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
  26 #include <net/dst.h>
  27 #include <net/dst_metadata.h>
  28
  29 struct fl_flow_key {
  30         int     indev_ifindex;
  31         struct flow_dissector_key_control control;
  32         struct flow_dissector_key_control enc_control;
  33         struct flow_dissector_key_basic basic;
  34         struct flow_dissector_key_eth_addrs eth;
  35         struct flow_dissector_key_vlan vlan;
  36         struct flow_dissector_key_vlan cvlan;
  37         union {
  38                 struct flow_dissector_key_ipv4_addrs ipv4;
  39                 struct flow_dissector_key_ipv6_addrs ipv6;
  40         };
  41         struct flow_dissector_key_ports tp;
  42         struct flow_dissector_key_icmp icmp;
  43         struct flow_dissector_key_arp arp;
  44         struct flow_dissector_key_keyid enc_key_id;
  45         union {
  46                 struct flow_dissector_key_ipv4_addrs enc_ipv4;
  47                 struct flow_dissector_key_ipv6_addrs enc_ipv6;
  48         };
  49         struct flow_dissector_key_ports enc_tp;
  50         struct flow_dissector_key_mpls mpls;
  51         struct flow_dissector_key_tcp tcp;
  52         struct flow_dissector_key_ip ip;
  53         struct flow_dissector_key_ip enc_ip;
  54         struct flow_dissector_key_enc_opts enc_opts;
  55         struct flow_dissector_key_ports tp_min;
  56         struct flow_dissector_key_ports tp_max;
  57 } __aligned(BITS_PER_LONG / 8); /* Ensure that we can do comparisons as longs. */
  58
  59 struct fl_flow_mask_range {
  60         unsigned short int start;
  61         unsigned short int end;
  62 };
  63
  64 struct fl_flow_mask {
  65         struct fl_flow_key key;
  66         struct fl_flow_mask_range range;
  67         u32 flags;
  68         struct rhash_head ht_node;
  69         struct rhashtable ht;
  70         struct rhashtable_params filter_ht_params;
  71         struct flow_dissector dissector;
  72         struct list_head filters;
  73         struct rcu_work rwork;
  74         struct list_head list;
  75         refcount_t refcnt;
  76 };
  77
  78 struct fl_flow_tmplt {
  79         struct fl_flow_key dummy_key;
  80         struct fl_flow_key mask;
  81         struct flow_dissector dissector;
  82         struct tcf_chain *chain;
  83 };
  84
  85 struct cls_fl_head {
  86         struct rhashtable ht;
  87         spinlock_t masks_lock; /* Protect masks list */
  88         struct list_head masks;
  89         struct list_head hw_filters;
  90         struct rcu_work rwork;
  91         struct idr handle_idr;
  92 };
  93
  94 struct cls_fl_filter {
  95         struct fl_flow_mask *mask;
  96         struct rhash_head ht_node;
  97         struct fl_flow_key mkey;
  98         struct tcf_exts exts;
  99         struct tcf_result res;
 100         struct fl_flow_key key;
 101         struct list_head list;
 102         struct list_head hw_list;
 103         u32 handle;
 104         u32 flags;
 105         u32 in_hw_count;
 106         struct rcu_work rwork;
 107         struct net_device *hw_dev;
 108         /* Flower classifier is unlocked, which means that its reference counter
 109          * can be changed concurrently without any kind of external
 110          * synchronization. Use atomic reference counter to be concurrency-safe.
 111          */
 112         refcount_t refcnt;
 113         bool deleted;
 114 };
 115
 116 static const struct rhashtable_params mask_ht_params = {
 117         .key_offset = offsetof(struct fl_flow_mask, key),
 118         .key_len = sizeof(struct fl_flow_key),
 119         .head_offset = offsetof(struct fl_flow_mask, ht_node),
 120         .automatic_shrinking = true,
 121 };
 122
 123 static unsigned short int fl_mask_range(const struct fl_flow_mask *mask)
 124 {
 125         return mask->range.end - mask->range.start;
 126 }
 127
 128 static void fl_mask_update_range(struct fl_flow_mask *mask)
 129 {
 130         const u8 *bytes = (const u8 *) &mask->key;
 131         size_t size = sizeof(mask->key);
 132         size_t i, first = 0, last;
 133
 134         for (i = 0; i < size; i++) {
 135                 if (bytes[i]) {
 136                         first = i;
 137                         break;
 138                 }
 139         }
 140         last = first;
 141         for (i = size - 1; i != first; i--) {
 142                 if (bytes[i]) {
 143                         last = i;
 144                         break;
 145                 }
 146         }
 147         mask->range.start = rounddown(first, sizeof(long));
 148         mask->range.end = roundup(last + 1, sizeof(long));
 149 }
 150
 151 static void *fl_key_get_start(struct fl_flow_key *key,
 152                               const struct fl_flow_mask *mask)
 153 {
 154         return (u8 *) key + mask->range.start;
 155 }
 156
 157 static void fl_set_masked_key(struct fl_flow_key *mkey, struct fl_flow_key *key,
 158                               struct fl_flow_mask *mask)
 159 {
 160         const long *lkey = fl_key_get_start(key, mask);
 161         const long *lmask = fl_key_get_start(&mask->key, mask);
 162         long *lmkey = fl_key_get_start(mkey, mask);
 163         int i;
 164
 165         for (i = 0; i < fl_mask_range(mask); i += sizeof(long))
 166                 *lmkey++ = *lkey++ & *lmask++;
 167 }
 168
 169 static bool fl_mask_fits_tmplt(struct fl_flow_tmplt *tmplt,
 170                                struct fl_flow_mask *mask)
 171 {
 172         const long *lmask = fl_key_get_start(&mask->key, mask);
 173         const long *ltmplt;
 174         int i;
 175
 176         if (!tmplt)
 177                 return true;
 178         ltmplt = fl_key_get_start(&tmplt->mask, mask);
 179         for (i = 0; i < fl_mask_range(mask); i += sizeof(long)) {
 180                 if (~*ltmplt++ & *lmask++)
 181                         return false;
 182         }
 183         return true;
 184 }
 185
 186 static void fl_clear_masked_range(struct fl_flow_key *key,
 187                                   struct fl_flow_mask *mask)
 188 {
 189         memset(fl_key_get_start(key, mask), 0, fl_mask_range(mask));
 190 }
 191
 192 static bool fl_range_port_dst_cmp(struct cls_fl_filter *filter,
 193                                   struct fl_flow_key *key,
 194                                   struct fl_flow_key *mkey)
 195 {
 196         __be16 min_mask, max_mask, min_val, max_val;
 197
 198         min_mask = htons(filter->mask->key.tp_min.dst);
 199         max_mask = htons(filter->mask->key.tp_max.dst);
 200         min_val = htons(filter->key.tp_min.dst);
 201         max_val = htons(filter->key.tp_max.dst);
 202
 203         if (min_mask && max_mask) {
 204                 if (htons(key->tp.dst) < min_val ||
 205                     htons(key->tp.dst) > max_val)
 206                         return false;
 207
 208                 /* skb does not have min and max values */
 209                 mkey->tp_min.dst = filter->mkey.tp_min.dst;
 210                 mkey->tp_max.dst = filter->mkey.tp_max.dst;
 211         }
 212         return true;
 213 }
 214
 215 static bool fl_range_port_src_cmp(struct cls_fl_filter *filter,
 216                                   struct fl_flow_key *key,
 217                                   struct fl_flow_key *mkey)
 218 {
 219         __be16 min_mask, max_mask, min_val, max_val;
 220
 221         min_mask = htons(filter->mask->key.tp_min.src);
 222         max_mask = htons(filter->mask->key.tp_max.src);
 223         min_val = htons(filter->key.tp_min.src);
 224         max_val = htons(filter->key.tp_max.src);
 225
 226         if (min_mask && max_mask) {
 227                 if (htons(key->tp.src) < min_val ||
 228                     htons(key->tp.src) > max_val)
 229                         return false;
 230
 231                 /* skb does not have min and max values */
 232                 mkey->tp_min.src = filter->mkey.tp_min.src;
 233                 mkey->tp_max.src = filter->mkey.tp_max.src;
 234         }
 235         return true;
 236 }
 237
 238 static struct cls_fl_filter *__fl_lookup(struct fl_flow_mask *mask,
 239                                          struct fl_flow_key *mkey)
 240 {
 241         return rhashtable_lookup_fast(&mask->ht, fl_key_get_start(mkey, mask),
 242                                       mask->filter_ht_params);
 243 }
 244
 245 static struct cls_fl_filter *fl_lookup_range(struct fl_flow_mask *mask,
 246                                              struct fl_flow_key *mkey,
 247                                              struct fl_flow_key *key)
 248 {
 249         struct cls_fl_filter *filter, *f;
 250
 251         list_for_each_entry_rcu(filter, &mask->filters, list) {
 252                 if (!fl_range_port_dst_cmp(filter, key, mkey))
 253                         continue;
 254
 255                 if (!fl_range_port_src_cmp(filter, key, mkey))
 256                         continue;
 257
 258                 f = __fl_lookup(mask, mkey);
 259                 if (f)
 260                         return f;
 261         }
 262         return NULL;
 263 }
 264
 265 static struct cls_fl_filter *fl_lookup(struct fl_flow_mask *mask,
 266                                        struct fl_flow_key *mkey,
 267                                        struct fl_flow_key *key)
 268 {
 269         if ((mask->flags & TCA_FLOWER_MASK_FLAGS_RANGE))
 270                 return fl_lookup_range(mask, mkey, key);
 271
 272         return __fl_lookup(mask, mkey);
 273 }
 274
 275 static int fl_classify(struct sk_buff *skb, const struct tcf_proto *tp,
 276                        struct tcf_result *res)
 277 {
 278         struct cls_fl_head *head = rcu_dereference_bh(tp->root);
 279         struct cls_fl_filter *f;
 280         struct fl_flow_mask *mask;
 281         struct fl_flow_key skb_key;
 282         struct fl_flow_key skb_mkey;
 283
 284         list_for_each_entry_rcu(mask, &head->masks, list) {
 285                 fl_clear_masked_range(&skb_key, mask);
 286
 287                 skb_key.indev_ifindex = skb->skb_iif;
 288                 /* skb_flow_dissect() does not set n_proto in case an unknown
 289                  * protocol, so do it rather here.
 290                  */
 291                 skb_key.basic.n_proto = skb->protocol;
 292                 skb_flow_dissect_tunnel_info(skb, &mask->dissector, &skb_key);
 293                 skb_flow_dissect(skb, &mask->dissector, &skb_key, 0);
 294
 295                 fl_set_masked_key(&skb_mkey, &skb_key, mask);
 296
 297                 f = fl_lookup(mask, &skb_mkey, &skb_key);
 298                 if (f && !tc_skip_sw(f->flags)) {
 299                         *res = f->res;
 300                         return tcf_exts_exec(skb, &f->exts, res);
 301                 }
 302         }
 303         return -1;
 304 }
 305
 306 static int fl_init(struct tcf_proto *tp)
 307 {
 308         struct cls_fl_head *head;
 309
 310         head = kzalloc(sizeof(*head), GFP_KERNEL);
 311         if (!head)
 312                 return -ENOBUFS;
 313
 314         spin_lock_init(&head->masks_lock);
 315         INIT_LIST_HEAD_RCU(&head->masks);
 316         INIT_LIST_HEAD(&head->hw_filters);
 317         rcu_assign_pointer(tp->root, head);
 318         idr_init(&head->handle_idr);
 319
 320         return rhashtable_init(&head->ht, &mask_ht_params);
 321 }
 322
 323 static void fl_mask_free(struct fl_flow_mask *mask)
 324 {
 325         WARN_ON(!list_empty(&mask->filters));
 326         rhashtable_destroy(&mask->ht);
 327         kfree(mask);
 328 }
 329
 330 static void fl_mask_free_work(struct work_struct *work)
 331 {
 332         struct fl_flow_mask *mask = container_of(to_rcu_work(work),
 333                                                  struct fl_flow_mask, rwork);
 334
 335         fl_mask_free(mask);
 336 }
 337
 338 static bool fl_mask_put(struct cls_fl_head *head, struct fl_flow_mask *mask)
 339 {
 340         if (!refcount_dec_and_test(&mask->refcnt))
 341                 return false;
 342
 343         rhashtable_remove_fast(&head->ht, &mask->ht_node, mask_ht_params);
 344
 345         spin_lock(&head->masks_lock);
 346         list_del_rcu(&mask->list);
 347         spin_unlock(&head->masks_lock);
 348
 349         tcf_queue_work(&mask->rwork, fl_mask_free_work);
 350
 351         return true;
 352 }
 353
 354 static struct cls_fl_head *fl_head_dereference(struct tcf_proto *tp)
 355 {
 356         /* Flower classifier only changes root pointer during init and destroy.
 357          * Users must obtain reference to tcf_proto instance before calling its
 358          * API, so tp->root pointer is protected from concurrent call to
 359          * fl_destroy() by reference counting.
