net/sched/cls_flower.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * net/sched/cls_flower.c               Flower classifier
   4  *
   5  * Copyright (c) 2015 Jiri Pirko <jiri@resnulli.us>
   6  */
   7
   8 #include <linux/kernel.h>
   9 #include <linux/init.h>
  10 #include <linux/module.h>
  11 #include <linux/rhashtable.h>
  12 #include <linux/workqueue.h>
  13 #include <linux/refcount.h>
  14
  15 #include <linux/if_ether.h>
  16 #include <linux/in6.h>
  17 #include <linux/ip.h>
  18 #include <linux/mpls.h>
  19
  20 #include <net/sch_generic.h>
  21 #include <net/pkt_cls.h>
  22 #include <net/pkt_sched.h>
  23 #include <net/ip.h>
  24 #include <net/flow_dissector.h>
  25 #include <net/geneve.h>
  26 #include <net/vxlan.h>
  27 #include <net/erspan.h>
  28 #include <net/gtp.h>
  29
  30 #include <net/dst.h>
  31 #include <net/dst_metadata.h>
  32
  33 #include <uapi/linux/netfilter/nf_conntrack_common.h>
  34
  35 #define TCA_FLOWER_KEY_CT_FLAGS_MAX \
  36                 ((__TCA_FLOWER_KEY_CT_FLAGS_MAX - 1) << 1)
  37 #define TCA_FLOWER_KEY_CT_FLAGS_MASK \
  38                 (TCA_FLOWER_KEY_CT_FLAGS_MAX - 1)
  39
  40 struct fl_flow_key {
  41         struct flow_dissector_key_meta meta;
  42         struct flow_dissector_key_control control;
  43         struct flow_dissector_key_control enc_control;
  44         struct flow_dissector_key_basic basic;
  45         struct flow_dissector_key_eth_addrs eth;
  46         struct flow_dissector_key_vlan vlan;
  47         struct flow_dissector_key_vlan cvlan;
  48         union {
  49                 struct flow_dissector_key_ipv4_addrs ipv4;
  50                 struct flow_dissector_key_ipv6_addrs ipv6;
  51         };
  52         struct flow_dissector_key_ports tp;
  53         struct flow_dissector_key_icmp icmp;
  54         struct flow_dissector_key_arp arp;
  55         struct flow_dissector_key_keyid enc_key_id;
  56         union {
  57                 struct flow_dissector_key_ipv4_addrs enc_ipv4;
  58                 struct flow_dissector_key_ipv6_addrs enc_ipv6;
  59         };
  60         struct flow_dissector_key_ports enc_tp;
  61         struct flow_dissector_key_mpls mpls;
  62         struct flow_dissector_key_tcp tcp;
  63         struct flow_dissector_key_ip ip;
  64         struct flow_dissector_key_ip enc_ip;
  65         struct flow_dissector_key_enc_opts enc_opts;
  66         union {
  67                 struct flow_dissector_key_ports tp;
  68                 struct {
  69                         struct flow_dissector_key_ports tp_min;
  70                         struct flow_dissector_key_ports tp_max;
  71                 };
  72         } tp_range;
  73         struct flow_dissector_key_ct ct;
  74         struct flow_dissector_key_hash hash;
  75         struct flow_dissector_key_num_of_vlans num_of_vlans;
  76 } __aligned(BITS_PER_LONG / 8); /* Ensure that we can do comparisons as longs. */
  77
  78 struct fl_flow_mask_range {
  79         unsigned short int start;
  80         unsigned short int end;
  81 };
  82
  83 struct fl_flow_mask {
  84         struct fl_flow_key key;
  85         struct fl_flow_mask_range range;
  86         u32 flags;
  87         struct rhash_head ht_node;
  88         struct rhashtable ht;
  89         struct rhashtable_params filter_ht_params;
  90         struct flow_dissector dissector;
  91         struct list_head filters;
  92         struct rcu_work rwork;
  93         struct list_head list;
  94         refcount_t refcnt;
  95 };
  96
  97 struct fl_flow_tmplt {
  98         struct fl_flow_key dummy_key;
  99         struct fl_flow_key mask;
 100         struct flow_dissector dissector;
 101         struct tcf_chain *chain;
 102 };
 103
 104 struct cls_fl_head {
 105         struct rhashtable ht;
 106         spinlock_t masks_lock; /* Protect masks list */
 107         struct list_head masks;
 108         struct list_head hw_filters;
 109         struct rcu_work rwork;
 110         struct idr handle_idr;
 111 };
 112
 113 struct cls_fl_filter {
 114         struct fl_flow_mask *mask;
 115         struct rhash_head ht_node;
 116         struct fl_flow_key mkey;
 117         struct tcf_exts exts;
 118         struct tcf_result res;
 119         struct fl_flow_key key;
 120         struct list_head list;
 121         struct list_head hw_list;
 122         u32 handle;
 123         u32 flags;
 124         u32 in_hw_count;
 125         struct rcu_work rwork;
 126         struct net_device *hw_dev;
 127         /* Flower classifier is unlocked, which means that its reference counter
 128          * can be changed concurrently without any kind of external
 129          * synchronization. Use atomic reference counter to be concurrency-safe.
 130          */
 131         refcount_t refcnt;
 132         bool deleted;
 133 };
 134
 135 static const struct rhashtable_params mask_ht_params = {
 136         .key_offset = offsetof(struct fl_flow_mask, key),
 137         .key_len = sizeof(struct fl_flow_key),
 138         .head_offset = offsetof(struct fl_flow_mask, ht_node),
 139         .automatic_shrinking = true,
 140 };
 141
 142 static unsigned short int fl_mask_range(const struct fl_flow_mask *mask)
 143 {
 144         return mask->range.end - mask->range.start;
 145 }
 146
 147 static void fl_mask_update_range(struct fl_flow_mask *mask)
 148 {
 149         const u8 *bytes = (const u8 *) &mask->key;
 150         size_t size = sizeof(mask->key);
 151         size_t i, first = 0, last;
 152
 153         for (i = 0; i < size; i++) {
 154                 if (bytes[i]) {
 155                         first = i;
 156                         break;
 157                 }
 158         }
 159         last = first;
 160         for (i = size - 1; i != first; i--) {
 161                 if (bytes[i]) {
 162                         last = i;
 163                         break;
 164                 }
 165         }
 166         mask->range.start = rounddown(first, sizeof(long));
 167         mask->range.end = roundup(last + 1, sizeof(long));
 168 }
 169
 170 static void *fl_key_get_start(struct fl_flow_key *key,
 171                               const struct fl_flow_mask *mask)
 172 {
 173         return (u8 *) key + mask->range.start;
 174 }
 175
 176 static void fl_set_masked_key(struct fl_flow_key *mkey, struct fl_flow_key *key,
 177                               struct fl_flow_mask *mask)
 178 {
 179         const long *lkey = fl_key_get_start(key, mask);
 180         const long *lmask = fl_key_get_start(&mask->key, mask);
 181         long *lmkey = fl_key_get_start(mkey, mask);
 182         int i;
 183
 184         for (i = 0; i < fl_mask_range(mask); i += sizeof(long))
 185                 *lmkey++ = *lkey++ & *lmask++;
 186 }
 187
 188 static bool fl_mask_fits_tmplt(struct fl_flow_tmplt *tmplt,
 189                                struct fl_flow_mask *mask)
 190 {
 191         const long *lmask = fl_key_get_start(&mask->key, mask);
 192         const long *ltmplt;
 193         int i;
 194
 195         if (!tmplt)
 196                 return true;
 197         ltmplt = fl_key_get_start(&tmplt->mask, mask);
 198         for (i = 0; i < fl_mask_range(mask); i += sizeof(long)) {
 199                 if (~*ltmplt++ & *lmask++)
 200                         return false;
 201         }
 202         return true;
 203 }
 204
 205 static void fl_clear_masked_range(struct fl_flow_key *key,
 206                                   struct fl_flow_mask *mask)
 207 {
 208         memset(fl_key_get_start(key, mask), 0, fl_mask_range(mask));
 209 }
 210
 211 static bool fl_range_port_dst_cmp(struct cls_fl_filter *filter,
 212                                   struct fl_flow_key *key,
 213                                   struct fl_flow_key *mkey)
 214 {
 215         u16 min_mask, max_mask, min_val, max_val;
 216
 217         min_mask = ntohs(filter->mask->key.tp_range.tp_min.dst);
 218         max_mask = ntohs(filter->mask->key.tp_range.tp_max.dst);
 219         min_val = ntohs(filter->key.tp_range.tp_min.dst);
 220         max_val = ntohs(filter->key.tp_range.tp_max.dst);
 221
 222         if (min_mask && max_mask) {
 223                 if (ntohs(key->tp_range.tp.dst) < min_val ||
 224                     ntohs(key->tp_range.tp.dst) > max_val)
 225                         return false;
 226
 227                 /* skb does not have min and max values */
 228                 mkey->tp_range.tp_min.dst = filter->mkey.tp_range.tp_min.dst;
 229                 mkey->tp_range.tp_max.dst = filter->mkey.tp_range.tp_max.dst;
 230         }
 231         return true;
 232 }
 233
 234 static bool fl_range_port_src_cmp(struct cls_fl_filter *filter,
 235                                   struct fl_flow_key *key,
 236                                   struct fl_flow_key *mkey)
 237 {
 238         u16 min_mask, max_mask, min_val, max_val;
 239
 240         min_mask = ntohs(filter->mask->key.tp_range.tp_min.src);
 241         max_mask = ntohs(filter->mask->key.tp_range.tp_max.src);
 242         min_val = ntohs(filter->key.tp_range.tp_min.src);
 243         max_val = ntohs(filter->key.tp_range.tp_max.src);
 244
 245         if (min_mask && max_mask) {
 246                 if (ntohs(key->tp_range.tp.src) < min_val ||
 247                     ntohs(key->tp_range.tp.src) > max_val)
 248                         return false;
 249
 250                 /* skb does not have min and max values */
 251                 mkey->tp_range.tp_min.src = filter->mkey.tp_range.tp_min.src;
 252                 mkey->tp_range.tp_max.src = filter->mkey.tp_range.tp_max.src;
 253         }
 254         return true;
 255 }
 256
 257 static struct cls_fl_filter *__fl_lookup(struct fl_flow_mask *mask,
 258                                          struct fl_flow_key *mkey)
 259 {
 260         return rhashtable_lookup_fast(&mask->ht, fl_key_get_start(mkey, mask),
 261                                       mask->filter_ht_params);
 262 }
 263
 264 static struct cls_fl_filter *fl_lookup_range(struct fl_flow_mask *mask,
 265                                              struct fl_flow_key *mkey,
 266                                              struct fl_flow_key *key)
 267 {
 268         struct cls_fl_filter *filter, *f;
 269
 270         list_for_each_entry_rcu(filter, &mask->filters, list) {
 271                 if (!fl_range_port_dst_cmp(filter, key, mkey))
 272                         continue;
 273
 274                 if (!fl_range_port_src_cmp(filter, key, mkey))
 275                         continue;
 276
 277                 f = __fl_lookup(mask, mkey);
 278                 if (f)
 279                         return f;
 280         }
 281         return NULL;
 282 }
 283
 284 static noinline_for_stack
 285 struct cls_fl_filter *fl_mask_lookup(struct fl_flow_mask *mask, struct fl_flow_key *key)
 286 {
 287         struct fl_flow_key mkey;
 288
 289         fl_set_masked_key(&mkey, key, mask);
 290         if ((mask->flags & TCA_FLOWER_MASK_FLAGS_RANGE))
 291                 return fl_lookup_range(mask, &mkey, key);
 292
 293         return __fl_lookup(mask, &mkey);
 294 }
 295
 296 static u16 fl_ct_info_to_flower_map[] = {
 297         [IP_CT_ESTABLISHED] =           TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 298                                         TCA_FLOWER_KEY_CT_FLAGS_ESTABLISHED,
 299         [IP_CT_RELATED] =               TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 300                                         TCA_FLOWER_KEY_CT_FLAGS_RELATED,
 301         [IP_CT_ESTABLISHED_REPLY] =     TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 302                                         TCA_FLOWER_KEY_CT_FLAGS_ESTABLISHED |
 303                                         TCA_FLOWER_KEY_CT_FLAGS_REPLY,
 304         [IP_CT_RELATED_REPLY] =         TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 305                                         TCA_FLOWER_KEY_CT_FLAGS_RELATED |
 306                                         TCA_FLOWER_KEY_CT_FLAGS_REPLY,
 307         [IP_CT_NEW] =                   TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 308                                         TCA_FLOWER_KEY_CT_FLAGS_NEW,
 309 };
 310
 311 static int fl_classify(struct sk_buff *skb, const struct tcf_proto *tp,
 312                        struct tcf_result *res)
 313 {
 314         struct cls_fl_head *head = rcu_dereference_bh(tp->root);
 315         bool post_ct = tc_skb_cb(skb)->post_ct;
 316         u16 zone = tc_skb_cb(skb)->zone;
 317         struct fl_flow_key skb_key;
 318         struct fl_flow_mask *mask;
 319         struct cls_fl_filter *f;
 320
 321         list_for_each_entry_rcu(mask, &head->masks, list) {
 322                 flow_dissector_init_keys(&skb_key.control, &skb_key.basic);
 323                 fl_clear_masked_range(&skb_key, mask);
 324
 325                 skb_flow_dissect_meta(skb, &mask->dissector, &skb_key);
 326                 /* skb_flow_dissect() does not set n_proto in case an unknown
 327                  * protocol, so do it rather here.
 328                  */
 329                 skb_key.basic.n_proto = skb_protocol(skb, false);
 330                 skb_flow_dissect_tunnel_info(skb, &mask->dissector, &skb_key);
 331                 skb_flow_dissect_ct(skb, &mask->dissector, &skb_key,
 332                                     fl_ct_info_to_flower_map,
 333                                     ARRAY_SIZE(fl_ct_info_to_flower_map),
 334                                     post_ct, zone);
 335                 skb_flow_dissect_hash(skb, &mask->dissector, &skb_key);
 336                 skb_flow_dissect(skb, &mask->dissector, &skb_key,
 337                                  FLOW_DISSECTOR_F_STOP_BEFORE_ENCAP);
 338
 339                 f = fl_mask_lookup(mask, &skb_key);
 340                 if (f && !tc_skip_sw(f->flags)) {
 341                         *res = f->res;
 342                         return tcf_exts_exec(skb, &f->exts, res);
 343                 }
 344         }
 345         return -1;
 346 }
 347
 348 static int fl_init(struct tcf_proto *tp)
 349 {
 350         struct cls_fl_head *head;
 351
 352         head = kzalloc(sizeof(*head), GFP_KERNEL);
 353         if (!head)
 354                 return -ENOBUFS;
 355
 356         spin_lock_init(&head->masks_lock);
 357         INIT_LIST_HEAD_RCU(&head->masks);
 358         INIT_LIST_HEAD(&head->hw_filters);
 359         rcu_assign_pointer(tp->root, head);
 360         idr_init(&head->handle_idr);
 361
 362         return rhashtable_init(&head->ht, &mask_ht_params);
 363 }
 364
 365 static void fl_mask_free(struct fl_flow_mask *mask, bool mask_init_done)
 366 {
 367         /* temporary masks don't have their filters list and ht initialized */
 368         if (mask_init_done) {
 369                 WARN_ON(!list_empty(&mask->filters));
 370                 rhashtable_destroy(&mask->ht);
 371         }
 372         kfree(mask);
 373 }
 374
 375 static void fl_mask_free_work(struct work_struct *work)
 376 {
 377         struct fl_flow_mask *mask = container_of(to_rcu_work(work),
 378                                                  struct fl_flow_mask, rwork);
 379
 380         fl_mask_free(mask, true);
 381 }
 382
 383 static void fl_uninit_mask_free_work(struct work_struct *work)
 384 {
 385         struct fl_flow_mask *mask = container_of(to_rcu_work(work),
 386                                                  struct fl_flow_mask, rwork);
 387
 388         fl_mask_free(mask, false);
 389 }
 390
 391 static bool fl_mask_put(struct cls_fl_head *head, struct fl_flow_mask *mask)
 392 {
 393         if (!refcount_dec_and_test(&mask->refcnt))
 394                 return false;
 395
 396         rhashtable_remove_fast(&head->ht, &mask->ht_node, mask_ht_params);
 397
 398         spin_lock(&head->masks_lock);
 399         list_del_rcu(&mask->list);
 400         spin_unlock(&head->masks_lock);
 401
 402         tcf_queue_work(&mask->rwork, fl_mask_free_work);
 403
 404         return true;
 405 }
 406
 407 static struct cls_fl_head *fl_head_dereference(struct tcf_proto *tp)
 408 {
 409         /* Flower classifier only changes root pointer during init and destroy.
