net/sched/cls_flower.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * net/sched/cls_flower.c               Flower classifier
   4  *
   5  * Copyright (c) 2015 Jiri Pirko <jiri@resnulli.us>
   6  */
   7
   8 #include <linux/kernel.h>
   9 #include <linux/init.h>
  10 #include <linux/module.h>
  11 #include <linux/rhashtable.h>
  12 #include <linux/workqueue.h>
  13 #include <linux/refcount.h>
  14
  15 #include <linux/if_ether.h>
  16 #include <linux/in6.h>
  17 #include <linux/ip.h>
  18 #include <linux/mpls.h>
  19 #include <linux/ppp_defs.h>
  20
  21 #include <net/sch_generic.h>
  22 #include <net/pkt_cls.h>
  23 #include <net/pkt_sched.h>
  24 #include <net/ip.h>
  25 #include <net/flow_dissector.h>
  26 #include <net/geneve.h>
  27 #include <net/vxlan.h>
  28 #include <net/erspan.h>
  29 #include <net/gtp.h>
  30 #include <net/tc_wrapper.h>
  31
  32 #include <net/dst.h>
  33 #include <net/dst_metadata.h>
  34
  35 #include <uapi/linux/netfilter/nf_conntrack_common.h>
  36
  37 #define TCA_FLOWER_KEY_CT_FLAGS_MAX \
  38                 ((__TCA_FLOWER_KEY_CT_FLAGS_MAX - 1) << 1)
  39 #define TCA_FLOWER_KEY_CT_FLAGS_MASK \
  40                 (TCA_FLOWER_KEY_CT_FLAGS_MAX - 1)
  41
  42 struct fl_flow_key {
  43         struct flow_dissector_key_meta meta;
  44         struct flow_dissector_key_control control;
  45         struct flow_dissector_key_control enc_control;
  46         struct flow_dissector_key_basic basic;
  47         struct flow_dissector_key_eth_addrs eth;
  48         struct flow_dissector_key_vlan vlan;
  49         struct flow_dissector_key_vlan cvlan;
  50         union {
  51                 struct flow_dissector_key_ipv4_addrs ipv4;
  52                 struct flow_dissector_key_ipv6_addrs ipv6;
  53         };
  54         struct flow_dissector_key_ports tp;
  55         struct flow_dissector_key_icmp icmp;
  56         struct flow_dissector_key_arp arp;
  57         struct flow_dissector_key_keyid enc_key_id;
  58         union {
  59                 struct flow_dissector_key_ipv4_addrs enc_ipv4;
  60                 struct flow_dissector_key_ipv6_addrs enc_ipv6;
  61         };
  62         struct flow_dissector_key_ports enc_tp;
  63         struct flow_dissector_key_mpls mpls;
  64         struct flow_dissector_key_tcp tcp;
  65         struct flow_dissector_key_ip ip;
  66         struct flow_dissector_key_ip enc_ip;
  67         struct flow_dissector_key_enc_opts enc_opts;
  68         struct flow_dissector_key_ports_range tp_range;
  69         struct flow_dissector_key_ct ct;
  70         struct flow_dissector_key_hash hash;
  71         struct flow_dissector_key_num_of_vlans num_of_vlans;
  72         struct flow_dissector_key_pppoe pppoe;
  73         struct flow_dissector_key_l2tpv3 l2tpv3;
  74 } __aligned(BITS_PER_LONG / 8); /* Ensure that we can do comparisons as longs. */
  75
  76 struct fl_flow_mask_range {
  77         unsigned short int start;
  78         unsigned short int end;
  79 };
  80
  81 struct fl_flow_mask {
  82         struct fl_flow_key key;
  83         struct fl_flow_mask_range range;
  84         u32 flags;
  85         struct rhash_head ht_node;
  86         struct rhashtable ht;
  87         struct rhashtable_params filter_ht_params;
  88         struct flow_dissector dissector;
  89         struct list_head filters;
  90         struct rcu_work rwork;
  91         struct list_head list;
  92         refcount_t refcnt;
  93 };
  94
  95 struct fl_flow_tmplt {
  96         struct fl_flow_key dummy_key;
  97         struct fl_flow_key mask;
  98         struct flow_dissector dissector;
  99         struct tcf_chain *chain;
 100 };
 101
 102 struct cls_fl_head {
 103         struct rhashtable ht;
 104         spinlock_t masks_lock; /* Protect masks list */
 105         struct list_head masks;
 106         struct list_head hw_filters;
 107         struct rcu_work rwork;
 108         struct idr handle_idr;
 109 };
 110
 111 struct cls_fl_filter {
 112         struct fl_flow_mask *mask;
 113         struct rhash_head ht_node;
 114         struct fl_flow_key mkey;
 115         struct tcf_exts exts;
 116         struct tcf_result res;
 117         struct fl_flow_key key;
 118         struct list_head list;
 119         struct list_head hw_list;
 120         u32 handle;
 121         u32 flags;
 122         u32 in_hw_count;
 123         struct rcu_work rwork;
 124         struct net_device *hw_dev;
 125         /* Flower classifier is unlocked, which means that its reference counter
 126          * can be changed concurrently without any kind of external
 127          * synchronization. Use atomic reference counter to be concurrency-safe.
 128          */
 129         refcount_t refcnt;
 130         bool deleted;
 131 };
 132
 133 static const struct rhashtable_params mask_ht_params = {
 134         .key_offset = offsetof(struct fl_flow_mask, key),
 135         .key_len = sizeof(struct fl_flow_key),
 136         .head_offset = offsetof(struct fl_flow_mask, ht_node),
 137         .automatic_shrinking = true,
 138 };
 139
 140 static unsigned short int fl_mask_range(const struct fl_flow_mask *mask)
 141 {
 142         return mask->range.end - mask->range.start;
 143 }
 144
 145 static void fl_mask_update_range(struct fl_flow_mask *mask)
 146 {
 147         const u8 *bytes = (const u8 *) &mask->key;
 148         size_t size = sizeof(mask->key);
 149         size_t i, first = 0, last;
 150
 151         for (i = 0; i < size; i++) {
 152                 if (bytes[i]) {
 153                         first = i;
 154                         break;
 155                 }
 156         }
 157         last = first;
 158         for (i = size - 1; i != first; i--) {
 159                 if (bytes[i]) {
 160                         last = i;
 161                         break;
 162                 }
 163         }
 164         mask->range.start = rounddown(first, sizeof(long));
 165         mask->range.end = roundup(last + 1, sizeof(long));
 166 }
 167
 168 static void *fl_key_get_start(struct fl_flow_key *key,
 169                               const struct fl_flow_mask *mask)
 170 {
 171         return (u8 *) key + mask->range.start;
 172 }
 173
 174 static void fl_set_masked_key(struct fl_flow_key *mkey, struct fl_flow_key *key,
 175                               struct fl_flow_mask *mask)
 176 {
 177         const long *lkey = fl_key_get_start(key, mask);
 178         const long *lmask = fl_key_get_start(&mask->key, mask);
 179         long *lmkey = fl_key_get_start(mkey, mask);
 180         int i;
 181
 182         for (i = 0; i < fl_mask_range(mask); i += sizeof(long))
 183                 *lmkey++ = *lkey++ & *lmask++;
 184 }
 185
 186 static bool fl_mask_fits_tmplt(struct fl_flow_tmplt *tmplt,
 187                                struct fl_flow_mask *mask)
 188 {
 189         const long *lmask = fl_key_get_start(&mask->key, mask);
 190         const long *ltmplt;
 191         int i;
 192
 193         if (!tmplt)
 194                 return true;
 195         ltmplt = fl_key_get_start(&tmplt->mask, mask);
 196         for (i = 0; i < fl_mask_range(mask); i += sizeof(long)) {
 197                 if (~*ltmplt++ & *lmask++)
 198                         return false;
 199         }
 200         return true;
 201 }
 202
 203 static void fl_clear_masked_range(struct fl_flow_key *key,
 204                                   struct fl_flow_mask *mask)
 205 {
 206         memset(fl_key_get_start(key, mask), 0, fl_mask_range(mask));
 207 }
 208
 209 static bool fl_range_port_dst_cmp(struct cls_fl_filter *filter,
 210                                   struct fl_flow_key *key,
 211                                   struct fl_flow_key *mkey)
 212 {
 213         u16 min_mask, max_mask, min_val, max_val;
 214
 215         min_mask = ntohs(filter->mask->key.tp_range.tp_min.dst);
 216         max_mask = ntohs(filter->mask->key.tp_range.tp_max.dst);
 217         min_val = ntohs(filter->key.tp_range.tp_min.dst);
 218         max_val = ntohs(filter->key.tp_range.tp_max.dst);
 219
 220         if (min_mask && max_mask) {
 221                 if (ntohs(key->tp_range.tp.dst) < min_val ||
 222                     ntohs(key->tp_range.tp.dst) > max_val)
 223                         return false;
 224
 225                 /* skb does not have min and max values */
 226                 mkey->tp_range.tp_min.dst = filter->mkey.tp_range.tp_min.dst;
 227                 mkey->tp_range.tp_max.dst = filter->mkey.tp_range.tp_max.dst;
 228         }
 229         return true;
 230 }
 231
 232 static bool fl_range_port_src_cmp(struct cls_fl_filter *filter,
 233                                   struct fl_flow_key *key,
 234                                   struct fl_flow_key *mkey)
 235 {
 236         u16 min_mask, max_mask, min_val, max_val;
 237
 238         min_mask = ntohs(filter->mask->key.tp_range.tp_min.src);
 239         max_mask = ntohs(filter->mask->key.tp_range.tp_max.src);
 240         min_val = ntohs(filter->key.tp_range.tp_min.src);
 241         max_val = ntohs(filter->key.tp_range.tp_max.src);
 242
 243         if (min_mask && max_mask) {
 244                 if (ntohs(key->tp_range.tp.src) < min_val ||
 245                     ntohs(key->tp_range.tp.src) > max_val)
 246                         return false;
 247
 248                 /* skb does not have min and max values */
 249                 mkey->tp_range.tp_min.src = filter->mkey.tp_range.tp_min.src;
 250                 mkey->tp_range.tp_max.src = filter->mkey.tp_range.tp_max.src;
 251         }
 252         return true;
 253 }
 254
 255 static struct cls_fl_filter *__fl_lookup(struct fl_flow_mask *mask,
 256                                          struct fl_flow_key *mkey)
 257 {
 258         return rhashtable_lookup_fast(&mask->ht, fl_key_get_start(mkey, mask),
 259                                       mask->filter_ht_params);
 260 }
 261
 262 static struct cls_fl_filter *fl_lookup_range(struct fl_flow_mask *mask,
 263                                              struct fl_flow_key *mkey,
 264                                              struct fl_flow_key *key)
 265 {
 266         struct cls_fl_filter *filter, *f;
 267
 268         list_for_each_entry_rcu(filter, &mask->filters, list) {
 269                 if (!fl_range_port_dst_cmp(filter, key, mkey))
 270                         continue;
 271
 272                 if (!fl_range_port_src_cmp(filter, key, mkey))
 273                         continue;
 274
 275                 f = __fl_lookup(mask, mkey);
 276                 if (f)
 277                         return f;
 278         }
 279         return NULL;
 280 }
 281
 282 static noinline_for_stack
 283 struct cls_fl_filter *fl_mask_lookup(struct fl_flow_mask *mask, struct fl_flow_key *key)
 284 {
 285         struct fl_flow_key mkey;
 286
 287         fl_set_masked_key(&mkey, key, mask);
 288         if ((mask->flags & TCA_FLOWER_MASK_FLAGS_RANGE))
 289                 return fl_lookup_range(mask, &mkey, key);
 290
 291         return __fl_lookup(mask, &mkey);
 292 }
 293
 294 static u16 fl_ct_info_to_flower_map[] = {
 295         [IP_CT_ESTABLISHED] =           TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 296                                         TCA_FLOWER_KEY_CT_FLAGS_ESTABLISHED,
 297         [IP_CT_RELATED] =               TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 298                                         TCA_FLOWER_KEY_CT_FLAGS_RELATED,
 299         [IP_CT_ESTABLISHED_REPLY] =     TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 300                                         TCA_FLOWER_KEY_CT_FLAGS_ESTABLISHED |
 301                                         TCA_FLOWER_KEY_CT_FLAGS_REPLY,
 302         [IP_CT_RELATED_REPLY] =         TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 303                                         TCA_FLOWER_KEY_CT_FLAGS_RELATED |
 304                                         TCA_FLOWER_KEY_CT_FLAGS_REPLY,
 305         [IP_CT_NEW] =                   TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
 306                                         TCA_FLOWER_KEY_CT_FLAGS_NEW,
 307 };
 308
 309 TC_INDIRECT_SCOPE int fl_classify(struct sk_buff *skb,
 310                                   const struct tcf_proto *tp,
 311                                   struct tcf_result *res)
 312 {
 313         struct cls_fl_head *head = rcu_dereference_bh(tp->root);
 314         bool post_ct = tc_skb_cb(skb)->post_ct;
 315         u16 zone = tc_skb_cb(skb)->zone;
 316         struct fl_flow_key skb_key;
 317         struct fl_flow_mask *mask;
 318         struct cls_fl_filter *f;
 319
 320         list_for_each_entry_rcu(mask, &head->masks, list) {
 321                 flow_dissector_init_keys(&skb_key.control, &skb_key.basic);
 322                 fl_clear_masked_range(&skb_key, mask);
 323
 324                 skb_flow_dissect_meta(skb, &mask->dissector, &skb_key);
 325                 /* skb_flow_dissect() does not set n_proto in case an unknown
 326                  * protocol, so do it rather here.
 327                  */
 328                 skb_key.basic.n_proto = skb_protocol(skb, false);
 329                 skb_flow_dissect_tunnel_info(skb, &mask->dissector, &skb_key);
 330                 skb_flow_dissect_ct(skb, &mask->dissector, &skb_key,
 331                                     fl_ct_info_to_flower_map,
 332                                     ARRAY_SIZE(fl_ct_info_to_flower_map),
 333                                     post_ct, zone);
 334                 skb_flow_dissect_hash(skb, &mask->dissector, &skb_key);
 335                 skb_flow_dissect(skb, &mask->dissector, &skb_key,
 336                                  FLOW_DISSECTOR_F_STOP_BEFORE_ENCAP);
 337
 338                 f = fl_mask_lookup(mask, &skb_key);
 339                 if (f && !tc_skip_sw(f->flags)) {
 340                         *res = f->res;
 341                         return tcf_exts_exec(skb, &f->exts, res);
 342                 }
 343         }
 344         return -1;
 345 }
 346
 347 static int fl_init(struct tcf_proto *tp)
 348 {
 349         struct cls_fl_head *head;
 350
 351         head = kzalloc(sizeof(*head), GFP_KERNEL);
 352         if (!head)
 353                 return -ENOBUFS;
 354
 355         spin_lock_init(&head->masks_lock);
 356         INIT_LIST_HEAD_RCU(&head->masks);
 357         INIT_LIST_HEAD(&head->hw_filters);
 358         rcu_assign_pointer(tp->root, head);
 359         idr_init(&head->handle_idr);
 360
 361         return rhashtable_init(&head->ht, &mask_ht_params);
 362 }
 363
 364 static void fl_mask_free(struct fl_flow_mask *mask, bool mask_init_done)
 365 {
 366         /* temporary masks don't have their filters list and ht initialized */
 367         if (mask_init_done) {
 368                 WARN_ON(!list_empty(&mask->filters));
 369                 rhashtable_destroy(&mask->ht);
 370         }
 371         kfree(mask);
 372 }
 373
 374 static void fl_mask_free_work(struct work_struct *work)
 375 {
 376         struct fl_flow_mask *mask = container_of(to_rcu_work(work),
 377                                                  struct fl_flow_mask, rwork);
 378
 379         fl_mask_free(mask, true);
 380 }
 381
 382 static void fl_uninit_mask_free_work(struct work_struct *work)
 383 {
 384         struct fl_flow_mask *mask = container_of(to_rcu_work(work),
 385                                                  struct fl_flow_mask, rwork);
 386
 387         fl_mask_free(mask, false);
 388 }
 389
 390 static bool fl_mask_put(struct cls_fl_head *head, struct fl_flow_mask *mask)
 391 {
 392         if (!refcount_dec_and_test(&mask->refcnt))
 393                 return false;
 394
 395         rhashtable_remove_fast(&head->ht, &mask->ht_node, mask_ht_params);
 396
 397         spin_lock(&head->masks_lock);
 398         list_del_rcu(&mask->list);
 399         spin_unlock(&head->masks_lock);
 400
 401         tcf_queue_work(&mask->rwork, fl_mask_free_work);
 402
 403         return true;
 404 }
 405
 406 static struct cls_fl_head *fl_head_dereference(struct tcf_proto *tp)
 407 {
 408         /* Flower classifier only changes root pointer during init and destroy.
 409          * Users must obtain reference to tcf_proto instance before calling its
 410          * API, so tp->root pointer is protected from concurrent call to
 411          * fl_destroy() by reference counting.