 360          */
 361         return rcu_dereference_raw(tp->root);
 362 }
 363
 364 static void __fl_destroy_filter(struct cls_fl_filter *f)
 365 {
 366         tcf_exts_destroy(&f->exts);
 367         tcf_exts_put_net(&f->exts);
 368         kfree(f);
 369 }
 370
 371 static void fl_destroy_filter_work(struct work_struct *work)
 372 {
 373         struct cls_fl_filter *f = container_of(to_rcu_work(work),
 374                                         struct cls_fl_filter, rwork);
 375
 376         __fl_destroy_filter(f);
 377 }
 378
 379 static void fl_hw_destroy_filter(struct tcf_proto *tp, struct cls_fl_filter *f,
 380                                  bool rtnl_held, struct netlink_ext_ack *extack)
 381 {
 382         struct tc_cls_flower_offload cls_flower = {};
 383         struct tcf_block *block = tp->chain->block;
 384
 385         if (!rtnl_held)
 386                 rtnl_lock();
 387
 388         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, extack);
 389         cls_flower.command = TC_CLSFLOWER_DESTROY;
 390         cls_flower.cookie = (unsigned long) f;
 391
 392         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false);
 393         spin_lock(&tp->lock);
 394         list_del_init(&f->hw_list);
 395         tcf_block_offload_dec(block, &f->flags);
 396         spin_unlock(&tp->lock);
 397
 398         if (!rtnl_held)
 399                 rtnl_unlock();
 400 }
 401
 402 static int fl_hw_replace_filter(struct tcf_proto *tp,
 403                                 struct cls_fl_filter *f, bool rtnl_held,
 404                                 struct netlink_ext_ack *extack)
 405 {
 406         struct cls_fl_head *head = fl_head_dereference(tp);
 407         struct tc_cls_flower_offload cls_flower = {};
 408         struct tcf_block *block = tp->chain->block;
 409         bool skip_sw = tc_skip_sw(f->flags);
 410         int err = 0;
 411
 412         if (!rtnl_held)
 413                 rtnl_lock();
 414
 415         cls_flower.rule = flow_rule_alloc(tcf_exts_num_actions(&f->exts));
 416         if (!cls_flower.rule) {
 417                 err = -ENOMEM;
 418                 goto errout;
 419         }
 420
 421         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, extack);
 422         cls_flower.command = TC_CLSFLOWER_REPLACE;
 423         cls_flower.cookie = (unsigned long) f;
 424         cls_flower.rule->match.dissector = &f->mask->dissector;
 425         cls_flower.rule->match.mask = &f->mask->key;
 426         cls_flower.rule->match.key = &f->mkey;
 427         cls_flower.classid = f->res.classid;
 428
 429         err = tc_setup_flow_action(&cls_flower.rule->action, &f->exts);
 430         if (err) {
 431                 kfree(cls_flower.rule);
 432                 if (skip_sw)
 433                         NL_SET_ERR_MSG_MOD(extack, "Failed to setup flow action");
 434                 else
 435                         err = 0;
 436                 goto errout;
 437         }
 438
 439         err = tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, skip_sw);
 440         kfree(cls_flower.rule);
 441
 442         if (err < 0) {
 443                 fl_hw_destroy_filter(tp, f, true, NULL);
 444                 goto errout;
 445         } else if (err > 0) {
 446                 f->in_hw_count = err;
 447                 err = 0;
 448                 spin_lock(&tp->lock);
 449                 tcf_block_offload_inc(block, &f->flags);
 450                 spin_unlock(&tp->lock);
 451         }
 452
 453         if (skip_sw && !(f->flags & TCA_CLS_FLAGS_IN_HW)) {
 454                 err = -EINVAL;
 455                 goto errout;
 456         }
 457
 458         spin_lock(&tp->lock);
 459         list_add(&f->hw_list, &head->hw_filters);
 460         spin_unlock(&tp->lock);
 461 errout:
 462         if (!rtnl_held)
 463                 rtnl_unlock();
 464
 465         return err;
 466 }
 467
 468 static void fl_hw_update_stats(struct tcf_proto *tp, struct cls_fl_filter *f,
 469                                bool rtnl_held)
 470 {
 471         struct tc_cls_flower_offload cls_flower = {};
 472         struct tcf_block *block = tp->chain->block;
 473
 474         if (!rtnl_held)
 475                 rtnl_lock();
 476
 477         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, NULL);
 478         cls_flower.command = TC_CLSFLOWER_STATS;
 479         cls_flower.cookie = (unsigned long) f;
 480         cls_flower.classid = f->res.classid;
 481
 482         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false);
 483
 484         tcf_exts_stats_update(&f->exts, cls_flower.stats.bytes,
 485                               cls_flower.stats.pkts,
 486                               cls_flower.stats.lastused);
 487
 488         if (!rtnl_held)
 489                 rtnl_unlock();
 490 }
 491
 492 static void __fl_put(struct cls_fl_filter *f)
 493 {
 494         if (!refcount_dec_and_test(&f->refcnt))
 495                 return;
 496
 497         if (tcf_exts_get_net(&f->exts))
 498                 tcf_queue_work(&f->rwork, fl_destroy_filter_work);
 499         else
 500                 __fl_destroy_filter(f);
 501 }
 502
 503 static struct cls_fl_filter *__fl_get(struct cls_fl_head *head, u32 handle)
 504 {
 505         struct cls_fl_filter *f;
 506
 507         rcu_read_lock();
 508         f = idr_find(&head->handle_idr, handle);
 509         if (f && !refcount_inc_not_zero(&f->refcnt))
 510                 f = NULL;
 511         rcu_read_unlock();
 512
 513         return f;
 514 }
 515
 516 static struct cls_fl_filter *fl_get_next_filter(struct tcf_proto *tp,
 517                                                 unsigned long *handle)
 518 {
 519         struct cls_fl_head *head = fl_head_dereference(tp);
 520         struct cls_fl_filter *f;
 521
 522         rcu_read_lock();
 523         while ((f = idr_get_next_ul(&head->handle_idr, handle))) {
 524                 /* don't return filters that are being deleted */
 525                 if (refcount_inc_not_zero(&f->refcnt))
 526                         break;
 527                 ++(*handle);
 528         }
 529         rcu_read_unlock();
 530
 531         return f;
 532 }
 533
 534 static int __fl_delete(struct tcf_proto *tp, struct cls_fl_filter *f,
 535                        bool *last, bool rtnl_held,
 536                        struct netlink_ext_ack *extack)
 537 {
 538         struct cls_fl_head *head = fl_head_dereference(tp);
 539
 540         *last = false;
 541
 542         spin_lock(&tp->lock);
 543         if (f->deleted) {
 544                 spin_unlock(&tp->lock);
 545                 return -ENOENT;
 546         }
 547
 548         f->deleted = true;
 549         rhashtable_remove_fast(&f->mask->ht, &f->ht_node,
 550                                f->mask->filter_ht_params);
 551         idr_remove(&head->handle_idr, f->handle);
 552         list_del_rcu(&f->list);
 553         spin_unlock(&tp->lock);
 554
 555         *last = fl_mask_put(head, f->mask);
 556         if (!tc_skip_hw(f->flags))
 557                 fl_hw_destroy_filter(tp, f, rtnl_held, extack);
 558         tcf_unbind_filter(tp, &f->res);
 559         __fl_put(f);
 560
 561         return 0;
 562 }
 563
 564 static void fl_destroy_sleepable(struct work_struct *work)
 565 {
 566         struct cls_fl_head *head = container_of(to_rcu_work(work),
 567                                                 struct cls_fl_head,
 568                                                 rwork);
 569
 570         rhashtable_destroy(&head->ht);
 571         kfree(head);
 572         module_put(THIS_MODULE);
 573 }
 574
 575 static void fl_destroy(struct tcf_proto *tp, bool rtnl_held,
 576                        struct netlink_ext_ack *extack)
 577 {
 578         struct cls_fl_head *head = fl_head_dereference(tp);
 579         struct fl_flow_mask *mask, *next_mask;
 580         struct cls_fl_filter *f, *next;
 581         bool last;
 582
 583         list_for_each_entry_safe(mask, next_mask, &head->masks, list) {
 584                 list_for_each_entry_safe(f, next, &mask->filters, list) {
 585                         __fl_delete(tp, f, &last, rtnl_held, extack);
 586                         if (last)
 587                                 break;
 588                 }
 589         }
 590         idr_destroy(&head->handle_idr);
 591
 592         __module_get(THIS_MODULE);
 593         tcf_queue_work(&head->rwork, fl_destroy_sleepable);
 594 }
 595
 596 static void fl_put(struct tcf_proto *tp, void *arg)
 597 {
 598         struct cls_fl_filter *f = arg;
 599
 600         __fl_put(f);
 601 }
 602
 603 static void *fl_get(struct tcf_proto *tp, u32 handle)
 604 {
 605         struct cls_fl_head *head = fl_head_dereference(tp);
 606
 607         return __fl_get(head, handle);
 608 }
 609
 610 static const struct nla_policy fl_policy[TCA_FLOWER_MAX + 1] = {
 611         [TCA_FLOWER_UNSPEC]             = { .type = NLA_UNSPEC },
 612         [TCA_FLOWER_CLASSID]            = { .type = NLA_U32 },
 613         [TCA_FLOWER_INDEV]              = { .type = NLA_STRING,
 614                                             .len = IFNAMSIZ },
 615         [TCA_FLOWER_KEY_ETH_DST]        = { .len = ETH_ALEN },
 616         [TCA_FLOWER_KEY_ETH_DST_MASK]   = { .len = ETH_ALEN },
 617         [TCA_FLOWER_KEY_ETH_SRC]        = { .len = ETH_ALEN },
 618         [TCA_FLOWER_KEY_ETH_SRC_MASK]   = { .len = ETH_ALEN },
 619         [TCA_FLOWER_KEY_ETH_TYPE]       = { .type = NLA_U16 },
 620         [TCA_FLOWER_KEY_IP_PROTO]       = { .type = NLA_U8 },
 621         [TCA_FLOWER_KEY_IPV4_SRC]       = { .type = NLA_U32 },
 622         [TCA_FLOWER_KEY_IPV4_SRC_MASK]  = { .type = NLA_U32 },
 623         [TCA_FLOWER_KEY_IPV4_DST]       = { .type = NLA_U32 },
 624         [TCA_FLOWER_KEY_IPV4_DST_MASK]  = { .type = NLA_U32 },
 625         [TCA_FLOWER_KEY_IPV6_SRC]       = { .len = sizeof(struct in6_addr) },
 626         [TCA_FLOWER_KEY_IPV6_SRC_MASK]  = { .len = sizeof(struct in6_addr) },
 627         [TCA_FLOWER_KEY_IPV6_DST]       = { .len = sizeof(struct in6_addr) },
 628         [TCA_FLOWER_KEY_IPV6_DST_MASK]  = { .len = sizeof(struct in6_addr) },
 629         [TCA_FLOWER_KEY_TCP_SRC]        = { .type = NLA_U16 },
 630         [TCA_FLOWER_KEY_TCP_DST]        = { .type = NLA_U16 },
 631         [TCA_FLOWER_KEY_UDP_SRC]        = { .type = NLA_U16 },
 632         [TCA_FLOWER_KEY_UDP_DST]        = { .type = NLA_U16 },
 633         [TCA_FLOWER_KEY_VLAN_ID]        = { .type = NLA_U16 },
 634         [TCA_FLOWER_KEY_VLAN_PRIO]      = { .type = NLA_U8 },
 635         [TCA_FLOWER_KEY_VLAN_ETH_TYPE]  = { .type = NLA_U16 },
 636         [TCA_FLOWER_KEY_ENC_KEY_ID]     = { .type = NLA_U32 },
 637         [TCA_FLOWER_KEY_ENC_IPV4_SRC]   = { .type = NLA_U32 },
 638         [TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK] = { .type = NLA_U32 },
 639         [TCA_FLOWER_KEY_ENC_IPV4_DST]   = { .type = NLA_U32 },
 640         [TCA_FLOWER_KEY_ENC_IPV4_DST_MASK] = { .type = NLA_U32 },
 641         [TCA_FLOWER_KEY_ENC_IPV6_SRC]   = { .len = sizeof(struct in6_addr) },
 642         [TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK] = { .len = sizeof(struct in6_addr) },
 643         [TCA_FLOWER_KEY_ENC_IPV6_DST]   = { .len = sizeof(struct in6_addr) },
 644         [TCA_FLOWER_KEY_ENC_IPV6_DST_MASK] = { .len = sizeof(struct in6_addr) },
 645         [TCA_FLOWER_KEY_TCP_SRC_MASK]   = { .type = NLA_U16 },
 646         [TCA_FLOWER_KEY_TCP_DST_MASK]   = { .type = NLA_U16 },
 647         [TCA_FLOWER_KEY_UDP_SRC_MASK]   = { .type = NLA_U16 },
 648         [TCA_FLOWER_KEY_UDP_DST_MASK]   = { .type = NLA_U16 },
 649         [TCA_FLOWER_KEY_SCTP_SRC_MASK]  = { .type = NLA_U16 },
 650         [TCA_FLOWER_KEY_SCTP_DST_MASK]  = { .type = NLA_U16 },
 651         [TCA_FLOWER_KEY_SCTP_SRC]       = { .type = NLA_U16 },
