NL_SET_ERR_MSG(extack, "Missing flags mask");
1056                 return -EINVAL;
1057         }
1058
1059         key = be32_to_cpu(nla_get_u32(tb[TCA_FLOWER_KEY_FLAGS]));
1060         mask = be32_to_cpu(nla_get_u32(tb[TCA_FLOWER_KEY_FLAGS_MASK]));
1061
1062         *flags_key  = 0;
1063         *flags_mask = 0;
1064
1065         fl_set_key_flag(key, mask, flags_key, flags_mask,
1066                         TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT, FLOW_DIS_IS_FRAGMENT);
1067         fl_set_key_flag(key, mask, flags_key, flags_mask,
1068                         TCA_FLOWER_KEY_FLAGS_FRAG_IS_FIRST,
1069                         FLOW_DIS_FIRST_FRAG);
1070
1071         return 0;
1072 }
1073
1074 static void fl_set_key_ip(struct nlattr **tb, bool encap,
1075                           struct flow_dissector_key_ip *key,
1076                           struct flow_dissector_key_ip *mask)
1077 {
1078         int tos_key = encap ? TCA_FLOWER_KEY_ENC_IP_TOS : TCA_FLOWER_KEY_IP_TOS;
1079         int ttl_key = encap ? TCA_FLOWER_KEY_ENC_IP_TTL : TCA_FLOWER_KEY_IP_TTL;
1080         int tos_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TOS_MASK : TCA_FLOWER_KEY_IP_TOS_MASK;
1081         int ttl_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TTL_MASK : TCA_FLOWER_KEY_IP_TTL_MASK;
1082
1083         fl_set_key_val(tb, &key->tos, tos_key, &mask->tos, tos_mask, sizeof(key->tos));
1084         fl_set_key_val(tb, &key->ttl, ttl_key, &mask->ttl, ttl_mask, sizeof(key->ttl));
1085 }
1086
1087 static int fl_set_geneve_opt(const struct nlattr *nla, struct fl_flow_key *key,
1088                              int depth, int option_len,
1089                              struct netlink_ext_ack *extack)
1090 {
1091         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX + 1];
1092         struct nlattr *class = NULL, *type = NULL, *data = NULL;
1093         struct geneve_opt *opt;
1094         int err, data_len = 0;
1095
1096         if (option_len > sizeof(struct geneve_opt))
1097                 data_len = option_len - sizeof(struct geneve_opt);
1098
1099         if (key->enc_opts.len > FLOW_DIS_TUN_OPTS_MAX - 4)
1100                 return -ERANGE;
1101
1102         opt = (struct geneve_opt *)&key->enc_opts.data[key->enc_opts.len];
1103         memset(opt, 0xff, option_len);
1104         opt->length = data_len / 4;
1105         opt->r1 = 0;
1106         opt->r2 = 0;
1107         opt->r3 = 0;
1108
1109         /* If no mask has been prodived we assume an exact match. */
1110         if (!depth)
1111                 return sizeof(struct geneve_opt) + data_len;
1112
1113         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_GENEVE) {
1114                 NL_SET_ERR_MSG(extack, "Non-geneve option type for mask");
1115                 return -EINVAL;
1116         }
1117
1118         err = nla_parse_nested_deprecated(tb,
1119                                           TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX,
1120                                           nla, geneve_opt_policy, extack);
1121         if (err < 0)
1122                 return err;
1123
1124         /* We are not allowed to omit any of CLASS, TYPE or DATA
1125          * fields from the key.
1126          */
1127         if (!option_len &&
1128             (!tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS] ||
1129              !tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE] ||
1130              !tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA])) {
1131                 NL_SET_ERR_MSG(extack, "Missing tunnel key geneve option class, type or data");
1132                 return -EINVAL;
1133         }
1134
1135         /* Omitting any of CLASS, TYPE or DATA fields is allowed
1136          * for the mask.
1137          */
1138         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA]) {
1139                 int new_len = key->enc_opts.len;
1140
1141                 data = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA];
1142                 data_len = nla_len(data);
1143                 if (data_len < 4) {
1144                         NL_SET_ERR_MSG(extack, "Tunnel key geneve option data is less than 4 bytes long");
1145                         return -ERANGE;
1146                 }
1147                 if (data_len % 4) {
1148                         NL_SET_ERR_MSG(extack, "Tunnel key geneve option data is not a multiple of 4 bytes long");
1149                         return -ERANGE;
1150                 }
1151
1152                 new_len += sizeof(struct geneve_opt) + data_len;
1153                 BUILD_BUG_ON(FLOW_DIS_TUN_OPTS_MAX != IP_TUNNEL_OPTS_MAX);
1154                 if (new_len > FLOW_DIS_TUN_OPTS_MAX) {
1155                         NL_SET_ERR_MSG(extack, "Tunnel options exceeds max size");
1156                         return -ERANGE;
1157                 }
1158                 opt->length = data_len / 4;
1159                 memcpy(opt->opt_data, nla_data(data), data_len);
1160         }
1161
1162         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS]) {
1163                 class = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS];
1164                 opt->opt_class = nla_get_be16(class);
1165         }
1166
1167         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE]) {
1168                 type = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE];
1169                 opt->type = nla_get_u8(type);
1170         }
1171
1172         return sizeof(struct geneve_opt) + data_len;
1173 }
1174
1175 static int fl_set_vxlan_opt(const struct nlattr *nla, struct fl_flow_key *key,
1176                             int depth, int option_len,
1177                             struct netlink_ext_ack *extack)
1178 {
1179         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_MAX + 1];
1180         struct vxlan_metadata *md;
1181         int err;
1182
1183         md = (struct vxlan_metadata *)&key->enc_opts.data[key->enc_opts.len];
1184         memset(md, 0xff, sizeof(*md));
1185
1186         if (!depth)
1187                 return sizeof(*md);
1188
1189         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_VXLAN) {
1190                 NL_SET_ERR_MSG(extack, "Non-vxlan option type for mask");
1191                 return -EINVAL;
1192         }
1193
1194         err = nla_parse_nested(tb, TCA_FLOWER_KEY_ENC_OPT_VXLAN_MAX, nla,
1195                                vxlan_opt_policy, extack);
1196         if (err < 0)
1197                 return err;
1198
1199         if (!option_len && !tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]) {
1200                 NL_SET_ERR_MSG(extack, "Missing tunnel key vxlan option gbp");
1201                 return -EINVAL;
1202         }
1203
1204         if (tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]) {
1205                 md->gbp = nla_get_u32(tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]);
1206                 md->gbp &= VXLAN_GBP_MASK;
1207         }
1208
1209         return sizeof(*md);
1210 }
1211
1212 static int fl_set_erspan_opt(const struct nlattr *nla, struct fl_flow_key *key,
1213                              int depth, int option_len,
1214                              struct netlink_ext_ack *extack)
1215 {
1216         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_MAX + 1];
1217         struct erspan_metadata *md;
1218         int err;
1219
1220         md = (struct erspan_metadata *)&key->enc_opts.data[key->enc_opts.len];
1221         memset(md, 0xff, sizeof(*md));
1222         md->version = 1;
1223
1224         if (!depth)
1225                 return sizeof(*md);
1226
1227         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_ERSPAN) {
1228                 NL_SET_ERR_MSG(extack, "Non-erspan option type for mask");
1229                 return -EINVAL;
1230         }
1231
1232         err = nla_parse_nested(tb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_MAX, nla,
1233                                erspan_opt_policy, extack);
1234         if (err < 0)
1235                 return err;
1236
1237         if (!option_len && !tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER]) {
1238                 NL_SET_ERR_MSG(extack, "Missing tunnel key erspan option ver");
1239                 return -EINVAL;
1240         }
1241
1242         if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER])
1243                 md->version = nla_get_u8(tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER]);
1244
1245         if (md->version == 1) {
1246                 if (!option_len && !tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX]) {
1247                         NL_SET_ERR_MSG(extack, "Missing tunnel key erspan option index");
1248                         return -EINVAL;
1249                 }
1250                 if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX]) {
1251                         nla = tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX];
1252                         memset(&md->u, 0x00, sizeof(md->u));
1253                         md->u.index = nla_get_be32(nla);
1254                 }
1255         } else if (md->version == 2) {
1256                 if (!option_len && (!tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR] ||
1257                                     !tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID])) {
1258                         NL_SET_ERR_MSG(extack, "Missing tunnel key erspan option dir or hwid");
1259                         return -EINVAL;
1260                 }
1261                 if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR]) {
1262                         nla = tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR];
1263                         md->u.md2.dir = nla_get_u8(nla);
1264                 }
1265                 if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID]) {
1266                         nla = tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID];
1267                         set_hwid(&md->u.md2, nla_get_u8(nla));
1268                 }
1269         } else {
1270                 NL_SET_ERR_MSG(extack, "Tunnel key erspan option ver is incorrect");
1271                 return -EINVAL;
1272         }
1273
1274         return sizeof(*md);
1275 }
1276
1277 static int fl_set_enc_opt(struct nlattr **tb, struct fl_flow_key *key,
1278                           struct fl_flow_key *mask,
1279                           struct netlink_ext_ack *extack)
1280 {
1281         const struct nlattr *nla_enc_key, *nla_opt_key, *nla_opt_msk = NULL;
1282         int err, option_len, key_depth, msk_depth = 0;
1283
1284         err = nla_validate_nested_deprecated(tb[TCA_FLOWER_KEY_ENC_OPTS],
1285                                              TCA_FLOWER_KEY_ENC_OPTS_MAX,
1286                                              enc_opts_policy, extack);
1287         if (err)
1288                 return err;
1289
1290         nla_enc_key = nla_data(tb[TCA_FLOWER_KEY_ENC_OPTS]);
1291
1292         if (tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]) {
1293                 err = nla_validate_nested_deprecated(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK],
1294                                                      TCA_FLOWER_KEY_ENC_OPTS_MAX,
1295                                                      enc_opts_policy, extack);
1296                 if (err)
1297                         return err;
1298
1299                 nla_opt_msk = nla_data(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]);
1300                 msk_depth = nla_len(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]);
1301                 if (!nla_ok(nla_opt_msk, msk_depth)) {
1302                         NL_SET_ERR_MSG(extack, "Invalid nested attribute for masks");
1303                         return -EINVAL;
1304                 }
1305         }
1306
1307         nla_for_each_attr(nla_opt_key, nla_enc_key,
1308                           nla_len(tb[TCA_FLOWER_KEY_ENC_OPTS]), key_depth) {
1309                 switch (nla_type(nla_opt_key)) {
1310                 case TCA_FLOWER_KEY_ENC_OPTS_GENEVE:
1311                         if (key->enc_opts.dst_opt_type &&
1312                             key->enc_opts.dst_opt_type != TUNNEL_GENEVE_OPT) {
1313                                 NL_SET_ERR_MSG(extack, "Duplicate type for geneve options");
1314                                 return -EINVAL;
1315                         }
1316                         option_len = 0;
1317                         key->enc_opts.dst_opt_type = TUNNEL_GENEVE_OPT;
1318                         option_len = fl_set_geneve_opt(nla_opt_key, key,
1319                                                        key_depth, option_len,
1320                                                        extack);
1321                         if (option_len < 0)
1322                                 return option_len;
1323
1324                         key->enc_opts.len += option_len;
1325                         /* At the same time we need to parse through the mask
1326                          * in order to verify exact and mask attribute lengths.