 412          */
 413         return rcu_dereference_raw(tp->root);
 414 }
 415
 416 static void __fl_destroy_filter(struct cls_fl_filter *f)
 417 {
 418         tcf_exts_destroy(&f->exts);
 419         tcf_exts_put_net(&f->exts);
 420         kfree(f);
 421 }
 422
 423 static void fl_destroy_filter_work(struct work_struct *work)
 424 {
 425         struct cls_fl_filter *f = container_of(to_rcu_work(work),
 426                                         struct cls_fl_filter, rwork);
 427
 428         __fl_destroy_filter(f);
 429 }
 430
 431 static void fl_hw_destroy_filter(struct tcf_proto *tp, struct cls_fl_filter *f,
 432                                  bool rtnl_held, struct netlink_ext_ack *extack)
 433 {
 434         struct tcf_block *block = tp->chain->block;
 435         struct flow_cls_offload cls_flower = {};
 436
 437         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, extack);
 438         cls_flower.command = FLOW_CLS_DESTROY;
 439         cls_flower.cookie = (unsigned long) f;
 440
 441         tc_setup_cb_destroy(block, tp, TC_SETUP_CLSFLOWER, &cls_flower, false,
 442                             &f->flags, &f->in_hw_count, rtnl_held);
 443
 444 }
 445
 446 static int fl_hw_replace_filter(struct tcf_proto *tp,
 447                                 struct cls_fl_filter *f, bool rtnl_held,
 448                                 struct netlink_ext_ack *extack)
 449 {
 450         struct tcf_block *block = tp->chain->block;
 451         struct flow_cls_offload cls_flower = {};
 452         bool skip_sw = tc_skip_sw(f->flags);
 453         int err = 0;
 454
 455         cls_flower.rule = flow_rule_alloc(tcf_exts_num_actions(&f->exts));
 456         if (!cls_flower.rule)
 457                 return -ENOMEM;
 458
 459         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, extack);
 460         cls_flower.command = FLOW_CLS_REPLACE;
 461         cls_flower.cookie = (unsigned long) f;
 462         cls_flower.rule->match.dissector = &f->mask->dissector;
 463         cls_flower.rule->match.mask = &f->mask->key;
 464         cls_flower.rule->match.key = &f->mkey;
 465         cls_flower.classid = f->res.classid;
 466
 467         err = tc_setup_offload_action(&cls_flower.rule->action, &f->exts,
 468                                       cls_flower.common.extack);
 469         if (err) {
 470                 kfree(cls_flower.rule);
 471
 472                 return skip_sw ? err : 0;
 473         }
 474
 475         err = tc_setup_cb_add(block, tp, TC_SETUP_CLSFLOWER, &cls_flower,
 476                               skip_sw, &f->flags, &f->in_hw_count, rtnl_held);
 477         tc_cleanup_offload_action(&cls_flower.rule->action);
 478         kfree(cls_flower.rule);
 479
 480         if (err) {
 481                 fl_hw_destroy_filter(tp, f, rtnl_held, NULL);
 482                 return err;
 483         }
 484
 485         if (skip_sw && !(f->flags & TCA_CLS_FLAGS_IN_HW))
 486                 return -EINVAL;
 487
 488         return 0;
 489 }
 490
 491 static void fl_hw_update_stats(struct tcf_proto *tp, struct cls_fl_filter *f,
 492                                bool rtnl_held)
 493 {
 494         struct tcf_block *block = tp->chain->block;
 495         struct flow_cls_offload cls_flower = {};
 496
 497         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, NULL);
 498         cls_flower.command = FLOW_CLS_STATS;
 499         cls_flower.cookie = (unsigned long) f;
 500         cls_flower.classid = f->res.classid;
 501
 502         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false,
 503                          rtnl_held);
 504
 505         tcf_exts_hw_stats_update(&f->exts, &cls_flower.stats, cls_flower.use_act_stats);
 506 }
 507
 508 static void __fl_put(struct cls_fl_filter *f)
 509 {
 510         if (!refcount_dec_and_test(&f->refcnt))
 511                 return;
 512
 513         if (tcf_exts_get_net(&f->exts))
 514                 tcf_queue_work(&f->rwork, fl_destroy_filter_work);
 515         else
 516                 __fl_destroy_filter(f);
 517 }
 518
 519 static struct cls_fl_filter *__fl_get(struct cls_fl_head *head, u32 handle)
 520 {
 521         struct cls_fl_filter *f;
 522
 523         rcu_read_lock();
 524         f = idr_find(&head->handle_idr, handle);
 525         if (f && !refcount_inc_not_zero(&f->refcnt))
 526                 f = NULL;
 527         rcu_read_unlock();
 528
 529         return f;
 530 }
 531
 532 static struct tcf_exts *fl_get_exts(const struct tcf_proto *tp, u32 handle)
 533 {
 534         struct cls_fl_head *head = rcu_dereference_bh(tp->root);
 535         struct cls_fl_filter *f;
 536
 537         f = idr_find(&head->handle_idr, handle);
 538         return f ? &f->exts : NULL;
 539 }
 540
 541 static int __fl_delete(struct tcf_proto *tp, struct cls_fl_filter *f,
 542                        bool *last, bool rtnl_held,
 543                        struct netlink_ext_ack *extack)
 544 {
 545         struct cls_fl_head *head = fl_head_dereference(tp);
 546
 547         *last = false;
 548
 549         spin_lock(&tp->lock);
 550         if (f->deleted) {
 551                 spin_unlock(&tp->lock);
 552                 return -ENOENT;
 553         }
 554
 555         f->deleted = true;
 556         rhashtable_remove_fast(&f->mask->ht, &f->ht_node,
 557                                f->mask->filter_ht_params);
 558         idr_remove(&head->handle_idr, f->handle);
 559         list_del_rcu(&f->list);
 560         spin_unlock(&tp->lock);
 561
 562         *last = fl_mask_put(head, f->mask);
 563         if (!tc_skip_hw(f->flags))
 564                 fl_hw_destroy_filter(tp, f, rtnl_held, extack);
 565         tcf_unbind_filter(tp, &f->res);
 566         __fl_put(f);
 567
 568         return 0;
 569 }
 570
 571 static void fl_destroy_sleepable(struct work_struct *work)
 572 {
 573         struct cls_fl_head *head = container_of(to_rcu_work(work),
 574                                                 struct cls_fl_head,
 575                                                 rwork);
 576
 577         rhashtable_destroy(&head->ht);
 578         kfree(head);
 579         module_put(THIS_MODULE);
 580 }
 581
 582 static void fl_destroy(struct tcf_proto *tp, bool rtnl_held,
 583                        struct netlink_ext_ack *extack)
 584 {
 585         struct cls_fl_head *head = fl_head_dereference(tp);
 586         struct fl_flow_mask *mask, *next_mask;
 587         struct cls_fl_filter *f, *next;
 588         bool last;
 589
 590         list_for_each_entry_safe(mask, next_mask, &head->masks, list) {
 591                 list_for_each_entry_safe(f, next, &mask->filters, list) {
 592                         __fl_delete(tp, f, &last, rtnl_held, extack);
 593                         if (last)
 594                                 break;
 595                 }
 596         }
 597         idr_destroy(&head->handle_idr);
 598
 599         __module_get(THIS_MODULE);
 600         tcf_queue_work(&head->rwork, fl_destroy_sleepable);
 601 }
 602
 603 static void fl_put(struct tcf_proto *tp, void *arg)
 604 {
 605         struct cls_fl_filter *f = arg;
 606
 607         __fl_put(f);
 608 }
 609
 610 static void *fl_get(struct tcf_proto *tp, u32 handle)
 611 {
 612         struct cls_fl_head *head = fl_head_dereference(tp);
 613
 614         return __fl_get(head, handle);
 615 }
 616
 617 static const struct nla_policy fl_policy[TCA_FLOWER_MAX + 1] = {
 618         [TCA_FLOWER_UNSPEC]             = { .type = NLA_UNSPEC },
 619         [TCA_FLOWER_CLASSID]            = { .type = NLA_U32 },
 620         [TCA_FLOWER_INDEV]              = { .type = NLA_STRING,
 621                                             .len = IFNAMSIZ },
 622         [TCA_FLOWER_KEY_ETH_DST]        = { .len = ETH_ALEN },
 623         [TCA_FLOWER_KEY_ETH_DST_MASK]   = { .len = ETH_ALEN },
 624         [TCA_FLOWER_KEY_ETH_SRC]        = { .len = ETH_ALEN },
 625         [TCA_FLOWER_KEY_ETH_SRC_MASK]   = { .len = ETH_ALEN },
 626         [TCA_FLOWER_KEY_ETH_TYPE]       = { .type = NLA_U16 },
 627         [TCA_FLOWER_KEY_IP_PROTO]       = { .type = NLA_U8 },
 628         [TCA_FLOWER_KEY_IPV4_SRC]       = { .type = NLA_U32 },
 629         [TCA_FLOWER_KEY_IPV4_SRC_MASK]  = { .type = NLA_U32 },
 630         [TCA_FLOWER_KEY_IPV4_DST]       = { .type = NLA_U32 },
 631         [TCA_FLOWER_KEY_IPV4_DST_MASK]  = { .type = NLA_U32 },
 632         [TCA_FLOWER_KEY_IPV6_SRC]       = { .len = sizeof(struct in6_addr) },
 633         [TCA_FLOWER_KEY_IPV6_SRC_MASK]  = { .len = sizeof(struct in6_addr) },
 634         [TCA_FLOWER_KEY_IPV6_DST]       = { .len = sizeof(struct in6_addr) },
 635         [TCA_FLOWER_KEY_IPV6_DST_MASK]  = { .len = sizeof(struct in6_addr) },
 636         [TCA_FLOWER_KEY_TCP_SRC]        = { .type = NLA_U16 },
 637         [TCA_FLOWER_KEY_TCP_DST]        = { .type = NLA_U16 },
 638         [TCA_FLOWER_KEY_UDP_SRC]        = { .type = NLA_U16 },
 639         [TCA_FLOWER_KEY_UDP_DST]        = { .type = NLA_U16 },
 640         [TCA_FLOWER_KEY_VLAN_ID]        = { .type = NLA_U16 },
 641         [TCA_FLOWER_KEY_VLAN_PRIO]      = { .type = NLA_U8 },
 642         [TCA_FLOWER_KEY_VLAN_ETH_TYPE]  = { .type = NLA_U16 },
 643         [TCA_FLOWER_KEY_ENC_KEY_ID]     = { .type = NLA_U32 },
 644         [TCA_FLOWER_KEY_ENC_IPV4_SRC]   = { .type = NLA_U32 },
 645         [TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK] = { .type = NLA_U32 },
 646         [TCA_FLOWER_KEY_ENC_IPV4_DST]   = { .type = NLA_U32 },
 647         [TCA_FLOWER_KEY_ENC_IPV4_DST_MASK] = { .type = NLA_U32 },
 648         [TCA_FLOWER_KEY_ENC_IPV6_SRC]   = { .len = sizeof(struct in6_addr) },
 649         [TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK] = { .len = sizeof(struct in6_addr) },
 650         [TCA_FLOWER_KEY_ENC_IPV6_DST]   = { .len = sizeof(struct in6_addr) },
 651         [TCA_FLOWER_KEY_ENC_IPV6_DST_MASK] = { .len = sizeof(struct in6_addr) },
 652         [TCA_FLOWER_KEY_TCP_SRC_MASK]   = { .type = NLA_U16 },
 653         [TCA_FLOWER_KEY_TCP_DST_MASK]   = { .type = NLA_U16 },
 654         [TCA_FLOWER_KEY_UDP_SRC_MASK]   = { .type = NLA_U16 },
 655         [TCA_FLOWER_KEY_UDP_DST_MASK]   = { .type = NLA_U16 },
 656         [TCA_FLOWER_KEY_SCTP_SRC_MASK]  = { .type = NLA_U16 },
 657         [TCA_FLOWER_KEY_SCTP_DST_MASK]  = { .type = NLA_U16 },
 658         [TCA_FLOWER_KEY_SCTP_SRC]       = { .type = NLA_U16 },
 659         [TCA_FLOWER_KEY_SCTP_DST]       = { .type = NLA_U16 },
 660         [TCA_FLOWER_KEY_ENC_UDP_SRC_PORT]       = { .type = NLA_U16 },
 661         [TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK]  = { .type = NLA_U16 },
 662         [TCA_FLOWER_KEY_ENC_UDP_DST_PORT]       = { .type = NLA_U16 },
 663         [TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK]  = { .type = NLA_U16 },
 664         [TCA_FLOWER_KEY_FLAGS]          = { .type = NLA_U32 },
 665         [TCA_FLOWER_KEY_FLAGS_MASK]     = { .type = NLA_U32 },
 666         [TCA_FLOWER_KEY_ICMPV4_TYPE]    = { .type = NLA_U8 },
 667         [TCA_FLOWER_KEY_ICMPV4_TYPE_MASK] = { .type = NLA_U8 },
 668         [TCA_FLOWER_KEY_ICMPV4_CODE]    = { .type = NLA_U8 },
 669         [TCA_FLOWER_KEY_ICMPV4_CODE_MASK] = { .type = NLA_U8 },
 670         [TCA_FLOWER_KEY_ICMPV6_TYPE]    = { .type = NLA_U8 },
 671         [TCA_FLOWER_KEY_ICMPV6_TYPE_MASK] = { .type = NLA_U8 },
 672         [TCA_FLOWER_KEY_ICMPV6_CODE]    = { .type = NLA_U8 },
 673         [TCA_FLOWER_KEY_ICMPV6_CODE_MASK] = { .type = NLA_U8 },
 674         [TCA_FLOWER_KEY_ARP_SIP]        = { .type = NLA_U32 },
 675         [TCA_FLOWER_KEY_ARP_SIP_MASK]   = { .type = NLA_U32 },
 676         [TCA_FLOWER_KEY_ARP_TIP]        = { .type = NLA_U32 },
 677         [TCA_FLOWER_KEY_ARP_TIP_MASK]   = { .type = NLA_U32 },
 678         [TCA_FLOWER_KEY_ARP_OP]         = { .type = NLA_U8 },
 679         [TCA_FLOWER_KEY_ARP_OP_MASK]    = { .type = NLA_U8 },
 680         [TCA_FLOWER_KEY_ARP_SHA]        = { .len = ETH_ALEN },
 681         [TCA_FLOWER_KEY_ARP_SHA_MASK]   = { .len = ETH_ALEN },
 682         [TCA_FLOWER_KEY_ARP_THA]        = { .len = ETH_ALEN },
 683         [TCA_FLOWER_KEY_ARP_THA_MASK]   = { .len = ETH_ALEN },
 684         [TCA_FLOWER_KEY_MPLS_TTL]       = { .type = NLA_U8 },
 685         [TCA_FLOWER_KEY_MPLS_BOS]       = { .type = NLA_U8 },
 686         [TCA_FLOWER_KEY_MPLS_TC]        = { .type = NLA_U8 },
 687         [TCA_FLOWER_KEY_MPLS_LABEL]     = { .type = NLA_U32 },
 688         [TCA_FLOWER_KEY_MPLS_OPTS]      = { .type = NLA_NESTED },
 689         [TCA_FLOWER_KEY_TCP_FLAGS]      = { .type = NLA_U16 },
 690         [TCA_FLOWER_KEY_TCP_FLAGS_MASK] = { .type = NLA_U16 },
 691         [TCA_FLOWER_KEY_IP_TOS]         = { .type = NLA_U8 },
 692         [TCA_FLOWER_KEY_IP_TOS_MASK]    = { .type = NLA_U8 },
 693         [TCA_FLOWER_KEY_IP_TTL]         = { .type = NLA_U8 },
 694         [TCA_FLOWER_KEY_IP_TTL_MASK]    = { .type = NLA_U8 },
 695         [TCA_FLOWER_KEY_CVLAN_ID]       = { .type = NLA_U16 },
 696         [TCA_FLOWER_KEY_CVLAN_PRIO]     = { .type = NLA_U8 },
 697         [TCA_FLOWER_KEY_CVLAN_ETH_TYPE] = { .type = NLA_U16 },
 698         [TCA_FLOWER_KEY_ENC_IP_TOS]     = { .type = NLA_U8 },
 699         [TCA_FLOWER_KEY_ENC_IP_TOS_MASK] = { .type = NLA_U8 },
 700         [TCA_FLOWER_KEY_ENC_IP_TTL]      = { .type = NLA_U8 },
 701         [TCA_FLOWER_KEY_ENC_IP_TTL_MASK] = { .type = NLA_U8 },
 702         [TCA_FLOWER_KEY_ENC_OPTS]       = { .type = NLA_NESTED },
 703         [TCA_FLOWER_KEY_ENC_OPTS_MASK]  = { .type = NLA_NESTED },
 704         [TCA_FLOWER_KEY_CT_STATE]       =
 705                 NLA_POLICY_MASK(NLA_U16, TCA_FLOWER_KEY_CT_FLAGS_MASK),
 706         [TCA_FLOWER_KEY_CT_STATE_MASK]  =
 707                 NLA_POLICY_MASK(NLA_U16, TCA_FLOWER_KEY_CT_FLAGS_MASK),
 708         [TCA_FLOWER_KEY_CT_ZONE]        = { .type = NLA_U16 },
 709         [TCA_FLOWER_KEY_CT_ZONE_MASK]   = { .type = NLA_U16 },
 710         [TCA_FLOWER_KEY_CT_MARK]        = { .type = NLA_U32 },
 711         [TCA_FLOWER_KEY_CT_MARK_MASK]   = { .type = NLA_U32 },
 712         [TCA_FLOWER_KEY_CT_LABELS]      = { .type = NLA_BINARY,
 713                                             .len = 128 / BITS_PER_BYTE },
 714         [TCA_FLOWER_KEY_CT_LABELS_MASK] = { .type = NLA_BINARY,
 715                                             .len = 128 / BITS_PER_BYTE },
 716         [TCA_FLOWER_FLAGS]              = { .type = NLA_U32 },
 717         [TCA_FLOWER_KEY_HASH]           = { .type = NLA_U32 },
 718         [TCA_FLOWER_KEY_HASH_MASK]      = { .type = NLA_U32 },
 719         [TCA_FLOWER_KEY_NUM_OF_VLANS]   = { .type = NLA_U8 },
 720         [TCA_FLOWER_KEY_PPPOE_SID]      = { .type = NLA_U16 },
 721         [TCA_FLOWER_KEY_PPP_PROTO]      = { .type = NLA_U16 },
 722         [TCA_FLOWER_KEY_L2TPV3_SID]     = { .type = NLA_U32 },
 723
 724 };
 725
 726 static const struct nla_policy
 727 enc_opts_policy[TCA_FLOWER_KEY_ENC_OPTS_MAX + 1] = {
 728         [TCA_FLOWER_KEY_ENC_OPTS_UNSPEC]        = {
 729                 .strict_start_type = TCA_FLOWER_KEY_ENC_OPTS_VXLAN },
 730         [TCA_FLOWER_KEY_ENC_OPTS_GENEVE]        = { .type = NLA_NESTED },
 731         [TCA_FLOWER_KEY_ENC_OPTS_VXLAN]         = { .type = NLA_NESTED },
 732         [TCA_FLOWER_KEY_ENC_OPTS_ERSPAN]        = { .type = NLA_NESTED },
 733         [TCA_FLOWER_KEY_ENC_OPTS_GTP]           = { .type = NLA_NESTED },
 734 };
 735
 736 static const struct nla_policy
 737 geneve_opt_policy[TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX + 1] = {
 738         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS]      = { .type = NLA_U16 },
 739         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE]       = { .type = NLA_U8 },
 740         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA]       = { .type = NLA_BINARY,
 741                                                        .len = 128 },
 742 };
 743
 744 static const struct nla_policy
 745 vxlan_opt_policy[TCA_FLOWER_KEY_ENC_OPT_VXLAN_MAX + 1] = {
 746         [TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]         = { .type = NLA_U32 },
 747 };
 748
 749 static const struct nla_policy
 750 erspan_opt_policy[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_MAX + 1] = {
 751         [TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER]        = { .type = NLA_U8 },
 752         [TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX]      = { .type = NLA_U32 },
 753         [TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR]        = { .type = NLA_U8 },
 754         [TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID]       = { .type = NLA_U8 },
 755 };
 756
 757 static const struct nla_policy
 758 gtp_opt_policy[TCA_FLOWER_KEY_ENC_OPT_GTP_MAX + 1] = {
 759         [TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE]      = { .type = NLA_U8 },
 760         [TCA_FLOWER_KEY_ENC_OPT_GTP_QFI]           = { .type = NLA_U8 },
 761 };
 762
 763 static const struct nla_policy
 764 mpls_stack_entry_policy[TCA_FLOWER_KEY_MPLS_OPT_LSE_MAX + 1] = {
 765         [TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH]    = { .type = NLA_U8 },
 766         [TCA_FLOWER_KEY_MPLS_OPT_LSE_TTL]      = { .type = NLA_U8 },
 767         [TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS]      = { .type = NLA_U8 },
 768         [TCA_FLOWER_KEY_MPLS_OPT_LSE_TC]       = { .type = NLA_U8 },
 769         [TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL]    = { .type = NLA_U32 },
 770 };
 771
 772 static void fl_set_key_val(struct nlattr **tb,
 773                            void *val, int val_type,
 774                            void *mask, int mask_type, int len)
 775 {
 776         if (!tb[val_type])
 777                 return;
 778         nla_memcpy(val, tb[val_type], len);
 779         if (mask_type == TCA_FLOWER_UNSPEC || !tb[mask_type])
 780                 memset(mask, 0xff, len);
 781         else
 782                 nla_memcpy(mask, tb[mask_type], len);
 783 }
 784
 785 static int fl_set_key_port_range(struct nlattr **tb, struct fl_flow_key *key,
 786                                  struct fl_flow_key *mask,
 787                                  struct netlink_ext_ack *extack)
 788 {
 789         fl_set_key_val(tb, &key->tp_range.tp_min.dst,
 790                        TCA_FLOWER_KEY_PORT_DST_MIN, &mask->tp_range.tp_min.dst,
 791                        TCA_FLOWER_UNSPEC, sizeof(key->tp_range.tp_min.dst));
 792         fl_set_key_val(tb, &key->tp_range.tp_max.dst,
 793                        TCA_FLOWER_KEY_PORT_DST_MAX, &mask->tp_range.tp_max.dst,
 794                        TCA_FLOWER_UNSPEC, sizeof(key->tp_range.tp_max.dst));
 795         fl_set_key_val(tb, &key->tp_range.tp_min.src,
 796                        TCA_FLOWER_KEY_PORT_SRC_MIN, &mask->tp_range.tp_min.src,
 797                        TCA_FLOWER_UNSPEC, sizeof(key->tp_range.tp_min.src));
 798         fl_set_key_val(tb, &key->tp_range.tp_max.src,
 799                        TCA_FLOWER_KEY_PORT_SRC_MAX, &mask->tp_range.tp_max.src,
 800                        TCA_FLOWER_UNSPEC, sizeof(key->tp_range.tp_max.src));
 801
 802         if (mask->tp_range.tp_min.dst && mask->tp_range.tp_max.dst &&
 803             ntohs(key->tp_range.tp_max.dst) <=
 804             ntohs(key->tp_range.tp_min.dst)) {
 805                 NL_SET_ERR_MSG_ATTR(extack,
 806                                     tb[TCA_FLOWER_KEY_PORT_DST_MIN],
 807                                     "Invalid destination port range (min must be strictly smaller than max)");
 808                 return -EINVAL;
 809         }
 810         if (mask->tp_range.tp_min.src && mask->tp_range.tp_max.src &&
 811             ntohs(key->tp_range.tp_max.src) <=
 812             ntohs(key->tp_range.tp_min.src)) {
 813                 NL_SET_ERR_MSG_ATTR(extack,
 814                                     tb[TCA_FLOWER_KEY_PORT_SRC_MIN],
 815                                     "Invalid source port range (min must be strictly smaller than max)");
 816                 return -EINVAL;
 817         }
 818
 819         return 0;
 820 }
 821
 822 static int fl_set_key_mpls_lse(const struct nlattr *nla_lse,
 823                                struct flow_dissector_key_mpls *key_val,
 824                                struct flow_dissector_key_mpls *key_mask,
 825                                struct netlink_ext_ack *extack)
 826 {
 827         struct nlattr *tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_MAX + 1];
 828         struct flow_dissector_mpls_lse *lse_mask;
 829         struct flow_dissector_mpls_lse *lse_val;
 830         u8 lse_index;
 831         u8 depth;
 832         int err;
 833
 834         err = nla_parse_nested(tb, TCA_FLOWER_KEY_MPLS_OPT_LSE_MAX, nla_lse,
 835                                mpls_stack_entry_policy, extack);
 836         if (err < 0)
 837                 return err;
 838
 839         if (!tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH]) {
 840                 NL_SET_ERR_MSG(extack, "Missing MPLS option \"depth\"");
 841                 return -EINVAL;
 842         }
 843
 844         depth = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH]);
 845
 846         /* LSE depth starts at 1, for consistency with terminology used by
 847          * RFC 3031 (section 3.9), where depth 0 refers to unlabeled packets.