 652         [TCA_FLOWER_KEY_SCTP_DST]       = { .type = NLA_U16 },
 653         [TCA_FLOWER_KEY_ENC_UDP_SRC_PORT]       = { .type = NLA_U16 },
 654         [TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK]  = { .type = NLA_U16 },
 655         [TCA_FLOWER_KEY_ENC_UDP_DST_PORT]       = { .type = NLA_U16 },
 656         [TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK]  = { .type = NLA_U16 },
 657         [TCA_FLOWER_KEY_FLAGS]          = { .type = NLA_U32 },
 658         [TCA_FLOWER_KEY_FLAGS_MASK]     = { .type = NLA_U32 },
 659         [TCA_FLOWER_KEY_ICMPV4_TYPE]    = { .type = NLA_U8 },
 660         [TCA_FLOWER_KEY_ICMPV4_TYPE_MASK] = { .type = NLA_U8 },
 661         [TCA_FLOWER_KEY_ICMPV4_CODE]    = { .type = NLA_U8 },
 662         [TCA_FLOWER_KEY_ICMPV4_CODE_MASK] = { .type = NLA_U8 },
 663         [TCA_FLOWER_KEY_ICMPV6_TYPE]    = { .type = NLA_U8 },
 664         [TCA_FLOWER_KEY_ICMPV6_TYPE_MASK] = { .type = NLA_U8 },
 665         [TCA_FLOWER_KEY_ICMPV6_CODE]    = { .type = NLA_U8 },
 666         [TCA_FLOWER_KEY_ICMPV6_CODE_MASK] = { .type = NLA_U8 },
 667         [TCA_FLOWER_KEY_ARP_SIP]        = { .type = NLA_U32 },
 668         [TCA_FLOWER_KEY_ARP_SIP_MASK]   = { .type = NLA_U32 },
 669         [TCA_FLOWER_KEY_ARP_TIP]        = { .type = NLA_U32 },
 670         [TCA_FLOWER_KEY_ARP_TIP_MASK]   = { .type = NLA_U32 },
 671         [TCA_FLOWER_KEY_ARP_OP]         = { .type = NLA_U8 },
 672         [TCA_FLOWER_KEY_ARP_OP_MASK]    = { .type = NLA_U8 },
 673         [TCA_FLOWER_KEY_ARP_SHA]        = { .len = ETH_ALEN },
 674         [TCA_FLOWER_KEY_ARP_SHA_MASK]   = { .len = ETH_ALEN },
 675         [TCA_FLOWER_KEY_ARP_THA]        = { .len = ETH_ALEN },
 676         [TCA_FLOWER_KEY_ARP_THA_MASK]   = { .len = ETH_ALEN },
 677         [TCA_FLOWER_KEY_MPLS_TTL]       = { .type = NLA_U8 },
 678         [TCA_FLOWER_KEY_MPLS_BOS]       = { .type = NLA_U8 },
 679         [TCA_FLOWER_KEY_MPLS_TC]        = { .type = NLA_U8 },
 680         [TCA_FLOWER_KEY_MPLS_LABEL]     = { .type = NLA_U32 },
 681         [TCA_FLOWER_KEY_TCP_FLAGS]      = { .type = NLA_U16 },
 682         [TCA_FLOWER_KEY_TCP_FLAGS_MASK] = { .type = NLA_U16 },
 683         [TCA_FLOWER_KEY_IP_TOS]         = { .type = NLA_U8 },
 684         [TCA_FLOWER_KEY_IP_TOS_MASK]    = { .type = NLA_U8 },
 685         [TCA_FLOWER_KEY_IP_TTL]         = { .type = NLA_U8 },
 686         [TCA_FLOWER_KEY_IP_TTL_MASK]    = { .type = NLA_U8 },
 687         [TCA_FLOWER_KEY_CVLAN_ID]       = { .type = NLA_U16 },
 688         [TCA_FLOWER_KEY_CVLAN_PRIO]     = { .type = NLA_U8 },
 689         [TCA_FLOWER_KEY_CVLAN_ETH_TYPE] = { .type = NLA_U16 },
 690         [TCA_FLOWER_KEY_ENC_IP_TOS]     = { .type = NLA_U8 },
 691         [TCA_FLOWER_KEY_ENC_IP_TOS_MASK] = { .type = NLA_U8 },
 692         [TCA_FLOWER_KEY_ENC_IP_TTL]      = { .type = NLA_U8 },
 693         [TCA_FLOWER_KEY_ENC_IP_TTL_MASK] = { .type = NLA_U8 },
 694         [TCA_FLOWER_KEY_ENC_OPTS]       = { .type = NLA_NESTED },
 695         [TCA_FLOWER_KEY_ENC_OPTS_MASK]  = { .type = NLA_NESTED },
 696 };
 697
 698 static const struct nla_policy
 699 enc_opts_policy[TCA_FLOWER_KEY_ENC_OPTS_MAX + 1] = {
 700         [TCA_FLOWER_KEY_ENC_OPTS_GENEVE]        = { .type = NLA_NESTED },
 701 };
 702
 703 static const struct nla_policy
 704 geneve_opt_policy[TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX + 1] = {
 705         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS]      = { .type = NLA_U16 },
 706         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE]       = { .type = NLA_U8 },
 707         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA]       = { .type = NLA_BINARY,
 708                                                        .len = 128 },
 709 };
 710
 711 static void fl_set_key_val(struct nlattr **tb,
 712                            void *val, int val_type,
 713                            void *mask, int mask_type, int len)
 714 {
 715         if (!tb[val_type])
 716                 return;
 717         memcpy(val, nla_data(tb[val_type]), len);
 718         if (mask_type == TCA_FLOWER_UNSPEC || !tb[mask_type])
 719                 memset(mask, 0xff, len);
 720         else
 721                 memcpy(mask, nla_data(tb[mask_type]), len);
 722 }
 723
 724 static int fl_set_key_port_range(struct nlattr **tb, struct fl_flow_key *key,
 725                                  struct fl_flow_key *mask)
 726 {
 727         fl_set_key_val(tb, &key->tp_min.dst,
 728                        TCA_FLOWER_KEY_PORT_DST_MIN, &mask->tp_min.dst,
 729                        TCA_FLOWER_UNSPEC, sizeof(key->tp_min.dst));
 730         fl_set_key_val(tb, &key->tp_max.dst,
 731                        TCA_FLOWER_KEY_PORT_DST_MAX, &mask->tp_max.dst,
 732                        TCA_FLOWER_UNSPEC, sizeof(key->tp_max.dst));
 733         fl_set_key_val(tb, &key->tp_min.src,
 734                        TCA_FLOWER_KEY_PORT_SRC_MIN, &mask->tp_min.src,
 735                        TCA_FLOWER_UNSPEC, sizeof(key->tp_min.src));
 736         fl_set_key_val(tb, &key->tp_max.src,
 737                        TCA_FLOWER_KEY_PORT_SRC_MAX, &mask->tp_max.src,
 738                        TCA_FLOWER_UNSPEC, sizeof(key->tp_max.src));
 739
 740         if ((mask->tp_min.dst && mask->tp_max.dst &&
 741              htons(key->tp_max.dst) <= htons(key->tp_min.dst)) ||
 742              (mask->tp_min.src && mask->tp_max.src &&
 743               htons(key->tp_max.src) <= htons(key->tp_min.src)))
 744                 return -EINVAL;
 745
 746         return 0;
 747 }
 748
 749 static int fl_set_key_mpls(struct nlattr **tb,
 750                            struct flow_dissector_key_mpls *key_val,
 751                            struct flow_dissector_key_mpls *key_mask)
 752 {
 753         if (tb[TCA_FLOWER_KEY_MPLS_TTL]) {
 754                 key_val->mpls_ttl = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_TTL]);
 755                 key_mask->mpls_ttl = MPLS_TTL_MASK;
 756         }
 757         if (tb[TCA_FLOWER_KEY_MPLS_BOS]) {
 758                 u8 bos = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_BOS]);
 759
 760                 if (bos & ~MPLS_BOS_MASK)
 761                         return -EINVAL;
 762                 key_val->mpls_bos = bos;
 763                 key_mask->mpls_bos = MPLS_BOS_MASK;
 764         }
 765         if (tb[TCA_FLOWER_KEY_MPLS_TC]) {
 766                 u8 tc = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_TC]);
 767
 768                 if (tc & ~MPLS_TC_MASK)
 769                         return -EINVAL;
 770                 key_val->mpls_tc = tc;
 771                 key_mask->mpls_tc = MPLS_TC_MASK;
 772         }
 773         if (tb[TCA_FLOWER_KEY_MPLS_LABEL]) {
 774                 u32 label = nla_get_u32(tb[TCA_FLOWER_KEY_MPLS_LABEL]);
 775
 776                 if (label & ~MPLS_LABEL_MASK)
 777                         return -EINVAL;
 778                 key_val->mpls_label = label;
 779                 key_mask->mpls_label = MPLS_LABEL_MASK;
 780         }
 781         return 0;
 782 }
 783
 784 static void fl_set_key_vlan(struct nlattr **tb,
 785                             __be16 ethertype,
 786                             int vlan_id_key, int vlan_prio_key,
 787                             struct flow_dissector_key_vlan *key_val,
 788                             struct flow_dissector_key_vlan *key_mask)
 789 {
 790 #define VLAN_PRIORITY_MASK      0x7
 791
 792         if (tb[vlan_id_key]) {
 793                 key_val->vlan_id =
 794                         nla_get_u16(tb[vlan_id_key]) & VLAN_VID_MASK;
 795                 key_mask->vlan_id = VLAN_VID_MASK;
 796         }
 797         if (tb[vlan_prio_key]) {
 798                 key_val->vlan_priority =
 799                         nla_get_u8(tb[vlan_prio_key]) &
 800                         VLAN_PRIORITY_MASK;
 801                 key_mask->vlan_priority = VLAN_PRIORITY_MASK;
 802         }
 803         key_val->vlan_tpid = ethertype;
 804         key_mask->vlan_tpid = cpu_to_be16(~0);
 805 }
 806
 807 static void fl_set_key_flag(u32 flower_key, u32 flower_mask,
 808                             u32 *dissector_key, u32 *dissector_mask,
 809                             u32 flower_flag_bit, u32 dissector_flag_bit)
 810 {
 811         if (flower_mask & flower_flag_bit) {
 812                 *dissector_mask |= dissector_flag_bit;
 813                 if (flower_key & flower_flag_bit)
 814                         *dissector_key |= dissector_flag_bit;
 815         }
 816 }
 817
 818 static int fl_set_key_flags(struct nlattr **tb,
 819                             u32 *flags_key, u32 *flags_mask)
 820 {
 821         u32 key, mask;
 822
 823         /* mask is mandatory for flags */
 824         if (!tb[TCA_FLOWER_KEY_FLAGS_MASK])
 825                 return -EINVAL;
 826
 827         key = be32_to_cpu(nla_get_u32(tb[TCA_FLOWER_KEY_FLAGS]));
 828         mask = be32_to_cpu(nla_get_u32(tb[TCA_FLOWER_KEY_FLAGS_MASK]));
 829
 830         *flags_key  = 0;
 831         *flags_mask = 0;
 832
 833         fl_set_key_flag(key, mask, flags_key, flags_mask,
 834                         TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT, FLOW_DIS_IS_FRAGMENT);
 835         fl_set_key_flag(key, mask, flags_key, flags_mask,
 836                         TCA_FLOWER_KEY_FLAGS_FRAG_IS_FIRST,
 837                         FLOW_DIS_FIRST_FRAG);
 838
 839         return 0;
 840 }
 841
 842 static void fl_set_key_ip(struct nlattr **tb, bool encap,
 843                           struct flow_dissector_key_ip *key,
 844                           struct flow_dissector_key_ip *mask)
 845 {
 846         int tos_key = encap ? TCA_FLOWER_KEY_ENC_IP_TOS : TCA_FLOWER_KEY_IP_TOS;
 847         int ttl_key = encap ? TCA_FLOWER_KEY_ENC_IP_TTL : TCA_FLOWER_KEY_IP_TTL;
 848         int tos_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TOS_MASK : TCA_FLOWER_KEY_IP_TOS_MASK;
 849         int ttl_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TTL_MASK : TCA_FLOWER_KEY_IP_TTL_MASK;
 850
 851         fl_set_key_val(tb, &key->tos, tos_key, &mask->tos, tos_mask, sizeof(key->tos));
 852         fl_set_key_val(tb, &key->ttl, ttl_key, &mask->ttl, ttl_mask, sizeof(key->ttl));
 853 }
 854
 855 static int fl_set_geneve_opt(const struct nlattr *nla, struct fl_flow_key *key,
 856                              int depth, int option_len,
 857                              struct netlink_ext_ack *extack)
 858 {
 859         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX + 1];
 860         struct nlattr *class = NULL, *type = NULL, *data = NULL;
 861         struct geneve_opt *opt;
 862         int err, data_len = 0;
 863
 864         if (option_len > sizeof(struct geneve_opt))
 865                 data_len = option_len - sizeof(struct geneve_opt);
 866
 867         opt = (struct geneve_opt *)&key->enc_opts.data[key->enc_opts.len];
 868         memset(opt, 0xff, option_len);
 869         opt->length = data_len / 4;
 870         opt->r1 = 0;
 871         opt->r2 = 0;
 872         opt->r3 = 0;
 873
 874         /* If no mask has been prodived we assume an exact match. */
 875         if (!depth)
 876                 return sizeof(struct geneve_opt) + data_len;
 877
 878         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_GENEVE) {
 879                 NL_SET_ERR_MSG(extack, "Non-geneve option type for mask");
 880                 return -EINVAL;
 881         }
 882
 883         err = nla_parse_nested_deprecated(tb,
 884                                           TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX,
 885                                           nla, geneve_opt_policy, extack);
 886         if (err < 0)
 887                 return err;
 888
 889         /* We are not allowed to omit any of CLASS, TYPE or DATA
 890          * fields from the key.