1327                          */
1328                         mask->enc_opts.dst_opt_type = TUNNEL_GENEVE_OPT;
1329                         option_len = fl_set_geneve_opt(nla_opt_msk, mask,
1330                                                        msk_depth, option_len,
1331                                                        extack);
1332                         if (option_len < 0)
1333                                 return option_len;
1334
1335                         mask->enc_opts.len += option_len;
1336                         if (key->enc_opts.len != mask->enc_opts.len) {
1337                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
1338                                 return -EINVAL;
1339                         }
1340                         break;
1341                 case TCA_FLOWER_KEY_ENC_OPTS_VXLAN:
1342                         if (key->enc_opts.dst_opt_type) {
1343                                 NL_SET_ERR_MSG(extack, "Duplicate type for vxlan options");
1344                                 return -EINVAL;
1345                         }
1346                         option_len = 0;
1347                         key->enc_opts.dst_opt_type = TUNNEL_VXLAN_OPT;
1348                         option_len = fl_set_vxlan_opt(nla_opt_key, key,
1349                                                       key_depth, option_len,
1350                                                       extack);
1351                         if (option_len < 0)
1352                                 return option_len;
1353
1354                         key->enc_opts.len += option_len;
1355                         /* At the same time we need to parse through the mask
1356                          * in order to verify exact and mask attribute lengths.
1357                          */
1358                         mask->enc_opts.dst_opt_type = TUNNEL_VXLAN_OPT;
1359                         option_len = fl_set_vxlan_opt(nla_opt_msk, mask,
1360                                                       msk_depth, option_len,
1361                                                       extack);
1362                         if (option_len < 0)
1363                                 return option_len;
1364
1365                         mask->enc_opts.len += option_len;
1366                         if (key->enc_opts.len != mask->enc_opts.len) {
1367                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
1368                                 return -EINVAL;
1369                         }
1370                         break;
1371                 case TCA_FLOWER_KEY_ENC_OPTS_ERSPAN:
1372                         if (key->enc_opts.dst_opt_type) {
1373                                 NL_SET_ERR_MSG(extack, "Duplicate type for erspan options");
1374                                 return -EINVAL;
1375                         }
1376                         option_len = 0;
1377                         key->enc_opts.dst_opt_type = TUNNEL_ERSPAN_OPT;
1378                         option_len = fl_set_erspan_opt(nla_opt_key, key,
1379                                                        key_depth, option_len,
1380                                                        extack);
1381                         if (option_len < 0)
1382                                 return option_len;
1383
1384                         key->enc_opts.len += option_len;
1385                         /* At the same time we need to parse through the mask
1386                          * in order to verify exact and mask attribute lengths.
1387                          */
1388                         mask->enc_opts.dst_opt_type = TUNNEL_ERSPAN_OPT;
1389                         option_len = fl_set_erspan_opt(nla_opt_msk, mask,
1390                                                        msk_depth, option_len,
1391                                                        extack);
1392                         if (option_len < 0)
1393                                 return option_len;
1394
1395                         mask->enc_opts.len += option_len;
1396                         if (key->enc_opts.len != mask->enc_opts.len) {
1397                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
1398                                 return -EINVAL;
1399                         }
1400                         break;
1401                 default:
1402                         NL_SET_ERR_MSG(extack, "Unknown tunnel option type");
1403                         return -EINVAL;
1404                 }
1405
1406                 if (!msk_depth)
1407                         continue;
1408
1409                 if (!nla_ok(nla_opt_msk, msk_depth)) {
1410                         NL_SET_ERR_MSG(extack, "A mask attribute is invalid");
1411                         return -EINVAL;
1412                 }
1413                 nla_opt_msk = nla_next(nla_opt_msk, &msk_depth);
1414         }
1415
1416         return 0;
1417 }
1418
1419 static int fl_validate_ct_state(u16 state, struct nlattr *tb,
1420                                 struct netlink_ext_ack *extack)
1421 {
1422         if (state && !(state & TCA_FLOWER_KEY_CT_FLAGS_TRACKED)) {
1423                 NL_SET_ERR_MSG_ATTR(extack, tb,
1424                                     "no trk, so no other flag can be set");
1425                 return -EINVAL;
1426         }
1427
1428         if (state & TCA_FLOWER_KEY_CT_FLAGS_NEW &&
1429             state & TCA_FLOWER_KEY_CT_FLAGS_ESTABLISHED) {
1430                 NL_SET_ERR_MSG_ATTR(extack, tb,
1431                                     "new and est are mutually exclusive");
1432                 return -EINVAL;
1433         }
1434
1435         return 0;
1436 }
1437
1438 static int fl_set_key_ct(struct nlattr **tb,
1439                          struct flow_dissector_key_ct *key,
1440                          struct flow_dissector_key_ct *mask,
1441                          struct netlink_ext_ack *extack)
1442 {
1443         if (tb[TCA_FLOWER_KEY_CT_STATE]) {
1444                 int err;
1445
1446                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK)) {
1447                         NL_SET_ERR_MSG(extack, "Conntrack isn't enabled");
1448                         return -EOPNOTSUPP;
1449                 }
1450                 fl_set_key_val(tb, &key->ct_state, TCA_FLOWER_KEY_CT_STATE,
1451                                &mask->ct_state, TCA_FLOWER_KEY_CT_STATE_MASK,
1452                                sizeof(key->ct_state));
1453
1454                 err = fl_validate_ct_state(key->ct_state & mask->ct_state,
1455                                            tb[TCA_FLOWER_KEY_CT_STATE_MASK],
1456                                            extack);
1457                 if (err)
1458                         return err;
1459
1460         }
1461         if (tb[TCA_FLOWER_KEY_CT_ZONE]) {
1462                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK_ZONES)) {
1463                         NL_SET_ERR_MSG(extack, "Conntrack zones isn't enabled");
1464                         return -EOPNOTSUPP;
1465                 }
1466                 fl_set_key_val(tb, &key->ct_zone, TCA_FLOWER_KEY_CT_ZONE,
1467                                &mask->ct_zone, TCA_FLOWER_KEY_CT_ZONE_MASK,
1468                                sizeof(key->ct_zone));
1469         }
1470         if (tb[TCA_FLOWER_KEY_CT_MARK]) {
1471                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK_MARK)) {
1472                         NL_SET_ERR_MSG(extack, "Conntrack mark isn't enabled");
1473                         return -EOPNOTSUPP;
1474                 }
1475                 fl_set_key_val(tb, &key->ct_mark, TCA_FLOWER_KEY_CT_MARK,
1476                                &mask->ct_mark, TCA_FLOWER_KEY_CT_MARK_MASK,
1477                                sizeof(key->ct_mark));
1478         }
1479         if (tb[TCA_FLOWER_KEY_CT_LABELS]) {
1480                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK_LABELS)) {
1481                         NL_SET_ERR_MSG(extack, "Conntrack labels aren't enabled");
1482                         return -EOPNOTSUPP;
1483                 }
1484                 fl_set_key_val(tb, key->ct_labels, TCA_FLOWER_KEY_CT_LABELS,
1485                                mask->ct_labels, TCA_FLOWER_KEY_CT_LABELS_MASK,
1486                                sizeof(key->ct_labels));
1487         }
1488
1489         return 0;
1490 }
1491
1492 static int fl_set_key(struct net *net, struct nlattr **tb,
1493                       struct fl_flow_key *key, struct fl_flow_key *mask,
1494                       struct netlink_ext_ack *extack)
1495 {
1496         __be16 ethertype;
1497         int ret = 0;
1498
1499         if (tb[TCA_FLOWER_INDEV]) {
1500                 int err = tcf_change_indev(net, tb[TCA_FLOWER_INDEV], extack);
1501                 if (err < 0)
1502                         return err;
1503                 key->meta.ingress_ifindex = err;
1504                 mask->meta.ingress_ifindex = 0xffffffff;
1505         }
1506
1507         fl_set_key_val(tb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
1508                        mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
1509                        sizeof(key->eth.dst));
1510         fl_set_key_val(tb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
1511                        mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
1512                        sizeof(key->eth.src));
1513
1514         if (tb[TCA_FLOWER_KEY_ETH_TYPE]) {
1515                 ethertype = nla_get_be16(tb[TCA_FLOWER_KEY_ETH_TYPE]);
1516
1517                 if (eth_type_vlan(ethertype)) {
1518                         fl_set_key_vlan(tb, ethertype, TCA_FLOWER_KEY_VLAN_ID,
1519                                         TCA_FLOWER_KEY_VLAN_PRIO,
1520                                         TCA_FLOWER_KEY_VLAN_ETH_TYPE,
1521                                         &key->vlan, &mask->vlan);
1522
1523                         if (tb[TCA_FLOWER_KEY_VLAN_ETH_TYPE]) {
1524                                 ethertype = nla_get_be16(tb[TCA_FLOWER_KEY_VLAN_ETH_TYPE]);
1525                                 if (eth_type_vlan(ethertype)) {
1526                                         fl_set_key_vlan(tb, ethertype,
1527                                                         TCA_FLOWER_KEY_CVLAN_ID,
1528                                                         TCA_FLOWER_KEY_CVLAN_PRIO,
1529                                                         TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
1530                                                         &key->cvlan, &mask->cvlan);
1531                                         fl_set_key_val(tb, &key->basic.n_proto,
1532                                                        TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
1533                                                        &mask->basic.n_proto,
1534                                                        TCA_FLOWER_UNSPEC,
1535                                                        sizeof(key->basic.n_proto));
1536                                 } else {
1537                                         key->basic.n_proto = ethertype;
1538                                         mask->basic.n_proto = cpu_to_be16(~0);
1539                                 }
1540                         }
1541                 } else {
1542                         key->basic.n_proto = ethertype;
1543                         mask->basic.n_proto = cpu_to_be16(~0);
1544                 }
1545         }
1546
1547         if (key->basic.n_proto == htons(ETH_P_IP) ||
1548             key->basic.n_proto == htons(ETH_P_IPV6)) {
1549                 fl_set_key_val(tb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
1550                                &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
1551                                sizeof(key->basic.ip_proto));
1552                 fl_set_key_ip(tb, false, &key->ip, &mask->ip);
1553         }
1554
1555         if (tb[TCA_FLOWER_KEY_IPV4_SRC] || tb[TCA_FLOWER_KEY_IPV4_DST]) {
1556                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
1557                 mask->control.addr_type = ~0;
1558                 fl_set_key_val(tb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
1559                                &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
1560                                sizeof(key->ipv4.src));
1561                 fl_set_key_val(tb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
1562                                &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
1563                                sizeof(key->ipv4.dst));
1564         } else if (tb[TCA_FLOWER_KEY_IPV6_SRC] || tb[TCA_FLOWER_KEY_IPV6_DST]) {
1565                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
1566                 mask->control.addr_type = ~0;
1567                 fl_set_key_val(tb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
1568                                &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
1569                                sizeof(key->ipv6.src));
1570                 fl_set_key_val(tb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
1571                                &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
1572                                sizeof(key->ipv6.dst));
1573         }
1574
1575         if (key->basic.ip_proto == IPPROTO_TCP) {
1576                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
1577                                &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
1578                                sizeof(key->tp.src));
1579                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
1580                                &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
1581                                sizeof(key->tp.dst));
1582                 fl_set_key_val(tb, &key->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS,
1583                                &mask->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS_MASK,
1584                                sizeof(key->tcp.flags));
1585         } else if (key->basic.ip_proto == IPPROTO_UDP) {
1586                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
1587                                &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
1588                                sizeof(key->tp.src));
1589                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
1590                                &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
1591                                sizeof(key->tp.dst));
1592         } else if (key->basic.ip_proto == IPPROTO_SCTP) {
1593                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_SCTP_SRC,
1594                                &mask->tp.src, TCA_FLOWER_KEY_SCTP_SRC_MASK,