 410          * Users must obtain reference to tcf_proto instance before calling its
 411          * API, so tp->root pointer is protected from concurrent call to
 412          * fl_destroy() by reference counting.
 413          */
 414         return rcu_dereference_raw(tp->root);
 415 }
 416
 417 static void __fl_destroy_filter(struct cls_fl_filter *f)
 418 {
 419         tcf_exts_destroy(&f->exts);
 420         tcf_exts_put_net(&f->exts);
 421         kfree(f);
 422 }
 423
 424 static void fl_destroy_filter_work(struct work_struct *work)
 425 {
 426         struct cls_fl_filter *f = container_of(to_rcu_work(work),
 427                                         struct cls_fl_filter, rwork);
 428
 429         __fl_destroy_filter(f);
 430 }
 431
 432 static void fl_hw_destroy_filter(struct tcf_proto *tp, struct cls_fl_filter *f,
 433                                  bool rtnl_held, struct netlink_ext_ack *extack)
 434 {
 435         struct tcf_block *block = tp->chain->block;
 436         struct flow_cls_offload cls_flower = {};
 437
 438         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, extack);
 439         cls_flower.command = FLOW_CLS_DESTROY;
 440         cls_flower.cookie = (unsigned long) f;
 441
 442         tc_setup_cb_destroy(block, tp, TC_SETUP_CLSFLOWER, &cls_flower, false,
 443                             &f->flags, &f->in_hw_count, rtnl_held);
 444
 445 }
 446
 447 static int fl_hw_replace_filter(struct tcf_proto *tp,
 448                                 struct cls_fl_filter *f, bool rtnl_held,
 449                                 struct netlink_ext_ack *extack)
 450 {
 451         struct tcf_block *block = tp->chain->block;
 452         struct flow_cls_offload cls_flower = {};
 453         bool skip_sw = tc_skip_sw(f->flags);
 454         int err = 0;
 455
 456         cls_flower.rule = flow_rule_alloc(tcf_exts_num_actions(&f->exts));
 457         if (!cls_flower.rule)
 458                 return -ENOMEM;
 459
 460         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, extack);
 461         cls_flower.command = FLOW_CLS_REPLACE;
 462         cls_flower.cookie = (unsigned long) f;
 463         cls_flower.rule->match.dissector = &f->mask->dissector;
 464         cls_flower.rule->match.mask = &f->mask->key;
 465         cls_flower.rule->match.key = &f->mkey;
 466         cls_flower.classid = f->res.classid;
 467
 468         err = tc_setup_offload_action(&cls_flower.rule->action, &f->exts,
 469                                       cls_flower.common.extack);
 470         if (err) {
 471                 kfree(cls_flower.rule);
 472
 473                 return skip_sw ? err : 0;
 474         }
 475
 476         err = tc_setup_cb_add(block, tp, TC_SETUP_CLSFLOWER, &cls_flower,
 477                               skip_sw, &f->flags, &f->in_hw_count, rtnl_held);
 478         tc_cleanup_offload_action(&cls_flower.rule->action);
 479         kfree(cls_flower.rule);
 480
 481         if (err) {
 482                 fl_hw_destroy_filter(tp, f, rtnl_held, NULL);
 483                 return err;
 484         }
 485
 486         if (skip_sw && !(f->flags & TCA_CLS_FLAGS_IN_HW))
 487                 return -EINVAL;
 488
 489         return 0;
 490 }
 491
 492 static void fl_hw_update_stats(struct tcf_proto *tp, struct cls_fl_filter *f,
 493                                bool rtnl_held)
 494 {
 495         struct tcf_block *block = tp->chain->block;
 496         struct flow_cls_offload cls_flower = {};
 497
 498         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, NULL);
 499         cls_flower.command = FLOW_CLS_STATS;
 500         cls_flower.cookie = (unsigned long) f;
 501         cls_flower.classid = f->res.classid;
 502
 503         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false,
 504                          rtnl_held);
 505
 506         tcf_exts_hw_stats_update(&f->exts, cls_flower.stats.bytes,
 507                                  cls_flower.stats.pkts,
 508                                  cls_flower.stats.drops,
 509                                  cls_flower.stats.lastused,
 510                                  cls_flower.stats.used_hw_stats,
 511                                  cls_flower.stats.used_hw_stats_valid);
 512 }
 513
 514 static void __fl_put(struct cls_fl_filter *f)
 515 {
 516         if (!refcount_dec_and_test(&f->refcnt))
 517                 return;
 518
 519         if (tcf_exts_get_net(&f->exts))
 520                 tcf_queue_work(&f->rwork, fl_destroy_filter_work);
 521         else
 522                 __fl_destroy_filter(f);
 523 }
 524
 525 static struct cls_fl_filter *__fl_get(struct cls_fl_head *head, u32 handle)
 526 {
 527         struct cls_fl_filter *f;
 528
 529         rcu_read_lock();
 530         f = idr_find(&head->handle_idr, handle);
 531         if (f && !refcount_inc_not_zero(&f->refcnt))
 532                 f = NULL;
 533         rcu_read_unlock();
 534
 535         return f;
 536 }
 537
 538 static int __fl_delete(struct tcf_proto *tp, struct cls_fl_filter *f,
 539                        bool *last, bool rtnl_held,
 540                        struct netlink_ext_ack *extack)
 541 {
 542         struct cls_fl_head *head = fl_head_dereference(tp);
 543
 544         *last = false;
 545
 546         spin_lock(&tp->lock);
 547         if (f->deleted) {
 548                 spin_unlock(&tp->lock);
 549                 return -ENOENT;
 550         }
 551
 552         f->deleted = true;
 553         rhashtable_remove_fast(&f->mask->ht, &f->ht_node,
 554                                f->mask->filter_ht_params);
 555         idr_remove(&head->handle_idr, f->handle);
 556         list_del_rcu(&f->list);
 557         spin_unlock(&tp->lock);
 558
 559         *last = fl_mask_put(head, f->mask);
 560         if (!tc_skip_hw(f->flags))
 561                 fl_hw_destroy_filter(tp, f, rtnl_held, extack);
 562         tcf_unbind_filter(tp, &f->res);
 563         __fl_put(f);
 564
 565         return 0;
 566 }
 567
 568 static void fl_destroy_sleepable(struct work_struct *work)
 569 {
 570         struct cls_fl_head *head = container_of(to_rcu_work(work),
 571                                                 struct cls_fl_head,
 572                                                 rwork);
 573
 574         rhashtable_destroy(&head->ht);
 575         kfree(head);
 576         module_put(THIS_MODULE);
 577 }
 578
 579 static void fl_destroy(struct tcf_proto *tp, bool rtnl_held,
 580                        struct netlink_ext_ack *extack)
 581 {
 582         struct cls_fl_head *head = fl_head_dereference(tp);
 583         struct fl_flow_mask *mask, *next_mask;
 584         struct cls_fl_filter *f, *next;
 585         bool last;
 586
 587         list_for_each_entry_safe(mask, next_mask, &head->masks, list) {
 588                 list_for_each_entry_safe(f, next, &mask->filters, list) {
 589                         __fl_delete(tp, f, &last, rtnl_held, extack);
 590                         if (last)
 591                                 break;
 592                 }
 593         }
 594         idr_destroy(&head->handle_idr);
 595
 596         __module_get(THIS_MODULE);
 597         tcf_queue_work(&head->rwork, fl_destroy_sleepable);
 598 }
 599
 600 static void fl_put(struct tcf_proto *tp, void *arg)
 601 {
 602         struct cls_fl_filter *f = arg;
 603
 604         __fl_put(f);
 605 }
 606
 607 static void *fl_get(struct tcf_proto *tp, u32 handle)
 608 {
 609         struct cls_fl_head *head = fl_head_dereference(tp);
 610
 611         return __fl_get(head, handle);
 612 }
 613
 614 static const struct nla_policy fl_policy[TCA_FLOWER_MAX + 1] = {
 615         [TCA_FLOWER_UNSPEC]             = { .type = NLA_UNSPEC },
 616         [TCA_FLOWER_CLASSID]            = { .type = NLA_U32 },
 617         [TCA_FLOWER_INDEV]              = { .type = NLA_STRING,
 618                                             .len = IFNAMSIZ },
 619         [TCA_FLOWER_KEY_ETH_DST]        = { .len = ETH_ALEN },
 620         [TCA_FLOWER_KEY_ETH_DST_MASK]   = { .len = ETH_ALEN },
 621         [TCA_FLOWER_KEY_ETH_SRC]        = { .len = ETH_ALEN },
 622         [TCA_FLOWER_KEY_ETH_SRC_MASK]   = { .len = ETH_ALEN },
 623         [TCA_FLOWER_KEY_ETH_TYPE]       = { .type = NLA_U16 },
 624         [TCA_FLOWER_KEY_IP_PROTO]       = { .type = NLA_U8 },
 625         [TCA_FLOWER_KEY_IPV4_SRC]       = { .type = NLA_U32 },
 626         [TCA_FLOWER_KEY_IPV4_SRC_MASK]  = { .type = NLA_U32 },
 627         [TCA_FLOWER_KEY_IPV4_DST]       = { .type = NLA_U32 },
 628         [TCA_FLOWER_KEY_IPV4_DST_MASK]  = { .type = NLA_U32 },
 629         [TCA_FLOWER_KEY_IPV6_SRC]       = { .len = sizeof(struct in6_addr) },
 630         [TCA_FLOWER_KEY_IPV6_SRC_MASK]  = { .len = sizeof(struct in6_addr) },
 631         [TCA_FLOWER_KEY_IPV6_DST]       = { .len = sizeof(struct in6_addr) },
 632         [TCA_FLOWER_KEY_IPV6_DST_MASK]  = { .len = sizeof(struct in6_addr) },
 633         [TCA_FLOWER_KEY_TCP_SRC]        = { .type = NLA_U16 },
 634         [TCA_FLOWER_KEY_TCP_DST]        = { .type = NLA_U16 },
 635         [TCA_FLOWER_KEY_UDP_SRC]        = { .type = NLA_U16 },
 636         [TCA_FLOWER_KEY_UDP_DST]        = { .type = NLA_U16 },
 637         [TCA_FLOWER_KEY_VLAN_ID]        = { .type = NLA_U16 },
 638         [TCA_FLOWER_KEY_VLAN_PRIO]      = { .type = NLA_U8 },
 639         [TCA_FLOWER_KEY_VLAN_ETH_TYPE]  = { .type = NLA_U16 },
 640         [TCA_FLOWER_KEY_ENC_KEY_ID]     = { .type = NLA_U32 },
 641         [TCA_FLOWER_KEY_ENC_IPV4_SRC]   = { .type = NLA_U32 },
 642         [TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK] = { .type = NLA_U32 },
 643         [TCA_FLOWER_KEY_ENC_IPV4_DST]   = { .type = NLA_U32 },
 644         [TCA_FLOWER_KEY_ENC_IPV4_DST_MASK] = { .type = NLA_U32 },
 645         [TCA_FLOWER_KEY_ENC_IPV6_SRC]   = { .len = sizeof(struct in6_addr) },
 646         [TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK] = { .len = sizeof(struct in6_addr) },
 647         [TCA_FLOWER_KEY_ENC_IPV6_DST]   = { .len = sizeof(struct in6_addr) },
 648         [TCA_FLOWER_KEY_ENC_IPV6_DST_MASK] = { .len = sizeof(struct in6_addr) },
 649         [TCA_FLOWER_KEY_TCP_SRC_MASK]   = { .type = NLA_U16 },
 650         [TCA_FLOWER_KEY_TCP_DST_MASK]   = { .type = NLA_U16 },
 651         [TCA_FLOWER_KEY_UDP_SRC_MASK]   = { .type = NLA_U16 },
 652         [TCA_FLOWER_KEY_UDP_DST_MASK]   = { .type = NLA_U16 },
 653         [TCA_FLOWER_KEY_SCTP_SRC_MASK]  = { .type = NLA_U16 },
 654         [TCA_FLOWER_KEY_SCTP_DST_MASK]  = { .type = NLA_U16 },
 655         [TCA_FLOWER_KEY_SCTP_SRC]       = { .type = NLA_U16 },
 656         [TCA_FLOWER_KEY_SCTP_DST]       = { .type = NLA_U16 },
 657         [TCA_FLOWER_KEY_ENC_UDP_SRC_PORT]       = { .type = NLA_U16 },
 658         [TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK]  = { .type = NLA_U16 },
 659         [TCA_FLOWER_KEY_ENC_UDP_DST_PORT]       = { .type = NLA_U16 },
 660         [TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK]  = { .type = NLA_U16 },
 661         [TCA_FLOWER_KEY_FLAGS]          = { .type = NLA_U32 },
 662         [TCA_FLOWER_KEY_FLAGS_MASK]     = { .type = NLA_U32 },
 663         [TCA_FLOWER_KEY_ICMPV4_TYPE]    = { .type = NLA_U8 },
 664         [TCA_FLOWER_KEY_ICMPV4_TYPE_MASK] = { .type = NLA_U8 },
 665         [TCA_FLOWER_KEY_ICMPV4_CODE]    = { .type = NLA_U8 },
 666         [TCA_FLOWER_KEY_ICMPV4_CODE_MASK] = { .type = NLA_U8 },
 667         [TCA_FLOWER_KEY_ICMPV6_TYPE]    = { .type = NLA_U8 },
 668         [TCA_FLOWER_KEY_ICMPV6_TYPE_MASK] = { .type = NLA_U8 },
 669         [TCA_FLOWER_KEY_ICMPV6_CODE]    = { .type = NLA_U8 },
 670         [TCA_FLOWER_KEY_ICMPV6_CODE_MASK] = { .type = NLA_U8 },
 671         [TCA_FLOWER_KEY_ARP_SIP]        = { .type = NLA_U32 },
 672         [TCA_FLOWER_KEY_ARP_SIP_MASK]   = { .type = NLA_U32 },
 673         [TCA_FLOWER_KEY_ARP_TIP]        = { .type = NLA_U32 },
 674         [TCA_FLOWER_KEY_ARP_TIP_MASK]   = { .type = NLA_U32 },
 675         [TCA_FLOWER_KEY_ARP_OP]         = { .type = NLA_U8 },
 676         [TCA_FLOWER_KEY_ARP_OP_MASK]    = { .type = NLA_U8 },
 677         [TCA_FLOWER_KEY_ARP_SHA]        = { .len = ETH_ALEN },
 678         [TCA_FLOWER_KEY_ARP_SHA_MASK]   = { .len = ETH_ALEN },
 679         [TCA_FLOWER_KEY_ARP_THA]        = { .len = ETH_ALEN },
 680         [TCA_FLOWER_KEY_ARP_THA_MASK]   = { .len = ETH_ALEN },
 681         [TCA_FLOWER_KEY_MPLS_TTL]       = { .type = NLA_U8 },
 682         [TCA_FLOWER_KEY_MPLS_BOS]       = { .type = NLA_U8 },
 683         [TCA_FLOWER_KEY_MPLS_TC]        = { .type = NLA_U8 },
 684         [TCA_FLOWER_KEY_MPLS_LABEL]     = { .type = NLA_U32 },
 685         [TCA_FLOWER_KEY_MPLS_OPTS]      = { .type = NLA_NESTED },
 686         [TCA_FLOWER_KEY_TCP_FLAGS]      = { .type = NLA_U16 },
 687         [TCA_FLOWER_KEY_TCP_FLAGS_MASK] = { .type = NLA_U16 },
 688         [TCA_FLOWER_KEY_IP_TOS]         = { .type = NLA_U8 },
 689         [TCA_FLOWER_KEY_IP_TOS_MASK]    = { .type = NLA_U8 },
 690         [TCA_FLOWER_KEY_IP_TTL]         = { .type = NLA_U8 },
 691         [TCA_FLOWER_KEY_IP_TTL_MASK]    = { .type = NLA_U8 },
 692         [TCA_FLOWER_KEY_CVLAN_ID]       = { .type = NLA_U16 },
 693         [TCA_FLOWER_KEY_CVLAN_PRIO]     = { .type = NLA_U8 },
 694         [TCA_FLOWER_KEY_CVLAN_ETH_TYPE] = { .type = NLA_U16 },
 695         [TCA_FLOWER_KEY_ENC_IP_TOS]     = { .type = NLA_U8 },
 696         [TCA_FLOWER_KEY_ENC_IP_TOS_MASK] = { .type = NLA_U8 },
 697         [TCA_FLOWER_KEY_ENC_IP_TTL]      = { .type = NLA_U8 },
 698         [TCA_FLOWER_KEY_ENC_IP_TTL_MASK] = { .type = NLA_U8 },
 699         [TCA_FLOWER_KEY_ENC_OPTS]       = { .type = NLA_NESTED },
 700         [TCA_FLOWER_KEY_ENC_OPTS_MASK]  = { .type = NLA_NESTED },
 701         [TCA_FLOWER_KEY_CT_STATE]       =
 702                 NLA_POLICY_MASK(NLA_U16, TCA_FLOWER_KEY_CT_FLAGS_MASK),
 703         [TCA_FLOWER_KEY_CT_STATE_MASK]  =
 704                 NLA_POLICY_MASK(NLA_U16, TCA_FLOWER_KEY_CT_FLAGS_MASK),
 705         [TCA_FLOWER_KEY_CT_ZONE]        = { .type = NLA_U16 },
 706         [TCA_FLOWER_KEY_CT_ZONE_MASK]   = { .type = NLA_U16 },
 707         [TCA_FLOWER_KEY_CT_MARK]        = { .type = NLA_U32 },
 708         [TCA_FLOWER_KEY_CT_MARK_MASK]   = { .type = NLA_U32 },
 709         [TCA_FLOWER_KEY_CT_LABELS]      = { .type = NLA_BINARY,
 710                                             .len = 128 / BITS_PER_BYTE },
 711         [TCA_FLOWER_KEY_CT_LABELS_MASK] = { .type = NLA_BINARY,
 712                                             .len = 128 / BITS_PER_BYTE },
 713         [TCA_FLOWER_FLAGS]              = { .type = NLA_U32 },
 714         [TCA_FLOWER_KEY_HASH]           = { .type = NLA_U32 },
 715         [TCA_FLOWER_KEY_HASH_MASK]      = { .type = NLA_U32 },
 716         [TCA_FLOWER_KEY_NUM_OF_VLANS]   = { .type = NLA_U8 },
 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_flag(u32 flower_key, u32 flower_mask,
1045                             u32 *dissector_key, u32 *dissector_mask,
1046                             u32 flower_flag_bit, u32 dissector_flag_bit)
1047 {
1048         if (flower_mask & flower_flag_bit) {
1049                 *dissector_mask |= dissector_flag_bit;
1050                 if (flower_key & flower_flag_bit)
1051                         *dissector_key |= dissector_flag_bit;
1052         }
1053 }
1054
1055 static int fl_set_key_flags(struct nlattr **tb, u32 *flags_key,
1056                             u32 *flags_mask, struct netlink_ext_ack *extack)
1057 {
1058         u32 key, mask;
1059
1060         /* mask is mandatory for flags */
1061         if (!tb[TCA_FLOWER_KEY_FLAGS_MASK]) {
1062                 NL_SET_ERR_MSG(extack, "Missing flags mask");
1063                 return -EINVAL;
1064         }
1065
1066         key = be32_to_cpu(nla_get_be32(tb[TCA_FLOWER_KEY_FLAGS]));
1067         mask = be32_to_cpu(nla_get_be32(tb[TCA_FLOWER_KEY_FLAGS_MASK]));
1068
1069         *flags_key  = 0;
1070         *flags_mask = 0;
1071
1072         fl_set_key_flag(key, mask, flags_key, flags_mask,
1073                         TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT, FLOW_DIS_IS_FRAGMENT);