 891          */
 892         if (!option_len &&
 893             (!tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS] ||
 894              !tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE] ||
 895              !tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA])) {
 896                 NL_SET_ERR_MSG(extack, "Missing tunnel key geneve option class, type or data");
 897                 return -EINVAL;
 898         }
 899
 900         /* Omitting any of CLASS, TYPE or DATA fields is allowed
 901          * for the mask.
 902          */
 903         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA]) {
 904                 int new_len = key->enc_opts.len;
 905
 906                 data = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA];
 907                 data_len = nla_len(data);
 908                 if (data_len < 4) {
 909                         NL_SET_ERR_MSG(extack, "Tunnel key geneve option data is less than 4 bytes long");
 910                         return -ERANGE;
 911                 }
 912                 if (data_len % 4) {
 913                         NL_SET_ERR_MSG(extack, "Tunnel key geneve option data is not a multiple of 4 bytes long");
 914                         return -ERANGE;
 915                 }
 916
 917                 new_len += sizeof(struct geneve_opt) + data_len;
 918                 BUILD_BUG_ON(FLOW_DIS_TUN_OPTS_MAX != IP_TUNNEL_OPTS_MAX);
 919                 if (new_len > FLOW_DIS_TUN_OPTS_MAX) {
 920                         NL_SET_ERR_MSG(extack, "Tunnel options exceeds max size");
 921                         return -ERANGE;
 922                 }
 923                 opt->length = data_len / 4;
 924                 memcpy(opt->opt_data, nla_data(data), data_len);
 925         }
 926
 927         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS]) {
 928                 class = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS];
 929                 opt->opt_class = nla_get_be16(class);
 930         }
 931
 932         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE]) {
 933                 type = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE];
 934                 opt->type = nla_get_u8(type);
 935         }
 936
 937         return sizeof(struct geneve_opt) + data_len;
 938 }
 939
 940 static int fl_set_enc_opt(struct nlattr **tb, struct fl_flow_key *key,
 941                           struct fl_flow_key *mask,
 942                           struct netlink_ext_ack *extack)
 943 {
 944         const struct nlattr *nla_enc_key, *nla_opt_key, *nla_opt_msk = NULL;
 945         int err, option_len, key_depth, msk_depth = 0;
 946
 947         err = nla_validate_nested_deprecated(tb[TCA_FLOWER_KEY_ENC_OPTS],
 948                                              TCA_FLOWER_KEY_ENC_OPTS_MAX,
 949                                              enc_opts_policy, extack);
 950         if (err)
 951                 return err;
 952
 953         nla_enc_key = nla_data(tb[TCA_FLOWER_KEY_ENC_OPTS]);
 954
 955         if (tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]) {
 956                 err = nla_validate_nested_deprecated(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK],
 957                                                      TCA_FLOWER_KEY_ENC_OPTS_MAX,
 958                                                      enc_opts_policy, extack);
 959                 if (err)
 960                         return err;
 961
 962                 nla_opt_msk = nla_data(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]);
 963                 msk_depth = nla_len(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]);
 964         }
 965
 966         nla_for_each_attr(nla_opt_key, nla_enc_key,
 967                           nla_len(tb[TCA_FLOWER_KEY_ENC_OPTS]), key_depth) {
 968                 switch (nla_type(nla_opt_key)) {
 969                 case TCA_FLOWER_KEY_ENC_OPTS_GENEVE:
 970                         option_len = 0;
 971                         key->enc_opts.dst_opt_type = TUNNEL_GENEVE_OPT;
 972                         option_len = fl_set_geneve_opt(nla_opt_key, key,
 973                                                        key_depth, option_len,
 974                                                        extack);
 975                         if (option_len < 0)
 976                                 return option_len;
 977
 978                         key->enc_opts.len += option_len;
 979                         /* At the same time we need to parse through the mask
 980                          * in order to verify exact and mask attribute lengths.
 981                          */
 982                         mask->enc_opts.dst_opt_type = TUNNEL_GENEVE_OPT;
 983                         option_len = fl_set_geneve_opt(nla_opt_msk, mask,
 984                                                        msk_depth, option_len,
 985                                                        extack);
 986                         if (option_len < 0)
 987                                 return option_len;
 988
 989                         mask->enc_opts.len += option_len;
 990                         if (key->enc_opts.len != mask->enc_opts.len) {
 991                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
 992                                 return -EINVAL;
 993                         }
 994
 995                         if (msk_depth)
 996                                 nla_opt_msk = nla_next(nla_opt_msk, &msk_depth);
 997                         break;
 998                 default:
 999                         NL_SET_ERR_MSG(extack, "Unknown tunnel option type");
1000                         return -EINVAL;
1001                 }
1002         }
1003
1004         return 0;
1005 }
1006
1007 static int fl_set_key(struct net *net, struct nlattr **tb,
1008                       struct fl_flow_key *key, struct fl_flow_key *mask,
1009                       struct netlink_ext_ack *extack)
1010 {
1011         __be16 ethertype;
1012         int ret = 0;
1013 #ifdef CONFIG_NET_CLS_IND
1014         if (tb[TCA_FLOWER_INDEV]) {
1015                 int err = tcf_change_indev(net, tb[TCA_FLOWER_INDEV], extack);
1016                 if (err < 0)
1017                         return err;
1018                 key->indev_ifindex = err;
1019                 mask->indev_ifindex = 0xffffffff;
1020         }
1021 #endif
1022
1023         fl_set_key_val(tb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
1024                        mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
1025                        sizeof(key->eth.dst));
1026         fl_set_key_val(tb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
1027                        mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
1028                        sizeof(key->eth.src));
1029
1030         if (tb[TCA_FLOWER_KEY_ETH_TYPE]) {
1031                 ethertype = nla_get_be16(tb[TCA_FLOWER_KEY_ETH_TYPE]);
1032
1033                 if (eth_type_vlan(ethertype)) {
1034                         fl_set_key_vlan(tb, ethertype, TCA_FLOWER_KEY_VLAN_ID,
1035                                         TCA_FLOWER_KEY_VLAN_PRIO, &key->vlan,
1036                                         &mask->vlan);
1037
1038                         if (tb[TCA_FLOWER_KEY_VLAN_ETH_TYPE]) {
1039                                 ethertype = nla_get_be16(tb[TCA_FLOWER_KEY_VLAN_ETH_TYPE]);
1040                                 if (eth_type_vlan(ethertype)) {
1041                                         fl_set_key_vlan(tb, ethertype,
1042                                                         TCA_FLOWER_KEY_CVLAN_ID,
1043                                                         TCA_FLOWER_KEY_CVLAN_PRIO,
1044                                                         &key->cvlan, &mask->cvlan);
1045                                         fl_set_key_val(tb, &key->basic.n_proto,
1046                                                        TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
1047                                                        &mask->basic.n_proto,
1048                                                        TCA_FLOWER_UNSPEC,
1049                                                        sizeof(key->basic.n_proto));
1050                                 } else {
1051                                         key->basic.n_proto = ethertype;
1052                                         mask->basic.n_proto = cpu_to_be16(~0);
1053                                 }
1054                         }
1055                 } else {
1056                         key->basic.n_proto = ethertype;
1057                         mask->basic.n_proto = cpu_to_be16(~0);
1058                 }
1059         }
1060
1061         if (key->basic.n_proto == htons(ETH_P_IP) ||
1062             key->basic.n_proto == htons(ETH_P_IPV6)) {
1063                 fl_set_key_val(tb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
1064                                &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
1065                                sizeof(key->basic.ip_proto));
1066                 fl_set_key_ip(tb, false, &key->ip, &mask->ip);
1067         }
1068
1069         if (tb[TCA_FLOWER_KEY_IPV4_SRC] || tb[TCA_FLOWER_KEY_IPV4_DST]) {
1070                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
1071                 mask->control.addr_type = ~0;
1072                 fl_set_key_val(tb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
1073                                &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
1074                                sizeof(key->ipv4.src));
1075                 fl_set_key_val(tb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
1076                                &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
1077                                sizeof(key->ipv4.dst));
1078         } else if (tb[TCA_FLOWER_KEY_IPV6_SRC] || tb[TCA_FLOWER_KEY_IPV6_DST]) {
1079                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
1080                 mask->control.addr_type = ~0;
1081                 fl_set_key_val(tb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
1082                                &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
1083                                sizeof(key->ipv6.src));
1084                 fl_set_key_val(tb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
1085                                &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
1086                                sizeof(key->ipv6.dst));
1087         }
1088
1089         if (key->basic.ip_proto == IPPROTO_TCP) {
1090                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
1091                                &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
1092                                sizeof(key->tp.src));
1093                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
1094                                &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
1095                                sizeof(key->tp.dst));
1096                 fl_set_key_val(tb, &key->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS,
1097                                &mask->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS_MASK,
1098                                sizeof(key->tcp.flags));
1099         } else if (key->basic.ip_proto == IPPROTO_UDP) {
1100                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
1101                                &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
1102                                sizeof(key->tp.src));
1103                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
1104                                &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
1105                                sizeof(key->tp.dst));
1106         } else if (key->basic.ip_proto == IPPROTO_SCTP) {
1107                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_SCTP_SRC,
1108                                &mask->tp.src, TCA_FLOWER_KEY_SCTP_SRC_MASK,
1109                                sizeof(key->tp.src));
1110                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_SCTP_DST,
1111                                &mask->tp.dst, TCA_FLOWER_KEY_SCTP_DST_MASK,
1112                                sizeof(key->tp.dst));
1113         } else if (key->basic.n_proto == htons(ETH_P_IP) &&
1114                    key->basic.ip_proto == IPPROTO_ICMP) {
1115                 fl_set_key_val(tb, &key->icmp.type, TCA_FLOWER_KEY_ICMPV4_TYPE,
1116                                &mask->icmp.type,
1117                                TCA_FLOWER_KEY_ICMPV4_TYPE_MASK,
1118                                sizeof(key->icmp.type));
1119                 fl_set_key_val(tb, &key->icmp.code, TCA_FLOWER_KEY_ICMPV4_CODE,
1120                                &mask->icmp.code,
1121                                TCA_FLOWER_KEY_ICMPV4_CODE_MASK,
1122                                sizeof(key->icmp.code));
1123         } else if (key->basic.n_proto == htons(ETH_P_IPV6) &&
1124                    key->basic.ip_proto == IPPROTO_ICMPV6) {
1125                 fl_set_key_val(tb, &key->icmp.type, TCA_FLOWER_KEY_ICMPV6_TYPE,
1126                                &mask->icmp.type,
1127                                TCA_FLOWER_KEY_ICMPV6_TYPE_MASK,
1128                                sizeof(key->icmp.type));
1129                 fl_set_key_val(tb, &key->icmp.code, TCA_FLOWER_KEY_ICMPV6_CODE,
1130                                &mask->icmp.code,
1131                                TCA_FLOWER_KEY_ICMPV6_CODE_MASK,
1132                                sizeof(key->icmp.code));
1133         } else if (key->basic.n_proto == htons(ETH_P_MPLS_UC) ||
1134                    key->basic.n_proto == htons(ETH_P_MPLS_MC)) {
1135                 ret = fl_set_key_mpls(tb, &key->mpls, &mask->mpls);
1136                 if (ret)
1137                         return ret;
1138         } else if (key->basic.n_proto == htons(ETH_P_ARP) ||
1139                    key->basic.n_proto == htons(ETH_P_RARP)) {
1140                 fl_set_key_val(tb, &key->arp.sip, TCA_FLOWER_KEY_ARP_SIP,
1141                                &mask->arp.sip, TCA_FLOWER_KEY_ARP_SIP_MASK,
1142                                sizeof(key->arp.sip));
1143                 fl_set_key_val(tb, &key->arp.tip, TCA_FLOWER_KEY_ARP_TIP,
1144                                &mask->arp.tip, TCA_FLOWER_KEY_ARP_TIP_MASK,
1145                                sizeof(key->arp.tip));
1146                 fl_set_key_val(tb, &key->arp.op, TCA_FLOWER_KEY_ARP_OP,
1147                                &mask->arp.op, TCA_FLOWER_KEY_ARP_OP_MASK,
1148                                sizeof(key->arp.op));
1149                 fl_set_key_val(tb, key->arp.sha, TCA_FLOWER_KEY_ARP_SHA,
1150                                mask->arp.sha, TCA_FLOWER_KEY_ARP_SHA_MASK,
1151                                sizeof(key->arp.sha));