 848          */
 849         if (depth < 1 || depth > FLOW_DIS_MPLS_MAX) {
 850                 NL_SET_ERR_MSG_ATTR(extack,
 851                                     tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH],
 852                                     "Invalid MPLS depth");
 853                 return -EINVAL;
 854         }
 855         lse_index = depth - 1;
 856
 857         dissector_set_mpls_lse(key_val, lse_index);
 858         dissector_set_mpls_lse(key_mask, lse_index);
 859
 860         lse_val = &key_val->ls[lse_index];
 861         lse_mask = &key_mask->ls[lse_index];
 862
 863         if (tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_TTL]) {
 864                 lse_val->mpls_ttl = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_TTL]);
 865                 lse_mask->mpls_ttl = MPLS_TTL_MASK;
 866         }
 867         if (tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS]) {
 868                 u8 bos = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS]);
 869
 870                 if (bos & ~MPLS_BOS_MASK) {
 871                         NL_SET_ERR_MSG_ATTR(extack,
 872                                             tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS],
 873                                             "Bottom Of Stack (BOS) must be 0 or 1");
 874                         return -EINVAL;
 875                 }
 876                 lse_val->mpls_bos = bos;
 877                 lse_mask->mpls_bos = MPLS_BOS_MASK;
 878         }
 879         if (tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_TC]) {
 880                 u8 tc = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_TC]);
 881
 882                 if (tc & ~MPLS_TC_MASK) {
 883                         NL_SET_ERR_MSG_ATTR(extack,
 884                                             tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_TC],
 885                                             "Traffic Class (TC) must be between 0 and 7");
 886                         return -EINVAL;
 887                 }
 888                 lse_val->mpls_tc = tc;
 889                 lse_mask->mpls_tc = MPLS_TC_MASK;
 890         }
 891         if (tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL]) {
 892                 u32 label = nla_get_u32(tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL]);
 893
 894                 if (label & ~MPLS_LABEL_MASK) {
 895                         NL_SET_ERR_MSG_ATTR(extack,
 896                                             tb[TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL],
 897                                             "Label must be between 0 and 1048575");
 898                         return -EINVAL;
 899                 }
 900                 lse_val->mpls_label = label;
 901                 lse_mask->mpls_label = MPLS_LABEL_MASK;
 902         }
 903
 904         return 0;
 905 }
 906
 907 static int fl_set_key_mpls_opts(const struct nlattr *nla_mpls_opts,
 908                                 struct flow_dissector_key_mpls *key_val,
 909                                 struct flow_dissector_key_mpls *key_mask,
 910                                 struct netlink_ext_ack *extack)
 911 {
 912         struct nlattr *nla_lse;
 913         int rem;
 914         int err;
 915
 916         if (!(nla_mpls_opts->nla_type & NLA_F_NESTED)) {
 917                 NL_SET_ERR_MSG_ATTR(extack, nla_mpls_opts,
 918                                     "NLA_F_NESTED is missing");
 919                 return -EINVAL;
 920         }
 921
 922         nla_for_each_nested(nla_lse, nla_mpls_opts, rem) {
 923                 if (nla_type(nla_lse) != TCA_FLOWER_KEY_MPLS_OPTS_LSE) {
 924                         NL_SET_ERR_MSG_ATTR(extack, nla_lse,
 925                                             "Invalid MPLS option type");
 926                         return -EINVAL;
 927                 }
 928
 929                 err = fl_set_key_mpls_lse(nla_lse, key_val, key_mask, extack);
 930                 if (err < 0)
 931                         return err;
 932         }
 933         if (rem) {
 934                 NL_SET_ERR_MSG(extack,
 935                                "Bytes leftover after parsing MPLS options");
 936                 return -EINVAL;
 937         }
 938
 939         return 0;
 940 }
 941
 942 static int fl_set_key_mpls(struct nlattr **tb,
 943                            struct flow_dissector_key_mpls *key_val,
 944                            struct flow_dissector_key_mpls *key_mask,
 945                            struct netlink_ext_ack *extack)
 946 {
 947         struct flow_dissector_mpls_lse *lse_mask;
 948         struct flow_dissector_mpls_lse *lse_val;
 949
 950         if (tb[TCA_FLOWER_KEY_MPLS_OPTS]) {
 951                 if (tb[TCA_FLOWER_KEY_MPLS_TTL] ||
 952                     tb[TCA_FLOWER_KEY_MPLS_BOS] ||
 953                     tb[TCA_FLOWER_KEY_MPLS_TC] ||
 954                     tb[TCA_FLOWER_KEY_MPLS_LABEL]) {
 955                         NL_SET_ERR_MSG_ATTR(extack,
 956                                             tb[TCA_FLOWER_KEY_MPLS_OPTS],
 957                                             "MPLS label, Traffic Class, Bottom Of Stack and Time To Live must be encapsulated in the MPLS options attribute");
 958                         return -EBADMSG;
 959                 }
 960
 961                 return fl_set_key_mpls_opts(tb[TCA_FLOWER_KEY_MPLS_OPTS],
 962                                             key_val, key_mask, extack);
 963         }
 964
 965         lse_val = &key_val->ls[0];
 966         lse_mask = &key_mask->ls[0];
 967
 968         if (tb[TCA_FLOWER_KEY_MPLS_TTL]) {
 969                 lse_val->mpls_ttl = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_TTL]);
 970                 lse_mask->mpls_ttl = MPLS_TTL_MASK;
 971                 dissector_set_mpls_lse(key_val, 0);
 972                 dissector_set_mpls_lse(key_mask, 0);
 973         }
 974         if (tb[TCA_FLOWER_KEY_MPLS_BOS]) {
 975                 u8 bos = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_BOS]);
 976
 977                 if (bos & ~MPLS_BOS_MASK) {
 978                         NL_SET_ERR_MSG_ATTR(extack,
 979                                             tb[TCA_FLOWER_KEY_MPLS_BOS],
 980                                             "Bottom Of Stack (BOS) must be 0 or 1");
 981                         return -EINVAL;
 982                 }
 983                 lse_val->mpls_bos = bos;
 984                 lse_mask->mpls_bos = MPLS_BOS_MASK;
 985                 dissector_set_mpls_lse(key_val, 0);
 986                 dissector_set_mpls_lse(key_mask, 0);
 987         }
 988         if (tb[TCA_FLOWER_KEY_MPLS_TC]) {
 989                 u8 tc = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_TC]);
 990
 991                 if (tc & ~MPLS_TC_MASK) {
 992                         NL_SET_ERR_MSG_ATTR(extack,
 993                                             tb[TCA_FLOWER_KEY_MPLS_TC],
 994                                             "Traffic Class (TC) must be between 0 and 7");
 995                         return -EINVAL;
 996                 }
 997                 lse_val->mpls_tc = tc;
 998                 lse_mask->mpls_tc = MPLS_TC_MASK;
 999                 dissector_set_mpls_lse(key_val, 0);
1000                 dissector_set_mpls_lse(key_mask, 0);
1001         }
1002         if (tb[TCA_FLOWER_KEY_MPLS_LABEL]) {
1003                 u32 label = nla_get_u32(tb[TCA_FLOWER_KEY_MPLS_LABEL]);
1004
1005                 if (label & ~MPLS_LABEL_MASK) {
1006                         NL_SET_ERR_MSG_ATTR(extack,
1007                                             tb[TCA_FLOWER_KEY_MPLS_LABEL],
1008                                             "Label must be between 0 and 1048575");
1009                         return -EINVAL;
1010                 }
1011                 lse_val->mpls_label = label;
1012                 lse_mask->mpls_label = MPLS_LABEL_MASK;
1013                 dissector_set_mpls_lse(key_val, 0);
1014                 dissector_set_mpls_lse(key_mask, 0);
1015         }
1016         return 0;
1017 }
1018
1019 static void fl_set_key_vlan(struct nlattr **tb,
1020                             __be16 ethertype,
1021                             int vlan_id_key, int vlan_prio_key,
1022                             int vlan_next_eth_type_key,
1023                             struct flow_dissector_key_vlan *key_val,
1024                             struct flow_dissector_key_vlan *key_mask)
1025 {
1026 #define VLAN_PRIORITY_MASK      0x7
1027
1028         if (tb[vlan_id_key]) {
1029                 key_val->vlan_id =
1030                         nla_get_u16(tb[vlan_id_key]) & VLAN_VID_MASK;
1031                 key_mask->vlan_id = VLAN_VID_MASK;
1032         }
1033         if (tb[vlan_prio_key]) {
1034                 key_val->vlan_priority =
1035                         nla_get_u8(tb[vlan_prio_key]) &
1036                         VLAN_PRIORITY_MASK;
1037                 key_mask->vlan_priority = VLAN_PRIORITY_MASK;
1038         }
1039         if (ethertype) {
1040                 key_val->vlan_tpid = ethertype;
1041                 key_mask->vlan_tpid = cpu_to_be16(~0);
1042         }
1043         if (tb[vlan_next_eth_type_key]) {
1044                 key_val->vlan_eth_type =
1045                         nla_get_be16(tb[vlan_next_eth_type_key]);
1046                 key_mask->vlan_eth_type = cpu_to_be16(~0);
1047         }
1048 }
1049
1050 static void fl_set_key_pppoe(struct nlattr **tb,
1051                              struct flow_dissector_key_pppoe *key_val,
1052                              struct flow_dissector_key_pppoe *key_mask,
1053                              struct fl_flow_key *key,
1054                              struct fl_flow_key *mask)
1055 {
1056         /* key_val::type must be set to ETH_P_PPP_SES
1057          * because ETH_P_PPP_SES was stored in basic.n_proto
1058          * which might get overwritten by ppp_proto
1059          * or might be set to 0, the role of key_val::type
1060          * is similar to vlan_key::tpid
1061          */
1062         key_val->type = htons(ETH_P_PPP_SES);
1063         key_mask->type = cpu_to_be16(~0);
1064
1065         if (tb[TCA_FLOWER_KEY_PPPOE_SID]) {
1066                 key_val->session_id =
1067                         nla_get_be16(tb[TCA_FLOWER_KEY_PPPOE_SID]);
1068                 key_mask->session_id = cpu_to_be16(~0);
1069         }
1070         if (tb[TCA_FLOWER_KEY_PPP_PROTO]) {
1071                 key_val->ppp_proto =
1072                         nla_get_be16(tb[TCA_FLOWER_KEY_PPP_PROTO]);
1073                 key_mask->ppp_proto = cpu_to_be16(~0);
1074
1075                 if (key_val->ppp_proto == htons(PPP_IP)) {
1076                         key->basic.n_proto = htons(ETH_P_IP);
1077                         mask->basic.n_proto = cpu_to_be16(~0);
1078                 } else if (key_val->ppp_proto == htons(PPP_IPV6)) {
1079                         key->basic.n_proto = htons(ETH_P_IPV6);
1080                         mask->basic.n_proto = cpu_to_be16(~0);
1081                 } else if (key_val->ppp_proto == htons(PPP_MPLS_UC)) {
1082                         key->basic.n_proto = htons(ETH_P_MPLS_UC);
1083                         mask->basic.n_proto = cpu_to_be16(~0);
1084                 } else if (key_val->ppp_proto == htons(PPP_MPLS_MC)) {
1085                         key->basic.n_proto = htons(ETH_P_MPLS_MC);
1086                         mask->basic.n_proto = cpu_to_be16(~0);
1087                 }
1088         } else {
1089                 key->basic.n_proto = 0;
1090                 mask->basic.n_proto = cpu_to_be16(0);
1091         }
1092 }
1093
1094 static void fl_set_key_flag(u32 flower_key, u32 flower_mask,
1095                             u32 *dissector_key, u32 *dissector_mask,
1096                             u32 flower_flag_bit, u32 dissector_flag_bit)
1097 {
1098         if (flower_mask & flower_flag_bit) {
1099                 *dissector_mask |= dissector_flag_bit;
1100                 if (flower_key & flower_flag_bit)
1101                         *dissector_key |= dissector_flag_bit;
1102         }
1103 }
1104
1105 static int fl_set_key_flags(struct nlattr **tb, u32 *flags_key,
1106                             u32 *flags_mask, struct netlink_ext_ack *extack)
1107 {
1108         u32 key, mask;
1109
1110         /* mask is mandatory for flags */
1111         if (!tb[TCA_FLOWER_KEY_FLAGS_MASK]) {
1112                 NL_SET_ERR_MSG(extack, "Missing flags mask");
1113                 return -EINVAL;
1114         }
1115
1116         key = be32_to_cpu(nla_get_be32(tb[TCA_FLOWER_KEY_FLAGS]));
1117         mask = be32_to_cpu(nla_get_be32(tb[TCA_FLOWER_KEY_FLAGS_MASK]));
1118
1119         *flags_key  = 0;
1120         *flags_mask = 0;
1121
1122         fl_set_key_flag(key, mask, flags_key, flags_mask,
1123                         TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT, FLOW_DIS_IS_FRAGMENT);
1124         fl_set_key_flag(key, mask, flags_key, flags_mask,
1125                         TCA_FLOWER_KEY_FLAGS_FRAG_IS_FIRST,
1126                         FLOW_DIS_FIRST_FRAG);
1127
1128         return 0;
1129 }
1130
1131 static void fl_set_key_ip(struct nlattr **tb, bool encap,
1132                           struct flow_dissector_key_ip *key,
1133                           struct flow_dissector_key_ip *mask)
1134 {
1135         int tos_key = encap ? TCA_FLOWER_KEY_ENC_IP_TOS : TCA_FLOWER_KEY_IP_TOS;
1136         int ttl_key = encap ? TCA_FLOWER_KEY_ENC_IP_TTL : TCA_FLOWER_KEY_IP_TTL;
1137         int tos_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TOS_MASK : TCA_FLOWER_KEY_IP_TOS_MASK;
1138         int ttl_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TTL_MASK : TCA_FLOWER_KEY_IP_TTL_MASK;
1139
1140         fl_set_key_val(tb, &key->tos, tos_key, &mask->tos, tos_mask, sizeof(key->tos));
1141         fl_set_key_val(tb, &key->ttl, ttl_key, &mask->ttl, ttl_mask, sizeof(key->ttl));
1142 }
1143
1144 static int fl_set_geneve_opt(const struct nlattr *nla, struct fl_flow_key *key,
1145                              int depth, int option_len,
1146                              struct netlink_ext_ack *extack)
1147 {
1148         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX + 1];
1149         struct nlattr *class = NULL, *type = NULL, *data = NULL;
1150         struct geneve_opt *opt;
1151         int err, data_len = 0;
1152
1153         if (option_len > sizeof(struct geneve_opt))
1154                 data_len = option_len - sizeof(struct geneve_opt);
1155
1156         if (key->enc_opts.len > FLOW_DIS_TUN_OPTS_MAX - 4)
1157                 return -ERANGE;
1158
1159         opt = (struct geneve_opt *)&key->enc_opts.data[key->enc_opts.len];
1160         memset(opt, 0xff, option_len);
1161         opt->length = data_len / 4;
1162         opt->r1 = 0;
1163         opt->r2 = 0;
1164         opt->r3 = 0;
1165
1166         /* If no mask has been prodived we assume an exact match. */
1167         if (!depth)
1168                 return sizeof(struct geneve_opt) + data_len;
1169
1170         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_GENEVE) {
1171                 NL_SET_ERR_MSG(extack, "Non-geneve option type for mask");
1172                 return -EINVAL;
1173         }
1174
1175         err = nla_parse_nested_deprecated(tb,
1176                                           TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX,
1177                                           nla, geneve_opt_policy, extack);
1178         if (err < 0)
1179                 return err;
1180
1181         /* We are not allowed to omit any of CLASS, TYPE or DATA
1182          * fields from the key.
