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 #include <linux/ppp_defs.h>
  20
  21 #include <net/sch_generic.h>
  22 #include <net/pkt_cls.h>
  23 #include <net/pkt_sched.h>
  24 #include <net/ip.h>
  25 #include <net/flow_dissector.h>
  26 #include <net/geneve.h>
  27 #include <net/vxlan.h>
  28 #include <net/erspan.h>
  29 #include <net/gtp.h>
  30
  31 #include <net/dst.h>
  32 #include <net/dst_metadata.h>
  33
  34 #include <uapi/linux/netfilter/nf_conntrack_common.h>
  35
  36 #define TCA_FLOWER_KEY_CT_FLAGS_MAX \
  37                 ((__TCA_FLOWER_KEY_CT_FLAGS_MAX - 1) << 1)
  38 #define TCA_FLOWER_KEY_CT_FLAGS_MASK \
  39                 (TCA_FLOWER_KEY_CT_FLAGS_MAX - 1)
  40
  41 struct fl_flow_key {
  42         struct flow_dissector_key_meta meta;
  43         struct flow_dissector_key_control control;
  44         struct flow_dissector_key_control enc_control;
  45         struct flow_dissector_key_basic basic;
  46         struct flow_dissector_key_eth_addrs eth;
  47         struct flow_dissector_key_vlan vlan;
  48         struct flow_dissector_key_vlan cvlan;
  49         union {
  50                 struct flow_dissector_key_ipv4_addrs ipv4;
  51                 struct flow_dissector_key_ipv6_addrs ipv6;
  52         };
  53         struct flow_dissector_key_ports tp;
  54         struct flow_dissector_key_icmp icmp;
  55         struct flow_dissector_key_arp arp;
  56         struct flow_dissector_key_keyid enc_key_id;
  57         union {
  58                 struct flow_dissector_key_ipv4_addrs enc_ipv4;
  59                 struct flow_dissector_key_ipv6_addrs enc_ipv6;
  60         };
  61         struct flow_dissector_key_ports enc_tp;
  62         struct flow_dissector_key_mpls mpls;
  63         struct flow_dissector_key_tcp tcp;
  64         struct flow_dissector_key_ip ip;
  65         struct flow_dissector_key_ip enc_ip;
  66         struct flow_dissector_key_enc_opts enc_opts;
  67         struct flow_dissector_key_ports_range tp_range;
  68         struct flow_dissector_key_ct ct;
  69         struct flow_dissector_key_hash hash;
  70         struct flow_dissector_key_num_of_vlans num_of_vlans;
  71         struct flow_dissector_key_pppoe pppoe;
  72         struct flow_dissector_key_l2tpv3 l2tpv3;
  73 } __aligned(BITS_PER_LONG / 8); /* Ensure that we can do comparisons as longs. */
  74
  75 struct fl_flow_mask_range {
  76         unsigned short int start;
  77         unsigned short int end;
  78 };
  79
  80 struct fl_flow_mask {
  81         struct fl_flow_key key;
  82         struct fl_flow_mask_range range;
  83         u32 flags;
  84         struct rhash_head ht_node;
  85         struct rhashtable ht;
  86         struct rhashtable_params filter_ht_params;
  87         struct flow_dissector dissector;
  88         struct list_head filters;
  89         struct rcu_work rwork;
  90         struct list_head list;
  91         refcount_t refcnt;
  92 };
  93
  94 struct fl_flow_tmplt {
  95         struct fl_flow_key dummy_key;
  96         struct fl_flow_key mask;
  97         struct flow_dissector dissector;
  98         struct tcf_chain *chain;
  99 };
 100
 101 struct cls_fl_head {
 102         struct rhashtable ht;
 103         spinlock_t masks_lock; /* Protect masks list */
 104         struct list_head masks;
 105         struct list_head hw_filters;
 106         struct rcu_work rwork;
 107         struct idr handle_idr;
 108 };
 109
 110 struct cls_fl_filter {
 111         struct fl_flow_mask *mask;
 112         struct rhash_head ht_node;
 113         struct fl_flow_key mkey;
 114         struct tcf_exts exts;
 115         struct tcf_result res;
 116         struct fl_flow_key key;
 117         struct list_head list;
 118         struct list_head hw_list;
 119         u32 handle;
 120         u32 flags;
 121         u32 in_hw_count;
 122         struct rcu_work rwork;
 123         struct net_device *hw_dev;
 124         /* Flower classifier is unlocked, which means that its reference counter
 125          * can be changed concurrently without any kind of external
 126          * synchronization. Use atomic reference counter to be concurrency-safe.
 127          */
 128         refcount_t refcnt;
 129         bool deleted;
 130 };
 131
 132 static const struct rhashtable_params mask_ht_params = {
 133         .key_offset = offsetof(struct fl_flow_mask, key),
 134         .key_len = sizeof(struct fl_flow_key),
 135         .head_offset = offsetof(struct fl_flow_mask, ht_node),
 136         .automatic_shrinking = true,
 137 };
 138
 139 static unsigned short int fl_mask_range(const struct fl_flow_mask *mask)
 140 {
 141         return mask->range.end - mask->range.start;
 142 }
 143
 144 static void fl_mask_update_range(struct fl_flow_mask *mask)
 145 {
 146         const u8 *bytes = (const u8 *) &mask->key;
 147         size_t size = sizeof(mask->key);
 148         size_t i, first = 0, last;
 149
 150         for (i = 0; i < size; i++) {
 151                 if (bytes[i]) {
 152                         first = i;
 153                         break;
 154                 }
 155         }
 156         last = first;
 157         for (i = size - 1; i != first; i--) {
 158                 if (bytes[i]) {
 159                         last = i;
 160                         break;
 161                 }
 162         }
 163         mask->range.start = rounddown(first, sizeof(long));
 164         mask->range.end = roundup(last + 1, sizeof(long));
 165 }
 166
 167 static void *fl_key_get_start(struct fl_flow_key *key,
 168                               const struct fl_flow_mask *mask)
 169 {
 170         return (u8 *) key + mask->range.start;
 171 }
 172
 173 static void fl_set_masked_key(struct fl_flow_key *mkey, struct fl_flow_key *key,
 174                               struct fl_flow_mask *mask)
 175 {
 176         const long *lkey = fl_key_get_start(key, mask);
 177         const long *lmask = fl_key_get_start(&mask->key, mask);
 178         long *lmkey = fl_key_get_start(mkey, mask);
 179         int i;
 180
 181         for (i = 0; i < fl_mask_range(mask); i += sizeof(long))
 182                 *lmkey++ = *lkey++ & *lmask++;
 183 }
 184
 185 static bool fl_mask_fits_tmplt(struct fl_flow_tmplt *tmplt,
 186                                struct fl_flow_mask *mask)
 187 {
 188         const long *lmask = fl_key_get_start(&mask->key, mask);
 189         const long *ltmplt;
 190         int i;
 191
 192         if (!tmplt)
 193                 return true;
 194         ltmplt = fl_key_get_start(&tmplt->mask, mask);
 195         for (i = 0; i < fl_mask_range(mask); i += sizeof(long)) {
 196                 if (~*ltmplt++ & *lmask++)
 197                         return false;
 198         }
 199         return true;
 200 }
 201
 202 static void fl_clear_masked_range(struct fl_flow_key *key,
 203                                   struct fl_flow_mask *mask)
 204 {
 205         memset(fl_key_get_start(key, mask), 0, fl_mask_range(mask));
 206 }
 207
 208 static bool fl_range_port_dst_cmp(struct cls_fl_filter *filter,
 209                                   struct fl_flow_key *key,
 210                                   struct fl_flow_key *mkey)
 211 {
 212         u16 min_mask, max_mask, min_val, max_val;
 213
 214         min_mask = ntohs(filter->mask->key.tp_range.tp_min.dst);
 215         max_mask = ntohs(filter->mask->key.tp_range.tp_max.dst);
 216         min_val = ntohs(filter->key.tp_range.tp_min.dst);
 217         max_val = ntohs(filter->key.tp_range.tp_max.dst);
 218
 219         if (min_mask && max_mask) {
 220                 if (ntohs(key->tp_range.tp.dst) < min_val ||
 221                     ntohs(key->tp_range.tp.dst) > max_val)
 222                         return false;
 223
 224                 /* skb does not have min and max values */
 225                 mkey->tp_range.tp_min.dst = filter->mkey.tp_range.tp_min.dst;
 226                 mkey->tp_range.tp_max.dst = filter->mkey.tp_range.tp_max.dst;
 227         }
 228         return true;
 229 }
 230
 231 static bool fl_range_port_src_cmp(struct cls_fl_filter *filter,
 232                                   struct fl_flow_key *key,
 233                                   struct fl_flow_key *mkey)
 234 {
 235         u16 min_mask, max_mask, min_val, max_val;
 236
 237         min_mask = ntohs(filter->mask->key.tp_range.tp_min.src);
 238         max_mask = ntohs(filter->mask->key.tp_range.tp_max.src);
 239         min_val = ntohs(filter->key.tp_range.tp_min.src);
 240         max_val = ntohs(filter->key.tp_range.tp_max.src);
 241
 242         if (min_mask && max_mask) {
 243                 if (ntohs(key->tp_range.tp.src) < min_val ||
 244                     ntohs(key->tp_range.tp.src) > max_val)
 245                         return false;
 246
 247                 /* skb does not have min and max values */
 248                 mkey->tp_range.tp_min.src = filter->mkey.tp_range.tp_min.src;
 249                 mkey->tp_range.tp_max.src = filter->mkey.tp_range.tp_max.src;
 250         }
 251         return true;
 252 }
 253
 254 static struct cls_fl_filter *__fl_lookup(struct fl_flow_mask *mask,
 255                                          struct fl_flow_key *mkey)
 256 {
 257         return rhashtable_lookup_fast(&mask->ht, fl_key_get_start(mkey, mask),
 258                                       mask->filter_ht_params);
 259 }
 260
 261 static struct cls_fl_filter *fl_lookup_range(struct fl_flow_mask *mask,
 262                                              struct fl_flow_key *mkey,
 263                                              struct fl_flow_key *key)
 264 {
 265         struct cls_fl_filter *filter, *f;
 266
 267         list_for_each_entry_rcu(filter, &mask->filters, list) {
 268                 if (!fl_range_port_dst_cmp(filter, key, mkey))
 269                         continue;
 270
 271                 if (!fl_range_port_src_cmp(filter, key, mkey))
 272                         continue;
 273
 274                 f = __fl_lookup(mask, mkey);
 275                 if (f)
 276                         return f;
 277         }
 278         return NULL;
 279 }
 280
 281 static noinline_for_stack
 282 struct cls_fl_filter *fl_mask_lookup(struct fl_flow_mask *mask, struct fl_flow_key *key)
 283 {
 284         struct fl_flow_key mkey;
 285
 286         fl_set_masked_key(&mkey, key, mask);
 287         if ((mask->flags & TCA_FLOWER_MASK_FLAGS_RANGE))
 288                 return fl_lookup_range(mask, &mkey, key);
 289
 290         return __fl_lookup(mask, &mkey);
 291 }
 292
 293 static u16 fl_ct_info_to_flower_map[] = {
 294         [IP_CT_ESTABLISHED] =           TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 295                                         TCA_FLOWER_KEY_CT_FLAGS_ESTABLISHED,
 296         [IP_CT_RELATED] =               TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 297                                         TCA_FLOWER_KEY_CT_FLAGS_RELATED,
 298         [IP_CT_ESTABLISHED_REPLY] =     TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 299                                         TCA_FLOWER_KEY_CT_FLAGS_ESTABLISHED |
 300                                         TCA_FLOWER_KEY_CT_FLAGS_REPLY,
 301         [IP_CT_RELATED_REPLY] =         TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 302                                         TCA_FLOWER_KEY_CT_FLAGS_RELATED |
 303                                         TCA_FLOWER_KEY_CT_FLAGS_REPLY,
 304         [IP_CT_NEW] =                   TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 305                                         TCA_FLOWER_KEY_CT_FLAGS_NEW,
 306 };
 307
 308 static int fl_classify(struct sk_buff *skb, const struct tcf_proto *tp,
 309                        struct tcf_result *res)
 310 {
 311         struct cls_fl_head *head = rcu_dereference_bh(tp->root);
 312         bool post_ct = tc_skb_cb(skb)->post_ct;
 313         u16 zone = tc_skb_cb(skb)->zone;
 314         struct fl_flow_key skb_key;
 315         struct fl_flow_mask *mask;
 316         struct cls_fl_filter *f;
 317
 318         list_for_each_entry_rcu(mask, &head->masks, list) {
 319                 flow_dissector_init_keys(&skb_key.control, &skb_key.basic);
 320                 fl_clear_masked_range(&skb_key, mask);
 321
 322                 skb_flow_dissect_meta(skb, &mask->dissector, &skb_key);
 323                 /* skb_flow_dissect() does not set n_proto in case an unknown
 324                  * protocol, so do it rather here.
 325                  */
 326                 skb_key.basic.n_proto = skb_protocol(skb, false);
 327                 skb_flow_dissect_tunnel_info(skb, &mask->dissector, &skb_key);
 328                 skb_flow_dissect_ct(skb, &mask->dissector, &skb_key,
 329                                     fl_ct_info_to_flower_map,
 330                                     ARRAY_SIZE(fl_ct_info_to_flower_map),
 331                                     post_ct, zone);
 332                 skb_flow_dissect_hash(skb, &mask->dissector, &skb_key);
 333                 skb_flow_dissect(skb, &mask->dissector, &skb_key,
 334                                  FLOW_DISSECTOR_F_STOP_BEFORE_ENCAP);
 335
 336                 f = fl_mask_lookup(mask, &skb_key);
 337                 if (f && !tc_skip_sw(f->flags)) {
 338                         *res = f->res;
 339                         return tcf_exts_exec(skb, &f->exts, res);
 340                 }
 341         }
 342         return -1;
 343 }
 344
 345 static int fl_init(struct tcf_proto *tp)
 346 {
 347         struct cls_fl_head *head;
 348
 349         head = kzalloc(sizeof(*head), GFP_KERNEL);
 350         if (!head)
 351                 return -ENOBUFS;
 352
 353         spin_lock_init(&head->masks_lock);
 354         INIT_LIST_HEAD_RCU(&head->masks);
 355         INIT_LIST_HEAD(&head->hw_filters);
 356         rcu_assign_pointer(tp->root, head);
 357         idr_init(&head->handle_idr);
 358
 359         return rhashtable_init(&head->ht, &mask_ht_params);
 360 }
 361
 362 static void fl_mask_free(struct fl_flow_mask *mask, bool mask_init_done)
 363 {
 364         /* temporary masks don't have their filters list and ht initialized */
 365         if (mask_init_done) {
 366                 WARN_ON(!list_empty(&mask->filters));
 367                 rhashtable_destroy(&mask->ht);
 368         }
 369         kfree(mask);
 370 }
 371
 372 static void fl_mask_free_work(struct work_struct *work)
 373 {
 374         struct fl_flow_mask *mask = container_of(to_rcu_work(work),
 375                                                  struct fl_flow_mask, rwork);
 376
 377         fl_mask_free(mask, true);
 378 }
 379
 380 static void fl_uninit_mask_free_work(struct work_struct *work)
 381 {
 382         struct fl_flow_mask *mask = container_of(to_rcu_work(work),
 383                                                  struct fl_flow_mask, rwork);
 384
 385         fl_mask_free(mask, false);
 386 }
 387
 388 static bool fl_mask_put(struct cls_fl_head *head, struct fl_flow_mask *mask)
 389 {
 390         if (!refcount_dec_and_test(&mask->refcnt))
 391                 return false;
 392
 393         rhashtable_remove_fast(&head->ht, &mask->ht_node, mask_ht_params);
 394
 395         spin_lock(&head->masks_lock);
 396         list_del_rcu(&mask->list);
 397         spin_unlock(&head->masks_lock);
 398
 399         tcf_queue_work(&mask->rwork, fl_mask_free_work);
 400
 401         return true;
 402 }
 403
 404 static struct cls_fl_head *fl_head_dereference(struct tcf_proto *tp)
 405 {
 406         /* Flower classifier only changes root pointer during init and destroy.
 407          * Users must obtain reference to tcf_proto instance before calling its
 408          * API, so tp->root pointer is protected from concurrent call to
 409          * fl_destroy() by reference counting.