1074         fl_set_key_flag(key, mask, flags_key, flags_mask,
1075                         TCA_FLOWER_KEY_FLAGS_FRAG_IS_FIRST,
1076                         FLOW_DIS_FIRST_FRAG);
1077
1078         return 0;
1079 }
1080
1081 static void fl_set_key_ip(struct nlattr **tb, bool encap,
1082                           struct flow_dissector_key_ip *key,
1083                           struct flow_dissector_key_ip *mask)
1084 {
1085         int tos_key = encap ? TCA_FLOWER_KEY_ENC_IP_TOS : TCA_FLOWER_KEY_IP_TOS;
1086         int ttl_key = encap ? TCA_FLOWER_KEY_ENC_IP_TTL : TCA_FLOWER_KEY_IP_TTL;
1087         int tos_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TOS_MASK : TCA_FLOWER_KEY_IP_TOS_MASK;
1088         int ttl_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TTL_MASK : TCA_FLOWER_KEY_IP_TTL_MASK;
1089
1090         fl_set_key_val(tb, &key->tos, tos_key, &mask->tos, tos_mask, sizeof(key->tos));
1091         fl_set_key_val(tb, &key->ttl, ttl_key, &mask->ttl, ttl_mask, sizeof(key->ttl));
1092 }
1093
1094 static int fl_set_geneve_opt(const struct nlattr *nla, struct fl_flow_key *key,
1095                              int depth, int option_len,
1096                              struct netlink_ext_ack *extack)
1097 {
1098         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX + 1];
1099         struct nlattr *class = NULL, *type = NULL, *data = NULL;
1100         struct geneve_opt *opt;
1101         int err, data_len = 0;
1102
1103         if (option_len > sizeof(struct geneve_opt))
1104                 data_len = option_len - sizeof(struct geneve_opt);
1105
1106         opt = (struct geneve_opt *)&key->enc_opts.data[key->enc_opts.len];
1107         memset(opt, 0xff, option_len);
1108         opt->length = data_len / 4;
1109         opt->r1 = 0;
1110         opt->r2 = 0;
1111         opt->r3 = 0;
1112
1113         /* If no mask has been prodived we assume an exact match. */
1114         if (!depth)
1115                 return sizeof(struct geneve_opt) + data_len;
1116
1117         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_GENEVE) {
1118                 NL_SET_ERR_MSG(extack, "Non-geneve option type for mask");
1119                 return -EINVAL;
1120         }
1121
1122         err = nla_parse_nested_deprecated(tb,
1123                                           TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX,
1124                                           nla, geneve_opt_policy, extack);
1125         if (err < 0)
1126                 return err;
1127
1128         /* We are not allowed to omit any of CLASS, TYPE or DATA
1129          * fields from the key.
1130          */
1131         if (!option_len &&
1132             (!tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS] ||
1133              !tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE] ||
1134              !tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA])) {
1135                 NL_SET_ERR_MSG(extack, "Missing tunnel key geneve option class, type or data");
1136                 return -EINVAL;
1137         }
1138
1139         /* Omitting any of CLASS, TYPE or DATA fields is allowed
1140          * for the mask.
1141          */
1142         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA]) {
1143                 int new_len = key->enc_opts.len;
1144
1145                 data = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA];
1146                 data_len = nla_len(data);
1147                 if (data_len < 4) {
1148                         NL_SET_ERR_MSG(extack, "Tunnel key geneve option data is less than 4 bytes long");
1149                         return -ERANGE;
1150                 }
1151                 if (data_len % 4) {
1152                         NL_SET_ERR_MSG(extack, "Tunnel key geneve option data is not a multiple of 4 bytes long");
1153                         return -ERANGE;
1154                 }
1155
1156                 new_len += sizeof(struct geneve_opt) + data_len;
1157                 BUILD_BUG_ON(FLOW_DIS_TUN_OPTS_MAX != IP_TUNNEL_OPTS_MAX);
1158                 if (new_len > FLOW_DIS_TUN_OPTS_MAX) {
1159                         NL_SET_ERR_MSG(extack, "Tunnel options exceeds max size");
1160                         return -ERANGE;
1161                 }
1162                 opt->length = data_len / 4;
1163                 memcpy(opt->opt_data, nla_data(data), data_len);
1164         }
1165
1166         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS]) {
1167                 class = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS];
1168                 opt->opt_class = nla_get_be16(class);
1169         }
1170
1171         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE]) {
1172                 type = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE];
1173                 opt->type = nla_get_u8(type);
1174         }
1175
1176         return sizeof(struct geneve_opt) + data_len;
1177 }
1178
1179 static int fl_set_vxlan_opt(const struct nlattr *nla, struct fl_flow_key *key,
1180                             int depth, int option_len,
1181                             struct netlink_ext_ack *extack)
1182 {
1183         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_MAX + 1];
1184         struct vxlan_metadata *md;
1185         int err;
1186
1187         md = (struct vxlan_metadata *)&key->enc_opts.data[key->enc_opts.len];
1188         memset(md, 0xff, sizeof(*md));
1189
1190         if (!depth)
1191                 return sizeof(*md);
1192
1193         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_VXLAN) {
1194                 NL_SET_ERR_MSG(extack, "Non-vxlan option type for mask");
1195                 return -EINVAL;
1196         }
1197
1198         err = nla_parse_nested(tb, TCA_FLOWER_KEY_ENC_OPT_VXLAN_MAX, nla,
1199                                vxlan_opt_policy, extack);
1200         if (err < 0)
1201                 return err;
1202
1203         if (!option_len && !tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]) {
1204                 NL_SET_ERR_MSG(extack, "Missing tunnel key vxlan option gbp");
1205                 return -EINVAL;
1206         }
1207
1208         if (tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]) {
1209                 md->gbp = nla_get_u32(tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]);
1210                 md->gbp &= VXLAN_GBP_MASK;
1211         }
1212
1213         return sizeof(*md);
1214 }
1215
1216 static int fl_set_erspan_opt(const struct nlattr *nla, struct fl_flow_key *key,
1217                              int depth, int option_len,
1218                              struct netlink_ext_ack *extack)
1219 {
1220         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_MAX + 1];
1221         struct erspan_metadata *md;
1222         int err;
1223
1224         md = (struct erspan_metadata *)&key->enc_opts.data[key->enc_opts.len];
1225         memset(md, 0xff, sizeof(*md));
1226         md->version = 1;
1227
1228         if (!depth)
1229                 return sizeof(*md);
1230
1231         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_ERSPAN) {
1232                 NL_SET_ERR_MSG(extack, "Non-erspan option type for mask");
1233                 return -EINVAL;
1234         }
1235
1236         err = nla_parse_nested(tb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_MAX, nla,
1237                                erspan_opt_policy, extack);
1238         if (err < 0)
1239                 return err;
1240
1241         if (!option_len && !tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER]) {
1242                 NL_SET_ERR_MSG(extack, "Missing tunnel key erspan option ver");
1243                 return -EINVAL;
1244         }
1245
1246         if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER])
1247                 md->version = nla_get_u8(tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER]);
1248
1249         if (md->version == 1) {
1250                 if (!option_len && !tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX]) {
1251                         NL_SET_ERR_MSG(extack, "Missing tunnel key erspan option index");
1252                         return -EINVAL;
1253                 }
1254                 if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX]) {
1255                         nla = tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX];
1256                         memset(&md->u, 0x00, sizeof(md->u));
1257                         md->u.index = nla_get_be32(nla);
1258                 }
1259         } else if (md->version == 2) {
1260                 if (!option_len && (!tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR] ||
1261                                     !tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID])) {
1262                         NL_SET_ERR_MSG(extack, "Missing tunnel key erspan option dir or hwid");
1263                         return -EINVAL;
1264                 }
1265                 if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR]) {
1266                         nla = tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR];
1267                         md->u.md2.dir = nla_get_u8(nla);
1268                 }
1269                 if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID]) {
1270                         nla = tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID];
1271                         set_hwid(&md->u.md2, nla_get_u8(nla));
1272                 }
1273         } else {
1274                 NL_SET_ERR_MSG(extack, "Tunnel key erspan option ver is incorrect");
1275                 return -EINVAL;
1276         }
1277
1278         return sizeof(*md);
1279 }
1280
1281 static int fl_set_gtp_opt(const struct nlattr *nla, struct fl_flow_key *key,
1282                           int depth, int option_len,
1283                           struct netlink_ext_ack *extack)
1284 {
1285         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_GTP_MAX + 1];
1286         struct gtp_pdu_session_info *sinfo;
1287         u8 len = key->enc_opts.len;
1288         int err;
1289
1290         sinfo = (struct gtp_pdu_session_info *)&key->enc_opts.data[len];
1291         memset(sinfo, 0xff, option_len);
1292
1293         if (!depth)
1294                 return sizeof(*sinfo);
1295
1296         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_GTP) {
1297                 NL_SET_ERR_MSG_MOD(extack, "Non-gtp option type for mask");
1298                 return -EINVAL;
1299         }
1300
1301         err = nla_parse_nested(tb, TCA_FLOWER_KEY_ENC_OPT_GTP_MAX, nla,
1302                                gtp_opt_policy, extack);
1303         if (err < 0)
1304                 return err;
1305
1306         if (!option_len &&
1307             (!tb[TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE] ||
1308              !tb[TCA_FLOWER_KEY_ENC_OPT_GTP_QFI])) {
1309                 NL_SET_ERR_MSG_MOD(extack,
1310                                    "Missing tunnel key gtp option pdu type or qfi");
1311                 return -EINVAL;
1312         }
1313
1314         if (tb[TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE])
1315                 sinfo->pdu_type =
1316                         nla_get_u8(tb[TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE]);
1317
1318         if (tb[TCA_FLOWER_KEY_ENC_OPT_GTP_QFI])
1319                 sinfo->qfi = nla_get_u8(tb[TCA_FLOWER_KEY_ENC_OPT_GTP_QFI]);
1320
1321         return sizeof(*sinfo);
1322 }
1323
1324 static int fl_set_enc_opt(struct nlattr **tb, struct fl_flow_key *key,
1325                           struct fl_flow_key *mask,
1326                           struct netlink_ext_ack *extack)
1327 {
1328         const struct nlattr *nla_enc_key, *nla_opt_key, *nla_opt_msk = NULL;
1329         int err, option_len, key_depth, msk_depth = 0;
1330
1331         err = nla_validate_nested_deprecated(tb[TCA_FLOWER_KEY_ENC_OPTS],
1332                                              TCA_FLOWER_KEY_ENC_OPTS_MAX,
1333                                              enc_opts_policy, extack);
1334         if (err)
1335                 return err;
1336
1337         nla_enc_key = nla_data(tb[TCA_FLOWER_KEY_ENC_OPTS]);
1338
1339         if (tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]) {
1340                 err = nla_validate_nested_deprecated(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK],
1341                                                      TCA_FLOWER_KEY_ENC_OPTS_MAX,
1342                                                      enc_opts_policy, extack);
1343                 if (err)
1344                         return err;
1345
1346                 nla_opt_msk = nla_data(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]);
1347                 msk_depth = nla_len(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]);
1348                 if (!nla_ok(nla_opt_msk, msk_depth)) {
1349                         NL_SET_ERR_MSG(extack, "Invalid nested attribute for masks");
1350                         return -EINVAL;
1351                 }
1352         }
1353
1354         nla_for_each_attr(nla_opt_key, nla_enc_key,
1355                           nla_len(tb[TCA_FLOWER_KEY_ENC_OPTS]), key_depth) {
1356                 switch (nla_type(nla_opt_key)) {
1357                 case TCA_FLOWER_KEY_ENC_OPTS_GENEVE:
1358                         if (key->enc_opts.dst_opt_type &&
1359                             key->enc_opts.dst_opt_type != TUNNEL_GENEVE_OPT) {
1360                                 NL_SET_ERR_MSG(extack, "Duplicate type for geneve options");
1361                                 return -EINVAL;
1362                         }
1363                         option_len = 0;
1364                         key->enc_opts.dst_opt_type = TUNNEL_GENEVE_OPT;
1365                         option_len = fl_set_geneve_opt(nla_opt_key, key,
1366                                                        key_depth, option_len,
1367                                                        extack);
1368                         if (option_len < 0)
1369                                 return option_len;
1370
1371                         key->enc_opts.len += option_len;
1372                         /* At the same time we need to parse through the mask
1373                          * in order to verify exact and mask attribute lengths.
1374                          */
1375                         mask->enc_opts.dst_opt_type = TUNNEL_GENEVE_OPT;
1376                         option_len = fl_set_geneve_opt(nla_opt_msk, mask,
1377                                                        msk_depth, option_len,
1378                                                        extack);
1379                         if (option_len < 0)
1380                                 return option_len;
1381
1382                         mask->enc_opts.len += option_len;
1383                         if (key->enc_opts.len != mask->enc_opts.len) {
1384                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
1385                                 return -EINVAL;
1386                         }
1387                         break;
1388                 case TCA_FLOWER_KEY_ENC_OPTS_VXLAN:
1389                         if (key->enc_opts.dst_opt_type) {
1390                                 NL_SET_ERR_MSG(extack, "Duplicate type for vxlan options");
1391                                 return -EINVAL;
1392                         }
1393                         option_len = 0;
1394                         key->enc_opts.dst_opt_type = TUNNEL_VXLAN_OPT;
1395                         option_len = fl_set_vxlan_opt(nla_opt_key, key,
1396                                                       key_depth, option_len,
1397                                                       extack);
1398                         if (option_len < 0)
1399                                 return option_len;
1400
1401                         key->enc_opts.len += option_len;
1402                         /* At the same time we need to parse through the mask
1403                          * in order to verify exact and mask attribute lengths.