1152                 fl_set_key_val(tb, key->arp.tha, TCA_FLOWER_KEY_ARP_THA,
1153                                mask->arp.tha, TCA_FLOWER_KEY_ARP_THA_MASK,
1154                                sizeof(key->arp.tha));
1155         }
1156
1157         if (key->basic.ip_proto == IPPROTO_TCP ||
1158             key->basic.ip_proto == IPPROTO_UDP ||
1159             key->basic.ip_proto == IPPROTO_SCTP) {
1160                 ret = fl_set_key_port_range(tb, key, mask);
1161                 if (ret)
1162                         return ret;
1163         }
1164
1165         if (tb[TCA_FLOWER_KEY_ENC_IPV4_SRC] ||
1166             tb[TCA_FLOWER_KEY_ENC_IPV4_DST]) {
1167                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
1168                 mask->enc_control.addr_type = ~0;
1169                 fl_set_key_val(tb, &key->enc_ipv4.src,
1170                                TCA_FLOWER_KEY_ENC_IPV4_SRC,
1171                                &mask->enc_ipv4.src,
1172                                TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
1173                                sizeof(key->enc_ipv4.src));
1174                 fl_set_key_val(tb, &key->enc_ipv4.dst,
1175                                TCA_FLOWER_KEY_ENC_IPV4_DST,
1176                                &mask->enc_ipv4.dst,
1177                                TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
1178                                sizeof(key->enc_ipv4.dst));
1179         }
1180
1181         if (tb[TCA_FLOWER_KEY_ENC_IPV6_SRC] ||
1182             tb[TCA_FLOWER_KEY_ENC_IPV6_DST]) {
1183                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
1184                 mask->enc_control.addr_type = ~0;
1185                 fl_set_key_val(tb, &key->enc_ipv6.src,
1186                                TCA_FLOWER_KEY_ENC_IPV6_SRC,
1187                                &mask->enc_ipv6.src,
1188                                TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
1189                                sizeof(key->enc_ipv6.src));
1190                 fl_set_key_val(tb, &key->enc_ipv6.dst,
1191                                TCA_FLOWER_KEY_ENC_IPV6_DST,
1192                                &mask->enc_ipv6.dst,
1193                                TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
1194                                sizeof(key->enc_ipv6.dst));
1195         }
1196
1197         fl_set_key_val(tb, &key->enc_key_id.keyid, TCA_FLOWER_KEY_ENC_KEY_ID,
1198                        &mask->enc_key_id.keyid, TCA_FLOWER_UNSPEC,
1199                        sizeof(key->enc_key_id.keyid));
1200
1201         fl_set_key_val(tb, &key->enc_tp.src, TCA_FLOWER_KEY_ENC_UDP_SRC_PORT,
1202                        &mask->enc_tp.src, TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK,
1203                        sizeof(key->enc_tp.src));
1204
1205         fl_set_key_val(tb, &key->enc_tp.dst, TCA_FLOWER_KEY_ENC_UDP_DST_PORT,
1206                        &mask->enc_tp.dst, TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK,
1207                        sizeof(key->enc_tp.dst));
1208
1209         fl_set_key_ip(tb, true, &key->enc_ip, &mask->enc_ip);
1210
1211         if (tb[TCA_FLOWER_KEY_ENC_OPTS]) {
1212                 ret = fl_set_enc_opt(tb, key, mask, extack);
1213                 if (ret)
1214                         return ret;
1215         }
1216
1217         if (tb[TCA_FLOWER_KEY_FLAGS])
1218                 ret = fl_set_key_flags(tb, &key->control.flags, &mask->control.flags);
1219
1220         return ret;
1221 }
1222
1223 static void fl_mask_copy(struct fl_flow_mask *dst,
1224                          struct fl_flow_mask *src)
1225 {
1226         const void *psrc = fl_key_get_start(&src->key, src);
1227         void *pdst = fl_key_get_start(&dst->key, src);
1228
1229         memcpy(pdst, psrc, fl_mask_range(src));
1230         dst->range = src->range;
1231 }
1232
1233 static const struct rhashtable_params fl_ht_params = {
1234         .key_offset = offsetof(struct cls_fl_filter, mkey), /* base offset */
1235         .head_offset = offsetof(struct cls_fl_filter, ht_node),
1236         .automatic_shrinking = true,
1237 };
1238
1239 static int fl_init_mask_hashtable(struct fl_flow_mask *mask)
1240 {
1241         mask->filter_ht_params = fl_ht_params;
1242         mask->filter_ht_params.key_len = fl_mask_range(mask);
1243         mask->filter_ht_params.key_offset += mask->range.start;
1244
1245         return rhashtable_init(&mask->ht, &mask->filter_ht_params);
1246 }
1247
1248 #define FL_KEY_MEMBER_OFFSET(member) offsetof(struct fl_flow_key, member)
1249 #define FL_KEY_MEMBER_SIZE(member) FIELD_SIZEOF(struct fl_flow_key, member)
1250
1251 #define FL_KEY_IS_MASKED(mask, member)                                          \
1252         memchr_inv(((char *)mask) + FL_KEY_MEMBER_OFFSET(member),               \
1253                    0, FL_KEY_MEMBER_SIZE(member))                               \
1254
1255 #define FL_KEY_SET(keys, cnt, id, member)                                       \
1256         do {                                                                    \
1257                 keys[cnt].key_id = id;                                          \
1258                 keys[cnt].offset = FL_KEY_MEMBER_OFFSET(member);                \
1259                 cnt++;                                                          \
1260         } while(0);
1261
1262 #define FL_KEY_SET_IF_MASKED(mask, keys, cnt, id, member)                       \
1263         do {                                                                    \
1264                 if (FL_KEY_IS_MASKED(mask, member))                             \
1265                         FL_KEY_SET(keys, cnt, id, member);                      \
1266         } while(0);
1267
1268 static void fl_init_dissector(struct flow_dissector *dissector,
1269                               struct fl_flow_key *mask)
1270 {
1271         struct flow_dissector_key keys[FLOW_DISSECTOR_KEY_MAX];
1272         size_t cnt = 0;
1273
1274         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_CONTROL, control);
1275         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_BASIC, basic);
1276         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1277                              FLOW_DISSECTOR_KEY_ETH_ADDRS, eth);
1278         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1279                              FLOW_DISSECTOR_KEY_IPV4_ADDRS, ipv4);
1280         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1281                              FLOW_DISSECTOR_KEY_IPV6_ADDRS, ipv6);
1282         if (FL_KEY_IS_MASKED(mask, tp) ||
1283             FL_KEY_IS_MASKED(mask, tp_min) || FL_KEY_IS_MASKED(mask, tp_max))
1284                 FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_PORTS, tp);
1285         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1286                              FLOW_DISSECTOR_KEY_IP, ip);
1287         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1288                              FLOW_DISSECTOR_KEY_TCP, tcp);
1289         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1290                              FLOW_DISSECTOR_KEY_ICMP, icmp);
1291         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1292                              FLOW_DISSECTOR_KEY_ARP, arp);
1293         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1294                              FLOW_DISSECTOR_KEY_MPLS, mpls);
1295         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1296                              FLOW_DISSECTOR_KEY_VLAN, vlan);
1297         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1298                              FLOW_DISSECTOR_KEY_CVLAN, cvlan);
1299         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1300                              FLOW_DISSECTOR_KEY_ENC_KEYID, enc_key_id);
1301         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1302                              FLOW_DISSECTOR_KEY_ENC_IPV4_ADDRS, enc_ipv4);
1303         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1304                              FLOW_DISSECTOR_KEY_ENC_IPV6_ADDRS, enc_ipv6);
1305         if (FL_KEY_IS_MASKED(mask, enc_ipv4) ||
1306             FL_KEY_IS_MASKED(mask, enc_ipv6))
1307                 FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_ENC_CONTROL,
1308                            enc_control);
1309         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1310                              FLOW_DISSECTOR_KEY_ENC_PORTS, enc_tp);
1311         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1312                              FLOW_DISSECTOR_KEY_ENC_IP, enc_ip);
1313         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1314                              FLOW_DISSECTOR_KEY_ENC_OPTS, enc_opts);
1315
1316         skb_flow_dissector_init(dissector, keys, cnt);
1317 }
1318
1319 static struct fl_flow_mask *fl_create_new_mask(struct cls_fl_head *head,
1320                                                struct fl_flow_mask *mask)
1321 {
1322         struct fl_flow_mask *newmask;
1323         int err;
1324
1325         newmask = kzalloc(sizeof(*newmask), GFP_KERNEL);
1326         if (!newmask)
1327                 return ERR_PTR(-ENOMEM);
1328
1329         fl_mask_copy(newmask, mask);
1330
1331         if ((newmask->key.tp_min.dst && newmask->key.tp_max.dst) ||
1332             (newmask->key.tp_min.src && newmask->key.tp_max.src))
1333                 newmask->flags |= TCA_FLOWER_MASK_FLAGS_RANGE;
1334
1335         err = fl_init_mask_hashtable(newmask);
1336         if (err)
1337                 goto errout_free;
1338
1339         fl_init_dissector(&newmask->dissector, &newmask->key);
1340
1341         INIT_LIST_HEAD_RCU(&newmask->filters);
1342
1343         refcount_set(&newmask->refcnt, 1);
1344         err = rhashtable_replace_fast(&head->ht, &mask->ht_node,
1345                                       &newmask->ht_node, mask_ht_params);
1346         if (err)
1347                 goto errout_destroy;
1348
1349         /* Wait until any potential concurrent users of mask are finished */
1350         synchronize_rcu();
1351
1352         spin_lock(&head->masks_lock);
1353         list_add_tail_rcu(&newmask->list, &head->masks);
1354         spin_unlock(&head->masks_lock);
1355
1356         return newmask;
1357
1358 errout_destroy:
1359         rhashtable_destroy(&newmask->ht);
1360 errout_free:
1361         kfree(newmask);
1362
1363         return ERR_PTR(err);
1364 }
1365
1366 static int fl_check_assign_mask(struct cls_fl_head *head,
1367                                 struct cls_fl_filter *fnew,
1368                                 struct cls_fl_filter *fold,
1369                                 struct fl_flow_mask *mask)
1370 {
1371         struct fl_flow_mask *newmask;
1372         int ret = 0;
1373
1374         rcu_read_lock();
1375
1376         /* Insert mask as temporary node to prevent concurrent creation of mask
1377          * with same key. Any concurrent lookups with same key will return
1378          * -EAGAIN because mask's refcnt is zero. It is safe to insert
1379          * stack-allocated 'mask' to masks hash table because we call
1380          * synchronize_rcu() before returning from this function (either in case
1381          * of error or after replacing it with heap-allocated mask in
1382          * fl_create_new_mask()).
1383          */
1384         fnew->mask = rhashtable_lookup_get_insert_fast(&head->ht,
1385                                                        &mask->ht_node,
1386                                                        mask_ht_params);
1387         if (!fnew->mask) {
1388                 rcu_read_unlock();
1389
1390                 if (fold) {
1391                         ret = -EINVAL;
1392                         goto errout_cleanup;
1393                 }
1394
1395                 newmask = fl_create_new_mask(head, mask);
1396                 if (IS_ERR(newmask)) {
1397                         ret = PTR_ERR(newmask);
1398                         goto errout_cleanup;
1399                 }
1400
1401                 fnew->mask = newmask;
1402                 return 0;
1403         } else if (IS_ERR(fnew->mask)) {
1404                 ret = PTR_ERR(fnew->mask);
1405         } else if (fold && fold->mask != fnew->mask) {
1406                 ret = -EINVAL;
1407         } else if (!refcount_inc_not_zero(&fnew->mask->refcnt)) {
1408                 /* Mask was deleted concurrently, try again */
1409                 ret = -EAGAIN;
1410         }
1411         rcu_read_unlock();
1412         return ret;
1413
1414 errout_cleanup:
1415         rhashtable_remove_fast(&head->ht, &mask->ht_node,
1416                                mask_ht_params);
1417         /* Wait until any potential concurrent users of mask are finished */
1418         synchronize_rcu();
1419         return ret;
1420 }
1421
1422 static int fl_set_parms(struct net *net, struct tcf_proto *tp,
1423                         struct cls_fl_filter *f, struct fl_flow_mask *mask,
1424                         unsigned long base, struct nlattr **tb,
1425                         struct nlattr *est, bool ovr,
1426                         struct fl_flow_tmplt *tmplt, bool rtnl_held,
1427                         struct netlink_ext_ack *extack)
1428 {
1429         int err;
1430
1431         err = tcf_exts_validate(net, tp, tb, est, &f->exts, ovr, rtnl_held,
1432                                 extack);
1433         if (err < 0)
1434                 return err;
1435
1436         if (tb[TCA_FLOWER_CLASSID]) {
1437                 f->res.classid = nla_get_u32(tb[TCA_FLOWER_CLASSID]);
1438                 if (!rtnl_held)
1439                         rtnl_lock();
1440                 tcf_bind_filter(tp, &f->res, base);
1441                 if (!rtnl_held)
1442                         rtnl_unlock();
1443         }
1444
1445         err = fl_set_key(net, tb, &f->key, &mask->key, extack);
1446         if (err)
1447                 return err;
1448
1449         fl_mask_update_range(mask);
1450         fl_set_masked_key(&f->mkey, &f->key, mask);
1451
1452         if (!fl_mask_fits_tmplt(tmplt, mask)) {
1453                 NL_SET_ERR_MSG_MOD(extack, "Mask does not fit the template");
1454                 return -EINVAL;
1455         }
1456
1457         return 0;
1458 }
1459
1460 static int fl_ht_insert_unique(struct cls_fl_filter *fnew,
1461                                struct cls_fl_filter *fold,
1462                                bool *in_ht)
1463 {
1464         struct fl_flow_mask *mask = fnew->mask;
1465         int err;
1466
1467         err = rhashtable_lookup_insert_fast(&mask->ht,
1468                                             &fnew->ht_node,
1469                                             mask->filter_ht_params);
1470         if (err) {
1471                 *in_ht = false;
1472                 /* It is okay if filter with same key exists when
1473                  * overwriting.