 410          */
 411         return rcu_dereference_raw(tp->root);
 412 }
 413
 414 static void __fl_destroy_filter(struct cls_fl_filter *f)
 415 {
 416         tcf_exts_destroy(&f->exts);
 417         tcf_exts_put_net(&f->exts);
 418         kfree(f);
 419 }
 420
 421 static void fl_destroy_filter_work(struct work_struct *work)
 422 {
 423         struct cls_fl_filter *f = container_of(to_rcu_work(work),
 424                                         struct cls_fl_filter, rwork);
 425
 426         __fl_destroy_filter(f);
 427 }
 428
 429 static void fl_hw_destroy_filter(struct tcf_proto *tp, struct cls_fl_filter *f,
 430                                  bool rtnl_held, struct netlink_ext_ack *extack)
 431 {
 432         struct tcf_block *block = tp->chain->block;
 433         struct flow_cls_offload cls_flower = {};
 434
 435         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, extack);
 436         cls_flower.command = FLOW_CLS_DESTROY;
 437         cls_flower.cookie = (unsigned long) f;
 438
 439         tc_setup_cb_destroy(block, tp, TC_SETUP_CLSFLOWER, &cls_flower, false,
 440                             &f->flags, &f->in_hw_count, rtnl_held);
 441
 442 }
 443
 444 static int fl_hw_replace_filter(struct tcf_proto *tp,
 445                                 struct cls_fl_filter *f, bool rtnl_held,
 446                                 struct netlink_ext_ack *extack)
 447 {
 448         struct tcf_block *block = tp->chain->block;
 449         struct flow_cls_offload cls_flower = {};
 450         bool skip_sw = tc_skip_sw(f->flags);
 451         int err = 0;
 452
 453         cls_flower.rule = flow_rule_alloc(tcf_exts_num_actions(&f->exts));
 454         if (!cls_flower.rule)
 455                 return -ENOMEM;
 456
 457         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, extack);
 458         cls_flower.command = FLOW_CLS_REPLACE;
 459         cls_flower.cookie = (unsigned long) f;
 460         cls_flower.rule->match.dissector = &f->mask->dissector;
 461         cls_flower.rule->match.mask = &f->mask->key;
 462         cls_flower.rule->match.key = &f->mkey;
 463         cls_flower.classid = f->res.classid;
 464
 465         err = tc_setup_offload_action(&cls_flower.rule->action, &f->exts,
 466                                       cls_flower.common.extack);
 467         if (err) {
 468                 kfree(cls_flower.rule);
 469
 470                 return skip_sw ? err : 0;
 471         }
 472
 473         err = tc_setup_cb_add(block, tp, TC_SETUP_CLSFLOWER, &cls_flower,
 474                               skip_sw, &f->flags, &f->in_hw_count, rtnl_held);
 475         tc_cleanup_offload_action(&cls_flower.rule->action);
 476         kfree(cls_flower.rule);
 477
 478         if (err) {
 479                 fl_hw_destroy_filter(tp, f, rtnl_held, NULL);
 480                 return err;
 481         }
 482
 483         if (skip_sw && !(f->flags & TCA_CLS_FLAGS_IN_HW))
 484                 return -EINVAL;
 485
 486         return 0;
 487 }
 488
 489 static void fl_hw_update_stats(struct tcf_proto *tp, struct cls_fl_filter *f,
 490                                bool rtnl_held)
 491 {
 492         struct tcf_block *block = tp->chain->block;
 493         struct flow_cls_offload cls_flower = {};
 494
 495         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, NULL);
 496         cls_flower.command = FLOW_CLS_STATS;
 497         cls_flower.cookie = (unsigned long) f;
 498         cls_flower.classid = f->res.classid;
 499
 500         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false,
 501                          rtnl_held);
 502
 503         tcf_exts_hw_stats_update(&f->exts, cls_flower.stats.bytes,
 504                                  cls_flower.stats.pkts,
 505                                  cls_flower.stats.drops,
 506                                  cls_flower.stats.lastused,
 507                                  cls_flower.stats.used_hw_stats,
 508                                  cls_flower.stats.used_hw_stats_valid);
 509 }
 510
 511 static void __fl_put(struct cls_fl_filter *f)
 512 {
 513         if (!refcount_dec_and_test(&f->refcnt))
 514                 return;
 515
 516         if (tcf_exts_get_net(&f->exts))
 517                 tcf_queue_work(&f->rwork, fl_destroy_filter_work);
 518         else
 519                 __fl_destroy_filter(f);
 520 }
 521
 522 static struct cls_fl_filter *__fl_get(struct cls_fl_head *head, u32 handle)
 523 {
 524         struct cls_fl_filter *f;
 525
 526         rcu_read_lock();
 527         f = idr_find(&head->handle_idr, handle);
 528         if (f && !refcount_inc_not_zero(&f->refcnt))
 529                 f = NULL;
 530         rcu_read_unlock();
 531
 532         return f;
 533 }
 534
 535 static int __fl_delete(struct tcf_proto *tp, struct cls_fl_filter *f,
 536                        bool *last, bool rtnl_held,
 537                        struct netlink_ext_ack *extack)
 538 {
 539         struct cls_fl_head *head = fl_head_dereference(tp);
 540
 541         *last = false;
 542
 543         spin_lock(&tp->lock);
 544         if (f->deleted) {
 545                 spin_unlock(&tp->lock);
 546                 return -ENOENT;
 547         }
 548
 549         f->deleted = true;
 550         rhashtable_remove_fast(&f->mask->ht, &f->ht_node,
 551                                f->mask->filter_ht_params);
 552         idr_remove(&head->handle_idr, f->handle);
 553         list_del_rcu(&f->list);
 554         spin_unlock(&tp->lock);
 555
 556         *last = fl_mask_put(head, f->mask);
 557         if (!tc_skip_hw(f->flags))
 558                 fl_hw_destroy_filter(tp, f, rtnl_held, extack);
 559         tcf_unbind_filter(tp, &f->res);
 560         __fl_put(f);
 561
 562         return 0;
 563 }
 564
 565 static void fl_destroy_sleepable(struct work_struct *work)
 566 {
 567         struct cls_fl_head *head = container_of(to_rcu_work(work),
 568                                                 struct cls_fl_head,
 569                                                 rwork);
 570
 571         rhashtable_destroy(&head->ht);
 572         kfree(head);
 573         module_put(THIS_MODULE);
 574 }
 575
 576 static void fl_destroy(struct tcf_proto *tp, bool rtnl_held,
 577                        struct netlink_ext_ack *extack)
 578 {
 579         struct cls_fl_head *head = fl_head_dereference(tp);
 580         struct fl_flow_mask *mask, *next_mask;
 581         struct cls_fl_filter *f, *next;
 582         bool last;
 583
 584         list_for_each_entry_safe(mask, next_mask, &head->masks, list) {
 585                 list_for_each_entry_safe(f, next, &mask->filters, list) {
 586                         __fl_delete(tp, f, &last, rtnl_held, extack);
 587                         if (last)
 588                                 break;
 589                 }
 590         }
 591         idr_destroy(&head->handle_idr);
 592
 593         __module_get(THIS_MODULE);
 594         tcf_queue_work(&head->rwork, fl_destroy_sleepable);
 595 }
 596
 597 static void fl_put(struct tcf_proto *tp, void *arg)
 598 {
 599         struct cls_fl_filter *f = arg;
 600
 601         __fl_put(f);
 602 }
 603
 604 static void *fl_get(struct tcf_proto *tp, u32 handle)
 605 {
 606         struct cls_fl_head *head = fl_head_dereference(tp);
 607
 608         return __fl_get(head, handle);
 609 }
 610
 611 static const struct nla_policy fl_policy[TCA_FLOWER_MAX + 1] = {
 612         [TCA_FLOWER_UNSPEC]             = { .type = NLA_UNSPEC },
 613         [TCA_FLOWER_CLASSID]            = { .type = NLA_U32 },
 614         [TCA_FLOWER_INDEV]              = { .type = NLA_STRING,
 615                                             .len = IFNAMSIZ },
 616         [TCA_FLOWER_KEY_ETH_DST]        = { .len = ETH_ALEN },
 617         [TCA_FLOWER_KEY_ETH_DST_MASK]   = { .len = ETH_ALEN },
 618         [TCA_FLOWER_KEY_ETH_SRC]        = { .len = ETH_ALEN },
 619         [TCA_FLOWER_KEY_ETH_SRC_MASK]   = { .len = ETH_ALEN },
 620         [TCA_FLOWER_KEY_ETH_TYPE]       = { .type = NLA_U16 },
 621         [TCA_FLOWER_KEY_IP_PROTO]       = { .type = NLA_U8 },
 622         [TCA_FLOWER_KEY_IPV4_SRC]       = { .type = NLA_U32 },
 623         [TCA_FLOWER_KEY_IPV4_SRC_MASK]  = { .type = NLA_U32 },
 624         [TCA_FLOWER_KEY_IPV4_DST]       = { .type = NLA_U32 },
 625         [TCA_FLOWER_KEY_IPV4_DST_MASK]  = { .type = NLA_U32 },
 626         [TCA_FLOWER_KEY_IPV6_SRC]       = { .len = sizeof(struct in6_addr) },
 627         [TCA_FLOWER_KEY_IPV6_SRC_MASK]  = { .len = sizeof(struct in6_addr) },
 628         [TCA_FLOWER_KEY_IPV6_DST]       = { .len = sizeof(struct in6_addr) },
 629         [TCA_FLOWER_KEY_IPV6_DST_MASK]  = { .len = sizeof(struct in6_addr) },
 630         [TCA_FLOWER_KEY_TCP_SRC]        = { .type = NLA_U16 },
 631         [TCA_FLOWER_KEY_TCP_DST]        = { .type = NLA_U16 },
 632         [TCA_FLOWER_KEY_UDP_SRC]        = { .type = NLA_U16 },
 633         [TCA_FLOWER_KEY_UDP_DST]        = { .type = NLA_U16 },
 634         [TCA_FLOWER_KEY_VLAN_ID]        = { .type = NLA_U16 },
 635         [TCA_FLOWER_KEY_VLAN_PRIO]      = { .type = NLA_U8 },
 636         [TCA_FLOWER_KEY_VLAN_ETH_TYPE]  = { .type = NLA_U16 },
 637         [TCA_FLOWER_KEY_ENC_KEY_ID]     = { .type = NLA_U32 },
 638         [TCA_FLOWER_KEY_ENC_IPV4_SRC]   = { .type = NLA_U32 },
 639         [TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK] = { .type = NLA_U32 },
 640         [TCA_FLOWER_KEY_ENC_IPV4_DST]   = { .type = NLA_U32 },
 641         [TCA_FLOWER_KEY_ENC_IPV4_DST_MASK] = { .type = NLA_U32 },
 642         [TCA_FLOWER_KEY_ENC_IPV6_SRC]   = { .len = sizeof(struct in6_addr) },
 643         [TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK] = { .len = sizeof(struct in6_addr) },
 644         [TCA_FLOWER_KEY_ENC_IPV6_DST]   = { .len = sizeof(struct in6_addr) },
 645         [TCA_FLOWER_KEY_ENC_IPV6_DST_MASK] = { .len = sizeof(struct in6_addr) },
 646         [TCA_FLOWER_KEY_TCP_SRC_MASK]   = { .type = NLA_U16 },
 647         [TCA_FLOWER_KEY_TCP_DST_MASK]   = { .type = NLA_U16 },
 648         [TCA_FLOWER_KEY_UDP_SRC_MASK]   = { .type = NLA_U16 },
 649         [TCA_FLOWER_KEY_UDP_DST_MASK]   = { .type = NLA_U16 },
 650         [TCA_FLOWER_KEY_SCTP_SRC_MASK]  = { .type = NLA_U16 },
 651         [TCA_FLOWER_KEY_SCTP_DST_MASK]  = { .type = NLA_U16 },
 652         [TCA_FLOWER_KEY_SCTP_SRC]       = { .type = NLA_U16 },
 653         [TCA_FLOWER_KEY_SCTP_DST]       = { .type = NLA_U16 },
 654         [TCA_FLOWER_KEY_ENC_UDP_SRC_PORT]       = { .type = NLA_U16 },
 655         [TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK]  = { .type = NLA_U16 },
 656         [TCA_FLOWER_KEY_ENC_UDP_DST_PORT]       = { .type = NLA_U16 },
 657         [TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK]  = { .type = NLA_U16 },
 658         [TCA_FLOWER_KEY_FLAGS]          = { .type = NLA_U32 },
 659         [TCA_FLOWER_KEY_FLAGS_MASK]     = { .type = NLA_U32 },
 660         [TCA_FLOWER_KEY_ICMPV4_TYPE]    = { .type = NLA_U8 },
 661         [TCA_FLOWER_KEY_ICMPV4_TYPE_MASK] = { .type = NLA_U8 },
 662         [TCA_FLOWER_KEY_ICMPV4_CODE]    = { .type = NLA_U8 },
 663         [TCA_FLOWER_KEY_ICMPV4_CODE_MASK] = { .type = NLA_U8 },
 664         [TCA_FLOWER_KEY_ICMPV6_TYPE]    = { .type = NLA_U8 },
 665         [TCA_FLOWER_KEY_ICMPV6_TYPE_MASK] = { .type = NLA_U8 },
 666         [TCA_FLOWER_KEY_ICMPV6_CODE]    = { .type = NLA_U8 },
 667         [TCA_FLOWER_KEY_ICMPV6_CODE_MASK] = { .type = NLA_U8 },
 668         [TCA_FLOWER_KEY_ARP_SIP]        = { .type = NLA_U32 },
 669         [TCA_FLOWER_KEY_ARP_SIP_MASK]   = { .type = NLA_U32 },
 670         [TCA_FLOWER_KEY_ARP_TIP]        = { .type = NLA_U32 },
 671         [TCA_FLOWER_KEY_ARP_TIP_MASK]   = { .type = NLA_U32 },
 672         [TCA_FLOWER_KEY_ARP_OP]         = { .type = NLA_U8 },
 673         [TCA_FLOWER_KEY_ARP_OP_MASK]    = { .type = NLA_U8 },
 674         [TCA_FLOWER_KEY_ARP_SHA]        = { .len = ETH_ALEN },
 675         [TCA_FLOWER_KEY_ARP_SHA_MASK]   = { .len = ETH_ALEN },
 676         [TCA_FLOWER_KEY_ARP_THA]        = { .len = ETH_ALEN },
 677         [TCA_FLOWER_KEY_ARP_THA_MASK]   = { .len = ETH_ALEN },
 678         [TCA_FLOWER_KEY_MPLS_TTL]       = { .type = NLA_U8 },
 679         [TCA_FLOWER_KEY_MPLS_BOS]       = { .type = NLA_U8 },
 680         [TCA_FLOWER_KEY_MPLS_TC]        = { .type = NLA_U8 },
 681         [TCA_FLOWER_KEY_MPLS_LABEL]     = { .type = NLA_U32 },
 682         [TCA_FLOWER_KEY_MPLS_OPTS]      = { .type = NLA_NESTED },
 683         [TCA_FLOWER_KEY_TCP_FLAGS]      = { .type = NLA_U16 },
 684         [TCA_FLOWER_KEY_TCP_FLAGS_MASK] = { .type = NLA_U16 },
 685         [TCA_FLOWER_KEY_IP_TOS]         = { .type = NLA_U8 },
 686         [TCA_FLOWER_KEY_IP_TOS_MASK]    = { .type = NLA_U8 },
 687         [TCA_FLOWER_KEY_IP_TTL]         = { .type = NLA_U8 },
 688         [TCA_FLOWER_KEY_IP_TTL_MASK]    = { .type = NLA_U8 },
 689         [TCA_FLOWER_KEY_CVLAN_ID]       = { .type = NLA_U16 },
 690         [TCA_FLOWER_KEY_CVLAN_PRIO]     = { .type = NLA_U8 },
 691         [TCA_FLOWER_KEY_CVLAN_ETH_TYPE] = { .type = NLA_U16 },
 692         [TCA_FLOWER_KEY_ENC_IP_TOS]     = { .type = NLA_U8 },
 693         [TCA_FLOWER_KEY_ENC_IP_TOS_MASK] = { .type = NLA_U8 },
 694         [TCA_FLOWER_KEY_ENC_IP_TTL]      = { .type = NLA_U8 },
 695         [TCA_FLOWER_KEY_ENC_IP_TTL_MASK] = { .type = NLA_U8 },
 696         [TCA_FLOWER_KEY_ENC_OPTS]       = { .type = NLA_NESTED },
 697         [TCA_FLOWER_KEY_ENC_OPTS_MASK]  = { .type = NLA_NESTED },
 698         [TCA_FLOWER_KEY_CT_STATE]       =
 699                 NLA_POLICY_MASK(NLA_U16, TCA_FLOWER_KEY_CT_FLAGS_MASK),
 700         [TCA_FLOWER_KEY_CT_STATE_MASK]  =
 701                 NLA_POLICY_MASK(NLA_U16, TCA_FLOWER_KEY_CT_FLAGS_MASK),
 702         [TCA_FLOWER_KEY_CT_ZONE]        = { .type = NLA_U16 },
 703         [TCA_FLOWER_KEY_CT_ZONE_MASK]   = { .type = NLA_U16 },
 704         [TCA_FLOWER_KEY_CT_MARK]        = { .type = NLA_U32 },
 705         [TCA_FLOWER_KEY_CT_MARK_MASK]   = { .type = NLA_U32 },
 706         [TCA_FLOWER_KEY_CT_LABELS]      = { .type = NLA_BINARY,
 707                                             .len = 128 / BITS_PER_BYTE },
 708         [TCA_FLOWER_KEY_CT_LABELS_MASK] = { .type = NLA_BINARY,
 709                                             .len = 128 / BITS_PER_BYTE },
 710         [TCA_FLOWER_FLAGS]              = { .type = NLA_U32 },
 711         [TCA_FLOWER_KEY_HASH]           = { .type = NLA_U32 },
 712         [TCA_FLOWER_KEY_HASH_MASK]      = { .type = NLA_U32 },
 713         [TCA_FLOWER_KEY_NUM_OF_VLANS]   = { .type = NLA_U8 },
 714         [TCA_FLOWER_KEY_PPPOE_SID]      = { .type = NLA_U16 },
 715         [TCA_FLOWER_KEY_PPP_PROTO]      = { .type = NLA_U16 },
 716         [TCA_FLOWER_KEY_L2TPV3_SID]     = { .type = NLA_U32 },
 717
 718 };
 719
 720 static const struct nla_policy
 721 enc_opts_policy[TCA_FLOWER_KEY_ENC_OPTS_MAX + 1] = {
 722         [TCA_FLOWER_KEY_ENC_OPTS_UNSPEC]        = {
 723                 .strict_start_type = TCA_FLOWER_KEY_ENC_OPTS_VXLAN },
 724         [TCA_FLOWER_KEY_ENC_OPTS_GENEVE]        = { .type = NLA_NESTED },
 725         [TCA_FLOWER_KEY_ENC_OPTS_VXLAN]         = { .type = NLA_NESTED },
 726         [TCA_FLOWER_KEY_ENC_OPTS_ERSPAN]        = { .type = NLA_NESTED },
 727         [TCA_FLOWER_KEY_ENC_OPTS_GTP]           = { .type = NLA_NESTED },
 728 };
 729
 730 static const struct nla_policy
 731 geneve_opt_policy[TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX + 1] = {
 732         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS]      = { .type = NLA_U16 },
 733         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE]       = { .type = NLA_U8 },
 734         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA]       = { .type = NLA_BINARY,
 735                                                        .len = 128 },
 736 };
 737
 738 static const struct nla_policy
 739 vxlan_opt_policy[TCA_FLOWER_KEY_ENC_OPT_VXLAN_MAX + 1] = {
 740         [TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]         = { .type = NLA_U32 },
 741 };
 742
 743 static const struct nla_policy
 744 erspan_opt_policy[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_MAX + 1] = {
 745         [TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER]        = { .type = NLA_U8 },
 746         [TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX]      = { .type = NLA_U32 },
 747         [TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR]        = { .type = NLA_U8 },
 748         [TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID]       = { .type = NLA_U8 },
 749 };
 750
 751 static const struct nla_policy
 752 gtp_opt_policy[TCA_FLOWER_KEY_ENC_OPT_GTP_MAX + 1] = {
 753         [TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE]      = { .type = NLA_U8 },
 754         [TCA_FLOWER_KEY_ENC_OPT_GTP_QFI]           = { .type = NLA_U8 },
 755 };
 756
 757 static const struct nla_policy
 758 mpls_stack_entry_policy[TCA_FLOWER_KEY_MPLS_OPT_LSE_MAX + 1] = {
 759         [TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH]    = { .type = NLA_U8 },
 760         [TCA_FLOWER_KEY_MPLS_OPT_LSE_TTL]      = { .type = NLA_U8 },
 761         [TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS]      = { .type = NLA_U8 },
 762         [TCA_FLOWER_KEY_MPLS_OPT_LSE_TC]       = { .type = NLA_U8 },
 763         [TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL]    = { .type = NLA_U32 },
 764 };
 765
 766 static void fl_set_key_val(struct nlattr **tb,
 767                            void *val, int val_type,
 768                            void *mask, int mask_type, int len)
 769 {
 770         if (!tb[val_type])
 771                 return;
 772         nla_memcpy(val, tb[val_type], len);
 773         if (mask_type == TCA_FLOWER_UNSPEC || !tb[mask_type])
 774                 memset(mask, 0xff, len);
 775         else
 776                 nla_memcpy(mask, tb[mask_type], len);
 777 }
 778
 779 static int fl_set_key_port_range(struct nlattr **tb, struct fl_flow_key *key,
 780                                  struct fl_flow_key *mask,
 781                                  struct netlink_ext_ack *extack)
 782 {
 783         fl_set_key_val(tb, &key->tp_range.tp_min.dst,
 784                        TCA_FLOWER_KEY_PORT_DST_MIN, &mask->tp_range.tp_min.dst,
 785                        TCA_FLOWER_UNSPEC, sizeof(key->tp_range.tp_min.dst));
 786         fl_set_key_val(tb, &key->tp_range.tp_max.dst,
 787                        TCA_FLOWER_KEY_PORT_DST_MAX, &mask->tp_range.tp_max.dst,
 788                        TCA_FLOWER_UNSPEC, sizeof(key->tp_range.tp_max.dst));
 789         fl_set_key_val(tb, &key->tp_range.tp_min.src,
 790                        TCA_FLOWER_KEY_PORT_SRC_MIN, &mask->tp_range.tp_min.src,
 791                        TCA_FLOWER_UNSPEC, sizeof(key->tp_range.tp_min.src));
 792         fl_set_key_val(tb, &key->tp_range.tp_max.src,
 793                        TCA_FLOWER_KEY_PORT_SRC_MAX, &mask->tp_range.tp_max.src,
 794                        TCA_FLOWER_UNSPEC, sizeof(key->tp_range.tp_max.src));
 795
 796         if (mask->tp_range.tp_min.dst && mask->tp_range.tp_max.dst &&
 797             ntohs(key->tp_range.tp_max.dst) <=
 798             ntohs(key->tp_range.tp_min.dst)) {
 799                 NL_SET_ERR_MSG_ATTR(extack,
 800                                     tb[TCA_FLOWER_KEY_PORT_DST_MIN],
 801                                     "Invalid destination port range (min must be strictly smaller than max)");
 802                 return -EINVAL;
 803         }
 804         if (mask->tp_range.tp_min.src && mask->tp_range.tp_max.src &&
 805             ntohs(key->tp_range.tp_max.src) <=
 806             ntohs(key->tp_range.tp_min.src)) {
 807                 NL_SET_ERR_MSG_ATTR(extack,
 808                                     tb[TCA_FLOWER_KEY_PORT_SRC_MIN],
 809                                     "Invalid source port range (min must be strictly smaller than max)");
 810                 return -EINVAL;
 811         }
 812
 813         return 0;
 814 }
 815
 816 static int fl_set_key_mpls_lse(const struct nlattr *nla_lse,
 817                                struct flow_dissector_key_mpls *key_val,
 818                                struct flow_dissector_key_mpls *key_mask,
 819                                struct netlink_ext_ack *extack)
 820 {
 821         struct nlattr *tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_MAX + 1];
 822         struct flow_dissector_mpls_lse *lse_mask;
 823         struct flow_dissector_mpls_lse *lse_val;
 824         u8 lse_index;
 825         u8 depth;
 826         int err;
 827
 828         err = nla_parse_nested(tb, TCA_FLOWER_KEY_MPLS_OPT_LSE_MAX, nla_lse,
 829                                mpls_stack_entry_policy, extack);
 830         if (err < 0)
 831                 return err;
 832
 833         if (!tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH]) {
 834                 NL_SET_ERR_MSG(extack, "Missing MPLS option \"depth\"");
 835                 return -EINVAL;
 836         }
 837
 838         depth = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH]);
 839
 840         /* LSE depth starts at 1, for consistency with terminology used by
 841          * RFC 3031 (section 3.9), where depth 0 refers to unlabeled packets.