1404                          */
1405                         mask->enc_opts.dst_opt_type = TUNNEL_VXLAN_OPT;
1406                         option_len = fl_set_vxlan_opt(nla_opt_msk, mask,
1407                                                       msk_depth, option_len,
1408                                                       extack);
1409                         if (option_len < 0)
1410                                 return option_len;
1411
1412                         mask->enc_opts.len += option_len;
1413                         if (key->enc_opts.len != mask->enc_opts.len) {
1414                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
1415                                 return -EINVAL;
1416                         }
1417                         break;
1418                 case TCA_FLOWER_KEY_ENC_OPTS_ERSPAN:
1419                         if (key->enc_opts.dst_opt_type) {
1420                                 NL_SET_ERR_MSG(extack, "Duplicate type for erspan options");
1421                                 return -EINVAL;
1422                         }
1423                         option_len = 0;
1424                         key->enc_opts.dst_opt_type = TUNNEL_ERSPAN_OPT;
1425                         option_len = fl_set_erspan_opt(nla_opt_key, key,
1426                                                        key_depth, option_len,
1427                                                        extack);
1428                         if (option_len < 0)
1429                                 return option_len;
1430
1431                         key->enc_opts.len += option_len;
1432                         /* At the same time we need to parse through the mask
1433                          * in order to verify exact and mask attribute lengths.
1434                          */
1435                         mask->enc_opts.dst_opt_type = TUNNEL_ERSPAN_OPT;
1436                         option_len = fl_set_erspan_opt(nla_opt_msk, mask,
1437                                                        msk_depth, option_len,
1438                                                        extack);
1439                         if (option_len < 0)
1440                                 return option_len;
1441
1442                         mask->enc_opts.len += option_len;
1443                         if (key->enc_opts.len != mask->enc_opts.len) {
1444                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
1445                                 return -EINVAL;
1446                         }
1447                         break;
1448                 case TCA_FLOWER_KEY_ENC_OPTS_GTP:
1449                         if (key->enc_opts.dst_opt_type) {
1450                                 NL_SET_ERR_MSG_MOD(extack,
1451                                                    "Duplicate type for gtp options");
1452                                 return -EINVAL;
1453                         }
1454                         option_len = 0;
1455                         key->enc_opts.dst_opt_type = TUNNEL_GTP_OPT;
1456                         option_len = fl_set_gtp_opt(nla_opt_key, key,
1457                                                     key_depth, option_len,
1458                                                     extack);
1459                         if (option_len < 0)
1460                                 return option_len;
1461
1462                         key->enc_opts.len += option_len;
1463                         /* At the same time we need to parse through the mask
1464                          * in order to verify exact and mask attribute lengths.
1465                          */
1466                         mask->enc_opts.dst_opt_type = TUNNEL_GTP_OPT;
1467                         option_len = fl_set_gtp_opt(nla_opt_msk, mask,
1468                                                     msk_depth, option_len,
1469                                                     extack);
1470                         if (option_len < 0)
1471                                 return option_len;
1472
1473                         mask->enc_opts.len += option_len;
1474                         if (key->enc_opts.len != mask->enc_opts.len) {
1475                                 NL_SET_ERR_MSG_MOD(extack,
1476                                                    "Key and mask miss aligned");
1477                                 return -EINVAL;
1478                         }
1479                         break;
1480                 default:
1481                         NL_SET_ERR_MSG(extack, "Unknown tunnel option type");
1482                         return -EINVAL;
1483                 }
1484
1485                 if (!msk_depth)
1486                         continue;
1487
1488                 if (!nla_ok(nla_opt_msk, msk_depth)) {
1489                         NL_SET_ERR_MSG(extack, "A mask attribute is invalid");
1490                         return -EINVAL;
1491                 }
1492                 nla_opt_msk = nla_next(nla_opt_msk, &msk_depth);
1493         }
1494
1495         return 0;
1496 }
1497
1498 static int fl_validate_ct_state(u16 state, struct nlattr *tb,
1499                                 struct netlink_ext_ack *extack)
1500 {
1501         if (state && !(state & TCA_FLOWER_KEY_CT_FLAGS_TRACKED)) {
1502                 NL_SET_ERR_MSG_ATTR(extack, tb,
1503                                     "no trk, so no other flag can be set");
1504                 return -EINVAL;
1505         }
1506
1507         if (state & TCA_FLOWER_KEY_CT_FLAGS_NEW &&
1508             state & TCA_FLOWER_KEY_CT_FLAGS_ESTABLISHED) {
1509                 NL_SET_ERR_MSG_ATTR(extack, tb,
1510                                     "new and est are mutually exclusive");
1511                 return -EINVAL;
1512         }
1513
1514         if (state & TCA_FLOWER_KEY_CT_FLAGS_INVALID &&
1515             state & ~(TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
1516                       TCA_FLOWER_KEY_CT_FLAGS_INVALID)) {
1517                 NL_SET_ERR_MSG_ATTR(extack, tb,
1518                                     "when inv is set, only trk may be set");
1519                 return -EINVAL;
1520         }
1521
1522         if (state & TCA_FLOWER_KEY_CT_FLAGS_NEW &&
1523             state & TCA_FLOWER_KEY_CT_FLAGS_REPLY) {
1524                 NL_SET_ERR_MSG_ATTR(extack, tb,
1525                                     "new and rpl are mutually exclusive");
1526                 return -EINVAL;
1527         }
1528
1529         return 0;
1530 }
1531
1532 static int fl_set_key_ct(struct nlattr **tb,
1533                          struct flow_dissector_key_ct *key,
1534                          struct flow_dissector_key_ct *mask,
1535                          struct netlink_ext_ack *extack)
1536 {
1537         if (tb[TCA_FLOWER_KEY_CT_STATE]) {
1538                 int err;
1539
1540                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK)) {
1541                         NL_SET_ERR_MSG(extack, "Conntrack isn't enabled");
1542                         return -EOPNOTSUPP;
1543                 }
1544                 fl_set_key_val(tb, &key->ct_state, TCA_FLOWER_KEY_CT_STATE,
1545                                &mask->ct_state, TCA_FLOWER_KEY_CT_STATE_MASK,
1546                                sizeof(key->ct_state));
1547
1548                 err = fl_validate_ct_state(key->ct_state & mask->ct_state,
1549                                            tb[TCA_FLOWER_KEY_CT_STATE_MASK],
1550                                            extack);
1551                 if (err)
1552                         return err;
1553
1554         }
1555         if (tb[TCA_FLOWER_KEY_CT_ZONE]) {
1556                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK_ZONES)) {
1557                         NL_SET_ERR_MSG(extack, "Conntrack zones isn't enabled");
1558                         return -EOPNOTSUPP;
1559                 }
1560                 fl_set_key_val(tb, &key->ct_zone, TCA_FLOWER_KEY_CT_ZONE,
1561                                &mask->ct_zone, TCA_FLOWER_KEY_CT_ZONE_MASK,
1562                                sizeof(key->ct_zone));
1563         }
1564         if (tb[TCA_FLOWER_KEY_CT_MARK]) {
1565                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK_MARK)) {
1566                         NL_SET_ERR_MSG(extack, "Conntrack mark isn't enabled");
1567                         return -EOPNOTSUPP;
1568                 }
1569                 fl_set_key_val(tb, &key->ct_mark, TCA_FLOWER_KEY_CT_MARK,
1570                                &mask->ct_mark, TCA_FLOWER_KEY_CT_MARK_MASK,
1571                                sizeof(key->ct_mark));
1572         }
1573         if (tb[TCA_FLOWER_KEY_CT_LABELS]) {
1574                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK_LABELS)) {
1575                         NL_SET_ERR_MSG(extack, "Conntrack labels aren't enabled");
1576                         return -EOPNOTSUPP;
1577                 }
1578                 fl_set_key_val(tb, key->ct_labels, TCA_FLOWER_KEY_CT_LABELS,
1579                                mask->ct_labels, TCA_FLOWER_KEY_CT_LABELS_MASK,
1580                                sizeof(key->ct_labels));
1581         }
1582
1583         return 0;
1584 }
1585
1586 static bool is_vlan_key(struct nlattr *tb, __be16 *ethertype,
1587                         struct fl_flow_key *key, struct fl_flow_key *mask,
1588                         int vthresh)
1589 {
1590         const bool good_num_of_vlans = key->num_of_vlans.num_of_vlans > vthresh;
1591
1592         if (!tb) {
1593                 *ethertype = 0;
1594                 return good_num_of_vlans;
1595         }
1596
1597         *ethertype = nla_get_be16(tb);
1598         if (good_num_of_vlans || eth_type_vlan(*ethertype))
1599                 return true;
1600
1601         key->basic.n_proto = *ethertype;
1602         mask->basic.n_proto = cpu_to_be16(~0);
1603         return false;
1604 }
1605
1606 static int fl_set_key(struct net *net, struct nlattr **tb,
1607                       struct fl_flow_key *key, struct fl_flow_key *mask,
1608                       struct netlink_ext_ack *extack)
1609 {
1610         __be16 ethertype;
1611         int ret = 0;
1612
1613         if (tb[TCA_FLOWER_INDEV]) {
1614                 int err = tcf_change_indev(net, tb[TCA_FLOWER_INDEV], extack);
1615                 if (err < 0)
1616                         return err;
1617                 key->meta.ingress_ifindex = err;
1618                 mask->meta.ingress_ifindex = 0xffffffff;
1619         }
1620
1621         fl_set_key_val(tb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
1622                        mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
1623                        sizeof(key->eth.dst));
1624         fl_set_key_val(tb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
1625                        mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
1626                        sizeof(key->eth.src));
1627         fl_set_key_val(tb, &key->num_of_vlans,
1628                        TCA_FLOWER_KEY_NUM_OF_VLANS,
1629                        &mask->num_of_vlans,
1630                        TCA_FLOWER_UNSPEC,
1631                        sizeof(key->num_of_vlans));
1632
1633         if (is_vlan_key(tb[TCA_FLOWER_KEY_ETH_TYPE], &ethertype, key, mask, 0)) {
1634                 fl_set_key_vlan(tb, ethertype, TCA_FLOWER_KEY_VLAN_ID,
1635                                 TCA_FLOWER_KEY_VLAN_PRIO,
1636                                 TCA_FLOWER_KEY_VLAN_ETH_TYPE,
1637                                 &key->vlan, &mask->vlan);
1638
1639                 if (is_vlan_key(tb[TCA_FLOWER_KEY_VLAN_ETH_TYPE],
1640                                 &ethertype, key, mask, 1)) {
1641                         fl_set_key_vlan(tb, ethertype,
1642                                         TCA_FLOWER_KEY_CVLAN_ID,
1643                                         TCA_FLOWER_KEY_CVLAN_PRIO,
1644                                         TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
1645                                         &key->cvlan, &mask->cvlan);
1646                         fl_set_key_val(tb, &key->basic.n_proto,
1647                                        TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
1648                                        &mask->basic.n_proto,
1649                                        TCA_FLOWER_UNSPEC,
1650                                        sizeof(key->basic.n_proto));
1651                 }
1652         }
1653
1654         if (key->basic.n_proto == htons(ETH_P_IP) ||
1655             key->basic.n_proto == htons(ETH_P_IPV6)) {
1656                 fl_set_key_val(tb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
1657                                &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
1658                                sizeof(key->basic.ip_proto));
1659                 fl_set_key_ip(tb, false, &key->ip, &mask->ip);
1660         }
1661
1662         if (tb[TCA_FLOWER_KEY_IPV4_SRC] || tb[TCA_FLOWER_KEY_IPV4_DST]) {
1663                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
1664                 mask->control.addr_type = ~0;
1665                 fl_set_key_val(tb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
1666                                &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
1667                                sizeof(key->ipv4.src));
1668                 fl_set_key_val(tb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
1669                                &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
1670                                sizeof(key->ipv4.dst));
1671         } else if (tb[TCA_FLOWER_KEY_IPV6_SRC] || tb[TCA_FLOWER_KEY_IPV6_DST]) {
1672                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
1673                 mask->control.addr_type = ~0;
1674                 fl_set_key_val(tb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
1675                                &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
1676                                sizeof(key->ipv6.src));
1677                 fl_set_key_val(tb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
1678                                &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
1679                                sizeof(key->ipv6.dst));
1680         }
1681
1682         if (key->basic.ip_proto == IPPROTO_TCP) {
1683                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
1684                                &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
1685                                sizeof(key->tp.src));
1686                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
1687                                &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
1688                                sizeof(key->tp.dst));
1689                 fl_set_key_val(tb, &key->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS,
1690                                &mask->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS_MASK,
1691                                sizeof(key->tcp.flags));
1692         } else if (key->basic.ip_proto == IPPROTO_UDP) {
1693                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
1694                                &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
1695                                sizeof(key->tp.src));
1696                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
1697                                &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
1698                                sizeof(key->tp.dst));
1699         } else if (key->basic.ip_proto == IPPROTO_SCTP) {
1700                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_SCTP_SRC,
1701                                &mask->tp.src, TCA_FLOWER_KEY_SCTP_SRC_MASK,
1702                                sizeof(key->tp.src));
1703                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_SCTP_DST,
1704                                &mask->tp.dst, TCA_FLOWER_KEY_SCTP_DST_MASK,
1705                                sizeof(key->tp.dst));
1706         } else if (key->basic.n_proto == htons(ETH_P_IP) &&
1707                    key->basic.ip_proto == IPPROTO_ICMP) {
1708                 fl_set_key_val(tb, &key->icmp.type, TCA_FLOWER_KEY_ICMPV4_TYPE,
1709                                &mask->icmp.type,
1710                                TCA_FLOWER_KEY_ICMPV4_TYPE_MASK,
1711                                sizeof(key->icmp.type));
1712                 fl_set_key_val(tb, &key->icmp.code, TCA_FLOWER_KEY_ICMPV4_CODE,
1713                                &mask->icmp.code,
1714                                TCA_FLOWER_KEY_ICMPV4_CODE_MASK,
1715                                sizeof(key->icmp.code));
1716         } else if (key->basic.n_proto == htons(ETH_P_IPV6) &&
1717                    key->basic.ip_proto == IPPROTO_ICMPV6) {
1718                 fl_set_key_val(tb, &key->icmp.type, TCA_FLOWER_KEY_ICMPV6_TYPE,
1719                                &mask->icmp.type,
1720                                TCA_FLOWER_KEY_ICMPV6_TYPE_MASK,
1721                                sizeof(key->icmp.type));
1722                 fl_set_key_val(tb, &key->icmp.code, TCA_FLOWER_KEY_ICMPV6_CODE,
1723                                &mask->icmp.code,
1724                                TCA_FLOWER_KEY_ICMPV6_CODE_MASK,
1725                                sizeof(key->icmp.code));
1726         } else if (key->basic.n_proto == htons(ETH_P_MPLS_UC) ||
1727                    key->basic.n_proto == htons(ETH_P_MPLS_MC)) {
1728                 ret = fl_set_key_mpls(tb, &key->mpls, &mask->mpls, extack);
1729                 if (ret)
1730                         return ret;
1731         } else if (key->basic.n_proto == htons(ETH_P_ARP) ||
1732                    key->basic.n_proto == htons(ETH_P_RARP)) {
1733                 fl_set_key_val(tb, &key->arp.sip, TCA_FLOWER_KEY_ARP_SIP,
1734                                &mask->arp.sip, TCA_FLOWER_KEY_ARP_SIP_MASK,
1735                                sizeof(key->arp.sip));
1736                 fl_set_key_val(tb, &key->arp.tip, TCA_FLOWER_KEY_ARP_TIP,
1737                                &mask->arp.tip, TCA_FLOWER_KEY_ARP_TIP_MASK,
1738                                sizeof(key->arp.tip));
1739                 fl_set_key_val(tb, &key->arp.op, TCA_FLOWER_KEY_ARP_OP,
1740                                &mask->arp.op, TCA_FLOWER_KEY_ARP_OP_MASK,
1741                                sizeof(key->arp.op));
1742                 fl_set_key_val(tb, key->arp.sha, TCA_FLOWER_KEY_ARP_SHA,
1743                                mask->arp.sha, TCA_FLOWER_KEY_ARP_SHA_MASK,
1744                                sizeof(key->arp.sha));
1745                 fl_set_key_val(tb, key->arp.tha, TCA_FLOWER_KEY_ARP_THA,
1746                                mask->arp.tha, TCA_FLOWER_KEY_ARP_THA_MASK,