1595                                sizeof(key->tp.src));
1596                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_SCTP_DST,
1597                                &mask->tp.dst, TCA_FLOWER_KEY_SCTP_DST_MASK,
1598                                sizeof(key->tp.dst));
1599         } else if (key->basic.n_proto == htons(ETH_P_IP) &&
1600                    key->basic.ip_proto == IPPROTO_ICMP) {
1601                 fl_set_key_val(tb, &key->icmp.type, TCA_FLOWER_KEY_ICMPV4_TYPE,
1602                                &mask->icmp.type,
1603                                TCA_FLOWER_KEY_ICMPV4_TYPE_MASK,
1604                                sizeof(key->icmp.type));
1605                 fl_set_key_val(tb, &key->icmp.code, TCA_FLOWER_KEY_ICMPV4_CODE,
1606                                &mask->icmp.code,
1607                                TCA_FLOWER_KEY_ICMPV4_CODE_MASK,
1608                                sizeof(key->icmp.code));
1609         } else if (key->basic.n_proto == htons(ETH_P_IPV6) &&
1610                    key->basic.ip_proto == IPPROTO_ICMPV6) {
1611                 fl_set_key_val(tb, &key->icmp.type, TCA_FLOWER_KEY_ICMPV6_TYPE,
1612                                &mask->icmp.type,
1613                                TCA_FLOWER_KEY_ICMPV6_TYPE_MASK,
1614                                sizeof(key->icmp.type));
1615                 fl_set_key_val(tb, &key->icmp.code, TCA_FLOWER_KEY_ICMPV6_CODE,
1616                                &mask->icmp.code,
1617                                TCA_FLOWER_KEY_ICMPV6_CODE_MASK,
1618                                sizeof(key->icmp.code));
1619         } else if (key->basic.n_proto == htons(ETH_P_MPLS_UC) ||
1620                    key->basic.n_proto == htons(ETH_P_MPLS_MC)) {
1621                 ret = fl_set_key_mpls(tb, &key->mpls, &mask->mpls, extack);
1622                 if (ret)
1623                         return ret;
1624         } else if (key->basic.n_proto == htons(ETH_P_ARP) ||
1625                    key->basic.n_proto == htons(ETH_P_RARP)) {
1626                 fl_set_key_val(tb, &key->arp.sip, TCA_FLOWER_KEY_ARP_SIP,
1627                                &mask->arp.sip, TCA_FLOWER_KEY_ARP_SIP_MASK,
1628                                sizeof(key->arp.sip));
1629                 fl_set_key_val(tb, &key->arp.tip, TCA_FLOWER_KEY_ARP_TIP,
1630                                &mask->arp.tip, TCA_FLOWER_KEY_ARP_TIP_MASK,
1631                                sizeof(key->arp.tip));
1632                 fl_set_key_val(tb, &key->arp.op, TCA_FLOWER_KEY_ARP_OP,
1633                                &mask->arp.op, TCA_FLOWER_KEY_ARP_OP_MASK,
1634                                sizeof(key->arp.op));
1635                 fl_set_key_val(tb, key->arp.sha, TCA_FLOWER_KEY_ARP_SHA,
1636                                mask->arp.sha, TCA_FLOWER_KEY_ARP_SHA_MASK,
1637                                sizeof(key->arp.sha));
1638                 fl_set_key_val(tb, key->arp.tha, TCA_FLOWER_KEY_ARP_THA,
1639                                mask->arp.tha, TCA_FLOWER_KEY_ARP_THA_MASK,
1640                                sizeof(key->arp.tha));
1641         }
1642
1643         if (key->basic.ip_proto == IPPROTO_TCP ||
1644             key->basic.ip_proto == IPPROTO_UDP ||
1645             key->basic.ip_proto == IPPROTO_SCTP) {
1646                 ret = fl_set_key_port_range(tb, key, mask, extack);
1647                 if (ret)
1648                         return ret;
1649         }
1650
1651         if (tb[TCA_FLOWER_KEY_ENC_IPV4_SRC] ||
1652             tb[TCA_FLOWER_KEY_ENC_IPV4_DST]) {
1653                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
1654                 mask->enc_control.addr_type = ~0;
1655                 fl_set_key_val(tb, &key->enc_ipv4.src,
1656                                TCA_FLOWER_KEY_ENC_IPV4_SRC,
1657                                &mask->enc_ipv4.src,
1658                                TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
1659                                sizeof(key->enc_ipv4.src));
1660                 fl_set_key_val(tb, &key->enc_ipv4.dst,
1661                                TCA_FLOWER_KEY_ENC_IPV4_DST,
1662                                &mask->enc_ipv4.dst,
1663                                TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
1664                                sizeof(key->enc_ipv4.dst));
1665         }
1666
1667         if (tb[TCA_FLOWER_KEY_ENC_IPV6_SRC] ||
1668             tb[TCA_FLOWER_KEY_ENC_IPV6_DST]) {
1669                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
1670                 mask->enc_control.addr_type = ~0;
1671                 fl_set_key_val(tb, &key->enc_ipv6.src,
1672                                TCA_FLOWER_KEY_ENC_IPV6_SRC,
1673                                &mask->enc_ipv6.src,
1674                                TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
1675                                sizeof(key->enc_ipv6.src));
1676                 fl_set_key_val(tb, &key->enc_ipv6.dst,
1677                                TCA_FLOWER_KEY_ENC_IPV6_DST,
1678                                &mask->enc_ipv6.dst,
1679                                TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
1680                                sizeof(key->enc_ipv6.dst));
1681         }
1682
1683         fl_set_key_val(tb, &key->enc_key_id.keyid, TCA_FLOWER_KEY_ENC_KEY_ID,
1684                        &mask->enc_key_id.keyid, TCA_FLOWER_UNSPEC,
1685                        sizeof(key->enc_key_id.keyid));
1686
1687         fl_set_key_val(tb, &key->enc_tp.src, TCA_FLOWER_KEY_ENC_UDP_SRC_PORT,
1688                        &mask->enc_tp.src, TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK,
1689                        sizeof(key->enc_tp.src));
1690
1691         fl_set_key_val(tb, &key->enc_tp.dst, TCA_FLOWER_KEY_ENC_UDP_DST_PORT,
1692                        &mask->enc_tp.dst, TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK,
1693                        sizeof(key->enc_tp.dst));
1694
1695         fl_set_key_ip(tb, true, &key->enc_ip, &mask->enc_ip);
1696
1697         fl_set_key_val(tb, &key->hash.hash, TCA_FLOWER_KEY_HASH,
1698                        &mask->hash.hash, TCA_FLOWER_KEY_HASH_MASK,
1699                        sizeof(key->hash.hash));
1700
1701         if (tb[TCA_FLOWER_KEY_ENC_OPTS]) {
1702                 ret = fl_set_enc_opt(tb, key, mask, extack);
1703                 if (ret)
1704                         return ret;
1705         }
1706
1707         ret = fl_set_key_ct(tb, &key->ct, &mask->ct, extack);
1708         if (ret)
1709                 return ret;
1710
1711         if (tb[TCA_FLOWER_KEY_FLAGS])
1712                 ret = fl_set_key_flags(tb, &key->control.flags,
1713                                        &mask->control.flags, extack);
1714
1715         return ret;
1716 }
1717
1718 static void fl_mask_copy(struct fl_flow_mask *dst,
1719                          struct fl_flow_mask *src)
1720 {
1721         const void *psrc = fl_key_get_start(&src->key, src);
1722         void *pdst = fl_key_get_start(&dst->key, src);
1723
1724         memcpy(pdst, psrc, fl_mask_range(src));
1725         dst->range = src->range;
1726 }
1727
1728 static const struct rhashtable_params fl_ht_params = {
1729         .key_offset = offsetof(struct cls_fl_filter, mkey), /* base offset */
1730         .head_offset = offsetof(struct cls_fl_filter, ht_node),
1731         .automatic_shrinking = true,
1732 };
1733
1734 static int fl_init_mask_hashtable(struct fl_flow_mask *mask)
1735 {
1736         mask->filter_ht_params = fl_ht_params;
1737         mask->filter_ht_params.key_len = fl_mask_range(mask);
1738         mask->filter_ht_params.key_offset += mask->range.start;
1739
1740         return rhashtable_init(&mask->ht, &mask->filter_ht_params);
1741 }
1742
1743 #define FL_KEY_MEMBER_OFFSET(member) offsetof(struct fl_flow_key, member)
1744 #define FL_KEY_MEMBER_SIZE(member) sizeof_field(struct fl_flow_key, member)
1745
1746 #define FL_KEY_IS_MASKED(mask, member)                                          \
1747         memchr_inv(((char *)mask) + FL_KEY_MEMBER_OFFSET(member),               \
1748                    0, FL_KEY_MEMBER_SIZE(member))                               \
1749
1750 #define FL_KEY_SET(keys, cnt, id, member)                                       \
1751         do {                                                                    \
1752                 keys[cnt].key_id = id;                                          \
1753                 keys[cnt].offset = FL_KEY_MEMBER_OFFSET(member);                \
1754                 cnt++;                                                          \
1755         } while(0);
1756
1757 #define FL_KEY_SET_IF_MASKED(mask, keys, cnt, id, member)                       \
1758         do {                                                                    \
1759                 if (FL_KEY_IS_MASKED(mask, member))                             \
1760                         FL_KEY_SET(keys, cnt, id, member);                      \
1761         } while(0);
1762
1763 static void fl_init_dissector(struct flow_dissector *dissector,
1764                               struct fl_flow_key *mask)
1765 {
1766         struct flow_dissector_key keys[FLOW_DISSECTOR_KEY_MAX];
1767         size_t cnt = 0;
1768
1769         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1770                              FLOW_DISSECTOR_KEY_META, meta);
1771         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_CONTROL, control);
1772         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_BASIC, basic);
1773         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1774                              FLOW_DISSECTOR_KEY_ETH_ADDRS, eth);
1775         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1776                              FLOW_DISSECTOR_KEY_IPV4_ADDRS, ipv4);
1777         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1778                              FLOW_DISSECTOR_KEY_IPV6_ADDRS, ipv6);
1779         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1780                              FLOW_DISSECTOR_KEY_PORTS, tp);
1781         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1782                              FLOW_DISSECTOR_KEY_PORTS_RANGE, tp_range);
1783         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1784                              FLOW_DISSECTOR_KEY_IP, ip);
1785         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1786                              FLOW_DISSECTOR_KEY_TCP, tcp);
1787         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1788                              FLOW_DISSECTOR_KEY_ICMP, icmp);
1789         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1790                              FLOW_DISSECTOR_KEY_ARP, arp);
1791         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1792                              FLOW_DISSECTOR_KEY_MPLS, mpls);
1793         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1794                              FLOW_DISSECTOR_KEY_VLAN, vlan);
1795         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1796                              FLOW_DISSECTOR_KEY_CVLAN, cvlan);
1797         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1798                              FLOW_DISSECTOR_KEY_ENC_KEYID, enc_key_id);
1799         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1800                              FLOW_DISSECTOR_KEY_ENC_IPV4_ADDRS, enc_ipv4);
1801         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1802                              FLOW_DISSECTOR_KEY_ENC_IPV6_ADDRS, enc_ipv6);
1803         if (FL_KEY_IS_MASKED(mask, enc_ipv4) ||
1804             FL_KEY_IS_MASKED(mask, enc_ipv6))
1805                 FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_ENC_CONTROL,
1806                            enc_control);
1807         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1808                              FLOW_DISSECTOR_KEY_ENC_PORTS, enc_tp);
1809         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1810                              FLOW_DISSECTOR_KEY_ENC_IP, enc_ip);
1811         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1812                              FLOW_DISSECTOR_KEY_ENC_OPTS, enc_opts);
1813         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1814                              FLOW_DISSECTOR_KEY_CT, ct);
1815         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1816                              FLOW_DISSECTOR_KEY_HASH, hash);
1817
1818         skb_flow_dissector_init(dissector, keys, cnt);
1819 }
1820
1821 static struct fl_flow_mask *fl_create_new_mask(struct cls_fl_head *head,
1822                                                struct fl_flow_mask *mask)
1823 {
1824         struct fl_flow_mask *newmask;
1825         int err;
1826
1827         newmask = kzalloc(sizeof(*newmask), GFP_KERNEL);
1828         if (!newmask)
1829                 return ERR_PTR(-ENOMEM);
1830
1831         fl_mask_copy(newmask, mask);
1832
1833         if ((newmask->key.tp_range.tp_min.dst &&
1834              newmask->key.tp_range.tp_max.dst) ||
1835             (newmask->key.tp_range.tp_min.src &&
1836              newmask->key.tp_range.tp_max.src))
1837                 newmask->flags |= TCA_FLOWER_MASK_FLAGS_RANGE;
1838
1839         err = fl_init_mask_hashtable(newmask);
1840         if (err)
1841                 goto errout_free;
1842
1843         fl_init_dissector(&newmask->dissector, &newmask->key);
1844
1845         INIT_LIST_HEAD_RCU(&newmask->filters);
1846
1847         refcount_set(&newmask->refcnt, 1);
1848         err = rhashtable_replace_fast(&head->ht, &mask->ht_node,
1849                                       &newmask->ht_node, mask_ht_params);
1850         if (err)
1851                 goto errout_destroy;
1852
1853         spin_lock(&head->masks_lock);
1854         list_add_tail_rcu(&newmask->list, &head->masks);
1855         spin_unlock(&head->masks_lock);
1856
1857         return newmask;
1858
1859 errout_destroy:
1860         rhashtable_destroy(&newmask->ht);
1861 errout_free:
1862         kfree(newmask);
1863
1864         return ERR_PTR(err);
1865 }
1866
1867 static int fl_check_assign_mask(struct cls_fl_head *head,
1868                                 struct cls_fl_filter *fnew,
1869                                 struct cls_fl_filter *fold,
1870                                 struct fl_flow_mask *mask)
1871 {
1872         struct fl_flow_mask *newmask;
1873         int ret = 0;
1874
1875         rcu_read_lock();
1876
1877         /* Insert mask as temporary node to prevent concurrent creation of mask
1878          * with same key. Any concurrent lookups with same key will return
1879          * -EAGAIN because mask's refcnt is zero.