 842          */
 843         if (depth < 1 || depth > FLOW_DIS_MPLS_MAX) {
 844                 NL_SET_ERR_MSG_ATTR(extack,
 845                                     tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH],
 846                                     "Invalid MPLS depth");
 847                 return -EINVAL;
 848         }
 849         lse_index = depth - 1;
 850
 851         dissector_set_mpls_lse(key_val, lse_index);
 852         dissector_set_mpls_lse(key_mask, lse_index);
 853
 854         lse_val = &key_val->ls[lse_index];
 855         lse_mask = &key_mask->ls[lse_index];
 856
 857         if (tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_TTL]) {
 858                 lse_val->mpls_ttl = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_TTL]);
 859                 lse_mask->mpls_ttl = MPLS_TTL_MASK;
 860         }
 861         if (tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS]) {
 862                 u8 bos = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS]);
 863
 864                 if (bos & ~MPLS_BOS_MASK) {
 865                         NL_SET_ERR_MSG_ATTR(extack,
 866                                             tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS],
 867                                             "Bottom Of Stack (BOS) must be 0 or 1");
 868                         return -EINVAL;
 869                 }
 870                 lse_val->mpls_bos = bos;
 871                 lse_mask->mpls_bos = MPLS_BOS_MASK;
 872         }
 873         if (tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_TC]) {
 874                 u8 tc = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_TC]);
 875
 876                 if (tc & ~MPLS_TC_MASK) {
 877                         NL_SET_ERR_MSG_ATTR(extack,
 878                                             tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_TC],
 879                                             "Traffic Class (TC) must be between 0 and 7");
 880                         return -EINVAL;
 881                 }
 882                 lse_val->mpls_tc = tc;
 883                 lse_mask->mpls_tc = MPLS_TC_MASK;
 884         }
 885         if (tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL]) {
 886                 u32 label = nla_get_u32(tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL]);
 887
 888                 if (label & ~MPLS_LABEL_MASK) {
 889                         NL_SET_ERR_MSG_ATTR(extack,
 890                                             tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL],
 891                                             "Label must be between 0 and 1048575");
 892                         return -EINVAL;
 893                 }
 894                 lse_val->mpls_label = label;
 895                 lse_mask->mpls_label = MPLS_LABEL_MASK;
 896         }
 897
 898         return 0;
 899 }
 900
 901 static int fl_set_key_mpls_opts(const struct nlattr *nla_mpls_opts,
 902                                 struct flow_dissector_key_mpls *key_val,
 903                                 struct flow_dissector_key_mpls *key_mask,
 904                                 struct netlink_ext_ack *extack)
 905 {
 906         struct nlattr *nla_lse;
 907         int rem;
 908         int err;
 909
 910         if (!(nla_mpls_opts->nla_type & NLA_F_NESTED)) {
 911                 NL_SET_ERR_MSG_ATTR(extack, nla_mpls_opts,
 912                                     "NLA_F_NESTED is missing");
 913                 return -EINVAL;
 914         }
 915
 916         nla_for_each_nested(nla_lse, nla_mpls_opts, rem) {
 917                 if (nla_type(nla_lse) != TCA_FLOWER_KEY_MPLS_OPTS_LSE) {
 918                         NL_SET_ERR_MSG_ATTR(extack, nla_lse,
 919                                             "Invalid MPLS option type");
 920                         return -EINVAL;
 921                 }
 922
 923                 err = fl_set_key_mpls_lse(nla_lse, key_val, key_mask, extack);
 924                 if (err < 0)
 925                         return err;
 926         }
 927         if (rem) {
 928                 NL_SET_ERR_MSG(extack,
 929                                "Bytes leftover after parsing MPLS options");
 930                 return -EINVAL;
 931         }
 932
 933         return 0;
 934 }
 935
 936 static int fl_set_key_mpls(struct nlattr **tb,
 937                            struct flow_dissector_key_mpls *key_val,
 938                            struct flow_dissector_key_mpls *key_mask,
 939                            struct netlink_ext_ack *extack)
 940 {
 941         struct flow_dissector_mpls_lse *lse_mask;
 942         struct flow_dissector_mpls_lse *lse_val;
 943
 944         if (tb[TCA_FLOWER_KEY_MPLS_OPTS]) {
 945                 if (tb[TCA_FLOWER_KEY_MPLS_TTL] ||
 946                     tb[TCA_FLOWER_KEY_MPLS_BOS] ||
 947                     tb[TCA_FLOWER_KEY_MPLS_TC] ||
 948                     tb[TCA_FLOWER_KEY_MPLS_LABEL]) {
 949                         NL_SET_ERR_MSG_ATTR(extack,
 950                                             tb[TCA_FLOWER_KEY_MPLS_OPTS],
 951                                             "MPLS label, Traffic Class, Bottom Of Stack and Time To Live must be encapsulated in the MPLS options attribute");
 952                         return -EBADMSG;
 953                 }
 954
 955                 return fl_set_key_mpls_opts(tb[TCA_FLOWER_KEY_MPLS_OPTS],
 956                                             key_val, key_mask, extack);
 957         }
 958
 959         lse_val = &key_val->ls[0];
 960         lse_mask = &key_mask->ls[0];
 961
 962         if (tb[TCA_FLOWER_KEY_MPLS_TTL]) {
 963                 lse_val->mpls_ttl = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_TTL]);
 964                 lse_mask->mpls_ttl = MPLS_TTL_MASK;
 965                 dissector_set_mpls_lse(key_val, 0);
 966                 dissector_set_mpls_lse(key_mask, 0);
 967         }
 968         if (tb[TCA_FLOWER_KEY_MPLS_BOS]) {
 969                 u8 bos = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_BOS]);
 970
 971                 if (bos & ~MPLS_BOS_MASK) {
 972                         NL_SET_ERR_MSG_ATTR(extack,
 973                                             tb[TCA_FLOWER_KEY_MPLS_BOS],
 974                                             "Bottom Of Stack (BOS) must be 0 or 1");
 975                         return -EINVAL;
 976                 }
 977                 lse_val->mpls_bos = bos;
 978                 lse_mask->mpls_bos = MPLS_BOS_MASK;
 979                 dissector_set_mpls_lse(key_val, 0);
 980                 dissector_set_mpls_lse(key_mask, 0);
 981         }
 982         if (tb[TCA_FLOWER_KEY_MPLS_TC]) {
 983                 u8 tc = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_TC]);
 984
 985                 if (tc & ~MPLS_TC_MASK) {
 986                         NL_SET_ERR_MSG_ATTR(extack,
 987                                             tb[TCA_FLOWER_KEY_MPLS_TC],
 988                                             "Traffic Class (TC) must be between 0 and 7");
 989                         return -EINVAL;
 990                 }
 991                 lse_val->mpls_tc = tc;
 992                 lse_mask->mpls_tc = MPLS_TC_MASK;
 993                 dissector_set_mpls_lse(key_val, 0);
 994                 dissector_set_mpls_lse(key_mask, 0);
 995         }
 996         if (tb[TCA_FLOWER_KEY_MPLS_LABEL]) {
 997                 u32 label = nla_get_u32(tb[TCA_FLOWER_KEY_MPLS_LABEL]);
 998
 999                 if (label & ~MPLS_LABEL_MASK) {
1000                         NL_SET_ERR_MSG_ATTR(extack,
1001                                             tb[TCA_FLOWER_KEY_MPLS_LABEL],
1002                                             "Label must be between 0 and 1048575");
1003                         return -EINVAL;
1004                 }
1005                 lse_val->mpls_label = label;
1006                 lse_mask->mpls_label = MPLS_LABEL_MASK;
1007                 dissector_set_mpls_lse(key_val, 0);
1008                 dissector_set_mpls_lse(key_mask, 0);
1009         }
1010         return 0;
1011 }
1012
1013 static void fl_set_key_vlan(struct nlattr **tb,
1014                             __be16 ethertype,
1015                             int vlan_id_key, int vlan_prio_key,
1016                             int vlan_next_eth_type_key,
1017                             struct flow_dissector_key_vlan *key_val,
1018                             struct flow_dissector_key_vlan *key_mask)
1019 {
1020 #define VLAN_PRIORITY_MASK      0x7
1021
1022         if (tb[vlan_id_key]) {
1023                 key_val->vlan_id =
1024                         nla_get_u16(tb[vlan_id_key]) & VLAN_VID_MASK;
1025                 key_mask->vlan_id = VLAN_VID_MASK;
1026         }
1027         if (tb[vlan_prio_key]) {
1028                 key_val->vlan_priority =
1029                         nla_get_u8(tb[vlan_prio_key]) &
1030                         VLAN_PRIORITY_MASK;
1031                 key_mask->vlan_priority = VLAN_PRIORITY_MASK;
1032         }
1033         if (ethertype) {
1034                 key_val->vlan_tpid = ethertype;
1035                 key_mask->vlan_tpid = cpu_to_be16(~0);
1036         }
1037         if (tb[vlan_next_eth_type_key]) {
1038                 key_val->vlan_eth_type =
1039                         nla_get_be16(tb[vlan_next_eth_type_key]);
1040                 key_mask->vlan_eth_type = cpu_to_be16(~0);
1041         }
1042 }
1043
1044 static void fl_set_key_pppoe(struct nlattr **tb,
1045                              struct flow_dissector_key_pppoe *key_val,
1046                              struct flow_dissector_key_pppoe *key_mask,
1047                              struct fl_flow_key *key,
1048                              struct fl_flow_key *mask)
1049 {
1050         /* key_val::type must be set to ETH_P_PPP_SES
1051          * because ETH_P_PPP_SES was stored in basic.n_proto
1052          * which might get overwritten by ppp_proto
1053          * or might be set to 0, the role of key_val::type
1054          * is simmilar to vlan_key::tpid
1055          */
1056         key_val->type = htons(ETH_P_PPP_SES);
1057         key_mask->type = cpu_to_be16(~0);
1058
1059         if (tb[TCA_FLOWER_KEY_PPPOE_SID]) {
1060                 key_val->session_id =
1061                         nla_get_be16(tb[TCA_FLOWER_KEY_PPPOE_SID]);
1062                 key_mask->session_id = cpu_to_be16(~0);
1063         }
1064         if (tb[TCA_FLOWER_KEY_PPP_PROTO]) {
1065                 key_val->ppp_proto =
1066                         nla_get_be16(tb[TCA_FLOWER_KEY_PPP_PROTO]);
1067                 key_mask->ppp_proto = cpu_to_be16(~0);
1068
1069                 if (key_val->ppp_proto == htons(PPP_IP)) {
1070                         key->basic.n_proto = htons(ETH_P_IP);
1071                         mask->basic.n_proto = cpu_to_be16(~0);
1072                 } else if (key_val->ppp_proto == htons(PPP_IPV6)) {
1073                         key->basic.n_proto = htons(ETH_P_IPV6);
1074                         mask->basic.n_proto = cpu_to_be16(~0);
1075                 } else if (key_val->ppp_proto == htons(PPP_MPLS_UC)) {
1076                         key->basic.n_proto = htons(ETH_P_MPLS_UC);
1077                         mask->basic.n_proto = cpu_to_be16(~0);
1078                 } else if (key_val->ppp_proto == htons(PPP_MPLS_MC)) {
1079                         key->basic.n_proto = htons(ETH_P_MPLS_MC);
1080                         mask->basic.n_proto = cpu_to_be16(~0);
1081                 }
1082         } else {
1083                 key->basic.n_proto = 0;
1084                 mask->basic.n_proto = cpu_to_be16(0);
1085         }
1086 }
1087
1088 static void fl_set_key_flag(u32 flower_key, u32 flower_mask,
1089                             u32 *dissector_key, u32 *dissector_mask,
1090                             u32 flower_flag_bit, u32 dissector_flag_bit)
1091 {
1092         if (flower_mask & flower_flag_bit) {
1093                 *dissector_mask |= dissector_flag_bit;
1094                 if (flower_key & flower_flag_bit)
1095                         *dissector_key |= dissector_flag_bit;
1096         }
1097 }
1098
1099 static int fl_set_key_flags(struct nlattr **tb, u32 *flags_key,
1100                             u32 *flags_mask, struct netlink_ext_ack *extack)
1101 {
1102         u32 key, mask;
1103
1104         /* mask is mandatory for flags */
1105         if (!tb[TCA_FLOWER_KEY_FLAGS_MASK]) {
1106                 NL_SET_ERR_MSG(extack, "Missing flags mask");
1107                 return -EINVAL;
1108         }
1109
1110         key = be32_to_cpu(nla_get_be32(tb[TCA_FLOWER_KEY_FLAGS]));
1111         mask = be32_to_cpu(nla_get_be32(tb[TCA_FLOWER_KEY_FLAGS_MASK]));
1112
1113         *flags_key  = 0;
1114         *flags_mask = 0;
1115
1116         fl_set_key_flag(key, mask, flags_key, flags_mask,
1117                         TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT, FLOW_DIS_IS_FRAGMENT);
1118         fl_set_key_flag(key, mask, flags_key, flags_mask,
1119                         TCA_FLOWER_KEY_FLAGS_FRAG_IS_FIRST,
1120                         FLOW_DIS_FIRST_FRAG);
1121
1122         return 0;
1123 }
1124
1125 static void fl_set_key_ip(struct nlattr **tb, bool encap,
1126                           struct flow_dissector_key_ip *key,
1127                           struct flow_dissector_key_ip *mask)
1128 {
1129         int tos_key = encap ? TCA_FLOWER_KEY_ENC_IP_TOS : TCA_FLOWER_KEY_IP_TOS;
1130         int ttl_key = encap ? TCA_FLOWER_KEY_ENC_IP_TTL : TCA_FLOWER_KEY_IP_TTL;
1131         int tos_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TOS_MASK : TCA_FLOWER_KEY_IP_TOS_MASK;
1132         int ttl_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TTL_MASK : TCA_FLOWER_KEY_IP_TTL_MASK;
1133
1134         fl_set_key_val(tb, &key->tos, tos_key, &mask->tos, tos_mask, sizeof(key->tos));
1135         fl_set_key_val(tb, &key->ttl, ttl_key, &mask->ttl, ttl_mask, sizeof(key->ttl));
1136 }
1137
1138 static int fl_set_geneve_opt(const struct nlattr *nla, struct fl_flow_key *key,
1139                              int depth, int option_len,
1140                              struct netlink_ext_ack *extack)
1141 {
1142         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX + 1];
1143         struct nlattr *class = NULL, *type = NULL, *data = NULL;
1144         struct geneve_opt *opt;
1145         int err, data_len = 0;
1146
1147         if (option_len > sizeof(struct geneve_opt))
1148                 data_len = option_len - sizeof(struct geneve_opt);
1149
1150         opt = (struct geneve_opt *)&key->enc_opts.data[key->enc_opts.len];
1151         memset(opt, 0xff, option_len);
1152         opt->length = data_len / 4;
1153         opt->r1 = 0;
1154         opt->r2 = 0;
1155         opt->r3 = 0;
1156
1157         /* If no mask has been prodived we assume an exact match. */
1158         if (!depth)
1159                 return sizeof(struct geneve_opt) + data_len;
1160
1161         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_GENEVE) {
1162                 NL_SET_ERR_MSG(extack, "Non-geneve option type for mask");
1163                 return -EINVAL;
1164         }
1165
1166         err = nla_parse_nested_deprecated(tb,
1167                                           TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX,
1168                                           nla, geneve_opt_policy, extack);
1169         if (err < 0)
1170                 return err;
1171
1172         /* We are not allowed to omit any of CLASS, TYPE or DATA
1173          * fields from the key.
1174          */
1175         if (!option_len &&
1176             (!tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS] ||
1177              !tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE] ||
1178              !tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA])) {
1179                 NL_SET_ERR_MSG(extack, "Missing tunnel key geneve option class, type or data");
1180                 return -EINVAL;
1181         }
1182
1183         /* Omitting any of CLASS, TYPE or DATA fields is allowed
1184          * for the mask.