1183          */
1184         if (!option_len &&
1185             (!tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS] ||
1186              !tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE] ||
1187              !tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA])) {
1188                 NL_SET_ERR_MSG(extack, "Missing tunnel key geneve option class, type or data");
1189                 return -EINVAL;
1190         }
1191
1192         /* Omitting any of CLASS, TYPE or DATA fields is allowed
1193          * for the mask.
1194          */
1195         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA]) {
1196                 int new_len = key->enc_opts.len;
1197
1198                 data = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA];
1199                 data_len = nla_len(data);
1200                 if (data_len < 4) {
1201                         NL_SET_ERR_MSG(extack, "Tunnel key geneve option data is less than 4 bytes long");
1202                         return -ERANGE;
1203                 }
1204                 if (data_len % 4) {
1205                         NL_SET_ERR_MSG(extack, "Tunnel key geneve option data is not a multiple of 4 bytes long");
1206                         return -ERANGE;
1207                 }
1208
1209                 new_len += sizeof(struct geneve_opt) + data_len;
1210                 BUILD_BUG_ON(FLOW_DIS_TUN_OPTS_MAX != IP_TUNNEL_OPTS_MAX);
1211                 if (new_len > FLOW_DIS_TUN_OPTS_MAX) {
1212                         NL_SET_ERR_MSG(extack, "Tunnel options exceeds max size");
1213                         return -ERANGE;
1214                 }
1215                 opt->length = data_len / 4;
1216                 memcpy(opt->opt_data, nla_data(data), data_len);
1217         }
1218
1219         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS]) {
1220                 class = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS];
1221                 opt->opt_class = nla_get_be16(class);
1222         }
1223
1224         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE]) {
1225                 type = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE];
1226                 opt->type = nla_get_u8(type);
1227         }
1228
1229         return sizeof(struct geneve_opt) + data_len;
1230 }
1231
1232 static int fl_set_vxlan_opt(const struct nlattr *nla, struct fl_flow_key *key,
1233                             int depth, int option_len,
1234                             struct netlink_ext_ack *extack)
1235 {
1236         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_MAX + 1];
1237         struct vxlan_metadata *md;
1238         int err;
1239
1240         md = (struct vxlan_metadata *)&key->enc_opts.data[key->enc_opts.len];
1241         memset(md, 0xff, sizeof(*md));
1242
1243         if (!depth)
1244                 return sizeof(*md);
1245
1246         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_VXLAN) {
1247                 NL_SET_ERR_MSG(extack, "Non-vxlan option type for mask");
1248                 return -EINVAL;
1249         }
1250
1251         err = nla_parse_nested(tb, TCA_FLOWER_KEY_ENC_OPT_VXLAN_MAX, nla,
1252                                vxlan_opt_policy, extack);
1253         if (err < 0)
1254                 return err;
1255
1256         if (!option_len && !tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]) {
1257                 NL_SET_ERR_MSG(extack, "Missing tunnel key vxlan option gbp");
1258                 return -EINVAL;
1259         }
1260
1261         if (tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]) {
1262                 md->gbp = nla_get_u32(tb[TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP]);
1263                 md->gbp &= VXLAN_GBP_MASK;
1264         }
1265
1266         return sizeof(*md);
1267 }
1268
1269 static int fl_set_erspan_opt(const struct nlattr *nla, struct fl_flow_key *key,
1270                              int depth, int option_len,
1271                              struct netlink_ext_ack *extack)
1272 {
1273         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_MAX + 1];
1274         struct erspan_metadata *md;
1275         int err;
1276
1277         md = (struct erspan_metadata *)&key->enc_opts.data[key->enc_opts.len];
1278         memset(md, 0xff, sizeof(*md));
1279         md->version = 1;
1280
1281         if (!depth)
1282                 return sizeof(*md);
1283
1284         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_ERSPAN) {
1285                 NL_SET_ERR_MSG(extack, "Non-erspan option type for mask");
1286                 return -EINVAL;
1287         }
1288
1289         err = nla_parse_nested(tb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_MAX, nla,
1290                                erspan_opt_policy, extack);
1291         if (err < 0)
1292                 return err;
1293
1294         if (!option_len && !tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER]) {
1295                 NL_SET_ERR_MSG(extack, "Missing tunnel key erspan option ver");
1296                 return -EINVAL;
1297         }
1298
1299         if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER])
1300                 md->version = nla_get_u8(tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER]);
1301
1302         if (md->version == 1) {
1303                 if (!option_len && !tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX]) {
1304                         NL_SET_ERR_MSG(extack, "Missing tunnel key erspan option index");
1305                         return -EINVAL;
1306                 }
1307                 if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX]) {
1308                         nla = tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX];
1309                         memset(&md->u, 0x00, sizeof(md->u));
1310                         md->u.index = nla_get_be32(nla);
1311                 }
1312         } else if (md->version == 2) {
1313                 if (!option_len && (!tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR] ||
1314                                     !tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID])) {
1315                         NL_SET_ERR_MSG(extack, "Missing tunnel key erspan option dir or hwid");
1316                         return -EINVAL;
1317                 }
1318                 if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR]) {
1319                         nla = tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR];
1320                         md->u.md2.dir = nla_get_u8(nla);
1321                 }
1322                 if (tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID]) {
1323                         nla = tb[TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID];
1324                         set_hwid(&md->u.md2, nla_get_u8(nla));
1325                 }
1326         } else {
1327                 NL_SET_ERR_MSG(extack, "Tunnel key erspan option ver is incorrect");
1328                 return -EINVAL;
1329         }
1330
1331         return sizeof(*md);
1332 }
1333
1334 static int fl_set_gtp_opt(const struct nlattr *nla, struct fl_flow_key *key,
1335                           int depth, int option_len,
1336                           struct netlink_ext_ack *extack)
1337 {
1338         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_GTP_MAX + 1];
1339         struct gtp_pdu_session_info *sinfo;
1340         u8 len = key->enc_opts.len;
1341         int err;
1342
1343         sinfo = (struct gtp_pdu_session_info *)&key->enc_opts.data[len];
1344         memset(sinfo, 0xff, option_len);
1345
1346         if (!depth)
1347                 return sizeof(*sinfo);
1348
1349         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_GTP) {
1350                 NL_SET_ERR_MSG_MOD(extack, "Non-gtp option type for mask");
1351                 return -EINVAL;
1352         }
1353
1354         err = nla_parse_nested(tb, TCA_FLOWER_KEY_ENC_OPT_GTP_MAX, nla,
1355                                gtp_opt_policy, extack);
1356         if (err < 0)
1357                 return err;
1358
1359         if (!option_len &&
1360             (!tb[TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE] ||
1361              !tb[TCA_FLOWER_KEY_ENC_OPT_GTP_QFI])) {
1362                 NL_SET_ERR_MSG_MOD(extack,
1363                                    "Missing tunnel key gtp option pdu type or qfi");
1364                 return -EINVAL;
1365         }
1366
1367         if (tb[TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE])
1368                 sinfo->pdu_type =
1369                         nla_get_u8(tb[TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE]);
1370
1371         if (tb[TCA_FLOWER_KEY_ENC_OPT_GTP_QFI])
1372                 sinfo->qfi = nla_get_u8(tb[TCA_FLOWER_KEY_ENC_OPT_GTP_QFI]);
1373
1374         return sizeof(*sinfo);
1375 }
1376
1377 static int fl_set_enc_opt(struct nlattr **tb, struct fl_flow_key *key,
1378                           struct fl_flow_key *mask,
1379                           struct netlink_ext_ack *extack)
1380 {
1381         const struct nlattr *nla_enc_key, *nla_opt_key, *nla_opt_msk = NULL;
1382         int err, option_len, key_depth, msk_depth = 0;
1383
1384         err = nla_validate_nested_deprecated(tb[TCA_FLOWER_KEY_ENC_OPTS],
1385                                              TCA_FLOWER_KEY_ENC_OPTS_MAX,
1386                                              enc_opts_policy, extack);
1387         if (err)
1388                 return err;
1389
1390         nla_enc_key = nla_data(tb[TCA_FLOWER_KEY_ENC_OPTS]);
1391
1392         if (tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]) {
1393                 err = nla_validate_nested_deprecated(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK],
1394                                                      TCA_FLOWER_KEY_ENC_OPTS_MAX,
1395                                                      enc_opts_policy, extack);
1396                 if (err)
1397                         return err;
1398
1399                 nla_opt_msk = nla_data(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]);
1400                 msk_depth = nla_len(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]);
1401                 if (!nla_ok(nla_opt_msk, msk_depth)) {
1402                         NL_SET_ERR_MSG(extack, "Invalid nested attribute for masks");
1403                         return -EINVAL;
1404                 }
1405         }
1406
1407         nla_for_each_attr(nla_opt_key, nla_enc_key,
1408                           nla_len(tb[TCA_FLOWER_KEY_ENC_OPTS]), key_depth) {
1409                 switch (nla_type(nla_opt_key)) {
1410                 case TCA_FLOWER_KEY_ENC_OPTS_GENEVE:
1411                         if (key->enc_opts.dst_opt_type &&
1412                             key->enc_opts.dst_opt_type != TUNNEL_GENEVE_OPT) {
1413                                 NL_SET_ERR_MSG(extack, "Duplicate type for geneve options");
1414                                 return -EINVAL;
1415                         }
1416                         option_len = 0;
1417                         key->enc_opts.dst_opt_type = TUNNEL_GENEVE_OPT;
1418                         option_len = fl_set_geneve_opt(nla_opt_key, key,
1419                                                        key_depth, option_len,
1420                                                        extack);
1421                         if (option_len < 0)
1422                                 return option_len;
1423
1424                         key->enc_opts.len += option_len;
1425                         /* At the same time we need to parse through the mask
1426                          * in order to verify exact and mask attribute lengths.
1427                          */
1428                         mask->enc_opts.dst_opt_type = TUNNEL_GENEVE_OPT;
1429                         option_len = fl_set_geneve_opt(nla_opt_msk, mask,
1430                                                        msk_depth, option_len,
1431                                                        extack);
1432                         if (option_len < 0)
1433                                 return option_len;
1434
1435                         mask->enc_opts.len += option_len;
1436                         if (key->enc_opts.len != mask->enc_opts.len) {
1437                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
1438                                 return -EINVAL;
1439                         }
1440                         break;
1441                 case TCA_FLOWER_KEY_ENC_OPTS_VXLAN:
1442                         if (key->enc_opts.dst_opt_type) {
1443                                 NL_SET_ERR_MSG(extack, "Duplicate type for vxlan options");
1444                                 return -EINVAL;
1445                         }
1446                         option_len = 0;
1447                         key->enc_opts.dst_opt_type = TUNNEL_VXLAN_OPT;
1448                         option_len = fl_set_vxlan_opt(nla_opt_key, key,
1449                                                       key_depth, option_len,
1450                                                       extack);
1451                         if (option_len < 0)
1452                                 return option_len;
1453
1454                         key->enc_opts.len += option_len;
1455                         /* At the same time we need to parse through the mask
1456                          * in order to verify exact and mask attribute lengths.
1457                          */
1458                         mask->enc_opts.dst_opt_type = TUNNEL_VXLAN_OPT;
1459                         option_len = fl_set_vxlan_opt(nla_opt_msk, mask,
1460                                                       msk_depth, option_len,
1461                                                       extack);
1462                         if (option_len < 0)
1463                                 return option_len;
1464
1465                         mask->enc_opts.len += option_len;
1466                         if (key->enc_opts.len != mask->enc_opts.len) {
1467                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
1468                                 return -EINVAL;
1469                         }
1470                         break;
1471                 case TCA_FLOWER_KEY_ENC_OPTS_ERSPAN:
1472                         if (key->enc_opts.dst_opt_type) {
1473                                 NL_SET_ERR_MSG(extack, "Duplicate type for erspan options");
1474                                 return -EINVAL;
1475                         }
1476                         option_len = 0;
1477                         key->enc_opts.dst_opt_type = TUNNEL_ERSPAN_OPT;
1478                         option_len = fl_set_erspan_opt(nla_opt_key, key,
1479                                                        key_depth, option_len,
1480                                                        extack);
1481                         if (option_len < 0)
1482                                 return option_len;
1483
1484                         key->enc_opts.len += option_len;
1485                         /* At the same time we need to parse through the mask
1486                          * in order to verify exact and mask attribute lengths.
1487                          */
1488                         mask->enc_opts.dst_opt_type = TUNNEL_ERSPAN_OPT;
1489                         option_len = fl_set_erspan_opt(nla_opt_msk, mask,
1490                                                        msk_depth, option_len,
1491                                                        extack);
1492                         if (option_len < 0)
1493                                 return option_len;
1494
1495                         mask->enc_opts.len += option_len;
1496                         if (key->enc_opts.len != mask->enc_opts.len) {
1497                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
1498                                 return -EINVAL;
1499                         }
1500                         break;
1501                 case TCA_FLOWER_KEY_ENC_OPTS_GTP:
1502                         if (key->enc_opts.dst_opt_type) {
1503                                 NL_SET_ERR_MSG_MOD(extack,
1504                                                    "Duplicate type for gtp options");
1505                                 return -EINVAL;
1506                         }
1507                         option_len = 0;
1508                         key->enc_opts.dst_opt_type = TUNNEL_GTP_OPT;
1509                         option_len = fl_set_gtp_opt(nla_opt_key, key,
1510                                                     key_depth, option_len,
1511                                                     extack);
1512                         if (option_len < 0)
1513                                 return option_len;
1514
1515                         key->enc_opts.len += option_len;
1516                         /* At the same time we need to parse through the mask
1517                          * in order to verify exact and mask attribute lengths.