1880          */
1881         fnew->mask = rhashtable_lookup_get_insert_fast(&head->ht,
1882                                                        &mask->ht_node,
1883                                                        mask_ht_params);
1884         if (!fnew->mask) {
1885                 rcu_read_unlock();
1886
1887                 if (fold) {
1888                         ret = -EINVAL;
1889                         goto errout_cleanup;
1890                 }
1891
1892                 newmask = fl_create_new_mask(head, mask);
1893                 if (IS_ERR(newmask)) {
1894                         ret = PTR_ERR(newmask);
1895                         goto errout_cleanup;
1896                 }
1897
1898                 fnew->mask = newmask;
1899                 return 0;
1900         } else if (IS_ERR(fnew->mask)) {
1901                 ret = PTR_ERR(fnew->mask);
1902         } else if (fold && fold->mask != fnew->mask) {
1903                 ret = -EINVAL;
1904         } else if (!refcount_inc_not_zero(&fnew->mask->refcnt)) {
1905                 /* Mask was deleted concurrently, try again */
1906                 ret = -EAGAIN;
1907         }
1908         rcu_read_unlock();
1909         return ret;
1910
1911 errout_cleanup:
1912         rhashtable_remove_fast(&head->ht, &mask->ht_node,
1913                                mask_ht_params);
1914         return ret;
1915 }
1916
1917 static int fl_set_parms(struct net *net, struct tcf_proto *tp,
1918                         struct cls_fl_filter *f, struct fl_flow_mask *mask,
1919                         unsigned long base, struct nlattr **tb,
1920                         struct nlattr *est, bool ovr,
1921                         struct fl_flow_tmplt *tmplt, bool rtnl_held,
1922                         struct netlink_ext_ack *extack)
1923 {
1924         int err;
1925
1926         err = tcf_exts_validate(net, tp, tb, est, &f->exts, ovr, rtnl_held,
1927                                 extack);
1928         if (err < 0)
1929                 return err;
1930
1931         if (tb[TCA_FLOWER_CLASSID]) {
1932                 f->res.classid = nla_get_u32(tb[TCA_FLOWER_CLASSID]);
1933                 if (!rtnl_held)
1934                         rtnl_lock();
1935                 tcf_bind_filter(tp, &f->res, base);
1936                 if (!rtnl_held)
1937                         rtnl_unlock();
1938         }
1939
1940         err = fl_set_key(net, tb, &f->key, &mask->key, extack);
1941         if (err)
1942                 return err;
1943
1944         fl_mask_update_range(mask);
1945         fl_set_masked_key(&f->mkey, &f->key, mask);
1946
1947         if (!fl_mask_fits_tmplt(tmplt, mask)) {
1948                 NL_SET_ERR_MSG_MOD(extack, "Mask does not fit the template");
1949                 return -EINVAL;
1950         }
1951
1952         return 0;
1953 }
1954
1955 static int fl_ht_insert_unique(struct cls_fl_filter *fnew,
1956                                struct cls_fl_filter *fold,
1957                                bool *in_ht)
1958 {
1959         struct fl_flow_mask *mask = fnew->mask;
1960         int err;
1961
1962         err = rhashtable_lookup_insert_fast(&mask->ht,
1963                                             &fnew->ht_node,
1964                                             mask->filter_ht_params);
1965         if (err) {
1966                 *in_ht = false;
1967                 /* It is okay if filter with same key exists when
1968                  * overwriting.
1969                  */
1970                 return fold && err == -EEXIST ? 0 : err;
1971         }
1972
1973         *in_ht = true;
1974         return 0;
1975 }
1976
1977 static int fl_change(struct net *net, struct sk_buff *in_skb,
1978                      struct tcf_proto *tp, unsigned long base,
1979                      u32 handle, struct nlattr **tca,
1980                      void **arg, bool ovr, bool rtnl_held,
1981                      struct netlink_ext_ack *extack)
1982 {
1983         struct cls_fl_head *head = fl_head_dereference(tp);
1984         struct cls_fl_filter *fold = *arg;
1985         struct cls_fl_filter *fnew;
1986         struct fl_flow_mask *mask;
1987         struct nlattr **tb;
1988         bool in_ht;
1989         int err;
1990
1991         if (!tca[TCA_OPTIONS]) {
1992                 err = -EINVAL;
1993                 goto errout_fold;
1994         }
1995
1996         mask = kzalloc(sizeof(struct fl_flow_mask), GFP_KERNEL);
1997         if (!mask) {
1998                 err = -ENOBUFS;
1999                 goto errout_fold;
2000         }
2001
2002         tb = kcalloc(TCA_FLOWER_MAX + 1, sizeof(struct nlattr *), GFP_KERNEL);
2003         if (!tb) {
2004                 err = -ENOBUFS;
2005                 goto errout_mask_alloc;
2006         }
2007
2008         err = nla_parse_nested_deprecated(tb, TCA_FLOWER_MAX,
2009                                           tca[TCA_OPTIONS], fl_policy, NULL);
2010         if (err < 0)
2011                 goto errout_tb;
2012
2013         if (fold && handle && fold->handle != handle) {
2014                 err = -EINVAL;
2015                 goto errout_tb;
2016         }
2017
2018         fnew = kzalloc(sizeof(*fnew), GFP_KERNEL);
2019         if (!fnew) {
2020                 err = -ENOBUFS;
2021                 goto errout_tb;
2022         }
2023         INIT_LIST_HEAD(&fnew->hw_list);
2024         refcount_set(&fnew->refcnt, 1);
2025
2026         err = tcf_exts_init(&fnew->exts, net, TCA_FLOWER_ACT, 0);
2027         if (err < 0)
2028                 goto errout;
2029
2030         if (tb[TCA_FLOWER_FLAGS]) {
2031                 fnew->flags = nla_get_u32(tb[TCA_FLOWER_FLAGS]);
2032
2033                 if (!tc_flags_valid(fnew->flags)) {
2034                         err = -EINVAL;
2035                         goto errout;
2036                 }
2037         }
2038
2039         err = fl_set_parms(net, tp, fnew, mask, base, tb, tca[TCA_RATE], ovr,
2040                            tp->chain->tmplt_priv, rtnl_held, extack);
2041         if (err)
2042                 goto errout;
2043
2044         err = fl_check_assign_mask(head, fnew, fold, mask);
2045         if (err)
2046                 goto errout;
2047
2048         err = fl_ht_insert_unique(fnew, fold, &in_ht);
2049         if (err)
2050                 goto errout_mask;
2051
2052         if (!tc_skip_hw(fnew->flags)) {
2053                 err = fl_hw_replace_filter(tp, fnew, rtnl_held, extack);
2054                 if (err)
2055                         goto errout_ht;
2056         }
2057
2058         if (!tc_in_hw(fnew->flags))
2059                 fnew->flags |= TCA_CLS_FLAGS_NOT_IN_HW;
2060
2061         spin_lock(&tp->lock);
2062
2063         /* tp was deleted concurrently. -EAGAIN will cause caller to lookup
2064          * proto again or create new one, if necessary.
2065          */
2066         if (tp->deleting) {
2067                 err = -EAGAIN;
2068                 goto errout_hw;
2069         }
2070
2071         if (fold) {
2072                 /* Fold filter was deleted concurrently. Retry lookup. */
2073                 if (fold->deleted) {
2074                         err = -EAGAIN;
2075                         goto errout_hw;
2076                 }
2077
2078                 fnew->handle = handle;
2079
2080                 if (!in_ht) {
2081                         struct rhashtable_params params =
2082                                 fnew->mask->filter_ht_params;
2083
2084                         err = rhashtable_insert_fast(&fnew->mask->ht,
2085                                                      &fnew->ht_node,
2086                                                      params);
2087                         if (err)
2088                                 goto errout_hw;
2089                         in_ht = true;
2090                 }
2091
2092                 refcount_inc(&fnew->refcnt);
2093                 rhashtable_remove_fast(&fold->mask->ht,
2094                                        &fold->ht_node,
2095                                        fold->mask->filter_ht_params);
2096                 idr_replace(&head->handle_idr, fnew, fnew->handle);
2097                 list_replace_rcu(&fold->list, &fnew->list);
2098                 fold->deleted = true;
2099
2100                 spin_unlock(&tp->lock);
2101
2102                 fl_mask_put(head, fold->mask);
2103                 if (!tc_skip_hw(fold->flags))
2104                         fl_hw_destroy_filter(tp, fold, rtnl_held, NULL);
2105                 tcf_unbind_filter(tp, &fold->res);
2106                 /* Caller holds reference to fold, so refcnt is always > 0
2107                  * after this.