1185          */
1186         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA]) {
1187                 int new_len = key->enc_opts.len;
1188
1189                 data = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA];
1190                 data_len = nla_len(data);
1191                 if (data_len < 4) {
1192                         NL_SET_ERR_MSG(extack, "Tunnel key geneve option data is less than 4 bytes long");
1193                         return -ERANGE;
1194                 }
1195                 if (data_len % 4) {
1196                         NL_SET_ERR_MSG(extack, "Tunnel key geneve option data is not a multiple of 4 bytes long");
1197                         return -ERANGE;
1198                 }
1199
1200                 new_len += sizeof(struct geneve_opt) + data_len;
1201                 BUILD_BUG_ON(FLOW_DIS_TUN_OPTS_MAX != IP_TUNNEL_OPTS_MAX);
1202                 if (new_len > FLOW_DIS_TUN_OPTS_MAX) {
1203                         NL_SET_ERR_MSG(extack, "Tunnel options exceeds max size");
1204                         return -ERANGE;
1205                 }
1206                 opt->length = data_len / 4;
1207                 memcpy(opt->opt_data, nla_data(data), data_len);
1208         }
1209
1210         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS]) {
1211                 class = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS];
1212                 opt->opt_class = nla_get_be16(class);
1213         }
1214
1215         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE]) {
1216                 type = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE];
1217                 opt->type = nla_get_u8(type);
1218         }
1219
1220         return sizeof(struct geneve_opt) + data_len;
1221 }
1222
1223 static int fl_set_vxlan_opt(const struct nlattr *nla, struct fl_flow_key *key,
1224                             int depth, int option_len,
1225                             struct netlink_ext_ack *extack)
1226 {
1227         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_MAX + 1];
1228         struct vxlan_metadata *md;
1229         int err;
1230
1231         md = (struct vxlan_metadata *)&key->enc_opts.data[key->enc_opts.len];
1232         memset(md, 0xff, sizeof(*md));
1233
1234         if (!depth)
1235                 return sizeof(*md);
1236
1237         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_VXLAN) {
1238                 NL_SET_ERR_MSG(extack, "Non-vxlan option type for mask");
1239                 return -EINVAL;
1240         }
1241
1242         err = nla_parse_nested(tb, TCA_FLOWER_KEY_ENC_OPT_VXLAN_MAX, nla,
1243                                vxlan_opt_policy, extack);
1244         if (err < 0)
1245                 return err;
1246
1247         if (!option_len && !tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]) {
1248                 NL_SET_ERR_MSG(extack, "Missing tunnel key vxlan option gbp");
1249                 return -EINVAL;
1250         }
1251
1252         if (tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]) {
1253                 md->gbp = nla_get_u32(tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]);
1254                 md->gbp &= VXLAN_GBP_MASK;
1255         }
1256
1257         return sizeof(*md);
1258 }
1259
1260 static int fl_set_erspan_opt(const struct nlattr *nla, struct fl_flow_key *key,
1261                              int depth, int option_len,
1262                              struct netlink_ext_ack *extack)
1263 {
1264         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_MAX + 1];
1265         struct erspan_metadata *md;
1266         int err;
1267
1268         md = (struct erspan_metadata *)&key->enc_opts.data[key->enc_opts.len];
1269         memset(md, 0xff, sizeof(*md));
1270         md->version = 1;
1271
1272         if (!depth)
1273                 return sizeof(*md);
1274
1275         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_ERSPAN) {
1276                 NL_SET_ERR_MSG(extack, "Non-erspan option type for mask");
1277                 return -EINVAL;
1278         }
1279
1280         err = nla_parse_nested(tb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_MAX, nla,
1281                                erspan_opt_policy, extack);
1282         if (err < 0)
1283                 return err;
1284
1285         if (!option_len && !tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER]) {
1286                 NL_SET_ERR_MSG(extack, "Missing tunnel key erspan option ver");
1287                 return -EINVAL;
1288         }
1289
1290         if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER])
1291                 md->version = nla_get_u8(tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER]);
1292
1293         if (md->version == 1) {
1294                 if (!option_len && !tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX]) {
1295                         NL_SET_ERR_MSG(extack, "Missing tunnel key erspan option index");
1296                         return -EINVAL;
1297                 }
1298                 if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX]) {
1299                         nla = tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX];
1300                         memset(&md->u, 0x00, sizeof(md->u));
1301                         md->u.index = nla_get_be32(nla);
1302                 }
1303         } else if (md->version == 2) {
1304                 if (!option_len && (!tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR] ||
1305                                     !tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID])) {
1306                         NL_SET_ERR_MSG(extack, "Missing tunnel key erspan option dir or hwid");
1307                         return -EINVAL;
1308                 }
1309                 if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR]) {
1310                         nla = tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR];
1311                         md->u.md2.dir = nla_get_u8(nla);
1312                 }
1313                 if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID]) {
1314                         nla = tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID];
1315                         set_hwid(&md->u.md2, nla_get_u8(nla));
1316                 }
1317         } else {
1318                 NL_SET_ERR_MSG(extack, "Tunnel key erspan option ver is incorrect");
1319                 return -EINVAL;
1320         }
1321
1322         return sizeof(*md);
1323 }
1324
1325 static int fl_set_gtp_opt(const struct nlattr *nla, struct fl_flow_key *key,
1326                           int depth, int option_len,
1327                           struct netlink_ext_ack *extack)
1328 {
1329         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_GTP_MAX + 1];
1330         struct gtp_pdu_session_info *sinfo;
1331         u8 len = key->enc_opts.len;
1332         int err;
1333
1334         sinfo = (struct gtp_pdu_session_info *)&key->enc_opts.data[len];
1335         memset(sinfo, 0xff, option_len);
1336
1337         if (!depth)
1338                 return sizeof(*sinfo);
1339
1340         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_GTP) {
1341                 NL_SET_ERR_MSG_MOD(extack, "Non-gtp option type for mask");
1342                 return -EINVAL;
1343         }
1344
1345         err = nla_parse_nested(tb, TCA_FLOWER_KEY_ENC_OPT_GTP_MAX, nla,
1346                                gtp_opt_policy, extack);
1347         if (err < 0)
1348                 return err;
1349
1350         if (!option_len &&
1351             (!tb[TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE] ||
1352              !tb[TCA_FLOWER_KEY_ENC_OPT_GTP_QFI])) {
1353                 NL_SET_ERR_MSG_MOD(extack,
1354                                    "Missing tunnel key gtp option pdu type or qfi");
1355                 return -EINVAL;
1356         }
1357
1358         if (tb[TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE])
1359                 sinfo->pdu_type =
1360                         nla_get_u8(tb[TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE]);
1361
1362         if (tb[TCA_FLOWER_KEY_ENC_OPT_GTP_QFI])
1363                 sinfo->qfi = nla_get_u8(tb[TCA_FLOWER_KEY_ENC_OPT_GTP_QFI]);
1364
1365         return sizeof(*sinfo);
1366 }
1367
1368 static int fl_set_enc_opt(struct nlattr **tb, struct fl_flow_key *key,
1369                           struct fl_flow_key *mask,
1370                           struct netlink_ext_ack *extack)
1371 {
1372         const struct nlattr *nla_enc_key, *nla_opt_key, *nla_opt_msk = NULL;
1373         int err, option_len, key_depth, msk_depth = 0;
1374
1375         err = nla_validate_nested_deprecated(tb[TCA_FLOWER_KEY_ENC_OPTS],
1376                                              TCA_FLOWER_KEY_ENC_OPTS_MAX,
1377                                              enc_opts_policy, extack);
1378         if (err)
1379                 return err;
1380
1381         nla_enc_key = nla_data(tb[TCA_FLOWER_KEY_ENC_OPTS]);
1382
1383         if (tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]) {
1384                 err = nla_validate_nested_deprecated(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK],
1385                                                      TCA_FLOWER_KEY_ENC_OPTS_MAX,
1386                                                      enc_opts_policy, extack);
1387                 if (err)
1388                         return err;
1389
1390                 nla_opt_msk = nla_data(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]);
1391                 msk_depth = nla_len(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]);
1392                 if (!nla_ok(nla_opt_msk, msk_depth)) {
1393                         NL_SET_ERR_MSG(extack, "Invalid nested attribute for masks");
1394                         return -EINVAL;
1395                 }
1396         }
1397
1398         nla_for_each_attr(nla_opt_key, nla_enc_key,
1399                           nla_len(tb[TCA_FLOWER_KEY_ENC_OPTS]), key_depth) {
1400                 switch (nla_type(nla_opt_key)) {
1401                 case TCA_FLOWER_KEY_ENC_OPTS_GENEVE:
1402                         if (key->enc_opts.dst_opt_type &&
1403                             key->enc_opts.dst_opt_type != TUNNEL_GENEVE_OPT) {
1404                                 NL_SET_ERR_MSG(extack, "Duplicate type for geneve options");
1405                                 return -EINVAL;
1406                         }
1407                         option_len = 0;
1408                         key->enc_opts.dst_opt_type = TUNNEL_GENEVE_OPT;
1409                         option_len = fl_set_geneve_opt(nla_opt_key, key,
1410                                                        key_depth, option_len,
1411                                                        extack);
1412                         if (option_len < 0)
1413                                 return option_len;
1414
1415                         key->enc_opts.len += option_len;
1416                         /* At the same time we need to parse through the mask
1417                          * in order to verify exact and mask attribute lengths.
1418                          */
1419                         mask->enc_opts.dst_opt_type = TUNNEL_GENEVE_OPT;
1420                         option_len = fl_set_geneve_opt(nla_opt_msk, mask,
1421                                                        msk_depth, option_len,
1422                                                        extack);
1423                         if (option_len < 0)
1424                                 return option_len;
1425
1426                         mask->enc_opts.len += option_len;
1427                         if (key->enc_opts.len != mask->enc_opts.len) {
1428                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
1429                                 return -EINVAL;
1430                         }
1431                         break;
1432                 case TCA_FLOWER_KEY_ENC_OPTS_VXLAN:
1433                         if (key->enc_opts.dst_opt_type) {
1434                                 NL_SET_ERR_MSG(extack, "Duplicate type for vxlan options");
1435                                 return -EINVAL;
1436                         }
1437                         option_len = 0;
1438                         key->enc_opts.dst_opt_type = TUNNEL_VXLAN_OPT;
1439                         option_len = fl_set_vxlan_opt(nla_opt_key, key,
1440                                                       key_depth, option_len,
1441                                                       extack);
1442                         if (option_len < 0)
1443                                 return option_len;
1444
1445                         key->enc_opts.len += option_len;
1446                         /* At the same time we need to parse through the mask
1447                          * in order to verify exact and mask attribute lengths.
1448                          */
1449                         mask->enc_opts.dst_opt_type = TUNNEL_VXLAN_OPT;
1450                         option_len = fl_set_vxlan_opt(nla_opt_msk, mask,
1451                                                       msk_depth, option_len,
1452                                                       extack);
1453                         if (option_len < 0)
1454                                 return option_len;
1455
1456                         mask->enc_opts.len += option_len;
1457                         if (key->enc_opts.len != mask->enc_opts.len) {
1458                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
1459                                 return -EINVAL;
1460                         }
1461                         break;
1462                 case TCA_FLOWER_KEY_ENC_OPTS_ERSPAN:
1463                         if (key->enc_opts.dst_opt_type) {
1464                                 NL_SET_ERR_MSG(extack, "Duplicate type for erspan options");
1465                                 return -EINVAL;
1466                         }
1467                         option_len = 0;
1468                         key->enc_opts.dst_opt_type = TUNNEL_ERSPAN_OPT;
1469                         option_len = fl_set_erspan_opt(nla_opt_key, key,
1470                                                        key_depth, option_len,
1471                                                        extack);
1472                         if (option_len < 0)
1473                                 return option_len;
1474
1475                         key->enc_opts.len += option_len;
1476                         /* At the same time we need to parse through the mask
1477                          * in order to verify exact and mask attribute lengths.
1478                          */
1479                         mask->enc_opts.dst_opt_type = TUNNEL_ERSPAN_OPT;
1480                         option_len = fl_set_erspan_opt(nla_opt_msk, mask,
1481                                                        msk_depth, option_len,
1482                                                        extack);
1483                         if (option_len < 0)
1484                                 return option_len;
1485
1486                         mask->enc_opts.len += option_len;
1487                         if (key->enc_opts.len != mask->enc_opts.len) {
1488                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
1489                                 return -EINVAL;
1490                         }
1491                         break;
1492                 case TCA_FLOWER_KEY_ENC_OPTS_GTP:
1493                         if (key->enc_opts.dst_opt_type) {
1494                                 NL_SET_ERR_MSG_MOD(extack,
1495                                                    "Duplicate type for gtp options");
1496                                 return -EINVAL;
1497                         }
1498                         option_len = 0;
1499                         key->enc_opts.dst_opt_type = TUNNEL_GTP_OPT;
1500                         option_len = fl_set_gtp_opt(nla_opt_key, key,
1501                                                     key_depth, option_len,
1502                                                     extack);
1503                         if (option_len < 0)
1504                                 return option_len;
1505
1506                         key->enc_opts.len += option_len;
1507                         /* At the same time we need to parse through the mask
1508                          * in order to verify exact and mask attribute lengths.
1509                          */
1510                         mask->enc_opts.dst_opt_type = TUNNEL_GTP_OPT;
1511                         option_len = fl_set_gtp_opt(nla_opt_msk, mask,
1512                                                     msk_depth, option_len,
1513                                                     extack);
1514                         if (option_len < 0)
1515                                 return option_len;
1516
1517                         mask->enc_opts.len += option_len;
1518                         if (key->enc_opts.len != mask->enc_opts.len) {
1519                                 NL_SET_ERR_MSG_MOD(extack,
1520                                                    "Key and mask miss aligned");
1521                                 return -EINVAL;
1522                         }
1523                         break;
1524                 default:
1525                         NL_SET_ERR_MSG(extack, "Unknown tunnel option type");
1526                         return -EINVAL;
1527                 }
1528
1529                 if (!msk_depth)
1530                         continue;
1531
1532                 if (!nla_ok(nla_opt_msk, msk_depth)) {
1533                         NL_SET_ERR_MSG(extack, "A mask attribute is invalid");
1534                         return -EINVAL;
1535                 }
1536                 nla_opt_msk = nla_next(nla_opt_msk, &msk_depth);
1537         }
1538
1539         return 0;
1540 }
1541
1542 static int fl_validate_ct_state(u16 state, struct nlattr *tb,
1543                                 struct netlink_ext_ack *extack)
1544 {
1545         if (state && !(state & TCA_FLOWER_KEY_CT_FLAGS_TRACKED)) {
1546                 NL_SET_ERR_MSG_ATTR(extack, tb,
1547                                     "no trk, so no other flag can be set");
1548                 return -EINVAL;
1549         }
1550
1551         if (state & TCA_FLOWER_KEY_CT_FLAGS_NEW &&
1552             state & TCA_FLOWER_KEY_CT_FLAGS_ESTABLISHED) {
1553                 NL_SET_ERR_MSG_ATTR(extack, tb,
1554                                     "new and est are mutually exclusive");
1555                 return -EINVAL;
1556         }
1557
1558         if (state & TCA_FLOWER_KEY_CT_FLAGS_INVALID &&
1559             state & ~(TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
1560                       TCA_FLOWER_KEY_CT_FLAGS_INVALID)) {
1561                 NL_SET_ERR_MSG_ATTR(extack, tb,
1562                                     "when inv is set, only trk may be set");
1563                 return -EINVAL;
1564         }
1565
1566         if (state & TCA_FLOWER_KEY_CT_FLAGS_NEW &&
1567             state & TCA_FLOWER_KEY_CT_FLAGS_REPLY) {
1568                 NL_SET_ERR_MSG_ATTR(extack, tb,
1569                                     "new and rpl are mutually exclusive");
1570                 return -EINVAL;
1571         }
1572
1573         return 0;
1574 }
1575
1576 static int fl_set_key_ct(struct nlattr **tb,
1577                          struct flow_dissector_key_ct *key,
1578                          struct flow_dissector_key_ct *mask,
1579                          struct netlink_ext_ack *extack)
1580 {
1581         if (tb[TCA_FLOWER_KEY_CT_STATE]) {
1582                 int err;
1583
1584                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK)) {
1585                         NL_SET_ERR_MSG(extack, "Conntrack isn't enabled");
1586                         return -EOPNOTSUPP;
1587                 }
1588                 fl_set_key_val(tb, &key->ct_state, TCA_FLOWER_KEY_CT_STATE,
1589                                &mask->ct_state, TCA_FLOWER_KEY_CT_STATE_MASK,
1590                                sizeof(key->ct_state));
1591
1592                 err = fl_validate_ct_state(key->ct_state & mask->ct_state,
1593                                            tb[TCA_FLOWER_KEY_CT_STATE_MASK],
1594                                            extack);
1595                 if (err)
1596                         return err;
1597
1598         }
1599         if (tb[TCA_FLOWER_KEY_CT_ZONE]) {
1600                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK_ZONES)) {
1601                         NL_SET_ERR_MSG(extack, "Conntrack zones isn't enabled");
1602                         return -EOPNOTSUPP;
1603                 }
1604                 fl_set_key_val(tb, &key->ct_zone, TCA_FLOWER_KEY_CT_ZONE,
1605                                &mask->ct_zone, TCA_FLOWER_KEY_CT_ZONE_MASK,
1606                                sizeof(key->ct_zone));
1607         }
1608         if (tb[TCA_FLOWER_KEY_CT_MARK]) {
1609                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK_MARK)) {
1610                         NL_SET_ERR_MSG(extack, "Conntrack mark isn't enabled");
1611                         return -EOPNOTSUPP;
1612                 }
1613                 fl_set_key_val(tb, &key->ct_mark, TCA_FLOWER_KEY_CT_MARK,
1614                                &mask->ct_mark, TCA_FLOWER_KEY_CT_MARK_MASK,
1615                                sizeof(key->ct_mark));
1616         }
1617         if (tb[TCA_FLOWER_KEY_CT_LABELS]) {
1618                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK_LABELS)) {
1619                         NL_SET_ERR_MSG(extack, "Conntrack labels aren't enabled");
1620                         return -EOPNOTSUPP;
1621                 }
1622                 fl_set_key_val(tb, key->ct_labels, TCA_FLOWER_KEY_CT_LABELS,
1623                                mask->ct_labels, TCA_FLOWER_KEY_CT_LABELS_MASK,
1624                                sizeof(key->ct_labels));
1625         }
1626
1627         return 0;
1628 }
1629
1630 static bool is_vlan_key(struct nlattr *tb, __be16 *ethertype,
1631                         struct fl_flow_key *key, struct fl_flow_key *mask,