1474                  */
1475                 return fold && err == -EEXIST ? 0 : err;
1476         }
1477
1478         *in_ht = true;
1479         return 0;
1480 }
1481
1482 static int fl_change(struct net *net, struct sk_buff *in_skb,
1483                      struct tcf_proto *tp, unsigned long base,
1484                      u32 handle, struct nlattr **tca,
1485                      void **arg, bool ovr, bool rtnl_held,
1486                      struct netlink_ext_ack *extack)
1487 {
1488         struct cls_fl_head *head = fl_head_dereference(tp);
1489         struct cls_fl_filter *fold = *arg;
1490         struct cls_fl_filter *fnew;
1491         struct fl_flow_mask *mask;
1492         struct nlattr **tb;
1493         bool in_ht;
1494         int err;
1495
1496         if (!tca[TCA_OPTIONS]) {
1497                 err = -EINVAL;
1498                 goto errout_fold;
1499         }
1500
1501         mask = kzalloc(sizeof(struct fl_flow_mask), GFP_KERNEL);
1502         if (!mask) {
1503                 err = -ENOBUFS;
1504                 goto errout_fold;
1505         }
1506
1507         tb = kcalloc(TCA_FLOWER_MAX + 1, sizeof(struct nlattr *), GFP_KERNEL);
1508         if (!tb) {
1509                 err = -ENOBUFS;
1510                 goto errout_mask_alloc;
1511         }
1512
1513         err = nla_parse_nested_deprecated(tb, TCA_FLOWER_MAX,
1514                                           tca[TCA_OPTIONS], fl_policy, NULL);
1515         if (err < 0)
1516                 goto errout_tb;
1517
1518         if (fold && handle && fold->handle != handle) {
1519                 err = -EINVAL;
1520                 goto errout_tb;
1521         }
1522
1523         fnew = kzalloc(sizeof(*fnew), GFP_KERNEL);
1524         if (!fnew) {
1525                 err = -ENOBUFS;
1526                 goto errout_tb;
1527         }
1528         INIT_LIST_HEAD(&fnew->hw_list);
1529         refcount_set(&fnew->refcnt, 1);
1530
1531         err = tcf_exts_init(&fnew->exts, net, TCA_FLOWER_ACT, 0);
1532         if (err < 0)
1533                 goto errout;
1534
1535         if (tb[TCA_FLOWER_FLAGS]) {
1536                 fnew->flags = nla_get_u32(tb[TCA_FLOWER_FLAGS]);
1537
1538                 if (!tc_flags_valid(fnew->flags)) {
1539                         err = -EINVAL;
1540                         goto errout;
1541                 }
1542         }
1543
1544         err = fl_set_parms(net, tp, fnew, mask, base, tb, tca[TCA_RATE], ovr,
1545                            tp->chain->tmplt_priv, rtnl_held, extack);
1546         if (err)
1547                 goto errout;
1548
1549         err = fl_check_assign_mask(head, fnew, fold, mask);
1550         if (err)
1551                 goto errout;
1552
1553         err = fl_ht_insert_unique(fnew, fold, &in_ht);
1554         if (err)
1555                 goto errout_mask;
1556
1557         if (!tc_skip_hw(fnew->flags)) {
1558                 err = fl_hw_replace_filter(tp, fnew, rtnl_held, extack);
1559                 if (err)
1560                         goto errout_ht;
1561         }
1562
1563         if (!tc_in_hw(fnew->flags))
1564                 fnew->flags |= TCA_CLS_FLAGS_NOT_IN_HW;
1565
1566         spin_lock(&tp->lock);
1567
1568         /* tp was deleted concurrently. -EAGAIN will cause caller to lookup
1569          * proto again or create new one, if necessary.
1570          */
1571         if (tp->deleting) {
1572                 err = -EAGAIN;
1573                 goto errout_hw;
1574         }
1575
1576         if (fold) {
1577                 /* Fold filter was deleted concurrently. Retry lookup. */
1578                 if (fold->deleted) {
1579                         err = -EAGAIN;
1580                         goto errout_hw;
1581                 }
1582
1583                 fnew->handle = handle;
1584
1585                 if (!in_ht) {
1586                         struct rhashtable_params params =
1587                                 fnew->mask->filter_ht_params;
1588
1589                         err = rhashtable_insert_fast(&fnew->mask->ht,
1590                                                      &fnew->ht_node,
1591                                                      params);
1592                         if (err)
1593                                 goto errout_hw;
1594                         in_ht = true;
1595                 }
1596
1597                 refcount_inc(&fnew->refcnt);
1598                 rhashtable_remove_fast(&fold->mask->ht,
1599                                        &fold->ht_node,
1600                                        fold->mask->filter_ht_params);
1601                 idr_replace(&head->handle_idr, fnew, fnew->handle);
1602                 list_replace_rcu(&fold->list, &fnew->list);
1603                 fold->deleted = true;
1604
1605                 spin_unlock(&tp->lock);
1606
1607                 fl_mask_put(head, fold->mask);
1608                 if (!tc_skip_hw(fold->flags))
1609                         fl_hw_destroy_filter(tp, fold, rtnl_held, NULL);
1610                 tcf_unbind_filter(tp, &fold->res);
1611                 /* Caller holds reference to fold, so refcnt is always > 0
1612                  * after this.
1613                  */
1614                 refcount_dec(&fold->refcnt);
1615                 __fl_put(fold);
1616         } else {
1617                 if (handle) {
1618                         /* user specifies a handle and it doesn't exist */
1619                         err = idr_alloc_u32(&head->handle_idr, fnew, &handle,
1620                                             handle, GFP_ATOMIC);
1621
1622                         /* Filter with specified handle was concurrently
1623                          * inserted after initial check in cls_api. This is not
1624                          * necessarily an error if NLM_F_EXCL is not set in
1625                          * message flags. Returning EAGAIN will cause cls_api to
1626                          * try to update concurrently inserted rule.
1627                          */
1628                         if (err == -ENOSPC)
1629                                 err = -EAGAIN;
1630                 } else {
1631                         handle = 1;
1632                         err = idr_alloc_u32(&head->handle_idr, fnew, &handle,
1633                                             INT_MAX, GFP_ATOMIC);
1634                 }
1635                 if (err)
1636                         goto errout_hw;
1637
1638                 refcount_inc(&fnew->refcnt);
1639                 fnew->handle = handle;
1640                 list_add_tail_rcu(&fnew->list, &fnew->mask->filters);
1641                 spin_unlock(&tp->lock);
1642         }
1643
1644         *arg = fnew;
1645
1646         kfree(tb);
1647         kfree(mask);
1648         return 0;
1649
1650 errout_ht:
1651         spin_lock(&tp->lock);
1652 errout_hw:
1653         fnew->deleted = true;
1654         spin_unlock(&tp->lock);
1655         if (!tc_skip_hw(fnew->flags))
1656                 fl_hw_destroy_filter(tp, fnew, rtnl_held, NULL);
1657         if (in_ht)
1658                 rhashtable_remove_fast(&fnew->mask->ht, &fnew->ht_node,
1659                                        fnew->mask->filter_ht_params);
1660 errout_mask:
1661         fl_mask_put(head, fnew->mask);
1662 errout:
1663         __fl_put(fnew);
1664 errout_tb:
1665         kfree(tb);
1666 errout_mask_alloc:
1667         kfree(mask);
1668 errout_fold:
1669         if (fold)
1670                 __fl_put(fold);
1671         return err;
1672 }
1673
1674 static int fl_delete(struct tcf_proto *tp, void *arg, bool *last,
1675                      bool rtnl_held, struct netlink_ext_ack *extack)
1676 {
1677         struct cls_fl_head *head = fl_head_dereference(tp);
1678         struct cls_fl_filter *f = arg;
1679         bool last_on_mask;
1680         int err = 0;
1681
1682         err = __fl_delete(tp, f, &last_on_mask, rtnl_held, extack);
1683         *last = list_empty(&head->masks);
1684         __fl_put(f);
1685
1686         return err;
1687 }
1688
1689 static void fl_walk(struct tcf_proto *tp, struct tcf_walker *arg,
1690                     bool rtnl_held)
1691 {
1692         struct cls_fl_filter *f;
1693
1694         arg->count = arg->skip;
1695
1696         while ((f = fl_get_next_filter(tp, &arg->cookie)) != NULL) {
1697                 if (arg->fn(tp, f, arg) < 0) {
1698                         __fl_put(f);
1699                         arg->stop = 1;
1700                         break;
1701                 }
1702                 __fl_put(f);
1703                 arg->cookie++;
1704                 arg->count++;
1705         }
1706 }
1707
1708 static struct cls_fl_filter *
1709 fl_get_next_hw_filter(struct tcf_proto *tp, struct cls_fl_filter *f, bool add)
1710 {
1711         struct cls_fl_head *head = fl_head_dereference(tp);
1712
1713         spin_lock(&tp->lock);
1714         if (list_empty(&head->hw_filters)) {
1715                 spin_unlock(&tp->lock);
1716                 return NULL;
1717         }
1718
1719         if (!f)
1720                 f = list_entry(&head->hw_filters, struct cls_fl_filter,
1721                                hw_list);
1722         list_for_each_entry_continue(f, &head->hw_filters, hw_list) {
1723                 if (!(add && f->deleted) && refcount_inc_not_zero(&f->refcnt)) {
1724                         spin_unlock(&tp->lock);
1725                         return f;
1726                 }
1727         }
1728
1729         spin_unlock(&tp->lock);
1730         return NULL;
1731 }
1732
1733 static int fl_reoffload(struct tcf_proto *tp, bool add, tc_setup_cb_t *cb,
1734                         void *cb_priv, struct netlink_ext_ack *extack)
1735 {
1736         struct tc_cls_flower_offload cls_flower = {};
1737         struct tcf_block *block = tp->chain->block;
1738         struct cls_fl_filter *f = NULL;
1739         int err;
1740
1741         /* hw_filters list can only be changed by hw offload functions after
1742          * obtaining rtnl lock. Make sure it is not changed while reoffload is
1743          * iterating it.
1744          */
1745         ASSERT_RTNL();
1746
1747         while ((f = fl_get_next_hw_filter(tp, f, add))) {
1748                 cls_flower.rule =
1749                         flow_rule_alloc(tcf_exts_num_actions(&f->exts));
1750                 if (!cls_flower.rule) {
1751                         __fl_put(f);
1752                         return -ENOMEM;
1753                 }
1754
1755                 tc_cls_common_offload_init(&cls_flower.common, tp, f->flags,
1756                                            extack);
1757                 cls_flower.command = add ?
1758                         TC_CLSFLOWER_REPLACE : TC_CLSFLOWER_DESTROY;
1759                 cls_flower.cookie = (unsigned long)f;
1760                 cls_flower.rule->match.dissector = &f->mask->dissector;
1761                 cls_flower.rule->match.mask = &f->mask->key;
1762                 cls_flower.rule->match.key = &f->mkey;
1763
1764                 err = tc_setup_flow_action(&cls_flower.rule->action, &f->exts);
1765                 if (err) {
1766                         kfree(cls_flower.rule);
1767                         if (tc_skip_sw(f->flags)) {
1768                                 NL_SET_ERR_MSG_MOD(extack, "Failed to setup flow action");
1769                                 __fl_put(f);
1770                                 return err;
1771                         }
1772                         goto next_flow;
1773                 }
1774
1775                 cls_flower.classid = f->res.classid;
1776
1777                 err = cb(TC_SETUP_CLSFLOWER, &cls_flower, cb_priv);
1778                 kfree(cls_flower.rule);
1779
1780                 if (err) {
1781                         if (add && tc_skip_sw(f->flags)) {
1782                                 __fl_put(f);
1783                                 return err;
1784                         }
1785                         goto next_flow;
1786                 }
1787
1788                 spin_lock(&tp->lock);
1789                 tc_cls_offload_cnt_update(block, &f->in_hw_count, &f->flags,
1790                                           add);
1791                 spin_unlock(&tp->lock);
1792 next_flow:
1793                 __fl_put(f);
1794         }
1795
1796         return 0;
1797 }
1798
1799 static int fl_hw_create_tmplt(struct tcf_chain *chain,
1800                               struct fl_flow_tmplt *tmplt)
1801 {
1802         struct tc_cls_flower_offload cls_flower = {};
1803         struct tcf_block *block = chain->block;
1804
1805         cls_flower.rule = flow_rule_alloc(0);
1806         if (!cls_flower.rule)
1807                 return -ENOMEM;
1808
1809         cls_flower.common.chain_index = chain->index;
1810         cls_flower.command = TC_CLSFLOWER_TMPLT_CREATE;
1811         cls_flower.cookie = (unsigned long) tmplt;
1812         cls_flower.rule->match.dissector = &tmplt->dissector;
1813         cls_flower.rule->match.mask = &tmplt->mask;
1814         cls_flower.rule->match.key = &tmplt->dummy_key;
1815
1816         /* We don't care if driver (any of them) fails to handle this
1817          * call. It serves just as a hint for it.