2108                  */
2109                 refcount_dec(&fold->refcnt);
2110                 __fl_put(fold);
2111         } else {
2112                 if (handle) {
2113                         /* user specifies a handle and it doesn't exist */
2114                         err = idr_alloc_u32(&head->handle_idr, fnew, &handle,
2115                                             handle, GFP_ATOMIC);
2116
2117                         /* Filter with specified handle was concurrently
2118                          * inserted after initial check in cls_api. This is not
2119                          * necessarily an error if NLM_F_EXCL is not set in
2120                          * message flags. Returning EAGAIN will cause cls_api to
2121                          * try to update concurrently inserted rule.
2122                          */
2123                         if (err == -ENOSPC)
2124                                 err = -EAGAIN;
2125                 } else {
2126                         handle = 1;
2127                         err = idr_alloc_u32(&head->handle_idr, fnew, &handle,
2128                                             INT_MAX, GFP_ATOMIC);
2129                 }
2130                 if (err)
2131                         goto errout_hw;
2132
2133                 refcount_inc(&fnew->refcnt);
2134                 fnew->handle = handle;
2135                 list_add_tail_rcu(&fnew->list, &fnew->mask->filters);
2136                 spin_unlock(&tp->lock);
2137         }
2138
2139         *arg = fnew;
2140
2141         kfree(tb);
2142         tcf_queue_work(&mask->rwork, fl_uninit_mask_free_work);
2143         return 0;
2144
2145 errout_ht:
2146         spin_lock(&tp->lock);
2147 errout_hw:
2148         fnew->deleted = true;
2149         spin_unlock(&tp->lock);
2150         if (!tc_skip_hw(fnew->flags))
2151                 fl_hw_destroy_filter(tp, fnew, rtnl_held, NULL);
2152         if (in_ht)
2153                 rhashtable_remove_fast(&fnew->mask->ht, &fnew->ht_node,
2154                                        fnew->mask->filter_ht_params);
2155 errout_mask:
2156         fl_mask_put(head, fnew->mask);
2157 errout:
2158         __fl_put(fnew);
2159 errout_tb:
2160         kfree(tb);
2161 errout_mask_alloc:
2162         tcf_queue_work(&mask->rwork, fl_uninit_mask_free_work);
2163 errout_fold:
2164         if (fold)
2165                 __fl_put(fold);
2166         return err;
2167 }
2168
2169 static int fl_delete(struct tcf_proto *tp, void *arg, bool *last,
2170                      bool rtnl_held, struct netlink_ext_ack *extack)
2171 {
2172         struct cls_fl_head *head = fl_head_dereference(tp);
2173         struct cls_fl_filter *f = arg;
2174         bool last_on_mask;
2175         int err = 0;
2176
2177         err = __fl_delete(tp, f, &last_on_mask, rtnl_held, extack);
2178         *last = list_empty(&head->masks);
2179         __fl_put(f);
2180
2181         return err;
2182 }
2183
2184 static void fl_walk(struct tcf_proto *tp, struct tcf_walker *arg,
2185                     bool rtnl_held)
2186 {
2187         struct cls_fl_head *head = fl_head_dereference(tp);
2188         unsigned long id = arg->cookie, tmp;
2189         struct cls_fl_filter *f;
2190
2191         arg->count = arg->skip;
2192
2193         rcu_read_lock();
2194         idr_for_each_entry_continue_ul(&head->handle_idr, f, tmp, id) {
2195                 /* don't return filters that are being deleted */
2196                 if (!refcount_inc_not_zero(&f->refcnt))
2197                         continue;
2198                 rcu_read_unlock();
2199
2200                 if (arg->fn(tp, f, arg) < 0) {
2201                         __fl_put(f);
2202                         arg->stop = 1;
2203                         rcu_read_lock();
2204                         break;
2205                 }
2206                 __fl_put(f);
2207                 arg->count++;
2208                 rcu_read_lock();
2209         }
2210         rcu_read_unlock();
2211         arg->cookie = id;
2212 }
2213
2214 static struct cls_fl_filter *
2215 fl_get_next_hw_filter(struct tcf_proto *tp, struct cls_fl_filter *f, bool add)
2216 {
2217         struct cls_fl_head *head = fl_head_dereference(tp);
2218
2219         spin_lock(&tp->lock);
2220         if (list_empty(&head->hw_filters)) {
2221                 spin_unlock(&tp->lock);
2222                 return NULL;
2223         }
2224
2225         if (!f)
2226                 f = list_entry(&head->hw_filters, struct cls_fl_filter,
2227                                hw_list);
2228         list_for_each_entry_continue(f, &head->hw_filters, hw_list) {
2229                 if (!(add && f->deleted) && refcount_inc_not_zero(&f->refcnt)) {
2230                         spin_unlock(&tp->lock);
2231                         return f;
2232                 }
2233         }
2234
2235         spin_unlock(&tp->lock);
2236         return NULL;
2237 }
2238
2239 static int fl_reoffload(struct tcf_proto *tp, bool add, flow_setup_cb_t *cb,
2240                         void *cb_priv, struct netlink_ext_ack *extack)
2241 {
2242         struct tcf_block *block = tp->chain->block;
2243         struct flow_cls_offload cls_flower = {};
2244         struct cls_fl_filter *f = NULL;
2245         int err;
2246
2247         /* hw_filters list can only be changed by hw offload functions after
2248          * obtaining rtnl lock. Make sure it is not changed while reoffload is
2249          * iterating it.
2250          */
2251         ASSERT_RTNL();
2252
2253         while ((f = fl_get_next_hw_filter(tp, f, add))) {
2254                 cls_flower.rule =
2255                         flow_rule_alloc(tcf_exts_num_actions(&f->exts));
2256                 if (!cls_flower.rule) {
2257                         __fl_put(f);
2258                         return -ENOMEM;
2259                 }
2260
2261                 tc_cls_common_offload_init(&cls_flower.common, tp, f->flags,
2262                                            extack);
2263                 cls_flower.command = add ?
2264                         FLOW_CLS_REPLACE : FLOW_CLS_DESTROY;
2265                 cls_flower.cookie = (unsigned long)f;
2266                 cls_flower.rule->match.dissector = &f->mask->dissector;
2267                 cls_flower.rule->match.mask = &f->mask->key;
2268                 cls_flower.rule->match.key = &f->mkey;
2269
2270                 err = tc_setup_flow_action(&cls_flower.rule->action, &f->exts);
2271                 if (err) {
2272                         kfree(cls_flower.rule);
2273                         if (tc_skip_sw(f->flags)) {
2274                                 NL_SET_ERR_MSG_MOD(extack, "Failed to setup flow action");
2275                                 __fl_put(f);
2276                                 return err;
2277                         }
2278                         goto next_flow;
2279                 }
2280
2281                 cls_flower.classid = f->res.classid;
2282
2283                 err = tc_setup_cb_reoffload(block, tp, add, cb,
2284                                             TC_SETUP_CLSFLOWER, &cls_flower,
2285                                             cb_priv, &f->flags,
2286                                             &f->in_hw_count);
2287                 tc_cleanup_flow_action(&cls_flower.rule->action);
2288                 kfree(cls_flower.rule);
2289
2290                 if (err) {
2291                         __fl_put(f);
2292                         return err;
2293                 }
2294 next_flow:
2295                 __fl_put(f);
2296         }
2297
2298         return 0;
2299 }
2300
2301 static void fl_hw_add(struct tcf_proto *tp, void *type_data)
2302 {
2303         struct flow_cls_offload *cls_flower = type_data;
2304         struct cls_fl_filter *f =
2305                 (struct cls_fl_filter *) cls_flower->cookie;
2306         struct cls_fl_head *head = fl_head_dereference(tp);
2307
2308         spin_lock(&tp->lock);
2309         list_add(&f->hw_list, &head->hw_filters);
2310         spin_unlock(&tp->lock);
2311 }
2312
2313 static void fl_hw_del(struct tcf_proto *tp, void *type_data)
2314 {
2315         struct flow_cls_offload *cls_flower = type_data;
2316         struct cls_fl_filter *f =
2317                 (struct cls_fl_filter *) cls_flower->cookie;
2318
2319         spin_lock(&tp->lock);
2320         if (!list_empty(&f->hw_list))
2321                 list_del_init(&f->hw_list);
2322         spin_unlock(&tp->lock);
2323 }
2324
2325 static int fl_hw_create_tmplt(struct tcf_chain *chain,
2326                               struct fl_flow_tmplt *tmplt)
2327 {
2328         struct flow_cls_offload cls_flower = {};
2329         struct tcf_block *block = chain->block;
2330
2331         cls_flower.rule = flow_rule_alloc(0);
2332         if (!cls_flower.rule)
2333                 return -ENOMEM;
2334
2335         cls_flower.common.chain_index = chain->index;
2336         cls_flower.command = FLOW_CLS_TMPLT_CREATE;
2337         cls_flower.cookie = (unsigned long) tmplt;
2338         cls_flower.rule->match.dissector = &tmplt->dissector;
2339         cls_flower.rule->match.mask = &tmplt->mask;
2340         cls_flower.rule->match.key = &tmplt->dummy_key;
2341
2342         /* We don't care if driver (any of them) fails to handle this
2343          * call. It serves just as a hint for it.