1632                         int vthresh)
1633 {
1634         const bool good_num_of_vlans = key->num_of_vlans.num_of_vlans > vthresh;
1635
1636         if (!tb) {
1637                 *ethertype = 0;
1638                 return good_num_of_vlans;
1639         }
1640
1641         *ethertype = nla_get_be16(tb);
1642         if (good_num_of_vlans || eth_type_vlan(*ethertype))
1643                 return true;
1644
1645         key->basic.n_proto = *ethertype;
1646         mask->basic.n_proto = cpu_to_be16(~0);
1647         return false;
1648 }
1649
1650 static int fl_set_key(struct net *net, struct nlattr **tb,
1651                       struct fl_flow_key *key, struct fl_flow_key *mask,
1652                       struct netlink_ext_ack *extack)
1653 {
1654         __be16 ethertype;
1655         int ret = 0;
1656
1657         if (tb[TCA_FLOWER_INDEV]) {
1658                 int err = tcf_change_indev(net, tb[TCA_FLOWER_INDEV], extack);
1659                 if (err < 0)
1660                         return err;
1661                 key->meta.ingress_ifindex = err;
1662                 mask->meta.ingress_ifindex = 0xffffffff;
1663         }
1664
1665         fl_set_key_val(tb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
1666                        mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
1667                        sizeof(key->eth.dst));
1668         fl_set_key_val(tb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
1669                        mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
1670                        sizeof(key->eth.src));
1671         fl_set_key_val(tb, &key->num_of_vlans,
1672                        TCA_FLOWER_KEY_NUM_OF_VLANS,
1673                        &mask->num_of_vlans,
1674                        TCA_FLOWER_UNSPEC,
1675                        sizeof(key->num_of_vlans));
1676
1677         if (is_vlan_key(tb[TCA_FLOWER_KEY_ETH_TYPE], &ethertype, key, mask, 0)) {
1678                 fl_set_key_vlan(tb, ethertype, TCA_FLOWER_KEY_VLAN_ID,
1679                                 TCA_FLOWER_KEY_VLAN_PRIO,
1680                                 TCA_FLOWER_KEY_VLAN_ETH_TYPE,
1681                                 &key->vlan, &mask->vlan);
1682
1683                 if (is_vlan_key(tb[TCA_FLOWER_KEY_VLAN_ETH_TYPE],
1684                                 &ethertype, key, mask, 1)) {
1685                         fl_set_key_vlan(tb, ethertype,
1686                                         TCA_FLOWER_KEY_CVLAN_ID,
1687                                         TCA_FLOWER_KEY_CVLAN_PRIO,
1688                                         TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
1689                                         &key->cvlan, &mask->cvlan);
1690                         fl_set_key_val(tb, &key->basic.n_proto,
1691                                        TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
1692                                        &mask->basic.n_proto,
1693                                        TCA_FLOWER_UNSPEC,
1694                                        sizeof(key->basic.n_proto));
1695                 }
1696         }
1697
1698         if (key->basic.n_proto == htons(ETH_P_PPP_SES))
1699                 fl_set_key_pppoe(tb, &key->pppoe, &mask->pppoe, key, mask);
1700
1701         if (key->basic.n_proto == htons(ETH_P_IP) ||
1702             key->basic.n_proto == htons(ETH_P_IPV6)) {
1703                 fl_set_key_val(tb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
1704                                &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
1705                                sizeof(key->basic.ip_proto));
1706                 fl_set_key_ip(tb, false, &key->ip, &mask->ip);
1707         }
1708
1709         if (tb[TCA_FLOWER_KEY_IPV4_SRC] || tb[TCA_FLOWER_KEY_IPV4_DST]) {
1710                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
1711                 mask->control.addr_type = ~0;
1712                 fl_set_key_val(tb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
1713                                &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
1714                                sizeof(key->ipv4.src));
1715                 fl_set_key_val(tb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
1716                                &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
1717                                sizeof(key->ipv4.dst));
1718         } else if (tb[TCA_FLOWER_KEY_IPV6_SRC] || tb[TCA_FLOWER_KEY_IPV6_DST]) {
1719                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
1720                 mask->control.addr_type = ~0;
1721                 fl_set_key_val(tb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
1722                                &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
1723                                sizeof(key->ipv6.src));
1724                 fl_set_key_val(tb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
1725                                &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
1726                                sizeof(key->ipv6.dst));
1727         }
1728
1729         if (key->basic.ip_proto == IPPROTO_TCP) {
1730                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
1731                                &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
1732                                sizeof(key->tp.src));
1733                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
1734                                &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
1735                                sizeof(key->tp.dst));
1736                 fl_set_key_val(tb, &key->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS,
1737                                &mask->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS_MASK,
1738                                sizeof(key->tcp.flags));
1739         } else if (key->basic.ip_proto == IPPROTO_UDP) {
1740                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
1741                                &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
1742                                sizeof(key->tp.src));
1743                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
1744                                &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
1745                                sizeof(key->tp.dst));
1746         } else if (key->basic.ip_proto == IPPROTO_SCTP) {
1747                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_SCTP_SRC,
1748                                &mask->tp.src, TCA_FLOWER_KEY_SCTP_SRC_MASK,
1749                                sizeof(key->tp.src));
1750                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_SCTP_DST,
1751                                &mask->tp.dst, TCA_FLOWER_KEY_SCTP_DST_MASK,
1752                                sizeof(key->tp.dst));
1753         } else if (key->basic.n_proto == htons(ETH_P_IP) &&
1754                    key->basic.ip_proto == IPPROTO_ICMP) {
1755                 fl_set_key_val(tb, &key->icmp.type, TCA_FLOWER_KEY_ICMPV4_TYPE,
1756                                &mask->icmp.type,
1757                                TCA_FLOWER_KEY_ICMPV4_TYPE_MASK,
1758                                sizeof(key->icmp.type));
1759                 fl_set_key_val(tb, &key->icmp.code, TCA_FLOWER_KEY_ICMPV4_CODE,
1760                                &mask->icmp.code,
1761                                TCA_FLOWER_KEY_ICMPV4_CODE_MASK,
1762                                sizeof(key->icmp.code));
1763         } else if (key->basic.n_proto == htons(ETH_P_IPV6) &&
1764                    key->basic.ip_proto == IPPROTO_ICMPV6) {
1765                 fl_set_key_val(tb, &key->icmp.type, TCA_FLOWER_KEY_ICMPV6_TYPE,
1766                                &mask->icmp.type,
1767                                TCA_FLOWER_KEY_ICMPV6_TYPE_MASK,
1768                                sizeof(key->icmp.type));
1769                 fl_set_key_val(tb, &key->icmp.code, TCA_FLOWER_KEY_ICMPV6_CODE,
1770                                &mask->icmp.code,
1771                                TCA_FLOWER_KEY_ICMPV6_CODE_MASK,
1772                                sizeof(key->icmp.code));
1773         } else if (key->basic.n_proto == htons(ETH_P_MPLS_UC) ||
1774                    key->basic.n_proto == htons(ETH_P_MPLS_MC)) {
1775                 ret = fl_set_key_mpls(tb, &key->mpls, &mask->mpls, extack);
1776                 if (ret)
1777                         return ret;
1778         } else if (key->basic.n_proto == htons(ETH_P_ARP) ||
1779                    key->basic.n_proto == htons(ETH_P_RARP)) {
1780                 fl_set_key_val(tb, &key->arp.sip, TCA_FLOWER_KEY_ARP_SIP,
1781                                &mask->arp.sip, TCA_FLOWER_KEY_ARP_SIP_MASK,
1782                                sizeof(key->arp.sip));
1783                 fl_set_key_val(tb, &key->arp.tip, TCA_FLOWER_KEY_ARP_TIP,
1784                                &mask->arp.tip, TCA_FLOWER_KEY_ARP_TIP_MASK,
1785                                sizeof(key->arp.tip));
1786                 fl_set_key_val(tb, &key->arp.op, TCA_FLOWER_KEY_ARP_OP,
1787                                &mask->arp.op, TCA_FLOWER_KEY_ARP_OP_MASK,
1788                                sizeof(key->arp.op));
1789                 fl_set_key_val(tb, key->arp.sha, TCA_FLOWER_KEY_ARP_SHA,
1790                                mask->arp.sha, TCA_FLOWER_KEY_ARP_SHA_MASK,
1791                                sizeof(key->arp.sha));
1792                 fl_set_key_val(tb, key->arp.tha, TCA_FLOWER_KEY_ARP_THA,
1793                                mask->arp.tha, TCA_FLOWER_KEY_ARP_THA_MASK,
1794                                sizeof(key->arp.tha));
1795         } else if (key->basic.ip_proto == IPPROTO_L2TP) {
1796                 fl_set_key_val(tb, &key->l2tpv3.session_id,
1797                                TCA_FLOWER_KEY_L2TPV3_SID,
1798                                &mask->l2tpv3.session_id, TCA_FLOWER_UNSPEC,
1799                                sizeof(key->l2tpv3.session_id));
1800         }
1801
1802         if (key->basic.ip_proto == IPPROTO_TCP ||
1803             key->basic.ip_proto == IPPROTO_UDP ||
1804             key->basic.ip_proto == IPPROTO_SCTP) {
1805                 ret = fl_set_key_port_range(tb, key, mask, extack);
1806                 if (ret)
1807                         return ret;
1808         }
1809
1810         if (tb[TCA_FLOWER_KEY_ENC_IPV4_SRC] ||
1811             tb[TCA_FLOWER_KEY_ENC_IPV4_DST]) {
1812                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
1813                 mask->enc_control.addr_type = ~0;
1814                 fl_set_key_val(tb, &key->enc_ipv4.src,
1815                                TCA_FLOWER_KEY_ENC_IPV4_SRC,
1816                                &mask->enc_ipv4.src,
1817                                TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
1818                                sizeof(key->enc_ipv4.src));
1819                 fl_set_key_val(tb, &key->enc_ipv4.dst,
1820                                TCA_FLOWER_KEY_ENC_IPV4_DST,
1821                                &mask->enc_ipv4.dst,
1822                                TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
1823                                sizeof(key->enc_ipv4.dst));
1824         }
1825
1826         if (tb[TCA_FLOWER_KEY_ENC_IPV6_SRC] ||
1827             tb[TCA_FLOWER_KEY_ENC_IPV6_DST]) {
1828                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
1829                 mask->enc_control.addr_type = ~0;
1830                 fl_set_key_val(tb, &key->enc_ipv6.src,
1831                                TCA_FLOWER_KEY_ENC_IPV6_SRC,
1832                                &mask->enc_ipv6.src,
1833                                TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
1834                                sizeof(key->enc_ipv6.src));
1835                 fl_set_key_val(tb, &key->enc_ipv6.dst,
1836                                TCA_FLOWER_KEY_ENC_IPV6_DST,
1837                                &mask->enc_ipv6.dst,
1838                                TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
1839                                sizeof(key->enc_ipv6.dst));
1840         }
1841
1842         fl_set_key_val(tb, &key->enc_key_id.keyid, TCA_FLOWER_KEY_ENC_KEY_ID,
1843                        &mask->enc_key_id.keyid, TCA_FLOWER_UNSPEC,
1844                        sizeof(key->enc_key_id.keyid));
1845
1846         fl_set_key_val(tb, &key->enc_tp.src, TCA_FLOWER_KEY_ENC_UDP_SRC_PORT,
1847                        &mask->enc_tp.src, TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK,
1848                        sizeof(key->enc_tp.src));
1849
1850         fl_set_key_val(tb, &key->enc_tp.dst, TCA_FLOWER_KEY_ENC_UDP_DST_PORT,
1851                        &mask->enc_tp.dst, TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK,
1852                        sizeof(key->enc_tp.dst));
1853
1854         fl_set_key_ip(tb, true, &key->enc_ip, &mask->enc_ip);
1855
1856         fl_set_key_val(tb, &key->hash.hash, TCA_FLOWER_KEY_HASH,
1857                        &mask->hash.hash, TCA_FLOWER_KEY_HASH_MASK,
1858                        sizeof(key->hash.hash));
1859
1860         if (tb[TCA_FLOWER_KEY_ENC_OPTS]) {
1861                 ret = fl_set_enc_opt(tb, key, mask, extack);
1862                 if (ret)
1863                         return ret;
1864         }
1865
1866         ret = fl_set_key_ct(tb, &key->ct, &mask->ct, extack);
1867         if (ret)
1868                 return ret;
1869
1870         if (tb[TCA_FLOWER_KEY_FLAGS])
1871                 ret = fl_set_key_flags(tb, &key->control.flags,
1872                                        &mask->control.flags, extack);
1873
1874         return ret;
1875 }
1876
1877 static void fl_mask_copy(struct fl_flow_mask *dst,
1878                          struct fl_flow_mask *src)
1879 {
1880         const void *psrc = fl_key_get_start(&src->key, src);
1881         void *pdst = fl_key_get_start(&dst->key, src);
1882
1883         memcpy(pdst, psrc, fl_mask_range(src));
1884         dst->range = src->range;
1885 }
1886
1887 static const struct rhashtable_params fl_ht_params = {
1888         .key_offset = offsetof(struct cls_fl_filter, mkey), /* base offset */
1889         .head_offset = offsetof(struct cls_fl_filter, ht_node),
1890         .automatic_shrinking = true,
1891 };
1892
1893 static int fl_init_mask_hashtable(struct fl_flow_mask *mask)
1894 {
1895         mask->filter_ht_params = fl_ht_params;
1896         mask->filter_ht_params.key_len = fl_mask_range(mask);
1897         mask->filter_ht_params.key_offset += mask->range.start;
1898
1899         return rhashtable_init(&mask->ht, &mask->filter_ht_params);
1900 }
1901
1902 #define FL_KEY_MEMBER_OFFSET(member) offsetof(struct fl_flow_key, member)
1903 #define FL_KEY_MEMBER_SIZE(member) sizeof_field(struct fl_flow_key, member)
1904
1905 #define FL_KEY_IS_MASKED(mask, member)                                          \
1906         memchr_inv(((char *)mask) + FL_KEY_MEMBER_OFFSET(member),               \
1907                    0, FL_KEY_MEMBER_SIZE(member))                               \
1908
1909 #define FL_KEY_SET(keys, cnt, id, member)                                       \
1910         do {                                                                    \
1911                 keys[cnt].key_id = id;                                          \
1912                 keys[cnt].offset = FL_KEY_MEMBER_OFFSET(member);                \
1913                 cnt++;                                                          \
1914         } while(0);
1915
1916 #define FL_KEY_SET_IF_MASKED(mask, keys, cnt, id, member)                       \
1917         do {                                                                    \
1918                 if (FL_KEY_IS_MASKED(mask, member))                             \
1919                         FL_KEY_SET(keys, cnt, id, member);                      \
1920         } while(0);
1921
1922 static void fl_init_dissector(struct flow_dissector *dissector,
1923                               struct fl_flow_key *mask)
1924 {
1925         struct flow_dissector_key keys[FLOW_DISSECTOR_KEY_MAX];
1926         size_t cnt = 0;
1927
1928         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1929                              FLOW_DISSECTOR_KEY_META, meta);
1930         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_CONTROL, control);
1931         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_BASIC, basic);
1932         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1933                              FLOW_DISSECTOR_KEY_ETH_ADDRS, eth);
1934         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1935                              FLOW_DISSECTOR_KEY_IPV4_ADDRS, ipv4);
1936         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1937                              FLOW_DISSECTOR_KEY_IPV6_ADDRS, ipv6);
1938         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1939                              FLOW_DISSECTOR_KEY_PORTS, tp);
1940         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1941                              FLOW_DISSECTOR_KEY_PORTS_RANGE, tp_range);
1942         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1943                              FLOW_DISSECTOR_KEY_IP, ip);
1944         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1945                              FLOW_DISSECTOR_KEY_TCP, tcp);
1946         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1947                              FLOW_DISSECTOR_KEY_ICMP, icmp);
1948         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1949                              FLOW_DISSECTOR_KEY_ARP, arp);
1950         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1951                              FLOW_DISSECTOR_KEY_MPLS, mpls);
1952         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1953                              FLOW_DISSECTOR_KEY_VLAN, vlan);
1954         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1955                              FLOW_DISSECTOR_KEY_CVLAN, cvlan);
1956         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1957                              FLOW_DISSECTOR_KEY_ENC_KEYID, enc_key_id);
1958         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1959                              FLOW_DISSECTOR_KEY_ENC_IPV4_ADDRS, enc_ipv4);
1960         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1961                              FLOW_DISSECTOR_KEY_ENC_IPV6_ADDRS, enc_ipv6);
1962         if (FL_KEY_IS_MASKED(mask, enc_ipv4) ||
1963             FL_KEY_IS_MASKED(mask, enc_ipv6))
1964                 FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_ENC_CONTROL,
1965                            enc_control);
1966         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1967                              FLOW_DISSECTOR_KEY_ENC_PORTS, enc_tp);
1968         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1969                              FLOW_DISSECTOR_KEY_ENC_IP, enc_ip);
1970         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1971                              FLOW_DISSECTOR_KEY_ENC_OPTS, enc_opts);
1972         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1973                              FLOW_DISSECTOR_KEY_CT, ct);
1974         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1975                              FLOW_DISSECTOR_KEY_HASH, hash);
1976         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1977                              FLOW_DISSECTOR_KEY_NUM_OF_VLANS, num_of_vlans);
1978         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1979                              FLOW_DISSECTOR_KEY_PPPOE, pppoe);
1980         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1981                              FLOW_DISSECTOR_KEY_L2TPV3, l2tpv3);
1982
1983         skb_flow_dissector_init(dissector, keys, cnt);
1984 }
1985
1986 static struct fl_flow_mask *fl_create_new_mask(struct cls_fl_head *head,
1987                                                struct fl_flow_mask *mask)
1988 {
1989         struct fl_flow_mask *newmask;
1990         int err;
1991
1992         newmask = kzalloc(sizeof(*newmask), GFP_KERNEL);
1993         if (!newmask)
1994                 return ERR_PTR(-ENOMEM);
1995
1996         fl_mask_copy(newmask, mask);
1997
1998         if ((newmask->key.tp_range.tp_min.dst &&
1999              newmask->key.tp_range.tp_max.dst) ||
2000             (newmask->key.tp_range.tp_min.src &&
2001              newmask->key.tp_range.tp_max.src))
2002                 newmask->flags |= TCA_FLOWER_MASK_FLAGS_RANGE;
2003
2004         err = fl_init_mask_hashtable(newmask);
2005         if (err)
2006                 goto errout_free;
2007
2008         fl_init_dissector(&newmask->dissector, &newmask->key);
2009
2010         INIT_LIST_HEAD_RCU(&newmask->filters);
2011
2012         refcount_set(&newmask->refcnt, 1);
2013         err = rhashtable_replace_fast(&head->ht, &mask->ht_node,
2014                                       &newmask->ht_node, mask_ht_params);
2015         if (err)
2016                 goto errout_destroy;
2017
2018         spin_lock(&head->masks_lock);
2019         list_add_tail_rcu(&newmask->list, &head->masks);
2020         spin_unlock(&head->masks_lock);
2021
2022         return newmask;
2023
2024 errout_destroy:
2025         rhashtable_destroy(&newmask->ht);
2026 errout_free:
2027         kfree(newmask);
2028
2029         return ERR_PTR(err);
2030 }
2031
2032 static int fl_check_assign_mask(struct cls_fl_head *head,
2033                                 struct cls_fl_filter *fnew,
2034                                 struct cls_fl_filter *fold,
2035                                 struct fl_flow_mask *mask)
2036 {
2037         struct fl_flow_mask *newmask;
2038         int ret = 0;
2039
2040         rcu_read_lock();
2041
2042         /* Insert mask as temporary node to prevent concurrent creation of mask
2043          * with same key. Any concurrent lookups with same key will return
2044          * -EAGAIN because mask's refcnt is zero.