1747                                sizeof(key->arp.tha));
1748         }
1749
1750         if (key->basic.ip_proto == IPPROTO_TCP ||
1751             key->basic.ip_proto == IPPROTO_UDP ||
1752             key->basic.ip_proto == IPPROTO_SCTP) {
1753                 ret = fl_set_key_port_range(tb, key, mask, extack);
1754                 if (ret)
1755                         return ret;
1756         }
1757
1758         if (tb[TCA_FLOWER_KEY_ENC_IPV4_SRC] ||
1759             tb[TCA_FLOWER_KEY_ENC_IPV4_DST]) {
1760                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
1761                 mask->enc_control.addr_type = ~0;
1762                 fl_set_key_val(tb, &key->enc_ipv4.src,
1763                                TCA_FLOWER_KEY_ENC_IPV4_SRC,
1764                                &mask->enc_ipv4.src,
1765                                TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
1766                                sizeof(key->enc_ipv4.src));
1767                 fl_set_key_val(tb, &key->enc_ipv4.dst,
1768                                TCA_FLOWER_KEY_ENC_IPV4_DST,
1769                                &mask->enc_ipv4.dst,
1770                                TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
1771                                sizeof(key->enc_ipv4.dst));
1772         }
1773
1774         if (tb[TCA_FLOWER_KEY_ENC_IPV6_SRC] ||
1775             tb[TCA_FLOWER_KEY_ENC_IPV6_DST]) {
1776                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
1777                 mask->enc_control.addr_type = ~0;
1778                 fl_set_key_val(tb, &key->enc_ipv6.src,
1779                                TCA_FLOWER_KEY_ENC_IPV6_SRC,
1780                                &mask->enc_ipv6.src,
1781                                TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
1782                                sizeof(key->enc_ipv6.src));
1783                 fl_set_key_val(tb, &key->enc_ipv6.dst,
1784                                TCA_FLOWER_KEY_ENC_IPV6_DST,
1785                                &mask->enc_ipv6.dst,
1786                                TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
1787                                sizeof(key->enc_ipv6.dst));
1788         }
1789
1790         fl_set_key_val(tb, &key->enc_key_id.keyid, TCA_FLOWER_KEY_ENC_KEY_ID,
1791                        &mask->enc_key_id.keyid, TCA_FLOWER_UNSPEC,
1792                        sizeof(key->enc_key_id.keyid));
1793
1794         fl_set_key_val(tb, &key->enc_tp.src, TCA_FLOWER_KEY_ENC_UDP_SRC_PORT,
1795                        &mask->enc_tp.src, TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK,
1796                        sizeof(key->enc_tp.src));
1797
1798         fl_set_key_val(tb, &key->enc_tp.dst, TCA_FLOWER_KEY_ENC_UDP_DST_PORT,
1799                        &mask->enc_tp.dst, TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK,
1800                        sizeof(key->enc_tp.dst));
1801
1802         fl_set_key_ip(tb, true, &key->enc_ip, &mask->enc_ip);
1803
1804         fl_set_key_val(tb, &key->hash.hash, TCA_FLOWER_KEY_HASH,
1805                        &mask->hash.hash, TCA_FLOWER_KEY_HASH_MASK,
1806                        sizeof(key->hash.hash));
1807
1808         if (tb[TCA_FLOWER_KEY_ENC_OPTS]) {
1809                 ret = fl_set_enc_opt(tb, key, mask, extack);
1810                 if (ret)
1811                         return ret;
1812         }
1813
1814         ret = fl_set_key_ct(tb, &key->ct, &mask->ct, extack);
1815         if (ret)
1816                 return ret;
1817
1818         if (tb[TCA_FLOWER_KEY_FLAGS])
1819                 ret = fl_set_key_flags(tb, &key->control.flags,
1820                                        &mask->control.flags, extack);
1821
1822         return ret;
1823 }
1824
1825 static void fl_mask_copy(struct fl_flow_mask *dst,
1826                          struct fl_flow_mask *src)
1827 {
1828         const void *psrc = fl_key_get_start(&src->key, src);
1829         void *pdst = fl_key_get_start(&dst->key, src);
1830
1831         memcpy(pdst, psrc, fl_mask_range(src));
1832         dst->range = src->range;
1833 }
1834
1835 static const struct rhashtable_params fl_ht_params = {
1836         .key_offset = offsetof(struct cls_fl_filter, mkey), /* base offset */
1837         .head_offset = offsetof(struct cls_fl_filter, ht_node),
1838         .automatic_shrinking = true,
1839 };
1840
1841 static int fl_init_mask_hashtable(struct fl_flow_mask *mask)
1842 {
1843         mask->filter_ht_params = fl_ht_params;
1844         mask->filter_ht_params.key_len = fl_mask_range(mask);
1845         mask->filter_ht_params.key_offset += mask->range.start;
1846
1847         return rhashtable_init(&mask->ht, &mask->filter_ht_params);
1848 }
1849
1850 #define FL_KEY_MEMBER_OFFSET(member) offsetof(struct fl_flow_key, member)
1851 #define FL_KEY_MEMBER_SIZE(member) sizeof_field(struct fl_flow_key, member)
1852
1853 #define FL_KEY_IS_MASKED(mask, member)                                          \
1854         memchr_inv(((char *)mask) + FL_KEY_MEMBER_OFFSET(member),               \
1855                    0, FL_KEY_MEMBER_SIZE(member))                               \
1856
1857 #define FL_KEY_SET(keys, cnt, id, member)                                       \
1858         do {                                                                    \
1859                 keys[cnt].key_id = id;                                          \
1860                 keys[cnt].offset = FL_KEY_MEMBER_OFFSET(member);                \
1861                 cnt++;                                                          \
1862         } while(0);
1863
1864 #define FL_KEY_SET_IF_MASKED(mask, keys, cnt, id, member)                       \
1865         do {                                                                    \
1866                 if (FL_KEY_IS_MASKED(mask, member))                             \
1867                         FL_KEY_SET(keys, cnt, id, member);                      \
1868         } while(0);
1869
1870 static void fl_init_dissector(struct flow_dissector *dissector,
1871                               struct fl_flow_key *mask)
1872 {
1873         struct flow_dissector_key keys[FLOW_DISSECTOR_KEY_MAX];
1874         size_t cnt = 0;
1875
1876         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1877                              FLOW_DISSECTOR_KEY_META, meta);
1878         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_CONTROL, control);
1879         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_BASIC, basic);
1880         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1881                              FLOW_DISSECTOR_KEY_ETH_ADDRS, eth);
1882         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1883                              FLOW_DISSECTOR_KEY_IPV4_ADDRS, ipv4);
1884         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1885                              FLOW_DISSECTOR_KEY_IPV6_ADDRS, ipv6);
1886         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1887                              FLOW_DISSECTOR_KEY_PORTS, tp);
1888         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1889                              FLOW_DISSECTOR_KEY_PORTS_RANGE, tp_range);
1890         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1891                              FLOW_DISSECTOR_KEY_IP, ip);
1892         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1893                              FLOW_DISSECTOR_KEY_TCP, tcp);
1894         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1895                              FLOW_DISSECTOR_KEY_ICMP, icmp);
1896         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1897                              FLOW_DISSECTOR_KEY_ARP, arp);
1898         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1899                              FLOW_DISSECTOR_KEY_MPLS, mpls);
1900         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1901                              FLOW_DISSECTOR_KEY_VLAN, vlan);
1902         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1903                              FLOW_DISSECTOR_KEY_CVLAN, cvlan);
1904         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1905                              FLOW_DISSECTOR_KEY_ENC_KEYID, enc_key_id);
1906         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1907                              FLOW_DISSECTOR_KEY_ENC_IPV4_ADDRS, enc_ipv4);
1908         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1909                              FLOW_DISSECTOR_KEY_ENC_IPV6_ADDRS, enc_ipv6);
1910         if (FL_KEY_IS_MASKED(mask, enc_ipv4) ||
1911             FL_KEY_IS_MASKED(mask, enc_ipv6))
1912                 FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_ENC_CONTROL,
1913                            enc_control);
1914         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1915                              FLOW_DISSECTOR_KEY_ENC_PORTS, enc_tp);
1916         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1917                              FLOW_DISSECTOR_KEY_ENC_IP, enc_ip);
1918         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1919                              FLOW_DISSECTOR_KEY_ENC_OPTS, enc_opts);
1920         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1921                              FLOW_DISSECTOR_KEY_CT, ct);
1922         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1923                              FLOW_DISSECTOR_KEY_HASH, hash);
1924         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1925                              FLOW_DISSECTOR_KEY_NUM_OF_VLANS, num_of_vlans);
1926
1927         skb_flow_dissector_init(dissector, keys, cnt);
1928 }
1929
1930 static struct fl_flow_mask *fl_create_new_mask(struct cls_fl_head *head,
1931                                                struct fl_flow_mask *mask)
1932 {
1933         struct fl_flow_mask *newmask;
1934         int err;
1935
1936         newmask = kzalloc(sizeof(*newmask), GFP_KERNEL);
1937         if (!newmask)
1938                 return ERR_PTR(-ENOMEM);
1939
1940         fl_mask_copy(newmask, mask);
1941
1942         if ((newmask->key.tp_range.tp_min.dst &&
1943              newmask->key.tp_range.tp_max.dst) ||
1944             (newmask->key.tp_range.tp_min.src &&
1945              newmask->key.tp_range.tp_max.src))
1946                 newmask->flags |= TCA_FLOWER_MASK_FLAGS_RANGE;
1947
1948         err = fl_init_mask_hashtable(newmask);
1949         if (err)
1950                 goto errout_free;
1951
1952         fl_init_dissector(&newmask->dissector, &newmask->key);
1953
1954         INIT_LIST_HEAD_RCU(&newmask->filters);
1955
1956         refcount_set(&newmask->refcnt, 1);
1957         err = rhashtable_replace_fast(&head->ht, &mask->ht_node,
1958                                       &newmask->ht_node, mask_ht_params);
1959         if (err)
1960                 goto errout_destroy;
1961
1962         spin_lock(&head->masks_lock);
1963         list_add_tail_rcu(&newmask->list, &head->masks);
1964         spin_unlock(&head->masks_lock);
1965
1966         return newmask;
1967
1968 errout_destroy:
1969         rhashtable_destroy(&newmask->ht);
1970 errout_free:
1971         kfree(newmask);
1972
1973         return ERR_PTR(err);
1974 }
1975
1976 static int fl_check_assign_mask(struct cls_fl_head *head,
1977                                 struct cls_fl_filter *fnew,
1978                                 struct cls_fl_filter *fold,
1979                                 struct fl_flow_mask *mask)
1980 {
1981         struct fl_flow_mask *newmask;
1982         int ret = 0;
1983
1984         rcu_read_lock();
1985
1986         /* Insert mask as temporary node to prevent concurrent creation of mask
1987          * with same key. Any concurrent lookups with same key will return
1988          * -EAGAIN because mask's refcnt is zero.
1989          */
1990         fnew->mask = rhashtable_lookup_get_insert_fast(&head->ht,
1991                                                        &mask->ht_node,
1992                                                        mask_ht_params);
1993         if (!fnew->mask) {
1994                 rcu_read_unlock();
1995
1996                 if (fold) {
1997                         ret = -EINVAL;
1998                         goto errout_cleanup;
1999                 }
2000
2001                 newmask = fl_create_new_mask(head, mask);
2002                 if (IS_ERR(newmask)) {
2003                         ret = PTR_ERR(newmask);
2004                         goto errout_cleanup;
2005                 }
2006
2007                 fnew->mask = newmask;
2008                 return 0;
2009         } else if (IS_ERR(fnew->mask)) {
2010                 ret = PTR_ERR(fnew->mask);
2011         } else if (fold && fold->mask != fnew->mask) {
2012                 ret = -EINVAL;
2013         } else if (!refcount_inc_not_zero(&fnew->mask->refcnt)) {
2014                 /* Mask was deleted concurrently, try again */
2015                 ret = -EAGAIN;
2016         }
2017         rcu_read_unlock();
2018         return ret;
2019
2020 errout_cleanup:
2021         rhashtable_remove_fast(&head->ht, &mask->ht_node,
2022                                mask_ht_params);
2023         return ret;
2024 }
2025
2026 static int fl_set_parms(struct net *net, struct tcf_proto *tp,
2027                         struct cls_fl_filter *f, struct fl_flow_mask *mask,
2028                         unsigned long base, struct nlattr **tb,
2029                         struct nlattr *est,
2030                         struct fl_flow_tmplt *tmplt,
2031                         u32 flags, u32 fl_flags,
2032                         struct netlink_ext_ack *extack)
2033 {
2034         int err;
2035
2036         err = tcf_exts_validate_ex(net, tp, tb, est, &f->exts, flags,
2037                                    fl_flags, extack);
2038         if (err < 0)
2039                 return err;
2040
2041         if (tb[TCA_FLOWER_CLASSID]) {
2042                 f->res.classid = nla_get_u32(tb[TCA_FLOWER_CLASSID]);
2043                 if (flags & TCA_ACT_FLAGS_NO_RTNL)
2044                         rtnl_lock();
2045                 tcf_bind_filter(tp, &f->res, base);
2046                 if (flags & TCA_ACT_FLAGS_NO_RTNL)
2047                         rtnl_unlock();
2048         }
2049
2050         err = fl_set_key(net, tb, &f->key, &mask->key, extack);
2051         if (err)
2052                 return err;
2053
2054         fl_mask_update_range(mask);
2055         fl_set_masked_key(&f->mkey, &f->key, mask);
2056
2057         if (!fl_mask_fits_tmplt(tmplt, mask)) {
2058                 NL_SET_ERR_MSG_MOD(extack, "Mask does not fit the template");
2059                 return -EINVAL;
2060         }
2061
2062         return 0;
2063 }
2064
2065 static int fl_ht_insert_unique(struct cls_fl_filter *fnew,
2066                                struct cls_fl_filter *fold,
2067                                bool *in_ht)
2068 {
2069         struct fl_flow_mask *mask = fnew->mask;
2070         int err;
2071
2072         err = rhashtable_lookup_insert_fast(&mask->ht,
2073                                             &fnew->ht_node,
2074                                             mask->filter_ht_params);
2075         if (err) {
2076                 *in_ht = false;
2077                 /* It is okay if filter with same key exists when
2078                  * overwriting.
2079                  */
2080                 return fold && err == -EEXIST ? 0 : err;
2081         }
2082
2083         *in_ht = true;
2084         return 0;
2085 }
2086
2087 static int fl_change(struct net *net, struct sk_buff *in_skb,
2088                      struct tcf_proto *tp, unsigned long base,
2089                      u32 handle, struct nlattr **tca,
2090                      void **arg, u32 flags,
2091                      struct netlink_ext_ack *extack)
2092 {
2093         struct cls_fl_head *head = fl_head_dereference(tp);
2094         bool rtnl_held = !(flags & TCA_ACT_FLAGS_NO_RTNL);
2095         struct cls_fl_filter *fold = *arg;
2096         struct cls_fl_filter *fnew;
2097         struct fl_flow_mask *mask;
2098         struct nlattr **tb;
2099         bool in_ht;
2100         int err;
2101
2102         if (!tca[TCA_OPTIONS]) {
2103                 err = -EINVAL;
2104                 goto errout_fold;
2105         }
2106
2107         mask = kzalloc(sizeof(struct fl_flow_mask), GFP_KERNEL);
2108         if (!mask) {
2109                 err = -ENOBUFS;
2110                 goto errout_fold;
2111         }
2112
2113         tb = kcalloc(TCA_FLOWER_MAX + 1, sizeof(struct nlattr *), GFP_KERNEL);
2114         if (!tb) {
2115                 err = -ENOBUFS;
2116                 goto errout_mask_alloc;
2117         }
2118
2119         err = nla_parse_nested_deprecated(tb, TCA_FLOWER_MAX,
2120                                           tca[TCA_OPTIONS], fl_policy, NULL);
2121         if (err < 0)
2122                 goto errout_tb;
2123
2124         if (fold && handle && fold->handle != handle) {
2125                 err = -EINVAL;
2126                 goto errout_tb;
2127         }
2128
2129         fnew = kzalloc(sizeof(*fnew), GFP_KERNEL);
2130         if (!fnew) {
2131                 err = -ENOBUFS;
2132                 goto errout_tb;
2133         }
2134         INIT_LIST_HEAD(&fnew->hw_list);
2135         refcount_set(&fnew->refcnt, 1);
2136
2137         err = tcf_exts_init(&fnew->exts, net, TCA_FLOWER_ACT, 0);
2138         if (err < 0)
2139                 goto errout;
2140
2141         if (tb[TCA_FLOWER_FLAGS]) {
2142                 fnew->flags = nla_get_u32(tb[TCA_FLOWER_FLAGS]);
2143
2144                 if (!tc_flags_valid(fnew->flags)) {
2145                         err = -EINVAL;
2146                         goto errout;
2147                 }
2148         }
2149
2150         err = fl_set_parms(net, tp, fnew, mask, base, tb, tca[TCA_RATE],
2151                            tp->chain->tmplt_priv, flags, fnew->flags,
2152                            extack);
2153         if (err)
2154                 goto errout;
2155
2156         err = fl_check_assign_mask(head, fnew, fold, mask);
2157         if (err)
2158                 goto errout;
2159
2160         err = fl_ht_insert_unique(fnew, fold, &in_ht);
2161         if (err)
2162                 goto errout_mask;
2163
2164         if (!tc_skip_hw(fnew->flags)) {
2165                 err = fl_hw_replace_filter(tp, fnew, rtnl_held, extack);
2166                 if (err)
2167                         goto errout_ht;
2168         }
2169
2170         if (!tc_in_hw(fnew->flags))
2171                 fnew->flags |= TCA_CLS_FLAGS_NOT_IN_HW;
2172
2173         spin_lock(&tp->lock);
2174
2175         /* tp was deleted concurrently. -EAGAIN will cause caller to lookup
2176          * proto again or create new one, if necessary.