2344          */
2345         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false, true);
2346         kfree(cls_flower.rule);
2347
2348         return 0;
2349 }
2350
2351 static void fl_hw_destroy_tmplt(struct tcf_chain *chain,
2352                                 struct fl_flow_tmplt *tmplt)
2353 {
2354         struct flow_cls_offload cls_flower = {};
2355         struct tcf_block *block = chain->block;
2356
2357         cls_flower.common.chain_index = chain->index;
2358         cls_flower.command = FLOW_CLS_TMPLT_DESTROY;
2359         cls_flower.cookie = (unsigned long) tmplt;
2360
2361         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false, true);
2362 }
2363
2364 static void *fl_tmplt_create(struct net *net, struct tcf_chain *chain,
2365                              struct nlattr **tca,
2366                              struct netlink_ext_ack *extack)
2367 {
2368         struct fl_flow_tmplt *tmplt;
2369         struct nlattr **tb;
2370         int err;
2371
2372         if (!tca[TCA_OPTIONS])
2373                 return ERR_PTR(-EINVAL);
2374
2375         tb = kcalloc(TCA_FLOWER_MAX + 1, sizeof(struct nlattr *), GFP_KERNEL);
2376         if (!tb)
2377                 return ERR_PTR(-ENOBUFS);
2378         err = nla_parse_nested_deprecated(tb, TCA_FLOWER_MAX,
2379                                           tca[TCA_OPTIONS], fl_policy, NULL);
2380         if (err)
2381                 goto errout_tb;
2382
2383         tmplt = kzalloc(sizeof(*tmplt), GFP_KERNEL);
2384         if (!tmplt) {
2385                 err = -ENOMEM;
2386                 goto errout_tb;
2387         }
2388         tmplt->chain = chain;
2389         err = fl_set_key(net, tb, &tmplt->dummy_key, &tmplt->mask, extack);
2390         if (err)
2391                 goto errout_tmplt;
2392
2393         fl_init_dissector(&tmplt->dissector, &tmplt->mask);
2394
2395         err = fl_hw_create_tmplt(chain, tmplt);
2396         if (err)
2397                 goto errout_tmplt;
2398
2399         kfree(tb);
2400         return tmplt;
2401
2402 errout_tmplt:
2403         kfree(tmplt);
2404 errout_tb:
2405         kfree(tb);
2406         return ERR_PTR(err);
2407 }
2408
2409 static void fl_tmplt_destroy(void *tmplt_priv)
2410 {
2411         struct fl_flow_tmplt *tmplt = tmplt_priv;
2412
2413         fl_hw_destroy_tmplt(tmplt->chain, tmplt);
2414         kfree(tmplt);
2415 }
2416
2417 static int fl_dump_key_val(struct sk_buff *skb,
2418                            void *val, int val_type,
2419                            void *mask, int mask_type, int len)
2420 {
2421         int err;
2422
2423         if (!memchr_inv(mask, 0, len))
2424                 return 0;
2425         err = nla_put(skb, val_type, len, val);
2426         if (err)
2427                 return err;
2428         if (mask_type != TCA_FLOWER_UNSPEC) {
2429                 err = nla_put(skb, mask_type, len, mask);
2430                 if (err)
2431                         return err;
2432         }
2433         return 0;
2434 }
2435
2436 static int fl_dump_key_port_range(struct sk_buff *skb, struct fl_flow_key *key,
2437                                   struct fl_flow_key *mask)
2438 {
2439         if (fl_dump_key_val(skb, &key->tp_range.tp_min.dst,
2440                             TCA_FLOWER_KEY_PORT_DST_MIN,
2441                             &mask->tp_range.tp_min.dst, TCA_FLOWER_UNSPEC,
2442                             sizeof(key->tp_range.tp_min.dst)) ||
2443             fl_dump_key_val(skb, &key->tp_range.tp_max.dst,
2444                             TCA_FLOWER_KEY_PORT_DST_MAX,
2445                             &mask->tp_range.tp_max.dst, TCA_FLOWER_UNSPEC,
2446                             sizeof(key->tp_range.tp_max.dst)) ||
2447             fl_dump_key_val(skb, &key->tp_range.tp_min.src,
2448                             TCA_FLOWER_KEY_PORT_SRC_MIN,
2449                             &mask->tp_range.tp_min.src, TCA_FLOWER_UNSPEC,
2450                             sizeof(key->tp_range.tp_min.src)) ||
2451             fl_dump_key_val(skb, &key->tp_range.tp_max.src,
2452                             TCA_FLOWER_KEY_PORT_SRC_MAX,
2453                             &mask->tp_range.tp_max.src, TCA_FLOWER_UNSPEC,
2454                             sizeof(key->tp_range.tp_max.src)))
2455                 return -1;
2456
2457         return 0;
2458 }
2459
2460 static int fl_dump_key_mpls_opt_lse(struct sk_buff *skb,
2461                                     struct flow_dissector_key_mpls *mpls_key,
2462                                     struct flow_dissector_key_mpls *mpls_mask,
2463                                     u8 lse_index)
2464 {
2465         struct flow_dissector_mpls_lse *lse_mask = &mpls_mask->ls[lse_index];
2466         struct flow_dissector_mpls_lse *lse_key = &mpls_key->ls[lse_index];
2467         int err;
2468
2469         err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH,
2470                          lse_index + 1);
2471         if (err)
2472                 return err;
2473
2474         if (lse_mask->mpls_ttl) {
2475                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_TTL,
2476                                  lse_key->mpls_ttl);
2477                 if (err)
2478                         return err;
2479         }
2480         if (lse_mask->mpls_bos) {
2481                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS,
2482                                  lse_key->mpls_bos);
2483                 if (err)
2484                         return err;
2485         }
2486         if (lse_mask->mpls_tc) {
2487                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_TC,
2488                                  lse_key->mpls_tc);
2489                 if (err)
2490                         return err;
2491         }
2492         if (lse_mask->mpls_label) {
2493                 err = nla_put_u32(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL,
2494                                   lse_key->mpls_label);
2495                 if (err)
2496                         return err;
2497         }
2498
2499         return 0;
2500 }
2501
2502 static int fl_dump_key_mpls_opts(struct sk_buff *skb,
2503                                  struct flow_dissector_key_mpls *mpls_key,
2504                                  struct flow_dissector_key_mpls *mpls_mask)
2505 {
2506         struct nlattr *opts;
2507         struct nlattr *lse;
2508         u8 lse_index;
2509         int err;
2510
2511         opts = nla_nest_start(skb, TCA_FLOWER_KEY_MPLS_OPTS);
2512         if (!opts)
2513                 return -EMSGSIZE;
2514
2515         for (lse_index = 0; lse_index < FLOW_DIS_MPLS_MAX; lse_index++) {
2516                 if (!(mpls_mask->used_lses & 1 << lse_index))
2517                         continue;
2518
2519                 lse = nla_nest_start(skb, TCA_FLOWER_KEY_MPLS_OPTS_LSE);
2520                 if (!lse) {
2521                         err = -EMSGSIZE;
2522                         goto err_opts;
2523                 }
2524
2525                 err = fl_dump_key_mpls_opt_lse(skb, mpls_key, mpls_mask,
2526                                                lse_index);
2527                 if (err)
2528                         goto err_opts_lse;
2529                 nla_nest_end(skb, lse);
2530         }
2531         nla_nest_end(skb, opts);
2532
2533         return 0;
2534
2535 err_opts_lse:
2536         nla_nest_cancel(skb, lse);
2537 err_opts:
2538         nla_nest_cancel(skb, opts);
2539
2540         return err;
2541 }
2542
2543 static int fl_dump_key_mpls(struct sk_buff *skb,
2544                             struct flow_dissector_key_mpls *mpls_key,
2545                             struct flow_dissector_key_mpls *mpls_mask)
2546 {
2547         struct flow_dissector_mpls_lse *lse_mask;
2548         struct flow_dissector_mpls_lse *lse_key;
2549         int err;
2550
2551         if (!mpls_mask->used_lses)
2552                 return 0;
2553
2554         lse_mask = &mpls_mask->ls[0];
2555         lse_key = &mpls_key->ls[0];
2556
2557         /* For backward compatibility, don't use the MPLS nested attributes if
2558          * the rule can be expressed using the old attributes.
2559          */
2560         if (mpls_mask->used_lses & ~1 ||
2561             (!lse_mask->mpls_ttl && !lse_mask->mpls_bos &&
2562              !lse_mask->mpls_tc && !lse_mask->mpls_label))
2563                 return fl_dump_key_mpls_opts(skb, mpls_key, mpls_mask);
2564
2565         if (lse_mask->mpls_ttl) {
2566                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_TTL,
2567                                  lse_key->mpls_ttl);
2568                 if (err)
2569                         return err;
2570         }
2571         if (lse_mask->mpls_tc) {
2572                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_TC,
2573                                  lse_key->mpls_tc);
2574                 if (err)
2575                         return err;
2576         }
2577         if (lse_mask->mpls_label) {
2578                 err = nla_put_u32(skb, TCA_FLOWER_KEY_MPLS_LABEL,
2579                                   lse_key->mpls_label);
2580                 if (err)
2581                         return err;
2582         }
2583         if (lse_mask->mpls_bos) {
2584                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_BOS,
2585                                  lse_key->mpls_bos);
2586                 if (err)
2587                         return err;
2588         }
2589         return 0;
2590 }
2591
2592 static int fl_dump_key_ip(struct sk_buff *skb, bool encap,
2593                           struct flow_dissector_key_ip *key,
2594                           struct flow_dissector_key_ip *mask)
2595 {
2596         int tos_key = encap ? TCA_FLOWER_KEY_ENC_IP_TOS : TCA_FLOWER_KEY_IP_TOS;
2597         int ttl_key = encap ? TCA_FLOWER_KEY_ENC_IP_TTL : TCA_FLOWER_KEY_IP_TTL;
2598         int tos_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TOS_MASK : TCA_FLOWER_KEY_IP_TOS_MASK;
2599         int ttl_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TTL_MASK : TCA_FLOWER_KEY_IP_TTL_MASK;
2600
2601         if (fl_dump_key_val(skb, &key->tos, tos_key, &mask->tos, tos_mask, sizeof(key->tos)) ||
2602             fl_dump_key_val(skb, &key->ttl, ttl_key, &mask->ttl, ttl_mask, sizeof(key->ttl)))
2603                 return -1;
2604
2605         return 0;
2606 }
2607
2608 static int fl_dump_key_vlan(struct sk_buff *skb,
2609                             int vlan_id_key, int vlan_prio_key,
2610                             struct flow_dissector_key_vlan *vlan_key,
2611                             struct flow_dissector_key_vlan *vlan_mask)
2612 {
2613         int err;
2614
2615         if (!memchr_inv(vlan_mask, 0, sizeof(*vlan_mask)))
2616                 return 0;
2617         if (vlan_mask->vlan_id) {
2618                 err = nla_put_u16(skb, vlan_id_key,
2619                                   vlan_key->vlan_id);
2620                 if (err)
2621                         return err;
2622         }
2623         if (vlan_mask->vlan_priority) {
2624                 err = nla_put_u8(skb, vlan_prio_key,
2625                                  vlan_key->vlan_priority);
2626                 if (err)
2627                         return err;
2628         }
2629         return 0;
2630 }
2631
2632 static void fl_get_key_flag(u32 dissector_key, u32 dissector_mask,
2633                             u32 *flower_key, u32 *flower_mask,
2634                             u32 flower_flag_bit, u32 dissector_flag_bit)
2635 {
2636         if (dissector_mask & dissector_flag_bit) {
2637                 *flower_mask |= flower_flag_bit;
2638                 if (dissector_key & dissector_flag_bit)
2639                         *flower_key |= flower_flag_bit;
2640         }
2641 }
2642
2643 static int fl_dump_key_flags(struct sk_buff *skb, u32 flags_key, u32 flags_mask)
2644 {
2645         u32 key, mask;
2646         __be32 _key, _mask;
2647         int err;
2648
2649         if (!memchr_inv(&flags_mask, 0, sizeof(flags_mask)))
2650                 return 0;
2651
2652         key = 0;
2653         mask = 0;
2654
2655         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
2656                         TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT, FLOW_DIS_IS_FRAGMENT);
2657         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
2658                         TCA_FLOWER_KEY_FLAGS_FRAG_IS_FIRST,
2659                         FLOW_DIS_FIRST_FRAG);
2660
2661         _key = cpu_to_be32(key);
2662         _mask = cpu_to_be32(mask);
2663
2664         err = nla_put(skb, TCA_FLOWER_KEY_FLAGS, 4, &_key);
2665         if (err)