2045          */
2046         fnew->mask = rhashtable_lookup_get_insert_fast(&head->ht,
2047                                                        &mask->ht_node,
2048                                                        mask_ht_params);
2049         if (!fnew->mask) {
2050                 rcu_read_unlock();
2051
2052                 if (fold) {
2053                         ret = -EINVAL;
2054                         goto errout_cleanup;
2055                 }
2056
2057                 newmask = fl_create_new_mask(head, mask);
2058                 if (IS_ERR(newmask)) {
2059                         ret = PTR_ERR(newmask);
2060                         goto errout_cleanup;
2061                 }
2062
2063                 fnew->mask = newmask;
2064                 return 0;
2065         } else if (IS_ERR(fnew->mask)) {
2066                 ret = PTR_ERR(fnew->mask);
2067         } else if (fold && fold->mask != fnew->mask) {
2068                 ret = -EINVAL;
2069         } else if (!refcount_inc_not_zero(&fnew->mask->refcnt)) {
2070                 /* Mask was deleted concurrently, try again */
2071                 ret = -EAGAIN;
2072         }
2073         rcu_read_unlock();
2074         return ret;
2075
2076 errout_cleanup:
2077         rhashtable_remove_fast(&head->ht, &mask->ht_node,
2078                                mask_ht_params);
2079         return ret;
2080 }
2081
2082 static int fl_set_parms(struct net *net, struct tcf_proto *tp,
2083                         struct cls_fl_filter *f, struct fl_flow_mask *mask,
2084                         unsigned long base, struct nlattr **tb,
2085                         struct nlattr *est,
2086                         struct fl_flow_tmplt *tmplt,
2087                         u32 flags, u32 fl_flags,
2088                         struct netlink_ext_ack *extack)
2089 {
2090         int err;
2091
2092         err = tcf_exts_validate_ex(net, tp, tb, est, &f->exts, flags,
2093                                    fl_flags, extack);
2094         if (err < 0)
2095                 return err;
2096
2097         if (tb[TCA_FLOWER_CLASSID]) {
2098                 f->res.classid = nla_get_u32(tb[TCA_FLOWER_CLASSID]);
2099                 if (flags & TCA_ACT_FLAGS_NO_RTNL)
2100                         rtnl_lock();
2101                 tcf_bind_filter(tp, &f->res, base);
2102                 if (flags & TCA_ACT_FLAGS_NO_RTNL)
2103                         rtnl_unlock();
2104         }
2105
2106         err = fl_set_key(net, tb, &f->key, &mask->key, extack);
2107         if (err)
2108                 return err;
2109
2110         fl_mask_update_range(mask);
2111         fl_set_masked_key(&f->mkey, &f->key, mask);
2112
2113         if (!fl_mask_fits_tmplt(tmplt, mask)) {
2114                 NL_SET_ERR_MSG_MOD(extack, "Mask does not fit the template");
2115                 return -EINVAL;
2116         }
2117
2118         return 0;
2119 }
2120
2121 static int fl_ht_insert_unique(struct cls_fl_filter *fnew,
2122                                struct cls_fl_filter *fold,
2123                                bool *in_ht)
2124 {
2125         struct fl_flow_mask *mask = fnew->mask;
2126         int err;
2127
2128         err = rhashtable_lookup_insert_fast(&mask->ht,
2129                                             &fnew->ht_node,
2130                                             mask->filter_ht_params);
2131         if (err) {
2132                 *in_ht = false;
2133                 /* It is okay if filter with same key exists when
2134                  * overwriting.
2135                  */
2136                 return fold && err == -EEXIST ? 0 : err;
2137         }
2138
2139         *in_ht = true;
2140         return 0;
2141 }
2142
2143 static int fl_change(struct net *net, struct sk_buff *in_skb,
2144                      struct tcf_proto *tp, unsigned long base,
2145                      u32 handle, struct nlattr **tca,
2146                      void **arg, u32 flags,
2147                      struct netlink_ext_ack *extack)
2148 {
2149         struct cls_fl_head *head = fl_head_dereference(tp);
2150         bool rtnl_held = !(flags & TCA_ACT_FLAGS_NO_RTNL);
2151         struct cls_fl_filter *fold = *arg;
2152         struct cls_fl_filter *fnew;
2153         struct fl_flow_mask *mask;
2154         struct nlattr **tb;
2155         bool in_ht;
2156         int err;
2157
2158         if (!tca[TCA_OPTIONS]) {
2159                 err = -EINVAL;
2160                 goto errout_fold;
2161         }
2162
2163         mask = kzalloc(sizeof(struct fl_flow_mask), GFP_KERNEL);
2164         if (!mask) {
2165                 err = -ENOBUFS;
2166                 goto errout_fold;
2167         }
2168
2169         tb = kcalloc(TCA_FLOWER_MAX + 1, sizeof(struct nlattr *), GFP_KERNEL);
2170         if (!tb) {
2171                 err = -ENOBUFS;
2172                 goto errout_mask_alloc;
2173         }
2174
2175         err = nla_parse_nested_deprecated(tb, TCA_FLOWER_MAX,
2176                                           tca[TCA_OPTIONS], fl_policy, NULL);
2177         if (err < 0)
2178                 goto errout_tb;
2179
2180         if (fold && handle && fold->handle != handle) {
2181                 err = -EINVAL;
2182                 goto errout_tb;
2183         }
2184
2185         fnew = kzalloc(sizeof(*fnew), GFP_KERNEL);
2186         if (!fnew) {
2187                 err = -ENOBUFS;
2188                 goto errout_tb;
2189         }
2190         INIT_LIST_HEAD(&fnew->hw_list);
2191         refcount_set(&fnew->refcnt, 1);
2192
2193         err = tcf_exts_init(&fnew->exts, net, TCA_FLOWER_ACT, 0);
2194         if (err < 0)
2195                 goto errout;
2196
2197         if (tb[TCA_FLOWER_FLAGS]) {
2198                 fnew->flags = nla_get_u32(tb[TCA_FLOWER_FLAGS]);
2199
2200                 if (!tc_flags_valid(fnew->flags)) {
2201                         err = -EINVAL;
2202                         goto errout;
2203                 }
2204         }
2205
2206         err = fl_set_parms(net, tp, fnew, mask, base, tb, tca[TCA_RATE],
2207                            tp->chain->tmplt_priv, flags, fnew->flags,
2208                            extack);
2209         if (err)
2210                 goto errout;
2211
2212         err = fl_check_assign_mask(head, fnew, fold, mask);
2213         if (err)
2214                 goto errout;
2215
2216         err = fl_ht_insert_unique(fnew, fold, &in_ht);
2217         if (err)
2218                 goto errout_mask;
2219
2220         if (!tc_skip_hw(fnew->flags)) {
2221                 err = fl_hw_replace_filter(tp, fnew, rtnl_held, extack);
2222                 if (err)
2223                         goto errout_ht;
2224         }
2225
2226         if (!tc_in_hw(fnew->flags))
2227                 fnew->flags |= TCA_CLS_FLAGS_NOT_IN_HW;
2228
2229         spin_lock(&tp->lock);
2230
2231         /* tp was deleted concurrently. -EAGAIN will cause caller to lookup
2232          * proto again or create new one, if necessary.
2233          */
2234         if (tp->deleting) {
2235                 err = -EAGAIN;
2236                 goto errout_hw;
2237         }
2238
2239         if (fold) {
2240                 /* Fold filter was deleted concurrently. Retry lookup. */
2241                 if (fold->deleted) {
2242                         err = -EAGAIN;
2243                         goto errout_hw;
2244                 }
2245
2246                 fnew->handle = handle;
2247
2248                 if (!in_ht) {
2249                         struct rhashtable_params params =
2250                                 fnew->mask->filter_ht_params;
2251
2252                         err = rhashtable_insert_fast(&fnew->mask->ht,
2253                                                      &fnew->ht_node,
2254                                                      params);
2255                         if (err)
2256                                 goto errout_hw;
2257                         in_ht = true;
2258                 }
2259
2260                 refcount_inc(&fnew->refcnt);
2261                 rhashtable_remove_fast(&fold->mask->ht,
2262                                        &fold->ht_node,
2263                                        fold->mask->filter_ht_params);
2264                 idr_replace(&head->handle_idr, fnew, fnew->handle);
2265                 list_replace_rcu(&fold->list, &fnew->list);
2266                 fold->deleted = true;
2267
2268                 spin_unlock(&tp->lock);
2269
2270                 fl_mask_put(head, fold->mask);
2271                 if (!tc_skip_hw(fold->flags))
2272                         fl_hw_destroy_filter(tp, fold, rtnl_held, NULL);
2273                 tcf_unbind_filter(tp, &fold->res);
2274                 /* Caller holds reference to fold, so refcnt is always > 0
2275                  * after this.
2276                  */
2277                 refcount_dec(&fold->refcnt);
2278                 __fl_put(fold);
2279         } else {
2280                 if (handle) {
2281                         /* user specifies a handle and it doesn't exist */
2282                         err = idr_alloc_u32(&head->handle_idr, fnew, &handle,
2283                                             handle, GFP_ATOMIC);
2284
2285                         /* Filter with specified handle was concurrently
2286                          * inserted after initial check in cls_api. This is not
2287                          * necessarily an error if NLM_F_EXCL is not set in
2288                          * message flags. Returning EAGAIN will cause cls_api to
2289                          * try to update concurrently inserted rule.
2290                          */
2291                         if (err == -ENOSPC)
2292                                 err = -EAGAIN;
2293                 } else {
2294                         handle = 1;
2295                         err = idr_alloc_u32(&head->handle_idr, fnew, &handle,
2296                                             INT_MAX, GFP_ATOMIC);
2297                 }
2298                 if (err)
2299                         goto errout_hw;
2300
2301                 refcount_inc(&fnew->refcnt);
2302                 fnew->handle = handle;
2303                 list_add_tail_rcu(&fnew->list, &fnew->mask->filters);
2304                 spin_unlock(&tp->lock);
2305         }
2306
2307         *arg = fnew;
2308
2309         kfree(tb);
2310         tcf_queue_work(&mask->rwork, fl_uninit_mask_free_work);
2311         return 0;
2312
2313 errout_ht:
2314         spin_lock(&tp->lock);
2315 errout_hw:
2316         fnew->deleted = true;
2317         spin_unlock(&tp->lock);
2318         if (!tc_skip_hw(fnew->flags))
2319                 fl_hw_destroy_filter(tp, fnew, rtnl_held, NULL);
2320         if (in_ht)
2321                 rhashtable_remove_fast(&fnew->mask->ht, &fnew->ht_node,
2322                                        fnew->mask->filter_ht_params);
2323 errout_mask:
2324         fl_mask_put(head, fnew->mask);
2325 errout:
2326         __fl_put(fnew);
2327 errout_tb:
2328         kfree(tb);
2329 errout_mask_alloc:
2330         tcf_queue_work(&mask->rwork, fl_uninit_mask_free_work);
2331 errout_fold:
2332         if (fold)
2333                 __fl_put(fold);
2334         return err;
2335 }
2336
2337 static int fl_delete(struct tcf_proto *tp, void *arg, bool *last,
2338                      bool rtnl_held, struct netlink_ext_ack *extack)
2339 {
2340         struct cls_fl_head *head = fl_head_dereference(tp);
2341         struct cls_fl_filter *f = arg;
2342         bool last_on_mask;
2343         int err = 0;
2344
2345         err = __fl_delete(tp, f, &last_on_mask, rtnl_held, extack);
2346         *last = list_empty(&head->masks);
2347         __fl_put(f);
2348
2349         return err;
2350 }
2351
2352 static void fl_walk(struct tcf_proto *tp, struct tcf_walker *arg,
2353                     bool rtnl_held)
2354 {
2355         struct cls_fl_head *head = fl_head_dereference(tp);
2356         unsigned long id = arg->cookie, tmp;
2357         struct cls_fl_filter *f;
2358
2359         arg->count = arg->skip;
2360
2361         rcu_read_lock();
2362         idr_for_each_entry_continue_ul(&head->handle_idr, f, tmp, id) {
2363                 /* don't return filters that are being deleted */
2364                 if (!refcount_inc_not_zero(&f->refcnt))
2365                         continue;
2366                 rcu_read_unlock();
2367
2368                 if (arg->fn(tp, f, arg) < 0) {
2369                         __fl_put(f);
2370                         arg->stop = 1;
2371                         rcu_read_lock();
2372                         break;
2373                 }
2374                 __fl_put(f);
2375                 arg->count++;
2376                 rcu_read_lock();
2377         }
2378         rcu_read_unlock();
2379         arg->cookie = id;
2380 }
2381
2382 static struct cls_fl_filter *
2383 fl_get_next_hw_filter(struct tcf_proto *tp, struct cls_fl_filter *f, bool add)
2384 {
2385         struct cls_fl_head *head = fl_head_dereference(tp);
2386
2387         spin_lock(&tp->lock);
2388         if (list_empty(&head->hw_filters)) {
2389                 spin_unlock(&tp->lock);
2390                 return NULL;
2391         }
2392
2393         if (!f)
2394                 f = list_entry(&head->hw_filters, struct cls_fl_filter,
2395                                hw_list);
2396         list_for_each_entry_continue(f, &head->hw_filters, hw_list) {
2397                 if (!(add && f->deleted) && refcount_inc_not_zero(&f->refcnt)) {
2398                         spin_unlock(&tp->lock);
2399                         return f;
2400                 }
2401         }
2402
2403         spin_unlock(&tp->lock);
2404         return NULL;
2405 }
2406
2407 static int fl_reoffload(struct tcf_proto *tp, bool add, flow_setup_cb_t *cb,
2408                         void *cb_priv, struct netlink_ext_ack *extack)
2409 {
2410         struct tcf_block *block = tp->chain->block;
2411         struct flow_cls_offload cls_flower = {};
2412         struct cls_fl_filter *f = NULL;
2413         int err;
2414
2415         /* hw_filters list can only be changed by hw offload functions after
2416          * obtaining rtnl lock. Make sure it is not changed while reoffload is
2417          * iterating it.
2418          */
2419         ASSERT_RTNL();
2420
2421         while ((f = fl_get_next_hw_filter(tp, f, add))) {
2422                 cls_flower.rule =
2423                         flow_rule_alloc(tcf_exts_num_actions(&f->exts));
2424                 if (!cls_flower.rule) {
2425                         __fl_put(f);
2426                         return -ENOMEM;
2427                 }
2428
2429                 tc_cls_common_offload_init(&cls_flower.common, tp, f->flags,
2430                                            extack);
2431                 cls_flower.command = add ?
2432                         FLOW_CLS_REPLACE : FLOW_CLS_DESTROY;
2433                 cls_flower.cookie = (unsigned long)f;
2434                 cls_flower.rule->match.dissector = &f->mask->dissector;
2435                 cls_flower.rule->match.mask = &f->mask->key;
2436                 cls_flower.rule->match.key = &f->mkey;
2437
2438                 err = tc_setup_offload_action(&cls_flower.rule->action, &f->exts,
2439                                               cls_flower.common.extack);
2440                 if (err) {
2441                         kfree(cls_flower.rule);
2442                         if (tc_skip_sw(f->flags)) {
2443                                 __fl_put(f);
2444                                 return err;
2445                         }
2446                         goto next_flow;
2447                 }
2448
2449                 cls_flower.classid = f->res.classid;
2450
2451                 err = tc_setup_cb_reoffload(block, tp, add, cb,
2452                                             TC_SETUP_CLSFLOWER, &cls_flower,
2453                                             cb_priv, &f->flags,
2454                                             &f->in_hw_count);
2455                 tc_cleanup_offload_action(&cls_flower.rule->action);
2456                 kfree(cls_flower.rule);
2457
2458                 if (err) {
2459                         __fl_put(f);
2460                         return err;
2461                 }
2462 next_flow:
2463                 __fl_put(f);
2464         }
2465
2466         return 0;
2467 }
2468
2469 static void fl_hw_add(struct tcf_proto *tp, void *type_data)
2470 {
2471         struct flow_cls_offload *cls_flower = type_data;
2472         struct cls_fl_filter *f =
2473                 (struct cls_fl_filter *) cls_flower->cookie;
2474         struct cls_fl_head *head = fl_head_dereference(tp);
2475
2476         spin_lock(&tp->lock);
2477         list_add(&f->hw_list, &head->hw_filters);
2478         spin_unlock(&tp->lock);
2479 }
2480
2481 static void fl_hw_del(struct tcf_proto *tp, void *type_data)
2482 {
2483         struct flow_cls_offload *cls_flower = type_data;
2484         struct cls_fl_filter *f =
2485                 (struct cls_fl_filter *) cls_flower->cookie;
2486
2487         spin_lock(&tp->lock);
2488         if (!list_empty(&f->hw_list))
2489                 list_del_init(&f->hw_list);
2490         spin_unlock(&tp->lock);
2491 }
2492
2493 static int fl_hw_create_tmplt(struct tcf_chain *chain,
2494                               struct fl_flow_tmplt *tmplt)
2495 {
2496         struct flow_cls_offload cls_flower = {};
2497         struct tcf_block *block = chain->block;
2498
2499         cls_flower.rule = flow_rule_alloc(0);
2500         if (!cls_flower.rule)
2501                 return -ENOMEM;
2502
2503         cls_flower.common.chain_index = chain->index;
2504         cls_flower.command = FLOW_CLS_TMPLT_CREATE;
2505         cls_flower.cookie = (unsigned long) tmplt;
2506         cls_flower.rule->match.dissector = &tmplt->dissector;
2507         cls_flower.rule->match.mask = &tmplt->mask;
2508         cls_flower.rule->match.key = &tmplt->dummy_key;
2509
2510         /* We don't care if driver (any of them) fails to handle this
2511          * call. It serves just as a hint for it.