2177          */
2178         if (tp->deleting) {
2179                 err = -EAGAIN;
2180                 goto errout_hw;
2181         }
2182
2183         if (fold) {
2184                 /* Fold filter was deleted concurrently. Retry lookup. */
2185                 if (fold->deleted) {
2186                         err = -EAGAIN;
2187                         goto errout_hw;
2188                 }
2189
2190                 fnew->handle = handle;
2191
2192                 if (!in_ht) {
2193                         struct rhashtable_params params =
2194                                 fnew->mask->filter_ht_params;
2195
2196                         err = rhashtable_insert_fast(&fnew->mask->ht,
2197                                                      &fnew->ht_node,
2198                                                      params);
2199                         if (err)
2200                                 goto errout_hw;
2201                         in_ht = true;
2202                 }
2203
2204                 refcount_inc(&fnew->refcnt);
2205                 rhashtable_remove_fast(&fold->mask->ht,
2206                                        &fold->ht_node,
2207                                        fold->mask->filter_ht_params);
2208                 idr_replace(&head->handle_idr, fnew, fnew->handle);
2209                 list_replace_rcu(&fold->list, &fnew->list);
2210                 fold->deleted = true;
2211
2212                 spin_unlock(&tp->lock);
2213
2214                 fl_mask_put(head, fold->mask);
2215                 if (!tc_skip_hw(fold->flags))
2216                         fl_hw_destroy_filter(tp, fold, rtnl_held, NULL);
2217                 tcf_unbind_filter(tp, &fold->res);
2218                 /* Caller holds reference to fold, so refcnt is always > 0
2219                  * after this.
2220                  */
2221                 refcount_dec(&fold->refcnt);
2222                 __fl_put(fold);
2223         } else {
2224                 if (handle) {
2225                         /* user specifies a handle and it doesn't exist */
2226                         err = idr_alloc_u32(&head->handle_idr, fnew, &handle,
2227                                             handle, GFP_ATOMIC);
2228
2229                         /* Filter with specified handle was concurrently
2230                          * inserted after initial check in cls_api. This is not
2231                          * necessarily an error if NLM_F_EXCL is not set in
2232                          * message flags. Returning EAGAIN will cause cls_api to
2233                          * try to update concurrently inserted rule.
2234                          */
2235                         if (err == -ENOSPC)
2236                                 err = -EAGAIN;
2237                 } else {
2238                         handle = 1;
2239                         err = idr_alloc_u32(&head->handle_idr, fnew, &handle,
2240                                             INT_MAX, GFP_ATOMIC);
2241                 }
2242                 if (err)
2243                         goto errout_hw;
2244
2245                 refcount_inc(&fnew->refcnt);
2246                 fnew->handle = handle;
2247                 list_add_tail_rcu(&fnew->list, &fnew->mask->filters);
2248                 spin_unlock(&tp->lock);
2249         }
2250
2251         *arg = fnew;
2252
2253         kfree(tb);
2254         tcf_queue_work(&mask->rwork, fl_uninit_mask_free_work);
2255         return 0;
2256
2257 errout_ht:
2258         spin_lock(&tp->lock);
2259 errout_hw:
2260         fnew->deleted = true;
2261         spin_unlock(&tp->lock);
2262         if (!tc_skip_hw(fnew->flags))
2263                 fl_hw_destroy_filter(tp, fnew, rtnl_held, NULL);
2264         if (in_ht)
2265                 rhashtable_remove_fast(&fnew->mask->ht, &fnew->ht_node,
2266                                        fnew->mask->filter_ht_params);
2267 errout_mask:
2268         fl_mask_put(head, fnew->mask);
2269 errout:
2270         __fl_put(fnew);
2271 errout_tb:
2272         kfree(tb);
2273 errout_mask_alloc:
2274         tcf_queue_work(&mask->rwork, fl_uninit_mask_free_work);
2275 errout_fold:
2276         if (fold)
2277                 __fl_put(fold);
2278         return err;
2279 }
2280
2281 static int fl_delete(struct tcf_proto *tp, void *arg, bool *last,
2282                      bool rtnl_held, struct netlink_ext_ack *extack)
2283 {
2284         struct cls_fl_head *head = fl_head_dereference(tp);
2285         struct cls_fl_filter *f = arg;
2286         bool last_on_mask;
2287         int err = 0;
2288
2289         err = __fl_delete(tp, f, &last_on_mask, rtnl_held, extack);
2290         *last = list_empty(&head->masks);
2291         __fl_put(f);
2292
2293         return err;
2294 }
2295
2296 static void fl_walk(struct tcf_proto *tp, struct tcf_walker *arg,
2297                     bool rtnl_held)
2298 {
2299         struct cls_fl_head *head = fl_head_dereference(tp);
2300         unsigned long id = arg->cookie, tmp;
2301         struct cls_fl_filter *f;
2302
2303         arg->count = arg->skip;
2304
2305         rcu_read_lock();
2306         idr_for_each_entry_continue_ul(&head->handle_idr, f, tmp, id) {
2307                 /* don't return filters that are being deleted */
2308                 if (!refcount_inc_not_zero(&f->refcnt))
2309                         continue;
2310                 rcu_read_unlock();
2311
2312                 if (arg->fn(tp, f, arg) < 0) {
2313                         __fl_put(f);
2314                         arg->stop = 1;
2315                         rcu_read_lock();
2316                         break;
2317                 }
2318                 __fl_put(f);
2319                 arg->count++;
2320                 rcu_read_lock();
2321         }
2322         rcu_read_unlock();
2323         arg->cookie = id;
2324 }
2325
2326 static struct cls_fl_filter *
2327 fl_get_next_hw_filter(struct tcf_proto *tp, struct cls_fl_filter *f, bool add)
2328 {
2329         struct cls_fl_head *head = fl_head_dereference(tp);
2330
2331         spin_lock(&tp->lock);
2332         if (list_empty(&head->hw_filters)) {
2333                 spin_unlock(&tp->lock);
2334                 return NULL;
2335         }
2336
2337         if (!f)
2338                 f = list_entry(&head->hw_filters, struct cls_fl_filter,
2339                                hw_list);
2340         list_for_each_entry_continue(f, &head->hw_filters, hw_list) {
2341                 if (!(add && f->deleted) && refcount_inc_not_zero(&f->refcnt)) {
2342                         spin_unlock(&tp->lock);
2343                         return f;
2344                 }
2345         }
2346
2347         spin_unlock(&tp->lock);
2348         return NULL;
2349 }
2350
2351 static int fl_reoffload(struct tcf_proto *tp, bool add, flow_setup_cb_t *cb,
2352                         void *cb_priv, struct netlink_ext_ack *extack)
2353 {
2354         struct tcf_block *block = tp->chain->block;
2355         struct flow_cls_offload cls_flower = {};
2356         struct cls_fl_filter *f = NULL;
2357         int err;
2358
2359         /* hw_filters list can only be changed by hw offload functions after
2360          * obtaining rtnl lock. Make sure it is not changed while reoffload is
2361          * iterating it.
2362          */
2363         ASSERT_RTNL();
2364
2365         while ((f = fl_get_next_hw_filter(tp, f, add))) {
2366                 cls_flower.rule =
2367                         flow_rule_alloc(tcf_exts_num_actions(&f->exts));
2368                 if (!cls_flower.rule) {
2369                         __fl_put(f);
2370                         return -ENOMEM;
2371                 }
2372
2373                 tc_cls_common_offload_init(&cls_flower.common, tp, f->flags,
2374                                            extack);
2375                 cls_flower.command = add ?
2376                         FLOW_CLS_REPLACE : FLOW_CLS_DESTROY;
2377                 cls_flower.cookie = (unsigned long)f;
2378                 cls_flower.rule->match.dissector = &f->mask->dissector;
2379                 cls_flower.rule->match.mask = &f->mask->key;
2380                 cls_flower.rule->match.key = &f->mkey;
2381
2382                 err = tc_setup_offload_action(&cls_flower.rule->action, &f->exts,
2383                                               cls_flower.common.extack);
2384                 if (err) {
2385                         kfree(cls_flower.rule);
2386                         if (tc_skip_sw(f->flags)) {
2387                                 __fl_put(f);
2388                                 return err;
2389                         }
2390                         goto next_flow;
2391                 }
2392
2393                 cls_flower.classid = f->res.classid;
2394
2395                 err = tc_setup_cb_reoffload(block, tp, add, cb,
2396                                             TC_SETUP_CLSFLOWER, &cls_flower,
2397                                             cb_priv, &f->flags,
2398                                             &f->in_hw_count);
2399                 tc_cleanup_offload_action(&cls_flower.rule->action);
2400                 kfree(cls_flower.rule);
2401
2402                 if (err) {
2403                         __fl_put(f);
2404                         return err;
2405                 }
2406 next_flow:
2407                 __fl_put(f);
2408         }
2409
2410         return 0;
2411 }
2412
2413 static void fl_hw_add(struct tcf_proto *tp, void *type_data)
2414 {
2415         struct flow_cls_offload *cls_flower = type_data;
2416         struct cls_fl_filter *f =
2417                 (struct cls_fl_filter *) cls_flower->cookie;
2418         struct cls_fl_head *head = fl_head_dereference(tp);
2419
2420         spin_lock(&tp->lock);
2421         list_add(&f->hw_list, &head->hw_filters);
2422         spin_unlock(&tp->lock);
2423 }
2424
2425 static void fl_hw_del(struct tcf_proto *tp, void *type_data)
2426 {
2427         struct flow_cls_offload *cls_flower = type_data;
2428         struct cls_fl_filter *f =
2429                 (struct cls_fl_filter *) cls_flower->cookie;
2430
2431         spin_lock(&tp->lock);
2432         if (!list_empty(&f->hw_list))
2433                 list_del_init(&f->hw_list);
2434         spin_unlock(&tp->lock);
2435 }
2436
2437 static int fl_hw_create_tmplt(struct tcf_chain *chain,
2438                               struct fl_flow_tmplt *tmplt)
2439 {
2440         struct flow_cls_offload cls_flower = {};
2441         struct tcf_block *block = chain->block;
2442
2443         cls_flower.rule = flow_rule_alloc(0);
2444         if (!cls_flower.rule)
2445                 return -ENOMEM;
2446
2447         cls_flower.common.chain_index = chain->index;
2448         cls_flower.command = FLOW_CLS_TMPLT_CREATE;
2449         cls_flower.cookie = (unsigned long) tmplt;
2450         cls_flower.rule->match.dissector = &tmplt->dissector;
2451         cls_flower.rule->match.mask = &tmplt->mask;
2452         cls_flower.rule->match.key = &tmplt->dummy_key;
2453
2454         /* We don't care if driver (any of them) fails to handle this
2455          * call. It serves just as a hint for it.
2456          */
2457         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false, true);
2458         kfree(cls_flower.rule);
2459
2460         return 0;
2461 }
2462
2463 static void fl_hw_destroy_tmplt(struct tcf_chain *chain,
2464                                 struct fl_flow_tmplt *tmplt)
2465 {
2466         struct flow_cls_offload cls_flower = {};
2467         struct tcf_block *block = chain->block;
2468
2469         cls_flower.common.chain_index = chain->index;
2470         cls_flower.command = FLOW_CLS_TMPLT_DESTROY;
2471         cls_flower.cookie = (unsigned long) tmplt;
2472
2473         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false, true);
2474 }
2475
2476 static void *fl_tmplt_create(struct net *net, struct tcf_chain *chain,
2477                              struct nlattr **tca,
2478                              struct netlink_ext_ack *extack)
2479 {
2480         struct fl_flow_tmplt *tmplt;
2481         struct nlattr **tb;
2482         int err;
2483
2484         if (!tca[TCA_OPTIONS])
2485                 return ERR_PTR(-EINVAL);
2486
2487         tb = kcalloc(TCA_FLOWER_MAX + 1, sizeof(struct nlattr *), GFP_KERNEL);
2488         if (!tb)
2489                 return ERR_PTR(-ENOBUFS);
2490         err = nla_parse_nested_deprecated(tb, TCA_FLOWER_MAX,
2491                                           tca[TCA_OPTIONS], fl_policy, NULL);
2492         if (err)
2493                 goto errout_tb;
2494
2495         tmplt = kzalloc(sizeof(*tmplt), GFP_KERNEL);
2496         if (!tmplt) {
2497                 err = -ENOMEM;
2498                 goto errout_tb;
2499         }
2500         tmplt->chain = chain;
2501         err = fl_set_key(net, tb, &tmplt->dummy_key, &tmplt->mask, extack);
2502         if (err)
2503                 goto errout_tmplt;
2504
2505         fl_init_dissector(&tmplt->dissector, &tmplt->mask);
2506
2507         err = fl_hw_create_tmplt(chain, tmplt);
2508         if (err)
2509                 goto errout_tmplt;
2510
2511         kfree(tb);
2512         return tmplt;
2513
2514 errout_tmplt:
2515         kfree(tmplt);
2516 errout_tb:
2517         kfree(tb);
2518         return ERR_PTR(err);
2519 }
2520
2521 static void fl_tmplt_destroy(void *tmplt_priv)
2522 {
2523         struct fl_flow_tmplt *tmplt = tmplt_priv;
2524
2525         fl_hw_destroy_tmplt(tmplt->chain, tmplt);
2526         kfree(tmplt);
2527 }
2528
2529 static int fl_dump_key_val(struct sk_buff *skb,
2530                            void *val, int val_type,
2531                            void *mask, int mask_type, int len)
2532 {
2533         int err;
2534
2535         if (!memchr_inv(mask, 0, len))
2536                 return 0;
2537         err = nla_put(skb, val_type, len, val);
2538         if (err)
2539                 return err;
2540         if (mask_type != TCA_FLOWER_UNSPEC) {
2541                 err = nla_put(skb, mask_type, len, mask);
2542                 if (err)
2543                         return err;
2544         }
2545         return 0;
2546 }
2547
2548 static int fl_dump_key_port_range(struct sk_buff *skb, struct fl_flow_key *key,
2549                                   struct fl_flow_key *mask)
2550 {
2551         if (fl_dump_key_val(skb, &key->tp_range.tp_min.dst,
2552                             TCA_FLOWER_KEY_PORT_DST_MIN,
2553                             &mask->tp_range.tp_min.dst, TCA_FLOWER_UNSPEC,
2554                             sizeof(key->tp_range.tp_min.dst)) ||
2555             fl_dump_key_val(skb, &key->tp_range.tp_max.dst,
2556                             TCA_FLOWER_KEY_PORT_DST_MAX,
2557                             &mask->tp_range.tp_max.dst, TCA_FLOWER_UNSPEC,
2558                             sizeof(key->tp_range.tp_max.dst)) ||
2559             fl_dump_key_val(skb, &key->tp_range.tp_min.src,
2560                             TCA_FLOWER_KEY_PORT_SRC_MIN,
2561                             &mask->tp_range.tp_min.src, TCA_FLOWER_UNSPEC,
2562                             sizeof(key->tp_range.tp_min.src)) ||
2563             fl_dump_key_val(skb, &key->tp_range.tp_max.src,
2564                             TCA_FLOWER_KEY_PORT_SRC_MAX,
2565                             &mask->tp_range.tp_max.src, TCA_FLOWER_UNSPEC,
2566                             sizeof(key->tp_range.tp_max.src)))
2567                 return -1;
2568
2569         return 0;
2570 }
2571
2572 static int fl_dump_key_mpls_opt_lse(struct sk_buff *skb,
2573                                     struct flow_dissector_key_mpls *mpls_key,
2574                                     struct flow_dissector_key_mpls *mpls_mask,
2575                                     u8 lse_index)
2576 {
2577         struct flow_dissector_mpls_lse *lse_mask = &mpls_mask->ls[lse_index];
2578         struct flow_dissector_mpls_lse *lse_key = &mpls_key->ls[lse_index];
2579         int err;
2580
2581         err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH,
2582                          lse_index + 1);
2583         if (err)
2584                 return err;
2585
2586         if (lse_mask->mpls_ttl) {
2587                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_TTL,
2588                                  lse_key->mpls_ttl);
2589                 if (err)
2590                         return err;
2591         }
2592         if (lse_mask->mpls_bos) {
2593                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS,
2594                                  lse_key->mpls_bos);
2595                 if (err)
2596                         return err;
2597         }
2598         if (lse_mask->mpls_tc) {
2599                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_TC,
2600                                  lse_key->mpls_tc);
2601                 if (err)
2602                         return err;
2603         }
2604         if (lse_mask->mpls_label) {
2605                 err = nla_put_u32(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL,
2606                                   lse_key->mpls_label);
2607                 if (err)
2608                         return err;
2609         }
2610
2611         return 0;
2612 }
2613
2614 static int fl_dump_key_mpls_opts(struct sk_buff *skb,
2615                                  struct flow_dissector_key_mpls *mpls_key,
2616                                  struct flow_dissector_key_mpls *mpls_mask)
2617 {
2618         struct nlattr *opts;
2619         struct nlattr *lse;
2620         u8 lse_index;
2621         int err;
2622
2623         opts = nla_nest_start(skb, TCA_FLOWER_KEY_MPLS_OPTS);
2624         if (!opts)
2625                 return -EMSGSIZE;
2626
2627         for (lse_index = 0; lse_index < FLOW_DIS_MPLS_MAX; lse_index++) {
2628                 if (!(mpls_mask->used_lses & 1 << lse_index))
2629                         continue;
2630
2631                 lse = nla_nest_start(skb, TCA_FLOWER_KEY_MPLS_OPTS_LSE);
2632                 if (!lse) {
2633                         err = -EMSGSIZE;
2634                         goto err_opts;
2635                 }
2636
2637                 err = fl_dump_key_mpls_opt_lse(skb, mpls_key, mpls_mask,
2638                                                lse_index);
2639                 if (err)
2640                         goto err_opts_lse;
2641                 nla_nest_end(skb, lse);
2642         }
2643         nla_nest_end(skb, opts);
2644
2645         return 0;
2646
2647 err_opts_lse:
2648         nla_nest_cancel(skb, lse);
2649 err_opts:
2650         nla_nest_cancel(skb, opts);
2651
2652         return err;
2653 }
2654
2655 static int fl_dump_key_mpls(struct sk_buff *skb,
2656                             struct flow_dissector_key_mpls *mpls_key,
2657                             struct flow_dissector_key_mpls *mpls_mask)
2658 {
2659         struct flow_dissector_mpls_lse *lse_mask;
2660         struct flow_dissector_mpls_lse *lse_key;
2661         int err;
2662
2663         if (!mpls_mask->used_lses)
2664                 return 0;
2665
2666         lse_mask = &mpls_mask->ls[0];
2667         lse_key = &mpls_key->ls[0];
2668
2669         /* For backward compatibility, don't use the MPLS nested attributes if
2670          * the rule can be expressed using the old attributes.