1818          */
1819         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false);
1820         kfree(cls_flower.rule);
1821
1822         return 0;
1823 }
1824
1825 static void fl_hw_destroy_tmplt(struct tcf_chain *chain,
1826                                 struct fl_flow_tmplt *tmplt)
1827 {
1828         struct tc_cls_flower_offload cls_flower = {};
1829         struct tcf_block *block = chain->block;
1830
1831         cls_flower.common.chain_index = chain->index;
1832         cls_flower.command = TC_CLSFLOWER_TMPLT_DESTROY;
1833         cls_flower.cookie = (unsigned long) tmplt;
1834
1835         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false);
1836 }
1837
1838 static void *fl_tmplt_create(struct net *net, struct tcf_chain *chain,
1839                              struct nlattr **tca,
1840                              struct netlink_ext_ack *extack)
1841 {
1842         struct fl_flow_tmplt *tmplt;
1843         struct nlattr **tb;
1844         int err;
1845
1846         if (!tca[TCA_OPTIONS])
1847                 return ERR_PTR(-EINVAL);
1848
1849         tb = kcalloc(TCA_FLOWER_MAX + 1, sizeof(struct nlattr *), GFP_KERNEL);
1850         if (!tb)
1851                 return ERR_PTR(-ENOBUFS);
1852         err = nla_parse_nested_deprecated(tb, TCA_FLOWER_MAX,
1853                                           tca[TCA_OPTIONS], fl_policy, NULL);
1854         if (err)
1855                 goto errout_tb;
1856
1857         tmplt = kzalloc(sizeof(*tmplt), GFP_KERNEL);
1858         if (!tmplt) {
1859                 err = -ENOMEM;
1860                 goto errout_tb;
1861         }
1862         tmplt->chain = chain;
1863         err = fl_set_key(net, tb, &tmplt->dummy_key, &tmplt->mask, extack);
1864         if (err)
1865                 goto errout_tmplt;
1866
1867         fl_init_dissector(&tmplt->dissector, &tmplt->mask);
1868
1869         err = fl_hw_create_tmplt(chain, tmplt);
1870         if (err)
1871                 goto errout_tmplt;
1872
1873         kfree(tb);
1874         return tmplt;
1875
1876 errout_tmplt:
1877         kfree(tmplt);
1878 errout_tb:
1879         kfree(tb);
1880         return ERR_PTR(err);
1881 }
1882
1883 static void fl_tmplt_destroy(void *tmplt_priv)
1884 {
1885         struct fl_flow_tmplt *tmplt = tmplt_priv;
1886
1887         fl_hw_destroy_tmplt(tmplt->chain, tmplt);
1888         kfree(tmplt);
1889 }
1890
1891 static int fl_dump_key_val(struct sk_buff *skb,
1892                            void *val, int val_type,
1893                            void *mask, int mask_type, int len)
1894 {
1895         int err;
1896
1897         if (!memchr_inv(mask, 0, len))
1898                 return 0;
1899         err = nla_put(skb, val_type, len, val);
1900         if (err)
1901                 return err;
1902         if (mask_type != TCA_FLOWER_UNSPEC) {
1903                 err = nla_put(skb, mask_type, len, mask);
1904                 if (err)
1905                         return err;
1906         }
1907         return 0;
1908 }
1909
1910 static int fl_dump_key_port_range(struct sk_buff *skb, struct fl_flow_key *key,
1911                                   struct fl_flow_key *mask)
1912 {
1913         if (fl_dump_key_val(skb, &key->tp_min.dst, TCA_FLOWER_KEY_PORT_DST_MIN,
1914                             &mask->tp_min.dst, TCA_FLOWER_UNSPEC,
1915                             sizeof(key->tp_min.dst)) ||
1916             fl_dump_key_val(skb, &key->tp_max.dst, TCA_FLOWER_KEY_PORT_DST_MAX,
1917                             &mask->tp_max.dst, TCA_FLOWER_UNSPEC,
1918                             sizeof(key->tp_max.dst)) ||
1919             fl_dump_key_val(skb, &key->tp_min.src, TCA_FLOWER_KEY_PORT_SRC_MIN,
1920                             &mask->tp_min.src, TCA_FLOWER_UNSPEC,
1921                             sizeof(key->tp_min.src)) ||
1922             fl_dump_key_val(skb, &key->tp_max.src, TCA_FLOWER_KEY_PORT_SRC_MAX,
1923                             &mask->tp_max.src, TCA_FLOWER_UNSPEC,
1924                             sizeof(key->tp_max.src)))
1925                 return -1;
1926
1927         return 0;
1928 }
1929
1930 static int fl_dump_key_mpls(struct sk_buff *skb,
1931                             struct flow_dissector_key_mpls *mpls_key,
1932                             struct flow_dissector_key_mpls *mpls_mask)
1933 {
1934         int err;
1935
1936         if (!memchr_inv(mpls_mask, 0, sizeof(*mpls_mask)))
1937                 return 0;
1938         if (mpls_mask->mpls_ttl) {
1939                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_TTL,
1940                                  mpls_key->mpls_ttl);
1941                 if (err)
1942                         return err;
1943         }
1944         if (mpls_mask->mpls_tc) {
1945                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_TC,
1946                                  mpls_key->mpls_tc);
1947                 if (err)
1948                         return err;
1949         }
1950         if (mpls_mask->mpls_label) {
1951                 err = nla_put_u32(skb, TCA_FLOWER_KEY_MPLS_LABEL,
1952                                   mpls_key->mpls_label);
1953                 if (err)
1954                         return err;
1955         }
1956         if (mpls_mask->mpls_bos) {
1957                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_BOS,
1958                                  mpls_key->mpls_bos);
1959                 if (err)
1960                         return err;
1961         }
1962         return 0;
1963 }
1964
1965 static int fl_dump_key_ip(struct sk_buff *skb, bool encap,
1966                           struct flow_dissector_key_ip *key,
1967                           struct flow_dissector_key_ip *mask)
1968 {
1969         int tos_key = encap ? TCA_FLOWER_KEY_ENC_IP_TOS : TCA_FLOWER_KEY_IP_TOS;
1970         int ttl_key = encap ? TCA_FLOWER_KEY_ENC_IP_TTL : TCA_FLOWER_KEY_IP_TTL;
1971         int tos_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TOS_MASK : TCA_FLOWER_KEY_IP_TOS_MASK;
1972         int ttl_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TTL_MASK : TCA_FLOWER_KEY_IP_TTL_MASK;
1973
1974         if (fl_dump_key_val(skb, &key->tos, tos_key, &mask->tos, tos_mask, sizeof(key->tos)) ||
1975             fl_dump_key_val(skb, &key->ttl, ttl_key, &mask->ttl, ttl_mask, sizeof(key->ttl)))
1976                 return -1;
1977
1978         return 0;
1979 }
1980
1981 static int fl_dump_key_vlan(struct sk_buff *skb,
1982                             int vlan_id_key, int vlan_prio_key,
1983                             struct flow_dissector_key_vlan *vlan_key,
1984                             struct flow_dissector_key_vlan *vlan_mask)
1985 {
1986         int err;
1987
1988         if (!memchr_inv(vlan_mask, 0, sizeof(*vlan_mask)))
1989                 return 0;
1990         if (vlan_mask->vlan_id) {
1991                 err = nla_put_u16(skb, vlan_id_key,
1992                                   vlan_key->vlan_id);
1993                 if (err)
1994                         return err;
1995         }
1996         if (vlan_mask->vlan_priority) {
1997                 err = nla_put_u8(skb, vlan_prio_key,
1998                                  vlan_key->vlan_priority);
1999                 if (err)
2000                         return err;
2001         }
2002         return 0;
2003 }
2004
2005 static void fl_get_key_flag(u32 dissector_key, u32 dissector_mask,
2006                             u32 *flower_key, u32 *flower_mask,
2007                             u32 flower_flag_bit, u32 dissector_flag_bit)
2008 {
2009         if (dissector_mask & dissector_flag_bit) {
2010                 *flower_mask |= flower_flag_bit;
2011                 if (dissector_key & dissector_flag_bit)
2012                         *flower_key |= flower_flag_bit;
2013         }
2014 }
2015
2016 static int fl_dump_key_flags(struct sk_buff *skb, u32 flags_key, u32 flags_mask)
2017 {
2018         u32 key, mask;
2019         __be32 _key, _mask;
2020         int err;
2021
2022         if (!memchr_inv(&flags_mask, 0, sizeof(flags_mask)))
2023                 return 0;
2024
2025         key = 0;
2026         mask = 0;
2027
2028         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
2029                         TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT, FLOW_DIS_IS_FRAGMENT);
2030         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
2031                         TCA_FLOWER_KEY_FLAGS_FRAG_IS_FIRST,
2032                         FLOW_DIS_FIRST_FRAG);
2033
2034         _key = cpu_to_be32(key);
2035         _mask = cpu_to_be32(mask);
2036
2037         err = nla_put(skb, TCA_FLOWER_KEY_FLAGS, 4, &_key);
2038         if (err)
2039                 return err;
2040
2041         return nla_put(skb, TCA_FLOWER_KEY_FLAGS_MASK, 4, &_mask);
2042 }
2043
2044 static int fl_dump_key_geneve_opt(struct sk_buff *skb,
2045                                   struct flow_dissector_key_enc_opts *enc_opts)
2046 {
2047         struct geneve_opt *opt;
2048         struct nlattr *nest;
2049         int opt_off = 0;
2050
2051         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_GENEVE);
2052         if (!nest)
2053                 goto nla_put_failure;
2054
2055         while (enc_opts->len > opt_off) {
2056                 opt = (struct geneve_opt *)&enc_opts->data[opt_off];
2057
2058                 if (nla_put_be16(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS,
2059                                  opt->opt_class))
2060                         goto nla_put_failure;
2061                 if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE,
2062                                opt->type))
2063                         goto nla_put_failure;
2064                 if (nla_put(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA,
2065                             opt->length * 4, opt->opt_data))
2066                         goto nla_put_failure;
2067
2068                 opt_off += sizeof(struct geneve_opt) + opt->length * 4;
2069         }
2070         nla_nest_end(skb, nest);
2071         return 0;
2072
2073 nla_put_failure:
2074         nla_nest_cancel(skb, nest);
2075         return -EMSGSIZE;
2076 }
2077
2078 static int fl_dump_key_options(struct sk_buff *skb, int enc_opt_type,
2079                                struct flow_dissector_key_enc_opts *enc_opts)
2080 {
2081         struct nlattr *nest;
2082         int err;
2083
2084         if (!enc_opts->len)
2085                 return 0;
2086
2087         nest = nla_nest_start_noflag(skb, enc_opt_type);
2088         if (!nest)
2089                 goto nla_put_failure;
2090
2091         switch (enc_opts->dst_opt_type) {
2092         case TUNNEL_GENEVE_OPT:
2093                 err = fl_dump_key_geneve_opt(skb, enc_opts);
2094                 if (err)
2095                         goto nla_put_failure;
2096                 break;
2097         default:
2098                 goto nla_put_failure;
2099         }
2100         nla_nest_end(skb, nest);
2101         return 0;
2102
2103 nla_put_failure:
2104         nla_nest_cancel(skb, nest);
2105         return -EMSGSIZE;
2106 }
2107
2108 static int fl_dump_key_enc_opt(struct sk_buff *skb,
2109                                struct flow_dissector_key_enc_opts *key_opts,
2110                                struct flow_dissector_key_enc_opts *msk_opts)
2111 {
2112         int err;
2113
2114         err = fl_dump_key_options(skb, TCA_FLOWER_KEY_ENC_OPTS, key_opts);
2115         if (err)
2116                 return err;
2117
2118         return fl_dump_key_options(skb, TCA_FLOWER_KEY_ENC_OPTS_MASK, msk_opts);
2119 }
2120
2121 static int fl_dump_key(struct sk_buff *skb, struct net *net,
2122                        struct fl_flow_key *key, struct fl_flow_key *mask)
2123 {
2124         if (mask->indev_ifindex) {
2125                 struct net_device *dev;
2126
2127                 dev = __dev_get_by_index(net, key->indev_ifindex);
2128                 if (dev && nla_put_string(skb, TCA_FLOWER_INDEV, dev->name))
2129                         goto nla_put_failure;
2130         }
2131
2132         if (fl_dump_key_val(skb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
2133                             mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
2134                             sizeof(key->eth.dst)) ||
2135             fl_dump_key_val(skb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
2136                             mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
2137                             sizeof(key->eth.src)) ||
2138             fl_dump_key_val(skb, &key->basic.n_proto, TCA_FLOWER_KEY_ETH_TYPE,
2139                             &mask->basic.n_proto, TCA_FLOWER_UNSPEC,
2140                             sizeof(key->basic.n_proto)))
2141                 goto nla_put_failure;
2142
2143         if (fl_dump_key_mpls(skb, &key->mpls, &mask->mpls))
2144                 goto nla_put_failure;
2145
2146         if (fl_dump_key_vlan(skb, TCA_FLOWER_KEY_VLAN_ID,
2147                              TCA_FLOWER_KEY_VLAN_PRIO, &key->vlan, &mask->vlan))
2148                 goto nla_put_failure;
2149
2150         if (fl_dump_key_vlan(skb, TCA_FLOWER_KEY_CVLAN_ID,
2151                              TCA_FLOWER_KEY_CVLAN_PRIO,
2152                              &key->cvlan, &mask->cvlan) ||
2153             (mask->cvlan.vlan_tpid &&
2154              nla_put_be16(skb, TCA_FLOWER_KEY_VLAN_ETH_TYPE,
2155                           key->cvlan.vlan_tpid)))
2156                 goto nla_put_failure;
2157
2158         if (mask->basic.n_proto) {
2159                 if (mask->cvlan.vlan_tpid) {
2160                         if (nla_put_be16(skb, TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
2161                                          key->basic.n_proto))
2162                                 goto nla_put_failure;