2512          */
2513         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false, true);
2514         kfree(cls_flower.rule);
2515
2516         return 0;
2517 }
2518
2519 static void fl_hw_destroy_tmplt(struct tcf_chain *chain,
2520                                 struct fl_flow_tmplt *tmplt)
2521 {
2522         struct flow_cls_offload cls_flower = {};
2523         struct tcf_block *block = chain->block;
2524
2525         cls_flower.common.chain_index = chain->index;
2526         cls_flower.command = FLOW_CLS_TMPLT_DESTROY;
2527         cls_flower.cookie = (unsigned long) tmplt;
2528
2529         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false, true);
2530 }
2531
2532 static void *fl_tmplt_create(struct net *net, struct tcf_chain *chain,
2533                              struct nlattr **tca,
2534                              struct netlink_ext_ack *extack)
2535 {
2536         struct fl_flow_tmplt *tmplt;
2537         struct nlattr **tb;
2538         int err;
2539
2540         if (!tca[TCA_OPTIONS])
2541                 return ERR_PTR(-EINVAL);
2542
2543         tb = kcalloc(TCA_FLOWER_MAX + 1, sizeof(struct nlattr *), GFP_KERNEL);
2544         if (!tb)
2545                 return ERR_PTR(-ENOBUFS);
2546         err = nla_parse_nested_deprecated(tb, TCA_FLOWER_MAX,
2547                                           tca[TCA_OPTIONS], fl_policy, NULL);
2548         if (err)
2549                 goto errout_tb;
2550
2551         tmplt = kzalloc(sizeof(*tmplt), GFP_KERNEL);
2552         if (!tmplt) {
2553                 err = -ENOMEM;
2554                 goto errout_tb;
2555         }
2556         tmplt->chain = chain;
2557         err = fl_set_key(net, tb, &tmplt->dummy_key, &tmplt->mask, extack);
2558         if (err)
2559                 goto errout_tmplt;
2560
2561         fl_init_dissector(&tmplt->dissector, &tmplt->mask);
2562
2563         err = fl_hw_create_tmplt(chain, tmplt);
2564         if (err)
2565                 goto errout_tmplt;
2566
2567         kfree(tb);
2568         return tmplt;
2569
2570 errout_tmplt:
2571         kfree(tmplt);
2572 errout_tb:
2573         kfree(tb);
2574         return ERR_PTR(err);
2575 }
2576
2577 static void fl_tmplt_destroy(void *tmplt_priv)
2578 {
2579         struct fl_flow_tmplt *tmplt = tmplt_priv;
2580
2581         fl_hw_destroy_tmplt(tmplt->chain, tmplt);
2582         kfree(tmplt);
2583 }
2584
2585 static int fl_dump_key_val(struct sk_buff *skb,
2586                            void *val, int val_type,
2587                            void *mask, int mask_type, int len)
2588 {
2589         int err;
2590
2591         if (!memchr_inv(mask, 0, len))
2592                 return 0;
2593         err = nla_put(skb, val_type, len, val);
2594         if (err)
2595                 return err;
2596         if (mask_type != TCA_FLOWER_UNSPEC) {
2597                 err = nla_put(skb, mask_type, len, mask);
2598                 if (err)
2599                         return err;
2600         }
2601         return 0;
2602 }
2603
2604 static int fl_dump_key_port_range(struct sk_buff *skb, struct fl_flow_key *key,
2605                                   struct fl_flow_key *mask)
2606 {
2607         if (fl_dump_key_val(skb, &key->tp_range.tp_min.dst,
2608                             TCA_FLOWER_KEY_PORT_DST_MIN,
2609                             &mask->tp_range.tp_min.dst, TCA_FLOWER_UNSPEC,
2610                             sizeof(key->tp_range.tp_min.dst)) ||
2611             fl_dump_key_val(skb, &key->tp_range.tp_max.dst,
2612                             TCA_FLOWER_KEY_PORT_DST_MAX,
2613                             &mask->tp_range.tp_max.dst, TCA_FLOWER_UNSPEC,
2614                             sizeof(key->tp_range.tp_max.dst)) ||
2615             fl_dump_key_val(skb, &key->tp_range.tp_min.src,
2616                             TCA_FLOWER_KEY_PORT_SRC_MIN,
2617                             &mask->tp_range.tp_min.src, TCA_FLOWER_UNSPEC,
2618                             sizeof(key->tp_range.tp_min.src)) ||
2619             fl_dump_key_val(skb, &key->tp_range.tp_max.src,
2620                             TCA_FLOWER_KEY_PORT_SRC_MAX,
2621                             &mask->tp_range.tp_max.src, TCA_FLOWER_UNSPEC,
2622                             sizeof(key->tp_range.tp_max.src)))
2623                 return -1;
2624
2625         return 0;
2626 }
2627
2628 static int fl_dump_key_mpls_opt_lse(struct sk_buff *skb,
2629                                     struct flow_dissector_key_mpls *mpls_key,
2630                                     struct flow_dissector_key_mpls *mpls_mask,
2631                                     u8 lse_index)
2632 {
2633         struct flow_dissector_mpls_lse *lse_mask = &mpls_mask->ls[lse_index];
2634         struct flow_dissector_mpls_lse *lse_key = &mpls_key->ls[lse_index];
2635         int err;
2636
2637         err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH,
2638                          lse_index + 1);
2639         if (err)
2640                 return err;
2641
2642         if (lse_mask->mpls_ttl) {
2643                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_TTL,
2644                                  lse_key->mpls_ttl);
2645                 if (err)
2646                         return err;
2647         }
2648         if (lse_mask->mpls_bos) {
2649                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS,
2650                                  lse_key->mpls_bos);
2651                 if (err)
2652                         return err;
2653         }
2654         if (lse_mask->mpls_tc) {
2655                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_TC,
2656                                  lse_key->mpls_tc);
2657                 if (err)
2658                         return err;
2659         }
2660         if (lse_mask->mpls_label) {
2661                 err = nla_put_u32(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL,
2662                                   lse_key->mpls_label);
2663                 if (err)
2664                         return err;
2665         }
2666
2667         return 0;
2668 }
2669
2670 static int fl_dump_key_mpls_opts(struct sk_buff *skb,
2671                                  struct flow_dissector_key_mpls *mpls_key,
2672                                  struct flow_dissector_key_mpls *mpls_mask)
2673 {
2674         struct nlattr *opts;
2675         struct nlattr *lse;
2676         u8 lse_index;
2677         int err;
2678
2679         opts = nla_nest_start(skb, TCA_FLOWER_KEY_MPLS_OPTS);
2680         if (!opts)
2681                 return -EMSGSIZE;
2682
2683         for (lse_index = 0; lse_index < FLOW_DIS_MPLS_MAX; lse_index++) {
2684                 if (!(mpls_mask->used_lses & 1 << lse_index))
2685                         continue;
2686
2687                 lse = nla_nest_start(skb, TCA_FLOWER_KEY_MPLS_OPTS_LSE);
2688                 if (!lse) {
2689                         err = -EMSGSIZE;
2690                         goto err_opts;
2691                 }
2692
2693                 err = fl_dump_key_mpls_opt_lse(skb, mpls_key, mpls_mask,
2694                                                lse_index);
2695                 if (err)
2696                         goto err_opts_lse;
2697                 nla_nest_end(skb, lse);
2698         }
2699         nla_nest_end(skb, opts);
2700
2701         return 0;
2702
2703 err_opts_lse:
2704         nla_nest_cancel(skb, lse);
2705 err_opts:
2706         nla_nest_cancel(skb, opts);
2707
2708         return err;
2709 }
2710
2711 static int fl_dump_key_mpls(struct sk_buff *skb,
2712                             struct flow_dissector_key_mpls *mpls_key,
2713                             struct flow_dissector_key_mpls *mpls_mask)
2714 {
2715         struct flow_dissector_mpls_lse *lse_mask;
2716         struct flow_dissector_mpls_lse *lse_key;
2717         int err;
2718
2719         if (!mpls_mask->used_lses)
2720                 return 0;
2721
2722         lse_mask = &mpls_mask->ls[0];
2723         lse_key = &mpls_key->ls[0];
2724
2725         /* For backward compatibility, don't use the MPLS nested attributes if
2726          * the rule can be expressed using the old attributes.
2727          */
2728         if (mpls_mask->used_lses & ~1 ||
2729             (!lse_mask->mpls_ttl && !lse_mask->mpls_bos &&
2730              !lse_mask->mpls_tc && !lse_mask->mpls_label))
2731                 return fl_dump_key_mpls_opts(skb, mpls_key, mpls_mask);
2732
2733         if (lse_mask->mpls_ttl) {
2734                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_TTL,
2735                                  lse_key->mpls_ttl);
2736                 if (err)
2737                         return err;
2738         }
2739         if (lse_mask->mpls_tc) {
2740                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_TC,
2741                                  lse_key->mpls_tc);
2742                 if (err)
2743                         return err;
2744         }
2745         if (lse_mask->mpls_label) {
2746                 err = nla_put_u32(skb, TCA_FLOWER_KEY_MPLS_LABEL,
2747                                   lse_key->mpls_label);
2748                 if (err)
2749                         return err;
2750         }
2751         if (lse_mask->mpls_bos) {
2752                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_BOS,
2753                                  lse_key->mpls_bos);
2754                 if (err)
2755                         return err;
2756         }
2757         return 0;
2758 }
2759
2760 static int fl_dump_key_ip(struct sk_buff *skb, bool encap,
2761                           struct flow_dissector_key_ip *key,
2762                           struct flow_dissector_key_ip *mask)
2763 {
2764         int tos_key = encap ? TCA_FLOWER_KEY_ENC_IP_TOS : TCA_FLOWER_KEY_IP_TOS;
2765         int ttl_key = encap ? TCA_FLOWER_KEY_ENC_IP_TTL : TCA_FLOWER_KEY_IP_TTL;
2766         int tos_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TOS_MASK : TCA_FLOWER_KEY_IP_TOS_MASK;
2767         int ttl_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TTL_MASK : TCA_FLOWER_KEY_IP_TTL_MASK;
2768
2769         if (fl_dump_key_val(skb, &key->tos, tos_key, &mask->tos, tos_mask, sizeof(key->tos)) ||
2770             fl_dump_key_val(skb, &key->ttl, ttl_key, &mask->ttl, ttl_mask, sizeof(key->ttl)))
2771                 return -1;
2772
2773         return 0;
2774 }
2775
2776 static int fl_dump_key_vlan(struct sk_buff *skb,
2777                             int vlan_id_key, int vlan_prio_key,
2778                             struct flow_dissector_key_vlan *vlan_key,
2779                             struct flow_dissector_key_vlan *vlan_mask)
2780 {
2781         int err;
2782
2783         if (!memchr_inv(vlan_mask, 0, sizeof(*vlan_mask)))
2784                 return 0;
2785         if (vlan_mask->vlan_id) {
2786                 err = nla_put_u16(skb, vlan_id_key,
2787                                   vlan_key->vlan_id);
2788                 if (err)
2789                         return err;
2790         }
2791         if (vlan_mask->vlan_priority) {
2792                 err = nla_put_u8(skb, vlan_prio_key,
2793                                  vlan_key->vlan_priority);
2794                 if (err)
2795                         return err;
2796         }
2797         return 0;
2798 }
2799
2800 static void fl_get_key_flag(u32 dissector_key, u32 dissector_mask,
2801                             u32 *flower_key, u32 *flower_mask,
2802                             u32 flower_flag_bit, u32 dissector_flag_bit)
2803 {
2804         if (dissector_mask & dissector_flag_bit) {
2805                 *flower_mask |= flower_flag_bit;
2806                 if (dissector_key & dissector_flag_bit)
2807                         *flower_key |= flower_flag_bit;
2808         }
2809 }
2810
2811 static int fl_dump_key_flags(struct sk_buff *skb, u32 flags_key, u32 flags_mask)
2812 {
2813         u32 key, mask;
2814         __be32 _key, _mask;
2815         int err;
2816
2817         if (!memchr_inv(&flags_mask, 0, sizeof(flags_mask)))
2818                 return 0;
2819
2820         key = 0;
2821         mask = 0;
2822
2823         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
2824                         TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT, FLOW_DIS_IS_FRAGMENT);
2825         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
2826                         TCA_FLOWER_KEY_FLAGS_FRAG_IS_FIRST,
2827                         FLOW_DIS_FIRST_FRAG);
2828
2829         _key = cpu_to_be32(key);
2830         _mask = cpu_to_be32(mask);
2831
2832         err = nla_put(skb, TCA_FLOWER_KEY_FLAGS, 4, &_key);
2833         if (err)
2834                 return err;
2835
2836         return nla_put(skb, TCA_FLOWER_KEY_FLAGS_MASK, 4, &_mask);
2837 }
2838
2839 static int fl_dump_key_geneve_opt(struct sk_buff *skb,
2840                                   struct flow_dissector_key_enc_opts *enc_opts)
2841 {
2842         struct geneve_opt *opt;
2843         struct nlattr *nest;
2844         int opt_off = 0;
2845
2846         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_GENEVE);
2847         if (!nest)
2848                 goto nla_put_failure;
2849
2850         while (enc_opts->len > opt_off) {
2851                 opt = (struct geneve_opt *)&enc_opts->data[opt_off];
2852
2853                 if (nla_put_be16(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS,
2854                                  opt->opt_class))
2855                         goto nla_put_failure;
2856                 if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE,
2857                                opt->type))
2858                         goto nla_put_failure;
2859                 if (nla_put(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA,
2860                             opt->length * 4, opt->opt_data))
2861                         goto nla_put_failure;
2862
2863                 opt_off += sizeof(struct geneve_opt) + opt->length * 4;
2864         }
2865         nla_nest_end(skb, nest);
2866         return 0;
2867
2868 nla_put_failure:
2869         nla_nest_cancel(skb, nest);
2870         return -EMSGSIZE;
2871 }
2872
2873 static int fl_dump_key_vxlan_opt(struct sk_buff *skb,
2874                                  struct flow_dissector_key_enc_opts *enc_opts)
2875 {
2876         struct vxlan_metadata *md;
2877         struct nlattr *nest;
2878
2879         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_VXLAN);
2880         if (!nest)
2881                 goto nla_put_failure;
2882
2883         md = (struct vxlan_metadata *)&enc_opts->data[0];
2884         if (nla_put_u32(skb, TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP, md->gbp))
2885                 goto nla_put_failure;
2886
2887         nla_nest_end(skb, nest);
2888         return 0;
2889
2890 nla_put_failure:
2891         nla_nest_cancel(skb, nest);
2892         return -EMSGSIZE;
2893 }
2894
2895 static int fl_dump_key_erspan_opt(struct sk_buff *skb,
2896                                   struct flow_dissector_key_enc_opts *enc_opts)
2897 {
2898         struct erspan_metadata *md;
2899         struct nlattr *nest;
2900
2901         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_ERSPAN);
2902         if (!nest)
2903                 goto nla_put_failure;
2904
2905         md = (struct erspan_metadata *)&enc_opts->data[0];
2906         if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER, md->version))
2907                 goto nla_put_failure;
2908
2909         if (md->version == 1 &&
2910             nla_put_be32(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX, md->u.index))
2911                 goto nla_put_failure;
2912
2913         if (md->version == 2 &&
2914             (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR,
2915                         md->u.md2.dir) ||
2916              nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID,
2917                         get_hwid(&md->u.md2))))
2918                 goto nla_put_failure;
2919
2920         nla_nest_end(skb, nest);
2921         return 0;
2922
2923 nla_put_failure:
2924         nla_nest_cancel(skb, nest);
2925         return -EMSGSIZE;
2926 }
2927
2928 static int fl_dump_key_gtp_opt(struct sk_buff *skb,
2929                                struct flow_dissector_key_enc_opts *enc_opts)
2930
2931 {
2932         struct gtp_pdu_session_info *session_info;
2933         struct nlattr *nest;
2934
2935         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_GTP);
2936         if (!nest)
2937                 goto nla_put_failure;
2938
2939         session_info = (struct gtp_pdu_session_info *)&enc_opts->data[0];
2940
2941         if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE,
2942                        session_info->pdu_type))
2943                 goto nla_put_failure;
2944
2945         if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_GTP_QFI, session_info->qfi))
2946                 goto nla_put_failure;
2947
2948         nla_nest_end(skb, nest);
2949         return 0;
2950
2951 nla_put_failure:
2952         nla_nest_cancel(skb, nest);
2953         return -EMSGSIZE;
2954 }
2955
2956 static int fl_dump_key_ct(struct sk_buff *skb,
2957                           struct flow_dissector_key_ct *key,
2958                           struct flow_dissector_key_ct *mask)
2959 {
2960         if (IS_ENABLED(CONFIG_NF_CONNTRACK) &&
2961             fl_dump_key_val(skb, &key->ct_state, TCA_FLOWER_KEY_CT_STATE,
2962                             &mask->ct_state, TCA_FLOWER_KEY_CT_STATE_MASK,
2963                             sizeof(key->ct_state)))
2964                 goto nla_put_failure;
2965
2966         if (IS_ENABLED(CONFIG_NF_CONNTRACK_ZONES) &&
2967             fl_dump_key_val(skb, &key->ct_zone, TCA_FLOWER_KEY_CT_ZONE,
2968                             &mask->ct_zone, TCA_FLOWER_KEY_CT_ZONE_MASK,
2969                             sizeof(key->ct_zone)))
2970                 goto nla_put_failure;
2971
2972         if (IS_ENABLED(CONFIG_NF_CONNTRACK_MARK) &&
2973             fl_dump_key_val(skb, &key->ct_mark, TCA_FLOWER_KEY_CT_MARK,
2974                             &mask->ct_mark, TCA_FLOWER_KEY_CT_MARK_MASK,
2975                             sizeof(key->ct_mark)))
2976                 goto nla_put_failure;
2977
2978         if (IS_ENABLED(CONFIG_NF_CONNTRACK_LABELS) &&
2979             fl_dump_key_val(skb, &key->ct_labels, TCA_FLOWER_KEY_CT_LABELS,
2980                             &mask->ct_labels, TCA_FLOWER_KEY_CT_LABELS_MASK,
2981                             sizeof(key->ct_labels)))
2982                 goto nla_put_failure;
2983
2984         return 0;
2985
2986 nla_put_failure:
2987         return -EMSGSIZE;
2988 }
2989
2990 static int fl_dump_key_options(struct sk_buff *skb, int enc_opt_type,
2991                                struct flow_dissector_key_enc_opts *enc_opts)
2992 {
2993         struct nlattr *nest;
2994         int err;
2995
2996         if (!enc_opts->len)
2997                 return 0;
2998
2999         nest = nla_nest_start_noflag(skb, enc_opt_type);
3000         if (!nest)
3001                 goto nla_put_failure;
3002
3003         switch (enc_opts->dst_opt_type) {
3004         case TUNNEL_GENEVE_OPT:
3005                 err = fl_dump_key_geneve_opt(skb, enc_opts);
3006                 if (err)
3007                         goto nla_put_failure;
3008                 break;
3009         case TUNNEL_VXLAN_OPT:
3010                 err = fl_dump_key_vxlan_opt(skb, enc_opts);
3011                 if (err)
3012                         goto nla_put_failure;
3013                 break;
3014         case TUNNEL_ERSPAN_OPT:
3015                 err = fl_dump_key_erspan_opt(skb, enc_opts);
3016                 if (err)
3017                         goto nla_put_failure;
3018                 break;
3019         case TUNNEL_GTP_OPT:
3020                 err = fl_dump_key_gtp_opt(skb, enc_opts);
3021                 if (err)
3022                         goto nla_put_failure;
3023                 break;
3024         default:
3025                 goto nla_put_failure;