2666                 return err;
2667
2668         return nla_put(skb, TCA_FLOWER_KEY_FLAGS_MASK, 4, &_mask);
2669 }
2670
2671 static int fl_dump_key_geneve_opt(struct sk_buff *skb,
2672                                   struct flow_dissector_key_enc_opts *enc_opts)
2673 {
2674         struct geneve_opt *opt;
2675         struct nlattr *nest;
2676         int opt_off = 0;
2677
2678         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_GENEVE);
2679         if (!nest)
2680                 goto nla_put_failure;
2681
2682         while (enc_opts->len > opt_off) {
2683                 opt = (struct geneve_opt *)&enc_opts->data[opt_off];
2684
2685                 if (nla_put_be16(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS,
2686                                  opt->opt_class))
2687                         goto nla_put_failure;
2688                 if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE,
2689                                opt->type))
2690                         goto nla_put_failure;
2691                 if (nla_put(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA,
2692                             opt->length * 4, opt->opt_data))
2693                         goto nla_put_failure;
2694
2695                 opt_off += sizeof(struct geneve_opt) + opt->length * 4;
2696         }
2697         nla_nest_end(skb, nest);
2698         return 0;
2699
2700 nla_put_failure:
2701         nla_nest_cancel(skb, nest);
2702         return -EMSGSIZE;
2703 }
2704
2705 static int fl_dump_key_vxlan_opt(struct sk_buff *skb,
2706                                  struct flow_dissector_key_enc_opts *enc_opts)
2707 {
2708         struct vxlan_metadata *md;
2709         struct nlattr *nest;
2710
2711         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_VXLAN);
2712         if (!nest)
2713                 goto nla_put_failure;
2714
2715         md = (struct vxlan_metadata *)&enc_opts->data[0];
2716         if (nla_put_u32(skb, TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP, md->gbp))
2717                 goto nla_put_failure;
2718
2719         nla_nest_end(skb, nest);
2720         return 0;
2721
2722 nla_put_failure:
2723         nla_nest_cancel(skb, nest);
2724         return -EMSGSIZE;
2725 }
2726
2727 static int fl_dump_key_erspan_opt(struct sk_buff *skb,
2728                                   struct flow_dissector_key_enc_opts *enc_opts)
2729 {
2730         struct erspan_metadata *md;
2731         struct nlattr *nest;
2732
2733         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_ERSPAN);
2734         if (!nest)
2735                 goto nla_put_failure;
2736
2737         md = (struct erspan_metadata *)&enc_opts->data[0];
2738         if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER, md->version))
2739                 goto nla_put_failure;
2740
2741         if (md->version == 1 &&
2742             nla_put_be32(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX, md->u.index))
2743                 goto nla_put_failure;
2744
2745         if (md->version == 2 &&
2746             (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR,
2747                         md->u.md2.dir) ||
2748              nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID,
2749                         get_hwid(&md->u.md2))))
2750                 goto nla_put_failure;
2751
2752         nla_nest_end(skb, nest);
2753         return 0;
2754
2755 nla_put_failure:
2756         nla_nest_cancel(skb, nest);
2757         return -EMSGSIZE;
2758 }
2759
2760 static int fl_dump_key_ct(struct sk_buff *skb,
2761                           struct flow_dissector_key_ct *key,
2762                           struct flow_dissector_key_ct *mask)
2763 {
2764         if (IS_ENABLED(CONFIG_NF_CONNTRACK) &&
2765             fl_dump_key_val(skb, &key->ct_state, TCA_FLOWER_KEY_CT_STATE,
2766                             &mask->ct_state, TCA_FLOWER_KEY_CT_STATE_MASK,
2767                             sizeof(key->ct_state)))
2768                 goto nla_put_failure;
2769
2770         if (IS_ENABLED(CONFIG_NF_CONNTRACK_ZONES) &&
2771             fl_dump_key_val(skb, &key->ct_zone, TCA_FLOWER_KEY_CT_ZONE,
2772                             &mask->ct_zone, TCA_FLOWER_KEY_CT_ZONE_MASK,
2773                             sizeof(key->ct_zone)))
2774                 goto nla_put_failure;
2775
2776         if (IS_ENABLED(CONFIG_NF_CONNTRACK_MARK) &&
2777             fl_dump_key_val(skb, &key->ct_mark, TCA_FLOWER_KEY_CT_MARK,
2778                             &mask->ct_mark, TCA_FLOWER_KEY_CT_MARK_MASK,
2779                             sizeof(key->ct_mark)))
2780                 goto nla_put_failure;
2781
2782         if (IS_ENABLED(CONFIG_NF_CONNTRACK_LABELS) &&
2783             fl_dump_key_val(skb, &key->ct_labels, TCA_FLOWER_KEY_CT_LABELS,
2784                             &mask->ct_labels, TCA_FLOWER_KEY_CT_LABELS_MASK,
2785                             sizeof(key->ct_labels)))
2786                 goto nla_put_failure;
2787
2788         return 0;
2789
2790 nla_put_failure:
2791         return -EMSGSIZE;
2792 }
2793
2794 static int fl_dump_key_options(struct sk_buff *skb, int enc_opt_type,
2795                                struct flow_dissector_key_enc_opts *enc_opts)
2796 {
2797         struct nlattr *nest;
2798         int err;
2799
2800         if (!enc_opts->len)
2801                 return 0;
2802
2803         nest = nla_nest_start_noflag(skb, enc_opt_type);
2804         if (!nest)
2805                 goto nla_put_failure;
2806
2807         switch (enc_opts->dst_opt_type) {
2808         case TUNNEL_GENEVE_OPT:
2809                 err = fl_dump_key_geneve_opt(skb, enc_opts);
2810                 if (err)
2811                         goto nla_put_failure;
2812                 break;
2813         case TUNNEL_VXLAN_OPT:
2814                 err = fl_dump_key_vxlan_opt(skb, enc_opts);
2815                 if (err)
2816                         goto nla_put_failure;
2817                 break;
2818         case TUNNEL_ERSPAN_OPT:
2819                 err = fl_dump_key_erspan_opt(skb, enc_opts);
2820                 if (err)
2821                         goto nla_put_failure;
2822                 break;
2823         default:
2824                 goto nla_put_failure;
2825         }
2826         nla_nest_end(skb, nest);
2827         return 0;
2828
2829 nla_put_failure:
2830         nla_nest_cancel(skb, nest);
2831         return -EMSGSIZE;
2832 }
2833
2834 static int fl_dump_key_enc_opt(struct sk_buff *skb,
2835                                struct flow_dissector_key_enc_opts *key_opts,
2836                                struct flow_dissector_key_enc_opts *msk_opts)
2837 {
2838         int err;
2839
2840         err = fl_dump_key_options(skb, TCA_FLOWER_KEY_ENC_OPTS, key_opts);
2841         if (err)
2842                 return err;
2843
2844         return fl_dump_key_options(skb, TCA_FLOWER_KEY_ENC_OPTS_MASK, msk_opts);
2845 }
2846
2847 static int fl_dump_key(struct sk_buff *skb, struct net *net,
2848                        struct fl_flow_key *key, struct fl_flow_key *mask)
2849 {
2850         if (mask->meta.ingress_ifindex) {
2851                 struct net_device *dev;
2852
2853                 dev = __dev_get_by_index(net, key->meta.ingress_ifindex);
2854                 if (dev && nla_put_string(skb, TCA_FLOWER_INDEV, dev->name))
2855                         goto nla_put_failure;
2856         }
2857
2858         if (fl_dump_key_val(skb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
2859                             mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
2860                             sizeof(key->eth.dst)) ||
2861             fl_dump_key_val(skb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
2862                             mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
2863                             sizeof(key->eth.src)) ||
2864             fl_dump_key_val(skb, &key->basic.n_proto, TCA_FLOWER_KEY_ETH_TYPE,
2865                             &mask->basic.n_proto, TCA_FLOWER_UNSPEC,
2866                             sizeof(key->basic.n_proto)))
2867                 goto nla_put_failure;
2868
2869         if (fl_dump_key_mpls(skb, &key->mpls, &mask->mpls))
2870                 goto nla_put_failure;
2871
2872         if (fl_dump_key_vlan(skb, TCA_FLOWER_KEY_VLAN_ID,
2873                              TCA_FLOWER_KEY_VLAN_PRIO, &key->vlan, &mask->vlan))
2874                 goto nla_put_failure;
2875
2876         if (fl_dump_key_vlan(skb, TCA_FLOWER_KEY_CVLAN_ID,
2877                              TCA_FLOWER_KEY_CVLAN_PRIO,
2878                              &key->cvlan, &mask->cvlan) ||
2879             (mask->cvlan.vlan_tpid &&
2880              nla_put_be16(skb, TCA_FLOWER_KEY_VLAN_ETH_TYPE,
2881                           key->cvlan.vlan_tpid)))
2882                 goto nla_put_failure;
2883
2884         if (mask->basic.n_proto) {
2885                 if (mask->cvlan.vlan_eth_type) {
2886                         if (nla_put_be16(skb, TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
2887                                          key->basic.n_proto))
2888                                 goto nla_put_failure;
2889                 } else if (mask->vlan.vlan_eth_type) {
2890                         if (nla_put_be16(skb, TCA_FLOWER_KEY_VLAN_ETH_TYPE,
2891                                          key->vlan.vlan_eth_type))
2892                                 goto nla_put_failure;
2893                 }
2894         }
2895
2896         if ((key->basic.n_proto == htons(ETH_P_IP) ||
2897              key->basic.n_proto == htons(ETH_P_IPV6)) &&
2898             (fl_dump_key_val(skb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
2899                             &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
2900                             sizeof(key->basic.ip_proto)) ||
2901             fl_dump_key_ip(skb, false, &key->ip, &mask->ip)))
2902                 goto nla_put_failure;
2903
2904         if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
2905             (fl_dump_key_val(skb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
2906                              &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
2907                              sizeof(key->ipv4.src)) ||
2908              fl_dump_key_val(skb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
2909                              &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
2910                              sizeof(key->ipv4.dst))))
2911                 goto nla_put_failure;
2912         else if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
2913                  (fl_dump_key_val(skb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
2914                                   &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
2915                                   sizeof(key->ipv6.src)) ||
2916                   fl_dump_key_val(skb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
2917                                   &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
2918                                   sizeof(key->ipv6.dst))))
2919                 goto nla_put_failure;
2920
2921         if (key->basic.ip_proto == IPPROTO_TCP &&
2922             (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
2923                              &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
2924                              sizeof(key->tp.src)) ||
2925              fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
2926                              &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
2927                              sizeof(key->tp.dst)) ||
2928              fl_dump_key_val(skb, &key->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS,
2929                              &mask->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS_MASK,
2930                              sizeof(key->tcp.flags))))
2931                 goto nla_put_failure;
2932         else if (key->basic.ip_proto == IPPROTO_UDP &&
2933                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
2934                                   &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
2935                                   sizeof(key->tp.src)) ||
2936                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
2937                                   &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
2938                                   sizeof(key->tp.dst))))
2939                 goto nla_put_failure;
2940         else if (key->basic.ip_proto == IPPROTO_SCTP &&
2941                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_SCTP_SRC,