2163                 } else if (mask->vlan.vlan_tpid) {
2164                         if (nla_put_be16(skb, TCA_FLOWER_KEY_VLAN_ETH_TYPE,
2165                                          key->basic.n_proto))
2166                                 goto nla_put_failure;
2167                 }
2168         }
2169
2170         if ((key->basic.n_proto == htons(ETH_P_IP) ||
2171              key->basic.n_proto == htons(ETH_P_IPV6)) &&
2172             (fl_dump_key_val(skb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
2173                             &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
2174                             sizeof(key->basic.ip_proto)) ||
2175             fl_dump_key_ip(skb, false, &key->ip, &mask->ip)))
2176                 goto nla_put_failure;
2177
2178         if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
2179             (fl_dump_key_val(skb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
2180                              &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
2181                              sizeof(key->ipv4.src)) ||
2182              fl_dump_key_val(skb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
2183                              &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
2184                              sizeof(key->ipv4.dst))))
2185                 goto nla_put_failure;
2186         else if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
2187                  (fl_dump_key_val(skb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
2188                                   &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
2189                                   sizeof(key->ipv6.src)) ||
2190                   fl_dump_key_val(skb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
2191                                   &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
2192                                   sizeof(key->ipv6.dst))))
2193                 goto nla_put_failure;
2194
2195         if (key->basic.ip_proto == IPPROTO_TCP &&
2196             (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
2197                              &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
2198                              sizeof(key->tp.src)) ||
2199              fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
2200                              &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
2201                              sizeof(key->tp.dst)) ||
2202              fl_dump_key_val(skb, &key->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS,
2203                              &mask->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS_MASK,
2204                              sizeof(key->tcp.flags))))
2205                 goto nla_put_failure;
2206         else if (key->basic.ip_proto == IPPROTO_UDP &&
2207                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
2208                                   &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
2209                                   sizeof(key->tp.src)) ||
2210                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
2211                                   &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
2212                                   sizeof(key->tp.dst))))
2213                 goto nla_put_failure;
2214         else if (key->basic.ip_proto == IPPROTO_SCTP &&
2215                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_SCTP_SRC,
2216                                   &mask->tp.src, TCA_FLOWER_KEY_SCTP_SRC_MASK,
2217                                   sizeof(key->tp.src)) ||
2218                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_SCTP_DST,
2219                                   &mask->tp.dst, TCA_FLOWER_KEY_SCTP_DST_MASK,
2220                                   sizeof(key->tp.dst))))
2221                 goto nla_put_failure;
2222         else if (key->basic.n_proto == htons(ETH_P_IP) &&
2223                  key->basic.ip_proto == IPPROTO_ICMP &&
2224                  (fl_dump_key_val(skb, &key->icmp.type,
2225                                   TCA_FLOWER_KEY_ICMPV4_TYPE, &mask->icmp.type,
2226                                   TCA_FLOWER_KEY_ICMPV4_TYPE_MASK,
2227                                   sizeof(key->icmp.type)) ||
2228                   fl_dump_key_val(skb, &key->icmp.code,
2229                                   TCA_FLOWER_KEY_ICMPV4_CODE, &mask->icmp.code,
2230                                   TCA_FLOWER_KEY_ICMPV4_CODE_MASK,
2231                                   sizeof(key->icmp.code))))
2232                 goto nla_put_failure;
2233         else if (key->basic.n_proto == htons(ETH_P_IPV6) &&
2234                  key->basic.ip_proto == IPPROTO_ICMPV6 &&
2235                  (fl_dump_key_val(skb, &key->icmp.type,
2236                                   TCA_FLOWER_KEY_ICMPV6_TYPE, &mask->icmp.type,
2237                                   TCA_FLOWER_KEY_ICMPV6_TYPE_MASK,
2238                                   sizeof(key->icmp.type)) ||
2239                   fl_dump_key_val(skb, &key->icmp.code,
2240                                   TCA_FLOWER_KEY_ICMPV6_CODE, &mask->icmp.code,
2241                                   TCA_FLOWER_KEY_ICMPV6_CODE_MASK,
2242                                   sizeof(key->icmp.code))))
2243                 goto nla_put_failure;
2244         else if ((key->basic.n_proto == htons(ETH_P_ARP) ||
2245                   key->basic.n_proto == htons(ETH_P_RARP)) &&
2246                  (fl_dump_key_val(skb, &key->arp.sip,
2247                                   TCA_FLOWER_KEY_ARP_SIP, &mask->arp.sip,
2248                                   TCA_FLOWER_KEY_ARP_SIP_MASK,
2249                                   sizeof(key->arp.sip)) ||
2250                   fl_dump_key_val(skb, &key->arp.tip,
2251                                   TCA_FLOWER_KEY_ARP_TIP, &mask->arp.tip,
2252                                   TCA_FLOWER_KEY_ARP_TIP_MASK,
2253                                   sizeof(key->arp.tip)) ||
2254                   fl_dump_key_val(skb, &key->arp.op,
2255                                   TCA_FLOWER_KEY_ARP_OP, &mask->arp.op,
2256                                   TCA_FLOWER_KEY_ARP_OP_MASK,
2257                                   sizeof(key->arp.op)) ||
2258                   fl_dump_key_val(skb, key->arp.sha, TCA_FLOWER_KEY_ARP_SHA,
2259                                   mask->arp.sha, TCA_FLOWER_KEY_ARP_SHA_MASK,
2260                                   sizeof(key->arp.sha)) ||
2261                   fl_dump_key_val(skb, key->arp.tha, TCA_FLOWER_KEY_ARP_THA,
2262                                   mask->arp.tha, TCA_FLOWER_KEY_ARP_THA_MASK,
2263                                   sizeof(key->arp.tha))))
2264                 goto nla_put_failure;
2265
2266         if ((key->basic.ip_proto == IPPROTO_TCP ||
2267              key->basic.ip_proto == IPPROTO_UDP ||
2268              key->basic.ip_proto == IPPROTO_SCTP) &&
2269              fl_dump_key_port_range(skb, key, mask))
2270                 goto nla_put_failure;
2271
2272         if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
2273             (fl_dump_key_val(skb, &key->enc_ipv4.src,
2274                             TCA_FLOWER_KEY_ENC_IPV4_SRC, &mask->enc_ipv4.src,
2275                             TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
2276                             sizeof(key->enc_ipv4.src)) ||
2277              fl_dump_key_val(skb, &key->enc_ipv4.dst,
2278                              TCA_FLOWER_KEY_ENC_IPV4_DST, &mask->enc_ipv4.dst,
2279                              TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
2280                              sizeof(key->enc_ipv4.dst))))
2281                 goto nla_put_failure;
2282         else if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
2283                  (fl_dump_key_val(skb, &key->enc_ipv6.src,
2284                             TCA_FLOWER_KEY_ENC_IPV6_SRC, &mask->enc_ipv6.src,
2285                             TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
2286                             sizeof(key->enc_ipv6.src)) ||
2287                  fl_dump_key_val(skb, &key->enc_ipv6.dst,
2288                                  TCA_FLOWER_KEY_ENC_IPV6_DST,
2289                                  &mask->enc_ipv6.dst,
2290                                  TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
2291                             sizeof(key->enc_ipv6.dst))))
2292                 goto nla_put_failure;
2293
2294         if (fl_dump_key_val(skb, &key->enc_key_id, TCA_FLOWER_KEY_ENC_KEY_ID,
2295                             &mask->enc_key_id, TCA_FLOWER_UNSPEC,
2296                             sizeof(key->enc_key_id)) ||
2297             fl_dump_key_val(skb, &key->enc_tp.src,
2298                             TCA_FLOWER_KEY_ENC_UDP_SRC_PORT,
2299                             &mask->enc_tp.src,
2300                             TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK,
2301                             sizeof(key->enc_tp.src)) ||
2302             fl_dump_key_val(skb, &key->enc_tp.dst,
2303                             TCA_FLOWER_KEY_ENC_UDP_DST_PORT,
2304                             &mask->enc_tp.dst,
2305                             TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK,
2306                             sizeof(key->enc_tp.dst)) ||
2307             fl_dump_key_ip(skb, true, &key->enc_ip, &mask->enc_ip) ||
2308             fl_dump_key_enc_opt(skb, &key->enc_opts, &mask->enc_opts))
2309                 goto nla_put_failure;
2310
2311         if (fl_dump_key_flags(skb, key->control.flags, mask->control.flags))
2312                 goto nla_put_failure;
2313
2314         return 0;
2315
2316 nla_put_failure:
2317         return -EMSGSIZE;
2318 }
2319
2320 static int fl_dump(struct net *net, struct tcf_proto *tp, void *fh,
2321                    struct sk_buff *skb, struct tcmsg *t, bool rtnl_held)
2322 {
2323         struct cls_fl_filter *f = fh;
2324         struct nlattr *nest;
2325         struct fl_flow_key *key, *mask;
2326         bool skip_hw;
2327
2328         if (!f)
2329                 return skb->len;
2330
2331         t->tcm_handle = f->handle;
2332
2333         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
2334         if (!nest)
2335                 goto nla_put_failure;
2336
2337         spin_lock(&tp->lock);
2338
2339         if (f->res.classid &&
2340             nla_put_u32(skb, TCA_FLOWER_CLASSID, f->res.classid))
2341                 goto nla_put_failure_locked;
2342
2343         key = &f->key;
2344         mask = &f->mask->key;
2345         skip_hw = tc_skip_hw(f->flags);
2346
2347         if (fl_dump_key(skb, net, key, mask))
2348                 goto nla_put_failure_locked;
2349
2350         if (f->flags && nla_put_u32(skb, TCA_FLOWER_FLAGS, f->flags))
2351                 goto nla_put_failure_locked;
2352
2353         spin_unlock(&tp->lock);
2354
2355         if (!skip_hw)
2356                 fl_hw_update_stats(tp, f, rtnl_held);
2357
2358         if (nla_put_u32(skb, TCA_FLOWER_IN_HW_COUNT, f->in_hw_count))
2359                 goto nla_put_failure;
2360
2361         if (tcf_exts_dump(skb, &f->exts))
2362                 goto nla_put_failure;
2363
2364         nla_nest_end(skb, nest);
2365
2366         if (tcf_exts_dump_stats(skb, &f->exts) < 0)
2367                 goto nla_put_failure;
2368
2369         return skb->len;
2370
2371 nla_put_failure_locked:
2372         spin_unlock(&tp->lock);
2373 nla_put_failure:
2374         nla_nest_cancel(skb, nest);
2375         return -1;
2376 }
2377
2378 static int fl_tmplt_dump(struct sk_buff *skb, struct net *net, void *tmplt_priv)
2379 {
2380         struct fl_flow_tmplt *tmplt = tmplt_priv;
2381         struct fl_flow_key *key, *mask;
2382         struct nlattr *nest;
2383
2384         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
2385         if (!nest)
2386                 goto nla_put_failure;
2387
2388         key = &tmplt->dummy_key;
2389         mask = &tmplt->mask;
2390
2391         if (fl_dump_key(skb, net, key, mask))
2392                 goto nla_put_failure;
2393
2394         nla_nest_end(skb, nest);
2395
2396         return skb->len;
2397
2398 nla_put_failure:
2399         nla_nest_cancel(skb, nest);
2400         return -EMSGSIZE;
2401 }
2402
2403 static void fl_bind_class(void *fh, u32 classid, unsigned long cl)
2404 {
2405         struct cls_fl_filter *f = fh;
2406
2407         if (f && f->res.classid == classid)
2408                 f->res.class = cl;
2409 }
2410
2411 static struct tcf_proto_ops cls_fl_ops __read_mostly = {
2412         .kind           = "flower",
2413         .classify       = fl_classify,
2414         .init           = fl_init,
2415         .destroy        = fl_destroy,
2416         .get            = fl_get,
2417         .put            = fl_put,
2418         .change         = fl_change,
2419         .delete         = fl_delete,
2420         .walk           = fl_walk,
2421         .reoffload      = fl_reoffload,
2422         .dump           = fl_dump,
2423         .bind_class     = fl_bind_class,
2424         .tmplt_create   = fl_tmplt_create,
2425         .tmplt_destroy  = fl_tmplt_destroy,
2426         .tmplt_dump     = fl_tmplt_dump,
2427         .owner          = THIS_MODULE,
2428         .flags          = TCF_PROTO_OPS_DOIT_UNLOCKED,
2429 };
2430
2431 static int __init cls_fl_init(void)
2432 {
2433         return register_tcf_proto_ops(&cls_fl_ops);
2434 }
2435
2436 static void __exit cls_fl_exit(void)
2437 {
2438         unregister_tcf_proto_ops(&cls_fl_ops);
2439 }
2440
2441 module_init(cls_fl_init);
2442 module_exit(cls_fl_exit);
2443
2444 MODULE_AUTHOR("Jiri Pirko <jiri@resnulli.us>");
2445 MODULE_DESCRIPTION("Flower classifier");
2446 MODULE_LICENSE("GPL v2");