1518                          */
1519                         mask->enc_opts.dst_opt_type = TUNNEL_GTP_OPT;
1520                         option_len = fl_set_gtp_opt(nla_opt_msk, mask,
1521                                                     msk_depth, option_len,
1522                                                     extack);
1523                         if (option_len < 0)
1524                                 return option_len;
1525
1526                         mask->enc_opts.len += option_len;
1527                         if (key->enc_opts.len != mask->enc_opts.len) {
1528                                 NL_SET_ERR_MSG_MOD(extack,
1529                                                    "Key and mask miss aligned");
1530                                 return -EINVAL;
1531                         }
1532                         break;
1533                 default:
1534                         NL_SET_ERR_MSG(extack, "Unknown tunnel option type");
1535                         return -EINVAL;
1536                 }
1537
1538                 if (!msk_depth)
1539                         continue;
1540
1541                 if (!nla_ok(nla_opt_msk, msk_depth)) {
1542                         NL_SET_ERR_MSG(extack, "A mask attribute is invalid");
1543                         return -EINVAL;
1544                 }
1545                 nla_opt_msk = nla_next(nla_opt_msk, &msk_depth);
1546         }
1547
1548         return 0;
1549 }
1550
1551 static int fl_validate_ct_state(u16 state, struct nlattr *tb,
1552                                 struct netlink_ext_ack *extack)
1553 {
1554         if (state && !(state & TCA_FLOWER_KEY_CT_FLAGS_TRACKED)) {
1555                 NL_SET_ERR_MSG_ATTR(extack, tb,
1556                                     "no trk, so no other flag can be set");
1557                 return -EINVAL;
1558         }
1559
1560         if (state & TCA_FLOWER_KEY_CT_FLAGS_NEW &&
1561             state & TCA_FLOWER_KEY_CT_FLAGS_ESTABLISHED) {
1562                 NL_SET_ERR_MSG_ATTR(extack, tb,
1563                                     "new and est are mutually exclusive");
1564                 return -EINVAL;
1565         }
1566
1567         if (state & TCA_FLOWER_KEY_CT_FLAGS_INVALID &&
1568             state & ~(TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
1569                       TCA_FLOWER_KEY_CT_FLAGS_INVALID)) {
1570                 NL_SET_ERR_MSG_ATTR(extack, tb,
1571                                     "when inv is set, only trk may be set");
1572                 return -EINVAL;
1573         }
1574
1575         if (state & TCA_FLOWER_KEY_CT_FLAGS_NEW &&
1576             state & TCA_FLOWER_KEY_CT_FLAGS_REPLY) {
1577                 NL_SET_ERR_MSG_ATTR(extack, tb,
1578                                     "new and rpl are mutually exclusive");
1579                 return -EINVAL;
1580         }
1581
1582         return 0;
1583 }
1584
1585 static int fl_set_key_ct(struct nlattr **tb,
1586                          struct flow_dissector_key_ct *key,
1587                          struct flow_dissector_key_ct *mask,
1588                          struct netlink_ext_ack *extack)
1589 {
1590         if (tb[TCA_FLOWER_KEY_CT_STATE]) {
1591                 int err;
1592
1593                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK)) {
1594                         NL_SET_ERR_MSG(extack, "Conntrack isn't enabled");
1595                         return -EOPNOTSUPP;
1596                 }
1597                 fl_set_key_val(tb, &key->ct_state, TCA_FLOWER_KEY_CT_STATE,
1598                                &mask->ct_state, TCA_FLOWER_KEY_CT_STATE_MASK,
1599                                sizeof(key->ct_state));
1600
1601                 err = fl_validate_ct_state(key->ct_state & mask->ct_state,
1602                                            tb[TCA_FLOWER_KEY_CT_STATE_MASK],
1603                                            extack);
1604                 if (err)
1605                         return err;
1606
1607         }
1608         if (tb[TCA_FLOWER_KEY_CT_ZONE]) {
1609                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK_ZONES)) {
1610                         NL_SET_ERR_MSG(extack, "Conntrack zones isn't enabled");
1611                         return -EOPNOTSUPP;
1612                 }
1613                 fl_set_key_val(tb, &key->ct_zone, TCA_FLOWER_KEY_CT_ZONE,
1614                                &mask->ct_zone, TCA_FLOWER_KEY_CT_ZONE_MASK,
1615                                sizeof(key->ct_zone));
1616         }
1617         if (tb[TCA_FLOWER_KEY_CT_MARK]) {
1618                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK_MARK)) {
1619                         NL_SET_ERR_MSG(extack, "Conntrack mark isn't enabled");
1620                         return -EOPNOTSUPP;
1621                 }
1622                 fl_set_key_val(tb, &key->ct_mark, TCA_FLOWER_KEY_CT_MARK,
1623                                &mask->ct_mark, TCA_FLOWER_KEY_CT_MARK_MASK,
1624                                sizeof(key->ct_mark));
1625         }
1626         if (tb[TCA_FLOWER_KEY_CT_LABELS]) {
1627                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK_LABELS)) {
1628                         NL_SET_ERR_MSG(extack, "Conntrack labels aren't enabled");
1629                         return -EOPNOTSUPP;
1630                 }
1631                 fl_set_key_val(tb, key->ct_labels, TCA_FLOWER_KEY_CT_LABELS,
1632                                mask->ct_labels, TCA_FLOWER_KEY_CT_LABELS_MASK,
1633                                sizeof(key->ct_labels));
1634         }
1635
1636         return 0;
1637 }
1638
1639 static bool is_vlan_key(struct nlattr *tb, __be16 *ethertype,
1640                         struct fl_flow_key *key, struct fl_flow_key *mask,
1641                         int vthresh)
1642 {
1643         const bool good_num_of_vlans = key->num_of_vlans.num_of_vlans > vthresh;
1644
1645         if (!tb) {
1646                 *ethertype = 0;
1647                 return good_num_of_vlans;
1648         }
1649
1650         *ethertype = nla_get_be16(tb);
1651         if (good_num_of_vlans || eth_type_vlan(*ethertype))
1652                 return true;
1653
1654         key->basic.n_proto = *ethertype;
1655         mask->basic.n_proto = cpu_to_be16(~0);
1656         return false;
1657 }
1658
1659 static int fl_set_key(struct net *net, struct nlattr **tb,
1660                       struct fl_flow_key *key, struct fl_flow_key *mask,
1661                       struct netlink_ext_ack *extack)
1662 {
1663         __be16 ethertype;
1664         int ret = 0;
1665
1666         if (tb[TCA_FLOWER_INDEV]) {
1667                 int err = tcf_change_indev(net, tb[TCA_FLOWER_INDEV], extack);
1668                 if (err < 0)
1669                         return err;
1670                 key->meta.ingress_ifindex = err;
1671                 mask->meta.ingress_ifindex = 0xffffffff;
1672         }
1673
1674         fl_set_key_val(tb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
1675                        mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
1676                        sizeof(key->eth.dst));
1677         fl_set_key_val(tb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
1678                        mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
1679                        sizeof(key->eth.src));
1680         fl_set_key_val(tb, &key->num_of_vlans,
1681                        TCA_FLOWER_KEY_NUM_OF_VLANS,
1682                        &mask->num_of_vlans,
1683                        TCA_FLOWER_UNSPEC,
1684                        sizeof(key->num_of_vlans));
1685
1686         if (is_vlan_key(tb[TCA_FLOWER_KEY_ETH_TYPE], &ethertype, key, mask, 0)) {
1687                 fl_set_key_vlan(tb, ethertype, TCA_FLOWER_KEY_VLAN_ID,
1688                                 TCA_FLOWER_KEY_VLAN_PRIO,
1689                                 TCA_FLOWER_KEY_VLAN_ETH_TYPE,
1690                                 &key->vlan, &mask->vlan);
1691
1692                 if (is_vlan_key(tb[TCA_FLOWER_KEY_VLAN_ETH_TYPE],
1693                                 &ethertype, key, mask, 1)) {
1694                         fl_set_key_vlan(tb, ethertype,
1695                                         TCA_FLOWER_KEY_CVLAN_ID,
1696                                         TCA_FLOWER_KEY_CVLAN_PRIO,
1697                                         TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
1698                                         &key->cvlan, &mask->cvlan);
1699                         fl_set_key_val(tb, &key->basic.n_proto,
1700                                        TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
1701                                        &mask->basic.n_proto,
1702                                        TCA_FLOWER_UNSPEC,
1703                                        sizeof(key->basic.n_proto));
1704                 }
1705         }
1706
1707         if (key->basic.n_proto == htons(ETH_P_PPP_SES))
1708                 fl_set_key_pppoe(tb, &key->pppoe, &mask->pppoe, key, mask);
1709
1710         if (key->basic.n_proto == htons(ETH_P_IP) ||
1711             key->basic.n_proto == htons(ETH_P_IPV6)) {
1712                 fl_set_key_val(tb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
1713                                &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
1714                                sizeof(key->basic.ip_proto));
1715                 fl_set_key_ip(tb, false, &key->ip, &mask->ip);
1716         }
1717
1718         if (tb[TCA_FLOWER_KEY_IPV4_SRC] || tb[TCA_FLOWER_KEY_IPV4_DST]) {
1719                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
1720                 mask->control.addr_type = ~0;
1721                 fl_set_key_val(tb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
1722                                &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
1723                                sizeof(key->ipv4.src));
1724                 fl_set_key_val(tb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
1725                                &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
1726                                sizeof(key->ipv4.dst));
1727         } else if (tb[TCA_FLOWER_KEY_IPV6_SRC] || tb[TCA_FLOWER_KEY_IPV6_DST]) {
1728                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
1729                 mask->control.addr_type = ~0;
1730                 fl_set_key_val(tb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
1731                                &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
1732                                sizeof(key->ipv6.src));
1733                 fl_set_key_val(tb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
1734                                &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
1735                                sizeof(key->ipv6.dst));
1736         }
1737
1738         if (key->basic.ip_proto == IPPROTO_TCP) {
1739                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
1740                                &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
1741                                sizeof(key->tp.src));
1742                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
1743                                &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
1744                                sizeof(key->tp.dst));
1745                 fl_set_key_val(tb, &key->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS,
1746                                &mask->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS_MASK,
1747                                sizeof(key->tcp.flags));
1748         } else if (key->basic.ip_proto == IPPROTO_UDP) {
1749                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
1750                                &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
1751                                sizeof(key->tp.src));
1752                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
1753                                &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
1754                                sizeof(key->tp.dst));
1755         } else if (key->basic.ip_proto == IPPROTO_SCTP) {
1756                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_SCTP_SRC,
1757                                &mask->tp.src, TCA_FLOWER_KEY_SCTP_SRC_MASK,
1758                                sizeof(key->tp.src));
1759                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_SCTP_DST,
1760                                &mask->tp.dst, TCA_FLOWER_KEY_SCTP_DST_MASK,
1761                                sizeof(key->tp.dst));
1762         } else if (key->basic.n_proto == htons(ETH_P_IP) &&
1763                    key->basic.ip_proto == IPPROTO_ICMP) {
1764                 fl_set_key_val(tb, &key->icmp.type, TCA_FLOWER_KEY_ICMPV4_TYPE,
1765                                &mask->icmp.type,
1766                                TCA_FLOWER_KEY_ICMPV4_TYPE_MASK,
1767                                sizeof(key->icmp.type));
1768                 fl_set_key_val(tb, &key->icmp.code, TCA_FLOWER_KEY_ICMPV4_CODE,
1769                                &mask->icmp.code,
1770                                TCA_FLOWER_KEY_ICMPV4_CODE_MASK,
1771                                sizeof(key->icmp.code));
1772         } else if (key->basic.n_proto == htons(ETH_P_IPV6) &&
1773                    key->basic.ip_proto == IPPROTO_ICMPV6) {
1774                 fl_set_key_val(tb, &key->icmp.type, TCA_FLOWER_KEY_ICMPV6_TYPE,
1775                                &mask->icmp.type,
1776                                TCA_FLOWER_KEY_ICMPV6_TYPE_MASK,
1777                                sizeof(key->icmp.type));
1778                 fl_set_key_val(tb, &key->icmp.code, TCA_FLOWER_KEY_ICMPV6_CODE,
1779                                &mask->icmp.code,
1780                                TCA_FLOWER_KEY_ICMPV6_CODE_MASK,
1781                                sizeof(key->icmp.code));
1782         } else if (key->basic.n_proto == htons(ETH_P_MPLS_UC) ||
1783                    key->basic.n_proto == htons(ETH_P_MPLS_MC)) {
1784                 ret = fl_set_key_mpls(tb, &key->mpls, &mask->mpls, extack);
1785                 if (ret)
1786                         return ret;
1787         } else if (key->basic.n_proto == htons(ETH_P_ARP) ||
1788                    key->basic.n_proto == htons(ETH_P_RARP)) {
1789                 fl_set_key_val(tb, &key->arp.sip, TCA_FLOWER_KEY_ARP_SIP,
1790                                &mask->arp.sip, TCA_FLOWER_KEY_ARP_SIP_MASK,
1791                                sizeof(key->arp.sip));
1792                 fl_set_key_val(tb, &key->arp.tip, TCA_FLOWER_KEY_ARP_TIP,
1793                                &mask->arp.tip, TCA_FLOWER_KEY_ARP_TIP_MASK,
1794                                sizeof(key->arp.tip));
1795                 fl_set_key_val(tb, &key->arp.op, TCA_FLOWER_KEY_ARP_OP,
1796                                &mask->arp.op, TCA_FLOWER_KEY_ARP_OP_MASK,
1797                                sizeof(key->arp.op));
1798                 fl_set_key_val(tb, key->arp.sha, TCA_FLOWER_KEY_ARP_SHA,
1799                                mask->arp.sha, TCA_FLOWER_KEY_ARP_SHA_MASK,
1800                                sizeof(key->arp.sha));
1801                 fl_set_key_val(tb, key->arp.tha, TCA_FLOWER_KEY_ARP_THA,
1802                                mask->arp.tha, TCA_FLOWER_KEY_ARP_THA_MASK,
1803                                sizeof(key->arp.tha));
1804         } else if (key->basic.ip_proto == IPPROTO_L2TP) {
1805                 fl_set_key_val(tb, &key->l2tpv3.session_id,
1806                                TCA_FLOWER_KEY_L2TPV3_SID,
1807                                &mask->l2tpv3.session_id, TCA_FLOWER_UNSPEC,
1808                                sizeof(key->l2tpv3.session_id));
1809         }
1810
1811         if (key->basic.ip_proto == IPPROTO_TCP ||
1812             key->basic.ip_proto == IPPROTO_UDP ||
1813             key->basic.ip_proto == IPPROTO_SCTP) {
1814                 ret = fl_set_key_port_range(tb, key, mask, extack);
1815                 if (ret)
1816                         return ret;
1817         }
1818
1819         if (tb[TCA_FLOWER_KEY_ENC_IPV4_SRC] ||
1820             tb[TCA_FLOWER_KEY_ENC_IPV4_DST]) {
1821                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
1822                 mask->enc_control.addr_type = ~0;
1823                 fl_set_key_val(tb, &key->enc_ipv4.src,
1824                                TCA_FLOWER_KEY_ENC_IPV4_SRC,
1825                                &mask->enc_ipv4.src,
1826                                TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
1827                                sizeof(key->enc_ipv4.src));
1828                 fl_set_key_val(tb, &key->enc_ipv4.dst,
1829                                TCA_FLOWER_KEY_ENC_IPV4_DST,
1830                                &mask->enc_ipv4.dst,
1831                                TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
1832                                sizeof(key->enc_ipv4.dst));
1833         }
1834
1835         if (tb[TCA_FLOWER_KEY_ENC_IPV6_SRC] ||
1836             tb[TCA_FLOWER_KEY_ENC_IPV6_DST]) {
1837                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
1838                 mask->enc_control.addr_type = ~0;
1839                 fl_set_key_val(tb, &key->enc_ipv6.src,
1840                                TCA_FLOWER_KEY_ENC_IPV6_SRC,
1841                                &mask->enc_ipv6.src,
1842                                TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
1843                                sizeof(key->enc_ipv6.src));
1844                 fl_set_key_val(tb, &key->enc_ipv6.dst,
1845                                TCA_FLOWER_KEY_ENC_IPV6_DST,
1846                                &mask->enc_ipv6.dst,
1847                                TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
1848                                sizeof(key->enc_ipv6.dst));
1849         }
1850
1851         fl_set_key_val(tb, &key->enc_key_id.keyid, TCA_FLOWER_KEY_ENC_KEY_ID,
1852                        &mask->enc_key_id.keyid, TCA_FLOWER_UNSPEC,
1853                        sizeof(key->enc_key_id.keyid));
1854
1855         fl_set_key_val(tb, &key->enc_tp.src, TCA_FLOWER_KEY_ENC_UDP_SRC_PORT,
1856                        &mask->enc_tp.src, TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK,
1857                        sizeof(key->enc_tp.src));
1858
1859         fl_set_key_val(tb, &key->enc_tp.dst, TCA_FLOWER_KEY_ENC_UDP_DST_PORT,
1860                        &mask->enc_tp.dst, TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK,
1861                        sizeof(key->enc_tp.dst));
1862
1863         fl_set_key_ip(tb, true, &key->enc_ip, &mask->enc_ip);
1864
1865         fl_set_key_val(tb, &key->hash.hash, TCA_FLOWER_KEY_HASH,
1866                        &mask->hash.hash, TCA_FLOWER_KEY_HASH_MASK,
1867                        sizeof(key->hash.hash));
1868
1869         if (tb[TCA_FLOWER_KEY_ENC_OPTS]) {
1870                 ret = fl_set_enc_opt(tb, key, mask, extack);
1871                 if (ret)
1872                         return ret;
1873         }
1874
1875         ret = fl_set_key_ct(tb, &key->ct, &mask->ct, extack);
1876         if (ret)
1877                 return ret;
1878
1879         if (tb[TCA_FLOWER_KEY_FLAGS])
1880                 ret = fl_set_key_flags(tb, &key->control.flags,
1881                                        &mask->control.flags, extack);
1882
1883         return ret;
1884 }
1885
1886 static void fl_mask_copy(struct fl_flow_mask *dst,
1887                          struct fl_flow_mask *src)
1888 {
1889         const void *psrc = fl_key_get_start(&src->key, src);
1890         void *pdst = fl_key_get_start(&dst->key, src);
1891
1892         memcpy(pdst, psrc, fl_mask_range(src));
1893         dst->range = src->range;
1894 }
1895
1896 static const struct rhashtable_params fl_ht_params = {
1897         .key_offset = offsetof(struct cls_fl_filter, mkey), /* base offset */
1898         .head_offset = offsetof(struct cls_fl_filter, ht_node),
1899         .automatic_shrinking = true,
1900 };
1901
1902 static int fl_init_mask_hashtable(struct fl_flow_mask *mask)
1903 {
1904         mask->filter_ht_params = fl_ht_params;
1905         mask->filter_ht_params.key_len = fl_mask_range(mask);
1906         mask->filter_ht_params.key_offset += mask->range.start;
1907
1908         return rhashtable_init(&mask->ht, &mask->filter_ht_params);
1909 }
1910
1911 #define FL_KEY_MEMBER_OFFSET(member) offsetof(struct fl_flow_key, member)
1912 #define FL_KEY_MEMBER_SIZE(member) sizeof_field(struct fl_flow_key, member)
1913
1914 #define FL_KEY_IS_MASKED(mask, member)                                          \
1915         memchr_inv(((char *)mask) + FL_KEY_MEMBER_OFFSET(member),               \
1916                    0, FL_KEY_MEMBER_SIZE(member))                               \
1917
1918 #define FL_KEY_SET(keys, cnt, id, member)                                       \
1919         do {                                                                    \
1920                 keys[cnt].key_id = id;                                          \
1921                 keys[cnt].offset = FL_KEY_MEMBER_OFFSET(member);                \
1922                 cnt++;                                                          \
1923         } while(0);
1924
1925 #define FL_KEY_SET_IF_MASKED(mask, keys, cnt, id, member)                       \
1926         do {                                                                    \
1927                 if (FL_KEY_IS_MASKED(mask, member))                             \
1928                         FL_KEY_SET(keys, cnt, id, member);                      \
1929         } while(0);
1930
1931 static void fl_init_dissector(struct flow_dissector *dissector,
1932                               struct fl_flow_key *mask)
1933 {
1934         struct flow_dissector_key keys[FLOW_DISSECTOR_KEY_MAX];
1935         size_t cnt = 0;
1936
1937         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1938                              FLOW_DISSECTOR_KEY_META, meta);
1939         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_CONTROL, control);
1940         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_BASIC, basic);
1941         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1942                              FLOW_DISSECTOR_KEY_ETH_ADDRS, eth);
1943         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1944                              FLOW_DISSECTOR_KEY_IPV4_ADDRS, ipv4);
1945         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1946                              FLOW_DISSECTOR_KEY_IPV6_ADDRS, ipv6);
1947         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1948                              FLOW_DISSECTOR_KEY_PORTS, tp);
1949         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1950                              FLOW_DISSECTOR_KEY_PORTS_RANGE, tp_range);
1951         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1952                              FLOW_DISSECTOR_KEY_IP, ip);
1953         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1954                              FLOW_DISSECTOR_KEY_TCP, tcp);
1955         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1956                              FLOW_DISSECTOR_KEY_ICMP, icmp);
1957         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1958                              FLOW_DISSECTOR_KEY_ARP, arp);
1959         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1960                              FLOW_DISSECTOR_KEY_MPLS, mpls);
1961         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1962                              FLOW_DISSECTOR_KEY_VLAN, vlan);
1963         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1964                              FLOW_DISSECTOR_KEY_CVLAN, cvlan);
1965         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1966                              FLOW_DISSECTOR_KEY_ENC_KEYID, enc_key_id);
1967         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1968                              FLOW_DISSECTOR_KEY_ENC_IPV4_ADDRS, enc_ipv4);
1969         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1970                              FLOW_DISSECTOR_KEY_ENC_IPV6_ADDRS, enc_ipv6);
1971         if (FL_KEY_IS_MASKED(mask, enc_ipv4) ||
1972             FL_KEY_IS_MASKED(mask, enc_ipv6))
1973                 FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_ENC_CONTROL,
1974                            enc_control);
1975         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1976                              FLOW_DISSECTOR_KEY_ENC_PORTS, enc_tp);
1977         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1978                              FLOW_DISSECTOR_KEY_ENC_IP, enc_ip);
1979         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1980                              FLOW_DISSECTOR_KEY_ENC_OPTS, enc_opts);
1981         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1982                              FLOW_DISSECTOR_KEY_CT, ct);
1983         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1984                              FLOW_DISSECTOR_KEY_HASH, hash);
1985         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1986                              FLOW_DISSECTOR_KEY_NUM_OF_VLANS, num_of_vlans);
1987         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1988                              FLOW_DISSECTOR_KEY_PPPOE, pppoe);
1989         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1990                              FLOW_DISSECTOR_KEY_L2TPV3, l2tpv3);
1991
1992         skb_flow_dissector_init(dissector, keys, cnt);
1993 }
1994
1995 static struct fl_flow_mask *fl_create_new_mask(struct cls_fl_head *head,
1996                                                struct fl_flow_mask *mask)
1997 {
1998         struct fl_flow_mask *newmask;
1999         int err;
2000
2001         newmask = kzalloc(sizeof(*newmask), GFP_KERNEL);
2002         if (!newmask)
2003                 return ERR_PTR(-ENOMEM);
2004
2005         fl_mask_copy(newmask, mask);
2006
2007         if ((newmask->key.tp_range.tp_min.dst &&
2008              newmask->key.tp_range.tp_max.dst) ||
2009             (newmask->key.tp_range.tp_min.src &&
2010              newmask->key.tp_range.tp_max.src))
2011                 newmask->flags |= TCA_FLOWER_MASK_FLAGS_RANGE;
2012
2013         err = fl_init_mask_hashtable(newmask);
2014         if (err)
2015                 goto errout_free;
2016
2017         fl_init_dissector(&newmask->dissector, &newmask->key);
2018
2019         INIT_LIST_HEAD_RCU(&newmask->filters);
2020
2021         refcount_set(&newmask->refcnt, 1);
2022         err = rhashtable_replace_fast(&head->ht, &mask->ht_node,
2023                                       &newmask->ht_node, mask_ht_params);
2024         if (err)
2025                 goto errout_destroy;
2026
2027         spin_lock(&head->masks_lock);
2028         list_add_tail_rcu(&newmask->list, &head->masks);
2029         spin_unlock(&head->masks_lock);
2030
2031         return newmask;
2032
2033 errout_destroy:
2034         rhashtable_destroy(&newmask->ht);
2035 errout_free:
2036         kfree(newmask);
2037
2038         return ERR_PTR(err);
2039 }
2040
2041 static int fl_check_assign_mask(struct cls_fl_head *head,
2042                                 struct cls_fl_filter *fnew,
2043                                 struct cls_fl_filter *fold,
2044                                 struct fl_flow_mask *mask)
2045 {
2046         struct fl_flow_mask *newmask;
2047         int ret = 0;
2048
2049         rcu_read_lock();
2050
2051         /* Insert mask as temporary node to prevent concurrent creation of mask
2052          * with same key. Any concurrent lookups with same key will return
2053          * -EAGAIN because mask's refcnt is zero.