3026         }
3027         nla_nest_end(skb, nest);
3028         return 0;
3029
3030 nla_put_failure:
3031         nla_nest_cancel(skb, nest);
3032         return -EMSGSIZE;
3033 }
3034
3035 static int fl_dump_key_enc_opt(struct sk_buff *skb,
3036                                struct flow_dissector_key_enc_opts *key_opts,
3037                                struct flow_dissector_key_enc_opts *msk_opts)
3038 {
3039         int err;
3040
3041         err = fl_dump_key_options(skb, TCA_FLOWER_KEY_ENC_OPTS, key_opts);
3042         if (err)
3043                 return err;
3044
3045         return fl_dump_key_options(skb, TCA_FLOWER_KEY_ENC_OPTS_MASK, msk_opts);
3046 }
3047
3048 static int fl_dump_key(struct sk_buff *skb, struct net *net,
3049                        struct fl_flow_key *key, struct fl_flow_key *mask)
3050 {
3051         if (mask->meta.ingress_ifindex) {
3052                 struct net_device *dev;
3053
3054                 dev = __dev_get_by_index(net, key->meta.ingress_ifindex);
3055                 if (dev && nla_put_string(skb, TCA_FLOWER_INDEV, dev->name))
3056                         goto nla_put_failure;
3057         }
3058
3059         if (fl_dump_key_val(skb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
3060                             mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
3061                             sizeof(key->eth.dst)) ||
3062             fl_dump_key_val(skb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
3063                             mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
3064                             sizeof(key->eth.src)) ||
3065             fl_dump_key_val(skb, &key->basic.n_proto, TCA_FLOWER_KEY_ETH_TYPE,
3066                             &mask->basic.n_proto, TCA_FLOWER_UNSPEC,
3067                             sizeof(key->basic.n_proto)))
3068                 goto nla_put_failure;
3069
3070         if (mask->num_of_vlans.num_of_vlans) {
3071                 if (nla_put_u8(skb, TCA_FLOWER_KEY_NUM_OF_VLANS, key->num_of_vlans.num_of_vlans))
3072                         goto nla_put_failure;
3073         }
3074
3075         if (fl_dump_key_mpls(skb, &key->mpls, &mask->mpls))
3076                 goto nla_put_failure;
3077
3078         if (fl_dump_key_vlan(skb, TCA_FLOWER_KEY_VLAN_ID,
3079                              TCA_FLOWER_KEY_VLAN_PRIO, &key->vlan, &mask->vlan))
3080                 goto nla_put_failure;
3081
3082         if (fl_dump_key_vlan(skb, TCA_FLOWER_KEY_CVLAN_ID,
3083                              TCA_FLOWER_KEY_CVLAN_PRIO,
3084                              &key->cvlan, &mask->cvlan) ||
3085             (mask->cvlan.vlan_tpid &&
3086              nla_put_be16(skb, TCA_FLOWER_KEY_VLAN_ETH_TYPE,
3087                           key->cvlan.vlan_tpid)))
3088                 goto nla_put_failure;
3089
3090         if (mask->basic.n_proto) {
3091                 if (mask->cvlan.vlan_eth_type) {
3092                         if (nla_put_be16(skb, TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
3093                                          key->basic.n_proto))
3094                                 goto nla_put_failure;
3095                 } else if (mask->vlan.vlan_eth_type) {
3096                         if (nla_put_be16(skb, TCA_FLOWER_KEY_VLAN_ETH_TYPE,
3097                                          key->vlan.vlan_eth_type))
3098                                 goto nla_put_failure;
3099                 }
3100         }
3101
3102         if ((key->basic.n_proto == htons(ETH_P_IP) ||
3103              key->basic.n_proto == htons(ETH_P_IPV6)) &&
3104             (fl_dump_key_val(skb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
3105                             &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
3106                             sizeof(key->basic.ip_proto)) ||
3107             fl_dump_key_ip(skb, false, &key->ip, &mask->ip)))
3108                 goto nla_put_failure;
3109
3110         if (mask->pppoe.session_id) {
3111                 if (nla_put_be16(skb, TCA_FLOWER_KEY_PPPOE_SID,
3112                                  key->pppoe.session_id))
3113                         goto nla_put_failure;
3114         }
3115         if (mask->basic.n_proto && mask->pppoe.ppp_proto) {
3116                 if (nla_put_be16(skb, TCA_FLOWER_KEY_PPP_PROTO,
3117                                  key->pppoe.ppp_proto))
3118                         goto nla_put_failure;
3119         }
3120
3121         if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
3122             (fl_dump_key_val(skb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
3123                              &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
3124                              sizeof(key->ipv4.src)) ||
3125              fl_dump_key_val(skb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
3126                              &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
3127                              sizeof(key->ipv4.dst))))
3128                 goto nla_put_failure;
3129         else if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
3130                  (fl_dump_key_val(skb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
3131                                   &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
3132                                   sizeof(key->ipv6.src)) ||
3133                   fl_dump_key_val(skb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
3134                                   &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
3135                                   sizeof(key->ipv6.dst))))
3136                 goto nla_put_failure;
3137
3138         if (key->basic.ip_proto == IPPROTO_TCP &&
3139             (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
3140                              &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
3141                              sizeof(key->tp.src)) ||
3142              fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
3143                              &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
3144                              sizeof(key->tp.dst)) ||
3145              fl_dump_key_val(skb, &key->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS,
3146                              &mask->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS_MASK,
3147                              sizeof(key->tcp.flags))))
3148                 goto nla_put_failure;
3149         else if (key->basic.ip_proto == IPPROTO_UDP &&
3150                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
3151                                   &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
3152                                   sizeof(key->tp.src)) ||
3153                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
3154                                   &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
3155                                   sizeof(key->tp.dst))))
3156                 goto nla_put_failure;
3157         else if (key->basic.ip_proto == IPPROTO_SCTP &&
3158                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_SCTP_SRC,
3159                                   &mask->tp.src, TCA_FLOWER_KEY_SCTP_SRC_MASK,
3160                                   sizeof(key->tp.src)) ||
3161                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_SCTP_DST,
3162                                   &mask->tp.dst, TCA_FLOWER_KEY_SCTP_DST_MASK,
3163                                   sizeof(key->tp.dst))))
3164                 goto nla_put_failure;
3165         else if (key->basic.n_proto == htons(ETH_P_IP) &&
3166                  key->basic.ip_proto == IPPROTO_ICMP &&
3167                  (fl_dump_key_val(skb, &key->icmp.type,
3168                                   TCA_FLOWER_KEY_ICMPV4_TYPE, &mask->icmp.type,
3169                                   TCA_FLOWER_KEY_ICMPV4_TYPE_MASK,
3170                                   sizeof(key->icmp.type)) ||
3171                   fl_dump_key_val(skb, &key->icmp.code,
3172                                   TCA_FLOWER_KEY_ICMPV4_CODE, &mask->icmp.code,
3173                                   TCA_FLOWER_KEY_ICMPV4_CODE_MASK,
3174                                   sizeof(key->icmp.code))))
3175                 goto nla_put_failure;
3176         else if (key->basic.n_proto == htons(ETH_P_IPV6) &&
3177                  key->basic.ip_proto == IPPROTO_ICMPV6 &&
3178                  (fl_dump_key_val(skb, &key->icmp.type,
3179                                   TCA_FLOWER_KEY_ICMPV6_TYPE, &mask->icmp.type,
3180                                   TCA_FLOWER_KEY_ICMPV6_TYPE_MASK,
3181                                   sizeof(key->icmp.type)) ||
3182                   fl_dump_key_val(skb, &key->icmp.code,
3183                                   TCA_FLOWER_KEY_ICMPV6_CODE, &mask->icmp.code,
3184                                   TCA_FLOWER_KEY_ICMPV6_CODE_MASK,
3185                                   sizeof(key->icmp.code))))
3186                 goto nla_put_failure;
3187         else if ((key->basic.n_proto == htons(ETH_P_ARP) ||
3188                   key->basic.n_proto == htons(ETH_P_RARP)) &&
3189                  (fl_dump_key_val(skb, &key->arp.sip,
3190                                   TCA_FLOWER_KEY_ARP_SIP, &mask->arp.sip,
3191                                   TCA_FLOWER_KEY_ARP_SIP_MASK,
3192                                   sizeof(key->arp.sip)) ||
3193                   fl_dump_key_val(skb, &key->arp.tip,
3194                                   TCA_FLOWER_KEY_ARP_TIP, &mask->arp.tip,
3195                                   TCA_FLOWER_KEY_ARP_TIP_MASK,
3196                                   sizeof(key->arp.tip)) ||
3197                   fl_dump_key_val(skb, &key->arp.op,
3198                                   TCA_FLOWER_KEY_ARP_OP, &mask->arp.op,
3199                                   TCA_FLOWER_KEY_ARP_OP_MASK,
3200                                   sizeof(key->arp.op)) ||
3201                   fl_dump_key_val(skb, key->arp.sha, TCA_FLOWER_KEY_ARP_SHA,
3202                                   mask->arp.sha, TCA_FLOWER_KEY_ARP_SHA_MASK,
3203                                   sizeof(key->arp.sha)) ||
3204                   fl_dump_key_val(skb, key->arp.tha, TCA_FLOWER_KEY_ARP_THA,
3205                                   mask->arp.tha, TCA_FLOWER_KEY_ARP_THA_MASK,
3206                                   sizeof(key->arp.tha))))
3207                 goto nla_put_failure;
3208         else if (key->basic.ip_proto == IPPROTO_L2TP &&
3209                  fl_dump_key_val(skb, &key->l2tpv3.session_id,
3210                                  TCA_FLOWER_KEY_L2TPV3_SID,
3211                                  &mask->l2tpv3.session_id,
3212                                  TCA_FLOWER_UNSPEC,
3213                                  sizeof(key->l2tpv3.session_id)))
3214                 goto nla_put_failure;
3215
3216         if ((key->basic.ip_proto == IPPROTO_TCP ||
3217              key->basic.ip_proto == IPPROTO_UDP ||
3218              key->basic.ip_proto == IPPROTO_SCTP) &&
3219              fl_dump_key_port_range(skb, key, mask))
3220                 goto nla_put_failure;
3221
3222         if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
3223             (fl_dump_key_val(skb, &key->enc_ipv4.src,
3224                             TCA_FLOWER_KEY_ENC_IPV4_SRC, &mask->enc_ipv4.src,
3225                             TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
3226                             sizeof(key->enc_ipv4.src)) ||
3227              fl_dump_key_val(skb, &key->enc_ipv4.dst,
3228                              TCA_FLOWER_KEY_ENC_IPV4_DST, &mask->enc_ipv4.dst,
3229                              TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
3230                              sizeof(key->enc_ipv4.dst))))
3231                 goto nla_put_failure;
3232         else if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
3233                  (fl_dump_key_val(skb, &key->enc_ipv6.src,
3234                             TCA_FLOWER_KEY_ENC_IPV6_SRC, &mask->enc_ipv6.src,
3235                             TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
3236                             sizeof(key->enc_ipv6.src)) ||
3237                  fl_dump_key_val(skb, &key->enc_ipv6.dst,
3238                                  TCA_FLOWER_KEY_ENC_IPV6_DST,
3239                                  &mask->enc_ipv6.dst,
3240                                  TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
3241                             sizeof(key->enc_ipv6.dst))))
3242                 goto nla_put_failure;
3243
3244         if (fl_dump_key_val(skb, &key->enc_key_id, TCA_FLOWER_KEY_ENC_KEY_ID,
3245                             &mask->enc_key_id, TCA_FLOWER_UNSPEC,
3246                             sizeof(key->enc_key_id)) ||
3247             fl_dump_key_val(skb, &key->enc_tp.src,
3248                             TCA_FLOWER_KEY_ENC_UDP_SRC_PORT,
3249                             &mask->enc_tp.src,
3250                             TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK,
3251                             sizeof(key->enc_tp.src)) ||
3252             fl_dump_key_val(skb, &key->enc_tp.dst,
3253                             TCA_FLOWER_KEY_ENC_UDP_DST_PORT,
3254                             &mask->enc_tp.dst,
3255                             TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK,
3256                             sizeof(key->enc_tp.dst)) ||
3257             fl_dump_key_ip(skb, true, &key->enc_ip, &mask->enc_ip) ||
3258             fl_dump_key_enc_opt(skb, &key->enc_opts, &mask->enc_opts))
3259                 goto nla_put_failure;
3260
3261         if (fl_dump_key_ct(skb, &key->ct, &mask->ct))
3262                 goto nla_put_failure;
3263
3264         if (fl_dump_key_flags(skb, key->control.flags, mask->control.flags))
3265                 goto nla_put_failure;
3266
3267         if (fl_dump_key_val(skb, &key->hash.hash, TCA_FLOWER_KEY_HASH,
3268                              &mask->hash.hash, TCA_FLOWER_KEY_HASH_MASK,
3269                              sizeof(key->hash.hash)))
3270                 goto nla_put_failure;
3271
3272         return 0;
3273
3274 nla_put_failure:
3275         return -EMSGSIZE;
3276 }
3277
3278 static int fl_dump(struct net *net, struct tcf_proto *tp, void *fh,
3279                    struct sk_buff *skb, struct tcmsg *t, bool rtnl_held)
3280 {
3281         struct cls_fl_filter *f = fh;
3282         struct nlattr *nest;
3283         struct fl_flow_key *key, *mask;
3284         bool skip_hw;
3285
3286         if (!f)
3287                 return skb->len;
3288
3289         t->tcm_handle = f->handle;
3290
3291         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
3292         if (!nest)
3293                 goto nla_put_failure;
3294
3295         spin_lock(&tp->lock);
3296
3297         if (f->res.classid &&
3298             nla_put_u32(skb, TCA_FLOWER_CLASSID, f->res.classid))
3299                 goto nla_put_failure_locked;
3300
3301         key = &f->key;
3302         mask = &f->mask->key;
3303         skip_hw = tc_skip_hw(f->flags);
3304
3305         if (fl_dump_key(skb, net, key, mask))
3306                 goto nla_put_failure_locked;
3307
3308         if (f->flags && nla_put_u32(skb, TCA_FLOWER_FLAGS, f->flags))
3309                 goto nla_put_failure_locked;
3310
3311         spin_unlock(&tp->lock);
3312
3313         if (!skip_hw)
3314                 fl_hw_update_stats(tp, f, rtnl_held);
3315
3316         if (nla_put_u32(skb, TCA_FLOWER_IN_HW_COUNT, f->in_hw_count))
3317                 goto nla_put_failure;
3318
3319         if (tcf_exts_dump(skb, &f->exts))
3320                 goto nla_put_failure;
3321
3322         nla_nest_end(skb, nest);
3323
3324         if (tcf_exts_dump_stats(skb, &f->exts) < 0)
3325                 goto nla_put_failure;
3326
3327         return skb->len;
3328
3329 nla_put_failure_locked:
3330         spin_unlock(&tp->lock);
3331 nla_put_failure:
3332         nla_nest_cancel(skb, nest);
3333         return -1;
3334 }
3335
3336 static int fl_terse_dump(struct net *net, struct tcf_proto *tp, void *fh,
3337                          struct sk_buff *skb, struct tcmsg *t, bool rtnl_held)
3338 {
3339         struct cls_fl_filter *f = fh;
3340         struct nlattr *nest;
3341         bool skip_hw;
3342
3343         if (!f)
3344                 return skb->len;
3345
3346         t->tcm_handle = f->handle;
3347
3348         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
3349         if (!nest)
3350                 goto nla_put_failure;
3351
3352         spin_lock(&tp->lock);
3353
3354         skip_hw = tc_skip_hw(f->flags);
3355
3356         if (f->flags && nla_put_u32(skb, TCA_FLOWER_FLAGS, f->flags))
3357                 goto nla_put_failure_locked;
3358
3359         spin_unlock(&tp->lock);
3360
3361         if (!skip_hw)
3362                 fl_hw_update_stats(tp, f, rtnl_held);
3363
3364         if (tcf_exts_terse_dump(skb, &f->exts))
3365                 goto nla_put_failure;
3366
3367         nla_nest_end(skb, nest);
3368
3369         return skb->len;
3370
3371 nla_put_failure_locked:
3372         spin_unlock(&tp->lock);
3373 nla_put_failure:
3374         nla_nest_cancel(skb, nest);
3375         return -1;
3376 }
3377
3378 static int fl_tmplt_dump(struct sk_buff *skb, struct net *net, void *tmplt_priv)
3379 {
3380         struct fl_flow_tmplt *tmplt = tmplt_priv;
3381         struct fl_flow_key *key, *mask;
3382         struct nlattr *nest;
3383
3384         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
3385         if (!nest)
3386                 goto nla_put_failure;
3387
3388         key = &tmplt->dummy_key;
3389         mask = &tmplt->mask;
3390
3391         if (fl_dump_key(skb, net, key, mask))
3392                 goto nla_put_failure;
3393
3394         nla_nest_end(skb, nest);
3395
3396         return skb->len;
3397
3398 nla_put_failure:
3399         nla_nest_cancel(skb, nest);
3400         return -EMSGSIZE;
3401 }
3402
3403 static void fl_bind_class(void *fh, u32 classid, unsigned long cl, void *q,
3404                           unsigned long base)
3405 {
3406         struct cls_fl_filter *f = fh;
3407
3408         tc_cls_bind_class(classid, cl, q, &f->res, base);
3409 }
3410
3411 static bool fl_delete_empty(struct tcf_proto *tp)
3412 {
3413         struct cls_fl_head *head = fl_head_dereference(tp);
3414
3415         spin_lock(&tp->lock);
3416         tp->deleting = idr_is_empty(&head->handle_idr);
3417         spin_unlock(&tp->lock);
3418
3419         return tp->deleting;
3420 }
3421
3422 static struct tcf_proto_ops cls_fl_ops __read_mostly = {
3423         .kind           = "flower",
3424         .classify       = fl_classify,
3425         .init           = fl_init,
3426         .destroy        = fl_destroy,
3427         .get            = fl_get,
3428         .put            = fl_put,
3429         .change         = fl_change,
3430         .delete         = fl_delete,
3431         .delete_empty   = fl_delete_empty,
3432         .walk           = fl_walk,
3433         .reoffload      = fl_reoffload,
3434         .hw_add         = fl_hw_add,
3435         .hw_del         = fl_hw_del,
3436         .dump           = fl_dump,
3437         .terse_dump     = fl_terse_dump,
3438         .bind_class     = fl_bind_class,
3439         .tmplt_create   = fl_tmplt_create,
3440         .tmplt_destroy  = fl_tmplt_destroy,
3441         .tmplt_dump     = fl_tmplt_dump,
3442         .owner          = THIS_MODULE,
3443         .flags          = TCF_PROTO_OPS_DOIT_UNLOCKED,
3444 };
3445
3446 static int __init cls_fl_init(void)
3447 {
3448         return register_tcf_proto_ops(&cls_fl_ops);
3449 }
3450
3451 static void __exit cls_fl_exit(void)
3452 {
3453         unregister_tcf_proto_ops(&cls_fl_ops);
3454 }
3455
3456 module_init(cls_fl_init);
3457 module_exit(cls_fl_exit);
3458
3459 MODULE_AUTHOR("Jiri Pirko <jiri@resnulli.us>");
3460 MODULE_DESCRIPTION("Flower classifier");
3461 MODULE_LICENSE("GPL v2");