2942                                   &mask->tp.src, TCA_FLOWER_KEY_SCTP_SRC_MASK,
2943                                   sizeof(key->tp.src)) ||
2944                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_SCTP_DST,
2945                                   &mask->tp.dst, TCA_FLOWER_KEY_SCTP_DST_MASK,
2946                                   sizeof(key->tp.dst))))
2947                 goto nla_put_failure;
2948         else if (key->basic.n_proto == htons(ETH_P_IP) &&
2949                  key->basic.ip_proto == IPPROTO_ICMP &&
2950                  (fl_dump_key_val(skb, &key->icmp.type,
2951                                   TCA_FLOWER_KEY_ICMPV4_TYPE, &mask->icmp.type,
2952                                   TCA_FLOWER_KEY_ICMPV4_TYPE_MASK,
2953                                   sizeof(key->icmp.type)) ||
2954                   fl_dump_key_val(skb, &key->icmp.code,
2955                                   TCA_FLOWER_KEY_ICMPV4_CODE, &mask->icmp.code,
2956                                   TCA_FLOWER_KEY_ICMPV4_CODE_MASK,
2957                                   sizeof(key->icmp.code))))
2958                 goto nla_put_failure;
2959         else if (key->basic.n_proto == htons(ETH_P_IPV6) &&
2960                  key->basic.ip_proto == IPPROTO_ICMPV6 &&
2961                  (fl_dump_key_val(skb, &key->icmp.type,
2962                                   TCA_FLOWER_KEY_ICMPV6_TYPE, &mask->icmp.type,
2963                                   TCA_FLOWER_KEY_ICMPV6_TYPE_MASK,
2964                                   sizeof(key->icmp.type)) ||
2965                   fl_dump_key_val(skb, &key->icmp.code,
2966                                   TCA_FLOWER_KEY_ICMPV6_CODE, &mask->icmp.code,
2967                                   TCA_FLOWER_KEY_ICMPV6_CODE_MASK,
2968                                   sizeof(key->icmp.code))))
2969                 goto nla_put_failure;
2970         else if ((key->basic.n_proto == htons(ETH_P_ARP) ||
2971                   key->basic.n_proto == htons(ETH_P_RARP)) &&
2972                  (fl_dump_key_val(skb, &key->arp.sip,
2973                                   TCA_FLOWER_KEY_ARP_SIP, &mask->arp.sip,
2974                                   TCA_FLOWER_KEY_ARP_SIP_MASK,
2975                                   sizeof(key->arp.sip)) ||
2976                   fl_dump_key_val(skb, &key->arp.tip,
2977                                   TCA_FLOWER_KEY_ARP_TIP, &mask->arp.tip,
2978                                   TCA_FLOWER_KEY_ARP_TIP_MASK,
2979                                   sizeof(key->arp.tip)) ||
2980                   fl_dump_key_val(skb, &key->arp.op,
2981                                   TCA_FLOWER_KEY_ARP_OP, &mask->arp.op,
2982                                   TCA_FLOWER_KEY_ARP_OP_MASK,
2983                                   sizeof(key->arp.op)) ||
2984                   fl_dump_key_val(skb, key->arp.sha, TCA_FLOWER_KEY_ARP_SHA,
2985                                   mask->arp.sha, TCA_FLOWER_KEY_ARP_SHA_MASK,
2986                                   sizeof(key->arp.sha)) ||
2987                   fl_dump_key_val(skb, key->arp.tha, TCA_FLOWER_KEY_ARP_THA,
2988                                   mask->arp.tha, TCA_FLOWER_KEY_ARP_THA_MASK,
2989                                   sizeof(key->arp.tha))))
2990                 goto nla_put_failure;
2991
2992         if ((key->basic.ip_proto == IPPROTO_TCP ||
2993              key->basic.ip_proto == IPPROTO_UDP ||
2994              key->basic.ip_proto == IPPROTO_SCTP) &&
2995              fl_dump_key_port_range(skb, key, mask))
2996                 goto nla_put_failure;
2997
2998         if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
2999             (fl_dump_key_val(skb, &key->enc_ipv4.src,
3000                             TCA_FLOWER_KEY_ENC_IPV4_SRC, &mask->enc_ipv4.src,
3001                             TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
3002                             sizeof(key->enc_ipv4.src)) ||
3003              fl_dump_key_val(skb, &key->enc_ipv4.dst,
3004                              TCA_FLOWER_KEY_ENC_IPV4_DST, &mask->enc_ipv4.dst,
3005                              TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
3006                              sizeof(key->enc_ipv4.dst))))
3007                 goto nla_put_failure;
3008         else if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
3009                  (fl_dump_key_val(skb, &key->enc_ipv6.src,
3010                             TCA_FLOWER_KEY_ENC_IPV6_SRC, &mask->enc_ipv6.src,
3011                             TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
3012                             sizeof(key->enc_ipv6.src)) ||
3013                  fl_dump_key_val(skb, &key->enc_ipv6.dst,
3014                                  TCA_FLOWER_KEY_ENC_IPV6_DST,
3015                                  &mask->enc_ipv6.dst,
3016                                  TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
3017                             sizeof(key->enc_ipv6.dst))))
3018                 goto nla_put_failure;
3019
3020         if (fl_dump_key_val(skb, &key->enc_key_id, TCA_FLOWER_KEY_ENC_KEY_ID,
3021                             &mask->enc_key_id, TCA_FLOWER_UNSPEC,
3022                             sizeof(key->enc_key_id)) ||
3023             fl_dump_key_val(skb, &key->enc_tp.src,
3024                             TCA_FLOWER_KEY_ENC_UDP_SRC_PORT,
3025                             &mask->enc_tp.src,
3026                             TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK,
3027                             sizeof(key->enc_tp.src)) ||
3028             fl_dump_key_val(skb, &key->enc_tp.dst,
3029                             TCA_FLOWER_KEY_ENC_UDP_DST_PORT,
3030                             &mask->enc_tp.dst,
3031                             TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK,
3032                             sizeof(key->enc_tp.dst)) ||
3033             fl_dump_key_ip(skb, true, &key->enc_ip, &mask->enc_ip) ||
3034             fl_dump_key_enc_opt(skb, &key->enc_opts, &mask->enc_opts))
3035                 goto nla_put_failure;
3036
3037         if (fl_dump_key_ct(skb, &key->ct, &mask->ct))
3038                 goto nla_put_failure;
3039
3040         if (fl_dump_key_flags(skb, key->control.flags, mask->control.flags))
3041                 goto nla_put_failure;
3042
3043         if (fl_dump_key_val(skb, &key->hash.hash, TCA_FLOWER_KEY_HASH,
3044                              &mask->hash.hash, TCA_FLOWER_KEY_HASH_MASK,
3045                              sizeof(key->hash.hash)))
3046                 goto nla_put_failure;
3047
3048         return 0;
3049
3050 nla_put_failure:
3051         return -EMSGSIZE;
3052 }
3053
3054 static int fl_dump(struct net *net, struct tcf_proto *tp, void *fh,
3055                    struct sk_buff *skb, struct tcmsg *t, bool rtnl_held)
3056 {
3057         struct cls_fl_filter *f = fh;
3058         struct nlattr *nest;
3059         struct fl_flow_key *key, *mask;
3060         bool skip_hw;
3061
3062         if (!f)
3063                 return skb->len;
3064
3065         t->tcm_handle = f->handle;
3066
3067         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
3068         if (!nest)
3069                 goto nla_put_failure;
3070
3071         spin_lock(&tp->lock);
3072
3073         if (f->res.classid &&
3074             nla_put_u32(skb, TCA_FLOWER_CLASSID, f->res.classid))
3075                 goto nla_put_failure_locked;
3076
3077         key = &f->key;
3078         mask = &f->mask->key;
3079         skip_hw = tc_skip_hw(f->flags);
3080
3081         if (fl_dump_key(skb, net, key, mask))
3082                 goto nla_put_failure_locked;
3083
3084         if (f->flags && nla_put_u32(skb, TCA_FLOWER_FLAGS, f->flags))
3085                 goto nla_put_failure_locked;
3086
3087         spin_unlock(&tp->lock);
3088
3089         if (!skip_hw)
3090                 fl_hw_update_stats(tp, f, rtnl_held);
3091
3092         if (nla_put_u32(skb, TCA_FLOWER_IN_HW_COUNT, f->in_hw_count))
3093                 goto nla_put_failure;
3094
3095         if (tcf_exts_dump(skb, &f->exts))
3096                 goto nla_put_failure;
3097
3098         nla_nest_end(skb, nest);
3099
3100         if (tcf_exts_dump_stats(skb, &f->exts) < 0)
3101                 goto nla_put_failure;
3102
3103         return skb->len;
3104
3105 nla_put_failure_locked:
3106         spin_unlock(&tp->lock);
3107 nla_put_failure:
3108         nla_nest_cancel(skb, nest);
3109         return -1;
3110 }
3111
3112 static int fl_terse_dump(struct net *net, struct tcf_proto *tp, void *fh,
3113                          struct sk_buff *skb, struct tcmsg *t, bool rtnl_held)
3114 {
3115         struct cls_fl_filter *f = fh;
3116         struct nlattr *nest;
3117         bool skip_hw;
3118
3119         if (!f)
3120                 return skb->len;
3121
3122         t->tcm_handle = f->handle;
3123
3124         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
3125         if (!nest)
3126                 goto nla_put_failure;
3127
3128         spin_lock(&tp->lock);
3129
3130         skip_hw = tc_skip_hw(f->flags);
3131
3132         if (f->flags && nla_put_u32(skb, TCA_FLOWER_FLAGS, f->flags))
3133                 goto nla_put_failure_locked;
3134
3135         spin_unlock(&tp->lock);
3136
3137         if (!skip_hw)
3138                 fl_hw_update_stats(tp, f, rtnl_held);
3139
3140         if (tcf_exts_terse_dump(skb, &f->exts))
3141                 goto nla_put_failure;
3142
3143         nla_nest_end(skb, nest);
3144
3145         return skb->len;
3146
3147 nla_put_failure_locked:
3148         spin_unlock(&tp->lock);
3149 nla_put_failure:
3150         nla_nest_cancel(skb, nest);
3151         return -1;
3152 }
3153
3154 static int fl_tmplt_dump(struct sk_buff *skb, struct net *net, void *tmplt_priv)
3155 {
3156         struct fl_flow_tmplt *tmplt = tmplt_priv;
3157         struct fl_flow_key *key, *mask;
3158         struct nlattr *nest;
3159
3160         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
3161         if (!nest)
3162                 goto nla_put_failure;
3163
3164         key = &tmplt->dummy_key;
3165         mask = &tmplt->mask;
3166
3167         if (fl_dump_key(skb, net, key, mask))
3168                 goto nla_put_failure;
3169
3170         nla_nest_end(skb, nest);
3171
3172         return skb->len;
3173
3174 nla_put_failure:
3175         nla_nest_cancel(skb, nest);
3176         return -EMSGSIZE;
3177 }
3178
3179 static void fl_bind_class(void *fh, u32 classid, unsigned long cl, void *q,
3180                           unsigned long base)
3181 {
3182         struct cls_fl_filter *f = fh;
3183
3184         if (f && f->res.classid == classid) {
3185                 if (cl)
3186                         __tcf_bind_filter(q, &f->res, base);
3187                 else
3188                         __tcf_unbind_filter(q, &f->res);
3189         }
3190 }
3191
3192 static bool fl_delete_empty(struct tcf_proto *tp)
3193 {
3194         struct cls_fl_head *head = fl_head_dereference(tp);
3195
3196         spin_lock(&tp->lock);
3197         tp->deleting = idr_is_empty(&head->handle_idr);
3198         spin_unlock(&tp->lock);
3199
3200         return tp->deleting;
3201 }
3202
3203 static struct tcf_proto_ops cls_fl_ops __read_mostly = {
3204         .kind           = "flower",
3205         .classify       = fl_classify,
3206         .init           = fl_init,
3207         .destroy        = fl_destroy,
3208         .get            = fl_get,
3209         .put            = fl_put,
3210         .change         = fl_change,
3211         .delete         = fl_delete,
3212         .delete_empty   = fl_delete_empty,
3213         .walk           = fl_walk,
3214         .reoffload      = fl_reoffload,
3215         .hw_add         = fl_hw_add,
3216         .hw_del         = fl_hw_del,
3217         .dump           = fl_dump,
3218         .terse_dump     = fl_terse_dump,
3219         .bind_class     = fl_bind_class,
3220         .tmplt_create   = fl_tmplt_create,
3221         .tmplt_destroy  = fl_tmplt_destroy,
3222         .tmplt_dump     = fl_tmplt_dump,
3223         .owner          = THIS_MODULE,
3224         .flags          = TCF_PROTO_OPS_DOIT_UNLOCKED,
3225 };
3226
3227 static int __init cls_fl_init(void)
3228 {
3229         return register_tcf_proto_ops(&cls_fl_ops);
3230 }
3231
3232 static void __exit cls_fl_exit(void)
3233 {
3234         unregister_tcf_proto_ops(&cls_fl_ops);
3235 }
3236
3237 module_init(cls_fl_init);
3238 module_exit(cls_fl_exit);
3239
3240 MODULE_AUTHOR("Jiri Pirko <jiri@resnulli.us>");
3241 MODULE_DESCRIPTION("Flower classifier");
3242 MODULE_LICENSE("GPL v2");