2054          */
2055         fnew->mask = rhashtable_lookup_get_insert_fast(&head->ht,
2056                                                        &mask->ht_node,
2057                                                        mask_ht_params);
2058         if (!fnew->mask) {
2059                 rcu_read_unlock();
2060
2061                 if (fold) {
2062                         ret = -EINVAL;
2063                         goto errout_cleanup;
2064                 }
2065
2066                 newmask = fl_create_new_mask(head, mask);
2067                 if (IS_ERR(newmask)) {
2068                         ret = PTR_ERR(newmask);
2069                         goto errout_cleanup;
2070                 }
2071
2072                 fnew->mask = newmask;
2073                 return 0;
2074         } else if (IS_ERR(fnew->mask)) {
2075                 ret = PTR_ERR(fnew->mask);
2076         } else if (fold && fold->mask != fnew->mask) {
2077                 ret = -EINVAL;
2078         } else if (!refcount_inc_not_zero(&fnew->mask->refcnt)) {
2079                 /* Mask was deleted concurrently, try again */
2080                 ret = -EAGAIN;
2081         }
2082         rcu_read_unlock();
2083         return ret;
2084
2085 errout_cleanup:
2086         rhashtable_remove_fast(&head->ht, &mask->ht_node,
2087                                mask_ht_params);
2088         return ret;
2089 }
2090
2091 static int fl_set_parms(struct net *net, struct tcf_proto *tp,
2092                         struct cls_fl_filter *f, struct fl_flow_mask *mask,
2093                         unsigned long base, struct nlattr **tb,
2094                         struct nlattr *est,
2095                         struct fl_flow_tmplt *tmplt,
2096                         u32 flags, u32 fl_flags,
2097                         struct netlink_ext_ack *extack)
2098 {
2099         int err;
2100
2101         err = tcf_exts_validate_ex(net, tp, tb, est, &f->exts, flags,
2102                                    fl_flags, extack);
2103         if (err < 0)
2104                 return err;
2105
2106         if (tb[TCA_FLOWER_CLASSID]) {
2107                 f->res.classid = nla_get_u32(tb[TCA_FLOWER_CLASSID]);
2108                 if (flags & TCA_ACT_FLAGS_NO_RTNL)
2109                         rtnl_lock();
2110                 tcf_bind_filter(tp, &f->res, base);
2111                 if (flags & TCA_ACT_FLAGS_NO_RTNL)
2112                         rtnl_unlock();
2113         }
2114
2115         err = fl_set_key(net, tb, &f->key, &mask->key, extack);
2116         if (err)
2117                 return err;
2118
2119         fl_mask_update_range(mask);
2120         fl_set_masked_key(&f->mkey, &f->key, mask);
2121
2122         if (!fl_mask_fits_tmplt(tmplt, mask)) {
2123                 NL_SET_ERR_MSG_MOD(extack, "Mask does not fit the template");
2124                 return -EINVAL;
2125         }
2126
2127         return 0;
2128 }
2129
2130 static int fl_ht_insert_unique(struct cls_fl_filter *fnew,
2131                                struct cls_fl_filter *fold,
2132                                bool *in_ht)
2133 {
2134         struct fl_flow_mask *mask = fnew->mask;
2135         int err;
2136
2137         err = rhashtable_lookup_insert_fast(&mask->ht,
2138                                             &fnew->ht_node,
2139                                             mask->filter_ht_params);
2140         if (err) {
2141                 *in_ht = false;
2142                 /* It is okay if filter with same key exists when
2143                  * overwriting.
2144                  */
2145                 return fold && err == -EEXIST ? 0 : err;
2146         }
2147
2148         *in_ht = true;
2149         return 0;
2150 }
2151
2152 static int fl_change(struct net *net, struct sk_buff *in_skb,
2153                      struct tcf_proto *tp, unsigned long base,
2154                      u32 handle, struct nlattr **tca,
2155                      void **arg, u32 flags,
2156                      struct netlink_ext_ack *extack)
2157 {
2158         struct cls_fl_head *head = fl_head_dereference(tp);
2159         bool rtnl_held = !(flags & TCA_ACT_FLAGS_NO_RTNL);
2160         struct cls_fl_filter *fold = *arg;
2161         struct cls_fl_filter *fnew;
2162         struct fl_flow_mask *mask;
2163         struct nlattr **tb;
2164         bool in_ht;
2165         int err;
2166
2167         if (!tca[TCA_OPTIONS]) {
2168                 err = -EINVAL;
2169                 goto errout_fold;
2170         }
2171
2172         mask = kzalloc(sizeof(struct fl_flow_mask), GFP_KERNEL);
2173         if (!mask) {
2174                 err = -ENOBUFS;
2175                 goto errout_fold;
2176         }
2177
2178         tb = kcalloc(TCA_FLOWER_MAX + 1, sizeof(struct nlattr *), GFP_KERNEL);
2179         if (!tb) {
2180                 err = -ENOBUFS;
2181                 goto errout_mask_alloc;
2182         }
2183
2184         err = nla_parse_nested_deprecated(tb, TCA_FLOWER_MAX,
2185                                           tca[TCA_OPTIONS], fl_policy, NULL);
2186         if (err < 0)
2187                 goto errout_tb;
2188
2189         if (fold && handle && fold->handle != handle) {
2190                 err = -EINVAL;
2191                 goto errout_tb;
2192         }
2193
2194         fnew = kzalloc(sizeof(*fnew), GFP_KERNEL);
2195         if (!fnew) {
2196                 err = -ENOBUFS;
2197                 goto errout_tb;
2198         }
2199         INIT_LIST_HEAD(&fnew->hw_list);
2200         refcount_set(&fnew->refcnt, 1);
2201
2202         if (tb[TCA_FLOWER_FLAGS]) {
2203                 fnew->flags = nla_get_u32(tb[TCA_FLOWER_FLAGS]);
2204
2205                 if (!tc_flags_valid(fnew->flags)) {
2206                         kfree(fnew);
2207                         err = -EINVAL;
2208                         goto errout_tb;
2209                 }
2210         }
2211
2212         if (!fold) {
2213                 spin_lock(&tp->lock);
2214                 if (!handle) {
2215                         handle = 1;
2216                         err = idr_alloc_u32(&head->handle_idr, NULL, &handle,
2217                                             INT_MAX, GFP_ATOMIC);
2218                 } else {
2219                         err = idr_alloc_u32(&head->handle_idr, NULL, &handle,
2220                                             handle, GFP_ATOMIC);
2221
2222                         /* Filter with specified handle was concurrently
2223                          * inserted after initial check in cls_api. This is not
2224                          * necessarily an error if NLM_F_EXCL is not set in
2225                          * message flags. Returning EAGAIN will cause cls_api to
2226                          * try to update concurrently inserted rule.
2227                          */
2228                         if (err == -ENOSPC)
2229                                 err = -EAGAIN;
2230                 }
2231                 spin_unlock(&tp->lock);
2232
2233                 if (err) {
2234                         kfree(fnew);
2235                         goto errout_tb;
2236                 }
2237         }
2238         fnew->handle = handle;
2239
2240         err = tcf_exts_init_ex(&fnew->exts, net, TCA_FLOWER_ACT, 0, tp, handle,
2241                                !tc_skip_hw(fnew->flags));
2242         if (err < 0)
2243                 goto errout_idr;
2244
2245         err = fl_set_parms(net, tp, fnew, mask, base, tb, tca[TCA_RATE],
2246                            tp->chain->tmplt_priv, flags, fnew->flags,
2247                            extack);
2248         if (err)
2249                 goto errout_idr;
2250
2251         err = fl_check_assign_mask(head, fnew, fold, mask);
2252         if (err)
2253                 goto errout_idr;
2254
2255         err = fl_ht_insert_unique(fnew, fold, &in_ht);
2256         if (err)
2257                 goto errout_mask;
2258
2259         if (!tc_skip_hw(fnew->flags)) {
2260                 err = fl_hw_replace_filter(tp, fnew, rtnl_held, extack);
2261                 if (err)
2262                         goto errout_ht;
2263         }
2264
2265         if (!tc_in_hw(fnew->flags))
2266                 fnew->flags |= TCA_CLS_FLAGS_NOT_IN_HW;
2267
2268         spin_lock(&tp->lock);
2269
2270         /* tp was deleted concurrently. -EAGAIN will cause caller to lookup
2271          * proto again or create new one, if necessary.
2272          */
2273         if (tp->deleting) {
2274                 err = -EAGAIN;
2275                 goto errout_hw;
2276         }
2277
2278         if (fold) {
2279                 /* Fold filter was deleted concurrently. Retry lookup. */
2280                 if (fold->deleted) {
2281                         err = -EAGAIN;
2282                         goto errout_hw;
2283                 }
2284
2285                 fnew->handle = handle;
2286
2287                 if (!in_ht) {
2288                         struct rhashtable_params params =
2289                                 fnew->mask->filter_ht_params;
2290
2291                         err = rhashtable_insert_fast(&fnew->mask->ht,
2292                                                      &fnew->ht_node,
2293                                                      params);
2294                         if (err)
2295                                 goto errout_hw;
2296                         in_ht = true;
2297                 }
2298
2299                 refcount_inc(&fnew->refcnt);
2300                 rhashtable_remove_fast(&fold->mask->ht,
2301                                        &fold->ht_node,
2302                                        fold->mask->filter_ht_params);
2303                 idr_replace(&head->handle_idr, fnew, fnew->handle);
2304                 list_replace_rcu(&fold->list, &fnew->list);
2305                 fold->deleted = true;
2306
2307                 spin_unlock(&tp->lock);
2308
2309                 fl_mask_put(head, fold->mask);
2310                 if (!tc_skip_hw(fold->flags))
2311                         fl_hw_destroy_filter(tp, fold, rtnl_held, NULL);
2312                 tcf_unbind_filter(tp, &fold->res);
2313                 /* Caller holds reference to fold, so refcnt is always > 0
2314                  * after this.
2315                  */
2316                 refcount_dec(&fold->refcnt);
2317                 __fl_put(fold);
2318         } else {
2319                 idr_replace(&head->handle_idr, fnew, fnew->handle);
2320
2321                 refcount_inc(&fnew->refcnt);
2322                 list_add_tail_rcu(&fnew->list, &fnew->mask->filters);
2323                 spin_unlock(&tp->lock);
2324         }
2325
2326         *arg = fnew;
2327
2328         kfree(tb);
2329         tcf_queue_work(&mask->rwork, fl_uninit_mask_free_work);
2330         return 0;
2331
2332 errout_ht:
2333         spin_lock(&tp->lock);
2334 errout_hw:
2335         fnew->deleted = true;
2336         spin_unlock(&tp->lock);
2337         if (!tc_skip_hw(fnew->flags))
2338                 fl_hw_destroy_filter(tp, fnew, rtnl_held, NULL);
2339         if (in_ht)
2340                 rhashtable_remove_fast(&fnew->mask->ht, &fnew->ht_node,
2341                                        fnew->mask->filter_ht_params);
2342 errout_mask:
2343         fl_mask_put(head, fnew->mask);
2344 errout_idr:
2345         if (!fold)
2346                 idr_remove(&head->handle_idr, fnew->handle);
2347         __fl_put(fnew);
2348 errout_tb:
2349         kfree(tb);
2350 errout_mask_alloc:
2351         tcf_queue_work(&mask->rwork, fl_uninit_mask_free_work);
2352 errout_fold:
2353         if (fold)
2354                 __fl_put(fold);
2355         return err;
2356 }
2357
2358 static int fl_delete(struct tcf_proto *tp, void *arg, bool *last,
2359                      bool rtnl_held, struct netlink_ext_ack *extack)
2360 {
2361         struct cls_fl_head *head = fl_head_dereference(tp);
2362         struct cls_fl_filter *f = arg;
2363         bool last_on_mask;
2364         int err = 0;
2365
2366         err = __fl_delete(tp, f, &last_on_mask, rtnl_held, extack);
2367         *last = list_empty(&head->masks);
2368         __fl_put(f);
2369
2370         return err;
2371 }
2372
2373 static void fl_walk(struct tcf_proto *tp, struct tcf_walker *arg,
2374                     bool rtnl_held)
2375 {
2376         struct cls_fl_head *head = fl_head_dereference(tp);
2377         unsigned long id = arg->cookie, tmp;
2378         struct cls_fl_filter *f;
2379
2380         arg->count = arg->skip;
2381
2382         rcu_read_lock();
2383         idr_for_each_entry_continue_ul(&head->handle_idr, f, tmp, id) {
2384                 /* don't return filters that are being deleted */
2385                 if (!f || !refcount_inc_not_zero(&f->refcnt))
2386                         continue;
2387                 rcu_read_unlock();
2388
2389                 if (arg->fn(tp, f, arg) < 0) {
2390                         __fl_put(f);
2391                         arg->stop = 1;
2392                         rcu_read_lock();
2393                         break;
2394                 }
2395                 __fl_put(f);
2396                 arg->count++;
2397                 rcu_read_lock();
2398         }
2399         rcu_read_unlock();
2400         arg->cookie = id;
2401 }
2402
2403 static struct cls_fl_filter *
2404 fl_get_next_hw_filter(struct tcf_proto *tp, struct cls_fl_filter *f, bool add)
2405 {
2406         struct cls_fl_head *head = fl_head_dereference(tp);
2407
2408         spin_lock(&tp->lock);
2409         if (list_empty(&head->hw_filters)) {
2410                 spin_unlock(&tp->lock);
2411                 return NULL;
2412         }
2413
2414         if (!f)
2415                 f = list_entry(&head->hw_filters, struct cls_fl_filter,
2416                                hw_list);
2417         list_for_each_entry_continue(f, &head->hw_filters, hw_list) {
2418                 if (!(add && f->deleted) && refcount_inc_not_zero(&f->refcnt)) {
2419                         spin_unlock(&tp->lock);
2420                         return f;
2421                 }
2422         }
2423
2424         spin_unlock(&tp->lock);
2425         return NULL;
2426 }
2427
2428 static int fl_reoffload(struct tcf_proto *tp, bool add, flow_setup_cb_t *cb,
2429                         void *cb_priv, struct netlink_ext_ack *extack)
2430 {
2431         struct tcf_block *block = tp->chain->block;
2432         struct flow_cls_offload cls_flower = {};
2433         struct cls_fl_filter *f = NULL;
2434         int err;
2435
2436         /* hw_filters list can only be changed by hw offload functions after
2437          * obtaining rtnl lock. Make sure it is not changed while reoffload is
2438          * iterating it.
2439          */
2440         ASSERT_RTNL();
2441
2442         while ((f = fl_get_next_hw_filter(tp, f, add))) {
2443                 cls_flower.rule =
2444                         flow_rule_alloc(tcf_exts_num_actions(&f->exts));
2445                 if (!cls_flower.rule) {
2446                         __fl_put(f);
2447                         return -ENOMEM;
2448                 }
2449
2450                 tc_cls_common_offload_init(&cls_flower.common, tp, f->flags,
2451                                            extack);
2452                 cls_flower.command = add ?
2453                         FLOW_CLS_REPLACE : FLOW_CLS_DESTROY;
2454                 cls_flower.cookie = (unsigned long)f;
2455                 cls_flower.rule->match.dissector = &f->mask->dissector;
2456                 cls_flower.rule->match.mask = &f->mask->key;
2457                 cls_flower.rule->match.key = &f->mkey;
2458
2459                 err = tc_setup_offload_action(&cls_flower.rule->action, &f->exts,
2460                                               cls_flower.common.extack);
2461                 if (err) {
2462                         kfree(cls_flower.rule);
2463                         if (tc_skip_sw(f->flags)) {
2464                                 __fl_put(f);
2465                                 return err;
2466                         }
2467                         goto next_flow;
2468                 }
2469
2470                 cls_flower.classid = f->res.classid;
2471
2472                 err = tc_setup_cb_reoffload(block, tp, add, cb,
2473                                             TC_SETUP_CLSFLOWER, &cls_flower,
2474                                             cb_priv, &f->flags,
2475                                             &f->in_hw_count);
2476                 tc_cleanup_offload_action(&cls_flower.rule->action);
2477                 kfree(cls_flower.rule);
2478
2479                 if (err) {
2480                         __fl_put(f);
2481                         return err;
2482                 }
2483 next_flow:
2484                 __fl_put(f);
2485         }
2486
2487         return 0;
2488 }
2489
2490 static void fl_hw_add(struct tcf_proto *tp, void *type_data)
2491 {
2492         struct flow_cls_offload *cls_flower = type_data;
2493         struct cls_fl_filter *f =
2494                 (struct cls_fl_filter *) cls_flower->cookie;
2495         struct cls_fl_head *head = fl_head_dereference(tp);
2496
2497         spin_lock(&tp->lock);
2498         list_add(&f->hw_list, &head->hw_filters);
2499         spin_unlock(&tp->lock);
2500 }
2501
2502 static void fl_hw_del(struct tcf_proto *tp, void *type_data)
2503 {
2504         struct flow_cls_offload *cls_flower = type_data;
2505         struct cls_fl_filter *f =
2506                 (struct cls_fl_filter *) cls_flower->cookie;
2507
2508         spin_lock(&tp->lock);
2509         if (!list_empty(&f->hw_list))
2510                 list_del_init(&f->hw_list);
2511         spin_unlock(&tp->lock);
2512 }
2513
2514 static int fl_hw_create_tmplt(struct tcf_chain *chain,
2515                               struct fl_flow_tmplt *tmplt)
2516 {
2517         struct flow_cls_offload cls_flower = {};
2518         struct tcf_block *block = chain->block;
2519
2520         cls_flower.rule = flow_rule_alloc(0);
2521         if (!cls_flower.rule)
2522                 return -ENOMEM;
2523
2524         cls_flower.common.chain_index = chain->index;
2525         cls_flower.command = FLOW_CLS_TMPLT_CREATE;
2526         cls_flower.cookie = (unsigned long) tmplt;
2527         cls_flower.rule->match.dissector = &tmplt->dissector;
2528         cls_flower.rule->match.mask = &tmplt->mask;
2529         cls_flower.rule->match.key = &tmplt->dummy_key;
2530
2531         /* We don't care if driver (any of them) fails to handle this
2532          * call. It serves just as a hint for it.