2671          */
2672         if (mpls_mask->used_lses & ~1 ||
2673             (!lse_mask->mpls_ttl && !lse_mask->mpls_bos &&
2674              !lse_mask->mpls_tc && !lse_mask->mpls_label))
2675                 return fl_dump_key_mpls_opts(skb, mpls_key, mpls_mask);
2676
2677         if (lse_mask->mpls_ttl) {
2678                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_TTL,
2679                                  lse_key->mpls_ttl);
2680                 if (err)
2681                         return err;
2682         }
2683         if (lse_mask->mpls_tc) {
2684                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_TC,
2685                                  lse_key->mpls_tc);
2686                 if (err)
2687                         return err;
2688         }
2689         if (lse_mask->mpls_label) {
2690                 err = nla_put_u32(skb, TCA_FLOWER_KEY_MPLS_LABEL,
2691                                   lse_key->mpls_label);
2692                 if (err)
2693                         return err;
2694         }
2695         if (lse_mask->mpls_bos) {
2696                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_BOS,
2697                                  lse_key->mpls_bos);
2698                 if (err)
2699                         return err;
2700         }
2701         return 0;
2702 }
2703
2704 static int fl_dump_key_ip(struct sk_buff *skb, bool encap,
2705                           struct flow_dissector_key_ip *key,
2706                           struct flow_dissector_key_ip *mask)
2707 {
2708         int tos_key = encap ? TCA_FLOWER_KEY_ENC_IP_TOS : TCA_FLOWER_KEY_IP_TOS;
2709         int ttl_key = encap ? TCA_FLOWER_KEY_ENC_IP_TTL : TCA_FLOWER_KEY_IP_TTL;
2710         int tos_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TOS_MASK : TCA_FLOWER_KEY_IP_TOS_MASK;
2711         int ttl_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TTL_MASK : TCA_FLOWER_KEY_IP_TTL_MASK;
2712
2713         if (fl_dump_key_val(skb, &key->tos, tos_key, &mask->tos, tos_mask, sizeof(key->tos)) ||
2714             fl_dump_key_val(skb, &key->ttl, ttl_key, &mask->ttl, ttl_mask, sizeof(key->ttl)))
2715                 return -1;
2716
2717         return 0;
2718 }
2719
2720 static int fl_dump_key_vlan(struct sk_buff *skb,
2721                             int vlan_id_key, int vlan_prio_key,
2722                             struct flow_dissector_key_vlan *vlan_key,
2723                             struct flow_dissector_key_vlan *vlan_mask)
2724 {
2725         int err;
2726
2727         if (!memchr_inv(vlan_mask, 0, sizeof(*vlan_mask)))
2728                 return 0;
2729         if (vlan_mask->vlan_id) {
2730                 err = nla_put_u16(skb, vlan_id_key,
2731                                   vlan_key->vlan_id);
2732                 if (err)
2733                         return err;
2734         }
2735         if (vlan_mask->vlan_priority) {
2736                 err = nla_put_u8(skb, vlan_prio_key,
2737                                  vlan_key->vlan_priority);
2738                 if (err)
2739                         return err;
2740         }
2741         return 0;
2742 }
2743
2744 static void fl_get_key_flag(u32 dissector_key, u32 dissector_mask,
2745                             u32 *flower_key, u32 *flower_mask,
2746                             u32 flower_flag_bit, u32 dissector_flag_bit)
2747 {
2748         if (dissector_mask & dissector_flag_bit) {
2749                 *flower_mask |= flower_flag_bit;
2750                 if (dissector_key & dissector_flag_bit)
2751                         *flower_key |= flower_flag_bit;
2752         }
2753 }
2754
2755 static int fl_dump_key_flags(struct sk_buff *skb, u32 flags_key, u32 flags_mask)
2756 {
2757         u32 key, mask;
2758         __be32 _key, _mask;
2759         int err;
2760
2761         if (!memchr_inv(&flags_mask, 0, sizeof(flags_mask)))
2762                 return 0;
2763
2764         key = 0;
2765         mask = 0;
2766
2767         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
2768                         TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT, FLOW_DIS_IS_FRAGMENT);
2769         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
2770                         TCA_FLOWER_KEY_FLAGS_FRAG_IS_FIRST,
2771                         FLOW_DIS_FIRST_FRAG);
2772
2773         _key = cpu_to_be32(key);
2774         _mask = cpu_to_be32(mask);
2775
2776         err = nla_put(skb, TCA_FLOWER_KEY_FLAGS, 4, &_key);
2777         if (err)
2778                 return err;
2779
2780         return nla_put(skb, TCA_FLOWER_KEY_FLAGS_MASK, 4, &_mask);
2781 }
2782
2783 static int fl_dump_key_geneve_opt(struct sk_buff *skb,
2784                                   struct flow_dissector_key_enc_opts *enc_opts)
2785 {
2786         struct geneve_opt *opt;
2787         struct nlattr *nest;
2788         int opt_off = 0;
2789
2790         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_GENEVE);
2791         if (!nest)
2792                 goto nla_put_failure;
2793
2794         while (enc_opts->len > opt_off) {
2795                 opt = (struct geneve_opt *)&enc_opts->data[opt_off];
2796
2797                 if (nla_put_be16(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS,
2798                                  opt->opt_class))
2799                         goto nla_put_failure;
2800                 if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE,
2801                                opt->type))
2802                         goto nla_put_failure;
2803                 if (nla_put(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA,
2804                             opt->length * 4, opt->opt_data))
2805                         goto nla_put_failure;
2806
2807                 opt_off += sizeof(struct geneve_opt) + opt->length * 4;
2808         }
2809         nla_nest_end(skb, nest);
2810         return 0;
2811
2812 nla_put_failure:
2813         nla_nest_cancel(skb, nest);
2814         return -EMSGSIZE;
2815 }
2816
2817 static int fl_dump_key_vxlan_opt(struct sk_buff *skb,
2818                                  struct flow_dissector_key_enc_opts *enc_opts)
2819 {
2820         struct vxlan_metadata *md;
2821         struct nlattr *nest;
2822
2823         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_VXLAN);
2824         if (!nest)
2825                 goto nla_put_failure;
2826
2827         md = (struct vxlan_metadata *)&enc_opts->data[0];
2828         if (nla_put_u32(skb, TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP, md->gbp))
2829                 goto nla_put_failure;
2830
2831         nla_nest_end(skb, nest);
2832         return 0;
2833
2834 nla_put_failure:
2835         nla_nest_cancel(skb, nest);
2836         return -EMSGSIZE;
2837 }
2838
2839 static int fl_dump_key_erspan_opt(struct sk_buff *skb,
2840                                   struct flow_dissector_key_enc_opts *enc_opts)
2841 {
2842         struct erspan_metadata *md;
2843         struct nlattr *nest;
2844
2845         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_ERSPAN);
2846         if (!nest)
2847                 goto nla_put_failure;
2848
2849         md = (struct erspan_metadata *)&enc_opts->data[0];
2850         if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER, md->version))
2851                 goto nla_put_failure;
2852
2853         if (md->version == 1 &&
2854             nla_put_be32(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX, md->u.index))
2855                 goto nla_put_failure;
2856
2857         if (md->version == 2 &&
2858             (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR,
2859                         md->u.md2.dir) ||
2860              nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID,
2861                         get_hwid(&md->u.md2))))
2862                 goto nla_put_failure;
2863
2864         nla_nest_end(skb, nest);
2865         return 0;
2866
2867 nla_put_failure:
2868         nla_nest_cancel(skb, nest);
2869         return -EMSGSIZE;
2870 }
2871
2872 static int fl_dump_key_gtp_opt(struct sk_buff *skb,
2873                                struct flow_dissector_key_enc_opts *enc_opts)
2874
2875 {
2876         struct gtp_pdu_session_info *session_info;
2877         struct nlattr *nest;
2878
2879         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_GTP);
2880         if (!nest)
2881                 goto nla_put_failure;
2882
2883         session_info = (struct gtp_pdu_session_info *)&enc_opts->data[0];
2884
2885         if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE,
2886                        session_info->pdu_type))
2887                 goto nla_put_failure;
2888
2889         if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_GTP_QFI, session_info->qfi))
2890                 goto nla_put_failure;
2891
2892         nla_nest_end(skb, nest);
2893         return 0;
2894
2895 nla_put_failure:
2896         nla_nest_cancel(skb, nest);
2897         return -EMSGSIZE;
2898 }
2899
2900 static int fl_dump_key_ct(struct sk_buff *skb,
2901                           struct flow_dissector_key_ct *key,
2902                           struct flow_dissector_key_ct *mask)
2903 {
2904         if (IS_ENABLED(CONFIG_NF_CONNTRACK) &&
2905             fl_dump_key_val(skb, &key->ct_state, TCA_FLOWER_KEY_CT_STATE,
2906                             &mask->ct_state, TCA_FLOWER_KEY_CT_STATE_MASK,
2907                             sizeof(key->ct_state)))
2908                 goto nla_put_failure;
2909
2910         if (IS_ENABLED(CONFIG_NF_CONNTRACK_ZONES) &&
2911             fl_dump_key_val(skb, &key->ct_zone, TCA_FLOWER_KEY_CT_ZONE,
2912                             &mask->ct_zone, TCA_FLOWER_KEY_CT_ZONE_MASK,
2913                             sizeof(key->ct_zone)))
2914                 goto nla_put_failure;
2915
2916         if (IS_ENABLED(CONFIG_NF_CONNTRACK_MARK) &&
2917             fl_dump_key_val(skb, &key->ct_mark, TCA_FLOWER_KEY_CT_MARK,
2918                             &mask->ct_mark, TCA_FLOWER_KEY_CT_MARK_MASK,
2919                             sizeof(key->ct_mark)))
2920                 goto nla_put_failure;
2921
2922         if (IS_ENABLED(CONFIG_NF_CONNTRACK_LABELS) &&
2923             fl_dump_key_val(skb, &key->ct_labels, TCA_FLOWER_KEY_CT_LABELS,
2924                             &mask->ct_labels, TCA_FLOWER_KEY_CT_LABELS_MASK,
2925                             sizeof(key->ct_labels)))
2926                 goto nla_put_failure;
2927
2928         return 0;
2929
2930 nla_put_failure:
2931         return -EMSGSIZE;
2932 }
2933
2934 static int fl_dump_key_options(struct sk_buff *skb, int enc_opt_type,
2935                                struct flow_dissector_key_enc_opts *enc_opts)
2936 {
2937         struct nlattr *nest;
2938         int err;
2939
2940         if (!enc_opts->len)
2941                 return 0;
2942
2943         nest = nla_nest_start_noflag(skb, enc_opt_type);
2944         if (!nest)
2945                 goto nla_put_failure;
2946
2947         switch (enc_opts->dst_opt_type) {
2948         case TUNNEL_GENEVE_OPT:
2949                 err = fl_dump_key_geneve_opt(skb, enc_opts);
2950                 if (err)
2951                         goto nla_put_failure;
2952                 break;
2953         case TUNNEL_VXLAN_OPT:
2954                 err = fl_dump_key_vxlan_opt(skb, enc_opts);
2955                 if (err)
2956                         goto nla_put_failure;
2957                 break;
2958         case TUNNEL_ERSPAN_OPT:
2959                 err = fl_dump_key_erspan_opt(skb, enc_opts);
2960                 if (err)
2961                         goto nla_put_failure;
2962                 break;
2963         case TUNNEL_GTP_OPT:
2964                 err = fl_dump_key_gtp_opt(skb, enc_opts);
2965                 if (err)
2966                         goto nla_put_failure;
2967                 break;
2968         default:
2969                 goto nla_put_failure;
2970         }
2971         nla_nest_end(skb, nest);
2972         return 0;
2973
2974 nla_put_failure:
2975         nla_nest_cancel(skb, nest);
2976         return -EMSGSIZE;
2977 }
2978
2979 static int fl_dump_key_enc_opt(struct sk_buff *skb,
2980                                struct flow_dissector_key_enc_opts *key_opts,
2981                                struct flow_dissector_key_enc_opts *msk_opts)
2982 {
2983         int err;
2984
2985         err = fl_dump_key_options(skb, TCA_FLOWER_KEY_ENC_OPTS, key_opts);
2986         if (err)
2987                 return err;
2988
2989         return fl_dump_key_options(skb, TCA_FLOWER_KEY_ENC_OPTS_MASK, msk_opts);
2990 }
2991
2992 static int fl_dump_key(struct sk_buff *skb, struct net *net,
2993                        struct fl_flow_key *key, struct fl_flow_key *mask)
2994 {
2995         if (mask->meta.ingress_ifindex) {
2996                 struct net_device *dev;
2997
2998                 dev = __dev_get_by_index(net, key->meta.ingress_ifindex);
2999                 if (dev && nla_put_string(skb, TCA_FLOWER_INDEV, dev->name))
3000                         goto nla_put_failure;
3001         }
3002
3003         if (fl_dump_key_val(skb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
3004                             mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
3005                             sizeof(key->eth.dst)) ||
3006             fl_dump_key_val(skb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
3007                             mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
3008                             sizeof(key->eth.src)) ||
3009             fl_dump_key_val(skb, &key->basic.n_proto, TCA_FLOWER_KEY_ETH_TYPE,
3010                             &mask->basic.n_proto, TCA_FLOWER_UNSPEC,
3011                             sizeof(key->basic.n_proto)))
3012                 goto nla_put_failure;
3013
3014         if (mask->num_of_vlans.num_of_vlans) {
3015                 if (nla_put_u8(skb, TCA_FLOWER_KEY_NUM_OF_VLANS, key->num_of_vlans.num_of_vlans))
3016                         goto nla_put_failure;
3017         }
3018
3019         if (fl_dump_key_mpls(skb, &key->mpls, &mask->mpls))
3020                 goto nla_put_failure;
3021
3022         if (fl_dump_key_vlan(skb, TCA_FLOWER_KEY_VLAN_ID,
3023                              TCA_FLOWER_KEY_VLAN_PRIO, &key->vlan, &mask->vlan))
3024                 goto nla_put_failure;
3025
3026         if (fl_dump_key_vlan(skb, TCA_FLOWER_KEY_CVLAN_ID,
3027                              TCA_FLOWER_KEY_CVLAN_PRIO,
3028                              &key->cvlan, &mask->cvlan) ||
3029             (mask->cvlan.vlan_tpid &&
3030              nla_put_be16(skb, TCA_FLOWER_KEY_VLAN_ETH_TYPE,
3031                           key->cvlan.vlan_tpid)))
3032                 goto nla_put_failure;
3033
3034         if (mask->basic.n_proto) {
3035                 if (mask->cvlan.vlan_eth_type) {
3036                         if (nla_put_be16(skb, TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
3037                                          key->basic.n_proto))
3038                                 goto nla_put_failure;
3039                 } else if (mask->vlan.vlan_eth_type) {
3040                         if (nla_put_be16(skb, TCA_FLOWER_KEY_VLAN_ETH_TYPE,
3041                                          key->vlan.vlan_eth_type))
3042                                 goto nla_put_failure;
3043                 }
3044         }
3045
3046         if ((key->basic.n_proto == htons(ETH_P_IP) ||
3047              key->basic.n_proto == htons(ETH_P_IPV6)) &&
3048             (fl_dump_key_val(skb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