2533          */
2534         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false, true);
2535         kfree(cls_flower.rule);
2536
2537         return 0;
2538 }
2539
2540 static void fl_hw_destroy_tmplt(struct tcf_chain *chain,
2541                                 struct fl_flow_tmplt *tmplt)
2542 {
2543         struct flow_cls_offload cls_flower = {};
2544         struct tcf_block *block = chain->block;
2545
2546         cls_flower.common.chain_index = chain->index;
2547         cls_flower.command = FLOW_CLS_TMPLT_DESTROY;
2548         cls_flower.cookie = (unsigned long) tmplt;
2549
2550         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false, true);
2551 }
2552
2553 static void *fl_tmplt_create(struct net *net, struct tcf_chain *chain,
2554                              struct nlattr **tca,
2555                              struct netlink_ext_ack *extack)
2556 {
2557         struct fl_flow_tmplt *tmplt;
2558         struct nlattr **tb;
2559         int err;
2560
2561         if (!tca[TCA_OPTIONS])
2562                 return ERR_PTR(-EINVAL);
2563
2564         tb = kcalloc(TCA_FLOWER_MAX + 1, sizeof(struct nlattr *), GFP_KERNEL);
2565         if (!tb)
2566                 return ERR_PTR(-ENOBUFS);
2567         err = nla_parse_nested_deprecated(tb, TCA_FLOWER_MAX,
2568                                           tca[TCA_OPTIONS], fl_policy, NULL);
2569         if (err)
2570                 goto errout_tb;
2571
2572         tmplt = kzalloc(sizeof(*tmplt), GFP_KERNEL);
2573         if (!tmplt) {
2574                 err = -ENOMEM;
2575                 goto errout_tb;
2576         }
2577         tmplt->chain = chain;
2578         err = fl_set_key(net, tb, &tmplt->dummy_key, &tmplt->mask, extack);
2579         if (err)
2580                 goto errout_tmplt;
2581
2582         fl_init_dissector(&tmplt->dissector, &tmplt->mask);
2583
2584         err = fl_hw_create_tmplt(chain, tmplt);
2585         if (err)
2586                 goto errout_tmplt;
2587
2588         kfree(tb);
2589         return tmplt;
2590
2591 errout_tmplt:
2592         kfree(tmplt);
2593 errout_tb:
2594         kfree(tb);
2595         return ERR_PTR(err);
2596 }
2597
2598 static void fl_tmplt_destroy(void *tmplt_priv)
2599 {
2600         struct fl_flow_tmplt *tmplt = tmplt_priv;
2601
2602         fl_hw_destroy_tmplt(tmplt->chain, tmplt);
2603         kfree(tmplt);
2604 }
2605
2606 static int fl_dump_key_val(struct sk_buff *skb,
2607                            void *val, int val_type,
2608                            void *mask, int mask_type, int len)
2609 {
2610         int err;
2611
2612         if (!memchr_inv(mask, 0, len))
2613                 return 0;
2614         err = nla_put(skb, val_type, len, val);
2615         if (err)
2616                 return err;
2617         if (mask_type != TCA_FLOWER_UNSPEC) {
2618                 err = nla_put(skb, mask_type, len, mask);
2619                 if (err)
2620                         return err;
2621         }
2622         return 0;
2623 }
2624
2625 static int fl_dump_key_port_range(struct sk_buff *skb, struct fl_flow_key *key,
2626                                   struct fl_flow_key *mask)
2627 {
2628         if (fl_dump_key_val(skb, &key->tp_range.tp_min.dst,
2629                             TCA_FLOWER_KEY_PORT_DST_MIN,
2630                             &mask->tp_range.tp_min.dst, TCA_FLOWER_UNSPEC,
2631                             sizeof(key->tp_range.tp_min.dst)) ||
2632             fl_dump_key_val(skb, &key->tp_range.tp_max.dst,
2633                             TCA_FLOWER_KEY_PORT_DST_MAX,
2634                             &mask->tp_range.tp_max.dst, TCA_FLOWER_UNSPEC,
2635                             sizeof(key->tp_range.tp_max.dst)) ||
2636             fl_dump_key_val(skb, &key->tp_range.tp_min.src,
2637                             TCA_FLOWER_KEY_PORT_SRC_MIN,
2638                             &mask->tp_range.tp_min.src, TCA_FLOWER_UNSPEC,
2639                             sizeof(key->tp_range.tp_min.src)) ||
2640             fl_dump_key_val(skb, &key->tp_range.tp_max.src,
2641                             TCA_FLOWER_KEY_PORT_SRC_MAX,
2642                             &mask->tp_range.tp_max.src, TCA_FLOWER_UNSPEC,
2643                             sizeof(key->tp_range.tp_max.src)))
2644                 return -1;
2645
2646         return 0;
2647 }
2648
2649 static int fl_dump_key_mpls_opt_lse(struct sk_buff *skb,
2650                                     struct flow_dissector_key_mpls *mpls_key,
2651                                     struct flow_dissector_key_mpls *mpls_mask,
2652                                     u8 lse_index)
2653 {
2654         struct flow_dissector_mpls_lse *lse_mask = &mpls_mask->ls[lse_index];
2655         struct flow_dissector_mpls_lse *lse_key = &mpls_key->ls[lse_index];
2656         int err;
2657
2658         err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_DEPTH,
2659                          lse_index + 1);
2660         if (err)
2661                 return err;
2662
2663         if (lse_mask->mpls_ttl) {
2664                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_TTL,
2665                                  lse_key->mpls_ttl);
2666                 if (err)
2667                         return err;
2668         }
2669         if (lse_mask->mpls_bos) {
2670                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_BOS,
2671                                  lse_key->mpls_bos);
2672                 if (err)
2673                         return err;
2674         }
2675         if (lse_mask->mpls_tc) {
2676                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_TC,
2677                                  lse_key->mpls_tc);
2678                 if (err)
2679                         return err;
2680         }
2681         if (lse_mask->mpls_label) {
2682                 err = nla_put_u32(skb, TCA_FLOWER_KEY_MPLS_OPT_LSE_LABEL,
2683                                   lse_key->mpls_label);
2684                 if (err)
2685                         return err;
2686         }
2687
2688         return 0;
2689 }
2690
2691 static int fl_dump_key_mpls_opts(struct sk_buff *skb,
2692                                  struct flow_dissector_key_mpls *mpls_key,
2693                                  struct flow_dissector_key_mpls *mpls_mask)
2694 {
2695         struct nlattr *opts;
2696         struct nlattr *lse;
2697         u8 lse_index;
2698         int err;
2699
2700         opts = nla_nest_start(skb, TCA_FLOWER_KEY_MPLS_OPTS);
2701         if (!opts)
2702                 return -EMSGSIZE;
2703
2704         for (lse_index = 0; lse_index < FLOW_DIS_MPLS_MAX; lse_index++) {
2705                 if (!(mpls_mask->used_lses & 1 << lse_index))
2706                         continue;
2707
2708                 lse = nla_nest_start(skb, TCA_FLOWER_KEY_MPLS_OPTS_LSE);
2709                 if (!lse) {
2710                         err = -EMSGSIZE;
2711                         goto err_opts;
2712                 }
2713
2714                 err = fl_dump_key_mpls_opt_lse(skb, mpls_key, mpls_mask,
2715                                                lse_index);
2716                 if (err)
2717                         goto err_opts_lse;
2718                 nla_nest_end(skb, lse);
2719         }
2720         nla_nest_end(skb, opts);
2721
2722         return 0;
2723
2724 err_opts_lse:
2725         nla_nest_cancel(skb, lse);
2726 err_opts:
2727         nla_nest_cancel(skb, opts);
2728
2729         return err;
2730 }
2731
2732 static int fl_dump_key_mpls(struct sk_buff *skb,
2733                             struct flow_dissector_key_mpls *mpls_key,
2734                             struct flow_dissector_key_mpls *mpls_mask)
2735 {
2736         struct flow_dissector_mpls_lse *lse_mask;
2737         struct flow_dissector_mpls_lse *lse_key;
2738         int err;
2739
2740         if (!mpls_mask->used_lses)
2741                 return 0;
2742
2743         lse_mask = &mpls_mask->ls[0];
2744         lse_key = &mpls_key->ls[0];
2745
2746         /* For backward compatibility, don't use the MPLS nested attributes if
2747          * the rule can be expressed using the old attributes.
2748          */
2749         if (mpls_mask->used_lses & ~1 ||
2750             (!lse_mask->mpls_ttl && !lse_mask->mpls_bos &&
2751              !lse_mask->mpls_tc && !lse_mask->mpls_label))
2752                 return fl_dump_key_mpls_opts(skb, mpls_key, mpls_mask);
2753
2754         if (lse_mask->mpls_ttl) {
2755                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_TTL,
2756                                  lse_key->mpls_ttl);
2757                 if (err)
2758                         return err;
2759         }
2760         if (lse_mask->mpls_tc) {
2761                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_TC,
2762                                  lse_key->mpls_tc);
2763                 if (err)
2764                         return err;
2765         }
2766         if (lse_mask->mpls_label) {
2767                 err = nla_put_u32(skb, TCA_FLOWER_KEY_MPLS_LABEL,
2768                                   lse_key->mpls_label);
2769                 if (err)
2770                         return err;
2771         }
2772         if (lse_mask->mpls_bos) {
2773                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_BOS,
2774                                  lse_key->mpls_bos);
2775                 if (err)
2776                         return err;
2777         }
2778         return 0;
2779 }
2780
2781 static int fl_dump_key_ip(struct sk_buff *skb, bool encap,
2782                           struct flow_dissector_key_ip *key,
2783                           struct flow_dissector_key_ip *mask)
2784 {
2785         int tos_key = encap ? TCA_FLOWER_KEY_ENC_IP_TOS : TCA_FLOWER_KEY_IP_TOS;
2786         int ttl_key = encap ? TCA_FLOWER_KEY_ENC_IP_TTL : TCA_FLOWER_KEY_IP_TTL;
2787         int tos_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TOS_MASK : TCA_FLOWER_KEY_IP_TOS_MASK;
2788         int ttl_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TTL_MASK : TCA_FLOWER_KEY_IP_TTL_MASK;
2789
2790         if (fl_dump_key_val(skb, &key->tos, tos_key, &mask->tos, tos_mask, sizeof(key->tos)) ||
2791             fl_dump_key_val(skb, &key->ttl, ttl_key, &mask->ttl, ttl_mask, sizeof(key->ttl)))
2792                 return -1;
2793
2794         return 0;
2795 }
2796
2797 static int fl_dump_key_vlan(struct sk_buff *skb,
2798                             int vlan_id_key, int vlan_prio_key,
2799                             struct flow_dissector_key_vlan *vlan_key,
2800                             struct flow_dissector_key_vlan *vlan_mask)
2801 {
2802         int err;
2803
2804         if (!memchr_inv(vlan_mask, 0, sizeof(*vlan_mask)))
2805                 return 0;
2806         if (vlan_mask->vlan_id) {
2807                 err = nla_put_u16(skb, vlan_id_key,
2808                                   vlan_key->vlan_id);
2809                 if (err)
2810                         return err;
2811         }
2812         if (vlan_mask->vlan_priority) {
2813                 err = nla_put_u8(skb, vlan_prio_key,
2814                                  vlan_key->vlan_priority);
2815                 if (err)
2816                         return err;
2817         }
2818         return 0;
2819 }
2820
2821 static void fl_get_key_flag(u32 dissector_key, u32 dissector_mask,
2822                             u32 *flower_key, u32 *flower_mask,
2823                             u32 flower_flag_bit, u32 dissector_flag_bit)
2824 {
2825         if (dissector_mask & dissector_flag_bit) {
2826                 *flower_mask |= flower_flag_bit;
2827                 if (dissector_key & dissector_flag_bit)
2828                         *flower_key |= flower_flag_bit;
2829         }
2830 }
2831
2832 static int fl_dump_key_flags(struct sk_buff *skb, u32 flags_key, u32 flags_mask)
2833 {
2834         u32 key, mask;
2835         __be32 _key, _mask;
2836         int err;
2837
2838         if (!memchr_inv(&flags_mask, 0, sizeof(flags_mask)))
2839                 return 0;
2840
2841         key = 0;
2842         mask = 0;
2843
2844         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
2845                         TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT, FLOW_DIS_IS_FRAGMENT);
2846         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
2847                         TCA_FLOWER_KEY_FLAGS_FRAG_IS_FIRST,
2848                         FLOW_DIS_FIRST_FRAG);
2849
2850         _key = cpu_to_be32(key);
2851         _mask = cpu_to_be32(mask);
2852
2853         err = nla_put(skb, TCA_FLOWER_KEY_FLAGS, 4, &_key);
2854         if (err)
2855                 return err;
2856
2857         return nla_put(skb, TCA_FLOWER_KEY_FLAGS_MASK, 4, &_mask);
2858 }
2859
2860 static int fl_dump_key_geneve_opt(struct sk_buff *skb,
2861                                   struct flow_dissector_key_enc_opts *enc_opts)
2862 {
2863         struct geneve_opt *opt;
2864         struct nlattr *nest;
2865         int opt_off = 0;
2866
2867         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_GENEVE);
2868         if (!nest)
2869                 goto nla_put_failure;
2870
2871         while (enc_opts->len > opt_off) {
2872                 opt = (struct geneve_opt *)&enc_opts->data[opt_off];
2873
2874                 if (nla_put_be16(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS,
2875                                  opt->opt_class))
2876                         goto nla_put_failure;
2877                 if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE,
2878                                opt->type))
2879                         goto nla_put_failure;
2880                 if (nla_put(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA,
2881                             opt->length * 4, opt->opt_data))
2882                         goto nla_put_failure;
2883
2884                 opt_off += sizeof(struct geneve_opt) + opt->length * 4;
2885         }
2886         nla_nest_end(skb, nest);
2887         return 0;
2888
2889 nla_put_failure:
2890         nla_nest_cancel(skb, nest);
2891         return -EMSGSIZE;
2892 }
2893
2894 static int fl_dump_key_vxlan_opt(struct sk_buff *skb,
2895                                  struct flow_dissector_key_enc_opts *enc_opts)
2896 {
2897         struct vxlan_metadata *md;
2898         struct nlattr *nest;
2899
2900         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_VXLAN);
2901         if (!nest)
2902                 goto nla_put_failure;
2903
2904         md = (struct vxlan_metadata *)&enc_opts->data[0];
2905         if (nla_put_u32(skb, TCA_FLOWER_KEY_ENC_OPT_VXLAN_GBP, md->gbp))
2906                 goto nla_put_failure;
2907
2908         nla_nest_end(skb, nest);
2909         return 0;
2910
2911 nla_put_failure:
2912         nla_nest_cancel(skb, nest);
2913         return -EMSGSIZE;
2914 }
2915
2916 static int fl_dump_key_erspan_opt(struct sk_buff *skb,
2917                                   struct flow_dissector_key_enc_opts *enc_opts)
2918 {
2919         struct erspan_metadata *md;
2920         struct nlattr *nest;
2921
2922         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_ERSPAN);
2923         if (!nest)
2924                 goto nla_put_failure;
2925
2926         md = (struct erspan_metadata *)&enc_opts->data[0];
2927         if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_VER, md->version))
2928                 goto nla_put_failure;
2929
2930         if (md->version == 1 &&
2931             nla_put_be32(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_INDEX, md->u.index))
2932                 goto nla_put_failure;
2933
2934         if (md->version == 2 &&
2935             (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_DIR,
2936                         md->u.md2.dir) ||
2937              nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_ERSPAN_HWID,
2938                         get_hwid(&md->u.md2))))
2939                 goto nla_put_failure;
2940
2941         nla_nest_end(skb, nest);
2942         return 0;
2943
2944 nla_put_failure:
2945         nla_nest_cancel(skb, nest);
2946         return -EMSGSIZE;
2947 }
2948
2949 static int fl_dump_key_gtp_opt(struct sk_buff *skb,
2950                                struct flow_dissector_key_enc_opts *enc_opts)
2951
2952 {
2953         struct gtp_pdu_session_info *session_info;
2954         struct nlattr *nest;
2955
2956         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_GTP);
2957         if (!nest)
2958                 goto nla_put_failure;
2959
2960         session_info = (struct gtp_pdu_session_info *)&enc_opts->data[0];
2961
2962         if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_GTP_PDU_TYPE,
2963                        session_info->pdu_type))
2964                 goto nla_put_failure;
2965
2966         if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_GTP_QFI, session_info->qfi))
2967                 goto nla_put_failure;
2968
2969         nla_nest_end(skb, nest);
2970         return 0;
2971
2972 nla_put_failure:
2973         nla_nest_cancel(skb, nest);
2974         return -EMSGSIZE;
2975 }
2976
2977 static int fl_dump_key_ct(struct sk_buff *skb,
2978                           struct flow_dissector_key_ct *key,
2979                           struct flow_dissector_key_ct *mask)
2980 {
2981         if (IS_ENABLED(CONFIG_NF_CONNTRACK) &&
2982             fl_dump_key_val(skb, &key->ct_state, TCA_FLOWER_KEY_CT_STATE,
2983                             &mask->ct_state, TCA_FLOWER_KEY_CT_STATE_MASK,
2984                             sizeof(key->ct_state)))
2985                 goto nla_put_failure;
2986
2987         if (IS_ENABLED(CONFIG_NF_CONNTRACK_ZONES) &&
2988             fl_dump_key_val(skb, &key->ct_zone, TCA_FLOWER_KEY_CT_ZONE,
2989                             &mask->ct_zone, TCA_FLOWER_KEY_CT_ZONE_MASK,
2990                             sizeof(key->ct_zone)))
2991                 goto nla_put_failure;
2992
2993         if (IS_ENABLED(CONFIG_NF_CONNTRACK_MARK) &&
2994             fl_dump_key_val(skb, &key->ct_mark, TCA_FLOWER_KEY_CT_MARK,
2995                             &mask->ct_mark, TCA_FLOWER_KEY_CT_MARK_MASK,
2996                             sizeof(key->ct_mark)))
2997                 goto nla_put_failure;
2998
2999         if (IS_ENABLED(CONFIG_NF_CONNTRACK_LABELS) &&
3000             fl_dump_key_val(skb, &key->ct_labels, TCA_FLOWER_KEY_CT_LABELS,
3001                             &mask->ct_labels, TCA_FLOWER_KEY_CT_LABELS_MASK,
3002                             sizeof(key->ct_labels)))
3003                 goto nla_put_failure;
3004
3005         return 0;
3006
3007 nla_put_failure:
3008         return -EMSGSIZE;
3009 }
3010
3011 static int fl_dump_key_options(struct sk_buff *skb, int enc_opt_type,
3012                                struct flow_dissector_key_enc_opts *enc_opts)
3013 {
3014         struct nlattr *nest;
3015         int err;
3016
3017         if (!enc_opts->len)
3018                 return 0;
3019
3020         nest = nla_nest_start_noflag(skb, enc_opt_type);
3021         if (!nest)
3022                 goto nla_put_failure;
3023
3024         switch (enc_opts->dst_opt_type) {
3025         case TUNNEL_GENEVE_OPT:
3026                 err = fl_dump_key_geneve_opt(skb, enc_opts);
3027                 if (err)
3028                         goto nla_put_failure;
3029                 break;
3030         case TUNNEL_VXLAN_OPT:
3031                 err = fl_dump_key_vxlan_opt(skb, enc_opts);
3032                 if (err)
3033                         goto nla_put_failure;
3034                 break;
3035         case TUNNEL_ERSPAN_OPT:
3036                 err = fl_dump_key_erspan_opt(skb, enc_opts);
3037                 if (err)
3038                         goto nla_put_failure;
3039                 break;
3040         case TUNNEL_GTP_OPT:
3041                 err = fl_dump_key_gtp_opt(skb, enc_opts);
3042                 if (err)
3043                         goto nla_put_failure;
3044                 break;
3045         default:
3046                 goto nla_put_failure;
3047         }
3048         nla_nest_end(skb, nest);
3049         return 0;
3050
3051 nla_put_failure:
3052         nla_nest_cancel(skb, nest);
3053         return -EMSGSIZE;
3054 }
3055
3056 static int fl_dump_key_enc_opt(struct sk_buff *skb,
3057                                struct flow_dissector_key_enc_opts *key_opts,
3058                                struct flow_dissector_key_enc_opts *msk_opts)
3059 {
3060         int err;
3061
3062         err = fl_dump_key_options(skb, TCA_FLOWER_KEY_ENC_OPTS, key_opts);
3063         if (err)
3064                 return err;
3065
3066         return fl_dump_key_options(skb, TCA_FLOWER_KEY_ENC_OPTS_MASK, msk_opts);
3067 }
3068
3069 static int fl_dump_key(struct sk_buff *skb, struct net *net,
3070                        struct fl_flow_key *key, struct fl_flow_key *mask)
3071 {
3072         if (mask->meta.ingress_ifindex) {
3073                 struct net_device *dev;
3074
3075                 dev = __dev_get_by_index(net, key->meta.ingress_ifindex);
3076                 if (dev && nla_put_string(skb, TCA_FLOWER_INDEV, dev->name))
3077                         goto nla_put_failure;
3078         }
3079
3080         if (fl_dump_key_val(skb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
3081                             mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
3082                             sizeof(key->eth.dst)) ||
3083             fl_dump_key_val(skb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
3084                             mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
3085                             sizeof(key->eth.src)) ||
3086             fl_dump_key_val(skb, &key->basic.n_proto, TCA_FLOWER_KEY_ETH_TYPE,
3087                             &mask->basic.n_proto, TCA_FLOWER_UNSPEC,
3088                             sizeof(key->basic.n_proto)))
3089                 goto nla_put_failure;
3090
3091         if (mask->num_of_vlans.num_of_vlans) {
3092                 if (nla_put_u8(skb, TCA_FLOWER_KEY_NUM_OF_VLANS, key->num_of_vlans.num_of_vlans))
3093                         goto nla_put_failure;
3094         }
3095
3096         if (fl_dump_key_mpls(skb, &key->mpls, &mask->mpls))
3097                 goto nla_put_failure;
3098
3099         if (fl_dump_key_vlan(skb, TCA_FLOWER_KEY_VLAN_ID,
3100                              TCA_FLOWER_KEY_VLAN_PRIO, &key->vlan, &mask->vlan))
3101                 goto nla_put_failure;
3102
3103         if (fl_dump_key_vlan(skb, TCA_FLOWER_KEY_CVLAN_ID,
3104                              TCA_FLOWER_KEY_CVLAN_PRIO,
3105                              &key->cvlan, &mask->cvlan) ||
3106             (mask->cvlan.vlan_tpid &&
3107              nla_put_be16(skb, TCA_FLOWER_KEY_VLAN_ETH_TYPE,
3108                           key->cvlan.vlan_tpid)))
3109                 goto nla_put_failure;
3110
3111         if (mask->basic.n_proto) {
3112                 if (mask->cvlan.vlan_eth_type) {
3113                         if (nla_put_be16(skb, TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
3114                                          key->basic.n_proto))
3115                                 goto nla_put_failure;
3116                 } else if (mask->vlan.vlan_eth_type) {
3117                         if (nla_put_be16(skb, TCA_FLOWER_KEY_VLAN_ETH_TYPE,
3118                                          key->vlan.vlan_eth_type))
3119                                 goto nla_put_failure;
3120                 }
3121         }
3122
3123         if ((key->basic.n_proto == htons(ETH_P_IP) ||
3124              key->basic.n_proto == htons(ETH_P_IPV6)) &&
3125             (fl_dump_key_val(skb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
3126                             &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
3127                             sizeof(key->basic.ip_proto)) ||
3128             fl_dump_key_ip(skb, false, &key->ip, &mask->ip)))
3129                 goto nla_put_failure;
3130
3131         if (mask->pppoe.session_id) {
3132                 if (nla_put_be16(skb, TCA_FLOWER_KEY_PPPOE_SID,
3133                                  key->pppoe.session_id))
3134                         goto nla_put_failure;
3135         }
3136         if (mask->basic.n_proto && mask->pppoe.ppp_proto) {
3137                 if (nla_put_be16(skb, TCA_FLOWER_KEY_PPP_PROTO,
3138                                  key->pppoe.ppp_proto))
3139                         goto nla_put_failure;
3140         }
3141
3142         if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
3143             (fl_dump_key_val(skb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
3144                              &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
3145                              sizeof(key->ipv4.src)) ||
3146              fl_dump_key_val(skb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
3147                              &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
3148                              sizeof(key->ipv4.dst))))
3149                 goto nla_put_failure;
3150         else if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
3151                  (fl_dump_key_val(skb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
3152                                   &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
3153                                   sizeof(key->ipv6.src)) ||
3154                   fl_dump_key_val(skb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
3155                                   &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
3156                                   sizeof(key->ipv6.dst))))
3157                 goto nla_put_failure;
3158
3159         if (key->basic.ip_proto == IPPROTO_TCP &&
3160             (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
3161                              &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
3162                              sizeof(key->tp.src)) ||
3163              fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
3164                              &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
3165                              sizeof(key->tp.dst)) ||
3166              fl_dump_key_val(skb, &key->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS,
3167                              &mask->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS_MASK,
3168                              sizeof(key->tcp.flags))))
3169                 goto nla_put_failure;
3170         else if (key->basic.ip_proto == IPPROTO_UDP &&
3171                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
3172                                   &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
3173                                   sizeof(key->tp.src)) ||
3174                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
3175                                   &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
3176                                   sizeof(key->tp.dst))))
3177                 goto nla_put_failure;
3178         else if (key->basic.ip_proto == IPPROTO_SCTP &&
3179                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_SCTP_SRC,
3180                                   &mask->tp.src, TCA_FLOWER_KEY_SCTP_SRC_MASK,
3181                                   sizeof(key->tp.src)) ||
3182                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_SCTP_DST,
3183                                   &mask->tp.dst, TCA_FLOWER_KEY_SCTP_DST_MASK,
3184                                   sizeof(key->tp.dst))))
3185                 goto nla_put_failure;
3186         else if (key->basic.n_proto == htons(ETH_P_IP) &&
3187                  key->basic.ip_proto == IPPROTO_ICMP &&
3188                  (fl_dump_key_val(skb, &key->icmp.type,
3189                                   TCA_FLOWER_KEY_ICMPV4_TYPE, &mask->icmp.type,
3190                                   TCA_FLOWER_KEY_ICMPV4_TYPE_MASK,
3191                                   sizeof(key->icmp.type)) ||
3192                   fl_dump_key_val(skb, &key->icmp.code,
3193                                   TCA_FLOWER_KEY_ICMPV4_CODE, &mask->icmp.code,
3194                                   TCA_FLOWER_KEY_ICMPV4_CODE_MASK,
3195                                   sizeof(key->icmp.code))))
3196                 goto nla_put_failure;
3197         else if (key->basic.n_proto == htons(ETH_P_IPV6) &&
3198                  key->basic.ip_proto == IPPROTO_ICMPV6 &&
3199                  (fl_dump_key_val(skb, &key->icmp.type,
3200                                   TCA_FLOWER_KEY_ICMPV6_TYPE, &mask->icmp.type,
3201                                   TCA_FLOWER_KEY_ICMPV6_TYPE_MASK,
3202                                   sizeof(key->icmp.type)) ||
3203                   fl_dump_key_val(skb, &key->icmp.code,
3204                                   TCA_FLOWER_KEY_ICMPV6_CODE, &mask->icmp.code,
3205                                   TCA_FLOWER_KEY_ICMPV6_CODE_MASK,
3206                                   sizeof(key->icmp.code))))
3207                 goto nla_put_failure;
3208         else if ((key->basic.n_proto == htons(ETH_P_ARP) ||
3209                   key->basic.n_proto == htons(ETH_P_RARP)) &&
3210                  (fl_dump_key_val(skb, &key->arp.sip,
3211                                   TCA_FLOWER_KEY_ARP_SIP, &mask->arp.sip,
3212                                   TCA_FLOWER_KEY_ARP_SIP_MASK,
3213                                   sizeof(key->arp.sip)) ||
3214                   fl_dump_key_val(skb, &key->arp.tip,
3215                                   TCA_FLOWER_KEY_ARP_TIP, &mask->arp.tip,
3216                                   TCA_FLOWER_KEY_ARP_TIP_MASK,
3217                                   sizeof(key->arp.tip)) ||
3218                   fl_dump_key_val(skb, &key->arp.op,
3219                                   TCA_FLOWER_KEY_ARP_OP, &mask->arp.op,
3220                                   TCA_FLOWER_KEY_ARP_OP_MASK,
3221                                   sizeof(key->arp.op)) ||
3222                   fl_dump_key_val(skb, key->arp.sha, TCA_FLOWER_KEY_ARP_SHA,
3223                                   mask->arp.sha, TCA_FLOWER_KEY_ARP_SHA_MASK,
3224                                   sizeof(key->arp.sha)) ||
3225                   fl_dump_key_val(skb, key->arp.tha, TCA_FLOWER_KEY_ARP_THA,
3226                                   mask->arp.tha, TCA_FLOWER_KEY_ARP_THA_MASK,
3227                                   sizeof(key->arp.tha))))
3228                 goto nla_put_failure;
3229         else if (key->basic.ip_proto == IPPROTO_L2TP &&
3230                  fl_dump_key_val(skb, &key->l2tpv3.session_id,
3231                                  TCA_FLOWER_KEY_L2TPV3_SID,
3232                                  &mask->l2tpv3.session_id,
3233                                  TCA_FLOWER_UNSPEC,
3234                                  sizeof(key->l2tpv3.session_id)))
3235                 goto nla_put_failure;
3236
3237         if ((key->basic.ip_proto == IPPROTO_TCP ||
3238              key->basic.ip_proto == IPPROTO_UDP ||
3239              key->basic.ip_proto == IPPROTO_SCTP) &&
3240              fl_dump_key_port_range(skb, key, mask))
3241                 goto nla_put_failure;
3242
3243         if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
3244             (fl_dump_key_val(skb, &key->enc_ipv4.src,
3245                             TCA_FLOWER_KEY_ENC_IPV4_SRC, &mask->enc_ipv4.src,
3246                             TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
3247                             sizeof(key->enc_ipv4.src)) ||
3248              fl_dump_key_val(skb, &key->enc_ipv4.dst,
3249                              TCA_FLOWER_KEY_ENC_IPV4_DST, &mask->enc_ipv4.dst,
3250                              TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
3251                              sizeof(key->enc_ipv4.dst))))
3252                 goto nla_put_failure;
3253         else if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
3254                  (fl_dump_key_val(skb, &key->enc_ipv6.src,
3255                             TCA_FLOWER_KEY_ENC_IPV6_SRC, &mask->enc_ipv6.src,
3256                             TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
3257                             sizeof(key->enc_ipv6.src)) ||
3258                  fl_dump_key_val(skb, &key->enc_ipv6.dst,
3259                                  TCA_FLOWER_KEY_ENC_IPV6_DST,
3260                                  &mask->enc_ipv6.dst,
3261                                  TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
3262                             sizeof(key->enc_ipv6.dst))))
3263                 goto nla_put_failure;
3264
3265         if (fl_dump_key_val(skb, &key->enc_key_id, TCA_FLOWER_KEY_ENC_KEY_ID,
3266                             &mask->enc_key_id, TCA_FLOWER_UNSPEC,
3267                             sizeof(key->enc_key_id)) ||
3268             fl_dump_key_val(skb, &key->enc_tp.src,
3269                             TCA_FLOWER_KEY_ENC_UDP_SRC_PORT,
3270                             &mask->enc_tp.src,
3271                             TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK,
3272                             sizeof(key->enc_tp.src)) ||
3273             fl_dump_key_val(skb, &key->enc_tp.dst,
3274                             TCA_FLOWER_KEY_ENC_UDP_DST_PORT,
3275                             &mask->enc_tp.dst,
3276                             TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK,
3277                             sizeof(key->enc_tp.dst)) ||
3278             fl_dump_key_ip(skb, true, &key->enc_ip, &mask->enc_ip) ||
3279             fl_dump_key_enc_opt(skb, &key->enc_opts, &mask->enc_opts))
3280                 goto nla_put_failure;
3281
3282         if (fl_dump_key_ct(skb, &key->ct, &mask->ct))
3283                 goto nla_put_failure;
3284
3285         if (fl_dump_key_flags(skb, key->control.flags, mask->control.flags))
3286                 goto nla_put_failure;
3287
3288         if (fl_dump_key_val(skb, &key->hash.hash, TCA_FLOWER_KEY_HASH,
3289                              &mask->hash.hash, TCA_FLOWER_KEY_HASH_MASK,
3290                              sizeof(key->hash.hash)))
3291                 goto nla_put_failure;
3292
3293         return 0;
3294
3295 nla_put_failure:
3296         return -EMSGSIZE;
3297 }
3298
3299 static int fl_dump(struct net *net, struct tcf_proto *tp, void *fh,
3300                    struct sk_buff *skb, struct tcmsg *t, bool rtnl_held)
3301 {
3302         struct cls_fl_filter *f = fh;
3303         struct nlattr *nest;
3304         struct fl_flow_key *key, *mask;
3305         bool skip_hw;
3306
3307         if (!f)
3308                 return skb->len;
3309
3310         t->tcm_handle = f->handle;
3311
3312         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
3313         if (!nest)
3314                 goto nla_put_failure;
3315
3316         spin_lock(&tp->lock);
3317
3318         if (f->res.classid &&
3319             nla_put_u32(skb, TCA_FLOWER_CLASSID, f->res.classid))
3320                 goto nla_put_failure_locked;
3321
3322         key = &f->key;
3323         mask = &f->mask->key;
3324         skip_hw = tc_skip_hw(f->flags);
3325
3326         if (fl_dump_key(skb, net, key, mask))
3327                 goto nla_put_failure_locked;
3328
3329         if (f->flags && nla_put_u32(skb, TCA_FLOWER_FLAGS, f->flags))
3330                 goto nla_put_failure_locked;
3331
3332         spin_unlock(&tp->lock);
3333
3334         if (!skip_hw)
3335                 fl_hw_update_stats(tp, f, rtnl_held);
3336
3337         if (nla_put_u32(skb, TCA_FLOWER_IN_HW_COUNT, f->in_hw_count))
3338                 goto nla_put_failure;
3339
3340         if (tcf_exts_dump(skb, &f->exts))
3341                 goto nla_put_failure;
3342
3343         nla_nest_end(skb, nest);
3344
3345         if (tcf_exts_dump_stats(skb, &f->exts) < 0)
3346                 goto nla_put_failure;
3347
3348         return skb->len;
3349
3350 nla_put_failure_locked:
3351         spin_unlock(&tp->lock);
3352 nla_put_failure:
3353         nla_nest_cancel(skb, nest);
3354         return -1;
3355 }
3356
3357 static int fl_terse_dump(struct net *net, struct tcf_proto *tp, void *fh,
3358                          struct sk_buff *skb, struct tcmsg *t, bool rtnl_held)
3359 {
3360         struct cls_fl_filter *f = fh;
3361         struct nlattr *nest;
3362         bool skip_hw;
3363
3364         if (!f)
3365                 return skb->len;
3366
3367         t->tcm_handle = f->handle;
3368
3369         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
3370         if (!nest)
3371                 goto nla_put_failure;
3372
3373         spin_lock(&tp->lock);
3374
3375         skip_hw = tc_skip_hw(f->flags);
3376
3377         if (f->flags && nla_put_u32(skb, TCA_FLOWER_FLAGS, f->flags))
3378                 goto nla_put_failure_locked;
3379
3380         spin_unlock(&tp->lock);
3381
3382         if (!skip_hw)
3383                 fl_hw_update_stats(tp, f, rtnl_held);
3384
3385         if (tcf_exts_terse_dump(skb, &f->exts))
3386                 goto nla_put_failure;
3387
3388         nla_nest_end(skb, nest);
3389
3390         return skb->len;
3391
3392 nla_put_failure_locked:
3393         spin_unlock(&tp->lock);
3394 nla_put_failure:
3395         nla_nest_cancel(skb, nest);
3396         return -1;
3397 }
3398
3399 static int fl_tmplt_dump(struct sk_buff *skb, struct net *net, void *tmplt_priv)
3400 {
3401         struct fl_flow_tmplt *tmplt = tmplt_priv;
3402         struct fl_flow_key *key, *mask;
3403         struct nlattr *nest;
3404
3405         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
3406         if (!nest)
3407                 goto nla_put_failure;
3408
3409         key = &tmplt->dummy_key;
3410         mask = &tmplt->mask;
3411
3412         if (fl_dump_key(skb, net, key, mask))
3413                 goto nla_put_failure;
3414
3415         nla_nest_end(skb, nest);
3416
3417         return skb->len;
3418
3419 nla_put_failure:
3420         nla_nest_cancel(skb, nest);
3421         return -EMSGSIZE;
3422 }
3423
3424 static void fl_bind_class(void *fh, u32 classid, unsigned long cl, void *q,
3425                           unsigned long base)
3426 {
3427         struct cls_fl_filter *f = fh;
3428
3429         tc_cls_bind_class(classid, cl, q, &f->res, base);
3430 }
3431
3432 static bool fl_delete_empty(struct tcf_proto *tp)
3433 {
3434         struct cls_fl_head *head = fl_head_dereference(tp);
3435
3436         spin_lock(&tp->lock);
3437         tp->deleting = idr_is_empty(&head->handle_idr);
3438         spin_unlock(&tp->lock);
3439
3440         return tp->deleting;
3441 }
3442
3443 static struct tcf_proto_ops cls_fl_ops __read_mostly = {
3444         .kind           = "flower",
3445         .classify       = fl_classify,
3446         .init           = fl_init,
3447         .destroy        = fl_destroy,
3448         .get            = fl_get,
3449         .put            = fl_put,
3450         .change         = fl_change,
3451         .delete         = fl_delete,
3452         .delete_empty   = fl_delete_empty,
3453         .walk           = fl_walk,
3454         .reoffload      = fl_reoffload,
3455         .hw_add         = fl_hw_add,
3456         .hw_del         = fl_hw_del,
3457         .dump           = fl_dump,
3458         .terse_dump     = fl_terse_dump,
3459         .bind_class     = fl_bind_class,
3460         .tmplt_create   = fl_tmplt_create,
3461         .tmplt_destroy  = fl_tmplt_destroy,
3462         .tmplt_dump     = fl_tmplt_dump,
3463         .get_exts       = fl_get_exts,
3464         .owner          = THIS_MODULE,
3465         .flags          = TCF_PROTO_OPS_DOIT_UNLOCKED,
3466 };
3467
3468 static int __init cls_fl_init(void)
3469 {
3470         return register_tcf_proto_ops(&cls_fl_ops);
3471 }
3472
3473 static void __exit cls_fl_exit(void)
3474 {
3475         unregister_tcf_proto_ops(&cls_fl_ops);
3476 }
3477
3478 module_init(cls_fl_init);
3479 module_exit(cls_fl_exit);
3480
3481 MODULE_AUTHOR("Jiri Pirko <jiri@resnulli.us>");
3482 MODULE_DESCRIPTION("Flower classifier");
3483 MODULE_LICENSE("GPL v2");