3049                             &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
3050                             sizeof(key->basic.ip_proto)) ||
3051             fl_dump_key_ip(skb, false, &key->ip, &mask->ip)))
3052                 goto nla_put_failure;
3053
3054         if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
3055             (fl_dump_key_val(skb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
3056                              &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
3057                              sizeof(key->ipv4.src)) ||
3058              fl_dump_key_val(skb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
3059                              &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
3060                              sizeof(key->ipv4.dst))))
3061                 goto nla_put_failure;
3062         else if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
3063                  (fl_dump_key_val(skb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
3064                                   &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
3065                                   sizeof(key->ipv6.src)) ||
3066                   fl_dump_key_val(skb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
3067                                   &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
3068                                   sizeof(key->ipv6.dst))))
3069                 goto nla_put_failure;
3070
3071         if (key->basic.ip_proto == IPPROTO_TCP &&
3072             (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
3073                              &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
3074                              sizeof(key->tp.src)) ||
3075              fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
3076                              &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
3077                              sizeof(key->tp.dst)) ||
3078              fl_dump_key_val(skb, &key->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS,
3079                              &mask->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS_MASK,
3080                              sizeof(key->tcp.flags))))
3081                 goto nla_put_failure;
3082         else if (key->basic.ip_proto == IPPROTO_UDP &&
3083                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
3084                                   &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
3085                                   sizeof(key->tp.src)) ||
3086                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
3087                                   &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
3088                                   sizeof(key->tp.dst))))
3089                 goto nla_put_failure;
3090         else if (key->basic.ip_proto == IPPROTO_SCTP &&
3091                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_SCTP_SRC,
3092                                   &mask->tp.src, TCA_FLOWER_KEY_SCTP_SRC_MASK,
3093                                   sizeof(key->tp.src)) ||
3094                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_SCTP_DST,
3095                                   &mask->tp.dst, TCA_FLOWER_KEY_SCTP_DST_MASK,
3096                                   sizeof(key->tp.dst))))
3097                 goto nla_put_failure;
3098         else if (key->basic.n_proto == htons(ETH_P_IP) &&
3099                  key->basic.ip_proto == IPPROTO_ICMP &&
3100                  (fl_dump_key_val(skb, &key->icmp.type,
3101                                   TCA_FLOWER_KEY_ICMPV4_TYPE, &mask->icmp.type,
3102                                   TCA_FLOWER_KEY_ICMPV4_TYPE_MASK,
3103                                   sizeof(key->icmp.type)) ||
3104                   fl_dump_key_val(skb, &key->icmp.code,
3105                                   TCA_FLOWER_KEY_ICMPV4_CODE, &mask->icmp.code,
3106                                   TCA_FLOWER_KEY_ICMPV4_CODE_MASK,
3107                                   sizeof(key->icmp.code))))
3108                 goto nla_put_failure;
3109         else if (key->basic.n_proto == htons(ETH_P_IPV6) &&
3110                  key->basic.ip_proto == IPPROTO_ICMPV6 &&
3111                  (fl_dump_key_val(skb, &key->icmp.type,
3112                                   TCA_FLOWER_KEY_ICMPV6_TYPE, &mask->icmp.type,
3113                                   TCA_FLOWER_KEY_ICMPV6_TYPE_MASK,
3114                                   sizeof(key->icmp.type)) ||
3115                   fl_dump_key_val(skb, &key->icmp.code,
3116                                   TCA_FLOWER_KEY_ICMPV6_CODE, &mask->icmp.code,
3117                                   TCA_FLOWER_KEY_ICMPV6_CODE_MASK,
3118                                   sizeof(key->icmp.code))))
3119                 goto nla_put_failure;
3120         else if ((key->basic.n_proto == htons(ETH_P_ARP) ||
3121                   key->basic.n_proto == htons(ETH_P_RARP)) &&
3122                  (fl_dump_key_val(skb, &key->arp.sip,
3123                                   TCA_FLOWER_KEY_ARP_SIP, &mask->arp.sip,
3124                                   TCA_FLOWER_KEY_ARP_SIP_MASK,
3125                                   sizeof(key->arp.sip)) ||
3126                   fl_dump_key_val(skb, &key->arp.tip,
3127                                   TCA_FLOWER_KEY_ARP_TIP, &mask->arp.tip,
3128                                   TCA_FLOWER_KEY_ARP_TIP_MASK,
3129                                   sizeof(key->arp.tip)) ||
3130                   fl_dump_key_val(skb, &key->arp.op,
3131                                   TCA_FLOWER_KEY_ARP_OP, &mask->arp.op,
3132                                   TCA_FLOWER_KEY_ARP_OP_MASK,
3133                                   sizeof(key->arp.op)) ||
3134                   fl_dump_key_val(skb, key->arp.sha, TCA_FLOWER_KEY_ARP_SHA,
3135                                   mask->arp.sha, TCA_FLOWER_KEY_ARP_SHA_MASK,
3136                                   sizeof(key->arp.sha)) ||
3137                   fl_dump_key_val(skb, key->arp.tha, TCA_FLOWER_KEY_ARP_THA,
3138                                   mask->arp.tha, TCA_FLOWER_KEY_ARP_THA_MASK,
3139                                   sizeof(key->arp.tha))))
3140                 goto nla_put_failure;
3141
3142         if ((key->basic.ip_proto == IPPROTO_TCP ||
3143              key->basic.ip_proto == IPPROTO_UDP ||
3144              key->basic.ip_proto == IPPROTO_SCTP) &&
3145              fl_dump_key_port_range(skb, key, mask))
3146                 goto nla_put_failure;
3147
3148         if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
3149             (fl_dump_key_val(skb, &key->enc_ipv4.src,
3150                             TCA_FLOWER_KEY_ENC_IPV4_SRC, &mask->enc_ipv4.src,
3151                             TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
3152                             sizeof(key->enc_ipv4.src)) ||
3153              fl_dump_key_val(skb, &key->enc_ipv4.dst,
3154                              TCA_FLOWER_KEY_ENC_IPV4_DST, &mask->enc_ipv4.dst,
3155                              TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
3156                              sizeof(key->enc_ipv4.dst))))
3157                 goto nla_put_failure;
3158         else if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
3159                  (fl_dump_key_val(skb, &key->enc_ipv6.src,
3160                             TCA_FLOWER_KEY_ENC_IPV6_SRC, &mask->enc_ipv6.src,
3161                             TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
3162                             sizeof(key->enc_ipv6.src)) ||
3163                  fl_dump_key_val(skb, &key->enc_ipv6.dst,
3164                                  TCA_FLOWER_KEY_ENC_IPV6_DST,
3165                                  &mask->enc_ipv6.dst,
3166                                  TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
3167                             sizeof(key->enc_ipv6.dst))))
3168                 goto nla_put_failure;
3169
3170         if (fl_dump_key_val(skb, &key->enc_key_id, TCA_FLOWER_KEY_ENC_KEY_ID,
3171                             &mask->enc_key_id, TCA_FLOWER_UNSPEC,
3172                             sizeof(key->enc_key_id)) ||
3173             fl_dump_key_val(skb, &key->enc_tp.src,
3174                             TCA_FLOWER_KEY_ENC_UDP_SRC_PORT,
3175                             &mask->enc_tp.src,
3176                             TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK,
3177                             sizeof(key->enc_tp.src)) ||
3178             fl_dump_key_val(skb, &key->enc_tp.dst,
3179                             TCA_FLOWER_KEY_ENC_UDP_DST_PORT,
3180                             &mask->enc_tp.dst,
3181                             TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK,
3182                             sizeof(key->enc_tp.dst)) ||
3183             fl_dump_key_ip(skb, true, &key->enc_ip, &mask->enc_ip) ||
3184             fl_dump_key_enc_opt(skb, &key->enc_opts, &mask->enc_opts))
3185                 goto nla_put_failure;
3186
3187         if (fl_dump_key_ct(skb, &key->ct, &mask->ct))
3188                 goto nla_put_failure;
3189
3190         if (fl_dump_key_flags(skb, key->control.flags, mask->control.flags))
3191                 goto nla_put_failure;
3192
3193         if (fl_dump_key_val(skb, &key->hash.hash, TCA_FLOWER_KEY_HASH,
3194                              &mask->hash.hash, TCA_FLOWER_KEY_HASH_MASK,
3195                              sizeof(key->hash.hash)))
3196                 goto nla_put_failure;
3197
3198         return 0;
3199
3200 nla_put_failure:
3201         return -EMSGSIZE;
3202 }
3203
3204 static int fl_dump(struct net *net, struct tcf_proto *tp, void *fh,
3205                    struct sk_buff *skb, struct tcmsg *t, bool rtnl_held)
3206 {
3207         struct cls_fl_filter *f = fh;
3208         struct nlattr *nest;
3209         struct fl_flow_key *key, *mask;
3210         bool skip_hw;
3211
3212         if (!f)
3213                 return skb->len;
3214
3215         t->tcm_handle = f->handle;
3216
3217         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
3218         if (!nest)
3219                 goto nla_put_failure;
3220
3221         spin_lock(&tp->lock);
3222
3223         if (f->res.classid &&
3224             nla_put_u32(skb, TCA_FLOWER_CLASSID, f->res.classid))
3225                 goto nla_put_failure_locked;
3226
3227         key = &f->key;
3228         mask = &f->mask->key;
3229         skip_hw = tc_skip_hw(f->flags);
3230
3231         if (fl_dump_key(skb, net, key, mask))
3232                 goto nla_put_failure_locked;
3233
3234         if (f->flags && nla_put_u32(skb, TCA_FLOWER_FLAGS, f->flags))
3235                 goto nla_put_failure_locked;
3236
3237         spin_unlock(&tp->lock);
3238
3239         if (!skip_hw)
3240                 fl_hw_update_stats(tp, f, rtnl_held);
3241
3242         if (nla_put_u32(skb, TCA_FLOWER_IN_HW_COUNT, f->in_hw_count))
3243                 goto nla_put_failure;
3244
3245         if (tcf_exts_dump(skb, &f->exts))
3246                 goto nla_put_failure;
3247
3248         nla_nest_end(skb, nest);
3249
3250         if (tcf_exts_dump_stats(skb, &f->exts) < 0)
3251                 goto nla_put_failure;
3252
3253         return skb->len;
3254
3255 nla_put_failure_locked:
3256         spin_unlock(&tp->lock);
3257 nla_put_failure:
3258         nla_nest_cancel(skb, nest);
3259         return -1;
3260 }
3261
3262 static int fl_terse_dump(struct net *net, struct tcf_proto *tp, void *fh,
3263                          struct sk_buff *skb, struct tcmsg *t, bool rtnl_held)
3264 {
3265         struct cls_fl_filter *f = fh;
3266         struct nlattr *nest;
3267         bool skip_hw;
3268
3269         if (!f)
3270                 return skb->len;
3271
3272         t->tcm_handle = f->handle;
3273
3274         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
3275         if (!nest)
3276                 goto nla_put_failure;
3277
3278         spin_lock(&tp->lock);
3279
3280         skip_hw = tc_skip_hw(f->flags);
3281
3282         if (f->flags && nla_put_u32(skb, TCA_FLOWER_FLAGS, f->flags))
3283                 goto nla_put_failure_locked;
3284
3285         spin_unlock(&tp->lock);
3286
3287         if (!skip_hw)
3288                 fl_hw_update_stats(tp, f, rtnl_held);
3289
3290         if (tcf_exts_terse_dump(skb, &f->exts))
3291                 goto nla_put_failure;
3292
3293         nla_nest_end(skb, nest);
3294
3295         return skb->len;
3296
3297 nla_put_failure_locked:
3298         spin_unlock(&tp->lock);
3299 nla_put_failure:
3300         nla_nest_cancel(skb, nest);
3301         return -1;
3302 }
3303
3304 static int fl_tmplt_dump(struct sk_buff *skb, struct net *net, void *tmplt_priv)
3305 {
3306         struct fl_flow_tmplt *tmplt = tmplt_priv;
3307         struct fl_flow_key *key, *mask;
3308         struct nlattr *nest;
3309
3310         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
3311         if (!nest)
3312                 goto nla_put_failure;
3313
3314         key = &tmplt->dummy_key;
3315         mask = &tmplt->mask;
3316
3317         if (fl_dump_key(skb, net, key, mask))
3318                 goto nla_put_failure;
3319
3320         nla_nest_end(skb, nest);
3321
3322         return skb->len;
3323
3324 nla_put_failure:
3325         nla_nest_cancel(skb, nest);
3326         return -EMSGSIZE;
3327 }
3328
3329 static void fl_bind_class(void *fh, u32 classid, unsigned long cl, void *q,
3330                           unsigned long base)
3331 {
3332         struct cls_fl_filter *f = fh;
3333
3334         if (f && f->res.classid == classid) {
3335                 if (cl)
3336                         __tcf_bind_filter(q, &f->res, base);
3337                 else
3338                         __tcf_unbind_filter(q, &f->res);
3339         }
3340 }
3341
3342 static bool fl_delete_empty(struct tcf_proto *tp)
3343 {
3344         struct cls_fl_head *head = fl_head_dereference(tp);
3345
3346         spin_lock(&tp->lock);
3347         tp->deleting = idr_is_empty(&head->handle_idr);
3348         spin_unlock(&tp->lock);
3349
3350         return tp->deleting;
3351 }
3352
3353 static struct tcf_proto_ops cls_fl_ops __read_mostly = {
3354         .kind           = "flower",
3355         .classify       = fl_classify,
3356         .init           = fl_init,
3357         .destroy        = fl_destroy,
3358         .get            = fl_get,
3359         .put            = fl_put,
3360         .change         = fl_change,
3361         .delete         = fl_delete,
3362         .delete_empty   = fl_delete_empty,
3363         .walk           = fl_walk,
3364         .reoffload      = fl_reoffload,
3365         .hw_add         = fl_hw_add,
3366         .hw_del         = fl_hw_del,
3367         .dump           = fl_dump,
3368         .terse_dump     = fl_terse_dump,
3369         .bind_class     = fl_bind_class,
3370         .tmplt_create   = fl_tmplt_create,
3371         .tmplt_destroy  = fl_tmplt_destroy,
3372         .tmplt_dump     = fl_tmplt_dump,
3373         .owner          = THIS_MODULE,
3374         .flags          = TCF_PROTO_OPS_DOIT_UNLOCKED,
3375 };
3376
3377 static int __init cls_fl_init(void)
3378 {
3379         return register_tcf_proto_ops(&cls_fl_ops);
3380 }
3381
3382 static void __exit cls_fl_exit(void)
3383 {
3384         unregister_tcf_proto_ops(&cls_fl_ops);
3385 }
3386
3387 module_init(cls_fl_init);
3388 module_exit(cls_fl_exit);
3389
3390 MODULE_AUTHOR("Jiri Pirko <jiri@resnulli.us>");
3391 MODULE_DESCRIPTION("Flower classifier");
3392 MODULE_LICENSE("GPL v2");